GENERALIZED HAMILTONIAN DYNAMICS

P. A. M. DIRAC

1. Introduction. The equations of dynamics were put into a general form

by Lagrange, who expressed them in terms of a set of generalized coordinates
and velocities. An alternative general form was later given by Hamilton, in
terms of coordinates and momenta. Let us consider the relative merits of the
two forms.
With the Lagrangian form the requirements of special relativity can very
easily be satisfied, simply by taking the action, i.e. the time integral of the
Lagrangian, to be Lorentz invariant. There is no such simple way of making
the Hamiltonian form relativistic.
For the purpose of setting up a quantum theory one must work from the
Hamiltonian form. There are well-established rules for passing from Hamilton's
dynamics to quantum dynamics, by making the coordinates and momenta
into linear operators. The rules lead to definite results in simple cases and,
although they cannot be applied to complicated examples without ambiguity,
they have proved to be adequate for practical purposes.
Thus both forms have their special values at the present time and one must
work with both. The two forms are closely connected. Starting with any
Lagrangian one can introduce the momenta and, in the case when the momenta
are independent functions of the velocities, one can obtain the Hamiltonian.
The present paper is concerned with setting up a more general theory which
can be applied also when the momenta are not independent functions of the
velocities. A more general form of Hamiltonian dynamics is obtained, which
can still be used for the purpose of quantization, and which turns out to be
specially well suited for a relativistic description of dynamical processes.

2. Strong and weak equations. We consider a dynamical system of N degrees

of freedom, described in terms of generalized coordinates qn(n = 1,2, . . . , N)
and velocities dqn/dt or qn. We assume a Lagrangian L, which for the present
can be any function of the coordinates and velocities
(1) L - L(q, q).
We define the momenta by
(2) pn = dL/dqn.
For the development of the theory we introduce a variation procedure,
varying each of the quantities qn, qny pn independently by a small quantity
èqn, dqn, opn of order e and working to the accuracy of e. As a result of this
This paper is based on the first half of a course of lectures given a t the Canadian Math-
ematical Seminar in Vancouver in August 1949.
129

https://doi.org/10.4153/CJM-1950-012-1 Published online by Cambridge University Press

 130 P . A. M . DIRAC

variation procedure equation (2) will get violated, as its left-hand side will be
made to differ from its right-hand side by a quantity of order e. We shall now
have to distinguish between two kinds of equations, equations such as (2)
which get violated by a quantity of order e when we apply the variation, and
equations which remain valid to the accuracy e under the variation. Equation
(1) will be of the latter kind, since the variation in L will equal, by definition,
the variation of the function L(q,q). The former kind of equation we shall call
a weak equation and write with the usual equality sign = , the latter we shall
call a strong equation and write with the sign = .
We have the following rules governing algebraic work with weak and strong
equations :
if il s 0 then ôA = 0;
if 1 = 0 then bX j* 0;
in general. From the weak equation X = 0 we can deduce
ÔX2 = 2XÔX = 0,
so we can deduce the strong equation

Similarly, from two weak equations X\ = 0 and X2 = 0 we can deduce the

strong equation
XiX 2 s 0.
It may be that the N quantities dL/dqn on the right-hand side of (2) are all
independent functions of the N velocities qn. In this case equations (2) deter-
mine each g as a function of the q's and p's. This case will be referred to as
the standard case, and is the only one usually considered in dynamical theory.
If the dL/dq's are not independent functions of the velocities, we can elimin-
ate the q's from equations (2) and obtain one or more equations
(3) <K<Z, P) =0
involving only q's and p's. We may suppose equation (3) to be written in
such a way that the variation procedure changes <j> by a quantity of order e,
since if it changes 0 by a quantity of order e*, we have only to replace <j> by
<j)l/h in (3) and the desired condition will be fulfilled. We now have equation (3)
violated by the order e when we apply the variation, so it is correctly written
as a weak equation.
We shall need to use a complete set of independent equations of the type (3),
say
(4) <t>m(q, p) = 0, m = 1,2, . . . , M.
The condition of independence means that none of the <£'s is expressible
linearly in terms of the others, with functions of the q's and p's as coefficients.
The condition of completeness means that any function of the q's and p's
which vanishes on account of equations (2) and changes by the order e with
the variation procedure is expressible as a linear function of the <f>m with
functions of the q's and p's as coefficients.

https://doi.org/10.4153/CJM-1950-012-1 Published online by Cambridge University Press

 GENERALIZED H A M I L T O N I A N DYNAMICS 131

We may picture the relationship of strong and weak equations in the fol-
lowing way. Take the 3N dimensional space with the q's, q's and p's as co-
ordinates. In this space there will be a 2 N dimensional region in which
equations (2) are satisfied. Call it the region R. Equations (4) will also be
satisfied in this region, as they are consequences of (2). Now consider all
points of the 3 N dimensional space which are within a distance of order e
from R. They will form a 3 N dimensional region like a shell with a thickness
of order e. Call this the region Re. A weak equation holds in the region Rt
a strong equation holds in the region R€.
3. The Hamiltonian. The Hamiltonian H is defined by
(5) H = pnqn - L,
where a summation is understood over all values for a repeated suffix in a
term. We have
bH = b(pnqn -L)
= Pn8qn + qnbpn ~ dL/dqn.ôqn - dL/dqn.ôqn
(6) = qnbpn - dL/dqn.ôqn.
We find that ôH does not depend on the bq's. This important result holds
whether we have the standard case or not.
Equation (5) gives a definition for H as a function of the q's, q's and p's,
holding throughout the 3 N dimensional space of q's, q's, and p's. We shall
use the definition only in the region Rt, and in this region the result (6)
holds, to the first order. This means that, if we keep the qs and p's con-
stant and make a first-order variation in the q's, the variation in H will be
of the second order. Thus if we keep the qs and p's constant and make a
finite variation in the q's, keeping all the time in the region Rt (which is
possible when we do not have the standard case), the variation in i^will be
of the first order. If we keep in the region R, the variation in H will be zero.
It follows that in the region R, H is a function of the q's and p's only. Calling
this function {£>(q, p), we have the weak equation
(7) H = $(q, p)
holding in the region R. In the standard case the function § is the ordinary
Hamiltonian.
Starting from a point in R and making a general variation, we have from

(6) W - $) - L - ^ W -(£.+ &) Sqn.

