Singular Lagrangians, Constrained Hamiltonian Systems and

Gauge Invariance: An Example of the Dirac–Bergmann Algorithm

J. David Brown1, ∗
1
Department of Physics, North Carolina State University, Raleigh, NC 27695
(Dated: March 10, 2022)
The Dirac–Bergmann algorithm is a recipe for converting a theory with a singular Lagrangian
into a constrained Hamiltonian system. Constrained Hamiltonian systems include gauge theories—
general relativity, electromagnetism, Yang–Mills, string theory, etc. The Dirac–Bergmann algorithm
is elegant but at the same time rather complicated. It consists of a large number of logical steps
linked together by a subtle chain of reasoning. Examples of the Dirac–Bergmann algorithm found
in the literature are designed to isolate and illustrate just one or two of those logical steps. In this
paper, I analyze a finite-dimensional system that exhibits all of the major steps in the algorithm. The
arXiv:2201.06558v3 [gr-qc] 9 Mar 2022

system includes primary and secondary constraints, first and second class constraints, restrictions
on Lagrange multipliers, and both physical and gauge degrees of freedom. This relatively simple
system provides a platform for discussing the Dirac conjecture, constructing Dirac brackets, and
applying gauge conditions.

I. INTRODUCTION

In the early 1950’s, Dirac and Bergmann independently developed the Hamiltonian formalism for systems with
singular Lagrangians [1–10]. These systems, often called “constrained Hamiltonian systems”, include gauge theories.
Gauge freedom is more clearly and more completely displayed in the Hamiltonian setting, with the generators of gauge
transformations expressed as functions on phase space. Historically, the main motivation for casting gauge theories
in Hamiltonian form was to facilitate their canonical quantization. Dirac and Bergmann were primarily motivated by
the prospect of developing a quantum theory of gravity based on a Hamiltonian formulation of general relativity.
Textbook treatments of Lagrangian and Hamiltonian mechanics invariably assume that the Lagrangian L(q, q̇)
is nonsingular; that is, that the matrix of second derivatives of L(q, q̇) with respect to the velocities is invertible.
In classical mechanics, the nonsingular case appears to be sufficient to cover problems of physical interest. However, one
might argue that textbooks avoid certain physically interesting problems simply because their Lagrangians are singular.
In field theory, the issue of singular Lagrangians and gauge freedom cannot be avoided. Nearly every field theory
of physical interest—electrodynamics, Yang–Mills theory, general relativity, relativistic string theory—has gauge free-
dom.
The Dirac–Bergmann algorithm transforms a singular Lagrangian system into a Hamiltonian system. The formalism
is elegant but at the same time rather complex. It consists of a large number of logical steps, linked together by a
chain of reasoning that can be difficult to keep straight. Of course, there are many examples in the literature in which
the Dirac–Bergmann algorithm is applied, converting a singular Lagrangian into Hamiltonian form. However, to
my knowledge, all of these examples are designed to illustrate just one or two of the logical steps in the algorithm.
The student of the subject is faced with the task of linking these examples together to create a complete picture of
the algorithm.
For those who learn by example, what is needed is a single example that illustrates all of the major logical steps
in the Dirac–Bergmann algorithm and shows how these steps are linked together. Such a “complete” example is not
easy to identify because there is no obvious way to predict, starting with a particular Lagrangian, which of the steps
in the algorithm will be needed.
The system analyzed in this paper is defined by the Lagrangian
1n o
L(q, q̇) = (q1 + q̇2 + q̇3 )2 + (q̇4 − q̇2 )2 + (q1 + 2q2 )(q1 + 2q4 ) (1.1)
2
where the dot denotes a time derivative. The matrix of second derivatives with respect to the velocities is
0 0 0 0
 
2
∂ L 0 2 1 −1
= . (1.2)
∂ q̇i ∂ q̇j 0 1 1 0
0 −1 0 1

∗ [email protected]
 2

This matrix is singular; it has rank 2.

As we will see, the system defined by the Lagrangian (1.1) is relatively complete.1 It contains both primary and
secondary constraints, both first and second class constraints, and restrictions on the Lagrange multipliers. The first
class constraints for this system are not all primary; this allows us to address the Dirac conjecture. The second class
constraints can be eliminated by introducing Dirac brackets. Finally, this system contains both physical and gauge
degrees of freedom. The gauge freedom can be eliminated with suitable gauge conditions.
One characteristic of any complete example such as (1.1) is that the configuration space, the space of q’s, must be
at least four-dimensional. Here is why: The number of physical degrees of freedom is equal to the dimension of the
configuration space, minus the number of first class constraints, minus half the number of second class constraints. If
the example is to have at least one physical degree of freedom, at least two first class constraints (one primary and
one secondary), and at least two second class constraints (the number of second class constraints must be even), then
the configuration space must be at least four–dimensional.
The study of constrained Hamiltonian systems predates Dirac and Bergmann with earlier work by Rosenfeld [12,
13]. Like Rosenfeld, Bergmann and his collaborators [1, 2, 4, 6, 7, 14] were focused on field theories such as general
relativity that are covariant with respect to general four-dimensional coordinate transformations. (For a historical
review, see [15].) Dirac took a more basic approach to the problem by considering a generic singular Lagrangian [3, 5, 8–
10]. He developed the algorithm for the case of systems with a finite number of degrees of freedom. His view was that
the generalization to field theory, with an infinite number of degrees of freedom, would be “merely a formal matter”.
Although the typical starting point for the Dirac–Bergmann analysis is a singular Lagrangian, not all gauge theories
can be expressed in terms of a Lagrangian that depends only on the q’s. General relativity is one such example.
The Einstein–Hilbert action is a functional of Lagrange multipliers (the lapse function and shift vector) as well as the
configuration variables (the spatial metric). Nevertheless, the Dirac–Bergmann algorithm provides the foundation for
our interpretation of general relativity as a constrained Hamiltonian system.
In this paper, I apply the Dirac–Bergmann algorithm to the Lagrangian (1.1), following closely the general treatment
given by Henneaux and Teitelboim [11]. In turn, the account of Henneaux and Teitelboim closely follows Dirac’s 1964
Lectures on Quantum Mechanics [10]. Presentations of the Dirac–Bergmann algorithm can also be found in books by
Hanson, Regge and Teitelboim [16], Sundermeyer [17], Rothe and Rothe [18], and Lusanna [19].
Throughout the paper, I attempt to explain the reasoning behind the logical steps of the Dirac–Bergmann algorithm,
but avoid general proofs. The reader is referred to references [11, 16–19] for more details.
We begin in Section II with a derivation of Lagrange’s equations for the singular Lagrangian (1.1). The general
solution is derived, and in Section III, we discuss the gauge freedom at the Lagrangian level. We begin construction
of the Hamiltonian theory in Section IV with a derivation of the primary constraints and canonical Hamiltonian.
In Section V, we introduce the primary Hamiltonian and the primary action. Section VI is devoted to a discussion
of the initial value problem and the need to go beyond the primary Hamiltonian. In Section VII, we apply Dirac’s
consistency conditions to derive the secondary constraints and restrictions on the Lagrange multipliers. The concept
of weak equality is introduced in Section VIII, along with a formal analysis of the restrictions on the Lagrange
multipliers. The total Hamiltonian is computed in Section IX, and in Section X we sort the constraints into first and
second class. The first class Hamiltonian and gauge generators are identified in Section XI, where we also introduce the
Dirac conjecture. In Sections XII and XIII, we define the extended Hamiltonian and extended action. Dirac brackets
are used in Section XIV to eliminate the second class constraints, which yields a partially reduced Hamiltonian.
The corresponding partially reduced action is derived in Section XV, and in Section XVI we eliminate the momenta
to obtain a partially reduced Lagrangian. Gauge conditions are introduced in Section XVII and Dirac brackets are
used to eliminate the constraints and gauge conditions. This yields a fully reduced Hamiltonian. The fully reduced
action is derived in Section XVIII. Finally, Section XIX contains a summary of the Dirac–Bergmann algorithm and
a discussion of the Einstein–Hilbert action for general relativity.

II. LAGRANGIAN ANALYSIS

The action is the integral of the Lagrangian (1.1):

Z T
S[q] = dt L(q, q̇) . (2.1)
0

The notation S[q] indicates that S is a functional of the complete set of coordinates, qi = {q1 , q2 , q3 , q4 }. The equations
of motion are obtained by extremizing the action. For this example, we are not concerned with boundary conditions

1 This example does not cover every contingency. In particular, it does not include redundant constraints [11].
 3

and integrate by parts freely. Lagrange’s equations are

δS
0= = q̇2 + q̇3 + 2q1 + q2 + q4 , (2.2a)
δq1
δS
0= = −2q̈2 − q̈3 + q̈4 − q̇1 + q1 + 2q4 , (2.2b)
δq2
δS
0= = −q̈2 − q̈3 − q̇1 , (2.2c)
δq3
δS
0= = q̈2 − q̈4 + q1 + 2q2 . (2.2d)
δq4
We can rewrite these as follows. First, add Equations (2.2b) and (2.2d), then subtract Equation (2.2c). This gives
q1 + q2 + q4 = 0 . (2.3a)
Next, subtract this result from Equation (2.2a) to obtain
q̇2 + q̇3 + q1 = 0 . (2.3b)
The time derivative of this equation yields Equation (2.2c). Finally, we find the result
q̈4 − q̈2 = q2 − q4 . (2.3c)
by solving Equation (2.3a) for q1 and using the equation of motion (2.2d).
Equations (2.3) are equivalent to Lagrange’s Equations (2.2). In particular, Equation (2.2a) is the sum of Equa-
tions (2.3a) and (2.3b); Equation (2.2b) is the sum of Equations (2.3a) and (2.3c) with the time derivative of (2.3b)
subtracted; Equation (2.2c) is the negative of the time derivative of (2.3b); Equation (2.2d) is obtained by subtracting
(2.3c) from (2.3a).
The equations of motion for this simple linear system are easily solved. Note that the combination q4 − q2 is
determined by (2.3c) along with initial or boundary data; thus, we have
q4 (t) − q2 (t) = A sin t + B cos t , (2.4)
where A and B are constants. Now, Equation (2.3a) gives
q1 (t) + 2q2 (t) = −A sin t − B cos t . (2.5)
If we knew q2 (t), we could solve the previous two equations for q1 (t) and q4 (t), then integrate Equation (2.3b) to
obtain q3 (t). Clearly, we do not have enough information to fully determine each of the q’s as functions of time. One
of the q’s must remain undetermined. For example, let us choose q2 arbitrarily by setting q2 (t) = −Ψ(t)/2 for some
function Ψ(t). We can then use the equations above to solve for q1 , q3 and q4 :
q1 (t) = −A sin t − B cos t + Ψ(t) , (2.6a)
q2 (t) = −Ψ(t)/2 , (2.6b)
Z t
q3 (t) = −A cos t + B sin t + Ψ(t)/2 − ds Ψ(s) + C , (2.6c)
0
q4 (t) = A sin t + B cos t − Ψ(t)/2 , (2.6d)
where C is an integration constant. This is the general solution of the equations of motion.

