ANNALS OF PHYSICS 169, 2947 (1986)

Covariant Poisson Brackets for Classical Fields

J. E. MARSDEN AND R. MONTGOMERY

Department of Mathematics, Universiry of California,

Berkeley, California 94720

P. J. MORRISON

Deparfmem of Physics and Institute jbr Fusion Studies,

University qf Texas, Ausrin, Texas 78712

AND

W. B. THOMPSON

Department of Physics, University of California.

San Diego, La Jolla, California 92093

Received January 8, 1985

Poisson brackets that are spacetime covariant are presented for a variety of relativistic field
theories. These theories include electromagnetism, general relativity, and general relativistic
fluids and plasmas in Eulerian representation. The examples presented suggest the develop-
ment of a general theory; the beginnings of such a theory are presented. Our covariant bracket
formalism provides a general setting for. amongst other things, clarifying the transition from
the covariant formalism to the dynamical 3 + 1 Hamiltonian formalism of Dirac and
Arnowitt, Deser, and Misner. We illustrate the relevant procedures with electromagnetism.
4? 1986 Academic Press. Inc.

1. INTRODUCTION

The purpose of this paper is to show how to write the equations of some specific
general relativistic field theories in covariant Poisson bracket form. Our approach is
to proceed from explicit examples to some speculations on the structure of the
underlying mathematical theory. For each of the examples, the field equations will
be shown to be equivalent to equations of the form

where F is an arbitrary function of the fields and S is an action integral. The

theories considered fall into two categories:
A. Purefields, typified by gauge fields, where F and S in (1.1) are functions of the
basic field variables 4” and their conjugate momenta ~5.
29
ooo3-4916/86 s7.50
Copyright 0 1986 by Academic Press. Inc.
All rights of reproduction in any form reserved.
30 MARSDEN ET AL.

B. Media fields, such as those describing relativistic fluids and plasmas in

Eulerian representation, where F and S are just functions of the basic fields
(without the addition of conjugate momenta).
In either case, the formalism has these features:
1. It involves an integration over both space and time.
2. Equation (1.1) satisfies the usual properties of Poisson brackets, such as
Jacobi’s identity, so the space of fields forms a Poisson manifold (see Dirac [9]).
3. S is the action (Lagrangian) suitably expressed as a function of the field
variables.
For pure fields, the bracket contains a spacetime vector field V’ which would
correspond to the choice of a slicing, were a 3 + 1 Dirac-ADM decomposition per-
formed (see Fischer and Marsden [ 111 and Isenberg and Nester [ 191 for reviews).
For media fields, the bracket is a covariant extension of brackets of Lie-Poisson
type that are now common for 3 + 1 relativistic and non-relativistic media fields
(Iwinski and Turski [20], Morrison [37], Marsden and Weinstein [34], etc.).
We reiterate: the aim of this paper is purely observational. We simply observe
that many relativistic field theories can be written in covariant Poisson bracket
form. The basic mathematical underpinnings of the present work are not claimed to
be worked out. To complete the basic theory, one should tie up the present results
with the multisymplectic approach (see for example, Kijowski and Tulczyjew [ZS]).
It is anticipated that the covariant Poisson bracket formalism will be useful for
calculation. In this direction, Kaufman and Holm [22] have used a covariant single
particle bracket (due to Ignatiev) with success. The present formalism would be
interesting to pursue along these and other lines.
To motivate some of our results, let us first consider the simple case of particle
mechanics. We recall that the canonical Hamilton equations

aff . ai
"'=j$ (1.2)
pi=-q'
can be written as 6s = 0. where the action

SCYI =J (Pi4’-ff(q,P)) dt (1.3)

is regarded as a functional on r, the space of paths y(t) = (q(t), p(t)) in phase space
with appropriate boundary conditions (see, e.g., Arnold [2, p. 2431). Let us rewrite
this variational principle in terms of a Poisson bracket on r. For functionals F and
G of paths y, set

(1.4)
 COVARIANT POISSON BRACKETS FOR CLASSICAL FIELDS 31

where the functional derivatives are defined by

for variations 6y vanishing at the endpoints of y. It is straightforward to check that

