The Gravitational Million-Body Problem

Douglas Heggie
Department of Mathematics and Statistics
University of Edinburgh, Edinburgh EH9 3JZ, Scotland

Piet Hut
School of Natural Sciences
Institute for Advanced Study, Princeton, NJ 08540, U.S.A.
For Eiko, Linda, Caroline and Alison
 Table of Contents

Preface iii

PART I INTRODUCTIONS 1
1 Astrophysics Introduction 3
2 Theoretical Physics Introduction 15
3 Computational Physics Introduction 25
4 Mathematical Introduction 35
PART II THE CONTINUUM LIMIT: N → ∞ 45
5 Paradoxical Thermodynamics 47
6 Statistical Mechanics 54
7 Motion in a Central Potential 63
8 Some Famous Models 74
9 Methods 84
PART III MEAN FIELD DYNAMICS: N = 106 98
10 Violent Relaxation 99
11 Internal Mass Loss 111
12 External Influences 119
PART IV MICROPHYSICS: N = 2 129
13 Exponential Orbit Instability 130
14 Two-Body Relaxation 138
15 From Kepler to Kustaanheimo 154

i
ii Table of Contents

PART V GRAVOTHERMODYNAMICS: N = 106 164

16 Escape and Mass Segregation 165
17 Gravothermal Instability 176
18 Core Collapse Rate for Star Clusters 185
PART VI GRAVITATIONAL SCATTERING: N = 3 193
19 Thought Experiments 195
20 Mathematical Three-Body Scattering 203
21 Analytical Approximations 215
22 Laboratory Experiments 229
23 Gravitational Burning and Transmutation 239
PART VII PRIMORDIAL BINARIES: N = 4 246
24 Binaries in Star Clusters 247
25 Triple Formation and Evolution 253
26 A Non-Renewable Energy Source 264
6
PART VIII POST-COLLAPSE EVOLUTION: N = 10 272
27 Surviving Core Collapse 273
28 Gravothermal Oscillations 283
29 Dissolution 294
PART IX STAR CLUSTER ECOLOGY 300
30 Stellar and Dynamical Evolution 302
31 Collisions and Capture 307
32 Binary Star Evolution and Blue Stragglers 315
33 Star Cluster Evolution 323
A Appendix – A Simple N-Body Integrator 331
B Appendix – Hints to solution of problems 341
References 354
Index 376
 Preface

Since this book is aimed at a broad audience within the physical sci-
ences, we expect most of our readers not to be experts in either astro-
physics or mathematics. For those readers, the title of this book may seem
puzzling at least. Why should they be interested in the gravitational at-
traction between bodies? What is so special about a millionindexmillion-
body problem, rather than a billion or a trillion bodies? What kind of
bodies do we have in mind? And finally, what is the problem with this
whole topic?
In physics, many complex systems can be modeled as an aggregate of
a large number of relatively simple entities with relatively simple interac-
tions between them. It is one of the most fascinating aspects of physics
that an enormous richness can be found in the collective phenomena that
emerge out of the interplay of the much simpler building blocks. Smoke
rings and turbulence in air, for example, are complex manifestations of
a system of air molecules with relatively simple interactions, strongly re-
pulsive at small scales and weakly attractive at larger scales. From the
spectrum of avalanches in sand piles to the instabilities in plasmas of
more than a million degrees in labs to study nuclear fusion, we deal with
one or a few constituents with simple prescribed forces. What is special
about gravitational interactions is the fact that gravity is the only force
that is mutually attractive. Unlike a handful of protons and electrons,
where like charges repel and opposite charges attract, a handful of stars
shows attraction between every pair of stars. As a result, a star cluster
holds itself together: there is no need for a container (as with a gas or
plasma in a lab) or a table (as with a sand pile). And for astronomers,
an extra reason to study star clusters is simply: because they are there.

iii
iv Preface

Now why are we interested in a million stars? Usually in physics we

analyze a system of interacting components in two limits: one in which
the number of components is one, or a few; and one in which the number
of components tends to infinity. And at first sight, a million and a billion
stars both seem to be ‘close enough’ to infinity to allow a common treat-
ment. However, there is a large difference in behavior between the two
types of systems. In a typical rich star cluster, with a million stars, each
star feels enough of the granularity of the gravitational field of the other
stars that the consequent perturbations lead to a total loss of memory
of the initial conditions of its orbit, within the life time of the Universe.
In contrast, in a typical galaxy, with a number of stars between a billion
and a trillion, an individual star for all intents and purposes feels only the
smooth average background field. This is the reason that galaxies still
retain much prettier shapes than the shapeless (indeed globular) clusters
for which orbital memories have been wiped out. In other words, in this
part of astronomy, when we count ‘one, two, a few, many’, a few means
a million.
The term body is just an archaic term for a material object, be it a
molecule or a stone or a star. In many cases in astronomy, when we
study the motion of a group of stars we can completely neglect the finite
size of the stars, and we may as well treat them as if they were simple
mass points, with a mass but without an extension. When stars come
closer, we may want to approximate them as finite-size bodies, with a
simple description of their mass distribution. Only when stars physically
collide, as they sometimes do in the dense inner regions of star clusters,
do we have to remind ourselves how complex these celestial bodies really
are∗ , in order to say something about the transformations that stars
undergo when they merge.
Finally, why do we call all this a problem? In mathematics, more
than in physics, we talk about specific topics as problems, as long as
they have not (yet) been solved. For example, the famous four-colour
problem, the question of how many different colours we need to introduce
to colour a map such that neighboring countries can always be given
different colours, remained a real problem until the seventies. Partly
with the help of computers, this problem was solved (the answer was:
four colours suffice), and then the four-body problem turned into the four-
colour theorem. In the days of Newton, at the beginning of mathematical
physics, the same term ‘problem’ was used to describe the challenge to
find the motions of two, three, or more bodies under the influence of their

∗
Addressing an astronomer who had just remarked that “after all, a star is a pretty
simple thing”, R.O. Redman pointed out, that, “at a distance of 10 parsecs you’d
look pretty simple!”
 Preface v

mutual gravitational attractions: hence the two-body problem, the three-

body problem, etc. Unlike the four-colour problem, though, we don’t
expect to ever ‘solve’ the gravitational N -body problem, for arbitrary
N – and even for N = 2, where we do have analytic solutions, we still
follow tradition in calling this the two-body problem. So apart from this
quaint piece of history, we could have called this book ‘the gravitational
million-body system’.

Why would someone want to study the gravitational million-body prob-

lem? There are at least four quite different motivations, centering on
the fields of astronomy, theoretical physics, computational physics, and
mathematics.
Let us first take the point of view of astronomy. Throughout the last
hundred years or more, astronomers have been studying an important
class of objects called globular star clusters, which are roughly spherical
collections of stars, each much smaller than a galaxy but much bigger than
the solar system. The number of stars they contain varies from one object
to another, but a million is the right order of magnitude. Astronomers
study globular clusters for many reasons, of which we can touch on only a
few. It is thought that they were among the first recognisable structures
that were born in galaxies like ours, and their age is a vital constraint on
that of the universe. All the stars in each cluster were born with different
masses at roughly the same time, and their present stellar population
gives a snapshot of the results of about 10 billion years of stellar evolution.
They are exceptionally rich in some of the more exotic kinds of star which
astronomers now study: binary X-ray sources, millisecond pulsars, etc.
Finally, the stars inside a cluster are sufficiently densely packed that they
can and do collide with each other, and so clusters provide us with a
sort of laboratory where we may hope to understand the more dramatic
effects of the dense stellar environment in the nuclei of galaxies. Thus
the gravitational million-body problem has an important place in modern
astronomy.
Now let us consider the point of view of theoretical physics. Whichever
way a physicist approaches them, gravitational problems pose particular
difficulties. Classical thermodynamics is poorly developed for such prob-
lems, because of the long-range nature of the gravitational force. Kinetic
theory requires ad hoc approximations, for the same reasons. The meth-
ods of plasma physics fail because gravitational forces, being attractive,
are unshielded. Gravitational systems are intractable with many tradi-
tional methods because the natural equilibria are not spatially uniform.
The gravitational many-body problem also has a wider significance in
physics, because of its historical roots. The science of mechanics be-
came firmly rooted, in large part, because of the success of Newton’s
vi Preface

programme for the study of planetary motion. The further development

of this theory led to the development of Lagrangian methods, and then,
in the hands of Poincaré, to the foundation of the study of chaos. Much
of the development of physics since the time of Newton has rested on
foundations set in place with the aid of the gravitational many-body
problem, which became the model for how a successful physical theory
should look. Finally, the most difficult models to predict and to interpret
their behavior are the ones where we are dealing with neither a very large
nor a very small number of particles. When there are huge numbers of
particles, as in a gas, there are successful approximate methods, like fluid
mechanics, which greatly simplify the problems. When there are very few
particles, especially one or two, the problem is either completely soluble
or can be understood approximately. When there are an intermediate
number (a million or less in our case), elements of both extremes are in
play simultaneously, thereby thwarting either type of approximation.
This brings us naturally to computational physics. Just as the theoret-
ical physicist does, the computational physicist immediately grasps the
fact that the gravitational million-body problem poses formidable prob-
lems whichever way it is approached. It is not even clear whether the
solution can be obtained reliably at all. Even if we lay such fundamental
issues aside, running an N -body simulation with N > 4
∼ 10 is exceptionally
time-consuming for even the fastest generally available computers. This
in turn has led to the development of special computers (the GRAPE
family), whose sole task is to solve this problem very quickly. Around
1995 these were the fastest computers in the world, a place they recap-
tured in 2001. This approach has become a model of the way in which
future work in computational science in other areas may be carried out
economically and quickly, given the right flair and ingenuity. Providing
the hardware is only half the problem though; developing the software
for running N -body simulations is equally challenging. In our particu-
lar case of globular cluster simulations, we are confronted with length
scales spanning the range from kilometers to parsecs (a factor of more
than 1013 ), and with time scales spanning the range from milliseconds to
the life time of the universe (a factor of more than 1020 ). As a result,
stellar dynamicists have developed special integration methods that are
not encountered in any text books on differential equations.
And finally we turn to mathematics. When Newton laid the founda-
tions of classical physics in the Principia, among the mathematical tools
he deployed was the infinitesimal calculus, which he had invented for
the purpose. This illustrates the rich potential for the invention of new
mathematics which results from intense scrutiny of gravitational prob-
lems, and other issues in dynamics. Newton set a precedent which has
repeated itself several times since. The work of Poincaré was developed
 Preface vii

in the context of celestial mechanics, and the famous theorem of Kol-

mogorov, Arnold and Moser, foreshadowed as it was by the work of C.L.
Siegel in his book “Vorlesungen über Himmelsmechanik”∗ , owes much to
this discipline, and the emphasis which it helped to place on Hamilto-
nian problems. And yet when one turns to the million-body problem, as
opposed to the few-body problems of the solar system, the flow of infor-
mation has been from mathematics to stellar dynamics, rather than the
other way around. A number of techniques introduced by mathemati-
cians for the solution of quite abstract problems have turned out to be
just what was needed to improve the numerical solution of our problem.
This flow of ideas has been sufficiently influential that the astronomer’s
understanding of the problem would not have advanced so far without it.
One of our purposes in writing this book, perhaps a far-fetched one, is
the hope that it may help reverse the flow, and stimulate the birth of new
mathematics. At the very least, we hope that mathematicians will enjoy
learning how their work has been put to use. As one of us has enjoyed a
place in the tolerant community of mathematicians for many years, it is
a way of giving something back.

We believe that the gravitational million-body problem has a seminal

role to play in all four areas: astronomy, physics, mathematics, computa-
tional science. But while the above remarks have sometimes emphasised
the difficulties it poses, we will have failed in one of our aims if, after read-
ing the book, the reader from any of these disciplines is not impressed
by how much of the problem can be understood, and how much it has to
offer. In other words, one of our goals is to convey some of the beauty and
simplicity underlying classical dynamics, as illustrated through the grav-
itational many-body problem. This is a book that is meant to be read
by a variety of scientists who share a curiosity for the roots of physics,
the recent fruits that have sprouted directly from those old roots, and its
interconnections with neighbouring sciences.
That having been said, it is mostly within the astronomical community
that this subject has been developed. And it has been another firm inten-
tion in writing the book that it may serve the role of a graduate textbook
on the theory of stellar dynamics of dense stellar systems. Wherever
possible, therefore, our statements and results are developed from first
principles. We include some exercises and problems, and hints for their
solution.
In writing the book we were aware that the reader may or may not be
an astronomer, and may or may not be interested in learning the details of

∗
Lectures on Celestial Mechanics
viii Preface

the theory. We realised also that the variety of topics which make up the
gravitational million-body problem are best treated at a variety of levels,
and that the interconnectedness of these topics forces interconnections
between the chapters. Some of the topics we address might even strike
some readers as being faintly whimsical. Recognising, therefore, that not
all chapters will be equally accessible or interesting to all readers, we
start each part of the book (each of which consists of several connected
chapters) with an outline of its contents. We hope that readers will thus
find their own optimal route between the contents page and the index.
For much the same variety of reasons we have relegated some material
to boxes. Often these contain details of some derivation or discussion, and
a reader in a hurry could avoid them. Sometimes, however, they contain
background which might help a reader over some difficulty. Sometimes
they collect some useful results which might act as a reference resource.
Sometimes they explore interesting or amusing by-ways that would oth-
erwise interrupt the text. Perhaps, therefore, they should not be passed
over too readily.
We hope the index will both help and intrigue the reader, but the
quirky and personal selection of names there is not meant as a compre-
hensive answer to the question of who did what. (Last time we looked, we
could not even find our own names there.) Even in the extensive list of
references to published work we have not attempted completeness. While
this risks antagonizing the reader for the omission of his or her own most
cherished paper, for which we apologise, it could be just as hazardous
if we tried to attribute every advance to its originator. Our main rea-
son for giving references is to give the reader an entry into the research
literature. Nowadays it is quite easy to move both forwards and back-
wards from a suitable entry point, using such invaluable bibliographical
resources as the Astronomy Abstract Service of the NASA Astrophysics
Data System.

And now we have a confession to make. Studying problems in the-

oretical physics, the solved as well as the unsolved, is interesting and
enjoyable all by itself. Finding new and unexpected insights, and extend-
ing the world’s body of knowledge about these problems, is an added
pleasure. Sharing these interests with others is still more rewarding. But
we have other reasons for our interest in the gravitational N -body prob-
lem, and among these is its status as one of the oldest unsolved problems
in the exact sciences, and one with an exceptionally distinguished pedi-
gree. Some of the great names of the subject have been mentioned in this
preface, but we would add some other much admired names of the past,
including those of Jeans, Chandrasekhar and Spitzer. With equal plea-
sure we call to mind the many contemporaries with whom we have worked
 Acknowledgements ix

and continue to work on these problems. The community is world-wide,

but it is not a large one, it is rather close-knit, and works together and
openly in the best tradition of the “community of scholars”. We hope
that the common enjoyment of this joint enterprise surfaces from time to
time throughout this book.

Acknowledgements

We have a number of these colleagues to thank with particular regard

to the help they have given with parts of the book. Large sections have
been read and commented on by Haldan Cohn, Herwig Dejonghe, Mirek
Giersz, Jeremy Goodman, Jarrod Hurley, Jun Makino, Simon Portegies
Zwart, Rainer Spurzem and Peter Teuben, and we thank them warmly
for their comments. If there are problems with the book, it is probably
because we did not always take their advice. Simon Portegies Zwart and
Chris Eilbeck also gave considerable help with appendix A, and we also
thank Maurizio Salaris and Max Ruffert for their input. For permission
to reproduce illustrations we would like to thank O. Gnedin, J. Good-
man, L. Hernquist, K. Johnston, S. Kulkarni, C. Lada, J. Makino, S.
McMillan, J. McVean, E. Milone, F. Rasio, the American Astronomical
Society, Blackwell Publishing (publishers of Monthly Notices of the Royal
Astronomical Society), and other individuals and institutions named be-
side individual figures. We are also most grateful to staff at Cambridge
University Press, especially A. Black (who has since moved on), J. Gar-
get and S. Mitton, for their advice, encouragement and patience. DCH
is deeply grateful to the Institute for Advanced Study, Princeton, for its
generous hospitality over several visits during which much of the planning
and drafting of the book was carried out.

Edinburgh and Princeton

2002
 PART I. Introduction

Newton’s equations for the gravitational N -body problem are the start-
ing point for all four chapters in Part I, but each time seen in a different
light. To the astrophysicist (Chapter 1) they represent an accurate model
for the dynamical aspects of systems of stars, which is the subject known
as stellar dynamics. We distinguish this from celestial mechanics, and
sketch the distinction between the two main flavours of stellar dynamics.
This book is largely devoted to what is often (but maybe misleadingly)
called collisional stellar dynamics. This does not refer to actual physical
collisions, though these can happen, but to the dominant role of gravita-
tional encounters of pairs of stars. In dense stellar systems their role is
a major one. In collisionless stellar dynamics, by contrast, motions are
dominated by the average gravitational force exerted by great numbers
of stars.
We lay particular emphasis on the stellar systems known as globular
star clusters. We survey the gross features of their dynamics, and also
the reasons for their importance within the wider field of astrophysics.
Though understanding the million-body problem is not among the most
urgent problems in astrophysics, through globular clusters it has close
connections with several areas which are. Another practical topic we
deal with here is that of units, which may be elementary, but is one area
where the numbers can easily get out of hand.
Chapter 2 looks at the N -body equations from the point of view of
theoretical physicists. To them, the gravitational many-body problem is
just one of many interesting N -body problems. In some ways it is simple
(being scale-free), and in other ways complicated (long and short ranges
have comparable importance, for example, unlike in plasmas). Theoreti-
cal physicists understand the two-body problem, just like astrophysicists
do, but for large N their usual methods don’t apply very well. Statistical

1
2 PART I. Introduction

mechanics is one casualty. The study of limiting cases is stymied by the

fact that there are no parameters to alter, except N .
The enormous range of length scales (and, consequently, time scales) is
also one reason why the N -body equations are a severe challenge to the
computational physicist (Chapter 3). There are good reasons why the
faster methods used in other (e.g. collisionless) problems are deprecated
in this field. But the penalty is that computational progress has been
slow: we are only now approaching the capability of directly modelling
even a small globular cluster. Though software advances are continually
playing a role, what has made this possible is the development of special-
purpose computers tailored precisely for this problem.
Mathematicians (Chapter 4) seek a qualitative understanding of the
N -body equations. They are concerned with the non-obvious question
of what one even means by a “solution”. They pay particular attention
to the (collision) singularities of the equations, and ask how solutions
behave in their vicinity, even if they occur only at complex values of the
time. The mathematician asks whether the equations can exhibit singu-
larities of other kinds. Historically, the fertile equations of the three-body
problem have bred a variety of topics which remain mainstream issues in
mathematics, foremost among them being questions of integrability.
 1
Astrophysics Introduction

When Newton studied dynamics and calculus he was motivated, at least

in part, by a desire to understand the movements of the planets. The
mathematical model which he reached may be described by the equations

X
j=N
ri − rj
r̈i = −G mj , (1)
j=1,j6=i
|ri − rj |3

where rj is the position vector of the jth body at time t, mj is its mass,
G is a constant, and a dot denotes differentiation with respect to t. Now,
over 300 years later, our motivation has a very different emphasis. The
study of the solar system has lost none of its importance or fascination,
but it now competes with the study of star clusters, galaxies, and even
the structure of the universe as a whole. What is remarkable is that
Newton’s model is as central to this extended field of study as it was in
his own time.

Stellar Dynamics

It may seem surprising that the single, simple set of equations (1) forms a
good first approximation for modeling many astrophysical systems, such
as the solar system, star clusters, whole galaxies as well as clusters of
galaxies (Fig.1). The reason is that gravity, being an attractive long-
range force, dominates everything else in the Universe. The only other
long-range force, electromagnetism, is generally not important on very
large scales, since positive and negative charges tend to screen each other.
Short-range forces, such as gas pressure, are usually only important on
small scales, such as in the interiors of stars. On large scales, comparable
to the size of a galaxy, pressure is rarely important. On intermediate

3
4 1 Astrophysics Introduction

scales we have giant gas clouds in which both pressure and magnetic
fields can play a role.
This dominance of gravity on cosmic scales is a fortunate feature of
our Universe. It implies that it is relatively simple to perform detailed
computer simulations of many astronomical systems. Mastery of the
much more complicated physics of other specializations in astronomy,
such as plasma astrophysics, radiative transfer, or nuclear astrophysics,
is often not immediately necessary. Nor, in applications to star clusters,
is it usually necessary to use the more refined theory of general relativity.
The velocities are too low, except in a few situations involving degenerate
stars (neutron stars). Even the motions of stars round the black hole at
the centre of our Galaxy (Ghez et al. 2000), though it takes place on a
human time scale, can be understood with classical dynamics.
Stellar dynamics can be defined as studying the consequences of Eq.
(1) in astrophysical contexts. Traditionally these equations were discov-
ered by studying the motions of the moon and planets, and for the next
few centuries they were applied mainly to planetary dynamics. Before
the advent of electronic computers, most effort went into developing an-
alytical approximations to the nearly regular motion of the planets. This
field, known as celestial mechanics, had an important influence on devel-
opments both in physics and mathematics. An example is the study of
chaos, which first arose as an annoying complexity barring attempts to
make long-time predictions in celestial mechanics.
The solar system, however, is a very regular system. All planets move
in orbits close to the ecliptic, and all revolve in the same direction. The or-
bits are well-separated, and consequently no close encounters take place.
When we look around us on larger scales, that of star clusters and galax-
ies, then no such regularity applies. In our galaxy, most stars move in
the galactic disk, revolving in the same sense as the sun, but close en-
counters are not excluded. In many globular star clusters, there is not
even a preferred direction of rotation, and all stars move as they please
in any direction. Does this mean that analytic approximations are not
very useful for such systems?
The answer is: “it depends on what you would like to know”. In
general, the less regular a situation is, the larger the need for a computer
simulation. For example, during the collision of two galaxies each star in
each of the galaxies is so strongly perturbed that it becomes very difficult
to predict the overall outcome with pen and paper. This is indeed an
area of research which had to wait for computers to even get started,
in the early seventies (Toomre & Toomre 1972; but see the reference to
Holmberg in Chapter 3). On the other hand, when we want to understand
 Stellar Dynamics 5

Fig. 1. Two extreme astrophysical many–body problems. Top A binary star,

ζ 1 Ursae Majoris, imaged interferometrically (Credit: J. Benson et al., NPOI
Group, USNO, NRL) Bottom The galaxy M87. This is not a straightforward
N -body problem, as much of the mass is not stellar. Many of the spots of light in
the fringes of the galaxy are individual globular clusters (Credit: NOAO/AURA/
NSF )
6 1 Astrophysics Introduction

the conditions in a relatively isolated galaxy, such as our own Milky Way,
then there is some scope for pen-and-paper work. For example, a rich
variety of analytic as well as semi-numerical models has been constructed
for a range of galaxies. However, even in this case we often have to switch
to numerical simulations if we want to obtain more precise results.

Two Flavours of Stellar Dynamics

The field of stellar dynamics can be divided into two subfields, tradi-
tionally called collisional and collisionless stellar dynamics. The word
‘collision’ is a bit misleading here, since it is used to describe a close en-
counter between two or more stars, not a physical collision. After all,
physical collisions are excluded in principle as soon as we have made the
approximation of point particles, as we will do throughout most of the
present book, with the exception of the last few chapters, where we will
take up the question of real stellar traffic accidents.
Collisional stellar dynamics is concerned about the long-term effects of
close (as well as not-so-close) stellar encounters. The evolution of a star
cluster is governed by the slow diffusion of “heat” through the system
from the inside towards the edge. The heat transport occurs through
the frequent interactions of pairs of stars, in a way similar to the heat
conduction in the air in a room, which is caused by collisions between
pairs of gas molecules. The main difference here is that individual stars
in a star cluster have mean free paths that are much longer than the size
of the system. In other words, little heat exchange takes place during a
system crossing time scale.
Collisionless stellar dynamics is the subfield of stellar dynamics in which
the heat flow due to pairwise interactions of stars is neglected. For small
systems this approximation is appropriate when we consider the evolution
of the star system on a time scale which does not exceed the crossing
time by a very large amount. For a system with very many particles,
such as a galaxy, the collisionless approximation is generally valid even
on timescales comparable to the age of the Universe (which is comparable
to the age of most galaxies).

Life in a Collisional Stellar System

So far, we have talked about the astrophysical many-body problem in a
rather generalised way. We have dealt with stars or planets moving in the
gravitational field of their mutually attractive forces, and have ignored
their internal structure, but we have not focused on any particular kind of
system to which these idealisations apply. Fortunately, there are systems
in nature which approach the idealizations of collisional stellar dynamics
to a remarkable degree. These are the globular clusters, among the oldest
 Life in a Collisional Stellar System 7

components of our galactic system, and going their own way in wide orbits
around the galaxy (cf. Webbink 1988). They are nearly spherical because
they don’t rotate much by astronomical standards (cf. Merritt et al. 1997,
Davoust & Prugniel 1990). They are largely isolated from perturbing
influences of the galactic disk, and therefore form ideal laboratories for
stellar dynamics. Typically they have about a million stars each (Table
1), with a range of individual stellar masses (e.g. Paresce & De Marchi
2000).
Other galaxies also have globular clusters (Fig.1, bottom; Ashman &
Zepf 1998). In some galaxies they are not the oldest but some of the
youngest stellar components (see Lançon & Boily 2000). For instance the
cluster NGC 1866, which lies in the Large Magellanic Cloud, has a mass
of about 105 M but is only about 108 yrs old (Fischer et al. 1992, Testa
et al. 1999). Even in our own galaxy, the most massive young clusters are
found in the apparently hostile environment of the Galactic Centre (Figer
et al. 1999), and they are as massive as a modest old globular cluster
(though they will certainly not last as long, cf. Kim et al. 2000, Portegies
Zwart et al. 2001). The realisation that the formation of globular clusters
is still going on in the universe around us has taken hold only relatively
recently, and has helped to rejuvinate the study of these stellar systems.
For the most part, though, we concentrate on the old and better observed
systems in our own Galaxy.

Table 1. Basic Facts about the Globular Clusters of the Galaxy

Number known 147

Median distance from Galactic Centre 9.3kpc
Median absolute V magnitude -7.27
Median concentration 1.50
Median core relaxation time 3.39 × 108 yr
Median relaxation time at the half-mass radius 1.17 × 109 yr
Median core radius 1.32pc
Median half-mass radius 3.08pc
Median tidal radius 34.5pc
Median mass 8.1 × 104 M
Median line-of-sight velocity dispersion 5.50km/s
The data are based on tables in Harris (1996; for updated online ver-
sions see http://physun.physics.mcmaster.ca/Globular.html), Mandu-
shev et al. 1991 and Pryor & Meylan (1993), and different data are
based on somewhat different samples. The terms concentration, re-
laxation time, and core and tidal radius are defined elsewhere in the
book. Units are discussed in Box 1.
8 1 Astrophysics Introduction

Box 1. Units
Consider a stellar system whose total mass and energy are M and E,
respectively. Then we may choose M for the unit of mass, and in simple
GM 2
cases a common and useful unit of length is the virial radius Rv = − .
4E
(For example, for a binary star this is roughly the diameter of the relative
orbit.) Then, if we want units in which the constant of gravitation is unity,
3/2 √
the unit of time is required to be Ut = Rv / GM . Another √ common
unit of time is the crossing time, which is defined to be 2 2Ut .
Now in the evaluation of these expressions Rv and M will often be in
units favoured by astrophysicists (e.g. parsecs and solar masses, respec-
tively), while G is most familiar in one of the standard physical systems,
such as SI. The evaluation of Ut then requires conversion to a common
system of units. At this point, because of the enormous distances and
time scales which prevail in astronomy, the formulae bristle with enor-
mous exponents, until the final result is re-expressed in astrophysically
sensible units, and the numbers become reasonably once again. Indeed
it is one of the minor pleasures and surprises of a theorist’s life, and a
clue that he is on the right lines, when sensible numbers emerge at the
end of a long and tiresome calculation studded with huge powers of ten.
Sir Harold Jeffreys once remarked that “incorrect geophysical hypotheses
usually fail by extremely large margins” (Jeffreys 1929). The same is true
in astrophysics.
To avoid this unpleasantness it is much simpler to calculate entirely
within the astrophysical system, provided that you know the value of G
in these units. The following remarks will therefore save much time. Con-
sider units such that masses are measured in M , speeds are measured in
km/sec, and distances are measured in parsecs. Then G = 1/232 approx-
imately. Also, the corresponding unit of time is 106 years approximately.
In integrating the N -body equations numerically, still other units are
used. Part of the reason for this is to avoid carrying around the values
of physical constants, such as G, that would be needed in any physical
system, partly to avoid the huge and unreadable numbers that would
result from an inappropriate choice of physical units, but mainly so as not
to obscure the intrinsic scalability of many simple N -body calculations.
In simulations it is natural to choose units such that G = 1, and,
following Hénon (1971), the units of length and mass are conventionally
chosen so that the total mass of the system is M = 1 and the total
energy is E = −1/4. These choices might seem bizarre, but it then
follows that, if the system is in virial equilibrium (chapter 8), the potential
energy is W = −1/2, and the virial radius is then Rvir = 1. These units
are referred to as N -body units, and are in widespread, but not quite
universal, use. They become rather ambiguous when external fields or
many binary stars are included in the computations.
 A Recipe for a Star Cluster 9

In order to get a feel for the physical presence of a globular cluster,

imagine that we would live in the very core of a dense globular (Fig.2).
The density of stars there can easily be as high as 106 M /pc3 . This is
a factor 106 higher than in our own neighbourhood, in the part of the
galactic disk where the Sun happens to reside. Therefore, we can get an
impression of the night sky by bringing each star that we normally see at
night closer to us by a factor 102 . Each star would thus become brighter
by a factor 104 , which corresponds to a difference of ten magnitudes (in
the astronomical system where a factor ten corresponds to ∆m = 2.5).
The brightest stars would thus appear at magnitudes at or above m ∼
−10, comparable to that of the full moon. They would be too bright to
look at directly, because their size would be so much smaller than that
of the moon (they would still look point-like). It would be easy to read
books by the light of the night sky. This is what the huge stellar density
in these systems means from the “human” point of view. Later on we
shall consider what effect it has on the stars.

A Recipe for a Star Cluster

Looking at Fig.2, we cannot immediately tell how such an assembly of
stars has got there, or what its fate will be. One might even guess, as
Eddington (1926a) once did, that the stars cannot “escape the fate of
ultimately condensing into one confused mass”. Nowadays, though many
details are unclear, we would argue that the clusters have survived for
almost as long as the universe itself, and that they are only now slowly

Fig. 2. The central region of the globular cluster M15. Image by courtesy of P.
Guhathakurta, UC Santa Cruz.
10 1 Astrophysics Introduction

dying off as their stars escape. Here we paint a few details on this picture,
showing what we think are the stages through which a cluster would have
to pass in order to arrive at something resembling those we see around
us.
To make a star cluster, in practice, one should start out with a large
gas cloud, which under the right conditions undergoes internal collapse on
several length and time scales, to produce a large number of stars. This is
a process which is being modelled with increasing sophistication (see, for
example, the beautiful movie at http://www.ukaff.ac.uk/movies.shtml, or
Klessen & Burkert 2001). Nevertheless the physical processes related to
star formation are very complex, and at present only partly understood.
Therefore, let us perform a thought experiment in which we will cheat
a bit. Let us take a bucket full of ready-made single stars, and sprinkle
them into a limited region of space, in order to watch just the stellar
dynamics between the stars at play.
The simplest way to place the stars in space, is to start from rest. As
soon as the stars are allowed to start moving, the whole system will begin
to collapse under the mutual gravitational attraction of all the stars. If
the stars are sprinkled in at random, the contraction will not proceed
very far, since the individual star-star interactions will cause deflections
from a purely radial infall. The stars will pass at some distance from the
centre, each one at a somewhat different time, and then move out again.
As a consequence, the whole system will breath a few times: globally
shrinking and expanding. However, phase coherence will be lost very
soon, and after a few breathing motions, the stars will be pretty well
mixed. Some stars will be lost, spilled into the surrounding space, never
to return, but most stars will remain bound to the system.
After those first few crossing times, not much will happen, for quite a
while. During a typical crossing time, a typical star will move through
the system on a rosette-shaped orbit, while almost conserving its energy
and angular momentum. On a longer time scale, though, the cumulative
effect of many distant encounters, and occasional closer encounters, will
affect the orbit. If we wait long enough, the memory of the original energy
and angular momentum of the star will be lost. This time scale is called
the relaxation time scale.
Even if the system was set off in a non-spherical distribution, such
as that of a pancake, say, the original order will be erased after a few
relaxation time scales, and the system will become spherical. There will
also be a steady trickle of stars that escape from the system. In addition,
energy will be transported from the centre to the outer regions, and as
a result the inner regions tend to contract. Before long, the centre will
gain in density, by many orders of magnitude, a phenomenon called core
collapse.
 Dynamics of Dense Stellar Systems 11

After core collapse, double stars will be formed, that provide a source
of energy for the system, empowering the ongoing evaporation of the
system, until the whole star cluster is dissolved. While the star cluster
slowly evaporates, the system will undergo core oscillations, with vast
swings in central density, if the number of stars is large enough, well
over 104 . These so-called gravothermal oscillations are the result of the
fact that the inner regions, having much shorter relaxation times than
the rest of the system, grow impatient in the post-collapse phase, and
start collapsing all over again. Interestingly, these oscillations exhibit
mathematical chaos, one more example of the gravitational many-body
problem providing a stage for some of the most modern forms of applied
mathematics.
This general picture, of a star cluster being formed, coming to rest,
undergoing a slow form of core ‘collapse’, followed by evaporation, will
be spun out in much greater detail throughout the rest of the book. But
before we take a closer look to the specific million-body problem, let us
see how much the picture can change if we treat the stars as stars and
not as point masses.

Dynamics of Dense Stellar Systems

Most stars do not interact much with their environment, after they leave
the cradle of the interstellar cloud they are born in. Some stars are born
single, and stay that way throughout their life, although they may have
acquired a planetary system during the late stages of their formation.
However, most stars are member of a double star or an even more complex
multiple system (triples, quadruples, etc.). In such a system, when two
stars are sufficiently close, all kind of interesting interactions may take
place including the transfer of mass from one star to another, and even
the spiral-in and eventual merging of two or more stars.
Although binary-star evolution is much more complex than single star
evolution, it can generally still be studied in isolation from its wider
environment. After all, the typical separation between stars in the solar
neighbourhood (itself typical for our galaxy) is some hundred million
times larger than the diameters of individual stars. The exception to this
rule occurs in unusually dense stellar systems. Examples are star clusters,
both in the disk of the galaxy as well as outside (the globular clusters),
and the nuclei of galaxies.
Especially in the cores of globular clusters and in the nuclei of galaxies,
the densities of stars can easily exceed that of the solar neighbourhood by
a factor of a million or more (Fig.2). At any given time, such a system is
still dilute enough to make physical collisions unlikely, even during many
crossing times, thus allowing point mass dynamics to provide useful first-
12 1 Astrophysics Introduction

order approximations. However, when viewed over a time scale of billions

of years, such collisions become unavoidable in many of these systems.
In recent years much progress has been made in the study of physical
collisions, both theoretically in computer simulations, and observationally
by looking for ‘star wrecks’ as tell-tale signs of violent encounters. For
example, observations with the Hubble Space Telescope have shown us
the presence of so-called blue stragglers, right down in the centre of the
most crowded star clusters. Blue stragglers are unusual types of stars
that are at least compatible with being the products of stellar collision.
Earlier, millisecond pulsars and X-ray binaries already have hinted to
us more indirectly about the sagas of their formation and subsequent
interactions.
In order to interpret this wealth of observational information, vigorous
attempts are being made by several groups to make theoretical models
for the evolution of dense stellar systems. The first step is to determine
the long-time behavior of a large system of point masses (a million or
so), a classical problem that is still far from solved, and that has given
rise to fascinating new insights, even over the last ten years, including
the surprising discovery of the presence of mathematical chaos in the late
stages of its evolution. It is this step that is the main theme for the
present book.
The second step is to integrate our understanding of the dynamics of
a system of point masses with the extra complexity introduced by the
non-point-behavior, in the form of stellar evolution, physical collisions
between stars, mass loss, etc. This integration gives rise to an extremely
complex picture of the ecology of star clusters, as we will briefly discuss
in the last part of this book.

Why study the dynamics of globular clusters?

The dynamical study of globular clusters has had a peculiar history,
which stretches over the entire 20th century and beyond. Even up to the
1960s, however, many astrophysicist regarded it as a quiet backwater,
even though it had attracted the attention of the best theorists of each
generation: Jeans, Eddington, Chandrasekhar, Spitzer, Hénon, Lynden-
Bell, Ostriker,... Perhaps they were initially attracted to it because of
what Lyman Spitzer called its “appealing but deceptive simplicity”.
This dynamical study of the clusters went in parallel with the astro-
physical study of the stars inside them, as these were seen to be excellent
test beds for checking the theory of stellar evolution, the chemical history
of the Galaxy, the age of the universe, and so on. In fact there was hardly
any major problem of astronomy – from star formation to cosmology –
 Why study the dynamics of globular clusters? 13

in which globular star clusters did not have something important to say,
and this remains true to the present day.
Until the early 1970s these parallel tracks in the study of globular
clusters – the dynamical and the astrophysical – proceeded entirely in-
dependently. The two communities had nothing in common and did not
even need to talk to each other. By the 1990s all this had changed. What
happened?
More than anything else, it was the discovery of variable X-ray sources
in globular clusters in the 1970s which brought down the barriers be-
tween the two communities (Fig.3). Though their origin is still not quite
clear, it was soon realised that dynamical processes were important. Dy-
namicists also realised that here was a new process which could have a
profound influence on their understanding of dynamical evolution. Sud-
denly the dynamics of globular clusters had rejoined the mainstream of
astrophysics, and in the 1970s there was not much that seemed more
glamorous than X-ray astronomy.
The integration of the dynamical and stellar studies of globular clusters
has progressed steadily in the intervening decades, and there is even a new
breed of “cross-over” astrophysicists who are at home in both camps. This
grand unification of cluster studies has been strengthened by a stream of
exciting discoveries: binary stars (through the rapid explosion of data on
radial velocities, and other techniques), millisecond pulsars, white dwarfs,
blue stragglers in dense clusters, core collapse, and so on. They are ideal

Fig. 3. A recent X-ray image of the Andromeda galaxy, M31. More than 5% of
the sources lie in globular clusters belonging to Andromeda. Original colour im-
age by courtesy of the ROSAT Mission (http://www.xray.mpe.mpg.de/) and the
Max-Planck-Institut für extraterrestrische Physik (http://www.mpe.mpg.de/)
14 1 Astrophysics Introduction

targets for searching for extrasolar planets, or should be (Gilliland et al.

2000).
The sheer visual beauty of globular clusters, and the richness of the
problems they present, must not lead us to lose sight of their place in
the bigger picture. In particular, they allow us to study the behaviour of
stars in dense environments, and will prove to be an important stepping
stone for the understanding of the even denser stellar systems in galactic
nuclei – surely one of the major problems of astrophysics for the 21st
century. Nor are we here restricting attention to active galactic nuclei.
For instance the nucleus of the quiescent nearby galaxy M33 can really be
regarded as a gigantic globular cluster sitting at the centre of its parent
galaxy, and may well have exhibited the kind of core collapse studied in
the context of globular clusters (Hernquist et al. 1991).
Sidney van den Bergh (1980) once expressed this philosophy very
nicely: “One of the dangers of living on a beautiful isle, such as Van-
couver Island, is that it is very easy to become insular. By the same
token it is all too easy for us, working in various exciting areas of cluster
studies, to forget about the broader impact that such work might have
on other areas of astronomy. The justification of our work on clusters
is not just that it is fun but also that such investigations often help to
illuminate other branches of science.”

Problems
1) Compute the virial radius (Box 1) of a binary consisting of two stars
of mass m1 and m2 in a circular relative orbit of radius a.
2) Compute the speed of a star in a circular orbit at the median tidal
radius (Table 1) around a cluster of median mass. Do this in astro-
physical, physical and N -body units.
3) Using the fact that the magnitude, m, and luminosity, L, of a source
at a distance d are related by
L d
m = −2.5 log10 + 5 log10 +M ,
L 10pc
where L and M are the luminosity and absolute magnitude of the
sun, find the mean surface brightness (within the tidal radius) of the
median globular cluster (using data from Table 1). Express your result
in magnitudes per square arc second.
 2
Theoretical Physics Introduction

A Single Coupling Constant

The gravitational N -body problem can be defined as the challenge to

understand the motion of N point masses, acted upon by their mutual
gravitational forces (Eq.[1.1]). From the physical point of view a fun-
damental feature of these equations is the presence of only one coupling
constant: the constant of gravitation, G = 6.67 × 10−8 cm3 g−1 sec−2
(see Seife 2000 for recent measurements). It is even possible to remove
this altogether by making a choice of units in which G = 1. Matters
would be more complicated if there existed some length scale at which
the gravitational interaction departed from the inverse square dependence
on distance. Despite continuing efforts, no such behaviour has been found
(Schwarzschild 2000).
The fact that a self-gravitating system of point masses is governed
by a law with only one coupling constant (or none, after scaling) has
important consequences. In contrast to most macroscopic systems, there
is no decoupling of scales. We do not have at our disposal separate
dials that can be set in order to study the behaviour of local and global
aspects separately. As a consequence, the only real freedom we have,
when modeling a self-gravitating system of point masses, is our choice of
the value of the dimensionless number N , the number of particles in the
system.
As we will see, the value of N determines a large number of seemingly
independent characteristics of the system: its granularity and thereby its
speed of internal heat transport and evolution; the size of the central
region of highest density after the system settles down in an asymptotic
state; the nature of the oscillations that may occur in this central region;
and to a surprisingly weak extent the rate of exponential divergence of
nearby trajectories in the system. The significance of the value of N is
underlined by the fact that N often makes its appearance in the very

15
16 2 Theoretical Physics Introduction

names of the problems we study, e.g. the famous “three-body problem”,

and in the titles of books like this one.

N = 2, 3
The three-body problem is so famous precisely because it is one of the
oldest problems that cannot be solved. In contrast, the two-body problem
is one of the oldest solved problems. It was Newton’s great triumph
that he was able to explain why planets move in elliptical orbits around
the sun, as had been discovered earlier by the brilliant Kepler, using
observational data by Tycho Brahe. The significance of Newton’s solution
can hardly be overestimated, since the result was a breaching of the
seemingly separated realms of the temporal, human, sub-lunar world and
the eternal world beyond the Moon’s orb.
Though Newton had shown how to solve the two-body problem, the
presence of even an infinitesimal third body (the so-called “restricted”
three-body problem) is insoluble in the usual sense.∗ We should not
forget, however, that Newton was able to solve some three-body problems
approximately. He knew, for instance, that it is better to treat a planet as
revolving around the barycentre of the inner solar system than about the
sun itself. If we do not insist on a precise solution, as with the two-body
problem, in other words, if we look at it from the point of view of physics
rather than mathematics, as Newton did, many aspects of the behaviour
of the three-body problem are not so hard to understand. Indeed it is one
of the quiet triumphs of recent decades that our intuitive understanding
of three-body motions has developed to the extent it has. Most physics
undergraduates are still exposed to the two-body problem, and can easily
develop a feel for the motion in ellipses and hyperbolae, especially when
the scattering problems of atomic physics are encountered. We think that
it is possible to arrive at a similar feel for the richer behaviour of three-
body systems, and one of our aims in writing this book is to demonstrate
this.
The mathematical development of the three-body problem continued
in the hands of Euler, Lagrange, Laplace and many others, who system-
atised the methods of approximate solution. These led to the remarkable
and successful development of the theory of planetary motion; the suc-
cesses of the recovery of Ceres and the discovery of Neptune; and the
familiar but no less remarkable ability of astronomers to predict the mo-

∗
As an aside, it is amusing to see the progress of physics reflected in our (in)ability to
solve N-body problems. In Newtonian gravity we cannot solve the 3-body problem.
In general relativity we cannot solve the 2-body problem. In quantum electrody-
namics we cannot solve the 1-body problem. And in quantum chromodynamics we
cannot even solve the 0-body problem, the vacuum.
 N →∞ 17

tions of the planets, satellites, asteroids, and comets. Equally, they led to
profound insights in the emerging fields of dynamical systems and chaos
(Chapter 4), which in turn have illuminated our understanding of N -
body problems. But they also led away from the sort of approach we
need if our goal is to find quantitative answers to questions about the
million-body problem.

N → ∞
Throughout classical physics there are several important analogues to
the million-body problem, and it is from this background, and not so
much from celestial mechanics, that the most fruitful approaches have
come. At first sight the closest analogy is with plasma physics, where
inverse square laws, as in Eq.(1.1), also arise, but the analogy has not
proved as fruitful as one might think.
In the first place plasmas are often nearly neutral, and for this reason
it makes sense to conceive of plasmas which are nearly uniform, nearly at
rest and of large spatial extent. By contrast, it is impossible to conceive
of an infinite uniform gravitational medium in equilibrium (Fig.1). New-
ton himself glimpsed this problem, and solved it by noting that the stars
had been placed at immense distances from each other, lest they should,
“by their gravity, fall on each other”. This was a satisfactory solution
if one supposed that the system of fixed stars were young enough. For
stellar systems the modern solution is that the random motions or circu-
lation of the stars maintain dynamic equilibrium, just as the motions of
the planets prevent their falling into the sun. Even so, we shall see that
the contraction of stellar systems in dynamic quasi-equilibrium is a real
and fascinating issue, though it is a much gentler process than the head-
long rush suggested by the phrase “gravitational collapse”. In problems
of cosmology, Newton’s dilemma is avoided by an overall expansion or
contraction of the entire universe, though a general relativistic treatment
is required.
The second basic feature of an infinite plasma is that several impor-
tant effects are localised within a Debye length, beyond which the rear-
rangement of charges effectively screens off the influence of an individual
charge (see, for example, Sturrock 1994). If one naively computes the
Debye length for a stellar system, it is found usually to be of the order of
the size of the system itself (Problem 2). Thus it is difficult to treat grav-
itational interactions as being localised within a stellar system, though it
is a difficulty that stellar dynamicists habitually ignore. In any event, all
these considerations help to explain why many of the well known proper-
ties of plasmas have little relevance to stellar systems. And yet there is
18 2 Theoretical Physics Introduction

-2

-2

-2

-2
-2 -1 0 1 2 -2 -1 0 1 2 -2 -1 0 1 2 -2 -1 0 1 2 -2 -1 0 1 2
Fig. 1. Collapse of a uniform “cold” sphere. Initially 2048 particles are dis-
tributed randomly within a sphere. In units such that G = 1, the initial radius
√
is 2.4, the total mass is 1, and successive frames are taken at intervals of 2 2/5
(cf. Box 1.1). Late in the collapse (frame 7 and 8) the distribution of particles
has becomes very irregular. In frame 9 the particles which arrived first have be-
gun to reexpand. By the last row of frames the remaining central condensation
has settled nearly into dynamical equilibrium.

one phrase that stellar dynamicists use all the time which betrays their
debt to plasma physics: the words “Coulomb logarithm”, which occur in
the theory of relaxation (Chapter 14).

Thermodynamics
The plasma analogy exploits the form of the force equation but not
the size of N . Physicists routinely study problems which great numbers
of particles using concepts of thermodynamics. How fruitful is such an
approach to the million-body problem?
From a formal point of view, unfortunately, the field of thermodynamics
excludes a description of self-gravitating systems. The reason is that
the existence of gravitational long-range forces violates the notion of an
asymptotic thermodynamic limit in which physical quantities are either
intensive or extensive. Gravity exhibits what is known in particle physics
 Thermodynamics 19

as an infra-red divergence. This means that the effect of long-distance

interactions cannot be neglected, even though gravitational forces fall off
with the inverse square of the distance (Problem 3).
Take a large box containing a homogeneous swarm of stars. Now en-
large the box, keeping the density and temperature of the star distribu-
tion constant. The total mass M of the stars will then scale with the
size R of the box as M ∝ R3 , and the total kinetic energy Ekin will
simply scale with the mass: Ekin ∝ M . The total potential energy Epot ,
however, will grow faster: Epot ∝ M 2 /R ∝ M 5/3 . Unlike intensive ther-
modynamic variables that stay constant when we enlarge the system, and
unlike extensive variables that grow linearly with the mass of the system,
Epot is a superextensive variable, growing faster than linear.
As a consequence, the specific gravitational potential energy of the
system, the total potential energy of the system divided by the particle
number N , grows without bounds when we increase N . This causes vari-
ous problems. For example, the kinetic energy of a stable self-gravitating
system is directly related to the gravitational potential energy through
the so-called virial theorem (Chapter 9). Therefore, we have to make a
choice when enlarging a self-gravitating system. Either we increase the
temperature steadily while increasing N , in order to increase the specific
kinetic energy enough to satisfy the virial theorem and guarantee sta-
bility. Or we keep the temperature constant, and quickly lose stability
when enlarging our system. In the latter case, the system will ‘curdle’: it
will fall apart in more and more subclumps, and the original homogeneity
will be lost quickly.
In conclusion, there is no way that we can reach an asymptotic ther-
modynamic limit, with the system size becoming arbitrarily large while
holding the intensive variables fixed∗ . Therefore, the traditional road to
equilibrium thermodynamics is blocked. There are no arbitrarily large
homogeneous distributions of stars. As soon as the Universe became old
and cold enough to let matter condense out of the original fire ball into
‘islands’ in the form of galaxy clusters and galaxies, the original homo-
geneity was lost. And each individual clump of self-gravitating material,
be it a galaxy or a star cluster, is ultimately unstable against evapora-
tion, and will fall apart into a bunch of escaping particles (Fig.2). Most
of these escapers will be single, some will escape as stable pairs, and a
few will even manage to form stable triples or higher-number multiples
of particles.

∗
A consistent theory is possible if we let R → ∞ with N/R fixed (de Vega & S’anchez
2000), though what relevance this might have to the million-body problem is unclear
20 2 Theoretical Physics Introduction

The presence of these few-body systems is a robust feature of N -body

systems, as we shall see throughout this book. From our present perspec-
tive, it is an indication of another problem: a short-range (“ultra-violet”)
divergence. The usual Boltzmann factor used in calculations of canonical
ensemble averages, i.e. exp(−E/kT ), gives divergent results in the limit
when two particles approach each other! within a small distance r. This
Gm2
factor then contains a term exp .
rkT

A Lack of Handles
Even though we cannot use thermodynamics in a formal way, when
dealing with a star cluster, we can still describe the motion of the stars
in a way that is analogous to the treatment of the motion of molecules in
a gas studied in a laboratory. One important difference is that a swarm
of stars forms an open system, while a body of gas in a lab has to be
contained. Typical textbook experiments in thermodynamics show the
gas to reside inside a cylinder, with a movable piston that allows the
experimenter to change the volume of the gas. In a star cluster, there are
no cylinder and piston. Instead, the stars are confined by their collective
gravitational field.
The structural simplicity of a star cluster thus allows far less experi-
mentation than is the case for a body of gas in a lab situation. Whether in
6

0
y

-2

-4

-6
-8 -6 -4 -2 0 2 4 6 8
x

Fig. 2. Escape from an N -body system. This computer model of a star cluster
shows a thin stream of escapers emerging at the left and right. The escapers
are channeled into these streams because of external forces, and the streams are
curved because of Coriolis forces.
 Towards an understanding of the million-body problem 21

thought experiments, computer simulations, or in actual table top exper-

iments, the macroscopic parameters of a laboratory gas can be changed
freely, independent of the microscopic parameters governing the attrac-
tion and repulsion between individual molecules. Temperature, density,
and size of the system all can be varied at will. In contrast, once the
number of particles in a self-gravitating system has been chosen, we are
left with no degree of freedom at all, apart from trivial scalings in the
choice of unit of length, time, and mass.
The fact that there are no dials that can be turned in a self-gravitating
experiment, apart from the choice of the total number of stars, is directly
related to the ultra-violet and infra-red divergences of classical gravity.
Having a simple shape for the gravitational potential energy well, with
an energy inversely proportional to distance, leaves no room for preferred
length scales. In contrast, molecular interactions show far more compli-
cated forces, typically strongly repulsive at shorter distances and weakly
attractive at larger separations between the molecules. This change in be-
haviour automatically specifies particular length scales, for example the
distance at which repulsion changes into attraction. In contrast, gravity is
attractive everywhere, at least in the classical Newtonian approximation.

Towards an understanding of the million-body problem

It is now being realised that the gravitational N -body problem is just one
of a growing list of known systems with long-range interactions where the
non-extensivity of energy looks like an obstacle (Cipriani & Pettini 2001).
(In other contexts these are known as “small systems”, to indicate that
their spatial extent is comparable with the range of the relevant interac-
tion.) It turns out that non-extensivity is only an issue if we insist on
treating these problems with the traditional tools of canonical ensembles
(corresponding to systems immersed in a heat bath). This hardly seems
natural in the stellar dynamical context. It is the microcanonical ensem-
ble which corresponds best to the isolated N -body systems which have
been studied so much in stellar dynamics. Indeed it is known that the
different ensembles one studies in thermodynamics, and which are usually
regarded as equivalent for many purposes, have very different properties
for self-gravitating systems (see, for example, Youngkins & Miller 2000).
The formal inability to apply traditional thermodynamic concepts,
then, does not seriously hinder us from thinking and working with them.
Gravity is just as important in the theory of stellar structure and evolu-
tion, where thermodynamics is just as much the stock-in-trade as in any
other area of gas dynamics. Indeed it was precisely the search for extrema
of the entropy of stellar systems that led V.A. Antonov (Antonov 1962)
22 2 Theoretical Physics Introduction

to one of the most profound insights into the behaviour of stellar systems
– gravothermal stability.
In addition to fundamental approaches like this, there are several other
routes by which the behaviour of N -body systems are explored (Chapter
9). One can construct toy models, one can borrow models from other
areas (such as the theory of stellar evolution, or the kinetic theory of
gases), and these can be studied analytically or by numerical methods.
Finally, simulations may be based, with the minimum of simplifying as-
sumptions, on the numerical integration of Eqs.(1.1). As a result of all
these types of approaches, over the last few decades we have developed
a relatively deep and accurate understanding of the many subtle aspects
of the evolution of a system of gravitating point masses, the topic of this
book. Indeed the study of the gravitational N -body problem must rank
as one of those mature areas of research where the variety of approaches
enrich each other like a community of craftsmen.
Let us highlight a few points of interest for a general physics point of
view.
Perhaps the most fascinating aspect of self-gravitating systems is their
instability∗ , exemplified by the fact that these systems have a negative
heat capacity (Chapter 5). Thermodynamic purists would shudder, but
the idea is intuitively simple to grasp and very powerful. Removing heat
from a system means reducing its kinetic energy of random motions. If
this is done to a self-gravitating system, its constituents fall towards each
other a little, and in doing so actually pick up more kinetic energy than
they lost in the first place. When this concept is applied to part of a
stellar system (as it easily can be in thought experiments) an instability
can result. Imagine what might happen to a block of material with a
negative specific heat, as heat flows from a hot spot: the more heat it
would lose, the hotter it would get, and it would quickly burst into flame.
Next, the N -body problem is a fascinating example of a system for
which no useful equilibrium exists. Particles can and do escape (Fig.2).
A Maxwellian distribution of velocities would always allow particles of
arbitrarily high speed, but in self-gravitating systems there is a finite es-
cape speed. Therefore the notion of “local thermodynamic equilibrium”,
which is such a powerful idea in many areas, has limited usefulness. Even
if we prevent particles from escaping (in a thought experiment), ther-
modynamic equilibrium is possible, but may be unstable (because of the
negative specific heat). There is certainly no equilibrium in the sense of
solutions of Eqs.(1.1) in which all particles are at rest. Even the con-

∗
We exclude N = 2, and certain kinds of stable larger systems, e.g. hierarchical
triples (Chapter 25)
 Towards an understanding of the million-body problem 23

cept of dynamic equilibrium (if we ignore for a moment the escape of

particles) brings with it the difficulty of showing why such equilibria are
stable. This in turn depends on an understanding of the global modes
of oscillation of a system in dynamic equilibrium, and for stellar systems
like globular clusters this area is still in its infancy, despite a great deal
of work by many experts. One of the obstacles, of course, is the severe
spatial inhomogeneity of gravitating systems.
The N -body problem exemplifies some of the most perplexing issues
in statistical mechanics. Even though the equations of motion are re-
versible, particles escape and are never captured. A stellar system with
a collapsing core (Chapter 17) will collapse, even if the velocities are re-
versed. In every reasonable sense it is a chaotic system, which rapidly
forgets its initial conditions, and collisions can almost always be treated
with Boltzmann’s classic “Stosszahlansatz”. The fact that there are ex-
ceptions, e.g. the mutual interaction of stellar orbits in the dominating
field of a central black hole (Rauch & Tremaine 1996), is an avenue (so
far unexplored) for investigating the foundations and limitations of the
stochastic treatment of collisions.

Problems
1) How would Newton have solved the problem of the collapse of the
system of fixed stars if he had had access to a GRAPE∗ ?
2) In plasma physics the Debye length is defined to be
s
kTe
λD = .
4πe2 Ne
Translate this into the language of stellar dynamics, bearing in mind
that the rms speed of electrons is related to the electron temperature
by
s
3kTe
vTe = .
me
If a stellar system is in dynamic equilibrium then it approximately
obeys the virial relation 2T + W = 0, where T is now the total kinetic
energy and W the total potential energy (Chapter 9). By making
suitable estimates for the density and other parameters of the system,
show that its radius is comparable with the Debye length.
3) Newton’s Theorems on the gravitational force due to a uniform spher-
ical shell imply that the force inside vanishes, and the force outside
is the same as that due to an equal point mass at the centre of the

∗
instead of an Apple
24 2 Theoretical Physics Introduction

shell. Hence show that the force at a point due to an infinite uniform
medium can take any value we please.
4) Using Newton’s Theorems (Problem 3) show that the acceleration of
a point at a distance r from the centre of a uniform sphere of finite
radius a and total mass M is
GM r
r̈ = − 3 .
a
Deduce that the sphere collapses homologously, and that collapse is
complete at time
s
π2 a30
t= ,
8GM
where a0 is the initial value. Compute this for the system displayed
in Fig.1.
 3
Computational
Physics Introduction

Following the evolution of a star cluster is among the most compute-

intensive and delicate problems in science, let alone stellar dynamics. The
main challenges are to deal with the extreme discrepancy of length- and
time-scales, the need to resolve the very small deviations from thermal
equilibrium that drive the evolution of the system, and the shear number
of computations involved. Though numerical algorithms of many kinds
are used, this is not an exercise in numerical analysis: the choice of al-
gorithm and accuracy are dictated by the need to simulate the physics
faithfully rather than to solve the equations of motion as exactly as pos-
sible.

Length/time-scale problem

Simultaneous close encounters between three or more stars have to be

modeled accurately, since they determine the exchange of energy and
angular momentum between internal and external degrees of freedom
(Chapter 23). Especially the energy flow is important, since the genera-
tion of energy by double stars provides the heat input needed to drive the
evolution of the whole system, at least in its later stages (Chapter 27f).
Unfortunately, the size of the stars is a factor 109 smaller than the size of
a typical star cluster. If neutron stars are taken into account, the prob-
lem is worse, and we have a factor of 1014 instead, for the discrepancy in
length scales.
The time scales involved are even worse, a close passage between two
stars taking place on a time scale of hours for normal stars, milliseconds
for neutron stars (Table 1). In contrast, the time scale on which star
clusters evolve can be as long as the age of the universe, of order ten

25
 26 3 Computational Physics Introduction

Table 1. Time Scales of the Million-Body Problem

Time Scale Stellar Dynamics Stellar Evolution Human Evolution

Seconds White dwarf collision Formation of neutron star Heartbeat

Years Hard/soft binary period1 Long-period variable Malt whisky
Myrs Crossing time2 Shortest stellar lifetime Human evolution
Gyrs Relaxation time3 Lifetime of the sun Life on Earth
1 Chapter 19; 2 Chapter 1; 3 Chapter 14

billion years, giving a discrepancy of time scales of a factor 1014 for normal
stars, and 1020 for neutron stars.
indextime scales
Sophisticated algorithms have been developed over the years to deal
with these problems, using individual time step schemes, local coordinate
patches, and even the introduction of mappings into four dimensions in
order to regularize the 3-D Kepler problem (through a Hopf map to a 4-D
harmonic oscillator, cf. Chapter 15). While these algorithms have been
crucial to make the problem tractable, they are still very time-consuming.

Near-equilibrium problem
In the central regions of a star cluster, the two-body relaxation time scale,
which determines the rate at which energy can be conducted
through the system, can be far shorter than the time scale of evolution for
the system as a whole, by several orders of magnitude. For example, in
globular clusters, density contrasts between the centre and the half-mass
radius can easily be as large as 104 , which implies a similar discrepancy
in relaxation time scales.
As a consequence, thermal equilibrium is maintained to a very high
degree. Since it is precisely the deviation from thermal equilibrium that
drives the evolution of the system (Chapter 2), it is extremely difficult to
cut corners in the calculation of close encounters. If any systematic type
of error would slip in here, even at the level of, say, 10−6 , the result could
easily invalidate the whole calculation. It is for this reason that none of
the recently developed fast methods for approximate force calculations
has been adopted in this area, e.g. tree codes, P3 M codes, etc (Barnes
& Hut 1986, Greengard 1990, Efstathiou et al. 1985). All such methods
gain speed at the expense of relatively large errors in the force computa-
tion. The result is that the N -body simulations of stellar dynamics can
boast far fewer particles than in, say cosmological simulations. This is a
pity, because N is often used as a crude “figure of merit” in the art of
simulation, whereas what really matters is the value of the science that
comes out.
 Computational requirements 27

Computational requirements
The cpu cost of a direct N -body calculation scales ∝ N 3 , where the inter-
particle forces contribute two powers in N (Problem 1) and the increased
time scale for heat conduction contributes the third factor of N . For
this reason the progress of N -body simulations of star clusters has been
painfully slow, from the earliest published work of S. von Hoerner in 1960
(Fig.1).
Almost none of this progress has been made by large general-purpose
supercomputers. The use of parallel computers, such as Cray-T3E, has
had less impact on this area than on many others, because of the com-
munication bottleneck. The force on each particle necessarily depends
on the position of every other, and therefore it is not usually efficient to
parallelise the force calculations (which are the main bottleneck in serial
codes). Another way of exploiting parallelism is to advance many parti-
cles simultaneously (Spurzem & Baumgardt 2001). While this works well
in simple cases, the enormous range of time scales can ruin the efficiency
of this approach also: individual time steps were introduced precisely so

100000 ?Makino

Makino

10000 Spurzem & Aarseth

Aarseth & Heggie
Inagaki
N

1000 Terlevich

Aarseth
100

von Hoerner
10
1960 1965 1970 1975 1980 1985 1990 1995 2000
Date

Fig. 1. The slow progress of N -body simulations of star clusters. Models com-
puted well into the late evolution are plotted against publication date. For a
human perspective, see Aarseth (1999) and von Hoerner (2001).
28 3 Computational Physics Introduction

that it should not be necessary to advance all particles with the tiny time
step required for one close binary!
Currently, with routine calculations, it is only feasible to model the
evolution of a globular cluster containing a few thousand stars, since this
requires some 1015 floating point calculations, equivalent to 10 Gflops-
day, or several months to a year on a typical workstation. Therefore, a
calculation with half a million stars, resembling a typical globular star
cluster, will require ∼ 10 Pflops-day.
In contrast, the memory requirements are and will remain very modest.
All that is needed is to keep track of N = 5 × 105 particles, each with
a mass, position, velocity, and a few higher derivatives for higher-order
integration algorithms. Adding a few extra diagnostics per particle still
will keep the total number of words per particle to about 25 or so. With
200 bytes per particle, the total memory requirement will be a mere 100
Mbytes.
Output requirements will not be severe either. A snapshot of the po-
sitions and velocities of all particles will only take 10 Mbytes. With, say,
105 snapshot outputs for a run, the total run worth 10 Pflops-day will
result in an output of only 1 Tbyte.

Special-purpose Hardware
While general-purpose supercomputers have not yet made much impact
in this field, from time to time it has attracted attention as a possible
application of special-purpose hardware (Fukushige et al. 1999). The
earliest idea along these lines was put into practice by Holmberg as long
ago as 1941 (Holmberg 1941; see also Tremaine 1981). He arranged a
set of light bulbs like the stars in a stellar system, and used photometers
to determine the illumination at each. Since light also obeys an inverse
square law, this provided an analogue estimate of the gravitational field.
The next step in astronomy took place not in stellar dynamics but
in celestial mechanics, with the building and development of the Digital
Orrery (Applegate et al. 1985). For several years it performed ground-
breaking calculations on the stability of the solar system, including the
discovery of chaos in the motion of Pluto (Sussman & Wisdom 1988), until
eventually being regretfully laid to rest in the Smithsonian Museum.
A significant step toward the modeling of globular star clusters was
made in 1995 with the completion of a special-purpose piece of hard-
ware, the GRAPE-4, by an ingenious group of astrophysicists at Tokyo
University (Makino & Taiji 1998). GRAPE, short for GRAvity PipE,
is the name of a family of pipeline processors that contain chips spe-
cially designed to calculate the Newtonian gravitational force between
particles. A GRAPE processor operates in cooperation with a general-
 Special-purpose Hardware 29

purpose host computer, typically a normal workstation. Just as a floating

point accelerator can improve the floating point speed of a personal com-
puter, without any need to modify the software on that computer, so the
GRAPE chips act as a form of Newtonian accelerator (Box 1).
The force integration and particle pushing are all done on the host
computer, and only the inter-particle force calculations are done on the
GRAPE. Since the latter require a computer power that scales with N 2 ,
while the former only require power ∝ N , load balance can always be
achieved by choosing N values large enough.
For example, the complete GRAPE-4 configuration, with a speed of
more than 1 Tflops, could be efficiently driven by a workstation of 100
Mflops. Although such a workstation operates at a speed that is lower
than that of the GRAPE by a factor of 104 , load balance could be achieved
for particle numbers of order N ∼ 5 × 105 . In practice, even with this
hardware, routine calculations did not greatly exceed a particle number
of about 104 , since much larger simulations could not be completed in less
than a few months, and it has been found scientifically more productive
to compute large numbers of relatively modest simulations rather than
to break records. (Note, by the way, that the extreme parallelism of

Fig. 2. The Grape-6 at the University of Tokyo, with J. Makino (right). With
permission.
30 3 Computational Physics Introduction

Box 1. GRAPE design

The design of a typical GRAPE chip (Fig.1) reflects the N -body equa-
tions (Eq.(1.1)). The position and mass of an attracting particle (index
j) are read from memory, the difference r = ri − rj is computed, then
r2 , then r3 , and finally one term on the right of the equations of motion.
Contributions from all attracting particles are summed. One reason for
the efficiency of GRAPE is the fact that, because of the “pipelined” de-
sign, one contribution is computed for each clock cycle. On a conventional
computer each arithmetic operation usually requires several clock cycles,
and each contribution requires about 30 such operations.

Fig. 1 Schematic of a GRAPE-3 chip (from Okumura et al. 1993).

This and other versions of the GRAPE chips perform a few other in-
tensive calculations at the same time. Also, several chips or pipelines are
installed, along with control hardware and memory for the particle data,
on one board, rather like the mother board of a conventional PC. This
board communicates with the host computer via a cable and an inter-
face board, such as a PCI card. In larger installations, several GRAPE
boards are arranged in a tree with suitable communications interfaces to
the single host computer.
In order to make use of an installed GRAPE, sections of a simulation
program are replaced by calls to software libraries which have been writ-
ten by the GRAPE team in Tokyo. For example, computation of forces
(and potential) on all n particles on the now-obsolete GRAPE-3 with 8
chips was computed as follows, using the library function g3frc:

the GRAPE does not allow the most efficient scalar algorithm to be
implemented.) In addition, cut-down versions of this computer can be
 Special-purpose Hardware 31

Box 1 (continued)
do 119 i=1,n,8
ii = 8
if(i+ii .gt. n+1) ii = n- i + 1
call g3frc(x(1,i),awork,pwork,ii)
do 1198 j = 1,ii
do 1197 k = 1,3
f(k,i+j-1)=awork(k,j)
1197 continue
pot(i+j-1) = pwork(j) + mass(i+j-1)*epsinv
1198 continue
119 continue
The particle data are loaded beforehand with similar instructions.

and have been used for simulations in a wide range of other problems
in astrophysics (Hut & Makino 1999) and, indeed, other fields of science,
such as plasma physics, molecular dynamics, the study of turbulence, and
even protein folding.
There are several reasons for GRAPE’s success. In the first place it was
developed quickly, always keeping ahead of general-purpose computers.
Secondly, the mathematical model of inverse square laws is quite fixed,
and can be “hard-wired”. Thirdly, the GRAPE group ensured that the
devices could be made available to potential users throughout the world,
and this maximised the scientific returns.
The introduction of special-purpose hardware has been a truly revolu-
tionary advance, and not just in speed. Before GRAPE and its prede-
cessor, the Digital Orrery, the hardware used by theorists was bought off
the shelf from a computer dealer. By contrast observers have always built
their own hardware (or have had it built to their own specification), even
back to the time of Galileo. From this perspective, GRAPE represents
a remarkable culture shift in the way theorists can do science. In retro-
spect it is not surprising that it is in the area of dynamical astronomy
that this has happened, as it is here that the governing equations and the
underlying physical model are most stable. And we are not yet at the
end of the road: GRAPE 5 is already at work (Kawai et al. 2000) and, as
we write, GRAPE 6 is coming on stream (Fig.2). It is about 100 times
faster than GRAPE 4.
32 3 Computational Physics Introduction

Software Environments
Generating data is only half the job in any simulation. The other half
of the work of a computational theorist parallels that of an observer, and
lies in the job of data reduction. As in the observational case, here too
a good set of tools is essential. And not only that: unless the tools can
be used in a flexible and coherent software environment, their usefulness
will still be limited.
Three requirements are central in handling the data flow from a full-
scale star cluster simulation: modularity, flexibility, and compatibility.
For example, to set up a major simulation, it is very useful to have a set
of model building tools that are sufficiently modular, so that they can be
combined in many different ways. If the data representation is flexible
enough, it will be possible to add new physical variables whose use may
not have been foreseen at the time that the software package was first
developed. And in order for those new variables not to interfere with
existing programs, compatibility is a vital issue as well.
The two main specially constructed environments in use are called
Nemo and starlab (Box 2). Both consist of large collections of software
and tools satisfying the above requirements. In addition a great deal of
N -body work is carried out in the unix-type environments used univer-
sally by computational scientists. The principal codes used in this way
are the suite of N -body programmes written by S.J. Aarseth (1985). An
extremely simplified N -body code is provided in Appendix A.

Problems
1) Use either N -body code in Appendix A to investigate how the CPU
time depends on the number of particles. Try to explain the depen-
dence you find.
2) Code the N -body equations (Eq.(1.1)) using a Runge-Kutta solver,
either one specially prepared for the purpose, or one drawn from any
available numerical library, such as Press et al. (1992). Try to ensure
that the accuracy (judged, for example, by energy conservation) is
comparable with that in Appendix A. Compare the timing with that
of Problem 1, and explain the difference.
3) Code the N -body equations in a symbolic computation package, such
as Maple or Mathematica. Compare the timing with that in Problem
2, again with comparable accuracy, and explain the difference.
 Software Environments 33

Box 2. Starlab

Starlab is a software package for simulating the evolution of dense stel-

lar systems and analyzing the resultant data. It is a collection of loosely
coupled programs (“tools”) linked at the level of the UNIX operating
system. The tools share a common data structure and can be combined
in arbitrarily complex ways to study the dynamics of star clusters and
galactic nuclei.
Starlab features the following basic modules:
• Three- and four-body automated scattering packages, constructed
around a time-symmetrized Hermite integration scheme.
• A collection of initialization and analysis routines for use with general
N -body systems.
• A general Kepler package for manipulation of two-body orbits.
• N -body integrators incorporating both 2nd-order leapfrog and 4th-
order Hermite integration algorithms.
• Kira, a general N -body integrator incorporating recursive coordinate
transformations, allowing uniform treatment of hierarchical systems of
arbitrary complexity within a general N -body framework.
In addition, starlab enables the use of stellar evolution packages such
as SeBa, which models the evolution of any star or binary from arbitrary
starting conditions.
A novel aspect of Starlab is its very flexible external data represen-
tation, which guarantees that tools can be combined in arbitrary ways,
without loss of data or internally-generated comments. Thus, two tools
connected by UNIX pipes may operate on different portions of the same
data set, even though neither understands the data structures, or even the
physical variables, used by the other. Unknown data are simply passed
through unchanged to the next tool in the chain.
Individual Starlab modules may be linked in the “traditional” way, as
function calls to C++ (the language in which most of the package is
written), C, or FORTRAN routines, or at a much higher level–as indi-
vidual programs connected by UNIX pipes. The former linkage is more
efficient, and allows finer control of the package’s capabilities; however,
the latter provides a quick and compact way of running test simulations
and managing production runs. The combination affords great flexibil-
ity to Starlab, allowing it to be used by both the novice and the expert
programmer with equal ease.
34 3 Computational Physics Introduction

Box 2 (continued)
To some extent, Starlab is modeled on NEMO, a stellar dynamics soft-
ware environment developed during the 1980s at the Institute for Ad-
vanced Study, Princeton, in large part by Josh Barnes, with input from
Peter Teuben and Piet Hut (and subsequently maintained and extended
by Peter Teuben). Starlab differs from NEMO mainly in its use of UNIX
pipes, rather than temporary files, its use of tree structures rather than ar-
rays to represent N -body systems, and its guarantee of data conservation–
data which are not understood by a given module are simply passed on
to the next rather than filtered out and lost. The original version of
Starlab was written by Piet Hut in 1989, while on sabbatical at Tokyo
University. From 1993 onwards, Steve McMillan has extended starlab,
with help from Piet Hut, Jun Makino, Simon Portegies Zwart and Peter
Teuben. Visualisation tools, such as partiview, have been added by Stu-
art Levy. Nemo, starlab and partiview are all available at the web site
<http://www.manybody.org>.
Recently, the concept of a ‘virtual observatory’ has been the topic of
several workshops and conferences. The idea is to connect the major
observational archives, from radio to optical to X-ray observations, to
make available a ‘digital sky’ online. Archives of large-scale simulations,
such as those provided by Starlab, will be connected with those virtual
observatories as well, facilitating comparisons between observations and
simulations (Teuben et al. 2001).
 4
Mathematical Introduction

For the mathematician, the gravitational N -body problem is the prob-

lem of understanding by pure thought the solutions of the set of differ-
ential equations

X
j=N
ri − rj
r̈i = −G mj , (1)
j=1,j6=i
|ri − rj |3

where rj is the position vector of the jth body at time t, mj is its

mass, G is a constant, and a dot denotes differentiation with respect to t.
Superficially, what distinguishes the work of a mathematician from that
of, say, an astrophysicist, is its organisation into theorems, lemmas, and
so on, but that is simply a matter of style. There are few formal theorems
in Poincaré’s “Les Méthodes Nouvelles de la Mécanique Céleste”, but it
is a rich vein of ideas. At a deeper level, the work of the mathematician
aims at a greater level of rigour.
Apparently it was Herman (1710; see Volk 1975) who first solved the
two-body problem using Eq.(1) (in component form). Since then this
manner of expressing the problem has proved remarkably resilient: much
the same form of equation for the general case can be found over 200 years
later in the book by Moser (1973). Though Eqs.(1) are usually referred to
as “Newtonian”, there is nothing like them in any of Newton’s published
works or writings. Instead, his expositions are dressed in the language of
geometry or infinitesimals. Curiously, the modern language of geometry
has taken an increasingly important role in recent decades: Box 1 shows
a recent statement of the two-body problem, in terms of a manifold M
and its canonical symplectic structure. Normally, however, we prefer to
work with Eq.(1).

35
36 4 Mathematical Introduction

Box 1. The modern two-body problem.

In relatively modern language the two-body problem may be defined

to be (Abraham & Marsden, 1978) the “system (M, H µ , m, µ), where:
(i) M = T ∗ W with canonical symplectic structure,
n o
W = R3 × R3 \∆, ∆ = (q, q)|q ∈ R3 ;
(ii) m ∈ M (initial conditions);
(iii) µ ∈ R, µ > 0 (mass ratio);
(iv) H µ ∈ F (M ) given by
kpk2 kp0 k2 1
H µ (q, q0 , p, p0 ) = + −
2µ 2µ kq − q0 k
where q, q0 ∈ R3 , p, p0 ∈ (R3 )∗ and k k denotes the Euclidean norm in
R3 .”

Our opening remark begs the question of what is meant by a “solution”.

The very existence of a solution, at least locally, is assured by the usual
undergraduate theorem in a course on ordinary differential equations (e.g.
Arnold 1978b). Globally, the obvious pitfalls are the numerous surfaces
(in phase space) where singularities of the differential equations occur,
i.e. two-body singularities (a “hypersurface” Sij where ri = rj for some
pair i, j), three-body singularities (where two two-body surfaces Sij and
Sik intersect), and so on.
Simply by counting dimensions it is easy to see that any given orbit in
phase-space is very unlikely to encounter one of these singularities. The
argument may be illustrated in a three-dimensional context. A single
condition such as x2 + y 2 + z 2 = 1 determines a two-dimensional surface.
An orbit is one-dimensional, and if a given orbit intersects the surface
then neighbouring orbits do (Fig.1). On the other hand if we have two
conditions to satisfy, each yields a surface, and both conditions are sat-
isfied only on their intersection, which is a curve (one-dimensional). It
is still possible for the orbit to intersect this curve, but neighbouring or-
bits do not, in general. This argument shows that curves which intersect
two surfaces simultaneously are rare in three dimensions, or “of measure
zero”.
Returning to the N -body problem, it is easy to show that orbits inter-
secting any of the surfaces Sij are rare, and so two-body and higher-order
collisions are rare in the same sense. The situation changes dramatically,
however, if we allow the position vectors ri to depend on a complex time
 4 Mathematical Introduction 37

Fig. 1. If an orbit in three dimensions intersects a two-dimensional surface,

neighbouring orbits do. If it intersects a one-dimensional curve, neighbouring
orbits do not (in general).

variable t (Box 2). Why should one do this? One answer is that it is
an important setting in which to discuss the analytical properties of the
solutions, not only because of its pure-mathematical significance, but for
“practical” reasons also. For example, the numerical treatment of the
equations of motion requires special care in the vicinity of singularities
(cf. Chapter 22), and in the N -body problem these are usually to be
found in the complex t-plane. Another application (Chapter 21) is in
certain problems of three-body scattering, when a third body temporar-
ily approaches a short-period binary star; it turns out that the change
in its energy is determined by the point in the complex plane where the
intruder collides with the binary.
We have treated the two-body singularities as though they are to be
avoided at all costs. In fact they are quite innocuous. For example, sup-
pose a collision occurs in a two-body problem when t = 0, and we try to
study this problem by numerical integration of the equations of motion,
starting at some negative value of t during the approach to the collision.
As t approaches 0 from below, it will be found that the numerical inte-
gration requires ever smaller time steps as we approach the singularity
at t = 0, and there is nothing that can be done to get past it. Our
local theorems about the existence of solutions also take us no further.
When we examine the same problem analytically, however, it turns out
that the coordinates of the two bodies are expressible, in the run-up to
the collision, as power series in the variable t2/3 (Box 2). Since this se-
ries contains non-integer powers of t we recover the fact that x is not an
analytic function at t = 0. But we also observe that the series can be
expressed in powers of (t2 )1/3 , which may be equally well evaluated when
38 4 Mathematical Introduction

Box 2. Singularities of the two-body problem.

For the planar two-body problem, complex singularities are easily lo-
cated using the exact solution. A negative-energy solution of the equation
r̈ = −r/r3 (1)
may be represented as
r = (a(cos E − e), b sin E), (2)
√
where a, b and e are constants such that b = a 1 − e2 , and E is deter-
mined by Kepler’s equation; this is
n(t − t0 ) = E − e sin E, (3)
where n and t0 are more constants, with n2 a3 = 1. From Eq.(2) it
follows that r = a(1 − e cos E), and so a two-body collision occurs where
cos E = 1/e, i.e. where E = ±i cosh−1 (1/e) + 2πm, in which m is an
√ this corresponds to the complex times t = t0 +(2πm±
integer. By Eq.(3),
i cosh−1 (1/e) − i e−2 − 1).
Now let e = 1 and t0 = 0 in Eqs.(2) and (3). Then a collision occurs
at t = 0, and the analytical solution may be written as r = (x, 0), where
x = a(cos E − 1) (4)
and nt = E − sin E. In the vicinity of the collision at t = 0 we have
nt = E 3 /6 + O(E 5 ) and x = −aE 2 /2 + O(E 4 ). It is not hard to see from
the first of these equations that E is expressible as a series in odd powers
of t1/3 , and then the second equation shows that x may be developed as
a series in powers of t2/3 .

t is positive as when t is negative. Furthermore, it turns out that this

observation is equally valid when we consider two-body collisions which
are perturbed by other bodies, where the argument used in Box 2 (which
is based on the exact solution) no longer holds. In any event, it turns
out that this approach shows that there is an analytic continuation of the
solution beyond the two-body singularity.
Some additional steps are needed before this idea can be turned to
practical use, and we shall have something to say about this important
technique later on (Chapter 15). In the meantime we may say that there
is a choice of independent variable (the quantity t2/3 in the above discus-
sion) which allows us to represent the solution of the N -body problem
as a function which is analytic on the real axis, even if there are two-
 4 Mathematical Introduction 39

body collisions. We may say that the collision singularities have been
regularised.
There is an artificial but mathematically important class of N -body
problems in which these techniques are essential. Two-body collisions
occur naturally if one studies the collinear N -body problem, but when
these collisions have been neutralised, other, higher-order singularities
come into view. The most obvious of these is the triple-collision singular-
ity, in which the coordinates of three bodies tend to coincidence as some
value of t is approached. One way of studying this problem is touched
on in Chapter 21, and it shows that these singularities cannot usually be
handled by the technique of analytic continuation which works so well for
two-body collisions. The problem is that the exponents which occur in
the corresponding power series are not usually simple rational numbers,
such as the power of 2/3 which occurs in the case of two-body encoun- √
ters. Instead, the exponents may involve irrational numbers such as 13
(Chapter 21), even in the simplest case of equal masses, and a series in-
volving such powers of t cannot be evaluated (with a real result) for both
positive and negative values of t. Nevertheless the study of these singu-
larities has progressed to remarkable lengths, and these investigations are
not without their consequences for the astrophysical applications of the
N -body problem.
The collinear N -body problem exhibits also other classes of singulari-
ties, in which no more than two bodies collide at one time, but the colli-
sions occur more and more frequently as a certain time approaches (see
Marchal 1990). In order to give an impression of how this can happen, we
have to anticipate some results of Chapter 21. Consider first the notion
of a binary. In the collinear problem this is a pair of stars exhibiting
a relative motion like that in Eq.(4) of Box 2: they periodically bounce
off each other (as we assume that the relevant two-body encounters have
been regularised). Just after one such bounce the right-hand component
moves to the right at high speed. Now suppose a third body of low mass
approaches from the right and collides with the right-hand component
(Fig.2). After this collision the third body recedes at high speed, its
energy having been gained at the expense of the binary, which becomes
“tighter”. Suppose finally there is a fourth body, to the right of the third
and moving off to the right. The third body, moving very fast, catches
up with it and, being of relatively low mass, bounces back towards the
binary. With sufficient care its next encounter can be arranged to occur,
once again, just after a collision between the binary components. (In
Chapter 20 we shall see in a little more detail how careful choice of initial
conditions can lead, in an analogous problem, to an orbit with desired
properties.) Now we have the design of a powerful four-body machine
40 4 Mathematical Introduction

which, it may be shown, can accelerate the middle body to arbitrarily

high speeds within a finite time.
When we return to a more reasonable number of dimensions it is not
hard to avoid triple collisions in the three-body problem. All that is
needed is to endow the system with non-zero angular momentum in its
barycentric frame. The essential reason is that, if all three particles could
be confined (however briefly) into a sphere of radius r, energy conserva-
tion shows that their speeds would scale as r −1/2 and so their angular
momentum would scale as r1/2 . Thus confinement within an arbitrarily
small volume is inconsistent with non-zero angular momentum.
Even though collisions are usually avoided in three dimensions, singu-
larities analogous to the one shown in Fig.2, though without collisions,
are still possible, at least for the 5-body problem. The story of how this
remarkable result was obtained (by J. Xia) is beautifully told in Diacu &
Holmes (1996); cf. also Saari & Xia (1995). Essentially, Xia’s example
consists of two Sitnikov machines (see Chapter 20) coupled end-to-end,
with cunningly contrived initial conditions.
Though these examples might seem like mathematical playthings, they
bear some resemblance to a curious idea with possible implications (ad-
mittedly, in the very long run) for mankind. As the sun expands and heats
up it may be possible (in principle) to keep the Earth cool by making its
orbit expand. This is done by repeated two-body encounters involving
the Earth and an asteroid, and Jupiter and the asteroid (Korycansky et
al. 2001).

time

1 2 3 4

Fig. 2. Design of a four-body machine for accelerating a particle to infinite

speed in a finite time.
 4 Mathematical Introduction 41

Examples like Xia’s are highly contrived and rare. In the N -body
problem there will usually be no singularities on the real time-axis. Even
if we regularise two-body collisions, however, there will usually be plenty
of singularities in the rest of the complex plane. These prevent us from
being able to express the solution of the N -body problem as power series
in t (or the appropriate regularising variable) which converge for all times.
However there is an amazingly simple transformation of the independent
variable (Box 3) which does allow us (in principle) to write the solution
as a series which converges for all times; it is not a power series in t,
however. Unfortunately the solution in this form has never been put to
practical use.
Since a solution of the 3-body problem may be expressed as a conver-
gent series, it is surprising to recall that this problem is often quoted as
one of the famous unsolved problems of applied mathematics. Clearly,
the issue hinges on what is meant by a “solution”. Though the series ex-
pression is a solution of sorts, it would be very hard to extract from this
series any information about the qualitative behaviour of the motion of
the N bodies (or even any quantitative results). Nor is it very useful for
numerical calculations as the rate of convergence is even more painfully
slow than an N -body simulation. One usually expects much more from
a satisfactory solution of a dynamical problem.
The best known class of soluble problems in dynamics are the so-called
“integrable” problems. We shall take this to mean a problem in which
sufficiently many constants of the motion can be found so that the solu-
tion may be written down in terms of “quadratures”, i.e. an integral of
a function of a single variable. How one finds these integrals is another
matter, and usually boils down to identifying a symmetry of the problem
at hand. For example the motion of a particle in any spherical poten-
tial may be integrated using the fact that the angular momentum and
energy are constant, and the existence of these integrals results from the
invariance of the potential under rotations and time-translation.
The question now arises whether the N -body problem is of this type.
For N = 2 the answer is affirmative, and indeed it may be reduced to
the problem of motion in a spherical potential (cf. Chapter 7). In fact in
this case the quadratures can be carried out analytically. Even for N =
3, however, insufficiently many integrals are known, and the search for
other integrals has even led to theorems proving their non-existence under
certain conditions (Whittaker 1927, Moser 1973). Chapter 20 describes a
particular kind of three-body problem where this question can be settled
rather directly. The existence of integrals was one of the questions which
motivated Poincaré to study the three-body problem and, in the process,
42 4 Mathematical Introduction

Box 3. A series solution of the three-body problem.

For some solution of the 3-body problem let us suppose that the sin-
gularity which is closest to the real t-axis is at a distance h > 0 (i.e. its
imaginary part is ±ih). Then the solution of the N -body equations is
analytic (without singularities) throughout a strip of width 2h (Fig.1).
Now the complex transformation
2h 1+τ
t= log
π 1−τ
maps the unit disk τ ≤ 1 into the strip just mentioned. For example, when
t is real the transformation is τ = tanh(πt/(4h)), and this varies from
τ = −1 to τ = 1 as t increases from −∞ to ∞. Since our solution of the
N -body problem is analytic in the stated strip in the t plane, it follows
that, if τ is used as independent variable, the solution of the N -body
problem has no singularities in the unit disk. Therefore, by elementary
theorems in the theory of analytic functions, the solution can be expanded
as a power series in τ for all τ in the unit disk. In principle, therefore,
a solution of the N -body problem may be represented as a convergent
power series for all time. See Saari (1990), Barrow-Green (1996).

Imt
singularity
h

Re t

-h

Fig. 1. A solution of the 3-body problem is analytic in a strip |Im t| < h, which
maps to the unit disc in the complex τ plane.

to uncover many of the foundations of current research in Hamiltonian

dynamical systems.
What one can do with the known integrals, however, is to reduce the
order of the problem, i.e. the dimensionality of the phase space. The
integrals associated with the motion of the centre of mass (or barycentre),
for instance, reduce the order by 6, but this is not much compared with
 4 Mathematical Introduction 43

6N . With all such tricks even the three-body problem can be reduced
only to order 7 (though one more order can be removed by transformation
of the independent variable). Another way of expressing this is to say
that a three-body system moves on a 7-dimensional subspace of phase
space. By studying the topology of this subspace one can sometimes
reach rigorous conclusions on the stability of three-body systems (the
“c2 h” criterion, cf. Marchal 1990). But this topological problem also
has a pure-mathematical life of its own which recently led to a complete
census of all possible types of topology in this context (cf. Diacu 2000).
The topics raised in this chapter, and even the title of Poincaré’s book,
illustrate the remarkably close links that have existed between mathe-
matics and dynamical astronomy ever since the time of Newton. It is
significant, however, that most of this has occurred within the subject
that is even now termed “celestial mechanics”. Loosely speaking, this is
the mathematical study of few-body problems, usually with one domi-
nant mass, such as those found within the solar system. We think that
there is equally fertile ground for such cross-fertilisation between mathe-
matics and the N -body problem of stellar dynamics, which is the subject
of this book. That this is less well-developed than in the area of celes-
tial mechanics can be traced to the fact that astronomy as a whole has
become largely a part of physics. And yet it will be found from certain
sections of this book that tools which have been developed by mathemati-
cians for their own inscrutable reasons have turned out to have important
applications in stellar dynamics, often several decades afterwards.

Problems
1) Show that the set of initial conditions of the N -body problem which
lead to a collision between a specific pair of particles has codimension
2 (i.e. 2 less than the dimension of the set of all initial conditions).
2) Verify that x = −q(1 − σ2 ), y = 2qσ is a solution of Box 2, Eq.(1),
provided that q 6= 0 and
1 t
σ + σ3 = p 3 .
3 2q
Write down the appropriate collision solution, corresponding to q = 0,
and show that the orbit varies smoothly with q as one passes through
the collision orbit. Determine the geometric nature of the orbits. By
treating the x,y plane as a√complex z- plane and applying the Levi
Civita transformation ζ = z, determine the geometric nature of the
transformed orbits, and verify that they vary smoothly through the
collision orbit.
44 4 Mathematical Introduction

Repeat this problem with the solution given in Box 2, Eq.(2), by

varying e and keeping a fixed.
3) Two particles of mass m move along the x-axis, and are located at
x1 = sin2 (E/2) and x2 = −x1 at time t = E − sin E. Verify that their
motion satisfies the two-body equations if G = 1 and m = 1/2.
A third particle moves on the y-axis, and is massless. Show that its
equation of motion may be written as the system
ẏ = v
y
v̇ = − .
(y 2 + x21 )3/2
Show that there are three possible values of the constant c such that
this system has the solution y = cx1 .
Change the independent variable in the system to E and code the
equations numerically. By taking initial conditions close to one of the
non-zero special solutions found previously, show that it is possible
for the particle to be ejected with very high speed. (For a comparable
problem with equal masses see Szebehely 1974.)
 PART II. THE CONTINUUM
LIMIT: N → ∞

Even though this is a book about dense stellar systems (i.e. what is
often called “collisional” stellar dynamics, though no physical collisions
need take place), it rests on a foundation of “collisionless” stellar dy-
namics, and the relevant aspects are surveyed in these five chapters. In
addition we outline the various ways in which the effects of gravitational
encounters can be incorporated, though the details are deferred to later
sections of the book.
Chapter 5 begins with a discussion of the main aspects of the thermo-
dynamic behaviour of N -body systems: how a stellar system responds
to being put in contact with a “heat bath”, for instance. In fact stellar
systems tend to cool down if heat is added; paradoxical though this might
seem, it helps us to understand even the motion of an earth satellite. A
toy model helps to explain what is happening.
Chapter 6 introduces the basic tools used for describing large numbers
of gravitating particles: phase space, the distribution function f , the
gravitational potential, and the equation governing the evolution of f (the
“collisionless Boltzmann equation”). We outline some of its solutions, and
aspects of the manner in which they evolve, especially phase mixing. We
also look at the development of Jeans’ instability.
For our purpose the most important distribution functions are those
exhibiting spherical spatial symmetry. Therefore Chapter 7 is devoted
to the motion of stars in spherical potentials, including constants of the
motion and their link with symmetry. An important example which we
shall exploit later in other contexts is the Kepler problem. We outline
the theory for more general potentials as well. Finally we explore briefly
the nature of nearly radial motions (touching on their importance for
the stability of stellar systems) and those in time-dependent (but still
spherical) potentials.

45
46 PART II. THE CONTINUUM LIMIT: N → ∞

Chapter 8 builds on Chapter 6, by examining those solutions of the

collisionless Boltzmann equation which are of particular relevance to the
million-body problem. These are the isothermal model, and King’s and
Plummer’s models. We take the opportunity of explaining why it is that
the last of these (which is actually one of the family of polytropic models)
has such a simple form. All these models have a few parameters, which
may be adjusted in order to attempt to fit a model to observations, but
the chapter closes with a brief account of non-parametric methods of
fitting theory to observations.
Chapter 9 continues to develop the dynamical theme of Chapter 6, by
showing how the evolution of a large N -body system may be modelled
approximately. At the crudest level it may be characterised by the mass
and energy of the system, especially if we exploit the virial theorem. At
the next (or, perhaps, the first) level of sophistication we can treat the
system like a fluid (a “star gas”), where the stars are regarded as the
atoms of the fluid. This introduces us to the Jeans equations of stellar
dynamics, but we also depart from the collisionless emphasis of these
chapters by mentioning the manner in which close gravitational encoun-
ters between pairs of stars are incorporated (by analogy with collisions
in a gas). In the same way we show how they modify the collisionless
Boltzmann equation to produce the Fokker-Planck equation (in the sense
in which this term is used in stellar dynamics). To complete the story we
remind ourselves of the most fundamental model: the N -body equations.
 5
Paradoxical Thermodynamics

A Strange Black Box

Compared to laboratory situations, a self-gravitating star cluster is a
very strange object. Imagine that you were handed a star cluster in a
closed box, so that you could only measure the temperature at the surface
of the box. Imagine also that you could change the conditions of the star
cluster from the outside in two ways: 1) you could put the box inside
a larger box with a different temperature, as an effective heat bath, in
order to change the temperature inside; 2) you could change the size of
the box, compressing or expanding its volume.
So far, there is nothing unusual, and we might still pretend that we
are about to carry out a textbook thermodynamics experiment. But
when we dip our box into a heat bath, something strange may occur:
depending on the exact conditions inside the box, the box may exhibit
a most bizarre behavior. When placed in a colder environment, the box
may actually heat up, without limit. The only way to cool the box back
to its original temperature would be to place it temporarily inside an
even hotter environment — but not for too long, otherwise it will cool to
below its original temperature.
This contrary tendency of self-gravitating systems corresponds to the
fact that such systems can exhibit a negative heat capacity. We will
return to this mysterious character later on, when we analyze its effects
in detail, both on macroscopic scales, governing the evolution of a star
cluster as a whole, as well as on a microscopic scale, when we deal with
few-body systems.

