Adv. Polym. Sci.

(2005) 173:105149
DOI:10.1007/b99427
Springer-Verlag Berlin Heidelberg 2005
Thermostat Algorithms for Molecular Dynamics
Simulations
Philippe H. Hnenberger
Laboratorium fr Physikalische Chemie, ETH Zrich, CH-8093 Zrich, Switzerland
phil@[Link]
Abstract Molecular dynamics simulations rely on integrating the classical (Newtonian)
equations of motion for a molecular system and thus, sample a microcanonical (constant-
energy) ensemble by default. However, for compatibility with experiment, it is often desirable
to sample congurations from a canonical (constant-temperature) ensemble instead. A modi-
cation of the basic molecular dynamics scheme with the purpose of maintaining the temper-
ature constant (on average) is called a thermostat algorithm. The present article reviews the
various thermostat algorithms proposed to date, their physical basis, their advantages and their
shortcomings.
Keywords Computer simulation, Molecular dynamics, Canonical ensemble, Thermostat al-
gorithm
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
2 Ensembles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
3 Thermostat Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
3.1 Temperature in the Monte Carlo Algorithm. . . . . . . . . . . . . . . . . . . . . . . . 118
3.2 Temperature Relaxation by Stochastic Dynamics . . . . . . . . . . . . . . . . . . . 120
3.3 Temperature Relaxation by Stochastic Coupling . . . . . . . . . . . . . . . . . . . 122
3.4 Temperature Constraining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
3.5 Temperature Relaxation by Weak Coupling . . . . . . . . . . . . . . . . . . . . . . . 127
3.6 Temperature Relaxation by the Extended-System Method . . . . . . . . . . . 129
3.7 Generalizations of the Previous Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 137
4 Practical Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
5 Appendix: Phase-Space Probability Distributions . . . . . . . . . . . . . . . . 138
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
List of Abbreviations and Symbols
CM center of mass
MC Monte Carlo
106 Philippe H. Hnenberger
MD molecular dynamics
NMR nuclear magnetic resonance
SD stochastic dynamics
Kronecker delta symbol or Dirac delta function
h Heaviside step function
k
B
Boltzmanns constant
U instantaneous potential energy
K instantaneous kinetic energy
H Hamiltonian
E energy (thermodynamical)
H enthalpy
L Hill energy
R Ray enthalpy
T temperature (instantaneous)
T temperature (thermodynamical)
T
o
reference temperature (heat bath)
= (k
B
T
o
)
1
V volume (instantaneous)
V volume (thermodynamical)
P pressure (instantaneous)
P pressure (thermodynamical)
NN number of particles (n-species, instantaneous)
N number of particles (n-species, thermodynamical)
chemical potential (n-species, instantaneous)
chemical potential (n-species, thermodynamical)
p
sys
system linear momentum
L
sys
system angular momentum
p
box
box linear momentum
L
box
box angular momentum
N
d f
number of internal degrees of freedom
N
c
number of geometrical constraints
N
r
number of external degrees of freedom
m
i
mass of atom i
r
o
i
real velocity of atom i
r
i
peculiar velocity of atom i
F
i
force on atom i
p
i
momentum of atom i
R
i
stochastic force on atom i

i
friction coefcient of atom i
velocity scaling factor
t timestep

T
temperature relaxation time
collision frequency (Andersen thermostat)

B
relaxation time (Berendsen thermostat)
Q mass of the time-scaling coordinate (Nos-Hoover thermostat)
Thermostat Algorithms 107

NH
effective relaxation time (Nos-Hoover thermostat)
L
e
extended-system Lagrangian (Nos-Hoover thermostat)
H
e
extended-system Hamiltonian (Nos-Hoover thermostat)
E
e
extended-system energy (Nos-Hoover thermostat)
1
Introduction
Classical atomistic simulations, and in particular molecular dynamics (MD) simu-
lations, have nowadays become a common tool for investigating the properties of
polymer [1] and (bio-)molecular systems [2, 3, 4, 5, 6, 7]. Due to their remarkable
resolution in space (single atom), time (femtosecond), and energy, they represent a
powerful complement to experimental techniques, providing mechanistic insight into
experimentally observed processes. However, direct comparison with experiment re-
quires that the boundary conditions imposed on the simulated system are in adequa-
tion with the experimental conditions. The term boundary condition is used here to
denote any geometrical or thermodynamical constraint enforced within the whole
system during the simulation. One may distinguish between hard and soft boundary
conditions. A hard boundary condition represents a constraint on a given instanta-
neous observable, i.e. it is satised exactly at any timepoint during the simulation.
A soft boundary condition represents a constraint on the average value of an observ-
able, i.e. the corresponding instantaneous value is allowed to uctuate around the
specied average. The denition of a soft boundary condition generally also requires
the specication of a timescale for which the average observable should match the
specied value. There exist four main types of boundary conditions in simulations:
1. Spatial boundary conditions include the denition of the shape of the simulated
system and the nature of its surroundings. In molecular simulations, one typi-
cally uses either: (i) vacuum boundary conditions (solute molecule surrounded
by vacuum); (ii) xed boundary conditions (solute-solvent system surrounded
by vacuum, e.g. droplet [8, 9, 10, 11, 12, 13, 14, 15]); (iii) periodic boundary
conditions (solute-solvent system in a space-lling box, surrounded by an in-
nite array of periodic copies of itself [16, 17]). In the two former cases, the effect
of a surrounding solvent can be reintroduced in an implicit fashion by a modi-
cation of the system Hamiltonian. Typical modications are the inclusion of: (i)
solvation forces accounting for the mean effect of the solvent [18, 19, 20, 21];
(ii) stochastic and frictional forces accounting for the effect of collisions with
solvent molecules [22, 23, 24, 25, 2]; (iii) forces at the system boundary to mim-
ick a system-solvent interface [8, 9, 10, 11, 12, 13, 14, 15]. Spatial boundary
conditions are hard boundary conditions, because they apply strictly to all con-
gurations during a simulation.
2. Thermodynamical boundary conditions include the denition of the n + 2
thermodynamical quantities characterizing the macroscopic state of a (mono-
plastic) n-component system (for systems under vacuum boundary conditions,
only n + 1 quantities are required because the volume is not dened while
108 Philippe H. Hnenberger
the thermodynamical pressure is zero). These quantities can be selected form
pairs of extensive and intensive quantities including: (i) the number of particles
(N {N
i
| i = 1...n}) or chemical potential ( {
i
| i = 1...n}) of all
species; (ii) the volume V or pressure P; (iii) the energy E (or a related ex-
tensive thermodynamical potential) or temperature T. The selected set of n +2
quantities, together with their reference (macroscopic) values, dene the thermo-
dynamical ensemble that is sampled during a simulation (Table 1). By default,
MD simulations sample microstates in the microcanonical (NV E) ensemble. By
applying specic modications to the system Hamiltonian or equations of mo-
tion, it is possible to maintain instead a constant temperature, pressure or chemi-
cal potential for the different species (or any combination of these changes). The
thermodynamical boundary conditions involving extensive quantities should be
treated as hard boundary conditions, while those involving intensive quantities
should be soft.
3. Experimentally derived boundary conditions are used to explicitly enforce agree-
ment between a simulation and some experimental result. These may be applied
to enforce, e.g., the reproduction of (average) electron density maps from X-ray
crystallography [27, 28, 29, 30], or the agreement with (average) interatomic
distances and J-coupling constants from NMR measurements [31, 32, 33, 30].
Since experiments always provide averages over a given time and number of
molecules, experimentally derived boundary conditions should be handled as
soft boundary conditions.
4. Geometrical constraints can also be considered as boundary conditions. A typ-
ical example is the use of bond-length constraints in simulations [34, 35, 36,
37, 38, 39], which represent a better approximation to the quantum-mechanical
behavior of high-frequency oscillators (h k
B
T) compared to the classical
treatment [40]. Since they are satised exactly at every timepoint during a simu-
lation, geometrical constraints represent hard boundary conditions.
The present article is concerned with one specic type of thermodynamical
boundary condition, namely the imposition of a constant (average) temperature dur-
ing MD simulations by means of thermostat algorithms. The simultaneous enforce-
ment of a constant (average) pressure [51, 52, 53, 54, 55, 46, 56, 57, 58, 59, 60,
61, 53, 62, 63, 64, 65, 66, 67, 68, 69, 70] or chemical potential [71, 72, 73, 74]
will not be considered here. The discussion is also restricted to systems under ei-
ther vacuum or periodic boundary conditions, i.e., isolated systems. This implies
that the Hamiltonian is time-independent, and invariant upon translation or rotation
of the whole system. This Hamiltonian may contain terms accounting for the mean
effect of the environment (e.g., implicit-solvation term), as long as it still satises the
above conditions. The only exception considered here (whenever explicitly stated) is
the possible inclusion of stochastic and frictional forces as applied in stochastic dy-
namics (SD) simulations, or of randomcollisional forces as applied in the stochastic-
coupling (Andersen) thermostat. Finally, it should be stressed that the inclusion of
geometrical constraints during a simulation affects the statistical mechanics of the
sampled microstates [75]. This is mainly because in the presence of such constraints,
Thermostat Algorithms 109
Table 1. The eight thermodynamical ensembles, and the corresponding independent and de-
pendent variables. Intensive variables are the chemical potential for all n species ( {
i
|
i = 1...n}), the pressure (P) and the temperature (T). Extensive variables are the number of
particles for all species (N {N
i
| i = 1...n}), the volume (V), and the energy (E), enthalpy
(H = E + PV), Hill energy (L = E

i
N
i
), or Ray enthalpy (R = E + PV

i
N
i
).
Note that grand-ensembles may be open with respect to a subset of species only (e.g., semi-
grand-canonical ensemble). The generalized ensemble is not a physical ensemble, because its
size is not specied (no independent extensive variable). Isothermal ensembles are discussed
in many standard textbooks. Specic references are given for the (less common) adiabatic
ensembles
Independent Dependent Ensemble
NV E PT Microcanonical [41, 42, 43, 44, 45]
NVT PE Canonical
NPH VT Isoenthalpic-isobaric [46, 47, 48, 45]
NPT V H Isothermal-isobaric (Gibbs)
V L NPT Grand-microcanonical [49, 45]
VT NPL Grand-canonical
PR NVT Grand-isothermal-isobaric [50, 45]
PT NV R Generalized
the kinetic energy of the system cannot be written in a conguration-independent
way (unless the constraints are exclusively involved in fully-rigid atom groups, e.g.,
rigid molecules). This restriction limits the validity of a number of equations pre-
sented in this article. However, many results are expected to remain approximately
valid for systems involving a small proportion of constrainted degrees of freedom,
and no attempt is made here to derive forms including explicitly the effect of geo-
metrical constraints.
2
Ensembles
An isolated system is characterized by a time-independent, translationally invariant
and rotationally invariant Hamiltonian. Integration of the classical equations of mo-
tion for such a system leads, in the limit of innite sampling, to a trajectory mapping
a microcanonical (NV E) ensemble of microstates
1
. Assuming an innite numerical
precision, this is also what a standard MD simulation will deliver.
The laws of classical mechanics also lead to two additional conserved quantities,
namely the linear momentump
sys
of the system, and the angular momentumL
sys
of
1
Thermodynamical ensembles are generally dened without the constraint of Hamiltonian
translational and rotational invariance, in which case the previous statement is not entirely
correct. In the present article, however, the terminology of Table 1 will be (loosely) retained
to encompass ensembles where this invariance is enforced. The statistical mechanics of
these latter ensembles must be adapted accordingly [76, 77, 78, 79, 80, 81]. This requires
in particular the introduction of a modied denition for the instantaneous temperature,
relying solely on internal degrees of freedom and kinetic energy (Sect. 3).
110 Philippe H. Hnenberger
the system around its center of mass (CM). In simulations under periodic boundary
conditions, the two quantities refer to the innite periodic system. However, in this
case, if the linear momentum p
box
of the computational box is also conserved, the
corresponding angular momentum L
box
is not. This is because correlated rotational
motion in two adjacent boxes exert friction on each other, leading to an exchange
of kinetic energy with the other (internal) degrees of freedom of the system. Note
that the physical properties of a molecular system are independent of p
sys
. However,
they depend on L
sys
, because the rotation of the system leads to centrifugal forces.
For this reason, L
sys
should be added to the list of independent variables dening the
ensemble sampled. Whenever L
sys
is not given, it generally implicitly means that
L
sys
= 0. The use of L
sys
= 0 in simulations under periodic boundary conditions
(overall uniform rotation of the innite periodic system) is actually impossible, be-
cause it would lead to non-periodic centrifugal forces. Finally, it should be specied
that the total energy E of the system is dened here so as to exclude the kinetic en-
ergy contributions corresponding to the overall translation and rotation of the system
(so that E is independent of p
sys
and L
sys
).
Because the independent variables of the microcanonical ensemble are all exten-
sive, they should be strictly conserved (i.e., time-independent) during the course of
a simulation. The corresponding dependent variables, namely the chemical potential
, the pressure P, and the temperature T, are not conserved. In a non-equilibrium
simulation, these quantities may undergo a systematic drift. In an equilibrium sim-
ulation, the corresponding instantaneous observables (denoted by , P, and T) will
uctuate around well-dened average values , P, and T. Two important comments
should be made concerning the previous statement. First, the instantaneous observ-
ables , P, and T are not uniquely dened. The instantaneous temperature is gener-
ally related to the total kinetic energy of the system (Eq. (8)), and the instantaneous
pressure to the total virial and kinetic energy. However, alternative denitions are
available (differing from the above by any quantity with a vanishing equilibrium
average), leading to identical average values in equilibrium situations, but to differ-
ent uctuations. Second, a microcanonical ensemble at equilibrium could equally
well be specied by stating that N, V, and E are conserved, and giving the val-
ues of instead of N, P instead of V, or T instead of E (as long as at least one
extensive variable is specied). However, such a specication would be rather un-
natural as well as inapplicable to non-equilibrium situations. Furthermore, only the
natural variables for dening a given thermodynamical ensemble (Table 1) are ei-
ther time-independent or characterized by vanishing uctuations in the limit of a
macroscopic system. Finally, it should be stressed that computer simulations can-
not be performed at innite numerical precision. As a consequence, quantities which
are formally time-independent in classical mechanics may still undergo a numerical
drift in simulations. In microcanonical simulations, this is typically the case for E,
as well as p
sys
and L
sys
(vacuum boundary conditions), or p
box
(periodic boundary
conditions).
Unfortunately, the microcanonical ensemble that comes out of a standard MD
simulation does not correspond to the conditions under which most experiments are
Thermostat Algorithms 111
carried out. For comparison with experiment, the following ensembles are more use-
ful, which involve one or more intensive independent variables (Table 1):
1. In the canonical ensemble (NV T), the temperature has a specied average
(macroscopic) value, while the instantaneous observable representing the total
energy of the system (i.e., the Hamiltonian H) can uctuate. At equilibrium, the
root-mean-square uctuations
E
of the Hamiltonian around its average value E
are related to the system isochoric heat capacity, c
V
, through [16]

