J.

DIFFERENTIAL GEOMETRY
18 (1983) 523-557

THE LOCAL STRUCTURE OF POISSON

MANIFOLDS

ALAN WEINSTEIN

Table of Contents
Historical Introduction 523
Acknowledgments 526
1. Poisson Manifolds and Mappings 527
2. Splitting 530
3. Linear Poisson Structures 533
4. Linear Approximation 535
5. Hamiltonian Systems 536
6. The Linearization Problem 537
7. Function Groups, Realizations, and Momentum Mappings 540
8. Dual Pairs and Gauge Groups 542
9. Existence of Realizations 544
10. Uniqueness of Realizations 549
11. The Restricted Three Body Problems and Other Examples 550
References 555

Historical Introduction
The classical Poisson bracket operation defined on functions on R2" is

(*) {/,*}= Σ
In the early nineteenth century, Poisson noticed that the vanishing of {/, g}
and {/, h] imply that of {/, {g, A}}; almost thirty years later Jacobi dis-
covered the identity {/, {g, h}} = {{/, g}, h) + {g, {/, h}} which "explains"
Poisson's theorem. In his study of general composition laws satisfying the
Jacobi identity, Lie [29] defined in local coordinate form what is now known as
a Poisson structure. On Rr such a structure is given by functions w^x^- ,jcr)
satisfying the identities
w.. +Wβ = o,

Received September 17, 1982. Research supported by the Miller Institute and National Science
Foundation Grant MCS 80-23356.
524 ALAN WEINSTEIN

which imply that the bilinear operation

is antisymmetric and satisfies the Jacobi identity; i.e., the algebra of functions
C°°(Rr) becomes a Lie algebra. An abstractly defined Lie algebra structure { , }
on C°°(Rr) arises in this way if and only if it satisfies the Leibniz identity
{F9 GH} = {F, G}H + G{F9 H}9 and this enables us to define a Poisson
structure on a manifold P to be a Lie algebra structure { , } on C°°(P) which
satisfies the Leibniz identity. The functions wtJ may then be seen as the
components in local coordinates of an antisymmetric contravariant 2-tensor w;
the Jacobi identity may be interpreted as the vanishing on w of a certain
natural quadratic differential operator of first order. (Berezin [5], Hermann
[19], Lichnerowicz [28], Tulczejew [44]).
Poisson structures have recently become interesting in connection with
completely integrable systems and for the hamiltonian formulation of field
theoreis in physical variables. (See the papers in Tabor and Treve [43] and
references therein.) The aim of this paper is to develop the theory of Poisson
manifolds with an eye toward these applications and also a new application—
the study of singular limits of hamiltonian systems.
The natural setting for hamiltonian systems is on symplectic manifolds.
These may be described as manifolds carrying a Poisson structure which is
locally isomorphic to the standard one on an R2rt; the coordinates
(<7i>* * *>#/!> PM' ' '>Pn)a r e ^en called canonical variables. (There is an exten-
sion of this definition to infinite dimensions, which we ignore for now.) A
criterion for a Poisson structure to be symplectic is that the tensor w (or the
matrix (wzy)) have rank everywhere equal to the dimension of the manifold.
The simplest example of a nonsymplectic Poisson structure is R2w+S with
variables (ql9 -9qn9 pl9 -9pn9 cl9- 9cs) and Poisson brackets given by the
standard formula (*). Functions of (cl9 -9cs) have zero Poisson bracket with
everything and are called by Lie distinguished functions. (Nowadays they are
called invariants or Casimir functions.) Each manifold on which all cj are
constant inherits a Poisson structure which is then nondegenerate, so R2n+S
may be thought of as foliated by symplectic manifolds. This turns out to be the
situation at the generic points of any Poisson manifold: Lie proved that any
Poisson structure on R2/1+s for which the rank of wtj is constant1 and equal to
2« admits s independent distinguished functions and so is locally isomorphic

'Actually, Lie assumed the constancy of rank without saying it explicitly.

 THE LOCAL STRUCTURE OF POISSON MANIFOLDS 52 5

to the standard example just described; furthermore, the rank of any Poisson
structure is constant (and even) on the open subset where it attains its maximal
2
value.
Another important example of a Poisson structure, also introduced by Lie,
3
arises when we are given the structure constants of a Lie algebra, i.e., r
constants cijk satisfying the identities
c c c
Uk + βk = °> Cijj + Cjki + ku = 0.
The functions w l 7 (x) = Σk=\CiJkxk then define a Poisson structure on Rr.3 The
rank of (w ; y ) is no longer constant (e.g. it is zero at x = 0), so Lie's theory
applies only at the "regular points" where the rank is maximal. This was
enough for Lie to prove his " third theorem" on the existence of a local group
with given structure constants and to classify locally the r 'tuples (φ{9 ,φ r ) of
functions on R2w satisfying the bracket relations (φ,, φj} = Σk=ι ciJkφk. In the
course of this classification, Lie observed that the leaves of the symplectic
foliation of Rr were just the "smallest invariant submanifolds for the dual of
the adjoint group" acting on RΓ.
Nowadays we think of RΓ as the dual space g* of the Lie algebra ( Q , [ , ] )
whose basis Xx, ,Xk satisfies [Xi9 Xj] = Σrk=ιcijkXk. The "dual of the ad-
joint group" is called the coadjoint representation, and the smallest invariant
manifolds are the coadjoint orbits.4 Their symplectic structure was redis-
covered in the 1960's by Kirillov [24], Kostant [26], and Souriau [41], whose
arguments covered the singular coadjoint orbits as well as the regular ones. In
fact it turns out (see Kirillov [25]) that through each point of every Poisson
manifold there passes a symplectic manifold whose dimension equals the rank
of the Poisson structure there, and the Poisson structure is built up out of the
Poisson structures on these symplectic leaves. For a g * carrying Lie's Poisson
structure, the symplectic leaves are just the coadjoint orbits.
This property of being smoothly decomposed into symplectic manifolds of
different dimensions seems to make Poisson manifolds an appropriate setting
for studying a phenomenon which is quite common in mechanics: if a
mechanical system is modeled by a symplectic manifold, then when a parame-
ter in the system reaches a limiting value (usually 0 or oo), the limiting system
also has a symplectic formulation, but with fewer degrees of freedom. Exam-
ples of this are:
(i) the restricted 3-body problem in celestial mechanics (mass -> 0);
2
These results were rediscovered by Lichnerowicz [28] and Hermann [20].
3
Introduced again by Berezin [4].
4
By the way, Lie also discovered and used, more or less as it is used now, the momentum
mapping and its equivariance under the coadjoint representation.
526 ALAN WEINSTEIN

(ii) the guiding center limit for a particle in an electromagnetic field

(charge/mass -> oo);
(iii) the limit of discrete vortices in the motion of incompressible fluids
(concentration of vorticity -> oo);
(iv) the classical limit of quantum mechanics (Λ -* 0).
There are also examples when the number of degrees of freedom remains the
same but the global structure of a symplectic manifold of group changes, as in
the newtonian limit of special relativity (c -* oo); these too should be accessi-
ble to study in terms of Poisson structures.
A good part of this paper then will consist of an extension of Lie's results to
the case of variable rank. Most of the methods are ones which Lie himself
would have used, but the geometric language developed since his time gives us
new insight and enables us to say some things more efficiently.
There is one new aspect to the theory which is peculiar to the case of
variable rank. At a point where the functions w^x) vanish, we can linearize
them to obtain a linear Poisson structure living on the tangent space. The
linearization question is whether the original Poisson structure is locally equiva-
lent to a linear one. The question has different answers according to whether
we work in the formal, analytic, or C 00 categories, and there remain some
unsolved problems.
Some interesting work on the singular structure of Poisson manifolds, from a
more algebraic viewpoint than ours, can be found in the papers of Vinogradov
and Krasilshchik [46] and Berger [6].
To end this introduction, the author wishes to point out that a great deal of
interest today in symplectic actions and coadjoint orbits lies in their relevance
for understanding unitary actions of Lie groups. This is truly a modern
development for which there does not seem to be any precedent in Lie's work.

Acknowledgments
Bob Hermann and Wilfrid Schmid, in conversations and their published
work (Hermann [20], Schmid [40]) pointed the author toward Lie's [29]
pioneering work on Poisson manifolds. Hans Duistermaat was responsible for
my recent education in the geometry of Lie groups, some of which found direct
applications in this paper. The author's interest in applications of Poisson
manifolds to the understanding of variational principles and Clebsch variables
was stimulated at a conference at the La Jolla Institute (December 1981)
organized by Michael Tabor and Yvain Treve. Conversations there with Frank
 THE LOCAL STRUCTURE OF POISSON MANIFOLDS 52 7

Henyey, Darryl Holm, John Hubbard, Boris Kupershmidt, and Robert Little-
john were especially helpful. Finally, Jerry Marsden has been a patient listener
and frequent collaborator in the work described here—our joint work described
in Marsden and Weinstein [32] and Marsden, Ratiu and Weinstein [33] is a
close companion to the present paper.

1. Poisson manifolds and mappings

A Poisson structure on a manifold P is defined as a Lie algebra structure
{ , } on C°°(P) satisfying the Liebniz identity {FG, H) = F{G, H) + (F, H}G.
The bracket operation {,} is thus a derivation in each entry, and so in
particular for each function H there is a vector field ξH such that i-H F =
{F, H] for all F. ξH is called the hamiltonian vector field generated by H. At
any point p E P, the value of {F, H} and hence of ξH depends only on the
differential of H, so there is a bundle map B: T*P -> TP such that ξff = B o dH
for all H. We may also think of B as defining a contravariant antisymmetric
2-tensor w in P, for which (F, G] = ((dF9 dG), w). The tensor w is sometimes
called a cosymplectic structure; the Jacobi identity for { , } is equivalent to the
vanishing of the so-called Schouten [42] bracket [w, w].
In local coordinates (*,,- -,xr), a Poisson structure is determined by the
component functions wtJ(x) of w. In terms of the bracket we have simply
{*,, Xj) = w/y(x); in other words, the Poisson structure is specified if we give
the bracket relations satisfied by the coordinate functions. This is exactly how
Lie thought of Poisson structures; he noted that the bracket