\ dpj \dqn dqn/
Thus ô(H — Jp) depends only on the 8qs and op's. If the variation is such
that we stay in the region R, then of course ô(H — § ) = 0 . Thus d(H — § )
vanishes for any variation of the q's and p's such that one can choose the ôq's
so as to preserve equations (2). The only restriction this imposes on the ôq's
and ô£'s is that they must preserve equations (4), i.e. they must lead to
ô<t>m = 0 for all m. Thus b(H — § ) is zero for any values ôq, ôp that make
ô(f>m = 0, and hence for arbitrary bq, bp

https://doi.org/10.4153/CJM-1950-012-1 Published online by Cambridge University Press

 132 P. A. M. DIRAC

(8) Ô(H - $ ) = vmô<t>m

with suitable coefficients vm. These coefficients will be functions of the qs,
^'s and p's, and with the help of (2) can be expressed as functions of the
qs and q's only. We now get
Ô(H - § - Vm<t>m) = S(H - § ) — VrrMm ~ <j>m^m = 0
from (8) and (4), and hence
(9) H s § + z,m0m.
We have here a strong equation, holding to the first order in the region Rt,
in contradistinction to the weak equation (7), which holds only in R.
Equation (8) gives
oil = ô£> + vmô<f>m

dp- <5pn + —- àqn + z>m ( dpn + dqn I .

dp
C o m p a r i n g this w i t h (6),, we g e t

d<t>m
(10) qn + vm
dpn dpn
ÔL d<t>m
(11) — + Vm i

dqn dqn dqn

Equations (10) give the q's in terms of the q's, p's and v's. They show
that the 2N variables qn,qn can be expressed in terms of the 2N + M variables
qn, pn, vm. Between these 2N + M variables there exist the M relations (4).
There cannot be any other relations between these variables, as otherwise
the 2N variables qn, qn would not be independent. Thus the v's must each
be independent of the qs, p's and other v's. The v's can be considered as a
kind of velocity variables, which serve to fix those qs that cannot be ex-
pressed in terms of qs and p's.
When we work with the Hamiltonian form of dynamics we use as basic
variables the qs, p's and v's, between which certain relations (4) are assumed
to exist, and which are otherwise independent. These variables will be called
the Hamiltonian variables.

4. The equations of motion. We assume the usual Lagrangian equations

of motion as weak equations,
(12) pn = dL/dqn.
By substituting for the p's in (12) their values given by (2), we get equations
involving the accelerations qn. In the standard case these equations will
determine all the q's in terms of the qs and qs. In the case with M equations
(4), the equations of motion will give us only N — M equations for the q's.
The remaining M equations of motion will tell us how the </>m's vary with time.

https://doi.org/10.4153/CJM-1950-012-1 Published online by Cambridge University Press

 GENERALIZED H A M I L T O N I A N DYNAMICS 133

For consistency the </>m's must remain zero. These consistency conditions
will be examined later.
With the help of (11) the equations of motion (12) take the form

dqn dqn
Equations (13) together with (10) constitute the Hamiltonian equations of
motion. They are fixed by the function § and the equations <t>m = 0. The
Hamiltonian equations of motion give us the qs and p's in terms of the
Hamiltonian variables q, p, v. They give us no direct information about the
v's, but will give us some information indirectly when we examine the con-
sistency conditions.
The Hamiltonian equations of motion can be expressed more easily with
the help of the Poisson bracket notation. Any two functions £ and 77 of the
qs and p's have a P. b. [£, 77], defined by
(i4) [ç,„]silii _ il i i .
dqn dpn dpn dqn
It is easily verified that the P. b. remains invariant under a transformation
to new q's and p's, in which the new qs are any independent functions of the
original qs and the new p's are defined by the new equations (2) with L
expressed in terms of the new qs and their time derivatives. This invariance
property gives the P. b. its importance.
P. b.'s are subject to the following laws, which are easily verified from the
definition:

(15) \ U,/(ui,i,2, - . . ) ] - -f & "l] + r ^ & * ! + ••• •

In the second of these l a w s / is any function of various quantities 771, 772, . . . ,

each of which is a function of the qs and p's. The third law, known as
Poisson's identity, applies to any three functions £, 77, f of the q's and £'s.
It is desirable to extend the notion of P. b.'s to include functions of the
qs which are not expressible in terms of the q's and p's. We assume these
more general P. b.'s are subject to the laws (15) but are otherwise arbitrary.
Alternatively, we may assume that the qs are arbitrary functions of the q's
and p's, and the laws (15) can then be deduced with J, the 77's and f involving
the qs.
From a strong equation ^ 4 = 0 we can infer the weak equations

£ £ = 0 ) ^ - = 0 , ^-=o,
dqn dqn dpn
and hence, by an application of the second of the laws (15),
K, ^4) = 0

https://doi.org/10.4153/CJM-1950-012-1 Published online by Cambridge University Press

 134 P. A. M. DIRAC

for any £. We may have [£, A] = 0, (for example when A = 0 by definition)

but this is not necessarily so. From a weak equation X = 0 we cannot infer
[£, X] = 0 in general.
If g is any function of the q's and p'sy we have from (10) and (13)
a àg(d^ d<t>m\ àg(d^ d<t>m\
dqn\dpn dpn/ dpn\dqn dqn/
(16) = [g, $\ + Vm \g, <t>m].
This is the general Hamiltonian equation of motion. It may also be written,
with the help of (4), as
(17) g = [g, $ ] + Vm \g, <t>m] + [g, Vm] 4>m = [g, H ] ,
when it is the same as the usual Hamiltonian equation of motion in P. b.
notation.