2
E
=
_
H
2
_
NVT
H
2
NVT
= k
B
T
2
c
V
. (1)
The uctuations
T
of the instantaneous temperature T (dened by Eq. (8)) in a
canonical ensemble are given by [16]

2
T
=
_
T
2
_
NVT
T
2
NVT
= 2N
1
d f
T
2
, (2)
where N
d f
is the number of internal degrees of freedom in the system (Eq. (9)).
These uctuations vanish in the limit of a macroscopic system, but are often
non-negligible for the system sizes typically considered in simulations.
2. In the isothermal-isobaric (Gibbs) ensemble (NPT), the pressure has (just as
the temperature) a specied average value, while the instantaneous volume Vof
the system can uctuate. At equilibrium, the root-mean-square uctuations
V
of the instantaneous volume around its average value V are related to the system
isothermal compressibility,
T
, through [16]

2
V
= V
2

NPT
V
2
NPT
= Vk
B
T
T
. (3)
The root-mean-square uctuations
H
of the instantaneous enthalpy H + PV
around its average value H are related to the system isobaric heat capacity, c
P
,
through [16]

2
H
= (H + PV)
2

NPT
H + PV
2
NPT
= k
B
T
2
c
P
. (4)
Both the instantaneous temperature T and the instantaneous pressure P will
uctuate around their corresponding macroscopic values, the magnitude of these
uctuations vanishing in the limit of a macroscopic system.
3. The grand-canonical ensemble (V T) has a constant volume and temperature
(as the canonical ensemble), but is open for exchanging particles with a sur-
rounding bath. In this case, the chemical potential of the different species has
a specied average, while the instantaneous value NN of the number of particles
can uctuate. For a one-component system at equilibrium, the uctuations
N
of the instantaneous number of particles around its average value N are related
to the system isothermal compressibility,
T
, through [16]

2
N
=
_
N
2
_
VT
N
2
VT
= N
2
V
1
k
B
T
T
. (5)
112 Philippe H. Hnenberger
The root-mean-square uctuations
L
of the instantaneous Hill energy HN
around its average value L are given by [16]

2
L
=
_
(H N)
2
_
VT
H N
2
VT
=k
B
T
2
_
L(, V, T)
T
_
V
. (6)
Three other combinations of variables are possible (Table 1), but the correspond-
ing ensembles [45] are of more limited practical relevance. The last combination
(generalized ensemble) is not physical, because its size is not specied (no indepen-
dent extensive variable). Note that although MD samples the microcanonical ensem-
ble by default, the basic Monte Carlo (MC; [82, 83, 84]) and stochastic dynamics
(SD; [22, 23, 24, 25, 2]) algorithms sample the canonical ensemble.
Performing a MDsimulation in an other ensemble than microcanonical requires a
means to keep at least one intensive quantity constant (on average) during the simula-
tion. This can be done either in a hard or in a soft manner. Applying a hard boundary
condition on an intensive macroscopic variable means constraining a corresponding
instantaneous observable to its specied macroscopic value at every timepoint dur-
ing the simulation (constraint method). Remember, however, that the choice of this
instantaneous observable is not unique. In contrast, the use of a soft boundary condi-
tion allows for uctuations in the instantaneous observable, only requiring its aver-
age to remain equal to the macroscopic value (on a given timescale). Typical meth-
ods for applying soft boundary conditions are the penalty-function, weak-coupling,
extended-system and stochastic-coupling methods [85]. These methods will be dis-
cussed in the following sections in the context of constant-temperature simulations.
Although there are many ways to ensure that the average of an instantaneous quan-
tity takes a specied value, ensuring that the simulation actually samples the correct
ensemble (and in particular provides the correct uctuations for the specic instan-
taneous observable in the given ensemble) is much more difcult.
3
Thermostat Algorithms
A modication of the Newtonian MD scheme with the purpose of generating a ther-
modynamical ensemble at constant temperature is called a thermostat algorithm. The
use of a thermostat can be motivated by one (or a number) of the following reasons:
(i) to match experimental conditions (most condensed-phase experiments are per-
formed on thermostatized rather than isolated systems); (ii) to study temperature-
dependent processes (e.g., determination of thermal coefcients, investigation of
temperature-dependent conformational or phase transitions); (iii) to evacuate the heat
in dissipative non-equilibrium MD simulations (e.g., computation of transport coef-
cients by viscous-ow or heat-ow simulations); (iv) to enhance the efciency of
a conformational search (e.g., high-temperature dynamics, simulated annealing); (v)
Thermostat Algorithms 113
to avoid steady energy drifts caused by the accumulation of numerical errors during
MD simulations
2
.
The use of a thermostat requires the denition of an instantaneous temperature.
This temperature will be compared to the reference temperature T
o
of the heat bath to
which the system is coupled. Following from the equipartition theorem, the average
internal kinetic energy K of a system is related to its macroscopic temperature T
through
K = K =
1
2
k
B
N
d f
T (7)
where k
B
is Boltzmanns constant, N
d f
the number of internal degrees of freedom
of the system, and K its instantaneous internal kinetic energy. Dening the instanta-
neous temperature T at any timepoint as
T =
2
k
B
N
d f
K , (8)
one ensures that the average temperature T is identical to the macroscopic temper-
ature T. This denition is commonly adopted, but by no means unique. For example,
the instantaneous temperature could be dened based on the equipartition principle
for only a subset of the internal degrees of freedom. It may also be dened purely on
the basis of conguration, without any reference to the kinetic energy [87, 88].
In the absence of stochastic and frictional forces (see below; Eq. (17)), a few
degrees of freedom are not coupled (i.e., do not exchange kinetic energy) with the
internal degrees of freedom of the system. These external degrees of freedom corre-
spond to the system rigid-body translation and, under vacuum boundary conditions,
rigid-body rotation. Because the kinetic energy associated with these external de-
grees of freedom can take an arbitrary (constant) value determined by the initial
atomic velocities, they must be removed from the denition of the system internal
temperature. Consequently, the number of internal degrees of freedom is calculated
as three times the total number N of atoms in the system, minus the number N
c
of
geometrical constraints, i.e.
N
d f
= 3N N
c
N
r
. (9)
The subtraction of constrained degrees of freedom is necessary because geometrical
constraints are characterized by a time-independent generalized coordinate associ-
ated with a vanishing generalized momentum (i.e., no kinetic energy). A more for-
mal statistical-mechanical justication for the subtraction of the external degrees of
2
A thermostat algorithm (involving explicit reference to a heat-bath temperature T
o
) will
avoid systematic energy drifts, because if the instantaneous temperature is forced to uctu-
ate within a limited range around T
o
, the energy will also uctuate within a limited range
around its corresponding equilibrium value. To perform long microcanonical simulations
(no thermostat), it is also advisable to employ an algorithm that will constrain the energy
to its reference value E
o
(ergostat algorithm [86]).
114 Philippe H. Hnenberger
freedom in the case of periodic boundary conditions can be found elsewhere [45].
A corresponding derivation for vacuum boundary conditions has, to our knowledge,
never been reported. When stochastic and frictional forces are applied, as in SD,
these forces will couple the rigid-body translational and rotational degrees of free-
domwith the internal ones. In this case all degrees of freedomare considered internal
to the system. Thus, Eq. (9) is to be used with N
r
= 0 in the presence of stochastic
and frictional forces, and otherwise with N
r
= 3 under periodic boundary conditions
or N
r
= 6 under vacuum boundary conditions. Similarly, the instantaneous internal
kinetic energy is dened as
K =
1
2
N

i=1
m
i
r
2
i
, (10)
where the internal (also called peculiar) velocities r
i
are obtained from the real
atomic velocities r
o
i
by excluding any component along the external degrees of free-
dom
3
. These corrected velocities are calculated as
r
i
=

r
o
i
if N
r
= 0
r
o
i
r
o
CM
if N
r
= 3
r
o
i
r
o
CM
I
1
CM
(r
o
) L
o
CM
(r
o
i
r
o
CM
) if N
r
= 6
, (11)
where r
o
CM
is the coordinate vector of the system center of mass (CM), L
o
CM
the
system angular momentum about the CM, and I
CM
is the (conguration-dependent)
inertia tensor of the system relative to the CM. The latter quantity is dened as
I
CM
(r) =
N

i=1
m
i
(r
i
r
CM
) (r
i
r
CM
) , (12)
where a b denotes the tensor with elements , equal to a

. Application of
Eq. (11) ensures that
N

i=1
m
i
r
i
= 0 for N
r
= 3 or 6 (13)
and (irrespective of the origin of the coordinate system)
N

i=1
m
i
r
i
r
i
= 0 for N
r
= 6. (14)
Equation (13) is a straightforward consequence of the denition of r
o
CM
. Equa-
tion (14) is proved by using
o
CM
(r
o
i
r
o
CM
) = r
o
i
r
o
CM
where
o
CM
=
I
1
CM
(r
o
) L
o
CM
is the angular velocity vector about the CM. Thus, the linear and an-
gular momenta of the internal velocities vanish, as expected.
3
It is assumed that the velocities r
o
i
are already exempt of any component along possible
geometrical constraints.
Thermostat Algorithms 115
Because the instantaneous temperature is directly related to the atomic internal
velocities (Eqs. (8) and (10)), maintaining the temperature constant (on average)
in MD simulations requires imposing some control on the rate of change of these
velocities. For this reason, thermostat algorithms require a modication of Newtons
second law
4
r
i
(t ) = m
1
i
F
i
(t ) . (15)
In the present context, this equation (and the thermostatized analogs discussed be-
low) should be viewed as providing the time-derivative of the internal velocity r
i
dened by Eq. (11). In turn, r
i
is related to the real atomic velocity r
o
i
through the
inverse of Eq. (11), namely
r
o
i
=

r
i
if N
r
= 0
r
i
+ r

CM
if N
r
= 3
r
i
+ r

CM
+I
1
CM
(r
o
) L

CM
(r
o
i
r
o
CM
) if N
r
= 6
, (16)
where r

CM
and L

CM
are constant parameters determined by the initial velocities
r
o
i
(0). This distinction between real and internal velocities is often ignored in stan-
dard simulation programs. Many programs completely disregard the problem, while
others only remove the velocity component along the external degrees of freedomfor
the computation of the temperature (but do not use internal velocities in the equations
of motion). However, as discussed in Sect. 4, this can have very unpleasant conse-
quences in practice. In the following discussion, it is assumed that the equation of
motion (Eq. (15) or any thermostatized modication) is applied to the internal veloci-
ties dened by Eq. (11), while the atomic coordinates are propagated simultaneously
in time using the real velocities r
o
i
dened by Eq. (16).
The prototype of most isothermal equations of motion is the Langevin equation
(as used in SD; see Sect. 3.2), i.e.
r
i
(t ) = m
1
i
F
i
(t )
i
(t ) r
i
(t ) +m
1
i
R
i
(t ) , (17)
where R
i
is a stochastic force and
i
a (positive) atomic friction coefcient. Many
thermostats avoid the stochastic force in Eq. (17) and use a single friction coefcient
for all atoms. This leads to the simplied form
r
i
(t ) = m
1
i
F
i
(t ) (t) r
i
(t ) . (18)
In this case, loses its physical meaning of a friction coefcient and is no longer
restricted to positive values. A positive value indicates that heat ows from the sys-
tem to the heat bath. A negative value indicates a heat ow in the opposite direction.
Note that if Eq. (18) was applied to the real velocities r
o
i
(as often done in simulation
programs) instead of the internal velocities r
i
, the linear and angular momenta of the
system would not be conserved (unless they exactly vanish).
4
It is assumed that the forces F
i
are exempt of any component along possible geometrical
constraints.
116 Philippe H. Hnenberger
Any algorithm relying on the equation of motion given by Eq. (18) is smooth
(i.e., generates a continuous velocity trajectory) and deterministic
5
. It is also time-
reversible if is antisymmetric with respect to time-reversal
6
.
Practical implementations of Eq. (18) often rely on the stepwise integration of
Newtons second law (Eq. (15)), altered by the scaling of the atomic velocities after
each iteration step. In the context of the leap-frog integrator
7
[89], this can be written
r
i
_
t +
t
2
_
= (t ; t) r

i
_
t +
t
2
_
= (t; t)
_
r
i
(t
t
2
) +m
1
i
F
i
(t )t
_
, (20)
where (t; t) is a time- and timestep-dependent velocity scaling factor. Imposing
the constraint
8
(t ; 0) = 1, one recovers Eq. (18) in the limit of an innitesimal
timestep t , with
(t) = lim
t 0
(t ; t ) 1
t
=
(t; t)
(t)
|
t =0
. (21)
Note that for a given equation of motion, i.e., a specied form of (t), Eq. (21) does
not uniquely specify the scaling factor (t; t). It can be shown that Eq. (20) retains
the original accuracy of the leap-frog algorithm if the velocity-scaling factor applied
to atom i is chosen as [90]