&S[r] = 0; i.e., y solves Hamilton’s equation if and only if
{F, S)(y)=0 (1.6)

for all functionals F. It is this variational principle for Hamilton’s equations that we
shall generalize and apply to field theory. The covariant theory does not, of course,
single out a time direction; rather space and time occur on equal footing, as will be
seen below.
For general covariant brackets, the operation FH (F, S> may be viewed as a
variation of the action S along a direction in function space determined by F. In
this sense, (1.1) can be viewed as a reformulation of the conventional variational
approach to field theory. However, it is also a generalization and unifying principle,
for conventional field theories treat electromagnetism and fluids, for example, in a
rather different way.
Field theories of the traditional Euler-Lagrange form have been analyzed by
symplectic methods in a fairly well-developed way. See, for example, Chernoff and
Marsden [7], Dedecker [8], Abraham and Marsden [l], Kijowski and Tulczy-
jew [25] and Gimmsy [14] and references therein. However, a number of impor-
tant field theories do not lit this mold, just as rigid body equations in body
representation, i.e., in terms of the body angular momenta, admit a simple Poisson
description, but not a symplectic one (since, for example, there are three equations);
see for example, Sudarshan and Mukunda [41] and Holmes and Marsden [ 181 for
these descriptions. There are also Poisson bracket formulations of the equations of
fluids and plasmas in Eulerian description, as is now well known. For reviews, see
Morrison [38], Holm and Kupershmidt [17], Marsden, Weinstein, Ratiu, Schmid
and Spencer [35], Marsden and Morrison [29], and Marsden [28]. We shall
show, however, that the relativistic version of these theories (either interacting with
gravity or with a fixed background) do admit a simple covariant Poisson bracket
description. Thus, by means of (1.1) we obtain a unifying principle for media as
well as pure fields. For relativistic media field theories written in either Eulerian or
Lagrangian description but in 3 + 1 dynamical formulation, a Poisson bracket for-
mulation is known; see Bialynicki-Birula and Iwinski [6], Iwinski and Turski [20],
Bialynicki-Birula, Hubbard and Turski [S], Kunzle and Nester [26], Tulczy-
jew [43], Holm and Kupershmidt [17], Bao, Marsden and Walton [3], and
Holm [ 151. We note that the idea of covariant canonid variables has been around
for awhile; it is presented, for example, in Barut [4].
Because of its generality, Eq. (1.1) is a natural starting point for obtaining 3 + 1
reductions that result in Hamiltonian formalisms. To see how this works for our
motivating particle mechanics example, we suppose that the F of (1.6) has the form

595/1fJ9:1-3
32 MARSDENETAL.

(1.7)

where n(t) is an arbitrary function of time and 9 is an arbitrary function of the q’s
and p’s. Upon inserting (1.7) and (1.3) into (1.6) we obtain

{F, S)(y) = j n(t)@ - (9, H}‘3’) dt = 0, (1.8)

where (9, H}‘3’ is the conventional Poisson bracket. Since n(t) is arbitrary,
9 = (9, H)(3). (1.9)
This is, of course, equivalent to (1.2).
Starting with a symplectic formulation of classical field theory for pure fields,
Gimmsy [ 141 shows how to obtain the 3 + 1 adjoint Hamiltonian form of ADM
and Fischer and Marsden [ 11, 121 (a field theoretic generalization of (1.9) with
arbitrary spacetime slicings). The results here give an alternative setting for the
same procedures. We illustrate this for electromagnetism in Section 2. The for-
mulation can also be shown to yield the 3 + 1 brackets of Bao, Marsden and
Walton [3] for general relativistic fluids. In addition, the incorporation of
covariant momentum maps should be possible for these covariant Poisson struc-
tures, as well as a covariant version of the reduction procedure (Marsden and
Weinstein [33]). The latter would enable one, for example, to pass directly from a
covariant Hamiltonian description of a relativistic fluid or plasma in material
representation to one in spacetime representation (see Holm [ 151 for some results
in this direction).
The plan of the paper is as follows. We will first present Maxwell’s equations, the
relativistic Maxwell-Vlasov system, general relativity and general relativistic fluids
as examples. The covariant Poisson bracket form is exhibited explicitly in each case
and the 3 + 1 transition for electromagnetism is given. (Other examples are similar;
the authors have treated additional cases, such as the Einstein-Maxwell,
Yang-Mills or relativistic Liouville equations. For the non-relativistic Liouville
equation, see Marsden, Morrison and Weinstein [30]). We conclude with some
remarks on how these results suggest a general formulation of classical field theory.

2. ELECTROMAGNETISM

To begin, we deal with Maxwell’s equations on Minkowsi spacetime. For the

usual Euler-Lagrange variational principles, see, for example, Jackson [21]. (The
results are generalizable to arbitrary background spacetimes and to general gauge
fields.) Let A denote the four vector potential, thought of as a one-form on
Minkowski space. Let

F = dA, i.e., Fpv = a,A, - d,,A, (2.1)

 COVARIANT POISSON BRACKETS FOR CLASSICAL FIELDS 33

be the electromagnetic field tensor, where ap = a/W; p = 0, 1, 2, 3; and x0, x1, x’, x3
are the usual Minkowski coordinates.
The standard Lagrangian for the theory with an external current density Jp is