{F,G}= 2 Wij-fo te
ΌX ΌX
i J = \ i j

satisfies the asymmetry and Jacobi conditions if and only if wtj = -wJi and

It follows from the Jacobi identity that the (local) flow of each ξH preserves
the Poisson structure, and also that £{#,#} = £#£// ~ %H%K — \%H> ^AΓ1» w e u s e
Arnold's [3] sign convention for the bracket of vector fields. Thus the Poisson
structure determines a Lie algebra homomorphism from C°°(P) to the infini-
tesimal automorphisms of the Poisson structure.
Automorphisms of a Poisson manifold are an example of Poisson mappings
defined in general as maps J: Pλ^> P2 between Poisson manifolds such that
{ F o / , G ° y } , = {F, G}2 © /, or equivalently by the condition J*wx{x) —
w2(J(x)). A Poisson submanfold is a submanifold Q in a Poisson manifold P
528 ALAN WEINSTEIN

with a Poisson structure for which the inclusion is a Poisson mapping. Such a
structure, if it exists, is unique.
Lemma 1.1. Q C P is a Poisson submanifold if and only if each tangent space
TXQ contains the image of Bx: T*P -> TXP, i.e., if and only if all hamiltonian
vector fields are tangent to Q.
Proof. Given functions F and G on Q, extend them to functions F and G
on P. (It will be enough to do this locally.) The tangency condition implies that
the restriction of {F, G} to Q depends only on F and G, so there is an induced
bracket operation on Q which is easily seen to make Q into a Poisson
submanifold. If the tangency condition fails, the bracket of extended functions
depends upon the extensions, so the restriction map C°°(P) -> C°°(Q) cannot
define a homomorphism for any Lie algebra structure on C°°(β).
A related fact is
Lemma 1.2. Let J: Px^> P2be a Poisson mapping, and H a function on P2.
Then the trajectories on P of the hamiltonian vector field ξH are the projections
under J of the trajectories on P2ofζH o j .
Proof. Let σ(t) be an integral curve of ξH o j . Then for any function G on

jt{Go(joa)) = - | ( ( G ° / ) ° σ ) = { G o / , H o J} o σ

= ({G,H] oj)oσ= {G,H} -(Joσ).

It follows that / ° σ is an integral curve of ζH. Clearly, all integral curves of ξH

arise in this way. q.e.d.
See Guillemin and Sternberg [15] and Mishchenko and Fomenko [34] for
applications of this lemma to the study of collective and invariant motion.
The rank of a Poisson structure at a point x E P is defined to be the rank of
Bx: T*P -> TXP. (In local coordinates it is the rank of the matrix w/y(jc).) The
invariance of the Poisson structure under hamiltonian flows implies the con-
stancy of rank along the orbits of such flows. From this it is not hard to derive
the following result, due to Kirillov [25] (see also Hermann [17]) in general but
to Lie [29] for the case of constant rank: Every Poisson manifold is essentially
a union of symplectic manifolds which fit together in a smooth way.
Proposition 1.3. Define a relation ~ on P by declaring x ~ y if y can be
reached from x by apiecewise smooth curve, each segment of which is a trajectory
of a hamiltonian vector field. Then — is an equivalence relation, and the
equivalence classes are Poisson submanifolds of P. The bracket of functions on P
is therefore determined by the brackets of their restrictions to these submanifolds.
The dimension of each such submanifold Q is equal to the rank of the Poisson
structure (of P or of Q) at each point of Q.
 THE LOCAL STRUCTURE OF POISSON MANIFOLDS 52 9

A Poisson structure for which the rank is everywhere equal to the dimension
of the manifold is called nondegenerate or symplectic. The symplectic structure
on such a manifold is the 2-form Ω defined by Ω(£, η) = w(B~ιξ, B~ Ί\)\ the
X

usual Poisson brackets and hamiltonian vector fields for this symplectic
structure are just the ones for the Poisson structure. We shall call the
equivalence classes Q in Proposition 1.3 the symplectic leaves of P, since they
form a foliation everywhere that the rank of the Poisson structure is locally
constant (for instance on the open dense subset of points with locally maximal
rank).
A Casimir function, or invariant, on a Poisson manifold is a function C such
that {C, F] = 0 for all functions F\ equivalent conditions are that C is
constant along the orbits of all hamiltonian vector fields or that the
hamiltonian vector field of C itself is zero. A Casimir function is constant
along each symplectic leaf, and in a region where the rank is constant, the
symplectic leaves are exactly the common level manifolds of the local Casimir
functions.
It is sometimes possible to define an induced Poisson structure on a
submanifold which is not a Poisson submanifold. For instance, only Poisson
submanifolds of a symplectic manifold are its connected components, but any
submanifold on which the symplectic form is nondegenerate has an induced
symplectic, hence Poisson structure. The following proposition will be useful in
the next section.
Proposition 1.4. Let Q be a submanifold of the Poisson manifold P such that
the following conditions are satisfied at each x E Q:

(ή)TxQ±nKeτBx={0}9
where TXQX is the annihilator of TXQ in T*P. Then there is a naturally induced
Poisson structure on Q.
Proof. Condition (ii) is equivalent to TXQ + Im Bx = TXP, i.e., Q intersects
each symplectic leaf transversely. Thus Q is a union of manifolds, each of
which is a submanifold of a symplectic manifold. Now the intersection
B^^Q^ Π TXQ is just the null space of the induced symplectic form on
Im Bx Π TXQ, so condition (i) tells us that the intersection manifolds making
up Q are in fact symplectic. Thus we can put together the Poisson brackets on
these manifolds to get a bracket on Q. To see that the resulting structure is
smooth, we note that conditions (i) and (ϋ) imply that B^T^) Θ TXQ = TXP
for each Q, giving a smooth bundle projection π from the restricted tangent
bundle TQP onto TQ. The induced Poisson structure on Q is then defined by
530 ALAN WEINSTEIN
B
Q

the composed map T*Q -> T£P -> TQP -* TQ. (The argument involving sub-
manifolds was needed to show that the induced structure satisfies the Jacobi
and Leibniz identities.)

2. Splitting
The product Px X P2 becomes a Poisson manifold in an obvious way so that
the projections ττ7: PXX P2-+ Pj are Poisson mappings, and πx(C°°(Px)) and
7Γ2(C°°(P2)) are commuting subalgebras of C°°(P). In terms of coordinates, if
bracket relations {xi9 Xj) = wtJ(x) and {yi9 yj} = vtj{y) are given, then these
define a bracket on functions of x and y when augmented by the relations

The main theorem of this section states that every Poisson manifold is
locally the product of a symplectic manifold and a Poisson manifold having a
point where the rank is zero.
Theorem 2.1 (Splitting theorem). Let x0 be any point in a Poisson manifold
P. Then there are a neighborhood U ofx0 in P and an isomorphism φ = φs X φN
from U to a product S X N such that S is symplectic and the rank ofN at φN(x0)
is zero. The factors S and N are unique up to local isomorphism.
Proof. (The following existence proof for the decomposition is essentially a
standard proof of Darboux's theorem for symplectic manifolds.) If the rank of
P at x0 is zero, we are done; otherwise, there are functions q[ andpx such that
{q[9 P\}(x0) ¥= 0. Then ξPι(x0) Φ 0, so by straightening out this vector field
near JC0 we can find a function qx such that ξpιqλ = 1, i.e., {qx, px) — 1. Now
the vector fields ξqχ and ξp^ commute, so we can find functions x 3 , ,xn such
that the Xj's commute with qλ and pl9 and (qx, pλ9 x3,- - ,xr) form a coordi-
nate system. By Poisson's theorem, the brackets (JC^ Xj) commute with qx and
px as well; thus
3{jcί, x }
0 = {{Xi,Xj},qι} =—g^—{P\99\} +zeros,

so {.X;, Xj) is independent of px and similarly of qx. Hence we have relations

{Λ:,, Xj) = ϋ / 7 (x), so P is locally the product of a two-dimensional symplectic
manifold and a Poisson manifold whose dimension and rank at each point are
two less than for P.
Repeating this process as often as necessary, we find "canonical" coordi-
nates (qι,~',qk,Pι9' ',Pk,yι, ' ,ys) with {qi9 qj) = {/?,., pj) = {q., y,) =
{Pi> yj} = 0, {qi9 Pj) = δiJt and {yi9 yj\ = v^y) with o i y = 0 at x0. This
completes the existence proof.
 THE LOCAL STRUCTURE OF POISSON MANIFOLDS 531

The uniqueness of the symplectic factor S follows from the fact that its
dimension is just the rank of P at JC0. A "natural" representative for S is the
symplectic leaf through x0. (In fact, the argument up to this point gives
another proof of the existence of the symplectic leaf.)
To prove the uniqueness of N, we will identify the functions vtJ(y) as the
induced Poisson structure on a submanifold transverse to the symplectic leaf,
using Proposition 1.4.
Suppose then that we have coordinates (/?, q, y) as just constructed. We
may assume that all the coordinates vanish at x 0 , so we may identify the
symplectic factor S with the set j> = 0 and the singular factor N with q = p = 0.
We shall show that the Poisson structure on N is just the induced structure
from Proposition 1.4, with N considered as a submanifold.
The tangent bundle TN is spanned by the vector fields 3/3^ and the
orthogonal bundle TQ^ by the 1-forms dqj and dpy Now B(dqj) = -3/3/?,
and B(dpj) = d/dqj, so the hypothesis of Proposition 1.4 are satisfied. The
projection TΓ from TNP to TN kills 3/3<77 and 3/3/?y and fixes d/dyi9 so TΓ*
takes dyi to dyt. Thus the induced Poisson structure maps

and is equal to the structure on the factor N.