i
(t ; t ) = 1 (t)t +
_

2
(t )
2
+
(t)F
i
(t )
2m
i
r
i
(t )
_
(t )
2
. (22)
From a thermodynamical point of view, some thermostats can be proved to gen-
erate (at constant volume and number of atoms) a canonical ensemble in the limit of
innite sampling times (and within the usual statistical-mechanical assumptions of
equal a priori probabilities and ergodicity). More precisely, some thermostats lead to
a canonical ensemble of microstates, i.e., microstates are sampled with a statistical
5
The advantages of deterministic algorithms are that (i) the results can be exactly re-
produced (in the absence of numerical errors), and (ii) there are well-dened conserved
quantities (constants of the motion). In the case of Eq. (18), the constant of the motion is
C = K(t ) +U(t ) +2
_
t
0
dt K(t ) (t ) . (19)
6
Considering a given microstate, time-reversibility is achieved if the change dt dt
(leading in particular to r r, r r, and r r) leaves the equation of motion for the
coordinates unaltered (while the velocities are reversed). Clearly, this condition is satised
for Eq. (18) only if the corresponding change for is .
7
The implementation of thermostats will only be discussed here in the context of the leap-
frog integrator. However, implementation with other integrators is generally straightfor-
ward.
8
An algorithm with (t ; 0) = 1 would involve a Dirac delta function in its equation of
motion.
Thermostat Algorithms 117
weight proportional to e
H
where = (k
B
T
o
)
1
. In this case and in the absence
of geometrical constraints, expressing the Hamiltonian in Cartesian coordinates as
H(r, p) = U(r) +K(p), U being the potential energy, the probability distribution
of microstates may be written
(r, p) =
e
H(r,p)
_
dr
_
dpe
H(r,p)
=
e
U(r)
_
dr e
U(r)
e
K(p)
_
dpe
K(p)
. (23)
Integrating this expression over either momenta or coordinates shows that the dis-
tribution is also canonical in both congurations (i.e., congurations are sampled
with a statistical weight proportional to e
U
) and momenta (i.e., momenta are sam-
pled with a statistical weight proportional to e
K
). In Cartesian coordinates, such a
canonical distribution of momenta reads

p
(p) =
e
K(p)
_
dpe
K(p)
=
3N

i
e
(2m
i
)
1
p
2
i
_
dp
i
e
(2m
i
)
1
p
2
i
=
3N

i
p( p
i
) , (24)
where Eq. (10) was used together with p
i
= m
i
r
i
. Noting that p( r
i
) = m
i
p( p
i
)
and evaluating the required Gaussian integral, this result shows that internal veloci-
ties obey a Maxwell-Boltzmann distribution, i.e., the velocity components r
i
appear
with the probability
p( r
i
) =
_
m
i
2
_
1/2
e
(1/2)m
i
r
2
i
. (25)
Note that the above statements do not formally hold in the presence of geometrical
constraints, but are generally assumed to provide a good approximation in this case.
Some other thermostats only generate a canonical ensemble of congurations, but
not of microstates and momenta. This is generally not a serious disadvantage for the
computation of thermodynamical properties, because the contribution of momenta to
thermodynamical quantities can be calculated analytically (ideal-gas contribution).
Finally, there also exists thermostats that generate distributions that are canonical
neither in congurations nor in momenta.
From a dynamical point of view, assessing the relative merits of different ther-
mostats is somewhat subjective
9
. Clearly, the congurational dynamics of a system
will be affected by the timescale of its instantaneous temperature uctuations, and a
good thermostat should reproduce this timescale at least qualitatively. However, the
direct comparison between experimental thermostats (e.g., a heat bath surrounding
9
An objective question, however, is whether the thermostat is able to produce correct time-
correlation functions (at least in the limit of a macroscopic system). Since transport coef-
cients (e.g., the diffusion constant) can be calculated either as ensemble averages (Einstein
formulation) or as integrals of a time-correlation function (Green-Kubo formulation), at
least such integrals should be correct if the thermostat leads to a canonical ensemble. When
this is the case, it has been shown that the correlation functions themselves are also correct
at least for some thermostats [91, 86].
118 Philippe H. Hnenberger
a macroscopic system, or the bulk medium around a microscopic sample of matter)
and thermostats used in simulations is not straightforward. The reason is that exper-
imental thermostats, because they involve the progressive diffusion of heat from the
system surface towards its center (or inversely), lead to inhomogeneities in the spa-
tial temperature distribution within the sample. On the other hand, the thermostats
used in simulations generally modify instantaneously and simultaneously the veloc-
ities of all atoms irrespective of their locations, and should lead to an essentially
homogeneous temperature distribution.
One may nevertheless try to quantify the timescale of the temperature uctuations
to be expected in a thermostatized simulation. This timescale can be estimated based
on a semi-macroscopic approach [55]. Consider a systemcharacterized by an average
temperature T, in contact with a heat bath at a different temperature T
o
. By average
temperature, it is meant that the quantity T is spacially-averaged over the entire
system and time-averaged over an interval that is short compared to the experimental
timescale, but long compared to the time separating atomic collisions. The difference
between T and T
o
may result, e.g., froma natural uctuation of T within the system.
From macroscopic principles, the rate of heat transfer from the heat bath to the the
systemshould be proportional to the temperature difference T
o
T and to the thermal
conductivity of the system. Thus, the rate of change in the average temperature can
be written (at constant volume)

T(t ) = c
1
v

E(t) =
1
T
[T
o
T(t )] (26)
with the denition

T
=
1
V
1/3
c
v

1
, (27)
where c
v
is the system isochoric heat capacity, V the system volume, and a dimen-
sionless constant depending on the system shape and on the temperature inhomo-
geneity within the system. For a given system geometry (e.g., spherical) and initial
temperature distribution (i.e., T(x, 0)), a reasonable value for could in principle be
estimated by solving simultaneously the ux equation
J(x, t) = T(x, t), (28)
where J(x, t) is the energy ux through a surface element perpendicular to the direc-
tion of the vector, and the conservation equation
T(x, t)
t
= Vc
1
v
J(x, t ) . (29)
Eq. (26) implies that, at equilibrium, the natural uctuations of T away from T
o
decay exponentially with a temperature relaxation time
T
, i.e.
T(t ) = T
o
+[T(0) T
o
] e

1
T
t
. (30)
Note that on a very short timescale (i.e., of the order of the time separating atomic
collisions), the instantaneous temperature T(t ) is also affected by important stochas-
tic variations (see Sect. 3.2). Only on an intermediate timescale does the mean effect
Thermostat Algorithms 119
of these stochastic uctuations result in an exponential relaxation for T(t ). However,
because stochastic variations contribute signicantly to the instantaneous tempera-
ture uctuations, a thermostat based solely on an exponential relaxation for T(t)
leads to incorrect (underestimated) temperature uctuations (see Sect. 3.5).
To summarize, although assessing whether one thermostat leads to a better de-
scription of the dynamics compared to another one is largely subjective, it seems
reasonable to assume that: (i) thermostats permitting temperature uctuations are
more likely to represent the dynamics correctly compared to thermostats constrain-
ing the temperature at a xed value; (ii) thermostats with temperature uctuations
are more likely to represent the dynamics correctly when these uctuations occur
at a timescale (measured in a simulation, e.g., as the decay time of the temperature
autocorrelation function) of the order of
T
(Eq. (26)), and when the dynamics is
smooth (continuous velocity trajectory). These differences will be more signicant
for small systems, where the temperature uctuations are of of larger magnitudes (the
corresponding root-mean-square uctuations scale as N
1/2
, see Eq. (2)) and higher
frequencies (the corresponding relaxation times scale as N
1/3
, see Eq. (27)).
A summary of the common thermostats used in MD simulations, together with
their main properties, is given in Table 2. The various algorithms are detailed in the
following sections.
3.1
Temperature in the Monte Carlo Algorithm
Although the present discussion mainly focuses on thermostatized MD, the simplest
way to generate a thermodynamical ensemble at constant temperature is to use the
MC algorithm [82, 83, 84]. This algorithm does not involve atomic velocities or
kinetic energy. Random trial moves are generated, and accepted with a probability
p = min{e
U
, 1} (31)
depending on the potential energy change U associated with the move and on
the reference temperature T
o
. Following this criterion, moves involving rigid-body
translation and, under vacuum boundary conditions, rigid-body rotation are always
accepted because they do not change the potential energy. For this reason, the corre-
sponding degrees of freedomare external to the system. Note also that under vacuum
boundary conditions, the centrifugal forces due to the rigid-body rotation of the sys-
tem, which would be included in a MD simulation, are absent in the MC procedure.
Therefore, MC samples by default an ensemble at zero angular momentum. It can be
shown that the ensemble generated by the MC procedure represents (at constant vol-
ume) a canonical distribution of congurations. The modication of the MC scheme
to sample other isothermal ensembles (including the grand-canonical ensemble [92])
is possible. Modications permitting the sampling of adiabatic ensembles (e.g., the
microcanonical ensemble [92, 93, 94]) have also been devised. The MC procedure is
non-smooth, non-deterministic, time-irreversible, and does not provide any dynami-
cal information.
120 Philippe H. Hnenberger
Table 2. Characteristics of the main thermostat algorithms used in MD simulations. MD:
molecular dynamics (generates a microcanonical ensemble, only shown for comparison); MC:
Monte Carlo (Sect. 3.1); SD: stochastic dynamics (with
i
> 0 for at least one atom; Sect. 3.2);
A: MD with Andersen thermostat (with > 0; Sect. 3.3); HE: MD with Hoover-Evans ther-
mostat (Sect. 3.4); W: MD with Woodcock thermostat (Sect. 3.4); HG: MD with Haile-Gupta
thermostat (Sect. 3.4); B: MD with Berendsen thermostat (with t <
B
< ; Sect. 3.5);
NH: MD with Nos-Hoover thermostat (with 0 < Q < ; Sect. 3.6). MD is a limiting case
of SD (with
i
= 0 for all atoms), A (with = 0), B (with
B
), and NH (with Q ,
(0) = 0). HE/Wis a limiting cases of B (with
B
= t ) and is a constrained formof NH. HG
is also a constrained form of NH. Deterministic: trajectory is deterministic; Time-reversible:
equation of motion is time-reversible; Smooth: velocity trajectory is available and continuous.
Energy drift: possible energy (and temperature) drift due to accumulation of numerical errors;
Oscillations: possible oscillatory behavior of the temperature dynamics; External d.o.f.: some
external degrees of freedom (rigid-body translation and, under vacuum boundary conditions,
rotation) are not coupled with the internal degrees of freedom. Constrained K: no kinetic en-
ergy uctuations; Canonical in H: generates a canonical distribution of microstates; Canonical
in U: generates a canonical distribution of congurations. Dynamics: dynamical information
on the system is either absent () or likely to be unrealistic (; constrained temperature
or non-smooth trajectory), moderately realistic (+; smooth trajectory, but temperature uc-
tuations of incorrect magnitude), or realistic (++; smooth trajectory, correct magnitude of
the temperature uctuations). The latter appreciation is rather subjective and depends on an
adequate choice of the adjustable parameters of the thermostat
MD MC SD A HE W HG B NH
Deterministic + + + + + +
Timereversible + + + + +
Smooth + + + + + + +
Energy drift + +
Oscillations +
External d.o.f. + + + + + + +
Constrained K + + +
Canonical in H + + +
Canonical in U + + + + + +
Dynamics ++ ++ + ++
Eqn. of motion 15 17 41 46 51 52 57 78,79
3.2
Temperature Relaxation by Stochastic Dynamics
The SD algorithm relies on the integration of the Langevin equation of motion [22,
23, 95, 96, 97, 98, 99, 100] as given by Eq. (17). The stochastic forces R
i
(t ) have the
following properties
10
: (i) they are uncorrelated with the velocities r(t

) and system-
atic forces F
i
(t

) at previous times t

< t; (ii) their time-averages are zero; (iii) their

mean-square components evaluate to 2m
i

i
k
B
T
o
; (iv) the force component R
i
(t )
10
More complex SD schemes can be used, which incorporate time or space correlations
in the stochastic forces. It is also assumed here that the friction coefcients
i
are time-
independent.
Thermostat Algorithms 121
along the Cartesian axis is uncorrelated with any component R
j
(t

) along axis
unless i = j , = , and t

= t. The two last conditions can be combined into the

relation
R
i
(t )R
j
(t

) = 2m
i

i
k
B
T
o

i j

(t

t) . (32)
It can be shown that a trajectory generated by integrating the Langevin equation
of motion (with at least one non-vanishing atomic friction coefcient
i
) maps (at
constant volume) a canonical distribution of microstates at temperature T
o
.
The Langevin equation of motion is smooth, non-deterministic and time-
irreversible. Under vacuum boundary conditions and aiming at reproducing bulk
properties, it may produce a reasonable picture of the dynamics if the mean effect
of the surrounding solvent is incorporated into the systematic forces, and if the fric-
tion coefcients are representative of the solvent viscosity (possibly weighted by
the solvent accessibility). If SD is merely used as a thermostat in explicit-solvent
simulations, as is the case, e.g., when applying stochastic boundary conditions to a
simulated system [101, 11, 12], some care must be taken in the choice of the atomic
friction coefcients
i
. On the one hand, too small values (loose coupling) may cause
a poor temperature control. Indeed, the limiting case of SD where all friction coef-
cients (and thus the stochastic forces) are set to zero is MD, which generates a
microcanonical ensemble. However, arbitrarily small atomic friction coefcients (or
even a non-vanishing coefcient for a single atom) are sufcient to guarantee in
principle the generation of a canonical ensemble. But if the friction coefcients are
chosen too low, the canonical distribution will only be obtained after very long sim-
ulation times. In this case, systematic energy drifts due to accumulation of numerical
errors may interfere with the thermostatization. On the other hand, too large values
of the friction coefcients (tight coupling) may cause the large stochastic and fric-
tional forces to perturb the dynamics of the system. In principle, the perturbation of
the dynamics due to stochastic forces will be minimal when the atomic friction co-
efcients
i
are made proportional to m
i
. In this case, Eqs. (17) and (32) show that
the root-mean-square acceleration due to stochastic forces is identical for all atoms.
In practice, however, it is often more convenient to set the friction coefcients to a
common value . The limiting case of SD for very large friction coefcients (i.e.,
when the acceleration r
i
can be neglected compared to the other terms in Eq. (17))
is Brownian dynamics (BD), with the equation of motion
r
i
(t ) =
1
i
m
1
i
[F
i
(t ) +R
i
(t )] . (33)
Although the magnitude of the temperature uctuations is in principle not af-
fected by the values of the friction coefcients (unless they are all zero), the timescale
of these uctuations strongly depends on the
i
coefcients. In fact, it can be
shown [54] that there is a close relationship between the friction coefcients in SD
(used as a mere thermostat) and the temperature relaxation time
T
in Eq. (26). Con-
sider the case where all coefcients
i
are set to a common value . Following from
Eqs. (8) and (10), the change T of the instantaneous temperature over a time inter-
val from t = 0 to can be written
122 Philippe H. Hnenberger
T =
2
k
B
N
d f
_