L[A]=jY:=j(-+Fp,F”“-A,,J’)d4x (2.2)

where indices are raised and lowered using the Minkowski metric. In order to
define a Legendre transformation, we introduce the covariant momentum variables,
71~”as follows:

The primary constraint manifold is defined to be the image of the map defined by
(2.3), where Y is regarded as defined on the space of A, and 8, A,‘s. This image
space is the space of pairs of fields (A,, C) with +“ skew symmetric, and is our
basic covariant phase space.
If F is a functional of A and r, one defines the functional derivatives as usual,
being cautious about the constraint on +‘l’ (just as one must be cautious about the
div B = 0 constraint in the MHD and Maxwell-Vlasov equations). Namely, 6F/6+”
is a skew tensor satisfying

for SrF a skew symmetric perturbation.

The covariant Poisson bracket of two functions F and G of A, and 7~~’is defined
by

(2.4)

where v” is an arbitrary vector field on spacetime, and functional derivatives are

defined as usual. (The vector field V is related to the passage to dynamical
equations-see Remark 3 in Section 6.) The bracket (2.4) can be written as

where (2.5)

is the associated density,

34 MARSDEN ET AL.

Let S be defined by the covariant analogue of (1.3), namely

S[A,n] = j [7f”AII*”
- H(A,7t)]d‘k, (2.6)

where

H(A, 7c)= $cJc~v + A,JP

= &‘“A jl.’ - 9. (2.7)
We claim that Maxwell’s equations are equivalent to

(F, S),. (A, n)=O (2.8)

for all V and F. The statement (2.8) is clearly equivalent to

ss
iis -0
pF- and (2.9)
6A,,=O’

i.e., to
52j”‘= -(aPAI’-$A”) and $“’
.v
= - JP (2.10)

which, together with F= dA, are the Maxwell equations. (We remark that the
choice s(A, n) = j [$T~“‘F,,,, - H(A, n)] d4x would have yielded skew symmetry of rc
as one of the consequences of (2.8), but (2.6) seems to be a more useful version for
the general theory; in fact one has, in general, a fair amount of freedom in the
choice of S. We have followed an analogue of the form (1.3). In other cases, for
example gravity, we do not follow such an analogue, but rather directly transcribe
the Lagrangian into phase space variables.)
Let us now see how this relates to the standard 3 + 1 canonical theory in which A
and -E are conjugate variables (see the earlier references or Marsden and
Weinstein [34], for example). We choose coordinates so the spacetime vector field
V is

To get a closed system, we choose J= 0 and rewrite S from (2.6) as

S = j { j [TC’~A,~,- &‘I d?} dt, (2.12)

where Latin indices run over 1, 2, 3 and where

is? = @,7t” + nio7ti0). (2.13)
 COVARIANT POISSON BRACKETS FOR CLASSICAL FIELDS 35

Choose

F[qS, n] = j” n(t) F[A,, do] dt, (2.14)

where F is a function of the 3 + 1 variables Ai and #‘. Clearly

-
;=n(t); A dt
I I

(2.15)

Hence

(2.16)

where

I?= Rd3s (2.17)

I

and {, j(3) is the usual canonical Poisson bracket for functionals of the canonically
conjugate variables A,, rci = rc”. Since n(t) is arbitrary (2.16) yields

(2.18)
In deriving (2.18) we regard Z? as a function of just Ai and xi. To do this, we re-
insert the relation

This is an indication of a general and well-known feature of 3 + l’ing: one must do

more than simply replace the rcs with 7cA=x+-the left out momenta rc> must be
regarded as functions of the 4” and xA through the Legendre transformation.
If we set rci = -E’, ?+ = &OkBk, we thus recover the usual canonical formalism for
electromagnetism. (The formalism with E and B as the basic variables-the
Pauli-Born-Infeld bracket-requires a reduction by the gauge group of elec-
tromagnetism. See Marsden and Weinstein [34]).
36 MARSDEN ETAL.

3. THE RELATIVISTIC MAXWELL-VLASOV EQUATIONS

A special relativistic particle moves in an external electromagnetic field F= dA

according to the Lorentz force law

dx”
- up; (3.1)
x--

where z is the particle’s proper time, e is its charge and m its rest mass. Declare

pP=mu,+eA (3.2)
c p

to be canonically conjugate to X~ and set

Thus (3.1) are equivalent to Hamilton’s equations

dxP
-=-; a* dPP = -a* = e uvY.
aA
(3.4)
dz ap, dz axp c axp

A relativistic plasma density f(x, p) d4xd4p is constant along its particles’ world
lines:

df af up +- eaf-
-=- up YdA = 0. (3.5)
dz axp cap, axp
We may rewrite this as

(3.6)