Since the induced structure on a submanifold is canonically defined, our
proof of the uniqueness theorem will be complete if we can prove the following
lemma.
Lemma 2.2. Let No and Nx be submanifolds of the Poisson manifold P having
complementary dimension to a symplectic leaf S. Suppose that each Nt intersects S
at a single point transversely. Then there is an automorphism of P which maps a
neighborhood of No Π S in No onto a neighborhood of Nx Π S in Nλ. (This
automorphism induces an isomorphism of the induced Poisson structures on the
neighborhoods.)
Proof. By integrating hamiltonian vector fields, we can map N0Γ) S to
Nx Π 5, so we may assume that these points of intersection are actually the
same. Next we interpolate between No and Nx by a family of manifolds which
can be defined by equations of the form qj - Qj(yv - ,ys,t) and p}-
Pj(y\>%"Λ» 0» where t E [0,1], and Pj and Qj are smooth functions.
We will find a family % of Poisson automorphisms such that % is the
identity, and % maps a neighborhood of No Π S in No onto Nr The % are
obtained by integrating a time-dependent hamiltonian vector field ξH. In order
532 ALAN WEINSTEIN

that the integral curves of ξHf "track" the Nt% we must have the equations

satisfied along Nr Substituting the coefficients of ξHt for qJ9 pj and yi9 we find

dPjP /£ dyt
,./£, ty°"
ty°" dt '

dqj .f=χdyi 3j,ϋ" dt •

In other words, the effect on H of vector fields of the form
8 ^ v 3 3 , «. 3

is prescribed along Nr On Nt Π 5, the functions vu and hence c^ and dβ are

zero, so the vector fields are transverse to Nn and the required function Ht can
be found. (For instance, one could begin by setting Ht — 0 along Nr) This
completes the proof of the lemma and, hence, of the splitting theorem, q.e.d.
An easy corollary of the existence part of the splitting theorem is the
following basic result of Lie [29] mentioned in the Introduction.
Corollary 2.3. Suppose that the rank of the Poisson manifold is constant near
x0. Then there are coordinates (qλ, 9 qk9 pλ9- -,pk, yλ9- ,ys) near x0 satisfy-
ing the canonical bracket relations {#,, qj) - {/?,, pj) - {qi9 yj) = {/?,, yj) =
{yi9 yj) = 0, {qi9 pj) = δ / y .
Proof. In the coordinates given by the splitting theorem, the rank of
v
ij(y) — {λ > yj) must be constant. Since it is zero at xθ9 it must be identically
zero, q.e.d.
As a consequence of the proof of the uniqueness part of the splitting
theorem, there is a well-defined notion of " transverse Poisson structure" along
any symplectic leaf: the induced Poisson structures on all cross sections to the
symplectic leaf are locally isomorphic, but there is no natural representative for
this transverse structure.
Another consequence of the splitting theorem is that the symplectic leaves in
P near x0 are locally products of S with symplectic leaves in the transverse
Poisson manifold. A global conclusion may be drawn if the transverse Poisson
manifold N is stable near N Π S in the sense that JV has arbitrarily small
neighborhoods of N Π S which are invariant under all hamiltonian vector
 THE LOCAL STRUCTURE OF POISSON MANIFOLDS 533

fields. A proof of the following result can be modeled on the proof of the Reeb
[38] stability theorem for foliations.
Corollary 2.4. Let S be a closed simply connected symplectic leaf in P such
that the transverse Poisson structure is stable. Then S has arbitrarily small
neighborhoods in P which are invariant under all hamiltonian vector fields. Each
symplectic leaf of P near S is a bundle over S whose fibre is a symplectic leaf in
the transverse Poisson structure.
If S is not simply connected, then we need a stronger stability condition: the
transverse Poisson structure should have small neighborhoods which are in-
variant under all automorphisms. In this case, the symplectic leaves near S are
again bundles over S with the fibre being a disjoint union of symplectic leaves
for the transverse structure. In general (for instance, if the transverse Poisson
structure is zero), there is nothing to prevent symplectic leaves from wandering
apart, so that the splitting theorem gives no global information.
We note that the space of Casimir functions on a neighborhood of a point is
isomorphic to the space of Casimirs for the transverse Poisson structure. The
same holds in the neighborhood of a simply connected symplectic leaf when
the transverse Poisson structure is stable.
We can think of a neighborhood of S in P as being a bundle of Poisson
manifolds with fibre N over the symplectic manifold S. It is not clear whether
every such bundle arises from a Poisson manifold in which S is embedded as a
symplectic leaf; this is related to the question of which bundles of symplectic
manifolds over a symplectic base admit a "compatible" symplectic structure on
the total space. (See Weinstein [48] and Gotay, Sniatycki and Weinstein [12].)

3. Linear Poisson structures

A Poisson structure on a vector space Fis called linear if the Poisson bracket
of any two linear functions on V is again linear. This makes the dual space V*
into a Lie algebra which we shall denote by g. V is then the dual g* of a Lie
algebra, and the Poisson bracket on Fis given by

f
where [, ] is the Lie algebra operation in g, ( , ) is the pairing of g* with g, and
8F/8μ is the differential of F considered as an element of g instead of g**.
The formula (1) defines a linear Poisson structure on the dual space g* of
any Lie algebra g. We shall call this the Lie-Poisson structure, since it was
discovered by Lie [29] who wrote down the structure in local coordinate form.
534 ALAN WEINSTEIN

If Xl9'-,Xr is a basis of g satisfying [Xi9 Xj] - Σk=ιCijkXk, then the A)

considered as coordinate functions xλ9- -,xr on g* have the Poisson bracket
relations {*,, jCy} = Σk==ιciJkxk, i.e., the components of the Poisson structure
r
are ^ij{x) — ^ k=lcijkxk. These are linear functions of x, hence the term
"linear Poisson structure."
Let G be the simply connected Lie group whose Lie algebra is g. The Lie
algebra homomorphism Xt H> £ X from g to the hamiltonian vector fields on g*
induces a homomorphism from G to the group of linear Poisson automor-
phisms of g*. This homomorphism is called the coadjoint representation, since
it is dual to the adjoint representation of G on g. (See Lie [29].) The coadjoint
representation is trivial on the center of G, so in fact it is defined for any Lie
group whose Lie algebra is G.
The hamiltonian vector field of any function F on g* is the linear combina-
tion iF = Σri=λ(dF/dXi)ξx., so the orbits of all hamiltonian vector fields lie on
orbits of the coadjoint representation. These coadjoint orbits are precisely the
symplectic leaves in g*. The Casimir functions are just the functions on g*
which are invariant under the coadjoint representation (the name Casimir
functions comes from this example).
For μ E g*, we denote its coadjoint isotropy subgroup by Gμ, the Lie
algebra of Gμ by g μ , and the coadjoint orbit by G μ » G/Gμ.
Theorem 3.1. The transverse Poisson structure at μ to the symplectic leaf
G μ is isomorphic to the Lie-Poisson structure on g*.
Proof. For any μ G g*, the map Bμ: g -> g* is defined by (Bμx, y) —
- ( μ , [x9 y]), which also equals (ad*μ, y >. Then Ker Bμ = {x \ ad*μ = 0} = g μ ,
and so the tangent space to the coadjoint orbit is Im Bμ = (KerBμ)± = g μ . Let
V C g* be a complement to g^, so that V1- is a complement to Qμ in g. Thus
the transverse Poisson structure at μ can be defined on the affine subspace
V + μ, which can be identified with g* by the isomorphism α : λ + μ H> λ |fl .
The cotangent spaces to V + μ are then identified with g μ .
Now the transverse Poisson structure on V + μ at the point λ -f- μ is given
by the bilinear form on g μ

(since ad*μ = 0), which equals ( α ( λ + μ),[;c, >>]) and so is just the natural
Poisson structure on g*.
Corollary 3.2. The coadjoint orbits near a simply connected orbit G μ are
bundles over G μ whose fibres are coadjoint orbits in g*.
We can also derive a well-known theorem of Duflo and Vergne [8]. In fact,
our Theorem 3.1 can be considered as a generalization of their result.
 THE LOCAL STRUCTURE OF POISSON MANIFOLDS 535

Corollary 3.3. If μ G Q* is a regular point in g* (i.e., a point where the rank

of the Poisson structure is locally constant), then g μ is abelian.
Proof. The transverse Poisson structure is always trivial at a regular
symplectic leaf.

4. The linear approximation

In view of the splitting theorem, the local study of Poisson manifolds can be
reduced to the case where the rank at a point is zero. We shall see that the
Poisson structure near any such point can be approximated by a linear one.
If x is any point in a manifold P, the cotangent space T*P may be identified
with the quotient xnx/xn2x, where mx is the ideal in C°°(P) consisting of all
functions which vanish at x together with all derivatives of order up to k — 1.
If P is a Poisson manifold, then m^is a Lie subalgebra of C°°(P). Further-
more, if the rank is zero at x, then mx is a Lie ideal in mx. In fact, since ml is
generated by products GH of elements in m^, it suffices to prove that (i% GH}
is a sum of such products when F9 G, and H are in xnx. But {F, GH} =
{F, G}H + {F, H}G is such a sum because all Poisson brackets are zero at x.
Now T*P = mx/m\ has the structure of a Lie algebra, which we shall
denote by QX, and so TXP = g* carries a Lie-Poisson structure. We call this the
linear approximation to the structure at x. If *,,• ,JCΓ are coordinates on P
which all vanish at x, then the functions wtJ = {xi9 Xj) vanish at 0, so we can
write wtJ(x) = Σk=ιciJkxk + 0(Λ; ) where cijk = dw^/dx^O). Then the cijk are
2

the structure constants of QX, and the coefficients of the linear approximation
are just the truncations Σk=\ cijk xk of wtj.
The linear approximation can also be described in terms of the bundle map
B: T*P -> TP. If Bx: T*P -> TXP is zero, there is a well-defined intrinsic
derivative (Golubitsky and Guillemin [11]) DXB which is a linear map from
TXP to Hom(T^P, TXP). Identifying the T%P and TXP inside the "Horn" with
tangent and cotangent spaces to TXP, we may consider DXB as a bundle map
from T*(TXP) to T(TXP), which is just the linear approximation Poisson
structure on TXP.
To approximate a Poisson structure at a point x where the rank is not zero,
we first apply the splitting theorem and then take the linear approximation to
the (transverse) Poisson structure on the factor N. This Poisson structure lives
in a natural way on the normal space Λ^. = TXP/Im Bx to the symplectic leaf
through x. One way to see this is by using the intrinsic derivative again. If Bx is
not necessarily zero, DXB maps TXP to Hom(KerBχ9 TxP/lm Bx). Since B is
skew symmetric, Ker Bx = (Im B^π (TxP/lm Bx)* = #*, so DXB maps TXP
536 ALAN WEINSTEIN

to Hom(N*9 Nx). Now one can see easily, for instance from the local coordi-
nate formulation of the splitting, that DXB annihilates the tangent space Im Bx
to the symplectic leaf through x9 so DXB induces a map from Nx to
Hom(N*, Nx); this map is just the linear approximation to the transverse
Poisson structure. The dual space N*9 called the conormal space, has the
structure of a Lie algebra which we may denote by QX as before. As we have
seen, if P = g* with the natural Poisson structure, then $x really is the isotropy
algebra of x, so the notations are consistent.
Notice that the linear approximation to the transverse Poisson structure lives
on a well-defined quotient space, while the transverse Poisson structure is just
an equivalence class of isomorphic structures with no natural representative. If
S is any symplectic leaf, the linear approximations to the transverse Poisson
structures make the normal bundle NS — U ^ C N V into a bundle of linear
Poisson manifolds, and the conormal bundle N*S becomes a bundle of Lie
algebras.