Reversing the Pedals

A simple way to gain some familiarity with the negativity of the heat
capacity of a star cluster is to consider a satellite in orbit around the
Earth. When the satellite undergoes some friction from the atmosphere,
it will tend to lose some energy, and as a consequence it will lose some

47
48 5 Paradoxical Thermodynamics

height. Dropping to a lower orbit, however, brings the satellite closer to

the Earth, where the gravitational attraction is larger. This means that
a higher velocity is required in order to stay in a circular orbit, balancing
the centrifugal forces of inertia with the centripetal forces of gravity (Box
1).
The immediate effect, then, of a loss of energy is a speeding up of
the motion of the satellite. Net energy is lost, of course, but even more

Box 1. How air drag speeds up a satellite

Instead of moving on a circular orbit, the satellite describes a tightly

wound spiral (Fig.1). Therefore its velocity has a tiny radial inward
component. Two forces do work on the satellite: air drag, which acts
to oppose the motion, and gravity, which has a tiny component in the
direction of motion. Which wins?
In fact the rate of change of kinetic energy is
GM m
Ṫ = −ṙ − vF, (1)
r2
where m and M are the masses of the satellite and the earth, respectively,
r is the distance between them, v is the speed of the satellite, and F
is the magnitude of the drag force. In nearly circular motion we have
GM m
approximately T = , where we have used the approximate equation
2r
v 2 GM GM m
of motion = 2 , and so Ṫ = − ṙ. Hence it is clear that the
r r 2r2
first (gravitational) term on the right side of Eq.(1) is twice the size of the
second (drag) term. In other words gravity wins and the satellite speeds
up.
F
Satellite orbit

Earth’s surface

To Earth’s Centre

Fig. 1. Motion of a satellite on a nearly circular orbit under the action of a

small drag force.
 Never at Rest 49

potential energy is lost, and therefore the kinetic part of the energy is
actually increased. It would be as if stepping on the brake in a car would
cause the car to speed up, and pushing the accelerator would slow the
car down (in a gravitational car, those pedals presumably would be called
the push and the decelerator, respectively).
Note the symmetry between both operations. Pushing the brake and
seeing the car speed up nonetheless could be accomplished when going
down a steep mountain road, with a brake that is not powerful enough.
In that case, however, pushing the accelerator would have a qualitatively
similar effect, of increasing the velocity (by an even larger amount), and
doing nothing would also lead to a rising speed. In the case of a satellite,
even a slight acceleration would lead to a higher orbit with a consequently
lower speed.
This property of orbits is also of importance in understanding the role
of resonances in the maintenance of spiral patterns in galaxies. In the
words of Lynden-Bell and Kalnajs (1972), “stars act like donkeys slowing
down when pulled forwards and speeding up when held back.” (Transport
analogies seem to be favoured in discussions of this behaviour.)
This paradoxical behaviour does not always occur for centripetal forces
like gravity. Consider, for instance, the artificial case of two-dimensional
gravity. This corresponds to the attraction of infinite parallel rods, which
varies as r−1 , like the force between two line charges in electrostatics.
A system of gravitating rods does not exhibit a negative specific heat
(Padmanabhan 1990, and Problem 2).

Never at Rest

The situation is actually more complicated. Doing nothing, while per-

haps difficult for human beings, is altogether excluded for a self-gravitat-
ing star system. Even if we just leave our box with stars alone, and occa-
sionally measure its temperature, we will find, after a long time, that the
temperature will begin to creep up. Something is going on inside, and all
the signs are pointing towards a perpetuum mobile!
What happens is that, if we wait long enough, a simultaneous close
three-body encounter will produce a tightly bound pair (Eq.(21.9)). It
can be shown (Chapter 19) that from that moment on, the probability
is overwhelmingly large that this pair will grow tighter and tighter, on
average, giving off more and more energy. This energy is converted to
kinetic energy of all the other particles, including the kinetic energy of
the center-of-mass motion of the bound pair (see Fig.1). So energy is still
conserved.
50 5 Paradoxical Thermodynamics

Fig. 1. Energies of escapers in isolated N -body systems (after Giersz & Heggie
1994). In the first half of the evolution there are no binaries; escapers are caused
by two-body encounters (Chapter 16) and have low energy. The escape rate
increases as the core of the system collapses (Chapter 18). Eventually the core is
so dense that binaries form. In their evolution they eject stars with much higher
energies (Chapter 23). If the system were enclosed in a box, as in the text, they
would be perceived through the increased temperature of the walls. Energy is
measured in units of 2/3 of the mean kinetic energy per particle. Results from
many simulations are collected in the one figure.

This is an example of a negative heat capacity showing up on a micro-

scopic level, in a self-gravitating system of particles (Chapter 19). It is
also an example of what is known in particle physics as an ultra-violet
divergence. Since the individual particles are mass points with no spatial
extension, they can come arbitrarily close, and therefore their negative
gravitational binding energy can become arbitrarily large. Just one pair
of particles can therefore provide an unlimited amount of positive energy
to the rest of the system.
In this thought experiment, after closing the lid of the box, we would
notice that the walls of the box would get hotter, without bound. Even
if we would slowly extract heat from the box, its temperature would keep
rising. The tightly bound pair of particles in our experiment will have an
orbital motion that is much faster than that of the single particles that
it encounters. In an attempt to reach equipartition, the bound particles
will try to convey some of their rapid motion to the single particles,
speeding the latter up in the process. The particles in the bound pair
 A Toy Model 51

themselves, however, while attempting to slow down, will find themselves

falling to an even tighter orbit about each other. A shorter distance in
the gravitational two-body problemtwo-body problem implies a higher
orbital velocity, so the net effect is that both the single particles and the
bound particles will speed up as a result of their interactions.

A Toy Model

Imagine a spherical enclosure of radius R with a fixed attracting mass

M at the centre. Another body, of unit mass, freely moves around inside,
and bounces elastically whenever it strikes the inner wall of the sphere.
When its energy is low enough, in fact when E < −GM/R, it can never
reach the sphere, and then its mean kinetic energy is hT i = −E (Problem
3). When its energy is high enough to allow collisions with the sphere,
the calculation of the mean kinetic energy is a good deal harder∗. At
very high energies, however, it is clear that the particle speeds past the
central mass almost without deflection, and its mean kinetic energy is

1.8

1.6

<T>
1.4

1.2

-2 -1.5 -1 -0.5 0 0.5 1

E

Fig. 2. Mean kinetic energy hT i against energy E for the toy model discussed
in the text. Both energies are scaled by GM/R.

∗
See Problem 6.3. The calculations are simpler in a related model discussed by
Padmanabhan (1990), but he takes a more advanced viewpoint thermodynamically.
Our model was suggested by a reading of this paper, which contains several other
highly illuminating toy models and discussions of the thermodynamic behaviour of
self-gravitating systems.
52 5 Paradoxical Thermodynamics

only a little bigger than at the boundary, i.e. hT i >∼ E + GM/R. Using
hT i as a surrogate for temperature, we see that, above some minimum
temperature, a system of a given temperature can be in either a high
energy or a low energy state (Fig.2). Furthermore the high energy state
has positive heat capacity dE/dhT i, and the low energy state has negative
heat capacity.
Now suppose we have many slightly interacting particles inside the
enclosure, one at low energy and high temperature (i.e. high up on the
left branch of the curve in Fig.2), and N others at high energy and low
temperature (i.e. lower down on the right branch). Because the slopes of
the two branches are comparable (though of opposite sign), it is easy to
see that a flow of heat from the hot particle to the remainder heats all the
particles, but the inner body is heated more, roughly in the ratio of N :1.
Thus the temperature difference is exacerbated, and the energy of the
single body decreases still further. The outer bodies strike the wall with
increasing energy, which is registered as an increase in the temperature
which the system presents to the outside world.
This is a toy model not just of paradoxical thermodynamic behaviour,
but of the million-body problem itself. If you turn Fig.2 upside down
it resembles part of Fig.17.2, which is concerned with the stability of
self-gravitating systems.

Problems
1) You are the pilot of a spacecraft about to dock with a satellite which
moves ahead of you on the same circular orbit. In which direction
should the thrusters of your spacecraft be fired in order to catch up
with your objective: forward or backward?
2rF
2) Using the theory in Box 1, show that Ṫ = vF and ṙ = − . If
mv
−n
the attractive force of the Earth were to vary as r , show that the
2rF n−1
results would become ṙ = − and Ṫ = vF . Deduce
mv(3 − n) 3−n
that paradoxical behaviour only occurs if 1 < n < 3.
(Harder) What happens if (a) n = 3, (b) n > 3, (c) n = 1, (d)
n < 1?
3) A particle of unit mass moves in the gravitational field of a fixed mass
M according to the equation of motion
GM r
r̈ = − .
r3
If the energy is E = v2 /2 − GM/r, show that r < R provided that
d2
E < −GM/R. Show that 2 r2 = 2T + 2E, where T is the kinetic
dt
 A Toy Model 53

energy. Assuming that the motion is periodic, deduce that hT i = −E.

(This is a version of the virial theorem (Chapter 9) for the Kepler
problem.)
 6
Statistical Mechanics

In order to progress from qualitative arguments and toy models it is nec-

essary to set up apparatus for describing a gravitational N -body system.
There are several ways in which this can be done.
One common approach is to employ the N position vectors ri and the
N velocity vectors vi of the stars at some time. Each of these vectors
has 3 components, and so the entire system can be described by a single
6N -dimensional vector, i.e. a single point in a 6N -dimensional space Γ.
This is a useful description, because it is sufficient to specify uniquely
the entire subsequent evolution of the system, as the equations of motion
are of second order; they describe the motion of this point through Γ.
Implicitly, therefore, this is the description adopted in N -body methods,
even though it is more natural to think of N particles moving in a 6-
dimensional phase space.
This description in a 6N -dimensional space can be turned into a sta-
tistical one if we imagine a collection of stellar systems, each described
by a distinct point in Γ. If their distribution is described by a probabil-
ity density function f , the evolution of f is determined by the equations
of motion, and indeed is equivalent to them. This description is almost
never used in stellar dynamics.
Another way of describing a stellar system is to represent each star
by a single point in a 6-dimensional space with coordinates r and v.
Thus a single system is represented by N such points. We call this space
“phase space”, though it is rather more usual to define phase space as
a space with coordinates and momenta, rather than velocities. We shall
see, however, that, in situations where this description is most useful,
the basic dynamics is independent of mass. This information, i.e. the
positions in phase space of all N stars, given at some time t, is sufficient
to determine the entire subsequent evolution.

54
 6 Statistical Mechanics 55

We can also use phase space for a statistical description. Either we

imagine that there are sufficiently many stars in the system that their
distribution can be adequately represented by a smooth function, or else
we can imagine the evolution of a collection of systems. In either case we
suppose that there is a function f (r, v, t) which can be interpreted either
as a probability density or, in more physical terms, the density of stars
in phase space, just as the density of a fluid describes the distribution of
atoms in ordinary space. It is often called the “1-particle” distribution, to
distinguish it from the distribution in Γ. There are several ways in which
f can be normalised, depending on its interpretation as a number density,
a mass density orR a probability density. Usually, however,R we adopt the
convention that f d3 rd3 v = N . This implies that f d3 v = n, the
number density in space.
The most important fact about f , however, is not these matters of con-
vention but its main limitation: it is insufficient for an exact description
of the dynamics (see also Fig.1). It cannot tell us what happens when two
stars come very close together, even though a close encounter will dras-
tically alter their orbits. In fact, phase space is a useful description only
when such events are sufficiently rare or insignificant: if the dynamics of
each star is principally determined by the large mass of relatively distant
stars, then the use of phase space is a useful simplification. Put another
way, this is the point at which we make a choice between “collisional”
and “collisionless” stellar dynamics. It is not that we are avoiding physi-
cal collisions; rather we are neglecting the special effects of gravitational
two-body encounters. Fortunately this assumption is justified in practice
in many situations of interest (cf. Chapter 14).
For the time being, then, we suppose that the dynamics of the stars
can be studied if we know only their distribution f . If all N stars have
the same mass m then their space density is
Z
ρ(r, t) = m f (r, v, t)d3 v. (1)

If stars have various masses, then we introduce an m-dependence into f

and modify this integral accordingly. In any event, from ρ we determine
the potential φ(r, t) from Poisson’s equation
∇2 φ = 4πGρ, (2)
where G is the constant of gravitation, and then the equation of motion
of an individual star is
r̈ = −∇φ. (3)
This illustrates two things: (i) that the motion of the star is independent
of its mass, and (ii) the motion of two neighbouring stars will be virtually
56 6 Statistical Mechanics

identical, and independent of their neighbour’s location. The first of these

remarks will play a significant role in Chapter 10; how to improve on the
second approximation is the subject of Chapter 14.
In order to complete the approximate statistical description of a stellar
system, we now show how the equation of motion, Eq.(3), determines
the evolution of f . The analogy with fluids is again helpful. If a fluid
of density ρ moves in 3-space with velocity v, then conservation of mass
leads to the “continuity equation” ∂ρ/∂t + ∇.(ρv) = 0. In our case the
space is 6-dimensional, with “coordinates” given by the 6-vector (r, v),
the corresponding “velocity” is the 6-vector (v, −∇φ), and f takes the
place of density. Thus the corresponding equation is

∂f /∂t + ∇r .(f v) − ∇v .(f ∇φ) = 0, (4)

where
X
3
∂
∇r .(f v) = (f vi ),
r
i=1 i

-0.5

-1

-1.5

-2
φ

-2.5

-3

-3.5

-4
-3 -2 -1 0 1 2 3
x
Fig. 1. The potential well of a Plummer model (Chapter 8) for various numbers
of particles N . The potential φ is plotted for points along the x-axis, for N =
1 (the deepest), 4, 16, 64, 256 and 1024, and for the exact Plummer model
(corresponding to infinite N ; this is the heavy dashed curve). The curve for each
successive N was obtained by adding the appropriate number of particles to the
previous system, i.e. the various systems are not independent realisations of a
Plummer model. This figure illustrates the fact that the potential in a phase-
space description (the heavy dashed curve) may differ substantially from that in
an N -body description.
 6 Statistical Mechanics 57

and ∇v is similarly defined. Because r and v are independent, and φ

does not depend on v, Eq.(4) simplifies to

∂f /∂t + v.∇r f − ∇φ.∇v f = 0. (5)

This equation is rather fundamental. It is often called the “collisionless

Boltzmann equation”, astrophysicists being rather reluctant to use the
term “Vlasov equation”, as is more common in plasma physics∗ . Another
way of writing it is in the form
∂f
+ [H, f ] = 0, (6)
∂t
where H = v2 /2 + φ is the Hamiltonian for the motion of a particle in the
potential field φ and [H, f ] denotes the Poisson bracket of H and f (see,
e.g., Goldstein 1980). This is very useful for problems where Hamiltonian
methods are appropriate, and in that context Eq.(6) is called Liouville’s
Theorem (cf. Goldstein 1980). To write the collisionless Boltzmann equa-
tion in curvilinear coordinates, however, it is better to write Eq.(5) as

∂f /∂t + ṙ.∇r f + v̇.∇v f = 0, (7)

and simply write down the corresponding equation in the new coordinates
(cf. Problem 1).
It is worth remarking that Eq.(6) is equivalent to Eq.(3), and not just
deducible from it: if we seek a solution of Eq.(6) in the form δ(r−r(t), v−
ṙ(t)) then it turns out that r(t) must obey Eq.(3). The link between the
two is also evident from a mathematical point of view. The standard
method of solving a partial differential equation like Eq.(6) is by the
method of characteristics (e.g. Garabedian 1986, Chapter 2), and it turns
out that the characteristics of Eq.(6) are just the solutions of Eq.(3).
The analogy with ordinary fluids remains helpful in the interpretation
of Eq.(5). If a fluid is incompressible, then its “convective derivative”
vanishes; the convective derivative is simply the derivative following the
motion of the fluid, and so ∂ρ/∂t + v.∇ρ = 0. The resemblance to
Eq.(5) shows that, in the phase space description of a stellar system, f
is constant if we follow the motion of a star in phase space; in other
words, phase “fluid” is incompressible. This property often allows us to
visualise a solution of the collisionless Boltzmann equation that would be
quite difficult to write down (Fig.2).

∗
Even here, however, the phrase “Vlasov equation” commonly refers to the coupled
system consisting of the collisionless Boltzmann equation and Maxwell’s equations.
Spelling also differs from one source to another; if you need to look it up in an index,
try also Wlasow. See also Hénon (1982).
58 . 6 Statistical Mechanics
r

-2

-2
-1
-1
0 0
r
1
1
2

Fig. 2. Visualising a solution of the collisionless Boltzmann equation. The one-

dimensional Kepler problem has conserved energy E = ṙ2 /2 − GM/r, and a
two-dimensional phase space with coordinates r, ṙ. (The collision singularity
at r = 0 is irrelevant to our purposes; besides, it could be removed by regu-
larisation (Chapter 15).) Particles move downwards (i.e. ṙ decreases, though r
may either increase or decrease) on curves E = constant. Particles at smaller
E have shorter periods and move more quickly. The figure labels the positions
of two particles at times . . . , −2, −1, 0, 1, 2, . . .. Suppose we have to solve the
collisionless Boltzmann equation with initial conditions such that f = constant
inside the region [−1, −1, 0, 0] and f = 0 outside. At time t = 2 the corners have
moved to [1, 1, 2, 2], and so f = constant inside this region and f = 0 outside.

The fact that f is invariant as we follow particles in phase space can

sometimes be used to impose useful constraints on very complicated mo-
tions in stellar dynamics, provided that the collisionless Boltzmann equa-
tion is valid. For example, in studying collisions between galaxies, which
usually lead to a single remnant, astrophysicists are interested in knowing
conditions near the centre of the remnant. Conservation of phase space
density places an upper bound on the phase density there. In fact, other
dynamical process such as “phase mixing” (Chapter 10) finely intersperse
regions of high and low phase density, and our incompressible phase fluid
becomes more like a foam.
As we have just seen, f is constant along orbits in phase space. In
some cases we know functions with this property. For example the energy
E is constant if φ is time-independent. If φ is spherically symmetric (cf.
Chapter 7) then the angular momentum is constant. Since any function of
these is also constant, it follows that any distribution of the form f (E, J)
solves the collisionless Boltzmann equation if φ is spherically symmetric
and static. In the astrophysical literature this is a special case of what is
 6 Statistical Mechanics 59

often called Jeans’ Theorem, after one of the founding fathers of stellar
dynamics. But Jeans was a fine applied mathematician, and in his own
writings (Jeans 1929) he attributed the result to Lagrange.
It is easy to be confused by the expression f (E, J), because this is not
the distribution function of E and J, but the distribution in phase space
expressed as a function of these two variables. To relate the two concepts,
let us define N (E, J)dEdJ as the number Rof stars with E and J in small
intervals dE, dJ , respectively. Then N = f (E, J)δ(v2 /2 + φ − E)δ(|r ×
v| − J)d3vd3 r. This is easily evaluated by first integrating with respect
to v, using r as polar axis. The result is that

N (E, J) = 8π 2 JP (E, J)f, (8)

where P (E, J) is the radial period of the orbit with parameters E and J
(cf. Chapter 7). Similarly, if f depends on E alone, then the number of
stars per unit energy is
Z rmax
2
N (E) = 16π f (E) r2 (2(E − φ))1/2 dr, (9)
0

where φ(rmax ) = E.
What determines f ? Beyond Jeans’ theorem, Eq.(6) is little guide,
and information on this question comes from other considerations, most
of which lie outside this chapter. First, the interpretation of f shows it
must be non-negative. Another consideration, which is not as irrelevant
as it might seem at first sight, is mathematical convenience. Much ef-
fort has gone into the construction of more-or-less simple distributions
which can be used as models for the interpretation of observations or
the construction of initial conditions for N -body models, and so on (cf.
Chapter 8). Of deeper significance are distribution functions f with phys-
ical meaning. Since Eq.(6) shows that f evolves in time, there is at least
a possibility that concepts of statistical mechanics might be applicable.
If so, it might be possible to suggest particular forms (e.g. entropy max-
ima) towards which f would evolve after sufficient time (cf. Chapter
10). Finally, there is the dynamics which is not embodied in Eq.(6),
i.e. the interactions between neighbouring particles. In a simple model
of an ordinary gas this leads to relaxation to a Maxwellian distribution
of velocities. In stellar dynamics the corresponding mechanism is called
“two-body relaxation” (Chapter 14). It is also known that three- and
four body interactions influence the distribution (Chapter 19, 26), but
these involve “internal degrees of freedom” which are not described by
our single-particle distribution function f .
60 6 Statistical Mechanics

Jeans’ Instability
One of the standard applications of the formalism introduced in this
chapter is the stellar-dynamic analogue of a problem in gas dynamics
studied by Jeans. The calculation is accessible in Binney and Tremaine
(1987), and we merely summarise the physical interpretation and touch
on a historical curiosity. This is a real digression: we have little further
use for the calculation in the remainder of the book.
The problem begins with the unlikely concept of a static, spatially
infinite and uniform medium, described by a potential φ0 , density ρ0 and
distribution function f0 , all of which are independent of position and
time. Next we suppose that these functions are subject to small (t- and
r-dependent) perturbations, denoted by subscript 1, and we proceed to
linearise Eqs.(2) and (5). We find that the terms of zeroth and first order
give the equations
∇2 φ0 = 4πGρ0 (10)
∇2 φ1 = 4πGρ1 (11)
−∇r φ0 .∇v f0 = 0, (12)
∂f1
+ v.∇r f1 − ∇r φ1 .∇v f0 = ∇r φ0 .∇v f1 . (13)
∂t
Unfortunately there is no potential φ0 satisfying both Eqs.(10) and
(12). This should not surprise us: as shown in Problem 2.3, the gravita-
tional force in an infinite uniform system can take any value we please.
We now have two choices. One choice is to revise the assumption of
a uniform, infinite medium in dynamic equilibrium. The second is to
invoke the ‘Jeans swindle’. Let us consider these in turn.
The initial assumptions may be modified in various ways which allow
progress. What we need is an unperturbed model whose potential does
not accelerate (or, more fundamentally, change the momentum of) a per-
turbed particle. (This is the interpretation of the troublesome term on
the right of Eq.(13).) An example is collapse of a “cold” uniform sphere
(Fig.2.1). If we use coordinates which collapse along with the sphere,
all particles are at rest in the absence of perturbation. There is a close
analogy here with the theory of gravitational instability in the expand-
ing universe, where the basic mechanism is well understood (cf. Peebles
1993). The stellar-dynamical problem is quite involved (Aarseth et al.
1988 and Problem 4)∗ , but again the system is unstable, which accounts
for the clumpiness visible in frames 6–8 in the cited figure. Another uni-
form model which does not accelerate a perturbed particle is described

∗
Even its numerical simulation has some subtleties (Theis & Spurzem 1999)
 6 Statistical Mechanics 61

in Box 7.3, though there it is the angular momentum of the perturbed

particle that is unaffected. Finally there is a very interesting (if highly
artificial) system in which the gravitating matter is confined to a ring
and the attraction is suitably doctored (Inagaki & Konishi 1993).
The other way of dealing with the troublesome term in Eq.(13) is simply
to ignore it. Certainly this makes the problem tractable, and the resulting
analysis leads to a stellar dynamical analogy to the Jeans instability of
self-gravitating gases. We shall say a little more about this in Chapter 9.
We do not know who invented the term ‘Jeans swindle’, which invites
the suspicion that Jeans cheated. To the best of our knowledge, Jeans
tackled only the gas-dynamical analogue of the above problem, towards
the end of a long paper (Jeans 1902) which also treated the stability of
both spherical and slowly rotating equilibria. In the latter cases the equi-
librium structure was not neglected, but when it came to the discussion
of the infinite case Jeans argued that, “If space has no boundary there is
presumably no need to satisfy a boundary condition at infinity, so that ρ
may have any value”, and simply asserted that φ0 = 0.

Problems

1) Suppose the distribution function depends only on r, vr and vt , where

r is the distance from the centre, vr is the radial component of velocity,
and vt is the transverse component. Observe, by analogy with Eq.(7),
that the collisionless Boltzmann equation takes the form

∂f ∂f ∂f
∂f /∂t + ṙ. + v̇r + v̇t = 0.
∂r ∂vr ∂vt

If the potential φ is spherically symmetric (Chapter 7), express ṙ, v̇r

and v̇t in terms of r, vr , vt and φ, and deduce that
!
∂f vt2 ∂φ ∂f vr vt ∂f
∂f /∂t + vr + − − = 0.
∂r r ∂r ∂vr r ∂vt

2) If the potential is due to the attraction of a mass M , i.e. φ = −GM/r,

find the relation between f (E) and N (E) for E < 0, using Eq.(9).
3) In the toy model of Chapter 5 (cf. Fig.5.2), a particle of unit mass
moves in the potential φ = −GM/r within a sphere of radius R.
Show that the mean kinetic energy, averaged in the natural way over
62 6 Statistical Mechanics

all particles of energy E, is

R
r≤R δ(v + φ − E)v2 d3 rd3 v
2 /2
hT i = R ,
2 r≤R δ(v2 /2 + φ − E)d3 rd3 v
R rmax  GM 3/2
0 r +E r2 dr
=R  1/2 ,
rmax GM
0 r +E r2 dr
where rmax = R if E > −GM/R, and −GM/E otherwise.
If E > 0 show that
hT iR
= x×
GM p √ √
3x + 17x2 + 22x3 + 8x4 + 3 x(1 + x) ln( 1 + x + x)
× p √ √ ,
−3x − x2 + 10x3 + 8x4 + 3 x(1 + x) ln( 1 + x + x)
where x = RE/(GM ). (We confess to using Maple for this.)
Find similar results in the cases (a) −GM/R < E < 0 and (b)
E < −GM/R. (They are plotted in Fig.5.2).
4) At time t = 0, N particles of mass m are distributed uniformly
inside a sphere of radius R, and are at rest. Show that, because of
Poisson fluctuations, two particles at the same initial radius
√ R have
accelerations which differ by a relative amount of order 1/ N . Deduce
that the times at which they reach some smaller radius R0 differ by a
similar relative amount, and hence √ that their radii at the same time
differ by an amount of order V 0 t0 / N , where V 0 is the infall speed
at time t0 . By arguing√that the minimum radius of the system is
given by Rmin > 0 0 0 0
∼ V t / N , and estimating V and t , deduce that
Rmin ∼ RN −1/3 .
 7
Motion in a Central Potential

There are at least two reasons for studying this classical problem in a
modern book on the gravitational N -body problem. First, it approxi-
mately describes the motion of individual stars in a nearly spherical sys-
tem. Second, it describes the relative motion of two stars which happen
to come close together. These applications are complementary. In the
first we approximate the potential due to the other stars in the system
by a smooth spherically symmetric field (Fig.6.1). Occasionally, however,
two stars come close together, and then the field of the rest of the cluster
can be neglected temporarily. In this case it is the relative motion of the
two stars that is described by a central force equation:
G(m1 + m2 )
r̈ = − r, (1)
r3
where their masses are m1 and m2 , and r is their relative position vector.
The individual motion of each star can then be constructed if the motion
of their barycentre is known.
In a central force problem motion is described by an equation of the
form
r̈ = −∇φ(r),
where r = |r|. Thus choosing φ = −G(m1 + m2 )/r yields Eq.(1). We
define the specific angular momentum, i.e. angular momentum per unit
mass, by J = r × ṙ. It is immediately seen by differentiation that this
is conserved during motion in a central field, even if the field is time-
dependent. (Physically, a central force exerts no torque.) In turn this
implies that motion is on a fixed plane passing through the centre of the
force field, since J.r = 0. The other important invariant (for a static
potential) is the specific energy, defined by
1
E = ṙ2 + φ(r).
2
63
64 7 Motion in a Central Potential

These two invariants reflect symmetries of the dynamical problem, and

are special cases of Noether’s Theorem (Box 1). The invariance of J
reflects only the spherical symmetry of φ, and is true even if the field
is time-dependent. The constancy of E depends only on the assumed
time-independence of φ, and not the spherical symmetry. Certain central
forces have further symmetries, and further invariant quantities can be
constructed. The Kepler problem (this chapter and Chapter 15) is one
of these, and the three-dimensional simple harmonic oscillator is another,
which is discussed further below.
The potentials corresponding to these two problems are approximately
applicable to the dynamics of stellar systems. Far outside the bulk of
the mass, φ ' −GM/r, where M is the total mass. (The potential is
arbitrary up to an additive constant, but in idealised models isolated
from all external matter, it is convenient to choose the constant so that
φ → 0 as r → ∞.) Near the centre, on the other hand, if the potential is
non-singular it can be expanded as
2πGρ0 2
φ ' φ0 + r , (2)
3
where ρ is the density of matter and the subscript 0 refers to the centre.
The coefficient here comes from the theory of the gravitational potential

Box 1. Noether’s Theorem

Suppose a system is described by a Lagrangian L(q, q̇), where q is a

vector of generalised coordinates. Let hs (q) be a one-parameter (param-
eter s) group of transformations such that h0 is the identity. Let the
∂hs
generalised velocities transform under the map hs∗ : q̇ → q̇. If L
∂q
is invariant under these transformations, i.e. L(q, q̇) = L(hs (q), hs∗ (q̇)),
∂L d s
then the quantity h (q)|s=0 is invariant.
∂ q̇ ds
1
For motion in a spherical potential the Lagrangian is L = ṙ2 − φ(r).
2
Let hs be rotation about the z-axis by angle s. Then ṙ is not invariant,
but ṙ2 is, and so is r. For small s, hs (x, y, z) ' (x − ys, y + xs, z), and
∂L d s
so h (q) = (ẋ, ẏ, ż) · (−y, x, 0), i.e. the z-component of angular
∂ q̇ ds
momentum.
Bringing the energy integral within this framework requires some mod-
est extensions (see Arnold 1978a).
 7 Motion in a Central Potential 65

in spherical systems, for which Newton’s Second Theorem (cf. Binney

and Tremaine 1987) asserts that dφ/dr = −GM (r)/r 2, where M (r) is the
mass within an imaginary sphere of radius r. Near the centre, M (r) '
4πρ0 r3 /3, whence Eq.(2) follows immediately.
From the properties of Keplerian and simple harmonic motions, several
useful results follow. For example, comparison of Eq.(2) with the stan-
dard simple harmonic potential ω 2 r2 /2 shows that the period of orbits
near the centre is given approximately by
q
P ' 3π/Gρ0 . (3)

Orbits which remain outside most of the mass will, on the other hand,
resemble a conic section, by Kepler’s First Law. If this is an ellipse of
semi-major axis a, it follows from the theory of Keplerian motion (Box
2) that the period is approximately
q
P ' 4π 2 a3 /(GM ). (4)

This can be written as

q
P ' 3π/Gρ̄, (5)

if ρ̄ is thought of as the mean density in a sphere of radius a containing

the entire mass of the cluster. The resemblance to Eq.(3) follows from
dimensional arguments, but the equality of the coefficients may look like
an undeserved bonus (Problem 1).
To investigate central orbits further we can use plane polar coordinates
r, θ in the plane of motion. The angular momentum is then J = r 2 θ̇, the
1
energy integral becomes E = (ṙ2 + r2 θ̇ 2 ) + φ, and so
2
1 J2
E = (ṙ2 + 2 ) + φ. (6)
2 r
We assume that φ has the form of a finite potential well (Fig.1), and
J2
that φ → 0 as r → ∞. The function φef f = 2 + φ, often called the
2r
“effective potential”, plays an important role in orbit theory. For a given
non-zero value of J, φef f has a single minimum at some radius a > 0,
where
J2
− + φ0 (a) = 0, (7)
a3
and a circular orbit of this radius and angular momentum is possible if
E = φef f (a), since Eq.(6) then implies that ṙ = 0.
66 7 Motion in a Central Potential

Box 2. Formulae for Keplerian Motion

Proofs of the following can be found in numerous texts on dynamics,

e.g. Goldstein (1980). We also like very much the vectorial treatment in
Pollard (1976).
µr
Consider the equation of planar motion r̈ = − 3 (cf. text Eq.(1)).
r
1 2 µ
Then the specific energy is E = ṙ − . If E < 0 then the motion
2 r
is on an ellipse with the origin at one focus. The semi-major axis is
µ 2π
given by E = − and the period by , where the “mean motion”
2a n
is given 2 3
p by n a = µ. The specific angular momentum has magnitude
J = µa(1 − e2 ), where e is the eccentricity of the ellipse. If the orbit
is aligned with pericentre along the positive x-axis,√ then the position
vector at time t is (a(cos ψ − e), b sin ψ), where b = a 1 − e2 is the semi-
minor axis, and the eccentric anomaly ψ is given by Kepler’s equation
nt = ψ − e sin ψ, if the origin of t is taken at the time of pericentre. (We
would prefer E for the eccentric anomaly, but that is already in use.) In
plane polar coordinates the orbit is given by
a(1 − e2 )
r= . (1)
1 + e cos θ
Similar formulae are obtained for hyperbolic motion (E > 0), but are
usually expressed in real form via hyperbolic functions instead of circular
ones, and it is sometimes helpful to redefine a to make it positive, i.e.
Gµ
E= . An important additional formula in the hyperbolic case is for
2a
the angle α between the ingoing and outgoing asymptotes:
√
tan α = µa/J. (2)
Again the case of parabolic motion (E = 0) leads to a different suite
of formulae. A unified treatment of all three types of motion is possible
if specially designed variables are used (Stumpff 1962).
When these equations describe the relative motion of two bodies with
masses m1 and m2 (cf. text), we have µ = G(m1 + m2 ), and the energy
and angular momentum in their barycentric frame are obtained by multi-
plying by the reduced mass m1 m2 /(m1 + m2 ). Thus the energy becomes
1 m1 m2 2 Gm1 m2
E= ṙ − .
2 m1 + m2 m1 + m2
 7 Motion in a Central Potential 67

φ
eff
a r
φ= E

Fig. 1. Radial motion in a spherical potential. See the text for an explanation
of the symbols. The case φef f < E < 0 is illustrated.

For a given angular momentum J, if E slightly exceeds φef f (a) then

motion in a small range of radii near radius a becomes possible. Differ-
entiating Eq.(6), cancelling ṙ, expanding about r = a, and using Eq.(7),
we find that r̈ = −κ2 (r − a) approximately, where κ2 = 3φ0 (a)/a + φ00 .
This describes the radial motion. In addition, conservation of angular
momentum shows that the star circulates at a rate which is nearly con-
stant, except for an oscillation at frequency κ. The composition of these
two motions can be represented as motion on a small ellipse whose centre
circulates round the centre of attraction. Because of the resemblance of
this description to the Ptolemaic model of the solar system, κ is referred
to as the “epicyclic” frequency.
The qualitative features of orbits which are not nearly circular can
be easily deduced from Fig.1. We continue to assume that J 6= 0. If
E sufficiently exceeds φef f (a) that the epicyclic approximation is too
inaccurate, but still E < 0, then Fig.1 shows that motion is confined
between two radii, which we denote by rmin and rmax , and is periodic
in radius. If E ≥ 0 then the star escapes to infinity. In the case of
the Keplerian potential the distinction between orbits of negative and
positive energy is reflected in the transition from bound, elliptic motion
to unbound motion on a hyperbola.
At the opposite extreme, as it were, from epicyclic orbits are nearly
radial orbits. Such orbits, which pass close to the centre of the potential,
are of particular interest for two reasons: first, stars which are ejected
from the dense central parts of a stellar system travel on such orbits, and,
second, they play a crucial role in one of the basic instabilities to which
68 7 Motion in a Central Potential

stellar systems are liable (see, for example, Antonov 1973, Fridman &
Polyachenko 1984, Palmer 1994).
Let us consider orbits of a given energy E and small angular momen-
tum. For J = 0 the orbit is purely radial, and successive maxima in r
are separated (in θ) by an angle exactly equal to π. If J pis small the star
passes close to the origin, and its speed then is nearly 2(E − φ0 ). By
angular momentum conservation it follows that the distance of closest
approach to the origin is nearly proportional to J. It is also easy to see
that the transverse deviation from a purely radial orbit is proportional
to J, and therefore successive maxima in r are separated (in θ) by an-
gles which differ from π by a small angle proportional to J (Fig.2b and
Problem 2).
It is this tiny angle by which nearly radial orbits rotate which allows
them to cooperate with a non-radial perturbation in the potential to give
rise to the “radial orbit instability”, especially in systems where such
orbits are heavily populated. In such a system the pressure tensor is
very anisotropic, and the transverse velocity dispersion is relatively low.
In this sense such systems have a low “transverse temperature”. We al-
ready remarked (Chapter 6) that cold stellar systems tend to clump, by a
process analogous to the Jeans instability, and in much the same way sys-
tems with a low transverse temperature tend to clump into a bar-shaped
configuration (cf. Barnes 1985). A proper understanding of the mecha-
nism by which this happens involves a little more orbit theory, however,
and is deferred to a later remark in this chapter about resonances, and
Box 3.
Because an analysis of general orbits, i.e. those which are neither
nearly radial nor nearly circular, can be reduced to motion in radius (cf.
Fig.1), the results can be expressed as quadratures. For example, the
time between successive maxima in r is
Z rmax Z q
2 dr/ṙ = 2 dr/ 2(E − φ(r)) − J 2 /r2 , (8)
rmin

by Eq.(6). Only a few cases are known where the appropriate integrals can
be expressed in terms of elementary functions. Aside from the familiar
Keplerian and simple harmonic cases, one potential with this property is
GM
the “isochrone” potential φ = − √ , which was so named by
b + b2 + r2
Hénon because it has the property that all orbits of the same energy have
the same radial period (Problem 5).
Let us denote the frequency of radial motion by ωr , and similarly use
ωθ to denote the mean angular velocity θ̇. Thus the angle through which
the star rotates between two successive maxima in r is 2πωθ /ωr . For
 7 Motion in a Central Potential 69

y(t) y(t)
1
2

0.5
1

-2 -1 0 1 2 -1 -0.5 0 0.5 1
x(t) x(t)
-1
-0.5

-2
-1
√
Fig. 2. Two orbits in the scaled Plummer potential φ = −1/ 1 + r2 . They start
at r = 1 on the x-axis, moving transversely upwards with speeds (a) 0.9 and (b)
0.1. Note that the angular separation of successive apocentres, when measured in
the direction of motion, lie between the values for the simple harmonic oscillator
(π) and Kepler motion (2π); see Problem 4.

nearly radial orbits this is nearly π, whence ωr ' 2ωθ for such orbits.
Therefore such orbits exhibit a resonance, which has an important role
in the “radial orbit instability”, as already mentioned. See Box 3.
We also mentioned that stars moving close to the centre ofpthe cluster
exhibit nearly simple harmonic motion with a certain period 3π/(Gρ0 ),
and it might have been thought that such orbits would also resonate.
However the number of such stars is small. Consider circular p orbits, for
example. The period of a circular orbit of radius r is 2π r/φ0 . Now φ
can be expressed as a power series in even powers of r (cf. Eq.(2)), and
so, therefore, can the period. Therefore the period varies approximately
linearly with the energy for low-energy orbits. Now in the most usual
models of star clusters, the volume of phase space available to stars of
energy less than E varies as (E − φ0 )3∗ , and so it follows that the number
of stars of a given period
p increases rapidly as the period increases above
the minimum value 3π/(Gρ0 ). Thus there is no significant population
of resonant stars near this frequency.
These remarks are relevant to the radial orbit instability, but there
are other instabilities which can afflict spherical stellar systems. For
example, though it is easy to build a spherical system with an almost

∗
Consider, for example, a three-dimensional simple harmonic oscillator. The energy
is E = (ṙ2 + ω 2 r2 )/2, and the right side is the square of the radius of an ellipsoid
in 6-dimensional phase space.
70 7 Motion in a Central Potential

Box 3. The radial orbit instability.

Imagine a spherically symmetric system, with a slight aspherical po-

tential perturbation. Does the response of the stars tend to reinforce this
perturbation, or to suppress it?
To capture the essence of the problem we consider a perturbing po-
tential with a plane of symmetry, so that a star initially orbiting in this
plane remains there. With plane polar coordinates r, θ, suppose the per-
turbing potential has a form like φ1 = −a(r) cos(2θ), where a(r) > 0
is some function and the subscript denotes a perturbation. (Note that
we have assumed that φ1 is invariant under rotation by π; the factor 2
in the argument of the cosine is often expressed by stating that we are
examining an “m = 2” perturbation.)
As already stated, there is a near-resonance between radial and az-
imuthal motions for nearly radial orbits, i.e. a nearly vanishing linear
combination of frequencies. If we think of frequency as rate of change
of an angle, it turns out that the corresponding slowly varying angle can
be interpreted as the orientation, ψ, of the long axis of the orbit (Fig.1).
The corresponding “canonical momentum” (in the sense of Hamiltonian
dynamics, cf. Goldstein 1980) is the angular momentum, J. As shown
in Problem 2, in a spherical potential ψ advances in each radial oscil-
lation by a an amount of order J, and so the average rate of change is
ψ̇ ' AJ, for some A > 0, plus a small contribution from the perturbing
potential. It is also evident that J is altered only by the perturbation. If
0 < ψ < π/2 the perturbation pulls the star towards the positive x-axis,
and the angular momentum decreases; the orbit is “pulled back” by the
perturbation. Considering also the other quadrants we see that in general
we may write J˙ ' −B sin 2ψ, for some B > 0.

Fig. 1 A nearly radial orbit in a slightly non-spherical potential.

 7 Motion in a Central Potential 71

Box 3 (continued)
Now let f (J, ψ) be the distribution of these two variables. Since ψ and
J are conjugate, f approximately obeys a Wlasow equation
∂f ∂f ∂f
+ ψ̇ + J˙ =0
∂t ∂ψ ∂J
(cf. Eq.(6.7)). We write f , ψ̇ and J˙ as a sum of a contribution from the
underlying spherical cluster model plus a contribution from the perturba-
tion, e.g. f = f0 (J) + f1 (J, ψ), and neglect all but linear terms involving
the perturbation. Then
∂f1 ∂f1 ∂f0
+ AJ ' B sin(2ψ) . (1)
∂t ∂ψ ∂J
(Notice, by the way, that the analogue of the term which gave us so much
trouble in the study of the Jeans instability (Eq.(6.13)) would have the
∂φ0 ∂f1
form , which vanishes because φ0 is spherically symmetric.)
∂ψ ∂J
The forced solution of Eq.(1) is
B ∂f0
f1 = − cos 2ψ. (2)
2AJ ∂J
If f1 is to be consistent with the assumed perturbation in the potential
then we need an enhancement of orbits along the x-axis (near ψ = 0).
By Eq.(2) we see that this will happen if J∂f0 /∂J < 0. It follows that,
for the radial-orbit instability to proceed, it is necessary that the distri-
bution function should increase as J decreases to zero from either side,
i.e. ∂f0 /∂J is positive (negative) if J is negative (positive). This means
that we expect to see the instability only in systems in which there is a
strong radial anisotropy.
There are some general lessons here for several kinds of instabilities in
stellar systems. In order for an instability to flourish, there must be a
resonance, and there must be a sufficient gradient of f0 across it (see, for
example, Palmer 1994, and the classic Lynden-Bell and Kalnajs 1972).

arbitrary mass pprofile M (r), by setting particles on circular orbits with

velocity v(r) = GM (r)/r, there is no such model which is known to be
stable (cf. Barnes et al. 1986).
Finally we turn to the behaviour of stars in central time-dependent
potentials. This is important in the astrophysical applications of the
million-body problem for two reasons. First, it is possible that stellar
systems breed massive central black holes, and the resulting evolving
72 7 Motion in a Central Potential

potential is an example of the time-dependence we have in mind. A

second example is the slow evolution of a stellar system caused by two-
body relaxation (Chapter 14).
In time-dependent central forces the angular momentum J is still con-
stant. All the interest is in the energy. Since dE/dt = ∂φ/∂t, it follows
that the change during half of one radial oscillation is
Z rmax ∂φ dr
δE = . (9)
rmin ∂t ṙ
The integrand is evaluated at the time when the particle reaches each
radius. If, however, φ changes only slowly (i.e. on a time scale much
longer than the radial orbital period), to good approximation we can
evaluate the right hand side at some fixed time. Multiplying by half the
radial period, we deduce that
Z rmax Z rmax
dr dr
δE = δφ(r) , (10)
rmin ṙ rmin ṙ
where δφ(r) is the change at fixed r in half a radial period. Bringing the
two sides together we now see that
Z rmax δE − δφ
s dr = 0,
rmin J2
2(E − φ) − 2
r
and so
s
Z rmax J2
δ 2(E − φ) − dr = 0. (11)
rmin r2
R
Equation (11) shows that the quantity Ir = ṙdr is invariant (except
for variations that cancel out over one period), provided that the potential
varies on a long time scale. It is, in fact, an example of a standard result
on the adiabatic invariance of action variables, since Ir is simply (except
for a factor π) the radial action (cf. Palmer 1994). We shall make use of
this property of Ir in a statistical description in Chapter 11.

Problems
1) Explain why Eqs.(3) and (5) have the same coefficient.
2) For motion close to the x-axis in the x,y plane of a spherically sym-
metric potential φ(x, y), show that we have ÿ = y(ẍ/x) exactly, and
∂
ẍ = − φ(x, 0) approximately. By considering orbits with initial con-
∂x
dition x = R, y = 0, ẋ = 0 and ẏ = J/R, where R is a constant, show
that the rate of rotation of the orbit is nearly proportional to J.
 7 Motion in a Central Potential 73

3) Consider motion of energy E and small angular momentum J in

a smooth spherically symmetric potential φ(r) such that φ(0) = 0.
Show that successive pericentres are separated by an angle which is
approximately
Z rmax √ p !
J 1 2E − 2(E − φ(r))
π+√ − + p dr ,
2E rmax 0 r2 2(E − φ(r))
where φ(rmax ) = E.
4) (Suggested by a remark of H. Dejonghe.) Assuming that ρ is non-
negative and is non-increasing with r, show that the angle between
successive apocentres lies between π and 2π.
5) Compute the dependence of r and θ on t for an orbit of energy E and
angular momentum J in the isochrone potential.
6) Compute the radial action (cf. Eq.(11)) for the bound Kepler prob-
lem, where φ = −GM/r and E < 0 . Consider a circular Keplerian
orbit. What happens if M is decreased by a factor f (0 < f < 1) (a)
slowly, and (b) instantaneously?
7) In the Kepler problem show that ṙ describes a circle in velocity space.
 8
Some Equilibrium Models

In Chapter 6 we spent some time discussing the statistical description

of an N -body system in terms of its one-particle distribution function,
f . We also introduced an evolution equation for f , the “collisionless”
Boltzmann equation. We did not, however, dwell on any solutions of this
equation, except to characterise them in terms of Jeans’ Theorem. In fact
equilibrium solutions have been known and studied for about 100 years,
and they are of enduring importance.
The specific choice of f may be made for a variety of reasons. Plum-
mer’s model, for instance, is often used for starting a numerical calcula-
tion, because of its analytical convenience. The isothermal model is of
importance in the study of thermodynamic stability (Chapter 17), while
King’s model has taken centre stage for many years in the interpretation
of observations. Another approach to this particular topic, also based on
the phase space description, deserves a section on its own at the end of
the chapter.
By Jeans’ Theorem, f is a collisionless equilibrium solution if (but not
only if) it depends on the energy, E. Most of the models we mention are
of this kind. In what follows we shall usually characterise a model by its
distribution function f (E), but that is only part of the story, because f
depends on the potential φ (via E), and we need to know how φ depends
on the position in space r in order to determine
R f at any point in phase
space. The space density is given by ρ = m f d3 v, and also depends on
φ. Therefore the potential must satisfy Poisson’s equation in the form
∇2 φ = 4πGρ(φ). If this equation can be solved, the end result is what is
called a “self-consistent” model of a stellar system.

74
 The Isothermal Model 75

The Isothermal Model

Let us suppose that
f = f0 exp(−2j 2 E), (1)
where f0 and j are constants. Then the distribution of velocities at any
point is Maxwellian, with one-dimensional dispersion σ 2 given by
σ2 = 1/(2j 2). (2)
This is the distribution that is set up by collisions in simple gases, and
two-body relaxation (Chapter 14) plays a similar role in the million-body
problem.
Computing the density at a point where the potential is φ, we find that
ρ/ρ0 = exp(−m(φ − φ0 )/kT ), (3)
where kT /m = σ2 , and ρ0 and φ0 are the density and potential (respec-
tively) at some point, often taken to be the centre. This equation already
teaches us something important: that the density of more massive stars
decreases with increasing φ more quickly than the density of less massive
stars. This behaviour is a simple example of mass segregation, which we
discuss more dynamically in Chapter 16. Note that it depends on the
implicit assumption that all stars have the same temperature.
Now we turn to Poisson’s equation ∇2 φ = 4πGρ, and introduce the
scaled potential
u = −2j 2 (φ − φ0 ) (4)
and the scaled radius
q
z = r 8πGρ0 j 2 . (5)
This leads to the isothermal equation
d2 u 2 du
+ + eu = 0, (6)
dz 2 z dz
which is a spherically symmetric version of one of Liouville’s equations∗ ,
by the way. Its solutions are thoroughly discussed in Chandrasekhar
(1939). (This book is also well worth looking at for the marvellous isocline
diagrams reproduced from Emden’s Gaskugeln (Emden 1907).)
If we assume as above that φ(0) is the central potential, at the bottom
of the potential well, the appropriate boundary conditions are
u(0) = 0, u0 (0) = 0. (7)

∗
i.e. ∇2 u + exp u = 0
76 8 Some Equilibrium Models

Then the solution is uniquely specified (Fig.1), and so we see that the
family of isothermal models has two free dimensional parameters: the
velocity dispersion σ2 (which is related to j by Eq.(2)), and the cen-
tral density. There is, however, also a limiting case, called the singular
isothermal solution, which has infinite central density (Problem 1). We
missed it by our implicit assumption that ρ0 exists. Finite solutions ap-
proach the singular solution asymptotically at large radius (Problem 2,
and Fig.1).
Though it is important to understand the properties of this model
because of its thermodynamic significance (Chapter 17), its practical ap-
plications are limited. If a model of a stellar system is in dynamic equi-
librium there can be no escaping stars, and so f must vanish above the
energy of escape. Now the isothermal distribution, Eq.(1), never van-
ishes, and this means that there is no escape energy. The reason for this
is that the model turns out to have infinite mass (Problems 1 and 2).

King’s Model

The standard way of improving the behaviour of the isothermal model,

at least in practical terms, is to “lower” the Maxwellian distribution of
Eq.(1) into
(  
f0 exp(−2j 2 E) − exp(−2j 2 E0 ) ifE < E0 ,
f= (8)
0 ifE > E0 ,

0.1

ρ/ρ0 0.01

0.001

1e–4

1e–5

0.1 1 10 100 1000

z
Fig. 1 The isothermal model. (See also Problem 2)
 King’s Model 77

where E0 is the escape energy. This is King’s model (King 1966)∗, and it
has several attractive features, aside from finite mass and radius. Deep
inside the system E  E0 , and so the distribution is nearly Maxwellian
where the system may be expected to be nearly relaxed (since the time
of relaxation is shortest where the density is highest, cf. Chapter 14).
Also, the distribution function Eq.(8) turns out to be a good approxima-
tion to a solution of the Fokker-Planck equation. What is occasionally
forgotten, however, is that this solution is not an equilibrium solution,
but one in which stars escape at a rate governed essentially by the two-
body relaxation time scale. Another pitfall is to imagine that the central
velocity dispersion is still given by Eq.(2), but this is only good in the
limit E0 → ∞.
King’s models are characterised by a third parameter in addition to the
two parameters needed for an isothermal model. This third parameter
is often taken to be the quantity W0 = 2j 2 (E0 − Ec ), where Ec is the
energy of a star at rest at the centre. Fig.2 illustrates the models with
W0 = 3 and 7. It is customary to characterise the size of the central
region by what is called the “core radius”, rc , though different definitions
exist. We adopt the definition given by the physically appealing relation

4πG
ρc rc2 = vc2 , (9)
3

where ρc and vc2 are the central density and mean square velocity, respec-
tively; i.e. vc2 = 3σc2 . For an isothermal model, rc is the radius at which
the space density drops to about one third of its central value, and the
projected (or surface) density falls to about one half of its central value.
This and other characteristic radii are plotted in Fig.3. A commonly
used measure of how centrally concentrated a model is is the quantity c
defined as c = log10 (rt /rc ), although King himself defined c without the
logarithm, and with a different definition of rc .
King’s model has remained a cornerstone of stellar dynamics for over 30
years. It has been extended to include anisotropy (Michie & Bodenheimer
1963), a mass spectrum (e.g. Gunn and Griffin 1979), and an external
gravitational field (Heggie and Ramamani 1995), though these extensions
do not have the same support from the Fokker-Planck equation. In ad-
dition, it quite unclear what is the appropriate theoretical definition of

∗
In another paper (King 1962) he proposed a simple mathematical function which
fits the profile of the surface brightness Σ of globular clusters, and it is sometimes

also referred to as King’s model, unfortunately. The formula is
2
Σ = k [1 + (r/rc )2 ]−1/2 − [1 + (rt /rc )2 ]−1/2 . We propose to call this King’s
Law, by analogy with de Vaucouleurs’ Law.
 78 8 Some Equilibrium Models

0.1

0.01

ρ/ρ0 1e–3

1e–4

1e–5

1e–6 1 10
r/r c

Fig. 2 The King models with scaled potential W0 = 3 (below) and W0 = 7.

10
Radius (N-Body Units)

0.1

0.01
0 2 4 6 8 10 12
W
0
Fig. 3. The tidal radius (solid), half-mass radius (dashed) and the core radius
(dotted) plotted against the scaled central potential W0 for the sequence of King
models, in N -body units (Box 1.1).

the core radius if stellar masses are unequal. Partly for this reason rather
more empirical measures are in common use (Casertano and Hut 1985).
 Plummer’s Model 79

On the theoretical side, substantial work has been done on the modes of
oscillation of King models (e.g. Sobouti 1985). In particular, Weinberg
(1994a) has pointed out how slowly some modes are damped.

Plummer’s Model
Our next exhibit is the world’s favourite theoretical model, simply be-
cause its structure can be written down in terms of very simple functions;
sufficiently simple, in fact, that any theorist can memorise some of the ba-
sic data without difficulty (Table 1). It is actually the stellar-dynamical
analogue of an n = 5 polytrope (Plummer 1911∗ ; Box 1). Less well known
is the fact that it is one of another series of analytical models which in-
cludes the isochrone model (Chapter 7; see Dejonghe 1984). Evolutionary
calculations starting with Plummer initial conditions have been carried
out by many people (e.g. Spitzer & Shull 1975).
The free parameters of the model are the total mass M and the “scale
radius” a. Table 1 gives several quantities which we have not introduced
before. These include the projected mass density, Σ(d), which is the
density per unit surface area of the sky at the distance of the object (so
that, in astrophysical units, it would be measured in solar masses per
square parsec.) The line of sight passes at a distance d from the centre
of the object. Analogous to σ2 , σz2 is the mean square component of the
velocity along the line of sight. The total mass is M , the mass within
a sphere of radius r is M (r), and M (d) is the mass within a cylinder of
radius d parallel to the line of sight.

Models of observational data

We have already mentioned that models such as these are used in the
interpretation of observations, and now we consider just how much of a
restriction the observations of a given stellar system actually place on a
model. Throughout, we suppose that the system is spherically symmetric
and in dynamical equilibrium, so that f can be regarded as a function of
E and J.
Let us suppose first that the observations consist of data on the sur-
face brightness of a stellar system. If it is supposed that “mass follows
light”, this yields (within a constant of proportionality) the projected
mass density, Σ, which can also be computed for any of the above theo-
retical models. The simplest approach which proves reasonably successful
(e.g. Peterson & Reed 1987, Trager et al. 1995) is to choose the three

∗
He attributes the solution to Schuster, however.
80 8 Some Equilibrium Models

Box 1. Plummer’s model

A polytrope of index n is a gaseous model in which the pressure-density

relation is of the form ρ ∝ φn . The n = 5 polytrope has some special
properties. It is the last of the series (in increasing n) with finite mass, and
the first with infinite radius. An examination of the polytropic equation
shows why it is so special, as follows.
If all the physical constants are scaled out from Poisson’s equation,
with the above relation between ρ and φ, the resulting form is y 00 +
(2/x)y 0 + y n = 0, where y is minus the scaled potential and x is the
scaled radius. Now it is easy to see that this equation is invariant under
the symmetry y → λy, x → λ(1−n)/2 x, and this suggests use of the
variable Y = yx−2/(1−n) , which is symmetry-invariant. The resulting
equation is
6 − 2n Y
x2 Y 00 + (xY 0 + ) + Y n = 0,
1−n 1−n
and the form of the derivatives further suggests the transformation from
x to t = ln x. Then the equation takes the transparent form
5−n 6 − 2n
Ÿ + Ẏ + Y + Y n = 0,
1−n (1 − n)2
where a dot means a t-derivative. Now it is clear why n = 5 is so special:
this is then the equation of a non-dissipative oscillator, and so
1 2 1 2 1 6
Ẏ − Y + Y = constant.
2 8 6
Even better, the boundary conditions y(0) finite and y 0 (0) = 0 transform
to the conditions Y → 0 and Ẏ → 0 as t → −∞, and so the constant
is zero. In fact the required solutions emanate from and return to the
equilibrium at Y = 0, and the solution
p which passes through Y = (3/4)1/4
at t = 0 is√easily found to be Y 2 = 3/4secht. This transforms back to
y = 31/4 / 1 + x2 , which is the potential of a Plummer model, suitably
scaled.

free parameters of King’s model (e.g. f0 , j 2 and E0 ) to optimise the

fit of the theoretical model to the observations. Enrico Fermi once said,
“Give me three parameters and I’ll fit an elephant”†, and indeed these
models fit astonishingly well. One reason for this is that the process of

†
Some people say it was four, but the point is the same
 Models of observational data 81

Table 1. Plummer’s Model

Quantity Symbol Expression or value

3.27/2 a2
Distribution function f (−E)7/2
7π G M 4 m
3 5
−5/2
3M r2
Mass density ρ 1 +
4πa3 a2
 −2
M d2
Projected mass density Σ 1 +
πa2 a2
 −3/2
a2
Mass within radius r M (r) M 1+ 2
r
 −1
a2
Mass within projected radius d M (d) M 1+ 2
d
 −1/2
GM r2
(Specific) Potential φ − 1+ 2
a a
1
One-dimensional velocity dispersion σ2 − φ(r)
6  −1/2
3π GM d2
Projected velocity dispersion σz2 1+ 2
64 a a
2
3π GM
Potential energy W −
32 a
W
Kinetic energy T −
2
W
Total energy E
2
a
Core radius rc √
2
16
Virial radius Rv a
3π
a
Half-mass radius Rh √ , ' 1.305a
2 −1
2/3
0.206N a3/2
Half-mass relaxation time trh √
GM ln Λ

turning a distribution function into a surface density is a sequence of in-

tegrations, and these tend to iron out differences between quite disparate
models. (From f we integrate to get ρ(φ), integrate again to obtain φ(r)
and hence ρ(r), and integrate yet again to construct Σ.) Another reason
may be that the profile of a star cluster is a lot simpler than that of an
elephant.
An alternative approach is to attempt to invert this series of integra-
tions more-or-less directly, and in principle this can be done. In other
words, from Σ it is possible to invert the appropriate integral equation
82 8 Some Equilibrium Models

to obtain ρ(r) (Merritt and Tremblay 1994), and then φ(r) comes from
solving Poisson’s equation. This gives ρ(φ), from which it is possible to
construct a distribution function f (E) depending on E alone, i.e. an
isotropic model (Problem 5). In turn this suggests that the anisotropy
is, in a sense, arbitrary, and that a whole sequence of models (differing
in the degree of anisotropy) can be constructed to fit a given observed
distribution Σ. Kinematic information is, therefore, essential to lift this
degeneracy, and provide more unambiguous information on the distribu-
tion f (E, J). Usually this is radial velocity data (of the sort catalogued
in Pryor & Meylan 1993) but useful internal proper motion data exist for
some clusters (e.g. van Leeuwen et al. 2000).
The problem of inverting the integral equations from observational data
rather than from mathematical functions is another issue. Consider, for
example, the problem of determining ρ(r). If information on the distribu-
tion of the stars is given in the form of star counts, or even the positions
of all the stars, the inversion usually magnifies the noisiness of the data.
One way to cope with this is to determine a function ρ(r) which best
fits the data with the constraint that it is not too noisy. In practice, one
represents ρ on a radial grid, writes down a corresponding expression for
Σ at the location of each star, and then maximises the likelihood, with an
additional penalty function which is large when the grid of values of ρ is
too noisy, as judged by the differences in the tabular values (in a suitable
sense.) This is not a purely academic discussion; Fischer & Gebhardt
(1995) have given some interesting applications of this kind of method.
There are merits and disadvantages in both methods of fitting models
to data. Parametric fitting can yield deceptively agreeable fits, while
the results of non-parametric fitting take insufficient account of what is
known on theoretical grounds about the distribution function, and the
result depends on the choice of penalty function, leaving considerable
freedom in the hands of the investigator. Nevertheless the velocity maps
of the rotating cluster ω Cen produced by Merritt et al. (1997) are a
striking demonstration of the power of this approach.

Problems
1) Verify that u = − ln(z 2 /2) is an exact solution of Eq.(6). Write down
the corresponding density, and compute the mass inside radius r, and
the projected density.
2) Transform Eq.(6) to variables v = u+log(z 2 /2) and t = log z. Deduce
that v → 0 as t → ∞, and hence show that all isothermal models have
infinite mass. By linearising about√ the equilibrium v = 0, show that
u(z) ' − ln(z 2 /2) + az −1/2 cos( 7(ln z)/2 + b), where a and b are
constants. The decaying oscillations are visible in Fig.1.
 Models of observational data 83

3) Solve the 1- and 2-dimensional isothermal equations, with initial con-

1 du
ditions (7). (The coefficient of in Eq.(6) is n − 1, where n is
z dz
the number of dimensions.) See Fanelli et al. (2001) for an interesting
evolutionary model of the 1-dimensional case.
4) Find the first few terms of a series solution of Eq.(6), with boundary
conditions (7).
5) If the mass density in phase space is f = f (E), show that the space
density at a point where the potential is φ is
Z ∞q
ρ(φ) = 4π 2(E − φ)f dE.
φ

Differentiate with respect to φ and solve the resulting Abel integral

equation for f .
6) Show that, in the limit W0 → 0, a King model is a polytrope with
index n = 2.5 (Chandrasekhar 1939). Deduce that the limiting radius
in N -body units (Box 1.1) is 12/5. (No galactic globular clusters
are known with very small W0 , and it might be thought that this is
because the depth of the potential well is insufficient to retain stars.
Remember, however, that W0 is a scaled potential. A King model
with low W0 enjoys as perfect a virial balance as any other.)
7) Consider the distribution function f = 2fP if Jz > 0, 0 otherwise,
where Jz is the z-component of angular momentum and fP is the dis-
tribution of Plummer’s model. Verify that this satisfies Jeans’ Theo-
rem (Chapter 6) if the potential is axisymmetric and stationary. Show
that the density is the same as in Plummer’s model. What is the an-
gular momentum?
8) Consider the distribution function

 3
(−2E − J 2 )3/2 if 2E + J 2 < 0
f= 4π 3 Gm

0 otherwise,
where E and J are, respectively, total energy and angular momen-
tum per unit mass. Deduce that the density at a point where the
3 φ3
radius is r and the potential is φ(r) is ρ = − . Deduce that
4πG 1 + r2
Poisson’s√equation is satisfied if φ is the scaled Plummer potential
φ = −1/ 1 + r2 . Show that the mean square radial and transverse
1 1 1 1
velocities are, respectively, √ and . See Dejonghe
4 1 + r2 2 (1 + r2 )3/2
(1987a).
 9
Methods

In stellar dynamics we do not really study stellar systems like globu-

lar clusters and galaxies. That is the job of astronomers. What we do
is study models of these systems. Just as in many branches of applied
mathematics, a model is nothing other than a mathematical structure
into which we try to incorporate our knowledge of the system at hand.
Sometimes “knowledge” is not the right word: it may be nothing bet-
ter than a hunch about how things might work. Often, however, our
knowledge will include physical laws, especially the ones we think are
relevant. So far in this book, for example, we have implicitly thrown out
almost everything we know about stellar systems except the gravitational
dynamics∗ .
In the context of stellar dynamics, a model consists of two kinds of
mathematical construct. One is the mathematical object used in the de-
scription of the system, and the other is the equation which determines
its evolution. So far, for example, we have introduced the N -body model,
where the system is described by N time-dependent vectors, which evolve
according to Newton’s law of motion. We have also introduced a statisti-
cal model of collisionless stellar dynamics, where the system is described
by the one-particle distribution function f , which evolves according to
the collisionless Boltzmann equation†
In this final chapter of Part II we give a foretaste of the full variety
of models for the dynamics of dense stellar systems. It is a bridge be-
tween the essentially simpler dynamics considered up to now and the

∗
R.A. Lyttleton used to say that constructing a mathematical model was a bit like
plucking a chicken, but at the end of the job the bird has still to be airworthy.
Einstein said that everything should be made as simple as possible but not simpler.
†
In the previous chapter we have also used the word “model” in a different but
traditional sense, to denote a specific solution.

84
 The scaling description: evaporative models 85

more colourful variety of dynamical processes on which we focus in the

rest of the book.
The four models we consider (Fig.1) are taken in order of increasing
detail and veracity, but also increasing complexity and difficulty. For any
particular problem at hand, the choice of model is dictated by the usual
balance of considerations: simplicity and veracity. The most important
problems in the dynamics of the million-body problem have been studied
using all four. Indeed it is a measure of the maturity of the subject that
so much can be understood from such a wide variety of viewpoints.

The scaling description: evaporative models

The simplest model of a stellar system characterises it in terms of three
fundamental parameters, which may be chosen to be its mass, its size
and the mean square speed of the stars in it (cf. Box 1.1). Simple as it
is, it can be used to discuss a variety of forms of dynamical evolution.
As an example, let us consider the virial theorem in the form

I¨ = 4E − 2W, (1)

where I is the “moment of inertia”, E is the total energy, and W is the

potential energy (Box 1). This is very often applied to systems in dynamic
equilibrium, when the average value of the left-hand side vanishes, and
so it follows that 4E − 2W = 0 on average. It is also possible, however,
to use the virial theorem in a dynamical way, via the scaling description.
If the system is specified by its total mass M and a characteristic radius
R∗ , and if the assumption is made that the evolution of the distribution
of mass is homologous, i.e. given simply by a change of scale, then there
are “form factors” α and β such that I = αM R2 and

W = −βGM 2 /R. (2)

Substituting in Eq.(1) we see that there is an equilibrium radius R0 =

βGM 2
− , and that small oscillations around this radius have frequency ω
2E
βGM
given by ω 2 = . These are called “virial oscillations” (Fig.2). Note
αR03
that the frequency involves the square root of G times a density, just as
in Chapter 7.

∗
For spherical systems it is common to take for R the half-mass radius, which contains
half the mass. Then for many models (cf Chapter 8) β ' 0.4. Another common
choice is the virial radius, defined by taking β = 1/2.
 86
GRAPE (Makino & Taiji 1998)
direct parallel (Spurzem 1999)
summation
serial (Aarseth 1999)
N−body model
tree code (McMillan & Aarseth 1993)

rotating (Einsel & Spurzem 1999)

isotropic finite difference (Cohn 1979)
A real star cluster Fokker−Planck
model finite difference (Takahashi 1995)
anisotropic
Monte Carlo Serial (Giersz 1998)

9 Methods
Parallel (Joshi et al 2000)
hybrid (Giersz & Spurzem 2000)
Fluid model anisotropic (Louis & Spurzem 1991)
gas

isotropic (Lynden−Bell & Eggleton 1980)

moment (Larson 1970)
Scaling model
evaporative (Chernoff & Shapiro 1987)

Fig.1. Models of the dynamics of dense stellar systems. We give what we judge to be the most informative introductory reference on
each method of some importance in the study of dense stellar systems; not necessarily the first, or most recent; nor do we include a num-
ber of hybrids, or methods used in other areas of stellar dynamics. The picture of M15 was obtained using NASA’s SkyView facility
(http://skyview.gsfc.nasa.gov) located at NASA Goddard Space Flight Center.
 The scaling description: evaporative models 87

It is clear that an N -body system has a tendency to collapse if the

speeds of the stars are too slow (Fig.2.1), and the above formulation of
the virial theorem makes this quantitative. In fact I¨ < 0 if 2T + W < 0,
i.e. if hv2 i < βGM/R, where hv2 i is the mean square speed. Aside from
numerical factors, this is the same as the condition one obtains from
naïve application of the Jeans stability criterion (Chapter 6) to a stellar
medium whose density is estimated from M and R.
Crude though such models might seem, they capture much basic phys-
ics, and the above model can be extended to include rotation, flattening,
external gravitational fields, loss of mass, escape of stars, and even several
aspects of internal evolution driven by relaxation. The phenomenon of
stellar escape was one of the first processes to be discussed using this
model (Ambartsumian 1938, Spitzer 1940). In this case the evolution
equation is different. It is assumed that mass-loss takes place on a two-
body relaxation time scale tr (Chapter 14), and then

dM M
= −µ , (3)
dt tr
Radius

0 1 2 3 4 5 6 7 8 9 10
Time
Fig. 2. Virial oscillations in an N -body model. Initial conditions: Plummer
model (Chapter 8) with N = 16384, and velocities increased to make 2T /W =
−1.4. Units: N -body units (Box 1.1). Each curve is a Lagrangian radius, i.e.
the radius of a sphere containing a fixed mass and centred at the densest part of
the cluster. From the bottom the mass enclosed is 0.02, 0.05, 0.1, 0.2, 0.3, 0.4,
0.5, 0.75, 0.9. Except at large radii, where the time scale is larger, there is no
evidence of more than one complete cycle: the oscillations are heavily damped,
unless two-body relaxation (Chapter 14) is suppressed (cf. David & Theuns
1989). The form factor α (see text) is infinite (logarithmically divergent) for the
Plummer model.
88 9 Methods

Box 1. The Virial Theorem.

This has to be proved from one of the more fundamental models coming
later in this chapter. As an illustration we start with the collisionless
Boltzmann equation
∂f ∂f ∂φ ∂φ
+ vi − = 0.
∂t ∂xi ∂xi ∂vi
If we define Z
I= mf r2 d3 rd3 v,

the integral being taken over all of phase space, then

Z  
∂f ∂φ ∂φ
I˙ = m −vi + r2 d3 rd3 v.
∂xi ∂xi ∂vi
Integration of the second term with respect to vi gives zero, if we assume
that f → 0 at infinity, while integration of the first term by parts with
respect to vi , with the same assumption, gives
Z
I˙ = 2 mr · vf d3 rd3 v.

Differentiating with respect to t again and applying similar tricks leads

to the result Z
I¨ = 2 m(v2 − r · ∇r φ)f d3 rd3 v.

The first of these is evidently four times the kinetic energy, 4T , while
integration with respect to v shows that the second term is
Z Z Z
r · (r − r0 ) 3 3 0
−2 ρ(r)r · ∇φd3 r = −2G ρ(r)ρ(r0 ) d rd r .
|r − r0 |
Interchanging r and r0 shows that either expression is
Z Z
ρ(r)ρ(r0 ) 3 3 0
−G d rd r = 2W.
|r − r0 |
Finally, E = T + W , and so the virial theorem follows:
I¨ = 4E − 2W.
For a system in dynamical equilibrium, I¨ = 0 on average, and so 4E −
2W = 0, or 2T + W = 0. These are common forms in which the virial
theorem is applied.
 The Fluid Description: gas models 89

where µ is a constant. It can be determined in various ways, e.g. by

solving the Fokker-Planck equation (see below) in an idealised poten-
tial well (Spitzer & Härm 1958). Another equation is required for the
evolution of E. For instance, if we suppose that stars escape following
small-angle scattering, as in the theory of relaxation, they do so with
negligible energy, and so we may suppose that E is constant (Problem
14.8). We shall not describe this problem further, but it gives its name
to this whole class of models, which are often referred to loosely as “the
evaporative model”. It is possible to add further properties of the cluster,
such as its concentration, which was defined for a King model in Chapter
8 (see, for example, Prata 1971, Chernoff & Shapiro 1987).

The Fluid Description: gas models

There is a close analogy between a cluster, regarded as a self-gravitating
ball of stars, and a star, modelled as a self-gravitating ball of gas, and
it leads to a more elaborate, but still highly simplified description of the
million-body problem (Hachisu et al. 1978, Lynden-Bell and Eggleton
1980, Box 2). In the spherically symmetric case the structure is de-
scribed by an r-dependent density and temperature, denoted by ρ and
T respectively. Temperature here is used as a surrogate for the velocity
dispersion σ2 , which endows the star-gas with a pressure p = ρσ 2 ; this
equation would be written as p = nkT in the kinetic theory of a per-
fect gas. A more elaborate version of the description allows the pressure
tensor to be anisotropic, corresponding to an anisotropic distribution of
velocities.
Now we illustrate a little of the dynamics that can be associated with
this description. For a stellar system in dynamic equilibrium the appro-
priate equation is
dp GM (r)
=− ρ, (4)
dr r2
where M (r) is the mass within radius r; thus dM/dr = 4πρr 2 . This
equation is familiar from hydrostatics, as −GM (r)/r 2 is simply the grav-
itational acceleration. It can also be regarded as a moment equation
(in the time-independent case), derived from the collisionless Boltzmann
equation by taking the first moment with respect to the distribution of
velocities (Problem 5). In the context of stellar dynamics it is usually
referred to as (one of) “Jeans’ Equations”.
Beyond the condition of local hydrostatic equilibrium, the gas model
can be extended to include a thermal conductivity (cf. Chapter 1) and
the resulting slow evolution from one quasi-equilibrium to another. The
90 9 Methods

0 0.1 0.2 0.3 0.4

Fig. 3. Four models of core collapse and gravothermal oscillations: (top left)
an extension of the evaporative model; (top right) the gaseous model; (bottom
left) the Fokker-Planck model; (bottom right) the N -body model. After Allen
& Heggie (1992)∗ , Heggie & Ramamani (1989), Cohn et al. (1989) and Makino
(1996). The density at the centre of a system is plotted against time. The initial
rise to the first maximum corresponds to the phenomenon of “core collapse”,
while the subsequent variations are referred to as “gravothermal oscillations”.

∗ ∗
With reference to Eqs.(17) in that paper, the model parameters used were T10 =
4.186, ξT = 0.00037 and Γ = 1.28, and the initial values were 3.437346, 3.442962
and 1.258366.
 The Fluid Description: gas models 91

Box 2. The Equations of Stellar Structure and Cluster Evolution.

In the simplest case of radiative transfer of energy in a spherical star

consisting of a perfect gas in mechanical equilibrium, the equations are
∂M
= 4πρr2
∂r
∂p GM (r)
=− ρ
∂r r2
∂L DS
= −4πρr2 (T − ) (1)
∂r Dt
∂T 3 κρ L(r)
=− ,
∂r 4ac T 3 4πr2
(e.g. Clayton 1983, Chapter 6), where most of the variables have the same
meaning as in the text, except that S is the specific entropy (∝ ln(T 3/2 /p)
for a perfect gas),  is the specific rate of energy generation, a is the
Stefan-Boltzmann constant, c the speed of light, and κ the opacity.
In the even simpler case of a star cluster, the equations are (Lynden-
Bell & Eggleton 1980)
∂M
= 4πρr2
∂r
∂p GM (r)
=− ρ
∂r r2
∂L DS
= −4πρr2 (T − )
∂r Dt
∂v2 1 v L(r)
=− ,
∂r 3GmC ln Λ ρ 4πr2
where v is the root mean square one-dimensional speed, m is the individ-
ual stellar mass, C is a numerical constant determined by the theory of
relaxation (Chapter 14, which also explains the Coulomb logarithm ln Λ),
S = ln(v3 /ρ), and  is the rate of generation of energy in three-body in-
teractions (Eq.(23.5)).

intention is to model the effects of two-body relaxation (Chapter 14),

and so the thermal conductivity must be chosen to ensure that transport
of thermal energy takes place on the relaxation time scale. (In particu-
lar, the conductivity appropriate to a non-magnetised plasma yields the
wrong timescale.)
92 9 Methods

Box 3. Gas and Moment Equations.

The difference between the gas equations, which incorporate relaxation

as thermal conduction, and the moment equations of the Fokker-Planck
equation, can be illustrated by the following discussion, in the context
of spherical symmetry. Both models involve a heat transport equation
which can be written in the form
∂L D σ3
= −4πρr2 σ2 ln , (1)
∂r Dt ρ
where D/Dt denotes a “convective” rate of change (i.e. following the
radial motion of the material), and L is the conductive flux, relative to
the material, of thermal energy across a sphere of radius r.
In the case of the gas model this is nothing other than the first law
of thermodynamics in the form dQ = T dS, where dQ is an increment
of thermal energy, and T and S are the temperature and entropy. For a
spherical shell with radii r and r + dr, we have dQ = (L(r) − L(r + dr))dt,
where dt is a short time interval. Similarly, using the expression for the
specific entropy of a perfect gas, the entropy of the material in this shell is
ρ σ3
4π r2 dr × k ln , where k is Boltzmann’s constant and m is the stellar
m ρ
mass. Then the gas model takes the flux per unit area, L/(4πr 2 ), to
be proportional the radial gradient of the temperature, or of the mean
square velocity dispersion in a single component, σ 2 .

The resulting model has been very successful in developing our under-
standing of the million-body problem. The phenomenon of core collapse
(Chapter 18) was first understood in detail using models of this gen-
eral kind. Even more impressively, it was with these models that the
phenomenon of gravothermal oscillations (Chapter 28) was actually dis-
covered. Fig.3 shows this behaviour modelled with the four different
methods discussed in this chapter.
Generally speaking the gas model produces results which compare very
well with those produced by other techniques (Aarseth & Lecar 1975),
though the reason for this success has remained obscure. For example,
it has not been demonstrated that all the equations may be obtained by
taking moments of the Boltzmann equation. If one does take moments,
however (Larson 1970, Louis 1990) it is possible to obtain a set of equa-
tions which are rather similar in complexity and effectiveness to the gas
equations, provided that a suitable closure assumption is made, but they
 The Fluid Description: gas models 93

Box 3 (continued)
In the case of the moment method, Eq.(1) is obtained from the Fokker-
Planck equation. For the present purpose this is best written in the form
 
∂f ∂f ∂φ ∂f ∂f
+ vi − = ,
∂t ∂xi ∂xi ∂vi ∂t enc
where the left-side is as in the collisionless Boltzmann equation, and the
right-hand side is the “encounter term”, the form of which is discussed
in Chapter 14. (Briefly, encounters alter the velocities of participating
stars, and this changes f at a rate described by this term.) Next, take
the moment with respect
R to v of m(v − hvi)2 , where the mean velocity is
defined by ρhvi = f vd3 v. The moment of the right-hand side vanishes,
as encounters do not change the total mass, momentum or kinetic energy
of stars in a small spatial volume. Computing the moment of the left-hand
side requires some care. It is also necessary to use the zeroth moment,
∂ρ
or continuity equation, i.e. + ∇ · (ρhvi) = 0, and the sort of tricks
∂t
such as integration by parts that were adopted in the derivation of the
virial theorem (Box 1). If it is assumed that the distribution function is
isotropic in a frame moving with velocity hvi, the result is exactly Eq.(1),
with an explicit form of L, i.e.
1
L = 4πr2 ρhr̂.(v − hvi)(v − hvi)2 i. (2)
2
The term on the right, which involves a third moment of the velocity
distribution, is clearly the flux of kinetic energy (in the rest frame of the
material) crossing a sphere of radius r.
The equation looks the same, but the essential difference with the gas
model is that, in the moment method, another equation must be written
down for the evolution of the third moment. This in turn involves a
fourth moment, whose evolution involves a fifth moment, and it is usual
at this point to close the chain by assuming that this fifth moment can
be expressed in terms of lower moments.

differ in detail (Box 3). One reason for this is that the mean free path
in a stellar system is very long (compared to the typical size of the sys-
tem; cf. Problem 14.5), whereas in an actual gas the mean free path is
short compared to the length scale of the system. Perhaps the success
of the gas equations may be explained if we suppose that so much of
the behaviour of stellar systems depends more on the thermodynamics of
94 9 Methods

self-gravitating spheres than on the details of the mechanism of energy

transport.
The foregoing discussion of the dynamics of the gas model has concen-
trated on the evolution due to relaxation processes. Stellar systems also
evolve on the much shorter time scale of the orbital motions, but the gas
model is not usually adopted for the study of such problems except in
one sense. The dynamical stability of stellar systems is a difficult topic in
general, but it is possible to show (under mild and reasonable hypotheses)
that a stellar system is stable if the corresponding gaseous model is, and
the latter question is often a good deal simpler (see Binney and Tremaine
1987 for an introduction to this area).

The phase space description

Boltzmann and Fokker-Planck equations

This approach has been introduced in Chapter 6, where we also described
one of the appropriate dynamical equations – the collisionless Boltzmann
equation. Like the gas equations, this can be extended to take account of
the effects of two-body relaxation, and then the result is either the Boltz-
mann equation or the Fokker-Planck equation. The essential difference
between the two models is whether or not one may neglect the occasional
close two-body encounter in comparison with the cumulative effect of the
more numerous distant encounters. Useful formulations including strong
encounters have been developed (e.g. Kaliberda & Petrovskaya 1970),
but it is the Fokker-Planck equation which is much the commoner in
stellar dynamics, even though the relative errors in the Fokker-Planck
description are no smaller than order 1/ log N (Chapter 14). Where it
has been compared with evolution according to the Boltzmann equation,
however (e.g. Goodman 1983), it has been found to yield very similar
results. The reason may, again, be the same as that mentioned in the
penultimate paragraph of the previous section.
The Fokker-Planck model is generally regarded as the most accurate
model of a stellar system, short of the N -body model, and it is discussed
in more detail in Chapter 14. Its general structure is illustrated by the
form it takes for a spherical system in quasi-dynamic equilibrium, with
an isotropic distribution of velocities. Here it is assumed that the distri-
bution function f depends only on E and t, and then the Fokker-Planck
equation is
∂f ∂s ∂f ∂s ∂ ∂2
− = (DE f ) + (DEE f ), (5)
∂t ∂E ∂E ∂t ∂E ∂E
 The N -body description: simulation 95

where DE and DEE are diffusion coefficients, representing aspects of the

rate of change of energy of stars in consequence of numerous small-angle
scatterings off other stars, and s is the volume in phase space inside the
hypersurface with energy E, i.e. the surface v2 /2 + φ(r, t) = E. The
second term on the left side represents the change in f resulting from
the fact that the energy of a star changes in consequence of the time-
dependence of the potential.
There are several other forms of the Fokker-Planck equation, e.g. in
the case where f is allowed to depend also on the angular momentum
(Chapter 14). There are also different ways of solving these equations
numerically, either using some flavour of a Monte Carlo approach, or
finite differences (Fig.1). The former has enough flexibility that it is
somewhat easier to add more complicated dynamical processes such as
physical collisions (Freitag & Benz 2001). In recent years the increasing
attention paid to N -body models has been at the expense of Fokker-
Planck models. And yet for the time being they remain the best source
of evolutionary models of individual globular clusters (e.g. Drukier 1995).

The N -body description: simulation

This, the most detailed model, takes as its data the masses and time-
dependent positions of the N stars. The dynamical equations are usually
Newton’s equations in the form
X Gmj (ri − rj )
r̈i = − . (6)
j6=i
|ri − rj |3

In some problems the right-hand side is approximated by the gradient

of a smooth potential (cf. Chapter 6). This is appropriate in the study of
collisionless phenomena, e.g. dynamical stability. In this context a sys-
tem may be simulated by an N -body model with a much smaller number
of particles, provided that some precautions
q are taken. A very common
trick is to replace |ri − rj | in Eq.(5) by |ri − rj |2 + 2 , where  is a cun-
ningly chosen smoothing parameter. The effect of this is to circumvent
the numerical errors which occur in close encounters of particles. The
error made by using the wrong force is smaller than that due to using
a small number of simulation particles to represent the behaviour of a
much larger system, provided that ε is not too large (cf. Hernquist et al.
1993).
Another possibility is to use a tree code. Though one might be con-
cerned whether the important process of two-body relaxation (Chapter
14) is correctly modelled in such codes, practical experience appears to
confirm their reliability (McMillan & Aarseth 1993, Arabadjis & Rich-
stone 1998). They may well be competitive for N > 4
∼ 10 (McMillan &
96 9 Methods

Aarseth 1993), though this depends on the formulation (cf. Jernigan &
Porter 1989).
For an honest solution of an N -body problem, however, the problems
are of a different order. As is evident from Problem 3.2, it is all too
easy to code these equations inefficiently, and the present state of the art
represents about 40 years of sustained improvement (Fig.3.1). Different
particles must be integrated with different time steps (or a hierarchy of
timesteps). Depending on the hardware, the force of distant particles
can be summed less often than that from near neighbours (Ahmad &
Cohen 1973). The long-range force itself is readily evaluated nowadays
using special-purpose hardware (GRAPE). Special care is still needed to
deal with the singularities of Eqs.(6) (cf. Chapter 15). Long-lived triple
systems pose another batch of awkward problems. And all this before
one has even figured out how to analyse the data (cf. Casertano & Hut
1985, Eisenstein & Hut 1990), or considered adding the effects of stellar
evolution to the problem.
For some purposes there is no substitute for an honest N -body simu-
lation. One shouldn’t forget, however, that for a wide range of simpler
and yet important problems, more approximate methods such as Fokker-
Planck techniques may be quite adequate, and certainly much faster. For
such problems all techniques give quite similar results (e.g. Aarseth et
al. 1974, Giersz & Spurzem 1994, Takahashi & Portegies Zwart 1998).

Problems

1) Use Eq.(4) to establish the virial theorem for a system enclosed in a

sphere of radius re , in the form 2T + W = 4πre3 pe , where pe is the
pressure at the boundary.
X
N
2) Defining I = mi ri2 , use Eq.(6) to establish the virial theorem.
i=1
Repeat the exercise for two-dimensional gravity, in which the potential
between two rods is 2Gmi mj ln |ri − rj |.
3) Use an N -body code in the Appendix to look for virial oscillations.
In the printed code the condition of virial balance is satisfied initially,
and so one possibility is to multiply all initial velocities by a constant
factor. The current virial radius, Rv , can be computed most easily
from the current kinetic energy and the initial conditions, since W =
−GM 2 /(2Rv ) = W0 + T0 − T . If only the initial velocities are altered
in the code, G = M = 1 and W0 = −1/2.
 The N -body description: simulation 97

4) Compute the half-mass and virial radii for Plummer’s model, i.e. a
model whose density is
ρ0
ρ= !5/2 ,
r2
1+ 2
a
where ρ0 is the central density and a is a constant.
5) Starting from the collisionless Boltzmann equation, either (a) fill out
the details of the calculation of Eqs.(1) and (2) in Box.3, or (b) derive
Eq.(4).
6) Estimate the conductivity in gas models by equating the time scale of
thermal energy transport with the relaxation time (Eq.(14.8)). Check
your estimate against the model of Lynden-Bell & Eggleton (1980),
who give L = −12πGmC ln Λr2 (ρ/σ)(∂σ2 /∂r), where C is a constant.
7) Suppose that the potential is dominated by a massive central black
hole, and that the density (ρ) and velocity dispersion are power laws in
r. Estimate the energy of stars between r and 2r, and the relaxation
time. If the luminosity L is constant (independent of radius) show that
ρ ∝ r−7/4 . (For a thorough treatment of this problem see Bahcall &
Wolf 1976.)
 PART III. MEAN FIELD
DYNAMICS: N = 106

We continue the emphasis on collective effects, i.e. those in which individ-

ual interactions between stars are of no importance, but we increasingly
focus on those effects that really matter in the million-body problem.
Chapter 10 opens with a brief discussion of the notions of equilibrium and
stability in this context, but is largely concerned with non-equilibrium
phenomena: phase mixing and “violent” relaxation. Another mechanism
for evolution of the distribution function, even in static potentials, is
diffusion by chaotic motions.
Chapter 11 introduces a variant with a strong astrophysical motiva-
tion: the behaviour of N -body systems consisting of particles with time-
dependent masses, and how this affects the energy and spatial scale of the
system. Much depends on whether the variation is rapid or slow, and in
the latter case we can easily study its effect on the distribution function
itself.
Again motivated by the astrophysical setting, Chapter 12 introduces
the effect of a steady external potential. The problem closely resembles an
important idealised version of the motion of the moon around the earth
under the external perturbing effect of the sun (Hill’s problem). We study
the non-integrable motions in this potential, and the important problem
of escape. The study is then extended to the case, even more important
in applications, of an unsteady external potential.

98
 10
Violent Relaxation

Equilibrium and Stability

Mathematicians classify equilibria in various ways. There are, for ex-
ample, unstable equilibria, which are rarely found in nature, but are
important in the theoretical understanding of a complicated dynamical
system. Of greater practical importance are stable equilibria. The defi-
nition of this concept amounts to saying that, if the system is disturbed
slightly from the equilibrium, then it remains in the vicinity of the equi-
librium. In nature, however, stable equilibria often exhibit a still stronger
behaviour, which mathematicians classify as asymptotic stability. This
means that the disturbed system returns to the equilibrium state from
which it was disturbed. This happens commonly in nature because of
dissipative forces. The process of returning to equilibrium is often re-
ferred to as relaxation, and it is one with which we are all familiar (late
at night).
With this background it is astonishing that relaxation plays such a
central role in stellar dynamics. Not only is there no dissipation in the
gravitational many-body problem, there is no equilibrium either. It is
true that one can think of some highly artificial solutions which can be
regarded as equilibria. The Euler-Lagrange solutions of the three-body
problem, in which the three stars appear to be at rest in a uniformly
rotating reference frame, come into this class, and, from a more general
point of view it may be fruitful to regard a periodic solution as a gen-
eralised equilibrium. But even where these solutions are stable, there is
no question of asymptotic stability. It is true that, if an equilibrium is
unstable, then there will be some special solutions which tend asymptoti-
cally to the equilibrium as t → ∞; the mathematical definition, however,
requires that the equilibrium be stable and that all solutions starting in
its neighbourhood should tend towards it.
The concepts of equilibrium and relaxation begin to take on meaning
when we replace an N -body system, described in terms of the position

99
100 10 Violent Relaxation

and velocity of the N stars, by a distribution function, f (Chapter 6).

The dynamics of the N -body problem is then replaced by an appropriate
evolution equation for f , such as the collisionless Boltzmann equation.
Then the concept of equilibrium refers simply to a solution which is in-
dependent of time, t. A “corresponding” N -body system (if we can refer
to such a thing without explaining precisely what is meant) is not in
equilibrium: the stars will be in constant motion, and we say that the
system is in “dynamic equilibrium”. Many such equilibria are known,
even though, in the stellar-dynamic case, the problem of constructing
self-consistent equilibria (Chapter 8) is non-trivial in general.
Observe that what we mean by “equilibrium” cannot be dissociated
from the way in which we model the system and its dynamics. The same
will be true of the term “relaxation”, since this refers to the process by
which a system evolves towards equilibrium. But, though we have now
established a framework in which the concept of equilibrium is meaning-
ful, we are no closer to seeing how a non-dissipative system can exhibit
anything like relaxation.

