Constraints, Dirac Brackets, and Symplectic Reduction

Constraints first select a submanifold of phase space; what happens next depends on the two-form pulled back to that submanifold. If the pullback is nondegenerate, the constraints are second class and its inverse Poisson structure is computed with the Dirac bracket. If the pullback has null directions and those directions represent gauge redundancy, the physical phase space is obtained by quotienting their leaves. Under regularity, constant-rank, and smooth-quotient hypotheses, either route produces a lower-dimensional symplectic phase space. Projectable observables and a projectable Hamiltonian then supply its physical quantities and evolution.

This answer is local and conditional. A rank-changing constraint set, a non-Hausdorff orbit space, or a non-differentiable field-theory generator can block the regular quotient. Even when the quotient exists, a Hamiltonian that is not constant along the null leaves does not define reduced dynamics.

Required background. Symplectic Forms, Hamiltonian Flows, and Poisson Brackets. It supplies symplectic forms, Hamiltonian vector fields, pullbacks, and both ordinary and equal-time functional Poisson brackets.

Constraint surface and symplectic sign convention

We retain its sign package:

$$\begin{gathered} \omega=\sum_i\mathrm dq^i\wedge\mathrm dp_i, \qquad \iota_{X_F}\omega=\mathrm dF,\\ \{F,G\}=\omega(X_F,X_G), \qquad X_F[G]=\{G,F\},\\ \left[X_F,X_G\right]=-X_{\{F,G\}}, \qquad \dot F=\partial_tF+\{F,H\}. \end{gathered}$$

Thus a generator $G$ acts as $\delta F=\{F,G\}$. Sources that instead use $\iota_{X_F}\omega=-\mathrm dF$ require a simultaneous sign translation; mixing individual formulas will reverse constraint flows or the Dirac correction.

Let $\mathcal C$ denote the final constraint surface. Weak equality means equality only after restriction to it:

$$F\approx G \quad\Longleftrightarrow\quad F|_{\mathcal C}=G|_{\mathcal C}.$$

One must compute an ordinary Poisson bracket before imposing weak equality. For example, $q\approx0$ and $p\approx0$ do not imply $\{q,p\}\approx0$.

From a singular Legendre map to the final surface

The Legendre map sends $(q,\dot q)$ to $(q,p)$ with

$$p_i=\frac{\partial L}{\partial\dot q^i}.$$

Its velocity derivative is the Lagrangian Hessian. When that Hessian is invertible, one solves for the velocities and obtains $H_c=p_i\dot q^i-L$. When it is singular, assume this energy descends to the image of the Legendre map; relations cutting out that image are the primary constraints $\phi_a^{(1)}\approx0$. Starting from the resulting canonical Hamiltonian $H_c$, form the total Hamiltonian

$$H_T=H_c+u^a(t)\phi_a^{(1)}.$$

Every constraint already found must be preserved:

$$0\approx\dot\phi_a =\partial_t\phi_a+\{\phi_a,H_T\}.$$

Each consistency equation can hold identically, determine one or more multipliers $u^a$, produce a secondary constraint, or expose an inconsistency. New constraints are appended and the test is repeated until no new condition appears. This is the stabilization step. Date 2010, Chapters 3–4, pp. 14–22, PDF develops this procedure, the weak-equality notation, the class split, and the Dirac bracket in one conventionally coherent treatment.

Two classifications must not be confused. Primary and secondary describe where a constraint entered the stabilization algorithm. First class and second class describe its Poisson geometry after the full constraint set is known. A secondary constraint may be first class, and a primary constraint may be second class.

Suppose the stabilized constraints are the components of a smooth map

$$\Phi=(\phi_1,\ldots,\phi_m):M\longrightarrow\mathbb R^m, \qquad \mathcal C=\Phi^{-1}(0).$$

Assume their differentials are independent on $\mathcal C$. Then $0$ is a regular value,

$$T_x\mathcal C =\bigcap_{a=1}^m\ker(\mathrm d\phi_a)_x, \qquad \dim\mathcal C=\dim M-m.$$

Reducible constraints, or constraints whose differential rank changes, do not satisfy this starting hypothesis. They require a new local description or a singular treatment rather than the regular formulas below.