The basic field for the Vlasov theory is the plasma phase space density function. As
in Iwinski and Turski [20] and in the non-relativistic case (Morrison [37] and
Marsden and Weinstein [34]) we define the bracket of two functionals F, G off to
be of Lie-Poisson form:

(3.7

Let

WI = {Ax, P) H(T PI &X&P (3.8

 COVARIANT POISSON BRACKETS FOR CLASSICAL FIELDS 37

so SS/Sf = H. An integration by parts shows that the covariant bracket equation

(6 SHf)=O (3.9)

is equivalent to the relativistic Vlasov equation (3.5) [or (3.6)]. (As in Kaufman
and Holm [22, p. 278), one must suitably restrict the fields and functionals so these
integrals converge.)
The basic fields for the relativistic Maxwell-Vlasov equations are triples
rP”, f ). The bracket of two functions of (A, n, f) is just the sum of (2.3) and

(3.10)

Let

S[A,n,f]=I(n”A,.-tn,,.n”)d4x

A,)(@ -f Ap) d4xd“p. (3.1 : )

The field equations are

{F,S}.(A,n,f)=O (3.12)

for all F and all V. These are obviously equivalent to

6S
-=6s 0 and
67c~” 6A,,=O
and (3.13)

If {$$}~y~d4xd4p=0.

These are, in turn, equivalent to the relativistic Maxwell-Vlasov equations

CISu”+eiifUv!%=o
axp cap, axp

a, Fpv = f s uyf(x, p) d4p (3.14)

E;, = 8, A, - a, A,
38 MARSDENETAL.

Remarks. 1. Here we have not mentioned the obvious physical constraint that f
vanishes unless u%, = - 1. This can be treated a posteriori since it can be shown
that iff(x, p) is a solution of the relativistic Vlasov equation (3.5) defined on all of
xp space, then g(x, p) =f(x, p) S(U%, + 1), where u, = (l/m)(p, - eA./c), also is a
solution. Alternatively, this constraint can be treated by restricting to density
functions concentrated on the mass shell at the outset.

2. The bracket (3.7) is literally the Lie-Poisson bracket for the group of
canonical transformations on xp space, the cotangent bundle of spacetime. Thus,
this part of the bracket can be regarded as the reduction from canonical coor-
dinates in Lagrangian representation by the particle relabeling group. In
Lagrangian representation, the bracket has a form similar to (2.3); the vector field
Y should disappear during reduction because one relabels by world lines, not by
points (I, p). This is part of a general covariant reduction process which is planned
for future development.

3. Another reduction process that we plan to pusue is the elimination of the

gauge freedom for electromagnetism via reduction. This should re-express the
bracket in terms of FPv and f alone and build in the div E constraint. When
expressed dynamically, this should reproduce the known bracket for relativistic
plasmas (Iwinski and Turki [20] and Bialynicki-Birula, Hubbard and Turski [S]),
and should coincide with the non-relativistic bracket (Morrison [37], Marsden and
Weinstein [ 341).

4. GENERAL RELATIVITY

The basic field variables we use for general relativity are the contravaiant sym-
metric two-tensor g@ representing the dual metric and the “conjugate momenta”
rc$ which are symmetric in c(and /I. We shall identify rr$ with the affine connection;
this is standard, although not strictly true from the point of view of the Legendre
transformation because of second derivatives of gap in the Lagrangian density (see
Misner, Thorne and Wheeler [36, Chap. 211, Kijowski and Szczyrba [24], and
Szczyrba [42] ).
The Poisson brackets are of the same form as (2.3) namely

(4.1)

Here functional derivatives are defined so that those with respect to g”” are tensors:
 COVARIANT POISSON BRACKETS FOR CLASSICAL FIELDS 39

whereas those with respect to 7~;~are tensor densities:

d
F(rc + 1611) = j” z hc$d4x.
z j.=O

The action is the usual one written in terms of gzp and ?rcD:

(4.2)

where ,,& d4x is the volume element on spacetime,

R,, = a,lT;, - aljnfj,- 7rfp:, + 7cfj,7r;,( (4.3)

and where

is an (externally imposed) stressxnergy tensor.

The covariant bracket equations are

{F, S>,,=O for all F, V:

i.e.,

(4.4)

The first equation yields the field equations

G,,j = gzTxp, (4.5)

where
G,, = R,,{ - ;gXp R”,

while the second equation can be shown to imply

(p 61:~ := a,(p &Qi) + n;,.,p vG

+ 7-$g”” Jq - n;,,g”fl& = 0, (4.6)

which implies that n is the Levi-Civita connection:

~if,~= wmg,,, + 66, - 4gd. (4.7)

(See Misner, Thorne, and Wheeler [36], Chap. 21, Sect. 21.)
40 MARSDEN ETAL.