0
dt

K(t) =
2
k
B
N
d f
N

i=1
m
i
_

0
dt r
i
(t ) r
i
(t ) . (34)
Inserting Eq. (17), this can be rewritten
T =
2
k
B
N
d f
N

i=1
__

0
dt
_
F
i
(t ) m
i
r
i
(t )
_
r
i
(t ) +
_

0
dt (35)
R
i
(t )
_
r
i
(0) +
_
t
0
dt

_
m
1
i
F
i
(t

) r
i
(t

) +m
1
i
R
i
(t

)
_
__
.
Using Eq. (32) and the fact that the stochastic force is uncorrelated with the velocities
and systematic forces at previous times, this simplies to
T =
2
k
B
N
d f
N

i=1
{
_

0
dt [F
i
(t ) r
i
(t ) m
i
r
2
i
(t )]
+ m
1
i
_

0
dt R
i
(t )
_
t
0
dt

R
i
(t

)}
=
2
k
B
N
d f
N

i=1
_

0
dt [F
i
(t ) r
i
(t ) m
i
r
2
i
(t )]
+ 6N
1
d f
N T
o
. (36)
This expression can be rewritten
11
T

=
2
k
B
N
d f
N

i=1
F
i
r
i
+2 [T
o
T] , (37)
where F
i
r
i
and T stand for averages over the interval . The rst term represents
the temperature change caused by the effect of the systematic forces, and would be
unaltered in the absence of thermostat (Newtonian MD simulation). Thus, the second
term can be identied with a temperature change arising from the coupling to a heat
bath. This means that on an intermediate timescale (as dened at the end of Sect. 3),
the mean effect of thermostatization can be written

T(t ) = 2 [T
o
T(t )] . (38)
Comparing with Eq. (26) allows to identify 2 with the inverse of the temperature
relaxation time
T
in Eq. (26), i.e., to suggest = (1/2)
1
T
as an appropriate value
for simulations. This discussion also shows that the semi-macroscopic expression
of Eq. (26) is only valid on an intermediate timescale, when the stochastic uctua-
tions occuring on a shorter timescale (i.e., of the order of the time separating atomic
collisions) are averaged out and only their mean effect is retained.
11
In the absence of constraints N
d f
= 3N due to Eq. (9) with N
c
= 0 and N
r
= 0 (as
appropriate for SD). In the presence of constraints, the derivation should include constraint
forces.
Thermostat Algorithms 123
3.3
Temperature Relaxation by Stochastic Coupling
The stochastic-coupling method was proposed by Andersen [55]. In this approach,
Newtons equation of motion (Eq. (15)) is integrated in time, with the modication
that at each timestep, the velocity of all atoms are conditionally reassigned from a
Maxwell-Boltzmann distribution. More precisely, if an atom i is selected for a ve-
locity reassignment, each Cartesian component of the new velocity is selected at
random according to the Maxwell-Boltzmann probability distribution of Eq. (25).
The selection procedure is such that the time intervals between two successive
velocity reassignments of a given atom are selected at random according to a proba-
bility p() = e

, where is a constant reassignment frequency. In principle, one

can select at random and for each atom a series of successive -values (obeying the
specied probability distribution) before starting the simulation. This series is then
used to determine when the particle is to undergo a velocity reassignment. In prac-
tice, a simpler procedure can be used when t
1
(infrequent reassignments).
At each timestep and for each atom in turn, one generates a randomnumber between
0 and 1. If this number is larger than t for a given atom, this atom undergoes a
velocity reassignment. This procedure leads to a probability distribution
p()t = (1 t )
/t
t (39)
for the intervals without velocity reassignment. This implies
ln
1
p() = (/t ) ln(1 t ) = + O[(t )
2
] . (40)
Thus, when t
1
, p() = e

, as expected. If the condition is not satised,

this second method will not work because the probability of multiple reassignments
within the same timestep becomes non-negligible.
The equation of motion for the Andersen thermostat can formally be written
r
i
(t ) = m
1
i
F
i
(t ) +

n=1

_
t
n

m=1

i,m
_
_
r

i,n
(t ) r
i
(t )
_
, (41)
where {
i,n
| n = 1, 2, . . .} is the series of intervals without reassignment for particle
i , and r

i,n
the randomly-reassigned velocity after the n
t h
interval. This approach
mimicks the effect irregularly-occurring stochastic collisions of randomly chosen
atoms with a bath of ctitious particles at a temperature T
o
. Because, the system
evolves at constant energy between the collisions, this method generates a succession
of microcanonical simulations, interrupted by small energy jumps corresponding to
each collision.
It can be shown [55] that the Andersen thermostat with non-zero collision fre-
quency leads to a canonical distribution of microstates. The proof [55] involves
similar arguments as the derivation of the probability distribution generated by the
MCprocedure. It is based on the fact that the Andersen algorithmgenerates a Markov
chain of microstates in phase space. The only required assumption is that every mi-
crostate is accessible from every other one within a nite time (ergodicity). Note
124 Philippe H. Hnenberger
also that the system total linear and angular momenta are affected by the velocity
reassignments, so that these degrees of freedom are internal to the system, as in SD.
The Andersen algorithm is non-deterministic and time-irreversible. Moreover, it
has the disadvantage of being non-smooth, i.e., generating a discontinuous velocity
trajectory where the randomly-occurring collisions may interfere with the natural
dynamics of the system.
Some care must be taken in the choice of the collision frequency [55, 102]. On
the one hand, too small values (loose coupling) may cause a poor temperature con-
trol. The same observations apply here as those made for SD. The limiting case of
the Andersen thermostat with a vanishing collision frequency is MD, which gener-
ates a microcanonical ensemble. Arbitrarily small collision frequencies are sufcient
to guarantee in principle the generation of a canonical ensemble. But if the collision
frequency is too low, the canonical distribution will only be obtained after very long
simulation times. In this case, systematic energy drifts due to accumulation of nu-
merical errors may interfere with the thermostatization. On the other hand, too large
values for the collision frequency (tight coupling) may cause the velocity reassign-
ments to perturb heavily the dynamics of the system. Although the magnitude of
the temperature uctuations is in principle not affected by the value of the collision
frequency (unless it is zero), the timescale of these uctuations strongly depends on
this parameter. In fact, it can be shown [55] that there is a close relationship between
the collision frequency and the temperature relaxation time
T
in Eq. (26). Each col-
lision changes the kinetic energy of the system by (3/2)k
B
[T
o
T(t )] on average,
and there are N such collisions per unit of time. Thus, one expects

T(t ) = c
1
v

E(t) = (3/2)c
1
v
Nk
B
[T
o
T(t )] , (42)
where T and E stand for averages over an intermediate timescale (as dened at the
end of Sect. 3). Comparing with Eq. (26) allows to identify (3/2)c
1
v
Nk
B
with
the inverse of the temperature relaxation time
T
in Eq. (26), i.e., to suggest =
(2/3)(Nk
B
)
1
c
v

1
T
as an appropriate value for simulations. Note that because
T
scales as N
1/3
(Eq. (27)), the collision frequency for any particle scales as N
2/3
,
so that the time each particle spends without reassignment increases with the system
size. On the other hand, the collision frequency for the whole system, N, scales as
N
1/3
, so that the length of each microcanonical sub-simulation decreases with the
system size.
3.4
Temperature Constraining
Temperature constraining aims at xing the instantaneous temperature T to the ref-
erence heat-bath value T
o
without allowing for any uctuations. In this sense, tem-
perature constraining represents a hard boundary condition, in constrast to the soft
boundary conditions employed by all other thermostats mentioned in this article.
Note that constraining the temperature, i.e., enforcing the relation T(t) = T
o
(or

T(t ) = 0) represents a non-holonomic constraint. Holonomic constraints are those

Thermostat Algorithms 125
which only involve generalized coordinates and time, excluding any dependence on
the generalized velocities. Two main temperature-constraining algorithms have been
proposed. The rst one is due to Woodcock [103], and the second one was simulta-
neously proposed by Hoover [104] and Evans [105].
In the Hoover-Evans algorithm [104, 52, 51, 105, 106], the quantity (t; t) in
Eq. (20) is found by imposing temperature conservation in the form T(t +
t
2
) =
T(t
t
2
). Using Eqs. (8) and (10), this leads to the condition

2
(t ; t )
k
B
N
d f
_
N

i=1
m
i
_
r
i
_
t
t
2
_
+m
1
i
F
i
(t )t
_
2
_
=
1
k
B
N
d f
N

i=1
m
i
r
2
i
_
t
t
2
_
. (43)
Solving for (t; t) gives
(t ; t ) =

N
i=1
m
i
r
2
i
(t
t
2
)

N
i=1
m
i
_
r
i
(t
t
2
) +m
1
i
F
i
(t )t
_
2

1/2
=
_
T(t
t
2
)
T

(t +
t
2
)
_
1/2
, (44)
where T(t
t
2
) and T

(t +
t
2
) are the instantaneous temperatures computed based
on the velocities r
i
(t
t
2
) and r

i
(t +
t
2
), see Eq. (20). Because this quantity satises
(t; 0) = 1, applying Eq. (21) gives
(t ) = [N
d f
k
B
T(t )]
1
N

i=1
r
i
(t ) F
i
(t ) . (45)
Inserting into Eq. (18) shows that the equation of motion corresponding to the
Hoover-Evans thermostat is
r
i
(t ) = m
1
i
F
i
(t )
_
N
d f
k
B
T(t )
_
1
_
N

i=1
r
i
(t ) F
i
(t )
_
r
i
(t ) . (46)
This equation of motion should sample an isothermal trajectory at a temperature
determined by the initial internal velocities. Eq. (46) can also be derived directly
from Eq. (18) by imposing

T(t ) = 0. Using Eqs. (8) and (10) this becomes

T(t ) =
d
dt
_
1
k
B
N
d f
N

i=1
m
i
r
2
i
(t )
_
=
2
k
B
N
d f
N

i=1
m
i
r
i
(t ) r
i
(t ) = 0 . (47)
Inserting Eq. (18) and solving for (t ) leads to Eq. (46).
126 Philippe H. Hnenberger
The Hoover-Evans algorithmshould in principle ensure a constant temperature at
all time points. However, this condition is only enforced by zeroing the temperature
derivative. Because the reference temperature T
o
does not appear explicitly in the
scaling factor of Eq. (44), numerical inaccuracies will inevitably prevent temperature
conservation, and cause the temperature to actually drift in simulations.
In the Woodcock algorithm [103], the quantity (t ; t ) in Eq. (20) is found by
imposing temperature conservation in the formT(t +
t
2
) =
g
N
d f
T
o
, thereby making
explicit use of the reference temperature. Although g = N
d f
seems to be the obvious
choice, it turns out that g = N
d f
1 is the approptiate choice for the algorithm to
generate a canonical ensemble of congurations at temperature T
o
(see below). Using
Eqs. (8) and (10), this leads to the condition

2
(t ; t )
k
B
N
d f
_
N

i=1
m
i
_
r
i
_
t
t
2
_
+m
1
i
F
i
(t )t
_
2
_
=
g
N
d f
T
o
. (48)
Solving for (t; t) gives
12
(t ; t ) =

gk
B
T
o

N
i=1
m
i
_
r
i
_
t
t
2
_
+m
1
i
F
i
(t )t
_
2

1/2
=
_
g
N
d f
T
o
T

_
t +
t
2
_
_
1/2
. (49)
If T(t
t
2
) =
g
N
d f
T
o
(i.e., if the simulation was started with internal velocities
corresponding to the reference temperature, or otherwise, after a rst equilibration
timestep), this quantity satises (t; 0) = 1. In this case, applying Eq. (21) gives
(t) = (gk
B
T
o
)
1
N

i=1
r
i
(t ) F
i
(t ) . (50)
Inserting into Eq. (18) shows that the equation of motion corresponding to the Wood-
cock thermostat is
r
i
(t ) = m
1
i
F
i
(t ) (gk
B
T
o
)
1
_
N

i=1
r
i
(t ) F
i
(t )
_
r
i
(t ) . (51)
This equation of motion is rigorously equivalent to the Hoover-Evans equation of
motion (Eq. (46)), provided that the initial internal velocities are appropriate for
12
Note that some simulation programs do not apply the scaling of the velocities by (t ; t ) at
every timestep, but perform the scaling on a periodic basis, or when the difference between
the instantaneous and reference temperatures is larger than a given tolerance [107].
Thermostat Algorithms 127
the temperature T
o
, i.e., that T(0) =
g
N
d f
T
o
. However, even in this case, the cor-
responding algorithms differ numerically. Because the Woodcock algorithm explic-
itly involves the reference temperature T
o
in the calculation of the scaling factor of
Eq. (49), its application removes the risk of a temperature drift.
Because maintaining the temperature constant represents a single constraint
equation involving a total of N
d f
velocity variables, it should not be surprising that
numerous other choices of equations of motion lead to an isothermal dynamics (also
satisfying the two constraints that the system linear and angular momenta are con-
stants of the motion). For example, Haile and Gupta [108] have shown how to con-
struct two general classes of isothermal equations of motion based on generalized
forces or generalized potentials. An example of the former class is the Hoover-Evans
thermostat. An example of the second class is a thermostat similar (but, contrary
to the claim of the authors [108], not identical) to the Woodcock thermostat. The
equations of motion of this Haile-Gupta thermostat are
r
i
(t ) =
_
g
N
d f
T
o
T