III. GAUGE INVARIANCE

The undetermined function Ψ(t) that appears in the general solution (2.6) can be freely specified. This is the gauge
freedom of the theory. We can express the gauge freedom in another way: the Lagrangian (1.1) and the equations of
motion (either (2.2) or (2.3)) are invariant under the replacements
q1 (t) → q1 (t) + Ψ(t) , (3.1a)
q2 (t) → q2 (t) − Ψ(t)/2 , (3.1b)
Z t
q3 (t) → q3 (t) + Ψ(t)/2 − ds Ψ(s) , (3.1c)
0
q4 (t) → q4 (t) − Ψ(t)/2 , (3.1d)
 4

where Ψ(t) is an arbitrary function of time.

Although each configuration of the system (that is, each set of q values) corresponds to a specific physical state of
the system, the converse is not true. Because of the gauge freedom, there are many sets of q’s that describe one and
the same physical state.
Let us examine the gauge freedom more closely, in anticipation of the Hamiltonian description of evolution. To begin,
choose the gauge Ψ(t) = 0 and consider the general solution (2.6). This solution describes the evolution of the system
from initial data

q1 (0) = −B , q̇1 (0) = −A , (3.2a)

q2 (0) = 0 , q̇2 (0) = 0 , (3.2b)
q3 (0) = C − A , q̇3 (0) = B , (3.2c)
q4 (0) = B , q̇4 (0) = A . (3.2d)

The configuration at some arbitrary final time t = T is

q1 (T ) = −A sin T − B cos T , (3.3a)

q2 (T ) = 0 , (3.3b)
q3 (T ) = C − A cos T + B sin T , (3.3c)
q4 (T ) = A sin T + B cos T . (3.3d)

This configuration corresponds to a particular state of the physical system.

We can choose a different gauge in Equation (2.6). As long as the new gauge satisfies Ψ(0) = Ψ̇(0) = 0, the solution
will describe evolution from the same initial data (3.2). For example, with

(π 2 − 4)ǫ πǫ
Ψ(t) = [cos(πt/T ) − 1] + [πt/T − sin(πt/T )] (3.4)
8 4
where ǫ = const, the configuration at t = T is

q1 (T ) = −A sin T − B cos T + ǫ , (3.5a)

q2 (T ) = −ǫ/2 , (3.5b)
q3 (T ) = C − A cos T + B sin T + ǫ/2 , (3.5c)
q4 (T ) = A sin T + B cos T − ǫ/2 . (3.5d)

The configurations (3.3) and (3.5) represent the same physical state of the system, since they evolve from the same
initial data.
We can express this result more compactly as

δq1 =ǫ, (3.6a)

δq2 = −ǫ/2 , (3.6b)
δq3 = ǫ/2 , (3.6c)
δq4 = −ǫ/2 . (3.6d)

Here, δqi denotes the change in qi at the generic time T , due to the change in gauge function Ψ(t).
Here is another example. With
π2 ǫ πǫ
Ψ(t) = [cos(πt/T ) − 1] + [πt/T − sin(πt/T )] (3.7)
4T 2T
we obtain a configuration that differs from Equation (3.3) by

δq1 =0, (3.8a)

δq2 =0, (3.8b)
δq3 =ǫ, (3.8c)
δq4 =0. (3.8d)

This configuration is also evolved from the initial data (3.2), and represents the same physical state as the configura-
tions (3.3) and (3.5).
 5

Although the gauge transformation (3.1) contains a single arbitrary function of time, the gauge invariance naturally
splits into two types. The first consists of variations subject to δq2 = −δq3 = δq4 = −δq1 /2. The second consists of
arbitrary variations in δq3 , with δq1 = δq2 = δq4 = 0. This apparent “doubling” of the gauge freedom arises because
the solution (3.1c) for q3 (t) (unlike the other variables) includes the integral of Ψ(t). There is enough freedom of
choice in Ψ(t) to allow variations in q3 that are independent of the variations among the other variables. Both types
of gauge transformations leave the physical state of the system unchanged.
The consequences of gauge invariance are most clearly expressed in the Hamiltonian formalism. The extended
Hamiltonian defined in Section XII includes phase space generators for both types of gauge transformations.

IV. PRIMARY CONSTRAINTS AND THE CANONICAL HAMILTONIAN

We now begin construction of the Hamiltonian description of the system. The conjugate momenta are defined as
usual by pi = ∂L/∂ q̇i . For the Lagrangian (1.1), we have

p1 =0 (4.1a)
p2 = 2q̇2 + q̇3 − q̇4 + q1 (4.1b)
p3 = q̇2 + q̇3 + q1 (4.1c)
p4 = q̇4 − q̇2 (4.1d)

Because the Lagrangian is singular, the matrix of second derivatives ∂ 2 L/∂ q̇i ∂ q̇j is not invertible and we cannot solve
Equations (4.1) for the velocities as functions of the coordinates and momenta. The definitions (4.1) yield two primary
constraints,

φ1 ≡ p1 = 0 , (4.2a)
φ2 ≡ p2 − p3 + p4 = 0 , (4.2b)

that restrict the phase space variables pi , qi . We will denote these constraints collectively by φa , where a = 1, 2. Note
that in this simple example, the primary constraints are independent of the q’s.
There is freedom in choosing how the constraints are written. For example, we could replace the φ’s above with

φ̃1 ≡ p1 , (4.3a)
φ̃2 ≡ −p1 + p2 − p3 + p4 . (4.3b)

In fact, any choice for the constraints is allowed, as long as they follow from the definitions pi = ∂L/∂ q̇i and satisfy the
regularity conditions. These conditions state that, roughly speaking, the constraints should have nonzero gradients
on the constraint surface. More precisely, the Jacobian matrix formed from the derivatives of the constraints with
respect to the p’s and q’s should have maximal rank on the constraint subspace [10, 11]. For both choices, (4.2) and
(4.3), the rank of the Jacobian matrix is 2. On the other hand, the set

φ̄1 ≡ p21 , (4.4a)

φ̄2 ≡ p2 − p3 + p4 , (4.4b)

is not permissible because the gradient of φ̄1 vanishes on the constraint surface where p1 = 0. Correspondingly,
the rank of the Jacobian matrix is less than 2 on the constraint surface.
The next step in constructing the Hamiltonian formalism is to compute the canonical Hamiltonian. The canonical
Hamiltonian HC is defined from the usual prescription by writing pi q̇i − L(q, q̇) in terms of p’s and q’s. Although we
cannot solve for all of the q̇’s in terms of p’s and q’s, it can be shown that the combination pi q̇i − L(q, q̇) depends only
on the phase space variables [10, 11]. For our example problem, the canonical Hamiltonian is

1h 2 i
HC = p3 + p24 − 2p3 q1 − (q1 + 2q2 )(q1 + 2q4 ) . (4.5)
2

Note that HC is ambiguous. For example, we could use the primary constraint (4.2b) to replace the term −2p3 q1
with −2(p2 + p4 )q1 .
 6

V. PRIMARY HAMILTONIAN AND THE PRIMARY ACTION

The primary Hamiltonian HP is obtained from the canonical Hamiltonian HC by adding the primary constraints
with Lagrange multipliers,

HP = HC + λa φa , (5.1)

where a sum over the repeated index a is implied. The primary action is built from the primary Hamiltonian in the
RT
usual way: SP [q, p, λ] = 0 dt{pi q̇i − HP }. Explicitly, we have
T 
1h
Z i
SP [q, p, λ] = dt pi q̇i − p23 + p24 − 2p3 q1 − (q1 + 2q2 )(q1 + 2q4 )
0 2

− λ1 p1 − λ2 (p2 − p3 + p4 ) . (5.2)

The primary action is a functional of the complete set of phase space coordinates, qi , pi , as well as the Lagrange
multipliers λ1 and λ2 .
The equations of motion are obtained by extremizing the primary action SP . Extremization with respect to the
momenta pi gives

q̇1 = λ1 , (5.3a)
q̇2 = λ2 , (5.3b)
q̇3 = p3 − q1 − λ2 , (5.3c)
q̇4 = p4 + λ2 , (5.3d)

while extremization with respect to the coordinates qi yields

ṗ1 = p3 + q1 + q2 + q4 , (5.3e)
ṗ2 = q1 + 2q4 , (5.3f)
ṗ3 =0, (5.3g)
ṗ4 = q1 + 2q2 . (5.3h)

Extremizing the action SP with respect to the Lagrange multipliers gives the constraints,

φ1 ≡ p1 = 0 , (5.3i)
φ2 ≡ p2 − p3 + p4 = 0 . (5.3j)

These equations of motion (5.3) are equivalent to Lagrange’s Equations (2.2). To show this, we first solve Equa-
tions (5.3c,d,i,j) for the momenta to obtain

p1 =0, (5.4a)
p2 = q̇3 − q̇4 + q1 + 2λ2 , (5.4b)
p3 = q̇3 + q1 + λ2 , (5.4c)
p4 = q̇4 − λ2 . (5.4d)

Using these results along with Equations (5.3a,b) for the Lagrange multipliers, we find that Equations (5.3e,f,g,h)
agree precisely with Lagrange’s Equations (2.2).