The constraint matrix reveals the geometry

Let $\iota:\mathcal C\hookrightarrow M$ be the inclusion and pull back the symplectic form:

$$\omega_{\mathcal C}=\iota^*\omega.$$

It is closed, but it need not be nondegenerate. At a regular point,

$$(T_x\mathcal C)^\omega := \left\{ w\in T_xM: \omega_x(w,v)=0 \ \text{for every }v\in T_x\mathcal C \right\}.$$

This symplectic orthogonal is

$$(T_x\mathcal C)^\omega =\operatorname{span}\{X_{\phi_1}(x),\ldots,X_{\phi_m}(x)\}.$$

Indeed, if $v\in T_x\mathcal C$, then $\omega(X_{\phi_a},v)=\mathrm d\phi_a(v)=0$. Independence of the differentials supplies the reverse dimension count.

Now define the antisymmetric constraint matrix

$$\Delta_{ab}(x)=\{\phi_a,\phi_b\}(x).$$

A vector $\lambda^aX_{\phi_a}(x)$ is also tangent to $\mathcal C$ precisely when

$$\Delta_{ba}(x)\lambda^a=0.$$

Consequently the characteristic kernel is

$$\begin{aligned} K_x &=\ker(\omega_{\mathcal C})_x\\ &=T_x\mathcal C\cap(T_x\mathcal C)^\omega, \end{aligned} \qquad \dim K_x=m-\operatorname{rank}\Delta(x).$$

This is the geometric source of the class split. A function $F$ is first class when

$$\{F,\phi_a\}\approx0 \qquad\text{for every final constraint }\phi_a.$$

Its Hamiltonian field is tangent to $\mathcal C$. If $F$ is itself a constraint, its restricted Hamiltonian field lies in $K$. By contrast, a regular chosen family $\chi_A$ is second class when

$$\Delta_{AB}=\{\chi_A,\chi_B\}$$

is invertible on a neighborhood of its zero set. An invertible antisymmetric matrix has even size. In a mixed system the split into first- and second-class combinations is generally local; it is not valid to label every constraint that fails an individual first-class test as an independent second-class constraint.

Second-class constraints and the Dirac bracket

Let $\chi_A$, $A=1,\ldots,s$, be an irreducible second-class set, where $s$ is even, and choose

$$\Delta^{AB}\Delta_{BC}=\delta^A{}_C.$$

The Dirac bracket is

$$\boxed{ \{F,G\}_D =\{F,G\} -\{F,\chi_A\}\Delta^{AB}\{\chi_B,G\}. }$$

It is again a Poisson bracket wherever $\Delta^{-1}$ exists. Its defining check is immediate:

$$\begin{aligned} \{F,\chi_C\}_D &=\{F,\chi_C\} -\{F,\chi_A\}\Delta^{AB}\Delta_{BC}\\ &=0. \end{aligned}$$

Thus the second-class constraints are strong Casimirs of the new bracket. After switching to $\{\cdot,\cdot\}_D$, they may be imposed before later Dirac-bracket calculations. They may not be imposed before the ordinary Poisson brackets used to build $\Delta$.

The corresponding projected Hamiltonian field is

$$X_F^D =X_F +X_{\chi_A}\Delta^{AB}\{F,\chi_B\}.$$

It is tangent to the second-class surface $\mathcal S$, and for every $v\in T\mathcal S$,

$$\omega(X_F^D,v)=\mathrm dF(v).$$

If the $\chi_A$ are the entire regular constraint set, invertibility of $\Delta$ is equivalent to nondegeneracy of $\iota_{\mathcal S}^*\omega$. For functions $f,g\in C^\infty(\mathcal S)$ and arbitrary extensions $F,G$ off the surface,

$$\{f,g\}_{\mathcal S} =\{F,G\}_D|_{\mathcal S}.$$

The right-hand side is independent of those extensions. The off-surface formula can depend on the chosen extensions of the constraints; the intrinsic bracket on $\mathcal S$ is the invariant result.

First-class constraints and the characteristic quotient

After any second-class subset has been eliminated, suppose the remaining regular surface $\mathcal C$ is coisotropic:

$$(T\mathcal C)^\omega\subseteq T\mathcal C.$$

Then

$$K=\ker\omega_{\mathcal C}=(T\mathcal C)^\omega$$

is spanned locally by the Hamiltonian fields of a complete first-class constraint set. Because $\omega_{\mathcal C}$ is closed, its constant-rank kernel is involutive. Indeed, for $Y,Z\in\Gamma(K)$, Cartan’s formula gives

$$\iota_{[Y,Z]}\omega_{\mathcal C}=0.$$

The leaves of $K$ are the characteristic, or locally gauge, directions. Calling them physical redundancies still requires the model’s boundary conditions and observable content; in field theory a nonzero boundary charge can prevent such a direction from being discarded.