Phase Mixing
As a first concrete illustration, let us consider a system which is described
by the collisionless Boltzmann equation (Chapter 6). In order to simplify
the discussion to its utmost, let us consider instead the corresponding
one-dimensional problem in a time-independent potential φ(x), i.e.
∂f ∂f ∂φ ∂f
+v − = 0, (1)
∂t ∂x ∂x ∂v
where x is the position, v is the velocity. Now the motion of a particle in
a one-dimensional, time-independent potential is integrable, and instead
of using coordinates x and v in phase space, we may be better to use
action-angle variables, J and θ (e.g. Goldstein 1980). These are canonical
variables such that J is constant and θ (the “phase angle”) increases
linearly with time, its rate of change being some J-dependent frequency
ω(J). Similarly, the distribution function may be written as f (θ, J, t), and
then the collisionless Boltzmann equation, Eq.(1), takes the very simple
form
∂f ∂f
+ ω(J) = 0. (2)
∂t ∂θ
(Note that Eqs.(1) and (2) are different forms of the general equation
f˙ + [H, f ] = 0, i.e. Eq.(6.6), where square brackets denote the usual
Poisson bracket and H is the Hamiltonian. In the variables of Eq.(1) this
is H = v2 /2 + φ, while ω = dH(J)/dJ if action-angle variables are used.)
 Phase Mixing 101

The equilibrium solutions of Eq.(2) are easy to describe: since ∂f /∂t =

0, Eq.(2) implies that ∂f /∂φ = 0 also, and so they are functions of J
alone; but how do other solutions relax to a solution of this form? If the
initial distribution (at t = 0) is f0 (θ, J), the solution of Eq.(2) is
f (θ, J, t) = f0 (θ − ωt, J). (3)
This never approaches an equilibrium solution (pointwise), unless f0 is
independent of θ, i.e. unless the system is already in equilibrium! At first
sight, then, the very idea of relaxation seems irrelevant.
The top three frames of Fig.1 illustrate the same points. They show
several snapshots of the θ, J plane if the initial distribution f0 is 1 (white)
for π < θ < 2π, and 0 (black) otherwise. The dynamics (in which J is
fixed) causes a horizontal motion of the boundary of the region in which f
is non-zero. As it moves off the right-hand side of the diagram at θ = 2π it
reappears on the left, as θ is just a phase. The rate of motion depends on
J (assuming that the potential is not simple harmonic), and so the region
where f 6= 0 becomes increasingly stretched. It also becomes increasingly
narrow, and the detail on this diagram reaches ever finer scales.
Let us recall that this diagram is meant to represent the evolution of a
system with a large but finite system of particles (like a million). Suppose

00 π 2π 0 π 2π 0 π 2π
ϑ ϑ ϑ
Fig. 1. Phase mixing. Here ω = J. The top three frames give the initial
conditions, as described in the text, and the solution at times t = 2 and 20. The
lower three frames give the same information from 100 initial points.
102 10 Violent Relaxation

initially that all particles are in the region where f0 6= 0, and that we
follow their motion in this diagram. At subsequent times the particles
will still occupy the region where f 6= 0. Eventually, however, the fine
detail which is visible in the top row of Fig.1 would not be discernible
in the distribution of points; certainly this would be true by the time
that the filaments are so fine that most do not contain any point. The
lower three frames of Fig.1 illustrate this. In the last frame the points
are confined to 10 or 11 nearly horizontal strips, as in the frame above,
but this is not evident unless one looks carefully for it (e.g. by looking
at the figure along the page). By this stage we see that the distribution
of points does not change perceptibly, even though the points themselves
are in rapid motion. There is a sense in which equilibrium is approached,
when we think in terms of particles rather than the distribution function;
it is a statistical equilibrium.
The same idea can be reexpressed in terms of the distribution function.
The distribution function is used to tell us how many particles there are
Rin regions of phase space. Let R be such a region. When we compute
R f dθdJ the fine filaments in f do not matter; we would obtain almost
the same answer if f were smoothed over little regions of phase space
which are small compared with R but large compared with the thickness
of the filaments.
It is not difficult to see that, after some time, this smoothed distri-
bution, which is often referred to as a “coarse-grained distribution” in
statistical mechanics, is given by averaging f0 with respect to θ. In other
words, in this smoothed or coarse-grained sense, f does indeed relax to
hf0 iθ . In fact such an equilibrium solution is asymptotically stable, in
the coarse-grained sense (Problem 1). And it is important to understand
the evolution of the coarse-grained density. When one attempts to infer
a phase density by looking at a galaxy with a telescope, or at an N -body
system with a computer, the best one can hope to do is to estimate a
coarse-grained density.
Curiously, the process we have discussed is not often referred to as
relaxation, but as “mixing”. In fact this word, like stability, is one which
is overloaded with meanings, only one or two of which we touch on in
this chapter. The particular type of mixing we have introduced is of a
relatively mild kind, more like setting the tea in a cup in steady circulation
(with a careful circular motion of the spoon) than churning it up with a
vigorous stir. In the dynamical context it is referred to as “phase mixing”,
for good reason. Consider two particles with neighbouring values of J and
the same phase θ at some time t. The original phases of these particles,
i.e. at t = 0, were θ − ω(J)t. Now ω depends on J in general, and so
if t is large the original phases differed greatly, even though the particles
now occupy neighbouring positions in phase space. The original phases
 The End Point of Violent Relaxation 103

have become mixed together, by the same mechanism which led to the
stretched, narrow filaments in Fig.1.
Notice that f in Eq.(3) depends on two integrals of motion (i.e. J and
θ − ωt, which illustrates Jeans’ Theorem (Chapter 6). The two integrals
have rather different character, however. One (the first argument of f )
is phase-dependent, and irrelevant for smooth solutions (at late times, in
this case). It is an example of a non-isolating integral, because curves on
which it is constant lie extremely close together (at late times); they do
not “isolate” gross regions of phase space from each other. In all these
respects it is quite different from J, which is an example of an isolating
integral.

The End Point of Violent Relaxation

The gravitational million-body problem differs in two essentials from the
simpler problem we have considered so far: it is three-dimensional, and
the potential is not fixed. Fig.2.1 gives an impression of what we have in
mind.
Let us consider the second point first. One consequence of the fixed
potential in the one-dimensional case is that the energy of each particle,
and hence the value of J, is fixed. The distribution of the energies of
the particles cannot vary. In an N -body system, however, the potential
depends, through Poisson’s equation, on the evolving spatial distribution
of the particles, and therefore the distribution of energies may be expected
to evolve. The process by which a system reaches equilibrium when the
potential is time-dependent is referred to as “violent relaxation”. The
picture to have in mind is that of a seething, pulsing distribution of stars,
with swarms of stars (but not the stars themselves) crashing into and
passing through each other (Fig.2).
There has been much discussion in the literature about what the end
result of this evolution may be. Certain rather general arguments can be
given, however. As mixing takes place, points where f is high become
closely interwoven with points at which f is low. The result is that the
coarse-grained distribution, which we denote by F , is a weighted average
of high and low values of f ; in this sense we write F = hf i. It follows that,
for any convex function C, C(F ) ≤ hC(f )i∗. By summing this result over
the whole of phase space, it is not hard to arrive at a result which looks
like an “H-theorem” (e.g. Reichl 1980): suppose that we equate f and F
at some time; then it may be shown that, at any later time, the value of

∗
As an example, let C(f ) = f 2 (a convex function), and suppose that f takes two
values f1 and f2 with equal probability. Then F = (f1 +f2 )/2, hC(f )i = (f12 +f22 )/2,
and C(F ) = hC(f )i − (f1 − f2 )2 /4
104 10 Violent Relaxation
R
C(F )dτ is less than (or at least never exceeds) its initial value, where
dτ is the element of phase volume (Tolman 1938, Tremaine et al. 1986).
Unfortunately, this is not quite the same thing as proving an H-theorem,
which states that there is a function H which decreases with t (Dejonghe
1987b, Soker 1996).
Unfortunately there are other reasons why this does not get us very
far: there is no compelling reason to prefer one form for C over any other,
and so we are no nearer finding the function to which F should tend. But
we can deploy a further argument. The evolution is constrained by the
fact that, as we follow the flow in phase space, the value of f around us
does not change. (Remember that this is what Eq.(1) means; the phase
fluid is incompressible (Chapter 6).) Therefore particles cannot crowd
together more closely than at time t = 0; they obey a kind of classical
exclusion principle. It is not too surprising, therefore, that the most
probable distribution of particles subject to this constraint is one which
resembles Fermi-Dirac statistics (Lynden-Bell 1967; see also Shu 1978).
Sadly, numerical experiments give at best ambiguous support to this
idea (e.g. Cuperman & Harten 1972), and it is not hard to see one
possible reason for this. We are relying on fluctuations in the potential
to drive the evolution. But the potential depends on the density, which is
an integral over the distribution function. We have seen, in the context of
phase mixing, that such integrals settle down to the integral of a steady,
coarse-grained distribution. Therefore one of the effects of the evolution
is to quench the very potential fluctuations which drive the relaxation.
It is therefore not surprising that the relaxation fails to complete its task
of bringing F to the expected equilibrium form.
None of this theory really vividly brings to mind the picture that as-
trophysicists see when “violent relaxation” is mentioned. They think of
galaxies colliding in what is little short of an explosion (Fig.2). The
galaxies are shattered, and some fragments may even be ejected from the
interaction region fast enough to escape. If this is so there will be almost
as many fragments with small positive energies as there are with small
negative energies. (The escape energy is a quantity set by conditions far
away from the place where the energies of the particles are being vio-
lently relaxed, and the relaxation process does not know about it.) In
other words, the number of stars per unit energy and angular momentum,
N (E, J), is almost independent of E near E = 0, at least for those stars
whose angular momenta are small enough for them to have come from
the ejection region. (If vesc is the escape speed from the edge of the in-
teraction region and R is its radius, the condition is J < Jmax = Rvesc .)
Stars with E ≥ 0 escape, whereas the remainder settle into dynamical
 The End Point of Violent Relaxation 105

Fig. 2. Two colliding galaxies (from Mihos & Hernquist 1996). The time (in
units of approximately 1.3 Myr) is given at top right of each frame.

equilibrium. We now set ourselves the task of determining their spatial

distribution (von Hoerner 1957; this long- forgotten result was rediscov-
ered decades later by Makino et al. 1990).
The orbits of these particles lie outside most of the mass, and are there-
fore nearly Keplerian, and so their period is P ∝ (−E)−3/2 when E is
slightly negative. By Eq.(6.8) it follows that the phase space distribu-
tion is f (E, J) ∝ E 3/2 /J when J < Jmax . By integrating over velocities
Z
(−E)3/2 v
we deduce that the density is ρ ∝ dvdθ, where θ is the an-
r
gle between v and r, and integration is restricted to the domain where
J = rv sin θ < Jmax and v2 < −2φ. When r is large the first condition
reduces approximately to θ < Jmax /(rv), whence ρ ∝ (−φ)2 Jmax /r2 ,
∝ r−4 . This result is borne out approximately in numerical experiments
(e.g. van Albada 1982).
106 10 Violent Relaxation

Motion in Nonintegrable Potentials

Let us return once more to the problem of evolution in a fixed potential,

but step up from the one-dimensional problem, where the only thing
that happens is phase mixing, to a more realistic number of dimensions.
In fact, two dimensions will do, as we shall then meet almost all the
points we need consider. Nor is it is such an unrealistic problem to
tackle, a non-trivial example being motion in the meridional plane of an
axisymmetric three-dimensional potential. Furthermore, it was precisely
in this astronomical context that the seminal paper of Hénon & Heiles
(1964) was set.
Let x, y be coordinates in this plane. Both coordinates oscillate, but in
general the oscillations couple∗. Poincaré gave a compact way of picturing
the corresponding motion of a particle in phase space. Every time y = 0,
plot the phase coordinates of the x-oscillation. We shall suppose that
we use the action and phase for this purpose. Then, if there were no
coupling between the two motions, our discussion of Fig.1 shows us what
to expect: the successive points representing one particle lie on horizontal
lines (Fig.3a).
Unless we are told the answer, it is much harder to guess what happens
when the coupling is non-zero. In fact, if the coupling is small, the picture
does not change much: successive points (representing the motion of one
particle) still lie on curves in the plane (Fig.3b). The curves are no
longer quite straight, and also some of them distort into ovals instead of
stretching all the way across the diagram (Problem 3). The latter are
associated with resonances between the x- and y-motions.
When the coupling is larger still, more dramatic changes become ap-
parent (Fig.3c). For some particles, the points still lie on curves or ovals.
In that case we talk of “regular motion”. For other particles, however,
the points are scattered on regions of the plane, and we speak of “chaotic
motion”. One orbit in Fig.3c has this character. Incidentally, the differ-
ence between Figs.3b and 3c is not as great as might be thought: it is
just that, when the coupling is small (Fig.3b), the chaotic regions are so
slender as to be indistinguishable from curves.
Let us look (metaphorically) at one of these chaotic regions, where
points appear to be scattered at random, and not on a curve. Instead of

∗ ∂φ
In cylindrical polar coordinates the equations are r̈ − J 2 /r3 = − (r, z),
∂r
∂φ
z̈ = − (r, z). Unless z = 0 the oscillation in z influences the oscillation in r, and
∂z
vice versa.
 Motion in Nonintegrable Potentials 107

1.8
1.6
1.4
1.2
1
J

0.8
0.6
0.4
0.2
0 1 2 3ϑ 4 5 6
1.8
1.6
1.4
1.2
1
J

0.8
0.6
0.4
0.2
0 1 2 3ϑ 4 5 6
1.8
1.6
1.4
1.2
1
J

0.8
0.6
0.4
0.2
0 1 2 3ϑ 4 5 6

Fig. 3. Surface of section for two coupled oscillators in 2 dimensions. Details

of the dynamics are given in Problem 3. For a problem better motivated by the
million-body problem, see Fig.12.2.
108 10 Violent Relaxation

studying the motion of a single particle, let us consider what happens if, at
time t = 0, we consider a set of particles occupying a little patch in phase
space, much as in our consideration of Fig.1. Again this patch becomes
stretched out as the evolution proceeds, but the rate at which it does so is
very different. In Fig.1, the stretching is caused by the linear increase of
the phase with time. In the corresponding motion in Fig.3c, however, the
factor by which the patch of phase is stretched increases approximately
exponentially with time† . There is mixing, and it is very different from
phase mixing, but it leads to the same end result: in the region where
this kind of behaviour occurs, the coarse-grained distribution function F
will approach an equilibrium form.
Different chaotic zones may well be separated by curves. In this case, a
particle which starts its evolution in one zone is forever separated from the
other. The consequence of this is that the equilibrium value of F in one
region may be quite different from that in another. We should picture a
coarse-grained distribution function which varies smoothly across regions
of regular motion, rather as in the one-dimensional case, but which is flat
in each chaotic zone.
We have spoken of the curves between neighbouring chaotic zones as
impermeable barriers. In fact some of them leak (and in three dimensions
all of them do). A few leak like sieves, with large holes which let particles
through readily; others are more like the paper in a coffee filter, and it
takes a long time before the leakage becomes noticeable. It follows that
the notion of an “equilibrium coarse-grained distribution” depends on the
interval of time, T , in which one is interested. If the boundary between
two chaotic regions is easily breached on a time scale much shorter than
T , then the appropriate distribution will be constant across both regions.
It may be different in two regions if the barrier between them is unlikely
to leak over a time of order T .
Unless the potential we are looking at is very simple, there will be
many regions of chaotic motions, and the barriers between them may be
breached on a wide variety of time scales. It would therefore follow that,
no matter how long we waited, the coarse-grained distribution function
would always be changing. There would always be barriers which could
only be breached on a time-scale comparable with the present age of the
system, and it is leakage across those barriers that would be the most
noticeable evolutionary process at that stage. The distribution function
would already have reached equilibrium across barriers which could be

†
This is different in character from the exponential divergence discussed in Chapter
13. A better picture to have in mind is Arnold’s famous “cat” map (cf. Lichtenberg
& Liebermann 1992, p.306).
 Relaxation, Stability and Evolution in Dynamical Astronomy 109

breached on much shorter time-scales. Those whose presence would only

become noticeable on much longer time scales would be effectively im-
permeable.
Some time scales for diffusion are exponentially long (cf. Lichtenberg
& Liebermann 1992, p.398f). The so-called “Arnold diffusion” appears
to have this character, and it has even been said that “if you can see it,
then it is not Arnold diffusion”!

Relaxation, Stability and Evolution in Dynamical Astronomy

The remarks we have made are rather speculative, at least in the context
of stellar dynamics, where their consequences are only just beginning to
be worked out. Part of the reason for this is that the study of motion in a
fixed potential, especially in the study of evolutionary phenomena, seems
so far from reality, where the potential must evolve with the system.
Nevertheless, the idea that a system will always be evolving on a time
scale comparable with its age does seem to be a rather attractive and
general one, when interpreted intelligently.
In recent years it has led to a reinterpretation of the age-old question of
the stability of the solar system (Laskar 1996), a problem which has led to
so much important dynamical theory. It is possible to argue that the solar
system is always subject to instabilities acting on a time scale comparable
to its lifetime. Instabilities acting on much shorter time scales will have
either reached equilibrium or else have led to the ejection of objects on
unstable orbits. Those acting on extremely long time scales will have had
(as yet) little bearing on the evolution of the solar system.
In the case of stellar dynamics, certainly when applied to globular star
clusters, it is usual to take a more simplistic point of view. It is usually
supposed that the action of processes such as phase mixing and violent
relaxation play themselves out early on in the lifetime of the system.
Thereafter the system is in dynamic equilibrium under the orbital motions
of the stars. If the system is isolated and nearly spherical, then the sort
of chaotic behaviour illustrated in Fig.3c can be neglected. It does play a
significant role in the escape of particles from clusters which are strongly
perturbed by external fields (Chapter 12). That apart, however, one
must look elsewhere for any mechanism which will cause the system to
evolve slowly (or “secularly”) from one quasi-equilibrium to another. The
mechanism usually invoked is two-body relaxation, and we turn to that
topic in Chapter 14.
110 10 Violent Relaxation

Problems
1) Determine the t-dependence of f at a given point in Fig.1. If f is
averaged over a small interval δJ in J, show that hf i → 1/2, and
determine a time after which |hf i − 1/2| <  for any given  > 0.
2) Prove the quasi-H-theorem stated in the text for any convex function
C, when evolution is governed by the collisionless Boltzmann equation.
3) Fig.3 deals with the system θ̇ = J, J˙ =  sin θ cos t, where  = 0, 0.02
and 0.2, respectively, in the three frames. Defining φ = θ − t, show
that φ varies slowly near J = 1, and use the method of averaging to
show that J and φ evolve approximately according to the equations
φ̇ = J − 1

J˙ = sin φ
2
if  is small. Within this approximation deduce that (J − 1)2 /2 +
(/2) cos φ is nearly constant, and hence determine the width (in J)
of the resonant region.
 11
Internal Mass Loss

A stellar system in dynamic equilibrium loses neither mass nor energy.

In fact the stellar systems in nature do both, and the reasons for this
are both external and internal. In this chapter we consider the latter;
that is, we consider a stellar system isolated from all external influences,
including gravitational ones.
We have two processes in mind. One is caused by the internal evolu-
tion of the stars. Note that this is the first occasion on which we have
abandoned the point mass model on which we have relied so far, at least
to the extent that we now consider time-dependent masses. The other
is caused by the gravitational interactions of pairs of stars, which is re-
ally the topic of Chapter 14, and will be discussed rather briefly in this
chapter. We also deal with the effects in two ways. One is the scaling
treatment (Chapter 9) and the other uses a phase space description.

Evolution of Length Scale

A single star evolves at a rate which is a rapidly increasing function

of its initial mass. Therefore, if we examine the stars in an old stellar
system, we find that only those with a sufficiently low mass are more-or-
less unevolved, with masses close to those they were born with. Those
which were born with higher masses will have evolved, and in the process
will have lost mass, leaving a remnant which may take the form of a
black hole, a neutron star, or a white dwarf. Simple prescriptions for
these aspects of stellar evolution have been in use in stellar dynamics
for a long time (cf. Terlevich 1987, Chernoff & Weinberg 1990, Problem
1.) The details, however, are constantly changing: the relation between
initial and final mass is an active area of research in stellar evolution
(see, for example, Weidemann (2000) for white dwarfs and Timmes et al.
(1996) for neutron stars and black holes).

111
112 11 Internal Mass Loss

In relation to the characteristic dynamical time scale in a stellar system,

i.e. the crossing time (Box 1.1), these processes of mass loss may be either
fast or slow. For example, the mass lost in a supernova event is expelled
at speeds which exceed the orbital speeds in a stellar system by a factor of
100 or more, and the time scale is proportionately shorter. Mass expelled
in a steady wind may be lost on a time scale much exceeding the crossing
time. In either case the mass lost by the system is lost on a long time
scale, since even the instantaneous loss of the entire mass of one star
represents only a slight loss of mass by the entire system. Fig.1 shows
the mass lost in several stellar populations representative of those often
used in the study of old stellar systems.
The following argument shows why the time scale on which mass is lost
is important (Hills 1980). Suppose a system is in dynamical equilibrium,
so that its total, kinetic and potential energies are related by T = −E and
W = 2E (cf. Chapter 9). If a fraction f of the mass is lost instantaneously
then the new kinetic and potential energies are T 0 = (1 − f )T and W 0 =
(1 − f )2 W , whence the new energy is E 0 = E(1 − f )(1 − 2f ). It follows
that at most half of the mass can be lost if the system is to remain bound.
In practice, if more mass is lost then the most bound stars remain as a
1

0.8
Fraction of mass remaining

0.6

0.4

0.2

0
7 7.5 8 8.5 9 9.5 10
log(time) (yr)
Fig. 1. Mass lost as a function of time for five possible model stellar populations.
The middle curve corresponds to an analytical approximation of a Miller-Scalo
mass function (Eggleton et al. 1989). The others give results for a power law
mass function with probability density f (m) ∝ m−α , with α = 3.35, 2.35, 1.35
and 0.35. (The value 2.35 comes from Salpeter 1955.) In each case the maximum
and minimum mass are 15 and 0.1 times the mass of the sun, respectively. The
prescription for stellar mass loss is taken from Chernoff & Weinberg (1990),
though they applied it to different mass functions.
 Evolution of Length Scale 113

Fig. 2. Remaining bound mass following mass loss (after Lada et al. 1984).
The time scale of mass loss is τR , and τC is a measure of the crossing time. Slow
mass loss is less disruptive than fast mass loss.

small, bound system. Even if less than half the mass is lost, some stars
escape (Fig.2).
The role of dynamic equilibrium in this can be assessed by introducing
the virial ratio q = −T /W . Thus q = 1/2 in equilibrium, and if the
system is bound initially we require 0 < q < 1. If the above calculation
is repeated the result is that E 0 = E(1 − f )(1 − f − q)/(1 − q), and so
the condition that the final system is bound is that f < 1 − q. Thus
a “cool” system can lose a relatively high fraction of its mass and still
remain bound.
Now we return to systems in dynamic equilibrium, and consider what
happens if only a tiny fraction of the mass is lost. Then the change
in mass is dM = −f M , and so dE = E 0 − E ' −3EdM/M . If the
loss of mass is sufficiently slow, the system is able to remain close to
virial equilibrium. Therefore this relation is true for subsequent mass-loss
events, and it follows that E ∝ M −3 . Thus the system can, in the end,
lose an arbitrarily large fraction of its mass without becoming unbound.
Also, the length scale of the system can be estimated by R ∼ −E/M 2 ,
and so R ∝ M −1 .
Before passing on, it is worth stressing that all three scenarios lead to
an increase in the energy of the system, and since the energy is initially
negative, this means that the magnitude of the energy decreases. It might
114 11 Internal Mass Loss

seem that this is obvious, as the mass of the system has also decreased. It
is easy to see, however, that even the energy per unit mass increases, i.e
becomes less negative. In every sense, then, these mechanisms of slow or
impulsive mass loss lead to a heating of the system. We shall see in Part
VIII that some heating mechanism is needed in order to lead a stellar
system to a peaceful death, and nothing more complicated than loss of
mass is required.
Now we turn our attention, more briefly, to another important mech-
anism of mass loss, whose effects on the energy of a system could not
be more different. It occurs because of interactions of pairs of stars. In
such interactions the two stars exchange energy, and in some cases the
final energy of one of the stars may exceed the escape energy. Since the
energy of this escaper is positive, the energy of the remaining (bound)
members of the system is more negative than before. Let us assume, fur-
thermore, that the average energy carried off by the escaper is a certain
fraction of the energy of a single member of the stellar system. There-
fore we may assume that the changes in mass and energy are related
by dE = −f dM (E/M ), where f is a constant (unrelated to the con-
stant f above). Note that dE, dM and E are all negative here, and so
f > 0. It follows that E ∝ M −f , and so R ∝ M 2+f . The results of this
mass loss are therefore qualitatively very different from what we discussed
previously: as the mass of the system decreases the system shrinks. His-
torically, this was the first suggestion of the evolution of stellar systems to
a singular endpoint (Ambartsumian 1938, Spitzer 1940), an issue which
will play a major role in later chapters.

Evolution of Structure
So far the discussion has dealt with gross issues of mass, energy and
radius, but has said nothing about how the detailed structure of a stellar
system is altered. For this purpose we restrict attention to systems which
remain spherical, and close to dynamical equilibrium. Then we may
suppose the distribution function f depends only on E, J and t, where
now E refers to the specific energy of a single star. How, then, does
f (E, J, t) change in response to the loss of mass, i.e. to the resulting
changes in the potential?
One way of arriving at the correct answer is to substitute the expression
f (E, J, t) into the collisionless Boltzmann equation (Eq.6.5), remember-
ing that we must express E and J as functions of r and v. For example
E = v2 /2 + φ(r). Note that E depends on t because φ is time-dependent.
It follows that
∂f dE ∂f
+ = 0. (1)
∂t dt ∂E
 Evolution of Structure 115

The argument we have just given is insecure. The rate of change of energy
dE/dt = ∂φ(r, t)/∂t, and so depends on the location of the star. Thus
Eq.(1) tells us that the rate of change of a function which depends on only
E, J and t depends also on r. It is natural, perhaps, to interpret the term
dE/dt in Eq.(1) as an average over the orbit of given energy and angular
momentum, and Box 1 confirms this guess with a better argument.

Box 1. Evolution of the Distribution Function

in a Slowly Evolving Potential

To describe the motion of a star in a smooth, slowly varying, spheri-

cally symmetric potential, we may introduce suitable action-angle vari-
ables. As actions we choose J, Jz and Ir , where Jz is the component of
angular momentum in the direction of any suitable z-axis, and Ir is the
radial action introduced at the end of Chapter 7 (see also Palmer 1994).
There are also three conjugate angles, but we do not need to concern
ourselves with them. The transformation from rectangular coordinates
and momenta to action-angle variables is canonical, and has unit Jaco-
bian, and it follows that the distribution of stars in action-angle space
is equal to f . Also, it is clear that E may be written (in principle) as a
function of Ir and J, and so the distribution function may be written as
f (E(Ir , J, t), J, t). Now two of the actions, viz. J and Jz , are constant,
while the third, i.e. Ir is an adiabatic invariant (Chapter 7). It follows
that f , regarded as a function of the actions and time, is independent of
t. (In fact this is just the collisionless Boltzmann equation in action-angle
variables.) Hence
∂f ∂E ∂f
+ = 0,
∂t ∂t ∂E
which is just Eq.(1) again. The adiabatic invariance of Ir does indeed
depend on averaging over the fast radial motion of a star, and so it is
now clear that ∂E ∂t here is to be interpreted in an orbit-averaged sense.
Incidentally, by Eq.(6.8), another way of stating this result is to write
 
∂ N (E, J)
= 0, (1)
∂t P (E, J) Ir

where N dEdJ is the number of stars with energy and angular momentum
in ranges dE and dJ , respectively, and P is their radial period; the sub-
script indicates that the time-derivative is to be taken at constant values
of Ir .
116 11 Internal Mass Loss

We shall not stop to solve these equations, but draw attention to one
or two qualitative points which their solutions would exhibit. We have
already seen that one consequence of mass loss may be the expansion of
the system. However, the angular momentum of each star is constant, and
determined by the size and velocity dispersion of the system initially. By
the time the system has expanded by a large factor, therefore, most orbits
will have low angular momentum for a system of such size. Even if the
initial distribution of stellar velocities had been isotropic, in the expanded
system nearly radial motions would predominate. It is also likely that this
is one mechanism by which the anisotropy of stellar systems grows in so-
called post-collapse evolution (Chapter 28), even though the mechanism
of the expansion looks rather different.
Despite these remarks frequent use is made in stellar dynamics of an
analogous equation for the evolution of f when the anisotropy is ne-
glected, viz.
∂f ∂s ∂f ∂s
− = 0, (2)
∂t ∂E ∂E ∂t
where s is the volume in phase space of the region in which the energy
is less than E (cf. Chapter 9 and this chapter, Problem 4.) Again this is
based on Eq.(1), except that the expression for Ė is obtained by averaging
over all points in phase space accessible to a star of energy E, instead of
the points accessible to a star of energy E and angular momentum J.

Problems
1) Suppose the main sequence lifetime of a star of mass m is tms =
1010 × (m/M )−2.5 yr (cf. Hansen 1999), that the entire mass of the
star is lost at this time, and that the mass function is as in the caption
to Fig.1. Compute the time scale of mass loss of the cluster, i.e.
−M/Ṁ. (For a more elaborate formula for the main sequence lifetime
see Cassisi & Castellani 1993.)
2) Assume that the mass in Keplerian motion varies as m(t) = mr +
(m0 − mr ) exp(−t/τ ), where m0 and mr are the initial and final (rem-
nant) masses, respectively, and τ is constant. The radial equation of
motion is
c2 Gm(t)
r̈ = 3 − ,
r r2
where c is the specific angular momentum; we assume that this is
constant. The orbit is initially circular with radius a, and we assume
that τ is much smaller than the orbital period, but not infinitesimal.
Show that, during the time when r does not vary significantly from
 Evolution of Structure 117

its initial value,

G
ṙ ' (m0 − mr )(t + τ [e−t/τ − 1]).
a2
We define the energy in terms of the potential of the remnant, i.e.
1 c2 Gmr
E = ṙ2 + 2 − .
2 2r r
Show that
G(m0 − mr ) −t/τ
Ė = − ṙe .
r2
Substituting the approximate value of ṙ deduce that the value of E
immediately after completion of mass loss is
c2 Gmr G2 τ 2 (m0 − mr )2
Er ' − − .
2a2 a 2a4
Deduce that the condition for remaining bound is that
m0 Gτ 2 m20
mr ≥ −
2 8a3
approximately.
For a more systematic solution of this problem see Hut & Verhulst
(1981).
3) Show that Eq.(2) may be written in the form
 
∂f
= 0, (3)
∂t s
where the subscript indicates that the derivative is taken at fixed
s(E, t).
4) The rate of change of potential, averaged over all phase space acces-
sible to a star of energy E, is given by
R ∂φ 1
δ(E − v2 − φ)d3 vd3 r
hφ̇i = R∂t 2 .
δ(E − 12 v2 − φ)d3 vd3 r
In a spherical potential show that this may be written as
∂s
hφ̇i = − ∂t ,
∂s
∂E
where
Z
16π2
s(E) = r2 {2(E − φ)}3/2 dr.
3 φ(r,t)≤E
118 11 Internal Mass Loss

5) Suppose a stellar system is subject to a tidal cutoff, i.e. there is

an absorbing spherical boundary with radius equal to the tidal radius
rt ∝ M 1/3 (cf. Chapter 12 below), and that it loses almost all of its
mass very slowly. Then the final range of values of s is tiny, and Eq.(3)
implies that f must be the same as in the initial model near s = 0.
What is the resulting model?
6) (unsolved): repeat the previous problem for Eq.(1), Box 1.
 12
External Influences

In this chapter we add one ingredient to the topics discussed in the

previous chapter. There we outlined what happens to a stellar system
when it loses mass, by whatever mechanism. Implicitly, however, we
assumed that the system was isolated. Now we add to the picture the
fact that the stars in a stellar system are also affected by surrounding
matter, and this is especially true of escaping stars. The picture we have
in mind is of a system like a globular cluster, orbiting inside a galaxy,
which is simply another stellar system, but much larger and more massive.
The way in which the galaxy affects the cluster depends on such factors
as the orbit of the cluster, and the distribution of mass within the galaxy.
We begin with the simplest non-trivial idealisation. We assume that the
orbit of the barycentre of the cluster is circular of radius R. Clearly, this
is possible only for certain types of galaxy, e.g. those with axisymmetric
potentials φg . We use an accelerating and rotating frame of reference
with origin at the barycentre of the cluster, such that the x-axis points
radially outward, and the y-axis points in the direction of motion of the
cluster. The acceleration of a star in the cluster has several terms, due
to: (i) the field of the galaxy; (ii) the gravitational field, φ, of the cluster;
(iii) inertial forces, i.e. Coriolis and centrifugal terms. As we show in Box
1 (see also Chandrasekhar 1942, Sec.5.5, or Ogorodnikov 1965, Sec.7.4)
the resulting equations are

GM x
ẍ − 2ω ẏ + (κ2 − 4ω 2 )x = −
r3
GM y
ÿ + 2ω ẋ = − (1)
r3
GM z
z̈ + ωz2 z = − .
r3

119
120 12 External Influences

Box 1. Equations of motion of a star in a tidally perturbed cluster.

We consider and model each of the forces in turn.

If we suppose the field of the galaxy is axisymmetric, with the cluster
orbiting in a plane of symmetry, then we can write
q
φg = φg ( (R + x)2 + y 2 , z).
If the cluster is small compared with the radius of its orbit, we may
approximate this by its Taylor expansion, i.e.
y 2 ∂φg (R, 0) x2 ∂ 2 z2 ∂2
φg ' φg (R, 0) + (x + ) + φg (R, 0) + φg (R, 0).
2R ∂R 2 ∂R2 2 ∂z 2
(1)
(Terms with a single derivative with respect to z do not appear because
of our assumption that the orbit is in a plane of symmetry.) By the
elementary equations of circular motion,
Rω 2 = ∂φg /∂R, (2)
where ω is the angular velocity of the cluster about the galaxy.
The field of the cluster we approximate by that of a point mass M at
the cluster centre. This is an inessential approximation, but it is adequate
for escaping stars which do indeed lie outside the bulk of the mass, even
if this will not be spherically symmetric because the tidal field, Eq.(1), is
itself not symmetric. Thus φ = −GM/r, where r 2 = x2 + y 2 + z 2 and M
is the mass of the cluster.
The centrifugal force is derived from the potential −(1/2)ω 2((R+x)2 +
2
y ). The other inertial force is the Coriolis force, which has components
2ω(ẏ, −ẋ, 0).
Now we can write down the equation of motion of an escaping star. By
Eq.(2) there is no linear term in the sum of the galactic and centrifugal
potentials, and the quadratic terms simplify to (κ2 − 4ω 2 )x2 /2 + ωz2z 2 /2,
where κ is the epicyclic frequency (cf. Chapter 7) associated with circular
∂ 2 φg
motion of radius R in the galactic potential, and ωz2 = .
∂z 2
Hence we obtain Eqs.(1) in the main text.

These are the equations of a famous problem in dynamical astronomy,

called Hill’s problem (Hill 1878). G.W. Hill used them, towards the end
of the 19th century, in pioneering studies of the motion of the moon
about the earth, as perturbed by the sun. Here, however, it describes
the motion of a star about a cluster, perturbed by the Galaxy. It also
 Zero-Velocity Surfaces 121

arises in a beautiful modern problem: that of the mutual interaction of

the coorbital satellites of Saturn when close to conjunction (cf. Petit &
Hénon 1986). Finally, an identical force law occurs in the problem of the
tides (with additional terms due to fluid forces), and this gives its name
to the approximation used here: the “tidal” approximation. Roughly
speaking, it is applicable if the size of an object (in this case, the cluster)
is small compared to the distance to the perturber. Thus a cluster affected
by a large neighbouring mass is said to be “tidally perturbed”. Just as
the tides are raised on earth in the directions towards and away from the
moon, in Eq.(1) we see that the tidal field accelerates stars away from
the cluster in the ±x directions.
It is not surprising that κ appears here. If we neglect the right sides
of Eqs.(1), the x, y motion is displaced simple harmonic motion with this
frequency, i.e.
(x, y) = (a0 + a1 cos κt, b0 + a0 (κ2 /ω − 4ω)t/2 − 2ωa1 sin κt/κ), (2)
where the subscripted quantities are arbitrary constants. This is a lin-
earised approximation of the epicyclic motion described in Chapter 7.
The motion is retrograde (with respect to the motion of the cluster about
the galaxy). Therefore it is easy for stars to remain in the vicinity of the
cluster, if in retrograde motion.

Zero-Velocity Surfaces
The analysis of Eqs.(1) begins easily enough. There are two equilibrium
points, on the x-axis at x = ±rL , where
rL ≡ (GM/(4ω 2 − κ2 ))1/3. (3)
These are called Lagrangian points, as they correspond approximately
to two exact solutions of the general three-body problem discovered by
Lagrange (1772).
There is also an integral, called the Jacobi integral, given by C = (ẋ2 +
ẏ + ż 2 )/2 + (κ2 − 4ω 2 )x2 /2 + ωz2 z 2 /2 − GM/r. Its existence stems from
2

the fact that the potential is steady in this rotating, accelerating frame.
For a star at rest at one of the Lagrange points, C = CL ≡ −3GM/2rL.
The equipotential surfaces
(κ2 − 4ω 2 )x2 /2 + ωz2 z 2 /2 − GM/r = const (4)
are called “zero-velocity surfaces”, and are generalisations of what are
called “Hill’s curves” in the two-dimensional problem. These surfaces
exclude the motion from certain regions of space (Fig.1). For values of
C < CL they consist of three pieces, the innermost of which confines a
star to the vicinity of the cluster. Therefore if a star is situated within
122 12 External Influences

the cluster and has C < CL , it cannot escape. This is analogous to the
existence of an escape energy for stars in an isolated cluster. The lowest
value of C at which the surface opens out to infinity is C = CL , and so
it is usual to adopt the “radius” of this last closed surface, i.e. rL , as the
limiting radius of a cluster in a galactic potential. It is usually called the
tidal radius, denoted by rt .
When C  CL the star is confined to the middle of the cluster, far from
the reach of the tidal field. It behaves almost like a star moving in the
potential of the cluster alone. As C increases the star can wander further
and the effect of the tidal perturbation becomes stronger. Fig.10.3 tells
us what to expect, and helps to understand the surface of section actually
at C = CL (Fig.2). This shows several planar orbits (z = 0), and plots
the phase space coordinates x and ẋ whenever y = 0 and ẏ > 0. The
points are confined by the Jacobi integral, and the orbits are a mixture
of chaotic and regular ones. The large area of regular orbits on the left
correspond to retrograde motions, i.e. the stars move around the cluster
in the opposite sense to the motion of the cluster around the galaxy. We
have already seen why it is that retrograde motions are so special, and
there is plenty of supporting numerical evidence (e.g. Keenan & Innanen
1975).
Now we return to questions of escape. We have just stated that there
is a limiting “energy” below which escape is impossible, just as in the
problem of escape from an isolated, spherical potential. There the resem-
blance ends. In an isolated cluster, a star with energy above the escape

0.6
0.4
0.2
z
0
-0.2
-0.4
-0.6
-1

-0.5

y0 -1
0.5 -0.5
0
0.5 x
1 1

Fig. 1 The critical zero-velocity surface for Hill’s problem.

 Zero-Velocity Surfaces 123

4
3
2
1
dx/dt

0
-1
-2
-3
-4
-0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8
x

Fig. 2. Surface of section for a modified form of Hill’s equations. Here the
potential of the cluster is a Plummer potential with a suitably chosen scale
√ radius.
The equations of motion are obtained from Eqs.(1) by replacing r by r2 + a2 ,
where a = 0.1 is a suitably chosen scale radius, and setting κ = ω = GM = 1.
For further explanation see text.

energy will almost certainly escape (unless, on its way out, it encounters
another star which once again reduces its energy sufficiently.) In a cluster
in the tidal field of a galaxy, however, this is not guaranteed. We have al-
ready seen that stars in retrograde motion may remain near (but outside)
the cluster indefinitely, and this may easily happen to cluster members in
retrograde motion with C > CL . (The link between epicyclic motion and
retrograde escapers was drawn to our attention by Jeremy Goodman.)
One might say that the tidal radius for retrograde orbits exceeds that for
prograde orbits (Danilov 1999).
Even more fascinating is the fate of stars in prograde motion with
C > CL , and is most easily discussed when C − CL is small. Then the
openings in the equipotential surface through which a star must pass in
order to escape are tiny. Stars which do escape must do so through these
openings, and it is clear that the time they take to find them depends
on the amount by which C exceeds CL (Fukushige & Heggie 2000). This
in turn depends on how much the energy can be changed by relaxation
effects (Chapter 14) during the time taken to find the gaps. The result
of this interplay is that the rate of escape from star clusters in a steady
tidal field depends on N in a quite unexpected way (Baumgardt 2001).
When stars do escape, they do so along the directions of the Lagrange
points, as expected (Fig.2.2). But the Coriolis acceleration (i.e. the
124 12 External Influences

terms with coefficient ±2ω in Eq.(1)), bends the orbits along the direction
of motion of the cluster round the galaxy (Fig.3, left). “Extratidal”
extensions of this general nature are now known around a number of
Galactic globular clusters (e.g. Grillmair et al. 1995) and can be modelled
quite successfully (Combes et al. 1999).
When C = CL the width of the gaps through which stars escape re-
duces to zero, and they lie at the Lagrangian points. These equilibrium
solutions exist only at this critical value of C. As C increases slightly,
they give birth to a pair of periodic orbits, called Liapounov orbits. These
orbits are unstable, but they control the way in which stars can escape.
(Incidentally, the analogous orbits near the Earth, perturbed by the sun,
lead to a series of orbits which have a role in space dynamics. On Decem-
ber 2 1995 the international satellite SOHO, for observations of the sun
and heliosphere, was placed on an orbit of this general kind, with occa-
sional adjustments to quench the instability (“station keeping”). An orbit
in this location is particularly suited for continuous solar observation.)
The manner in which the Liapounov orbit controls escape is a little
complicated, but is most easily understood for orbits in the x, y plane.
For such orbits the Liapounov orbit behaves like a saddle point. An
escaper, which necessarily has to come close to this orbit if C − CL is
small and positive, must pass on one side or the other. On one side it
is deflected by the Liapounov orbit out of the cluster, and it escapes in
every sense. If it approaches on the other side it is deflected back into the
cluster, and must try again later, when the same alternatives apply. Even
a star which does escape may, if it approaches too close to the Liapounov
orbit, spend a long time there before being deflected one way or the other.
Furthermore, the number of times a star has to approach this orbit before
escaping is an extremely irregular function of its position.
Incidentally, the Liapounov orbits (there is one near each Lagrangian
point) also play a critical role in the behaviour of the coorbital satel-
lites of Saturn, and there also give rise to sensitive dependence on initial
conditions. It is an example of “irregular scattering” (Hénon 1988).

Time-Dependent Tides

These details can be fascinating, but they live in a highly idealised

problem, and the general lesson they should teach us is that escape is a
complicated problem as soon as we step away from the spherical poten-
tial of an isolated cluster. And if we take a more realistic cluster orbit
than the circular orbit discussed above, it is clear that almost nothing is
understood about it.
For a cluster in an elongated galactic orbit, it is usual to assume that its
limiting radius will be the minimum value of rL around its orbit, which is
 Time-Dependent Tides 125

at perigalacticon, where the tidal field is strongest. But we already know

that this is an oversimplification for a circular orbit, and for an elongated
orbit there is not even any analogue of the Jacobi constant, because the
tidal field is now time-dependent, even in a rotating frame. Another way
of expressing this is to say that the tide changes the energy of individual
stars in a cluster on an elongated orbit, a mechanism referred to as “tidal
heating”. It should be clear from our explanation that this does not
happen if the cluster is on a circular orbit, though it is a not uncommon
misconception to suppose that tidal heating occurs in that situation also.
Two kinds of time-dependent tidal fields are of importance in the study
of star clusters. One is associated with clusters on elongated orbits as
they pass perigalacticon (Fig.3). If this occurs close to the dense central
parts of the galaxy the dynamical effects are referred to as bulge shocking.
A local and short-lived effect is also caused by passage through the disk
of the galaxy, and this is what we now study.
To see how different this is from the steady tides acting on a cluster in
a circular orbit, consider the timescales involved. The scale height of the
disk potential is of order 100pc, but a cluster passes through the plane

200 200
150 150
100 100
50 50
0 0
-50 -50
-100 -100
-150 -150
-200 -200
-150
-100
-50
0
50
100
150

-150
-100
-50
0
50
100
150

Fig. 3. Mass loss by a cluster in a tidal field. On the left a King model with
concentration W0 = 3 (Chapter 8) is placed on a circular galactic orbit, and is
shown after one revolution. Stars are lost steadily. On the right the same cluster
has been placed on an elliptical orbit of eccentricity 0.5; it starts at apogalacticon
with a radius corresponding to the tidal radius at perigalacticon, and is shown
after one revolution. In this case stars are lost mostly near perigalacticon, and
there is a gap in the distribution of escapers near the cluster. Galactic centre is
to the right.
126 12 External Influences

with a normal speed of order 100kms−1 . On the other hand the radius
of a typical cluster is of order 10pc, while typical speeds for stars within
the cluster are of order 10kms−1 . Therefore the timescales of passage
through the disk and orbital motions within the cluster are comparable.
This is why the process is often referred to as disk shocking.
While a cluster is traversing the disk, the relative effect of the disk
and the cluster on the internal motions within the cluster are propor-
tional to the differential accelerations, across the extent of the cluster,
of the fields of the disk and cluster, respectively. Since the density in
the disk (∼ 0.1M pc−3 ) is much smaller than that in a typical clus-
ter (∼ 103 M pc−3 ), and since the event is so short-lived, the effect of
the disk is relatively minor. Nevertheless, clusters experience of order
100 such shocks in their lifetime, and so their cumulative effect is not so
negligible.
Box 2 shows that each shock tends to heat the cluster (until it reviri-
alises, once the shock is over), and finally gives an estimate of the time
scale on which repeated shock heating becomes effective. Notice that none
of the parameters of the cluster appear in that expression. Therefore, in
a galaxy with a massive disk, all clusters will succumb to destruction by
disk shocking on a comparable time scale, unless some other process acts
even faster∗ . This is, admittedly, an oversimplification. The strength of
shock heating does depend in detail on the structure, size and mass of the
cluster, and the point in the plane where the crossing occurs: the den-
sity within the disk increases significantly towards the centre of a galaxy.
Also, our estimate of the importance of this effect was a global one; in
fact the effect of disk shocking is strongest on the outer stars of a cluster.
Indeed for many clusters this is the dominant mechanism for changing
the energies of the outermost stars, and has therefore been referred to as
“shock relaxation” (Kundić & Ostriker 1995). Detailed models suggest
that it has a substantial effect on the escape rate (Spitzer & Chevalier
1973).
There is one circumstance in which this effect is suppressed, and that is
when we are considering a cluster where the time scale of the disk crossing
considerably exceeds that of orbital motions within the cluster. One
might expect then that the energy of a star should behave as an adiabatic
invariant, but in fact the situation is more complicated and interesting. In
the first place the typical orbital time scales are longest in the outer parts
of a cluster, and so the effects of a moderately slow disk crossing continue
to be significant there. Secondly, in a general spherical potential the

∗
J.P. Ostriker pointed this out to us
 Time-Dependent Tides 127

Box 2 Shock Heating

Imagine that the disk of the galaxy is modelled as an infinite sheet of

matter of surface density Σ. Then the acceleration on opposite sides of
the disk is 2πGΣ, just as in the standard calculation of the electrostatic
field due to a charge sheet.
Suppose now that a star in the cluster is on the opposite side of the
plane from the centre of the cluster. It is accelerated relative to the centre
by an acceleration of order 4πGΣ, and this lasts for a time of order r/V ,
where r is the displacement of the star from the centre of the cluster
and V is the speed of the cluster through the galaxy. (In fact we should
measure both quantities in a direction perpendicular to the galactic plane,
but the above estimate is good enough for order-of-magnitude purpose.)
It follows that the internal speed of the star is altered by an amount
δv given by δv ∼ 4πGΣr/V , and that the kinetic energy of the cluster is
increased by an amount of order Mc (δv)2 /2, ∼ 8π 2 G2 Mc Σ2 r2 /V 2 , where
Mc is the mass of the cluster.
Now tidal shocks occur at intervals of order 2πR/V , where R is the
radius of the orbit of the cluster around the galaxy. Therefore, if we write
the kinetic energy of the cluster as M v2 /2, where v is the mass-weighted
root mean square speed of its members, we find that the time on which
tidal shocks may destroy a cluster is of order
v2 RV
tsh = . (1)
8πG2 Σ2 r2
Next, for a virialised cluster we may estimate that v2 ∼ GMc /r (Chap-
ter 9). Also, if the cluster is tidally truncated, its size is given by the tidal
radius, in order of magnitude. Therefore by Eq.(3) in the text, where we
may treat the galactic field as due to that of a point mass of mass Mg ,
we may write
r ∼ R(Mc /Mg )1/3 . (2)
It follows that
Mg V
tsh ∼ . (3)
GΣ2 R2

motion of a star is governed by two frequencies (of radial and azimuthal

motion in the orbital plane). We have already seen (Chapter 7) that near-
resonances can then occur, which means that there are combinations of
frequencies corresponding to long time scales. Provided that the external
128 12 External Influences

potential is significantly coupled to these combinations of frequencies,

even a slow crossing can have a substantial effect. This observation is
due to Martin Weinberg (Weinberg 1994; see also Gnedin & Ostriker
1999).

Problems
1) Show that the mean density within a tidally limited star cluster is
constant, and comparable with that of the galaxy.
2) In a galaxy with a flat rotation curve, show that the tidal radius (for a
cluster of given mass) varies with galactocentric distance as rt ∝ R2/3 .
(See Surdin (1995) for a comparison with observational data.)
3) Explain qualitatively why seas on Earth exhibit two tides each day.
4) Consider a very large, planar epicycle, given by Eqs.(2), which encir-
cles the cluster. By using the method of variation of parameters on the
full Eqs.(1), and averaging the resulting equations over the epicyclic
motion, show that the centre of the epicycle oscillates stably.
5) By linearising Eqs.(1) in the vicinity of the Lagrangian point, find
an approximation to the Liapounov orbits there, and examine their
stability, all within the linear approximation.
 PART IV.
MICROPHYSICS: N = 2

The following three chapters are devoted to two-body interactions in the

context of the million-body problem. Chapter 13 shows that these cause
neighbouring orbits to diverge with an approximately exponential time
dependence. Beginning with the 3-body problem, we go on to investigate
the N -dependence of the e-folding time scale for the divergence. We
relate the phenomenon to the exponential divergence of geodesics in an
alternative formulation of the problem.
Chapter 14 is quintessential collisional stellar dynamics. Here we con-
sider the cumulative effect of many two-body encounters on the motion
of a single star: the theory of two-body relaxation. We develop a number
of standard formula for the first and second moments of the cumulative
change in its velocity. The first moment corresponds to the phenomenon
of dynamical friction. We then go on to incorporate this theory into
an evolution equation (the Fokker-Planck equation) for the distribution
function. We approximate this equation in a form appropriate to the
situation in stellar dynamics, when the time scale of relaxation is much
longer than that of orbital motions. This also incorporates the evolution
which may result from slow changes in the potential.
Chapter 15 takes a close look at the two-body problem itself. We show,
in particular, that the two-body collision singularity is a removable sin-
gularity. This introduces a number of topics which might seem surprising
in the context of the million-body problem: the Lenz vector, quaternions,
the Hopf map, the simple harmonic oscillator, and even a transformation
into four dimensions. The significance of all this theory is that it leads
to an important computational tool: KS regularisation.

129
 13
Exponential Orbit Instability

Up to this point in the book we have largely turned our back on the
microscopic character of the million-body problem. Usually we have ap-
proximated the gravitational field by that of a smooth distribution of
matter. Now we concentrate on the interactions between small numbers
of stars in the system, often only two or three stars at a time. In later
parts of the book we shall see how these microscopic processes influence
the large-scale behaviour.
The purpose of this chapter is to look at the question of sensitivity
to initial conditions in the million-body problem. Much current research
in other dynamical problems is devoted to this question, because of its
importance for prediction, for the foundations of statistical mechanics,
and perhaps even for the survival of life on Earth. What lies behind this
remark is the fact that the question of sensitivity is linked to stability,
and the stability of the solar system is something we rely on implicitly.
But the collision of comet Shoemaker-Levy with Jupiter in 1994 reminded
us that the solar system is not the well regulated clock we often take it
for. Less well known is the recent realisation that the rotation of the
Earth (which itself influences climate strongly) appears to be stabilised
by the presence of the moon (see Laskar 1996).
Consider the 1-body problem. The star proceeds with uniform recti-
linear motion r1 (t). Another single star, not interacting with the first,
and starting with a similar initial velocity and position, exhibits sim-
ilar motion r2 (t). The distance between
p p the two stars in space, i.e.
(r1 − r2 )2 , or in phase space, i.e. (r1 − r2 )2 + (ṙ1 − ṙ2 )2 , eventu-
ally depends nearly linearly on t, after an initial transient, during which
they may actually approach. Their separation in momentum space is
constant, however.

130
 13 Exponential Orbit Instability 131

To bring gravity into the picture, consider the relative motion of two
stars, i.e. the Kepler problem, with equation of motion
r̈ = −µr/r3 , (1)
where µ is constant. Now we play the same game of considering two Kep-
lerian motions started with similar initial conditions, but not influencing
each other. Assuming they have negative energy, in general these two mo-
tions will have slightly different periods, and so their separation in space
will grow nearly linearly while the separation remains small, except for
an initial transient, and any periodic oscillation. While their separation
δr is small, δr approximately obeys the variational equation
 
µ 3(r.δr)r
δ̈r = − 3
δr − , (2)
r r2
where r is a solution of Eq.(1). Note that the coefficient on the right-hand
side involves the derivative of the Keplerian acceleration, i.e. the “tidal”
acceleration of Chapter 12. (Observe the resemblance between the form
of these equations in Problem 1 and Eqs.(12.1)).
For a while, Eq.(2) describes approximately the evolution of the dis-
placement between the position on one orbit (r) and that on the other
(r + δr). For large times, however, both orbits remain bounded (being
Keplerian ellipses), and so, therefore, does their separation. The approx-
imation which leads to Eq.(2) breaks down, however, and the solution of
this equation continues to increase nearly linearly (Problem 1).
The three-body problem is non-integrable (cf. Moser 1973), and it is
harder to determine the rate at which solutions separate without recourse
to numerical methods. One problem that has been studied analytically is
Sitnikov’s problem (see Chapter 20). It is a special case of the “restricted”
three-body problem, in which one star, denoted by m3 , is massless. The
other two stars have equal masses and move as a binary on an elliptical
relative orbit. As m3 swings back and forth through the binary it is
subject to the time-dependent acceleration of the binary components.
Sometimes it gains energy and sometimes it loses energy, depending on
the phase of the binary as m3 passes through. It is easy to see that, if the
amplitude of its motion is large, the time between successive passages is
very large. In this case a small difference in the initial conditions leads
to a large difference in the phase of the binary at the time of closest
approach, and a large difference in the change of energy. The subsequent
passage through the binary will then be very different for the two motions.
The motion is very sensitive to initial conditions.
In fact for this example it is plausible to suppose that the difference
between the two motions, while it is still small, is increased roughly by
some factor during each passage through the binary, and therefore grows
132 13 Exponential Orbit Instability

roughly exponentially with time. This growth does not necessarily per-
sist, however, after the time at which the two motions lose all resemblance.
Moreover, in the long run it is likely that one and then the other motion
results in escape. By that stage their separation grows no longer exponen-
tially but nearly linearly (Fig.1). The exponential growth of separation
is also inapplicable at early times: before that first passage through the
binary, the motions separate less rapidly.
While this example might seem ludicrously artificial, it encapsulates
every essential aspect of these questions, even for the million-body prob-
lem. Here too the separation of solutions starting from neighbouring ini-
tial conditions is rather modest at first, and then there is a phase during
which they separate exponentially, until all the stars in the two motions
are separated by a distance of order the size of the system∗ . At still
later times, when stars escape, it may be conjectured that the separation
increases nearly linearly again. Overall, the factor by which the initial
displacement is magnified up until this time can be enormous, even for a
3-body system; factors of order 10200 have been measured in numerical
experiments (Dejonghe & Hut 1986).
In applying this idea to larger N -body systems, all that is missing is an
estimate of the timescale on which the exponential separation acts. After
pioneering numerical work by Miller (1964) there was a long gap before

2 10
1.5 5
1
log|z1 - z2|

0
0.5
0 -5
z

-0.5 -10
-1 -15
-1.5
-20
-2
0
100
200
300
400
500
600
700
800
900
1000
0
50
100
150
200
250
300
350
400

Time Time

Fig. 1. Exponential divergence in Sitnikov’s problem (see text and Chapter 20).
The left hand plot shows one solution of the equations of motion of this problem,
and the right hand plot shows (on a logarithmic scale) the absolute difference
between two initially neighbouring solutions. Up to the time when the third
body escapes (t ' 350) the growth of the difference is roughly exponential.

∗
√
More precisely, at a distance of order the size of the system divided by N ; cf. Box
1. Between that point and the time when √the separation is of order the size of the
system, the separation grows roughly as t.
 13 Exponential Orbit Instability 133

Box 1. Estimate of the Liapounov exponent for an N -body system.

Suppose one star in an N -body system moves with relative speed V

past another star, their distance of closest approach being p (Fig.1). If
m is a typical individual stellar mass, the change in velocity is of order
Gm/(pV ), and the path of the first star is deflected through an angle
of order Gm/(pV 2 ) (cf. Box 7.2, Eq.(2), in the case α  1). Suppose
this star now travels a distance D before its next encounter. The angu-
lar deflection caused by the first encounter changes its position during
this encounter, and in particular its distance of closest approach, by an
amount of order GmD/(pV 2 ).

Gm/(pV2 )

p D

Fig. 1 Two-body deflection in an N -body system.

Now let us rerun these two encounters, but with a value of p differing
by an amount δp. Then the distance of closest approach in the second
encounter is changed by an amount of order δp + GmDδp/(pV )2 . (The
first term here occurs because the closest encounter distance would have
been affected even without the occurrence of the first encounter. The sec-
ond term is the differential of the displacement caused by the encounter.)
Thus the separation between the two orbits is magnified by a factor of
order 1 + GmD/(pV )2 per encounter, or by a factor 1 + Gmδt/(p2 V ),
where δt ∼ D/V is the interval between encounters.
Now let’s focus on the effect of repeated encounters at about this value
of the impact parameter. To estimate how often these encounters occur,
we use an “nσv” argument: if we travel at speed v in a medium where
the spatial number-density of stars is σ, and we are trying to hit a target
of cross sectional area σ around each star, the rate at which this happens
is nσv. Here σ ∼ p2 and v = V , and so the time between encounters is
δt ∼ 1/(nV p2 ). If we denote by s the separation between the two orbits,
it follows that s increases by a factor of order 1 + Gmnδt2 between each
encounter. After a time t  δt, then, we see that
log s ∼ log s0 + t log(1 + Gmnδt2 )/δt, (1)
where s0 is the initial value.
Now we can see which encounters are most effective in this process.
√
The second term on the right of Eq.(1) is largest when δt ∼ 1/ Gmn.
This time scale is nothing other than the crossing time scale, which is of
order the time scale of virial oscillations (Chapter 8) or typical orbital
periods (Chapter 7).
134 13 Exponential Orbit Instability

Box 1 (continued)
Finally we discuss what this result means. First, it is applicable while
the separation between the orbits is small compared with the distance
of closest approach, p. (Recall that the foregoing estimate was based on
taking the differential of the expression GmD/(pV 2 ).) The appropriate
value of p here corresponds to the typical closest distance of approach
in a time of order a crossing time. During this time a star traverses
the entire system. If we project the positions of all stars onto a plane
orthogonal to its trajectory we obtain N stars within a disc of radius R,
i.e. the radius of the cluster. It is√clear, therefore, that the typical closest
encounter distance is of order R/ N . (This is another “nσv” argument!)
Because the growth of displacement is exponential while the displacement
remains small enough, it takes only a few crossing times for any reasonable
initial displacement to reach a size of this order; the dependence on the
initial displacement is logarithmic. After the displacement becomes of
this order, the exponential growth of s slows down (Problem 4).

the first theoretical attack on the problem. This was the work of V.G.
Gurzadyan and G.K. Savvidi (1986), who used a formulation of the prob-
lem which goes back to work of N. Krylov. In this theory the trajectory
of a set of N particles can be treated as a geodesic in their configuration
space when it is endowed with a certain metric (Maupertuis’ Principle).
Then the divergence between two trajectories is analysed with the help
of Jacobi’s equation of geodesic deviation. The timescale on which tra-
jectories diverge may be estimated by computing the scalar curvature of
the manifold, which in turn involves the square of the acceleration ex-
perienced by each body. Gurzadyan & Savvidi estimated this by using
the mean square acceleration, though this is formally divergent, and it is
necessary to impose a short-length cutoff, denoted by rc . Unfortunately
1/2
the estimate of the timescale depends sensitively (as rc ) on this choice,
and the choice made by Gurzadyan & Savvidi leads to a time scale which
subsequent work has not confirmed.
The actual acceleration is never infinite, unless a collision occurs, and
an alternative approach is to recognise that the acceleration on each par-
ticle varies with time, and to consider its cumulative effect. A rough
argument is given in Box 1. It indicates that the encounters which are
most effective in sustaining exponential divergence of neighbouring solu-
tions are those which take place typically once per crossing time. Those
which are closer are more effective, but too rare, while more distant en-
counters are just not powerful enough.
 13 Exponential Orbit Instability 135

The calculation in Box 1 is so rough that it is not even clear that a

fully three-dimensional calculation would not change the result. Still,
more careful studies do indicate that the result is correct except for a
factor of order log(log N ). Indeed for typical models such as Plummer’s
model (Chapter 8) the e-folding time for the separation of orbits is ap-
proximately 0.7/ ln(1.1 ln N ) in standard units (Goodman, Heggie & Hut
1993). Incidentally, logs of logs are unusual in science, and it is not often
that such a slow dependence on a parameter appears in a theory: for
instance the time scale decreases by less than 35% between N = 103 and
N = 106 . It is virtually impossible to verify such a slow dependence em-
pirically over the restricted range of N for which numerical calculations
are feasible.
It is important to notice that the time scale we have estimated is not
a relaxation time scale in the sense which has been customary in stellar
dynamics. Until recently (see the introduction to Chapter 14), relax-
ation has been a name given to any process which changes the energies
of the stars (among other things). Close encounters do indeed change
the energies of the stars, and this process is analysed in the next chapter.
What we have considered in the present chapter, on the other hand, is
how the evolution of two neighbouring systems differ, and not how each
evolves. Still, the exponential divergence continues only for as long as
the linearised approximation (represented by the variational equations)
remains valid. When the difference between two neighbouring orbits be-
comes large enough, the energies of the particles in the two systems evolve
more independently, on a time scale which is indeed given by the relax-
ation time scale (Chapter 14). The distinction between relaxation and
exponential divergence has been made particularly clear in the case of
one-dimensional gravity, i.e. when we consider the perpendicular motion
of infinite parallel sheets (Tsuchiya & Gouda 2000).
The fact that the motions in an N -body system exhibit such sensitivity
to initial conditions does have important consequences, however. Though
the following arguments have no rigorous foundation for the gravitational
N -body problem, in each case there are analogous systems which also
exhibit sensitive dependence and for which something rigorous can be
said.
In the first place, sensitive dependence on initial conditions may imply
that successive encounters between the same pair of stars (which may
occur after the lapse of several crossing times) are effectively independent.
The issue here is the decay of correlations, which are set up in a pair of
stars each time they are involved in a close interaction. As we have
seen, sensitive dependence is a symptom of the exponential divergence of
neighbouring solutions. In the rigorous treatment of the foundations of
statistical mechanics, this in turn is an essential element of what is termed
136 13 Exponential Orbit Instability

“hyperbolic” behaviour. In certain well studied models exhibiting this

type of behaviour, it can be proved that correlations decay quickly (see,
for example, the review by Liverani 2000). By analogy, then, such an
argument may help to justify the statistical treatment of the cumulative
effect of many encounters, which is the foundation for the use of the
Fokker-Planck equation (Chapter 14).
Sensitive dependence also implies that the accurate computer simula-
tion of N -body systems can never be guaranteed for more than a few
crossing times. Indeed attempts at the numerical integration of the same
problem on different computers never yielded reproducible results on time
scales of interest (e.g. Hayli 1970), except for the smallest N . Such inte-
grations are often referred to as “experiments”, but reproducibility is of
first importance in other areas of experimental science, and seems to be
missing here. For a long time it has been assumed that, even though the
positions and velocities of the stars in a simulation are certainly wrong,
nevertheless the statistical results of a simulation are reliable. In a sense,
each simulation is a running Monte Carlo simulation.
The demonstration of numerical shadowing theorems has done some-
thing to underpin this assumption (Quinlan & Tremaine 1992). What
has been shown is that, even though numerical errors are made at each
step of a simulation, nevertheless there exists an exact solution of the N -
body problem which stays close to the computed solution, for a limited
period of time at least. In practice, also, the statistical results of N -body
calculations seem remarkably robust to errors (Smith 1979, Heggie 1991).

Problems
1) In Eq.(2), suppose that r corresponds to planar circular motion. In a
frame rotating with the corresponding angular velocity, show that the
coplanar components of Eq.(2) become
ẍ − 2ω ẏ − 3ω 2 x = 0
ÿ + 2ω ẋ = 0,
where ω 2 = µ/r3 . Solve these equations.
2) Show that the variational equations for the N -body problem (analo-
gous to Eq.(2)) are
!
X
N
Gmj 3([ri − rj ].[δri − δrj ])(ri − rj )
δr̈i = − δri − δrj − .
|r − rj |3
j=1,j6=i i
|ri − rj |2

By estimating the contribution from the typical nearest neighbour in

a system with density
√ n, show that the corresponding e-folding time
scale is of order 1/ Gmn, where m is a typical stellar mass.
 13 Exponential Orbit Instability 137

3)
X Gm2
i) Let V = − be the potential function for the gravitational
a<b
rab
N -body problem with equal masses m. (This restriction is inessen-
tial; cf. Gurzadyan & Savvidi 1986). Suppose the 3N -dimensional
configuration space, with coordinates xi (1 ≤ i ≤ 3N ), is given the
metric ds2 = (E − V )dxi dxi , where E is a constant, and summation
convention is in use. Show that geodesics are orbits of the N -body
problem with energy E.
ii) Show that the scalar curvature is
3W,i W,i W,ii
R= (3N − 1)(N − 2) + (3N − 1),
4W 3 W2
where W ≡ E − V . Show that this vanishes if N = 2 (except at a sin-
gularity). Does this have anything to do with the lack of exponential
divergence in the 2-body problem?
iii) For the reduced planar Kepler problem in plane polar coordinates
we have
ds2 = (E + 1/r)(dr 2 + r2 dθ 2 ). (3)
Now consider an axisymmetric surface in R3 of the form (ρ(r), θ, z(r)),
where (ρ, θ, z) are cylindrical polar coordinates. Try to find functions
ρ(r), z(r), so that the metric induced on the surface by the Euclidean
metric in R3 is as in Eq.(3).
4) In the language of Box 1, suppose the separation s exceeds the impact
parameter p of the most effective encounters (p ∼ RN −1/2 , where R
is the size of the N -body system). Ignoring encounters with p 
s, show that Eq.(1) in Box 1 is replaced by ln s = ln s0 + Gmntδt,
where δt ∼ 1/(nV s2 ). pUsing this to obtain a differential equation
for s(t), show that s ∼ Gmt/V asymptotically. Assuming that the
separation in energy,
p δE, may be estimated by δE/E ∼ s/R, deduce
that δE/E ∼ t/ts , where ts ∼ V 3 /(G2 m2 n) for a system in virial
equilibrium (Chapter 9).
 14
Two-Body Relaxation

As first introduced by Maxwell, the term “relaxation” meant the process

by which a deformed elastic body returned to equilibrium. It was then
extended to the dynamical theory of gases, where equilibrium is a statis-
tical equilibrium characterised by the particular form of the distribution
of energies, and then transferred by Jeans to stellar dynamics. In stellar
dynamics equilibrium is never achieved, because particles escape, but one
can still think of a “quasi-equilibrium” on the time scale of many crossing
times. Even so, in due course the term “relaxation” gradually became
applied to several mechanisms which alter such properties as the energies
of the stars, whether or not they have anything to do with the approach
to equilibrium. More recently it has been argued (Merritt 1999) that
the term should be extended further to apply to any one of a variety of
mechanisms which cause evolution of the distribution function, whether
or not the quantities like energy or angular momentum are altered. The
history of the word reflects the development of the subject, from its initial
concern with equilibrium models to its modern concern with dynamical
evolution.
This chapter deals with one mechanism of relaxation, in which the
energy of one star is altered by its interaction with one other. It is
often called “collisional” relaxation, though the interaction is entirely
gravitational; real collisions we do not discuss until Chapter 31. Chapter
10 describes “violent” relaxation, in which the energy of a star is altered
by the collective effect of many stars in the system, while Chapter 12
mentions “shock” relaxation, which is caused by external fields.

Theory of two-body encounters

The link between two-body relaxation and orbital motions is quite sub-
tle, and initially we consider a spatially uniform distribution of stars, in
which somehow the stars would all describe rectilinear orbits, except for

138
 Theory of two-body encounters 139

occasional two-body encounters. We concentrate our attention on one

star, of mass m1 and velocity ṙ1 , say, and consider how it is affected by
encounters with other stars which it meets as it moves along. First, we
consider just one encounter with another star, of mass m2 and velocity
ṙ2 , say, assuming that the impact parameter (Box 1, Fig.1) is sufficiently
great that the deflection of the star is slight; this is referred to as the
“weak scattering approximation”, or “small-angle scattering”. (It turns
out that it is the cumulative effect of many weak encounters that domi-
nates, rather than the occasional stronger encounter.) Using the theory
of Keplerian motion we can compute the change in velocity of the star,
δ ṙ1 ; cf. Box 1, Eqs.(3) and (4).
The next step in the calculation is to sum the result over all encounters
which the star suffers in a time interval t, say, characterised by various
values of the impact parameter p and the velocity ṙ2 . In fact the star may
experience several encounters simultaneously, but in the weak scattering
approximation it is reasonable to suppose that successive (and even si-
multaneous) encounters can be treated as being independent. (After all,
the result for a single encounter is obtained by using unperturbed motions
of the two stars.)
The result (Box 1) is that
X Z
4πG2 m2 (m1 + m2 )
δ ṙ1 = −t (ṙ1 − ṙ2 )f (ṙ2 )dpd3ṙ2 , (1)
p|ṙ1 − ṙ2 |3
where p is the impact parameter and f (ṙ2 ) is the number density, in phase
space, of stars of mass m2 and velocity ṙ2 . (Because of the assumption
of spatial homogeneity we need not mention any dependence on r2 .) We
shall see later that the above expression is not sufficient to characterise
all important aspects of the encounters. Different encounters contribute
different δ ṙ with some probability density, and Eq.(1) gives an expression
for the first moment of this distribution. The second moment will also
be needed in due course.
We proceed with the further evaluation
R of the right side of Eq.(1). The
integration with respect to p yields dp/p = log(pmax /pmin ), where pmax
and pmin are hitherto unspecified limits on p. The choice of both limits
exposes significant issues, which we now consider.
The minimum is easy; the small-angle approximation is valid only if
|α|  1 (cf. Box 1), i.e. if p  Gm/v2 , where m and v are typical masses
and speeds. It is not hard to include large-angle scattering, and in fact
a more accurate theory shows that, when p is small, the value of δ ṙ1 is
smaller than the value given by the small-angle approximation. Thus this
singularity in Eq.(1) as p → 0 is just an artefact of the weak scattering
approximation, and in a more exact theory the corresponding integral
is no longer singular. We shall take care of this by simply cutting the
140 14 Two-Body Relaxation

integral off at p ∼ Gm/v2 ; it turns out that, when these results are applied
to a star cluster with N stars in approximate dynamical equilibrium, the
term neglected is smaller than the term we shall retain, by a factor of
order log N .

Box 1. Approximate theory of two-body relaxation

Let r = r1 − r2 be the relative position vector of the two stars. If we

neglect all other stars, it evolves according to
r̈ = −G(m1 + m2 )r/r3 . (1)
Let V be their relative speed while still far apart, and p their distance
of closest approach. (For small-angle scattering we need not distinguish
between the impact parameter and the distance of closest approach.) In
the weak scattering approximation the change in ṙ over the encounter
may be evaluated by substituting rectilinear motion in the right side of
Eq.(1) and integrating. The result is
δ ṙ = −2G(m1 + m2 )p̂/(pV ), (2)
where p̂ is the unit vector in the direction of r at the instant of closest
approach on the rectilinear motion (Fig.1).

^
v

^
p

Fig. 1 Motion of an encounter in a frame in which particle 2 is fixed.

 Theory of two-body encounters 141

Box 1 (continued)
Eq.(2) is not quite accurate enough, even for small-angle scattering,
but can be improved by the following trick. From Eq.(2) it follows that
the angle of deflection is α ' 2G(m1 + m2 )/(pV 2 ), (cf. Box 7.2, Eq.(2)),
while energy conservation shows that the magnitude of ṙ equals V long
after the encounter. Thus a vector of this magnitude has simply been
rotated by this small angle, and so
2G(m1 + m2 ) 2G2 (m1 + m2 )2
δ ṙ = − p̂ − V̂, (3)
pV p2 V 3
where V̂ is a unit vector in the direction of the original relative velocity,
and the second term is a small-angle approximation of −(1 − cos α)V V̂.
The effect of the encounter on the velocity of m1 is
δ ṙ1 = m2 δ ṙ/(m1 + m2 ), (4)
since the barycentre of the two stars moves with constant velocity.
Now we sum the result over many encounters, assuming that different
encounters are independent, as discussed in the text. It is evident that
the average value of p̂ is zero, and so Eqs.(3) and (4) lead to
X X 2G2 m2 (m1 + m2 )
δ ṙ1 = − (ṙ1 − ṙ2 ).
p2 |ṙ1 − ṙ2 |4
In order to perform the sum, let f (ṙ2 ) be the number density, in phase
space, of stars of mass m2 and velocity ṙ2 . The rate at which encounters
occur with stars moving with a range of velocities d3 ṙ2 , with values of
p in the range (p, p + dp), is 2πp|ṙ1 − ṙ2 |f (ṙ2 )dpd3 ṙ2 , and so the mean
value of the sum, over all encounters in time t, is given by Eq.(1) in the
text.

The choice of pmax has given greater difficulty. At one time (Jeans
1929, Chandrasekhar 1942) it was thought that encounters must take
place one at a time, so that only the nearest neighbour of m1 could
contribute to its relaxation. Following more recent treatments, however
(beginning with Cohen et al. 1950 and Hénon 1958), we have assumed im-
plicitly that many relaxing encounters may occur simultaneously. This is
supported by numerical experiments (Farouki & Salpeter 1994, Fukushige
1995). In a finite stellar system it is assumed now that encounters con-
tribute up to a distance p at which the stellar density falls significantly
142 14 Two-Body Relaxation

below its local value∗ . Therefore for a star in the core, i.e. the densest
central part of the system, pmax may be of order the radius of the core.
For a typical star in a nearly spherical system, it may be of the order of
a suitably defined radius, such as the half-mass radius, rh , which is the
radius of an imaginary
R sphere containing the innermost half of the mass.
With this choice, dp/p = log(rh v2 /Gm). Furthermore, for a system in
R v ∼ GM/rh , where M is the total
virial equilibrium, we have typically 2
mass, and so it is usual to choose dp/p = log(γN ), where γ is a constant
of order unity.
By a refinement of the above argument, it was customary for a long
time to take γ = 0.4. More recently, however, comparison with N -body
simulations has consistently suggested a smaller value, around γ = 0.11,
for systems with equal masses (Giersz & Heggie 1994). When there is a
distribution of masses, considerably smaller values are suggested by both
theoretical arguments (Hénon 1975) and numerical experiments (Giersz
& Heggie 1996). Whatever the value of γ, this factor is often referred to
as the Coulomb logarithm, which betrays the analogy with the kinetic
theory of plasmas, where the interaction is electrostatic.
By this stage we have arrived at the result
X Z
4πG2 m2 (m1 + m2 )
δ ṙ1 = −t log γN (ṙ1 − ṙ2 )f (ṙ2)d3 ṙ2 . (2)
|ṙ1 − ṙ2 |3
The remaining integral has a familiar look. Apart from a factor, it is
the same as the formula for the gravitational acceleration due to a mass
distribution with spatial density f . Therefore for an isotropic distribution
of velocities (corresponding to a spherically symmetric distribution of
matter), the result may be written
Z
1X ṙ1
δ ṙ1 = −4πG2 m2 (m1 + m2 ) log γN f (ṙ2 )d3 ṙ2 . (3)
t |ṙ1 |3 |ṙ2 |<|ṙ1 |

The left side of this result is often written dṙ1 /dt, though it must
be remembered that it has a statistical meaning. At any rate it is an
important result. It implies that stars are subject to a deceleration,
which is called “dynamical friction”. Note that, in the case of isotropy,
it is caused by stars moving more slowly than m1 . The strength of this
deceleration increases with the mass m1 . This is unlike the gravitational
acceleration due to a static external body, which would be independent
of m1 , even though dynamical friction is caused by the gravitational
attraction of other stars. To understand this, and to gain a physical
feel for this phenomenon, imagine we move with a star whose velocity is

∗
One could, however, imagine a system, however unrealistic, in which a vast low-
density halo contributed most to relaxation
 Theory of two-body encounters 143

high (Figs.1,2). Then the other stars approach predominantly from the
direction in which we are moving. They are deflected by our mass (m1 )
and this forms an enhancement in their density behind us. The larger
our mass the larger the enhancement, and it is this density enhancement
which decelerates us.
Since all stars are subject to dynamical friction, it would seem that all
stars must slow down. Recall, however, that these results are statistical,
and we shall see that they are not even sufficient to describe the main
statistical effect of encounters (cf. Eq.(5), below). It should be added
that the response of a finite, non-uniform stellar system can be a lot
more complicated than Eq.(3), which is a local approximation. Even if
an object orbits a stellar system outside all the members of the system,
so that the local value of f is zero, it will still experience an effective
deceleration caused by the response of the system. A theory dealing with
this and other issues was developed by Tremaine & Weinberg (1984).
Just as Eq.(2) sums the changes in ṙ1 , we may compute the sum of their
P
squares, or, more generally, the tensor δ ṙ1i δ ṙ1j . Looking at Eq.(3) in
Box 1, we see that this involves terms which vary with various powers of
p. Only terms in p−2 yield the Coulomb logarithm; terms which decrease
even more rapidly do not do so, and in fact are smaller by a factor of
this order. It is, however, necessary to treat large-angle scatterings more
exactly to establish this. Such terms are called “non-dominant” terms,
and are usually neglected in the standard theory.

Fig. 1 Creation of a wake behind a fast-moving star.

144 14 Two-Body Relaxation

0
y

-1

-2

-4 -3 -2 -1 0
x

Fig. 2. Density contours in the wake behind a fast-moving star. (See Problem
2)

The one result about the tensor which we shall quote is its trace. For
an isotropic distribution f (ṙ2 ) this is given by
Z
1X 1
(δ ṙ1 )2 = 8πG2 m22 log γN ( f (ṙ2 )d3 ṙ2 +
t v1 |ṙ2 |<|ṙ1 |
Z (4)
1
f (ṙ2 )d3 ṙ2 ).
|ṙ2 |>|ṙ1 | v2

Unlike dynamical friction, this result does not involve the mass m1 . This
will be important in Chapter 16. (By the way, just as the integral in
Eq.(3) stems from one analogous to gravitational acceleration, that in
Eq.(4) resembles the expression for the gravitational potential in a spher-
ical distribution. Indeed these integrals are sometimes referred to as
“Rosenbluth Potentials” (cf. Rosenbluth et al. 1957).)
One may continue to characterise the statistical effect of small-angle
scattering by computing the sums of products of degree 3 (and even
higher) in δ ṙ1 . The corresponding integrands decrease sufficiently rapidly
with increasing p, however, that all such terms are non-dominant, in the
sense introduced above.
With Eq.(3) and similar results, other important effects of relaxation
can be analysed. For example the sum of the changes in the specific
 Kinetic Equations 145

kinetic energy of m1 caused by encounters is

X X X
δE1 = ṙ1 . δ ṙ1 + (δ ṙ1 )2 /2, (5)
and so the sum may be computed from expressions which have already
been given. Notice that the two terms have opposite sign. The dynamical
friction term leads to a decrease in energy, which can be compensated sta-
tistically by the second. It is the balance between these two terms which
attempts to bring m1 into thermal equilibrium with the distribution of
stars m2 .

Kinetic Equations
The information we have obtained describes the statistical effect of re-
laxation due to encounters on a single star. Now we must consider how
this information can be extended to describe its effect on a distribution
of stars. In practice this may be done by starting with the Kolmogorov-
Feller equation
Z
∂
f (ṙ1 , t) = (f (ṙ1 − δ ṙ1 )R(ṙ1 − δ ṙ1 , δ ṙ1 ) − f (ṙ1 )R(ṙ1 , δ ṙ1 )) d3 (δ ṙ1 ),
∂t
(6)
where R(ṙ1 , δ ṙ1 ) denotes the rate at which a star with velocity ṙ1 suffers
encounters which change its velocity by δ ṙ1 . The first term on the right
shows how the distribution function at a particular velocity is increased
by encounters which cause the velocity of a star to change from some
other value, while the second gives the rate at which it decreases as stars
with velocity ṙ1 are scattered to other values of the velocity.
Intuitively obvious though the meaning of this equation is, some fun-
damental assumptions are needed to justify it. For example, it describes
relaxation as a (continuous) Markov process, in which the evolution over
an infinitesimal time interval depends only on the state of f at the begin-
ning of that interval. The considerations of the previous chapter provide
some intuitive foundation for this assumption, though without rigour. On
the other hand the Kolmogorov-Feller equation clearly does not apply on
time scales shorter than the duration of a typical encounter.
It is not yet clear how our statistical information on the cumulative
effect of many encounters may be exploited in Eq.(6), but in fact this is
rather easy. Our small-angle approximation is satisfactory if δ ṙ1 is small,
and a truncated Taylor expansion of Eq.(6) leads to
Z
∂ ∂
f (ṙ1 , t) = − δ ṙ1i f (ṙ1 )R(ṙ1 , δ ṙ1 )d3 (δ ṙ1 )+
∂t ∂ ṙ1i
Z
1 ∂2
δ ṙ1i δ ṙ1j f (ṙ1 )R(ṙ1 , δ ṙ1 )d3 (δ ṙ1 ).
2 ∂ ṙ1i ∂ ṙ1j
146 14 Two-Body Relaxation

Now, except for the factor f (ṙ1 ), the integrals here are the statistical
sums calculated earlier, i.e this equation may be written
 X 
∂ ∂ 1
f (ṙ1 , t) = − δ ṙ1i f (ṙ1 ) +
∂t ∂ ṙ1i t
 X  (7)
1 ∂2 1
(δ ṙ1i δ ṙ1j )f (ṙ1 ) .
2 ∂ ṙ1i ∂ ṙ1j t

Note that the terms we have neglected in the Taylor expansion (of de-
gree 3 and higher) are smaller than these terms by a factor of order the
Coulomb logarithm. This is not a very big factor, even for the million-
body problem: for γ = 0.11, ln(γN ) varies from about 6 to about 12
for N in the range from 104 to 106 . Nevertheless, formulations exist
where some of the neglected effects can be accounted for (e.g. Ipser &
Semenzato 1983). Where it has been possible to check, the omission of
these terms seems to make little difference to the resulting behaviour of
a stellar system (Goodman 1983).
Eq.(7) is a form of Fokker-Planck equation. It is a diffusion equation
with a drift term, the first term on the right side resulting from the
action of dynamical friction. An important property of this equation
is that a Maxwellian distribution is a solution. This solution does not,
however, provide a thermal equilibrium solution to the structure of a
stellar system, for two reasons. One is that, as the system relaxes, the
evolution of the distribution of velocities leads to changes in the spatial
distribution of matter; in turn these cause the potential to vary, and
this changes the energies of the stars in addition to the changes due to
two-body relaxation. The problem of incorporating the effect of this
process on the evolution of the distribution function will be discussed in
a moment. The second reason why the Maxwellian distribution fails to
provide an equilibrium model is the boundary conditions: in a stellar
system stars may escape, especially those with high energy. The way in
which a system reconciles these two inconsistent requirements, thermal
equilibrium and boundary conditions, is the subject of chapter 17.
The most important lesson to learn from Eq.(7) is the time scale on
which two-body relaxation acts. In fact, from Eqs.(3), (4) and (7) it can
be seen that this is of order v3 /(G2 m2 n log γN ), where v is a typical
speed, m is a typical mass, and n is the number density of stars∗ . This is
an estimate of what is called the “time of relaxation”, though the precise
definition is a matter of convention; most authors follow Lyman Spitzer,

∗
R
Recall that f is the number density in phase space, so that n = f d3 ṙ
 Kinetic Equations 147

who introduced the definition

0.065v3
tr = (8)
G2 m2 n ln Λ
(cf. Spitzer & Hart 1971), where v is the root mean square speed, m is the
mean stellar mass, and Λ is written for the argument of the Coulomb log-
arithm. The curious numerical coefficient here comes from an assumption
that the distribution of stellar velocities is Maxwellian, and from other
conventions (see Problem 4), but the expression itself is often used even
when the distribution is different from Maxwellian.
The numerical factor in Eq.(8) gives a spurious impression of precision:
in fact this formula is only a guide. In particular, in systems in which
stars of widely differing mass occur, the time scale on which two-body
encounters operate may vary widely. For example, though dynamical
friction on a star of mass m1 acts on a time scale comparable with tr if
m1 ' m, where m is the average mass, we have already seen (in the dis-
cussion of Eq.(3)) that its strength increases with m1 . Therefore the time
scale on which it acts varies roughly as mtr /m1 for m1  m. This helps
to explain why mass segregation occurs so rapidly in N -body systems
(Chapter 16; see, for example, Fig.16.1).
Eq.(8) makes it clear that tr varies a lot from place to place, and it
is often useful to have some more global measure of the relaxation time.
Following Spitzer (1987), it is usual to try to evaluate Eq.(8) for average
quantities inside the half-mass radius, rh , which yields what is called the
half-mass relaxation time
3/2
0.138N 1/2 rh
trh '
(Gm)1/2 ln Λ.
(9)
This result is boxed. Anyone who uses this book for reference will consult
this formula more than any other.
The relaxation time is very long compared with the other fundamental
time scale in a stellar system, i.e. the crossing time, tcr (Box 1.1). In
fact for a system in dynamic equilibrium it is easily shown from the virial
theorem that tr /tcr ∼ N/ log Λ (Problem 5). This implies that the mean
free path of the stars greatly exceeds the size of the system. Equivalently,
the stars make many orbits almost unaffected by two-body encounters.
If the distribution of density is nearly spherical, for instance, the theory
of Chapter 7 is applicable: stars describe rosette-shaped planar orbits at
nearly constant angular momentum and energy. Gradually, however, the
effects of two-body encounters make themselves felt, and the orbit of a
star will be seen to evolve slowly and stochastically. (The N -dependence
of this process is illustrated in Kandrup & Sideris 2001.) Such behaviour
148 14 Two-Body Relaxation

contrasts strikingly with the effect of collisions in an ordinary gas, where

the mean free path is tiny.

Relaxation and Orbital Motion

These considerations reveal a complication. Our theory of relaxation
was based on a spatially uniform distribution of stars. In fact a typical
star moves between regions of high and low density, where relaxation is
(respectively) rapid and slow, before the net effect of relaxation becomes
significant. We shall assume that the theory we have described applies
locally, but must take into account the fact that the Fokker-Planck coef-
ficients vary widely as the star moves. Furthermore, as the distribution
function evolves by relaxation, so does the density in space. In turn this
alters the potential, and there is a further resulting evolution of stellar
orbits and the distribution function in phase space. We have to consider
how to incorporate these factors into the Fokker-Planck equation.
We have seen that the relaxation time much exceeds the crossing time.
If, therefore, the distribution function is evolving on the timescale tr , it is
nearly constant on the crossing timescale. Therefore it is nearly a solution
of the collisionless Boltzmann equation (Chapter 6). If, moreover, the
potential is nearly stationary, we may exploit Jeans’ Theorem. Thus in
the case of a spherically symmetric system we suppose that f = f (E, J, t),
where the explicit time-dependent evolution takes place on the time scale
tr . This deals with the orbital motions.
We have already seen in Chapter 11 how the distribution function
evolves in response to slow changes in the potential. It turns out to
be most convenient to express the result in terms of N (E1 , J1 , t), which
was defined to be the number density of stars with energy E1 and angular
momentum J1 . It is related to f by Eq.(6.8), and it evolves according to
Eq.(1) of Box 11.1, which may be written in the form
   
∂N N ∂P
− = 0. (10)
∂t Ir P ∂t Ir

Here P is the period of radial motion of a star with these parameters, and
the subscript indicates that the time derivative is taken with the radial
action Ir fixed.
The effect of two-body relaxation is simply to contribute an additional
encounter term to the right-hand side of Eq.(10). We cannot simply
express Eq.(10) in terms of f and then add the right side of Eq.(7),
however, as Eq.(7) deals with functions of velocity and position, whereas
the left side of Eq.(10) involves functions of E1 and J1 . One way of
proceeding is to average the collision term around an orbit of a star with
energy E1 and angular momentum J1 . It is, however, easier to approach
 Relaxation and Orbital Motion 149

the problem less directly (Box 2), though the result is the same. It
is referred to as the orbit-averaged Fokker-Planck equation, though the
essential idea goes back to Kuzmin (1957).
There are two formulations of the resulting equation which look rela-
tively elegant and memorable. One, which we learned about from Donald

Box 2. The Fokker-Planck equation in a slowly varying potential.

If we include encounters only, N (E, J) evolves in an analogous way to

f in Eq.(7), i.e.
 X   X 
∂N ∂ 1 ∂ 1
=− hδE1 iN − hδJ1 iN +
∂t ∂E1 t ∂J1 t
 X   X 
1 ∂2 1 ∂2 1
hδE12 iN + hδE1 δJ1 iN +
2 ∂E12 t ∂E1 ∂J1 t
 X 
1 ∂2 1
hδJ12 iN . (1)
2 ∂J12 t
The angle brackets here signify that the moments of the changes in energy
and angular momentum are obtained by orbit-averaging expressions like
Eq.(5). Then all we have to do is replace the left side of this equation by
the left side of Eq.(10), and we are done.
The case of an isotropic distribution of velocities, where we suppose
that f = f (E, t), is handled similarly. The analogue of Eq.(10), when
expressed in terms of f , is Eq.(11.2), and the analogue of Eq.(1) above is
Eq.(9.5), which is important enough to repeat:
∂f ∂s ∂f ∂s ∂ 1 ∂2
− =− (DE f ) + (DEE f ). (2)
∂t ∂E ∂E ∂t ∂E 2 ∂E 2
Here s(E, t) is the volume in phase space in which the energy is less
than E, and DE and DEE are diffusion coefficients analogous to those of
Eq.(1), but incorporating the conversion between f and the number of
∂s
stars per unit energy; this is N (E) = f (E) , and
∂E
1X ∂s
DE = hδE1 i . (3)
t ∂E
This coefficient is an orbit-averaged version, for equal masses, of
Eq.(16.7).
150 14 Two-Body Relaxation

Box 3. Outline derivation of the orbit-averaged isotropic Fokker-Planck

equation for spherical systems with equal masses.

Our task is to compute the diffusion coefficients DE and DEE in Box

2, Eq.(2). We shall give a fairly complete derivation for the first of these
and leave the second as an exercise. We assume that all masses are equal
(m1 = m2 = m), that the distribution f (ṙ2 ) is isotropic in velocity, and
that the system is spherically symmetric with potential φ(r).
Substituting text Eqs.(3) and (4) into text Eq.(5) we readily find that
 Z Z 
1X 1 v1 ∞
δE1 = 16π 2 G2 m2 ln Λ − v22 f (v2 )dv2 + v2 f (v2 )dv2 ,
t v1 0 v1

where Λ ≡ γN and vi ≡ |ṙi |. This expression depends on v1 and also on

the position in the cluster, since f (v2 ) depends on position. As we intend
to average over all parts of phase space accessible to a star with a given
energy, we reexpress the result in terms of Ei ≡ vi2 /2 + φ, i = 1, 2. The
result is
1X 1
δE1 = 16π 2 G2 m2 ln Λ(− √
t E1 − φ
Z E1 p Z ∞ (1)
E2 − φf (E2 )dE2 + f (E2 )dE2 ).
φ E1

The effective upper limit of the second integral will usually be finite,
because f will vanish above a certain energy.
Averaging a function F (r) at energy E1 (i.e. over an energy hypersur-
face in phase space) means performing the integral
R
F δ(E1 − v12 /2 − φ)d3 r1 d3 v1
hF i = R .
δ(E1 − v12 /2 − φ)d3 r1 d3 v1
The denominator is just ∂s/∂E1 , and cancels in the expression for DE
(Box 2, Eq.(3)). Integrating with respect to v1 and then leaving only the
integration with respect to r ≡ |r1 | results in
Z φ−1 (E1 ) q
∂s
hF i = 16π 2 2(E1 − φ)F r2 dr.
∂E1 0
∂s
Setting F = 1 gives a convenient expression for .
∂E1
Applying
√ this averaging procedure to Eq.(1) we readily find that the
factor E1 − φ cancels out, and the first term in Eq.(1) yields the ex-
∂s
pression for , i.e. the derivative of the phase-space volume at energy
∂E2
E2 . The result is the surprisingly compact and symmetric expression
Z Z !
E1 ∂s ∂s ∞
DE = 16π G m ln Λ −
2 2 2
f (E2 )dE2 + f (E2 )dE2 .
φ(0) ∂E2 ∂E1 E1
 Relaxation and Orbital Motion 151

Lynden-Bell, is the form appropriate to a spherically symmetric system

with an isotropic distribution of velocities and equal masses, for which
the Fokker-Planck equation reduces to
  Z  
∂f1 ∂ f10 f0
2 2
= 16π G m ln Λ 2
min(s1 , s2 )f1 f2 − 2 dE2 , (11)
∂t s ∂s1 f1 f2
where si ≡ s(Ei ), defined in Box 2, and fi ≡ f (Ei ), i = 1, 2. Box 3
presents an outline of a derivation of this equation. The beauty of the
end result makes up for the ugliness of the derivation, but also hints at
the presence of some as yet undiscovered insight. Incidentally, the result
is quite reminiscent of Landau’s form of the Fokker-Planck equation in
plasma kinetic theory (cf. Thompson 1962). As well as being elegant
and memorable it is also useful, as it is the basis of much of the Fokker-
Planck industry which, starting with Hénon’s classic study (Hénon 1961),
has been the foundation of so much that we know about the million-body
problem.
The other setting in which the orbit-averaged Fokker-Planck equation
looks relatively elegant invokes action-angle variables. If we assume that
f is a slowly varying function of the actions Ii , i = 1, 2, 3, the result is
 X   X 
∂f ∂ 1 1 ∂2 1
=− hδIi if + hδIi δIj if .
∂t ∂Ii t 2 ∂Ii Ij t

Box 3 (continued)
A similar sequence of calculations needs to be performed for the coef-
ficient DEE . There are no surprises, but it helps to have the formula
Z −1
16π2 φ (E)
s(E) = {2(E − φ)}3/2 r2 dr (2)
3 0
at hand. The result is that
(Z Z )
E1 ∞
2 2 2
DEE = 32π G m ln Λ s(E2 )f (E2 )dE2 + s(E1 ) f (E2 )dE2 .
φ(0) E1
(3)
These results are now to be substituted into Box 2, Eq.(2). It is again
surprising how much the right-hand side simplifies, if one performs inte-
grations by parts to remove derivatives of s, and a little further
 creative

∂f ∂s
manipulation. Also the left hand side may be written as ,
∂t s ∂E
where the subscript denotes a derivative at fixed s, but then we divide
by the second factor to reach Eq.(11).
152 14 Two-Body Relaxation

There is no term on the left caused by the slowly varying potential, as

the actions are adiabatic invariants, and the phase-space volume factor
analogous to the partial derivative in Box 2, Eq.(3) can be ignored; action-
angle variables are canonical, and so the factor is a constant.