VI. HAMILTON’S EQUATIONS AND THE INITIAL VALUE PROBLEM

At this point one might ask whether the task of expressing the singular system (1.1) in Hamiltonian form is complete.
After all, the primary action (5.2) provides the correct equations of motion for the phase space variables qi and pi .
In fact, we can obtain the time evolution of any phase space function F from Ḟ = [F, HP ] where HP is the primary
Hamiltonian and [ · , · ] denotes Poisson brackets. Hamilton’s equations for the coordinates and momenta, q̇i = [qi , HP ]
and ṗi = [pi , HP ], coincide with Equations (5.3a–h).
 7

Our task of expressing the singular system in Hamiltonian form is not yet complete because we still need to interpret
Hamilton’s Equations (5.3a–h) as an initial value problem. That is, Hamilton’s equations should determine the future
history of the system solely from initial data. In contrast, the primary action (5.2) defines a boundary value problem
in which the configuration variables, the q’s, are specified at initial and final times.
The key difference between the equations of motion δSP = 0 and Hamilton’s equations Ḟ = [F, HP ] is that
the former include the primary constraints, Equations (5.3i,j), whereas the latter do not. Thus, the phase space
trajectories that extremize the action SP must lie entirely in the primary constraint surface. (The primary constraint
“surface” is the subspace of phase space that satisfies the primary constraints.) In contrast, the trajectories obtained
from Hamiltonian evolution Ḟ = [F, HP ] are defined throughout the entire phase space. Note that we cannot simply
append the primary constraint equations to Hamilton’s equations, because in that case the complete system would
not be in Hamiltonian form.
Of course, the physically allowed phase space trajectories must satisfy the primary constraints. With an initial
value interpretation of Hamilton’s equations, we can try to enforce the primary constraints with appropriate choices
of initial data and Lagrange multipliers. In particular, we can choose initial data that lie on the primary constraint
surface φ1 (0) ≡ p1 (0) = 0 and φ2 (0) ≡ p2 (0) − p3 (0) + p4 (0) = 0. However, this is not enough, because the primary
constraints are not necessarily satisfied at later times as the system evolves into the future.
We can describe the situation as follows. The trajectories that extremize the action SP , the physical trajectories,
do not necessarily fill the entire primary constraint surface. Instead they might span only a subspace of the primary
constraint surface. If the initial data lie in the primary constraint surface but outside the subspace of physical
trajectories, then the primary constraints will not be preserved as the data are evolved.
How should the initial data and Lagrange multipliers be restricted such that the primary constraints hold throughout
the evolution? The primary constraints will hold for all time if they hold initially and their time derivatives (to all
orders) also vanish initially. In the general case, this leads to a hierarchy of restrictions on the initial data in the form
of secondary, tertiary, etc., constraints.2 It can also lead to restrictions on the Lagrange multipliers.
The higher order (secondary, tertiary, etc.) constraints and restrictions on the Lagrange multipliers are not new—
imposing them does not change the content or predictions of the physical theory. This is because the higher-order
constraints and restrictions on the Lagrange multipliers are direct consequences of the equations of motion (5.3) that
follow from the primary action (5.2). They are simply “hidden” in those equations. The process of identifying the
higher-order constraints and restrictions on Lagrange multipliers reveals these hidden conditions.

VII. CONSISTENCY CONDITIONS, SECONDARY CONSTRAINTS AND RESTRICTIONS ON THE

LAGRANGE MULTIPLIERS

We can ensure that the primary constraints hold for all time by applying Dirac’s consistency conditions [10]. Begin
by computing the time derivatives of the primary constraints with the primary Hamiltonian, φ̇a = [φa , HP ]. Now set
these equal to zero:
[φa , HP ] = 0 . (7.1)
For each value of the index a, there are three possibilities.3 First, [φa , HP ] might vanish on the constraint surface
φa = 0, so that the consistency condition (7.1) reduces to the identity 0 = 0. Second, [φa , HP ] could be a (non-
constant) phase space function that is independent of the Lagrange multipliers. In this case, Equation (7.1) is a
secondary constraint. Finally, [φa , HP ] might depend on the Lagrange multipliers. Then, Equation (7.1) fixes one of
the Lagrange multipliers in terms of the phase space variables and the other Lagrange multipliers.
The secondary constraints that arise from this process must themselves satisfy the consistency conditions. This can
lead to tertiary constraints and more restrictions on the Lagrange multipliers. In turn, the tertiary constraints can
lead to quaternary constraints, and so forth. We must continue to apply the consistency conditions until the process
naturally stops.
For our example, the primary constraints are
φ1 ≡ p1 = 0 , (7.2a)
φ2 ≡ p2 − p3 + p4 = 0 , (7.2b)

2 Some authors use the term “secondary constraints” to refer to all higher–order constraints—that is, all constraints beyond the primary
level.
3 This assumes Lagrange’s equations are self–consistent. Otherwise, the consistency conditions could lead to a contradiction such as 1 = 0.
 8

and their time derivatives are

φ̇1 = [φ1 , HP ] = p3 + q1 + q2 + q4 , (7.3a)

φ̇2 = [φ2 , HP ] = 2(q1 + q2 + q4 ) . (7.3b)

Thus, we find the secondary constraints

ψ1 ≡ p3 + q1 + q2 + q4 = 0 , (7.4a)
ψ2 ≡ 2(q1 + q2 + q4 ) = 0 . (7.4b)

These will be denoted collectively by ψa .

Applying the consistency conditions to the secondary constraints gives

ψ̇1 = [ψ1 , HP ] = p4 + λ1 + 2λ2 = 0 , (7.5a)

ψ̇2 = [ψ2 , HP ] = 2(p4 + λ1 + 2λ2 ) = 0 . (7.5b)

These equations restrict the Lagrange multipliers to satisfy

p4 + λ1 + 2λ2 = 0 . (7.6)

The process has now terminated. In this example, there are no tertiary or higher-order constraints.
Recall from the previous section that our goal was to restrict the initial data and Lagrange multipliers such that the
primary constraints vanish for all times under the Hamiltonian evolution defined by HP . We achieve this by imposing
the primary constraints at the initial time,

φ1 (0) ≡ p1 (0) = 0 , (7.7a)

φ2 (0) ≡ p2 (0) − p3 (0) + p4 (0) = 0 , (7.7b)

the secondary constraints at the initial time,

ψ1 (0) ≡ p3 (0) + q1 (0) + q2 (0) + q4 (0) = 0 , (7.8a)

ψ2 (0) ≡ 2[q1 (0) + q2 (0) + q4 (0)] = 0 , (7.8b)

and restricting the Lagrange multipliers to satisfy Equation (7.6) for all time t.
Let us review the reasoning. From Equations (7.5), the restriction (7.6) on the Lagrange multipliers tells us that
ψ̇a = 0 for all time. By Equations (7.8), ψa vanishes initially, so we see that ψa must vanish for all time. Now we
use Equations (7.3) to conclude that φ̇a must vanish for all time. Since φa vanishes initially, by Equations (7.7), it
follows that the primary constraints φa = 0 must hold for all time t.

VIII. WEAK EQUALITY AND LAGRANGE MULTIPLIER ANALYSIS

It will be useful to follow the general Dirac–Bergmann algorithm closely and carry out a formal analysis of the
restriction (7.6) on the Lagrange multipliers [10, 11]. We begin with the concept of weak equality.
Let CA denote the complete set of (primary, secondary, tertiary, etc.) constraints. For our example,

φ1 p1
   
 φ   p2 − p3 + p4 
CA ≡  2  =  (8.1)
ψ1 p3 + q1 + q2 + q4 
ψ2 2(q1 + q2 + q4 )

where the index A runs from 1 to 4.

Two phase space functions F and G are weakly equal if they are equal when the (primary, secondary, tertiary, etc.)
constraints hold. In other words, F and G are weakly equal if they coincide on the constraint surface, the subspace
of phase space defined by CA = 0. Weak equality is written as F ≈ G.
Functions F and G are strongly equal if they agree throughout phase space. Strong equality is written as F = G.
 9

Now we turn to the formal analysis of the restriction (7.6) on the Lagrange multipliers. This restriction can be
expressed as the weak equality C˙A ≈ 0. From Equations (7.3) and (7.5), we have
p3 + q1 + q2 + q4 0
   
2(q + q + q ) 0
C˙A = [C, HP ] =  1 2 4
≈ (8.2)
  
p4 + λ1 + 2λ2  0
2(p4 + λ2 + 2λ2 ) 0
which simplifies to
0 0 0   0
     
 0  0 0 λ1 0
 p  + 1 ≈  . (8.3)
4 2 λ2 0
2p4 2 4 0
This is a system of inhomogeneous linear equations for the Lagrange multipliers. A particular solution is
   
λ1 0
= , (8.4)
λ2 particular −p4 /2

and the homogeneous solutions are

   
λ1 1
= λ, (8.5)
λ2 homogeneous
−1/2

where λ is arbitrary. The general solution is the sum of particular and homogeneous solutions:
   
λ1 λ
= . (8.6)
λ2 general −(λ + p4 )/2

Thus, the restriction (7.6) on the Lagrange multipliers yields λ1 = λ and λ2 = −(λ + p4 )/2, where λ is an arbitrary
function of time.