We may think of (4.2) as depending parametrically on a set of matter and

radiation fields 4” through an additional Lagrangian L*. To couple these fields to
the gravitational fields (gap, nip) we need a covariant bracket for the 4. Then the
equations

ought to be equivalent to

and the field equations for the 4. This suggestion is followed in the next section.
Our treatment of general relativity is of course a reformulation of the standard
Palatini variational principle. One interesting feature of our bracket formulation is
that it allows an interesting coupling with media fields, as we shall see.
On the negative side, our choice of how to write S in (4.2) is somewhat ad hoc
following the Palatini formalism and not as close to the form (1.3) as one might
like. Also, the correspondence between the covariant bracket (4.1) and the
canonical Dirac-ADM bracket through a 3 + 1 process is suggestive, but it has not
been worked out. Difficulties of this or equivalent sorts are common to all the
canonical or symplectic treatments of general relativity we know of.

5. GENERAL RELATIVISTIC FLUIDS

We consider a perfect adiabatic fluid coupled to gravity; see Misner, Thorne, and
Wheeler [36, Chap. 221 for background. One can similarly treat, we presume,
plasmas coupled to general relativity (the Maxwell-Einstein-Vlasov system) or
charged general relativistic fluids or general relativistic MHD.
The basic fluid quantities are the following scalar fields:
p = fluid mass-energy per unit rest three volume
n = baryon number density per unit rest three volume
(T= entropy per unit rest three volume
p = pressure in a rest frame
s = entropy per baryon
p = relativistic inertial mass per unit rest three volume.
We have the relations

0 = ns and u=p+p. (5.1)

The equation of state has the form

P = dn, 0) (5.2)
 COVARIANT POISSON BRACKETS FOR CLASSICAL FIELDS 41

and the pressure is determined by the Legendre transform

ap ap
p=n~+aej-yp. (5.3)

The basic fluid variable are taken to be

n, 0, and M, = pu,.

Here u’ is the four velocity of the fluid, which satisfies u’u, = - 1, i.e., M”A4, = pL?.
This constraint is to be imposed after functional derivatives, i.e., variations are
taken. Here indices are raised using the Lorentz dual metric gap. The constraint
A, = - 1 can either be imposed directly, as we do, or be viewed as a constraint in
the sense of Dirac associated to the gauge symmetry of curve reparametrizations.
(The latter requires some work on covariant momentum maps-see Section 6
below.)
The fluid brackets are taken to be Lie-Poisson with a structure similar to that in
the non-relativistic case (Morrison and Greene [39], Dzyaloshinskii and
Volovick [lo] ):

The Lie algebra underlying this Lie-Poisson bracket is a semi-direct product of vec-
tor fields and (densities x densities), similar to the nonrelativistic case (see
Marsden [27], Holm and Kupershmidt [16] and Marsden ef al. [35]). Here,
functional derivatives are defined to be vectors or scalars, not densities:

-$,- F(M+dsM)=[$SM,fid4x
,. -0 ?
d
F(n+Mn)={$Sn&d4x
a i.=o
and
d
F(o + ASo) = j ~SO fi d4x.
z j,=O

We note that the two minus signs in (5.4) are in apparent disagreement with the
non-relativistic and 3 + 1 version of the theory (see the above references and Bao,
Marsden and Walton [3, Eq. (lC.131). However, when the covariant theory is
42 MARSDEN ET AL.

decomposed into its 3 + 1 parts, this discrepancy should disappear (for example,
when a bracket of vector fields on spacetime is decomposed, the result looks like a
semi-direct product bracket, but with a relative sign switch due to the signature
( + + + -) of the spacetime metric; cf. Fischer and Marsden [12, Appendix II].)
For the coupled system we use the variables

( $3 n:,, M,, 4 0)
and use the bracket (5.4) plus (4.1). For the action we take

-8742 g”“M, M, + V( n, 0) fi d4x, (5.5)

where R,, is given by (4.3) and

Un,o) = i[p(n, 0) - p(n, a)]. (5.4)

We note that the fluid term in (5.5), when evaluated on the constraint set
M”M, = p2 is proportional to the integral of the pressure. The covariant bracket
equations are

{F, s>,,=o (5.7)

for all F and I’. Choosing F= F( g, 71)gives

rc$ = Levi-Civita connection of g

and
G,, = 87cT,,i,
where
T,, = PU,U~ +m,pt (5.8)

as in the previous section. In getting &S,,,,/6gED = 87cTzB,we used the constraint

g”“M,M, = - 11’ after taking the variation. Choosing F = F(n ) and F = F(o) gives