5. Homogeneous velocities. The theory takes a specially simple form in

the case when the Lagrangian is homogeneous of the first degree in the veloci-
ties. The momenta defined by (2) are then homogeneous of degree zero in the
q's and so depend only on the ratios of the q's. Since there are N p's and only
N — 1 independent ratios of the q'sf the p's now cannot be independent func-
tions of the q's and there must be at least one relation (4) connecting the q's
and p's. The case when there is only one relation between the q's and p's may
now be considered as the standard case.
From Euler's theorem we have
(18) L s qn dL/dqn
and hence L = qnpn
so that
(19) H = 0.
This weak equation holding in the region R allows us to take § = 0, so that
(9) becomes
(20) H ^Vmïrn-
The general equation of motion (16) is now
(21) g = »m[g, 0m].
The Hamiltonian equations of motion are now fixed entirely by the equations
<t>m = 0 .
Equation (21) is homogeneous in the v's on the right-hand side. Given any
solution of the equations of motion, one can obtain another solution from it
by multiplying all the v's by a factor 7, which may vary arbitrarily with the
time. The new solution will have the time rate of change of all dynamical
variables multiplied by the factor 7. The new solution would be obtained
from the previous solution if we replaced the time ^ by a new independent
variable T such that dt/dr = 7. The new independent variable is completely
arbitrary: it can be any function of t and the q's and <z's. Thus, given any
solution of the equations of motion, we can get another solution from it by

https://doi.org/10.4153/CJM-1950-012-1 Published online by Cambridge University Press

 GENERALIZED HAMILTONIAN DYNAMICS 135

replacing t by an arbitrary r, so the equations of motion give us no information

about the independent variable. This is an important feature of dynamical
theory with homogeneous velocities, and makes it specially convenient for a
relativistic treatment.
The Lagrangian for any dynamical system can be made to satisfy the condition
for homogeneous velocities by taking the time t to be an extra coordinate qo
and using the equation qo = 1 to make the Lagrangian homogeneous of the first
degree in all the velocities, including q0. The new Lagrangian equations of
motion for all the q's can then be deduced, as has been shown by the author [1].
In this way we can get a new formulation for a general dynamical system in
terms of homogeneous velocities. The new formulation gives all the equations
of the old formulation except the equation q0= 1. If we want to have this
equation in the new formulation we may assume it as a supplementary condi-
tion, not derivable from the equations of motion but consistent with them. We
can, however, very well dispense with it, as its only effect is to fix the indepen-
dent variable, which would otherwise be arbitrary in the homogeneous
velocity formulation.
Thus we may confine ourselves to the homogeneous velocity theory without
losing any generality. We shall do this in future as it leads to somewhat
simpler equations, and use the dot to denote differentiation with respect to an
arbitrary independent variable r.
6. The consistency conditions. For consistency the equations of motion
must make each of the <j>m remain zero. Thus, putting 4>m* for g in (21),
we get
(22) Vm [0m, <M = 0.
Let us suppose the equations (22) to be reduced as far as possible with the
help of the set of equations (4). The reduction may involve the cancellation
of factors wherrwe can assume these factors do not vanish. The resulting
equations must each be of one of four types.
Type 1. It involves some of the variables v.
Type 2. It is independent of the v's but involves some of the variables p and
q. It is thus of the form
(23) x(ff, P) = 0
and is independent of the equations (4).
Type 3. It reduces to 0 = 0.
Type 4. It reduces to 1 = 0.
An equation of type 2 leads to a further consistency condition, since we
must have x remaining zero. Putting x for g in (21), we get
(24) *U0m, x] = 0.
This equation, reduced as far as possible with the help of equations (4) and,
any equations (23) that we already have, will again be of one of the four types.

https://doi.org/10.4153/CJM-1950-012-1 Published online by Cambridge University Press

 136 P. A. M. DIRAC

If it is of type 2 it will lead to yet another consistency condition. We continue

in this way with each equation of type 2 until it leads to an equation of another
type.
If any of the equations obtained in this way is of type 4, the equations of
motion are inconsistent. This case is of no interest and will be excluded in
future. Equations of type 3 are automatically satisfied. We are left with the
equations of types 1 and 2 to fit into the theory.
Let us call the complete set of equations of type 2
(25) Xk(q,P) =0, k = 1,2, . . . , K .
We may suppose the functions xk to be chosen, like the 4>m in (4), so that their
variations are of order e. Equations (25) are then correctly written as weak
equations. These further weak equations will reduce the region R, in which
all weak equations hold, so as to have only 2N — K dimensions. The region
Rt will also get reduced, as it will now consist of all points within a distance of
order e from the new region R.
For studying the equations of type 1, it is convenient to introduce some new
concepts. We define one of the quantities 4>m to be a first class <>/ if its P. b.
with every <f> and % vanishes. Thus <t>m> is first class if
,2Q\ [<t>m'y <t>™\ = 0, m = 1, 2, . . . , M,
[0m', Xk] = 0, k = 1,2, . . . , K.
T h e s e equations need hold only in t h e weak sense, which m e a n s t h a t t h e y need
hold only a s consequences of t h e equations 4>m — 0, xk = 0. T h u s t h e left-
h a n d sides of (26) m u s t each equal in t h e strong sense some linear function of
t h e <t)m a n d xk- A <j> which does n o t satisfy all these conditions we call a second
class <j>.
We can make a linear transformation of the <£'s of the form
(27) 0*m = 7mm'0m',
where the 7's are any functions of the g's and p's such that their determinant
does not vanish in the weak sense. The </>*'s are then equivalent to the <j>'s
for all the purposes of the theory.
Let us make a transformation of this kind so as to bring as many 0's as
possible into the first class. Let us call the first class </>'s that we then have
</>a's and the second class ones c^'s, with /3 = 1, 2, . . . B and a = B + 1,
B + 2, . . . M.
If <j>m> is first class, equation (22) is automatically satisfied. Further, in
equations (22) and (24) we can restrict <t>m to be second class, as first class </>w's
contribute zero. Thus the surviving equations (22) and (24) will read
/OQN Vf>\4>e,4>r\ = 0 , 0, p = 1, 2, . . . , B,
K }
vfi[4>0,xk] = 0 , k = 1,2, ...,K.
These are all the equations of type 1. They show that either all the ^ ' s vanish
or the matrix

https://doi.org/10.4153/CJM-1950-012-1 Published online by Cambridge University Press