IX. TOTAL HAMILTONIAN

The total Hamiltonian HT is obtained from the primary Hamiltonian HP by inserting the general solution for the
Lagrange multipliers:

H T = HP = HC + λφ1 − (λ + p4 )φ2 /2
λ1 =λ, λ2 =−(λ+p4 )/2
1 2 
= p3 − p2 p4 + p3 p4 − 2p3 q1 − (q1 + 2q2 )(q1 + 2q4 )
2
+ λ(p1 − p2 /2 + p3 /2 − p4 /2) . (9.1)

Physical phase space trajectories are defined by the total Hamiltonian as the weak equality Ḟ ≈ [F, HT ], with initial
data that satisfy the complete set of constraints, CA = 0.
Hamilton’s equations for the total Hamiltonian HT are
q̇1 ≈λ, (9.2a)
q̇2 ≈ −p4 /2 − λ/2 , (9.2b)
q̇3 ≈ p3 + p4 /2 − q1 + λ/2 , (9.2c)
q̇4 ≈ p3 /2 − p2 /2 − λ/2 , (9.2d)
and
ṗ1 ≈ p3 + q1 + q2 + q4 , (9.2e)
ṗ2 ≈ q1 + 2q4 , (9.2f)
ṗ3 ≈0, (9.2g)
ṗ4 ≈ q1 + 2q2 . (9.2h)
 10

Since these are weak equalities, we can use the constraints to simplify the results. Observe that the constraints
CA = 0 imply p1 = p3 = 0, p2 + p4 = 0 and q1 + q2 + q4 = 0. Therefore, we can set p1 and p3 to zero, replace p4 with
−p2 , and replace q4 with −q1 − q2 . Then Equations (9.2d,e,g,h) are either redundant or vacuous, and the remaining
equations are

q̇1 ≈λ, (9.3a)

q̇2 ≈ p2 /2 − λ/2 , (9.3b)
q̇3 ≈ −p2 /2 − q1 + λ/2 , (9.3c)
ṗ2 ≈ −q1 − 2q2 . (9.3d)

These equations, along with the constraints CA = 0, give a complete description of the physical system.
Let us check the results. The Lagrange multiplier λ can be eliminated from Equations (9.3a,b) to give q̇1 + 2q̇2 ≈ p2 .
Now differentiate this equation and eliminate ṗ2 with Equation (9.3d) to obtain q̈1 + 2q̈2 ≈ −q1 − 2q2 . The constraints
allow us to set q1 ≈ −q2 − q4 , which gives

q̈4 − q̈2 ≈ q2 − q4 . (9.4)

This is Equation (2.3c), which follows directly from Lagrange’s equations. The result (2.3a) from Lagrange’s equations
is simply the secondary constraint C4 ≡ 2(q1 + q2 + q4 ) = 0. Finally, the result (2.3b) is obtained by summing
Equations (9.3b) and (9.3c).
Recall that Equations (2.3a–c) are equivalent to Lagrange’s equations. Thus, we have verified that Hamilton’s
Equations (9.2), along with the primary and secondary constraints CA = 0, are equivalent to Lagrange’s equations.

X. FIRST AND SECOND CLASS CONSTRAINTS

A first class function F is a phase space function that has weakly vanishing Poisson brackets with all primary and
secondary constraints:

[F, CA ] ≈ 0 ⇐⇒ F is first class. (10.1)

It can be shown that the Poisson bracket of any two first class functions is itself a first class function [10, 11].
The constraints themselves can be first class; constraints that are not first class are called second class. The
constraints are separated into first and second class by examining the matrix of Poisson brackets:

0 0 −1 −2
 
0 0 −2 −4
[CA , CB ] =  . (10.2)
1 2 0 0
2 4 0 0

The rank of this 4 × 4 matrix is 2, and its nullity is 4 − 2 = 2. It follows that there are 2 independent eigenvectors with
eigenvalues equal to zero; for example, uA = (1, −1/2, 0, 0) and v A = (0, 0, 1, −1/2). Then there are two independent
combinations of constraints that are first class, namely uA CA and v A CA . (A sum over the repeated index A is implied.)
The first class constraints are
(f c)
C1 ≡ φ1 − φ2 /2 = p1 − p2 /2 + p3 /2 − p4 /2 , (10.3a)
(f c)
C2 ≡ ψ1 − ψ2 /2 = p3 . (10.3b)
(f c) (f c)
One can check that the first class conditions [C1 , CB ] = 0 and [C2 , CB ] = 0 hold. The most general first class
(f c) (f c)
constraint is a linear combination of C1 and C2 .
There are two remaining linear combinations of constraints, which we take to be
(sc)
C1 ≡ (φ1 + φ2 )/3 = (p1 + p2 − p3 + p4 )/3 , (10.4a)
(sc)
C2 ≡ (ψ1 + ψ2 )/3 = p3 /3 + q1 + q2 + q4 . (10.4b)

These are the second class constraints. They have nonvanishing Poisson brackets with each other,
(sc) (sc)
[C2 , C1 ]=1. (10.5)
 11

(sc) (sc) (f c) (f c)
The most general second class constraint is a linear combination of C1 , C2 , C1 and C2 , with nonzero coefficients
(sc) (sc)
on one or both of C1 and C2 .
The splitting of constraints into first and second class is independent of the splitting into primary and secondary.
In this example, the first class constraints are mixtures of primary and secondary constraints. Likewise, the second
class constraints are mixtures of primary and secondary constraints.

XI. FIRST CLASS HAMILTONIAN, GAUGE GENERATORS AND THE DIRAC CONJECTURE

The total Hamiltonian (9.1) includes the product of an arbitrary Lagrange multiplier λ with the first class con-
(f c) (f c)
straint C1 ≡ φ1 − φ2 /2. We refer to C1 as a primary first class constraint, since it is constructed entirely from
primary constraints.
If we remove the primary first class constraint from the total Hamiltonian, what remains is the first class Hamiltonian
Hf c . That is, the total Hamiltonian can be written as
(f c)
HT = Hf c + λC1 , (11.1)

where
1h 2 i
Hf c = p3 − p2 p4 + p3 p4 − 2p3 q1 − (q1 + 2q2 )(q1 + 2q4 ) (11.2)
2
is the first class Hamiltonian. A common notation for Hf c , the notation used by Dirac [10], is H ′ .
We can check directly that the first class Hamiltonian (11.2) is a first class function. However, this is not necessary,
because we know that the constraints are preserved under the time evolution defined by HT . That is, C˙A = [CA , HT ] ≈
(f c)
0. Thus, the total Hamiltonian must be first class, [HT , CA ] ≈ 0. Of course, the primary first class constraint C1 is
first class. It then follows from the definition (11.1) that Hf c must also be a first class function.
The splitting (11.1) of the total Hamiltonian into the first class Hamiltonian and the primary first class constraint
is not special to our example. This splitting will occur for any constrained Hamiltonian system [10, 11]. In general,
HT will include the products of every primary first class constraint with an arbitrary multiplier.
Primary first class constraints generate gauge transformations. Consider the change in a phase space function F
(f c)
generated by the primary first class constraint C1 ,
(f c)
δF = δǫ[F, C1 ]. (11.3)

This transformation does not change the physical state of the system. We can see this by considering F to be evaluated
as a function of the q’s and p’s at some particular time t. At an infinitesimally later time t + δt, this function becomes
F (t + δt) = F (t) + [F, HT ]δt. In terms of the first class Hamiltonian, we have
n o
(f c)
F (t + δt) = F (t) + [F, Hf c ] + λ[F, C1 ] δt . (11.4)

The Lagrange multiplier is arbitrary, so we can make a different choice during the time interval from t to t + δt, say,
λ̃. Then, the function F at time t + δt will be
n o
(f c)
F̃ (t + δt) = F (t) + [F, Hf c ] + λ̃[F, C1 ] δt . (11.5)

The physical state of the system at t + δt should not depend on our choice of Lagrange multiplier, so F (t + δt) and
F̃ (t + δt) must represent the same physical state. The result (11.3) is obtained by subtracting Equation (11.4) from
Equation (11.5) and defining δF ≡ F̃ − F and δǫ ≡ (λ̃ − λ)δt.
(f c)
For the phase space coordinates, the gauge transformation generated by the primary first class constraint C1 is

δq1 = δǫ , (11.6a)
δq2 = −δǫ/2 , (11.6b)
δq3 = δǫ/2 , (11.6c)
δq4 = −δǫ/2 . (11.6d)

The transformations of the p’s all vanish. This result agrees with the gauge transformation from Equation (3.6),
with the change of notation ǫ ↔ δǫ. Here, we denote the gauge parameter by δǫ because the transformation is
 12

infinitesimal; in Equation (3.6), we used ǫ because the transformation was finite. It is clear that the infinitesimal
transformation (11.6) can be iterated to obtain the finite transformation (3.6).
In general, a gauge transformation is defined as a transformation δF = δǫ[F, G] that does not alter the physical state
of the system. The function G is the gauge generator. We have seen that the primary first class constraints generate
gauge transformations. However, not all gauge transformations are generated by primary first class constraints.
In fact, it can be shown [10, 11] that the Poisson bracket between any primary first class constraint and the first class
Hamiltonian is itself a first class constraint that generates a gauge transformation.4
(f c)
For our example problem, the Poisson bracket of the primary first class constraint C1 and the first class Hamil-
tonian Hf c is
(f c)
[C1 , Hf c ] = p 3 . (11.7)
(f c)
This is the secondary first class constraint, C2 ≡ ψ1 − ψ2 /2 = p3 . Thus, we see that in this example both the
primary and secondary first class constraints are generators of gauge transformations. Explicitly, the transformation
(f c)
δF = δǫ[F, C2 ] is
δq1 =0, (11.8a)
δq2 =0, (11.8b)
δq3 = δǫ , (11.8c)
δq4 =0, (11.8d)
with the transformations of the p’s all vanishing. We can iterate this infinitesimal gauge transformation (11.8) to
obtain the finite transformation (3.8).
The “doubling” of the gauge freedom identified in Section III appears quite naturally in the Hamiltonian formalism.
(f c) (f c)
The two types of gauge transformation are generated by the two first class constraints, C1 and C2 .
The Dirac conjecture [10] says that all first class constraints (whether they are primary, secondary, etc., or a
combination of primary, secondary, etc.) generate gauge transformations. This conjecture does not hold as a general
theorem—there are known examples in which the transformation generated by a secondary first class constraint does
not coincide with any invariance of the original Lagrangian system.5 Nevertheless, the Dirac conjecture is usually
taken as an assumption. It appears that in practice, for systems of physical interest, all first class constraints generate
gauge transformations.