(t )
_
1/2
r

i
(t ) and r

i
(t ) = m
1
i
F
i
(t ) . (52)
In words, the auxiliary velocities r

i
are propagated independently in time according
to Newtons second law, and the true velocities obtained by multiplying these by the
appropriate scaling factor at each timestep. In contrast, in the Woodcock thermo-
stat, the auxiliary velocities r

i
are obtained at each timestep by increasing the true
velocities r
i
by m
1
i
F
i
t.
It can be shown that the ensemble generated by the (identical) Woodcock and
Hoover-Evans equations of motion represents a canonical distribution of congura-
tions (though obviously not of momenta) at temperature T
o
, provided that one sets
g = N
d f
1 ([104, 53]; see Appendix). This may seem surprizing at rst sight, but
canonical sampling of congurations is only achieved with T(t) =
N
d f
1
N
d f
T
o
= T
o
,
i.e., when simulating at a slightly lower internal temperature. The reason is that con-
straining the temperature effectively removes one degree of freedom from the sys-
tem. A more consistent approach would be to alter the denition of the instantaneous
temperature (Eq. (8)) by changing N
d f
to N
d f
1 in this case. However, since other
thermostats may involve different values for the factor g (see Sect. 3.6), it is more
convenient here to stick to a single denition of T. On the other hand (and countrary
to the authors initial claim [108]), the Haile-Gupta thermostat does not generate a
canonical ensemble of congurations ([53, 109]; see Appendix). The equations of
motion of temperature constraining are smooth, deterministic and time-reversible.
However, the absence of kinetic energy uctuations may lead to inaccurate dynam-
ics, especially in the context of the microscopic systems typically considered in sim-
ulations.
128 Philippe H. Hnenberger
3.5
Temperature Relaxation by Weak Coupling
The idea of a thermostat based on a rst-order relaxation equation is due to Berend-
sen [54]. As discussed at the end of Sect. 3, when a system at a given average tem-
perature T is in contact with a heat bath at a different temperature T
o
, the rate of
temperature change is given by Eq. (26). As discussed in Sect. 3.2, this equation is
only valid when the average temperature is calculated on an intermediate timescale
(short compared to the experimental timescale, but long compared to the time sep-
arating atomic collisions). On this timescale, only the mean effect of the stochastic
forces acting in SD needs to be considered, leading to the rst-order temperature
relaxation law of Eq. (26).
The idea behind the Berendsen thermostat is to modify the Langevin equation of
motion (Eq. (17)) in the sense of removing the local temperature coupling through
stochastic collisions (random noise), while retaining the global coupling (principle
of least local perturbation). This prescription is equivalent to assuming that Eq. (26)
also applies to the instantaneous temperature T, i.e., that

T(t) =
1
B
[T
o
T(t )] , (53)
where the appropriate value for
B
should be the temperature relaxation time
T
.
In this case, the quantity (t; t) in Eq. (20) is found by imposing T(t +
t
2
) =
T(t
t
2
) +
1
B
t
g
N
d f
[T
o
T(t
t
2
)], where in principle g = N
d f
. Using
Eqs. (8) and (10), this leads to the condition

2
(t ; t )T

_
t +
t
2
_
= T
_
t
t
2
_
+
1
B
t
_
g
N
d f
T
o
T
_
t
t
2
__
.
(54)
Solving for (t ; t ) gives
(t; t) =
_
T
_
t
t
2
_
T

_
t +
t
2
_ +
1
B
t
g
N
d f
T
o
T
_
t
t
2
_
T

_
t +
t
2
_
_
1/2

_
1 +
1
B
t
_
g
N
d f
T
o
T

_
t +
t
2
_ 1
__
1/2
. (55)
In general, the algorithmis implemented following the second (approximate) expres-
sion. For either of the two expressions, Eq. (21) gives
(t ) =
1
2

1
B
_
g
N
d f
T
o
T(t )
1
_
. (56)
Inserting into Eq. (18) shows that the equation of motion corresponding to the
Berendsen thermostat is
r
i
(t ) = m
1
i
F
i
(t )
1
2

1
B
_
g
N
d f
T
o
T(t )
1
_
r
i
(t ) . (57)
Thermostat Algorithms 129
In practice,
B
is used as an empirical parameter to adjust the strength of the cou-
pling. Its value should be chosen in a appropriate range. On the one hand, a too large
value (loose coupling) may cause a systematic temperature drift. Indeed, in the limit

B
, the Berendsen thermostat is inactive leading to the MD equation of mo-
tion, which samples a microcanonical ensemble. Thus, the temperature uctuations
will increase with
B
until they reach the appropriate value for a microcanonical en-
semble. However, they will never reach the appropriate value for a canonical ensem-
ble, which are larger. For large values of
B
, a systematic energy (and thus tempera-
ture) drift due to numerical errors may also occur, just as in MD. On the other hand,
a too small value (tight coupling) will cause unrealistically low temperature uctua-
tions. Indeed, the special case of the Berendsen algorithm (Eq. (55)) with
B
= t
is the Woodcock thermostat (Eq. (49)), which does not allow for temperature uc-
tuations. This shows that the limiting case of the Berendsen equation of motion for

B
0 is the Woodcock/Hoover-Evans equation of motion. Values of
B
0.1 ps
are typically used in MD simulations of condensed-phase systems. Note, however,
that this choice generally leads to uctuations close to those of the microcanonical
ensemble. With this choice, the Berendsen thermostat merely removes energy drifts
froma MD simulation, without signicantly altering the ensemble sampled (and thus
rather plays the role of an ergostat algorithm).
The Berendsen equation of motion is smooth and deterministic, but time-
irreversible. The ensemble generated by the Berendsen equations of motion is not
a canonical ensemble ([109]; see Appendix). Only in the limit
B
0 (or in prac-
tice
B
= t), when the Berendsen equation of motion becomes identical to the
Woodcock/Hoover-Evans equation of motion, does it generate a canonical distribu-
tion of congurations. In the limit
B
, the microcanonical ensemble is recov-
ered. All intermediate situations correspond to the sampling of an unusual weak-
coupling ensemble
13
, which is neither canonical nor microcanonical [109]. The
reason why the Berendsen thermostat systematically (for all values of
B
) under-
estimates temperature uctuations (and thus does not give the correct thermodynam-
ical ensemble) resides in the transition from Eq. (26) to Eq. (53), corresponding to
the neglect of the stochastic contribution to these uctuations on the microscopic
timescale.
3.6
Temperature Relaxation by the Extended-System Method
The idea of a thermostat based on an extended-systemmethod is due to Nos [61]. A
simpler formulation of the equations of motion was later proposed simultaneously
14
13
Assuming a relationship of the form of Eq. (115), it is possible to derive the congurational
partition function of the weak-coupling ensemble as a function of ([109]; see Appendix)
The limiting cases = 0 (
B
0; canonical) and = 1 (
B
; microcanonical) are
reproduced. Note that the Haile-Gupta thermostat generates congurations with the same
probability distribution as the Berendsen thermostat with = 1/2.
14
Eqs. (2.24) and (2.25) in [53] are equivalent to Eq. (6) in [63], provided that one identies
= s

/s

. These equations are Eqs. (78) and (79) of the present article.
130 Philippe H. Hnenberger
by Nos [53] and Hoover [63], so that this algorithm is generally referred to as the
Nos-Hoover thermostat.
The idea behind the original Nos [61] algorithm is to extend the real system by
addition of an articial (N
d f
+1)
t h
dynamical variable s (associated with a mass"
Q > 0, with actual units of energy(time)
2
, as well as a velocity

s, and satisfying
s > 0) that plays the role of a time-scaling parameter
15
. More precisely, the timescale
in the extended system is stretched by the factor s, i.e., an innitesimal time interval
d

t at time

t in the extended system corresponds to a time interval dt = s
1
(

t ) d

t
in the real system
16
. Consequently, although the atomic coordinates are identical in
both systems, the extended-systemvelocities are amplied by a factor s
1
compared
to the real-system velocities, i.e.
r = r ,

r = s
1
r , s = s , and

s = s
1
s . (58)
The Lagrangian for the extended system is chosen to be
L
e
( r,

r, s,

s) =
1
2
N

i=1
m
i
s
2

r
2
i
U( r) +
1
2
Q

s
2
gk
B
T
o
ln s , (59)
where g is equal to the number of degrees of freedom N
d f
in the real system, possibly
increased by one (see below). The rst two terms of the Lagrangian represent the
kinetic energy minus the potential energy of the real system (the extended-system
velocities are multiplied by s to recover the real-system ones). The third and fourth
terms represent the kinetic energy minus the potential energy associated with the s-
variable. The form of the last term is chosen to ensure that the algorithm produces a
canonical ensemble of microstates (see below). The Lagrangian equations of motion
derived from Eq. (59) read

r
i
= m
1
i
s
2

F
i
2 s
1

s

r
i
(60)
for the physical variables, and
17
15
All extended-system variables will be noted with a tilde overscript, to distinguish them
from the real-system variables (the real-system variable corresponding to s is noted s).
16
To simplify the notation, explicit dependence of the different functions on time is generally
omitted in this section. The time-dependent functions are

F(

t ), r(

t ), p(

t ), s(

t ) and p
s
(

t )
(together with their rst and second time derivatives) for the extended system, and F(t ),
r(t ), p(t ), s(t ), p
s
(t ), (t ) and T(t ) (together with their rst and second time derivatives)
for the real system. The dot overscripts indicate differentiation with respect to the extended-
system time

t for the extended-system variables, and with respect to the real-system time t
for the real-system variables.
17
Because the time-average of the time-derivative of a bounded quantity (for example,

s)
vanishes, this equation implies
_
s
1
N

i=1
m
i
s
2

r
2
i
_
e,v
=
_
s
1
N

i=1
m
i
r
2
i
_
e,v
=
_
s
1
_
e,v
gk
B
T
o
, (61)
Thermostat Algorithms 131

s = Q
1
s
1
_
N

i=1
m
i
s
2

r
2
i
gk
B
T
o
_
(63)
for the s-variable. These two second-order differential equations can be discretized
(based on a timestep

t in the extended system) and integrated simultaneously dur-

ing the simulation. The successive values of r = r and r = s

r describe the evolution

of the atomic coordinates and velocities in the real system at successive time points
separated by t = s
1
(

t )

t. This means that the algorithm implemented in this

form (referred to as virtual-time sampling) leads to sampling of the real-system tra-
jectory at uneven time intervals. Atrajectory with real-time sampling can be achieved
either by interpolation at evenly spaced real-time points of the coordinates and ve-
locities issued from virtual-time sampling, or by rewriting the equations of motion in
terms of the real-system variables (Nos-Hoover formulation [53, 63]; see below).
As an alternative to Eqs. (60) and (63), the Nos equations of motion can be
equivalently formulated using a Hamiltonian formalism. In this case, the extended-
system conjugate momenta p
i
and p
s
associated with the physical degrees of free-
dom and with the s-variable are dened as
p
i
=
L
e
( r,

r, s,

s)

r
= m
i
s
2

r
i
and p
s
=
L
e
( r,

r, s,

s)

s
= Q

s . (64)
Comparison with the corresponding real-system momenta
18
, dened as
p
i
= m
i
r
i
and p
s
= Qs
2
s , (66)
shows that the extended-system momenta are amplied by a factor s compared to
the real-system momenta. The extended-system Hamiltonian corresponding to the
Lagrangian of Eq. (59) can now be written
where ...
e,v
denotes ensemble averaging over the extended system (with virtual-time
sampling). Considering Eqs. (8) and (10), this result already suggests that the average
temperature of the real system coincides with T
o
. Using Eq. (101) and p
i
= m
i
r
i
, the
above equation can indeed be rewritten
T
e,r
=
_
1
k
B
N
d f
N

i=1
m
i
r
i
2
_
e,r
=
g
N
d f
T
o
, (62)
where ...
e,r
denotes ensemble averaging over the extended system (with real-time sam-
pling). From Eq. (102), this latter ensemble average can be identied with a canonical one
when g = N
d f
.
18
In the absence of thermostat (i.e., when the variable s is uncoupled from the system), the
real-system Lagrangian may be written
L(r, r, s, s) =
1
2
N

i=1
m
i
r
2
i
U(r) +
1
2
Qs
2
s
2
. (65)
The momenta of Eq. (66) are derived from this Lagrangian.
132 Philippe H. Hnenberger
H
e
( r, p, s, p
s
) =
1
2
N

i=1
m
1
i
s
2
p
2
i
+U( r) +
1
2
Q
1
p
2
s
+ gk
B
T
o
ln s . (67)
This function is a constant of the motion and evaluates to E
e
, the total energy of the
extended system. The corresponding Hamiltonian equations of motion read

p
i
=

F
i
and

r
i
= m
1
i
s
2
p
i
(68)
for the physical variables, and

p
s
= s
1
_
N

i=1
m
1
i
s
2
p
2
i
gk
B
T
o
_
and

s = Q
1
p
s
(69)
for the s-variable.
The Nos equations of motion sample a microcanonical ensemble in the extended
system ( r, p,

t ), with a constant total energy E

e
. However, the energy of the real sys-
tem (r = r, p = s
1
p, t =
_
s
1
d

t) is not constant. Accompanying the uctuations

of s, heat transfers occur between the system and a heat bath, which regulate the
system temperature. As will be seen below (Eq. (78) where = can be identied
with

s), the sign of

s determines the direction of the heat ow. When

s < 0, heat
ows into the real system. When

s > 0, heat ows out of the real system. It can be
proved ([61]; see Appendix) that the Nos equations of motion sample a canonical
ensemble of microstates in the real system, provided that g = N
d f
+1 (virtual-time
sampling) or g = N
d f
(real-time sampling), and that Q is nite, this irrespective of
the actual values of Q and E
e
. If the potential energy U( r) does not involve terms
giving rise to external forces, the total linear and angular momenta associated with
the physical degrees of freedom in the extended system, namely
N

i=1
p
i
=
N

i=1
m
i
s
2

r
i
and
N

i=1
r
i
p
i
=
N

i=1
m
i
s
2
r
i


r
i
, (70)
are also conserved. Because

r
i
= s
1
r
i
(Eq. (58)), this implies that the total linear
and angular momenta of the real system are linearly related to s
1
and thus not
conserved, unless they vanish. This should be the case if the components of the real
velocities r
o
i
along the external degrees of freedomhave been removed by application
of Eq. (11).
The Nos equations of motion are smooth, deterministic and time-reversible.
However, because the time-evolution of the variable s is described by a second-order
equation (Eq. (63)), heat may ow in and out of the system in an oscillatory fash-
ion [110], leading to nearly-periodic temperature uctuations. However, from the
discussion of Sects. 3 and 3.2, the dynamics of the temperature evolution should not
be oscillatory, but rather result from a combination of stochastic uctuations and ex-
ponential relaxation. At equilibrium, the approximate frequency of these oscillations
can be estimated in the following way [61]. Consider small deviations s of s away
from the equilibriumvalue s
e,r
, where ...
e,r
denotes ensemble averaging over the
Thermostat Algorithms 133
extended system with real-time sampling. Assuming that the interatomic forces have
a weak effect on the temperature dynamics (as, e.g., in a perfect gas), the quantity