 GENERALIZED H A M I L T O N I A N D Y N A M I C S 137

0 [01, 02] [01,03] • •• [01, 0B] [01, Xl] • • [01, XK

[02,0l] 0 [02, 03] • • [02,0B] [02, Xl] • • [02, XK
(29)
[0B>0l] [0B> 02] [0B,03] •• • 0 [0B,Xl] [0B,XK
is of rank less than B, in the weak sense.
It will now be proved that the first alternative is the correct one. Assume
the matrix (29) is of rank u < B. Form the determinant
01 0 [01,02] [01,03] ••• [01,0u]
02 [02,0l] 0 [02,03] ••• [02,0u]
(30) D
0U+1 [0U+l>0l] [0U+l>02] [0U+l>03] [0U+l>0u]
D is a linear function of the 0^'s and so vanishes in the weak sense. The P. b.
of D with any quantity / equals the sum of the determinants formed by taking
the P. b. of each column of (30) with./. All these determinants, except the
one formed by taking the P. b. of the first column with / , will vanish in the
weak sense, as the elements of their first column all vanish in the weak sense.
Thus
[01,/] 0 [01,02] [01,03] ... [01,0,;]
[02,/] [02,0l] 0 [02,03] ... [02.0J
(31) [D,f\ =
[0U+1>/] [0U+l>0l] [0U+l>02] [0U+l>03] [0u+i>0u]
If we t a k e / to be any 0 a , the first column of (31) vanishes and so [D, <f>a] = 0.
If we t a k e / to be any 0^ or x, the determinant (31) either has two columns
identical and so vanishes, or it is a minor of the matrix (29) with u + 1 rows
and columns, and vanishes because this matrix is assumed to be of rank u.
Thus D has zero P. b. with all the 0's and x's.
It may be that D vanishes in the strong sense on account of the co-factors
of the elements of its first column all vanishing in the weak sense. If this is
the case, we take a different determinant D, with its columns after the first
one corresponding to any u of the columns of (29) and its rows corresponding
to any u + 1 of the rows of (29). We can always choose such a determinant D
so that the co-factors of the elements of its first column do not all vanish,
from the assumption that (29) is of rank U. We get in this way a D which is
a first class 0 and is a linear function of the 0^'s. This contradicts the assump-
tion that we had previously put as many 0's as possible in the first class.
We can conclude that if we have put as many <t>'s as possible in the first class,
the v's associated with the second class 0's all vanish. The Hamiltonian (20)
then reduces to
(32) H s va<t>a1
and the general equation of motion (21) becomes
(33) g = va[g, 0 J .

https://doi.org/10.4153/CJM-1950-012-1 Published online by Cambridge University Press

 138 P. A. M. DIRAC

The vanishing of the v$s together with the equations (25) ensure that the
consistency conditions are all satisfied. The va's remain completely arbitrary.
Each of them gives rise to a freedom of motion—an arbitrary function in the
general solution of the equations of motion. In the standard case there is
just one </>, which is necessarily first class, and thus there is one arbitrary
function in the general solution of the equations of motion. This is connected
with the arbitrary character of the independent variable T.

7. Supplementary conditions. In dealing with a particular dynamical sys-

tem, we may wish to impose equations on the coordinates and velocities addi-
tional to the equations of motion that follow from the Lagrangian. Such
supplementary conditions must be introduced as further weak equations in
the theory.
With the help of equations (10) (with ^ = 0) the supplementary conditions
can be expressed as relations between the q's, p's and v's. They may lead to
equations between the q's and p's only. Such equations must be treated as
extra x equations, to be joined on to the set (25). They will give rise to further
consistency conditions, which are to be handled in the same way as the pre-
ceding ones and may lead to still more x equations. A first class <£ must now
be defined to have zero P. b. also with these new x's, so the number of first
class <t>'s may be reduced by the supplementary conditions. This would cause
a reduction in the number of freedoms of motion.
Those of the supplementary conditions that do not give x equations will
give conditions on the va variables. These conditions will usually be of a more
complicated kind than merely the vanishing of certain v's, like all the conditions
on the v's that follow from consistency conditions. They will make a further
reduction in the number of freedoms of motion, reducing it to less than the
number of first class <£'s.

8. Transformations of the Hamiltonian form. Take a set of functions

6s(s — 1, 2, . . . s) of the q's and p's such that the determinant
I 0 [0i,0 2 ] [01,0s] ... [0i,0 s ]
(34) A s [ft, 0J 0 [0 2 ,0 8 ] ... [0 2 , 0S]

I [0s, 0i] [0s, 0d [0s, 0a] . . . 0

does not vanish in the weak sense. This implies that s must be even. Let
c8B* denote the co-factor of [0S, 06>'] divided by A, so that

and
(35) cssf[6s, 0S/'] = 5 S ' S " .
Then we can define a new P. b. [£, rj]* for any two quantities £ and 77 by the
formula
(36) [«,i?]*=[£,i?] + [ £ , 0 j ^ [0^,17].

https://doi.org/10.4153/CJM-1950-012-1 Published online by Cambridge University Press