XII. EXTENDED HAMILTONIAN

The Dirac conjecture tells us that all first class constraints generate gauge transformations and should be treated
(f c)
on an equal footing. The extended Hamiltonian HE is defined by adding all first class constraints Ca with Lagrange
multipliers γa to the first class Hamiltonian:
HE = Hf c + γa Ca(f c) . (12.1)
(A sum over the index a is implied.)
For our example, the extended Hamiltonian is
1h i
HE = p23 − p2 p4 + p3 p4 − 2p3 q1 − (q1 + 2q2 )(q1 + 2q4 )
2
+ γ1 (p1 − p2 /2 + p3 /2 − p4 /2) + γ2 p3 . (12.2)
The equations of motion Ḟ ≈ [F, HE ] are
q̇1 ≈ γ1 , (12.3a)
q̇2 ≈ −p4 /2 − γ1 /2 , (12.3b)
q̇3 ≈ p3 + p4 /2 − q1 + γ1 /2 + γ2 , (12.3c)
q̇4 ≈ p3 /2 − p2 /2 − γ1 /2 , (12.3d)

4 For systems with more than one primary first class constraint, the Poisson brackets of any two primary first class constraints is also a
first class constraint that generates a gauge transformation [10, 11].
5 Counterexamples to the Dirac conjecture are discussed in Refs. [11, 20–26] and elsewhere. Proofs of the conjecture have been constructed
by adopting various simplifying assumptions [11, 27, 28]. The status of the conjecture is a subtle issue; see for example Refs. [18, 29, 30].
 13

and

ṗ1 ≈ p3 + q1 + q2 + q4 , (12.3e)
ṗ2 ≈ q1 + 2q4 , (12.3f)
ṗ3 ≈0, (12.3g)
ṗ4 ≈ q1 + 2q2 . (12.3h)

Let us compare these results to the equations of motion (9.2) obtained from the total Hamiltonian HT . There are
just two differences. The first is trivial: the Lagrange multiplier λ in Equations (9.2) has changed names to γ1 in
Equations (12.3). The second difference is significant: the equation for q̇3 has an extra term γ2 on the right-hand
side. This is a new feature of the extended Hamiltonian. It makes explicit the fact that the gauge freedom allows q3
to be changed arbitrarily, and independently, from the other variables.
We can check the equations of motion for HE following the same reasoning that was applied to the equations
of motion for HT . First, recall that the (first and second class) constraints imply p1 = p3 = 0, p4 = −p2 and
q4 = −q1 − q2 . Then Equations (12.3e) and (12.3g) are vacuous, and Equations (12.3f) and (12.3h) are redundant. It
also follows that with the constraints imposed, Equation (12.3d) is a consequence of Equations (12.3a) and (12.3b).
The remaining equations are

q̇1 ≈ γ1 , (12.4a)
q̇2 ≈ p2 /2 − γ1 /2 , (12.4b)
q̇3 ≈ −p2 /2 − q1 + γ1 /2 + γ2 , (12.4c)
ṗ2 ≈ −q1 − 2q2 . (12.4d)

These agree with Equations (9.3), apart from the change of notation λ → γ1 and the extra term γ2 on the right-hand
side of the q̇3 equation.
By eliminating γ1 , the equations of motion generated by the extended Hamiltonian HE become

q̇1 + 2q̇2 ≈ p2 , (12.5a)

q̇2 + q̇3 ≈ −q1 + γ2 , (12.5b)
ṗ2 ≈ −q1 − 2q2 . (12.5c)

If we differentiate the first equation, combine with the second, and use the constraint q1 + q2 + q4 = 0, we obtain
(q̈4 − q̈2 ) ≈ −(q4 −q2 ). This is the expected result (2.3c). In fact, the only difference between Hamilton’s equations Ḟ ≈
[F, HE ] and the results (2.3) (which are equivalent to Lagrange’s equations) is the extra term γ2 in Equation (12.5b)
above. That term does not appear in the corresponding Lagrangian Equation (2.3b).
Note that we can use the first class constraints to simplify the extended Hamiltonian HE . For example, using
(f c)
C2 = p3 , we can set p3 = 0 everywhere in Equation (12.2), except of course in the term γ2 p3 . The extended
Hamiltonian becomes
1h i
HE = −p2 p4 − (q1 + 2q2 )(q1 + 2q4 ) + γ1 (p1 − p2 /2 − p4 /2) + γ2 p3 . (12.6)
2
This amounts to replacing the Lagrange multiplier γ2 in Equation (12.2) by

γ2 → γ2 − p3 /2 − p4 /2 + q1 − γ1 /2 . (12.7)

This replacement does not change the physical content of the theory, since the Lagrange multiplier γ2 is arbitrary.

XIII. EXTENDED ACTION

The equations of motion for the extended theory can be derived from the action [11]
Z T n o
(sc) (sc)
SE [q, p, γ, σ] = dt pi q̇i − HE − σ1 C1 − σ2 C2 , (13.1)
0

(f c)
which includes the second class constraints with Lagrange multipliers σa . Recall that the first class constraints Ca
are included in the extended Hamiltonian HE with multipliers γa , so SE includes all four constraints.
 14

We can use either form of the extended Hamiltonian, Equation (12.2) or (12.6), in the extended action. Let us use
Equation (12.2). Then, the equations of motion that follow from extremizing SE with respect to the momenta pi are

q̇1 = γ1 + σ1 /3 , (13.2a)
q̇2 = −p4 /2 − γ1 /2 + σ1 /3 , (13.2b)
q̇3 = p3 + p4 /2 − q1 + γ1 /2 + γ2 − σ1 /3 + σ2 /3 , (13.2c)
q̇4 = p3 /2 − p2 /2 − γ1 /2 + σ1 /3 . (13.2d)

Extremizing SE with respect to the coordinates yields

ṗ1 = p3 + q1 + q2 + q4 − σ2 , (13.2e)
ṗ2 = q1 + 2q4 − σ2 , (13.2f)
ṗ3 =0, (13.2g)
ṗ4 = q1 + 2q2 − σ2 , (13.2h)

and the constraints

(f c)
C1 ≡ p1 − p2 /2 + p3 /2 − p4 /2 = 0 , (13.2i)
(f c)
C2 ≡ p3 = 0 , (13.2j)
(sc)
C1 ≡ (p1 + p2 − p3 + p4 )/3 = 0 , (13.2k)
(sc)
C2 ≡ p3 /3 + q1 + q2 + q4 = 0 , (13.2l)

follow from extremizing SE with respect to the Lagrange multipliers γa and σa .

Let us check these equations of motion. The constraints imply p1 = p3 = 0, p4 = −p2 and q1 + q2 + q4 = 0. Then the
equation of motion (13.2e) gives σ2 ≈ 0. The constraints also imply q̇1 + q̇2 + q̇4 = 0. The sum of Equations (13.2a),
(13.2b) and (13.2d) then yields σ1 ≈ 0. Now, if we set σ1 and σ2 to zero, the equations of motion (13.2a–h) agree
precisely with Hamilton’s Equations (12.3) for the extended Hamiltonian HE . In Section XII, we showed that the
equations generated by HE agree with Lagrange’s equations apart from the extra term γ2 in the equation for q̇3 . This
term extends the original Lagrangian theory by making explicit the fact that the gauge freedom allows for independent
transformations of q3 .
Finally, we note that the extended action is invariant under the transformation defined by
(f c) (f c)
δF = ǫ1 [F, C1 ] + ǫ2 [F, C2 ] (13.3a)

for the phase space variables and

δγ1 = ǫ̇1 , (13.3b)

δγ2 = ǫ̇2 + ǫ1 , (13.3c)
δσ1 =0, (13.3d)
δσ2 =0, (13.3e)

for the Lagrange multipliers. Here, the gauge parameters ǫ1 and ǫ2 are functions of time. These equations express
the gauge invariance at the level of the action SE .