Problems
1) In Fig.1 let r, θ, φ be spherical polar coordinates centred on the mas-
sive body, which has mass m1 . Suppose stars are treated as test
particles and approach at speed V on paths parallel to the axis θ = 0
when at a large distance. Show that the path of a star with impact
parameter p is given by
p2
r= ,
p sin θ + a(1 − cos θ)
where a = Gm1 /V 2 . (Note that two values of p, of opposite sign, cor-
respond to each value of the pair (r, θ).) By using conservation of flux
between paths with impact parameter p and p + dp, and conservation
of energy, show that the density at r, θ is given by
s  2
1 ∂r
|p| 1 +
ρ r ∂θ
=s ,
ρ0 2a ∂r
1+ r sin θ
r ∂p
where ρ0 is the density far upstream. Adding the results for the two
values of p, show that
ρ r2 sin2 θ + 2ra(1 − cos θ)
= q .
ρ0 r sin θ r2 sin2 θ + 4ra(1 − cos θ)
X
2) Write down an expression for (δE1 )2 which is analogous to Eq.(5).
By means of Box 1, Eqs.(3) and (4), remove all non-dominant terms.
For the case in which the distribution f (ṙ2 ) is isotropic and both
masses are equal, deduce that
Z
1X 32π 2 G2 m2 ln Λ 1 v1
(δE1 )2 = ( v24 f (v2 )dv2
t 3 v1 0
Z ∞
+ v12 v2 f (v2 )dv2 ).
v1

3) Show that the length scale in Fig.1 is proportional to the mass, m1 ,

of the fast-moving star (when this is large). Deduce that the mass
of the wake is proportional to m31 and that the deceleration varies as
m1 . What would the corresponding result be for particles moving in a
 Relaxation and Orbital Motion 153

plane, or for a system of rods extended orthogonally to their common

plane of motion? By a simple extension of the argument, obtain the
asymptotic form for the dependence on v1 , when this is large. Compare
with Eq.(3).
4) Compute the result of Problem 2 for a Maxwellian distribution
!
n2 v22
f (v2 ) = exp − .
(2πσ2 )3/2 2σ2
Note also that, when non-dominant terms are neglected, the result
1X
can be used to establish the value of (δv1k )2 , where δv1k is the
t P
component of δ ṙ1 parallel to ṙ1 . Hence find the time when (δv1k )2 =
σ2 , assuming that v12 = 3σ2 . Compare your result with Eq.(8).
5) For a system in virial equilibrium, estimate the mean square speed by
using the fact that, for a wide range of models, the potential energy is
approximately −0.4GM 2 /rh , where rh is the half-mass radius (Spitzer
1987, p.13f; cf. also Fig.8.3 and Table 8.1). By adopting the mean
density inside rh as a typical density, use Eq.(8) to derive Eq.(9). De-
N
duce that trh ∼ tcr , where tcr is the crossing time. (This implies
ln Λ
that a star moves a distance much greater than the size of the system
before its energy is significantly altered.)
6) Starting from the result of Problem 2, obtain the formula for DEE
(cf. Box 3, Eq.(3)).
7) Verify that a Maxwell-Boltzmann distribution f1 = f2 ∝ exp(−βE)
is an equilibrium solution of Eq.(11).
8) Suppose that a system evolves at constant total energy according to
the evaporative model (Eq.(9.3)). Using results of Problem 5, show
that the current relaxation time may be written as
 7/2
N ln Λ0
tr = tr0 ,
N0 ln Λ
where a subscript 0 denotes an initial value. If the variation of the
2
Coulomb logarithm is neglected, show that N → 0 as t → tr (0).
7µ
 15
From Kepler to Kustaanheimo

The previous chapter was concerned with the consequences of two-body

interactions, but made use of nothing more than an approximate solution
of the two-body problem. Here we consider the classical two-body prob-
lem without approximations. It is one of the oldest solved problems of
dynamics, and so, as we mentioned in the preface, it is no longer really a
problem. Yet its structure is of enduring interest, and offers new surprises
each time we view it from a fresh angle.
Along with the simple harmonic oscillator, the Kepler problem is to
dynamics what the Platonic solids are to geometry. And, just as there
is a duality among the latter (for example the cube, with 6 faces and 8
vertices, is dual to the octahedron, with 6 vertices and 8 faces), we shall
see that there is an intimate link between these two dynamical problems.
This chapter may look self-indulgent compared with the serious issues of
stellar dynamics in the surrounding chapters, and should perhaps be in
a box of its own, but in fact some of the results we shall survey have im-
portant applications to the million-body problem. The reason is that we
shall be taking a close look at the singularity of the two-body equations,
where numerical methods cause a lot of trouble.

Removing the collision singularity

Consider first the one-dimensional Kepler problem. With a suitable
scaling, the equation of motion is
ẍ = −1/x2 . (1)
This equation is singular, corresponding to a collision in the Kepler prob-
p is routine. The energy is h = ẋ /2 − 1/x if
lem. Even so, its solution 2

x > 0, whence ẋ = ± 2(h + 1/x). This is still singular when x = 0,

but we can sidestepR the singularity
p in the following way. The solution re-
quires the integral dx/ h + 1/x, and if h < 0 this suggests the change

154
 Removing the collision singularity 155

of variable
x = − sin2 θ/h, (2)
and then we find that
(−2h)3/2
θ̇ = . (3)
4 sin2 θ
if we make an appropriate choice of sign. Hence
2θ + sin 2θ = (−2h)3/2(t − t0 ), (4)
where t0 is a constant. (This is a form of Kepler’s equation.)
Eqs.(2) and (4) are a parametric form of solution. They may not look
very elegant, but this solution has some remarkable features. First, the
singularity at x = 0 is no longer evident: the solution can be computed
for any time. Second, if we think of θ as the independent variable, we
may also write the solution (Eq.(2)) as x = −(1 − cos 2θ)/(2h), which is
simple harmonic motion with the centre of attraction displaced from the
origin. Therefore, if we transform from t to θ in the equation of motion
we may expect to arrive at the equation of a displaced simple harmonic
oscillator. In fact Eqs.(2) and (3) show
p that this transformation is given
in differential form by dt/dθ = x −2/h. Though the square root is
unpleasant, we can easily scale θ to remove it.
All this suggests the introduction of a new independent variable τ de-
fined by
dt/dτ = r, (5)
where r = |x|. Though τ looks like a scaled version of θ its status is ac-
tually quite different: τ will play the role of a transformed time-variable,
whereas θ was a transformed spatial coordinate. Anyway, the transfor-
mation of Eq.(1) is very simple, and indeed leads to an equation of the
expected form:
x00 = 2hx + 1,
where a prime means d/dτ . This also holds when h ≥ 0, and again the
remarkable feature of this equation is the removal of the singularity at
x = 0. Though the expression for h is singular at x = 0, it need not be
evaluated anywhere except for the initial conditions, since it is constant.
The procedure we have outlined is easily extended to the perturbed
case, and to three dimensions. Indeed it yields a practical equation for
the solution of the perturbed two-body problem (Burdet 1967). If we
simply transform from t to τ , defined by Eq.(5), it is a straightforward
exercise to show that the three-dimensional unperturbed Kepler problem
transforms to
r00 = 2hr − e, (6)
156 15 From Kepler to Kustaanheimo

where
e = ṙ × (r × ṙ) − r/r. (7)
This vector, like h, is a constant of the motion. It is usually called the
Lenz or Runge-Lenz vector. Again we notice that Eq.(6) is displaced
simple harmonic motion.
Most of this chapter is concerned with a different link between the
Kepler problem and the simple harmonic oscillator. Again it is suggested
by thepsolution of the one-dimensional problem, Eq.(2), where we notice
that x(θ) also yields simple harmonic oscillation (this time without
displacement) if θ is used as independent variable. This suggests the
introduction of a new dependent variable X defined by
x = X 2, (8)
in addition to the transformation from t to τ defined by Eq.(5). The
transformations are most easily carried out in the energy equation, and
then differentiation leads to the expected equation
d2 X 1
2
= hX, (9)
dτ 2
Note that the frequency is half that obtained in the previous transfor-
mation. In a way this is obvious from Eq.(8): one oscillation of X gives
two oscillations of x. There is also, however, a geometric significance to
this observation, as will become clear as we proceed. It also accounts for
the fact that Eqs.(6) and (7) are not the preferred route for dealing with
two-body encounters in N -body simulations: Eq.(9) and its descendents
are, in this sense, twice as efficient.
The tricky step with the transformation leading to Eq.(9) is its exten-
sion to three dimensions, and we proceed by way of the planar case. In
two dimensions Eq.(8) is easily generalised if we think of a 2-vector as
a complex number, z. Then the equation of motion is z̈ = −z/|z|3 , the
energy is h = |ż|2 /2 − 1/|z|, and we define new variables by z = Z 2 and
dt/dτ = |z|. Direct substitution yields the rather unpromising equation
2
d2 Z Z 1 dZ
=− + ,
dτ 2 2|Z|2 Z̄ dτ
but we see that the right side simplifies if we transform also the expression
for h. In fact if a prime denotes differentiation with respect to τ , we find
that Z 00 = hZ/2, in precise analogy with Eq.(9).
The transformation we introduced (a step due to Levi Civita 1906)
has a strong geometrical flavour. Without this transformation, Kepler
orbits passing close to the origin are nearly parabolas. The singularity
in the transformation z = Z 2 at z = 0 approximately straightens these
 KS regularisation 157

curves out∗ . Indeed when h = 0 the curves are exactly parabolas, and the
motion in the Z-plane is uniform and rectilinear (Fig.1). Furthermore,
if the orbit has h < 0, one revolution in the Z plane corresponds to two
in the z plane: after a single revolution in the z-plane the Z-orbit moves
from one Riemann sheet of the transformation to the other, and after two
revolutions it moves back again.

KS regularisation
It took about 50 years before Levi Civita’s transformation was extended
to the practical case of the 3-dimensional Kepler problem. This delay
seems remarkable when one learns that the requisite tool (quaternions)
had been invented as long ago as 1843. Furthermore, they were invented
by Hamilton (1853) in order to provide a means of performing algebra
with 3-vectors; our requirements here are not dissimilar, i.e. to be able
to take the square root of a 3-vector. Another point of historical interest
is that, when the 3-dimensional extension eventually was carried out, it
was not done in the language of quaternions, but in terms of spinors.
So strongly did this approach hold sway that it was even said that the
requisite transformations could not be expressed in quaternions. Only in
1991† was the full story told, almost 150 years after Hamilton’s discovery.

z plane Z plane
c
c
C
BF D B

E
CG b AE b A

DH F H
G
a a
Fig. 1. Keplerian orbits in two dimensions (left) and in the transformed plane
(right). One circular and one parabolic orbit are shown, and the letters in the
two diagrams show corresponding points

∗
We might say that it “unfolds” the orbit, but this word is loaded mathematically.
It would be more accurate to say that it “blows up” the singularity, as McGehee
(1974) did so famously for the collinear 3-body problem. See also Waldvogel (1982)
for a similar approach to the non-collinear problem.
†
Neutsch (1991). The account in Vivarelli (1985) omits only the role of our Eq.(15)
158 15 From Kepler to Kustaanheimo

In the following account of the theory we shall try to convince the reader
that he could have found his way there with only the most rudimentary
knowledge of quaternions (as in the outline solution to Problem 1, cf.
Appendix B).
The usual convention is that a three-vector (x, y, z) corresponds to the
quaternion ix + jy + kz, where {1, i, j, k} are four basis vectors of the
space of quaternions. Now the real form of Levi Civita’s transform, in
obvious notation, is x = X 2 − Y 2 , y = 2XY . Let us try to treat the
space of quaternions spanned by i and j in the same way, i.e. we write
ix + jy = i(X 2 − Y 2 ) + 2jXY . The right side is easily factored as
i(X − kY )2 , which can be written in the form

ix + jy = (X + kY )i(X − kY ). (10)

The latter form is very convenient for handling rotations, and, now
that we know how to handle one plane with quaternions, we can handle
any plane, simply by rotating it into position. For example, to rotate a
quaternion ix + jy into the quaternion ix + ky (in the plane spanned by
i and k) we pre- and post-multiply as follows (cf. Problem 1):

1+i 1−i
√ (ix + jy) √ = ix + ky. (11)
2 2

This fits beautifully with Eq.(10), as we deduce that

1 1
ix + ky = √ (X + iX − jY + kY )i √ (X − iX + jY − kY ).
2 2

In general, this suggests that any quaternion representing a vector should

be written as

ix + jy + kz = qiq̄, (12)

where q is a quaternion and q̄ its conjugate (obtained by changing the

sign of the i-, j- and k-components).
There is a difficulty here, because this map from quaternions to physical
space is many-to-one. This implies that the initial conditions of a Kepler
orbit do not specify unique initial conditions for motion in the space
of quaternions. For example, each point in real space corresponds to a
“circle” of quaternions, given by

q = q0 exp(iφ), (13)
 KS regularisation 159

where φ is any angle and q0 any quaternion satisfying Eq.(12)∗ . Further-

more, the relation between the velocities is
iẋ + j ẏ + k ż = q̇iq̄ + qiq̄˙ , (14)
which has to be solved for the four components of q˙ in terms of the three
quantities ẋ, ẏ and ż. Another condition would be needed in order to
specify q̇ uniquely. Again we can use the special case of motion in the
x, y plane as a guide. In this case we may take q = X + kY , and it is
easily seen that the two terms on the right of Eq.(14) are equal. This
suggests that the extra condition to be added is
q̇iq̄ − qiq̄˙ = 0. (15)
There are three other reasons why this choice is appealing. First, it
has a lovely geometrical interpretation, as it implies that q˙ is orthogonal
to the tangent to the curve given parametrically by Eq.(13). Second, we
are hoping that the equation of motion in q-space will be
q 00 = hq/2, (16)
and it turns out that the expression on the left of Eq.(15) is a first integral
of this equation; thus if Eq.(15) holds initially it holds forever. Third,
Eq.(15) simplifies Eq.(14).
Now it is straightforward, if slightly tedious, to demonstrate the link
between the simple harmonic oscillator equation in quaternions and the
Kepler problem in real space. We assume that (15) and (16) hold, and
transform from q to r = (x, y, z) using Eq.(12), and from τ to t using
the obvious definition dt/dτ = |r|. The result is the familiar equation of
motion of the Kepler problem.
This transformation has many interesting properties (Stiefel & Scheifele
1971), of which we mention two. In two dimensions we saw that the
transform variable Z moved from one Riemann sheet to another (and
back again) as z traversed one orbit. In the 3-dimensional case the cor-
responding relation has some added complexity. After one complete rev-
olution in space, the moving ring of quaternions returns to its original
position, corresponding to the quaternion q(0). However at this time the
quaternion that is the solution of Eq.(16) is −q(0), and it is only after a
second revolution in space that it returns to its original position. Fig.2
(Problem 3) shows two of the loops for motion in space on a circle of
unit radius. In this case the quaternions sit on the unit sphere, which
is three-dimensional, and there is a familiar map which takes this sphere
into ordinary 3-dimensional real space.

∗
Note the analogy with the local Abelian U(1)-gauge invariance of quantum electro-
dynamics.
160 15 From Kepler to Kustaanheimo

The second interesting point we shall mention stems from an obvious

integral of Eq.(16), viz. e = hqiq̄ − 2q 0 iq̄ 0 . This resembles the energy
integral of the harmonic oscillator, but the latter is obtained (up to a
numerical factor) by omitting the factors i. At any rate it is only another
tedious but mechanical exercise to show that this is equivalent to the
famous Runge-Lenz vector defined in Eq.(7). Its appearance in the Kepler
problem is a good deal less obvious.
While we have Eq.(12) in front of us let us write it out in coordinates.
If ui , i = 1...4 are the four coordinates of q we find immediately that
x = u21 + u22 − u23 − u24 ,
y = 2(u1 u4 + u2 u3 ), (17)
z = 2(−u1 u3 + u2 u4 ).
These differ in only one or two inconsequential signs from the so-called
KS transformation, named after its discoverers Kustaanheimo and Stiefel
(1965). They found their transformation via spinors, which are quater-
nions in modern dress. Spinors entered physics in the 1930s, to describe
elementary particles with spin one half. One of the remarkable prop-
erties of this theory is that a 360◦ rotation in space corresponds to a
sign-reversal of the spin amplitudes, i.e. a rotation by 180◦ (cf. Feynman
1965). This is reminiscent of a property of Levi Civita’s regularisation to
which we already drew attention. In fact even the prolific Euler would
have recognised these formulae and their characteristics; they occur in the
kinematics of a rigid body when expressed in terms of the four Euler pa-
rameters (Euler 1776; cf. Goldstein 1980, who spells out the relation with
spinors). Truly, each generation must reinvent for itself the discoveries of
the past.
If the next generation wants to try something new it might look be-
yond the quaternions. Complex numbers and quaternions are examples
of Clifford algebras (see Peacock 1999 for a snappy introduction). After
quaternions the next in line is the 8-dimensional algebra of Cayley num-
bers, where multiplication is not even associative (i.e. it matters where
you place the brackets; see, for example, Lambert & Kibler 1988). No-
one has yet put them to good use in stellar dynamics, though there is
a known regularisation of the five-dimensional Kepler problem using 8
regularising variables (see Mladenov 1991)!

Numerical regularisation
Though collisions of point masses should not occur in computer simu-
lations of the million-problem, it is a common experience that numerical
methods behave badly even in the vicinity of a singularity. In N -body
simulations a moderately eccentric close binary is close enough to cause
 Numerical regularisation 161

trouble. The most common approach is to replace the position vectors of

the two components, r1 and r2 , by their relative position vector r = r1 −r2
and the position vector of their barycentre, R, and then to apply the KS
transformation to r. Binaries form and disintegrate, however, and so
codes include appropriate procedures for deciding when regularisation is
desirable and when it should terminate.
This technique is very successful, but it has never been clear just which
aspects of the technique are the vital ones. Even the transformation to
r and R alone does much to combat rounding error. It has been found
that simply ensuring that the integration of a close binary is carried out
by a time-reversible algorithm (without the KS transformation) produces
roughly comparable performance (Hut et al. 1995).
Another property of an integrator which seems relevant is symplectic-
ness (Chapter 22). This is a property of the N -body equations which not
all numerical integrators preserve. But there do exist very simple sym-
plectic integrators which can integrate even a collision orbit (Mikkola
& Tanikawa 1999), albeit with phase errors. In their method it is not
a coordinate transformation that is pulled out of the hat but a whole
new Hamiltonian. Let r be the position vector for Kepler motion, p
the conjugate momentum and h the energy. Then, for suitable scal-
ing, the usual Hamiltonian is H = p2 /2 − 1/r and h is the value of H.
Now it is easily seen (but nevertheless astonishing) that the Hamilto-
nian Λ = ln(p2 /2 − h) + ln r gives canonical equations which correspond
to the same Kepler motion, with an independent variable τ defined by
dτ /dt = p2 /2 − h. A leapfrog integrator of these equations (Chapter 22)
preserves everything except phase, even for a collision orbit.

Problems
1) If α = ix + jy, show that x = −(αi + iα)/2 and ky = (iα − αi)/2.
Hence obtain Eq. (11).
2) Let qa = ia1 + ja2 + ka3 , where a1 , a2 and a3 are real, and similarly
define qb = ib1 + jb2 + kb3 . Show that qa qb = −a.b + [a, b], where a
is the three vector (a1 , a2 , a3 ), b is defined similarly, a.b is the usual
dot (scalar) product, and [a, b] is the quaternion corresponding to the
usual vector product.
3) A quaternionic view of circular motion
(i) By squaring Eq.(12) show that |r| = kqk2 , with obvious notation.
(This shows that motion on the unit sphere |r| = 1 in space maps to
the sphere kqk = 1 of unit quaternions. Now we construct a way of
visualising the latter.)
(ii) Show that stereographic projection from the quaternion 1 to the
space of imaginary quaternions (spanned by i, j, k) maps q0 +iq1 +jq2 +
162 15 From Kepler to Kustaanheimo

kq3 to (iq1 +jq2 +kq3 )/(1−q0 ). If the unit sphere (q02 +q12 )+(q22 +q32 ) =
1 is parameterised in an obvious way as

q0 = cos θ cos φ
q1 = cos θ sin φ
(18)
q2 = sin θ cos ψ
q3 = sin θ sin ψ,

show that this point maps to the quaternion equivalent to the 3-vector

R = (cos θ sin φ, sin θ cos ψ, sin θ sin ψ)/(1 − cos θ cos φ). (19)

This gives us a way of visualising a unit quaternion, but note that the
coordinates are singular when θ is a multiple of π/2.
(iii) Show that the quaternion given by Eq.(18) maps, via Eq.(12),
to the quaternion equivalent to the 3-vector

r = (cos 2θ, sin 2θ sin(φ + ψ), − sin 2θ cos(φ + ψ)).

(Thus for unit vectors and quaternions, the KS map can be thought
of as the transformation from R to r, given by these formulae.)
(iv) Show that the point

r = (cos 2θ, sin 2θ, 0) (20)

maps (under the inverse KS map) to the curve C defined by

R = (cos θ sin φ, sin θ sin φ, sin θ cos φ)/(1 − cos θ cos φ), 0 < φ < 2π

(cf. Fig.2. Incidentally, the final result is correct even where θ is a

multiple of π/2, though a separate argument is needed.)
4) Any quaternion may be represented as in Eq.(18) if the factors cos θ
and sin θ are replaced by variables r and ρ, respectively. Hence es-
tablish a link between the KS-like transformation Eq.(17) and the
parabolic coordinates used in the quantum mechanical treatment of
the Stark effect:
p
x= ξη cos φ,
p
y= ξη sin φ,
1
z = (ξ − η)
2
(cf. Landau & Lifshitz 1977). Taken in reverse this is a pleasing route
to the KS transformation (Yoshida 1982).
 Numerical regularisation 163

Fig. 2. Circular Kepler motion in the space of unit quaternions, represented in

3-dimensional space as in Eq.(19). Each point on a Kepler motion corresponds
to a circle in the space of unit quaternions, and the figure shows how this circle
sweeps out a surface as the Kepler motion describes one quarter of an orbit. The
edges of the surface shown here correspond to Eq.(20) for θ = π/4 and 3π/4.
Note that these two curves loop through one another.

5) For the three-body problem in the plane with equal masses, let the
complex numbers xi , i = 1, 2, 3, be the position vectors of the three
bodies. Interpret geometrically the numbers
z1 = x3 − x2
1
z2 = x1 − (x2 + x3 ).
2
Finally define three real numbers ui by
u1 = |z1 |2 − |z2 |2
u2 + iu3 = 2z̄1 z2 .
Show that the ui are independent of rotation (and are therefore an
appropriate parameterisation of the shape of the triangle formed by the
three bodies), and that the map (z1 , z2 ) → (u1 , u2 , u3 ) is essentially
the KS map.
 PART V.
GRAVOTHERMODYNAMICS:
N = 106

The following three chapters begin the application of earlier results to

the million-body problem itself. Chapter 16 discusses two effects of two-
body gravitational encounters: escape and mass segregation. The first
of these actually develops the theory of two-body relaxation further, as
we cannot, in this context, approximate encounters by any small-angle
scattering approximation. This approach is, however, applicable to mass
segregation, which is an effect of the tendency to equipartition of energies
in two-body encounters. It also has an important influence on the stability
of the million-body problem (the “mass stratification instability”).
Chapter 17 is also concerned with instability, but an instability which
even exhibits itself in systems with equal masses. It was first discovered
through a remarkable thermodynamic result obtained by Antonov, which
helps to explain the relevance of the term “gravothermodynamics”. This
chapter deals with extrema of the entropy, and the stability of linear
series of equilibria.
Chapter 18 follows up the previous two chapters by tracing the conse-
quences of the mass stratification and gravothermal instabilities. This is
the process referred to as core collapse. In other contexts this would be
referred to as an example of “finite-time blow-up” and, in common with
other examples of this behaviour, it can be described asymptotically by
approximate self-similar solutions of the governing equations.

164
 16
Escape and Mass Segregation

This chapter deals with two effects of two-body encounters. In a general

way this process was discussed in Chapter 14, but now we begin to study
the effects on the system itself. Furthermore, the theory described there
is applicable only to one of the two topics of this chapter. That theory
describes the cumulative effects of many weak scatterings, and is perfectly
adequate for an understanding of mass segregation. The escape of stars
from an isolated stellar system, however, is controlled by single, more
energetic encounters, and a better theory is necessary. The theory we
shall describe is illustrative of a whole body of theory which improves on
that of Chapter 14, though for most purposes (e.g. mass segregation) the
improvements are unimportant.

Escape

We consider the case of an isolated stellar system. For this case a star
with speed v will almost certainly escape if v2 /2 + φ > 0, where φ is the
smoothed potential of the system at the location of the star, with the
convention that φ → 0 at infinity. The exceptions are binary components
(for which the true potential differs significantly from the smoothed po-
tential), and an escaper which, on its way out, interacts with another
star in such a way that its energy once again becomes negative. The
latter possibility is rare in large systems (King 1959), precisely because
two-body relaxation takes place on a much longer time scale than orbital
motions (Chapter 14).
Several mechanisms may be responsible for the generation of an esca-
per. It may result from three- or four-body encounters involving binaries
(Chapter 23). It may happen that loss of mass by stars through their
internal evolution makes the potential well of the cluster sufficiently shal-
low that weakly bound stars become escapers. Here, however, we consider

165
166 16 Escape and Mass Segregation

the most classical escape mechanism: two-body interactions, in which the

slingshot effect of an encounter leads to escape of one participant.
The condition for escape is that
v1 .δv1 + (δv1 )2 /2 > −E1 , (1)
where v1 is the initial velocity of the star, δv1 is the change in this
velocity over the encounter, and E1 is its initial energy. The change δv1
is given by Eq.(4) of Box 14.1, where the quantity δ ṙ on the right side, in
the small-angle scattering approximation, is given by Eq.(3) in the same
place. In the derivation of that equation it was stated that the relative
velocity ṙ rotates through an angle α, for which an approximate formula
was given, but a more accurate formula (which is exact for two-body
Keplerian motion) is
tan(α/2) = G(m1 + m2 )/(pV 2 ). (2)
It follows by simple geometry that
2m2 V α
δv1 = sin δ̂v1 , (3)
m1 + m2 2
where
α α
δ̂v1 = −p̂ cos − V̂ sin . (4)
2 2
(Recall that a hat denotes a unit vector, V is the initial relative velocity
of the participants, and p̂ is a unit vector orthogonal to V.)
An important result can be deduced right away, though it is also ob-
vious intuitively. If m2 is negligible, then so is δv1 , and so escape of the
star with mass m1 is unlikely. More generally, the likelihood of escape is
greater, the greater is the mass m2 . This already shows that two-body
encounters are likely to lead to the preferential escape of the stars of lower
mass, which is an extreme form of mass segregation.
Now we consider the case of equal masses. In order to derive the escape
rate a slightly intricate sequence of transformations, due to Michel Hénon
(1960), must be carried out. As usual we relegate most of the details to
a box (Box 1), but we would like to draw the attention of even the most
hurried reader to Eq.(4) there, which is an amazingly simple formula for
the rate at which changes of velocity of any given magnitude occur. The
next step of the calculation is to add the condition that the change must
be big enough, and eventually we obtain the result
Z
8π 2 G2 m2
Q= √ (2E1 + v22 )3/2 v2 f (v2 )dv2 (5)
3v1 E12 −2E1

for the rate of escape of a star of energy E1 and speed v1 .

 Escape 167

Box 1: Computing the escape rate from a star cluster

We first compute the rate at which a star of given mass, velocity and
position experiences encounters which would cause it to escape. This is
Z
Q= V f (v2 )|dp̂|pdpd3 v2 , (1)

the integral being taken over the domain defined in a very implicit way
by text Eq.(1); here f is the distribution of stellar velocities, normalised
to the space density of stars. To make progress, we notice first (from
text Eqs.(2) and (4)) that, for given initial velocities of the two stars,
δ̂v1 is determined by p̂ and p. In fact in Eq.(1) we can transform the
integration with respect to these two variables to give
Z
2 2 f (v2 ) 2
Q = 4G m d δ̂v1 d3 v2 , (2)
(δv1 )3
where we have also used text Eq.(3).
Escape obviously depends on the magnitude of δv1 , which we denote
by δv1 , though it is not the change in v1 . Now from text Eqs.(3) and (4)
it follows that
α
(v1 − v2 ).δv1 = −V 2 sin2 ,
2
whence
v2k = v1 .δ̂v1 + δv1 , (3)

where the subscript k denotes the component parallel to δ̂v1 . This sug-
gests the use of cylindrical polar coordinates for v2 , with δ̂v1 as polar
axis. But since we assume that the distribution of v2 is isotropic it is also
advisable to adopt v2 as one of the variables. In fact an unorthodox but
useful choice is v2k (or δv1 ), v2 and φ, where φ defines the orientation of
v2 around δ̂v1 (i.e. the cylindrical polar angle), and so from Eqs.(2) and
(3) we find that
Z
f (v2 ) 2
Q = 4G2 m2 d δ̂v1 dδv1 dφv2 dv2 , (4)
(δv1 )3
though we have to add the restriction v2 > |v2k |, i.e.

|v1 .δ̂v1 + δv1 | < v2 (5)

by Eq.(3). The integration with respect to φ is trivial.
168 16 Escape and Mass Segregation

Box 1 (continued)
So far we have not used the fact that we wish to compute the escape
rate. In fact this expression gives a general result for the rate of encoun-
ters leading to a given change in the velocity of the star. Now we add
the condition, text Eq.(1), and the integral with respect to δv1 is most
easily carried out using spherical polar coordinates with v̂1 as polar axis.
If the polar angle is denoted by θ, the conditions (5) and text Eq.(1) may
be expressed as
q
−v1 cos θ + v12 cos2 θ − 2E1 < δv1 < −v1 cos θ + v2 . (6)
Integration with respect to δv1 now yields
Z !
1 1
Q = 4πG m 2 2
f (v2 )v2 dv2 2π sin θdθ 2 − 2 ,
δv1min δv1max
where δv1min/max are the limits in Eq.(6), and the integration with re-
spect to θ is restricted to the range cos2 θ < (2E1 + v22 )/v12 . In this way
we arrive at text Eq.(5).

In order to evaluate the total escape rate over the entire cluster, we
must integrate over the initial velocity and position of the escaper. These
integrations hold no terrors, and the result may be expressed as
dN 256π 4 G2 m2 √
=− 2
dt 3
Z Z (6)
2 f1 f2 (E1 + E2 − φ)3/2
r dr dE1 dE2 ,
E1 +E2 >φ E12
where the distribution function f has been written as a function of the
energy E for both stars, and φ(r) is now the potential at radius r.
This result is more elegant than its proof, which quickly becomes in-
tractable if one sets a foot wrong anywhere along the route. It also reveals
the competing effects which control the net escape rate. The denominator
favours stars which are nearly unbound, as does the large spatial volume
which such stars occupy. On the other hand f is very small near E = 0.
Nevertheless, in models for which the calculations have been carried out,
the bulk of escapers are already weakly bound just prior to the encounter
which ejects them.
Hénon’s formula appears to work well when the assumptions behind it
are satisfied, i.e. the distribution of velocities is isotropic. It differs qual-
itatively, however, from many other formulae for the escape rate (those
 Mass segregation 169

summarised in Giersz & Heggie 1994a, for example), which are based on
the time of relaxation. For example, for a Plummer model Hénon’s result
is Ṅ = −0.00088N/(trh ln Λ), while the classic result of Spitzer & Härm
(1958) for a square-well potential is Ṅ = −0.0085N/tr (Spitzer 1987).
The difference in the coefficient (even though this cannot reasonably be
explained by different definitions of the relaxation time) is not the issue.
The issue of principle is indicated by the fact that the Coulomb loga-
rithm only appears in one of the formulae. Hénon argued that escape
by diffusion does not work, because an increasingly weakly bound star
spends longer and longer on an orbit well outside the core, and relaxation
is ineffective; therefore escape is dominated by single-encounter events.
In fact a synthesis of the two basic ideas seems to work best: relaxation
gives rise to a halo of loosely bound particles (on the relaxation time
scale), which can then become unbound in a single passage through the
core (Spitzer & Shapiro 1972).
What is certain is that, in evolving systems, the spatial structure and
the velocity distribution of the stars both change with time. Even if we
continue to assume that the distribution of velocities is isotropic, it is
found that the rate of escape increases as the core contracts and becomes
denser (Chapter 17; see Giersz & Heggie 1994a). In fact, the distribution
also becomes increasingly anisotropic (if it was isotropic initially), and
this enhances the escape rate, in the sense that, in a series of models
which have the same spatial structure but different anisotropy, the escape
rate increases with the amount of radial anisotropy. For these reasons,
if one computes Hénon’s escape rate for a Plummer model (Problem 1),
it is a poor guide to the escape rate from a system which starts life as a
Plummer model.
This discussion summarises the case of isolated systems of equal masses.
As we have seen, escape favours stars of low mass. For a system consisting
of equal numbers of stars of some fixed mass and some very small mass,
the low-mass stars escape about 30 times as fast as the high-mass stars
(Hénon 1969). The total escape rate is enhanced by a similar factor
(compared to the case of equal masses) for more realistic mass functions.
When we place the cluster in a tidal field the rate of escape is further
enhanced, but it also becomes a more complicated dynamical problem
(Chapter 12).

Mass segregation
The preferential escape of stars of low mass is an extreme manifestation
of a tendency in all two-body encounters between stars of different mass,
even when neither escapes. It is most easily understood as the tendency
towards equipartition of kinetic energies, which ideally leads to the con-
170 16 Escape and Mass Segregation

dition m1 hv12 i = m2 hv22 i: heavy stars move more slowly, on average. In

stellar systems, the massive stars then drop lower in the potential well,
while the stars of smaller mass move out and may even escape. In other
words, the primary mechanism is the tendency to equipartition of kinetic
energies, which leads to segregation of masses by energy, and this in turn
leads to spatial segregation (Fig.1).
In more detail, the tendency to equipartition of kinetic energies may be
seen as the result of the competing tendencies of relaxation and dynam-
ical friction, which give rise to the two terms on the right of Eq.(14.5).
The first term, which decreases the energy on average, is proportional to
m2 (m1 + m2 ), by Eq.(14.3), whereas by Eq.(14.4) the second varies as
m22 ; it is this term which causes the kinetic energy to increase on average.
A balance between these two effects can only be struck if stars of high
mass move more slowly on average.
This idea may be quantified, in terms of the kinetic energy per unit
mass, E1 ∗ , using these three equations from Chapter 14. They lead to
the result
 m Z
1X 1
δE1 = 4πG2 m2 log γN − f (v2 )d3 v2
t v1 v2 <v1
Z  (7)
1
+ m2 f (v2 )d3 v2 .
v2 >v1 v2

Multiplying by m1 and the distribution of velocities of stars of mass m1 ,

and integrating with respect to v1 , we obtain an expression for the mean
rate of change of the mean kinetic energy of stars of mass m1 , which is
  √
d 4 3πG2 m2 n2
(m1 E1 ) = (m2 hE2 i − m1 hE1 i)
dt (hE1 i + hE2 i)3/2

in the case of Maxwellian distributions; n2 is the number density of stars

of the second species. This strikingly simple result shows that encounters
drive the species towards equipartition – if m1 hE1 i is smaller than m2 hE2 i
then the former increases and the latter decreases. We can also easily
estimate the time scale for this to happen; by computing the rate of
change of m2 hE2 i − m1 hE1 i) we easily find that the e-folding time for the
tendency to equipartition is

(hE1 i + hE2 i)3/2

te = √ . (8)
4 3πG2 m1 m2 log(γN )(n1 + n2 )

∗
Earlier in the chapter E denoted the total energy (per unit mass)
 Mass segregation 171

Fig. 1. Mass segregation in an isolated N -body simulation. Each graph gives

the mean stellar mass within a “Lagrangian shell”, i.e. a shell containing a fixed
proportion of the total mass. This plot shows “projected” data, i.e. the shell is
a cylinder along the line of sight; its axis passes through the densest part of the
system. Results are averaged over a large number of simulations, each having
500 particles. The initial model is a Plummer model, the distribution of masses
is dN (m) ∝ m−2.5 dm, and the ratio of the extreme masses is 37.5. Initially
all masses have the same spatial distribution, and the mean mass is the same
(within statistical fluctuations) in all shells. Very quickly, however, the mean
mass becomes much higher within the central part of the cluster, and somewhat
smaller further out. After this phase of rapid mass segregation, however, the
subsequent evolution is very slow, for reasons which are not understood. Even
at late times there are significant departures from equipartition well within the
half-mass radius, and these do not diminish. From Giersz & Heggie (1996).
172 16 Escape and Mass Segregation

This result bears an interesting resemblance to the relaxation time

(Chapter 14). For many galactic globular clusters the latter is consider-
ably shorter than the age of the cluster, and so we might expect to see
evidence for mass segregation in them. For a long time the observational
difficulty of counting low-mass (and therefore faint) stars in globular clus-
ters was the main obstacle, but nowadays the evidence is robust (e.g. King
et al. 1995).
These results show that encounters tend to drive the system towards
equipartition, which means that the “temperatures” of all species are
the same. It does not follow, however, that this is what happens, be-
cause the distribution of velocities is affected by another process: the
change of potential caused by evolution of the spatial distribution of the
stars (cf. Eq.(11.2)). Calculations with various models (e.g. Inagaki
& Wiyanto 1984) show that different species can either move towards
or away from equipartition in different circumstances. Indeed there are
results which show that equipartition is impossible, under reasonable as-
sumptions (Saslaw & De Young 1971).
To see physically what the difficulty is, let us turn first to an analogous
problem for a system of equal masses. Here encounters tend to set up
a Maxwellian distribution of velocities, but the system cannot reach the
necessary isothermal equilibrium. This can be understood from several
points of view. On the one hand we know that isothermal systems must
have infinite mass (Chapter 8). On the other, a Maxwellian distribu-
tion must allow arbitrarily high speeds for a small fraction of stars, but
stars with speeds above the escape speed cannot remain in the system.
Finally, it is known that isothermal systems are unstable (Chapter 17).
How a system of equal gravitating point masses reconciles these features
with the demands placed on it by two-body encounters is the subject of
Chapter 18, but it already illustrates that the behaviour is likely to be
less straightforward than that of a perfect gas in a box.
Let us return to a system with stars of two different masses, and ask
how it responds to the encounters which are driving it towards equipar-
tition. For simplicity we imagine that, initially, the distribution of stellar
velocities at a given location in the system is the same for both species.
Let v2 be the mean square speed of both species at this time, and R
the “radius” of the system (e.g. the half-mass radius). We also suppose
that the heavier species (of individual mass m1 ) is a minor constituent of
the system, i.e. M1  M2 , where Mi denotes the total mass of the ith
species.
As encounters take place, the stars of the heavier species tend to lose
energy, and sink in the potential well of the lighter stars. They will have
reached equipartition by the time their mean square speed has fallen to
a value of order m2 v2 /m1 . If the potential well of the lighter stars is
 Mass segregation 173

modelled by a parabolic profile, as in a simple harmonic oscillator, it is

easily seen that the spatial distribution ofpthe heavier stars is confined
within a region whose size is of order R m2 /m1 . The total mass of
the two species within this region is thus M1 for the heavier species and
M2 (m2 /m1 )3/2 for the lighter species. Now it is no longer clear that the
lighter stars are the dominant mass in this region. If they are, then our
estimates are valid in order of magnitude, and approximate equipartition
may well be possible. If they are not, then the stars of the heavier species
will not only be moving in the potential well of the lighter species, but will
also be increasingly affected by their own self-gravity. If this happens, it
is less clear that they can contract sufficiently to bring themselves into
equipartition with the lighter stars. Indeed, a self-gravitating system
heats as it contracts (Chapter 5), and it may not be possible to reach the
low mean square speed required (cf. Fig.2).
These arguments suggest that it may not be possible to achieve equipar-
tition if
3/2
M1 >∼ M2 (m2 /m1 ) . (9)
More refined calculations confirm that a condition of this form (with
a suitable numerical coefficient) closely represents a real limit on the
composition of a self-gravitating system which can achieve equipartition.
The really interesting issue remaining is to find out what happens if this
criterion is violated. In that case the heavier species already becomes self-
gravitating before it has reached the low speeds required by equipartition.
It therefore heats up: interactions with the light species cause the heavy
stars to lose energy, but the resulting contraction in their own potential
well causes an increase in kinetic energy which exceeds the loss to the
lighter stars. This simply exacerbates the absence of equipartition, which
is thus self-perpetuating. Eventually, the heavy stars have contracted so
much, expelling lighter stars as they do so, that the latter no longer have
a significant role. From this point on the heavier species behave like a
self-gravitating system with stars of the same large mass. How such a
system behaves is the topic of the next chapter.
Before going on, however, it is worth remarking that we have (as usual)
considered an idealised problem: a stellar system with no initial (or “pri-
mordial”) mass segregation. There is evidence that the dynamical pro-
cesses we have discussed do not act fast enough to explain the mass
segregation observed in young clusters in Nature (e.g. Kontizas et al.
1998, Fischer et al. 1998). Presumably they were born that way.
174 16 Escape and Mass Segregation

Fig. 2. The tendency to equipartition in a Fokker-Planck simulation. The initial

model is a Plummer model, with an initial mass spectrum dN (m) ∝ m−3.5 dm
within the range 0.1M < m < 0.75M approximately, though it is represented
by a number of discrete components; in addition there is a population of heavier
“remnants” of mass 1.2M . The central “temperature” of each component,
which is proportional to the mean kinetic energy, is plotted against time. Initially
all components have the same mean square speed, and so T ∝ m. Just as in
Fig.1 there is an initial phase of rapid evolution (note the logarithmic scale of
t), after which the evolution is nearly self-similar, i.e. the ratio of temperature
between two species is nearly constant. Note that the heavier components nearly
reach equipartition in the initial phase, but the lighter particles remain cooler.
From Murphy et al. (1990)

Problems

1) Use Eq.(6) to compute the escape rate from a Plummer model of

mass M (Table 8.1). Deduce that the fraction lost per half-mass re-
laxation time (assuming that the escape rate does not change) is about
0.00088/ ln Λ.
2) Explain why the effective relaxation time at a distance r in the halo of
an isolated system varies as r−1/2 . Hence estimate the mass flux there
if the density is n (cf. Problem 9.7). If the mass flux is independent
of r, obtain the density profile.
3) Use the virial theorem to show that the mean square escape speed
is related to the mean square speed by hve2 i = hv2 i. Using this as an
 Mass segregation 175

estimate of the actual escape speed, find what fraction of a Maxwellian

distribution of velocities lies above it.
4) Consider a King model consisting of stars of mass m1 with distribution
function

exp(−m1 E) − 1 if E < 0,
f1 (E) =
0 if E > 0,
and central potential φc < 0. Consider adding a few stars of lower
mass m2 < m1 with distribution function

η(exp(−m2 E) − 1) if E < 0,
f2 (E) =
0 if E > 0,
where η is a small constant. Explain why (i) the two species are
nearly in equipartition if |m2 φc | is large enough, and (ii) the heavier
species dominates the mass density everywhere provided that η is small
enough. (Though this appears to contradict Eq.(9), that equation
depended on the assumption M1  M2 , which is broken here; cf.
Merritt 1981.)
 17
Gravothermal Instability

Can a million-body system be in equilibrium? More precisely, can a

model of a million-body system exhibit equilibrium? The answer de-
pends on the model and other conditions. But we already saw in Chap-
ter 8 that equilibrium models of gravitational many-body systems can be
constructed. Thus if it is modelled as a self-gravitating perfect gas it will
be in thermal equilibrium if its temperature is uniform. If it is modelled
by a Fokker-Planck or Boltzmann equation, then the equivalent condition
is that the single-particle distribution be Maxwellian. In both cases the
system is required to have infinite mass and extent.
These isothermal models are, in a strict sense, artificial, but they are of
great importance conceptually, and for other reasons. In order to make
progress in understanding them we shall replace one form of artificiality
with another. Instead of dealing with infinite systems, we shall enclose
our isothermal system in a spherical enclosure, which at least has the
merit of implying that our systems have finite mass and radius. We
shall suppose that the enclosure is rigid and spherical; in the N -body
model this means that stars bounce off it without loss of energy, while in
the gas or phase-space models, it implies that the enclosure is adiabatic.
We assume spherical symmetry, and that stars all have the same fixed
individual mass. We work entirely, however, with a perfect gas model.
As we saw in Chapter 8, an infinite isothermal model may be specified
by two parameters, and to specify a finite model we add a third, e.g. the
radius of the enclosure, re . For thermodynamic purposes it is sensible to
exchange the first two of these parameters for the total mass and energy,
denoted by M and E; these are constant, because we are interested in
systems within an impermeable, adiabatic enclosure. Thus for fixed M
and E there is a single-parameter family of solutions with parameter re .
All are in thermal equilibrium. Another useful way of parameterising the
models is in terms of the one dimensionless way of combining these three

176
 17 Gravothermal Instability 177

parameters, or something equivalent. In Chapter 8 we showed that the

isothermal model could be made dimensionless in terms of a scaled radius
which we called z. As our dimensionless parameter, therefore, we may
use ze , i.e. the value of z corresponding to re .
So far it is only the existence and structure of the equilibria that we
have been looking at. Now we take the obvious but rather far-reaching
step of looking at the stability of these models, by which we mean sta-
bility in the thermodynamic sense. Before we do so, it is perhaps worth
while making some reassuring remarks about the meaning of entropy in
this context, because one sometimes hears the assertion that “thermody-
namics doesn’t apply to self-gravitating systems” (Chapter 2). In fact
astrophysicists apply it all the time, in the context of stellar structure,
for example. Indeed it is not hard to show that the above equilibria
are determined by extremising the entropy, subject to the conditions of
constant mass, energy and volume, and the hydrostatic condition (Box
1).
As in conventional thermodynamics, the appropriate criterion for ther-
modynamic stability (for the case of a rigid adiabatic boundary) is based
on entropy considerations. As we have seen, equilibrium is characterised
by the vanishing of the first variation of the entropy, δS. The series of
models we have been discussing are precisely those determined by this
condition. Stability depends on the second variation of S in the vicin-
ity of an equilibrium. If this is negative definite, then the equilibrium is
stable, since any dissipative process can only increase the entropy, and
therefore no such process can cause the system to move indefinitely far
from the initial equilibrium. If there are variations in the structure of the
model for which the second variation δ 2 S is positive, then the equilibrium
is unstable. Our aim, then will be to examine the thermal stability of the
series of equilibria we have found.
There are some situations where the stability of an enclosed isother-
mal model is easily determined. Consider, for instance, the case ze  1.
Fig.8.1 shows that the density is nearly uniform; the reason is that the
effect of gravity is tiny in this case, and the gas behaves just like a per-
fect gas in a box. This is most certainly stable thermodynamically, and
therefore the same must be true of the isothermal model in the weakly
self-gravitating limit ze → 0∗ . For later reference we note that the energy

∗
Strictly, the system will evolve very slowly by the formation and hardening of a
binary (Chapter 5). The time scale is very long, however, of order N relaxation
times at least (cf. Problem 21.3).
178 17 Gravothermal Instability

Box 1. Entropy extrema.

We shall assume spherical symmetry, and so we may use the mass

within Rradius r as an alternative radial coordinate. Since entropy is S =
(k/m) ln(T 3/2 /ρ)dM its first variation is
Z  
k 3 δT δρ
δS = − dM,
m 2 T ρ
where δT and δρ are Lagrangian variations, i.e. following the mass. Sim-
ilarly, the energy is
Z Z
3k M
E= T dM − G dM,
2m r
and its first variation is
Z Z
3k M
0 = δE = δT dM + G δrdM. (1)
2m r2
dp
Now the hydrostatic condition tells us that GM/r2 = −4πr2 , and so
dM
we can integrate the second term in Eq.(1) by parts to remove the dp/dM
providedZthat we know how to differentiate the rest of the integrand. In
fact r = dM/(4πρr2), and so
Z
δr = −(4π)−1 (δρ/ρ − 2δr/r)dM/(ρr2).

Differentiation of this expression followed by a little rearrangement quick-

ly leads to the relation
d δρ
(r2 δr) = − ,
dM 4πρ2
which turns out to be just what we need. Using also the fact that δr = 0
at the outer boundary, we find that Eq.(1) may be rewritten as
Z Z
3k kT δρ
0 = δE = δT dM − dM.
2m m ρ
At this point we suddenly see that, if T is constant in the unperturbed
system, then the expression on the right of this equation is just T δS. In
other words the first variation of S vanishes if the system is in thermal
equilibrium; that is to say, the isothermal self-gravitating systems we
have been studying extremise the entropy.
 17 Gravothermal Instability 179

of the model in this limit is

3M
E= kT, (1)
2m
where m is the individual particle mass, and the other symbols have their
usual thermodynamic meaning.
Starting from this stable limiting case, we now travel along the one-
parameter family of equilibria looking for instability. In fact ze will do
very nicely as our parameter. As ze increases we see that gravity plays an
increasingly important part in determining the structure of the model.
Its density becomes increasingly inhomogeneous (cf. Fig.8.1), and we
can expect the paradoxical thermodynamic behaviour of self-gravitating
systems (Chapter 5) to make an appearance.
This is by now an old problem, and has been investigated quantitatively
using several different techniques (Antonov 1962, Lynden-Bell & Wood
1968, Horwitz & Katz 1978, Chavanis 2002, Chavanis et al. 2001). A
common, direct approach is to compute the second variation of S or
some other appropriate variable, but this can be somewhat laborious,
and here instead we employ what would nowadays be called bifurcation
theory. Actually, the method we adopt, which loosely follows that of
Katz (1978), has its origins in Poincaré’s theory of linear series (cf. Jeans
1929). For a simple analogue, consider a one-dimensional conservative
oscillator whose equation of motion is given by
∂
ẍ = − U (x, λ), (2)
∂x
where the potential U = x3 /3 − λx and λ is a parameter. The analogy
is that equilibria of our thermodynamic problem are given by a one-
parameter family of extrema of the entropy, whereas in Eq.(2) the equi-
libria are given by a one-parameter family of extrema of U . The main dif-
ference is that the dynamical problem has one degree of freedom, whereas
there are infinitely many in the √ problem of isothermal spheres. The equi-
libria for Eq.(2) occur at x = ± λ if λ ≥ 0, while there are none if λ < 0.
This is depicted in the “bifurcation diagram”, Fig.1, which shows that,
as λ is decreased, the two equilibria approach, fuse and disappear. Such
a feature is referred to as a “saddle-node bifurcation”.
Now let us look at stability. In Fig.1 the upper equilibrium (with x > 0)
is stable, and the other is unstable. The fact to concentrate on is that
stability changes precisely at the point where the equilibria disappear.
Now it could easily be claimed that this is a special feature of this toy
problem, but in fact it is “generic”, i.e. the systems for which it does not
happen are the special ones.
180 17 Gravothermal Instability

0.6
0.4
0.2
0
x
-0.2
-0.4

-0.6 -0.1 0 0.1 0.2 0.3 0.4

λ
Fig. 1 Bifurcation diagram for Eq.(2) with U = x3 /3 − λx.

In order to see this, consider Eq.(2) with a general potential about

which we assume only that there is a sequence of equilibria x(λ). Thus

(∂/∂x)U (x(λ), λ) = 0. (3)

Differentiating with respect to λ we see that

∂2 dx ∂2
U (x(λ), λ) + U (x(λ), λ) = 0. (4)
∂x2 dλ ∂λ∂x
Now suppose that there is a saddle-node bifurcation. At that point dx/dλ
is infinite, and it changes sign there as we pass along the series of equi-
libria. Provided, then, that
∂2
U (x(λ), λ) 6= 0 (5)
∂λ∂x
there, it follows that (∂ 2 /∂x2 )U changes sign there, and hence that the
equilibrium changes stability. Now, without an example at hand, one
might suspect that conditions at a saddle-node bifurcation are so special
that Eq.(5) is never satisfied. But in fact our toy problem illustrates that
this is not the case, and it is clear that it is systems in which Eq.(5)
is not satisfied that are special. In general, then, we should expect a
saddle-node bifurcation to be associated with a change of stability.
Now there are two more technicalities to deal with before all this ma-
chinery can be applied to the gravothermal problem, which has infinitely
many degrees of freedom, while the toy problem has only one. In the first
instance one would like a criterion which does not require plotting of x
against λ, because there are infinitely many x’s to choose from. One way
of dealing with this is to look at the variation of U along the series of
equilibria. In other words, define W (λ) = U (x(λ), λ). Then it is easily
 17 Gravothermal Instability 181

shown, using Eq.(3), that

d2 W ∂ 2 U dx ∂ 2 U
= + .
dλ2 ∂x∂λ dλ ∂λ2
If this is combined with Eq.(4) it follows that
!2
∂2U
 
d dW ∂x∂λ ∂2U
=− + ,
dλ dλ ∂2U ∂λ2
∂x2
and so the change of equilibrium can be detected by plotting dW/dλ
against λ and looking for a vertical tangent. If there are many degrees
of freedom with generalised coordinates {xi }, then the same conclusion
follows, the calculation being most easily carried out in coordinates which
diagonalise the second variation of U (x, λ) at fixed λ.
Now how do we apply all this to the gravothermal problem? In fact
almost everything is in place. Let us consider the series of equilibria
corresponding to a fixed total mass (M ) within an enclosure of fixed
radius (and therefore fixed volume, V ). For U we take the entropy S,
and the series of equilibria can be parameterised as we please, but it is
convenient to use the total energy E for the parameter λ. The reason
for this is that the quantity to be plotted, i.e. (∂S/∂E)M,V , then has a
simple interpretation: it is the inverse of the temperature, T . (Recall our
proof in Box 1 that δS = 0 in thermal equilibrium, which followed from
the result δE = T δS if T is uniform.) Therefore we plot 1/T against E
and look for vertical tangents in the resulting curve.
The result is shown in Fig.2. Several parts of this result can be un-
derstood without numerical computation. In the almost uniform, weakly
gravitating limit, it follows from Eq.(1) that T −1 ' 3M k/(2mE), which
yields the part of the curve going off to the right. Similarly the singular
isothermal solution and its asymptotic form (Problem 8.2) give rise to
the spiral.
Now we proceed from the right-hand end of the series of equilibria,
where the models are certainly stable, looking for a vertical tangent. It
is already evident that such points exits, and we need numerical methods
only to locate the first one. In fact it occurs when the scaled radius of
the enclosure is ze = 34.2, approximately. Stellar dynamicists remember
more easily the corresponding value of exp(ue ) = 1/709, cf. Eq.(8.4),
which is the ratio of the space density at the edge to the central value.
This is the last stable isothermal model inside a spherical, adiabatic con-
tainer. Models with a larger density contrast are unstable thermally.
182 17 Gravothermal Instability

2.4
2.2
1/T 2
1.8
1.6
1.4
1.2
1
-0.2 0 0.2 0.4 0.6 0.8
E

Fig. 2. The linear series of isothermal models of fixed mass and volume, in
suitable scaled units. Compare with Chapter 4, Fig.1.

After all this heavy analysis and theory, it is a relief to realise that the
physical origin of the instability is very easy to understand, provided that
one reads the appropriate paper (Lynden-Bell 1968). The arguments have
already essentially been deployed in Chapter 5. Imagine what happens if
a little thermal energy is transferred from the inner part of an enclosed
isothermal system to the outer part. The inner part should cool, but
the gravitational field causes the stars there to drop into lower orbits,
where they move faster than before. Just like a satellite spiralling slowly
to Earth under atmospheric friction, the net effect is that the stars are
moving faster, and the system is hotter, hotter even than at a time before
the heat was transferred. The thermal energy transferred to the outer
parts has a more familiar effect: these parts are held in not by gravity
but by the wall of the enclosure. Just as in a non-self-gravitating gas
in an enclosure, addition of thermal energy pushes up the temperature.
How much the temperature goes up depends on the heat capacity of
the outer part of the system. The larger and more extensive the outer
halo, the smaller the rise in temperature. For a sufficiently large halo
(sufficiently large re ) the rise in temperature of the outer part is smaller
than that of the inner part. Then a temperature gradient has been set
up, more heat flows from the inside to the outside (provided that the
gas is conducting), and the temperature gradient is enhanced. There is
a runaway, the famous “gravothermal catastrophe”.
This simple physical picture is suspicious, because it now seems as
though the boundary condition, an adiabatic rigid enclosure, plays a cru-
cial role. To show qualitatively that the boundary condition is not playing
an essential role, let us consider an isothermal model with a very large
value of re , i.e. very close to the centre of the spiral pattern in Fig.2.
There is a huge range of densities in this model. Therefore there is a
 17 Gravothermal Instability 183

huge range in the time scale for the convective transport of heat. On the
time scale for heat transport in the inner parts, the outer parts act effec-
tively as an insulating envelope. Also, the amount of mass in the inner
parts is so small that adjustments in the structure there have little effect
on the outer parts, which are also effectively a rigid barrier. In other
words, the outer parts of the system have the properties of the boundary
which we artificially inserted in order to make our model precise. It fol-
lows that the central parts of an isothermal system will be subject to the
gravothermal instability, and that the boundary conditions far out in the
system are irrelevant.
Precisely how the growth of the instability takes place is beyond the
scope of the above analysis. In particular, it depends on the details of
the mechanism by which heat is transported (Makino & Hut 1991). But
various calculations confirm that the “first” isothermal model in which
thermal inhomogeneities can grow is one in which the density contrast is
approximately the famous 709. This has been shown not only for mod-
els which treat the system as a self-gravitating, conducting, perfect gas
(Nakada 1978), but also for Fokker-Planck calculations (Inagaki 1980).
As we have said, however, the precise boundary conditions are not strictly
relevant to the behaviour of self-gravitating systems, and the critical value
of the density contrast is not the main issue.
There is another way of looking at the role of boundary conditions. So
far we have imagined that energy can be exchanged only between one
part of the system and another. Suppose now that the entire system may
exchange energy with its surroundings. We imagine that the enclosure is
still rigid and allows no mass in or out, but that the system can exchange
energy with its surroundings, which are held at constant temperature.
Then the heat capacity of the system is simply ∂E/∂T . It is therefore
subject to the kind of instability outlined in Chapter 5 as soon as the
tangent to the curve in Fig.2 becomes horizontal. Not surprisingly, this
occurs for a smaller density contrast than in the previous case, because
now the enclosure is taking on the thermodynamic role previously played
by the outer parts of the system. The distinction between the two situa-
tions can also be described in terms of the appropriate ensemble, in the
sense of statistical physics: the microcanonical ensemble for systems of
fixed energy (the first case, in which the enclosure is adiabatic), and the
canonical ensemble for systems of fixed temperature (the second case.)
There is one final twist to this tale. We gave a physical explanation for
the gravothermal catastrophe by contemplating a transfer of heat from
the inside to the outside. But instabilities can grow in either direction;
what happens if heat is transferred from the outside to the inside? It is
not hard to see that both the inside and the outside cool down, but again
the cooling of the outer parts is slight if they are sufficiently massive and
184 17 Gravothermal Instability

extensive. Then heat starts to flow inwards, driving the gravothermal

instability in reverse. We shall see in Chapter 28 that this is no mere
thought experiment.

Problems
1) A simple approximate model of an isothermal system in an enclosure
of radius re is defined by the density


 ρ0 , if r < r0 ;
 2
ρ(r) = r0
 ρ0
 , if r0 < r < re .
r
Compute M (r), i.e. the mass within radius r, and deduce that the
potential energy is
x GMe2
W =− (45x − 30 ln x − 42) ,
5(3x − 2)2 re
where Me is the total mass within the radius re and x = re /r0 > 1.
Using the appropriate form of virial theorem (Problem 9.1, where T
denotes the kinetic energy), compute the temperature and the total
energy. Hence plot the diagram analogous to Fig.2. Show that the
sequence includes models with positive and negative heat capacity.
2) In an enclosed isothermal model with enclosure radius re , the kinetic
3M
energy is kT , where T is the temperature. Use the appropriate
2m
form of the virial theorem, as in Problem 1, to show that the energy
is given by
 0
Ere 1 u+ 3u ,
= e
GM 2 u02 2z
in the notation of the isothermal equation (Eq.(8.6)). By solving the
latter equation numerically, and locating the minimum of Ere /(GM 2 )
for varying z, show that the first unstable model corresponds to a
density contrast e−u = 709 approximately.
3) Does a two-dimensional self-gravitating system, in which the interac-
tion potential between two masses m is 2Gm2 ln r, exhibit gravother-
mal instability? Think about this first from the point of view of the
virial theorem (Problem 9.2). Then use the analytical solution (Prob-
lem 8.3) to carry out an analysis as in Fig.2. See Katz & Lynden-Bell
(1978).
 18
Core Collapse Rate
for Star Clusters

The cruel fate of a system forever striving to be what it can never be –

in thermal equilibrium. – J. Goodman

The last two chapters have assembled most of the qualitative arguments
by which the evolution of the core of a stellar system can be understood.
In summary, the tendency towards equipartition drives the more mas-
sive stars to smaller radii. Unless their total mass is sufficiently small,
equipartition cannot be reached by the time the heavier stars become es-
sentially self-gravitating. When that happens they are eventually subject
to the gravothermal instability. It is the purpose of the present chapter
to flesh out this outline, but we shall do so in two passes, as it were.
First we shall examine the timescales on which these processes act, and a
number of factors which modify the simple picture; and we shall explain
the qualitative nature of the resulting evolution. Then we turn to a more
detailed description of one case which has been studied in great detail:
self-similar collapse in systems with stars of equal mass.

The big picture

The time scale for equipartition, te , was discussed in Chapter 16 (cf.
Eq.(16.8)). It is useful to compare it with the standard relaxation time
tr (Eq.(14.8)). For this purpose we evaluate the mean kinetic energy
per unit mass for each species by hEi i = v2 /2, independent of mass, i.e.
we assume equipartition of velocities. In Eq.(14.8) we write the factor
m2 n as m̄ρ, where m̄ denotes the mean stellar mass, and ρ is the mass-
density. We also identify the Coulomb logarithm in both timescales. For
a two-component system it follows that te /tr = 1.25m̄2 /(m1 m2 ).
In applications, we are often interested in models with many particles
of low mass (m1 , say) and a few of high mass m2  m1 . In this case
m̄ ' m1 , and so te /tr ' m1 /m2 ,  1. In other words the fastest pro-

185
186 18 Core Collapse Rate for Star Clusters

cesses are associated with equipartition. Indeed, if one uses some detailed
model (Chapter 9) to follow the evolution of a multi-component model
(i.e. one in which the particles do not all have the same mass), the first
thing one observes is that the heaviest particles drift towards the centre,
increasing the central density and the central proportion of massive par-
ticles (Fig.16.1). The greater the individual masses of the most massive
particles, the faster the evolution.
There are factors which modify this picture in more realistic models
of stellar systems. In the first place, the most massive stars also evolve
(internally) most rapidly (Chapter 11). In the very rapid, late stages of
their evolution, they lose most or all of their mass. Therefore the range
of individual stellar masses shrinks, and this in turn affects the relevant
timescale te . The second point to notice is that the most massive species
may well consist of binaries with massive companions. Though the effects
of close interactions between binaries and single stars are complicated
(Parts VI and VII), in distant encounters a binary behaves essentially
like a single star with a mass equal to the total mass of its components.
Eventually the tendency towards equipartition slows down drastically.
As already mentioned, this does not mean that equipartition has been
achieved; rather, the central parts of the system are now dominated by
the most massive stars, and, as a first approximation, we may consider
that the role of the less massive stars is unimportant. As a step towards
understanding this new situation, we proceed to discuss the evolution of
the core in a single-component system, in which all stars have the same
mass. This problem is of obvious interest in its own right, and it has been
rather well studied.
First let us consider the situation discussed in Chapter 17, i.e. an
isothermal model enclosed in a spherical container. As we saw, this equi-
librium is unstable only if the density contrast (between the centre and
the enclosure radius) exceeds a factor of about 709. At the enclosure
radius the relaxation time is trc /709, where trc denotes the central value.
Therefore we may expect that the time scale on which the instability acts
is at least of this order. In fact it may be much longer, as the system is
nearly isothermal, which reduces the thermal flux.
Now we should consider a more normal stellar system, still consisting
of stars of equal mass, but bound only by their self-gravity, and not by
any external wall. Such a system cannot be isothermal, and so the flux
of heat which drives evolution of the core may be stronger than in the
near-isothermal situation considered previously. Qualitatively, however,
the result is the same: the centre contracts and becomes hotter. At the
higher densities which result, the relaxation time decreases, and in time it
becomes much shorter than at the larger radii where the heat flux drove
the initial contraction. By the time this happens, the hot, central region
 The big picture 187

of the cluster is effectively isolated (thermally) from the regions at larger

radius. Thus there is a transition from a collapse which is driven by the
heat flux at large radii to a collapse, analogous to that in an enclosed
isothermal sphere, driven by the heat flux in a smaller, nearly isothermal
region. As this region collapses further, the time scale becomes shorter
still, and there is a “runaway” evolution towards high central densities
and temperatures (Fig.1).
The phenomena we have been describing are referred to collectively
as “core collapse”. Its study dominated much theoretical work on the
evolution of dense stellar systems in the 1970s. The time it takes to
reach completion is determined mainly by the early phase, in which the
central parts are not yet thermally isolated from the outer parts of the
system, and when the central relaxation time scale is longest. In fact,
detailed numerical experiments of various kinds, using Plummer’s model
for initial conditions, and equal masses, shows that the time taken is
about 15trh , where trh is the initial half-mass relaxation time (Chapter
14). For a system with a distribution of individual stellar masses, the
time to completion of core collapse is determined by the time scale of
mass segregation, and we have already seen that this is smaller than
the relaxation time scale of an equivalent equal-mass system. Detailed
experiments confirm this. Consider, for example, a system in which the
ratio of the largest to the smallest mass is 15:0.4 (values which do not
look so strange to an astrophysicist who interprets the unit of mass as
the mass of the sun), with a distribution of masses f (m) ∝ m−2.5 in this
range. Then the time to the end of core collapse is as short as about
0.9trh (Chernoff & Weinberg 1990). The time scale of core collapse is

Fig. 1. Core collapse from a King model with W0 = 3, N = 1285. The initial
model is shown on the left, and the view on the right shows the same model to
the same scale after core collapse. The model has a tidal cutoff.
188 18 Core Collapse Rate for Star Clusters

also altered if the system is rotating (Lagoute & Longaretti 1996, Einsel
& Spurzem 1999, Kim et al. 2002).

Self-similar core collapse with equal masses.

So far our discussion of core collapse has been short of quantitative detail
on anything except time scales. We have arrived at a broad-brush picture
in which the core of the cluster becomes smaller, hotter and denser. In
addition, when a mixture of stellar masses is present, the core is increas-
ingly dominated by the heaviest species. But it is not yet at all clear
what this core will look like. In the remainder of this chapter we shall try
to flesh out the portrait. We concentrate on systems in which all stars
have the same mass, and analyse the late evolution, i.e. when the core
is no longer in thermal contact with the outer parts of the system. The
approach we adopt was pioneered by Lynden-Bell & Eggleton (1980).
It is clear that the central parts of the collapsing core must be nearly
isothermal. If not, relaxation would quickly thermalise the largest region
around the centre which is gravothermally stable. Therefore the core can
be nearly characterised, like an isothermal model, by two parameters, e.g.
the central density ρc and the central one-dimensional velocity dispersion
σc (cf. Chapter 17).
We now make the assumption that the structure around the core, while
not being precisely isothermal, is nevertheless still determined by ρc and
σc . In other words, at each time during the collapse of the core the struc-
ture may be obtained from that at any other time by a simple rescaling.
Often such evolution is referred to as “homologous” or “self-similar”.
This apparently drastic assumption has been of great benefit in stellar
dynamics. Indeed we shall see in Chapter 33, where we summarise the
whole of the evolution of a star cluster, that much of it can be described
by patching together various kinds of self-similar evolution.
There are other areas of applied mathematics where such behaviour
is common. In particular, in problems of “finite-time blow-up”, i.e. so-
lutions of partial differential equations which tend to infinity at some
point in space as t approaches some finite value, it is often found that
the form of solution at late times, and in the vicinity of the singular-
ity, is self-similar. A familiar example is the “backward heat equation”
∂u/∂t = −∂ 2 u/∂x2 , which has the self-similar solution
!
1 x2
u= √ exp
−4πt 4t
for t < 0. It is perhaps not surprising that such behaviour should arise
also in core collapse in stellar dynamics. This is governed, as we saw
in Chapter 17, by a property of stellar systems analogous to a negative
 Self-similar core collapse with equal masses. 189

heat capacity. If we were to derive the standard heat equation for such
a material, we would obtain the backward heat equation.
For the star cluster problem, derivation of the self-similar solution re-
quires a modicum of numerical integration, and our treatment will confine
itself to some basic observations and a brief description of the results. In
the first place let us note that the time scale on which core parameters
such as ρc vary is trc , the central relaxation time. In self-similar evolution
it follows that trc ρ̇c /ρc is constant, and the same is true for other core
parameters. It follows that these parameters will vary with each other
as power laws (the powers being determined by the various constants)
and, if we ignore variations of the Coulomb logarithm, that trc is a linear
function of time.
Let us suppose, then, that
ρc ∝ rc−α , (1)
where α is some constant. Since we are supposing that the evolution is
self-similar, the density profile at any time (i.e. the function ρ(r, t) at
some t) may be obtained from the profile at any other time by scaling.
In other words,
 
ρc (t1 ) rc (t0 )
ρ(r, t1 ) = ρ r, t0 . (2)
ρc (t0 ) rc (t1 )
Now at large radii, where the relaxation timescale is huge compared with
that in the core (or even with the evolution time scale), the density profile
must be almost stationary. By substituting (1) into Eq.(2) it follows that,
at large radii,
ρ(r) ∝ r−α . (3)
This is the density profile that must be left behind by the outlying ma-
terial of the contracting core.
Now we are in a position to say something about α on physical grounds.
From Eq.(3) it follows that the mass at large radii varies as r 3−α , and
so the gravitational acceleration F varies as r 1−α . By the condition of
hydrostatic balance, i.e.
∂
(ρσ2 ) = ρF,
∂r
it is easy to deduce that the velocity dispersion must vary as r 2−α . Since
the collapse of the core is driven by a negative temperature gradient it
follows that α > 2.
Now the velocity dispersion in the core is related to ρc and rc by the
definition of the core radius (Eq.(8.9)), and so it follows that
σc2 ∝ rc2−α . (4)
190 18 Core Collapse Rate for Star Clusters

Also, the energy contained within the core is of order ρc rc3 σc2 , and by
Eqs.(1) and (4) we see that this varies as rc5−2α . It is physically reasonable
to suppose that this must remain finite as the core radius contracts, and
so α < 5/2.
It has been noted (Lancellotti & Kiessling 2001) that such a solution
would not be a self-similar solution of the full Fokker-Planck equation.
But this requirement is too strong, for it requires that the orbital motions,
as well as the collapse of the density profile, take place on the same time
scale. Here we are studying collapse, on a relaxation time scale, of a
system where orbital motions (on the crossing time scale) are much faster.
Then it is appropriate to solve only the orbit-averaged Fokker-Planck
equation, for which a wider class of self-similar solutions are possible.
Since collapse is driven by relaxation, we may estimate that the time
to collapse, τ , is proportional to the relaxation time, i.e. τ ∝ σc3 /ρc
(Eq.(14.8), except that we ignore the Coulomb logarithm). By Eqs.(1)
2
and (4) we deduce that rc ∝ τ 6 − α . For our range of estimates of α the
power here lies between 0.5 and about 0.57.
It does not appear to be easy to sharpen the possible range of values
of α further without numerical methods. Indeed more detailed models
show that α ' 2.2 or 2.3, and Fig.2 shows the numerically computed
density profile for a self-similar gaseous model. For the corresponding
Fokker-Planck model the e-folding time scale for the central density is
approximately 280trc , which is not so far from the estimate of 709trc
obtained previously on the basis of physical arguments. It should be
pointed out, though, that this homologous phase of core collapse is a
good approximation only very late on. It is preceded by a much longer
phase where the collapse time scale is a smaller multiple of the central
relaxation time.
Finally it should be stressed that this process of core collapse is not
quite the catastrophic process which the words tend to convey. This is
no headlong rush like the collapse to a black hole. It takes place on a
timescale which is very long even relative to the central relaxation time.
It also involves a rapidly decreasing fraction of the total mass of a cluster.
But it does imply that the central density reaches arbitrarily high values
within a finite time. For many galactic globular clusters it is a time
considerably smaller than their age. Even at the present day, a couple of
clusters approach this endpoint of core collapse every billion years in our
galaxy (Hut & Djorgovski 1992). It is accompanied by a marked increase
in the rate at which stars escape (cf. Fig.5.1), and in the rate at which
 Self-similar core collapse with equal masses. 191

Fig. 2. The self-similar model of Lynden-Bell & Eggleton (1980). The solid and
dashed straight lines correspond to the asymptotic isothermal relation ρ ∝ r−2
and the asymptotic profile of the collapsing model, ρ ∝ r−α . The variables
plotted are the scaled radius and density, r/rc and ρ/ρc , respectively.

interesting encounters take place (Chapter 31). What happens next is

the point at which Chapter 27 takes up the story.

Problems
1) Show from the isotropic Fokker-Planck equation (Eq.14.11) that the
rate of change of f at the bottom of the potential well (energy φ(0))
is given by
Z ∞
f˙ = 16π 2 G2 m2 ln Λ(f 0 f (E)dE + f 2 ),
φ(0)

where f and f0 are evaluated at φ(0), except in the integral. Hence

show that the time scale of the initial collapse of a Plummer model
(Table 8.1) is of order 0.88trh .
2) Globular star clusters in orbit around our Galaxy contain only older
stars, with an age of order ten billion years. This implies that nearly
all single objects in the cluster are less massive than the sun, with
the exception of neutron stars and black holes. During intermediate
stages of core collapse, the lighter stars tend to leave the core, through
mass segregation, with the result that the typical mass of stars in the
core will lie in a relatively narrow range, somewhere between 0.5 and
0.8M , with neutron stars with a mass around 1.4M only slightly
heavier. In contrast, black holes can easily have masses of 10M or
more. Let us investigate what happens with a population of black
192 18 Core Collapse Rate for Star Clusters

holes, remnants of the heaviest stars that burned up early on in the

history of a globular cluster.
a) argue that the black holes will form a ‘core within the core’ of
the globular cluster, on a time scale of a half-mass relaxation time
or shorter, and show that the radius of that inner core is at least
three times smaller than that of core of the cluster. (Assume that the
total number of black holes is small, so that the other stars are hardly
affected by the presence of the black holes, initially.)
b) argue that the central density of the black holes will increase
relative to the central density of the other stars, and make an estimate
of the number of stars left in the core by the time the densities between
holes and stars become comparable.
c) from this point on, the Spitzer mass segregation instability will
set in, if the number of black holes is sufficiently large. Argue that,
subsequently, the black holes will undergo core collapse on a shorter
timescale than the normal stars in the core. In particular, show that
the ratio of timescales is of order µ5/2 , where µ is the mass ratio
between a black hole and a typical single star.
(This problem continues in Problem 20.3.)
3) Using the estimate in Chapter 13 for the e-folding time for the growth
of errors, show that the number of decimal places required for an
accurate numerical integration to core collapse is roughly proportional
to N .
 PART VI. GRAVITATIONAL
SCATTERING: N = 3

Core collapse leads to high stellar densities, where interactions may in-
volve more than just two stars at a time. The chapters in this section
are therefore devoted to three-body interactions, especially interactions
between a binary star and a single star. One of our aims in these chap-
ters is to show that important aspects of the three-body problem can be
understood from various points of view, even though the problem itself
lacks a general mathematical solution.
Chapter 19 takes a phenomenological approach, applying notions of
equipartition and energy conservation. This already classifies encounters
according to whether the binary is hard or soft. In some interactions with
hard binaries the result (temporarily) is like a miniature star cluster of
three stars, and our previous knowledge of the behaviour of star clusters
can suggest how this evolves.
Chapter 20 takes an informal mathematical view of the same phenom-
ena. We see that the breakup of triple star clusters exhibits a sensitive
dependence on initial conditions, partly justifying a statistical treatment.
One of the standard examples in which this is most readily understood is
Sitnikov’s problem, which we use to introduce the Smale horseshoe. Fi-
nally we prove informally a theorem which shows that permanent capture
into a triple configuration is (practically) impossible, and end with some
recent surprising discoveries about permanently bound triple systems.
Chapter 21 takes a course half way between the previous ones, exploit-
ing a mixture of approximate analytical tools and physical arguments to
develop a theoretical results on the outcome of three-body interactions.
This is expressed in the language of cross sections, in analogy with simi-
lar problems in atomic scattering theory. Here the analogue of Coulomb
repulsion is gravitational focusing. We estimate the rate at which bi-
naries form within a million-body system. We treat in some detail the
case of adiabatic encounters between a hard binary and a single star,

193
194 PART VI. GRAVITATIONAL SCATTERING: N = 3

when the single star remains “outside” the binary and moves on a time
scale long compared with the period of the binary. In the opposite ex-
treme we analyse cases in which all three stars come within a comparable
small distance from each other, which produces results akin to those of
“threshold” scattering in the atomic context.
Far as these analytical techniques go, it is hard to make them com-
pletely quantitative and sufficiently accurate, which may be accomplished
by computer simulation (Chapter 22). We describe some of the issues in-
volved in integrating few-body systems efficiently (i.e. accurately and
quickly), and the manner in which the calculations may be organised and
automated.
Chapter 23 takes the first steps in applying these results to the evo-
lution of the million-body problem. The cross sections are turned into
quantities representing the rate of occurrence of three-body interactions,
and the mean rate of change of energy of a binary in three-body interac-
tions. One effect of these interactions is that binaries are driven from the
dense central core of an N -body system into the surrounding halo. The
most important results concern the rate at which energy is imparted to
the single stars, as it is this process which halts core collapse and allows
the system to evolve after that.
 19
Thought Experiments

As has been seen in Chapter 18, two-body relaxation predicts its own
downfall. It leads to the collapse of the core and, at the level of simplified
models, infinite central density. Clearly, some new dynamical processes,
beyond two-body encounters, must come into play. The very high density
is the clue, for it suggests that a third body may, with increasing prob-
ability, intervene in the two-body encounters which mediate relaxation.
In Chapter 27 it will be seen that three-body encounters do indeed act
on a sufficiently short time scale, late in core collapse, to have a decisive
influence on events. As we note there, this is not the only mechanism
that can work, but we concentrate on it for the time being.
The mechanism is a two-stage one, both stages involving three-body
encounters∗. In the first stage, which we consider in Chapter 21, a three-
body interaction leads to the formation of a binary star and a single third
body (which acts as a kind of catalyst). In the second stage, this binary
interacts with other single stars (again in three-body reactions). In this
chapter we shall study three-body encounters in isolation, in order to
uncover those properties which allow them to play their crucial role in
rescuing the cluster from collapse. Clusters get into this difficulty because
of their negative heat capacity, and in fact it is the negative heat capacity
of binaries which comes to their rescue.

Energetics of Three-Body Scattering

The energy of a binary star is

Eb = (m1 v12 + m2 v22 )/2 − Gm1 m2 /r,

∗
We have in mind systems which initially have no binaries. If “primordial binaries”
(Part VII) are present, the first stage is not needed

195
196 19 Thought Experiments

where mi and vi are, respectively, the mass and speed of the ith compo-
nent (i = 1, 2), and r is their separation. It is more convenient to write
this as
Eb = ε + (m1 + m2 )V 2 /2, (1)
where V is the speed of the centre of mass of the binary and
1 m1 m2 2
ε= v − Gm1 m2 /r; (2)
2 m1 + m2
here v is the relative speed of the components. The “binding energy” of
the binary is the positive quantity −ε; ε, which is the quantity we prefer
to work with, is referred to as the “internal energy” of the binary. The
second term on the right of Eq.(1) can be thought of as the kinetic energy
of the barycentre. It is often called the “external energy” of the binary,
though if the binary is immersed in an N body system the potential
due to the other members would also be included. Note that the kinetic
energy of internal motion, when expressed in terms of the relative speed
v, involves the reduced mass m1 m2 /(m1 + m2 ) of the two components.
Now we add a third body, of mass m3 , and at first we suppose that it
is well separated from the binary. Then the energy of this triple system
is obtained from Eb by adding the kinetic energy of m3 . The result takes
its simplest form in the frame of the barycentre of the triple system, and
is then
1 (m1 + m2 )m3 2
E3 = ε + V , (3)
2 m1 + m2 + m3
where V is the speed of m3 relative to the barycentre of the binary. The
second term on the right is the combined kinetic energy of the third body
and the barycentre of the binary.
Now we suppose that the third body approaches the binary, interacts
with it briefly, and then departs once again to a great distance. (This
is what physicists refer to as a “scattering problem”.) Suppose that
the energy of the binary and the relative speed at a time long after the
encounter are, respectively, ε0 and V 0 . Then it is convenient to gauge the
effect of the interaction by the relative change in the internal energy of
the binary, i.e. ∆ = (ε0 − ε)/ε. If ∆ > 0 the internal energy of the binary
has become more negative, i.e. it has become tighter, or more bound. By
conservation of energy E3 it follows that V 0 > V in this case. It is this
possibility which is all-important in the termination of core collapse, and
so in this chapter we place considerable emphasis on deciding whether ∆
is greater or less than 0 on average.
There are three related qualitative arguments which suggest what the
net effect of such encounters may be. One is based on the negative specific
heat of self-gravitating systems (Chapter 5). We already saw in the last
 Energetics of Three-Body Scattering 197

chapter that this property leads to increasing density in the most usual
situation. More precisely, this is what happens to a core in thermal
contact with a cooler halo, and we also saw that, if the surroundings are
hot, the evolution could run in reverse. One might guess that the same is
true of binaries, i.e. that binaries get tighter (more bound) as they react
with other, slow moving stars in three-body interactions, and looser (i.e.
less bound) if the stars it encounters move fast.
The second argument, based on the tendency to equipartition of ener-
gies, strengthens and clarifies these conclusions, and runs as follows. We
begin by noticing that there are degrees of freedom associated with both
terms on the right of Eq.(3). If the second term is much bigger than the
first (i.e. we consider fast encounters, or binaries of low binding energy),
the kinetic energy of relative motion will tend to diminish and the inter-
nal energy of the binary will tend to increase (i.e. it will tend to become
unbound, with ∆ < 0). Similarly, a binary with high binding energy will
tend to become even more bound, and V 0 > V on average.
For the third argument we assume that the interaction can be broken
up into an encounter between the third body and each of the binary com-
ponents separately. If the binary is hard, its components move fast, and
impart some of their energy to the incomer, rather like a stone deflected
by a rotary lawnmower. In this case energy is extracted from the binary,
which hardens. Similarly, if the incomer is moving fast, it transmits some
of this energy to the binary components, and the energy of the binary
increases.
The conclusions of all three arguments suggest that there is a watershed
binding energy, near

|ε| ∼ (m1 + m2 )m3 V 2 /(m1 + m2 + m3 ), (4)

above which binaries tend to become more bound, and below which
they tend to become less bound. Though we shall mention some re-
finements shortly, we note here that this kind of result is amply sup-
ported by numerical experiments (Chapter 23; see, for example, Fig.23.1).
The distinction is often expressed by saying that binaries with |ε| <
(m1 + m2 )m3 V 2 /(m1 + m2 + m3 ) are “soft”, which indicates their rela-
tive fragility to encounters, while those with |ε| > (m1 +m2 )m3 V 2 /(m1 +
m2 + m3 ) are called “hard”. In practice the distinction is not quite that
simple. The important point is to distinguish those binaries which harden
on average from those which tend to soften. For equal masses, it turns
out that our expression for the threshold energy has the correct form, and
there is a numerical coefficient near unity. If the masses are equal and
there is a distribution of velocities V , then Eq.(4) is correct if we add a
different numerical coefficient and replace V 2 by its mean square value.
198 19 Thought Experiments

When the masses are different, rather less is known about the correct
threshold energy. When m3 is tiny, i.e. the incomer is a test particle, it
is clear that the critical value is independent of m3 , unlike an estimate
based on Eq.(4). For this case it has been shown that, if the two members
of the binary have equal mass, then the critical value of V is of order the
orbital speed of the binary components about each other (Hills 1990; see
also Fullerton & Hills 1982). Sigurdsson & Phinney (1993) have given
useful data for masses relevant to globular clusters.

Soft Binaries
The equipartition argument is attractively simple, provided one comes
to it with a sufficiently developed physical intuition. But it is not hard
to confirm its prediction with somewhat more concrete considerations, as
we now show. We first consider the case of a fast encounter, i.e. a very
soft binary, in which the first term on the right of Eq.(3) is much smaller
than the second. Then the encounter is nearly impulsive, the relative
position of the two components of the binary hardly alters, and the main
effect of the encounter is on their relative velocity v. If this changes by
an amount δv, then Eq.(2) shows that ε changes by
m1 m2 1
δε ' (v.δv + δv2 ). (5)
m1 + m2 2
In an impulsive encounter δv depends only on the position of the com-
ponents of the binary and not on v. Furthermore, in any reasonable sta-
tistical distribution of binaries, velocities ±v will be present with equal
probabilities, and so the mean value of v.δv will be zero. It follows that
the mean value of δε will be positive, i.e. that soft binaries soften on
average, as expected.
Qualitatively different encounters are possible. If m3 comes sufficiently
close to one component of the binary, then the velocity of that component
may change by so much that its new velocity exceeds the velocity of escape
from its companion. Such an encounter is referred to as an ionisation,
by analogy with atomic scattering. In the usual case, when the binary
survives, we may refer to the encounter as a “fly-by” (which is a name
with definite space-age connotations). There is a third possibility for
these fast encounters, and it is most easily understood in the case of equal
masses. In a very close two-body encounter between equal masses the two
masses almost exactly exchange velocities, just as when a billiard ball hits
another stationary one. If, therefore, in a fast encounter between m3 and
a binary, the intruder scores almost a direct hit on one component, that
component will leave the binary with almost the speed of approach of
m3 , and the latter takes its place as one component. Such an encounter
is an example of an exchange.
 Hard Binaries 199

As an aside, it is worth remarking that the theory of exchange encoun-

ters has a curious history. The earliest numerical examples (Fig.1) were
given by Becker (1920) (working at Glasgow University, incidentally), but
there was a community of theorists who preferred to settle such questions
by pure thought. One of them (Chazy) conjectured that exchange was
impossible, and such was the suspicion of numerical methods among this
community, and such was Chazy’s influence, that his view held sway for
over half a century. According to Marchal (1990), who gives a lively
account of all this history, it was only in 1975 that the possibility of ex-
change was convincingly established. Meanwhile astrophysicists, largely
ignorant of these niceties, had already spent some years estimating the
first cross sections for exchange, both numerically and by approximate
analysis. As it happens these were published in the same year (Hills 1975,
Heggie 1975).

Hard Binaries
Now let us consider the opposite extreme, i.e. encounters in which the
second term on the right of Eq.(3) is small. We shall also suppose that
the encounter takes place in such a way that the minimum distance of
the intruder (m3 ) from the binary is not much greater than the size
of the binary (technically, the length of its semi-major axis). Typical
energies in the interaction are therefore of order ε. If a positive energy of
order |ε| is donated to the third body (or, more precisely, to the kinetic
energy of its motion relative to the binary), then the third body will
recede with much higher speed than its original speed of approach. (This

Fig. 1. One of the exchange orbits computed by Becker in 1920. The position
of one particle is used as the origin of coordinates, and the initial conditions
were chosen so that particles B and C are symmetric under time reversal and a
reflection.
200 19 Thought Experiments

follows from our assumption that |ε| much exceeds the second term on
the right of Eq.(3).) Also, by energy conservation, the binding energy of
the binary has increased by an amount of order |ε|, i.e. it has become
harder, as expected. What if, however, an energy of order |ε| is extracted
from the kinetic energy of the third body? Since |ε| much exceeds the
kinetic energy of the third body, the third body cannot escape, at least
not right away. What happens is that it must return to the vicinity of
the binary and there exchange another packet of energy. Again there are
two possibilities: either m3 can now escape or it cannot. It can be seen
that, in general, the only way in which the encounter can end is with the
eventual escape, after some large or small number of interactions, of one
of the participating stars (Fig.2). When it escapes, it will do so with a
kinetic energy of order ε, which much exceeds its initial kinetic energy,
and so again the net effect of the interaction will have been to harden the
binary.

Fig. 2. A three-body scattering event. It is a resonant exchange, and the se-

quence of interactions is summarised in the pictogram at upper left. From Hut &
Bahcall (1983); the authors had to look quite hard to find an example as simple
as this!
 Hard Binaries 201

From the above discussion it can be seen that the encounter may be over
quite promptly. If, on the other hand, there is a complicated interplay of
the three participants, then the encounter is described as a “resonance”.
The name again comes from atomic or nuclear scattering, where it refers
to the unusually large probability of interactions which can occur if the
participants can form a temporary bound state. In the gravitational case
what we see is a temporary, bound triple system. As we shall see in the
next chapter, such a configuration cannot last for ever, but while it lasts
it behaves very much like a star cluster with N = 3. In particular, the
motion exhibits sensitive dependence on initial conditions (Chapter 13),
which suggests that the behaviour of an ensemble of such systems may
exhibit some statistical features. Indeed the distribution of lifetimes of
these systems is nearly exponential (Agekyan et al. 1983), like the random
decay of an atomic nucleus. It turns out that this statistical approach is
a fruitful way of attempting a theoretical prediction of the distribution
of the energy of the eventual escaper (Heggie 1975, Nash & Monaghan
1978). More obviously, it suggests that, in the case of equal masses,
the probability that any of the three participants is the escaper is 1/3;
thus the probability that the resonance leads to exchange should be 2/3.
These resonances in which all three particles interact closely are therefore
sometimes called “democratic”, even in the case of unequal masses.
There is another interesting result which follows from thinking of a
resonance as a small star cluster. In the case of unequal masses our
knowledge of the escape rate in N -body systems (Chapter 16) suggests
that the body of lowest mass is most likely to escape. It follows that
a resonant encounter of a binary with a star whose mass considerably
exceeds that of one of the components is liable to lead to exchange of
that component. In other words, such encounters tend to increase the
mass of the components of a hard binary.
Finally we must turn briefly to the case in which the distance of closest
approach of the third body, denoted by rp , remains considerably greater
than a, the semi-major axis of the binary. Now the energy exchanged
between the two terms in Eq.(3) may be quite small. In fact, it may be
exponentially small, so that − ln |∆| varies roughly as a positive power of
rp /a. The reason for this is that, unless the masses are very disparate,
the timescale on which the third binary is within a distance of a few rp of
the binary much exceeds the period of the binary. Then the perturbation
by the intruder is adiabatic, and an exponentially small net perturbation
is to be expected (Chapter 21). Still, the encounter may result in a fly-by
or in formation of a resonance. Now, however, the third body has too
much angular momentum relative to the barycentre of the binary for a
close interaction, and the resonance is very unlikely to lead to exchange.
202 19 Thought Experiments

Such resonances are referred to as “hierarchical”. They usually resolve,

sooner or later, with the ejection of the original intruder.
When we contemplate Fig.2 it is not hard to see why the three-body
problem is not analytically soluble in general. Yet it has simple analytical
properties which can be very useful. We have already used energy conser-
vation effectively in the above discussion, and another valuable property
is the invariance of the equations of motion under time reversal. For
example, time-reversal of an ionisation encounter yields an encounter in
which three single stars interact to yield one binary and a single star
which carries off the excess energy. The same argument applied statisti-
cally also tells us how the rate at which binary stars are formed is related
to the rate at which they are destroyed. The general 3-body problem
may well be insoluble from any strict point of view, but there is a large
amount that may be understood from quite simple considerations. And
our study of the three-body problem in this book has just begun.

Problems
1) Answer the following riddle: How do you make a binary star with a
grain of sand? What this means is the following. Two stars approach
on a hyperbolic relative orbit, and your task is to throw in a grain of
sand so as to extract enough of their kinetic energy to bind them into
a binary. (This is classical mechanics of point masses.)
2) For an N -body system of energy E in virial equilibrium, show that
the threshold energy between hard and soft binaries is of order E/N .
3) Suppose single stars of mass m are distributed uniformly in space
with space density n and have a Maxwellian distribution of velocities.
Assuming that the two stars are uncorrelated, show that the number
density of very soft binaries of energy ε is
π3/2 G3 m6 n2 −5/2
nb ' |ε| .
4(kT )3/2
4) In quantum statistical mechanics, the equilibrium distribution of en-
ergies of hydrogen atoms is given by
f (E) = w(E)ρ(E) exp(−E/kT ),
where the first factor on the right is the statistical weight (multiplic-
ity) of one energy level, the second counts the density of levels per
unit energy, and the third is the usual Boltzmann factor. By applying
the correspondence principle to high-energy states, obtain the corre-
sponding result for the distribution of the binding energy of binaries
in statistical equilibrium, i.e. f (ε) ∝ |ε|−5/2 exp(−ε/kT ).
 20
Mathematical Three-
Body Scattering

While the last chapter roughed out a picture of what happens in an

interaction between a binary star and a third body, there are three very
different ways in which the picture can be sharpened. One is to develop
approximate analytical results on the outcome of an encounter, and that
is successful in various limiting cases, e.g. very distant encounters, very
hard binaries, and so on. This is the approach of Chapter 21. To cover
the middle ground between these extremes, there is no substitute for
computational studies (Chapter 22). In the present chapter, however,
we push the analytical methods in the opposite direction, and examine
minute corners of parameter space which may be of no conceivable value
in applications. The merit of this approach is that rigorous statements
become possible, at least in expert hands, and the resulting ideas help
to develop our intuition of what can happen in more realistic situations.
Our approach is quite informal, and places emphasis on the ideas behind
the proofs, without any technical details.