Assume now that the leaf space

$$M_{\mathrm{red}}=\mathcal C/K$$

is a smooth Hausdorff manifold and that the projection $\pi:\mathcal C\to M_{\mathrm{red}}$ is a surjective submersion. The pulled-back form is horizontal because $\iota_Y\omega_{\mathcal C}=0$ for $Y\in K$, and it is invariant because

$$\mathcal L_Y\omega_{\mathcal C} =\mathrm d(\iota_Y\omega_{\mathcal C}) +\iota_Y\mathrm d\omega_{\mathcal C}=0.$$

It therefore descends to the unique two-form satisfying

$$\boxed{ \pi^*\omega_{\mathrm{red}}=\omega_{\mathcal C}. }$$

Closedness follows by pulling back $\mathrm d\omega_{\mathrm{red}}$, and nondegeneracy follows because the only vectors removed by $\pi_*$ are precisely those in $K$. This is symplectic reduction.

A function on $\mathcal C$ descends exactly when it is constant on the characteristic leaves. If $F$ and $G$ are first-class extensions of two such functions,

$$\{f_{\mathrm{red}},g_{\mathrm{red}}\}_{\mathrm{red}}\circ\pi =\{F,G\}|_{\mathcal C}.$$

The restricted result is unchanged by adding functions that vanish on $\mathcal C$. In particular, a Hamiltonian must satisfy

$$\mathrm dH_{\mathcal C}(Y)=0 \qquad\text{for every }Y\in K$$

before there can be a reduced Hamiltonian $H_{\mathcal C}=H_{\mathrm{red}}\circ\pi$. The constrained equation

$$\iota_X\omega_{\mathcal C}=\mathrm dH_{\mathcal C}$$

is solvable under the same condition. Its solutions differ by a vector in $K$, but they induce one vector field on $M_{\mathrm{red}}$.

A method for mixed systems

For a regular problem, the complete procedure is:

1. Start with $(M,\omega,H_c)$ and the primary constraints supplied by the singular Legendre map.
2. Stabilize with $H_T$, adding every secondary constraint and solving only the multipliers that consistency determines.
3. Verify independence and constant rank on the region being studied.
4. Choose a local split into first- and second-class combinations.
5. Eliminate the second-class subset, either by solving it and pulling back $\omega$ or by using the Dirac bracket. These routes must give the same intrinsic bracket.
6. Find the remaining characteristic distribution and quotient its leaves only if the leaf space passes the smoothness and Hausdorff tests.
7. Check that the Hamiltonian and proposed observables are constant along the leaves, then compare their reduced brackets and evolution.

If $\dim M=2n$, there are $r$ independent first-class constraints and $s$ independent second-class constraints, and all regular quotient hypotheses hold, then

$$\dim M_{\mathrm{red}}=2n-2r-s.$$

Each constraint removes one dimension by restriction; each first-class constraint removes one further characteristic direction. The number $s$ is even. This count is invalid for reducible constraints, changing rank, or singular orbit strata.

A gauge condition may pair locally with a first-class constraint to create an invertible second-class matrix. That gives a local slice, not a theorem that every orbit is met once globally. Residual symmetries and Gribov-type multiple intersections are stop conditions for the naive global argument.

Moment-map reduction as a special case

This optional bridge connects the general quotient to Hamiltonian Group Actions and Moment Maps; no earlier step assumes that page. Let a Lie group $G$ act on $M$. For $\xi\in\mathfrak g$, write $\xi_M$ for its infinitesimal action field and $\mu_\xi=\langle\mu,\xi\rangle$. Assume

$$\mathrm d\mu_\xi=\iota_{\xi_M}\omega, \qquad \mu(g\mathbin{\cdot}x)=\operatorname{Ad}_g^*\mu(x).$$