5. Hamiltonian systems
If H is a function on a Poisson manifold, the flow exp tξH of the hamiltonian
vector field ξH leaves each symplectic leaf S invariant. Since exptξH is a
Poisson automorphism, it lifts to a bundle mapping φ' from NS to itself such
that each mapping Nx -* iV(exp tξH)M is a linear Poisson mapping. If we are given
some identification of the Poisson manifolds N^xpt^H)M with a fixed linear
Poisson manifold N9 then the maps φx determine a time-dependent linear
hamiltonian system on N which is a linear approximation to the original
hamiltonian system along the trajectory (exp tξH)(x).
Suppose now that there is a symmetry for i/, i.e., a function K such that
{K, H) = 0, and suppose that the "dynamical" trajectory (exptξH)(x) coin-
cides with the symmetry orbit (exptξκ)(x). Then the 1-parameter group
(exp tζκ)~\exp tξH) = exp/ξ ( ^_ A:) leaves the point x fixed, so its linear ap-
proximation at x gives a 1-parameter group of linear Poisson mappings on the
transverse space Nx = g*. The hamiltonian is obtained from the differential of
H — K at x, which annihilates TXS and so passes to a linear function on the
quotient Nx. This linear hamiltonian may also be considered as an element of
the Lie algebra g, and its flow is a 1-parameter subgroup of the corresponding
Lie group, acting on g* by the coadjoint representation.
Applications of this linear approximation procedure to some mechanical
systems will be given at the end of the paper.
 THE LOCAL STRUCTURE OF POISSON MANIFOLDS 537

6. The linearization problem

Motivated by the linearization results for other geometric structures (vector
fields, differential forms, group actions, etc.), we may ask whether a given
Poisson structure near a point of rank 0 is isomorphic to its linear approxima-
tion at that point. If so, we shall say that the Poisson structure can be
linearized. In other words, if a Poisson structure is of the form u> =
Σrk=\cijkxk + O(χ2\ does dropping the higher order term 0(x2) change the
qualitative nature of the Poisson structure, or can these terms be removed by a
suitable change of coordinates?5 The linearization problem is the deepest and
most difficult question in our study of Poisson manifolds; in fact, we can give
only partial answers.
First of all, we note that as a consequence of Theorem 3.1, the Poisson
structure on g* can be linearized at every point, and the set of points at which
any Poisson structure can be linearized is open. It is also dense, since the
structure at any regular point can obviously be linearized.
As with other geometric structures, the answer to the linearization question
depends on the nature of the linear approximation, as well as the category
(formal, analytic, C 00 ) in which the linearization is to be carried out.6 We shall
call a Lie algebra g (formally, analytically, C°°) nondegenerate if any Poisson
manifold whose linear approximation at a point x is isomorphic to g* is itself
isomorphic to g* by a (formal, analytic, C°°) local equivalence.
It is easy to see that not every Lie algebra is nondegenerate. The Poisson
structure on R3 defined by

{χ]9χ2} =\χ\2χ3>
{x 2 ,x 3 } =\x\2xu
{x^x^ =\χ\2χ2
is not trivial, but its linear approximation at zero is. Similarly, one can show
that no abelian Lie algebra is nondegenerate in any of our three senses.

5
Note that we consistently use the term linearization to refer to the removal of higher order
terms by change of coordinates, while their removal by simple truncation is called linear
approximation. At a point where the rank is not zero, the linearizability will refer to the transverse
Poisson structure.
6
For instance, a function Σrij=\auxixj + O(JC3) is necessarily equivalent to its quadratic part
only if the symmetric matrix atj is nonsingular (Morse lemma). A differential form ω =
Σ" j=\aijxi dxj + 0 ( x 2 ) satisfying ω Λ dω = 0 (Pfaffian form) is necessarily of the form gdf in
the formal or analytic category only if the symmetric matrix ai} is nonsingular. In the C 0 0 cateogry,
one must add the extra condition that the index of (α / 7 ) not be equal to 1 or r - 1 (Reeb [38],
Moussu [37]).
538 ALAN WEINSTEIN

On the other hand, we can show

Theorem 6.1. Any semisimple Lie algebra g is formally nondegenerate.
Proof. If we are given formal power series wiJ(xι,' ,x r ) with wιy(0) = 0,
satisfying the antisymmetry and Jacobi identities, then they define a Lie
algebra structure on the ring R[[JC 19 - ,xτ]] of formal power series. The powers
k 2
m of the maximal ideal are Lie subalgebras, and m is a Lie ideal in m, so
2
m / m is an r-dimensional Lie algebra g, and the Poisson manifold g* is the
linear approximation to our original formal Poisson structure.
Let Ak be the Lie algebra m/m*. Then we have a commutative diagram

and the linearization problem is to lift the isomorphism a λ through the tower
of Aks and so up to m. In other words, we have to find formal power series
φ1? ,φΓ in m such that {φ,, φ j = Σrk= ι cijk φk, where the ciJks are determined
by M * , ) , π(xj)} = Σ cijk π(xk) in g. '
The kernels of the homomorphisms πk_ι are easily seen to be abelian. Since
g is semisimple, all cohomology groups i J 2 ( g , V) are zero, and by standard
results in the extension theory of Lie algebras (Jacobson [22]), the homomor-
phism α" 1 can be lifted to each Ak and so up to the inverse limit m. q.e.d.
We may now use BoreΓs theorem on the existence of C 0 0 functions with
given Taylor expansions to conclude
Corollary 6.2. // x is a point of rank 0 in a Poisson manifold such that the
linear approximation is the dual of a semisimple Lie algebra, then there are
coordinates (JC,, ,xr) about x and constants cijk such that {JC,, Xj) = Σ cijkxk
+ 0(x°°).
The problem of C 0 0 nondegeneracy is also a lifting problem, from formal
power series to C 0 0 functions, but now the kernel of the homomorphism (flat
functions) is infinite dimensional and nonabelian so the cohomology theory
does not apply. For analytic nondegeneracy, the problem is to show that the
formal power series constructed in Theorem 6.1 are convergent; none of the
available theory (e.g., Malgrange [30]) on analytic solutions of partial differen-
tial equations seems to apply. At this point we then must be content with some
examples and conjectures.
A useful guide, especially in three dimensions, is the theory of Pfaffian
forms, already mentioned in a footnote earlier in this section. Also suggestive is
 THE LOCAL STRUCTURE OF POISSON MANIFOLDS 539

the theory of linearization of group actions (Hermann [18], Guillemin and

Sternberg [14]). Study of some examples in these theories leads, in particular,
to the following result.
Proposition 6 3 . The semisimple algebra £ 1(2, R) is not C°° nondegenerate.
Proof. The Poisson structure for SX(2,R) is given by {xv x2} = -*2>
{x2, x3] = xl9 {x3f xx} — x2. The function x\ + x\ — x\ commutes with ev-
erything, and the components of its level surfaces (cone, hyperboloids of one
and two sheets) are symplectic leaves. Now we will perturb the Poisson
structure to unwrap the hyperboloids of one sheet into surfaces which spiral
toward the cone. To do this, we choose a C°° function l(t) with /(/) = 0 for
t < 0 and l(t) > 0 for t > 0, and let
1 ί Λ 2 i / 9 ι 9 9 \
{
~ Λ 3 ' l A 2 » λr3)
Λ Λ Λ Λ
\' 2J l 2 i 2 v 1 2 3/'
JCi

The hamiltonian vector field ξX3 is

θ 9
x2 + r 2
Λ] i X2

whose integral curves outside the cone x2 H- jcf — jcf = 0 spiral in slowly
toward the cone. This behavior distinguishes the given Poisson manifold from
its linear approximation § 1(2, R)*.
The difficulties with £l(2,R)* seem to involve the nonsimple connectivity of
the symplectic leaves. By contrast, it seems likely that one can prove the C 0 0
and analytic nondegeneracy of £o(3). (By analytic continuation as in Guille-
min and Sternberg [14], one could then derive the analytic nondegeneracy of
% 1(2, R).) The first step would be to use the theorems of Reeb [38] and Moussu
[37] to "linearize" the foliation by symplectic leaves. Next, a "volume-preserv-
ing Morse lemma" (Vey [45], Guillemin [13]) would be used to put in standard
form the function which measures the symplectic area of the leaves. Finally,
the deformation method of Moser [36] and Weinstein [47] would have to be
applied to each symplectic leaf, with care taken to assure regularity at the
origin.
We conclude this section with
Conjecture 6.4. (a) Q is formally or analytically nondegenerate if and only
if Q is semisimple.
(b) If Q is semisimple, then g is C 0 0 nondegenerate if and only if the index of
the Killing form of g is not 1 or (dim g — 1); in particular, if g is of compact
type, then it is C 0 0 nondegenerate.
540 ALAN WEINSTEIN

These conjectures are not asserted with great conviction but are meant more
as questions. In particular, the condition in the noncompact case of (b) may
need to be modified.

7
7. Function groups, realizations, and momentum mappings
A function group as defined by Lie [29] is a collection ^of functions of the
canonical variables (qv ,qn, px, ,pn) such that
(i) ^is a Lie algebra under Poisson bracket,
(ii) if FX9 ,FS G <$ and G: Rs -> R, then G(Fl9- -,FS) G 3F. Lie always
assumed that ^ was generated (in the sense of the functional composition
operation (ii)) by a finite number φj, ,φr of independent functions, so that
(i) implies (φ,, φy} = wfyίφj,' ,φr) for functions wtj on Rr. Given a function
group with such generators, the Jacobi identity for the canonical Poisson
bracket implies that the wtj define a Poisson structure on RΓ for which the
mapping (φ l 5 ,φ r ): R2w -+ Rr is a Poisson mapping. Lie assumed implicitly
that this Poisson structure had constant rank.
The realization problem in local terms may then be stated as follows: given a
Poisson structure on an open subset % of Rr, find a Poisson mapping from the
canonical R2n onto % for some n. We will call such a mapping a realization of
the Poisson structure.8 In the case of constant rank, Lie proved that every
Poisson structure has a realization, and classified the realizations up to
canonical transformations.
In global terms, we may define a function group by specifying an equiva-
lence relation, or foliation, Φ on a symplectic manifold (5, Ω) such that the
quotient space S/Ω is a manifold, and letting ^φ be the space of functions
constant on the leaves (equivalence classes) of Φ. The condition that ^φ be
closed under Poisson bracket can then be expressed geometrically in the
following way. Let TΦ C TS be the subbundle of vectors tangent to the leaves
of Φ, and let ΓΦ X be its orthogonal complement under the symplectic
structure Ω. If F G % and υ G ΓΦ, then Ω(t>, ξF) = v F = 0, so the hamilto-
nian vector field ξF lies in ΓΦ" 1 . By dimension counting one sees now that
ΓΦ"1 is filled at each point by the hamiltonian vector fields of functions in ^φ.
If 9^ is a Lie algebra, then these hamiltonian vector fields are closed under