 GENERALIZED HAMILTONIAN DYNAMICS 139

It is easily seen that the new P. b.'s obey the first two of the laws (15), and
after some calculation one finds that they also obey the third, Poisson's iden-
tity. (See Appendix.) The new P. b.'s make
[£, 0J* = fc oa] + [f, es,] cMeM
s & *J - [f, 0S>] «•.
(37) = 0
for any £.
To understand the significance of the new P. b.'s, let us take the case when
the 0's consist of | s coordinates q and their conjugate p's. We then see that
the new P. b.'s are obtained by omitting the terms involving differentiations
with respect to these q's and p's from the summation over n in the definition
(14). Thus the new P. b.'s refer to a system of N — Js degrees of freedom.
If, instead of taking the 0's to be just certain g's and p's, we take them to be
any independent functions of these q's and p's, we get the same new P. b.'s.
With general 0's the new P. b.'s will still refer to a system of N — Js degrees
of freedom, but the reduction of the degrees of freedom is made in a more
complicated way than the mere omission of certain q's and p's.
Let us suppose the 0's are all <t>'s or x's. (The <f>'s must be second class, as
otherwise A = 0.) We then have [0S, H] = 0 for all s, and hence
(38) \g,H\* = [g,H\ = g
for g any function of the q's and p's. Thus the new P. b.'s may be used to give
the Hamiltonian equations of motion. We get in this way a new form for the
equations of motion, which is simpler because the number of effective degrees
of freedom is reduced.
Each of the 0's now vanishes in the weak sense. If we work only with the
new P. b.'s, we can assume each of the 0's vanishes in the strong sense without
getting a contradiction, because from (37) the new P. b. of a 0 with anything
vanishes. We can then use the strong equations 0S = 0 to simplify the
Hamiltonian.
Let us define a % to be first class it if has zero P. b. with all the <£'s and x's
and to be second class otherwise. We can make a linear transformation of the
x's of the form
(39) x** = ykk'Xw + y'km<l>m,
where the 7's and 7"s are any functions of the q's and p's such that the deter-
minant of the 7's does not vanish in the weak sense, and the new x's are then
equivalent to the old ones for all the purposes of the theory. Let us make a
transformation of this kind so as to bring as many x's as possible into the first
class, and let us call the first class x's that we then have Xa's and the second
class ones XA'S.
We may take the 0's to consist of all the 0^'s and X/s's. The determinant
A then does not vanish. The proof of this result is similar to the proof that
the matrix (29) is of rank B, and consists in assuming that A is of rank T < s
and constructing a determinant like

https://doi.org/10.4153/CJM-1950-012-1 Published online by Cambridge University Press

 140 P. A. M. DIRAC

I 0i 0 [0X,02] ... [0i,0 T ] I

(4Q\ 02 [02, 0l] 0 ... [0 2 , 0 T ]

I 0T+1 [0T+1) 0l] [0T+1» 02] ••• [0T+1> 0 T ] I

which is then seen to be a first class 0 or x and is a linear function of the fo's
an
d X/s's, so it contradicts the assumption that as many <£'s and x's as possible
have been put in the first class.
With this choice of 0's we get the maximum simplification of the Hamiltonian
equations of motion by this method. We get a new scheme in which all the
4>p and xp equations are strong equations. We may be able to use these
equations to eliminate some of the g's and p's entirely from the theory.
The form of the new scheme is not unique, because the fa's and xa's are not
unique. If we merely replace the c^'s and X/s's by linear functions of them-
selves, we do not change the final form. We can, however, add to the 0^'s
any linear functions of the <£a's, and to the xs's any linear functions of the
# 0 's and Xa's> which does not change A or the css>, but does in general change
[£> >?]*, and so the form of the Hamiltonian scheme is altered. The different
forms must, of course, be equivalent, as they all give the same equations of
motion.
As an application of the above method, let us consider the case of a Lagrang-
ian that does not involve some of the velocities. Suppose L does not involve
q}- (j = 1, 2, . . . , j < N ) . Then each pj equals zero in the weak sense and equals
a <t> in the strong sense. Suppose that no linear combination of the p/s is first
class. Then we can take the p/s to be ^ ' s . Let us now take half the 0's to
be the p/s and the other half to be suitable second class <£'s or x's so that A
does not vanish. Call these other 0's 0y. With this choice of 0's one easily
sees that the new P. b.'s are just what one would get if one applied the defini-
tion (14) to those degrees of freedom for which q is not a qj, with each pj
reckoned as strongly equal to zero and each qj reckoned as strongly equal to
a function of the other g's and p's given by the equations 0j = 0. We get in
this way a new Hamiltonian scheme (not necessarily with the maximum
simplification, as there may be other ^ ' s and xp's not included in the 0's)
in which the qj and pj do not appear as independent dynamical variables.
The new scheme could be obtained in a more direct way by not counting
the g/s as coordinates right from the beginning, and not introducing momenta
conjugate to them at all. Let us see what modifications this would bring into
the development of the theory.
Define n so as to take on only those values for which q is not a qj, i.e. the
values j + 1, J + 2 , . . . , N. Then equations (2) and (5) still hold and equation
(6) must be replaced by
(41) 8H = qnàpn — dL/dqn.àqn — dL/dqj.ôqj,
as we allow the g/s to vary. We may assume the equations
(42) dL/dqj = 0

https://doi.org/10.4153/CJM-1950-012-1 Published online by Cambridge University Press

 GENERALIZED HAMILTONIAN DYNAMICS 141

as supplementary conditions with this method. Equation (41) then reduces to

precisely (6). We can infer that H is of the form (20), where the <j>m are func-
tions of the qnys and pnJs} independent of the g/s, that vanish on account of
equations (2). The remainder of the theory can be developed as before, in
terms of <j>'s and x's that do not involve the g/s . Those of the <>
/ or x equations
that do involve the g/s can be looked upon as defining the q/s in terms of
the other variables, and play no further role in the theory.
With this form of the theory we have the Lagrangian containing variables
q} that involve momenta. The appearance of momentum variables in the La-
grangian is analogous to the appearance of the velocity variables va in the
Hamiltonian.