XIV. DIRAC BRACKETS AND THE PARTIALLY REDUCED HAMILTONIAN

We now return to the evolution defined by the extended Hamiltonian HE of Equation (12.2), and Hamilton’s
Equations (12.3). To obtain a physically allowed trajectory, we must choose initial data that satisfy the four constraints
(f c) (sc)
Ca = 0 and Ca = 0. Apart from restricting the initial data, the second class constraints play no role in the
formalism. It would be convenient if we could restrict the variables from the outset such that the second class
(sc) (sc)
constraints are automatically satisfied. For example, we could use C1 = 0 and C2 = 0 from Equations (13.2k,l) to
replace q1 with −q2 − q4 − p3 /3 and replace p2 with −p1 + p3 − p4 .
We are not allowed to apply the second class constraints in this way. For example, consider the Poisson brackets
[q1 , p1 ] = 1. If we were to replace q1 with −q2 − q4 − p3 /3, we would find a different answer: [−q2 − q4 − p3 /3, p1 ] = 0.
The second class constraints cannot be imposed before Poisson brackets are computed.
 15

Dirac devised a way to allow the second class constraints to be imposed from the outset by modifying the Poisson
brackets [10]. The result is the Dirac brackets.
To construct Dirac brackets, we first compute the matrix of Poisson brackets among the second class constraints:
 
(sc) 0 −1
Mab ≡ [Ca(sc) , Cb ] = . (14.1)
1 0

Let
 
0 1
M ab = (14.2)
−1 0

denote the inverse of Mab . Then, the Dirac brackets [F, G]∗ of two phase space functions F and G are defined by
(sc)
[F, G]∗ ≡ [F, G] − [F, Ca(sc) ]M ab [Cb , G] . (14.3)

Dirac brackets, like Poisson brackets, are antisymmetric and satisfy the Jacobi identity [10, 11].
Explicitly, the Dirac brackets among the coordinates are

0 0 1 0
 
1 0 0 1 0 
[qi , qj ]∗ =   , (14.4)
9 −1 −1 0 −1
0 0 1 0

and the Dirac brackets between the q’s and p’s are

2 −1 0 −1
 
∗ 1 −1 2 0 −1
[qi , pj ] =  (14.5)
3 1 1 3 1
−1 −1 0 2

For our example, the Dirac brackets among the momenta all vanish: [pi , pj ]∗ = 0.
There are two key properties that make Dirac brackets relevant. First, the Dirac brackets agree weakly with
Poisson brackets if one of the two functions is first class. Since the extended Hamiltonian is first class, we have
[F, HE ]∗ ≈ [F, HE ] for any F . It follows that we can write the equations of motion as

Ḟ ≈ [F, HE ]∗ , (14.6)

using Dirac brackets.

The second key property of the Dirac brackets is that they weakly vanish if one of the functions is a second
(sc)
class constraint: [F, Ca ]∗ ≈ 0. This allows us to apply the second class constraints before computing brackets.
For example, we can use either q1 or −q2 − q4 − p3 /3 to compute Dirac brackets with p1 :

[q1 , p1 ]∗ = [−q2 − q4 − p3 /3, p1 ]∗ = 2/3 . (14.7)

With Dirac brackets, the second class constraints can be treated as strong equations and imposed before computing
the equations of motion.
Let us use the second class constraints (13.2k,l) to eliminate q1 and p2 and write the extended Hamiltonian in terms
of the smaller set of variables q2 , q3 , q4 , p1 , p3 and p4 . Setting

q1 = −q2 − q4 − p3 /3 , (14.8a)
p2 = −p1 + p3 − p4 , (14.8b)

we have
1 2
HP R = 7p23 /9 + p4 + p1 p4 + 2p3 (q2 + q4 ) + (q2 − q4 )2 + γ1 (3p1 /2) + γ2 p3 .

(14.9)
2
This is the partially reduced Hamiltonian, obtained from the extended Hamiltonian by applying the second class con-
straints.
Of course, the partially reduced Hamiltonian is not unique. We could use the second class constraints to eliminate
some other pair of variables instead of q1 and p2 .
 16

The equations of motion generated by the partially reduced Hamiltonian, Ḟ ≈ [F, HR ]∗ , are
q̇2 ≈ −p1 /6 − p4 /2 − γ1 /2 , (14.10a)
q̇3 ≈ 4p3 /3 + p1 /6 + p4 /2 + q2 + q4 + γ1 /2 + γ2 , (14.10b)
q̇4 ≈ p1 /3 + p4 /2 − γ1 /2 , (14.10c)
ṗ1 ≈ 2p3 /3 , (14.10d)
ṗ3 ≈0, (14.10e)
ṗ4 ≈ −p3 /3 + q2 − q4 . (14.10f)
We can also use HP R and the Dirac brackets to compute q̇1 and ṗ2 . The results are equivalent to those obtained by
differentiating the right-hand sides of Equations (14.8) and using the equations of motion (14.10).
Let us check the equations of motion. With the second class constraints applied, the first class constraints imply
p1 = p3 = 0 and p4 = −p2 . Thus, Equations (14.10d) and (14.10e) are vacuous and the remaining equations become
q̇2 ≈ p2 /2 − γ1 /2 , (14.11a)
q̇3 ≈ −p2 /2 + q2 + q4 + γ1 /2 + γ2 , (14.11b)
q̇4 ≈ −p2 /2 − γ1 /2 , (14.11c)
ṗ2 ≈ −q2 + q4 , (14.11d)
Compare these to the independent Equations (12.4) that follow from the extended Hamiltonian. Equations (12.4b,c,d)
agree with Equations (14.11a,b,d) once we use q1 = −q2 − q4 . The final Equation (12.4a) is obtained by differentiating
q1 = −q2 − q3 in time and using Equations (14.11a) and (14.11b).

XV. PARTIALLY REDUCED ACTION

The partially reduced equations of motion (14.10) can be obtained from the extended action SE by eliminating the
superfluous variables. Note that the equations of motion obtained by varying SE with respect to p2 , q1 , σ1 and σ2
are Equations (13.2b,e,k,l), respectively. We can eliminate these variables by solving these equations and substituting
the results into the action.6 The results are:
q1 = −p3 /3 − q2 − q4 , (15.1a)
p2 = −p1 + p3 − p4 , (15.1b)
σ1 = 3q̇2 + 3p4 /2 + 3γ1 /2 , (15.1c)
σ2 = −ṗ1 + 2p3 /3 . (15.1d)
Inserting these into the extended action (13.1), we find
Z T 
SP R [q2 , q3 , q4 , p1 , p3 , p4 , γ1 , γ2 ] = dt − p1 ṗ3 /3 + (p3 − p4 − 2p1 )q̇2
0
+ p3 q̇3 + (p4 − p1 )q̇4 − HP R . (15.2)
This is the partially reduced action.
The equations of motion obtained from varying SP R with respect to the phase space variables are
2ṗ1 − ṗ3 + ṗ4 − p3 − q2 + q4 = 0 , (15.3a)
−ṗ3 = 0 , (15.3b)
ṗ1 − ṗ4 − p3 + q2 − q4 = 0 , (15.3c)
−2q̇2 − q̇4 − ṗ3 /3 − p4 /2 − 3γ1 /2 = 0 , (15.3d)
q̇2 + q̇3 + ṗ1 /3 − 14p3 /9 − q2 − q4 − γ2 = 0 , (15.3e)
q̇2 + q̇4 − p1 /2 − p4 = 0 , (15.3f)

6 Any action can be reduced by using the equations of motion obtained by varying with respect to a subset of variables, solving those
equations for the same subset of variables, then substituting the results into the action. In general, it is not permissible to reduce an
action by using the equations obtained by varying with respect to one subset of variables but solving those equations for a different
subset of variables.
 17

and the equations obtained by varying with respect to the Lagrange multipliers γ1 and γ2 are
−3p1 /2 = 0 , (15.4a)
−p3 = 0 . (15.4b)
These are, of course, the first class constraints, reduced by using the second class constraints to eliminate q1 and p2 .
We can now solve Equation (15.3) for the time derivatives of q2 , q3 , q4 , p1 , p3 and p4 . The result coincides with the
equations of motion (14.10) obtained from the partially reduced Hamiltonian HP R and the Dirac brackets.
The partially reduced action SP R is invariant under the transformation defined by
(f c) ∗ (f c) ∗
δF = ǫ1 [F, C1 ] + ǫ2 [F, C2 ] (15.5a)
for the phase space variables and
δγ1 = ǫ̇1 , (15.5b)
δγ2 = ǫ̇2 + ǫ1 , (15.5c)
for the Lagrange multipliers. These equations express the gauge invariance of the theory at the level of the action
principle with the second class constraints eliminated.

XVI. PARTIALLY REDUCED LAGRANGIAN

It is not too difficult to find a change of variables that will bring SP R into “canonical form”. For example, let
q2 = Q1 − P2 /9 , (16.1a)
q3 = Q2 + P2 /9 , (16.1b)
q4 = Q3 − P2 /9 , (16.1c)
p1 = (−P1 + P2 − P3 )/3 , (16.1d)
p3 = P2 , (16.1e)
p4 = (−P1 + P2 + 2P3 )/3 , (16.1f)
define a new set of variables Qα , Pα for the secondary constraint surface. (The index α ranges over 1, 2 and 3.) The
partially reduced action becomes
Z T n o
SP R [Q, P, γ] = dt Pα Q̇α − HP R (16.2)
0

with
1  2
2P1 + 12P22 + 2P32 − 4P1 P2 + 5(P2 − P1 )P3 + 18P2 (Q1 + Q3 ) + 9(Q1 − Q3 )2

HP R =
18
+ γ1 (−P1 + P2 − P3 )/2 + γ2 P2 . (16.3)

The equations of motion δSP R = 0 include Q̇α = [Qα , HP R ] and Ṗα = [Pα , HP R ], where [ · , · ] are the usual Poisson
brackets. The partially reduced action (16.2) is invariant under the transformation
(f c) (f c)
δF = ǫ1 [F, C1 ] + ǫ2 [F, C2 ], (16.4a)
δγ1 = ǫ̇1 , (16.4b)
δγ2 = ǫ̇2 + ǫ1 , (16.4c)
(f c) (f c)
where C1 = (−P1 + P2 − P3 )/2 and C2 = P2 are the first class constraints.
We can use the equations of motion δSP R /δPα = 0 to eliminate the momenta Pα from the partially reduced action
(16.2). These equations are

Q̇1 = (4P1 − 4P2 − 5P3 )/18 − γ1 /2 , (16.5a)

Q̇2 = −(4P1 − 24P2 − 5P3 )/18 + Q1 + Q3 + γ1 /2 + γ2 , (16.5b)
Q̇3 = −(5P1 − 5P2 − 4P3 )/18 − γ1 /2 , (16.5c)
 18

with solutions

P1 = (−71Q̇1 + 9Q̇2 − 100Q̇3 )/10 − 9(Q1 + Q3 )/10 − 9γ1 − 9γ2 /10 , (16.6a)
P2 = 9(Q̇1 + Q̇2 − Q1 − Q3 − γ2 )/10 , (16.6b)
P3 = −10Q̇1 − 8Q̇3 − 9γ1 . (16.6c)

Inserting these results into the partially reduced action, we find

Z T 
SP R [Q, γ] = dt (−71Q̇21 + 9Q̇22 − 80Q̇23 )/20 + (9Q̇2 − 100Q̇3)Q̇1 /10
0
− 9(Q1 + Q3 )(Q̇1 + Q̇2 )/10 − (Q21 + Q23 − 38Q1 Q3 )/20
− 9(Q̇1 + Q̇3 )γ1 − 9(Q̇1 + Q̇2 − Q1 − Q3 )γ2 /10

2 2
− 9(10γ1 − γ2 )/20 . (16.7)

The partially reduced Lagrangian is the integrand of this action.