7
Most of this section is essentially a reworking of Lie [29]. See also Hermann [20].
8
In Marsden and Weinstein [32] and Marsden, Ratiu and Weinstein [33], the coordinates on R2n
are called Clebsch variables for the Poisson manifold %. An infinite-dimensional version is studied
in those papers.
 THE LOCAL STRUCTURE OF POISSON MANIFOLDS 541

bracket, so the bundle TΦ± is involutive and so is tangent to another foliation

which we may call Φ± . On the other hand, if we know TΦ± to be involutive,
then the Poisson bracket of any two functions in % has its hamiltonian vector
field in TΦ± and hence lies in S^. Then we have proven
Proposition 7.1. Let Φ be a foliation on a symplectic manifold S. Then the
9
space S^ of functions along the leaves ofΦ is a function group if and only if the
X
subbundle TΦ symplectically orthogonal to the leaves ofΦ is involutive.
If the hypotheses of Proposition 7.1 are satisfied, the functions constant
along the leaves of Φ± form another function group %± , which Lie called the
polar group of ^φ. Thus function groups always come in pairs. The quotient
spaces S/Φ and 5/Φ" 1 are Poisson manifolds, and in §8 we shall show that
their Poisson structures are very similar to one another.
It is easily seen that the polar function group %± consists precisely of the
functions which Poisson commute with all the elements of ^φ\ this, in fact, is
the condition by which Lie defined polarity. The intersection % Π %± is thus
equal to the center of either of the Lie algebras % and %±. Lie called the
elements of ^φ Π ^φ± the distinguished functions of *$φ (or ^φ±). He also
observed that ^φ Π ^φ± forms a new commutative function group, but here the
assumption of constant rank of the Poisson manifold S/Φ must be invoked to
assure that *% Π ^φ± is associated with a smooth foliation. As functions on the
Poisson manifold S/Φ, the distinguished functions are just the center of the
Poisson algebra C°°(S/Φ\ i.e., the Casimir functions.
A realization of the Poisson manifold P is defined to be a Poisson mapping /
from a symplectic manifold S to P; the realization will be called full if / is a
submersion. By Lemma 1.2, hamiltonian flows on P can be put in canonical
form with the aid of realizations; this is the essential idea behind the use of
Clebsch variables in continuum mechanics.
A simple example of a realization is the inclusion map of any symplectic
leaf, but this realization is generally not full. We shall show in §9 that full
realizations always exist locally, and that they are essentially unique.
A realization S ->g*, where g* has its Lie-Poisson structure, corresponds to
a Lie algebra homomorphism from g to C°°(S) and hence to a hamiltonian
action on S of the corresponding Lie group G. In this context, the map / is
called the momentum mapping of the group action. Since the linear functions
on g* generate the coadjoint action of G, it follows from Lemma 1.2 that the
momentum mapping is G-equivariant with respect to the coadjoint action, a

9
One should speak of a sheaf Ίi the quotient space S/Φ is not a manifold.
542 ALAN WEINSTEIN

fact known already to Lie [29, Chapter 19] and rediscovered in the mid 1960's
by Kostant [26] and Souriau [41].

8. Dual pairs and gauge groups

J\ h
If S is a symplectic manifold, we call a pair of Poisson mappings Px <-S -> P 2
a dual pair™ if the function groups fx = JX*(C°°(PX)) and % = J$(C°°(P2)) are
polar to one another. We call the dual pair full if Jx and J2 are submersions
onto Px and P2. A dual pair may also be defined as one for which the
subbundles Ker TJλ and Ker TJ2 tangent to the level manifolds of Jλ and J2 are
symplectic orthogonal complements of one another.
If the maps /, and J2 have constant rank, then the images JX(PX) and J2(P2)
are Poisson submanifolds, and JX(PX) <-S -*J2(P2) is a full dual pair, so for
many purposes we can restrict our attention to this case.
Given a full dual pair, the spaces of Casimir functions on Px and P2 are both
in 1-1 correspondence with the space of distinguished functions in the polar
function groups 3^ and ^F2, so Px and P2 have the "same" Casimir functions.
We also note that for xx E Px each connected component of J~ι(xx) maps
under J2 to a symplectic leaf in P2, and vice versa. This gives a correspondence
between symplectic leaves in Px and P 2 , which is bijective locally (and globally
if Jx and J2 have connected fibres).
These observations, which have already been made in a purely symplectic
context by Kazhdan, Kostant, and Sternberg [23] and Guillemin and Sternberg
[15], suggest that the Poisson manifolds occurring in a full dual pair are very
closely related. This we shall now make precise and prove.
Theorem 8.1. Let PX<-S-*P2 be a full dual pair. For each x E 5, the
transverse Poisson structures on Px and P2 at Jx(x) andJ2(x) are anti-isomorphic
(i.e., isomorphic up to a reversal of sign of the bracket). Consequently, if
dim Px < dim P2, P2 is locally anti-isomorphic to the product of Px with a
symplectic manifold.
Proof. We begin by reducing the problem to the case where Px and P2 have
rank zero at Jx(x) and J2(x). To accomplish this, let (ql9- -9qk, px, -,pk,
yx,- -,ys) be the local coordinates on Px near Jx(x) obtained from the splitting
theorem (Theorem 2.1), and let (q[, ,q' h p\, ,p\, y[, ,>>/) be such coordi-
nates on P2 near J2(x). Pulling back via /, and J2, we can consider all these
coordinates as functions on S.

10
The term dual pair was introduced in representation theory by Roger Howe in an unpublished
manuscript. We are dealing here with the "classical" analog of Howe's "quantum" notion.
 THE LOCAL STRUCTURE OF POISSON MANIFOLDS 543

N o w (qu >9qk9 q[, 9

q'/9 px, -9pk, p\9- 9
p\) satisfy t h e canonical com-
mutation relations, and so they are functionally independent. By the method
used in the proof of Theorem 2.1, we can find further functions q'{9- -9q'^9
p", - - 9p'ή such that (q9 p9 q'9 pr9 q"9 p") is a canonical coordinate system on S.
Since the y9s and y"s commute with (q9 p9 q\ /?'), they must be functions of
(#"> P") alone. Thus we have split S as a product 5" X S" such that the
transverse parts of the Poisson structures on Px and P2 really live on S". It is
easy to check that these transverse parts form a full dual pair on S"', so our
reduction is complete. Thus we assume at this point that Px and P2 have rank
zero at /,(*) and J2(x).
We continue to use the coordinate functions (yl9- -9ys) and (y[9- •,)>/),.
which we may take to vanish at Jx(x) and J2(x) respectively. The conditions
{#» yj}(x) — 0 and {y{9 yj}(x) = 0 imply that the tangent spaces Ker TXJX
and Ker TXJ2 to the level manifolds of Jλ and J2 through x are coisotropic. By
duality, these tangent spaces are orthogonal complements of one another, so
they must in fact be lagrangian and equal. In particular, / = dim P{ = dim P2
= s9 and dim S = 2s.
Choose a lagrangian submanifold A Q S through x which is transverse to
Ker TXJX = Ker TXJ2. The functions yl9- -,ys restricted to Λ are independent
and may be extended to commuting functions qv -,qs (not the ^'s above)
near Λ in S. Next we can find functions p]9- -9ps vanishing on Λ such that
(#i>*' *>#5> P\>'' ΊPS) f ° r m s a canonical coordinate system.
We may write yt — qt + Lajlpι where the α^'s are functions on S. By
choosing coordinates on P29 we can arrange it so that y[ — qt on Λ as well, so
that >>/ = ?,- + Σ bjfPj. Now

and so {yi9 yj) = aJi — atj on Λ. Thus if we set {yχ9 yj) — wij9 we have
wfj — a}i — atj (identifying Px with Λ via / , ) . Similarly, {y(9 yj} = w{j implies
w/y = bjΊ — btj. Finally the duality gives

0 - {al9βj} = {qt + ^aikPk9 9j + ΣbJkPk),

which on Λ gives aki — bik — 0. Then w/y = ajΊ — atj — bu — bjt = w(J9 so the
identifications of Px and P2 with Λ give an anti-isomorphism between their
Poisson structures, q.e.d.

An important way to realize Poisson manifolds is as quotients of symplectic

manifolds by group actions. Suppose that a Lie group G acts freely on a
symplectic manifold S with momentum map / : S -> g*, and that the quotient
space S/G is a manifold. Then the functions on S/G may be identified with
544 ALAN WEINSTEIN

the function group of G-invariant functions, so S/G has a Poisson structure for
which the projection π: S -» S/G is a Poisson mapping. Since the π-level
J 1ΐ

manifolds are G orbits, it is immediate that g* <-S -+S/G is a dual pair. The
symplectic leaves in S/G are the manifolds π(J~ι(G μ)) = J~ι(G μ)G,
called the reduced manifolds of 5 by the G-action. (They are isomorphic to the
reduced manifolds J~\μ)/Gμ introduced by Marsden and Weinstein [31]; see
Kazhdan, Kostant, and Sternberg [23]).
In physical applications, the symplectic manifold S may represent a collec-
tion of states where elements of S in the same G-orbit are considered to be
physically indistinguishable. Thus the set of "true physical states" is the
Poisson manifold S/G, and the group G is called a gauge group. Often a
Poisson manifold P of physical states is given in advance, and one seeks to
realize P as a quotient S/G.
The point to be made here is that the gauge group G is essentially
determined by the transverse Poisson structures in P. By Theorem 8.1, if y is
any element of P — S/G with y — τr(jc), then the isotropy algebra QJM C g
must be such that the Lie-Poisson structure on Q*M is the normal Poisson
structure for P at y. In particular, only linearizable Poisson manifolds can be
realized as quotients of symplectic manifolds by Lie group actions, and only
regular Poisson manifolds can be realized with abelian gauge groups.
Some examples of dual pairs in mechanics may be found in Marsden and
Weinstein [32] and Marsden, Ratiu and Weinstein [33].