N
i=1
m
1
i
p
2
i
, which is solely altered by the action of the forces (Eq. (68)), can be
assumed nearly constant. In this case, one may write (with g = N
d f
)
1
2
N

i=1
m
i
s
2

r
2
i
=
1
2
N

i=1
m
1
i
s
2
p
2
i

1
2
s
2
e,r
s
2
N
d f
k
B
T
o
. (71)
Using this result, Eq. (63) may be written (for small s)

s = N
d f
k
B
T
o
Q
1
s
1
( s
2
e,r
s
2
1) 2N
d f
k
B
T
o
Q
1
s
2
e,r
s . (72)
This corresponds to a harmonic oscillator with frequency
= (2)
1
(2N
d f
k
B
T
o
)
1/2
Q
1/2
s
1
e,r
, (73)
where
19
s
e,r
=
_
N
d f
N
d f
+1
_
1/2 _
exp
_
_
N
d f
k
B
T
o
_
1
_
E
e
H(r, p)
_
__
, (75)
... denoting a canonical ensemble average. Comparison of the approximate oscil-
lation frequency with the inverse of the temperature relaxation time
T
(Eq. (26))
may guide the choice of parameters Q and E
e
leading to a realistic timescale of
temperature uctuations. If the number of degrees of freedom is large and E
e
is
close to H(r, p), the average canonical energy corresponding to the real system,
Eq. (75) becomes s
e,r
= 1. The latter condition will be satised if the algorithm is
initiated using real-system velocities taken from a Maxwell-Boltzmann distribution
(Eq. (25)), together with s(0) = 1 and

s(0) = 0. In this case, comparing Eq. (73)
with Eq. (26) allows to identify
20
1.2(2N
d f
k
B
T
o
)
1/2
Q
1/2
with the temperature re-
laxation time
T
, i.e., to suggest Q 1.4N
d f
k
B
T
o

2
T
as an appropriate value for
simulations. In fact, it may make sense to use an effective relaxation time

NH
= (N
d f
k
B
T
o
)
1/2
Q
1/2
(76)
instead of the (less intuitive) effective mass Q to characterize the strength of the
coupling to the heat bath.
19
Considering Eqs. (98) and (101), one has
s
e,r
=
_
dp
_
dr
_
d p
s
_
d s s
N
d f
+1
[ s s
o
]
_
dp
_
dr
_
d p
s
_
d s s
N
d f
[ s s
o
]
. (74)
Inserting Eq. (94), integrating over p
s
for the numerator and denominator, and setting
g = N
d f
leads to Eq. (75).
20
Because cos(1.2) exp(1), it is assumed here that the exponential relaxation time is
approximately given by 1.2/(2)
1
.
134 Philippe H. Hnenberger
The use of an extended systemwith a stretched timescale is not very intuitive, and
the sampling of a trajectory at uneven time intervals is rather impractical for the in-
vestigation of the dynamical properties of a system. However, as shown by Nos [53]
and Hoover [63], the Nos equations of motion can be reformulated in terms of
real-system variables (together with real-time sampling) so as to avoid these prob-
lems. The transformation from extended-system to real-system variables is achieved
through
s = s , s = s

s , s = s
2

s + s

s
2
,
r = r , r = s

r , r = s
2

r + s

r ,
p
s
= s
1
p
s
, p
s
=

p
s
Q
1
s
1
p
2
s
, (77)
p = s
1
p , p =

p Q
1
s
1
p
s
p ,
and F =

F .
Because dt = s
1
d

t, these equations are derived using d/dt = s d/d

t, together
with the denition of the real-system (Eq. (66)) and extended-system (Eq. (64)) mo-
menta. Based on these expressions, and dening the quantity = s
1
s = Q
1
sp
s
,
the Lagrangian equations of motion (Eqs. (60) and (63)) can be rewritten
r
i
= m
1
i
F
i
r
i
(78)
and
= k
B
N
d f
Q
1
T
_
g
N
d f
T
o
T
1
_
. (79)
Note that the variable is a real-system variable (i.e., a function of the real-system
time t ). The equivalent extended-system variable is = =

s. If the effective
coupling time
NH
(Eq. (76)) is used instead of Q, Eq. (79) becomes
=
2
NH
T
T
o
_
g
N
d f
T
o
T
1
_
. (80)
In a similar way, the Hamiltonian equations of motion (Eqs. (68) and (69)) can be
rewritten
p
i
= F
i
p
i
and r
i
= m
1
i
p
i
(81)
and
p
s
= k
B
N
d f
Ts
1
_
g
N
d f
T
o
T
1
_
p
s
and s = s . (82)
Equation (81) is easily identied with Eq. (78), and Eq. (82) with Eq. (79). The vari-
able s is absent from the rst set of equations, i.e., its dynamics has been decoupled.
In the second set of equations, the evolution of the real-system variables is indepen-
dent of the actual value of s, i.e., any choice of the initial value s(0) will lead to the
Thermostat Algorithms 135
same dynamics. Finally, it should be stressed that these equations of motion are no
longer Hamiltonian, although the quantity (Eq. (67))
H(r, p, s, p
s
) =
1
2
N

i=1
m
1
i
p
2
i
+U(r) +
1
2
Q
1
s
2
p
2
s
+ gk
B
T
o
ln s (83)
is still a constant of the motion (evaluating to E
e
). On the other hand, Eqs. (78) and
(79) are still Lagrangian, the corresponding Lagrangian being
L(r, r, s, s) = s [ L
e
(r, s
1
r, s, s
1
s) + E
e
] . (84)
To obtain the corresponding Lagrangian equations of motion, E
e
is initially treated
as a constant and later expanded using Eq. (83). It appears that Eq. (78) has exactly
the form of Eq. (18), i.e., the Nos-Hoover thermostat has one equation of motion
in common with both the Woodcock/Hoover-Evans and the Berendsen thermostats.
However, in contrast to these other thermostats where the value of was uniquely
determined by the instantaneous microstate of the system (compare Eq. (79) with
Eqs. (45), (50), and (56)), is here a dynamical variable which derivative (Eq. (79))
is determined by this instantaneous microstate. Accompanying the uctuations of
, heat transfers occur between the system and a heat bath, which regulate the sys-
tem temperature. Because = s
1
s = =

s (Eq. (77)), the variable in the
Nos-Hoover formulation plays the same role as

s in the Nos formulation. When
(or

s) is negative, heat ows from the heat bath into the system due to Eq. (78) (or
Eq. (60)). When the system temperature increases above T
o
, the time derivative of
(or

s) becomes positive due to Eq. (79) (or Eq. (63)) and the heat ow is progres-
sively reduced (feedback mechanism). Conversely, when (or

s) is positive, heat is
removed from the system until the system temperature decreases below T
o
and the
heat transfer is slowed down.
The second- and rst-order Eqs. (78) and (79) can be discretized (based on a
timestep t in the real system) and integrated simultaneously during the simulation.
Note that the Nos thermostat with g = N
d f
+ 1 and virtual-time sampling and
the Nos-Hoover thermostat with g = N
d f
formally sample the same trajectory. In
practice, however, the trajectories are sampled at different real-system time points
and will numerically diverge for nite timestep sizes. It can be proved ([63]; see
Appendix) that the Nos-Hoover equations of motion sample a canonical ensemble.
in the real system provided that g = N
d f
and that Q is nite, this irrespective of the
actual values of Q and E
e
. Though interesting, such a proof is not really necessary
since the Nos and Nos-Hoover formalisms are equivalent. As a by-product of this
proof, it is shown that the probability distribution of the variable is a Gaussian
of width determined by the parameter Q (Eq. (112)). The Nos-Hoover equations
of motion are smooth, deterministic and time-reversible. However, just as the Nos
algorithm, Nos-Hoover dynamics may lead to temperature oscillations.
In both algorithms, Some care must be taken in the choice of the ctitious mass
Q and extended-system energy E
e
. On the one hand, too large values of Q (loose
coupling) may cause a poor temperature control. Indeed, the limiting case of the
136 Philippe H. Hnenberger
Nos-Hoover thermostat with Q and (0) = 0 is MD, which generates a mi-
crocanonical ensemble. Although any nite (positive) mass is sufcient to guarantee
in principle the generation of a canonical ensemble, if Q is too large, the canonical
distribution will only be obtained after very long simulation times. In this case, sys-
tematic energy drifts due to accumulation of numerical errors may interfere with the
thermostatization. On the other hand, too small values (tight coupling) may cause
high-frequency temperature oscillations (Eq. (73)) leading to the same effect. This
is because if the s variable oscillates at a very high frequency, it will tend to be
off-resonance with the characteristic frequencies of the real system, and effectively
decouple from the physical degrees of freedom (slow exchange of kinetic energy).
The choice of the parameters Q and E
e
can be guided by comparison of the fre-
quency (Eq. (73)) with the inverse of the temperature relaxation time
T
(Eq. (26)).
Note that if a simulation is initiated with s(0) = 1 and (0) = 0, which seems the
most reasonable choice, the value of E
e
will match the initial energy of the real sys-
tem. The numerical integration of the Nos and Nos-Hoover equations will not be
discussed here. A number of alternative schemes have been proposed in the litera-
ture [111, 112, 113, 114, 90].
The constant-temperature Woodcock/Hoover-Evans equation of motion can be
retrieved from the the Nos-Hoover formalism by a slight modication of the
extended-system Hamiltonian. This is done by introducing the constraints
s = (gk
B
T
o
)
1/2
_
N

i=1
m
1
i
p
2
i
_
1/2
and p
s
= 0 (85)
into Eq. (67), leading to the modied Hamiltonian
H
c
( r, p) =
1
2
gk
B
T
o
+U( r) +
1
2
gk
B
T
o
ln
_
(gk
B
T
o
)
1
N

i=1
m
1
i
p
2
i
_
. (86)
The corresponding Hamiltonian equations of motion for the physical variables in the
extended system are still given by Eq. (68). It can be proved ([61]; see Appendix) that
these equations of motion sample a canonical ensemble of congurations in the real
system (r = r, r = s
1
p, t =
_
s
1
d

t ) with s given by Eq. (85), provided that g =

N
d f
(virtual-time sampling) or g = N
d f
1 (real-time sampling), irrespective of the
constant value E
e
of H
c
. To show that this situation matches the Woodcock/Hoover-
Evans equation of motion, the equation of motion must be rewritten in terms of
real-system variables. Evaluating

s based on Eq. (85) gives

s = (gk
B
T
o
)
1
s
1
N

i=1
m
1
i
p
i


p
i
. (87)
Applying the transformations of Eq. (77), setting = s
1
s, and inserting Eq. (68)
for

p, it is easily seen that the equation of motion in terms of the real-systemvariables
is Eq. (78) together with
Thermostat Algorithms 137
= (gk
B
T
o
)
1
N

i=1
r
i
F
i
, (88)
which is identical to the corresponding factor for the Woodcock (Eq. (50)) and
Hoover-Evans (Eq. (45), when T(0) =
g
N
d f
T
o
) equations of motion. This also shows
that the appropriate choice for g so as to generate a canonical ensemble of congu-
rations using either of these two thermostats is g = N
d f
1.
If, in addition to incorporating the constraint of Eq. (85), the Hamiltonian is
changed to
H
h
( r, p) =
_
gk
B
T
o
N

i=1
m
1
i
p
2
i
_
1/2
+U( r) , (89)
the corresponding Hamiltonian equations of motion for the physical variables in the
extended system become