Thus $\mu:M\to\mathfrak g^*$ is an equivariant moment map. Let $\nu$ be a regular value and define its coadjoint stabilizer by

$$G_\nu =\{g\in G:\operatorname{Ad}_g^*\nu=\nu\}.$$

Only $G_\nu$ necessarily preserves $\mu^{-1}(\nu)$. If its action on that level is free and proper, then

$$M_\nu=\mu^{-1}(\nu)/G_\nu, \qquad \pi^*\omega_\nu=\iota^*\omega.$$

If $\mu$ is a submersion along the level, the dimension is

$$\dim M_\nu =\dim M-\dim G-\dim G_\nu.$$

At a coadjoint-fixed value, including $\nu=0$, one has $G_\nu=G$, so

$$\dim M_0=\dim M-2\dim G.$$

At a general $\nu$, it is wrong to replace $G_\nu$ by all of $G$, or to declare every component of $\mu-\nu$ first class. Locally free actions can produce orbifolds; stabilizer jumps and nonregular values can produce stratified reduced spaces. Cannas da Silva 2006, Chapters 23–24, pp. 141–150, PDF proves the regular zero-level construction and treats other levels and finite stabilizers.

Two finite checks

A first-class null direction

Consider

$$L(q,\dot q)=\frac12\dot q_1^2-V(q_1),$$

where $q_2$ is absent. Then

$$p_1=\dot q_1, \qquad p_2=0, \qquad H_T=\frac12p_1^2+V(q_1)+u(t)p_2.$$

The primary constraint $\phi=p_2\approx0$ is preserved identically. Its multiplier remains arbitrary:

$$\dot p_2=0, \qquad \dot q_2=\{q_2,H_T\}=u(t).$$

Moreover, $p_2$ generates $\delta q_2=\{q_2,\epsilon p_2\}=\epsilon$. On $\mathcal C=\{p_2=0\}$,

$$\omega_{\mathcal C}=\mathrm dq_1\wedge\mathrm dp_1, \qquad K=\operatorname{span}\{\partial_{q_2}\}.$$

Taking the quotient by the $q_2$-translation leaves gives $M_{\mathrm{red}}\simeq T^*\mathbb R$ with coordinates $(q_1,p_1)$. Equivalently, the gauge condition $q_2=0$ pairs with $p_2=0$ because $\{q_2,p_2\}=1$; their Dirac bracket leaves $\{q_1,p_1\}_D=1$. The quotient and gauge-slice calculations agree in this global toy model.

A second-class sphere

Let $M=T^*\mathbb R^N$ and constrain a particle to the sphere of radius $R>0$ by

$$\chi_1=\frac12(q^2-R^2), \qquad \chi_2=q\cdot p.$$

The first equation fixes the position. For the kinetic Hamiltonian $H_0=p^2/(2m)$, its preservation supplies the second because $\dot\chi_1=\{\chi_1,H_0\}=\chi_2/m$; that condition makes the momentum tangent. Their matrix and inverse are

$$\Delta_{AB} = \begin{pmatrix} 0&q^2\\ -q^2&0 \end{pmatrix}, \qquad \Delta^{AB} = \begin{pmatrix} 0&-q^{-2}\\ q^{-2}&0 \end{pmatrix}.$$

It is invertible near the surface because $q^2=R^2>0$. Substitution in the Dirac formula gives

$$\begin{aligned} \{q_i,q_j\}_D&=0,\\ \{q_i,p_j\}_D &=\delta_{ij}-\frac{q_iq_j}{q^2},\\ \{p_i,p_j\}_D &=\frac{p_iq_j-q_ip_j}{q^2}. \end{aligned}$$

On the surface, the middle bracket is the tangent projector $P_{ij}=\delta_{ij}-q_iq_j/R^2$: it obeys $P_{ij}q_j=0$ and $P_{ik}P_{kj}=P_{ij}$. This independently identifies the intrinsic space as $T^*S^{N-1}$ and explains why its dimension, $2N-2$, is even. No quotient is needed because the pulled-back form is already nondegenerate.