9. Existence of realizations
The most important realization of a Lie-Poisson manifold g* is by the
momentum mapping g* -> T*G associated with the action of the corresponding
Lie group G on T*G by the lifts of left translations. / is just the operation of
right translation to the cotangent space at the identity element of G. The
quotient T*G/G can again be identified with g* but with the negative of the
usual Poisson structure, and the projection T*G -> g* is given by left transla-
J IT

tion. Thus we have the dual pair g* <- T*G -> g*, in which the roles of J and π
are completely symmetric; π is the momentum mapping for the action of G by
the lifts of right translations, and / is the quotient mapping for that action.
By the splitting theorem, we can use the basic example just described to find
a (local) realization of any Poisson structure which is linearizable. Of course,
this assumes the existence of the Lie group G whose Lie algebra is g. But
linearization at a regular element of any Poisson manifold (including an
 THE LOCAL STRUCTURE OF POISSON MANIFOLDS 545

arbitrary g*) produces an abelian Lie algebra, and in this case the existence of
a group is obvious; this idea was used by Lie [29] to prove the existence of a
local Lie group of diffeomorphisms corresponding to any given Lie algebra.
For a given Poisson manifold which may not be linearizable, constructing a
realization is tantamount to finding a substitute for the corresponding Lie
group. This is the idea behind the proof of the following theorem.
Theorem 9.1. // x is any point in a Poisson manifold P, there is a full
realization of a neighborhood of x by a symplectic manifold of dimension

Proof. By the splitting theorem, we can reduce the problem to the case
where the rank of P at x is zero, which we shall now assume. Let xu ,xn be a
coordinate system centered at x9 so that {xi9 Xj} = w^x^- 9xn) with w/y(0)

= 0. We shall take the symplectic manifold S to be R 2 w with coordinates

(JC,, -9xn9 yX9- -9yn) and the obvious projection of S onto P9 but with a
symplectic structure to be chosen in order that the Poisson bracket relations
{JC.9 Xj) = Wij(x) hold in S.
The manifold x = 0 will be lagrangian, so it will be transversal to some
lagrangian foliation; thus there is no less of generality in assuming that the
required symplectic structure be of the form dφ9 where φ = Σ?= x Φ, (x, y) dyt.
With φ = Σ? = i Φ, (*, 7 ) Φ/, we have
n Λ n

Ω —•
' dφ — }. a • dx A dy ~\~ ~τ ^ . b- dy- A dy 9
lJ J
ij=\ i,7=l
where atj — dφJ/dxi and btj = θφ^/θ^ — θφ /θj^ . Then the map Ω from
tangent vectors to cotangent vectors is given by
n n n
' 7\ \ I 7\ \
+
j- = Σfl/7*7» δ
U Γ =- Σ s ^ y Σ ^Φy
1 ' j=\ \ Si I j=\ j=\

The Poisson bracket {xf., Xj} is Ω " ^ ^ ) ίfcc,.. The inverse of the block
matrix
(O -a
\a b
has p = a~ιb(a')~ι in the upper left-hand corner, so {xi9 Xj) = — ρiJm Thus
-a~ιb(a')~ι — w, where wtj was the given value of {xi9 Xj} in the Poisson
structure. Finally, we have the equation b = -awa* which is a system of
\n(n — 1) first-order partial differential equations for the unknown functions
Φi(x9 y). In coordinates, we have
546 ALAN WEINSTEIN

Let us think of the φ/s not as scalar functions of x and y, but rather as
functions of y with values in the Lie algebra C°°(P) of functions of x.11
The form φ - Σφ, έfy, is then a C°°(P)-valued 1-form in y, and (1) is the
component form of the equation
(2) dφ = -±{4>,4>}>
where the bracket is the standard bracket on Lie algebra-valued 1-forms, as in
Bishop and Crittenden [7].
We may recognize (2) as the structure equation for the Maurer-Cartan form
ω on the group § whose Lie algebra is C°°(P\ or alternately as the equation
for the connection form of a flat connection on a principal β-bundle over y
space. Such a connection form can be found in the form γ*ω, where γ is a map
fromj-spacetoβ.
Now we do not have the group § at our disposal, but fortunately there is a
formula for the pullback of ω to the Lie algebra C°°(P) under the exponential
map, which is expressible purely in terms of the Lie algebra and makes sense in
the present context. Namely,12
(3) exp*(ω)(/) = fcxp(-s ad/)Λ.
•'o
Here / is an element of the Lie algebra C°°(P)9 ad / is the map g ι-> {/, g},
and hence exp(-s a d / ) can be interpreted as the time s map Ψs of the
hamiltonian flow ξf on P generated by /. Now if we identify each element of
>>-space with the linear function fy(x) — Σ/=1.x/.yz, we have constructed the
form φ as Σ Φ, dyi9 where
(4) φi(x,y)=fxi°*;ds.

(We write Φ* for the flow of ξf.) In other words, φ, is the average value of xt
over the trajectory from s — 0 to 1 of the hamiltonian flow generated by^,.
To complete the proof, we shall verify directly that the functions φ, defined
by (4) satisfy the differential equations (1) and incidentally that the matrix
atj — dφj/dx; is nonsingular so that dφ really is a symplectic structure. We will
also see explicitly that the Jacobi identity for the Poisson structure must be
used, although (4) makes sense without it.
We recall the " variation of parameters" formula for the derivative of a flow
with respect to the generating vector field. If ξ and η are vector fields on a

1
Arguments involving this infinite-dimensional Lie algebra may be thought of as merely
heuristic, since in the end we will produce an explicit formula which will be verified directly.
12
The right-hand side of (3) is usually written in the symbolic form (1 — e ~ 5 a d y r ) / ( s a d / ) . See
Helgason [16] or Freudenthal and de Vries [9].
 THE LOCAL STRUCTURE OF POISSON MANIFOLDS 547

manifold, then

(5) — exp[s(£ + εη)]

(See Abraham and Robbin [2, 32.1].) Here, as above, we are thinking of the
vector fields and their flows as operators on functions.
Now we differentiate (4) with respect to y . Since (d/dyj)ξf = ζx , we get
from (5),

— = / -r— expljξ, JXfds

dyj Jo dyj ^ h>

( ) ( y . ) x / dt ds

Now the Jacobi identity implies that Φf, preserves Poisson brackets, so

ΊΓ= If
dyj „ JJ

Interchanging i andy, we get

where in the last step we simply interchanged the dummy indices s and T. We
combine the two integrals to get an integral over the unit square

$-$ =- //

which establishes (1).

548 ALAN WEINSTEIN

Finally, we note that, for>> = 0, Ψ* is the identity for all s, so φf (x,0) = xi9
and hence the matrix atj — dφj/dxt is invertible in a neighborhood of
y — 0. q.e.d.
Some remarks on the proof may be of interest. First of all, since Ω =
diΣφidyj) = Σdφi Λ dy.9 the functions (φ,(x, y%- ,Φn(*, >>)> yl9- ,jrt) are
canonical coordinates on S. Thus to solve the problem of realizing the bracket
relations [xi9 Xj} — wtJ(x) in terms of canonical variables, it suffices to invert
the transformation (x, y) h-> (φ(x, y), y) and solve for x in terms of φ andy.
Next we note a special case. If the given Poisson structure is abelian, then all
the transformations Ψf, are the identity, and so we get the obvious realization
with φ, = xr More generally, if the Poisson structure is linear, then the
transformations Ψf, are linear, so the functions φ,(x, y) are linear in x. What
we have obtained then is the usual local expression for the symplectic structure
on the cotangent bundle of a Lie group in terms of the trivialization by
left-invariant 1-forms, them's being coordinates on the group. For a general
Poisson structure, them's are coordinates on a submanifold (but not subgroup)
of the "group" whose Lie algebra is C°°(P).
We saw in the beginning of this section section that every Lie-Poisson
manifold g* has a full realization T*G -> g*, and it is easy to see that this is the
"minimal" possible realization of g* since the rank at the origin is zero.
However, if we are working near a nonzero point ft G g*, a full realization is
possible, according to Theorem 9.1, by a symplectic manifold of dimension
2(dim g- ^(dim g — dim g^)) = dim g + dim gμ. An explicit construction of
this realization may sometimes be made as follows. We need to find a free
action of G on a smaller manifold than Γ*G, so it is natural to look for a
symplectic submanifold invariant under left translations. Let π: Γ*G -• g* be
the left translation map, and let W C g* be a piece of submanifold through μ
whose tangent space is complementary to the tangent space to the coadjoint
orbit (i.e., g^). Now π~\W) is a left-invariant submanifold of Γ*G, and one
can check the following:
(i) π~\W) is symplectic near π~\μ).
ι
(ii) The momentum mapping for the action of G on π~ (W) by left
translations is just /1 W, so the image contains μ.
Since the G action is free, the momentum map /1 W is a submersion, and
hence we have a full realization of a neighborhood of μ.
An application of this construction in fluid dynamics is given in Marsden
and Weinstein [32].
 THE LOCAL STRUCTURE OF POISSON MANIFOLDS 549

10. Uniqueness of realizations

The problem of uniqueness of realizations was already posed and solved in a
special case by Lie [29]. Given r-tuples of functions (al9 ,ar) and (βl9- ,βr)
2w
of the canonical coordinates (ql9- -,qn, pλ, ••,/?„) on R , when is there a
n
canonical transformation φ on R with aioφ = β.'> If (au ,ar) and
(/?,,- ,/?r) each generate a function group, then an obvious necessary condi-
tion for the existence of φ is that the "structure functions" wtj and vtj be equal,
where {«„ aj) = w^α,,- , α r ) and ( β , β,) = ©,-,•(/?!, •,#.). Under the as-
sumption that the matrices wtJ and vtj have constant rank, Lie proved that this
necessary condition is also sufficient. He also extended the result to the case
where ( α 1 } , α r ) and (βl9- -,βr) are defined on spaces of different dimen-
sions. The following theorem is an extension of Lie's result to the case of
variable rank. The proof is related to an argument in Roels and Weinstein [39].