(nu’)t, = 0 and (mz);, = 0, (5.9)

i.e., conservation of baryon number and entropy. (The apparent discrepancy in

(5.8) by a factor of 2 is discussed in Bao, Marsden and Walton [3].) Finally,
choosing F= F(M,) gives V. T= 0, which, of course, also follows from (5.8) and
the Bianchi identity.
 COVARIANT POISSON BRACKETS FOR CLASSICAL FIELDS 43

6. GENERAL CANONICAL FIELD THEORIES

We sketch here a framework in which the canonical brackets (2.5) and (4.1) can
be constructed and in which the Euler-Lagrange equations for a pure field are
equivalent to the covariant bracket equations. As we have remarked this also
covers, in principle, fluids and plasmas by reduction of this structure from
Lagrangian (material) representation to Eulerian (spatial) representation. (In 3 + 1)
form, the connection between these is discussed in Holm [15].) General relativity,
as usual, is anomalous: it is formally similar, but does not quite fit the scheme
presented here.
Our fields are assumed to be sections of a vector bundle rr: Y + X over a base
manifold X (we take X to be spacetime-but for plasmas it is T*(spacetime) or the
mass hyperboloid therein). We suspect that most of what we describe also works
for a general fiber bundle, but we have restricted to the vector bundle case for sim-
plicity. The fields are described in local coordinates by dA(xp), where A is a multi-
index for field components and .x?’ are spacetime coordinates. Let Y be a given
Lagrangian density defined on J’(Y), the first jet bundle of Y. Recall that the fiber
.I:( Y) of J’(Y) over a point ?: E T, is

J,‘.(Y)= Y,@T~;X=Y(T,X, Y,). (6.1)

The Lagrangian density of a field C$is locally given by U(&“, 8,d”). The field
equations are the usual Euler-Lagrange equations for 9,

(6.2)

and we set

We now describe (6.3) intrinsically (cf. Kijowski and Tulczyjew [25]). Let A4X be
the bundle of four forms (densities) over X so

9 : J'( Y) -+ A"(X). (6.4)

Let P be the bundle over X whose fiber at x is

P,=(Y,0T.~X)*On4,X~TT,XOY:o/14,X. (6.5)
Describe P by local coordinates (@, z;). The Legendre transformation is the fiber
derivative of Y:
44 MARSDEN ET AL.

given locally by

where n; is given by (6.3).

Let F and G be functionals of sections of P. Then we have a Poisson bracket

where ~F/c%$~ is a section of Y* 0 A4X + X and 6G/6n$ is a section of

Y@ T*X-+ X. They pair together to give a section of T*X@ A4X + X which can be
contracted with the vector field V and the resulting four form integrated over X.
For each fixed V, the bracket makes the sections of P into a Poisson manifold.
The primary constraint set C is the image of the Legendre transformation. We
will assume that it is a vector subbundle of P. This will be the case, for example, if
2’ is quadratic in 8,4” and if its “kinetic matrix” 8’2/a(a,4”) 8(8,@) has con-
stant rank.
Let 1: P + P be a smooth vector bundle projection with im I = C. For example, in
electromagnetism, 1 would project any tensor density onto its skew-symmetric part.
Then l*: Y@ T*X + Y@ T*X. Set C* = im I*, a subbundle of Y@ T*x. This is a
bundle dual to C, so that functional derivatives with respect to the constrained
covariant momenta naturally take values in C*:

where F is a functional of sections of C, &r is a variation in C and c?F/Sz~ takes

values in C*; thus (6F/h$). SK: is a density on X.
One may now define brackets of functionals on T(C) (sections of C) by the same
formula as before (6.6), however, where the 8/&r: are interpreted as sections of C*.
These brackets satisfy all the conditions for Poisson brackets. The only non-obvious
condition is the Jacobi identity. To check this, we extend functionals F on f(C) to
functionals F= i*F on r(P) as follows:

%A xl= F(hln).
For a general extension r’, we have

by the fiber linearity of 1. However, for our extension,

 COVARIANT F’OISSON BRACKETS FOR CLASSICAL FIELDS 45

since /y(S&%r) = 0, where I, = id - 1. Note that the bracket of two such extensions
is again such an extension and hence
{F,iC}={F,G}-. (6.7)
Now Jacobi’s identity follows since it holds for the extended functions.
Remarks. 1. Equation (6.7) is expressed by saying that (4, X) I-+ (4, /7~) is a
Poisson map. Thus the constrained brackets are the pull back of the full brackets
by the projection T(P) + r(C). Note that the injection F(C) + r(P) is not a
Poisson map for C # P. In fact,

for arbitrary extensions.