We can go one step further and eliminate the Lagrange multipliers using the equations of motion δS/δγ1 = 0 and
δS/δγ2 = 0. These equations have solutions

γ1 = −Q̇1 − Q̇3 , (16.8a)

γ2 = Q̇1 + Q̇2 − Q1 − Q3 . (16.8b)

Inserting these results back into the action yields

Z T  
1 2 1 2
SP R [Q] = dt (Q̇1 − Q̇3 ) − (Q1 − Q3 ) . (16.9)
0 2 2

This is the action for a harmonic oscillator in the variable Q1 − Q3 . Note that the coordinate transformation (16.1)
implies (Q1 − Q3 ) = (q2 − q4 ), so once again, we find that the coordinate combination q2 − q4 describes a simple
harmonic oscillator. In addition, we observe that the action (16.9) leaves the variable Q2 completely unspecified. This
(f c)
expresses the gauge freedom generated by the first class constraint C2 = P2 = p3 .

XVII. GAUGE CONDITIONS AND THE FULLY REDUCED HAMILTONIAN

Let us return to the theory described by the extended Hamiltonian, prior to the elimination of the second class con-
straints.
For our example problem, phase space is eight-dimensional. The physical trajectories fill the constraint “surface”,
which is the four-dimensional subspace where all first and second class constraints hold. Each point in the constraint
surface can be mapped into a physically equivalent state by the gauge generators, namely, the first class constraints
(f c)
Ca . Since there are two independent gauge generators, each physical state of the system corresponds to a two-
dimensional subspace of the constraint surface. The constraint surface is foliated by these two-dimensional slices,
referred to as gauge “orbits”.
We can select a single phase space point on each gauge orbit to represent the physical state. We do this by applying
gauge conditions. In particular, we will consider a canonical gauge7 which takes the form Ga (q, p) ≈ 0 with a = 1, 2.
A good canonical gauge condition must not be gauge invariant, otherwise it would allow more than one point on
the gauge orbit to represent the physical state of the system. To be precise, the matrix of Poisson brackets of gauge
(f c)
conditions and gauge generators, [Ga , Cb ], must be nonsingular [11].
As an example, let us choose

G1 = q1 − q2 , (17.1a)
G2 = q3 + p4 (17.1b)

7 Canonical gauges restrict the phase space variables. Noncanonical gauges [11] involve the Lagrange multipliers.
 19

as our gauge conditions. This is a good gauge:

(f c) 3/2 0
det[Ga , Cb ]= = 3/2 . (17.2)
1/2 1
(f c)
The matrix [Ga , Cb ] is nonsingular, as required.
The gauge conditions Ga = 0, like the first and second class constraints, restrict the phase space variables. The full
set of restrictions
(all) (f c) (f c) (sc) (sc)
CA = {G1 , G2 , C1 , C2 , C1 , C2 } (17.3)

reduces the available phase space from eight dimensions to two dimensions. (Here, the index A ranges from 1 to 6.)
(all)
Taken as a whole, the six conditions CA are second class. We see this by computing the Poisson brackets

0 0 9 0 0 0
 
0 0 3 6 −2 −4
(all) (all) 1
−9 −3 0 0 0 0

MAB ≡ [CA , CB ] =   . (17.4)
6 0 −6 0 0 0 0
0 2 0 0 0 −6
0 4 0 0 6 0

This matrix has a nonzero determinant, det(M ) = 9/4, which is the condition for the set of constraints and gauge
conditions to be second class.
We can eliminate the constraints and gauge conditions by constructing Dirac brackets. The inverse of MAB is

0 0 −2 1 0 0
 
 0 0 0 −3 0 0
1  2 0 0 0 0 0
 
M AB =   , (17.5)
3 −1 3 0 0 −2 1
 0 0 0 2 0 3
0 0 0 −1 −3 0

and the Dirac brackets are defined by

(all) (all)
[F, G]∗ = [F, G] − [F, CA ]M AB [CB , G] . (17.6)

The Dirac brackets among the phase space variables are

∗
[q1 , q3 ] = [q1 , p2 ]∗ = −[q1 , p4 ]∗ = 1/3 , (17.7a)
∗ ∗ ∗
[q2 , q3 ] = [q2 , p2 ] = −[q2 , p4 ] = 1/3 , (17.7b)
∗ ∗ ∗
[q3 , q4 ] = −[q4 , p2 ] = [q4 , p4 ] = 2/3 , (17.7c)

with all other brackets vanishing.

The constraints can be solved in various ways and the results can be used freely, either before or after computing
Dirac brackets. For example, the constraints imply

q1 = −q4 /2 , (17.8a)
q2 = −q4 /2 , (17.8b)
q3 = p2 , (17.8c)
p1 =0, (17.8d)
p3 =0, (17.8e)
p4 = −p2 . (17.8f)

We can use these to eliminate the variables q1 , q2 , q3 , p1 , p3 and p4 . Then the extended Hamiltonian HE becomes
the fully reduced Hamiltonian
 
1 2 9 2
HF R = p + q , (17.9)
2 2 4 4
 20

which depends only on q4 and p2 .

The Dirac brackets of the variables that remain are [q4 , p2 ]∗ = −2/3. Thus, the equations of motion become
q̇4 = [q4 , HF R ]∗ = −2p2 /3 , (17.10a)
ṗ2 = [p4 , HF R ]∗ = 3q4 /2 . (17.10b)
These are the equations for a simple harmonic oscillator with solution
q4 (t) = α sin t + β cos t , (17.11a)
p2 (t) = −(3α/2) cos t + (3β/2) sin t , (17.11b)
where α and β are arbitrary constants.
With the gauge fixed, the dynamics take place on the fully reduced phase space, the two-dimensional surface defined
(all)
by the constraints and gauge conditions CA = 0. There are many different choices of coordinates for this surface.
Instead of solving the constraints for q4 and p2 , we could solve them for q2 and q3 . In that case, the fully reduced
Hamiltonian is
9 1
HF R = q22 + q32 (17.12)
2 2
and Hamilton’s equations are
q̇2 = [q2 , HF R ]∗ = q3 /3 , (17.13a)
q̇3 = [q3 , HF R ]∗ = −3q2 (17.13b)
Again, this describes the simple harmonic oscillator.
We can choose other coordinates on the fully reduced phase space. For example, let q4 = Q + q2 and p2 = −P ,
then use the constraints to eliminate q1 , q2 , q3 , p1 , p3 and p4 . The fully reduced Hamiltonian becomes
1 2
HF R = (Q + P 2 ) . (17.14)
2
The nonzero Dirac brackets are [Q, P ]∗ = 1, and the equations of motion are simply Q̇ = P and Ṗ = −Q.
In each case, the fully reduced theory exhibits the single physical degree of freedom that we expect.

XVIII. FULLY REDUCED ACTION

The fully reduced equations of motion can be derived from the action that includes all of the constraints and gauge
conditions. For lack of a better name, let us denote this action with the subscript “all”:
Z T n o
Sall = dt pi q̇i − Hf c − γa Ca(f c) − σa Ca(sc) − ρa Ga . (18.1)
0
Now extremize Sall with respect to variations in p1 , p3 , p4 , q1 , q2 , q3 and q4 :
q̇1 = γ1 + σ1 /3 , (18.2a)
q̇3 = −q1 + p3 + p4 /2 + γ1 /2 + γ2 − σ1 /3 + σ2 /3 , (18.2b)
q̇4 = −p2 /2 + p3 /2 − γ1 /2 + σ1 /3 + ρ2 , (18.2c)
ṗ1 = q1 + q2 + q4 + p3 − σ2 − ρ1 , (18.2d)
ṗ2 = q1 + 2q4 + σ2 − ρ1 , (18.2e)
ṗ3 = −ρ2 . (18.2f)
(all)
In addition, vary Sall with respect to the Lagrange multipliers to obtain the constraints CA = 0. The solution of
(all)
the full set of equations, (18.2) and CA = 0, is given by Equations (17.8) along with
γ1 = −p2 /3 − 2ṗ3 /2 + 2q̇1 /3 − 2q̇4 /3 , (18.3a)
γ2 = p2 − 3q4 /4 + ṗ1 /6 + ṗ2 /6 + ṗ3 + q̇3 + q̇4 , (18.3b)
σ1 = p2 + 2ṗ3 + q̇1 + 2q̇4 , (18.3c)
σ2 = 3q4 /4 − ṗ1 /2 − ṗ2 /2 , (18.3d)
ρ1 = −3q4 /4 − ṗ1 /2 + ṗ2 /2 , (18.3e)
ρ2 = −ṗ3 . (18.3f)
 21

We now insert these results into Sall to obtain the fully reduced action
Z T   
3 1 2 9 2
SF R [q4 , p2 ] = dt − p2 q̇4 − p + q , (18.4)
0 2 2 2 4 4
which is a functional of q4 and p2 . The equations of motion δSF R = 0 are
δSF R 3 9
0= = ṗ2 − q4 , (18.5a)
δq4 2 4
δSF R 3
0= = − q̇4 − p2 . (18.5b)
δp2 2
These are equivalent to Hamilton’s Equations (17.10) for the fully reduced Hamiltonian.
We can place SF R into “canonical form” by defining new variables P = −p2 and Q = 3q4 /2. Then
Z T  
1 2 2
SF R [Q, P ] = dt P Q̇ − [P + Q ] , (18.6)
0 2
which is the familiar action for the harmonic oscillator.