9. The Hamiltonian as starting point. Instead of starting with the La-

grangian and obtaining the Hamiltonian from it, one can start with the Hamil-
tonian. One begins by assuming certain dynamical variables qn and pn (n = 1,
2 , . . . , N ) , or maybe other dynamical variables between which there are definite
P. b. relations satisfying the laws (15), and assuming certain weak equations as
<>
/ equations connecting them. There is no point in distinguishing </>'s and x's
with this method. At least one of the #'s must be first class, i.e. must have
zero P. b. with all the </>'s, or there can be no consistent motion. One then
assumes the Hamiltonian to be a linear function of the first class <£'s <£a, with
new variables va as coefficients, and assumes the Hamiltonian equations of
motion (17) or (33). The vJs can vary arbitrarily with the independent
variable r.
The previous scheme of equations of motion, derived from a Lagrangian and
involving possibly x's as well as </>'s, is to be looked upon as an example of the
present scheme with some of the v's restricted to be zero by supplementary
conditions. The c^'s corresponding to these va's are then the first class x's of
the previous scheme. Such supplementary conditions, or any supplementary
conditions involving the v's, are of no value for the application of the theory to
relativistic dynamics given in the next section and cannot be taken over into
the quantum theory, so they will not be included in the further work. Supple-
mentary conditions not involving the v's are just <j> equations.
The P. b. of two first class <t> s is a first class 0, as may be verified in the follow-
ing way. The P. b. [0 a , <f>a'] vanishes weakly and so is strongly equal to a
linear function of the #'s, these being the only quantities that are weakly zero
in the present scheme. We have to show that its P. b. with an arbitrary <j> is
weakly zero. From Poisson's identity
(43) [0, foal 4>a,]] s [[</>, « J , «.,] - [[</>, 4v], 0 J .
Since <f)a is first class, [<j>, <f>a] vanishes weakly and so is strongly equal to a linear
function of the 0's, and hence its P. b. with the first class 4>a' vanishes weakly.
Similarly the second term on the right-hand side of (43) vanishes weakly,
so the required result is proved.

https://doi.org/10.4153/CJM-1950-012-1 Published online by Cambridge University Press

 142 P. A. M. DIRAC

Suppose there are A independent first class <£'s and M independent </>'s
altogether. In the phase space (the 2N-dimensional space of the qn and pn
variables) there is a space of (2N — M) dimensions in which all the 4> equations
are satisfied. Call it the (2N — M)-space. The state of the dynamical system
for a particular r value is fixed by giving values to the qJs and p's satisfying
all the <t> equations and is thus represented by a point P in the (2N — M)-space.
The motion of the system ensuing from this state is represented by a curve in
the (2N — M)-space starting from P. On account of the A variables va being
arbitrary, this curve may start out in any direction in a small space of A
dimensions surrounding P . There is one of these small spaces of A dimensions
surrounding every point of the (2N — M)-space. It will now be shown that
these small spaces are integrable.
Suppose that for an interval of T, 8T = €1, all the v's vanish except vat, which
is equal to 1, and that for the following T interval, 5r = e2, all the v's vanish
except »a", which is equal to 1. Then any function g of the q's and p's is changed
at the end of the first interval to
g + €l[g,0.'].
It is changed at the end of the second interval, with the accuracy €ie2 but with
neglect of ei2 and e22, to
(44) g + €i[g, 4>A + *\g + €i[g, 0.,], 4>a»l
If the two kinds of motion are made in the reverse order, g changes to
(45) g + e2[g, 4>A + €i[g + e2[g, 0 a "], <M-
The difference between (44) and (45) is, by Poisson's identity,
(46) ei€2[g,[4v, 4>A\.
It was shown above that [0a/, <t>a»] is a first class </>, so that (46) is a possible
change in g arising from the equations of motion with a suitable choice of the
z>'s, and thus corresponds to a motion in the small A-dimensional space round
the starting point. This is the condition for integrability.
If there are supplementary conditions involving the v's, this integrability
may get spoiled. Thus the integrability does not necessarily hold for the
equations of motion derived from a Lagrangian.
The integration of the small spaces will provide a set of A-dimensional spaces
lying in the (2N—M)-space such that the motion always takes place entirely
in one of them. Call these spaces ^.-spaces. Every curve in an A -space represents
a possible solution of the equations of motion. Every point of the (2N—M)-
space lies in an A -space, which contains all the motions starting from that
point. It would be permissible to consider the A -space itself as the complete
solution of the equations of motion, rather than a general curve in it.
Given a particular A -space, we can fix a point of it by A coordinates, each
of which is some function of the ç's and p's. Call these coordinates ta (a = 1,
2, . . . , A). They will play the role of time variables. The A -space itself can

https://doi.org/10.4153/CJM-1950-012-1 Published online by Cambridge University Press

 GENERALIZED HAMILTONIAN DYNAMICS 143

be described by giving all the q's and p's as functions of the ta. If g is any q or
p, or a function of the q's and p's, we have
(47) g = ia dg/dta,
since the r variation of g may be looked upon as arising from the T variation
of the ta's. Using the Hamiltonian equations of motion (33) for g and i a , we
get
»afe, 0 a ] = Va[ta, 0 a ] dg/dta.