2. For general fiber bundles, or if C is not a vector bundle, the results just
described require extension. This development should be done in conjunction with
examples such as general relativistic fluids written in the Lagrangian (material) pic-
ture.
The “Hamiltonian density” is uniquely defined on C by
H(4, n) = WrdA - w$A, Q”). (6.7)
At first, the right-hand side is defined on J’(Y) x C. However, the partial derivative
with respect to a,$” is zero, so we get a well-defined density on C. Set

and note that

6S
and (6.9)
jg=O,
i.e.,
(F, S} “=O for all F, V (6.10)

reproduce respectively

a 6H
and (6.11)
8XP
-nn:,= -v
which are equivalent to the Euler-Lagrange equations (6.2).
Remarks. 1. In the above setting, only canonical brackets are described. Non-
canonical brackets, such as those for fluids and plasmas, are expected to come from
canonical brackets in Lagrangian representation as in the non-relativistic case by a
covariant version of the reduction process. See Marsden, Ratiu and Weinstein
[31, 321.
46 MARSDEN ETAL.

2. We conjecture that covariant momentum maps associated with a group

action G on P should be defined to be maps J: P --f g* Q TXQ A4P where 9 is the
Lie algebra of G. These should be consistent with the covariant momentum maps
defined in Gimmsy Cl43 and should include standard Noether identities.
As in Gimmsy [14], one can presumably show that for an appropriately
covariant localized theory, J vanishes on solutions of the field equations and that
these conditions J= 0 correspond to first class constraints in the sense of Dirac [9].
The momentum maps should play a key role in the reduction process, as in the
nonrelativistic case (Marsden et al. [35].
3. As already noted, the mathematical development of a systematic 3 + 1 analysis
is incomplete. This requires further development of the theory along the lines of
Remarks 1 and 2. Once this is done, the 3 + 1 analysis should proceed as in
Gimmsy [ 141. In particular, the 3 + 1 procedure applied to the covariant brackets
and field equations should directly yield the dynamical Poisson brackets and the
evolution equations in bracket form (which is equivalent to the adjoint form of
Fischer and Marsden [ 11, 123). As we saw in Section 2, the vector V in the
bracket (6.6) plays an important role in the 3 + 1 process. It corresponds to the
arbitrariness in the choice of the direction of time and to the lapse and shift which
appear in the dynamical formulation. Forming the variables 4” and nA = ~5 V, is a
first step in constructing conjugate variables for the 3 + 1 formalism. Subsequently,
one must also eliminate the so-called “atlas fields” (such as the temporal com-
ponent of A in electromagnetism), as in Gimmsy [ 141.
4. The results of this paper also need to be studied with a view towards
understanding limits and averaging (see, for example, Weinstein [44] and Similon,
Kaufman and Holm [40]). For example, we presume that the fluid bracket (5.4)
can be derived from the plasma bracket (3.10) in the cold plasma limit (Gibbons,
Holm and Kuperschmidt [13]) and that, as in Marsden et al. [35], taking
moments via reduction gives a Poisson map between these structures.

ACKNOWLEDGMENTS

We are grateful to D. Bao, G. Gimmsy, D. Helm, A. Kaufman and T. Ratiu for helpful discussions and
remarks.

REFERENCES

1. R. ABRAHAM AND J. MARSDEN, “Foundations of Mechanics,” 2”“ ed., Addison-Wesley,

2. V. ARNOLD, “Mathematical Methods of Classical Mechanics,” Springer-Verlag, Berlin/New York/
Heidelberg, 1978.
3. D. BAO. J. MARSDEN AND R. WALTON, Comm. Math. Ph.vs. 99 (1985), 319-345.
4. A. 0. BARU~, “Electrodynamics and Classical Theory of Fields and Particles,” Dover, New York.
1904.
5. I. BIALYNICKI-BIRULA, J. C. HUBBARD, AND L. A. TURSKI, Physica A 128 (1984). 509-519.
 COVARIANT POISSON BRACKETS FOR CLASSICAL FIELDS 47