p
i
=

F
i
and

r
i
= m
1
i
s
1
p
i
. (90)
Setting r
i
=

r
i
and p
i
= m
i
r
i
, one recovers (using Eq. (85) and identifying r

i
=
m
1
i
p
i
) the Haile-Gupta equations of motion (Eq. (52)). Note that s no longer acts
as a scaling parameter and the extended-systemLagragian has been changed, so that
Eqs. (58) and (64) no longer apply. It can be proved ([53]; see Appendix) that this
equation of motion does not sample a canonical ensemble of congurations in the
real system, irrespective of the choice of g.
3.7
Generalizations of the Previous Methods
An interesting extension of the SD thermostat (Sect. 3.2) is the so-called dissipative-
particle-dynamics (DPD) thermostat [115, 116, 117, 118]. This scheme retains a key
advantage of the SD thermostat (shared with the Andersen thermostat), namely that
it couples atomic velocities to the heat bath on a local basis (as opposed to the global
coupling applied by all other thermostats discussed in this article). Local coupling
leads to an intrinsically more efcient thermostatization and in turn, permits the use
of longer timesteps to integrate the equations of motion (in applications where this
timestep is not further limited by the curvature of the interaction function, i.e., when
using soft intermolecular potentials). On the other hand, the DPD scheme alleviates
two drawbacks of the SD scheme, namely (i) the non-conservation of the system
linear and angular momentum, and (ii) the loss of local hydronamic correlations be-
tween particles. In practice, this is achieved by replacing the frictional and stochastic
forces acting on individual atoms in SD (Eq. (17)), by corresponding central pairwise
forces acting on atom pairs within a given cutoff distance.
A number of extensions or generalizations of the Nos or Nos-Hoover ap-
proaches (Sect. 3.6) have also been reported in the literature, following three main
directions.
138 Philippe H. Hnenberger
First, more general extended Hamiltonians (including the Nos Hamiltonian as
a particular case) have been shown to also produce canonical phase-space sam-
pling [119, 120, 121, 122, 123, 124, 125, 126, 127]. The additional exibility of-
fered by these new schemes may be used to overcome the non-ergodic behaviour of
the Nos-Hoover thermostat in the context of small or stiff systems (e.g., single har-
monic oscillator) or systems at low temperatures [63, 128, 129, 121, 130, 131, 132].
A number of these variants can indeed produce the correct canonical distribution for
a single harmonic oscillator [121, 122, 123, 125, 126]. The most popuar of these
schemes is probably the Nos-Hoover chain thermostat [123], where the single ther-
mostatting variable of the Nos-Hoover scheme is replaced by a series of variables
thermostatting each other in sequence.
Second, alternatives have been proposed for the Nos-Hoover scheme, which are
phase-space conserving [133] or even symplectic [127]. One of the latter schemes,
referred to as the Nos-Poincar thermostat, leads to the same phase-space trajectory
as the Nos-Hoover thermostat with real time sampling, but has the advantage of
being Hamiltonian.
Third, generalized equations of motion have been proposed to sample arbitrary
(i.e., not necessarily canonical) probability distributions [134, 135, 136, 137]. Such
methods can be used, e.g., to optimize the efciency of conformational searches [134,
135, 137] or for generating Tsallis distributions of microstates [136].
4
Practical Considerations
This section briey mentions two practical aspects related to the use of thermostats
in MD simulations of (bio-)molecular systems.
The rst problem is encountered when simulating molecular systems involving
distinct sets of degrees of freedom with either (i) very different characteristic fre-
quencies or (ii) very different heating rates caused by algorithmic noise. In this case,
the joint coupling of all degrees of freedom to a thermostat may lead to different
effective temperatures for the distinct subsets of degrees of freedom, due to a too
slow exchange of kinetic energy between them. A typical example is the so-called
hot solvent cold solute problem in simulations of macromolecules. Because the
solvent is more signicantly affected by algorithmic noise (e.g., due to the use of
an electrostatic cutoff), the coupling of the whole system to a single thermostat may
cause the average solute temperature to be signicantly lower than the average sol-
vent temperature. A solution to this problem is to couple separately the solute and
solvent degrees of freedom to two different thermostats.
The second problemis encountered when using a simulation programthat (incor-
rectly) applies the thermostatization directly to the atomic velocities rather than to the
internal (peculiar) velocities (Sect. 3). In this case, the system linear and (under vac-
uum boundary conditions) angular momenta are not conserved, unless they exactly
vanish. However, even if these quantities are set to zero at the beginning of a simu-
lation, numerical errors will unavoidably alter these initial values, permitting a ow
Thermostat Algorithms 139
of kinetic energy between internal and external degrees of freedom. Unfortunately,
it appears that at least some thermostats tend to pump kinetic energy from high-
frequency to low-frequency degrees of freedom (thereby violating equipartition). In
this case, uniform translational and (under vacuum boundary conditions) rotational
motion will tend to build up. For simulations started with vanishing overall momenta,
one generally observes a very slow initial rise (often taking nanoseconds) followed
by a very sudden burst of translational and rotational kinetic energy. The accumu-
lation of kinetic energy in these degrees of freedom will effectively cool down the
internal ones, giving rise to the so-called ying ice cube effect [138, 139]. The
most obvious remedy to this problem is to remove the overall center of mass motion
from the atomic velocities at regular interval during the simulation. However, the
application of thermostatization on the basis of internal velocities (Sect. 3) should
probably be preferred, because it is more consistent and avoids the indrotuction of
discontinuities in the generated velocity trajectory.
5
Appendix: Phase-Space Probability Distributions
Here, the phase-space probability distributions are derived that correspond to the
Woodcock/Hoover-Evans (Sect. 3.4), Haile-Gupta (Sect. 3.4), Berendsen
(Sect. 3.5), Nos (Sect. 3.6) and Nos-Hoover (Sect. 3.6) thermostats. The deriva-
tions are given for all but the Berendsen thermostat, for which the result is merely
quoted.
The proof that the Nos thermostat samples a canonical ensemble of microstates,
provided that g = N
d f
+1 (virtual-time sampling) or g = N
d f
(real-time sampling),
is as follows [53]. The partition function of the microcanonical ensemble generated
for the extended system using virtual-time sampling (i.e., using the natural time evo-
lution of the extended system) reads
Z
e,v
= C
_
d p
_
d r
_
d p
s
_
d s [H
e
( r, p, s, p
s
) E
e
] , (91)
where C is a normalization constant, E
e
the (constant) extended-systemenergy, the
Dirac delta function, and H
e
the extended-systemHamiltonian (Eq. (67)). Recasting
this expression in terms of the real-system momenta p
i
= s
1
p
i
and substituting
r = r leads to
Z
e,v
= C
_
dp
_
dr
_
d p
s
_
d s s
N
d f

_
H(r, p) +
1
2
Q
1
p
2
s
+ g
1
ln s E
e
_
, (92)
where H is the real-system Hamiltonian, i.e.
H(r, p) =
1
2
N

i=1
m
1
i
p
2
i
+U(r) . (93)
140 Philippe H. Hnenberger
The argument of the delta function in Eq. (92) has a single zero with respect to the s
variable, namely
s
o
= exp
_
g
1

_
H(r, p) +
1
2
Q
1
p
2
s
E
e
__
. (94)
Using the relationship [ f ( s)] =| f

( s
o
) |
1
( s s
o
), one obtains
Z
e,v
= Cg
1

_
dp
_
dr
_
d p
s
_
d s s
N
d f
+1
[ s s
o
] (95)
= Cg
1

_
dp
_
dr
_
d p
s
exp
_
(N
d f
+1)g
1

_
H(r, p) +
1
2
Q
1
p
2
s
E
e
__
.
Integrating with respect to the variable p
s
and using the appropriate Gaussian integral
gives
Z
e,v
= C

_
dp
_
dr exp[(N
d f
+1)g
1
H(r, p)] (96)
with
C

= C[(N
d f
+1)]
1/2
(2gQ)
1/2
exp[(N
d f
+1)g
1
E
e
] . (97)
Equation (96) shows that the virtual-time extended-system ensemble average of any
quantity A depending on the real-system coordinates r = r and momenta p = s
1
p
(and also possibly on s), dened as
A(r, p)
e,v
=
_
dp
_
dr
_
d p
s
_
d s s
N
d f
+1
A(r, p)[ s s
o
]
_
dp
_
dr
_
d p
s
_
d s s
N
d f
+1
[ s s
o
]
, (98)
is equivalent to a canonical ensemble average, i.e.
A(r, p)
e,v
= A(r, p) when g = N
d f
+1. (99)
If real-time sampling is used instead, the probability of any microstate in the en-
semble is amplied by a factor s
1
due to the contraction of the timescale. For ex-
ample, ten microstates at 1 ps interval in the extended system represent 10 ps of
real-system time if s = 1 but only 5 ps if s = 2. Thus, the larger s, the lower the
real-system weight. Consequently, the real-time ensemble average of the quantity A,
dened from Eqs. (98) and (101) as
A(r, p)
e,r
=
_
dp
_
dr
_
d p
s
_
d s s
N
d f
A(r, p)[ s s
o
]
_
dp
_
dr
_
d p
s
_
d s s
N
d f
[ s s
o
]
, (100)
satises
A(r, p)
e,r
= s
1

1
e,v
s
1
A(r, p)
e,v
, (101)
Thermostat Algorithms 141
irrespective of the value of g. A straightforward consequence of Eqs. (98) and (101)
is that the real-time extended-system ensemble average of any quantity A is equiva-
lent to a canonical ensemble average, i.e.
A(r, p)
e,r
= A(r, p) when g = N
d f
. (102)
From Eq. (96), the real-system phase-space probability density for the Nos thermo-
stat (virtual-time sampling) can be written

v
(r, p) =
exp[(N
d f
+1)g
1
H(r, p)]
_
dp
_
dr exp[(N
d f
+1)g
1
H(r, p)]
. (103)
The proof that the Woodcock/Hoover-Evans thermostat samples a canonical en-
semble of congurations provided that g = N
d f
1 follows similar lines [53]. It has
been seen that this thermostat is identical to the Nos thermostat with the constraints
of Eq. (85) and the Hamiltonian of Eq. (86). In this case, the analog of Eq. (92) reads
Z
e,v
= C
_
dp
_
dr
_
d s s
N
d f

g
1/2

1/2
s
_
N

i=1
m
1
i
p
2
i
_
1/2
s

_
H
c
_
r, sp
_
E
e
_
,
= Cg
1
_
dp
_
1
2
N

i=1
m
1
i
p
2
i

1
2
g
1
_

_
dr
_
d s s
N
d f
1

_
1
2
g
1
+U(r) + g
1
ln s E
e
_
= C
_
dp
_
1
2
N

i=1
m
1
i
p
2
i

1
2
g
1
_

_
dr
_
d s s
N
d f

_
s exp
_
g
1

_
1
2
g
1
+U(r) E
e
___
= C

_
dp
_
1
2
N

i=1
m
1
i
p
2
i

1
2
g
1
_
_
dr exp
_
N
d f
g
1
U(r)
_
. (104)
The second equality follows from the relationship [ f (x)] =| f

(x
o
) |
1
(x x
o
),
with x =
1
2

N
i=1
m
1
i
p
2
i
and x
o
=
1
2
g
1
and inserting Eqs. (85) and (86). The
third equality follows from the relationship [ f ( s)] =| f

( s
o
) |
1
( s s
o
). This
partition function is canonical in the congurations if g = N
d f
(virtual-time sam-
pling). Considering Eq. (101), Eq. (104) shows that the real-time extended-system
ensemble average of any quantity A depending solely on the coordinates is equiva-
lent to a canonical ensemble average if g = N
d f
1. FromEq. (104), the real-system
phase-space probability density for the Woodcock/Hoover-Evans thermostat can be
written (real-time sampling)

r
(r, p) =
(
1
2

N
i=1
m
1
i
p
2
i

1
2
g
1
) exp[(N
d f
1)g
1
U(r)]
_
dr exp[(N
d f
1)g
1
U(r)]
. (105)
142 Philippe H. Hnenberger
The derivation of the phase-space probability distribution for the Haile-Gupta
thermostat follows similar lines [53]. It has been seen that this thermostat is identi-
cal to the Nos thermostat with the constraints of Eq. (85) and the Hamiltonian of
Eq. (89). In this case, the analog of Eq. (104) reads
Z
e,v
= C
_
dp
_
dr
_
d s s
N
d f

g
1/2

1/2
s
_
N

i=1
m
1
i
p
2
i
_
1/2
s

[H
h
(r, sp) E
e
]
= Cg
1
_
dp
_
1
2
N

i=1
m
1
i
p
2
i

1
2
g
1
_

_
dr
_
d s s
N
d f
1

_
g
1
s +U(r) E
e
_
= C

_
dp
_
1
2
N

i=1
m
1
i
p
2
i

1
2
g
1
_

_
dr{g
1
[E
e
U(r)]}
N
d f
1
h[E
e
U(r)] , (106)
where h is the Heaviside function. This function arises because s 0, so that E
e
<
U(r) leads to no solution for s. Thus, the real-systemphase-space probability density
corresponding to the Haile-Gupta thermostat (virtual-time sampling) is

v
(r, p) =
(
1
2

N
i=1
m
1
i
p
2
i

1
2
g
1
){g
1
[E
e
U(r)]}
N
d f
1
h[E
e
U(r)]
_
dr {g
1
[E
e
U(r)]}
N
d f
1
h[E
e
U(r)]
.
(107)
Using Eq. (106), it is easily seen that
s
e,v
= g
1
[E
e
U(r)]
e,v
. (108)
Thus, E
e
in Eq. (107) can be evaluated as
E
e
= g
1
s
e,v
+U(r)
e,v
. (109)
This distribution function is not canonical, irrespective of the value of g (an alterna-
tive derivation of this result can be found in [109]).
The proof that the Nos-Hoover thermostat samples a canonical ensemble of
microstates provided that g = N
d f
is as follows [63]. Consider the Nos-Hoover
equations of motion, Eqs. (79) and Eq. (81). Because the variables r, p, and are
independent, the ow of the (2N
d f
+ 1)-dimensional probability density (r, p, )
is given by the generalized (non-Hamiltonian) analog of the Liouville equation
21

t
=

r
r

p
p

_

r
r +

p
p +

_
. (111)
21
The generalized Liouville equation states the conservation of the total number of systems
in an ensmble. If = (q, p), this conservation law can be written
Thermostat Algorithms 143
One can postulate the following extended-system phase-space distribution function

e,r
(r, p, ) = C exp
_

_
U(r) +
1
2
N

i=1
m
1
i
p
2
i
+
1
2
Q
2
__
, (112)
where C is a normalization factor. Using Eqs. (58), (79), and (81) the derivatives
involved in Eq. (111) are

r
r =
N

i=1
m
1
i
F
i
p
i

p
p =
N

i=1
m
1
i
p
i
(F
i
p
i
)

= k
B
N
d f
T
_
g
N
d f
T
o
T
1
_

r
r = 0

p
p = N
d f

= 0 . (113)
Using these results and the denition of T, it is easily shown that /t = 0 in
Eq. (111) provided that g = N
d f
. This shows that the extended-system phase-space
density
e,r
(r, p, ) is a stationary (equilibrium) solution of Eq. (111) corresponding
to the Nos-Hoover equations of motion. Integrating out the variable leads to the
real-system phase-space probability density for the Nos-Hoover thermostat (real-
time sampling)

r
(r, p) =
exp[N
d f
g
1
H(r, p)]
_
dp
_
dr exp[N
d f
g
1
H(r, p)]
. (114)
which is the canonical probability density if g = N
d f
. The Nos-Hoover equations of
motion are unique in leading to the stationary extended-system phase-space density
of Eq. (112). However, they are not unique in leading to the real-system phase-space
density, because the distribution (Gaussian in Eq. (112)) is irrelevant here. Finally,
it should be mentionned that because the Nos-Hoover thermostat can be derived
from the Nos thermostat with real-time sampling, the above proof is not really nec-
essary. It is given anyway as a nice illustration of the use of the generalized Liouville
equation to derive probability distribution functions for thermodynamical ensembles.