XIX. SUMMARY AND DISCUSSION

Here is the Dirac–Bergmann algorithm:

• Compute the conjugate momenta pi = ∂L/∂ q̇i and define the canonical Hamiltonian HC as pi q̇ i −L(q, q̇), written
in terms of p’s and q’s.
• Identify the primary constraints. The primary Hamiltonian HP is obtained from HC by adding the primary
constraints with Lagrange multipliers.
• Apply Dirac’s consistency conditions to identify higher-order constraints and restrictions on the Lagrange mul-
tipliers.
• The total Hamiltonian HT is found from HP by incorporating the restrictions on the Lagrange multipliers.
• Separate the primary, secondary, and higher-order constraints into first and second class.
• The first class Hamiltonian Hf c is the part of HT with the primary first class constraints removed.
• The extended Hamiltonian HE is obtained from the first class Hamiltonian Hf c by adding all of the first class
constraints with Lagrange multipliers.
• The partially reduced Hamiltonian HP R is found from HE by using Dirac brackets to eliminate the second class
constraints.
• Gauge freedom is removed by assigning gauge conditions. The fully reduced Hamiltonian HF R is obtained from
HE by using Dirac brackets to impose all constraints and gauge conditions.
The theory defined by the singular Lagrangian (1.1) provides a relatively complete example of each step in the algo-
rithm.
One reason the Dirac–Bergmann algorithm can be confusing is that typical examples are chosen for simplicity,
allowing some of the logical steps to be skipped. This causes the distinction between Hamiltonians to become blurred.
For example, if there are no restrictions on the Lagrange multipliers, then the primary Hamiltonian HP and the total
Hamiltonian HT coincide. Likewise, if there are no secondary (or higher-order) first class constraints, then the total
Hamiltonian HT and the extended Hamiltonian HE coincide. Moreover, for theories with no second class constraints
and no gauge conditions imposed, Dirac brackets and the reduction process are not needed.
Another confusing aspect of the Dirac–Bergmann algorithm is that for many important theories, the Lagrangian is
given to us in a form that contains Lagrange multipliers. For example, consider the Einstein–Hilbert action of general
relativity. The Lagrangian density is the spacetime curvature scalar. A 3+1 splitting of the spacetime metric [31, 32]
allows us to write the Lagrangian density as

L = R + Kij (g ik g jℓ − g ij g kℓ )Kkℓ , (19.1)

 22

apart from a total derivative term that integrates to the boundary. Here, R and gij are the spatial scalar curvature
and spatial metric. In addition, Kij is the extrinsic curvature of space, built from the lapse function N , shift vector
N i and spatial derivatives of gij . The Lagrangian density depends on the Lagrange multipliers N and N i as well as
the configuration space coordinates gij .
Because the Einstein–Hilbert Lagrangian (19.1) depends on Lagrange multipliers, it is not analogous to the singular
Lagrangian (1.1) of our example problem. Rather, it is analogous to the partially reduced Lagrangian that appears
in the integrand of the action SP R [Q, γ] of Equation (16.7). Recall that the partially reduced Hamiltonian HP R
contains only first class constraints and no restrictions on the Lagrange multipliers. Likewise, the Hamiltonian for
general relativity [32] is constructed from first class constraints (the Hamiltonian and momentum constraints), and the
Lagrange multipliers (the lapse function and shift vector) are unrestricted.
We can attempt to eliminate the lapse and shift from the Einstein–Hilbert action, just as we eliminated the γ’s
from SP R [Q, γ] and obtained the result SP R [Q] in Equation (16.9). It is straightforward to eliminate the lapse N —
the result is the Baierlein–Sharp–Wheeler action [33]. However, the shift vector cannot be eliminated algebraically
because the equations of motion obtained by varying the action with respect to the lapse and shift depend on spatial
derivatives of N i .
So, for general relativity, we do not have a singular Lagrangian analogous to Equation (1.1), and we cannot expect
to apply the Dirac–Bergmann algorithm from beginning to end as laid out by Dirac [10]. Nevertheless, the general
Dirac–Bergmann algorithm serves as the foundation for our understanding and interpretation of the Hamiltonian form
of the theory.

ACKNOWLEDGMENTS

I would like to thank Claudio Bunster and Marc Henneaux for sparking my interest in this subject many years ago.
I also thank Ashiqul Islam Dip for helpful discussions.

[1] P.G. Bergmann. Non–Linear Field Theories. Physical Review, 75:680, 1949.
[2] P.G. Bergmann and J.H.M. Brunings. Non–Linear Field Theories II. Canonical Equations and Quantization. Rev. Mod.
Phys., 21:480, 1949.
[3] P.A.M. Dirac. Generalized Hamiltonian dynamics. Canadian Journal of Mathematics, 2:129, 1950.
[4] P.G. Bergmann, R. Penfield, R. Schiller, and H. Zatzkis. The Hamiltonian of the General Theory of Relativity with
Electromagnetic Field. Physical Review, 80:81, 1950.
[5] P.A.M. Dirac. The Hamiltonian form of field dynamics. Canadian Journal of Mathematics, 3:1, 1951.
[6] J.L. Anderson and P.G. Bergmann. Constraints in Covariant Field Theories. Physical Review, 83:1018, 1951.
[7] P.G. Bergmann and I. Goldberg. Dirac bracket transformations in phase space. Physical Review, 98:531, 1955.
[8] P.A.M. Dirac. Generalized Hamiltonian dynamics. Proc. Roy. Soc. (London), 246:326, 1958.
[9] P.A.M. Dirac. The theory of gravitation in Hamiltonian form. Proc. Roy. Soc. (London), 246:333, 1958.
[10] P.A.M. Dirac. Lectures on Quantum Mechanics. Belfer Graduate School of Sciences, Yeshiva University, New York, 1964.
[11] M. Henneaux and C. Teitelboim. Quantization of Gauge Systems. Princeton University Press, Princeton, New Jersey,
1992.
[12] L. Rosenfeld. Zur Quantelung der Wellenfelder. Annalen der Physik, 397:113, 1930.
[13] L. Rosenfeld. On the quantization of wave fields. Eur. Phys. J. H, 42:63, 2017.
[14] R.H. Penfeld. Hamiltonians without Parametrization. Physical Review, 84:737, 1951.
[15] D. C. Salisbury. Peter Bergmann and the invention of constrained Hamiltonian dynamics. Einstein Stud., 12:247–257,
2012.
[16] A. Hanson, T. Regge, and C. Teitelboim. Constrained Hamiltonian Systems. Accademia Nazionale dei Lincei, Roma, 1976.
[17] K. Sundermeyer. Constrained Dynamics. Springer–Verlag, Berlin, 1982.
[18] H.J. Rothe and K.D. Rothe. Classical and Quantum Dynamics of Constrained Hamiltonian Systems. World Scientific,
Singapore, 2010.
[19] L. Lusanna. Non-Inertial Frames and Dirac Observables in Relativity. Cambridge University Press, 6 2019.
[20] R. Cawley. Determination of the Hamiltonian in the Presence of Constraints. Phys. Rev. Lett., 42:413, 1979.
[21] Ziping Li. On the invalidity of a conjecture of dirac. Chinese Physics Letters, 10(2):68–70, feb 1993.
[22] Y.-L. Wang, Z.-P. Wang, and K. Wang. Gauge Symmetries and the Dirac Conjecture. Int. J. Theor. Phys., 48:1894–1904,
2009.
[23] A. Frenkel. Comment on Cawley’s counterexample to a conjecture of Dirac. Phys. Rev. D, 21:2986–2987, 1980.
[24] Reiji Sugano and Toshiei Kimura. Comments on Constraints of Gauge Theories. J. Phys. A, 16:4417, 1983.
[25] Bi-Chu Wu. Dirac’s conjecture. Int. J. Theor. Phys., 33:1529–1533, 1994.
 23

[26] Olivera Miskovic and Jorge Zanelli. Dynamical structure of irregular constrained systems. J. Math. Phys., 44:3876–3887,
2003.
[27] M. E. V. Costa, H. O. Girotti, and T. J. M. Simoes. Dynamics of Gauge Systems and Dirac’s Conjecture. Phys. Rev. D,
32:405, 1985.
[28] A. Cabo and D. Louis-Martinez. On Dirac’s conjecture for Hamiltonian systems with first and second class constraints.
Phys. Rev. D, 42:2726–2735, 1990.
[29] H.J. Rothe and K.D. Rothe. Gauge identities and the Dirac conjecture. Annals Phys., 313:479–496, 2004.
[30] J.M. Pons. On Dirac’s incomplete analysis of gauge transformations. Stud. Hist. Phil. Sci. B, 36:491–518, 2005.
[31] C.W. Misner, K.S. Thorne, and J.A. Wheeler. Gravitation. W.H. Freeman, San Francisco, 1973.
[32] Richard L. Arnowitt, Stanley Deser, and Charles W. Misner. The Dynamics of general relativity. Gen. Rel. Grav.,
40:1997–2027, 2008.
[33] R.F. Baierlein, D.H. Sharp, and J.A. Wheeler. Three-dimensional geometry as carrier of information about time. Phys.
Rev., 126:1864–1865, Jun 1962.