Fractals and Chaos

We first turn to resonances, those long-lived but temporarily bound
triple systems that often arise in scattering events. Our first aim here is
to discuss one situation which makes it particularly clear why the outcome
of a resonance depends sensitively on the initial data. We can then argue,
at least physically, that the system forgets details of the initial conditions
of its formation, and that the breakup is determined (in a statistical
sense) only by the quantities which are preserved in the evolution, i.e.
energy and angular momentum.
It is not hard to see how sensitive dependence can arise. Consider first
a resonant encounter in which the eventual escaper only just escapes.
Now let us vary one of the initial conditions, which we shall call x, and
suppose the condition that the escape occurs is that x > x∗ , where x∗ is

203
204 20 Mathematical Three-Body Scattering

some critical value. If x is just below x∗ , the would-be escaper moves to

an enormous distance from the binary before returning, and then what
happens depends on the phase of the binary at the moment when the
would-be escaper returns to its vicinity. But as x → x∗ , the phase of the
binary changes by 2π infinitely often, and in each corresponding interval
of x (i.e. corresponding to a given interval of phase of the binary, modulo
2π) virtually the same sequence of eventual outcomes will be observed
(Hut 1983b). Furthermore, if we look at one such sequence, we may be
able to find another critical value of x, say x∗∗ , where another near-escape
occurs. In its vicinity there will be a similar accumulation of patterns of
outcome, and furthermore this accumulation will occur infinitely often as
x → x∗ . Evidently this complicated dependence on the initial parameter
x may be visible at arbitrarily many levels of refinement, as in a fractal
pattern. Contemplation of this intricate structure gives a vivid impression
of how sensitively the outcome of resonant three-body encounters depends
on initial conditions (Fig.1).
The foregoing discussion refers, as always in this chapter, to three-body
interactions in isolation. In the context of the million-body problem, how-
ever, distant excursions of one member of the triple would be disturbed
by passing stars. The distant third body makes a very soft binary with
the centre of mass of the other two stars, and this would tend to be
disrupted, by analogy with “pressure ionisation” in gases.
In order to lend some precision to these ideas we shall consider one
of those very simple but amazingly rich problems with which dynamical
astronomy abounds. It is a special case of the restricted problem of
three bodies (i.e. the problem in which one of the stars is massless), and
is usually called “Sitnikov’s problem” (see also Chapters 4 and 13). The
two massive stars have equal masses m, and orbit in the x, y-plane (Fig.2)
with position vectors r and −r, where r is given by Eq.(2), Box 4.2. The
third body moves on the z-axis, and the symmetry of the positions of the
massive stars ensures that it stays there, though the problem is unstable
to off-axis displacements. We shall use units in which the period of the
binary is 2π. Our treatment follows Moser (1973) but omits many details.
Now, following Poincaré, we construct a surface of section for this prob-
lem, which may be done as follows. Each time the third body crosses the
plane of the binary (z = 0) we note the time t and its speed v = |ż|. Next,
using v and t as radial and angular polar coordinates, respectively, we
plot the corresponding point in a plane ∗ . Thus any point on this diagram

∗
Note that we seem to have lost information because values of t differing by a multiple
of 2π occupy the same place, but in fact this does not matter: if two test particles
 Fractals and Chaos 205

Fig. 1. Angle of deflection of the third body in a set of scattering events. The
system is coplanar, the masses are equal, the initial orbit of the binary is circular
and the initial speed of the third body is fixed. The initial conditions are then
determined by the two parameters b (which is the closest distance to which the
third body would approach the centre of the binary if it were undeflected) and φ,
which is the phase of the binary. For each pair of values the angle of deflection
of the third body is characterised by the shade of grey. There is one episode in
which the third body is almost ejected to infinity, and the “parallel” stripes arise
because the subsequent outcome is similar for motions in which the third body
arrives back at the binary after times which differ by a multiple of the binary
period. (From Boyd & McMillan 1992, copyright 1992 by the American Physical
Society)

arrive at z = 0 with the same value ż but at times differing by a multiple of 2π,
206 20 Mathematical Three-Body Scattering

1
0
0
1

1
0
0
1
11
00
00
11

Fig. 2 Sitnikov’s problem

specifies the time t at which the third body crosses the plane z = 0, and
its speed, for some solution of the problem. Following this solution, we
wait until it crosses the same plane the next time, again measure t and
v, and plot the corresponding point. In this way we have constructed a
map of the plane, called the Poincaré map, which takes the initial point
to the new point. We shall call this map φ, and say that it takes a point
(v0 , t0 ) to the point (v1 , t1 ) = φ(v0 , t0 ). Actually it is defined only on
a subset of the plane, consisting of those points which lead to motions
which do eventually recross the plane z = 0.
If e = 0 it is easy to visualise this map. In this case the two massive
bodies stay at a constant distance from the z-axis and the force experi-
enced by m3 is time-independent, even though the massive stars are in
motion. Hence it is clear that v1 = v0 . Also, the larger the value of v0
(below the escape speed) the longer the time t1 − t0 until m3 returns.
The result of these two facts is that φ rotates points along circles, the
angle of rotation being larger for larger circles (Fig.3). In fact in this
case φ is an example of a “twist” map, as the figure suggests. It can be
used to illustrate a number of concepts, such as “phase mixing”, and the
distinction between “isolating” and “non-isolating” integrals, which were
introduced in Chapter 10.
Now we go on to consider the case in which e is non-zero but small.
In this case the force experienced at a fixed point on the z-axis is time-
dependent. Therefore the energy of the massless particle is no longer

they follow the same subsequent motion with a constant time delay, because the
perturbing effect of the binary is 2π-periodic. For a similar reason it does not
matter that we have dropped the sign of ż by using v instead.
 Fractals and Chaos 207

Box 1. Sitnikov’s problem in general.

We outline some theory of the Poincaré map for Sitnikov’s problem

when e > 0. Because the potential is time dependent, the escape speed
from the plane z = 0 depends on the time at which a particle crosses this
plane. It follows that the set of points on which φ can be defined is not a
circular disk (as in the case e = 0), and we shall denote it by D0 (Fig.1).
Let its image under φ be denoted by D1 = φ(D0 ). Points in D1 which
are outside D0 give motions which escape, never returning to the plane
z = 0.

D
0

R
P

D =φ(D )
1 0
φ(R)
Fig.1. The domain of definition of the Poincaré map in the eccentric
Sitnikov problem, and its image.
Now draw a little curvilinear rectangle R in the corner beside P , where
the boundaries of D0 and D1 intersect, and consider what happens to R
under φ. As in the case e = 0, the image of R spirals infinitely many times
just inside the boundary of D1 . Therefore R itself contains infinitely many
pieces of φ(R). Different pieces correspond to different times at which m3
returns to the plane z = 0; for neighbouring pieces the times differ by
nearly one revolution of the binary.
Now a wonderful thing becomes apparent if we apply φ to R, with its
pattern of stripes, once again. Then it is apparent that, not only does
φ(R) cross R in these stripes, but within each stripe is a complete replica
of the infinite pattern of stripes. If we choose a point within a tiny stripe
inside a large stripe, we can then tell how many binary periods elapsed
between each pair of three successive crossings of the plane z = 0 (Fig.2).

constant, and it is rather clear that points corresponding to a single orbit

no longer lie on a circle in the surface of section. One might think they
208 20 Mathematical Three-Body Scattering

Box 1 continued
Once this point is understood it is also clear that, if these numbers
of periods are large enough, they can be chosen arbitrarily, simply by
choosing an arbitrary stripe within an arbitrary stripe. And we need not
stop at three successive crossings. We can repeat the process of applying
φ indefinitely, obtaining a fractal pattern of stripes within stripes within
stripes....ad infinitum. And we can choose initial conditions leading to
an arbitrary sequence of times between successive crossings (as measured,
say, in the nearest whole number of binary periods).

φ(R)
R

S S2
1

φ(S )
φ(S ) 2
1
Fig.2. A piece of the surface of section and parts of its image under φ
and φ2 (i.e. φ applied twice). Only two of the infinitely many strips are
show.
Now it is also possible to extend the argument into the past. This
time, however, we get a fractal set of stripes which cross the first set
transversely. There is a very important set of points which are obtained
by taking the points which lie on both sets of stripes at all levels of
the fractal structure. These correspond to orbits which never escape, no
matter how far forward or back in time we go. Now we can construct
orbits which exhibit any sequence of times between successive crossings,
not only into the future, but into the past as well. Examples are given in
the text.

simply lie on a slightly distorted curve, especially if e is small. Indeed this

is the kind of result that can often be established by using the famous
 Fractals and Chaos 209

φ(l )
l

Fig. 3. Sitnikov’s problem for e = 0. The figure shows how several points on
a radial line l are taken, by the action of the Poincaré map φ (thin arcs with
arrowheads) to corresponding points on a spiral curve, which is φ(l).

KAM Theorem of Hamiltonian dynamics. In the present case, however,

it would only be applicable usefully to a proportion of initial conditions
in the vicinity of the origin, i.e. to orbits of low energy. For other orbits
the behaviour of the points on the surface of section can be much more
complicated. In fact it can be shown for certain orbits that the points
sit on a fractal set (Box 1). Not surprisingly, the theory turns on the
neighbourhood of a point (P in Box 1, Fig.1) where an escaper just
escapes.
This structure has some remarkable consequences. Thus suppose we
characterise an orbit by the numbers of binary periods which elapse be-
tween each successive crossing of the plane of the binary, i.e. by a se-
quence such as . . . 2, 5, 314159, 27182818, 5 . . .. Then for each such se-

Box 1 (continued)
Now consider two points which lie on the same component of the fractal
of horizontal strips, and on the same component of φn (R), for some largest
positive n. These points both lie on I, and are very close together. On the
other hand they lie on different components of φn+1 (R), by assumption.
Therefore after n crossings of the plane z = 0 the two orbits differ greatly
in the number of orbital periods before the next crossing. This illustrates
sensitive dependence on initial data.
210 20 Mathematical Three-Body Scattering

quence∗ there is a corresponding orbit! This result is very powerful, and

produces a whole zoo of fascinating orbits. You can construct periodic
orbits using a constant sequence, capture or escape orbits by terminat-
ing the sequence with ∞ at one end or the other, and orbits which are
captured and then escape again (by terminating it at both ends). You
can even construct orbits which stay together for arbitrary long times
and then go their own separate ways. This is what we mean by sensitive
dependence on initial conditions. And similar examples can be used to
show that, when e 6= 0, not only is the energy not conserved, there is no
smooth function which is; in other words, the problem is non-integrable.

Capture

At several points in the previous section we tacitly assumed something

about the eventual fate of triple interactions, which is that they eventu-
ally result in escape of one of the particles. This is what is referred to as
“hyperbolic-elliptic” motion, as the escaper moves on a nearly hyperbolic
orbit relative to the centre of mass of the binary, while the orbit of the bi-
nary is nearly elliptic. But there are other possibilities (see Alekseev 1981,
or Arnold et al. 1997), and we even mentioned one: elliptic-parabolic mo-
tion, in which the escaper only just escapes. But why should escape, of
whatever kind, be regarded as inevitable. Couldn’t an approaching third
body be permanently captured by a binary?
We already touched on this question in the last chapter with a phys-
ical argument. But it is not hard to present, at least in outline (Box
2), a mathematical proof which shows that, while permanent capture is
not impossible, its probability is zero: among all initial conditions, the
set leading to permanent capture “has measure zero”, provided that the
interaction between the stars is non-dissipative† . The same is true of
hyperbolic-parabolic motion: almost any change in initial conditions will
tip the balance from parabolic escape to either hyperbolic escape or a
return to the vicinity of the binary.
The content of the mathematical proof bears some resemblance to the
fact that, if you keep filling a bottle, it eventually overflows. In both
situations volumes and incompressibility lie at the heart of the argument.
The other critical factor in the mathematical proof is the invariance of
the equations of motion under time translation.

∗
For technical reasons the numbers must be chosen larger than some fixed number.
This is understood in all such statements hereafter.
†
Capture even by a single star is quite possible if the interaction raises tides on either
star; cf. Chapter 27
 Capture 211

Box 2. The improbability of permanent capture.

Following J.E. Littlewood, we shall say that the third body has been
captured if, after some time, all three bodies remain forever within some
sphere S (in space) centred on the barycentre of the three bodies. Now
consider the phase space of the three-body problem, and let x(t) denote an
orbit in this space, so that x is really a vector of positions and momenta.
Let V0 be the set of initial conditions x(0) such that all three bodies lie
within S for all positive times t ≥ 0 (Fig.1). Next, let V1 be the set of
points x(1) which these orbits, originating in V0 at time t = 0, occupy
at time t = 1. It follows from Liouville’s Theorem that the volumes (or
measures) of V0 and V1 are equal. It is also easy to see that V1 lies within
V0 : for any point P in V1 lies on an orbit which started at a point in
V0 , and on the resulting orbit all three bodies subsequently lie within S;
therefore P could be used as an initial condition for an orbit lying entirely
within S; therefore P lies in V0 . To summarise so far, what we have now
established is that V1 is contained within V0 and has the same volume.
Now we concentrate on the set of points which are in V0 but not in V1 .
By what has just been proved this set has zero volume. Let us use one of
these points as initial conditions. On this orbit, all three bodies lie within
S for t ≥ 0, since x(0) lies within V0 . If, in addition, all three bodies lay
within S during the interval from t = −1 to t = 0, then the point x(−1)
could be used as initial conditions for an orbit with the bodies inside S for
t ≥ 0; in other words, x(−1) would lie in V0 , and so x(0) would lie within
V1 . This contradiction establishes that the points in V0 which are not in
V1 include those orbits which are captured during the interval between
t = −1 and t = 0. (There may be others, including orbits in which one
of the bodies is temporarily ejected from S.) Therefore the measure of
capture orbits is zero.

Invariance under time reversal does not have a role in the proof, but
can be used to extend it. By reversing time, the corresponding result
on capture shows that orbits which were bound∗ forever in the past, but
eventually escape, are also rare. The only orbits which one expects to find
are of three sorts: (i) triples which are permanently bound, (ii) triples
which never become bound, and (iii) triples which become temporarily
bound (perhaps more than once, perhaps many times).

∗
Remember that we are here defining “bound” in a particular sense, based on the
locality of the three bodies.
212 20 Mathematical Three-Body Scattering

Box 2 (continued)

V
V1 0

x(1)

x(0)
x(-1)

Fig.1 V0 is the set of initial conditions leading to capture for t ≥ 0, and

V1 is where these points are at time t = 1. The curve is an impossible
trajectory (see text).

Of these three types, the last two have become so familiar that many
people must have concluded that nothing else can happen. But stable
permanently bound triples do exist. Despite an earlier example inves-
tigated by Hénon (1976), one of the astonishing developments in this
subject took place around the start of 2000, when Chenciner & Mont-
gomery (2000) discovered, more or less by pure thought, a remarkable
stable periodic solution of the three-body problem with equal masses∗ .
The three particles chase each other round an orbit resembling a figure
8. Even more wonderful stable periodic motions are possible (Fig.4; see
also Montgomery 2001, and other papers and web references therein).

Problems
1) In a hierarchical triple the third body orbits well “outside” the inner
binary. Suppose the inner binary has period 2π and that the orbit of
the third body takes it close to the inner binary at times t0 < t1 <
t2 < . . .. We suppose that the energy of the outer binary remains
constant between close approaches, and that it changes at each close
approach by an amount that depends periodically on the phase of the

∗
It had been found previously by Moore (1993) using a numerical technique, but few
had paid much attention at that time.
 Capture 213

Fig. 4. Some periodic motions of the 3-body problem (M. Ruffert). Top left is
the orbit of Chenciner & Montgomery, with numerical initial conditions found by
C. Simó. All three particles traverse the same orbit, and the motion is periodic in
an inertial frame. The other orbits are periodic in a rotating frame. At top right,
two bodies traverse the thicker curve, while one traverses the thinner curve. In
the remaining four frames the three bodies traverse different curves. All motions
are stable for equal masses, and have zero angular momentum.
214 20 Mathematical Three-Body Scattering

inner binary at that instant, i.e. that En+1 = En +  cos tn , where En

is the energy of the outer binary between times tn−1 and tn . Show,
in a suitable scaling, that we may take tn+1 = tn + (−En+1 )−3/2
approximately. Show that the map (En , tn ) → (En+1 , tn+1 ) is area
preserving. Write a computer program to iterate this map, given , t0
and E0 to find the time tn at which the third body escapes, i.e. En+1 >
0. (You may have to insert a line to stop the program if E becomes
too small or t becomes too large.) By plotting the escape time against
E0 for fixed  and t0 at various resolutions, try to find fractal structure
in the plot.
2) Parabolic escape is improbable, and so parabolic capture followed by
parabolic escape seems doubly improbable. But is it possible?
3) “Oscillatory motion” as t → ∞ is a class of motions in which, roughly
speaking, the size of the system eventually exceeds any bound, but
there are always times in the future at which the size of the system lies
within some fixed bound. Using Sitnikov’s example, make plausible
the existence of a motion with this property both in the future and in
the past.
4) In the collinear parabolic Sitnikov problem the two masses lie at
x1 (t) = −x2 (t) = (3t)2/3/2 in suitable units, and the equation of
2z
motion of the test particle is z̈ = − 2 (Chernin & Valtonen
(z + x21 )3/2
1998). By introducing a new time variable τ = ln t and scaling z by
t2/3 , show that the equation of motion becomes that of a damped
oscillator. Hence determine and interpret the asymptotic motion of
the body as t → ∞.
 21
Analytical Approximations

The previous two chapters were intended to develop a qualitative un-

derstanding of the nature of the interactions between binary and sin-
gle stars, with no more than order-of-magnitude estimates. The present
chapter attempts to sharpen these ideas with some approximate quanti-
tative results. We imagine that binaries and single stars are distributed
throughout some region of space, such as a part of a star cluster, and we
want to know how frequently three-body interactions are taking place.

Cross Sections
What is important in applications (Chapters 23f) is the energetics of
these interactions, and that is why such stress was laid on the distinction
between soft and hard pairs in Chapter 19. In the present chapter this
consideration implies that we may be interested in interactions with bina-
ries of a given energy. Encounters with such binaries, however, are taking
place all the time with stars which approach from random directions and
random distances. Therefore, besides the energy of the binary, we usu-
ally do not know or care about the other properties of the participants,
except for their statistical distribution. This is true of the approach path
of the third body, and also usually it is true of the other 5 parameters
(besides the energy) which determine the relative motion of the binary
components∗.
An analogy with atomic scattering may now be clear. There also we
may be concerned with binaries (atoms) occupying a specific energy level
and orbital angular momentum (analogous to the energy and eccentricity
of a binary). In the atomic context interactions are characterised by the

∗
One exception may be the eccentricity; an important special case is that of binaries
in initially circular orbits.

215
216 21 Analytical Approximations

cross section of a specific type of interaction, and in stellar scattering

theory we do the same.
Consider a third body, of mass m3 , approaching a binary with relative
velocity V. The initial orbit of the intruder is characterised by the point
at which its orbit, had it been unaffected by the binary, would have
intersected an imaginary plane Π passing through the barycentre of the
binary and orthogonal to V. Suppose now we wish to calculate the cross
section for exchange reactions. We simply measure the area, in the plane
Π, corresponding to all encounters which lead to the required outcome.
Having done this for a given initial configuration of the binary, we now
repeat the process for the appropriate distribution of its parameters (such
as the orientation of its orbit). The resulting averaged area is the cross
section for the required process.
In applications we next have to convert the cross section to a reaction
rate, i.e. use expressions for the rate at which incomers strike the special
region on the plane Π. In time δt this is just the number of stars in a
volume σV δt. The relative velocity V itself has some probability den-
sity f (V), which can be computed from the distributions of velocity of
binaries and single stars, and so the reaction rate is
Z
R=n V σf (V)d3 V, (1)

where n is the space density of single stars. In order of magnitude work

R is approximated by nσv, where σ is the cross section for some typical
velocity v – what is called literally an “n-sigma-v” estimate (cf. Box
13.1). For the time being, however, we concentrate on σ.
As an illustration we shall estimate the cross section for close en-
counters, defined to be those in which the minimum distance of the in-
truder from the barycentre of the binary, Rp , is equal to a, the initial
semi-major axis of the binary. We use a Keplerian approximation for
the relative motion of the intruder and barycentre, and conservation of
energy and angular momentum lead to
pV = Rp Vp (2)
and
1 2 1 2 GM123
V = Vp − , (3)
2 2 Rp
where Vp is the relative speed at closest approach and M123 is the total
mass of all three stars. Setting Rp = a and solving for p yields the
following estimate for the cross section σ = πp2 :
 
2 2GM123
σ = πa 1+ .
aV 2
 Cross Sections 217

This formula has a simple interpretation. The geometrical cross sec-

tion of the relative orbit of the binary components is of order πa2 , but
their gravitational action on the intruder causes the intruder to be drawn
towards the binary, and this gives rise to the factor in brackets. It is
often called the “gravitational focussing” factor, where we are thinking
of an imaginary “beam” of intruders∗ . In the hard binary limit this term
dominates, and
2πGM123 a
σ' . (4)
V2
As an application of these formulae, let us estimate the rate at which
binary stars are formed in a stellar system, ignoring all numerical factors.
If a new binary is to survive it must be hard, but very hard binaries
can only be formed if three stars enter a very tiny volume, for which the
probability is tiny. Therefore we assume that a new binary is only slightly
hard. For its energy we take |ε| ∼ mv2 in the case of equal masses, where
v is the root mean square speed (cf. Eq.(19.4)). This energy corresponds
to a semi-major axis a ∼ Gm/v2 . Using Eqs.(1) and (4), we see that
the rate at which a given star encounters another at this distance is of
order G2 m2 n/v3 , where n is the number of stars per unit volume. The
probability that a third star is within this volume (of order a3 ) at the
same time is of order na3 , and so the rate of formation of triple systems
involving a given star is of order G5 m5 n2 /v9 . Since there are n stars per
unit volume, the rate is of order G5 m5 n3 /v9 per unit volume. In fact
much more elaborate calculations yield a value

G5 m5 n3
Ṅb = 0.75 , (5)
σv9

where v2 = 3σv2 (Goodman & Hut 1993). (The subscript should help
avoid confusion with a cross section.)
It has already been said that we can define cross sections for encounters
of a given type. For example, if we are interested in the amount by which
the binding energy of the binary changes we may proceed as follows. Let
∆ be the relative change in the binding energy, and let σ(∆) be the
cross section of encounters in which the relative change exceeds ∆. In
fact we already know (from a discussion in the previous chapter) that
a substantial fraction of “close” encounters changes the energy of the
binary by an amount comparable with its original value, and so Eq.(4)
is an order-of-magnitude estimate of σ(1). Equivalent information on
the behaviour of interactions is contained in the differential cross section

∗
Fig.14.1 gives a picture of what this means
218 21 Analytical Approximations

dσ/d∆. This is especially useful for the interpretation of numerical data

on cross sections (Chapter 22).
There are several ways in which the determination of cross sections
may be approached analytically. In the first place the nature of the equa-
tions of three-body motion places some non-trivial constraints on the
cross sections. We have in mind here the principle of detailed balance,
which relates the cross sections of inverse processes, i.e. reactions which
are related to each other by time-reversal. A simple physical way of un-
derstanding detailed balance is to imagine a stellar system of single and
binary stars in dynamic equilibrium, when the rates of inverse processes
must balance exactly. This allows us to relate the cross sections of in-
verse processes; these cross sections do not depend on the assumption
of dynamic equilibrium, and so these relations are valid more generally
(Problem 2).
The second analytical approach to the determination of cross sections is
to use approximate solutions of the equations of three-body motion. Dif-
ferent approximations are applicable in different regimes (Fig.1). Some
involve relatively routine calculations; high-speed encounters, for exam-
ple, can be handled by an impulsive approximation. This is relevant to
soft binaries, and these are energetically of little importance. We shall
now consider two other kinds of analytical approximation which are less
straightforward but rather more significant.

Adiabatic Encounters
Let us consider first a “wide” encounter with a hard binary, i.e. one in
which Rp considerably exceeds a. Then the angular velocity of the third
body about the binary, when at closest approach, is Vp /Rp . Now the
mean angular velocity of the binary components is given by Kepler’s 3rd
Law in the form
q
n= GM12 /a3 . (6)

(Note the change of meaning of n.) By use of Eqs.(2) and (3), it follows
that the ratio of these frequencies is
 3/2 !−1/2
nRp Rp 2M123 Rp aV 2
= + . (7)
Vp a M12 a GM12
We shall suppose that the masses are not very disparate. Since the binary
is hard, aV 2 /(GM12 )  1. It follows that, if Rp considerably exceeds
a, the time scale of the encounter will be long compared with the period
of the binary. In this case the encounter is “adiabatic”. It does not
follow that the binary exhibits adiabatic invariance, with exponentially
 Adiabatic Encounters 219

Fig. 1. Numerical and analytic estimates of cross sections for ionisation, ex-
change and resonance. The initial eccentricity of the binary is 0, and the relative
velocity of the incoming star is given in units of the minimum speed required
energetically to ionise the binary. Masses are equal. The dotted line gives the
theoretical estimate for exchange in the soft limit (v  1). The dashed line is the
theoretical estimate for ionisation, which occurs only if v ≥ 1. For equal masses
it happens to coincide very nearly with Eq.(4), the cross section for “close” en-
counters with a hard binary. This is why the numerical points follow this line
approximately for v  1. From Hut & Bahcall (1983).

small changes as in systems with one degree of freedom, because it is

a degenerate system with several (Arnold 1983, Goldstein 1980). For
example it is easily shown that, if the masses are comparable, the change
in eccentricity is of order (a/Rp )3/2 , i.e. of order the ratio of time scales
(Fig.2), and not exponentially small. But it is known that the change in
energy of the binary vanishes (as Rp /a → ∞) faster than any power of
this ratio (Heggie 1975).
It is not hard to give a calculation which shows (non-rigorously) that
the change in energy in fact declines exponentially (Box 1). To lowest
order the change in energy is given by an integral of rapidly oscillating
terms with slowly varying coefficients. (These two features arise because
of the two time scales in the problem.) Such integrals may be handled by
the method of steepest descents, which involves deforming the contour
of integration through a saddle point. It is an interesting feature of this
particular problem that the saddle is located (in the complex plane) at the
220 21 Analytical Approximations

0.008
0.006
0.004
0.002
δa/a
0
-0.002
-0.004
-0.006
-0.008
0 100 200 300 400
t

0.018
0.016
0.014
0.012
0.01
δe

0.008
0.006
0.004
0.002
0
0 100 200 300 400
t
Fig. 2. Relative change in the semi-major axis a (top) and the eccentricity e in
an adiabatic encounter, plotted against time. The binary is circular initially.

point where the third body would experience a collision with the centre of
mass of the binary, another instance of the curious role of singularities in
this subject. At this point the oscillatory terms of the integrand turn into
exponentials, and these determine the dependence of the energy change
(cf. Box 1, Eq.[7]). Not surprisingly, the exponent is essentially the ratio
of the time scales in the problem: the short period of the binary, and
the much longer time scale of the encounter. This is rather common
behaviour for the effect of adiabatic perturbations.
Now we show how to estimate the corresponding cross section. We use
Eq.(8) of Box 1 as an estimate of the relative energy change ∆, though
it is pointed out in the accompanying text that the coefficient of the
exponential also depends on the orbital parameters. When we solve for
p, however, this coefficient enters as a logarithm, and the resulting effect
on the cross section is weak. Aside from this logarithm, then, we estimate
 Adiabatic Encounters 221

Box 1. Theory of adiabatic encounters

Let ε be the energy of a binary, and let r denote the position of one
component relative to the other. Thus
m1 m2 2 Gm1 m2
ε= ṙ − ,
2M12 r
where M12 = m1 + m2 . Also,
m1 m2
ε̇ = ṙ.f , (1)
M12
where f is the perturbative acceleration due to m3 , i.e. its contribution
to the equation for r̈. The exact expression for this is
 
R − µ2 r R + µ1 r
f = Gm3 − ,
|R − µ2 r|3 |R + µ1 r|3
where µi = m1 /M12 and R is the position of m3 relative to the barycentre
of the binary. For encounters which are not too close we approximate this
at lowest (quadrupole) order in a multipole expansion, i.e. write
 
Gm3 3r.R
f '− 3
r− R . (2)
R R2
In first order perturbation theory we substitute unperturbed motion
for r and R in Eqs.(1) and (2). Terms involving r are periodic with
period 2π/n, and so the right hand side of Eq.(1) may be written as a
Fourier series (with R-dependent coefficients). By averaging Eq.(1) at
fixed R, we see that this series has no constant term, and the harmonics
of lowest order involve a factor proportional to exp(±int). (Harmonics
of higher order oscillate more rapidly, and from what follows it would be
seen that they would yield a negligible contribution to our result.) We
use a parabolic approximation for the motion of the third body (i.e. R).
The solution is standard, if relatively unfamiliar:
R = q(1 + σ2 ), (3)
where q is the distance of closest approach, and σ (not a cross section!)
is related to the time via the equation
s
1 GM123
σ + σ3 = t. (4)
3 2q 3
Here we have taken t = 0 at the instant of closest approach of the third
body, and M123 = m1 + m2 + m3 .
The relative change in energy is given by integrating Eq.(1). This yields
several terms, of which we take the integral
222 21 Analytical Approximations

Box 1 (continued)
Z ∞
I= R−3 exp(int)dt (5)
−∞

as a prototype. The above discussion of timescales implies that the ex-

ponential is rapidly oscillating compared with the remainder of the inte-
grand. It is therefore suitable for evaluation by the method of steepest
descents. In preparation for this we first transform from t to σ via Eq.(4),
which, with the aid of Eq.(3), yields
!
2inRp
Z ∞
exp [σ + σ3 /3]
2 Vp
I= dσ. (6)
Vp Rp2 −∞ (1 + σ2 )2
The method proceeds by deforming the contour of integration from the
real axis to one passing through a suitable saddle of the exponent. Dif-
ferentiating the exponent, we find that the appropriate saddle is located
at σ = i. The fact that this coincides with a singularity of the integrand
complicates the asymptotic evaluation of the integral, but does not alter
the exponential dependence of the result. Evaluating the exponent at the
saddle point we find, in fact, that
!
4 nRp
I ∝ exp − , (7)
3 Vp
though the “proportionality” involves a further power-law dependence on
the orbital parameters (cf. Problem 6). As we have seen, the exponent
here has a simple interpretation in terms of the two time scales of the
problem.
Finally we express the result in terms of the impact parameter p and
the relative speed at infinity, V . We also assume that the encounter is
not too distant, compared with the hardness of the binary, and neglect
the second term in the bracket in text Eq.(7). We can connect with the
parabolic approximation (for which p would be √ infinite) by equating the
specific angular momentum, which yields pV = 2GM123 q, and so
 1/2 !
2 (pV )3 M12
I ∝ exp − . (8)
3 (GM123a)3/2 M123
 Close Triple Encounters 223

that
4/3
πGM123 a
σ = πp2 ∼ 1/3
(3 ln |∆|)2/3 . (8)
M12 V 2
The differential cross section dσ/d∆ varies essentially as 1/|∆|. Note the
similarity to Eq.(4).

Close Triple Encounters

Now we consider the opposite extreme, i.e. encounters which can lead to
very large (positive) values of ∆. The semi-major axis of the final binary,
a1 , must be much smaller than the initial value a, and this requires the
interaction of all three stars in a region of space whose size is of order a1 .
We first outline an argument which attempts to estimate the cross sec-
tion for such processes, and then we discuss a refinement which introduces
some interesting fresh links with the theory of the three-body problem
and atomic scattering theory.
First we estimate the probability that the components of the binary
lie within a distance a1  a at the time of the encounter. The phase
space for their relative motion has coordinates r and v, and the vol-
Gm1 m2
ume of phase space accessible by a binary of energy ε = − is
Z 2a
m1 m2 2 Gm1 m2 3 3
δ(ε − v + )d rd v. After integration with respect to v
2M12 r
we see that this is
3/2 Z q
G1/2 M12
4π 2/r − 1/ad3 r.
m1 m2 r<2a

The fraction of this volume for which r < a1  a is easily seen to be

of order (a1 /a)5/2 , and this is an estimate of the required probability.
Note, however, that it is to be interpreted as applying to an ensemble of
binaries. A single binary has a specific pericentre distance, and smaller
values of r never occur.
If we assume that the binary is hard, we may easily estimate the cross
section of this tiny target by replacing a by a1 in Eq.(4). Multiplying by
the foregoing estimate of the probability that the binary components are
sufficiently close together, we find that the cross section for encounters
leading to such a small value of the semi-major axis is of order
 7/2
GM123 a a1
σ∼ , (9)
V2 a
i.e. σ ∼ GM123 a(1 + ∆)−7/2 /V 2 . Hence the differential cross section
varies as dσ/d∆ ∼ GM123 a(1 + ∆)−9/2 /V 2 .
224 21 Analytical Approximations

Now for the complication. Our previous treatment is really too rough,
because all three stars must arrive simultaneously within the small in-
teraction region of size of order a1 , and as they approach they interact
with each other in a way that cannot, as we have implicitly assumed, be
broken down into two-body approximations. Another way of saying the
same thing is that three-body gravitational focussing is different from the
two-body kind that we introduced earlier in the chapter.
A better way to handle the simultaneous close approach of three bodies
is to treat it as a perturbation from an exact triple collision. Solutions for
a triple collision can be written down exactly, and then linearisation about
one of these solutions can be adapted to provide the required information.
In the context of atomic scattering theory, and threshold scattering in
particular, this approach was initiated by G.H. Wannier (1953). It is
equally successful in the gravitational problem under consideration here.
It would take too long to analyse three-body gravitational scattering
in detail from this point of view, but we can illustrate the ideas involved
by studying close two-body encounters. In particular, we rederive a con-
sequence of Eq.(4) by looking at the relative motion of the third star and
the barycentre of the binary as a perturbation of a two-body collision or-
bit (Box 2). The trick is to find the time at which the deviation from the
collision orbit grows to become of the same order as the spatial size of the
collision solution itself. At this time the collision solution breaks down as
an approximation, and one can interpret the corresponding length scale
as the distance of closest approach.
It turns out that a similar calculation is possible for three-body mo-
tions (Simonovic & Grujic 1987, Heggie & Sweatman 1991), and it leads
to a similar relation which shows how the cross section for a close three-
body interaction within a small distance r depends on r. The calculation
is complicated because there are several configurations which yield exact
collision orbits, and because the resulting eigenvalue problems are some-
what more involved. The kind of result obtained may be illustrated by
the case of equal masses, which is quite easily worked out. In this case it
turns out that the cross section given in Eq.(9) does not √ have quite the
correct dependence on a1 : the exponent should be 2 + 13/2, ' 3.802....
While this discussion dealt with the case in which a binary is greatly
hardened by an encounter (∆  1), similar considerations occur in en-
counters in which a binary is almost destroyed. In this case all three
participants must emerge from the encounter on orbits close to an exact
triple expansion solution. When the size of the configuration has enlarged
sufficiently, one pair “peel off” to reform a binary, leaving the single star
to escape. The way in which the cross section for such events depends on
 Close Triple Encounters 225

Box 2. Close two-body encounters as perturbed collisions.

Consider the planar two-body equations

x
ẍ = − 3
r
y
ÿ = − 3 ,
r
where r2 = x2 + y 2 . An example of an exact collision orbit is given by
x = C(−t)2/3, y = 0, where C = (9/2)1/3. Linearisation about this
motion gives the following equations for the small deviations δx and δy:
4
δ̈x = δx,
9(−t)2
2
δ̈y = − δy.
9(−t)2
Searching for a power-law solution proportional to tλ leads to a simple
eigenvalue problem, and the general solution is
δx = C1 (−t)−1/3 + C2 (−t)4/3
(1)
δy = C3 (−t)1/3 + C4 (−t)2/3 ,
where the Ci are constants.
These terms have simple interpretations. The term in C1 corresponds
to a slight shift in the origin of time in the collision solution x = C(−t)2/3,
and the term with C4 corresponds to a small rotation of the collision orbit
off the x-axis. The term in C3 results from adding some angular momen-
tum to the rectilinear collision solution, as can be seen by computing the
angular momentum, and the remaining term (in C2 ) to changing its en-
ergy. We concentrate on the term in C3 . It determines both the impact
parameter p and the distance of closest approach. To make the discussion
as simple as possible we take C1 = C2 = C4 = 0.
At fixed large negative t, we see that the term in C3 ∝ p. As t in-
creases towards the time of collision (t = 0) this term becomes com-
parable with the value of x on the unperturbed (exact collision) orbit
when C3 (−t)1/3 ∼ (−t)2/3, i.e. when (−t)1/3 ∼ C3 . At this time
r ' C(−t)2/3 ∼ C32 ∝ p2 . In fact this relation shows how the min-
imum distance in the exact two-body motion is related to the impact
parameter. It follows that, for encounters in three dimensions, the cross
section for encounters at distance r is proportional to r, which we have
already established in Eq.(4).
226 21 Analytical Approximations

∆ can be obtained either by similar calculations, or by detailed balance.

Therefore similar curious exponents occur. They also occur in “ionisa-
tion” events, in which the binary is just destroyed. Analytically, the
cross section varies
√ with E, the energy of the triple system in a barycen-
( 13−1)/2
tric frame, as E when E is small and positive. In fact it had
been noticed by Hut and Bahcall (1983), on the basis of about 104 nu-
merical experiments, that the cross section varies as about E 1.3 . This
was surely one of the most bizarre numerical algorithms for computing
the square root of 13.
These numbers are closely related to the “Siegel exponents” of the
three-body problem, which are named after the mathematician who first
exhaustively analysed triple collisions (Siegel & Moser 1971). In the
neighbourhood of triple collisions an analysis like that in Box 2 gives
solutions proportional to (−t)λ√, and the possible values of λ (the Siegel
exponents) include λ = (−1 ± 13)/6 (Problem 7). The fact that some
of the exponents in this case are irrational points to the non-existence of
any technique for regularising triple collisions in general. This hinges on
whether the powers of t can be evaluated for both positive and negative
t. The contrast with two-body collisions, which are regularisable (Sund-
man 1913), is made clear by noting the form of the near-collision solution
in Eq.(1), Box 2. Those forms can be evaluated (with a real result) for
both positive and√negative t. This would not be true for irrational expo-
nents like (−1 ± 13)/6, which is the typical situation in the three-body
problem.

This completes our exploration of the analytical approximations that

may be employed in gravitational three-body scattering. Much has been
omitted, including the important question of exchange cross sections (e.g.
Hut 1983a). Enough has been done, however, to arrive at a picture of
the differential scattering cross section dσ/d∆ for hard binaries, which
is the most fundamental problem. If the masses are not too different,
Eq.(8) leads approximately to the relation dσ/d∆ ∼ Gma/(V 2 ∆) in the
limit |∆|  1, while the result we have just discussed leads to the com-
plementary result dσ/d∆ ∼ Gma/(V 2 ∆4.802...) for equal masses in the
limit that ∆  1.

Problems
1) A rapidly moving third body approaches one component of a soft
binary much more closely than the other. Use Eqs.(2) and (4) in Box
14.1, to estimate the change in velocity of this component. Substi-
tute this result into Eq.(19.5), using the statistical argument in the
associated text, and hence estimate the rate of softening of the binary.
(Impose a cutoff in impact parameter p at the semi-major axis (a) and
 Close Triple Encounters 227

at the 90◦ -deflection distance.) Hence estimate the lifetime of a soft

binary, in terms of the relaxation time.
2) Detailed balance
Let E3 be the energy of three particles (a binary and a single star
involved in an interaction), in their barycentric frame. Using Eq.(1),
show that the rate of encounters (per binary) in which E3 lies in a
range of width δE3 is 4πnV 2 σf (V)δE3 /µ, where µ is the reduced mass
of the binary and single star. If binaries have energy distribution f (ε),
i.e. the spatial number density of binaries with binding energy in a
range of width δε is f (ε)δε, deduce that the rate of encounters in which
binaries in this energy range are converted into binaries of energy
ε0 in a range δε0 , and the triples involved have energy in the stated
4πnV 2 d
range, is σ(ε, ∆)f (V)f (ε)δE3δεδε0 , where the arguments of
µ d∆
σ remind us of the parameters of the interaction, and ∆ = (ε0 − ε)/ε.
In dynamic equilibrium this is balanced by the rate of the inverse
process. Deduce that
4πnV 2 d 4πnV 02 d
σ(ε, ∆)f (V)f (ε) = σ(ε0 , ∆0 )f (V0 )f (ε0),
µ d∆ µ0 d∆0
where V 0 and ∆0 correspond to the inverse process. (Thus ∆0 =
(ε − ε0 )/ε0 , = −∆/(1 + ∆).) Using the result of Problem 19.3, and
taking f (V) to be Maxwellian, deduce that
d d
V2 σ(ε, ∆)ε−7/2 = V 02 σ(ε0 , ∆0 )ε0−7/2 .
d∆ d∆0
3) Use Eq.(5) to show that the mean time to formation of a hard binary
in a virialised system is of order N tr . (In a “hot” system enclosed by
a box it is even longer.)
4) Let t1 be the time for a particle to move a distance equal to the
90◦ -deflection distance, t2 the crossing time, t3 the relaxation time,
t4 the time to form the first binary (Problem 3), and t5 the time
to convert the entire cluster into binaries. Neglecting the Coulomb
logarithm, show that the N -dependence of these times is in geometric
progression. √
5) Obtain the estimate δh/h ∼ (m3 / M12 M123 )(a/Rp )3/2 for the rela-
tive change in angular momentum of a hard binary of semi-major axis
a which is subject to a nearly parabolic encounter with a third star
at minimum distance Rp  a. (Because the change in a decreases
exponentially with some power of Rp , as shown in Box 1, it follows
that the change in eccentricity can be much larger than the relative
change in energy. For numerical results on this, see Hut & Paczynski
1984.)
228 21 Analytical Approximations

6) For rectilinear motion of the third body in an adiabatic qencounter its

distance from the centre of mass of the binary is R = Rp2 + V 2 t2 .
Deforming the contour in Box 1, Eq.(5), to avoid t = 0, use integration
by parts to show that
Z  
1 exp(int) in 1
I= 2 q − 2 dt.
V Rp2 + V 2 t2 t t

For a very hard binary, for which n  V /Rp , the second term in the
bracket may be neglected. By deforming the contour further round
the branch point at t = iRp /V and a cut from there to t = i∞, show
that s  
2πn nRp
I exp −
Rp3 V 3 V
asymptotically.
7) An “isosceles” triple system in the plane consists of three stars of equal
mass m = 1 at the points (x, y), (−x, y) and (0, −2y). Write down
equations of motion in units such that G = 1, and verify that these are
satisfied by the collision√solution x = A(−t)2/3 , y = B(−t)2/3, where
A = 3/24/3 and B = 3/24/3 . Obtain the linearised equations for
small relative perturbations δx, δy. Show that these have solutions in
which δx, δy vary as (−t)λ , for four possible values of λ (which you
should find).
 22
Laboratory Experiments

The analytical approaches sampled in the previous chapter have some

good uses, but providing accurate useful numbers is not always one of
them. They may provide suitable scaling laws, showing how the statistics
of three-body scattering depends on the masses involved, but there is
usually an overall coefficient that must be determined in some other way.
We now turn to a technique which can fill in such gaps. Actually it
is marvellous how complementary the two techniques are. Numerical
methods are not good at determining cross sections of very rare events,
e.g. very close triples approaches, but it is often precisely these little
corners of parameter space where analytical methods are feasible.
Numerical methods offer astrophysicists a tool quite analogous to the
kinds of particle colliders in use by high-energy physicists. There beams
of particles are fired at targets (or other particles), and the relative fre-
quencies of different kinds of collision debris can be observed. Using
numerical methods we can see what happens when a binary (the target)
is fired on by a single particle, and the experiment may be repeated as
often as we care.
Numerical studies of three-body encounters go back almost one hun-
dred years. In 1920 L. Becker published results on exchange encounters
(see Chapter 19) which were carried out with the aid of “mechanical
quadrature”. But it was not until the era of electronic computing that
the investigation of triple scattering orbits became a sizable industry. In
1970 Harrington reported results from a few hundred encounters (Har-
rington 1970). By 1975, however, it was possible for Jack Hills (Hills 1975)
to publish some very useful data on the basis of about 14000 numerical
experiments, albeit with some significant restrictions on the number of
free parameters (see below). These restrictions were removed by Piet
Hut and John Bahcall in research which began to be published in 1983,
by which time they had accumulated data on over a million experiments

229
230 22 Laboratory Experiments

(Hut & Bahcall 1983). Much of the organisation of these runs was still
done by hand, and the process was eventually fully automated by Piet
Hut and Steve McMillan by 1995 (McMillan & Hut 1996).
There are two issues to be confronted with numerical techniques for the
scattering problem. One is the question of how to integrate the equations
of motion, and the other is the matter of how to organise the results.
These methodological issues are the themes of this chapter.

Numerical Integration

At first sight it is a simple matter to integrate the three-body equations

numerically. One can take an integration method off the shelf, like the
4th-order Runge-Kutta method. Indeed this usually works quite effec-
tively, though a sensible method of choosing the time step is a big help.
One effective method is to repeat each step with two steps of half the
size, which allows an estimate of the truncation error (cf. Press et al.
1992) . The time step is controlled by keeping this error within pre-
scribed bounds. The error estimate also allows the error to be corrected
approximately.
As an example we may consider the well-known case of the three-body
problem posed by Burrau (1913). It takes particles of mass 3, 4 and 5
initially at rest at the vertices of a 3−4 −5 Pythagorean triangle. Burrau
conjectured that this configuration would lead to a periodic orbit of the
general three-body problem. During its evolution there is one very close
approach of two of the bodies, and it was thought at one time (Szebehely
& Peters 1967) that a special treatment of close approaches (Chapter 15)
was needed to cope with this, but in fact the Runge-Kutta method with
adaptive step size control works well.
Other standard numerical methods which are often adopted are higher-
order Runge-Kutta methods, Hermite integration (Makino & Aarseth
1992, and Problem A.1), which requires computation of both the acceler-
ation and its derivative∗ , and the method of Bulirsch and Stoer (cf. Press
et al. 1992). The latter repeats the technique of interval-halving several
times for each integration step.
None of these methods take account of several important features of
the equations of motion of the three-body problem. In the first place the

∗
The derivative of the acceleration is sometimes called jerk. Perhaps the position and
its successive derivatives should be referred to by the letters x, v, a, j, s, c and p
(up to the 6th derivative), as we have been given to understand that some physicists
refer to the last three as snap, crackle and pop. See
http://www.weburbia.com/physics/jerk.html.
 Numerical Integration 231

equations may be written in the form

r̈i = ai (r1 , r2 , r3 ), (i = 1, 2, 3) (1)
where ai is the acceleration of the ith star. The usual forms of the above
methods force the user to rewrite the equations as first-order equations
ṙi = vi ,
(2)
v̇i = ai (r1 , r2 , r3 ),
where vi is the velocity of the ith star. The methods can, however, be
specially adapted to equations of the form (1). In the case of Runge-Kutta
algorithms the corresponding methods are referred to as “Runge-Kutta-
Nystrom” methods.
Next, none of these methods pay attention to the fact that Eqs.(2) are
canonical, i.e. they are Hamilton’s equations for an appropriate Hamil-
tonian. Now one of the analytical properties of Hamilton’s equations is
that they are a “symplectic” dynamical system. This is a generalisation
to systems with n degrees of freedom of the idea of area-preservation for
systems with 1 degree of freedom. For example, the one-dimensional sim-
ple harmonic oscillator has equations of motion which can be written in
Hamiltonian form as
ẋ = v,
(3)
v̇ = −x.
Now a solution of these equations can be plotted as a moving point in
phase space (with coordinates x and v), and the velocity of this point has
components given by the right-hand sides of Eqs.(3). Thus phase space is
equipped with a velocity field, and the important point to notice is that
this field is divergence-free: ∂ ẋ/∂x + ∂ v̇/∂v = 0. Thus if we follow the
motion of a two-dimensional region in phase space as it is carried along
by the flow, its volume does not change.
It might seem desirable to invent an integration algorithm which also
exhibited this property of area preservation (or, in the case of interest
to us, its generalisation to the concept of symplecticness). Runge-Kutta
methods are not symplectic in general, but there is another class of algo-
rithms which is. The simplest example is the “leapfrog” algorithm, often
called the Verlet algorithm in computational physics. For equations of
the form (2) this may be written as
n+1/2
rn+1
i = rn
i + hvi ,
n+3/2 n+1/2
(4)
vi = vi + han+1
i .
Here the superscript indicates the time, in multiples of the stepsize h
after some initial time. In the case of the simple harmonic oscillator the
232 22 Laboratory Experiments

equations reduce to

xn+1 = xn + hvn+1/2 ,
(5)
vn+3/2 = vn+1/2 − hxn+1 .

∂(xn+1 , vn+3/2 )
By computing the Jacobian determinant it is easy to
∂(xn , vn+1/2 )
show that this is area-preserving∗
It might seem to have nothing to do with area preservation or sym-
plecticness, but in practice an algorithm like the leapfrog method is very
good for long-term energy conservation. The fact that the method is
area-preserving essentially shows that the numerically generated phase
path does not spiral out from (or in to) the origin, which is precisely
the behaviour exhibited by the Euler method (Problem 1 and Fig.1). In
terms of order, i.e. the way in which the local truncation error depends
on h, the leapfrog is very primitive. It is not hard, however, to invent
symplectic methods of higher order (Kinoshita et al. 1991).
Our discussion of the symplectic property of the leapfrog integrator
made the unstated assumption that h was constant. In practice, as we
have already seen, h should be allowed to vary, and then it is not clear
that the symplectic property of the algorithm is preserved. This stum-
bling block to the implementation of effective symplectic integrators was
removed by Hut, Makino and McMillan (1995), who realised that the
way out was to incorporate another exact property of equations (1): their
time-reversal invariance. Observe that, if ri (t) is a solution of Eq.(1) then
so is ri (−t). One consequence of this fact is that, if the velocities of three
stars in a triple system are reversed at some instant, then the stars will
exactly retrace their orbits. Here is another property which we should
perhaps try to incorporate into a numerical integrator.
Let us look at Eqs.(4) from this point of view. By inverting these we
immediately find

n+1 n+1/2
rn
i = ri + (−h)vi ,
n+1/2 n+3/2
vi = vi + (−h)an+1
i .

∗
Strictly speaking an integrator for a problem with one-degree of freedom is sym-
plectic if the map (xn , vn ) → (xn+1 , vn+1 ) is area-preserving. The fact that v is
discretised at intermediate times complicates the issue, but we shall not explore this
detail here.
 Numerical Integration 233
2
1.5
1
0.5
0
-0.5
y

-1
-1.5
-2
-2.5
-3
-1 0 1 2 3 4 5 6
x

Fig. 1. Three numerical computations of a Kepler orbit with e = 0.9 and a = 1:

(a, dotted) Euler (cf. Problem 1), with time step proportional to r3/2 , (b, short-
dashed) leapfrog, with fixed time step, and (c, solid) using the method of Mikkola
& Tanikawa (1999). (This is a leapfrog algorithm based on a slightly more
complicated Hamiltonian than the usual Kepler Hamiltonian, cf. Chapter 15.)
Runs (b) and (c) used similar numbers of steps, and run (a) used about three
times as many. The purpose of the figure is not, however, to compare efficiency,
but to illustrate the properties of different algorithms. Euler appears to preserve
the direction of pericentre but not the energy, and with leapfrog the behaviour
is reversed. The third, i.e. (c), preserves both, but not the phase (though this
is not evident in this plot).

Since these are true at each step, we can reduce the superscript on the
second equation by one, and obtain
n+1 n+1/2
rn
i = ri + (−h)vi ,
n−1/2 n+1/2
vi = vi + (−h)an
i.

Note that these equations are of exactly the same form as Eqs.(4), except
that h is replaced by −h and the superscript is running downwards. Thus
if we stop a leapfrog integration at some point, reverse h, and continue
using the leapfrog algorithm, we recover the previous results (except for
rounding error). This is the analogue (for algorithms) of the time-reversal
invariance of the equations of motion.
Now the above argument is ruined if we allow the time step to vary, for
example as a function of the accelerations ai , unless we ensure that the
time-step respects the time-reversal invariance as well. One possibility is
to ensure that h depends on a symmetric function of the values at the
234 22 Laboratory Experiments

beginning and the end of the time step. It turns out that such a procedure
works very well.
It is still possible to break a time-symmetric, symplectic integrator
with sufficiently close approaches between two or three of the bodies.
There is a formulation, based on a different Hamiltonian, which will ac-
curately compute even a collision orbit (Mikkola & Tanikawa 1999, see
Chapter 15). The traditional way out, however, is the technique of two-
body regularisation, also discussed in Chapter 15. Its extension to three-
and other few-body systems is best carried out with the “chain regular-
isation” of Mikkola & Aarseth (1993), which is an extension of an old
3-body method devised by Aarseth & Zare (1974)∗. We recommend the
survey by Mikkola (1997) for a fascinating glimpse of the history behind
these methods. In recent years, however, it has even become possible to
combine these regularisation techniques with time-symmetrisation (Fu-
nato et al. 1996, Leimkuhler 1999).
The final point to be mentioned in connection with numerical methods
is the accuracy which it is feasible or desirable to insist on. This is really
the same question for N = 3 as we discussed in general in Chapter 13.
There we already saw that there are situations in which the slightest
change in the initial conditions of a three-body interaction can lead to
drastic changes in the final result. Dejonghe & Hut (1987) have discovered
examples in which an error in initial conditions may be magnified by over
100 orders of magnitude before the final fate of the triple system is settled.
These considerations show that one cannot hope to devise any method
which will produce an accurate solution for all triple interactions.

Organisation of the calculations

Each triple encounter requires the specification of initial conditions. As
the three-body problem has nine degrees of freedom, it might be thought
that no fewer than 18 parameters are required. However, it is possible to
work in a barycentric frame, reducing the tally to 12 parameters. One
more may be removed by choice of a unit of length (e.g. the initial
semi-major axis of the binary). We can rotate coordinates so that the
binary initially sits in the x, y plane with its major axis along the x-axis,
reducing the count by three (recall that the orientation of the orbit could
be specified by the three Euler angles). Next, it is usual to start the
integration with the third body at some fixed distance from the binary,
sufficiently large that the effect on the binary of the incomer up to that
point are negligible; this leaves only 7 parameters to be specified. These

∗
For N = 3 there is an alternative due to G. Lemaitre (1955), though he is better
known nowadays for his work in cosmology.
 Organisation of the calculations 235

may be the initial eccentricity and phase of the binary, and all but one
of the initial coordinates and all velocity components of the incomer.
Against these simplifications, however, we must not forget that the result
depends on the masses, though scaling (e.g. so that the total mass is
unity) removes one.
In a given problem it is usually clear how these parameters are to
be distributed. In a simple situation, for instance, we are interested in
encounters between stars of given masses, and the initial speed of the
intruder and eccentricity of the binary are specified. Then the initial
phase of the binary corresponds to a uniformly random choice of the
initial time, while the remaining parameters of the intruder correspond to
a uniform distribution in space and an isotropic distribution of velocities.
It is customary to select such initial conditions using a well designed
random number generator, but there is some advantage in these problems
in using quasi-random numbers (Press et al. 1992),which aim to sample
the domain of initial conditions more purposefully.
One of the parameters commonly used to specify the initial conditions
of the intruder is the impact parameter, p. i.e. the distance of closest
approach between it and the barycentre of the binary, if it was unde-
flected by the binary. This differs from all other parameters in that it is
unbounded, and it is the most troublesome parameter of all to deal with.
Some upper limit must be imposed, and it is legitimate to do so, for if
p is too large the intruder flies wide of the binary and does not affect
it significantly. But where should the limit be set? The problem with
the choice of maximum impact parameter is this: allowing too large an
impact parameter can imply a large waste of computer time on uninter-
esting orbits; while choosing too small an impact parameter (e.g. one
much smaller than the initial semi-major axis of the binary) will yield a
systematic underestimate of some cross sections, since some encounters
of interest will be missed.
One manual way of handling this situation is to carry out a pilot study
to determine the range of impact parameters in which the encounters of
interest are observed to take place. Only then does one embark on pro-
duction runs, intended to compute cross sections with desired statistical
accuracy (Fig.2).
It is now possible to automate the entire process (McMillan & Hut
1996). Rather than relying on human inspection of pilot calculations,
software has been written which includes an automatic feedback system
ensuring near-optimal coverage of the range of impact parameters while
guaranteeing completeness. The basic idea is to maintain a safety zone
outside the ‘bull’s eye’ region, defined as the area where interesting reac-
236 22 Laboratory Experiments

Fig. 2. Numerically determined cross section for hard binary scattering for equal
v 2 dσ
masses, and a semi-analytic fit. The scaled differential cross section P =
πa2 d∆
is plotted, with error bars, against the relative increase in binding energy, ∆,
where v is the initial relative speed of the third body and the binary, in units of
the minimum speed required for ionisation. From Heggie & Hut (1993).

tions take place (where ‘interesting’ is defined according to the particular

purpose underlying a given set of experiments). Allowing for the possibil-
ity of dynamically enlarging this safety zone guarantees rapid convergence
towards accurate cross sections.
The automated scattering software we are describing was designed in
a multi-layer object-oriented approach and is implemented in C++. At
the lowest layer, the orbit integration engine consists of a fourth-order,
time-symmetric, variable–time-step Hermite integrator. This technique
provides excellent long-term numerical stability, something that is essen-
tial for the treatment of long-lasting resonance scattering events.
On top of this lowest layer, there are several layers that contain: 1)
checks to determine whether a given scattering experiment has reached
its final outgoing state; 2) checks to allow optimization features to be
activated, such as analytical integration of inner and outer orbits of hi-
erarchical triple systems in which the outer orbital period vastly exceeds
the inner orbital period; 3) diagnostic functions to store information de-
scribing the build-up of energy errors; 4) various bookkeeping functions
that chart the overall character of the orbits (e. g. democratic versus
hierarchical resonance states); and 5) checks for overlap of stellar radii,
in which case merging routines are invoked that can replace two colliding
stars with a single merger product.
 Organisation of the calculations 237

On top of these layers, the first user-accessible layer contains a single-

scattering command, with a large number of options. The masses and
radii of the stars can be specified, as well as the orbital parameters of
the binary, the impact parameter of the encounter, and the relative ve-
locity, asymptotically far before the encounter. The initial distance from
which the integration starts is determined automatically (and will be
much larger for, say, a 10M black hole approaching a given binary com-
pared to a 0.5M dwarf). In addition, an overall accuracy parameter
gives a handle on the cost/performance tradeoff. In practice, typical rel-
ative energy errors can be easily kept as small as 10−10 ; production runs,
however, usually aim at errors of order 10−6 , allowing a speed-up of a fac-
tor ten in computer time with respect to the most accurate integrations
attainable.
The next layer contains all the management software to conduct a series
of scattering experiments. Depending on the type of total or differential
cross section requested, the user can choose an appropriate command
to activate a ‘beam’ of single stars aimed at the ‘target plate’ of binary
stars. After a few minutes, a preliminary report appears on the screen,
with estimates of all relevant cross sections plus their corresponding error
bars (Box 1). Thereafter, subsequent reports appear, each one after a
four times longer interval. Because of the Monte Carlo nature of the
orbit parameter sampling, each following report carries error bars that
are half the size of the corresponding ones in the previous report. Thus
reasonable estimates can often be obtained in ten or fifteen minutes,
with more accurate results following in an hour (on a fast modern work
station). This is in marked contrast with the duration of experiments in
a particle collider in high-energy physics!

Problems
1) Solve Eqs.(5) exactly, for initial conditions x0 = 1, v1/2 = 0. Find
out in what sense the solution is superior to that of the corresponding
Euler algorithm
xn+1 = xn + hvn ,
vn+1 = vn − hxn
with initial conditions x0 = 1, v0 = 0.
2) Compare the total of the cross sections at the bottom of Box 1 with
the estimate for the cross section of close encounters (Eq.(21.4)).
3) For fixed masses, how many parameters are needed to specify a three-
body scattering experiment (a) in two dimensions, (b) in one dimen-
sion?
238 22 Laboratory Experiments

Box 1 Example of computational determination of scattering

cross sections.

The command sigma3 -d 128 -v 1.2 -m 0.3 uses part of the “star-
lab” package (http://www.manybody.org/starlab.html) to estimate the
cross section for various outcomes of encounters between a single star
and a binary. The binary component masses are 1 − m and m, and the
mass of the third star may be set with a further option. The other two
parameters are d, which is a measure of the density of experiments in
a central trial √
zone (the fractional error in the largest cross section will
be of order 1/ d), and v, the speed of the third body when at a large
distance, in units of the value at which the total barycentric energy of
the triple system is zero. Here is a summary of the results. It may not
look beautiful to the reader, but it does to anyone who has organised this
kind of thing by hand.
scatter profile:
m1=0.7 m2=0.3 m3=0.5 r1=0 r2=0 r3=0 v inf=1.2
----------------------------------------------------
total scatterings=25, CPU time=7.45
exch1 exch2 ion error stop
2.220 0.888 3.108 0.000 0.000
+-1.83 +-.628 +-2.03 +-.000 +-.000
----------------------------------------------------
total scatterings=60, CPU time=16.94
exch1 exch2 ion error stop
0.444 0.999 0.777 0.000 0.000
+-.222 +-.333 +-.294 +-.000 +-.000
----------------------------------------------------
total scatterings=312, CPU time=105.6
exch1 exch2 ion error stop
0.768 1.224 0.712 0.000 0.000
+-.158 +-.199 +-.153 +-.000 +-.000
----------------------------------------------------
total scatterings=1560, CPU time=487.97
exch1 exch2 ion error stop
1.025 1.181 0.861 0.000 0.000
+-.101 +-.105 +-.089 +-.000 +-.000
v2 σ
The values are those of , where v is defined as above. Notice that
πa2
the cross section for exchanging the heavier binary companion is smaller.
Organisation of the calculations 239
 22
Laboratory Experiments

The analytical approaches sampled in the previous chapter have some

good uses, but providing accurate useful numbers is not always one of
them. They may provide suitable scaling laws, showing how the statistics
of three-body scattering depends on the masses involved, but there is
usually an overall coefficient that must be determined in some other way.
We now turn to a technique which can fill in such gaps. Actually it
is marvellous how complementary the two techniques are. Numerical
methods are not good at determining cross sections of very rare events,
e.g. very close triples approaches, but it is often precisely these little
corners of parameter space where analytical methods are feasible.
Numerical methods offer astrophysicists a tool quite analogous to the
kinds of particle colliders in use by high-energy physicists. There beams
of particles are fired at targets (or other particles), and the relative fre-
quencies of different kinds of collision debris can be observed. Using
numerical methods we can see what happens when a binary (the target)
is fired on by a single particle, and the experiment may be repeated as
often as we care.
Numerical studies of three-body encounters go back almost one hun-
dred years. In 1920 L. Becker published results on exchange encounters
(see Chapter 19) which were carried out with the aid of “mechanical
quadrature”. But it was not until the era of electronic computing that
the investigation of triple scattering orbits became a sizable industry. In
1970 Harrington reported results from a few hundred encounters (Har-
rington 1970). By 1975, however, it was possible for Jack Hills (Hills 1975)
to publish some very useful data on the basis of about 14000 numerical
experiments, albeit with some significant restrictions on the number of
free parameters (see below). These restrictions were removed by Piet
Hut and John Bahcall in research which began to be published in 1983,
by which time they had accumulated data on over a million experiments

229
230 22 Laboratory Experiments

(Hut & Bahcall 1983). Much of the organisation of these runs was still
done by hand, and the process was eventually fully automated by Piet
Hut and Steve McMillan by 1995 (McMillan & Hut 1996).
There are two issues to be confronted with numerical techniques for the
scattering problem. One is the question of how to integrate the equations
of motion, and the other is the matter of how to organise the results.
These methodological issues are the themes of this chapter.

Numerical Integration

At first sight it is a simple matter to integrate the three-body equations

numerically. One can take an integration method off the shelf, like the
4th-order Runge-Kutta method. Indeed this usually works quite effec-
tively, though a sensible method of choosing the time step is a big help.
One effective method is to repeat each step with two steps of half the
size, which allows an estimate of the truncation error (cf. Press et al.
1992) . The time step is controlled by keeping this error within pre-
scribed bounds. The error estimate also allows the error to be corrected
approximately.
As an example we may consider the well-known case of the three-body
problem posed by Burrau (1913). It takes particles of mass 3, 4 and 5
initially at rest at the vertices of a 3−4 −5 Pythagorean triangle. Burrau
conjectured that this configuration would lead to a periodic orbit of the
general three-body problem. During its evolution there is one very close
approach of two of the bodies, and it was thought at one time (Szebehely
& Peters 1967) that a special treatment of close approaches (Chapter 15)
was needed to cope with this, but in fact the Runge-Kutta method with
adaptive step size control works well.
Other standard numerical methods which are often adopted are higher-
order Runge-Kutta methods, Hermite integration (Makino & Aarseth
1992, and Problem A.1), which requires computation of both the acceler-
ation and its derivative∗ , and the method of Bulirsch and Stoer (cf. Press
et al. 1992). The latter repeats the technique of interval-halving several
times for each integration step.
None of these methods take account of several important features of
the equations of motion of the three-body problem. In the first place the

∗
The derivative of the acceleration is sometimes called jerk. Perhaps the position and
its successive derivatives should be referred to by the letters x, v, a, j, s, c and p
(up to the 6th derivative), as we have been given to understand that some physicists
refer to the last three as snap, crackle and pop. See
http://www.weburbia.com/physics/jerk.html.
 Numerical Integration 231

equations may be written in the form

r̈i = ai (r1 , r2 , r3 ), (i = 1, 2, 3) (1)
where ai is the acceleration of the ith star. The usual forms of the above
methods force the user to rewrite the equations as first-order equations
ṙi = vi ,
(2)
v̇i = ai (r1 , r2 , r3 ),
where vi is the velocity of the ith star. The methods can, however, be
specially adapted to equations of the form (1). In the case of Runge-Kutta
algorithms the corresponding methods are referred to as “Runge-Kutta-
Nystrom” methods.
Next, none of these methods pay attention to the fact that Eqs.(2) are
canonical, i.e. they are Hamilton’s equations for an appropriate Hamil-
tonian. Now one of the analytical properties of Hamilton’s equations is
that they are a “symplectic” dynamical system. This is a generalisation
to systems with n degrees of freedom of the idea of area-preservation for
systems with 1 degree of freedom. For example, the one-dimensional sim-
ple harmonic oscillator has equations of motion which can be written in
Hamiltonian form as
ẋ = v,
(3)
v̇ = −x.
Now a solution of these equations can be plotted as a moving point in
phase space (with coordinates x and v), and the velocity of this point has
components given by the right-hand sides of Eqs.(3). Thus phase space is
equipped with a velocity field, and the important point to notice is that
this field is divergence-free: ∂ ẋ/∂x + ∂ v̇/∂v = 0. Thus if we follow the
motion of a two-dimensional region in phase space as it is carried along
by the flow, its volume does not change.
It might seem desirable to invent an integration algorithm which also
exhibited this property of area preservation (or, in the case of interest
to us, its generalisation to the concept of symplecticness). Runge-Kutta
methods are not symplectic in general, but there is another class of algo-
rithms which is. The simplest example is the “leapfrog” algorithm, often
called the Verlet algorithm in computational physics. For equations of
the form (2) this may be written as
n+1/2
rn+1
i = rn
i + hvi ,
n+3/2 n+1/2
(4)
vi = vi + han+1
i .
Here the superscript indicates the time, in multiples of the stepsize h
after some initial time. In the case of the simple harmonic oscillator the
232 22 Laboratory Experiments

equations reduce to

xn+1 = xn + hvn+1/2 ,
(5)
vn+3/2 = vn+1/2 − hxn+1 .

∂(xn+1 , vn+3/2 )
By computing the Jacobian determinant it is easy to
∂(xn , vn+1/2 )
show that this is area-preserving∗
It might seem to have nothing to do with area preservation or sym-
plecticness, but in practice an algorithm like the leapfrog method is very
good for long-term energy conservation. The fact that the method is
area-preserving essentially shows that the numerically generated phase
path does not spiral out from (or in to) the origin, which is precisely
the behaviour exhibited by the Euler method (Problem 1 and Fig.1). In
terms of order, i.e. the way in which the local truncation error depends
on h, the leapfrog is very primitive. It is not hard, however, to invent
symplectic methods of higher order (Kinoshita et al. 1991).
Our discussion of the symplectic property of the leapfrog integrator
made the unstated assumption that h was constant. In practice, as we
have already seen, h should be allowed to vary, and then it is not clear
that the symplectic property of the algorithm is preserved. This stum-
bling block to the implementation of effective symplectic integrators was
removed by Hut, Makino and McMillan (1995), who realised that the
way out was to incorporate another exact property of equations (1): their
time-reversal invariance. Observe that, if ri (t) is a solution of Eq.(1) then
so is ri (−t). One consequence of this fact is that, if the velocities of three
stars in a triple system are reversed at some instant, then the stars will
exactly retrace their orbits. Here is another property which we should
perhaps try to incorporate into a numerical integrator.
Let us look at Eqs.(4) from this point of view. By inverting these we
immediately find

n+1 n+1/2
rn
i = ri + (−h)vi ,
n+1/2 n+3/2
vi = vi + (−h)an+1
i .

∗
Strictly speaking an integrator for a problem with one-degree of freedom is sym-
plectic if the map (xn , vn ) → (xn+1 , vn+1 ) is area-preserving. The fact that v is
discretised at intermediate times complicates the issue, but we shall not explore this
detail here.
 Numerical Integration 233
2
1.5
1
0.5
0
-0.5
y

-1
-1.5
-2
-2.5
-3
-1 0 1 2 3 4 5 6
x

Fig. 1. Three numerical computations of a Kepler orbit with e = 0.9 and a = 1:

(a, dotted) Euler (cf. Problem 1), with time step proportional to r3/2 , (b, short-
dashed) leapfrog, with fixed time step, and (c, solid) using the method of Mikkola
& Tanikawa (1999). (This is a leapfrog algorithm based on a slightly more
complicated Hamiltonian than the usual Kepler Hamiltonian, cf. Chapter 15.)
Runs (b) and (c) used similar numbers of steps, and run (a) used about three
times as many. The purpose of the figure is not, however, to compare efficiency,
but to illustrate the properties of different algorithms. Euler appears to preserve
the direction of pericentre but not the energy, and with leapfrog the behaviour
is reversed. The third, i.e. (c), preserves both, but not the phase (though this
is not evident in this plot).

Since these are true at each step, we can reduce the superscript on the
second equation by one, and obtain
n+1 n+1/2
rn
i = ri + (−h)vi ,
n−1/2 n+1/2
vi = vi + (−h)an
i.

Note that these equations are of exactly the same form as Eqs.(4), except
that h is replaced by −h and the superscript is running downwards. Thus
if we stop a leapfrog integration at some point, reverse h, and continue
using the leapfrog algorithm, we recover the previous results (except for
rounding error). This is the analogue (for algorithms) of the time-reversal
invariance of the equations of motion.
Now the above argument is ruined if we allow the time step to vary, for
example as a function of the accelerations ai , unless we ensure that the
time-step respects the time-reversal invariance as well. One possibility is
to ensure that h depends on a symmetric function of the values at the
234 22 Laboratory Experiments

beginning and the end of the time step. It turns out that such a procedure
works very well.
It is still possible to break a time-symmetric, symplectic integrator
with sufficiently close approaches between two or three of the bodies.
There is a formulation, based on a different Hamiltonian, which will ac-
curately compute even a collision orbit (Mikkola & Tanikawa 1999, see
Chapter 15). The traditional way out, however, is the technique of two-
body regularisation, also discussed in Chapter 15. Its extension to three-
and other few-body systems is best carried out with the “chain regular-
isation” of Mikkola & Aarseth (1993), which is an extension of an old
3-body method devised by Aarseth & Zare (1974)∗. We recommend the
survey by Mikkola (1997) for a fascinating glimpse of the history behind
these methods. In recent years, however, it has even become possible to
combine these regularisation techniques with time-symmetrisation (Fu-
nato et al. 1996, Leimkuhler 1999).
The final point to be mentioned in connection with numerical methods
is the accuracy which it is feasible or desirable to insist on. This is really
the same question for N = 3 as we discussed in general in Chapter 13.
There we already saw that there are situations in which the slightest
change in the initial conditions of a three-body interaction can lead to
drastic changes in the final result. Dejonghe & Hut (1987) have discovered
examples in which an error in initial conditions may be magnified by over
100 orders of magnitude before the final fate of the triple system is settled.
These considerations show that one cannot hope to devise any method
which will produce an accurate solution for all triple interactions.

Organisation of the calculations

Each triple encounter requires the specification of initial conditions. As
the three-body problem has nine degrees of freedom, it might be thought
that no fewer than 18 parameters are required. However, it is possible to
work in a barycentric frame, reducing the tally to 12 parameters. One
more may be removed by choice of a unit of length (e.g. the initial
semi-major axis of the binary). We can rotate coordinates so that the
binary initially sits in the x, y plane with its major axis along the x-axis,
reducing the count by three (recall that the orientation of the orbit could
be specified by the three Euler angles). Next, it is usual to start the
integration with the third body at some fixed distance from the binary,
sufficiently large that the effect on the binary of the incomer up to that
point are negligible; this leaves only 7 parameters to be specified. These

∗
For N = 3 there is an alternative due to G. Lemaitre (1955), though he is better
known nowadays for his work in cosmology.
 Organisation of the calculations 235

may be the initial eccentricity and phase of the binary, and all but one
of the initial coordinates and all velocity components of the incomer.
Against these simplifications, however, we must not forget that the result
depends on the masses, though scaling (e.g. so that the total mass is
unity) removes one.
In a given problem it is usually clear how these parameters are to
be distributed. In a simple situation, for instance, we are interested in
encounters between stars of given masses, and the initial speed of the
intruder and eccentricity of the binary are specified. Then the initial
phase of the binary corresponds to a uniformly random choice of the
initial time, while the remaining parameters of the intruder correspond to
a uniform distribution in space and an isotropic distribution of velocities.
It is customary to select such initial conditions using a well designed
random number generator, but there is some advantage in these problems
in using quasi-random numbers (Press et al. 1992),which aim to sample
the domain of initial conditions more purposefully.
One of the parameters commonly used to specify the initial conditions
of the intruder is the impact parameter, p. i.e. the distance of closest
approach between it and the barycentre of the binary, if it was unde-
flected by the binary. This differs from all other parameters in that it is
unbounded, and it is the most troublesome parameter of all to deal with.
Some upper limit must be imposed, and it is legitimate to do so, for if
p is too large the intruder flies wide of the binary and does not affect
it significantly. But where should the limit be set? The problem with
the choice of maximum impact parameter is this: allowing too large an
impact parameter can imply a large waste of computer time on uninter-
esting orbits; while choosing too small an impact parameter (e.g. one
much smaller than the initial semi-major axis of the binary) will yield a
systematic underestimate of some cross sections, since some encounters
of interest will be missed.
One manual way of handling this situation is to carry out a pilot study
to determine the range of impact parameters in which the encounters of
interest are observed to take place. Only then does one embark on pro-
duction runs, intended to compute cross sections with desired statistical
accuracy (Fig.2).
It is now possible to automate the entire process (McMillan & Hut
1996). Rather than relying on human inspection of pilot calculations,
software has been written which includes an automatic feedback system
ensuring near-optimal coverage of the range of impact parameters while
guaranteeing completeness. The basic idea is to maintain a safety zone
outside the ‘bull’s eye’ region, defined as the area where interesting reac-
236 22 Laboratory Experiments

Fig. 2. Numerically determined cross section for hard binary scattering for equal
v 2 dσ
masses, and a semi-analytic fit. The scaled differential cross section P =
πa2 d∆
is plotted, with error bars, against the relative increase in binding energy, ∆,
where v is the initial relative speed of the third body and the binary, in units of
the minimum speed required for ionisation. From Heggie & Hut (1993).

tions take place (where ‘interesting’ is defined according to the particular

purpose underlying a given set of experiments). Allowing for the possibil-
ity of dynamically enlarging this safety zone guarantees rapid convergence
towards accurate cross sections.
The automated scattering software we are describing was designed in
a multi-layer object-oriented approach and is implemented in C++. At
the lowest layer, the orbit integration engine consists of a fourth-order,
time-symmetric, variable–time-step Hermite integrator. This technique
provides excellent long-term numerical stability, something that is essen-
tial for the treatment of long-lasting resonance scattering events.
On top of this lowest layer, there are several layers that contain: 1)
checks to determine whether a given scattering experiment has reached
its final outgoing state; 2) checks to allow optimization features to be
activated, such as analytical integration of inner and outer orbits of hi-
erarchical triple systems in which the outer orbital period vastly exceeds
the inner orbital period; 3) diagnostic functions to store information de-
scribing the build-up of energy errors; 4) various bookkeeping functions
that chart the overall character of the orbits (e. g. democratic versus
hierarchical resonance states); and 5) checks for overlap of stellar radii,
in which case merging routines are invoked that can replace two colliding
stars with a single merger product.
 Organisation of the calculations 237

On top of these layers, the first user-accessible layer contains a single-

scattering command, with a large number of options. The masses and
radii of the stars can be specified, as well as the orbital parameters of
the binary, the impact parameter of the encounter, and the relative ve-
locity, asymptotically far before the encounter. The initial distance from
which the integration starts is determined automatically (and will be
much larger for, say, a 10M black hole approaching a given binary com-
pared to a 0.5M dwarf). In addition, an overall accuracy parameter
gives a handle on the cost/performance tradeoff. In practice, typical rel-
ative energy errors can be easily kept as small as 10−10 ; production runs,
however, usually aim at errors of order 10−6 , allowing a speed-up of a fac-
tor ten in computer time with respect to the most accurate integrations
attainable.
The next layer contains all the management software to conduct a series
of scattering experiments. Depending on the type of total or differential
cross section requested, the user can choose an appropriate command
to activate a ‘beam’ of single stars aimed at the ‘target plate’ of binary
stars. After a few minutes, a preliminary report appears on the screen,
with estimates of all relevant cross sections plus their corresponding error
bars (Box 1). Thereafter, subsequent reports appear, each one after a
four times longer interval. Because of the Monte Carlo nature of the
orbit parameter sampling, each following report carries error bars that
are half the size of the corresponding ones in the previous report. Thus
reasonable estimates can often be obtained in ten or fifteen minutes,
with more accurate results following in an hour (on a fast modern work
station). This is in marked contrast with the duration of experiments in
a particle collider in high-energy physics!

Problems
1) Solve Eqs.(5) exactly, for initial conditions x0 = 1, v1/2 = 0. Find
out in what sense the solution is superior to that of the corresponding
Euler algorithm
xn+1 = xn + hvn ,
vn+1 = vn − hxn
with initial conditions x0 = 1, v0 = 0.
2) Compare the total of the cross sections at the bottom of Box 1 with
the estimate for the cross section of close encounters (Eq.(21.4)).
3) For fixed masses, how many parameters are needed to specify a three-
body scattering experiment (a) in two dimensions, (b) in one dimen-
sion?
238 22 Laboratory Experiments

Box 1 Example of computational determination of scattering

cross sections.

The command sigma3 -d 128 -v 1.2 -m 0.3 uses part of the “star-
lab” package (http://www.manybody.org/starlab.html) to estimate the
cross section for various outcomes of encounters between a single star
and a binary. The binary component masses are 1 − m and m, and the
mass of the third star may be set with a further option. The other two
parameters are d, which is a measure of the density of experiments in
a central trial √
zone (the fractional error in the largest cross section will
be of order 1/ d), and v, the speed of the third body when at a large
distance, in units of the value at which the total barycentric energy of
the triple system is zero. Here is a summary of the results. It may not
look beautiful to the reader, but it does to anyone who has organised this
kind of thing by hand.
scatter profile:
m1=0.7 m2=0.3 m3=0.5 r1=0 r2=0 r3=0 v inf=1.2
----------------------------------------------------
total scatterings=25, CPU time=7.45
exch1 exch2 ion error stop
2.220 0.888 3.108 0.000 0.000
+-1.83 +-.628 +-2.03 +-.000 +-.000
----------------------------------------------------
total scatterings=60, CPU time=16.94
exch1 exch2 ion error stop
0.444 0.999 0.777 0.000 0.000
+-.222 +-.333 +-.294 +-.000 +-.000
----------------------------------------------------
total scatterings=312, CPU time=105.6
exch1 exch2 ion error stop
0.768 1.224 0.712 0.000 0.000
+-.158 +-.199 +-.153 +-.000 +-.000
----------------------------------------------------
total scatterings=1560, CPU time=487.97
exch1 exch2 ion error stop
1.025 1.181 0.861 0.000 0.000
+-.101 +-.105 +-.089 +-.000 +-.000
v2 σ
The values are those of , where v is defined as above. Notice that
πa2
the cross section for exchanging the heavier binary companion is smaller.
Organisation of the calculations 239
 23
Gravitational Burning
and Transmutation

The arguments of the last few chapters apply to three-body stellar

interactions anywhere. Now we have to take into account the specific
environment in which we are interested, i.e. to study the behaviour of
binaries and triples inside much larger stellar systems. For these purposes
it is very useful to have an idea of the mean rate of change in the energy
of a binary, and also the effect of a single encounter on the motion of the
binary.
Let us consider the mean rate of change first. Relative to the current
energy of the binary, this may be defined by
Z
h∆Ri = ∆R(∆)d∆, (1)

where R is the rate (number per unit time) at which the binary experi-
ences encounters in which the relative change of its binding energy is ∆.
The function R becomes infinite as ∆ → 0, but in practice we can impose
a small positive cutoff on ∆, at the point where positive and negative
energy changes nearly cancel.
As an example, consider the formula
R ∝ ∆−α (1 + ∆)−4.5+α , (2)
which, with α ' 0.54, has been shown to provide a satisfactory fit to
numerical data for ∆ > 0.01 for equal masses in the very hard binary limit
(Heggie & Hut 1993). Then the integral in Eq.(1) is quite insensitive to
the choice of cutoff, and the result is h∆Ri ' 0.21. This result accounts
for the values at the extreme right in Fig.1, which plots h∆Ri, normalised
by the binding energy. Eq.(2) can also be used to show that about half
the hardening rate of a hard binary is accounted for by encounters in
which
∆ > ∆∗ ≡ 0.53. (3)

239
240 23 Gravitational Burning and Transmutation

Fig. 1. The mean rate of change of energy of a binary subject to encounters,

from Heggie & Hut (1993). The initial relative speed of the binary and the single
star have a Maxwellian distribution corresponding to temperature T , and x is the
ratio between the binding energy of the binary and kT , where k is Boltzmann’s
constant. Thus hard binaries are at the right, and soft binaries are at the left.
The reaction rate R in the axis label differs (by scaling) from the quantity R
defined in the text, and the ordinate is actually proportional to the mean rate
of change of energy. This is constant in the hard binary limit, as shown (see
Problem 1). The circles and crosses show the result if “ionising” encounters,
i.e. those leading to disruption of the binary, are excluded. While it may be
useful to consider ionising encounters separately, it leads to a graph with a more
complicated shape than in the case where such encounters are retained (symbol
+).

Keeping this result for later, let us now turn to the effect of an encounter
on the motion of the binary. Let the initial and final velocity of the binary
relative to the intruder be V and V0 , respectively, and let ε be the initial
energy of the binary. By energy conservation (in the rest-frame of the
barycentre of the three bodies, assumed to have equal mass m) it follows
that
1 1
mV 2 + ε = mV 02 + ε(1 + ∆).
3 3
(The factor 1/3 arises because the reduced mass of a system consisting of a
binary of mass 2m and a single star of mass m is 2m/3.) For a sufficiently
hard binary and an average encounter, ∆ε  mV 2 , and so this reduces
approximately to V 02 ' −3∆ε/m. In this same frame the velocities of the
binary and single star are Vb0 = V0 /3 and Vs0 = −2V0 /3, respectively,
since the relative velocity must be V0 and the total momentum in the
 23 Gravitational Burning and Transmutation 241

barycentric frame must vanish. Finally, to compute the new speed of the
binary and single star in the rest frame of the whole cluster we must add
the velocity of the triple system, but again this is negligible for an average
encounter with a sufficiently hard binary. Thus the final p speeds of the
binary
p and single star in the cluster are approximately −∆ε/(3m) and
2 −∆ε/(3m), respectively. (Remember that ε < 0 and ∆ > 0 typically.)
Armed with this information, let us now follow the fate of a three-body
binary as it experiences a succession of encounters, starting from the time
of its formation in the core of a stellar system. Initially we do not expect
the predicted speeds to be significantly in excess of the velocity dispersion
of the core. The calculations of the preceding paragraph are very approx-
imate in this case, but the reaction products continue to move around in
the core, and their kinetic energy (in which, for a binary, we include only
the kinetic energy of the motion of its barycentre) has increased by |∆ε|
per encounter.
As time goes on and further encounters take place, the binary steadily
hardens, and typical values for the predicted speeds of the reaction prod-
ucts increase. The first thing to happen is that the single star can now
escape from the core, and eventually the time comes when the binary
achieves enough speed to do so also (Fig.2; see also Giersz & Spurzem
2000). This is no dream. There is a known binary (with an 8-hour period)
in the outskirts of the globular cluster M15 which probably got there by
ejection from the core in a three-body encounter (Phinney & Sigurdsson
1991). At least one of the components is a neutron star, and the binary
is sufficiently close to display a relativistic perihelion advance, just as the
planet Mercury does, but about 30, 000 times faster.
When the binary is ejected from the core it finds itself in a region
of lower density, and encounters are less frequent. If, as an example, we
consider a binary moving on a radial orbit, the total number of encounters
during one orbit (from the core to the maximum radius which it can reach
and back to the core again) varies as
Z
Nenc ∼ nσvdt, (4)

where n is the number-density of single stars, σ is the cross section for

close encounters, and v is the mean encounter speed. (Note the use of
an “n-sigma-v” argument; cf. Box 13.1.) Now for a hard binary we may
estimate that σ ∼ Gma/v2 (cf. Eq.(21.4)), where a is the semi-major axis
of the binary, and it turns out, for any reasonable cluster model, that the
dominant variation in the integrand in Eq.(4) comes from the density
profile n(r). This causes the rate of encounters to diminish rapidly as
the size of the orbit of the binary increases. Therefore the subsequent
hardening of the binary depends vitally on its orbit shrinking back to the
242 23 Gravitational Burning and Transmutation

core as quickly as possible. In fact mass segregation will accomplish this

in due course (on a time scale no longer than the two-body relaxation
time).
At long last the binary will become so hard that the single star it
interacts with will leave not only the core but the whole cluster. As it
climbs out of the potential well of the cluster it loses speed, emerging with
2
a kinetic energy − ∆ε + mφc , where the first term is an estimate of the
3
kinetic energy of the star after the encounter, and φc is the potential in
the core (assuming φ → ∞ at large radius). Since the encounter itself
contributed an energy −∆ε, the change in energy of the bound system
1
is the difference of these terms, i.e. − ∆ε − mφc . One can think of this
3
as consisting of two contributions; the first is the normal contribution of
the binary, which we still assume to remain bound, and the second is the
escape energy of the single star. The latter is the maximum contribution
which the single star can make, no matter how energetically it escapes.
A fruitful way of looking at this is that the term −mφc is the work
done on the cluster as the single star moves out of the potential well.
This interpretation also suggests that the energy contributed to each

100

10
Radius

0.1

0.01
1e-06 1e-05 0.0001 0.001 0.01 0.1
Energy

Fig. 2. Radius and binding energy of all hard binaries in an N -body simulation
of a star cluster, with N = 8192. The horizontal banding at the bottom results
from the fact that the radius was given to only two decimal places; similarly the
vertical banding at the left. The vertical banding at the right results from a few
persistent hard binaries escaping to large radii and (in most cases) drifting back
to the core. The model was a initially a Plummer model, with equal masses and
no tide.
 23 Gravitational Burning and Transmutation 243

part of the cluster equals the contribution which this part makes to the
central potential. Thus for a shell of material of mass M and radius r,
the heating is GM m/r. By this argument we also see that, if the star
escapes from the core but not from the cluster, it does not heat the core
alone, but instead all the material between its initial and final radii.
Though it is tangential to our present purpose, it is worth stating that
similar considerations apply to other forms of heating through mass loss.
S. Portegies Zwart has pointed out to us that mass loss from stellar evo-
lution is more important than one might think, because the mass is lost
by the most massive stars (which evolve fastest), and mass segregation
concentrates these in the core. It is here that the maximum energy is
generated per unit mass ejected.
Back to the binary. Eventually it is so hard that, not only does it
eject the single stars it interacts with, but it also ejects itself. Thus the
influence of each binary is self-limiting. Though a binary consisting of
two point masses can in principle release an infinite amount of energy,
the process by which it releases energy is its own bottleneck.

Box 1 Energy generation by a single hard binary.

Until the single star is capable of escaping, binary heating is 100%

efficient: the entire change in energy of the binary goes to heating the
cluster. This phase comes to an end when the new energy of the single
star is comparable with the escape energy, i.e. when 2∆ε/3 ' mφc . We
may suppose that this loss becomes important when half of the hardening
of the binary is caused by encounters that eject the single star, i.e. when
ε ' 3mφc /(2∆∗ ) (see Eq.(3)). Since the initial energy of the binary is
comparatively negligible, it follows that the energy contributed in this
first phase is itself about −3mφc /(2∆∗ ).
By a similar argument the binary (treated as a particle of mass 2m)
escapes at the last encounter when its energy is of order ε ' 6mφc /∆∗ .
During the second phase (i.e. up to this point) all energy given to the
binary is held within the cluster; this is one third of the change in in-
ternal energy of the binary, i.e. −3mφc /(2∆∗ ). It takes approximately
log(4)/ log(1 + ∆∗ ) ' log(4)/∆∗ encounters to increase the binding en-
ergy of the binary by the necessary factor of 4, and each contributes an
energy mφc via the single star (see text). Adding all these contributions
together we find that the energy given by a hard binary to the cluster
before escaping is roughly −mφc (3 + log(4))/∆∗.
244 23 Gravitational Burning and Transmutation

How much energy does the binary contribute before retiring? The
result is fairly easily estimated by adding up the contributions from the
various phases in its evolution (Box 1), which evaluates to about −24mφc .
It is usual to express the central potential in terms of the central one-
dimensional velocity dispersion, σc2 ∗ , though strictly there is no upper
bound on the ratio −φc /σc2 . Usually it is assumed that energy generation
by binaries is important in systems with small cores of very high density,
especially in late core collapse. In this case, as we have seen (Chapter 18)
the evolution of the core is self-similar, and in this case the above ratio is
constant: −φc /σ2 ' 13.5 in anisotropic Fokker-Planck models with stars
of equal mass (Takahashi 1995).
Putting our various estimates together, we find that each newly formed
hard binary eventually contributes an energy of approximately 110mσc2 .
Combining with an estimate for the rate at which such pairs form (see
Eq.(21.5)), we readily find that the rate of heating (per unit volume) is
given by

G5 m6 n3
ε ' 85 . (5)
σc7

The surprisingly large coefficient is a warning against reliance on dimen-

sional analysis alone. Nevertheless a considerable number of complexities
are still swept under the carpet if we use such a formula (and indeed it
is necessary to use a formula of this type when modelling the effect of
binaries in the evolution of stellar systems, unless one uses an N -body
method.) Quite apart from the fact that it would be better to express
the result in terms of the local potential rather than the local velocity
dispersion, we are assuming (i) that all the energy yielded by a binary is
injected at the location where it was born; (ii) that the energy is inserted
instantaneously, the moment it is born; and (iii) that binaries form con-
tinuously (in time). All these assumptions are wrong, and something can
be done to patch them up. For instance, it is possible, in the simulation
of cluster evolution, to inject binaries at times determined stochastically,
but in such a way that the mean rate of formation is as predicted. It is
also possible to arrange for the binaries to emit energy in discrete events,
just as in reality. And it is also possible to mimic the effect of each en-
counter on the spatial location of each binary, and, in particular, their
ejection from the core following particularly energetic interactions (Giersz
& Spurzem 2000). We stress again, however, that these difficulties are
entirely circumvented in an N -body simulation.

∗
The subscript should avoid confusion with a cross section
 23 Gravitational Burning and Transmutation 245

One final aspect of this problem which is less well understood and less
easy to model is how to handle the formation and evolution of binaries
in systems with a distribution of stellar masses. One interesting contrast
with the equal-mass case is the behaviour of the cross section for close
encounters (Eq.(21.4)). For stars of equal mass, as a binary hardens its
semi-major axis shrinks and so σ decreases. For stars of unequal mass,
we have already seen that a binary with one component of relatively low
mass is liable to have this component exchanged with a more massive
incomer. Now the energy of the binary need not change much, and an
increase in the mass of one component may cause a to increase. Thus the
close encounter cross section increases, because of both the increase of a
and the increase of the masses. Binary evolution has a runaway tendency
in these circumstances, at least until the components are about as heavy
as they can be.

Problems
1) For a binary on the hard/soft threshold, moving in a cluster in virial
equilibrium, use Eq.(4) and reasonable estimates for n, σ and v to
show that the number of close encounters in a relaxation time is of or-
der 1/ log Λ, i.e. the reciprocal of the Coulomb logarithm. How does
the result depend on the energy of the binary if it is hard? Use your
answer to show that the mean rate of hardening of a hard binary is in-
dependent of the semi-major axis a. Use this result in turn to estimate
the shape of the distribution of binding energies of hard binaries.
2) Assume that a binary is born with small energy, that its energy in-
creases in increments which are distributed as in Eq.(2), that the re-
sulting energy heats a cluster (with or without ejection of the single
star, as appropriate), and that the binary remains in the core until
it escapes. Perform Monte Carlo simulations to compute the mean
energy yield per binary, and compare with the result of Box 1.
3) Problem 18.2 showed that stellar-mass black holes would form a dense
core within the larger core of normal stars. In this core, black hole
binaries form by three-body encounters. Using arguments of Box 1,
estimate the number of single black holes which are ejected in the hard-
ening of each black hole binary, and hence estimate the approximate
fraction of black holes which are ejected as binaries.
Note: the black holes in some of those escaping binaries will later spi-
ral in to each other, under emission of gravitational radiation. These
binaries form a promising category for detection by LIGO, a pair of
laser-based gravitational wave detectors (Portegies Zwart & McMillan
2000).
 PART VII. PRIMORDIAL
BINARIES: N = 4

The following three chapters complicate the million-body problem for

astronomically motivated reasons. Chapter 24 explains these by tracing
the history of the discovery of binary stars in star clusters, in numbers
which imply that they are primordial, i.e. they were born along with
the cluster itself. They are associated with several of the remarkable
phenomena which help to explain why globular star clusters are so im-
portant to astrophysicists, such as the sources of X-rays within them. We
contrast their behaviour in star clusters with the much milder behaviour
of binaries in less extreme environments.
In systems with many binaries, four-body encounters between two bi-
naries are common. Chapter 25 discusses in detail one of the commoner
outcomes: hierarchical triple systems. They are one class of three-body
problem where the motion is both non-trivial and amenable to detailed
calculation. Since these systems are stable and very long-lived, but may
have tiny orbital timescales, such results are important for efficient com-
puter simulation of N -body systems with many primordial binaries.
Chapter 26 discusses the effect of binary-binary encounters on the rest
of the system. In important ways they can dominate the effect of the
three-body encounters discussed in earlier chapters, though not forever,
as binaries are also destroyed in these encounters. The outcomes of the
interactions are also more complicated than in three-body encounters,
and we show how to classify these. We also show that, in astrophysical
applications, the fact that stars have finite diameter has an important
role to play in encounters, when several stars are temporarily confined to
a small volume of space.

246
 24
Binaries in Star Clusters

In the original Star Wars movie there is a brief but memorable scene
of two stars shining down from the afternoon sky. We find it striking
because we are so used to seeing just a single star in the sky, but in fact
it is the sun that is unusual. In the neighbourhood of the sun most stars
are binaries, and some belong to triple systems or even little groups with
still larger numbers of stars (Duquennoy & Mayor 1991). Our nearest
neighbours, for example, are a binary (α Centauri) with a distant third
companion (Proxima Centauri, cf. Matthews & Gilmore 1993). Such
systems are unlikely to arise by chance encounters (Problem 1), and so
the abundance of binaries and triples suggests that most stars are born
that way. Indeed binaries and other multiples are most common among
the youngest stars (cf. Kroupa 1995).