Theorem 10.1. Let Sλ^>P and Sλ-+P be full realizations of the Poisson
manifold P. Let xt E St be such that Jx(xλ) = «/2(x2) Then there exist:
(i) Neighborhoods Gllx andell2 ofsx ands2.
(ii) Splittings 9 l , » S, X % and Gll2 = ξ>2X % such that Jt\Uj depends only
on the %\ components, via a Poisson map, Sy -> P.
(iii) A canonical transformation φ : S j -> S 2 such that J\—J2° Φ
In other words, any two full realizations of a Poisson manifold are equivalent up
to canonical transformations and the addition of inessential canonical variables.
Proof. We begin by splitting off symplectic factors. For the realization
5, ->P, we find functions (#„• •,?*,/>„• -,pk, q[9- ',qί,p\>- ,/?/,>Ί, -,ys,
y[,- -,y't) on a neighborhood 6llλ of xλ as in the proof of Theorem 8.1. The
functions (q, p, q\ p') map elll onto an open set Ύ, in R 2 <* +/ ). If we take S, to
be the quotient of eHl by the hamiltonian flows of q, p, q' and p\ there is a
natural isomorphism of 6ίl] with S, X Ύ p and the functions y]9 ,ys and
y{,- - -,y's all live on § , , where they generate polar function groups. After doing
the same for S2^>P, we have reduced the problem to the case where P has
rank zero at /-(x,-), and dim S, = dim S2 = 2 dim P, i.e., we are dealing with
realizations of minimal dimension.
Now we may choose for each i a lagrangian section σz of Ji though xi9 i.e., a
lagrangian embedding P -* St such that ^ © σz = identity and σ,.(^.(JC,-)) = JC# .
Next let ( j , , ,j 5 ) be coordinates on P, and ( z l 9 -,zs) the dual coordi-
nates on the dual space P*. For each z, a function ^ : S -> R may be defined
by ^ f.( j ) = (z, j ) © j.-9 by Lemma 2.1 its hamiltonian flow Ψ^t covers the
hamiltonian flow χ'z of (z, y) on P. We map P X P * (locally) to S, by
550 ALAN WEINSTEIN

1
λ,: (y, z) -* Ψ^^σ^y)), and set φ = λ 2 ° λ"; : Sx -* S2. It follows from our
construction that J2 © φ = Jχ.
To show that φ is canonical, we will show that λ^Qx = λ^Ω 2 , where Ω, is the
symplectic structure on Sr First we calculate the pulled back forms along the
"zero section" z — 0. There we have

Thus

since σ, is lagrangian. Similarly,

4 ° 0/ = - {yj, yk),

We note that all these results are independent of i, so λf Ω, = λ^Ω 2 along the
zero section. But the map λi at all points of P X P * is obtained by starting
along the zero section and following trajectories of hamiltonian vector fields
with respect to Ωf , so λfΩ, = λ^Ω 2 everywhere.

11. The restricted three-body problem and other examples

In this final section we shall give several illustrations of how Poisson
manifolds provide a context for describing naturally arising situations in which
symplectic manifolds of different topological types fit together.
A. Contractions of Lie groups. For any value of the parameter ε, the
relations
[X39 Xλ] = X29 [X2, X3] = Xl9 [Xλ9 X2] = εX3

describe a 3-dimensional Lie algebra. The algebra is isomorphic to §o(3) for

ε > 0 and to §>ϊ(2, R) for ε < 0, and is the algebra e(2) of the Euclidean motion
 THE LOCAL STRUCTURE OF POISSON MANIFOLDS 551

group of the plane for ε = 0. The corresponding bracket relations

(1) {x3, *,} = x29 {x2, x3} = xl9 {χl9 x2] = εx3
describe a 1-parameter family of Lie-Poisson structures on R3.
There is another interpretation of (1). If we consider ε as a fourth variable in
addition to (xx, x2, x3\ and we add the relations
(2) {ε,x,} = {ε,x 2 } = {ε, JC3} = 0 ,
4
then (1) and (2) together describe a Poisson structure on R , which is not of
Lie-Poisson type but includes as Poisson submanifolds all the members of the
family described by (1). In particular, the symplectic leaves of (1) for various
values of ε (spheres for ε > 0, cylinders for ε = 0, and hyperboloids for ε > 0)
are all included as the symplectic leaves for (1) and (2). The two Casimir
functions whose level manifolds are those symplectic leaves are ε and x\ Λ- x\
+ εx\.
It is interesting to look at the linearized structures at points where the rank is
zero. (Points of rank two are all regular.) Denoting by {yλ9 y2, y3, 8) the
coordinates on the tangent space, we may consider two cases. In each case, δ is
a Casimir function.
(i) JCJ = x2 = x3 — 0. We have

Since ε is again a parameter here, the Lie algebra is the product of R with
3o(3),3l(2,R),ore(2).
(ii) xx = x2 = ε = 0, x3 φ 0. We have
{JWI}
=
yi> { Λ » Λ } = J Ί > {y\>yi) = * 3 δ
This time the algebra is the four-dimensional "oscillator algebra" with the
parameter x3 playing the role of Planck's constant.
Similar constructions and analysis can be carried out whenever one has a
family of Lie algebra structures depending upon one or more parameters, such
as the degeneration or contraction of the Poincare group to the Galilean group
as the parameter (speed of light)"1 approaches zero.
B. The light top. The motion of a rigid body, anchored at one point, in a
gravitational field is given by a hamiltonian system on the Lie-Poisson mani-
fold e(3)*, where E(3) is the Euclidean motion group in R3. The bracket
relations on e(3)* are
{M 2 , M3) = MX9 {M 3 , Mx) = M29 {Mx, M2) = M 3 ,
{M 2 , F3) = Fλ9 {M 3 , Fx) = F29 {Ml9 F2) = F3,
{F 2 , M3) = FX9 {F3, Mx) = F29 {Fx, M2) = F 3 ,

{Fi9FJ} = {Fi9Mi}=0.
552 ALAN WEINSTEIN

The vector M = (M,, M 2 , M3) is the angular momentum in body coordinates;

F = (Fl9 F2, F3) is the gravitational acceleration vector times the mass of the
body, again in body coordinates.
The hamiltonian for the Euler-Poisson equations of rigid body motion is
H — jΣ M}/li + Σ XiFi9 where /,, I2,I3 are the principal moments of inertia,
and (x,, x 2 , x3) is the vector to the body's center of mass from the anchor
point. (See Guillemin and Sternberg [15] or Holmes and Marsden [21].)
The Casimir functions on e(3)* are generated by ||F|| 2 = Fx + F2 + F3 and
F M = FXMX + F2M2 + F3M3. The regular symplectic leaves (coadjoint
orbits), given by specifying a nonzero value for ||F|| 2 and any value for F M,
are diffeomorphic to T*S2 but with varying symplectic structures. The singular
symplectic leaves he in the 3-dimensional Poisson submanifold F = 0 and are
just the SO(3) coadjoint orbits ||M|| 2 = constant. Physically, these leaves
represent the situation where the force is turned off (or the gravitational mass
of the body is zero.) As explained in Arnold [3], this limit is equivalent to the
limit of infinitely fast rotation.
If we linearize at the leaf F = 0, ||M|| = M, we get a transverse Poisson
structure equal to the Lie-Poisson structure for the coadjoint isotropy group of
(0, M o ). This isotropy group turns out to consist of rotations about the M-axis
and all translations, so it is isomorphic to E(2) X R.
For simplicity we shall consider the case where (0, M o ) is a stationary point,
i.e., a critical point of H restricted to the symplectic leaf. Without loss of
generality we may assume that M o = (M, 0,0).
As coordinates on the transverse Lie-Poisson manifold we have
{Nx, Gx, G2, G3} with relations
{Nι,G2}=G3,{Ni,G3}=-G2,

and all other brackets zero. The linearized hamiltonian is H' = MNι/Iι +
x2G2 4- x3G3 yielding the equations of motion

Gx =09G2 = -MG3/IX, G3 = MG2/IX,NX = x2G3 - x3G2.

Each motion is confined to an E(2) X R coadjoint orbit, which is a cylinder

Gx — constant, G2 + G3 — constant, and it lies in a plane W — constant, so
the trajectories are circles, around which the phase point moves with the
precession-nutation frequency ω = M/Ix. Whether the center of mass is on the
rotation axis or not (xf + *f = or φ 0) determines whether Nx oscillates with
(G 2 , G3), but the motion is qualitatively the same in either case.
C. The restricted three-body problem. The phase space for the gravitational
6N
rt-body problem is the standard symplectic space R with coordinates
 THE LOCAL STRUCTURE OF POISSON MANIFOLDS 553

(q 1 ? ,qN, p l 5 ,P;v), each q, and pz being a vector in R3. The hamiltonian is

The restricted three-body problem is the "limit" in the case N = 3, as the

mass m3 goes to zero. The first two ("primary") bodies move as in a two-body
problem, while the third body moves in the time-varying gravitational field of
the first two. If the two primary bodies are in circular orbits, the motion of the
third body is a time-independent hamiltonian system when viewed in a
coordinate system which rotates with the primaries; the hamiltonian for this
motion is called the Jacobi integral (see Abraham and Marsden [1]).
Our aim here is to show how the symplectic structure for the restricted
three-body problem is related to that of the unrestricted problem. Curiously it
is necessary for us to do this via an infinite-dimensional system called the
Poisson-Vlasov equation. 13 This happens because no finite-dimensional Lie
group of canonical transformations is triply transitive on R6.
The basic phase space for the Poisson-Vlasov equation (Gibbons [10],
Marsden and Weinstein [32], Morrison [35]) is the Lie-Poisson algebra ^ ' ( R 6 )
of densities f(q9 v)dqdv on R6, dual to the Lie algebra C°°(R6) of functions
under the canonical Poisson bracket. 14 Denoting the bracket on ^ ' ( R 6 ) by
{{ , }} to distinguish it from the canonical bracket { , }, we have

where δF/δfis the "functional derivative" defined by

lim | [ F ( / + eg) - F(f)] = f^g{x, υ)^(x, υ) dxdv.

The corresponding Lie group is the group CanίR 6 ) of canonical transforma-

6 6
tions on R , and the symplectic leaves in ^ ' ( R ) are the classes of densities
15
which are equivalent under canonical transformations.
The typical symplectic leaf is infinite dimensional, but there are also
6Λf-dimensional leaves S(nϊ]...mN) consisting of delta densities of the form

13
The equation is also called the collisionless Boltzmann equation, or the equation of stellar
dynamics. For the electrostatic (repulsive) problem, the sign of the potential in H should be
reversed.
14
Our "fundamental" particles will be taken to have unit mass. "Bodies" will be (not necessarily
integer) multiplies of these.
15
Actually a group whose Lie algebra is C°°(R6) must be a one-dimensional central extension of
Can(R 6 ), but the coadjoint representation factors through Can(R 6 ) anyway.
554 ALAN WEINSTEIN

Σ ^ w ^ q - q,, v - vz), where (q,, Vj),- -,(q^, v^) are N distinct points in R6
and mx,— 9mN are given positive real numbers. Then CanR 6 acts on the
product R6N with symplectic structure ^=λmidί\i Λ d\t; the momentum map-
ping of this action, restricted to the set S(m^... mN) of non-repeating iV-tuples in
RβN is an iVΊ-fold covering map onto S(m^...mNy Writing p, for m^v,, we find
that S(nii ... m ) may be identified with the phase space of the iV-body problem.
The hamiltonian for the Poisson-Vlasov equation is the functional

= qι
(1) f#χJ ~ ^2IΓ Y(<11, v, )/(q 2 , v2 )rfq2 ί/v2.