6. I. BIALYNICKI-BIRULA AND Z. IWINSKI, Rep. Math. Ph.w. 4 (1973), 139-151.

7. P. CHERNOFF AND MARSDEN, “Properties of Infinite Dimensional Hamiltonian Systems,” Lecture
Notes in Mathematics Vol. 425, Springer-Verlag, Berlin/New York/Heidelberg, 1974.
8. P. DEDECKER, “On the Generalizations of Symplectic Geometry to Multiple Integrals in Calculus of
Variations,” Lecture Notes in Mathematics Vol. 570. pp. 395-456. Springer-Verlag.
Berlin/New York/Heidelberg, 1977.
9. P. A. M. DIRAC, “Lectures on Quantum Mechanics,” Belfer Graduate School of Science,
Monograph Series, Vol. 2, Yeshiva University.
IO. I. E. DZYALOSHINSKII AND G. E. VOLOVICK. Ann. Phys. (N.Y.) 125 (1980), 67-97.
11. A. FISCHER AND J. MARSDEN, Gen. Relafiv. Gruu. 12 (1970), 915-20.
12. A. FISCHER AND J. MARSDEN, in “Isolated Gravitating Systems in General Relativity” (J. Ehlers,
Ed.). pp. 322-395, Italian Physical Society. 1979.
13. J. GIBBONS, D. D. HOLM, AND B. A. KUPERSCHMIDT. Phxs. D 6 (1983), 179-194.
14. G. GIMMSY. Constraints, momentum maps. and the dynamics of classical relativistic field theories,
preprint, 1985.
15. D. D. HOLM, Physica D 17 (1985), l-36.
16. D. D. HOLM AND B. A. KLJPERSHMIDT, Physica D 6 (1983), 347-363.
17. D. D. HOLM AND B. A. KLJPERSHMIDT. Phy.7 Ler. A 101 (1984), 23-26.
18. P. J. HOLMES AND J. E. MARSDEN, Indiana Unir. Marh. J. 32 (1983), 273-310.
19. J. ISENBERG AND J. NESTER. in “General Relativity and Gravitation. Vol. 1” (A. Held, Ed.), Plenum.
New York, 1980.
20. Z. R. IWINSKI AND L. A. TURSKI, Lett. Appl. Eng. Sci. 4 (1976). 179-191.
21. J. D. JACKSON, “Classical Electrodynamics,” 2nd ed., Wiley, New York. 1975.
22. A. KAUFMAN AND D. HOLM, Phys. Letr. A 105 (1984), 277-279.
23. G. W. KENTWELL. Phys. Lett. A, in press.
24. J. KIJOWSKI AND W. SZCZYRBA, Comm. Math. Phys. 46 (1976), 183-205.
25. J. KIJOWSKI AND W. TULCZYJEW. “A Symplectic Framework for Filed Theories.” Lecture Notes in
Physics Vol. 107, Springer-Verlag, Berlin/New York/Heidelberg. 1979.
26. H. P. KUNZLE AND J. M. NESTER, Hamiltonian formulation of gravitating perfect fluids and the
Newtonian limit, J. Math. Phys. 25 (1984), 1009%1018.
27. J. E. MARSDEN, A group theoretic approach to the equtions of plasma physics. Can. Math. Bull. 25
(1982). 129-142.
28. J. E. MARSDEN, Nonlinear stability in fluids and plasmas, preprint, 1984.
29. J. E. MARSDEN AND P. J. MORRISON, Conl. Math. Amer. Math. Sot. 28 (1980). 133-150.
30. J. E. MARSDEN. P. H. MORRISON, AND A. WEINSTEIN, Conr. Math. Amer. Math. Sot., 28, 115-124.
31. J. E. MARSDEN.,T. RATIU, AND A. WEINSTEIN, Semidirect products and reduction in mechanics.
Trans. Amer. Math. Sot. 281 (1984), 147-177.
32. J. E. MARSDEN. T. RATIU, AND A. WEINSTEIN, Cont. Maih. Amer. Math. Sot. 28 (1984), 55-100.
33. J. E. MARSDEN ANU A. WEINSTEIN. Rep. Math. Phys. 5 (1974), 121-130.
34. J. E. MARSDEN AND A. WEINSTEIN, Physica D 4 (1982), 394406.
35. J. E. MARSDEN, A. WEINSTEIN, T. RATIU. R. SCHMID, ANV R. G. SPENCER, Atri Acad. Sci. Torino 117
(1983). 289-340.
36. C. W. MISNER. K. THORNE, AND J. A. WHEELER,“Gravitation,” W. H. Freeman, San Francisco.
37. P. J. MORRISON, P&w. Lert. AS0 (1980), 383-386.
38. P. J. MORRISON, in “Mathematical Methods in Hydrodynamics and Integrability in Related
Dynamical Systems,” (M. Tabor, Ed.), AIP Conf., #88 La Jolla, 1982.
39. P. J. MORRISON AND J. M. GREENE. Phys. Reu. Lert. 45 (1980), 79-794.
40. P. L. SIMILON, A. N. KAUFMAN, AND D. D. HOLM. Phys. Left. 106 (1984), 29-33.
41. E. C. G. SUDARSHAN AND N. MUKUNDA, “Classical Mechanics-A Modern Perspective,” Wiley.
New York, 1974 (Second edition, Krieger. Huntington, N.Y., 1983).
42. W. SZCZYRBA, Comm. Math. Phys. 51 (1976), 163-182.
43. W. M. TULCZYJEW. “Geemetrodynamics Proceedings” (A. Prastraro. Ed.) pp. 91-99, 1983.
44. A. WEINSTEIN, J. Differential Geom. 18 (1983), 523-557.

595/169/l-4