A brief observational history

With this background, astronomers were perplexed by how difficult it
was to find any binaries in globular star clusters. Admittedly wide vi-
sual pairs were not expected, as such binaries would be destroyed by
encounters with other stars in the dense environment of a globular clus-
ter (Problem 21.1). Indeed there is observational evidence that this has
happened (Cote et al. 1996). But it should have been possible to detect
closer binaries by observing periodic variations in their radial velocity,
or else through the discovery of eclipsing variable stars. Nothing was
found by these methods, and it was even being argued that close binaries
must have destroyed themselves. One possibility was that a close binary,
even though it does not tend to be destroyed by encounters with other
stars, may have its eccentricity significantly changed. If the eccentricity
becomes high enough (and arguments of statistical mechanics show that
this is not unlikely), the separation of the stars at pericentre may become
so low that the stars will collide and merge (cf. Chapter 30). Note that,

247
248 24 Binaries in Star Clusters

though this argument turned out not to be needed, it takes the dynamics
of star clusters beyond the point-mass approximation for the first time
in this book.
Firm evidence of binaries eventually arrived, but it came from a most
unexpected direction: the discovery of variable sources of X-rays (cf.
Verbunt & Hut 1987, Grindlay 1988). Such sources were already known
elsewhere in the Galaxy, and those discovered in globular clusters were
interpreted in the same way; i.e. as binaries containing neutron stars.
The standard model is that the other star in the binary is a more-or-
less normal star whose surface layers are just on the point of becoming
gravitationally unbound by the attraction of the neutron star∗ . As this
normal star evolves its material expands slowly, and so there is a small
but steady stream of gas which leaves its surface and, by conservation of
angular momentum, orbits the neutron star in a disk. Viscosity in this
disk then slowly drags the material of the disk towards the neutron star,
and the conversion of the potential energy of the gas, as it sinks in the
gravitational field of the neutron star, into other forms of energy, gives
rise to the X-ray emission that makes these objects observable.
The discovery of these sources was the first clear evidence for the pres-
ence of neutron stars as well as for binaries, and was a major factor in
turning astrophysicists’ attention and interest to the globular clusters.
Though the inferred number of binaries was not yet large by galactic
standards, and though it was not yet clear whether they were born in
the clusters or manufactured there, the numbers of X-ray sources were
somehow far higher in globular clusters than in the rest of the Galaxy.
Suddenly these old and placid members of the Milky Way community
turned out to be fertile breeding grounds for some of the most interesting
and energetic objects in the nearby universe.
In due course a few more binaries were eventually discovered by the
radial velocity technique, though only a few percent of all observed stars
were found to be binaries (cf. Hut et al. 1992 for early developments
in this subject). Perhaps the first observers (Gunn & Griffin 1979) had
just been unlucky. At any rate it was realised that there were strong
selection effects which prevented more than a fraction of the binaries from
exhibiting observable changes in their radial velocity. One reason for this
is that it was only possible to observe the brighter stars. These are giants,
with radii considerably exceeding those of most normal stars. Therefore

∗
This process is referred to as Roche lobe overflow. The donor star is surrounded by
an imaginary surface, outside which its gravitational attraction is overcome by that
of the neutron star. The equipotential surfaces closely resemble those which control
escape from a star cluster on a circular orbit about a galaxy (cf. Fig.12.1, which
can be thought of as depicting the “Roche lobe” of a star cluster).
 A brief observational history 249

very close binaries could not be expected to survive this giant stage of
stellar evolution, for the same reason that highly eccentric binaries would
not be expected. There was thus a rather limited range of separations
(between the two stars in the binary) where radial velocity binaries with
giant companions could be expected: too small, and the stars collide;
too large, and the binary would be destroyed by encounters. Difficult
though it is to allow accurately for such selection effects, by now it was
no longer at all certain that the proportion of binaries in globular clusters
was very much smaller than in the neighbourhood of the sun, except for
the destruction of wide binaries by encounters.
The hurdle having once been cleared, the evidence for binaries in glob-
ular clusters mounted steadily. Pulsars were observed in binary systems
(again by detecting a periodic Doppler shift, this time in the pulse fre-
quency; cf. Phinney 1996). One of these is even a triple system (Thorsett
et al. 1999). X-ray sources of lower luminosity, thought to be binaries
where a white dwarf takes the place of the neutron star in the brighter
sources, were discovered (cf. Livio 1996). In some clusters it may be
possible to detect binaries in colour-magnitude diagrams: unresolved bi-
naries have similar colours to the other stars but are noticeably brighter
(cf. Rubenstein & Bailyn 1997). Eclipsing binaries were discovered at
last (e.g. Fig.1), many as a by-product of searches for gravitational lenses
(e.g. Kaluzny et al. 1998)! And in recent years it has become possible to
find evidence for radial velocity binaries among the normal stars which
have not yet evolved into giants (Côté & Fischer 1996).
In short, there is now abundant evidence that a substantial fraction of
stars in globular clusters are binaries. Satisfactory as this may be to the
astrophysics community as a whole, stellar dynamicists have been a little
less wholehearted in their response. It is still easier to understand and
model clusters containing single stars, while the introduction of binaries
brings in its wake a host of new issues. Simplest, but still irksome, is
the need to decide, when constructing a dynamical model, what is the
distribution of semi-major axes, eccentricities and mass ratios of the bi-
naries. Much more complicated are the issues involving the finite radii
of the stars, the changes in their internal structure caused by their gravi-
tational interaction, and finally a whole host of questions concerning the
internal evolution of stars in binaries.
For the time being we shall continue to treat the stars in a binary as
point masses, and study the dynamical consequences of the existence of
primordial binaries. The other complications we defer to the final part
of the book. Of course it is by no means clear that these complications
really are so superficial that they can be treated as a minor perturbation
250 24 Binaries in Star Clusters

Fig. 1. The light curve of an eclipsing binary star in the globular cluster M71
(after McVean et al. 1997).

of the picture provided by the point-mass approximation. But, as Donald

Lynden-Bell once remarked, “Only a fool tries to solve a complicated
problem when he does not even understand the simplest idealisation.”

Where the Action is

Why should we care, from the point of view of dynamics, that many stars
are binary? After all, in studying the dynamics of a galaxy, the formation
of spiral arms or bars, infall of a companion galaxy, and many other
problems, the duplicity of the stellar populations is ignored. The fact
is that, nearly everywhere in our galaxy, the local dynamics of multiple
star systems and the global dynamics of the galaxy as a whole can be
neatly separated. In the solar neighbourhood, a double star (or triple
or other multiple star system) with a diameter of a few AU will have a
negligible chance to interact with neighbouring stars, even on a time scale
comparable to the current age of the galaxy. A rough estimate suffices
here. With a density n of one star per cubic parsec, a relative velocity v
of 50 km/s, and a target area σ with a radius r of 5 AU, the rate of close
encounters is nσv = πr2 vn = 10−13 /yr. This gives a probability of only
0.1% for such an event to occur during the next Hubble time.
More distant encounters do occur more frequently of course, but a sin-
gle encounter is relatively harmless, and it typically takes many encoun-
ters to unbind a wide binary (Fig.2). As a result, only the widest binaries
in the solar neighbourhood, with separations of order 103 − 104 AU, are
 Where the Action is 251

potentially subject to dissolution by encounters with passing stars and

molecular clouds. However, even in this case the local-global interaction
is a one-way street: the effect of the environment on the binary can be
dramatic, but the feedback effects on the environment are negligible. A
binary with a separation of 103 AU typically has a binding energy that is
three orders of magnitude lower than the kinetic energy associated with
the relative motion of single stars in the solar neighbourhood (and even
less compared with the energy of motion around the Galactic Centre).
Whether or not such a binary is dissolved hardly makes a mark on the
motion of the field stars. Clearly, the dynamical interactions with bina-
ries and multiple star systems form an utterly negligible component in
the energy budget of the galaxy as a whole.
Even in the densest parts of the galaxy, in the inner parsec of the
galactic nucleus, the energy locked up in the internal motions of binaries
is likely to be at least an order of magnitude less than the energy available
in the motions of the single stars and the centre-of-mass motions of the
binaries and higher multiple star systems. With a velocity dispersion of
more than 100 km/s, only the tightest binaries will have an orbital speed
exceeding the typical centre-of-mass motion, and these binaries will only
contain a small fraction of all the stars in the field.
The situation is just the opposite in the case of star clusters. Both open
clusters and globular clusters have a much lower velocity dispersion than
the galactic centre, and a much higher density than the solar neighbour-
hood. The combined effect gives a situation in which a typical binary
can easily have an orbital velocity far exceeding the velocity dispersion
of the cluster, and therefore have an energy  1kT. Combined with the
fact that observations show us that a fair fraction of the stars in clusters

Fig. 2. Evolution of the semimajor axis of a soft binary in the Galaxy. From
Bahcall et al. (1985).
252 24 Binaries in Star Clusters

have been formed as binaries, as we have seen, it is clear that binaries

play an important role in the dynamics of star clusters.
As a result, the total energy locked up in binary binding energy is
at least comparable to, and in some cases may well exceed, the total
energy of the cluster as a whole (in the form of the kinetic and potential
energy of the single stars and of the centres of mass of the binaries).
The binding energy of a single binary may easily dominate that of the
core. Given this situation, changes in binary properties that take place
during the course of normal stellar evolution will have a repercussion
on the dynamical evolution of a star cluster as a whole. While it is
the macroscopic energy budget that drives this dynamical evolution, this
budget can be significantly modified through the strong coupling with
the comparable microscopic energy budget of internal degrees of freedom
of binary stars. In addition, close encounters involving a combination of
single stars and binaries can affect the parameters of the binaries in very
complex ways. Either type of process, internal evolution in relatively
isolated binaries, or three-body and four-body encounters, will modify
the balance between the two energy budgets of a cluster, governing the
external and internal degrees of freedom (bulk energy and total binary
binding energy, respectively).

Problems
1) Estimate the rate of formation of hard binaries in the Galaxy, using
Eq.(21.5).
2) Suppose that the giants in a globular star cluster have mass 0.8M
and radius 0.3AU (cf. Harris & McClure 1983), and that the threshold
between hard and soft binaries occurs when the mean square relative
speed of the components is 10kms−1 . For binaries with one giant
component and one small main sequence component of comparable
mass, and eccentricity e, compare the semi-major axis for contact of
the components at pericentre with the semi-major axis of a binary on
the soft/hard threshold.
3) Binaries are distributed uniformly on the sky with surface density
Σ0 per unit area. Their semi-major axes a are distributed uniformly
in log a. If distance r is measured from the location of a particular
star, the mean surface density of companions, Σ(r), is defined to be
the average value of δN (r)/(2πrδr), where δN (r) is the number of
stars with distances between r and r + δr. Show that Σ(r) ∝ r −2 ,
approximately, for small r, and Σ(r) ' Σ0 for sufficiently large r.
 25
Triple Formation and Evolution

Triple systems are very familiar. The motion of the earth and moon
around the sun is lightly perturbed by the other planets, and if such effects
are neglected it is a nice example of a triple system. Furthermore the
distance between the earth and moon is much smaller than their distance
from the sun, and so it is an example of what is called a hierarchical
triple system. The dynamics of such a system can be understood, to a
satisfactory first approximation, as two Keplerian motions. In the case
of the earth-moon-sun system, one of these is the familiar motion of the
moon relative to the earth, and the other is the motion of the barycentre
of the earth-moon system around the sun. The barycentre lies within
the earth, in fact, and we are more familiar with the picture that the
earth orbits around the sun, but it is more accurate to say that it is
the motion of the barycentre that is approximately Keplerian. This was
realised by Newton (Principia, Book I, Prop LXV), and it was he who
really originated the study of hierarchical triples.
The mass ratios in the earth-moon-sun system are rather extreme.
Even though the sun is so distant, its mass is so great that it exerts
a much greater force on the moon than the earth does. Therefore it is
not immediately clear why the moon continues to orbit the earth, and
why the Keplerian approximations are successful (Problem 1). In the
same way the gravitational pull of the sun on the earth greatly exceeds
that of the moon on the earth, and so it is strange that the tides in the
earth’s oceans generally rise and fall in step with the diurnal motion of
the moon and not that of the sun.
Beyond the solar system there are triple systems in which the masses
are more comparable. In the globular cluster Messier 4, for example,
there is a triple consisting of a neutron star in a binary with what is
presumed to be white dwarf, and a more distant substellar object (a

253
254 25 Triple Formation and Evolution

massive planet) in orbit about the inner binary. The masses are thought
to be about 1.35M , 0.35M and ∼ 0.01M (Thorsett et al. 1999).

Triple Formation
How can such a system arise? Well, maybe they were born like that
(cf. Clarke & Pringle 1991). Otherwise the obvious answer, at first sight,
is capture: a third star approaches a binary, and its perturbing action
slightly unbinds the binary and binds it to the binary. Unfortunately we
have already seen a mathematical argument (Chapter 20) that three-body
capture will not work. This is also readily understood physically (Chapter
19): if the perturbation of the third star is strong enough to bind it, on
successive orbits the energy of the third star will perform a random walk
and eventually become unbound again. But if binaries are sufficiently
abundant in a stellar system, binary-binary interactions are as common
as binary-single interactions. It is then natural to ask whether binary-
binary encounters could lead to the formation of hierarchical triples with
non-zero probability.
Consider first the case in which one of the binaries is far harder than the
other. It is then very unlikely that either component of the wider pair will
make a close approach with the close pair; close, that is, compared with
the smaller semi-major axis. Therefore in most such encounters the close
pair will act like a single object with mass equal to the total mass of its two
components. Since we are assuming that the masses of all the single stars
are comparable, this binary behaves like a relatively massive object. Now
we have already seen, in our study of binary-single interactions (Chapter
19), that an exchange interaction occurs with substantial probability, and
that a massive intruder is very likely to displace a less massive component
of the binary. In the case of binary-binary encounters, then, we may
expect that the close pair will eject one component of the wider pair,
probably the less massive one. The end result is that the tight binary
pairs up with the other component, which acts as a third body orbiting
the close pair in a hierarchical triple.
It turns out, as numerical scattering calculations show, that the prob-
ability of formation of a hierarchical triple system in these circumstances
is substantial. Even for equal masses it occurs in roughly 50% of all en-
counters in which the minimum distance of the two binaries is of order
the sum of their original semi-major axes. And even when the original
semi-major axes are equal, so that neither pair can be thought of as acting
as a single object, the probability is reduced by a factor of only two or so
(Mikkola 1984). Nor is hierarchical triple formation in binary-binary en-
counters of only theoretical interest: it is plausible that the triple system
in the globular cluster M4 formed in this way (Rasio et al. 1995).
 Triple Evolution 255

The formation of triples seems relatively straightforward and common.

It is, therefore, worth investigating what happens to them subsequently.
Some are not quite hierarchical enough to survive for long: the mutual
perturbations of the three stars leads to a destruction of the hierarchy or
the escape of one of the stars. Finding the boundary line between stable
and unstable hierarchies is an important problem. Eggleton & Kiseleva
(1995) have given a rather reliable formula∗ determined empirically from
many numerical integrations. The analytical approach has proved harder,
generally providing criteria that may be rigorous but are wide of the
mark in practice. Recently, however, Mardling (2001) has developed a
new approach (based essentially on Fig.10.3) which gives a sharp criterion
with a sound theoretical foundation.
Now we turn to the hierarchical triples which pass the stability test.
Though there are many kinds of three-body problem, this is the one where
the traditional methods of celestial mechanics are most useful. We shall
see that the results justify the effort.

Triple Evolution
As before we describe a hierarchical triple as a combination of two Kep-
lerian motions: the relative motion of the “inner” binary, and the motion
of the “outer” star relative to the barycentre, B, of the inner pair (Fig.1).
Let the masses of the components of the inner pair be m1 and m2 , and
let r be the position vector of the first relative to the second. Let the
third star have mass m3 and position vector R relative to B. Then the
energy of the triple, in the rest of frame of its centre of mass, is

E = Ei + Eo + V, (1)

where Ei and Eo are the energies of the inner and outer Keplerian mo-
tions, and V is a correction for the fact that the outer star is not actually
in orbit about a star of mass m1 + m2 at B; thus

Gm1 m3 Gm2 m3 G(m1 + m2 )m3

V =− − + ,
|R − µ2 r| |R + µ1 r| R

where µi = mi /(m1 + m2 ) and R = |R|. Now the most systematic way of

handling a problem like this is some technique of canonical perturbation
theory. We, however, shall proceed by more elementary methods which
at least make the results plausible.

∗
except for a misprint: in their Eq.(2), for 2.2 read −2.2
256 25 Triple Formation and Evolution

B r
2

Fig. 1 Notation for hierarchical triple systems

First let us consider the evolution of Ei . Its rate of change is

∂
Ėi = −ṙ. V (r, R). (2)
∂r
In a significantly hierarchical triple with comparable masses, the time
scale on which r varies is much smaller than the time scale for the varia-
tion of R. Therefore, when integrating Eq.(2), we may hold R constant
to good approximation. At lowest order, then, the evolution of Ei is given
by Ei (t) = Ei (0) − V (r(t), R) + V (r(0), R). The right-hand side here is
periodic at lowest order, and so there is no secular (long term) evolution
in Ei and hence in the semi-major axis, ai , of the inner binary.
When it comes to the perturbations in the motion of the third body, we
may average V over the fast motion of the inner binary, as the difference
between V and its averaged value has negligible effect when integrated
over the period of the inner binary∗ . The averaged value depends on the
elements of the inner binary, which vary very slowly, even by comparison
with the outer period, as we shall see. It follows, then, from an argument
similar to that applied to Eq.(2), that Eo also has no long-term variation,
and so neither has the outer semi-major axis, ao .
Now we know that the left-hand side of Eq.(1) is constant and that Ei
and Eo have only periodic perturbations, but do not evolve secularly. It
follows that the orbit-averaged value of V , i.e. averaged over the motion
of both the inner and outer binaries, is constant. The averaging itself is
carried out in Box 1, and it turns out that the result is independent of

∗
Care is needed if the inner and outer binaries are in resonance. In practice this
exception is important only when the ratio of the outer and inner semi-major axes
is small enough that the hierarchy is close to being unstable (cf. Kiseleva et al.
1994)
 Triple Evolution 257

Box 1 Averaging the perturbation potential.

In order to evaluate the average of V , we first approximate it by the

first non-zero term in a multipole expansion, which gives
Gm1 m2 m3  
V '− 3(r.R) 2
− r 2 2
R . (1)
2(m1 + m2 )R5
In order to perform the orbit average over the motion of the outer binary,
we shall choose a coordinate system aligned with the major and minor
axes of the outer orbit. In this frame we write R in components as
R = R(cos θ, sin θ, 0). Then it is easily seen that, except for factors
depending on G, r and the masses, a typical term to be averaged is of the
form Vij = Xi Xj /R5 , where Xi and Xj are two of the components of R.
R −3
−1
For example if i = j = 1 its average is hV11 i = Po R cos2 θdt, where
Po is the outer period. From standard formulae of Keplerian motion we
have q
R2 θ̇ = G(m1 + m2 + m3 )ao (1 − e2o ) (2)
(specific angular momentum) and R = ao (1−e2o )/(1+eo cos θ) (the equa-
tion of an ellipse). It follows that
1
hV11 i = , (3)
2a3o (1 − e2o )3/2
where we have used a standard Keplerian result for the orbital period in
3/2 p
the form Po = 2πao / G(m1 + m2 + m3 ).
Similar calculations show that the only other non-zero average is hV22 i
and that
hV22 i = hV11 i. (4)
This demonstrates that the averaged perturbation is independent of the
orientation of the orbit of the third body within its plane of motion.

the orientation of the orbit of the third body within its plane. This is
remarkable, because it implies that the magnitude of the orbit-averaged
angular momentum of the outer binary, co , is constant. The argument
leading to this conclusion is analogous to the familiar fact that the angular
momentum of a system is conserved if the Lagrangian of the system
is invariant under rotations. It may also be viewed as an application
of Noether’s Theorem (Box 7.1), or as resulting from the fact that the
angular momentum of the outer orbit is conjugate to the orientation of
the orbit, in the sense of Hamiltonian dynamics.
258 25 Triple Formation and Evolution

Now we are in a position to exploit the fact that the total angular
momentum of the whole triple system, c, is constant, for c = co + ci , and
so
c2 = c2o + 2co ci cos i + c2i (3)
is constant, where i is the relative inclination of the inner and outer orbital
planes. Now, as we have seen, co is constant on average, and so the sum
of the last two terms on the right of Eq.(3) is also constant on average.
However, we see from Eq.(2) in Box 1, and the corresponding expression
for ci , that the last term on the right of Eq.(3) is negligible, since ai  ao
in a hierarchical triple, and so the average value of co ci cos i is constant.
Finally, using the definition of ci in terms of the elements of the inner
q as the fact that ai and co are constant on average, we see
orbit, as well
that c⊥ = 1 − e2i cos i is constant on average. The notation reminds us
that c⊥ is proportional to the component of the inner angular momentum
perpendicular to the outer orbit, but we have omitted a factor involving
several quantities which are either constant or constant on average.
This is our first interesting non-trivial result. To investigate its con-
sequences, let us denote initial values by (0). Then, neglecting periodic
variations, what we have found is that
q q
1 − e2i cos i = 1 − e2i (0) cos i(0). (4)
Suppose, for example, that the inner and outer orbital planes are nearly
orthogonal, so that the right side is nearly zero. It follows that, as i and ei
evolve, ei may reach values close to ei = 1, (depending on the maximum
value of cos i). Then the distance of closest approach of the components
of the inner binary may be much smaller than their semi-major axis,
and may even be smaller than the sum of their physical radii. (We are
remembering now that these are not point masses, but stars.) This has
obvious astrophysical consequences, but of course it does not follow, just
because high eccentricities are possible, that they are inevitable. We still
have some work to do.
Before we go on it is interesting to mention another context in which
similar phenomena arise. It concerns a fascinating group of comets named
the Kreutz group, after the 19th century astronomer who first drew at-
tention to them. They are more usually called the “sungrazers”, because
they approach the sun’s surface to within a distance of about a solar ra-
dius. It has been proposed (Bailey et al. 1992) that they originated in
orbits of very high inclination and much larger (i.e. more normal) peri-
helion distance, and that the perturbations by the planets have evolved
their orbits so that they now have relatively low inclination and very high
eccentricity.
 Triple Evolution 259

We return to the averaging of V . In Box 1 we substitute Eqs.(3) and

(4) into Eq.(1), and soon find that V , averaged over the outer motion,
takes the form
Gm1 m2 m3
hV i = − (r2 − 3z 2 ),
4(m1 + m2 )a3o (1 − e2o )3/2
where z is the component of r orthogonal to the outer orbital plane. Now
it is straightforward, but a shade tedious, to perform the orbit average
over the inner motion. As variable of integration it is best to use the
eccentric anomaly E ∗ , and all that is needed are some formulae of Kep-
lerian motion (Box 7.2), and a little three-dimensional geometry (Fig.2).
We simply state the result, which is that the doubly orbit-averaged ex-
pression for V is
Gm1 m2 m3 a2i 
hhV ii = − 2 + 3e2i
8(m1 + m2 )a3o (1 − e2o )3/2 (5)

−3 sin i
2
(5e2i sin2 ω +1 − e2i ) ,

where ω (the ‘argument of pericentre’) is the angle between the major axis
of the inner orbit and the line in which the two orbital planes intersect
(the ‘line of nodes’).
Now we can visualise the secular evolution of the triple system. In
Eq.(5) there are really only three variables: ei , i and ω. Also, hhV ii is
a constant, which can be evaluated from the initial conditions, and so
Eq.(5) states that the evolution takes place on a certain surface in the
space defined by these three variables (Fig.3). Eq.(4) defines another
surface in the same space, and so the evolution must take place on their
intersection, which is a curve. It can be seen from the intersections that,
though the evolution takes two rather different forms, in both forms the
evolution of the eccentricity and the inclination are oscillatory. What

ω
i

Fig. 2 Geometry of a hierarchical triple system.

∗
Not to be confused with energy!
260 25 Triple Formation and Evolution

distinguishes the two forms is the evolution of the pericentre ω, which

may either oscillate or “circulate” (Fig.3), rather as a simple pendulum
may either oscillate or swing through 360◦ . In either event the periodic
oscillations in ei and i are often called Kozai cycles (Kozai 1962).
We also notice from Fig.3 that the maximum value reached by the ec-
centricity is less, in general, than would be predicted naïvely from Eq.(4),
because the values are also constrained by Eq.(5). In our previous dis-
cussion it was assumed that the maximum value occurs when i = 0◦ or
180◦ , but in fact the values of i are constrained within narrower limits,
in general, by Eq.(5). Nevertheless, it is still true that large eccentricities
may occur if the initial inclination is high, i.e. near 90◦ (Fig.4).
It is worth making an observation which is, at first sight, very surpris-
ing, even to experts, and that is that these variations in i, ei and ω do not
depend on the masses or the dimensions of the orbits. One might have
thought that if the third body were sufficiently far away or of sufficiently
low mass its influence would remain very small. In fact, these affect only
the time scale on which the evolution takes place, and not the size of its
effects.
We have not quite finished, for Eq.(5) is the potential which the inner
orbit experiences, on average. It depends on two angles, and so the inner
orbit experiences a torque. Variations of ω correspond to rotation about
an axis normal to the inner orbital plane, and so the corresponding torque

6
5
6
4

4 ω3
ω 2
2 1

0 1
1 0.5
3 e
2 i 0.5 0 0
2.5 2 1.5 1 0.5
1 e 3
i
0 0

Fig. 3. Evolution of a hierarchical triple system according to the method of

averaging. The evolution takes place along a curve of intersection of the two
surfaces, which are given by Eqs.(4), solid, and (5), wire frame. The argument
of pericentre, ω, may travel through 360◦ (left) or oscillate (right).
 Triple Evolution 261

100
90

Inclination
80
70
60
50
40
30
0 5000 10000
Time

1
0.9
0.8
Eccentricity

0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 5000 10000
Time
Fig. 4. Evolution of the inclination (degrees) and inner eccentricity in a hier-
archical triple system in which the initial inclination was 90◦ . Both inner and
outer orbits were initially circular.

causes variations in the magnitude of the angular momentum. It is clear

that this torque vanishes if ei = 0, as is easily verified by computation
of ∂hhV ii/∂ω. Variations of i correspond to rotation about the “line
of nodes”, which is the imaginary line in which the two orbital planes
intersect. It is easy to see that this torque causes the angular momentum
of the inner orbit, which is orthogonal to the inner orbital plane, to precess
about the normal to the outer orbital plane (at least, at the present level
of approximation).
Though we have obtained this result by simplified perturbation theory,
Newton gave a very simple qualitative explanation of this precession∗. In
Fig.2, suppose that m3 is at the extreme left-hand side of its orbit. After
m1 moves up through the plane of motion of m3 , it is attracted to m3
less than m2 is. Hence m1 descends though the plane of m3 at a point
displaced (in a sense opposite to the motion of m1 ) from the original line

∗
Principia, Book I, Prop. LXVI, Cor.XI. See also Airy 1884, Moulton 1914.
262 25 Triple Formation and Evolution

Box 2 Precession of an inner circular orbit in a hierarchical triple system.

When ei = 0, the torque is simply

∂hhV ii 3Gm1 m2 m3 a2i
= sin i cos i,
∂i 4(m1 + m2 )a3o (1 − e2o )3/2
acting parallel to the outer orbital plane. Now the component of the
inner orbital angular momentum parallel to this plane is
m1 m2 q
ck = G(m1 + m2 )ai sin i
m1 + m2
if ei is negligible. The angular velocity of precession, Ω, is obtained by
dividing this into the torque, and so
s  3/2
3m3 G ai
Ω= .
4a3o m1 + m2 1 − e2o
This is rather unmemorable, and it is better to express the result in terms
of the periods (see text, Eq.(6)).

of nodes. The same thing happens when m1 passes below the orbital
plane of m3 , and when m3 itself is in other positions.
We can easily estimate the angular velocity of precession, although in
Box 2 we do so only in the case in which ei ' 0, by way of illustration.
The result is that the period of precession is

2π 4 m1 + m2 + m3 Po2
= (1 − e2o )3/2 hcos ii, (6)
Ω 3 m3 Pi
where Pi and Po are the inner and outer periods, respectively, and the
average of cos i is taken over one Kozai cycle. This result implies that,
for given masses, eo and mean inclination, the periods are in geometric
progression: the outer period is of order the geometric mean of the inner
and precession periods. This fact, which was known to Newton∗, is a
general rule of thumb for the time scale of long-period perturbations in
hierarchical triples, and for some observed triple systems it leads to a
time scale short enough that the effects are detectable within a few years
(Jha et al. 2000). It also explains the sense in which the perturbations by
a distant or low-mass companion are weak: the evolutionary effects are

∗
Principia, Book III, Prop.XXIII
 Triple Evolution 263

just as large as for a close or massive companion, but take much longer
to develop.
Anyone who has taken more than a casual interest in eclipses of the
sun and moon will be aware of our last result on time scales. Eclipses
occur at two times of the year, called “eclipse seasons”, when the sun lies
close to the line of nodes, i.e. the line of intersection of the orbital planes
of the moon (about the earth) and the earth (about the sun.) Because of
the perturbation of the sun on the earth-moon system, the orbital plane
of the moon precesses, the line of nodes rotates in the plane of motion of
the earth, and so the time of year at which eclipse seasons occur slowly
changes. The period of the precession in this case is 18.6 yr. Since the
ratio of the month to the year is about 13.4, Eq.(6) predicts 17.8 yr.
It is appropriate to end where we began, with the most familiar of
all hierarchical triple systems. It was here, moreover, that the peculiar
property of high-inclination orbits found one of its earliest applications
(Lidov 1963). It is little wonder that the moon is not on a high-inclination
orbit!

Problems
1) Explain why the moon continues to orbit the earth even though the
force exerted by the sun on the moon is much larger than that exerted
by the earth on the moon. Explain why the lunar tide is higher than
the solar tide, even though the gravitational attraction of the moon
on the earth is much weaker than that of the sun.
2) Use Eqs.(4) and (5) to compute the maximum eccentricity of the inner
orbit if it is initially nearly circular.
3) Complete the derivation of Eq.(5).
 26
A Non-Renewable Energy Source

As we have seen, primordial binaries exist in some abundance in globular

clusters, and so there is good reason to study the behaviour of a million-
body problem in which there are many binaries. How do such clusters
evolve?
To some extent the binaries may be treated as heavy point masses,
provided we are interested in interactions with other stars at distances
much greater than the typical semi-major axis of a binary. Thus the
binaries behave like a heavy species, and the process of mass segregation
(Chapter 16) ensures that they become heavily concentrated towards the
core of the cluster. It follows that, even if we start with a cluster in
which only ten percent of the stars are binaries, we quickly find ourselves
looking at part of the system (the core) where the abundance of binaries
is more like 90%. Incidentally it has even been argued that young star
clusters may consist entirely of binaries. For one reason or another, then,
the study of stellar systems in which essentially all stars are binaries is
an important one.
Something was already said about the interaction between two binaries
in the previous chapter. There we quickly focused on the formation of
hierarchical triple systems in such encounters. In the present chapter we
focus on the energetics of these interactions. We shall see in the next
chapter that this is the main way in which these interactions feed back
into the overall dynamics of a stellar system.
It is very instructive to compare the energetics of binary-binary inter-
actions with that of three-body interactions (Chapter 23). Conservation
of energy immediately tells us that the formation of binaries in three-
body interactions is necessarily exothermic, because a bound system is
formed from three unbound particles. Primordial binaries, on the other
hand, are initially endowed with negative energy, and it is only in their
subsequent interactions that this can influence the evolution of the N -

264
 26 A Non-Renewable Energy Source 265

body system. Again, once a hard three-body binary has been formed it
is very difficult to destroy: it keeps hardening and heating the stellar sys-
tem until it is ejected. By contrast we have seen in the previous chapter
that binary-binary encounters can easily destroy one of the participants,
perhaps forming a hierarchical triple system, and this suggests that the
useful energy inside a population of primordial binaries is limited. On
the other hand, it is clear that the cross section for an encounter between
one binary and another exceeds that for interaction between the same
binary and a single star: gravitational focusing is somewhat stronger,
because the participating masses are larger, and the physical size of the
participants is somewhat greater. When we weigh these considerations
together, it is not very clear whether or not binary-binary interactions
will be as effective a source of heat as three-body interactions.
Let us look at these interactions in a little more detail. We saw in
Chapter 22 that, even for three-body interactions, the number of param-
eters required to specify an interesting set of encounters is already rather
substantial, and this problem is much worse for binary-binary encounters.
Therefore we shall make matters as simple as possible by considering stars
of equal mass. It might be thought also simplest to consider binaries with
equal initial energy or semi-major axis, but for our present purposes it is
helpful to return to the rather extreme form of binary-binary encounter
considered in the previous chapter, in which one binary is much harder
than the other.
Broadly speaking, the outcome of a binary-binary encounter may be
classified in the same way as three-body interactions: fly-by, resonance
and exchange (Box 1). (With our choice of parameters the closest analogy
in three-body scattering is encounters between a binary and a single star
in which the mass ratios are 1 : 1 : 2, the last of these being the harder
binary.) Indeed numerical scattering experiments show that the three
outcomes occur in roughly comparable amounts, for impact parameters
in the range in which exchange and resonance encounters can occur.
A fly-by is the easiest to understand by three-body analogies. The
less hard binary is hardened (by an amount of order its initial energy, or
h∆iε on average, cf. Chapter 23) and both reaction products retire from
the encounter at increased speed. Thus the encounter is exothermic (on
average), and both binaries survive. The result is the same if the harder
binary is captured in a “hierarchical resonance”, which will eventually
decay again.
Prompt exchanges are also easy to understand. The massive incomer
displaces one component of the initially wider binary, without necessarily
altering its semi-major axis much, and increases the binding energy of
266 26 A Non-Renewable Energy Source

Box 1. Interactions between two binaries.

Fig.1 shows the various outcomes of binary-binary scattering when one

binary is so hard that its energy is unaffected by the encounter. In effect,
this is three-body scattering, and the state referred to as “Democratic
quadruple” is not truly democratic.

Hierarchical
Resonance
Resonant
Decay Flyby

Binary-binary Flyby ‘Democratic’

quadruple
Hierarchical
Resonance
Capture

Pre encounter Resonant

Capture

Resonant
Exchange

Ionisation
Prompt
Exchange

Hierarchical Triple

Fig. 1. Some of the reaction routes for interactions between one binary and
another much harder one. Initial state: double box; intermediate state: dashed
box; final state: solid box.

If we allow the harder binary to be slightly affected by the encounter,

but not to be disrupted, the number of possibilities increases. Now it is
possible for a binary-binary configuration to evolve into a “democratic
quadruple”, which can now be truly democratic. Another possible tem-
porary outcome of the encounter is a four-body hierarchy, consisting of
an inner hierarchical triple surrounded by a distant fourth body.
 26 A Non-Renewable Energy Source 267

Box 1 (continued)
In general (when there is no restriction on the hardness of either binary)
the outcomes may be classified as follows. Let x, x denote the components
of the first binary, and y, y those of the second binary. We use notation in
which stars which are bound into a binary are enclosed in square brackets,
whereas braces (i.e. {}) are used to group objects which are not bound to
each other. Thus the initial configuration of a binary-binary encounter is
represented by {[x, x], [y, y]}. With this notation, Table 1 lists the various
possible final outcomes. For those with an asterisk, a new configuration is
obtained by interchanging x and y. Some listed outcomes would be very
rare, e.g. those involving a bound triple system such as [x, x, y] (Fig.2).
Table 1
Stable outcomes of a binary-binary encounter

{x, x, y, y} Complete ionisation

∗
{[x, x], y, y} Ionisation of one pair
{[x, y], x, y} Ionisation and exchange
{[x, x], [y, y]} Flyby
{[x, y], [x, y]} Complete exchange
∗
{[[x, x], y], y} Hierarchical triple with escape
∗
{[[x, y], x], y} Hierarchical triple with exchange and escape
∗
[[[x, x], y], y] Hierarchical quadruple
∗
[[[x, y], x], y] Hierarchical quadruple with exchange
∗
{[x, x, y], y} Bound triple with escape
∗
[[x, x, y], y] Hierarchy with bound triple
The list can be extended to include temporary intermediate states (as
in Fig.1) by replacing braces with square brackets. Thus the first item
becomes [x, x, y, y] (democratic quadruple). The number of outcomes is
also greater if we wish to distinguish the two components of each binary, in
which case we must introduce different letters for each component. (Table
1 only distinguishes the two binaries.) With some effort, the notation can
even be extended to provide a systematic classification of the outcome of
encounters between arbitrary systems.

the binary by roughly a factor of two. This reaction is also exothermic,

268 26 A Non-Renewable Energy Source

Box 1 (continued)

Fig. 2. A rare outcome of binary-binary scattering; formation of a stable triple

system with one escaper (from Heggie 2001).

therefore. The difference between this kind of encounter and a flyby is

that the reaction products are now a single star and a hierarchical triple
system. Though this may be perectly stable in isolation, it now presents
a large cross section to interactions with other single stars. The effects of
these will tend to harden the motion of the outer star in the triple, and
eventually destroy the temporary stability of the hierarchy. The outcome
is the same if the harder incoming binary is insufficiently hard, for then
the triple is not hierarchical enough to be stable. (These refinements are
not shown in Fig.1 in Box 1, which restricts itself entirely to an isolated
binary-binary interaction in which one binary is arbitrarily hard.)
Finally we turn to democratic resonant encounters, i.e. the formation of
a temporary bound quadruple system. Eventually, just like a temporarily
bound triple, this is expected to resolve with the ejection of one star
(in which we continue to think of the harder binary as a single massive
object). The ejected star may be the binary, in which case the outcome is
just like a flyby. In fact, however, the escaper is likely to be one of the stars
of lower mass, i.e. one of the components of the initially wider binary. In
 26 A Non-Renewable Energy Source 269

that case the result is a hierarchical triple and a single escaper. But we
must pause before concluding that the result is like that of an exchange
encounter. During the evolution of the triple system, the three stars (one
of which is a close binary) have many opportunities for a close encounter.
Indeed numerical studies show that it is quite likely that, at some stage
before the evolution of a triple system is complete, the smallest separation
of some pair of stars is no more than about 1% of the initial separation
of the binary (Hut & Inagaki 1985), which here means the wider binary.
Now we did assume that the other binary was much harder than the wider
binary, but we now see that even a factor of 100 in hardness may not be
enough to ensure that the binary-binary interaction behaves like a triple
interaction, if a democratic resonance is formed. If indeed there is a close
interaction between the hard binary and one component of the wider pair,
this is likely to lead to the ejection of that component, and possibly also
of the binary. If the latter happens, the result of the encounter is a single
binary and two single stars; the energy released will be comparable with
that of the initially harder binary. If, however, the binary remains bound
to the other component of the originally wider binary, then the resulting
triple system may be either hierarchical or unstable.
This has been a somewhat elaborate discussion, which, incidentally,
illustrates that the problem of classifying these events automatically is
non-trivial. The bottom line, however, is that binary-binary interactions
release energy comparable with that of the less hard component (and
sometimes that of the harder component), but that the probability of
destroying one of the binaries is high, certainly above 50%. It follows
that the maximum energy that may be extracted from a binary, up to
the point of its destruction, is of order its binding energy. This is very
different from the behaviour of three-body binaries, whose energy yield
is determined by the central potential (Chapter 23). In other respects,
however, much the same considerations apply as with three-body binaries.
In particular, reaction products of binary-binary interactions are often
found well outside the core, from which they drizzle back into the core
by dynamical friction (Hut et al. 1992).
With these conclusions in mind let us return to the problem of a stellar
system containing a modest proportion of binaries initially, say f by
number∗ . The simplest way of estimating their effect is by comparing the
energy that they may release with the energy of the whole cluster. The
latter consists of two contributions: the internal energy of the binaries,
Eint , and all the rest, which by contrast we refer to as “external” energy,

∗
This means that, if the cluster contains N objects, where an “object” may be either
a binary or a single star, there are f N binaries
270 26 A Non-Renewable Energy Source

Eext (cf. Chapter 19). Approximately, Eext is the energy we would

compute by thinking of each binary as a single object, i.e by replacing it by
its total mass located at its barycentre. It is this energy which determines
the virial radius of the cluster. Now a binary which is on the border
between being soft and hard has an internal energy of approximately
Eext /N , where N is the total number of objects in the system (Chapter
19). Denoting by x the “hardness” of each binary, i.e. its energy in units
of this threshold energy, we see that Eint ' f xEext , since the number of
binaries is f N . Now we have already seen that the energy which these
binaries may yield is of order Eint itself, and therefore this may have a
significant effect on the cluster as a whole only if f x >∼ 1.
From this calculation we see that, if f  1/x, the primordial binaries
can have only a modest influence on the overall evolution of the cluster.
They may well burn up rapidly and yield all the energy they can, but
they are a non-renewable energy source, and if more energy is needed to
sustain the evolution of the system (Chapters 27, 29) it must come from
elsewhere.
From what has been said, it seems that primordial binaries may have a
dominant influence on the energetics of a stellar system if, either, they are
of modest hardness and substantial abundance (e.g. x ∼ 5 and f ∼ 20%),
or, there are a few very energetic binaries. In the latter case, however, it
is not enough simply to consider the energetics of the situation; the time
scale on which the energy is released is also relevant. Roughly speaking,
the time scale for close interactions of a binary of hardness x varies as
xtr , where tr is the relaxation time (Problem 23.1). Therefore it is clear
that, for extremely hard binaries, the time scale greatly exceeds that on
which other dynamical processes, and the evolution of the system as a
whole, are taking place. Indeed there are some non-trivial dimensionless
factors in the expression for the time scale of binary hardening, and even
binaries with x > ∼ 100 are relatively inefficient at powering the evolution
of a stellar system. The most important binaries are those with a hardness
of order 10 (plus or minus about half a dex), and it is clear that their
abundance must be not much less than about 10% by number if they are
to provide enough energy to feed the evolution of a stellar system. It is
rather curious that the observations of binaries in globular clusters imply
a population of roughly this size in the appropriate range of semi-major
axes.
All of our discussion has been based on the point mass approximation.
Close binary-binary interactions with stars of finite radius can be studied
with suitable simulations, and reveal a variety of important differences
(Goodman & Hernquist 1991).
 26 A Non-Renewable Energy Source 271

Problems
1) Following the guidance in Box 1, construct a list of temporary (in-
termediate) outcomes of a binary-binary encounter. Hence try to con-
struct pathways, as in Box 1 (Fig.1), leading to each permanent out-
come on the list in Table 1. (For some practical examples, see Aarseth
& Mardling 2001.)
2) Estimate the cross section for close interactions between two hard
binaries. If one is much harder than the other, how is the result
altered if one seeks the cross section for significant perturbation of
both binaries?
3) Obtain the starlab package
(http://www.sns.ias.edu/∼starlab/starlab.html and Box 22.1). The
command
scatter -i "-M 1 -r 0 -v 1 -t -a 1 -e 0 -p -a 1 -e 0"\
-t 100 -D 0.01|xstarplot
will generate a random head-on, equal-mass, binary-binary encounter.
By viewing the plot and inspecting the analysis of the final configura-
tion, try to follow the interaction via Box 1, Fig.1, or its extension to
binaries with comparable energies (Problem 1).
 PART VIII. POST-COLLAPSE
EVOLUTION: N = 106

The following three chapters complete the story of the evolution of a

million-body system, in its purely stellar dynamical form. Chapter 27
begins by estimating the rate at which the formation and evolution of
(non-primordial) binaries effectively generates energy within the system.
The first application is to show that this is sufficient to halt core collapse.
Then we consider other ways of generating the energy: binaries formed
in dissipative two-body encounters between single stars, and primordial
binaries; we quantify the extent to which the effectiveness of primordial
binaries depends on their abundance and their energy.
Chapter 28 we consider how a balance can be struck between the cre-
ation of energy (by binary interactions) deep in the core and the large-
scale structure of the rest of the cluster. We first describe a standard
argument which implies that conditions in the core, where the energy is
generated, are governed by the overall structure. We outline the core
parameters and overall evolution which this argument implies. Next we
give arguments to show that this balance can be unstable, and describe
the phenomenon of temperature inversion which is associated with the
generation of gravothermal oscillations. The manner in which they de-
pend on N (in idealised models) is an example of a Feigenbaum sequence
of “period-doubling bifurcations” in this context.
Chapter 29 rounds off the evolution of a million-body system by fo-
cusing on the evolution of gross structural parameters: total mass, and
a measure of the overall radius. The result depends on the boundary
conditions, and is discussed separately for an isolated system and one
limited by a tidal field.

272
 27
Surviving Core Collapse

The last 8 chapters, dealing as they have done with interactions between
only three or four stars, might seem a long digression away from the
subject implied by the title of this book. Yet we shall see, as we take up
the thread of the million-body problem where we broke off at the end of
Chapter 18, that an understanding of the behaviour of few-body systems
is crucial in following the evolution of the system through core collapse
and beyond.
We left the system rushing towards core collapse, its central density
rising inexorably, so that it would reach infinite values in finite time.
How is this catastrophe averted? In fact there is no shortage of choices,
for at least five different mechanisms have been proposed over the years.
Admittedly, two are rather out of favour at present: a central black hole
(e.g. Marchant & Shapiro 1980), or runaway coalescence and evolution of
massive stars (Lee 1987a, and Problem 1). The other three involve binary
stars in one guise or another, and it is not hard to see why this is attrac-
tive. After all, the mechanism responsible for core collapse is a two-body
one (Chapter 14). Therefore it is clear that higher-order interactions,
which we have neglected so far, might in principle eventually compete
with two-body relaxation when the density becomes high enough. And
three-body interactions can create binaries (Chapter 21 and Fig.1).
While the creation of binaries in three-body interactions at least brings
binaries into the picture, it is not the principal mechanism by which
core collapse is halted. Nor is this the only way in which binaries may
enter the scene. In this chapter we examine three possibilities: besides
“three-body binaries”, which is the name given to binaries formed in
the way we have just mentioned, we consider “tidal capture binaries”,
formed in dissipative encounters between two single stars (Chapter 31)
and “primordial binaries”, formed along with the single stars at the birth
of the stellar system (Chapter 24).

273
274 27 Surviving Core Collapse

From the astrophysical point of view the second of these is an “also-

ran”. In fact, since the discovery of primordial binaries, the third option
is the hot favourite. But there is plenty of room for a side bet on three-
body binaries, and the reason for this is that there are some observed
clusters whose structure may be hard to understand if we rely on pri-
mordial binaries to halt core collapse. As we shall see, in the presence of
primordial binaries, the core does not shrink that much before the col-
lapse is halted. But there is a sizable fraction of the globular clusters in
our galaxy whose cores are unobservably small (Trager et al. 1995). The
classic example is M15 (Fig.1.2; Guhathakurta et al. 1996; for a more
extended history of ideas on this cluster, see Meylan & Heggie 1997). At
one time the dense core in this object was explained by supposing that
it surrounded a black hole (Newell et al. 1976). Then it was explained
by a static model containing a dense core of neutron stars (Illingworth &
King 1977). More recently still it has been modelled successfully by an
evolving Fokker-Planck simulation (e.g. Dull et al. 1997), in which the
core is dominated by neutron stars, and the post-collapse evolution of
the cluster is powered by three-body binaries (Lee 1987). The number of
neutron stars lies well within observationally based limits (Wijers & van
Paradijs 1991). None of these models have a need for primordial binaries,
and it may be that primordial binaries are somehow ineffective in clusters
like M15.

Three-Body Binaries

The formation of three-body binaries is an exothermic reaction, in the

sense described in Chapter 23. Two single stars are replaced by a binary,
a third single star acting as a catalyst (Fig.1). Conservation of energy in
this reaction leads to the relation
1 1 1 1 1 Gm1 m2
+ m3 v30 −
2
m1 v12 + m2 v22 + m3 v32 = M12 v12
2
2 2 2 2 2 2a
where the subscript 12 refers to the barycentre of the binary (i.e. M12 =
m1 +m2 ), a is its semi-major axis, and v30 is the new speed of the catalyst.
The negative energy locked up in the relative motion of the components
of the binary (the last term) is compensated by increased kinetic energy
of the catalyst and the barycentre of the binary.
Formation of the binary is only a small part of the story. We have
already seen (Chapters 19, 23) that a hard binary, once formed, continues
to harden, and this further increases the kinetic energy of the single stars
and the barycentre of the binary. Intuitively it seems plausible that the
 Three-Body Binaries 275

3
2
2 1
1
0
-1
-2
-3
-4
-5
-6
-7 3
-8
-14 -12 -10 -8 -6 -4 -2 0

Fig. 1. Formation of a binary in a triple encounter. The stars are labelled at the
start of the motion. After a close encounter between 2 and 3, particle 1 absorbs
enough energy to bind them.

resulting enhancement of the kinetic energy of the particles in the core

should tend to resist continued collapse.
The rate at which energy is generated by binaries is not at all the same
thing as the rate of hardening. Though the mean rate of hardening of
an individual binary is nearly independent of its energy, provided that
it is hard (Chapter 23), a given binary is an effective source of energy
only for a limited time. The increasing kinetic energy imparted to its
barycentre gradually expels it to larger and larger distances from the core
(Fig.23.2), and then its further interactions become less and less frequent.
In fact each binary yields an energy of order a multiple of the individual
stellar kinetic energy in the core before it is effectively switched off by
its increasingly far-ranging orbit (if it is not expelled from the cluster
altogether).
The rate at which energy is created by the formation and evolution of
“three-body binaries” was estimated in Eq.(23.5). It is important to note,
however, that, although binaries are predominantly formed in the core,
relatively little of the energy they release is deposited there. As soon as
the energy of a binary significantly exceeds the mean kinetic energy of
single stars in the core, the products of energetic three-body interactions
are expelled from the core, and deposit most of their energy elsewhere.
In fact the energy added to the core (per binary) may be only of order
a few times the mean kinetic energy of one core member, i.e. of order a
few times mσc2 (cf. Chapter 23 and Spitzer 1987, Sect.7.1), where σc is
276 27 Surviving Core Collapse

the rms one-dimensional velocity of stars in the core. Therefore we shall

simply neglect numerical factors, and, denoting by rc the radius of the
core (Eq.(8.9)), we see that
Ė ∼ G5 m6 n3 rc3 σc−7 . (1)
In order to establish that the role of binaries can be important we
must compare this with the energetics of core collapse. One approach
is to think of the star cluster as a conducting stellar gas (Eq.(1) in Box
9.3∗ ), and apply simple thermodynamics to the core. In the first law
(“T dS = dQ”), we recall that the velocity dispersion is proportional
to the temperature, T , while the change in entropy, dS, is the change
in specific entropy times the mass. The change in thermal energy, dQ,
results from conduction (denoted by a conductive luminosity in the gas
model) and from energy generation (caused by formation and evolution
of binaries). Applying the first law to the matter inside radius rc we see
that
Z rc
D σ3
4π mnσ2 ln r2 dr = Ė − L, (2)
0 Dt n
where the logarithm is essentially the specific entropy of the gas, the sym-
bol D/Dt denotes a Lagrangian time-derivative, and L is the conductive
flux of energy across the Lagrangian shell of instantaneous radius rc .
Until energy generation by binaries becomes important, the balance of
this equation is struck between the left side and the flux L. Estimating
the Lagrangian derivative factor by the reciprocal of the current time
scale of core collapse, tcc , which shrinks as the core collapses, we deduce
the estimate
L ∼ mnc rc3 σc2 /tcc , (3)
where nc is the number density of stars in the core. (This estimate may
be equally well obtained by equating L to the rate of change of kinetic
energy of the core.) During core collapse tcc is related to the two-body
relaxation time in the core, and in late collapse the relation is close to
proportionality (Chapter 18). Thus we estimate
L ∼ G2 m2 nc ln Λrc3 /σc , (4)
where Λ is the argument of the “Coulomb logarithm” (Chapter 14).
Now it is clear that generation of energy by binary formation and
hardening becomes important if the two terms on the right of Eq.(2)
become comparable, i.e. if L ∼ Ė. It follows from Eqs.(1) and (4)

∗
A term 4πr2 ε would be added to the right hand side, where ε is the rate of generation
of energy per unit volume, much as in Eq.(1), Box 9.2.
 Two-Body Binaries 277

that this leads to a relation of the form (Gm)3 nc ∼ σc6 ln Λ. To make

this more meaningful, we recall (Eq.(8.9)) that the core radius is often
defined by 8πGmnc rc2 = 9σc2 . It follows, √then, that the energy supplied
by binaries is important when nc rc3 ∼ 1/ ln Λ. The dependence on the
Coulomb logarithm here is unimportant. Also, in our derivation we have
ignored all numerical factors. Therefore our conclusion should be that
binary heating becomes important when the number of stars in the core
has shrunk to a certain value. More careful estimates summarised by
Hut (1996) suggest that the “certain value” is actually in the range 50
to 80, whereas numerical models, such as those of Heggie & Ramamani
(1989), yield smaller values. For N -body models with unequal masses
(unpublished) the value is around 25, with no noticeable N -dependence
in the range in which it has been tested, i.e. 4096 ≤ N ≤ 32768.
Less important than the precise value is the fact that, within the lim-
itations of our estimates, the value is independent of the initial number
of stars in the cluster. Thus a small star cluster, with 100 stars, say, will
not collapse very far before three-body binaries become important. In
our million-body problem, however, binaries can influence the collapse of
the core only when the core has reached very high densities. This depen-
dence on N can be observed in the models displayed in the lower right
frame of Fig.9.3.

Two-Body Binaries
In the problem we have just been discussing, the stars are effectively point
masses, and all are single stars initially. This is the cleanest scenario for
the investigation of “core bounce”, but it is essentially unrealistic. In fact
stars have finite radii (though the point-mass approximation is adequate
for the degenerate stars, such as neutron stars), and the proportion of
primordial binaries is non-negligible (Part VII).
Consider first a system consisting of single stars with finite radii, i.e.
we continue to ignore primordial binaries. Binaries can now form by tidal
interactions, and no third body is needed to carry off the energy required
to bind them (Chapter 31). The important point is that formation of
binaries by this process is a two-body process. It is not immediately clear,
therefore, that it becomes increasingly important as the core collapses,
because collisional relaxation (which is driving the collapse) is also a two-
body process. Nevertheless it turns out that tidal capture does eventually
dominate. Indeed studies show that, in almost all star clusters, this
process would become important well before binary formation in three-
body interactions had a chance to halt core collapse (Inagaki 1984). We
shall not pursue these questions further, however, partly because the
subsequent fate of a binary formed by tidal capture is a controversial area
278 27 Surviving Core Collapse

(see Chapter 31 again), but also because this process is itself dominated
by interactions involving primordial binaries.

Primordial Binaries
Now we return to the assumption that the radii of the stars can be
neglected, and consider what happens if a certain proportion of the stars
present initially are binaries. As in the previous case, interaction between
binaries, or between binaries and single stars, is really a two-body process,
in the sense that its rate is proportional to the square of the space density.
Just as with tidal-capture binaries, therefore, it is not quite obvious that
the energetics of these reactions can compete with that of relaxation, but
in fact it does, at least if the proportion of binaries is high enough.
To understand why this is so we have to consider two effects which
enter the picture but work in opposite directions. The first effect is that
primordial binaries act in many respects like a species of heavy star. By
the process of mass segregation, therefore, primordial binaries accumulate
in the core, where their abundance steadily rises above its initial value as
the core collapses. This greatly enhances the efficiency of binary-binary
encounters in particular. The second effect is the destructive nature of
binary-binary encounters, which often convert the two binaries into one
binary and two single stars (Chapter 26). It turns out, however, that this
is insufficient to prevent the binaries from halting core collapse.
It seems very likely that it is primordial binaries which would bring core
collapse to a halt in a real star cluster, rather than either three-body or
tidal-capture binaries. Therefore it is worth re-estimating the conditions
under which core collapse will cease if we now consider binary-binary
heating rather than that due to the formation and hardening of three-
body binaries. The result (Box 1) is a condition of the form

f 2 /(η log Λ) ∼ 1, (5)

where η is a constant (the central relaxation time expressed in units of

the time until collapse tcc ) and f is the proportion of objects which are
binaries; if the quantity on the left is less than unity then core collapse
wins (and must be arrested by some other process, such as formation of
binaries in triple encounters), whereas heating wins in the opposite case.
Before we discuss this formula, we should consider whether sufficiently
many of the binaries can survive the carnage. Doing this precisely is
bedevilled by the difficulty of estimating how much of their energy is de-
posited in the core. If the binaries are not too hard, however, we may
estimate that this is a large proportion of all the energy they may emit
 Primordial Binaries 279

Box 1. Core bounce and primordial binaries.

In order to decide whether primordial binaries can arrest the collapse

of the core we need to estimate the rate of generation of energy in binary-
binary encounters. We assume that all individual stellar masses are com-
parable, of order m, and that the binaries are hard, with semi-major
axes of order a. As an estimate of the cross section for close encounters
(within a distance of order a) we may adopt an expression of the form of
Eq.(21.4). Without numerical factors this is simply Gma/v2 , where v is
a typical stellar speed in the core. (Strictly, we mean the typical speed
of a binary.) Let n be the number density of all objects, each binary
being regarded as a single object, and let f be the proportion of objects
which are binary. Each close encounter between two binaries will release
an energy of order Gm2 /a, and so the rate at which energy is created
is of order v(f n)2 (Gma/v2 )(Gm2 /a) per unit volume. Estimating the
volume of the core in terms of its radius rc , as before, we easily obtain
an estimate of the total rate, Ė, at which energy is generated. It is, if
anything, an underestimate, as we have ignored binary-single interactions
(see Goodman & Hut 1989).
As in the case of three-body binaries, the condition for core bounce is
obtained by equating Ė with the luminosity of the core, text Eq.(3). For
heating by binary-binary encounters this leads to the condition tcc ∼
v3 /(f 2G2 m2 n). Now tcc , which was defined to be the time scale of
core collapse, may be expressed as a multiple, (1/η, say,) of the cen-
tral relaxation time (cf. Chapter 18). (The quantity η is the one nu-
merical coefficient we keep track of, because it varies with time.) Thus
tcc ∼ v3 /(ηG2 m2 n log Λ), where the argument of the Coulomb logarithm
is Λ ∝ N (Chapter 14). Putting the estimates for energy generation and
the luminosity of the core together, we obtain the relation (5) in the text.

before destruction. Even if the softer binary is destroyed in every interac-

tion, the energy each binary emits will be of order Gm2 /(2a) per binary.
Finally, we err on the safe side by ignoring the substantial contribution
made in interactions between binaries and single stars in the core.
In view of these considerations we may safely estimate that the total
energy which the binaries in the core can release is of order f Nc Gm2 /a,
where Nc is the number of objects in the core. This is sufficient (in
amount) to halt core collapse if it exceeds the energy released by the core
on its collapse timescale, which in turn is of the order of the energy of
280 27 Surviving Core Collapse

the core itself, ∼ Nc mσc2 . Thus enough binaries survive if

GM 2 /a >
f 1. (6)
mσc2 ∼
The coefficient of f here is a measure of the hardness of the binaries.
The most effective binaries are not too hard, otherwise the time scale
on which they interact is too long compared with that of core collapse,
and they deposit little of their energy in the core, where it is needed.
For somewhat hard binaries, then, it is not implausible that enough will
survive to halt core collapse provided that f is not too low.
Though this discussion gives the impression that consumption of bina-
ries is a bad thing, it has interesting observable consequences. The single
stars which are created in binary-binary interactions may be expelled
from the system with high speeds, and this has often been proposed as
an explanation of “high-velocity stars” (e.g. Leonard & Duncan 1988).
It has also been argued that the only reason we see single stars at all is
that they were created (or released) in binary-binary interactions in star
clusters (Kroupa 1995).
Now we return to the discussion of Eq.(5), which is the condition for
the rate of heating to be adequate. Notice that n, the number density
of stars, is not involved, because both collisional relaxation and binary-
binary heating are two-body processes, in the sense already explained.
It follows that the arrest of core collapse by binary-binary heating is a
slightly delicate matter. Of the three quantities in Eq.(5), all work in the
required direction. As core collapse proceeds, f increases because of mass
segregation, as already explained. Also, η decreases as core collapse ad-
vances (as mentioned briefly in Chapter 18), as the thermal coupling be-
tween the core and the cooler, outer parts of the cluster becomes weaker.
Finally, if we assume that the effective value of Λ is proportional to the
number of stars in the core, as has been argued (Spitzer 1987), then this
factor decreases also. However, the variation of this term is the weakest.
A crucial point to notice is that all these factors vary on a time scale
which scales with the relaxation time. Suppose, then, that we have a
system with a given N in which core collapse is terminated at some
point, measured by a certain increase in the central density. Now consider
a system with quite different N , but the same initial structure and the
same initial proportion of binaries. (The binaries must also be equally
hard in both systems initially.) In the second system the same increase in
central density occurs after the same number of initial relaxation times.
And at this point the increase in f and the decrease in η will be the same
in both systems. Even the fractional decrease in the number of stars in
the core will be the same. It follows that the two systems will reach core
bounce at about the same point, i.e. after the same increase in the central
 Primordial Binaries 281

density of stars. This is in marked contrast with the arrest of core collapse
by the production of three-body binaries, where the number of stars in
the core at core bounce is independent of N . In that case, the depth
at which core collapse is arrested (as measured by the increase in the
central density) increases with N . When primordial binaries are present,
however, the depth of collapse at core bounce is almost independent of N ,
and deep collapses are prevented (Fig.2). One may say that the essential
physical reason for the difference in the two cases is that binary-binary
heating and mass segregation are two-body processes, while binary-single
heating is a three-body process.
There are, it is true, considerable approximations and simplifications
behind this assertion. In particular, primordial binaries do not last for-
ever. Some are consumed in arresting core collapse, and we must consider
whether enough survive to prevent a later recollapse before the cluster
dies (Chapter 29). For the meantime, however, core collapse has been
brought to a halt, and in the next chapter we consider what options the
system has for its subsequent behaviour.
Core radius

0.1

0.01
0 100 200 300 400 500 600 700 800
Time

Fig. 2. Core collapse with primordial binaries. Two N-body models contain
2425 objects. In one model (dashed line) 75 of these these are single stars with
twice the mass of the other stars, in the other (solid line) they are binaries. For
further details see Heggie & Aarseth 1992, where they are referred to as models
I and S, respectively.
282 27 Surviving Core Collapse

Problems
1) Loss of mass m from the core leads to an effective heating of the
cluster by an amount mφc , where φc is the central potential (Chapter
23). Suppose that some kind of stellar explosion leads to mass loss at
a rate Ṁ < 0 from the core. Neglecting numerical factors, show that
this is sufficient to arrest core collapse if −Ṁ ∼ Mc /trc , where Mc
is the mass of stars in the core and trc is the central relaxation time.
If the mass ejection is associated with stellar evolution, we may sup-
−Ṁ
pose that the associated luminosity of the core is Lc ∼ L ,
M /1010 yr
since the lifetime of the sun is of order 1010 yr. If the cluster is al-
most isothermal (Chapter! 8) out to the half-mass radius, deduce that
10
10 yr
Lc ∼ L rh /rc , where L is the luminosity of the cluster (in
trh
the usual astrophysical sense, but not as in Eq.(3)). (This would imply
a huge core luminosity; cf Goodman (1989), who considers the same
issue after core collapse.)
2) Write down the equilibrium equations in the gaseous model (Box
9.2). Show that no reasonable solution is possible if the rate of energy
generation, ε, takes the form in Eq.(23.5). Try to formulate a physical
understanding of this result.
 28
Gravothermal Oscillations

In the last chapter we saw that core collapse ends in what is (by stellar
dynamics standards) a blaze of energy, which is emitted in interactions
involving primordial binaries. Dramatic though that sounds, the real
climax of the whole book is reached in the present chapter, where for the
first time we catch a glimpse of the entire lifespan of a cluster. Admittedly
we concentrate here on a highly idealised cluster, isolated from the rest
of the universe, consisting of stars of equal mass, and totally devoid of
primordial binaries. In the next chapter we shall relax some of these
idealisations.

Steady Post-Collapse Expansion

What happens to the energy generated at the close of core collapse? To

answer this question it is easiest to think of the cluster as a conduct-
ing, self-gravitating mass of gas, with its temperature decreasing from
the core to the tenuous boundary. Because of the assumed temperature
gradient, the heat flux is outwards, and the time scale for the transport
of thermal energy is of order the local relaxation time, tr . Therefore,
after an interval of order trh , i.e. the value of tr at the radius rh which
contains half the mass, the thermal energy generated at core bounce has
diffused throughout the bulk of the cluster. In the same time interval M ,
the mass of the cluster, has hardly altered (cf. Problem 16.1). Therefore
the effect of the generation of energy follows from the estimate

rh ∝ GM 2 /|E|, (1)

283
284 28 Gravothermal Oscillations

where E is the energy of the cluster∗ . Because of energy released in binary

formation and evolution this increases, i.e. it becomes less negative, and
so the cluster expands.
There is an interesting analogy here with what happens to stars as
they settle on to the main sequence. They have contracted sufficiently
that the core is hot enough to sustain equilibrium through burning fuel
in nuclear reactions. In much the same way, the core of a cluster has
contracted sufficiently to supply enough energy by the consumption of
primordial binary stars† . The main difference is that a star is able to
radiate away the energy it generates. Star clusters, on the other hand,
can only conduct heat (relaxation) or convect it (escape), but conduction
is poor where the density is low, and convection is poor because the
escape rate is so slow, and so the heat cannot get out. The only option
open to the cluster is expansion.
Though this conclusion is correct, the argument gives the impression
that the expansion of the cluster is a consequence of the generation of
energy. It is easy to imagine that the binaries are the engine which
drives the subsequent evolution of the cluster, rather like a rocket engine
under the launch vehicle of a payload, and that the cluster simply has
to respond to whatever energy the binaries choose to provide. But more
careful consideration of what happens during core collapse will show us
that it is the behaviour of the core that is governed by the flow of energy
across the half-mass radius, and not the other way around.
The idea that the core responds to the energy needs of the rest of the
cluster was a remarkable insight gained by Michel Hénon (1975). It was
a vital breakthrough which allowed theorists to overcome the impasse of
core collapse.
The mechanism by which the core responds to the cluster as a whole can
be summarised as follows. As we have seen, the temperature gradient and
relaxation time at rh determine the energy flux there. This flux must be
supplied from within – somehow. If there are no binaries in the core then
the core will tend to collapse, just as in the collapse phase we studied in
Chapter 18. This drives the stellar density sufficiently high that the stars
are forced to produce enough power, by any of the mechanisms outlined
in Chapter 27, e.g. by making and hardening binaries. If, on the other
hand, the core is over-productive, then the energy it produces will heat
the core faster than the heat can be transported towards rh ; the core
expands, reducing the stellar density, and the rate at which the stars

∗
More precisely it is what is sometimes referred to (e.g. Giersz & Spurzem 2000) as
the external energy, i.e. the part not locked up inside binary stars (which is referred
to as the internal energy).
†
Perhaps a better analogy is the burning of deuterium.
 Steady Post-Collapse Expansion 285

are interacting. Therefore there are self-regulating mechanisms which

prevent either too much or too little energy being released by binaries in
the core.
Actually, the argument we have just given was already well known
in the context of a problem we have often used as an analogy for star
cluster evolution: the theory of stellar structure. Stars are “secularly
stable” because the core of a star automatically adjusts itself to produce
as much energy as is required by conditions of energy transport at larger
radii (e.g. Schwarzschild 1958, Sec.II.6; Prialnik 2000).
We return to stellar dynamics. Before going on, however, we should
point out where the energy comes from during the core collapse that we
studied in Chapter 18, when there may be no binaries. In this phase
of evolution the temperature gradient at rh , and the value of trh , are
quite comparable to the values immediately after core bounce, and so
the energy requirement of the cluster are very similar. The difference is
that, during core collapse, this energy is supplied by the collapse itself.
This form of energy supply also differs from binary heating in that it does
not alter the external∗ energy of the cluster, and so there is no overall
expansion of the cluster; the half-mass radius is almost static, in fact.
Now we return to what happens after core bounce. Hénon’s argument
suggests that the core now reaches a kind of equilibrium, settling into
a state where it provides just the energy required by the rest of the
cluster. On this basis it is not hard to estimate the conditions in the
core. The first point to observe is that the rms speed of stars cannot be
very different from that at rh : we have asserted that the flux from the
core is comparable with that at rh , where the time scale of heat transport
is trh , but the value of tr is very much shorter in the core. Therefore the
vicinity of the core must be nearly isothermal, otherwise the flux from
the core would be far too high. Besides the rms speed, only one other
parameter is needed to specify conditions in the core, such as its mass or
radius, and its value can be determined if we balance the heat generated
by binaries in the core with the flux at the half mass radius. It turns
out (Box 1) that the number of stars in the core in steady post-collapse
expansion is of order N 1/3 . The average number of binaries present in
the core varies as N −1/3 (Goodman 1984).
On the basis of Hénon’s picture the entire evolution of the cluster, and
not just of the core, now falls into place. First, the time scale of the
expansion is of order trh , and so it follows that trh ∼ t − tcoll , i.e. the

∗
See the first footnote on the previous page.
286 28 Gravothermal Oscillations

time since the end of core collapse. Using Eq.(14.9), we may deduce that

3/2
M 1/2 r
t − tcoll ∼ 1/2 h , (2)
G m log Λ

where m is the individual stellar mass and Λ is the argument of the

Coulomb logarithm. Actually this estimate should really be read back-
wards: M varies on a rather longer timescale, and Eq.(2) shows that
rh ∝ (t − tcoll )2/3 , roughly.
The slow escape of stars in post-collapsing isolated models is an in-
teresting and little explored problem. Some time ago N -body models
showed that M ∝ (t − tcoll )−ν , where ν ' 0.085, and this alters the time-
dependence of the expansion to nearly rh ∝ (t − tcoll )(2+ν)/3 (Giersz &
Heggie 1994). By today’s standards the values of N in these simulations
were modest, and this research did not address the N -dependence of the
escape rate. If mass is lost on the relaxation time scale then the time for
a fixed fraction of mass to escape should vary nearly in proportion to N ,
like the relaxation time. Recent N -body models by H. Baumgardt, how-
ever, show that the time taken to lose about 70% of the mass is almost
independent of N . The tentative reason for this appears to be the higher
concentration of models with larger N (Box 1), which leads to a larger
escape rate per relaxation time in larger models.
It is interesting that the dependence of rh on t follows from dimensional
analysis, given only G, M and rh . The same expansion law is obeyed
in two-body collisions (Box 21.2), cold collapse (Fig.2.1) and even the
Einstein-de Sitter cosmological model (e.g. Peebles 1993). There is, of
course, a large dimensionless number (N ) in our problem, and this can
make a big difference in the coefficient.
At last we can glimpse the eventual fate of the million-body problem.
It expands on an increasing time scale, slowly losing mass as it does so∗ .
After much time has elapsed it looks like a small N -body system with just
a few tens of stars, but of immense size, and evolving almost impercep-
tibly. Moving away, and dominating at great distances, are the escapers,
not only single stars but also binaries and sometimes even triples and
higher multiples. Eventually the last escaper is ejected, leaving behind
one last binary or triple.

∗
Continuing an earlier remark, it is tempting to draw an analogy with the evolution
of red giants. That is too controversial a topic, however, for the analogy to be
enlightening.
 Unsteady Post-Collapse Evolution 287

Box 1 Core radius in steady expansion with three-body binaries.

The flux of energy from the core is given by Eq.(27.1). In steady

expansion this must be comparable with the flux at rh , which is of or-
der |E|/trh . Using text Eq.(1), and Eq.(14.8), and assuming that the
velocity dispersion v2 does not vary much inside rh , we deduce that
G2 m2 ρ3c rc3 rh ∼ M 2 v4 ρh log Λ, where ρc and ρh are the density in the
core and at rh , respectively, m is the individual stellar mass and Λ is the
argument of the Coulomb logarithm.
This estimate is not very illuminating, but we note that the core radius
is defined so that Gρc rc2 ∼ v2 (Eq.(8.9)), and so our condition may be
written as
 2
m ρc rh 1
∼ 1.
M ρh rc log Λ
Finally, from the fact that the conditions inside rh are nearly isothermal,
it follows that ρ ∝ r−2 between rc and rh (Chapter 8). Therefore if we
neglect the slow N -dependence of Λ, we conclude that rh /rc ∝ N 2/3 .
Since the mass inside a given radius is nearly proportional to r in an
isothermal model, it follows also that the number of stars inside the core
in this post-collapse expansion varies as Nc ∝ N 1/3 .

Unsteady Post-Collapse Evolution

Apart from the title of this section, perhaps the reader has spotted one
or two other veiled clues that our picture of postcollapse evolution has
some gaps. One of these is the mismatch between the number of stars
in the core after core collapse (Box 1) and the number at core bounce,
where we estimated (Chapter 27) that Nc is independent of N . It follows
that, in a sufficiently rich system, the core must reexpand in the early
phase of post-collapse evolution. Core bounce really is a bounce, and
there must be a period of readjustment if the core is to settle down to
the equilibrium size required after core bounce.
The second clue is that Hénon’s argument does not, despite appear-
ances, imply that the required equilibrium is ever reached. Within the
terms of this argument it is equally possible that the balance is reached
only in a time-averaged sense, and that the core really does alternate
between high- and low-density phases.
The third clue is, in physical terms, the most telling. We mentioned
that the inner parts of the cluster must be nearly isothermal if the en-
ergy produced in the core is to balance that required at rh . Now we
288 28 Gravothermal Oscillations

saw in Chapter 17 that such a system is gravothermally unstable, and so

we should expect a time-dependent behaviour different from the steady
expansion we have assumed. It was Daiichiro Sugimoto and Erich Bet-
twieser who first understood the consequences of this (Sugimoto & Bet-
twieser 1983, Bettwieser & Sugimoto 1984).
The first and third clues are related. At the end of core bounce the core
is producing enough energy to halt the collapse, and this is too much for
the subsequent expansion of the cluster as a whole, which is determined
by the heat flux at rh . Therefore the thermal energy generated builds up
in and around the core faster than it can be conducted away. This causes
an expansion and cooling of the core and its immediate surroundings,
because of their negative specific heat. If core collapse is sufficiently
deep, which simply requires sufficiently large N , the expanding core can
actually cool to temperatures below that of its surroundings. At this point
there is a temperature inversion, i.e. there is a range of radii in which the
temperature increases outwards (Fig.1). In this region the thermal flux
is actually inwards. This flux enhances the expansion of the core and its
cooling, which simply reinforces the driving force behind the expansion.
It is the gravothermal instability again, this time working in reverse.
Eventually, however, the expanding core comes in thermal contact with
the cooler parts of the cluster around rh . Now the heat generated in the
core can flow out as required, and the temperature inversion disappears.
But by this point the core is so distended that it is producing insufficient
energy. By Hénon’s argument it is forced to recollapse so that its density
is sufficiently high to produce the required energy. But this collapse is
more-or-less like the first core collapse: the core overshoots, creates too
much energy, and a temperature inversion, and the cycle recurs.
What we have outlined is a picture of what are now called gravothermal
oscillations (Box 2). The detailed nature of these oscillations can be stud-
ied in various ways, but analogies with other problems suggest what we
might expect to find. We have argued that it is steady expansion which
is unstable, but it is easy to transform the problem into a problem on
the stability of equilibrium (Goodman 1987). In steady post-collapse ex-
pansion of an isolated model of constant mass we expect that radii scale
as (t − tcoll )2/3 , by Eq.(2). Therefore if we write r̃ = r(t − tcoll )−2/3 ,
and treat the other variables such as the space density similarly, we shall
turn what is really a steady expansion into what looks like an equilib-
rium, just as the expansion of the universe can be frozen out in comoving
coordinates.
Now the arguments of the present section of this chapter imply that
this equilibrium is unstable, for N sufficiently large. Thus we are studying
 Unsteady Post-Collapse Evolution 289

0.5

Log velocity dispersion 0.49

0.48

0.47

0.46

0.45

0.44
-3.5 -3 -2.5 -2 -1.5 -1 -0.5 0
Log radius

Fig. 1. Velocity dispersion (“temperature”) profile in a gas model undergoing

a gravothermal expansion (after Heggie et al. 1994).

a system in which an equilibrium becomes unstable as some parameter

(N in this case) changes. One of the commonest scenarios in such prob-
lems is that the instability takes place by a sequence of period-doubling
bifurcations, which accumulate at a parameter value at which chaotic os-
cillations set in (e.g. Cvitanović 1989). This is roughly what happens in
detailed models of the post-collapse behaviour of star clusters, both gas
models (Fig.2) and Fokker-Planck models (Breeden et al. 1994). When
N is small enough (less than a few thousand, in fact, if all stars have the
same mass), the post-collapse expansion is stable and steady. Just above
some value of N the expansion exhibits a simple periodic behaviour. At
somewhat larger N , alternate cycles have different amplitudes: the pe-
riod has doubled. This period-doubling presumably cascades rapidly as
N increases, until at a somewhat higher critical value of N one sees for
the first time non-periodic oscillations. Tests show that these have the
hallmarks of chaotic variations.
The analogies with other systems exhibiting this behaviour can be
pushed further. In particular, although the core, being a continuum, is a
system with infinitely many degrees of freedom, it is found that its evolu-
tion is nearly restricted to a low-dimensional (actually two-dimensional)
subspace (Breeden & Cohn 1995). Further, inspection of the flow on this
subspace shows that it can be represented by a one-dimensional first re-
turn map with the same qualitative properties as the famous logistic map,
290 28 Gravothermal Oscillations

Box 2 A brief history of gravothermal oscillations

These were discovered by Bettwieser and Sugimoto using a gaseous

model (Chapter 9). At that time, however, other gaseous models (com-
puted by one of us [Heggie 1984]) and Fokker-Planck models (Cohn 1985)
were showing no indication of this behaviour, and so the discovery was
greeted with much scepticism when it was announced and discussed at a
conference in 1984 (which was organised by the other of us [Goodman &
Hut 1985]). Behind-the-scenes discussions at that meeting drew attention
to the importance of the choice of time steps in the numerical integra-
tions. In particular, M. Schwarzschild mentioned to one of the authors
the resemblance to the numerical problems which had alerted him to the
He core flash in a Population II giant (Schwarzschild & Härm 1962). Not
long after the Symposium, further work by the main disputants confirmed
the oscillations. Jeremy Goodman (Goodman 1987) then showed that the
steady expansion, though possible, was unstable if N was large enough.
All of this work was conducted with simplified models, and there was
considerable doubt whether N -body models could exhibit such behaviour.
It was argued that the fluctuating temperature in a discrete model could
mask the subtle temperature inversion that plays such a crucial role in
the expansion phase. Limited N -body modelling (of a specially tailored
system in a reflecting sphere [Heggie et al. 1994]) at least removed this
objection.
It was still not possible to check that the full cycle of oscillations would
actually occur; computers were not powerful enough. A detailed analysis,
however (Hut et al. 1988), showed that the computer power necessary to
exhibit gravothermal oscillations, if they existed, was of order 1 Teraflop.
Eventually the GRAPE-4 computer was finished at the University of
Tokyo (Chapter 3). It had been specifically constructed to settle this
point, and was the first computer in the world to reach the required speed.
Its first results were presented at another conference (Hut & Makino
1996), this time in Tokyo and organised by Daiichiro Sugimoto, just over
ten years after the first announcement of gravothermal oscillations. The
existence of gravothermal oscillations in sufficiently large N -body systems
was now established (Makino 1996; and Fig.9.3, lower right panel).

in which so much of the basic theory of period-doubling was developed

(cf. Cvitanović 1989).
Admittedly, there are some important respects in which this picture
must be qualified. For one thing, it is complicated by the fact that the
 Unsteady Post-Collapse Evolution 291

Fig. 2. Period doubling and chaos in post-collapse expansion. Central density

is plotted against time for gas models with N = 6000 (top left), 8000 (bottom
left), 10000 (top right) and (bottom right) 50000 stars (after Heggie & Ramamani
1989).

generation of energy in N -body systems is stochastic in nature, as it de-

pends on the accidental formation and evolution of individual binaries.
It is not clear that period-doubling could be observed in N -body simula-
tions in these circumstances, though it is certain that there is a change
in character in the oscillations of the core at around the values predicted
by the simple models (Fig.9.3, lower right panel). Below those values
the fluctuations look simply random, and some are clearly associated
with energetic few-body interactions. Above the predicted values, the
oscillations, while chaotic, have a systematic character, independent of
specific identifiable binary events. The second qualification is that, when
we consider slightly more realistic models of stellar clusters, e.g. with a
spectrum of masses or primordial binaries, it turns out that conditions
for the onset of gravothermal oscillations are less favourable. The next
chapter explores some of these less idealised models, and little more will
be said of unsteady post-collapse evolution.
292 28 Gravothermal Oscillations

Problems
1) According to an isolated model of Hénon (1965) the time scale
of evolution in post-collapse expansion is given by rh /ṙh ' 10.8trh ,
where trh is the current half-mass relaxation time. N -body models of
this phase show that the number of particles diminishes (by escape)
roughly according to d ln N/d ln rh ' −0.12 (Giersz & Heggie 1994),
while the loss of mass in the precollapse phase is a few percent and
may be neglected here. Fokker-Planck models show that the duration
of core collapse for a Plummer model is about 18trh (0) (Spitzer &
Thuan 1972, Takahashi 1995), where trh (0) is the initial value. Es-
timate the entire lifetime of the model (i.e. until N ' 2) in units of
trh (0).
2) Several toy models illustrating gravothermal behaviour were investi-
gated by Allen & Heggie (1992). Among them was the simple system

dT1 Γ (4−α)/(α−2)
= ξ(T1 − T2 ) − cT1
dt Γ−1
dT2 1 −(6−α)/[2(α−2)]
= ξ(T1 − T2 ) − (T2 − 1)T1
dt Γ−1
where T1 is the temperature of the core, T2 is the temperature of
the region where a temperature inversion may develop, and Γ ' 1.28,
α ' 2.23 and ξ ' 0.00037 are constants. The parameter c ∝ N −2 , and
this term corresponds to binary heating. The forms of this and the
other terms were devised by simple phenomenological considerations
of gravothermal behaviour.
Using appropriate numerical or analytical tools, demonstrate the
following: (a) for sufficiently large c there are no equilibria; (b) for
small enough c there are two equilibria; (c) as c decreases, the hotter
equilibrium first changes from a stable node to a stable focus; (d) as c
decreases further, the hotter equilibrium loses stability and throws off
a stable limit cycle in a Hopf bifurcation.
These changes correspond physically to the following: (a) when N
is very small, the presence of a binary leads to the rapid expansion of
the core; (b) for larger N a stable balance is possible between binary
heating and gravothermal cooling; (c) for still larger N , the approach
to this stable balance has the nature of a damped oscillation (as is
observed in gas models); (d) for large N the balance between heating
and cooling is a dynamic (oscillatory) one.
A further variable is needed before a simple model like this can
demonstrate chaotic behaviour similar to well developed gravothermal
oscillations.
 Unsteady Post-Collapse Evolution 293

3) Detailed models show that gravothermal oscillations mainly affect

stars inside the radius r.01 containing the innermost 1% of the mass
(Cohn et al. 1989). If the region from there out to the half-mass radius
is nearly isothermal, show that the relaxation time there is tr,.01 '
10−4 trh . If tr,.01 governs the time scale of gravothermal oscillations,
and if each binary releases an energy amounting to about 320mσc2
(Chapter 23), show that about one binary per oscillation is needed for
N = 106 .
 29
Dissolution

The previous chapter dealt mostly with a highly idealised model. All
stars were single and had the same mass, and the system was isolated.
As we saw, even the presence of a spectrum of stellar masses changes
the picture, as it is found that gravothermal oscillations set in only for
considerably larger values of N . In the present chapter we shall also see
that the presence of primordial binaries further weakens their probable
relevance. Even when gravothermal oscillations do occur, they seriously
affect the structure of only the innermost 1% or so of a cluster. Therefore
in this chapter we concentrate once more on the steady post-collapse evo-
lution of a stellar system. Also, we mainly have in mind a system with a
significant population of primordial binaries, and boundary conditions set
by the tidal field of the surrounding galaxy. First, however, we consider
the simpler case of an isolated cluster.

Isolated Clusters
The first thing that is changed in post-collapse evolution when we add
primordial binaries is the radius of the core. A similar argument to that
of Box 28.1, shows that the ratio rh /rc is now almost independent of N
(cf. Problem 1). In fact the ratio depends more on the proportion of
binaries (which decreases as the binaries are consumed).
These statements greatly weaken the clues to the occurrence of gravo-
thermal oscillations which we discussed in the case of post-collapse evo-
lution powered by three-body binaries (Chapter 28). In the first place,
the ratio rh /rc at core bounce is also independent of N , if core collapse
is halted by primordial binaries (Chapter 27). Therefore, if there has to
be a period of adjustment to post-collapse conditions, at least it does
not get more serious as N increases. Second, the post-collapse core does
not get smaller and smaller as we consider systems with larger and larger
N . Therefore there is no need to posit a large (and unstable) isother-

294
 Isolated Clusters 295

mal region within the half-mass radius. Finally, it is quite possible that
this post-collapse core is as large as that in a modest N -body system
powered by three-body binaries, and we know that these do not exhibit
gravothermal oscillations.
The only worry is that a system may not have enough primordial bi-
naries to sustain a sufficiently long period of post-collapse expansion. A
substantial fraction were already burned simply in halting core collapse
(Chapter 27), and the residual population is gradually used up in pow-
ering the post-collapse evolution. After that, the only effective heating
mechanism left would be the formation and hardening of three-body bi-
naries. The point at which this first happens clearly depends on the
initial proportion of binaries. In early simulations (Spitzer & Mathieu
1980) it was found that core collapse was delayed by less than the time of
core-collapse, even for large binary fractions. Nowadays, however, there
is a consensus (with various methods and authors) that, if the initial
proportion of binaries is of order 10% (depending on the distribution of
binary parameters), then the binaries can keep gravothermal oscillations
at bay for a time about three times the length of the phase of core col-
lapse (Fig.1; see also McMillan & Hut 1994). Even this may, however, be
less than the lifetime of a number of globular clusters.

Fig. 1. Evolution of a cluster with primordial binaries, after Gao et al. (1991).
From top to bottom the curves are the half-mass radius of single stars, all stars
and binaries, and the core radius. All are plotted against time. Mass segregation
during core collapse is followed by a relatively steady post-collapse phase until
the binaries are exhausted, at which point gravothermal oscillations begin.
296 29 Dissolution

With this caveat we assume that the evolution of the core need not
distract us, and turn to the evolution of the cluster as a whole. We
already saw in Chapter 28 what happens if the system is isolated and
has stars of equal mass. When a spectrum of masses is present we might
expect our simple picture to be spoiled by the continued action of mass
segregation. In fact, however, N -body simulations suggest that our simple
picture holds up remarkably well. They show that mass segregation slows
dramatically towards the end of core collapse, and that thereafter the
profile of mass segregation does not alter very much (Fig.16.1). Simple
relations like
3/2
M 1/2 r
t − tcoll ∼ 1/2 h , (1)
G m log Λ
(cf. Eq.(28.2)) apply with unexpected accuracy even when there is a
broad spectrum of stellar masses, though the reasons for our good fortune
in this respect are not understood.

Tidally bound clusters

Next, let us replace our idealised picture of an isolated system by one
with a somewhat more realistic boundary condition. Tidal truncation
(Chapter 12) has the characteristic that the radius of the system, rt ,
must vary with its mass as
rt ∝ M 1/3 , (2)
which implies that the mean density of the system is constant. We have
seen, however, that an isolated cluster expands in post-collapse evolu-
tion, and thus its density reduces. What happens with a tidal boundary
condition is that the expansion pushes the tenuous outer layer over the
tidal boundary, leaving behind the denser material at smaller radii. A
balance is struck between the truncation by the tide, which leaves behind
denser material, and the expansion, which reduces the density, and the
net effect is that the mean density remains constant.
In isolated systems the mass is nearly constant, and so E ∝ −1/rh . In
tidally bound systems, however, the half-mass radius will be comparable
with (but smaller than) the tidal radius, and so it follows from Eq.(2)
that E ∝ −Grh5 . Thus the increase in E caused by binary interactions in
the core is accompanied by a decrease in rh : the cluster shrinks as it is
heated. This is quite unintuitive for someone brought up on the evolution
of isolated clusters.
Now we consider the time scale of this evolution. The right hand side
of Eq.(1) is an estimate of trh , and so Eq.(2) shows that that this is
proportional to M . Arguing as before that trh determines the time scale
 Tidally bound clusters 297

of evolution, we see that the evolution speeds up as the system loses mass,
and in fact should come to an end at a finite time in the future, tend , say.
The time dependence is therefore obtained by equating the right-hand
side of Eq.(1) to the time remaining. Using Eq.(2) again, it follows that
M ∝ tend − t (cf. Fig.2), and that the remaining lifetime of the cluster
is proportional to the current half-mass relaxation time. In the model of
Hénon (1961) the constant of proportionality is approximately 22.4trh .
The tidal truncation progressively removes the stars of highest energy.
By mass segregation these tend to be the stars of lowest mass. The
result is that a dying cluster contains an increasing concentration of the
more massive objects. As in panning for gold (or neutron stars), the tide
washes out the lighter particles. Incidentally this helps to compensate for
the destruction of binaries in the binary-binary reactions that we have
relied on to sustain the post-collapse evolution.
Stars which escape from the cluster, or which are washed across the
tidal boundary, do not necessarily leave the vicinity of the cluster quickly.
Unless of high energy, e.g. escapers resulting from three-body interac-

60000

50000
Mass (solar masses)

40000

30000

20000

10000

0
0 5000 10000 15000 20000 25000
Time (Myr)

Fig. 2. Loss of mass in tidally limited N -body models. From right to left the
initial number of stars is N = 1024, 2048, 4096, 8192, 16384 and 32768. The
models have been scaled to a real cluster with initial mass 6 × 104 M in a tidal
field of a certain strength, and with the same mass function and initial structure
(see Heggie et al. 1998). The scaling is imperfect, for reasons which are discussed
in Baumgardt (2001). The modest increase in the escape rate after about half the
mass has escaped is associated with the change from pre-collapse to post-collapse
evolution.
298 29 Dissolution

tions, they cross the tidal boundary at low speed in the vicinity of the
Lagrangian points (Chapter 12). Thereafter they move off under the ac-
tion of tidal and inertial forces to form “tidal tails” of escaping matter
(Fig.12.3). When the cluster as a whole dissolves, in the sense that its
tidal radius shrinks to zero, there is still a residual concentration of stars
– the smile on the face of the Cheshire cat.
It is remarkable that post-collapse evolution in a tidal field was the first
evolutionary problem in the subject to be understood in detail (Hénon
1961). At that stage the theory was developed for the case of a system
in which all stars have the same mass, the distribution of velocities is
isotropic, and the tidal field is modelled as a cutoff at the tidal radius.
Empirically, however, it has been found that much the same conclusions
apply to much more realistic systems: unequal masses, anisotropic ve-
locity distribution, anisotropic tidal field, primordial binaries, etc. The
nearly linear dependence of the mass on time is one of the most robust
and memorable results in the whole of this subject, and a nice exam-
ple by J. Stodólkiewicz appears prominently on the front cover of the
proceedings of IAU Symposium No.113 (Stodólkiewicz 1985).