The hamiltonian flow for H is given by

where

Vf(q) = -fj\qx -
(2) is a Liouville equation for the evolution of the density / in phase space
under the flow of the/-dependent hamiltonian i||v|| 2 + Vf. In particular, the
delta densities on the orbits Smχ...mN evolve under the flow of the N-body
problem with masses ml9' -,mN. Thus we have embedded all the JV-body
problems in a single equation-the Poisson-Vlasov equation.
Although the flow (2) restricts without difficulty to the orbits S Wlt ... f#Wjv ,
things are more complicated when we try to restrict the hamiltonian (1).
Formally, we obtain

( Σ^ίq-q.v-vjhy Σ^iwi' +j
N N
N \ λ 1

Σ ^^K-qyll"1.
i=l / i=\ i,y=l
The problem with (3) lies in the infinite "self-energy" terms with / =j in the
second sum. If we "renormalize" the hamiltonian by simply deleting these
terms, the result is the usual hamiltonian for the iV-body problem. It is not
clear how best to justify this renormalization, but the essential idea should be
6
that a hamiltonian flow on ^ ' ( R ) is unaltered if we subtract a Casimir
function from the hamiltonian. The Casimir function (a typical one is
fR6[f(x>v)V dxdυ) should be selected so as to approximate the large self-en-
ergy terms as the delta densities are approximated by smooth bump functions.
 THE LOCAL STRUCTURE OF POISSON MANIFOLDS 555

Putting aside the renormalization question, we are ready to take some

limits. If we let the last mass mN approach zero, then the 6iV-dimensional
leaves Sm ... m close in on the 6(N — l)-dimensional leaf Smχ...m _λ. Thus,
in particular, the phase space of the two-body problem is exhibited as a limit of
phase spaces of three-body problems.
To obtain the restricted three-body problem, we must linearize about Smχmi.
According to the theory in §4, the transverse Poisson structure at the point
Σ ^ ^ ^ q — q,., v — vf.) is the Lie-Poisson manifold g*, where g C C°°(R6) is
the subalgebra consisting of those functions having a critical point at (q^Vj)
and (q 2 , v 2 ). Thus g* consists of densities on R6 modulo some loss of informa-
tion at (q19V]) and (q 2 ,v 2 ). In particular, g* contains a coadjoint orbit of the
form Sm = {mδ(q - q 3 ,v - v 3 )|(q3,v 3 ) is different from (ql9\x) and (q 2 ,v 2 )}.
Sm is just the phase space for the restricted three-body problem in which the
third body has "infinitesimal mass mΓ Its symplectic structure is mdq Λ dx.
Now we apply the theory of §5. If we linearize along an arbitrary trajectory
of the two-body problem, the result is a time-dependent hamiltonian system in
Sm. If the motion of the two bodies is circular, though, it can be identified with
a trajectory of a symmetry group generated by the functional

K(f) = / R / ( X , V ) [ ( ( X - xo) X v) ω] dxdυ,

where x 0 is the center of mass of the two-body system, and ω is its angular
velocity vector. Subtracting K(f) from H(f) (renormalized), linearizing in the
transverse space, and restricting to the coadjoint orbit Sm, we obtain precisely
the Jacobi integral. Thus we have recovered the full hamiltonian structure of
the circular restricted three-body problem.

References

[1] R. Abraham & J. Marsden, Foundations of mechanics, 2nd edition, Benjamin Cummings,
Reading, Massaschusetts, 1978.
[2] R. Abraham & J. Robbin, Transversal mappings and flows, Benjamin, New York, 1967.
[3] V. I. Arnold, Mathematical methods of classical mechanics, Graduate Text in Math., Vol. 60,
Springer, New York, 1978.
[4] F. A. Berezin, Some remarks about the associated envelope of a Lie algebra, Funct. Anal.
Appl. 1(1967)91-102.
[5] , Quantization, Izv. Akad. Nauk. SSSR, Ser. Matem. 38 (1974) 1116-1175 ( = Math.
USSR Izv. 38 (1974) 1109-1164).
[6] R. Berger, Gέomέtrie algebrique de Poisson et deformations, Publ. Dept. Math. Univ. Claude
Bernard, Lyon 16, fasc. 2, 1979, 1-69.
[7] R. L. Bishop & R. J. Crittenden, Geometry of manifolds, Academic Press, New York, 1964.
[8] M. Duflo & M. Vergne, Une propriete de la representation coadjointe d'une algebre de Lie, C.
R. Acad. Sci. Paris, Ser. A-B, 268 (1969) A583-A585.
[9] H. Freudenthal & H. deVries, Linear Lie groups, Academic Press, New York, 1969.
556 ALAN WEINSTEIN

[10] J. Gibbons, Collisionless Boltzmann equations and integrable moment equations, Physica 3D
(1981)503-511.
[11] M. Golubitsky & V. Guillemin, Stable mappings and their singularities, Graduate Texts in
Math., Vol. 14, Springer, New York, 1973.
[12] M. Gotay, R. Lashof, J. Sniatycki & A. Weinstein, Closed forms on symplectic fibre bundles,
Comment. Math. Helv., to appear.
[13] V. W. Guillemin, Band asymptotics in two dimensions, Advances in Math. 42 (1981) 248-282.
[14] V. Guillemin & S. Sternberg, Remarks on a paper of Hermann, Trans. Amer. Math. Soc. 130
(1968)110-116.
[ 15] , The moment map and collective motion, Ann. Physics 127 (1980) 220-253.
[16] S. Helgason, Differential geometry and symmetric spaces, Academic Press, New York, 1962.
[17] R. Hermann, Cartan connections and the equivalence problem for geometric structures,
Contributions to Differential Equations 3 (1964) 199-248.
[18] , The formal linearization of a semisimple Lie algebra of vector fields about a singular
point, Trans. Amer. Math. Soc. 130 (1968) 105-109.
[19] , Gauge fields and Cartan-Ehresmann connections. Part A: Interdisciplinary mathe-
matics, Vol. X, Math. Sci. Press, Brookline, 1975.
[20] , Toda lattices, cosymplectic manifolds, Bάcklund transformations and kinks. Part A:
Interdisciplinary mathematics, Vol. XV, Math. Sci. Press, Brookline, 1977.
[21] P. J. Holmes & J. E. Marsden, Horseshoes and Arnold diffusion for Hamiltonian systems on Lie
groups, Indiana Univ. Math. 32 (1983) 273-309.
[22] N. Jacobson, Lie algebras, Interscience, New York, 1962.
[23] D. Kazhdan, B. Kostant & S. Sternberg, Hamiltonian group actions and dynamical systems of
Calogero type, Comm. Pure Appl. Math. 31 (1978) 481-508.
[24] A. A. Kirillov, Unitary representations of nilpotent Lie groups, Russian Math. Surveys 17
(1962) 53-104 ( = Usp. Mat. Nauk. 17 (1962) 57-110).
[25] , Local Lie algebras, Russian Math. Surveys 31 (1976) 56-75 ( = Uspekhi Mat. Nauk.
31(1976)57-76).
[26] B. Kostant, Orbits, symplectic structures, and representation theory, Proc. U.S.-Japan Seminar
in Differential Geometry, Kyoto, Japan, 1965, Nippon Hyoronisha Tokyo, 1965, 71.
[27] , Quantization and unitary representations. Part I: Prequantization, Lecture Notes in
Math. Vol. 170, Springer, Berlin, 1970, 87-208.
[28] A. Lichnerowicz, Les varieties de Poisson et leurs algebres de Lie associees, J. Differential
Geometry 12 (1977) 253-300.
[29] S. Lie, Theorie der transformationsgruppen, (Zweiter Abschnitt, unter mitwirkung von Prof.
Dr. Friedrich Engel), Teubner, Leipzig, 1890.
[30] B. Malgrange, Equation de Lie. II, J. Differential Geometry 7 (1972) 117-141.
[31] J. Marsden & A. Weinstein, Reduction of symplectic manifolds with symmetry, Rep. Mathe-
matical Phys. 5 (1974) 121-130.
[32] , Coadjoint orbits, vortices, and Clebsch variables for incompressible fluids, Proc. of the
Conference on Order in Chaos, Los Alamos, 1982, A. Scott, ed., Physica, to appear.
[33] J. Marsden, T. Ratiu & A. Weinstein, Semi-direct products and reduction in mechanics, to
appear in Trans. Amer. Math. Soc.
[34] A. S. Mishchenko & A. T. Fomenko, Generalized Liouville method of integration of Hamilto-
nian systems, Funct. Anal. Appl. 12 (1978) 113-121 ( = Funkt. Anal. Prilozh. 12 (1978)
46-56).
[35] P. J. Morrison, The Maxwell- Vlasov equations as a continuous Hamiltonian system, Phys. Lett.
80A (1980) 383-386.
[36] J. Moser, On the volume elements on a manifold, Trans. Amer. Math. Soc. 120 (1965)
286-294.
 THE LOCAL STRUCTURE OF POISSON MANIFOLDS 557

[37] R. Moussu, Sur I'existence d 'integrates premieres pour un germe de forme de Pfaff, Ann. Inst.
Fourier (Grenoble) 26 (1976) 171-120.
[38] G. Reeb, Sur certaines proprietes topologiques des varietes feuilletees, Hermann, Paris, 1952.
[39] J. Roels & A. Weinstein, On functions whose Poisson brackets are constant, J. Mathematical
Phys. 12(1971) 1482-1486.
[40] W. Schmid, Poincare and Lie groups, Bull. Amer. Math. Soc. 6 (1982) 175-186.
[41] J. M. Souriau, Quantification geometrique, Comm. Math. Phys. 1 (1966) 374-398.
[42] J. A. Schouten, Uber differentialkomitanten zweier kontraυarianter Grδssen, Nederl. Akad.
Wetensch. Proc. Ser. A, 43 (1940) 449-452.
[43] M. Tabor & Y. M. Treve (eds.), Mathematical methods in hydro-dynamics and integrability in
dynamical systems, Proc. Amer. Inst. Phys. Conference 88, 1982.
[44] W. M. Tulczyjew, Poisson brackets and canonical manifolds, Bull. Acad. Polon. Sci. Ser. Sci.
Math. Astronom. 22 (1974) 931-934.
[45] J. Vey, Sur le lemme de Morse, Invent. Math. 40 (1977) 1-10.
[46] A. M. Vinogradov & I. S. Krasilshchik, What is the Hamiltonian formalism! Russian Math.
Surveys 30 (1975) 117-202 ( = Uspehkhi Mat. Nauk. 30 (1975) 173-198).
[47] A. Weinstein, Symplectic manifolds and their lagrangian submanifolds, Advances in Math. 6
(1971)329-346.
[48] , Fat bundles and symplectic manifolds, Advances in Math. 37 (1971) 239-250.

UNIVERSITY OF CALIFORNIA, BERKELEY