This equation holds for arbitrary va, so

(48) [g, 0 j = [ t , <*>«] ag/a^0.
Equations (48) may be looked upon as the general equations of motion that
fix an A -space. They are the closest equations in the theory with homogeneous
velocities to the usual Hamiltonian equations of motion. If A = 1, we may
take the one variable ta to be the time and (48) then reduces to precisely the
usual Hamiltonian equations of motion.
To pass from the Hamiltonian to the Lagrangian, we introduce the velocities
qn by the equations
(49) qn = vad<f>Jdpn,
and then define L by
(50) L = pnqn - H = pnqn - va<t>a.
This gives L as a function of the q's, q's, p's and v's, linear in the q's and v's.
Making independent variations ôq} ôq, ôp> ôv, we get
ôL — qnàpn + pn8qn — <f>aôva — Va(d<t>a/dqn.ôqn + d<t>a/dpn.dpn)
(51) = pnàqn - Va d(j>a/dqn.ôqn.
Thus ôL does not depend on the ôpn and ôva. This result is to be compared
with (6).
If the equations (49) together with the <t> equations give the qs as indepen-
dent functions of the p's and v's, so that they allow the p's and v's to be con-
sidered as functions of the q's and g's, then (51) shows that L is strongly equal
to a function of the q's and q's only. This function must be homogeneous of
the first degree in the q's. Differentiating it partially with respect to a g or g,
we find
dL/dqn = p n
dL/dqn = — Va d<f>a/dqn = p n .
These are the usual Lagrangian equations.
If the equations (49) together with the <t> equations do not give the q's as
independent functions of the p's and v's, they lead to certain equations
between the q's and q's only, say
(53) Ri(qfq) = 0 , j = 1,2, . . . , j .
The R's are homogeneous in the q's and we arrange them to be of the first
degree. We now proceed by a method analogous to that of §3 with the role of
p's and q's interchanged. We obtain a result analogous to (9),

https://doi.org/10.4153/CJM-1950-012-1 Published online by Cambridge University Press

 144 r. A. M. DIRAC

(54) L s g + Xy R,;
where 8 is a function of the qs and qs only, which must be homogeneous of
the first degree in the qs, and the coefficients Xy are functions of the qs, p's
and v's.
We have again equations (52) if the X's are counted as independent variables
in the partial differentiations of L, and L is then homogeneous of the first degree
in the qs. Thus we have a Lagrangian containing momentum variables of
the type considered at the end of the preceding section, with the previous qj
corresponding to the present Xy and the supplementary conditions (42) giving
the equations (53).

10. Application to relativistic dynamics. In the ordinary non-relativistic

dynamics one works with the state of a dynamical system at a particular
instant of time, this state being specified by giving values to the qs and p's.
One has equations of motion which enable one, given the state at one instant,
to calculate the state at another instant. These equations of motion, written
in the Hamiltonian form with homogeneous velocities, need only one first class
<t>.
To get a dynamical theory which satisfies restricted relativity, we must set
up a scheme of equations which applies equally to observers with all velocities.
If we work with instants, we must include instants with respect to all observers.
An instant is then any flat three-dimensional surface in space-time having a
normal in a direction within the light-cone. A general instant needs four para-
meters to describe it, three to fix the direction of the normal, or the velocity
of the observer, and the fourth to distinguish different instants for the same
observer.
A relativistic dynamics that involves instants must enable one, given the
state at any of these instants, to calculate the state at any other. We must
have equations of motion showing how the dynamical variables vary as the
instant varies. We can allow the instant to vary arbitrarily, with a trans-
lational motion in space-time as well as the direction of its normal varying,
and the equations of motion must always apply. Thus we need four first
class (f)Js to give rise to the four freedoms of motion of the instant.
The four parameters that describe the instant are to be treated as qs,
subject to the equations of motion (17) or (33) along with the ether qs and
p's. They are distinguished from the other qs and p's in that it is specially-
convenient to take them as the / variables of the equations of motion (48).
These equations then show directly how any q or p varies for a given variation
of the instant.
There are other forms of relativistic dynamics not involving instants, which
have been discussed by the author [2]. There is the point form, in which a
state is defined with reference to a point in space-time. This form also needs
four first class <£'s, corresponding to the four freedoms of motion of the point.
Then there is the front form, which needs three first class $'s, corresponding

https://doi.org/10.4153/CJM-1950-012-1 Published online by Cambridge University Press

 GENERALIZED HAMILTONIAN DYNAMICS 145

to the three freedoms of motion of a front. Finally, we may take a state to

be defined on a general three-dimensional space-like surface in space-time.
There must then be infinitely many first class c/>'s, corresponding to all the
deformations that may be made in such a surface. With each of these forms
the variables that describe the point, front, or general space-like surface are
to be treated as qs, subject to the equations of motion (17) or (33), and are
specially convenient to be taken as the t variables of equations (48).
The first class </>'s discussed above are the fewest with which one can construct
a relativistic dynamics in the respective forms. There may be additional
ones. For example, an electrodynamics which allows gauge transformations
to be made after one has fixed the initial values of all the g's and p's must con-
tain extra freedoms of motion, which will need extra first class #'s to give
rise to them.

11. Quantization. In order to quantize a dynamical system which one has

worked out in the classical theory, one must set up a scheme of linear operators
corresponding to the classical dynamical variables q and p, and to functions of
them. There are no operators corresponding to the classical variables v, or to
velocity variables in general, or to anything involving r. The operators all
operate on the vectors i> of a Hilbert space, whose representatives in any
representation are the wave functions which specify states in the quantum
theory. Real classical variables correspond to self-adjoint operators.
The linear operators must be analogous to their classical counterparts in
accordance with two general principles. Using the same letter to denote two
things that are counterparts, the principles are
(i) P. b. relations between the classical variables correspond to commutation
relations between the operators, according to the formula
[£, r]] corresponds to 27r(£?7 — rj£)/ih.
(ii) Weak equations between the classical variables correspond to linear con-
ditions on the vectors \{/, according to the formula
-X'Gz» P) — 0 corresponds to X\[/ = 0.
The procedure of passing from the classical to the quantum theory is not
mathematically well-denned, because whenever a classical-quantity involves a
product of two factors whose P. b. does not vanish, there is an ambiguity in
the order in which the two factors should appear in the corresponding quan-
tum expression. In practice with simple examples one finds no difficulty in
deciding what the order should be. With complicated examples it may be im-
possible to choose the order in each case so as to make all the quantum equa-
tions consistent, and then one would not know how to quantize the theory.
The present-day methods of quantization are all of the nature of practical
rules, whose application depends on considerations of simplicity.
There are certain general features of the passage to the quantum theory
that one must pay attention to, in order that the consistency of the quantum

https://doi.org/10.4153/CJM-1950-012-1 Published online by Cambridge University Press