Problems
1) Use estimates in Box 27.1, to show that the total heating rate of pri-
mordial binaries in the core is of order f 2 mσc3 /rc , where f is the binary
fraction, m is the individual stellar mass, σc is the central velocity dis-
persion, and rc is the core radius. Deduce that, in steady post-collapse
expansion powered by primordial binaries, the ratio rc /rh is almost
independent of N .
2) (Spitzer 1987, p.59) In homologous evolution in a tidal field show that
the total energy varies with mass according to
d ln |E| 5
= .
d ln M 3
If stars escape across the tidal boundary, modelled as a cutoff at
the tidal radius rt , with negligible excess energy, show that Ė =
−(GM/rt )Ṁ, and deduce that E = −0.6GM 2 /rt . Infer that rt = 2.4
in N -body units (Box 1.1). (In the limit W0 → 0 King models have
this limiting radius [cf. Problem 8.6]. Most models have larger tidal
radii, however [cf. Fig.8.3] and require an additional source of energy
for homologous evolution.)
3) Suppose the estimate in Problem 1, with rc replaced by rh and σc the
rms 1-dimensional speed, can be used to estimate the rate of heating
in an isolated uncollapsed cluster with half-mass radius rh , and that
mass is lost by evaporation (Eq.(9.3), where tr is estimated by the half-
mass relaxation time trh ). Using standard estimates for a virialised
 Tidally bound clusters 299

cluster, deduce that

d ln rh f2
−2'− ,
d ln M µ ln(λM/m)
where µ and λ are constants. Sketch solutions of this differential equa-
tion, assuming that f is also constant. (Cf. Hills 1975, where heating
between single stars and primordial binaries was considered.)
 PART IX. STAR
CLUSTER ECOLOGY

The remaining chapters of the book go beyond the N -body problem as

it is understood outside astrophysics. Here the fact that the bodies are
stars is essential.
Chapter 30 sets the scene by summarising the various dynamical pro-
cesses that have been introduced so far, the various kinds of stars and
other relevant topics which are of interest in astrophysics (such as colour-
magnitude diagrams) and, most importantly, the relations between these
two sides of the problem. Special attention is paid to those kinds of stars
which are readily observed and where dynamical processes are most im-
mediately relevant: blue stragglers, millisecond pulsars and X-ray sources.
Chapter 31 analyses in some detail the simplest process where the stel-
lar nature of the bodies is vital: collisions and other encounters between
two individual stars, where the gravitational interaction is not the whole
story. We estimate the rate of collisions, and how it depends on the stellar
density and the kinds of stars present. Non-gravitational interactions are
also vital in understanding the role and evolution of binary stars, espe-
cially when interactions with other (single) stars occur frequently enough.
The effects of collisions on the participating stars are outlined, and we
consider the dynamics of near-collisions, where non-gravitational effects
are important (“tidal capture”).
The dynamics and evolution of binary stars is taken up in detail in
Chapter 32. Special attention is paid to the ways in which blue stragglers
can arise from interactions involving binaries. We give examples of the
effects of encounters between a binary and a single star, and those with
another binary. This is a growth area of the subject, and our outline of
its current status is likely to be out of date by the time this book is in
print.
The final chapter, Chapter 33, summarises the astrophysical million-
body problem. We recap the various dynamical processes we have dis-

300
 PART IX. STAR CLUSTER ECOLOGY 301

cussed and their interrelations, and present the overall evolution of a star
cluster, from birth to death, in a simplified diagram which encapsulates
the gross aspects of its spatial structure. We consider again the timescales
on which the evolution takes place. Finally we consider how it is that the
clusters we see have managed to survive for most of the lifetime of our
Galaxy, and the rate at which these last survivors are perishing.
 30
Stellar and Dynamical Evolution

Globular star clusters have an important place in modern astrophysics

for several reasons, but let us mention just two here. Firstly, they are
a laboratory for the study of gravitational interactions and dynamical
evolution, and this is the motivation for much of the research that we
have written about in this book. Secondly, however, each cluster is also
a sample of stars of very similar age and composition, and are an ideal
test-bed for theories of stellar evolution.
Over the years these two aspects of cluster studies built up their own
communities of theorists and observers, and their own suites of problems
(Fig.1). Now what is remarkable is that, for many years, research in
these two areas proceeded in almost total isolation from each other. It
was possible to pursue an active and successful research career on one side
of this diagram (Fig.1) without even being aware of the existence of the
people working on the other side. Whenever one did need something from
the other side, the most primitive tool for the job was used. Dynamicists
would use mass functions which to any observational astronomer would
seem distinctly bizarre, while observers, if they ever needed a theoretical
model, would dust off an old one which ignored decades of subsequent
theoretical development. Dynamicists were fascinated by the problems
of stellar systems with stars of only two different possible mass, while
fitting a Fokker-Planck model was something that no non-theorist ever
attempted.
Within about the last decade all that has changed. The volume of
proceedings of a meeting held in 1990 (Janes 1991) gives a snapshot of
this shift in outlook. It was increasingly being realised that dynamics
has an influence on the kinds of stars that are to be found in globular
clusters, and that these in turn influence the dynamics that can occur.
The feedback mechanisms between stellar dynamics and stellar evolution

302
 30 Stellar and Dynamical Evolution 303

Internal
Dynamics Stellar
Constituents
Core collapse
Mass segregation
Variable stars
Stellar escape
Mass spectrum
4-body encounters
Binaries
3-body encounters
LMXB
2-body encounters
msec pulsars
Stellar velocities
blue stragglers
Anisotropy
BHB stars
Gravothermal oscillations
Abundances
CM diagrams
External
Colours
Effects Black holes
Tidal truncation Stellar evolution
Hierarchical triples
Disk shocking
Bulge shocking

Fig. 1. Some of the problems of globular star clusters, and the interrelations
between them. Arrows between topics on the same side of the diagram are not
shown. Not all arrows imply cause and effect; the “stellar velocities” to “black
holes” means that the former could be an observational indicator of the latter;
in some cases the link is indirect. The diagram is considerably simplified: in fact
the mass spectrum affects almost everything on the left, and any of the first six
categories on the left could affect the CM diagram. Acronyms include: low mass
X-ray binary (LMXB), blue horizontal branch stars (BHB stars, also known as
“faint” or “hot” horizontal branch stars) and colour-magnitude diagrams (CM
diagrams).

therefore play a vital role in the evolution of star clusters. The term ‘ecol-
ogy’, introduced by Heggie (1992), captures the essence of this interplay.
We have already seen one way in which the evolution of the stars affects
the dynamics of a cluster. Towards the end of a star’s life on the main
sequence it loses much of its mass, and, as many stars complete their evo-
lution, the resulting unbinding of the cluster may even cause its complete
dissolution (Chapter 11). Other influences of stellar evolution on stellar
dynamics may be more subtle. According to point-mass dynamics, for
example, a hard binary tends remorselessly to get harder (Chapter 19f).
When we bear in mind that stars have finite radii, however, and that these
radii increase dramatically during certain phases of stellar evolution, the
resulting shrinking of the binary orbit cannot continue indefinitely. This
304 30 Stellar and Dynamical Evolution

may limit the energy which each binary may release, and its potential
importance for dynamical evolution.
Astronomers were also slow to realise the existence of links in the op-
posite direction, i.e. the role of dynamics in shaping stellar evolution.
But the evidence that stellar evolution in star clusters was different from
stellar evolution elsewhere gradually built up, and forced a paradigm shift
that is still not complete.
The earliest evidence of environmental factors in stellar evolution was
the discovery almost half a century ago by Sandage (1953) of “blue strag-
glers”, stars that seem to be substantially younger, and hence bluer, than
the bulk of the stars in a star cluster (Fig.2). The implications of this
discovery were not absorbed by the community at large, and for the next
twenty years, globular clusters still kept most of their secrets hidden.
Both authors of this book remember how, when we took our first course
in astronomy, globular clusters were portrayed as some of the oldest and
dullest objects in astrophysics. Even as we began our research careers,
the dynamics of star clusters was regarded as a quiet backwater in as-
trophysics, while all the excitement was being generated elsewhere; in
high-energy astrophysics, for instance. Paradoxically, it was precisely
from this direction that attention was once more directed to these “old
and boring” objects. In the mid seventies star-cluster research received an
enormous boost from the unexpected discovery of globular cluster X-ray
sources (cf. Fig.1.3).
It soon became clear that these sources had a much higher relative
abundance in globular clusters than in the galaxy as a whole (see Katz
1975, Hut & Verbunt 1983, Bailyn 1996). While the family of globular
clusters contain less than 0.1% of the stars in the galaxy, they harbor
more than 10% of the low-mass X-ray sources throughout the galaxy. The
search was on for an explanation, and an early suspect was what makes
globulars so special, dynamically: their high density. Various scenarios
were proposed, including binaries formed in stellar collisions (Sutantyo
1975), but quickly the dominant explanation arose that neutron stars had
been tidally captured by normal stars (Fabian et al. 1975), to form new
binary stars. In these binaries, mass overflow from the normal star could
then provide the fuel to turn the neutron star into an X-ray source.
More than ten years later, from 1987 onwards, another exciting type of
object was discovered in globular clusters, in rapidly growing numbers:
pulsars, rapidly spinning neutron stars that are discovered through the
pulsed radiation they emit (Fig.3; Kulkarni & Anderson 1996). Many
of these pulsars were found to be a member of a binary system, and
 30 Stellar and Dynamical Evolution 305

Fig. 2. A synthetic colour-magnitude diagram for an old stellar population (after

Portegies Zwart et al. 1997). The simulation starts at 10Gyr with a population of
degenerate stars (neutron stars and white dwarfs), and main sequence stars with
a relatively flat mass function, but no binaries. The results of stellar evolution
and collisions are shown at about 12Gyr. Besides the main and giant sequences,
there is a substantial population of blue stragglers (to the left of the turnoff point)
and yellow stragglers (to the left of the giant branch). For further examples of
synthetic “observational” data, see Portegies Zwart et al. (2001).

tidal capture was once again proposed as the main mechanism for the
formation of this other class of unusual binaries.
Soon afterwards, it is true, two important aspects challenged the use
of tidal capture as a panacea to explain any type of weird star system
in globular clusters. On theoretical grounds, closer investigation of the
mechanism of tidal capture cast doubt on its efficiency (mainly Kochanek
1992, but see also Mardling 1995, Kumar & Goodman 1996). On obser-
vational grounds, however, a new mechanism presented itself naturally,
based on the discovery of large numbers of primordial binaries (Chapter
24).
Still, the shift in outlook was permanent. It was now clear that one
could not understand the kinds of stars that were found in globular clus-
ters without taking the stellar environment into account, and, equally,
that the study of dynamics in isolation from stellar evolution led, at best,
to an incomplete picture. Nor was the evidence for this confined to exotic
kinds of star: even the relative abundance of the most ordinary stars ap-
peared to depend on the stellar density, causing observational variations
306 30 Stellar and Dynamical Evolution

Fig. 3. A high-energy astrophysicist’s view of the core of M15 (from Kulkarni

& Anderson 1996). Eight pulsars are indicated with letter designations, and one
low mass X-ray binary (AC211). Compare Fig.1.2.

in the colour of a star cluster from centre to halo (e.g. Djorgovski et al.
1991). What we shall do in the next two chapters is to describe some
aspects of this interplay between dynamics and the stellar nature of the
bodies involved.
Despite the weakening of the original motivation for considering tidal
capture, we begin there in the next chapter. This process is one of the
simplest, best studied and most robust aspects of cluster dynamics in
which the finite radii of the stars must be taken into account. Tidal
captures and stellar collisions between single stars do take place at sig-
nificant rates in globular star clusters, and are also an important aspect
of interactions involving binaries, though by no means the whole story.
Those interactions are the subject of Chapter 32.

Problems
1) In a system in which all stars have solar mass and radius, and the
(three-dimensional) velocity dispersion is 20kms−1 , compute the en-
ergy of the most energetic binary, in units of 1kT. (The solar radius is
about 7 × 105 km.) Compare the energy with that of a whole globular
star cluster.
2) Consider the arrows in Fig.1, and the mechanisms by which these
influences occur.
 31
Collisions and Capture

Once upon a time it was thought that novae arose from stellar colli-
sions. In a normal galactic environment, however, where most novae are
found, collisions are rare, in the sense that a given star is very unlikely to
experience a collision during the age of the universe. The stars are sim-
ply too far apart. In very dense environments, however, such as galactic
nuclei, collisions may even be sufficiently frequent to dominate dynami-
cal evolution. Globular clusters lie in between these extremes. In them,
collisions do affect large numbers of stars, but are not normally thought
to predominate over purely gravitational interactions (i.e. two-body re-
laxation). And the restriction to single stars is an unrealistic idealisation.
Globular clusters possess large populations of binaries, which automati-
cally place stars in close proximity. All in all, it is estimated that there is
a collision between two stars somewhere in the universe every ten seconds.

The rate of collisions

For a stellar system consisting of stars of equal mass m, radius R and
number-density n, the collision time scale (whose reciprocal gives the rate
at which a given star collides with any other) is given approximately by

t−1
coll = 8πGmnRhV
−1
i, (1)
where V is the relative speed of the two stars, and the average is taken
over all stars that may be encountered. (This formula is based on Ke-
plerian relative motion of the two stars, and assumes that gravitational
focusing is large, for which the condition is that V 2  2Gm/R; this is
satisfied in practice in globular star clusters.) For stars of solar type, the
formula may be reexpressed more intelligibly as
  
n km/s
t−1 ' 2 × 10 −15
yr−1 .
coll
pc−3 V

307
308 31 Collisions and Capture

That very small coefficient shows why collisions are so rare in the neigh-
bourhood of the sun, where n ∼ 1pc−3 and V ∼ 20km/s. In globular
clusters V is not so different, but the density is much higher, typically by
a factor of about 104 at the cluster centre. Hence the collision time scale
is reduced to something of order 1012 yr. Since a cluster lives for roughly
1010 yr, this means that of order 1% of stars in the core of the cluster will
have experienced a direct collision within its lifetime.
No one number can cover the complexity of this issue. In the first place
the collision timescale varies with the stellar radius and mass. Almost
all stars in globular clusters have smaller radii and masses than that
of the sun, and so the collision rate is smaller than we have estimated.
The main exception is the evolved giant stars, whose radii may exceed
that of the sun by a factor of 100. Collisions are less effective than
one might think, because of their low-density envelopes. Even so, in a
collision at 10km/s between a star of mass 1M and a giant of mass 2M ,
the maximum distance of closest approach for capture is almost as large
as the radius of the giant (Bailey & Davies 1999). More importantly,
this phase of a star’s life is relatively short, roughly 15% of its entire
lifetime (Eggleton et al. 1989), and so the contribution of these stars is
not dominant. For example, in models of the star clusters 47 Tuc and ω
Centauri, Davies & Benz (1995) have estimated that the rate of collisions
between two main sequence stars exceeds the rate of collisions between a
main-sequence star and a red giant by at least a factor of two. Such results
depend on the adopted mass function, but the required computations are
easily performed using a simple extension of Eq.(1). Furthermore, there
are now convenient simple formulae for the evolution of single stars, of
any relevant mass, metallicity and age, including formulae for the radius
(Hurley et al. 2000).
The second complicating factor is that the collision timescale depends
on the stellar density. This is smaller outside the core, but on the other
hand there are more stars there. Still, collisions are a two-body
R process,
and so the number of collisions is roughly proportional to ρ2 r2 dr, where
ρ(r) is the density at radius r. For any reasonable density profile this
integral is dominated by the vicinity of the core (cf. Problem 3).
A third point for consideration is the fact that the stellar density and
mass function, and therefore the collision rate, vary enormously because
of the dynamical evolution of the cluster, especially during collapse of
the core (e.g. Statler et al. 1987). The total number of collisions is given
by the time integral of the rate, and an argument along the lines of our
discussion of the spatial variation of the collision rate shows that late
core collapse does not dominate the total number of collisions (Sugimoto
1996); though the collision rate is very high then, this phase is over
too quickly for this to matter. On the other hand mass segregation is
 The rate of collisions 309

a more-or-less permanent consequence of dynamical evolution, and in

young, compact clusters (at least), it has a major effect on the collision
rate (Portegies Zwart et al. 1999).
Fourthly, the collision rate varies very much from one star cluster to
another. It has been found that the largest collision rate among the
galactic globular clusters is roughly 104 times larger than the smallest
(Hills & Day 1976). Put another way, half the collisions in the entire
cluster system of the galaxy (about 150 clusters) occur in just about six
clusters (Verbunt & Hut 1987, Johnston & Verbunt 1996). Such facts
are important in trying to understand why the abundance of possible
products of collisions (cf. Chapter 30) varies so much from one cluster to
another.
The discussion so far is typical of much discussion of the dynamics of
globular clusters in assuming that all stars are single, forgetting that it
has now been established that a sizable proportion are binaries (Chapter
24). The existence of binaries immediately puts stars in close proximity
and it is obvious that this may enhance the rate of collisions. There are
at least two mechanisms by which the components of a binary may be
brought into contact.
The first mechanism is purely internal. We have already remarked that
stars expand greatly when they become red giants, and when this happens
to one component of a binary it is very likely that the two components
will touch. More precisely, what happens is that the material at the
surface of the large component exceeds the “Roche radius”, at which
point it is more strongly attracted to its companion. The dynamics is
very similar to escape of stars in the tidal field of the galaxy. Fig.12.1
shows the Roche surface of the cluster in this case. For stars, however,
the process is referred to as Roche lobe overflow. It is really one aspect
of the incredibly complicated problem of stellar evolution in the context
of binary stars, which will be treated in the next chapter.
The second mechanism is external. When a binary is disturbed by a
passing star, its eccentricity may increase so much that the new pericen-
tric distance of the components is less than the sum of their radii. An
ordinary fly-by gives the intruder one chance to work on the eccentricity
of the binary, and the probability of a resulting collision need not be high.
If, however, the intruder is temporarily captured by the binary in a reso-
nance (Chapter 19), it is as if the eccentricity is constantly changing; for
this reason, during such an encounter the minimum distance between two
stars (and it may be attained by one binary component and the intruder
as easily as by the two components) is about 1% of the initial semi-major
axis of the binary (Hut & Inagaki 1985). Thus stars in binaries with
semi-major axes less than about 100 stellar radii are seriously at risk. If
310 31 Collisions and Capture

a hard binary with a large red giant forms a resonance, coalescence is

almost certain, since the initial semi-major axis cannot be large enough.
Now it is obvious qualitatively that this is a major source of collisions.
The time scale for such encounters between a binary and a third star is
much shorter than that for collisions between two single stars, because
the radius (semi-major axis) of a binary so much exceeds the radius of
a single star (cf. Eq.1). This more than compensates for the fact that
the proportion of stars in binaries of the required size is not very large.
Besides, mass segregation increases their concentration in the core, and
therefore their interaction rate.

The effect of collisions

What happens to stars when they collide? It is obvious from everyday

experience that speed has something to do with it. Place an egg on the
kitchen table and you have a useful egg; drop it on the table and you
have a mess.
If two stars, again of mass m and radius R, are dropped onto each
other, i.e. we neglect their relative speed when far apart, then at the
moment when they touch the kinetic energy of their relative motion is
mV 2 /4 = Gm2 /(2R), where V is their relative speed at the instant when
they come R
in contact∗. Now the gravitational binding energy of a single
star is G (M/r)dM , where M is the mass within radius r. For a star of
uniform density we have M ∝ r3 , and so the integrand varies as M 2/3 .
3 Gm2
Therefore the binding energy is . Comparison with the energy of
5 R
relative motion shows that, in a hypothetical star cluster with solar-type
stars, there is not enough kinetic energy to unbind even one star, let alone
two. Furthermore in a normal star the density increases inwards, and so
the binding energy required would be even larger.
Thus direct collisions in globular clusters cannot destroy stars. Instead,
the two stars fuse, after a period of violent internal motions (Fig.1).
Beyond this broad result, the details of the merged remnant depend on
the nature of the stars involved, and are somewhat controversial. If one
star is a white dwarf, for example, by its high density it sinks more-or-
less unscathed to the centre of its fellow-victim (see, for example, Ruffert
1992). Another possible kind of collision product, considered in the next
chapter, is a blue straggler. Quite general recipes for the collisions of a

∗
The factor of 4 on the left-hand side results from half the reduced mass of two stars
of equal mass; cf. Chapter 19.
 The effect of collisions 311

Fig. 1. Collision between two main sequence stars (after Sills et al. 2001). Ma-
terial in different layers of the stars is coarsely distinguished by different shades
of gray.

variety of common types of star are given, quite independently, by Tout

et al. (1997) and Portegies Zwart et al. (1997).
In the previous discussion we were concerned with a head-on collision,
as the stars were dropped together from a great separation. In general,
however, the stars will not be aimed directly at each other. If they still
physically touch, and if we assume that they still merge, it is clear that
one consequence will be that the remnant is rapidly spinning, rather like
slicing a golf ball, or putting spin on a snooker ball.
What, though, if the two stars do not even touch when at their closest
approach? They can still influence each other gravitationally. To under-
stand what happens it is instructive to consider a famous old problem of
dynamical astronomy which was first understood by Newton, that of the
tides raised on the oceans of the earth by the sun and moon.
Let us consider just the lunar tide, for simplicity. There are two tides
each day at most locations. To a first approximation, there is one high
tide when the moon is high in the sky, and another about 12 hours later.
In fact it is unusual for one of the high tides to coincide with the time
when the moon is highest in the sky. In order for the tide to rise under
312 31 Collisions and Capture

the moon, the water of the oceans must move across the bed of the ocean.
The resistance to this motion is especially strong near narrow channels
in shallow seas, as in the English Channel separating France from the
British Isles. This motion also takes time, and so the maximum height
of the water takes place some time after the moon is highest in the sky.
There is an analogy here with a mildly damped harmonic oscillator which
is forced at a frequency much smaller than its natural frequency. The
forced oscillation lags slightly behind the forcing.
Now return to the encounter between two passing stars, and let us think
of one star as taking the place of the earth with its oceans, while the other
plays the role of the moon. Again the second star raises tides on the first,
and again the high tide at some location on the first star occurs some
time after the second star has passed above that spot. Looking down at
the first star, the second star sees a nearly spherical object, except for
the tides he has raised. The nearer of these is behind him, in the sense
that he has to look in the opposite sense from his orbital motion to see it
(Fig.2). Now this tidal material exerts a weak gravitational pull on the
second star, in the sense opposite to his orbital motion around the first
star. Therefore it slows down his orbital motion.
What this argument shows is that the tidal interaction between the
two stars removes energy from their relative motion. It is clear that most
of it is removed around the time of closest approach. We are also now
imagining that the two stars have approached from a large distance where
they had a non-zero relative speed, of order 20km/sec, say, and therefore a
non-zero kinetic energy of relative motion. If this kinetic energy exceeds
the energy removed tidally at closest approach then the two stars will
eventually separate again. If the reverse is true, however, the two stars
are subsequently bound in a binary. This happens if the loss of energy is
large enough, i.e. if the encounter is sufficiently close. Tidal capture has
occurred.
What happens next is somewhat controversial. Certainly, because the
two stars are bound to each other, further close encounters will occur.

Fig. 2. The tidal bulge created by a passing star. Distortions of star 2 are not
shown. See also the upper right panel of Fig.1.
 The effect of collisions 313

What effect these have depends on whether the tide raised by the first
encounter has had time to subside (Kochanek 1992, Mardling 1995). If
not, at the time of the next close approach the old tide could be in advance
of the passing star, and then its effect is to unbind the pair; at least that
would be the sense in which the interaction worked, though it might not
be strong enough to succeed. In this scenario, the newly formed binary
enters a phase of chaotic evolution, sometimes becoming less bound and
sometimes more bound. The rotation of the stars presumably further
complicates the interaction, however (see, for example, Witte & Savonije
1999).
At the same time the internal structure of the stars is being altered.
The tidal interaction (except in some chaotic phases) is extracting energy
from the relative motion of the two stars and transferring it to the tide.
The dissipation of each tide further contributes this energy to the internal
energy of the stars, which should expand somewhat in consequence (cf.
McMillan et al. 1987, Tout & Kembhavi 1993). This clearly has the
effect of enhancing the strength of the tidal interaction, and of leading to
a situation where the two stars can be brought into contact. The upshot
is that a single tidal encounter between two stars, even if it did not lead
directly to physical contact between the two stars, can eventually lead
to their merger. In other words the time scale for encounters leading to
mergers is somewhat shorter than the collision timescale given by Eq.(1).
The numerical size of the enhancement is, however, not large, as the
effects of tidal dissipation decrease as a rather high power of the distance
of closest approach (Problem 4).

Problems
1) Derive Eq.(1).
2) Two stars of mass m and radius R have relative speed V initially,
when they are far apart. Estimate the mass loss in an encounter, by
computing how much mass can be ejected (from the surface of one star
to infinity) by the initial energy of relative motion of the two stars.
(The result is of course zero if V = 0, in contradiction with results of
numerical simulations; cf. Table 2 in Lombardi et al. (1996).)
3) In an isothermal model (Chapter 8) the density at large radius varies
approximately as ρ ∝ r−2 . Show that the rate of collisions is domi-
nated by the contribution from the core.
More precisely, recall that the scaled potential φ satisfies the equa-
tion
2
φ00 + φ0 = 9 exp(−φ),
r
314 31 Collisions and Capture

where r is the radius in units of the core radius and φ is assumed to

vanish at r = 0 (Eq.(8.6)). In units of the central density, the density
at a point where the potential is φ is exp(−φ). In these units show
(by numerical integration) that the integral
Z ∞ 4π
4πr2 ρ2 dr ' × 0.55.
0 3
(In general the result is ρ2c rc3 times this factor, where rc is the core
radius and ρc is the central density; Spitzer 1987, p.149.)
4) A star of radius r is disturbed by another star of equal mass passing
on a parabolic orbit at a pericentric distance R. Assuming that the
surface of the distorted star is an equipotential of the tidal field (cf.
Eq.(12.4)), show that the relative difference between its largest and
smallest radii is δr/r ' 6(r/R)3. By treating the distortions as two
small equal masses added to the star, deduce that the work W done
on the passing star varies as W/E ∝ (r/R)8 , where E is the binding
energy of one star.
 32
Binary Star Evolution
and Blue Stragglers

Once thought to be virtually devoid of binaries, globular clusters are now

known to contain binaries in abundances not very much less than that
of the galactic disk. Binaries in such large numbers, containing at least
ten percent of the stars in a typical globular cluster, cannot have resulted
from dynamical interactions, and therefore must have formed at the same
time that the bulk of the stars were formed. With so many binary stars
around, all kinds of interesting reaction channels are possible, in three-
body as well as four-body interactions. Binaries containing pulsars are
just one example of the unusual objects that can result.
Even without dynamical interactions with other objects, binary star
evolution in isolation is quite complicated enough. Compared to the
evolution of single stars, a wide variety of new kinds of binaries and
single stars can be created, through mass overflow from one star to the
other, or through mass loss from the system, at various stages in their
combined evolution. The stars can form a common envelope for a while,
or one of the stars can explode as a supernova. Even if the explosion
is symmetric the binary may or not survive, as in the impulsive loss of
mass in any stellar system (Chapter 11). Disruption is even more likely
if the the remnant receives a “kick” (see Hills 1983), and if you find a
neutron star in a binary in a star cluster it is likely to have got there by
dynamical interactions (e.g. Kalogera 1996).
Complicated as such processes are they have been studied for many
years, and much is understood. Suitable introductions can be found, for
example, in Iben (1985) and Iben & Tutukov (1997, 1998). For our pur-
poses, what is even more useful is that those aspects of binary evolution
that are reasonably well understood can be codified as relatively simple
recipes (Portegies Zwart & Verbunt 1996, Tout et al. 1997). What would
be even nicer is some confirmation that these recipes produce the same
answers.

315
316 32 Binary Star Evolution and Blue Stragglers

Even without internal evolution of the components, tidal interactions

between them can affect the dynamics of a binary. In the same way as
tidal interactions occurring in close encounters of single stars (Chapter
31), these extract orbital energy without changing the angular momentum
of the binary, and the result is circularisation of the binary orbit (Problem
1). The spins of the stars are also affected. Despite the low density in
the envelope of a giant, these processes are especially important when the
stellar radii are large.
The interaction between dynamical and stellar evolution of binaries is
surely destined to be one of the growth areas in this whole subject in
coming years. As yet few definitive results have emerged, and so our
treatment serves only to give the flavour of what is possible. We concen-
trate on blue stragglers, as these occur in relatively great abundance, and
even in small star clusters which are within our computational grasp at
the present time.

Recycling Stars: An Example

In the dense cores of globular clusters, a typical star has a significant
chance to undergo a collision during a Hubble time (Chapter 31). Since
the velocity dispersion in globular clusters is one or two orders of magni-
tude less than the escape velocity at the surface of a typical main sequence
star, almost all of the mass of two colliding stars is retained to form a
merger product (Problem 31.2).
In a collision between two stars of relatively low mass, the merger rem-
nant may have a mass below that of the main sequence turn-off, 0.8M .
In that case, it will be rather difficult to distinguish the merger product
from an ordinary star of that mass. For a given candidate, detailed inves-
tigation might show unusual abundances, but spotting such stars against
the background of normal stars would be akin to seeking a needle in a
haystack.
In a collision where the sum of the masses of the original stars signif-
icantly exceeds that of the turn-off, by being larger than 1M say, the
merger remnant will stand out in a HR diagram, since it will be posi-
tioned on or near the main sequence, but bluewards of the turn-off. Such
a star, if born at the same time as the rest of the stars of the cluster,
should have evolved away from the main sequence. Its presence thus gives
the impression of having straggled, and with its relatively bluer colour
such a star is called a blue straggler (cf. Fig.30.2).
The presence of such stars in old star clusters has been known for a
long time. Now they are studied so intensively that a whole conference
has been devoted to them (Saffer 1993; see also the review by Leonard
[1996]). Even on observational grounds alone there is evidence that there
 Recycling Stars: An Example 317

is more than one way to make a blue straggler (Ferraro et al. 1993), and
theorists are fortunately not short of ways of doing so too.
Many, if not most, of the collisions between single main sequence stars
in star clusters are likely to produce blue stragglers. One reason is that
mass segregation will produce an excess of heavier stars in the central
regions of a cluster. Other reasons are the larger radii of heavier stars,
together with their enhanced gravitational focusing, both of which favour
collisions for heavier, rather than lighter stars.
An even more efficient mechanism for inducing collisions is provided
by binary–single-star and binary–binary encounters, especially those in-
volving hard binaries, which have a significant chance to lead to resonant
encounters where three or four stars are temporarily captured in a small
area of space. Similar arguments again predict a larger collision probabil-
ity for heavier stars. A special feature of 3-body or 4-body channels for
blue straggler formation is that the mass of the final blue straggler may
well be significantly larger than twice the turn-off mass, if more than two
stars are involved in producing a single merger (Fig.1).
A third way of using binaries to produce blue stragglers is through mass
(Roche lobe) overflow from one component of a suitable binary onto the
other member. This process comes in several flavours: coalescence, and
mass transfer of various kinds (Leonard 1996). None of these mechanisms,
however, require a high density of neighbouring stars, and so they are not
special to the million-body problem.
As a simultaneous illustration of the last two types of blue straggler
formation events, a calculation by Simon Portegies Zwart is reproduced in
Fig.2. The calculations were performed in the Starlab environment, using
a version of the automated three-body scattering package described in
Chapter 22, integrated with a simplified recipes for binary-star evolution.
There is nothing contrived about this particular set of initial condi-
tions. Both the single star and the binary were drawn at random from
a prescribed cluster core population of single stars and binaries. At each
timestep of a thousand years, the probabilities for close encounters be-
tween different objects were calculated, and actual collisions were carried
out in Monte Carlo fashion by spinning a random number generator to
see which stars would actually undergo an encounter. At each timestep,
those stars and binaries that did not undergo an interaction (the vast
majority) were simply evolved with stellar evolution prescriptions: fit-
ting formulas for stellar evolution tracks in the case of single stars, and
simplified recipes in the case of binaries (for more details, see Portegies
Zwart 1996, Portegies Zwart & Verbunt 1996).
318 32 Binary Star Evolution and Blue Stragglers

Fig. 1. Collision between two binaries (after Goodman & Hernquist 1991).
Eventually all four stars merge into a single massive and extended stellar ob-
ject.

Fig.2 shows an example of an interplay between three stars which even-

tually leads to the formation of two blue stragglers. The first blue strag-
gler is formed in isolation, as the result of binary star evolution (Roche
lobe overflow), nearly 10 Gyr after its formation. At this time, the 1M
star attempts to climb the giant track. The first result is that its increased
size leads to rapid tidal circularisation, erasing the initial eccentricity of
0.3 around T = 9.97 Gyr. Soon thereafter, the primary fills its Roche
lobe, and around T = 9.98 Gyr dumps most of its mass on the secondary.
However, only 0.1M of this supply of matter can be accreted onto the
companion, and 0.5M leaves the system. The reason for the inefficiency
of mass transfer is that the donor star provides mass on its own thermal
time scale, while the other star in turn can only accept mass on its thermal
time scale, which is longer. By T = 9.99 Gyr, the primary turns into a
low-mass helium star which then cools to become a white dwarf. From
this time on, the secondary shows up as a blue straggler: it resembles a
main sequence star, but has a mass and luminosity greater than that of a
 Recycling Stars: An Example 319

Fig. 2. Evolution of a binary involving one triple interaction (from Portegies

Zwart 1996). For an explanation, see the text.
320 32 Binary Star Evolution and Blue Stragglers

turnoff star. The secondary now evolves rather more quickly, but before
it can reach the turn-off, a third star happens to pass through the system,
2 Gyr later, and is captured into a three-body resonance scattering event,
displayed with time moving horizontally from left to right, for a duration
of about a hundred years.
The outcome of the event, not surprisingly, leads to the lightest star
being ejected, in this case the white dwarf, the end product of the original
primary. The incoming 0.8M star now takes the place of the primary.
The orbit has tightened significantly, as could also be expected (hard
binaries tend to get harder; cf. Chapter 19). After another 2.4 Gyr, the
original secondary, saddled with the extra mass from the original primary,
at last begins to climb the giant track. This leads to a second phase of
tidal circularisation, in which the eccentricity induced in the three-body
encounter is erased. After another 15 Myr, mass overflow takes place,
again 0.6M is lost, of which this time 0.2M is accepted by the star
which entered as a result of the exchange. This increases the mass of the
latter from 0.8M to 1.0M , thereby turning this star into a second blue
straggler. The original secondary soon turns into a dwarf, orbiting the
1M star. This second straggler phase finally ends around T = 17.8 Gyr,
during the last phase of giant evolution.
The final state is not displayed in this figure as it takes place in the
future: 17.8Gyr exceeds almost all estimates of the age of the universe.
But we can guess what will happen. After the second blue straggler fills its
Roche lobe, the binary will shrink significantly, and most likely a common
envelope system will be formed. This is likely to lead to the merger of the
two cores, resulting in a late type giant of 1.3M . A bit later, a white
dwarf will then be left behind, with a mass of around 0.7M .
Observationally, the first blue straggler formed by the transfer of mass
in a close binary will not be distinguishable from those formed by the
“classical” route, i.e. in collisions between single stars (Fig.30.2). Only
after the dynamical exchange interaction does the system become clearly
different from “ordinary” blue stragglers: the observation of a blue strag-
gler in an eccentric orbit or in a detached binary with a main-sequence
companion is a direct indication for formation by an exchange reaction.

The Binary Zoo

What we have seen is just one particular case study of the formation of
blue stragglers. While it includes a stellar environment (which provided
the interloper for the three-body interaction), there is no surrounding
star cluster undergoing the many dynamical processes described in other
chapters of this book. Such studies are now under way, and confirm
 The Binary Zoo 321

that interactions involving both single stars and binaries are routes which
greatly enhance the proportion of blue stragglers.
An example is the old open cluster M67, for which Hurley et al. (2001)
have constructed an N -body model which includes the full panoply of
stellar and dynamical evolution. Though open clusters are often regarded
as rather low-grade fare in the world of N -body simulations, a 15000-body
simulation was needed to simulate the last half of its evolution. Twice
that number might have been needed to cover the entire evolution, which
would not have been feasible at present for a stellar system with a rich
complement of active binaries. Blue stragglers are present in M67 at an
abundance roughly 10 times higher than would occur in a population
of isolated binaries and single stars. As this simulation showed, it is the
dense environment in M67 which accounts for its wealth of blue stragglers.
Blue stragglers are just one of the unusual types of object that cannot
form through single star evolution, but require either the presence of pri-
mordial binaries, or the interaction between two, three, or more stars in
a dense stellar system. In addition, X-ray sources, pulsars, spectroscopic
binaries and eclipsing variables have been detected in significant numbers,
some of them right in the middle of the densest central regions of globular
clusters. This diverse collection of objects forms a gold mine for dynam-
icists modelling the evolution of globular clusters. Another metaphor,
coined by Melvyn Davies (Davies 1995) is also apt: the zoo of possible
objects that can be formed in few-body interactions is bewildering, and
rather little has been studied yet in this field — for those daring to brave
this jungle, many species still remain to be discovered.
While all these binaries and binary products provide useful diagnostics
of binaries and their interactions, some of them are even directly involved
in the physical processes of energy generation which drive the expansion
of the core after core collapse (Chapters 27–29). In this sense we are
beginning to get a direct look at the central engines which power the
later phases in the evolution of globular clusters.
Here is a final conceit. One may say that star cluster dynamics is
like a ballet with about a million performers, but it is the binary stars
which take centre stage. It is in their pas de deux that most of the action
develops, and it is through their interactions with other players, in the
form of occasional pas de trois and pas de quatre, that fascinating new
patterns (and even new characters) arise.

Problems

1) A binary evolves at constant angular momentum but decreasing or-

bital energy. Deduce that the eccentricity decreases.
322 32 Binary Star Evolution and Blue Stragglers

2) A stellar system has n single stars per unit volume, all with radius R.
Assuming that two stars merge if and only if their closest distance is
less than 2R, show that the rate of formation of mergers in two-body
encounters is 4πGmRn2 h1/V i, where V is the relative speed of the
two stars.
The system also has f n hard binaries per unit volume, all with semi-
major axis a, and all components have radius R. The cross section
for binary-single encounters in which at least two participants come
within a distance d is known to be approximately
   0.4
πa2 3 Gm d
σ ' 8.5 2
V 2 a a
(Hut & Inagaki 1985). Compute the rate of mergers in binary-single
encounters, and compare with the rate of single-single mergers.
 33
Star Cluster Evolution

It often happens in science that progress is made, not so much by the

discovery of new facts, but by organising known facts in a new way. An
illustration from chemistry is the periodic table, and in astronomy the
HR (Hertzprung-Russell) diagram is a perfect example. In its modern
form this diagram, now called a CM (colour-magnitude) diagram, is a
scatter plot of the luminosity or absolute magnitude of a sample of stars
plotted against their colour (cf. Fig.30.2). It is an immensely powerful
tool, so familiar that its power is taken for granted, and it is invaluable
for studying the evolution of stars. In this chapter our aim is to provide
something comparable for star clusters, though our goal is the much more
modest one of helping the reader to grasp in a few pictures the essentials
of what has been described in greater detail in earlier chapters of this
book.
The links between many of the dynamical processes we have discussed,
and one or two others, are summarised in Fig.1. Centre stage is mass loss,
which occurs through the agency of several processes. Meanwhile, as the
system is losing mass and heading towards oblivion, various processes are
also causing its internal structure to evolve. Chief among these is two-
body relaxation, which causes more massive stars to segregate inwards,
and the core to collapse. Both mechanisms enhance the importance of
interactions between primordial binaries, which eventually bring the col-
lapse to a halt. Interactions between pairs of stars, and even triples and
quadruples, also affect the kinds of stars we find in clusters, and the way in
which they evolve. For astronomers, one of the interesting consequences
is that the mass function changes. Often it is said that one simply needs
sufficient destruction mechanisms (such as disk shocking) to achieve this.
But disk shocking does not care about the mass of a star, and can only

323
324 33 Star Cluster Evolution

Bulge and Violent Dynamical

Disk relaxation Friction
Shocks ch.10 ch.14
ch.12

Tidal Cluster
Overflow Death
ch.12 ch.29,33

Mass
Loss Mass Function
from Evolution
Stellar
Evolution Cluster
ch.11,30-2

Mass
Segregation
ch.16
Two-body Collisional
Encounters Relaxation
ch.14
Core
Evolution
ch.17,18,27,28
3- and 4-Body Binary
Encounters Heating
ch.19-22,25 ch.23,26

Fig. 1 Dynamical processes in the evolution of the million-body problem.

influence the mass function if two-body relaxation has already segregated

stars of different mass.

Structural Evolution
Next we shall try to picture how the system evolves under these processes,
though we shall assume that certain of them are dominant. In particular,
we imagine the evolution of a star cluster bathed in the tidal field of a
parent galaxy and dominated by the following dynamical processes. First
is the internal evolution of the stars, which causes them to lose mass; this
weakens the potential well of the cluster, which may lead to much further
loss of mass. Next, there is two-body relaxation, which leads to mass
segregation and collapse of the core. Finally, collapse is arrested by the
intervention of primordial binaries; the cluster enters the final phase of
its life, continually shedding mass across the shrinking tidal boundary.
Throughout, it is shaken intermittently by tidal shocks.
We shall summarise these results in a diagram which reduces the struc-
ture of the cluster to its characteristic radii: the core radius, rc , the half-
 Structural Evolution 325

mass radius, rh , and the tidal radius, rt . To reduce the problem to one
that can be plotted on paper, we take the ratios rh /rt and rc /rh . The
first of these tells us how liable the cluster is to tidal disruption, and the
second states how evolved it is dynamically.

Phase I: Stellar Evolution

We begin with the early evolution, and assume that the time scale of
relaxation is much longer than the time scale for major mass loss by the
internal evolution of the stars. Provided that the latter is longer than
the typical orbital periods of stars in the cluster, the cluster expands
nearly homologously, unless it is already tidally limited. If it is not tidally
limited, it behaves like an isolated cluster: rc /rh remains nearly constant,
but rt decreases a little with the loss of mass, and so rh /rt increases. If
the cluster is tidally limited then the core may expand while the stars are
losing mass, but rh and rt will contract because of the more copious loss
of mass across rt ; thus rc /rh increases while rh /rt changes little. These
two possibilities are illustrated in Fig.2a as cases A and B. There is also
an intermediate case (C), where the cluster is not quite tidally limited
initially, but becomes so because of the expansion (Engle 1999).
We can place some numbers on this diagram. At the point where a
cluster is tidally limited, the value of rh /rt is roughly 0.4 (Giersz & Heggie
1997). Also, how much evolution there is depends on the proportions of
stars of different mass: if there are few stars of high mass, there is little

rc C
_
r
h A B

B´
A´

B˝
A˝

rh
_
rt

Fig. 2a Schematic evolutionary tracks of star clusters. See text.

326 33 Star Cluster Evolution

0.6

0.5

0.4
rc/rh

0.3

0.2

0.1

0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35
rh/rt

Fig. 2b. Evolutionary track of an N -body simulation (N = 4096). It starts

from a King model with W0 = 5 whose limiting radius is 0.25 times the tidal
radius. The initial mass function is a power law n(m)dm ∝ m−2 in the range
0.1M < m < 15M . Stellar evolution uses the simple prescription introduced
by Chernoff & Weinberg (1990). Plotting stops when the mass inside the tidal
radius is less than 5% of its initial value.

evolution, and the cluster expands by as little as 10%. If there are many,
the cluster may dissipate altogether. Roughly speaking, the fraction of
mass lost also gives the factor by which a cluster would expand, though
if it starts life nearly tidally limited then the mass loss may well result in
the death of the cluster by tidal dissipation. What this fraction is may be
seen in Fig.11.1, though this completely ignores loss of mass by expulsion
of gas, which is likely to dominate the first few million years.

Phase II: Dynamical Evolution

Eventually (and in quite a short time compared with the ages of glob-
ular clusters) the rate of mass loss caused directly by the evolution of
individual stars becomes negligible. Those clusters that survive now en-
ter a phase of evolution dominated by tidal stripping, relaxation, mass
segregation, and core collapse, which means that rc /rh decreases. If the
cluster already fills its tidal radius, it evolves with little change in the ra-
tio rh /rt (which can hardly get bigger without the total dissolution of the
cluster) but loses mass relatively copiously. Otherwise, little mass is lost
in these phases, which mainly affect the core, and again rh /rt remains
virtually unchanged. The two clusters shown in Fig.2a now move along
 Destruction mechanisms 327

the evolutionary paths denoted by A0 and B 0 . The time scale taken for
this stage of the evolution depends on a number of factors. For a sys-
tem which starts this stage of the evolution at a point where it is tidally
limited the relaxation time scale is proportional to the number of stars
times the galactic orbital period at the location of the cluster. For other
systems, the relevant time scale for the evolution of the core is smaller,
i.e. the evolution of the core is faster.
As we have mentioned, stars escape as rc /rh drops. If the cluster lasts
long enough, however, and assuming that it was born with a sufficient
proportion of primordial binary stars, these begin to release energy in
three- and four-body reactions when the mass inside the core has shrunk
to a certain fraction of the whole, i.e. when rc /rh is small enough. What
happens next depends on whether the cluster is living comfortably inside
its tidal boundary (case AA0 ), or already tidally limited (case BB 0 ). In
the second case the cluster sheds mass, evolving in a self-similar manner
first described many years ago by Hénon (B 00 ). Otherwise the release of
energy drives an expansion of the cluster until it joins its tidally bound
neighbour. This expansion is homologous out to the radius rh , and so the
evolution in our diagram is horizontal (A00 ) until the evolutionary track
arrives at B 00 .
Though these tracks are schematic, more realistic treatments confirm
the gross features. For example, Fig.2b illustrates an N -body simulation
on a diagram with the same axes. One can readily discern stages corre-
sponding to the phases A, A0 and A00 in Fig.2a, even though this cluster
is powered by 3-body binaries rather than primordial binaries. Stage A
is quite short.

Destruction mechanisms
Finally we would like to give a feel for the time scales on which all
this action takes place, with special attention to the lifetime of a stellar
system. A cluster is under attack both from the outside (tidal disruption,
disk shocking, etc.) and from its own internal diseases (stellar evolution,
two-body relaxation). To make matters simpler (without losing relevance
to the old star clusters around us, which must have survived this phase)
we shall ignore destruction on the time scale of stellar evolution, which
must have become ineffective long ago. It is a disease of infancy.
Essentially the task is to estimate the time scales of the main processes
which cause a cluster to lose mass. We ignore tidal overflow in the sense
that this is not a primary cause of loss of mass; it would seem quite
possible for a cluster to survive indefinitely in a tidal field if there were
no other mechanism causing loss of mass (Muzzio et al. 2001). It is true
that we have no reason to suppose that clusters are born in equilibrium
328 33 Star Cluster Evolution

with their tidal field, but again any outlying matter will be stripped on
a short time scale.
The first process to consider, and the one on which much emphasis has
been placed in this book, is two-body relaxation. For its time scale we
take the half-mass relaxation time (Eq.(14.9)), i.e.
1/2 3/2
0.138Mc rh
trh '
G1/2 m ln Λ,
where m is the individual stellar mass, and Mc , rh are the mass and
half-mass radius of the cluster, respectively. If we set this equal to the
ages of the typical old galactic globular star clusters we obtain a feel for
the gross parameters of clusters which can survive the destructive effects
of this process to the present day. The result (with some refinements) is
a line at the position of the lower left boundary in Fig.3, which is a plot
of cluster mass against half-mass radius.
The next process to consider is disk shocking (Chapter 12; see also
Gnedin et al. 1999). This does different things on different time scales.

R<12 kpc

R<8 kpc
R<5 kpc

R<3 kpc

Fig. 3. The survival of galactic globular clusters (from Gnedin & Ostriker 1997).
The masses of 119 galactic globular clusters are plotted against their half-mass
radius. Different symbols are used for clusters within different ranges of galac-
tocentric distance. The three stated destruction mechanisms are thought to
control their distribution. The key at top left refers to the assumed models for
the galactic potential and the distribution of the globular clusters.
 Destruction mechanisms 329

The overall energy of the cluster is altered on the time scale given in Box
12.2, Eq.(3). But the energies of individual stars are altered on the time
scale of shock relaxation, which may, like ordinary two-body relaxation,
lead to escape. Therefore we adopt a time scale of the form
Mc RV
tsr ∼ ,
GΣ2 rh3
where R is the galactocentric distance of the cluster, V is its orbital
speed round the galaxy, and Σ is the surface density of the galactic disk.
Setting this (with a suitable estimate of the numerical coefficient) equal
to the lifetime of the globular clusters gives lines at the positions of the
lower right boundaries in Fig.3. Note that the time scale depends on
location within the galaxy, especially on R: it decreases with decreasing
galactocentric radius, not only because of the explicit factor of R, but
also because Σ increases as R decreases.
Now we see that the observed clusters in our galaxy fit rather comfort-
ably into the pocket between these two constraints. They do not, it is
true, explain the absence of clusters of very high mass. This could well
be primordial, i.e. it may be that no such clusters ever existed. There
is, however, another mechanism that might have depopulated this part
of the diagram. We introduced dynamical friction (Chapter 14) as a pro-
cess which slows down stars moving through clusters, especially massive
stars. In the same way it acts on clusters moving through the galaxy, and

Fig. 4. A dying globular cluster: E3. Image obtained by use of NASA’s SkyView
facility (http://skyview.gsfc.nasa.gov) located at NASA Goddard Space Flight
Center.
330 33 Star Cluster Evolution

acts more quickly on more massive objects. Estimates show that it would
have dragged most clusters of high mass (above the upper boundary on
Fig.3) into the Galactic Centre. Together with the two processes of disk
shocking and two-body relaxation, this shows that there is a rather well
constrained triangle of survival in this diagram, where almost all of the
present-day clusters are found.
The present distribution of cluster masses is well explained by a combi-
nation of these dynamical processes and a suitably chosen initial condition
(e.g. a power law). It resembles a log-normal distribution (i.e. normal in
the logarithm of the mass), and can equally well be approximated by a
flat distribution at low masses and a power law at high masses (cf. Baum-
gardt 1998, Vesperini 1998 and references therein, Fall & Zhang 2001).
The radial distribution of clusters is also affected (e.g. Capuzzo-Dolcetta
& Tesseri 1999).
Estimates have been made of the rate at which clusters are dying
(Aguilar et al. 1988, Hut & Djorgovski 1992, Murali & Weinberg 1997,
Johnston et al. 1999). It is likely that about half of the present popu-
lation of the globular clusters in our galaxy will perish within the next
1010 years, which is about how long the present population has lasted.
There is nothing surprising in this, any more than in the fact that the
age of most individuals in the human population of the world is com-
parable with the span of a human life. The clusters which will expire
within the next billion years are sparse objects. Likely candidates are the
clusters E3 (Fig.4) and Palomar 13 (Siegel et al. 2001). Even the names
suggest that these are miserable objects, and not in the same class as the
brilliant clusters that made themselves known to Messier. Almost all of
the low-mass stars have left such an object, leaving behind the heavier
populations which escape least easily: binary stars, white dwarfs. Soon,
by galactic standards, nothing will be left.

Problems
1) Add the line in Fig.3 corresponding to the tidal boundary condition,
assuming that the rotation curve of the galaxy is flat, with rotation
speed 220kms−1 , and that rh < ∼ 0.4rt . (There is a different line for
each galactic radius R.)
2) Suppose that the relaxation time tr of a cluster is always proportional
to the time until its dissolution (Eq.(29.1)). Suppose the initial distri-
bution of tr is a power law f (tr ) ∝ t−α
r , with α ∼ 2, say. Show that
the distribution of log tr eventually evolves into a form resembling a
normal distribution.
 Appendix A
A Simple N -Body Integrator

N -body codes used in research run to many thousands of lines and

hundreds of subroutines. The two main collisional codes/packages are
1) Starlab See http://www.manybody.org
2) NBODYx A series of N -body codes of differing sophistication and
adapted to different hardware. See
ftp://ftp.ast.cam.ac.uk/pub/sverre/nbody6/
The codes that follow are much more modest. We give both Matlab and
C codes, but not a FORTRAN code (for which see Binney & Tremaine 1987).
Both codes may be found at http://www.manybody.org/codes.html .

Matlab
This code is presented mainly because of the integrated graphics capa-
bilities of Matlab (Fig.1). To run the code, enter ‘nbody(25,100,10)’ in
a matlab session, assuming that a file containing the following code is in
the matlab search path. The input data are the number of particles, the
number of outputs and the end time, respectively.
The first function initialises, runs at intervals of length dt (collecting
data in y), and outputs a final result.
function nbody(n,nout,tmax)
cpu = cputime;dt=tmax/nout;t=0;
[h,x,xdot,f,fdot,step,tlast,m,y] = initialise(n);
nsteps=0;
while t<tmax
[x,xdot,f,fdot,step,tlast,m,t,nsteps] = ...
hermite(x,xdot,f,fdot,step,tlast,m,t,t+dt,nsteps);
[y]=runtime output(h,x,y,t,n);
end

331
332 Appendix A

Fig. 1 Screen view after running the Matlab code.

final output(t,cpu,nsteps,y)
The next function initialises various scalars, and chooses n positions
and velocities with a uniform distribution inside spheres whose radii are
chosen to give N -body units approximately. It then sets masses, and
times when particles were most recently updated (i.e. 0, initially). Forces,
force derivatives and initial timesteps are selected (cf. Makino & Aarseth
1992). Finally a plot window for showing a movie is initialised.
function [h,x,xdot,f,fdot,step,tlast,m,y] = ...
initialise(n)
y(1)=0;figure(1)
x = uniform(n,6/5);
xdot = uniform(n,sqrt(5/6));
m = ones(1,n)./n;
tlast = zeros(1,n);
 A Simple N -Body Integrator 333

for i = 1:n
[f(1:3,i),fdot(1:3,i)] = ffdot(x,xdot,m,i);
step(i) = 0.01*sqrt(sum(f(1:3,i).^2)/sum(fdot(1:3,i).^2));
end
h=plot(x(1,:),x(2,:),’o’);
axis([-2 2 -2 2])
axis square
grid off
set(h,’EraseMode’,’xor’,’MarkerSize’,2)
function [x] = uniform(n,a)
r = a*rand(1,n).^(1/3);
cos theta = 2*rand(1,n)-1;
sin theta = sqrt(1-cos theta.^2);
phi = 2*pi*rand(1,n);
x(1,1:n) = r.*sin theta.*cos(phi);
x(2,1:n) = r.*sin theta.*sin(phi);
x(3,1:n) = r.*cos theta;
Particles are updated in this next function. In the main loop, the next
particle for updating is found (i), positions and velocities of all particles
are predicted, and force and force derivative on i are computed by a
call to ffdot (see below). Higher order Hermite corrections a2 and a3
are computed and the corrections applied. Finally the next timestep is
estimated and some variables are updated.
function [x,xdot,f,fdot,step,tlast,m,t,nsteps] = ...
hermite (x, xdot, f, fdot, step, tlast, m, t, tend,nsteps)
n = size(m,2);
while t < tend
[t,i] = min(step + tlast);
dt = t - tlast;
dts = dt.^2;
dtc = dts.*dt;
dt3 = dt(ones(1,3),:);
dts3 = dts(ones(1,3),:);
dtc3 = dtc(ones(1,3),:);
xtemp = x + xdot.*dt3 + f.*dts3/2 + fdot.*dtc3/6;
xdottemp = xdot + f.*dt3 + fdot.*dts3/2;
[fi,fidot] = ffdot(xtemp,xdottemp,m,i);
a2 = (-6*(f(1:3,i) - fi) - step(i)*(4*fdot(1:3,i) + ...
2*fidot))/step(i)^2;
a3 = (12*(f(1:3,i) - fi) + 6*step(i)*(fdot(1:3,i) + ...
fidot))/step(i)^3;
x(1:3,i) = xtemp(1:3,i) + step(i)^4*a2/24 + ...
334 Appendix A

step(i)^5*a3/120;
xdot(1:3,i) = xdottemp(1:3,i) + step(i)^3*a2/6 + ...
step(i)^4*a3/24;
f(1:3,i) = fi;
fdot(1:3,i) = fidot;
modfi = sum(fi.^2);
modfidot = sum(fidot.^2);
moda2 = sum(a2.^2);
moda3 = sum(a3.^2);
step(i) = min(1.2*step(i),sqrt(0.02*(sqrt(modfi*moda2) ...
+ modfidot)/(sqrt(modfidot*moda3) + moda2)));
tlast(i) = t;
nsteps = nsteps + 1;
end

function [fi,fidot] = ffdot(x,xdot,m,i)

n = size(m,2);
j = [1:i-1,i+1:n];
rij = x(:,i)*ones(1,n-1) - x(:,j);
rijdot = xdot(:,i)*ones(1,n-1) - xdot(:,j);
r2 = sum(rij.^2);
rrdot = sum(rij.*rijdot);
r3 = r2.^1.5;
r5 = r3.*r2;
fi = - rij*(m(j)./r3)’;
fidot = - rijdot*(m(j)./r3)’ + 3*rij*(rrdot.*m(j)./r5)’;

This customisable function updates the movie and collects data for
plotting the rms distance of all particles against time.
function [y]=runtime output(h,x,y,t,n)
set(h,’XData’,x(1,:),’YData’,x(2,:));
drawnow
m = size(y,1);
y(m+1,1) = sqrt(sum(sum(x(:,:).^2)))/n;
y(m+1,2) = t;

function final output(t,cpu,nsteps,y)

t
tcpu=cputime-cpu
nsteps
figure(2)
plot(y(2:end,2),y(2:end,1),’-’)
 A Simple N -Body Integrator 335

C
The following code lacks comments, but is structured like the preceding
Matlab code, so that the comments there should be helpful.
The code may be compiled with gcc∗ nbody.c -o nbody -lm, and
run with nbody 25 100 10 (say), where the data are the number of stars,
the number of outputs and the end time. It runs about six times faster
than the Matlab code. No movie is produced. To plot the accumulated
data y, gnuplot (http://www.gnuplot.org/) could be used.
#include <stdio.h>
#include <math.h>
#include <stdlib.h>
#include <time.h>
#include <assert.h>
#define NMAX 1024
#define NDIM 3
#define loop(idx,last) for (idx = 0; idx < last ; idx++)
#define min(x, y) (x<y?x:y)
typedef double real;
static int i, j, k;
void uniform(int n, real a, real temp[NMAX][NDIM]) {
real r, cos theta, sin theta, phi;
loop(i, n) {
r = a*pow(drand48(), 1./3.);
cos theta = 2.*rand()/((real)RAND MAX+1)-1;
sin theta = sqrt(1-pow(cos theta, 2));
phi = 2*M PI*rand()/((real)RAND MAX+1);
temp[i][0] = r*sin theta*cos(phi);
temp[i][1] = r*sin theta*sin(phi);
temp[i][2] = r*cos theta;
}
}
void ffdot(real x[NMAX][NDIM], real xdot[NMAX][NDIM],
real m[ ], int i, int n, real fi[ ], real fidot[ ]) {
real rij[NDIM], rijdot[NDIM], r2, r3, r5, rrdot;
loop(k, NDIM) { fi[k] = fidot[k] = 0; }
loop(j, n) {
if (j!=i) {
r2 = rrdot = 0;
loop(k, NDIM) {

∗
We used version 2.95 on a Sun running Solaris 2.7, for example.
336 Appendix A

rij[k] = x[i][k] - x[j][k];

rijdot[k] = xdot[i][k] - xdot[j][k];
r2 += pow(rij[k],2);
rrdot += rij[k]*rijdot[k];
}
r3 = pow(r2,1.5);
r5 = r3*r2;
assert(r3!=0&&r5!=0);
loop(k, NDIM) {
fi[k] = fi[k] - m[j]*rij[k]/r3;
fidot[k] = fidot[k] - m[j]*(rijdot[k]/r3 -
3*rrdot*rij[k]/r5);
}
}
}
}
int hermite(real x[NMAX][NDIM], real xdot[NMAX][NDIM],
real f[NMAX][NDIM], real fdot[NMAX][NDIM],
real step[ ], real tlast[ ], real m[ ], real *t,
real tend, int n)
{
real xtemp[NMAX][NDIM], xdottemp[NMAX][NDIM], fi[NDIM],
fidot[NDIM];
real a2[NDIM], a3[NDIM], modfi, modfidot, moda2, moda3;
real tmin, dt, dts, dtc, dnmtr, temp;
int imin, nsteps;
assert(n<=NMAX);
nsteps=0;
do {
tmin = pow(10,6);
imin = 0;
loop(i, n) {
if (step[i] + tlast[i] < tmin) {
tmin = step[i] + tlast[i];
imin = i;
}
}
assert(imin>=0);
*t = tmin;
i = imin;
loop(j, n) {
dt = *t - tlast[j];
 A Simple N -Body Integrator 337

dts = pow(dt, 2);

dtc = dts*dt;
loop(k, NDIM) {
xtemp[j][k] = x[j][k] + xdot[j][k]*dt + f[j][k]*dts/2
+ fdot[j][k]*dtc/6;
xdottemp[j][k] = xdot[j][k] + f[j][k]*dt +
fdot[j][k]*dts/2;
}
}
ffdot(xtemp,xdottemp,m,i,n,fi,fidot);
modfi = modfidot = moda2 = moda3 = 0;
assert(step[i]!=0);
loop(k, NDIM) {
a2[k] = (-6*(f[i][k] - fi[k]) - step[i]*(4*fdot[i][k]
+ 2*fidot[k]))/pow(step[i],2);
a3[k] = (12*(f[i][k] - fi[k]) + 6*step[i]*(fdot[i][k]
+ fidot[k]))/pow(step[i],3);
x[i][k] = xtemp[i][k] + pow(step[i],4)*a2[k]/24
+ pow(step[i],5)*a3[k]/120;
xdot[i][k] = xdottemp[i][k] + pow(step[i],3)*a2[k]/6
+ pow(step[i],4)*a3[k]/24;
f[i][k] = fi[k];
fdot[i][k] = fidot[k];
modfi += pow(fi[k],2);
modfidot += pow(fidot[k],2);
moda2 += pow(a2[k],2);
moda3 += pow(a3[k],2);
}
dnmtr = moda2 + sqrt(modfidot*moda3);
assert(dnmtr!=0);
temp = sqrt(0.02*(sqrt(modfi*moda2) + modfidot)/dnmtr);
step[i] = min(1.2*step[i], temp);
tlast[i] = *t;
nsteps++;
}
while(*t<tend);
return nsteps;
}
void initialise(int n, real x[NMAX][NDIM],
real xdot[NMAX][NDIM], real f[NMAX][NDIM],
real fdot[NMAX][NDIM], real step[NMAX],
real tlast[NMAX], real m[ ]) {
338 Appendix A

real fi[NDIM],fidot[NDIM], radius, fmod, fdotmod;

radius = 6./5.;
uniform(n,radius,x);
radius = sqrt(5./6.);
uniform(n,radius,xdot);
loop(i, n) { m[i] = 1./(real)n; }
loop(i, n) {
tlast[i] = 0;
ffdot(x,xdot,m,i,n,fi,fidot);
loop(k, NDIM) {
f[i][k] = fi[k];
fdot[i][k] = fidot[k];
}
fmod = fdotmod = 0;
loop(k, NDIM) {
fmod += pow(fi[k],2);
fdotmod += pow(fidot[k],2);
}
step[i] = 0.01*sqrt(fmod/fdotmod);
}
}
void runtime output(real x[NMAX][NDIM], real t,
long int n){
real ss=0;
loop (i,n) ss += x[i][0]*x[i][0] + x[i][1]*x[i][1] +
x[i][2]*x[i][2];
printf("%lf %lf\n",t,sqrt(ss));
}
void final output(real t, long int cpu, long int nsteps,
long int noutp){
printf("\n#%lf, %ld, %ld\n",
time, (clock()-cpu)/CLOCKS PER SEC, nsteps);
}
void main(int argc, char **argv) {
real x[NMAX][NDIM], xdot[NMAX][NDIM], f[NMAX][NDIM],
fdot[NMAX][NDIM];
real step[NMAX], tlast[NMAX], m[NMAX], t, dt;
long int j, i, n, nsteps, noutp, cpu;
cpu = clock();
printf("%ld\n",cpu);
srand(cpu);
 A Simple N -Body Integrator 339

assert(argc==4);
n = atoi(argv[1]);
noutp = atoi(argv[2]);
dt = atof(argv[3])/(real)noutp;
t = 0;
initialise(n, x, xdot, f, fdot, step, tlast, m);
nsteps=0;
loop(j, noutp) {
nsteps += hermite(x, xdot, f, fdot, step, tlast, m, &t,
t+dt, n);
runtime output(x,t,n);
}
final output(t,cpu,nsteps,noutp);
}

Problems
1) Given acceleration a and its derivative at the beginning and end
(subscripts 0 and 1, respectively) of a time step of length h, obtain
the approximate formulae

(2) −6(a0 − a1 ) − h(4ȧ0 + 2ȧ1 )

a0 =
h2
(3) 12(a0 − a1 ) + 6h(ȧ0 + ȧ1 )
a0 =
h3
for the second and third derivatives at the start of the time step. (The
Hermite integrator first takes the step using terms in a Taylor series
up to ȧ0 , i.e.
1 1
x01 = x0 + hv0 + h2 a0 + h3 ȧ0
2 6
0 1 2
v1 = v0 + ha0 + h ȧ0 ,
2
(2) (3)
computes a1 and ȧ1 , and then adds the terms in a0 and a0 , i.e.

1 4 (2) 1 5 (3)
x1 = x01 + h a0 + h a0
24 120
1 1
v1 = v10 + h3 a0 + h4 a0 .)
(2) (3)
6 24
2) Convert the code in Binney & Tremaine 1987 to the Hermite integra-
tor.
340 Appendix A

3) Modify the initial conditions of either N -body code to integrate the 3-

body problem with masses mi = 1 and the following initial conditions:
r1 = (−0.97000436, 0.24308753, 0) = −r2 ,
v1 = (−0.46620368, −0.43236573, 0) = v2 ,
r3 = 0, v3 = −2v1
(cf. Fig.20.4, top left panel).
4) At present the output routines use data corresponding to the last
update time. Modify so that it uses extrapolated data (as in the
hermite function/routine).
5) Perform the same kind of N -body calculation as the simple Matlab
and C codes (with the same initial conditions) using starlab.
 Appendix B
Hints to Solution of Problems

B.1 Part I

Chapter 1

1) See Box 7.2.

2) v2 = GM/rt , of course. For N -body units one could convert from c
to rt using Fig.8.3.
3) At distance d the area within a line of sight of solid angle 1 arcsec2
is (πd/648000)2.

Chapter 2

1) See Chapter 3.
2) For the radius of a stellar system, see Eq.(9.2), and footnote.
3) Choose an arbitrary centre O, and apply Newton’s Theorems to shells
with centre O.
4) r̈ = −(4π/3)Gρr, which is linear in r, and so collapse is homologous.
Solve the equation of motion at the edge of the sphere (r = a) using
formulae of Box 7.2, in case e = 1. Fig.1: π(6/5)3/2.

341
342 Appendix B

Chapter 3

1) Delete all code generating output, and run to (say) t = 1. Repeat

each run at least once to assess variance. Change N by factor 2.
2) Such a code will use the same time step for all masses (“lockstep”).
3) Even the circular two-body problem in the plane for one period with
minimal output will be illuminating. Maple performs many checks on
the validity of each arithmetic operation.

Chapter 4

1) The collision set has codimension 3, each orbit is one-dimensional.

For example, in the two-body problem the condition is (r1 − r2 ) ∧
(v1 − v2 ) = 0.
2 1/3
2) First part:
√ x = (9t /2) , y = 0; parabola; straight line.
3) c = 0, ± 3

Chapter 5

2) For n ≥ 3, circular motion is unstable (even without friction). For

n = 1 the speed in circular motion is independent of the radius of
the circle (isothermal case, cf. Chapter 8), and for n < 1 the speed
decreases with decreasing radius.

Chapter 6

1) In plane polars in the plane of motion the equations of motion are

−∂φ/∂r, rθ̈ + 2ṙθ̇ = 0, while vr = ṙ and vt = |rθ̇|.
r̈ − rθ̇ 2 =√
2) N (E) = 2(πGM )3f (E)/|E|5/2 for E < 0.
3) (a)
p
32x4 + 88x3 + 68x2 + 12x + 3 −x(x + 1)(π − 2 arcsin(2x + 1))
−x p
−32x4 − 40x3 + 4x2 + 12x + 3 −x(x + 1)(π − 2 arcsin(2x + 1))
(b) −x. It is a good idea to check continuity at x = 0 and −1.
4) Actually the variance of the 1/r2 force is infinite, so suppose there
is a short-distance cutoff
p independent of N . The collapse time scale
varies with force F as R/F .
 Hints to Solution of Problems 343

Chapter 7

1) Eq.(4) is true for a circular orbit of radius a around a mass M in

spherical symmetry. For both Keplerian and simple harmonic motions
the period is independent of eccentricity.
2) The approximate differential equation for y is linear.
3) This exercise is more challenging than useful. Anyway, from pericen-
tre (r0 ) to apocentre (r1 ) we have
q
ṙ = 2(E − φ) − J 2 /r2 ,
r2 θ̇ = J.
Hence the change in polar angle is
Z !
r1 J dr J
p = arccos p +
r0 r 2(E − φ) − J 2 /r2
2 r1 2(E − φ0 )
Z !
r1 dr 1 1
+J p − p ,
r0 r2 2(E − φ) − J /r
2 2 2(E − φ0 ) − J 2 /r2
where φ0 = φ(r0 ). To deal with the last integral, plot the integrand for
some model potential (e.g. Plummer, Table 8.1) to convince yourself
that the limit as J → 0 (and r0 → 0) poses no problems.
4) For a given orbit, construct a Keplerian potential φK = A − B/r
which agrees with φ at rmin and rmax . Prove that φK > φ for rmin <
r < rmax using the fact that M (r) is non-increasing, and then apply
the formula in the previous hint. For the other result, fit a harmonic
potential.
5) Use Eq.(8)
s and the transformation r = b sinh θ, where cosh θ =
GM J2
− + + (1 + GM/2bE)2 sin φ.
2bE 2b2 E √
6) Ans: π(−J + GM/ −2E). The pedestrian way is to use a trigono-
metric substitution for 1/r and then integrate by parts, leaving an
R π/2
integral of the form −π/2 dθ/(a + b sin θ). The smart way is to read
Goldstein (1980). (a) Ir remains close to 0, and so the orbit remains
nearly circular. Use Ė = ∂φ/∂t. (b) The orbit is Keplerian and starts
at an apse (ṙ = 0) with a specific velocity.
7) If so, the centre of the circle must be the midpoint of the values of ṙ
at peri- and apocentre.
344 Appendix B

Chapter 8

1) Ans: 1/(4πGj 2 r2 ), r/(Gj 2 ), 1/(4Gj 2d) on a line of sight passing at

distance d from the centre.
2) v̈ + v̇ +2(exp(v)−1) = 0, which is the equation of motion of a damped
oscillator with equilibrium at v = 0. (We learned this argument from
Donald Lynden-Bell.) Linearised equation: v̈ + v̇ + 2v = 0. √
3) 1d: find “energy” integral for u00 +exp(u) = 0. Ans: −2 ln cosh(z/ 2).
2d: transform to v(t) as in Problem 2, and then proceed as in 1d. Ans:
−2 ln(1 + z 2 /8). See Ostriker (1964).
4) In general in spherical symmetry one has φ ' φ(0) − 2πGρ(0)r
R∞ 2 p
2 /3.

5) ρ = 4π 0 f v dv. After differentiation, multiply by 1/ 2(φ − E 0 )

and integrate (with respect to φ) from E 0 to ∞. Reverse the or-
der of integration in the right side, and use the transformation φ =
E 0 cos2 θ + E sin2 θ. Finally, differentiate with respect to E 0 .
6) Eq.(8) is approximately f = 2j 2 f0 (E0 − E) if E < E0 . Then use the
first result of Problem 5 to show that

ρ = C(φe − φ)5/2 , (1)

where C is a constant R and e denotes the edge.R Next, the potential

energy is W = 2π ρφr2 dr, = M φe /2 − 2π ρ(φe − φ)r2 dr. Use
Eq.(1), integrate by parts,
R and use (1) again, which gives an integral
proportional to W = − GM dM/r. There is a similar result for any
finite polytrope (cf. Eddington 1926b, Sect.60).
7) Integrals are performed easily enough in spherical polar coordinates
with the corresponding velocity components vr , vθ and vφ . Ans:
221/2 G1/2 M 3/2 a1/2 /4851.
8) Use radial and transverse velocity components vr , vt for all the inte-
grals.

Chapter 9

R
1) T =R 6π pr2 dr. Integrate by parts, and use W =
−G
P P
M (r)dM (r)/r. P P
2) i P ri .(ri − rj )/|ri − rj |3 = j i6=j rj .(rj − ri )/|rj − ri |3 =
j6=iP
(1/2) i j6=i (ri − rj ).(ri − rj )/|ri − rj |3 .
4) Answers in Table 8.1.
 Hints to Solution of Problems 345

5) (a)
Z Z
∂f ∂
m(v − hvi)2 d3 v = f m(v − hvi)2 d3 v+
∂t ∂t Z
+2 ˙ 3v
f m(v − hvi).hvid
∂
= 3ρσ2 + 0.
∂t
Similarly the second term gives
 
∇ · ρh(v − hvi)(v − hvi)2 i + 3ρσ2 hvi + 2ρσ2 ∇ · hvi

and the last term gives nothing.

6) L ∼ M σ2 /tr , ∼ r2 k∂σ2 /∂r.
7) See Lightman & Shapiro (1977).

Chapter 10

1) fR is periodic in J with period 2π/t. t > π/(2δJ) is sufficient.

2) C(f )dτ is invariant, by Liouville’s Theorem.
3) J =  {sin(φ + 2t) +
˙
√ sin φ } before averaging. The fact that the width
is proportional to  is fundamental to the understanding of reso-
nance.

Chapter 11

1) For a power law with extreme masses mmin and mmax M is pro-
max − mmin if α 6= 2. The quoted approximation to
2−α
portional to m2−α
the Miller-Scalo mass function (see caption to Fig.1) implies that the
fraction of stars with mass less than m is F , where
0.19F
m= .
(1 − F )0.75 + 0.032(1 − F )0.25
Show that Ṁ = mṁdF/dm, where the right side is evaluated at the
turnoff mass.
2) Set c2 = Gm0 a, r ' a in the right side of the equation of motion.
(∂E/∂t)s
3) (∂s/∂t)E = − .
(∂E/∂s)t
5) f ' constant, and so ρ ∝ (−φ)3/2 . Cf. Box 8.1.
346 Appendix B

Chapter 12

√
1) In Eq.(3) ω ∼ κ ∼ GρG (cf. Eq.(7.5)).
3) Consider the acceleration (due to the sun) on (i) the sea on the sunlit
surface, (ii) the solid earth, (iii) the sea on the unlit surface.
4) Replace t in the trigonometric terms by t − t0 . Then
ẋ = −a1 κ sin[κ(t−t0 )] if ȧ0 + ȧ1 cos[κ(t−t0 )]+κa1 ṫ0 sin[κ(t−t0 )] = 0.
Three more conditions come from the corresponding treatment of y
and the first two equations of motion (1). The averaging has to be
done numerically. See Ross et al. (1997).
5) See Fukushige & Heggie (2000).

Chapter 13

1) See any dynamics book (e.g. Goldstein 1980) for the transformation
to rotating coordinates.
2) Nearest neighbour distance is of order n−1/3 .
3) (i) Rather than looking up general formulae for geodesics in arbi-
trary metrics, it is quicker to derive the equation with bare hands,
starting from the Euler-Lagrange equation. One obtains
 0
0 1 ∂V
f xi =− ,
2f ∂xi
q
where f = (E − V )/(xi0 xi0 ) and 0 = d/ds. (ii) The space is confor-
mally flat, and the expressions for Christoffel symbols and the Ricci
tensor take fairly simple forms. Cf. Synge (1964), though he considers
only the case of a four-dimensional space-time. For the last part con-
sider the two-dimensional two-body problem. (iii) The induced metric
is (ρ02 + z 02 )dr2 + ρ2 dθ 2 . A global solution seems not to be possible if
E < 0.

Chapter 14

1) In suitably rotated coordinates the orbit is given by Eq.(1), Box 7.2.

The first formula for ρ/ρ0 is given q
by Bisnovatyi-Kogan et al. (1979).
Eventually it reduces to p2 /(r sin θ r2 sin2 θ + 4ra(1 − cos θ)).
 Hints to Solution of Problems 347

2) A reassuring intermediate result is

Z
1X (p̂.v1 )2
(δE1 )2 = 4G2 m2 f (v2 )dp̂dpd3 v2 ,
t pV
where the integration with respect to p̂ is round a circle orthogonal
to V = v1 − v2 . To perform the integration with respect to v2 use
spherical polar coordinates with polar axis along v1 , and the generat-
ing function of Legendre polynomials.
3) Dimensional analysis
4) The intermediate result is
 s 
 
1X G2 m2 n2 σ2 ln Λ  2 v1 v2 v
(δE1 )2 = 8π − exp(− 12 ) + erf √1 .
t v1 π σ 2σ 2σ
Your final coefficient should be the same as the one in Eq.(8).
5) Strictly the coefficient should be 0.139.
8) To exploit the assumption of constant total energy, use the virial
theorem and the approximation for the potential energy in Problem
5.

Chapter 15

1) Simple algebra, using the quaternion commutation relations i2 = j 2 =

k 2 = −1, ij = −ji = k, jk = −kj = i and ki = −ki = j.
2) As 1.
3) (ii) Write down the condition that λ + (1 − λ)q is imaginary.
4) Some relabelling and further transformation of variables is needed.
5) Cf. Box 21.1, and Chenciner & Montgomery (2001).

Chapter 16

1) The integral with respect to E1 can be performed with the substi-

tution E1 = (φ − E2 ) sin2 θ, and the integral with respect to E2 sim-
ilarly. The r−integration gives s a Beta function. The exact result is
2 37/2 Γ(3/4) GM
Ṅ = − 2 . See Hénon (1960).
3.5 .7.11.13 .17Γ(1/4) πa3
2
2) Relaxation is dominated by passage through the core, at intervals
∝ r3/2 (Kepler’s
q 3rd law). Treating the effect as a random walk,
δE(t) ∝ t/r3/2 . Finally, E ∝ 1/r. See Lightman & Shapiro (1978,
§V.A.2), and also Spitzer (1987, p.89f).
348 Appendix B
P
3) T = M hv2 i/2, W = 2 mφ, ve2 = −2φ. Ans: 0.0074.
4) If |m2 φc | is big enough, then f1 and f2 are nearly of Maxwell-
Boltzmann form almost everywhere.

Chapter 17

1 (3x − 2)W (x)

1) The temperature is T = − . The heat capacity
6N k x−1
changes sign where dT /dE = 0 (cf. Binney & Tremaine 1987).
3) In 3d the virial theorem gives T = −E, and so loss of energy in-
creases temperature. What happens in 2d? In using the analytic
solution assume that the potential vanishes at the edge radius re .
Then the work done (per unit length)
Z in constructing the model from
re
material at this radius is W = 2GM (r) ln(r/re )dM. Then plot
0
GM m z2
= against
kT 4(1 + z 2 /8)
!
E (1 + z 2 /8)2 z2
2
= −64 ln(1 + z /8) + 12
2
GM z4 1 + z 2 /8
and look for a vertical tangent.

Chapter 18

1) By Eq.(2), Box 14.3, s(E) ∝ (E − φ(0))3 at energies E just above the

central potential φ(0).
2) (a) See Chapter 16 on mass segregation
(b) The density of black holes varies as rs−3 , where rs is the core
radius of the stars, while the density of stars varies as rs−α (Eq.(1)).
(c) Eq.(14.8). See Kulkarni et al. (1993) and Sigurdsson & Hernquist
(1993) for more detailed treatments.
3) Use results of Chapter 11, and Problem 14.5, or see Hut (1997).

Chapter 19

1) Apply time reversal to the required encounter: in that case a binary,

consisting of stars of mass M , is destroyed by a grain of sand of mass
m. To see how this can be done, consider a binary of high eccentric-
ity, and let the two stars be at apocentre, their separation being r.
 Hints to Solution of Problems 349

Their relative velocity may be made as small as we please, by suitable

adjustment of the orbital parameters. The sand grain approaches one
star with speed V , and the maximum speed which it can impart to
the star is of order mV /M . By choice of V this may exceed the escape
speed from the other star, and indeed may be made as large as we
please. Now reverse the velocities at the end of the encounter.
2) N mhv2 i/2 = −E.
3) The joint distribution of the two stars is n2 (2πkT /m)−3 exp(−(v12 +
v22 )/(2kT )). Convert to relative and barycentric coordinates and ve-
locities (cf. Eq.(1)), add a factor δ(mv2 /4 − Gm2 /r − ε) (cf. Eq.(2))
and integrate over all variables except the position of the barycentre.
A factor of 1/2 is needed in order not to count pairs twice.
4) See Hut (1985). A classical calculation is given by Jeans (1929).

Chapter 20

1) Kepler’s 3rd Law (cf. Petrosky 1986, who calls the resulting map the
“Kepler” map). Area-preservation: compute the Jacobian. If  = −0.1
and t0 = 0, there is interesting structure near E0 = 0.1, 0.24, 0.2344
and 0.23415. (The number of decimal places indicates the care with
which you have to look.) These structures are easier to see if the log
of the escape time is plotted against E0 .
2) Drop a particle parabolically into a Sitnikov machine, and imagine
finding a value for the initial phase at which the particle escapes hy-
perbolically after a number of interactions. Now vary the initial phase.
3) Adjust the phase of each visit to the binary so that the particle
more and more nearly escapes parabolically each time. One gets an
impression of what is needed from the case z(0) = 0, ż(0) = 2.81989
to t ' 1000, if the binary has e = 0.6, P = 2π and a = 1.
4) The equation u00 + (1/3)u0 − (2/9)u + 2u(u2 + 34/3 /4)3/2 = 0, where u
is the scaled z, has equilibria at u = 0 and u = ±37/6 /2. As the energy
decreases the oscillator either escapes or is trapped. In the latter case
the amplitude of the z motion increases without limit in general.

Chapter 21

Z
G2 m3 n a dp
1) ε̇ ∼ , using notation of Eq.(4), Box 14.1, and
V Gm/V 2 p
Eq.(5).
3) Compute Ṅb R3 tr .
350 Appendix B

4) See Hut (1989).

5) h.ḣ = µh.r×f , where µ is the reduced mass of the binary. Use Eq.(2),
Box 1, and estimate the average over binary motion (though this is
zero if the binary is circular). Hence δh ∼ Gm3 µa2 T /Rp3 , where T is
the time scale of the encounter.
6) If t = i(t0 + s) along the cut, where t0 = Rp /V , the exact result is
Z
2n ∞ exp(−n[t0 + s])ds
I= 3 √ .
V 0 (t0 + s) 2t0 s + s2
7) The linearised equations are
4 1 1
(−t)2 δ ẍ − (−t)δ ẋ = δx + δy
3 2 6 .
4 1 1
(−t)2 δ ÿ − (−t)δ ẏ = δx + δy
3 6 2

Chapter 22

1) Seek solutions ∝ λn . Consider |λ|.

3) See Hut (1983b).

Chapter 23

1) Use Eq.(21.4), and an argument in the discussion of hard binaries in

Chapter 19. For the last part see Goodman & Hut (1993), or Retterer
(1980).
2) Other results of such experiments are described by Giersz & Heggie
(1994).
3) Cf. Kulkarni et al. (1993) and Sigurdsson & Hernquist (1993).

Chapter 24

1) To help as much as possible, take conditions in the disk: ρ ∼ 0.1M

pc−3 , σ ∼ 40kms−1 , total mass ∼ 6 × 1010 M (adapted from Binney
& Tremaine 1987) and m ∼ 1M .
2) The number of AU in a parsec is the same as the number of arc
seconds in a radian.
3) See Klessen & Kroupa (2001) and references therein.
 Hints to Solution of Problems 351

Chapter 25

1) The relative motion of the moon and earth is governed by the differ-
ence in their acceleration.
2) Put ei (0) = 0 in Eqs.(4) and (5), which imply cos2 i = cos2 i(0)/(1 −
e2i ) and e2 (2 − 5 sin2 i sin2 ω) = 0. Hence e = 0 unless sin2 i ≥ 2/5.
3) z = (a[1 − e cos E] sin ω + b sin E cos ω) sin i; hf (E)i =
Z
1 2π
f (E)(1−e cos E)dE.
2π 0

Chapter 26

2) In the second case, one component of the wider binary must interact
closely with the harder pair.
3) In the input string, M is the projectile mass, v is the initial relative
velocity of the two binaries, and a and e refer to the projectile (p)
and target (t), respectively. The output shows how the particles are
grouped, along with position and velocity components, masses and
energy control.

Chapter 27

1) Use Eq.(3) to obtain the estimate of Ṁ . See Problem 8.1, for the
mass-radius relation in an isothermal cluster.
2) Unless ρ and σ decrease very slowly with increasing r, the third
equation shows that L → 0 as r → ∞, which is inconsistent with the
fourth equation. Relaxation is too slow in the outer halo of a cluster
to remove all the energy generated.

Chapter 28

1) Use the fact that trh ∝ (N rh3 )1/2 (Eq.14.9). Neglecting variation of
the Coulomb logarithm, the unreduced result is approximately
 k
t 10.8 N (0)
' + 18,
trh (0) 0.12k 2
where k = 3/0.24 − 1/2.
2) Good values are c = (a) 10−3 , (b) 10−5 , (c) 10−7 and (d) 10−8 .
352 Appendix B

3) The expansion of the cluster requires energy generation at a rate of

order N mσ2 /trh .

Chapter 29

1) Use Eq.(8.9). Equate the estimate to the heat flux at rh , and use a
virial estimate of σc , assuming the cluster is nearly isothermal inside
rh .
2) For the first step use an estimate like Eq.(2), Box 12.2, and virial
estimates.
3) If the initial total mass is large enough, rh begins by decreasing with
M.

Chapter 30

1) The energy of the binary is −GM 2 /(4R ).

Chapter 31

1) Reinterpret Eqs.(21.1) and (21.4).

Gmδm 1 m2 2
2) = V .
R 2 2m
4) Suppose the constant in Eq.(12.4), is −GM/a, and develop a simple
approximation to r − a. Put κ2 ' ωz2 ' ω 2 ' V 2 /R2 ' 4Gm/R3 . The
quadrupole field of lumps δm is of order Gδmr 2 /R3 . For a detailed
analysis of equilibrium tides see Eggleton et al. (1998).

Chapter 32

1) See Box 7.2.

2) Integrating Eq.(31.1) over all stars counts each collision twice.

Chapter 33

1) For a flat rotation curve φ ∝ ln r, whence κ can be calculated for

insertion into Eq.(12.3).
2) Show that tr (t) = tr (0) − kt, where k is a constant.
 Hints to Solution of Problems 353

Appendix A

(2) (3)
1) Use a1 = a0 + hȧ0 + h2 a0 /2 + h3 a0 /6 and a similar expression for
ȧ1 .
5) The command
mksphere -n 25|flat hermite -D 0.1 -t 1|xstarplot
almost does this, but initial velocities are uniform in a cube instead of
in a sphere.
 References

Aarseth S.J. (1985). Direct methods for N -body simulations, in Multiple Time
Scales, ed. Brackbill J.U. & Cohen B.I. (Academic Press, New York).
Aarseth S.J. (1999). From NBODY1 to NBODY6: the growth of an industry,
PASP 111, 1333.
Aarseth S.J., Hénon M., Wielen R. (1974). A comparison of numerical methods
for the study of star cluster dynamics, A&A 37, 183.
Aarseth, S.J., Lin, D.N.C. and Papaloizou, J.C.B. (1988). On the collapse and
violent relaxation of protoglobular clusters, ApJ 324, 288-310.
Aarseth S.J., Lecar M. (1975). Computer simulations of stellar systems,
ARA&A 13, 1.
Aarseth S.J., Mardling R.A. (2001). Evolution of binary and multiple star
systems, in “Eggletonfest”, a Meeting in Celebration of Peter Eggleton’s 60th
Birthday, eds. P. Podsiadlowski, S. Rappaport, A.R. King, F. D’Antona, L.
Burder, ASP Conf. Ser. 229 (ASP: San Francisco)
Aarseth S.J., Zare K. (1974). A regularization of the three-body problem, Celes.
Mech. 10, 185.
Abraham R., Marsden J.E., (1978). Foundations of Mechanics, 2nd ed.
(Benjamin/Cummings, Reading)
Agekyan T.A., Anosova Z.P., Orlov, V.V. (1983). Decay time of triple systems,
Astrophysics 19, 66.
Aguilar L., Hut P., Ostriker, J.P. (1988). On the evolution of globular cluster
systems. I - Present characteristics and rate of destruction in our Galaxy,
ApJ 335, 720.
Ahmad A., Cohen L. (1973). A numerical integration scheme for the N-body
gravitational problem, J. Comp. Phys. 12, 389.
Airy G.B. (1884). Gravitation : an elementary explanation of the principal
perturbations in the solar system, Macmillan, London
Alekseev V.M. (1981). Final motions in the three-body problem and symbolic
dynamics, Russ. Math. Surv. 36, 181.