t
=

(

) or
d
dt
=

. (110)
The two forms can be interconverted by expressing d/dt as a total derivative. If the
equations of motion are Hamiltonian, one shows easily that d/dt = 0. If ( ) is a
stationary (equilibrium) solution, one has /t = 0.
144 Philippe H. Hnenberger
The derivation of the phase-space probability distribution for the Berendsen ther-
mostat follows again similar lines [109]. The nal expression relies on the assump-
tion of a relationship
[K
2
K
2
]
1/2
= (
B
)[U
2
U
2
]
1/2
, (115)
between the uctuations in the kinetic and potential energies in simulations with the
Berendsen thermostat. Clearly, such a relationship exists for any system. However,
it is unclear is whether a common (
B
) applies to all systems, irrespective of their
composition and size. The derivation is then based on nding a stationary solution
for the generalized Liouville equation (Eq. (111)). The nal (approximate) result is
(with g = N
d f
)

b
(r, p)=

p
(p) exp{[U(r) N
1
d f

2
[U(r)
2
U(r)
2
b
]]}
_
dp
p
(p)
_
dr exp{[U(r) N
1
d f

2
[U(r)
2
U(r)
2
b
]]}
,
(116)
where
p
(p) is the (unknown) momentum probability distribution. Note that the
Haile-Gupta thermostat generates congurations with the same probability distrib-
ution as the Berendsen thermostat with = 1/2 [109].
References
1. Baschnagel J, Binder K, Doruker P, Gusev AA, Hahn O, Kremer K, Mattice WL, Mller-
Plathe F, Murat M, Paul W, Santos S, Suter UW, Tries W (2000) Adv Polymer Sci
152:41156 106
2. van Gunsteren WF, Berendsen HJC (1990) Angew Chem Int Ed Engl 29:9921023 106,
107, 111
3. Karplus M, McCammon JA (2002) Nature Struct Biol 9:646652 106
4. Daura X, Glttli A, Gee P, Peter C, van Gunsteren WF (2002) Adv Prot Chem 62:341
360 106
5. Cheatham III TE, Kollman PA (2000) Annu Rev Phys Chem 51:435471 106
6. Saiz L, Bandyopadhyay S, Klein ML (2002) Bioscience Reports 22:151173 106
7. Engelsen SB, Monteiro C, Herv de Penhoat C, Prez S (2001) Biophys Chem 93:103
127 106
8. Warshel A, King G (1985) Chem Phys Lett 121:124129 107
9. King G, Warshel A (1989) J Chem Phys 91:36473661 107
10. Juffer AH, Berendsen HJC (1993) Mol Phys 79:623644 107
11. Brooks III CL, Karplus M (1983) J Chem Phys 79:63126325 107, 120
12. Brnger A, Brooks III CL, Karplus M (1984) Chem Phys Lett 105:495500 107, 120
13. Beutler TC, Mark AE, van Schaik R, Gerber PR, van Gunsteren WF (1994) Chem Phys
Lett 222:529539 107
14. Essex JW, Jorgensen WL (1995) J Comput Chem 16:951972 107
15. Im W, Berbche S, Roux B (2001) J Chem Phys 117:29242937 107
16. Allen MP, Tildesley DJ (1987) Computer simulation of liquids. Oxford University Press,
New York 107, 110, 111
Thermostat Algorithms 145
17. Bekker H (1997) J Comput Chem 18:19301942 107
18. Tomasi J, Persico M (1994) Chem Rev 94:20272094 107
19. Smith PE, Pettitt BM (1994) J Phys Chem 98:97009711 107
20. Roux B, Simonson T (1999) Biophys Chem 78:120 107
21. Orozco M, Luque FJ (2000) Chem Rev 100:41874225 107
22. Schneider T, Stoll E (1978) Phys Rev B 17:13021322 107, 111, 120
23. Schneider T, Stoll E (1978) Phys Rev B 18:64686482 107, 111, 120
24. van Gunsteren WF, Berendsen HJC, Rullmann JAC (1981) Mol Phys 44:6995 107, 111
25. kesson T, Jnsson B (1985) Mol Phys 54:369381 107, 111
26. Brjesson U, Hnenberger PH (2001) J Chem Phys 114:97069719
27. Gros P, van Gunsteren WF, Hol WGJ (1990) Science 249:11491152 108
28. Gros P, van Gunsteren WF (1993) Mol Simul 10:377395 108
29. Schiffer CA, Gros P, van Gunsteren WF (1995) Acta Crystallogr D D51:8592 108
30. van Gunsteren WF, Brunne RM, Gros P, van Schaik RC, Schiffer CA, Torda AE (1994)
Accounting for molecular mobility in structure determination based on nuclear magnetic
resonance spectroscopic and X-ray diffraction data. In: James TL, Oppenheimer NJ (eds)
Methods in enzymology: nuclear magnetic resonance, Vol. 239, Academic Press, New
York, pp 619654 108
31. Torda AE, Scheek RM, van Gunsteren WF (1989) Chem Phys Lett 157:289294 108
32. Torda AE, Scheek RM, van Gunsteren WF (1990) J Mol Biol 214:223235 108
33. Torda AE, Brunne RM, Huber T, Kessler H, van Gunsteren WF (1993) J Biomol NMR
3:5566 108
34. van Gunsteren WF, Berendsen HJC (1977) Mol Phys 34:13111327 108
35. Ryckaert JP, Ciccotti G, Berendsen HJC (1977) J Comput Phys 23:327341 108
36. Andersen HC (1983) J Comput Phys 52:2434 108
37. Miyamoto S, Kollman PA (1992) J Comput Chem 13:952962 108
38. Hess B, Bekker H, Berendsen HJC, Fraaije JGEM(1997) J Comput Chem 18:14631472
108
39. Krutler V, van Gunsteren WF, Hnenberger PH (2001) J Comput Chem 22:501508
108
40. Tironi IG, Brunne RM, van Gunsteren WF (1996) Chem Phys Lett 250:1924 108
41. Lebowitz JL, Percus JK, Verlet L (1967) Phys Rev 153:250254 108
42. Cheung PSY (1977) Mol Phys 33:519526 108
43. Ray JR, Graben HW (1981) Mol Phys 43:12931297 108
44. Ray JR, Graben HW (1991) Phys Rev A 44:69056908 108
45. Graben HW, Ray JR (1991) Phys Rev A 43:41004103 108, 111, 113
46. Haile JM, Graben HW (1980) J Chem Phys 73:24122419 108
47. Haile JM, Graben HW (1980) Mol Phys 40:14331439 108
48. Ray JR, Graben HW (1986) Phys Rev A 34:25172519 108
49. Ray JR, Graben HW, Haile JM (1981) J Chem Phys 75:40774079 108
50. Ray JR, Graben HW (1990) J Chem Phys 93:42964298 108
51. Evans DJ, Morriss GP (1983) Chem Phys 77:6366 108, 124
52. Evans DJ, Morriss GP (1983) Phys Lett 98A 98A:433436 108, 124
53. Nos S (1984) J Chem Phys 81:511519 108, 127, 129, 131, 133, 136, 138, 140, 141
54. Berendsen HJC, Postma JPM, van Gunsteren WF, DiNola A, Haak JR (1984) J Chem
Phys 81:36843690 108, 121, 127
55. Andersen HC (1980) J Chem Phys 72:23842393 108, 117, 122, 123
56. Parrinello M, Rahman A (1980) Phys Rev Lett 45:11961199 108
57. Parrinello M, Rahman A (1982) J Phys Chem 76:26622666 108
146 Philippe H. Hnenberger
58. Ryckaert JP, Ciccotti G (1983) J Chem Phys 78:73687374 108
59. Nos S, Klein ML (1983) Mol Phys 50:10551076 108
60. Heyes DM (1983) Chem Phys 82:285301 108
61. Nos S (1984) Mol Phys 52:255268 108, 129, 132, 136
62. Brown D, Clarke JHR (1984) Mol Phys 51:12431252 108
63. Hoover WG (1985) Phys Rev A 31:16951697 108, 129, 131, 133, 135, 137, 142
64. Hoover WG (1986) Phys Rev A 34:24992500 108
65. Ferrario M, Ryckaert JP (1985) Mol Phys 54:587603 108
66. Ray JR, Rahman A (1985) J Chem Phys 82:42434247 108
67. Melchionna S, Ciccotti G, Holian BL (1993) Mol Phys 78:533544 108
68. Martyna GJ, Tobias DJ, Klein ML (1994) J Chem Phys 101:41774189 108
69. Melchionna S, Ciccotti G (1997) J Chem Phys 106:195199 108
70. Hnenberger PH (2002) J Chem Phys 116:68806897 108
71. a gin T, Pettitt BM (1991) Mol Phys 72:169175 108
72. Ji J, a gin T, Pettitt BM (1992) J Chem Phys 96:13331342 108
73. Lo C, Palmer B (1995) J Chem Phys 102:925931 108
74. Lynch GC, Pettitt BM (1997) J Chem Phys 107:85948610 108
75. Fixman M (1974) Proc Natl Acad Sci USA 71:30503053 109
76. Lado F (1981) J Chem Phys 75:54615463 109
77. Wallace DC, Straub GK (1983) Phys Rev A 27:22012205 109
78. a gin T, Ray JR (1988) Phys Rev A 37:247251 109
79. a gin T, Ray JR (1988) Phys Rev A 37:45104513 109
80. Lustig R (1994) J Chem Phys 100:30483059 109
81. Ray JR, Zhang H (1999) Phys Rev E 59:47814785 109
82. Metropolis N, Rosenbluth AW, Rosenbluth MN, Teller AH, Teller E (1953) J Chem Phys
21:10871092 111, 119
83. Valleau JP, Whittington SG (1977) A guide to Monde Carlo for statistical mechanics:
2. Byways. In: Berne BJ (ed) Modern theoretical chemistry, Vol. 5, Plenum Press, New
York, pp 137168 111, 119
84. Frenkel D (1993) Monte Carlo simulations: A primer. In: van Gunsteren, WF, Weiner PK,
Wilkinson AJ (eds) Computer simulation of biomolecular systems, theoretical and exper-
imental applications, Vol. 2, ESCOM Science Publishers, B.V., Leiden, The Netherlands,
pp 3766 111, 119
85. van Gunsteren WF, Nanzer AP, Torda AE (1995) Molecular simulation methods for gen-
erating ensembles or trajectories consistent with experimental data. In: Binder K, Ciccotti
G (eds) Monte Carlo and molecular dynamics of condensed matter systems, Proceedings
of the Euroconference, Vol. 49, SIF, Bologna, Italy, pp 777788 112
86. Evans DJ, Sarman S (1993) Phys Rev E 48:6570 112, 117
87. Rugh HH (1997) Phys Rev Lett 78:772774 113
88. Butler BD, Ayton G, Jepps OG, Evans DJ (1998) J Chem Phys 109:65196522 113
89. Hockney RW (1970) Methods Comput Phys 9:136211 115
90. Morishita T (2003) Mol Simul 29:6369 116, 135
91. Holian BL, Evans DJ (1983) J Chem Phys 78:51475150 117
92. Adams DJ (1975) Mol Phys 29:307311 120
93. Creutz M (1983) Phys Rev Lett 50:14111414 120
94. Ray JR (1991) Phys Rev A 44:40614064 120
95. Ciccotti G, Ferrario M, Ryckaert JP (1982) Mol Phys 46:875889 120
96. Berkowitz M, Morgan JD, McCammon JA (1983) J Chem Phys 78:32563261 120
97. Vesely FJ (1984) Mol Phys 53:505524 120
Thermostat Algorithms 147
98. Gurdia E, Padr JA (1985) J Chem Phys 83:19171920 120
99. van Gunsteren WF, Berendsen HJC (1988) Mol Simul 1:173185 120
100. Wang W, Skeel RD (2003) Mol Phys 101:21492156 120
101. Berkowitz M, McCammon JA (1982) Chem Phys Lett 90:215217 120
102. Tanaka H, Nakanishi K, Watanabe N (1983) J Chem Phys 78:26262634 123
103. Woodcock LV (1971) Chem Phys Lett 10:257261 124, 125
104. Hoover WG, Ladd AJC, Moran B (1982) Phys Rev Lett 48:18181820 124, 127
105. Evans DJ (1983) J Chem Phys 78:32973302 124
106. Evans DJ, Hoover WG, Failor BH, Moran B, Ladd AJC (1983) Phys Rev A 28:1016
1020 124
107. Broughton JQ, Gilmer GH, Weeks JD (1981) J Chem Phys 75:51285132 126
108. Haile JM, Gupta S (1983) J Chem Phys 79:30673076 126, 127
109. Morishita T (2000) J Chem Phys 113:29762982 127, 129, 142, 143, 144
110. Nos S (1993) Phys Rev E 47:164177 132
111. Holian BL, De Groot AJ, Hoover WG, Hoover CG (1990) Phys Rev A 41:45524553
135
112. Toxvaerd S (1991) Mol Phys 72:159168 135
113. Martyna GJ, Tuckerman ME, Tobias DJ, Klein ML (1996) Mol Phys 87:11171157 135
114. Jang S, Voth GA (1997) J Chem Phys 107:95149526 135
115. Hoogerbrugge PJ, Koelman JMVA (1992) Europhys Lett 19:155160 137
116. Espaol P, Warren P (1995) Europhys Lett 30:191196 137
117. Groot RD, Warren PB (1997) J Chem Phys 107:44234435 137
118. Soddemann T, Dnweg B, Kremer K (2003) Phys Rev E 68:046702(1)046702(8) 137
119. Jellinek J, Berry SR (1988) Phys Rev A 38:30693072 137
120. Hoover WG (1989) Phys Rev A 40:28142815 137
121. Hamilton IP (1990) Phys Rev A 42:74677470 137
122. Winkler RG (1992) Phys Rev A 45:22502255 137
123. Martyna GJ, Klein ML, Tuckerman M (1992) J Chem Phys 97:26352643 137
124. Holian BL, Posch HA, Hoover WG (1993) Phys Rev E 47:38523861 137
125. Bra nka AC, Wojciechowski KW (2000) Phys Rev E 62:32813292 137
126. Bra nka AC, Kowalik M, Wojciechowski KW (2003) J Chem Phys 119:19291936 137
127. Laird BB, Leimkuhler BJ (2003) Phys Rev E 68:016704(1)016704(6) 137
128. Posch HA, Hoover WG, Vesely FJ (1986) Phys Rev A 33:42534265 137
129. Jellinek J, Berry SR (1989) Phys Rev A 40:28162818 137
130. Toxvaerd S (1990) Ber Bunsenges Phys Chem 94:274278 137
131. Calvo F, Galindez J, Gada FX (2002) J Phys Chem A 106:41454152 137
132. DAlessandro M, Tenenbaum A, Amadei A (2002) J Phys Chem B 106:50505057 137
133. Winkler RG, Kraus V, Reineker P (1995) J Chem Phys 102:90189025 137
134. Pak Y, Wang S (1999) J Chem Phys 111:43594361 137
135. Fukuda I (2001) Phys Rev E 64:016203(1)016203(8) 137
136. Fukuda I, Nakamura H (2002) Phys Rev E 65:026105(1)026105(5) 137
137. Barth EJ, Laird BB, Leimkuhler BJ (2003) J Chem Phys 118:57595768 137
138. Harvey SC, Tan RKZ, Cheatham III TE (1998) J Comput Chem 19:726740 138
139. Chiu SW, Clark M, Subramaniam S, Jakobsson E (2000) J Comput Chem 21:121131
138
Index
boundary condition, 107
hard, 107
soft, 107
spatial, 107
thermodynamical, 108
Ensemble, 108
ensemble
canonical, 108
generalized, 108
grand-canonical, 108
grand-isothermal-isobaric, 108
grand-microcanonical, 108
isoenthalpic-isobaric, 108
isothermal-isobaric, 108
microcanonical, 108
thermodynamical, 108
force
frictional, 107, 108
stochastic, 107, 108
Hamiltonian
rotational-invariance, 108
time-independence, 108
translational-invariance, 108
molecular dynamics, 107