 146 P. A. M. DIRAC

equations shall not go wrong in an elementary way. We have in the classical

theory a number of 0 equations (counting % equations also as 0 equations),
which are to be used in the quantum theory according to the principle (ii).
We can transform the classical 0's linearly by the transformation (27) and the
new 0's are just as good as the old ones. If we make the corresponding trans-
formation in the quantum theory, we must take care to put the coefficients y
all to the left of the 0's. A general 0 in the quantum theory is a linear function
of the given 0's with coefficients on the left.
From two quantum equations obtained from 0 equations by principle (ii)
0i* = 0, 0 2 * = 0,
we can infer
0201* = 0, 0102* = 0,
and hence from principle (i),
[01, 02,] * = 0.
This corresponds to the classical weak equation
[01, 02] = 0.
We can infer that all the 0's must be first class if the passage to the quantum theory
is possible.
Given a classical theory with second class 0's, one can get a quantum theory
from it by first applying a transformation of the type described in §8, which
converts all the 0/s equations into strong equations. The strong equations will
correspond in the quantum theory to equations between operators, which serve
to define some of them in terms of the others.
The quantum equations 0* = 0, obtained by applying principle (ii) to the
first class 0 equations of the classical theory, are the Schroedinger wave equa-
tions. The usual classical dynamics with only one first class 0 leads to only
one Schroedinger equation. In the general theory there is one Schroedinger
equation for each classical freedom of motion. The operators in these equations
all correspond to classical dynamical variables for one r value. The operators
referring to a different r value do not belong to the same algebraic scheme,
and there does not seem to be anything in the quantum theory analogous to
the r dependence of the classical variables.
However, the dependence of the classical variables on the parameters /,
given by equations (48), does have a quantum analogue, provided the Vs are
chosen so as to have zero P. b.'s with one another, so that they can be given
numerical values simultaneously in the quantum theory. The specially con-
venient /'s for the various forms of relativistic dynamics discussed in the pre-
ceding section do satisfy this condition. We cannot immediately take over
equations (48) into the quantum theory because, as easily verified, the equa-
tions that we should get would not be invariant under a general linear trans-
formation of the 0's (27). We must first put equations (48) in a standard
form. By a transformation (27) we introduce a new set of 0's, 0 a say, in one-
one correspondence with the / a 's, so that

https://doi.org/10.4153/CJM-1950-012-1 Published online by Cambridge University Press

 GENERALIZED HAMILTONIAN DYNAMICS 147

(55) [ta, 4>af] = < W -

With these <j>s equations (48) reduce to
(56) [g, 0O] = dg/dta.
These equations can be taken over into the quantum theory, and are then
Heisenberg's quantum equations of motion for the present generalized
dynamics.

Appendix. Proof of Poisson's identity for the new P. b.'s defined by (36).
Use the suffixes r, s, t, . . . to distinguish different 0's. We have by the
definition
IK,*]*, n*
= [tt,ij] + [f ,«r]c,.[fl„ij],f] + [K.ij] + K,tf,]cr.[ff„ij]f 6»(]c(ttK,r]
(57) = [[{,,], f] + [[ÇA], flcr.[0.,ij] + fc,0,][c..,fl[0..'j] + [tA]crS[[8s,vU]
+ [[*,*], «Jc(„[tf„,r] + [[M 6t]crS[9s,v}ctu[du,{]
+ [i,er][Crs,Bt][es,ri\Ctu[Bu,ï\ + [ZA]Cr.\[9.,Ti\, ff|]C|«[tf«,f].

Let the operator £ denote the application of the three cyclic permutations
of £, 17, f and the summation of the three results. Then we have to prove that
E t t E d V l * = 0.
£ applied to the first term of (57) gives zero, from the ordinary Poisson's
identity. 2Z applied to the second, fourth and fifth terms gives
T,Crs[ee,li\{[[M f] + [fer,?],*] + [[te]A]} = 0,

from the ordinary Poisson's identity again. X applied to the sixth and eighth
terms of (57) gives, with a cyclic permutation of r, u, s, t in the latter,
(58) CrsCtuY.[es^][eu^]{[[iAiet] + [[*„& dr]} = -cr8ctuj:[e8,ii][ouMOrA]&
From (35) we can infer
[Ctu[0r,6tU] = 0
or
(59) [ctuMOM + ctu[[dr,0t],& = 0.
Thus (58) reduces to
Crs[er,dt]j:iOS,v][Ou,t][Ctuà] = ?L[0t,li\[0u,t][Ctu,Z],

with a further use of (35). This cancels with ]£ applied to the third term of
(57). £ applied to the remaining term of (57), the seventh, gives
(60) [ÇA][nA][Ç,eu]{ctu[Cr8,0t] + Ctr[csufit] + Ct8[CurA]}.

Using J^rsu to denote the operation of applying the three cyclic permutations
of r, s, u and simultaneously of r', s', u' and adding the three results, we have
from the ordinary Poisson's identity

https://doi.org/10.4153/CJM-1950-012-1 Published online by Cambridge University Press

 148 P. A. M. DIRAC

(61) Ylrsu Cr'rCsrsCuru[[dr',Os']iQu'] = 0.

From (59) with £ replaced by dut.

[Cr'r,0u'][0r',0s'] + ^r , r[[^r / ,^s']^u / ] = 0,
so (61) gives

With the help of (35), this reduces to

f ==
E rsu Cu'u \.Crsi"u \ v,

which shows that (60) vanishes. This completes the proof. All the above
equations may be written as strong equations, as no weak equations are used
in the proof.

REFERENCES

[1] P. A. M. Dirac, Homogeneous variables in classical dynamics, Proc. Camb. Phil. Soc,
vol. 29 (1933), 389.
[2] , Forms of relativistic dynamics, Rev. Mod. Phys., vol. 21 (1949), 392.

St. John's College, Cambridge

https://doi.org/10.4153/CJM-1950-012-1 Published online by Cambridge University Press