354
 References 355

Allen, F.S. and Heggie D.C. (1992). A model gravothermal oscillator, MNRAS
257, 245-56.
Ambartsumian, V.A. (1938). On the dynamics of open clusters, Uch. Zap.
L.G.U., No.22, p.19; transl. in Dynamics of Star Clusters, eds. J. Goodman
and P. Hut (D. Reidel, Dordrecht).
Antonov, V.A. (1962). (Most probable phase distribution in spherical star
systems and conditions for its existence), Vest. Leningrad Univ. 7, 135;
translated in Dynamics of Star Clusters, eds Goodman J., Hut P., (D. Reidel,
Dordrecht)
Antonov V.A. (1973). (On the instability of stationary spherical models with
purely radial motion), in T.B. Omarov, ed, The Dynamics of Galaxies and
Star Clusters, (Nauka Kazakh S.S.R., Alma Ata); transl. in P.T. de Zeeuw,
ed, Structure and Dynamics of Elliptical Galaxies, IAU Symp. 127 (Reidel,
Dordrecht)
Applegate J.H., Douglas M.R., Gürsel Y., Hunter P., Seitz C.L., Sussman G.J.
(1985). A digital orrery, IEEE Trans. Comp. C-34, 822.
Arabadjis J.S., Richstone D.O. (1998). The dynamical evolution of dense
rotating systems: paper I. Two-body relaxation effects, astro-ph/9810192
Arnold, V.I. (1978a). Mathematical Methods of Classical Mechanics
(Springer-Verlag, New York)
Arnold, V.I. (1978b). Ordinary Differential Equations (MIT Press, Cambridge)
Arnold V.I. (1983) Geometrical Methods in the Theory of Ordinary Differential
Equations (Springer, Berlin)
Arnold V.I., Kozlov V.V., Neishtadt A.I. (1997) Mathematical Aspects of
Classical and Celestial Mechanics, 2nd ed (Springer-Verlag, Berlin)
Ashman K.M., Zepf S.E. (1998). Globular cluster systems (Cambridge
University Press, Cambridge)
Bahcall J.N., Hut P., Tremaine S. (1985). Maximum mass of objects that
constitute unseen disk material, ApJ 290, 15.
Bahcall J.N., Wolf R.A. (1976). Star distribution around a massive black hole
in a globular cluster, ApJ 209, 214.
Bailey M.E., Chambers J.E., Hahn G. (1992). Origin of sungrazers - a frequent
cometary end-state, A&A 257, 315.
Bailey V.C., Davies M.B. (1999). Red giant collisions in the Galactic Centre,
MNRAS 308, 257.
Bailyn C.D. (1996). X-ray binaries in globular clusters, in The Origins,
Evolutions, and Destinies of Binary Stars in Clusters, ed. E.F. Milone and
J.-C. Mermilliod (ASP (Conf. Ser. 90), San Francisco).
Barnes J.E. (1985). Dynamical instabilities in spherical stellar systems, in
Dynamics of Star Clusters (IAU Symp. 113), ed. Goodman J., Hut P.
(Reidel, Dordrecht).
Barnes & Hut (1986). A hierarchical O(N log N ) force-calculation algorithm,
Nature 324, 446.
356 References

Barnes J., Goodman J., Hut P. (1986). Dynamical instabilities of spherical

stellar systems. I. Numerical experiments for generalized polytropes, ApJ
300, 112.
Barrow-Green J. (1996). Poincaré and the Three-Body Problem (American and
London Mathematical Societies, Providence and London)
Baumgardt H. (1998). The initial distribution and evolution of globular cluster
systems, A&A 330, 480.
Baumgardt H. (2001). Scaling of N -body calculations, MNRAS 325, 1323.
Becker L. (1920). On capture orbits, MNRAS 80, 590.
Bettwieser E., Sugimoto D. (1984). Post-collapse evolution and gravothermal
oscillation of globular clusters, MNRAS 208, 493.
Binney, J. and Tremaine, S. (1987). Galactic Dynamics (Princeton University
Press, Princeton)
Bisnovatyi-Kogan G.S., Kazhdan Ya.M., Klypin A.A., Lutskii A.E., Shakura
N.I. (1979). Accretion onto a rapidly moving gravitational center, Sov.
Astron. 23, 201.
Boyd P.T., McMillan S.L.W. (1992). Initial-value space structure in irregular
gravitational scattering, Phys. Rev. A46, 6277.
Breeden J.L., Cohn H.N. (1995). Chaos in core oscillations of globular clusters,
ApJ 448, 672.
Breeden J.L., Cohn H.N., Hut P. (1994). The onset of gravothermal oscillations
in globular cluster evolution, ApJ 421, 195.
Burdet C.A. (1967). Regularization of the two body problem, ZAMP 18, 434.
Burrau C. (1913). Numerische Berechnung eines Spezialfalles des
Dreikörperproblems, Astron. Nach. 195, 113.
Capuzzo-Dolcetta R., Tesseri A. (1999). Globular cluster system erosion and
nucleus formation in elliptical galaxies, MNRAS 308, 961.
Casertano, S. and Hut, P. (1985). Core radius and density measurements
in N -body experiments: Connections with theoretical and observational
definitions, ApJ 298, 80-94.
Cassisi S., Castellani V. (1993). An evolutionary scenario for primeval stellar
populations, ApJS 88, 509.
Chandrasekhar, S. (1939). Stellar Structure (University of Chicago Press,
Chicago); also Dover, New York
Chandrasekhar, S. (1942). Principles of Stellar Dynamics (University of Chicago
Press, Chicago) p.55f
Chavanis P.-H. (2002). Gravitational instability of finite isothermal spheres,
A&A 381, 340.
Chavanis P.-H., Rosier C., Sire C. (2001). Thermodynamics of self-gravitating
systems, cond-mat/0107345
Chenciner A., Montgomery R. (2000). A remarkable periodic solution of the
three-body problem in the case of equal masses, Ann. Math. 152, 881.
Chernin A.D., Valtonen M.J. (1998). Intermittent chaos in three-body
dynamics, N. Astron. Rev. 42, 41.
 References 357

Chernoff D.F., Shapiro S.L. (1987). Globular cluster evolution in the Galaxy -
a global view, ApJ 322, 113.
Chernoff, D.F. and Weinberg, M.D. (1990). Evolution of globular clusters in
the galaxy, ApJ 351, 121.
Cipriani P., Pettini M. (2001). Strong chaos in N -body problem and
microcanonical thermodynamics of collisionless self-gravitating systems,
astro-ph/0102143
Clarke C.J., Pringle J.E. (1991). The role of discs in the formation of binary
and multiple star systems, MNRAS 249, 588.
Clayton D.D. (1983). Principles of Stellar Evolution and Nucleosynthesis,
(University of Chicago Press, Chicago)
Cohen R.S., Spitzer L., Jr., Routly P.McR. (1950). The electrical conductivity
of an ionized gas, Phys. Rev. 80, 230.
Cohn H. (1979). Numerical integration of the Fokker-Planck equation and the
evolution of star clusters, ApJ 234, 1036.
Cohn H.N. (1985). Direct Fokker-Planck calculations, in Dynamics of star
clusters, IAU Symp. 113, ed. Goodman J., Hut P. (Reidel, Dordrecht).
Cohn, H., Hut, P. and Wise, M. (1989). Gravothermal oscillations after core
collapse in globular cluster evolution, ApJ 342, 814-22.
Combes F., Leon S., Meylan G. (1999). N -body simulations of globular cluster
tides, A&A 352, 149.
Côté P., Fischer P. (1996). Spectroscopic binaries in globular clusters. I. A
search for ultra-hard binaries on the main sequence in M4, AJ 112, 565.
Côté P., Pryor C., McClure R.D., Fletcher J.M., Hesser J.E. (1996).
Spectroscopic binaries in globular clusters. II. A search for long-period
binaries in M22, AJ 112, 574.
Cuperman S., Harten A. (1972). The evolution of a multi-phase space density
collisionless one-dimensional stellar system, A&A 16, 13.
Cvitanović P. (1989). Universality in Chaos, 2nd edition (Adam Hilger, Bristol)
Danilov V.M. (1997). Numerical experiments simulating the dynamics of open
clusters in the Galactic field, Astron. Rep. 41, 163.
David M., Theuns T. (1989). Numerical experiments on radial virial oscillations
in N -body systems, MNRAS 240, 957.
Davies M.B., Benz W. (1995). A stellar audit: the computation of encounter
rates for 47 Tucanae and ω Centauri, MNRAS 276, 876.
Davoust E., Prugniel P. (1990). On the flattening of globular clusters, A&A
230, 67.
Dejonghe H. (1984). The construction of analytical models for globular clusters,
A&A 133, 225.
Dejonghe H. (1987a). A completely analytical family of anisotropic Plummer
models, MNRAS 224, 13.
Dejonghe, H. (1987b). On entropy and stellar systems, ApJ 320, 477-81.
Dejonghe H., Hut P. (1986). Round-off sensitivity in the N -body problem, in
The Use of Supercomputers in Stellar Dynamics, ed. Hut P., McMillan S.
(Lect. Notes Phys., 267, Springer-Verlag, Berlin).
358 References

de Vega H.J., S’anchez N. (2000). The statistical mechanics of the

self-gravitating gas: equation of state and fractal dimension, Phys. Lett. B
490, 180.
Diacu F. (2000). A century-long loop, Math. Intelligencer 22, No.2, 19
Diacu, F. and Holmes, P. (1996). Celestial Encounters (Princeton University
Press, Princeton)
Djorgovski S., Piotto G., Phinney E.S., Chernoff D. (1991). Modification of
stellar populations in post-core-collapse globular clusters, ApJ 372L, 41.
Drukier G.A. (1995). Fokker-Planck models of NGC 6397, ApJ Supp 100, 347.
Dull J.D., Cohn H.N., Lugger P.M., Murphy B.W., Seitzer P.O., Callanan P.J.,
Rutten R.G.M., Charles P.A. (1997). The dynamics of M15: observations of
the velocity dispersion profile and Fokker-Planck models, ApJ 481, 267.
Duquennoy A., Mayor M. (1991). Multiplicity among solar-type stars in the
solar neighbourhood. II - Distribution of the orbital elements in an unbiased
sample, A&A 248, 485.
Eddington, A.S. (1926a). Star, in Encyclopaedia Brittanica, 13th ed. (London),
Vol.25, p.787
Eddington A.S. (1926b). The Internal Constitution of the Stars (Cambridge
University Press, Cambridge); also Dover Publications, New York, (1959)
Efstathiou G., Davis M., White S.D.M., Frenk C.S. (1985). Numerical
techniques for large cosmological N -body simulations, ApJS 57, 241.
Eggleton P., Kiseleva L. (1995). An empirical condition for stability of
hierarchical triple systems, ApJ 455, 640.
Eggleton P.P., Kiseleva L.G., Hut P. (1998). The equilibrium tide model for
tidal friction, ApJ 499, 853.
Eggleton, P.P., Tout, C.A. and Fitchett, M.J. (1989). The distribution of visual
binaries with two bright components, ApJ 347, 998-1011.
Einsel C., Spurzem R. (1999). Dynamical evolution of rotating stellar systems –
I. Pre-collapse, equal-mass system, MNRAS 302, 81.
Eisenstein D.J., Hut P. (1998). HOP: a new group-finding algorithm for N -body
simulations, ApJ, 498, 137
Emden R. (1907). Gaskugeln (Teubner, Leipzig)
Engle K. (1999), N -Body Simulations of Star Clusters (PhD Thesis, Drexel
University)
Euler L. (1776). Nova Methodus Motum Corporum Rigidorum Determinandi,
Novi comment. acad. sc. Petrop. 20, 208.
Fabian A.C., Pringle J.E., Rees M.J. (1975). Tidal capture formation of binary
systems and X-ray sources in globular clusters, MNRAS 172P, 15.
Fall S.M., Rees, M.J. (1977). Survival and disruption of galactic substructure,
MNRAS 181P, 37.
Fall S.M., Zhang Q. (2001). Dynamical Evolution of the Mass Function of
Globular Star Clusters, ApJ 561, 751.
Fanelli D., Merafina M., Ruffo S. (2001). A one-dimensional toy model of
globular clusters, Phys. Rev. E 63, 066614-1.
 References 359

Farouki R.T., Salpeter E.E. (1994). Mass segregation, relaxation, and the
Couloumb logarithm in N -body systems (again), ApJ 427, 676.
Ferraro F.R., Fusi Pecci F., Cacciari C., Corsi C., Buonanno R., Fahlman G.G.,
Richer H.B. (1993). Blue stragglers in the Galactic globular clusters M3:
evidence for two populations, AJ 106, 2324.
Feynman R.P. (1965). Lectures on Physics, Vol.III. (Addison-Wesley, Reading)
Chapter 6
Figer D.F., Kim S.S., Morris M., Serabyn E., Rich R.M., McLean I.S. (1999).
Hubble Space Telescope/NICMOS Observations of Massive Stellar Clusters
near the Galactic Center, ApJ 525, 750.
Fischer P., Welch D.L., Cote P., Mateo M., Madore B.F. (1992). Dynamics of
the young LMC cluster NGC 1866, AJ 103, 857.
Fischer P., Pryor C., Murray S., Mateo M., Richtler T. (1998). Mass segregation
in young Large Magellanic Cloud clusters. I - NGC 2157, AJ 115, 592.
Fridman A.M., Polyachenko V.L. (1984). Physics of Gravitating Systems,
(Springer Verlag, Berlin, 2 Vols)
Gebhardt K., Fischer P. (1995). Nonparametric dynamical analysis of globular
clusters: M15, 47 TUC, NGC 362, and NGC 3201, AJ 109, 209.
Freitag M., Benz W. (2001). A Monte Carlo code to investigate stellar
collisions in dense galactic nuclei, in Stellar Collisions, Mergers and their
Consequences, ed. Shara M., ASP Conf. Ser. 263 (ASP, San Francisco).
Fukushige T. (1995). Gravitational scattering experiments in infinite
homogeneous N -body systems, in Chaos in Gravitational N -body Systems,
eds. Muzzio J.C., Ferraz-Mello S., Henrard J. (Kluwer, Dordrecht)
Fukushige T., Heggie D.C. (2000). The time-scale of escape from star clusters,
MNRAS 318, 753.
Fukushige T., Hut P., Makino J. (1999). High-performance special-purpose
computers in science, IEEE Computing in Science and Engineering, 1, No. 2,
12
Fullerton L.W., Hills J.G. (1982). Computer simulations of close encounters
between binary and single stars - the effect of the impact velocity and the
stellar masses, AJ 87, 175.
Funato Y., Hut P., McMillan S., Makino J. (1996). Time-symmetrized
Kustaanheimo-Stiefel regularization, AJ 112, 1697.
Gao B., Goodman J., Cohn H., Murphy B. (1991). Fokker-Planck calculations
of star clusters with primordial binaries, ApJ 370, 567.
Garabedian P.R. (1986). Partial Differential Equations, 2nd edition (Chelsea,
New York)
Ghez A.M., Morris M., Becklin E.E., Tanner A., Kremenek T. (2000). The
accelerations of stars orbiting the Milky Way’s central black hole, Nature
407, 349.
Giersz M. (1998). Monte Carlo simulations of star clusters - I. First Results,
MNRAS 298, 1239.
Giersz M., Heggie D.C. (1994a). Statistics of N -body simulations – I. Equal
masses before core collapse, MNRAS 268, 257.
360 References

Giersz M., Heggie D.C. (1994b). Statistics of N -body simulations – II. Equal
masses after core collapse, MNRAS 270, 298.
Giersz M., Heggie D.C. (1996). Statistics of N -body simulations – III. Unequal
masses, MNRAS 279, 1037.
Giersz M., Heggie D.C. (1997). Statistics of N -body simulations – IV. Unequal
masses with a tidal field, MNRAS 286, 709-31.
Giersz M., Spurzem R. (1994). Comparing direct N -body integration with
anisotropic gaseous models of star clusters, MNRAS 269, 241.
Giersz M., Spurzem R. (2000). A stochastic Monte Carlo approach to model
real star cluster evolution - II. Self-consistent models and primordial binaries,
MNRAS 317, 581.
Gilliland R.L., Brown T.M., Guhathakurta P., Sarajedini A., Milone E.F.,
Albrow M.D., Baliber N.R., Bruntt H., Burrows A., Charbonneau D., Choi
P., Cochran W.D., Edmonds P.D., Frandsen S., Howell J.H., Lin D.N.C.,
Marcy G.W., Mayor M., Naef D., Sigurdsson S., Stagg C.R., Vandenberg
D.A., Vogt S.S., Williams M.D. (2000). A lack of planets in 47 Tucanae from
a Hubble Space Telescope search, ApJ 545L, 47.
Gnedin O.Y., Lee H.M., Ostriker J.P. (1999). Effects of tidal shocks on the
evolution of globular clusters, ApJ 522, 935.
Gnedin O.Y., Ostriker J.P. (1997). Destruction of the Galactic globular cluster
system, ApJ 474, 223.
Gnedin O.Y., Ostriker J.P. (1999). On the self-consistent response of stellar
systems to gravitational shocks, ApJ 513, 626.
Goldstein, H. (1980). Classical Mechanics, 2nd edition (Addison Wesley,
Reading, MA)
Goodman, J. (1983). Core collapse with strong encounters, ApJ 270, 700-10.
Goodman J. (1984). Homologous evolution of stellar systems after core collapse,
ApJ 280, 298.
Goodman J. (1987). On gravothermal oscillations, ApJ 313, 576.
Goodman J. (1989). Late evolution of globular clusters, in Dynamics of Dense
Stellar Systems, ed. Merritt D. (Cambridge University Press, Cambridge).
Goodman J., Heggie D.C., Hut P. (1993). On the exponential instability of
N -body systems, ApJ 415, 715.
Goodman J., Hernquist L. (1991). Hydrodynamics of collisions between binary
stars, ApJ 378, 637.
Goodman J., Hut P., eds (1985). Dynamics of Star Clusters, IAU Symp. 113
(Reidel, Dordrecht)
Goodman J., Hut P. (1989). Primordial binaries & globular cluster evolution,
Nature, 339, 40
Goodman J., Hut P. (1993). Binary-single-star scattering. V - Steady state
binary distribution in a homogeneous static background of single stars, ApJ
403, 271.
Greengard L. (1990). The numerical solution of the N -body problem,
Computers in Phys. Mar/Apr, 142-52
 References 361

Grillmair C.J., Freeman K.C., Irwin M., Quinn P.J. (1995). Globular clusters
with tidal tails: deep two-color star counts, AJ 109, 2553.
Grindlay J.E. (1988). X-ray binaries in globular clusters, in The Harlow-Shapley
Symposium on Globular Cluster Systems in Galaxies (Proc. IAU Symp. 126),
ed. Grindlay J.E., Philip A.G.D. (Kluwer, Dordrecth).
Guhathakurta P., Yanny B., Schneider D.P., Bahcall J.N. (1996). Globular
cluster photometry with the Hubble Space Telescope. V. WFPC study of
M15’s central density cusp, AJ 111, 267.
Gunn, J.E. and Griffin, R.F. (1979). Dynamical studies of globular clusters
based on photoelectric radial velocities of individual stars. I - M3, AJ 84,
752-73.
Gurzadyan V.G., Savvidi G.K. (1986). Collective relaxation of stellar systems,
A&A 160, 203.
Hachisu I., Nakada Y., Nomoto K., Sugimoto, D. (1978). Gravothermal
catastrophe of finite amplitude, Prog. Theor. Phys. 60, 393.
Hamilton W.R. (1853). Lectures on quaternions (Hodges and Smith, Dublin)
Hansen B.M.S. (1999). Cooling models for old white dwarfs, ApJ 520, 680.
Harris H.C., McClure R.D. (1983). Radial velocities of a random sample of K
giant stars and implications concerning multiplicity among giant stars in
clusters, ApJ 265L, 77.
Harrington R.S. (1970). Encounter phenomena in triple stars, AJ 75, 1140.
Harris, W.E. (1996). A catalog of parameters for globular clusters in the Milky
Way, AJ 112, 1487.
Hayli A. (1970). More comparisons of numerical integrations in the gravitational
N -body problem, A&A 7, 249.
Heggie D.C. (1975). Binary evolution in stellar dynamics, MNRAS 173, 729.
Heggie D.C. (1984). Post-collapse evolution of a gaseous cluster model, MNRAS
206, 179.
Heggie D.C. (1991). Chaos in the N -body problem of stellar dynamics, in
Predictability, Stability, and Chaos in N -Body Dynamical Systems, ed. Roy
A.E. (Plenum Press, New York).
Heggie D.C. (2000). A new outcome of binary-binary scattering, MNRAS 318,
61.
Heggie D.C., Aarseth S.J. (1992). Dynamical effects of primordial binaries in
star clusters. I - Equal masses, MNRAS 257, 513.
Heggie D.C., Hut P. (1993). Binary-single-star scattering. IV - Analytic
approximations and fitting formulae for cross sections and reaction rates,
ApJS 85, 347.
Heggie D.C., Inagaki S., McMillan S.L.W. (1994). Gravothermal expansion in
an N -body system, MNRAS 271, 706.
Heggie, D.C. and Ramamani, N. (1989). Evolution of star clusters after core
collapse, MNRAS 237, 757.
Heggie, D.C. and Ramamani, N. (1995). Approximate self-consistent models for
tidally truncated star clusters, MNRAS 272, 317-22.
362 References

Heggie D.C., Sweatman W.L. (1991). Three-body scattering near triple collision
or expansion, MNRAS 250, 555.
Hénon M. (1958). Un calcul amélioré des perturbations des vitesses stellaires,
Ann. d’Astr. 21, 186.
Hénon M. (1960). L’evasion des étoiles hors des amas isolés , Ann.Astr. 23,
668.
Hénon M. (1961). Sur l’evolution dynamique des amas globulaires, Ann. d’Astr.
24, 369.
Hénon M. (1965). Sur l’evolution dynamique des amas globulaires II. – Amas
isolé, Ann. d’Astr. 28, 62.
Hénon M. (1969). Rates of escape from isolated clusters with an arbitrary mass
distribution, A&A 2, 151.
Hénon M. (1971). The Monte Carlo method, ApSpSci 14, 151.
Hénon, M. (1975). Two recent developments concerning the Monte Carlo
method, in Dynamics of Stellar Systems, ed. Hayli A. (Proc. IAU Coll. 69,
D. Reidel, Dordrecht).
Hénon M. (1976). A complete family of periodic solutions of the planar
three-body problem, and their stability, Cel.Mech. 13, 267.
Hénon M. (1982). Vlasov equation, A&A 114, 211.
Hénon M. (1988). Chaotic scattering modelled by an inclined billiard, Physica
D 33, 132-156.
Hénon M., Heiles C. (1964). The applicability of the third integral of motion:
Some numerical experiments, AJ 69, 73.
Herman J. (1710). Metodo d’investigare l’Orbite de’Pianeti, nell’ ipotesi che le
forze centrali o pure le gravita degli stessi Pianeti sono in regione reciproca
de’ quadrati delle distanze, che i medesimi tengono dal Centro, a cui si
dirigono le forze stesse, Giorn. Lett. d’Italia 2, 447.
Hernquist L., Hut P., Kormendy J. (1991). A post-collapse model for the
nucleus of M33, Nature, 354, 376
Hernquist L., Hut P., Makino J. (1993). Discreteness noise versus force errors
in N -body simulations, ApJ 402, L85.
Hill G.W. (1878). Researches in the lunar theory, Am.J.Math. 1, 5, 129 and
245.
Hills J.G. (1975). Encounters between binary and single stars and their effect
on the dynamical evolution of stellar systems, AJ 80, 809.
Hills, J.G. (1980). The effect of mass loss on the dynamical evolution of a stellar
system: analytic approximations, ApJ 235, 986-91.
Hills J.G. (1983). The effects of sudden mass loss and a random kick velocity
produced in a supernova explosion on the dynamics of a binary star of
arbitrary orbital eccentricity - applications to X-ray binaries and to the
binary pulsars, ApJ 267, 322.
Hills J.G. (1990). Encounters between single and binary stars - The effect of
intruder mass on the maximum impact velocity for which the mean change
in binding energy is positive, AJ 99, 979.
 References 363

Hills J.G., Day C.A. (1976). Stellar collisions in globular clusters, Astrophys.
Lett. 17, 87.
Holmberg E. (1941). On the clustering tendencies among the nebulae. II. A
study of encounters between laboratory models of stellar systems by a new
integration procedure, ApJ 94, 385.
Horwitz G., Katz J. (1978). Steepest descent technique and stellar equilibrium
statistical mechanics. III Stability of various ensembles, ApJ 222, 941.
Hurley J., Pols O.R., Tout C.A. (2000). Comprehensive analytic formulae for
stellar evolution as a function of mass and metallicity, MNRAS 315, 543.
Hurley J.R., Tout C.A., Aarseth S.J., Pols O. (2001). Direct N -body modelling
of stellar populations: blue stragglers in M67, MNRAS 323, 630.
Hut P. (1983a). Binary-single star scattering. II - Analytic approximations for
high velocity, ApJ 268, 342.
Hut P. (1983b). The topology of three-body scattering, AJ 88, 1549.
Hut P. (1985). Binary formation and interactions with field stars, in Dynamics
of Star Clusters (Proc. IAU Symp. 113), ed. Hut P., Goodman J. (Reidel,
Dordrecht).
Hut P. (1989). Fundamental timescales in star cluster evolution, in Dynamics
of Dense Stellar Systems, ed. D. Merritt (Cambridge University Press), p.229
Hut P. (1996). Binary formation and interactions with field stars, in Dynamical
Evolution of Star Clusters, IAU Symp. 174, ed. Hut P., Makino J. (Kluwer,
Dordrecht).
Hut P. (1997). Gravitational thermodynamics, Complexity, 3, 38
Hut P., Bahcall J.N. (1983). Binary-single star scattering. I - Numerical
experiments for equal masses, ApJ 268, 319.
Hut P., Djorgovski S. (1992). Rates of collapse and evaporation of globular
clusters, Nature 359, 806.
Hut P., Inagaki S. (1985). Globular cluster evolution with finite-size stars -
Cross sections and reaction rates, ApJ 298, 502.
Hut P., McMillan S., Goodman J., Mateo M., Phinney E.S., Pryor C., Richer
H.B., Verbunt F., Weinberg M. (1992). Binaries in globular clusters, PASP
104, 981.
Hut P., Makino J., eds (1995). Dynamical Evolution of Star Clusters –
Confrontation of Theory and Observations, IAU Symp. 174 (Kluwer,
Dordrecht)
Hut P., Makino J. (1999). Astrophysics on the GRAPE family of special
purpose computers, Science 283, 501.
Hut P., Makino J., McMillan S. (1988). Modelling the evolution of globular
clusters, Nature 336, 31.
Hut P., Makino J., McMillan S. (1995). Building a better leapfrog, ApJ 443L,
93.
Hut P., McMillan S., Romani R.W. (1992). The evolution of a primordial
binary population in a globular cluster, ApJ 389, 527.
Hut P., Paczynski B. (1984). Effects of encounters with field stars on the
evolution of low mass semidetached binaries, ApJ 284, 675.
364 References

Hut P., Verbunt F. (1983). White dwarfs and neutron stars in globular cluster
X-ray sources, Nature 301, 587.
Hut, P. and Verhulst, F. (1981). Explosive mass loss in binary stars - the two
time-scale method, A&A 101, 134-7.
Iben I., Jr. (1985). The life and times of an intermediate mass star - In
isolation/in a close binary, QJRAS 26, 1.
Iben I., Jr., Tutukov A.V. (1997). The lives of stars: from birth to death and
beyond (Part 1), S&T 94, 36.
Iben I., Jr., Tutukov A.V. (1998). How close is close?, S&T 95, 47.
Illingworth G., King I.R. (1977). Dynamical models for M15 without a black
hole, ApJ 218L, 109.
Inagaki S. (1980). The gravothermal catastrophe of stellar systems, PASJ 32,
213.
Inagaki S. (1984). The effects of binaries on the evolution of globular clusters,
MNRAS 206, 149.
Inagaki S., Konishi T. (1993). Dynamical stability of a simple model similar to
self-gravitating systems, PASJ 45, 733.
Inagaki S., Wiyanto P. (1984). On equipartition of kinetic energies in
two-component star clusters, PASJ 36, 391.
Ipser J.R., Semenzato R. (1983). On the effects of strong encounters in stellar
systems. I - A basis for treating anisotropic systems, ApJ 271, 294.
Janes K., ed (1991) The Formation and Evolution of Star Clusters, ASP Conf.
Ser. 13 (ASP, San Francisco)
Jeans J.H. (1902). The stability of a spherical nebula, Phil. Trans. R. Soc. A
199, 49.
Jeans, J.H. (1929). Astronomy and Cosmogony (second edition, Cambridge
University Press, Cambridge); also Dover, New York, 1961
Jeffreys, H. (1929). The Earth, 2nd edition (Cambridge University Press,
Cambridge)
Jernigan J.G., Porter D.H. (1989). A tree code with logarithmic reduction of
force terms, hierarchical regularization of all variables, and explicit accuracy
controls, ApJS 71, 871.
Jha S., Torres G., Stefanik R.P., Latham D., Mazeh T. (2000). Studies
of multiple stellar systems – III. Modulations of orbital elements in the
triple-lined system HD 109648, MNRAS 317, 375.
Johnston H.M., Verbunt F. (1996). The globular cluster population of
low-luminosity X-ray sources, A&A 312, 80.
Johnston K.V., Sigurdsson S., Hernquist L. (1999). Measuring mass-loss rates
from Galactic satellites, MNRAS 302, 771.
Joshi K.J., Rasio F.A., Portegies Zwart S. (2000). Monte Carlo simulations of
globular cluster evolution. I. Method and test calculations, ApJ 540, 969.
Kaliberda V.S., Petrovskaya I.V. (1970). The velocity variation of a star as a
purely discontinuous random process. I. Zero mass stars, Astrofiz. 6, 135.
Kalogera V. (1996). Orbital characteristics of binary systems after asymmetric
supernova explosions, ApJ 471, 352.
 References 365

Kaluzny J., Kubiak M., Szymanski M., Udalski A., Krzeminski W., Mateo M.,
Stanek K.Z. (1998). The optical gravitational lensing experiment. Variable
stars in globular clusters. IV. Fields 104A-E in 47 Tucanae, A&AS 128, 19.
Kandrup H.E., Sideris I.V. (2001). Chaos and the continuum limit in the
gravitational N -body problem. I. Integrable potentials, Phys. Rev. E 64,
056209-1.
Katz J. (1978). On the number of unstable modes of an equilibrium, MNRAS
183, 765.
Katz J.I. (1975). Two kinds of stellar collapse, Nature 253, 698.
Katz J., Lynden-Bell D. (1978). The gravothermal instability in two dimensions,
MNRAS 184, 709.
Kawai A., Fukushige T., Makino J., Taiji M. (2000). GRAPE-5: a
special-purpose computer for N -body simulations, PASJ 52, 659.
Keenan D.W., Innanen K.A. (1975). Numerical investigation of galactic tidal
effects on spherical stellar systems, AJ 80, 290.
Kepler, J. (1609). Astronomia Nova (Prague); Johannes Kepler Gesammelte
Werke, Vol.3, 2nd ed (Munich), 1990; english translation: New Astronomy,
trans. William H. Donohue (Cambridge: Cambridge University Press, 1992)
Kim E., Einsel C., Lee H.-M., Spurzem R., Lee M.G. (2002), Dynamical
evolution of rotating stellar systems – II. Post-collapse, equal mass system,
MNRAS, in press
Kim S.S., Figer D.F., Lee H.M., Morris M. (2000). N -body simulations of
compact young clusters near the Galactic Center, ApJ 545, 301.
King I. (1959). The escape of stars from clusters IV. The retardation of escaping
stars, AJ 64, 351.
King, I.R. (1962). The structure of star clusters. I. An empirical density law,
AJ 67, 471-85.
King, I.R. (1966). The structure of star clusters. III. Some simple dynamical
models, AJ 71, 64-75.
King I.R., Sosin C., Cool A.M. (1995). Mass segregation in the globular cluster
NGC 6397, ApJ 452L, 33.
Kinoshita H., Yoshida H., Nakai H. (1991). Symplectic integrators and their
application to dynamical astronomy, CeMDA 50, 59.
Kiseleva L.G., Eggleton P.P., Orlov V.V. (1994). Instability of close triple
systems with coplanar initial doubly circular motion, MNRAS 270, 936.
Klessen R.S., Burkert A. (2001). The formation of stellar clusters: Gaussian
cloud conditions. II., ApJ 549, 386.
Klessen R., Kroupa P. (2001). The mean surface density of companions in a
stellar-dynamical context, A&A 372, 105.
Kochanek C.S. (1992). The dynamical evolution of tidal capture binaries, ApJ
385, 604.
Kontizas M., Hatzidimitriou D., Bellas-Velidis I., Gouliermis D., Kontizas E.,
Cannon R.D. (1998). Mass segregation in two young clusters in the Large
Magellanic Cloud: SL 666 and NGC 2098, A&A 336, 503.
366 References

Korycansky D.G., Laughlin G., Adams F.C. (2001). Astronomical engineering:

a strategy for modifying planetary orbits, ApSS 275, 349.
Kozai Y. (1962). Secular perturbations of asteroids with high inclination and
eccentricity, AJ 67, 591.
Kroupa P. (1995). Inverse dynamical population synthesis and star formation,
MNRAS 277, 1491.
Kulkarni S., Anderson S.B. (1996). Pulsars in globular clusters, in Dynamical
Evolution of Star Clusters, ed. P. Hut and J. Makino (Kluwer (IAUS 174),
Dordrecht).
Kulkarni S.R., Hut P., McMillan S. (1993). Stellar black holes in globular
clusters, Nature 364, 421.
Kumar P., Goodman J. (1996). Nonlinear damping of oscillations in
tidal-capture binaries, ApJ 466, 946.
Kundić T., Ostriker J.P. (1995). Tidal-shock relaxation: a reexamination of
tidal shocks in stellar systems, ApJ 438, 702-7.
Kustaanheimo P., Stiefel E. (1965). Perturbation theory of Kepler motion based
on spinor regularization, J. Reine Angew. Math. 218, 204.
Kuzmin G.G. (1957). (The effect of stellar encounters and the evolution of
stellar clusters), Tartu Astr. Obs. Publ. 33, 75.
Lada, C.J., Margulis, M. and Dearborn, D. (1984). The formation and early
dynamical evolution of bound stellar systems, ApJ 285, 141-52.
Lagoute C., Longaretti P.-Y. (1996). Rotating globular clusters. I. Onset of the
gravothermal instability, A&A 308, 441.
Lagrange J. (1772). Essai sur le Problème des Trois Corps, in J.-A. Serret, ed,
Oeuvres de Lagrange, 6, 231 (Gauthier-Villars, Paris)
Lambert D., Kibler M. (1988). An algebraic and geometric approach to
non-bijective quadratic transformations, J. Phys. A: Math. Gen. 21, 307.
Lancellotti C., Kiessling M. (2001). Self-similar gravitational collapse in stellar
dynamics, ApJ 549L, 93.
Lançon A., Boily C., eds (2000). Massive Star Clusters, ASP Conf. Ser. 211
(ASP, San Francisco)
Landau L.D., Lifshitz E.M. (1977). Quantum Mechanics (Pergamon, Oxford)
Larson, R.B. (1970). A method for computing the evolution of star clusters,
MNRAS 147, 323-37.
Laskar, J. (1996). Large scale chaos and marginal stability in the solar system,
in J.C. Muzzio, S. Ferraz-Mello and J. Henrard, eds, Chaos in Gravitational
N -Body Systems (Kluwer, Dordrecht); and CeMDA, 64, 115
Lee H.-M. (1987a). Dynamical effects of successive mergers on the evolution of
spherical stellar systems, ApJ 319, 801.
Lee H.-M. (1987b). Evolution of globular clusters including a degenerate
component, ApJ 319, 772.
Leimkuhler B. (1999). Reversible adaptive regularization: perturbed Kepler
motion and classical atomic trajectories, Phil. Trans. R. Soc. A 357, 1101.
Lemaitre G. (1955). Regularization of the three-body problem, Vistas Astron.
1, 207.
 References 367

Leonard P.J.T. (1996). Blue stragglers in star clusters, in The Origins,

Evolution and Destinies of Binary Stars in Clusters, eds. E.F. Milone and
J.-C. Mermilliod, ASP Conf. Ser. 90 (ASP, San Francisco)
Leonard P.J.T., Duncan M.J. (1988). Runaway stars from young star clusters
containing initial binaries. I - Equal-mass, equal-energy binaries, AJ 96, 222.
Levi Civita T. (1906). Sur la résolution qualitative du problème des trois corps,
Acta Math 30, 305.
Lichtenberg A.J., Lieberman M.A. (1992). Regular and Stochastic Motion, 2nd
edition (Springer-Verlag, Berlin)
Lidov M.L. (1963). On the approximate analysis of the evolution of orbits of
artificial satellites, in Problems of the Motion of Artificial Spacecraft (Izdat.
Akad. Nauk. SSSR, Moscow)
—
Lightman A.P., Shapiro S.L. (1977). The distribution and consumption rate of
stars around a massive, collapsed object, ApJ 211, 244.
Lightman A.P., Shapiro S.L. (1978). The dynamical evolution of globular
clusters, Rev. Mod. Phys. 50, 437.
Littlewood J.E. (1986). A Mathematician’s Miscellany, Rev.Ed. (CUP,
Cambridge)
Liverani C. (2000). Interacting particles, in Hard Ball Systems and the Lorentz
Gas, ed. Szasz D. (Springer-Verlag, Berlin).
Livio M. (1996). Cataclysmic variables in globular clusters, in The Origins,
Evolution, and Destinies of Binary Stars in Clusters, ed. Milone E.F.,
Mermilliod J.-C., ASP Conf.Ser. 90 (ASP, San Francisco).
Lombardi J.C., Jr., Rasio F.A., Shapiro S.L. (1996). Collisions of main-sequence
stars and the formation of blue stragglers in globular clusters, ApJ 468, 797.
Louis, P.D. (1990). An anisotropic homological model for core collapse in star
clusters, MNRAS 244, 478-92.
Louis P.D., Spurzem R. (1991). Anisotropic gaseous models for the evolution of
star clusters, MNRAS 251, 408.
Lynden-Bell, D. (1967). Statistical mechanics of violent relaxation in stellar
systems, MNRAS 136, 101-21.
Lynden-Bell D. (1968). Runaway centres, Bull. Astron. 3, 305.
Lynden-Bell, D. and Eggleton, P.P. (1980). On the consequences of the
gravothermal catastrophe, MNRAS 191, 483.
Lynden-Bell, D. and Kalnajs, A.J. (1972). On the generating mechanism of
spiral structure, MNRAS 157, 1-30.
Lynden-Bell D., Wood R. (1968). The gravo-thermal catastrophe in isothermal
spheres and the onset of red-giant structure for stellar systems, MNRAS
138, 495.
Makino, J. (1996). Postcollapse evolution of globular clusters, ApJ 471, 796.
Makino J., Aarseth S.J. (1992). On a Hermite integrator with Ahmad-Cohen
scheme for gravitational many-body problems, PASJ 44, 141.
368 References

Makino, J., Akiyama, K. and Sugimoto, D. (1990). On the apparent

universality of the r1/4 law for brightness distribution in galaxies, PASP 42,
205-15.
Makino J., Hut P. (1991). On core collapse, ApJ 383, 181.
Makino, J. and Taiji M. (1998). Scientific Simulations with Special-Purpose
Computers (Wiley, Chichester)
Mandushev, G., Spassova, N. and Staneva, A. (1991). Dynamical masses for
Galactic globular clusters, A&A 252, 94-9.
Marchal, C. (1990). The Three-Body Problem (Elsevier, Amsterdam). This
reference is an unbeatable source of information about the general three-body
problem.
Marchant A.B., Shapiro S.L. (1980). Star clusters containing massive, central
black holes. III – Evolution calculations, ApJ 239, 685.
Mardling R.A. (1995). The role of chaos in the circularization of tidal capture
binaries. I. The chaos boundary; II. Long-time evolution, ApJ 450, 722 and
732.
Mardling R.A. (2001). Stability in the general three-body problem, in Evolution
of Binary and Multiple Star Systems; A Meeting in Celebration of Peter
Eggleton’s 60th Birthday, eds Ph. Podsiadlowski, S. Rappaport, A.R. King,
F. D’Antona, and L. Burder, ASP Conf. Ser. 229 (ASP, San Francisco)
McGehee R. (1974). Triple collision in the collinear three-body problem, Inv.
Math. 27, 191.
McMillan S.L.W., Aarseth S.J. (1993). An O(N log N ) integration scheme for
collisional stellar systems, ApJ 414, 200.
McMillan S., Hut P. (1994). Star cluster evolution with primordial binaries. 3:
Effect of the Galactic tidal field, ApJ 427, 793.
McMillan S.L.W., Hut P. (1996). Binary–single-star scattering. VI. Automatic
determination of interaction cross sections, ApJ 467, 348.
McMillan S.L.W., McDermott P.N., Taam R.E. (1987). The formation and
evolution of tidal binary systems, ApJ 318, 261.
McVean J.R., Milone E.F., Mateo M., Yan L. (1997). Analyses of the light
curves of the eclipsing binaries in the globular cluster M71, ApJ 481, 782.
Matthews R., Gilmore G. (1993). Is Proxima really in orbit about Alpha CEN
A/B?, MNRAS 261L, 5.
Merritt D. (1981). Two-component stellar systems in thermal and dynamical
equilibrium, ApJ 86, 318.
Merritt D. (1999). Elliptical galaxy dynamics, PASP 111, 129.
Merritt D., Meylan G., Mayor M. (1997). The stellar dynamics of omega
centauri, AJ 114, 1074.
Merritt, D. and Tremblay, B. (1994). Nonparametric estimation of density
profiles, AJ 108, 514-37.
Meylan G., Heggie D.C. (1997). Internal dynamics of globular clusters, Astron.
Astrophys. Rev. 8, 1.
Michie R.W., Bodenheimer P.H. (1963). The dynamics of spherical stellar
systems, II, MNRAS 126, 267.
 References 369

Mihos J.C., Hernquist L. (1996). Gasdynamics and starbursts in major mergers,

ApJ 464, 641.
Mikkola S. (1984). Encounters of binaries. II - Unequal energies, MNRAS 207,
115.
Mikkola S., Aarseth S.J. (1993). An implementation of N -body chain
regularization, CeMDA 57, 439.
Mikkola S., Tanikawa K. (1999). Algorithmic regularization of the few-body
problem, MNRAS 310, 745.
Miller R.H. (1964). Irreversibility in small stellar systems, ApJ 140, 250.
Mladenov I.M. (1991). Geometric quantization of the five-dimensional Kepler
problem, in Bartocci C., Bruzzo U., Cianci R., eds, Differential Geometric
Methods in Theoretical Physics, LNP 375 (Springer-Verlag, Berlin), p.387
Montgomery R. (2001). A new solution to the three-body problem, Not. AMS
48, 471.
Moser, J. (1973). Stable and Random Motions in Dynamical Systems (Princeton
University Press, Princeton)
Moulton F.R. (1914). An Introduction to Celestial Mechanics, Macmillan; also
Dover, New York (1970)
Murali C., Weinberg M.D. (1997). Evolution of the Galactic globular cluster
system, MNRAS 291, 717.
Murphy B.W., Cohn H.N., Hut P. (1990). Realistic models for evolving globular
clusters – II. Post core collapse with a mass spectrum, MNRAS 245, 335.
Muzzio J.C., Vergne M.M., Wachlin F.C., Carpintero D.D. (2001). Stellar
motions in Galactic satellites, in Dynamics of Natural and Artificial Celestial
Bodies (Proc. US - European Celestial Mechanics Workshop, Poznan,
Poland, July 3-7, 2000), ed. H. Pretka-Ziomek, E. Wnuk, P.K. Seidelmann,
D. Richardson (Kluwer, Dordrecht).
Nakada Y. (1978). On the gravothermal catastrophe, PASJ 30, 57.
Nash P.E., Monaghan J.J. (1978). A statistical theory of the disruption of
three-body systems. III - Three-dimensional motion, MNRAS 184, 119.
Neutsch W. (1991). Quaternionic regularisation of perturbed Kepler motion,
Max-Planck-Inst. Math. Bonn, Preprint MPI/91-7
Newell B., Dacosta G.S., Norris J. (1976). Evidence for a central massive object
in the X-ray cluster M15, ApJ 208L, 55.
Newton, Sir I. Mathematical Principles of Natural Philosophy and System
of the World, trans. A. Motte (1729), rev. F. Cajori, 1930. (University of
California Press, 1962)
Ogorodnikov K.F. (1965). Dynamics of Stellar Systems (Pergamon Press,
Oxford)
Okumura S.K., Makino J., Ebisuzaki T., Fukushige T., Ito T., Sugimoto
D., Hashimoto E., Tomida K., f Miyakawa N. (1993). Highly parallelized
special-purpose computer, GRAPE-3, PASJ 45, 329.
Ostriker J. (1964). The equilibrium of polytropic and isothermal cylinders, ApJ
140, 1056.
370 References

Padmanabhan, T. (1990). Statistical mechanics of gravitating systems, Phys.

Rep. 188, 285-362.
Palmer, P.L. (1994). Stability of Collisionless Stellar Systems, Astrophysics and
Space Science Library, Vol.185 (Kluwer, Dordrecht)
Paresce F., De Marchi G. (2000). On the globular cluster initial mass function
below 1M , ApJ 534, 870.
Peacock J.A. (1999). Cosmological Physics (Cambridge University Press,
Cambridge)
Peebles, P.J.E. (1993). Principles of Physical Cosmology (Princeton University
Press, Princeton)
Peterson C.J., Reed B.C. (1987). Structural parameters and luminosities of
globular clusters, PASP 99, 20.
Petit J.-M, Hénon M. (1986). Satellite encounters, Icarus 66, 536-55.
Petrosky T.Y. (1986). Chaos and cometary clouds in the solar system,
Phys.Lett.A 117, 328.
Phinney E.S. (1996). Binaries and pulsars in globular clusters, in The origins,
evolution, and destinies of binary stars in clusters, ed. Milone E.F.,
Mermilliod J.-C., ASP Conf. Ser. 90 (ASP, San Francisco).
Phinney E.S., Sigurdsson S. (1991). Ejection of pulsars and binaries to the
outskirts of globular clusters, Nature 349, 220.
Plummer H.C. (1911). On the problem of distribution in globular star clusters,
MNRAS 71, 460.
Poincaré, H. (1892-9). Les Méthodes Nouvelles de la Mécanique Céleste
(Gauthier-Villars, Paris)
Pollard, H. (1976). Celestial Mechanics (Math. Assoc. Amer., Carus Math.
Mon., 18)
Portegies Zwart S.F. (1996). Interacting Stars, PhD Thesis, Univ. Utrecht
Portegies Zwart S.F., Hut P., Verbunt F. (1997). Star cluster ecology I. A
cluster core with encounters between single stars, A&A 328, 130-42.
Portegies Zwart S.F., McMillan S.L.W. (2000). Black hole mergers in the
universe, ApJ 528, L17.
Portegies Zwart S.F., McMillan S.L.W., Hut P., Makino J. (2001). Star cluster
ecology - IV. Dissection of an open star cluster: photometry, MNRAS 321,
199.
Portegies Zwart S.F., Makino J., McMillan S.L.W., Hut P. (1999). Star cluster
ecology. III. Runaway collisions in young compact star clusters, A&A 348,
117.
Portegies Zwart S.F., Makino J., McMillan S.L.W., Hut P. (2001). How many
young star clusters exist in the Galactic Center?, ApJ 546L, 101.
Portegies Zwart S.F., Verbunt F. (1996). Population synthesis of high-mass
binaries, A&A 309, 179.
Prata S.W. (1971). Dynamic evolution of rich Galactic star clusters. II, AJ 76,
1029.
 References 371

Press, W.H., Teukolsky, S.A., Vetterling, W.T. and Flannery B.P. (1992).
Numerical Recipes in FORTRAN, 2nd edition (Cambridge University Press,
Cambridge)
Prialnik D. (2000). An Introduction to the Theory of Stellar Structure and
Evolution, (Cambridge University Press, Cambridge)
Pryor, C. and Meylan, G. (1993). Velocity dispersions for Galactic globular
clusters, in Structure and Dynamics of Globular Clusters, eds S. Djorgovski
and G. Meylan, (ASP, San Francisco), Conf. Ser., vol. 50, p. 357
Quinlan G., Tremaine S. (1992). On the reliability of gravitational N -body
integrations, MNRAS 259, 505.
Rasio F.R., McMillan S., Hut P. (1995). Binary-binary interactions and the
formation of the PSR B1620-26 triple system in M4, ApJL, 438, L33
Rauch, K.P. and Tremaine, S. (1996). Resonant relaxation in stellar systems,
New Astron. 1, 149–70.
Reichl L.E. (1980). A Modern Course in Statistical Physics (University of Texas
Press, Austin)
Retterer J.M. (1980). The binding-energy distribution of the binaries in a star
cluster, AJ 85, 249.
Rosenbluth M.N., MacDonald W.M., Judd D.L. (1957). Fokker-Planck equation
for an inverse-square force, Phys. Rev. 107, 1.
Ross D.J., Mennim A., Heggie D.C. (1997). Escape from a tidally limited star
cluster, MNRAS 284, 811.
Rubenstein E.P., Bailyn C.D. (1997). Hubble Space Telescope observations of
the post–core-collapse globular cluster NGC 6752. II. A large main-sequence
binary population, ApJ 474, 701.
Ruffert M. (1992). Collisions between a white dwarf and a main-sequence star.
II - Simulations using multiple-nested refined grids, A&A 265, 82.
Saari D.G. (1990). A visit to the Newtonian N -body problem via elementary
complex-variables, Amer. Math. Mon. 97, 105.
Saari, D.G., Xia, Z. (1995). Off to infinity in finite time, Not. Amer. Math. Soc.
42, 538.
Saffer R.A., ed (1993). Blue Stragglers, Proc. Stars Journal Club Miniworkshop
(STScI, Baltimore)
Salpeter, E.E. (1955). The luminosity function and stellar evolution, ApJ 121,
161-7.
Sandage A.R. (1953). The color-magnitude diagram for the globular cluster
M3, AJ 58, 61.
Saslaw W.C., De Young D.S. (1971). On equipartition in galactic nuclei and
gravitating systems, ApJ 170, 423.
Schwarzschild, B. (2000). Theorists and experimenters seek to learn why gravity
is so weak, Phys. Today, 53, 9, 22
Schwarzschild M. (1958). Structure and Evolution of the Stars (Princeton
University Press, Princeton)
Schwarzschild M., Härm R. (1962). Red giants of Population II. II, ApJ 136,
158.
372 References

Seife, C. (2000). A slow carousel ride gauges gravity’s pull, Science, 288, 944
Shu F.H. (1978). On the statistical mechanics of violent relaxation, ApJ 225,
83.
Siegel M.H., Majewski S.R., Cudworth K.M., Takamiya, M. (2001). A cluster’s
last stand: the death of Palomar 13, AJ 121, 935.
Siegel C.L., Moser J.K. (1971). Lectures on Celestial Mechanics
(Springer-Verlag, Berlin)
Sigurdsson S., Hernquist L. (1993). Primordial black holes in globular clusters,
Nature 364, 423.
Sigurdsson S., Phinney E.S. (1993). Binary–single star interactions in globular
clusters, ApJ 415, 631.
Sills A., Faber J.A., Lombardi J.C., Jr., Rasio F.A., Warren A.R. (2001).
Evolution of stellar collision products in globular clusters. II. Off-axis
collisions, ApJ 548, 323.
Simonovic N., Grujic P. (1987). Small-energy three-body systems: I threshold
laws for the Coulomb interaction, J. Phys. B 20, 3427.
Smith H., Jr. (1979). The dependence of statistical results from N -body
calculations on N , A&A 76, 192.
Sobouti Y. (1985). Linear oscillations of isotropic stellar systems. II - Radial
modes of energy-truncated models, A&A 147, 61.
Soker N. (1996). H-function evolution during violent relaxation, ApJ 457,
287-90.
Spitzer, L., Jr. (1940). The stability of isolated clusters, MNRAS 100, 396.
Spitzer L.Jr. (1969). Equipartition and the formation of compact nuclei in
spherical stellar systems, ApJ 158, L139.
Spitzer, L., Jr. (1987). Dynamical Evolution of Globular Clusters (Princeton
University Press, Princeton).
Spitzer L., Jr., Chevalier R.A. (1973). Random gravitational encounters and
the evolution of spherical systems. V. Gravitational shocks, ApJ 183, 565.
Spitzer L., Jr., Härm R. (1958). Evaporation of stars from isolated clusters,
ApJ 127, 544.
Spitzer L., Jr., Hart M.H. (1971). Random gravitational encounters and the
evolution of spherical systems. I. Method, ApJ 164, 399.
Spitzer L., Jr., Mathieu R.D. (1980). Random gravitational encounters and the
evolution of spherical systems. VIII. Clusters with an initial distribution of
binaries, ApJ 241, 618.
Spitzer L., Jr., Shapiro S.L. (1972). Random gravitational encounters and the
evolution of spherical systems. III. Halo, ApJ 172, 529.
Spitzer L., Jr., Shull J.M. (1975). Random gravitational encounters and the
evolution of spherical systems. VI. Plummer’s model, ApJ 200, 339.
Spitzer L., Jr., Thuan T.X. (1972). Random gravitational encounters and the
evolution of spherical systems. IV. Isolated systems of identical stars, ApJ
175, 31.
 References 373

Spurzem R. (1999). Direct N -body simulations, in H. Riffert and K. Werner

(eds), Computational Astrophysics, J. Comp. Appl. Math., (Elsevier,
Amsterdam)
Spurzem R., Baumgardt H. (2001). A parallel implementation of an aarseth
N -body integrator on general and special purpose supercomputers,
MNRAS, submitted
Statler T.S., Ostriker J.P., Cohn H.N. (1987). Evolution of N -body systems
with tidally captured binaries through the core collapse phase, ApJ 316, 626.
Stiefel E.L., Scheifele G. (1971). Linear and Regular Celestial Mechanics,
(Springer-Verlag, Berlin)
Stodólkiewicz J.S. (1985). Monte-Carlo calculations, in Dynamics of Star
Clusters, ed. J. Goodman and P. Hut (Reidel (IAUS 113), Dordrecht).
Stumpff, K. (1962). Himmelsmechanik, Band I (VEB Deutscher Verlag der
Wissenschaften, Berlin)
Sturrock P.A. (1994). Plasma Physics (Cambridge University Press, Cambridge)
Sugimoto D. (1996). A comparative study of globular clusters, in Dynamical
Evolution of Star Clusters, ed. P. Hut and J. Makino, IAS 174 (Kluwer,
Dordrecht).
Sugimoto D., Bettwieser E. (1983). Post-collapse evolution of globular clusters,
MNRAS 204P, 19.
Sundman K.F. (1913). Mémoire sur le problème des trois corps, Acta Math. 36,
105.
Surdin V.G. (1995). Dynamical properties of globular clusters: primordial or
evolutional?, Astron. & Astrophys. Trans. 7, 147.
Sussman G.J., Wisdom J. (1988). Numerical evidence that the motion of Pluto
is chaotic, Sci 241, 433.
Sutantyo W. (1975). The formation of globular cluster X-ray sources through
neutron star-giant collisions, A&A 44, 227.
Synge J.L. (1964). Relativity: The General Theory (North-Holland, Amsterdam)
Szebehely V. (1974). Numerical investigation of a one-parameter family of
triple close approaches occurring in stellar systems, AJ 79, 1449.
Szebehely V., Peters C.F. (1967). Complete solution of a general problem of
three bodies, AJ 72, 876.
Takahashi K. (1995). Fokker-Planck models of star clusters with anisotropic
velocity distributions I. Pre-collapse evolution, PASJ 47, 561.
Takahashi K., Portegies Zwart S.F. (1998). The disruption of globular star
clusters in the Galaxy: a comparative analysis between Fokker-Planck and
N -body models, ApJ 503, 49.
Terlevich E. (1987). Evolution of N -body open clusters, MNRAS 224, 193.
Testa V., Ferraro F.R., Chieffi A., Straniero O., Limongi M., Fusi Pecci F.
(1999). The Large Magellanic Cloud cluster NGC 1866: new data, new
models, new analysis, AJ 118, 2839.
Teuben P.J., Hut P., Levy S., Makino J., McMillan S., Portegies Zwart S.,
Shara M., Emmart C. (2001). Immersive 4D interactive visualization of
large-scale simulations, in Astronomical Data Analysis Software and Systems
374 References

X, ed. Primini F.A., Harnden Jr. F.R., Payne H.E., ASP Conf. Ser. 238
(ASP, San Francisco).
Theis C., Spurzem R. (1999). On the evolution of shape in N -body simulations,
A&A 341, 361.
Thompson, W.B. (1962). An Introduction to Plasma Physics (Pergamon Press,
Oxford)
Thorsett S.E., Arzoumanian Z., Camilo F., Lyne A.G. (1999). The triple pulsar
system PSR B1620-26 in M4, ApJ 523, 763.
Timmes F.X., Woosley S.E., Weaver T.A. (1996). The neutron star and black
hole initial mass function, ApJ 457, 834.
Tolman, R.C. (1936) The Principles of Statistical Mechanics (Clarendon Press,
Oxford)
Toomre A., Toomre J. (1972). Galactic bridges and tails, ApJ 178, 623.
Tout C.A., Aarseth S.J., Pols O.R. (1997). Rapid binary star evolution for
N -body simulations and population synthesis, MNRAS 291, 732.
Tout C.A., Kembhavi A. (1993). Structural changes in stars following tidal
capture, in The Globular Cluster-Galaxy Connection, eds. G.H. Smith and
J.P. Brodie, ASP Conf. Ser. 48 (ASP, San Francisco)
Trager S.C., King I.R., Djorgovski S. (1995). Catalogue of Galactic
globular-cluster surface-brightness profiles, AJ 109, 218.
Tremaine, S. (1981). Galaxy mergers, in The Structure and Evolution of
Normal Galaxies, ed. Fall, S.M. and Lynden-Bell, D. (Cambridge University
Press, Cambridge).
Tremaine, S., Hénon M. and Lynden-Bell, D. (1986). H-functions and mixing
in violent relaxation, MNRAS 219, 285-97.
Tremaine S., Weinberg M.D. (1984). Dynamical friction in spherical systems,
MNRAS 209, 729.
Tsuchiya T., Gouda N. (2000). Relaxation and Lyapunov time scales in a
one-dimensional gravitating sheet system, Phys. Rev. E 61, 948.
van Albada, T.S. (1982). Dissipationless galaxy formation and the r1/4 law,
MNRAS 201, 939-955.
van den Bergh, S. (1980). Star clusters as touchstones for theories of galactic
evolution – a few examples, in Star Clusters, ed. Hesser J.E., IAU Symp. 85
(D. Reidel, Dordrecht)
van Leeuwen F., Le Poole R.S., Reijns R.A., Freeman K.C., de Zeeuw, P.T.
(2000). A proper motion study of the globular cluster ω Centauri, A&A 360,
472.
Verbunt F., Hut P. (1987). The globular cluster population of X-ray binaries,
in The Origin and Evolution of Neutron Stars, ed. D.J. Helfand and J.-H.
Huang, IAUS 125 (Reidel, Dordrecht).
Vesperini E. (1998). Evolution of the mass function of the Galactic globular
cluster system, MNRAS 299, 1019.
Volk O. (1975). Zur Geschichte der Himmelsmechanik: Johannes Kepler,
Leonhard Euler und die Regularisierung, Preprint No.4, Math. Inst., Univ.
Würzburg.
 References 375

von Hoerner, S. (1957). Internal structure of globular clusters, ApJ 125, 451.
von Hoerner S. (2001). How it all started, in Dynamics of Star Clusters and
the Milky Way, ed. S. Deiters, B. Fuchs, A. Just, R. Spurzem and R. Wielen
(ASP (Conf. Ser. 228), San Francisco).
Vivarelli M.D. (1985). The KS-transformation in hypercomplex form and the
quantization of the negative-energy orbit manifold of the Kepler problem,
Celes. Mech. 36, 349.
Waldvogel J. (1982). Symmetric and regularized coordinates on the plane triple
collision manifold, Celes. Mech. 28, 69.
Wannier G.H. (1953). The threshold law for single ionization of atoms or ions
by electrons, Phys. Rev. 90, 817.
Webbink R.F. (1988). Kinematics of the Galactic globular cluster system, in
The Harlow-Shapley Symposium on Globular Cluster Systems in Galaxies
IAU Symp. 126 (Kluwer, Dordrecht), p. 49
Weidemann V. (2000). Revision of the initial-to-final mass relation, A&A 363,
647.
Weinberg M.D. (1994a). Weakly damped modes in star clusters and galaxies,
ApJ 421, 481.
Weinberg M.D. (1994b). Adiabatic invariants in stellar dynamics. 1: Basic
concepts, AJ 108, 1398-1402.
Whittaker, Sir E.T. (1927). A Treatise on Analytical Dynamics, 3rd edition
(Cambridge University Press, Cambridge)
Wijers R.A.M.J., van Paradijs J. (1991). An upper limit to the number of
pulsars in globular clusters, A&A 241L, 37.
Witte M.G., Savonije G.J. (1999). Tidal evolution of eccentric orbits in massive
binary systems. A study of resonance locking, A&A 350, 129.
Yoshida H. (1982). A new derivation of the Kustaanheimo-Stiefel variables,
Celes. Mech. 28, 239.
Youngkins V.P., Miller B.N. (2000). Gravitational phase transitions in a
one-dimensional spherical system, Phys. Rev. E 62, 4583.
 Index

90◦ -deflection distance, 227 apogalacticon, 125

∆, 217 apple, 23
α Cen, 247 applied mathematics, 84, 188
ω
√ Cen, 82, 308 apse, 343
13, 224, 226 area preservation, 214, 231, 232, 349
709, 181, 183, 184, 186, 190 associativity, 160
8, 212 Astrophysics Data System, viii
47 Tuc, 308 asymptote, 66
atom, 215
Aarseth, 32, 234 atomic nucleus, 201
abundances, 302, 308, 316 averaging, 110, 221, 256, 257, 259,
AC211, 306 260
accuracy, 25, 234 axisymmetric potential, 83, 106
action, 72, 152
radial, 72, 73, 115, 148 Bahcall, 229
action-angle variables, 100, 115, 151, Bailey, 258
152 ballet, 321
adiabatic invariant, 72, 115, 126, 152, barycentre, 42, 63, 119, 141, 161
218 energy of motion of, 196
adiabatic perturbation, 201, 220 barycentric frame, 40, 66
air drag, 48 Baumgardt, 297
analytic continuation, 38, 39 Becker, 199, 229
angular momentum, 25, 40, 41, 58, Beta function, 347
64, 65, 70, 72, 83, 95, 116 Bettwieser, 288, 290
conservation of, 68, 257 bifurcation theory, 179
specific, 63, 66, 116 billiards, 198
angular velocity, galactocentric, 120 binaries, 5, 8, 11, 13, 14, 19, 25, 28,
anisotropy, 68, 77, 82, 89, 116, 302 37, 39, 50, 165, 177, 186, 194,
radial, 71, 169 197, 229, 239, 244, 245, 247, 248,
Antonov, 164 252–254, 273, 284–286, 292, 293,

376
 Index 377

300, 302, 303, 305, 307, 309, 315, tidal capture, 273, 277, 278, 300,
321, 330, See also X-ray binaries 304–306, 312
binding energy of, 195–197, 252 two-body, 277
circular, 215, 220 unresolved, 249
circularisation of, 316, 318 visual, 247
contact, 252 black holes, 4, 23, 71, 97, 111, 190,
destruction of, 227, 240, 246, 247, 191, 237, 245, 273, 274, 302, 303,
249, 251, 265, 297 348
distribution of, 198, 202, 245, 249 blow-up, 157, 164, 188
eccentricity of, 160, 215, 219, 227, blue stragglers, 12, 13, 300, 302, 304,
247, 309, 318, 320, 321 305, 310, 315–318, 320, 321
eclipsing, 247, 249, 250, 321 abundance of, 321
energy of, 223 body, iii, iv
evolution of, 11, 245, 272, 276, 284, Boltzmann equation, 92, 94, 176
291, 315–319 collisionless, 45, 46, 57, 58, 61, 74,
84, 88, 89, 93, 94, 97, 100, 110,
formation of, 49, 195, 264, 273,
114, 115, 148
275–277, 284, 291
linearised, 60
rate of, 193, 202, 217, 227, 244,
Boltzmann factor, 20, 202
252, 272
Boltzmann’s constant, 92, 240
stochastic, 244
branch point, 228
hard, 26, 193, 197, 199, 203, 215,
Bulirsch-Stoer method, 230
217–219, 223, 226, 227, 236,
Burrau, 230
239–245, 265, 270, 274, 279, 303,
310, 317, 320, 322, 350 C, 331, 335
hard/soft threshold of, 26, 197, 198, C++, 236
202, 245, 252, 270 c2 h criterion, 43
hardness of, 270, 280 calculus, vi
motion of, 240 canonical ensemble, 20, 21, 183
number of, 285 canonical equations, 231
primordial, 195, 246, 249, 264, 265, canonical momentum, 70
270, 272, 273, 277–279, 281, 283, canonical transformation, 115
284, 291, 294, 295, 298, 305, 321, canonical variables, 100
323, 324, 327 capture, 23, 193, 210–212, 214, 254,
proportion of, 264, 269, 278–280, 266, 307, 308
294, 295, 298 temporary, 309, 317
rate of hardening of, 239, 240, 275, cat map, 108
276 catalyst, 274
soft, 26, 193, 197, 198, 202, 204, catastrophe, 273
215, 218, 226, 227, 240, 251 Cayley numbers, 160
spectroscopic, 321 celestial mechanics, 1, 4, 17, 28, 43
three-body, 269, 273–275, 277–279, central force, 63, 65
281, 287, 294, 295, 327 centrifugal force, 48, 119, 120
378 Index

centripetal force, 48 colour-magnitude diagram, See CM

Ceres, 16 diagram
Chandrasekhar, viii comet Shoemaker-Levy, 130
chaos, vi, 4, 11, 17, 23, 28, 98, 106, comets, 258
109, 122, 203, 289, 291, 292, 313 common envelope, 315, 320
characteristics, 57 commutation relations, 347
Chazy, 199 complex numbers, 156
Chenciner, 212, 213 complex plane, 219
Chernin, 214 complex time, 36–38
Cheshire cat, 298 computational physics, vi, 25, 231
chicken, 84 computer simulation, 4, 6, 21, 136,
circular motion, 65, 69, 71, 73 160, 194, 246, See also N -body
stability of, 136, 342 simulation
circulation, 260 computers
Clifford algebras, 160 analogue, 28
climate, 130 general-purpose, 31
clock cycle, 30 parallel, 27
closest approach, distance of, 133, special-purpose, 2, 28, 96
134, 140, 225, 235, 258 super-, 27, 28
CM diagram, 249, 300, 302, 303, 305, concentration, 7, 89
323 central, 77
coalescence, 273, 310, 317 configuration space, 134, 137
Cohn, ix conformal flatness, 346
collapse, 17, 18, 24, 87 conic section, 65
cold, 286, 342 continuity equation, 56, 93
collective effects, 20, 98, 138 continuum limit, 45
collisions, 36, 37, 43, 45, 46, 75, 148, contour of integration, 219, 222, 228
154, 160, 161, 220, 225, 228, 300, convection, 284
305, 307–310, 342, See also convective derivative, 57, 92
encounters; mergers convergence, 41
atomic, 215 convex function, 103, 110
binary-binary, 318 coordinates
effect of, 310 barycentric, 234
near-, 300 comoving, 288
stellar, iv, 1, 6, 11, 12, 55, 95, 247, curvilinear, 57
304, 306, 307, 311, 316, 317, 320, cylindrical polar, 106
352 generalised, 64, 181
rate of, 300, 307–309, 313 parabolic, 162
three-body, 226 plane polar, 137
triple, 40 rotating, 99, 119, 346
two-body, 38, 39, 129, 224, 226, core, 7, 11, 142, 169, 194, 197, 241,
234, 286 243, 245, 252, 264, 269, 272, 275,
 Index 379

277–280, 282, 284, See also degeneracy, 82, 219

radius, core degrees of freedom, 179–181, 219,
luminosity of, 279, 282 231, 234, 289
volume of, 279 external, 25, 252
core bounce, 277, 279, 280, 285, 287, internal, 25, 59, 252
288, 294 Dejonghe, ix, 73
core collapse, 10, 13, 14, 23, 50, 90, density, 55, 75, 81, 89
92, 164, 185, 187–196, 244, central, 186, 190
272–275, 277–279, 281–288, 292, mass, 55
294, 295, 297, 302, 308, 321, 323, mean, 128
324, 326 phase space, 55, 58, 102, 105, 139,
self-similar model of, 191 141
time scale of, 278, 279 probability, 54, 55
core evolution, 324 projected, 77, 79
Coriolis force, 20, 119, 120, 123 space, 55
correlations, decay of, 135, 136 stellar, 9
correspondence principle, 202 surface, 77, 252
cosmological simulations, 26 density contrast, 186
cosmology, 17, 234, 286 critical, 183
Coulomb logarithm, 18, 91, 142, 143, density profile, 189, 308
146, 147, 153, 169, 185, 189, 190, detailed balance, 218, 226, 227
227, 245, 276, 277, 279, 286, 287 deuterium, 284
Coulomb repulsion, 193 diffusion, 6, 98, 109, 169
coupling constant, 15 Arnold, 109
cpu cost, 27 diffusion coefficient, 95, 149, 150
crackle, 230 diffusion equation, 146
Cray-T3E, 27 Digital Orrery, 28, 31
cross section, 133, 193, 194, 199, dimensional analysis, 244
215–220, 223–225, 235, 236, 238, direct summation, 86
245, 265, 271, 279, 322 disk, circumstellar, 248
differential, 217, 223, 226, 236, 237 dissipation, 99, 177
geometrical, 217 distribution function, 45, 59, 61,
crossing time, 6, 8, 26, 112, 113, 133, 81–83, 93, 98, 100, 102, 114, 129,
134, 138, 147, 148, 153, 190, 227 138
curvature, scalar, 134, 137 coarse-grained, 102–104, 108
cut, 350 evolution of, 98, 115, 148
cutoff, 134, 226, 239, 342 Maxwell-Boltzmann, 153, 348
Maxwellian, 22, 59, 75, 77, 146,
data reduction, 32 147, 153, 170, 172, 175, 176, 202,
Davies, 321 227, 240
de Vaucouleurs Law, 77 lowered, 76
Debye length, 17, 23 one-particle, 55, 59, 74, 84
deflection angle, 133, 141 divergence-free, 231
380 Index

donkeys, 49 close, 4, 6, 10, 25, 55, 135, 216, 217,

Doppler shift, 249 219, 245, 252, 271, 317
drift, 146 cross section of, 237, 241, 279
duality, 154 time scale for, 270
dwarf, 237, 320 dissipative, 272, 273, 300
dynamical astronomy, 43 distant, 10, 94
dynamical friction, 129, 142–147, 170, fast, 197
269, 324, 329 few-body, 291, 321
dynamical systems, 17 four-body, 59, 165, 246, 252, 302,
315, 323, 324, 327
e-folding time, 129, 135, 136, 190, 192 impulsive, 198
E3, 329 independence of, 141
earth, life on, 26, 130 mass loss in, 313
earth-moon-sun system, 253 non-gravitational, 300
eccentric anomaly, 66, 259 parabolic, 227
eccentricity, 66 rate of, 167, 194
eclipses, 263 specification of, 234, 237
ecology, 12, 300, 303 statistics of, 143
effective potential, 65 three-body, 49, 59, 91, 165,
eigenvalue, 224, 225 193–195, 200, 203, 204, 210, 224,
Eilbeck, ix 226, 229, 239, 241, 245, 246, 252,
Einstein, 84 264–266, 269, 273, 275, 277, 298,
Einstein-de Sitter model, 286 302, 319–321, 323, 324, 327
elasticity, 138 close, 223, 224, 229
electromagnetism, 3 two-body, 39, 50, 55, 94, 111, 114,
elephant, 80, 81 129, 138, 139, 147, 154, 164–166,
ellipse, 16, 66, 131, 257 169, 172, 195, 198, 224, 302, 307,
elliptic motion, 67, See also Kepler 312, 316, 321–324
motion, etc close, 225
elliptic-parabolic motion, 210 energy, 25, 41, 46, 58, 74, 85, See also
enclosure, 186 heat transport, thermal
adiabatic, 176, 177, 181, 182, 184 conduction
encounter term, 93, 148 conduction of, 26, 276
encounters, 1, 45, 46, 98, 191, 227, conservation of, 32, 40, 141, 193,
307, 315, 321, See also scattering; 196, 200, 232, 264, 274
exchange; fly-by; ionisation; external, 196, 269, 284, 285
resonance internal, 196, 243, 269, 284
adiabatic, 193, 218, 220, 221, 228 kinetic, 19, 22, 23, 48, 49, 61, 79,
binary-binary, 246, 254, 264–271, 88, 96, 127, 184
278–280, 297, 300, 317 potential, 8, 19, 23, 49, 79, 85, 153,
binary-single, 186, 193, 197, 203, 184
215, 216, 240, 254, 279, 300, 315, specific, 63, 66
317, 320, 322 thermal, 91, 92, 182, 276, 283
 Index 381

energy flux, 92, 93, 284, See also heat equipotential, 121, 122, 248, 314
flux escape, 10, 19, 20, 22, 23, 44, 50, 67,
energy generation, 91, 244, 264, 276, 76, 77, 87, 98, 104, 109, 113, 114,
296, 321, See also heating 119, 120, 122, 124–126, 132, 138,
binary-binary, 278, 280, 281, 298 146, 164–167, 169, 170, 190, 200,
rate of, 91, 194, 244, 272, 275, 279, 201, 203, 207–211, 214, 224, 241,
282, 298, 352 243, 245, 248, 255, 267–269, 284,
self-regulating, 285 286, 297, 298, 302, 309, 327, 329,
stochastic, 291 349
energy hypersurface, 150 of binary, 243
energy level, 202, 215 preferential, 166, 169
energy transport, 10, 94, 97, 183, See escape energy, 77, 104, 114, 122, 123,
also heat transport 242, 243
English Channel, 312 escape rate, 123, 167–169, 174, 201,
entropy, 21, 59, 92, 164, 177–179, 286, 297
181, 276 escape speed, 22, 104, 172, 198, 206,
specific, 91 207, 316
environment, 47, 304, 320 Euclidean norm, 36
galactic, 307 Euler, 160
epicycle, 67, 128 Euler angles, 234
epicyclic frequency, 67, 120 Euler method, 232, 233, 237
epicyclic motion, 121, 123 Euler parameters, 160
equations of motion, 54, 55, 342 Euler-Lagrange equation, 346
equilibrium, 17, 22, 74, 91, 98–101, Euler-Lagrange solutions, 99
103, 108, 109, 121, 138, 153, 164, evaporation, 11, 19, 298
176, 177, 179, 181, 282, 284, 287, evaporative model, 85, 86, 89, 90, 153
288, 292 evolution equation, 84, 129
collisionless, 74 evolutionary track, 325–327
dynamic, 17, 18, 23, 79, 85, 88, 89, exchange, 198, 201, 219, 226, 238,
94, 100, 105, 109, 111–114, 140, 254, 265, 267, 269, 320
147, 227 prompt, 265, 266
hydrostatic, 89 resonant, 200
isothermal, 172 exclusion principle, 104
local thermodynamic, 22 existence, 36
quasi-, 17, 89, 109, 138 exothermic reaction, 264, 265, 267,
statistical, 102, 138 274
thermal, 25, 26, 146, 176, 178, 181, exponential divergence, 15, 108, 129,
185 132, 134, 135, 137
unstable, 99, See also instability external effects, 20, 77, 87, 98, 109,
virial, 8, 83, 96, 113, 137, 142, 153, 119, 138, 302, 309
202, 245
equipartition, 50, 164, 169, 172–174, fast motion, 256
185, 186, 193, 197, 198 Feigenbaum sequence, 272
382 Index

Fermi, 80 gas dynamics, 60

Fermi-Dirac statistics, 104 gas model, 86, 89, 90, 94, 97, 176,
few-body systems, 43, 47, 194, 273 190, 282, 283, 289–292
finite differences, 95 gas pressure, 3
five-body problem, 40 gauge invariance, 159
flattening, 87 general relativity, 4, 16, 17
floating point accelerator, 29 geodesic, 129, 134, 137, 346
fluid, 46, 55–57 geodesic deviation, 134
fluid mechanics, vi geometric progression, 227, 262
fluid models, 86, 89 geometry, 35
fly-by, 198, 201, 265–268, 309 giant, 248, 252, 286, 290, 309, 318,
Fokker-Planck coefficients, 148 320
Fokker-Planck equation, 46, 77, 89, radius of, 308
92–95, 129, 136, 146, 148, 149, Giersz, ix
151, 190, 191 globular clusters, See star clusters,
orbit-averaged, 149, 150, 190 globular
Fokker-Planck model, 86, 90, 96, 174, Gnedin, ix
176, 183, 244, 274, 289, 290, 292, gold, 297, 321
302 Goodman, ix, 123, 185, 290
form factor, 85, 87 GRAPE, vi, 23, 28, 30, 86, 96, 290
FORTRAN, 331 graphics, 331
four-body problem, 39, 40 gravitation, iii, 3, 4
four-colour problem, iv constant of, 15, 55
Fourier series, 221 one-dimensional, 135
fractal, 203, 204, 208, 209, 214 two-dimensional, 49, 96, 184
frequencies of motion, 127 gravitational burning, 239
friction, 47, 182 gravitational focusing, 193, 217, 265,
307, 317
galactic bars, 250 three-body, 224
Galactic Centre, 7, 251, 330 gravitational lenses, 249
galactic disk, 4, 7, 9, 125 gravitational transmutation, 239
galactic nuclei, 11, 14, 251, 307 gravitational waves, 245
galactocentric radius, 329 gravothermal catastrophe, 182, 183
galaxies, iv, 4, 6, 12, 13 gravothermal oscillations, 11, 15, 90,
axisymmetric, 119, 120 92, 272, 283, 288, 290, 292–295,
collisions of, 4, 58, 104, 105 302
rotation of, 128, 330, 352 gravothermodynamics, 18, 51, 52,
spiral, 49 164, 173, 179, 181
Galileo, 31
gas, vi, 20, 46, 59, 61, 75, 80, 89, 93, H-theorem, 103, 104, 110
94, 138, 148, 172, 182, 248, 326 Härm, 169
perfect, 89, 91, 92 Hénon, 106, 151, 166, 212, 284, 288,
stellar, 89, 276 292
 Index 383

halo, 194, 197, 351 hydrogen atom, 202

Hamilton, 157 hydrostatics, 89, 177, 178, 189
Hamiltonian, 100, 161, 231, 233 hyperbola, 67
Hamiltonian dynamics, 42, 57, 70, hyperbolic motion, 66
257 hyperbolic-elliptic motion, 210
hardware, vi hyperbolicity, 136
heartbeat, 26
heat, 22 impact parameter, 137, 139, 140, 222,
heat bath, 21, 45, 47 225, 226, 235
heat capacity, 182–184, 348 impulse approximation, 218
negative, 22, 47, 50, 52, 189, 195 inclination, 258
positive, 52 incompressibility, 57, 58, 104, 210
heat equation, backward, 188 independent variable, 38
heat flux, 52, 186, 283, 288, 352 index, viii
heat transport, 6, 15, 92, 285 individual timesteps, 27
heating, 114, 125, 277, 278, 282, 298 inertial force, 119, 120
binary, 243, 285, 292, 324 infall, 250
binary-binary, 280, 281 infra-red divergence, 19, 21
binary-single, 281 inhomogeneity, 23
rate of, 298 inner binary, 255
Heiles, 106 instability, 22, 60, 67, 71, 109, 172,
helium core flash, 290 182, 183, 204, 256, 272, 288, 290,
helium star, 318 See also stability
Herman, 35 exponential, 130
Hermite integrator, 230, 236, 333, 339 gravothermal, 176, 180, 183–185,
Hernquist, ix 288, See also isothermal model,
high-energy astrophysics, 304, 306 instability of
high-energy physics, 229, 237 Jeans, 45, 60, 61, 68, 71, 87
Hill’s curves, 121 mass segregation/stratification,
Hill’s equations, 123 164, 173, 192
Hill’s problem, 98, 120, 122 radial orbit, 68–70
Hills, 229, 299 integrability, 2, 41, 98, 100, 131, 210
Holmberg, 28 integral, 41, 159
homogeneity, 19 (non-)isolating, 103, 206
homologous evolution, 85, 188, 190, integral equation, 81, 82
298, 325, 327, 341, See also Abel, 83
self-similar evolution integrals of motion, 41, 42, 45, 65,
Hopf bifurcation, 292 103
Hopf map, 26, 129 interaction region, 224
HR diagram, 316, 323 interstellar clouds, 10, 11, 251
human evolution, 26 invariant, 41, 58, 63, 64, 72, 80, 257
Hurley, ix, 321 inverse process, 218, 227
hybrid model, 86 inverse square law, 15, 17, 19, 31
384 Index

ionisation, 198, 202, 219, 226, 236, Kepler’s 3rd Law, 218, 349
240, 266, 267 Kepler’s equation, 38, 66, 155
irrational numbers, 39 Keplerian approximation, 216
isochrone model, 79 kick, 315
isochrone potential, 68, 73 kinematic data, 82
isocline, 75 kinetic theory, v, 22, 89, 145
isolated systems, 64, 109, 111, 119, King model, 46, 74, 76–80, 83, 89,
122, 124, 165, 169, 272, 283, 286, 125, 175, 187, 298, 326
294, 298, 325 multi-mass, 77
lifetime of, 292 King’s Law, 77
isothermal distribution, 76 Kolmogorov-Feller equation, 145
isothermal equation, 75, 184 Kozai cycle, 260, 262
one- and two-dimensional, 83 Kreutz group, 258
isothermal model, 46, 74–77, 82, 176, KS map, 162, 163
177, 181–183, 186–188, 191, 351 KS regularisation, 129, 157
density in, 287, 313, 314 KS transformation, 160, 161
instability of, 295 Kulkarni, ix
singular, 76, 181 Kustaanheimo, 160
isothermal system, 178, 179, 184 Kuzmin, 149
isotropic model, 82
Lada, ix
Jacobi integral, 121, 122, 125 Lagrange, 59
Jacobi’s equation, 134 Lagrange points, 121, 123, 128, 298
Jacobian, 232, 349 Lagrangian, 64, 257
Jeans, viii, 60, 138, 349 Lagrangian shell, 171, 276
Jeans equations, 46, 89 Lagrangian time-derivative, 276
Jeans swindle, 60, 61 Lagrangian variation, 178
Jeans’ Theorem, 59, 74, 83, 103, 148 Landau, 151
jerk, 230 Large Magellanic Cloud, 7
Johnston, ix leapfrog, 161, 231–233
Legendre polynomial, 347
KAM Theorem, vii, 209 Lemaitre, 234
Kepler, 16, 154 Lenz vector, 129, 156
Kepler map, 349 Levi Civita, 43, 156, 158, 160
Kepler motion, 65, 66, 69, 105, 116, Liapounov exponent, 133
131, 139, 161, 163, 166, 253, 255, Liapounov orbit, 124, 128
257, 259, 307, 343 light, 28
Kepler orbit, 156, 158, 233 LIGO, 245
Kepler potential, 67, 68, 343 likelihood, 82
Kepler problem, 26, 45, 53, 58, 64, 73, limit cycle, 292
131, 137, 154, 155, 157, 159, 160 line of sight, 79, 341, 344
five-dimensional, 160 linear series, 164, 179, 182
Kepler’s 1st Law, 65 Liouville’s equation, 75
 Index 385

Liouville’s Theorem, 57, 211, 345 mass transfer, 11, 317, 320
Littlewood, 211 Mathematica, 32
LMXB, 302, 303 Matlab, 331, 332, 335
lockstep, 342 Maupertuis’ Principle, 134
log-normal distribution, 330 Maxwell, 138
logistic map, 289 Maxwell’s equations, 57
long-range force, 3, 18, 19, 21, 96 Maxwellian, See distribution
luminosity, 14, 97, 276 function, Maxwellian
Lynden-Bell, 151, 250, 344 McGehee, 157
Lyttleton, 84 McMillan, ix, 230, 232
McVean, ix
M15, 241, 274, 306 mean field, 98
M31, 13 mean free path, 6, 93, 147, 148
M33, 14 mean motion, 66
M4, 253, 254 measure, 36, 210, 211
M67, 321 memory, 28, 30
M71, 250 Mercury, 241
magnetic fields, 4 mergers, 11, 247, 313, 316, 318, 320,
magnitude, 14 322
Makino, ix, 232 rate of, 322
manifold, 134 Messier, 330
Maple, 32, 62, 342 metric, 134, 137, 346
Marchal, 199, 368 Euclidean, 137
Markov process, 145 microcanonical ensemble, 21, 183
mass microphysics, 129
conservation of, 56 Mikkola, 234
time-dependent, 98 Mikkola & Tanikawa, method of, 233,
mass function, 7, 294, 296, 297, 302, 234
305, 308, 323, 324, 326 million-body problem, iv, v, vii, 300,
Miller-Scalo, 112, 345 324
power law, 112 Milone, ix
mass loss, 87, 112, 113, 117, 119, 125, mixing, 102, 103, 108
165, 243, 282, 303, 323–327 model, 84
impulsive, 315 models of observational data, 46, 79
internal, 111 modes, 23, 79
stellar, 12 molecular dynamics, 31
mass overflow, 304, 315, 317, 320 molecules, 6, 20, 21
mass segregation, 75, 147, 164–166, moment, 129, 139, 149
169–173, 187, 191, 242, 243, 264, moment equations, 89, 92
278, 280, 281, 295–297, 302, 308, moment method, 86, 93
310, 317, 323, 324, 326, 348 moment of inertia, 85
primordial, 173 momentum, 54
mass spectrum, 291, 303 Monte Carlo models, 86, 95
386 Index

Monte Carlo sampling, 136, 237, 245, Noether’s Theorem, 64, 257
317 noise, 82
Montgomery, 212, 213 non-dominant terms, 143, 144, 152
moon, 98, 130, 263, 311, 351 non-parametric methods, 46, 82
Moore, 212 novae, 307
Moser, 35, 204 nuclear reactions, 284
mother board, 30 numerical analysis, 25
movie, 332, 334 numerical errors, 192, 230, 232, 233,
multi-component model, 186 237
multiples, 286
multiplicity, 202 one-body problem, 130
multipole expansion, 221, 257 opacity, 91
open systems, 20
N -body codes, 331 orbit average, 115, 149
N -body equations, 3, 30, 35, 46 orbital motion and relaxation, 148
N -body integration, 28, 331 order (of truncation error), 232
N -body methods, 26, 30, 54, 95, 244 oscillations
N -body models, 84, 86, 100, 290, 321 coupled, 106, 107
N -body problem, 1, 15, 35, 42, 63, damped, 87, 292
85, 98, 99, 129, 130, 137, 246, forced, 312
264, 286, 300, See also one-, two-, oscillator, 80, 349, See also simple
few-, many-, million-body harmonic oscillator
problem, etc damped, 214, 312, 344
collinear, 39 one-dimensional, 179
N -body simulations, 26, 27, 32, 90, oscillatory motion, 214
95, 171, 242, 296, 326, 327 Ostriker, 126
N -body systems, description of, 54, output, 28
74
N -body units, 8, 83, 298, 341 P3 M code, 26
nσv argument, 133, 134, 216, 241 Padmanabhan, 51
NBODYx, 331 Pal 13, 330
nearest neighbour, 136, 141, 346 parabolic motion, 66, 221
negative definiteness, 177 parallelisation, 27
Nemo, 32 parallelism, 29
Neptune, 16 parameters, free, 76, 80
neutron stars, 4, 25, 26, 111, 191, parametric methods, 82
241, 248, 249, 253, 274, 277, 297, partial differential equation, 57, 188
304, 305, 315, See also pulsars particle collider, 229, 237
Newton, 3, 16, 17, 35, 253, 261, 311 PCI, 30
Newton’s equations, 95 penalty function, 82
Newton’s laws of motion, 84 pericentre, 66, 233
Newton’s Theorems, 23, 24, 65, 341 argument of, 259, 260
nodes, 259, 261–263 perigalacticon, 125
 Index 387

perihelion advance, 241 Poisson’s equation, 55, 74, 75, 80, 82,
period, 65, 257 83, 103
orbital, 69, 133, 325, 327 polytrope, 46, 79, 80, 83, 344
radial, 59, 68, 115 pop, 230
period-doubling, 272, 289–291 population, old, 305
periodic orbit, 124, 210, 213, 230 Portegies Zwart, ix, 243, 317
periodic table, 323 position vector, 35, 54
personal computer, 29 post-collapse evolution, 11, 116, 272,
perturbation, 221 274, 286, 291, 294, 296–298
perturbation theory steady, 295
canonical, 255 unsteady, 287
first order, 221 post-collapse expansion, 283, 285,
phase (of binary), 131, 204, 205, 212, 288, 289, 291, 292, 295
233, 235, 349 steady, 298
phase fluid, 57, 58, 104 potential, 45, 57, 64, 74, 75, 81, 243,
phase mixing, 45, 58, 98, 100, 101, 244, See also spherical, Kepler,
104, 106, 108, 109, 206 isochrone, square well potential
phase space, 42, 45, 54, 56–58, 94, central, 78, 269, 282
111, 211 fixed, 63, 83, 98, 103, 106
phase space average, 117 nonintegrable, 106
phase space volume, 69, 95, 116, 149, scaled, 78
152, 223 slowly varying, 115, 129, 148, 149,
physical laws, 84 152
pipeline, 28, 30 time-dependent, 71, 95, 103
plane of motion, 65 potential well, 56, 65, 75, 83, 165,
planet, 254 170, 172, 191, 242, 324
planetary motion, 16 power series, 37, 41, 42
planetary system, 11 precession, 261–263
planets, extrasolar, 14 pressure, 4, 89, 96
plasma, iii, v, 1, 4, 17, 18, 23, 31, 57, pressure ionisation, 204
91, 142, 151 pressure tensor, 68, 89
Platonic solids, 154 prograde motion, 123
Plummer model, 46, 56, 74, 79, 80, proper motion, 82
83, 87, 97, 135, 169, 171, 174, protein folding, 31
187, 191, 242, 292 Proxima Centauri, 247
Plummer potential, 69, 83, 123 Ptolemaic model, 67
Pluto, 28 pulsars, 249, 304, 306, 315, 321, See
Poincaré, vi, 35, 41, 106, 179, 204 also neutron stars
Poincaré map, 206 millisecond, v, 12, 13, 300, 302
point mass, 6, 11, 15, 23, 111, 120, Pythagorean problem, 230
248, 249
Poisson bracket, 57, 100 quadrature, 41, 68, 229
Poisson fluctuations, 62 quadruple, 11
388 Index

democratic, 266, 267 relaxation, 87, 89, 91, 94, 99–102,

hierarchical, 266, 267 104, 109, 123, 135, 138, 144, 145,
temporarily bound, 268 170, 188, 190, 278, 284, 326, 347,
quadrupole approximation, 221, 352 351, See also shock relaxation
qualitative behaviour, 41 collisional, 138
quantum chromodynamics, 16 two-body, 59, 72, 75, 87, 91, 94, 95,
quantum electrodynamics, 16, 159 109, 129, 138, 140, 146, 148, 164,
quantum mechanics, 202 165, 195, 273, 277, 280, 323, 324,
quaternions, 129, 157, 160–162, 347 327, 328, 330
violent, 98, 99, 103, 104, 109, 138,
radial motion, 116, 148, 241, See also 324
period, radial relaxation time, 7, 10, 11, 26, 77, 91,
nearly, 45, 67, 69, 70 97, 129, 135, 146–148, 153, 169,
radiation, stellar, 284 172, 174, 177, 185, 186, 189, 190,
radiative transfer, 91 227, 242, 245, 270, 276, 278–280,
radius 282–284, 286, 325, 327, 330
core, 7, 77–79, 142, 189, 276, 277, central, 187, 189, 190
279, 285, 287, 294, 295, 298, 324 half-mass, 79, 147, 174, 187, 192,
292, 297, 328
half-mass, 7, 26, 78, 79, 85, 97, 142,
remnant, 111, 174, 315
147, 153, 171, 172, 284, 285, 295,
repulsive force, 21
296, 325
resonance, 201, 203, 204, 219, 236,
Lagrangian, 87
265, 309, 310, 317, 320
stellar, 25, 236, 246, 249, 258, 270,
democratic, 201, 236, 268, 269
277, 303, 306, 308–310, 314, 316,
hierarchical, 202, 236, 265, 266
317
resonance (of frequencies), 49, 68, 69,
tidal, 7, 14, 78, 118, 122, 123, 127,
71, 106, 110, 127, 256, 345
128, 296, 298, 325, 326
retrograde motion, 121–123
virial, 8, 14, 79, 85, 96, 97, 270
Ricci tensor, 346
random numbers, 235, 317
Riemann sheet, 157, 159
quasi-, 235
rigour, 35
random walk, 254, 347 Roche lobe, 318, 320
range of scales, 2, 25, 26 Roche lobe overflow, 248, 309, 317,
Rasio, ix 318
rational numbers, 39 Roche radius, 309
reaction rate, 216, 239, 240 Roche surface, 309
rectilinear motion, 130, 228 Rosenbluth potentials, 144
Redman, iv rosette, 10, 147
reduced mass, 66, 196, 240, 310 rotations (geometric), 158
regular motion, 106, 108, 122 Ruffert, ix, 213
regularisation, 26, 39, 58, 226, 234 runaway, 182, 187
chain, 234 Runge-Kutta methods, 32, 230
numerical, 160 Runge-Kutta-Nystrom methods, 231
 Index 389

Runge-Lenz vector, 156, 160 bulge, 125, 302, 324

disk, 125–127, 302, 323, 324, 327,
saddle point, 219, 222 328, 330
saddle-node bifurcation, 179 tidal, 127, 324
Salaris, ix short-range force, 3
Salpeter, 112 Siegel, vii
sand grain, 202, 348 Siegel exponent, 226
Sandage, 304 simple harmonic motion, 65, 155,
satellites, 45, 47, 48, 182 156, 343
coorbital, 121, 124 simple harmonic oscillator, 26, 64, 69,
scale radius, 79 121, 129, 154, 155, 159, 160, 173,
scaled potential, 78, 80, 313 231
scaled radius, 177 simple harmonic potential, 68, 101
scaling, 8, 15, 21, 85, 111, 189, 229, Simó, 213
235 singularity, 36–40, 42, 129, 137, 154,
scaling model, 86 155, 157, 160, 188, 220, 222
scattering, 16, 37, 205, 216, 236, 238 collision, 2, 36, 38, 39, 58, 96, 154
atomic, 193, 194, 198, 201, 223, 224 Sitnikov’s problem, 40, 131, 132, 193,
gravitational, 193 204, 206, 207, 209, 214, 349
irregular, 124 slingshot, 166
nuclear, 201 Smale horseshoe, 193
numerical, 230, 236, 254, 265, 317 small systems, 21
scattering problem, 196 small-angle scattering, 89, 95, 139,
Schuster, 79 141, 144, 145, 164, 166
Schwarzschild, 290 Smithsonian Museum, 28
screening, 3, 17 smooth potential, 95
sea, 346 smoothing parameter, 95
secular evolution, 109, 256 snap, 230
self-consistent model, 74, 100 snapshot, 28
self-similar behaviour, 164, 174, 185, software, vi, 2, 30, 32
188, 190, 327 SOHO, 124
self-similar evolution, 188, 189, 244, solar neighbourhood, 11, 251, 308
See also homologous evolution solar system, 3, 4, 43, 253
semi-major axis, 65, 66 stability of, 28, 109
semi-minor axis, 66 solution, 2, 36, 41, 42, 341
sensitive dependence, 124, 130, 131, exact, 136
135, 136, 193, 201, 203, 204, 209, specific heat, negative, 49, 196, 288,
210 See also heat capacity
serial code, 27 spherical potential, 41, 45, 58, 61, 63,
shadowing, 136 64, 67, 71, 72, 115, 117, 124, 126,
shock heating, 127 144
shock relaxation, 126, 138, 329 non-, 120
shocking spin, 160
390 Index

spinors, 157, 160 rotation of, 7

spiral arms, 250 system of, 304, 309, 328, 330
Spitzer, viii, 146, 169 open, 251, 321
Spurzem, ix young, 7, 173, 264, 309
square-well potential, 169 star counts, 82
stability, 19, 23, 45, 52, 61, 98, 99, Star Wars, 247
102, 109, 128, 130, 177, 179, 180, Stark effect, 162
186, 212, 213, 246, 292, See also starlab, 32, 238, 271, 317, 331, 340
instability stars, 300, See also abundances,
asymptotic, 99, 102 dwarf, neutron stars, blue
dynamical, 94, 95 stragglers, yellow stragglers,
gravothermal, 22, 188 white dwarfs
numerical, 236 age of, 302
secular, 285 BHB, 302
thermal, 177 blue horizontal branch, 303
thermodynamic, 74, 177 degenerate, 277, 305
stable focus, 292 evolution of, v, 11, 12, 21, 22, 26,
stable node, 292 96, 111, 165, 186, 243, 252, 282,
star clusters, iii, iv, 3, 4, 11, 27, 47, 302, 304, 305, 308, 309, 315–317,
81, 89, 185, 193, 201, 215, 251, 323–327
252, 280, 284, 291, 304, 310, 316, faint horizontal branch, 303
See also stellar systems formation of, 10
binaries in, 246, 247, See also high-velocity, 280
binaries hot horizontal branch, 303
death rate of, 330 internal energy of, 310, 313
dissolution of, 9, 294, 303, 324, 326, lifetime of, 26, 116, 308
330 long-period variable, 26
evolution of, 6, 25, 47, 188, 301, main sequence, 284, 303, 305, 308,
302, 323, 324 311, 316–318
evolutionary track of, 325 orbits of, 10
globular, iv–vi, 1, 2, 4–6, 9, 11, 13, radius of, See radius, stellar
14, 26, 28, 77, 83, 84, 109, 119, rotation of, 311, 313, 316
124, 172, 190, 198, 246, 248, structure of, 21, 89, 91, 177, 285,
251–254, 264, 274, 295, 302, 303, 313
305–310, 316, 321, 329 variable, 302
ages of, 326, 328 young, 247
basic facts about, 7 station keeping, 124
binaries in, 247, 249, 270, 315 statistical accuracy, 235
colour gradients in, 306 statistical description, 74, 84
destruction of, 327, 328 statistical mechanics, 2, 23, 54, 59,
lifetime of, 327 102, 130, 135, 247
masses of, 330 statistical physics, 183
radial distribution of, 330 statistical results, 136
 Index 391

statistical weight, 202 temperature, 19, 47, 49, 50, 52, 75,
steepest descents, 219, 222 89, 92, 172, 174, 176, 181–184,
Stefan-Boltzmann constant, 91 187, 240, 276, 289, 348
stellar dynamics, 1, 4, 6, 10, 21, 23, temperature fluctuations, 290
25, 26, 28, 43, 46, 84, 86, 89, 99, temperature gradient, 182, 189,
109, 135, 138 283–285
collisional, 1, 6, 45, 55, 129 temperature inversion, 272, 288, 290,
collisionless, 1, 6, 45, 55, 84, 95 292
stellar systems, 63, 64, 84, 114, See test particle, 198
also star clusters Teuben, ix
bound, 112 theorems, 35
cold, 68 theoretical physics, 15
cool, 113 thermal conduction, 6, 27, 92, 283,
dense, 11, 14, 45, 84, 86, 187 284
expansion of, 111, 116, 285–290 thermal conductivity, 89, 91, 97
gross evolution of, 46, 94, 114, 272, thermal time scale, 318
See also post-collapse evolution thermodynamics, v, 18–22, 45, 47,
51, 76, 93, 164, 176, 179, 276
hot, 227
first law of, 92, 276
lifetime of, 297
three-body problem, v, 2, 16, 40–43,
one-component, 186
99, 121, 129, 131, 132, 163, 193,
radius of, 341
202, 223, 226, 230, 234, 255, See
rotating, 87, 188
also triples
spherical, 148
analytic solution of, 42
stereographic projection, 161
isosceles, 228
Stiefel, 160 numerical integration of, 230
Stodólkiewicz, 298 periodic solutions of, 99, 212
Stosszahlansatz, 23 restricted, 16, 131, 204
Stumpff variables, 66 three-body processes, 281, See also
Sugimoto, 288, 290 collisions; encounters
sun, 124, 247, 282 threshold scattering, 194
lifetime of, 26 tidal approximation, 121
radius of, 306 tidal boundary, 297, 324, 327, 330
sungrazers, 258 tidal bulge, 312
supernova, 112, 315 tidal circularisation, 320
surface brightness, 14, 77, 79 tidal cutoff, 118, 187, 298
surface of section, 122, 204, 207, 208 tidal disruption, 325, 327
survival triangle, 330 tidal dissipation, 313, 326
symmetry, 41, 64, 80, 120 tidal encounters, 277, 316
symplecticness, 35, 36, 161, 231, 232, tidal field, 122, 123, 125, 131, 169,
234 298, 309, 314, 324, 327
tidal heating, 125
target, 229, 237 tidal interaction, 313
392 Index

tidal overflow, 324, 327 two-body processes, 277, 278, 280,

tidal radius, See radius, tidal 281, 308, 321, See also collisions;
tidal stripping, 326 encounters
tidal tails, 124, 125, 298 Tycho, 16
tidal truncation, 127, 272, 294, 296,
ultra-violet divergence, 20, 21, 50
297, 302, 325, 327
unfolding, 157
tide, 312, 313, 352
uniform medium, 60
lunar, 121, 128, 253, 263, 311
uniform sphere, 18, 24
stellar, 210
cold, 60
time reversal, 23, 161, 199, 202, 211,
units, 1, 8, 79, 234
218, 232–234, 348
universe
time-translation invariance, 210 age of, 6, 25
timesteps, 26, 37, 96, 231, 290, 332, expansion of, 17, 60, 288
333, See also individual timesteps unix, 32
hierarchical, 96
topology, 43 Valtonen, 214
torque, 63, 260, 262 variables
toy model, 22, 292 conjugate, 257
tree, 30 extensive, 18, 19, 21
tree code, 26, 86, 95 intensive, 18, 19
triple expansion, 224 variation of parameters, 128
triples, 11, 19, 96, 201, 239, 241, 247, variation, first and second, 177–179,
249, 250, 253, 286 181
bound, 211, 267 variational equations, 131, 135, 136
evolution of, 253, 255, 259, 269 velocity
radial, 13, 82, 83, 247–249
formation of, 253–255
transverse, 83
hierarchical, 212, 236, 246, 253–256,
velocity dispersion, 7, 75, 76, 92, 189,
258, 261–263, 265–269, 302
289
evolution of, 259, 260
central, 77, 188, 244
formation of, 264
one-dimensional, 79, 188
lifetime of, 201
projected, 79
stability of, 255, 268, 269 transverse, 68
temporarily bound, 201, 203, 268 velocity distribution, 75, 167
turbulence, 31 isotropic, 94, 116, 142, 168, 235
turnoff, 305, 320 velocity space, 73
twist map, 206 velocity vector, 54
two-body approximation, 224 Verlet algorithm, 231
two-body problem, v, 16, 35, 36, 38, virial equilibrium, See equilibrium,
63, 129, 154, See also Kepler virial
problem virial estimate, 352
perturbed, 155 virial oscillations, 85, 87, 96, 133
 Index 393

virial radius, See radius, virial wind, stellar, 112

virial ratio, 113 Wlasow equation, 71
virial theorem, 19, 23, 46, 53, 85, 87, work, 242
88, 93, 96, 147, 174, 184, 347, 348 workstation, 28, 29
virialisation, 126, 127, 227
viscosity, 248 X-ray binaries, 12, 302, 306
Vlasov equation, 57 low-mass, 304
von Hoerner, 27 X-ray sources, v, 13, 246, 248, 249,
300, 304, 321
wake, 143, 144, 152 Xia, 40, 41
Wannier, 224
weak scattering, 139, 140, 165, See yellow stragglers, 305
also small-angle scattering
Zare, 234
Weinberg, 128
zero-velocity surface, 121, 122
whisky, 26
zoo, 321
white dwarfs, 13, 26, 111, 249, 253,
binary, 320
305, 310, 318, 320, 330