Constrained gauge systems before quantization

Take source-free Maxwell theory on the fixed slice $\Sigma=\mathbb R^3$, with fields and variations decaying rapidly and gauge parameters $\epsilon$ of compact support. We take the canonical Gauss constraint as input rather than repeating its complete Lagrangian derivation. Spatial indices are raised with the positive Euclidean slice metric, so $\partial^i=\delta^{ij}\partial_j$:

$$\Omega =\int_\Sigma\mathrm d^3x\, \boldsymbol{\delta}A_i\wedge\boldsymbol{\delta}E^i, \qquad \mathcal G=\partial_iE^i\approx0.$$

The smeared generator is

$$\begin{aligned} G[\epsilon] &=-\int_\Sigma\mathrm d^3x\, \epsilon\,\partial_iE^i\\ &=\int_\Sigma\mathrm d^3x\, E^i\partial_i\epsilon. \end{aligned}$$

The second line discards $-\int_{\partial\Sigma}\epsilon E^in_i$; the stated support and falloff conditions make it vanish. With

$$\{A_i(\mathbf x),E^j(\mathbf y)\} =\delta_i{}^j\delta^{(3)}(\mathbf x-\mathbf y),$$

one finds

$$\{A_i,G[\epsilon]\}=\partial_i\epsilon, \qquad \{E^i,G[\epsilon]\}=0, \qquad \{G[\epsilon],G[\eta]\}=0.$$

For a tangent variation satisfying $\partial_i\boldsymbol{\delta}E^i=0$,

$$\begin{aligned} \Omega\bigl((\partial\epsilon,0), (\boldsymbol{\delta}A,\boldsymbol{\delta}E)\bigr) &=\int_\Sigma\mathrm d^3x\, \partial_i\epsilon\,\boldsymbol{\delta}E^i\\ &=0. \end{aligned}$$

Thus the Gauss-generated directions are characteristic null directions in this controlled setting. The Maxwell Hamiltonian

$$H=\frac12\int_\Sigma\mathrm d^3x\, (\mathbf E^2+\mathbf B^2)$$

is constant along them because $\mathbf B=\boldsymbol\nabla\times\mathbf A$ is unchanged by $\mathbf A\mapsto\mathbf A+\boldsymbol\nabla\epsilon$. It therefore descends to the quotient.

The local gauge condition $\kappa=\partial^iA_i\approx0$ pairs with $\gamma=\partial_iE^i\approx0$:

$$\{\gamma(\mathbf x),\kappa(\mathbf y)\} =\nabla_{\mathbf x}^2\delta^{(3)}(\mathbf x-\mathbf y).$$

On a function space where the selected falloff conditions make $\nabla^2$ invertible, choose its Green inverse. The resulting Dirac bracket is

$$\{A_i(\mathbf x),E^j(\mathbf y)\}_D = \left( \delta_i{}^j-\partial_i\nabla^{-2}\partial^j \right) \delta^{(3)}(\mathbf x-\mathbf y).$$

For nonzero momentum this is the rank-two transverse projector

$$P_i{}^j(\mathbf k) =\delta_i{}^j-\frac{k_ik^j}{\mathbf k^2}.$$

This agreement between quotienting and a local gauge-fixed Dirac bracket is the promised QFT-facing transfer. It is classical: it neither quantizes Maxwell theory nor proves that reduction commutes with quantization. Tong 2006–2007, §§ 6.2–6.2.1, pp. 127–130, PDF supplies the canonical role of $A_0$, Gauss’s law, Coulomb gauge, and the transverse projector.

If $\Sigma$ has a boundary, if $\epsilon$ does not vanish there, or if topology produces zero modes, the discarded surface term or Laplacian kernel can change the conclusion. Developed orbit structure, stabilizers, singular strata, and boundary-qualified gauge transformations belong to Gauge Orbits, Gauss Constraints, and Stabilizers.

Stop conditions and common pitfalls

Stop when stabilization is inconsistent. A consistency equation with no solution for the multipliers and no admissible new constraint means the proposed constrained dynamics is inconsistent.

Stop before inverting a singular matrix. If $\det\Delta=0$, its rank changes, or the constraint set is redundant, the displayed Dirac bracket is not defined. Reclassify locally or use a singular method.

Restriction is not quotienting. Imposing first-class constraints leaves the characteristic directions in $\mathcal C$. Conversely, quotienting is not the treatment of a second-class surface, whose pulled-back form is already nondegenerate.

A first-class constraint is not automatically the full gauge generator. In a Lagrangian gauge theory, the transformation of every variable can require a tuned combination of primary and secondary constraints and time derivatives of gauge parameters.

A local slice is not a global slice. Invertibility of a gauge-fixing matrix proves a local transverse intersection only. Residual transformations or multiple intersections can still obstruct a global representative.

A singular quotient is still an answer, but not a manifold. Changing stabilizers, nonfree or nonproper actions, and non-Hausdorff leaf spaces must be reported as orbifold, stratified, or otherwise singular outcomes rather than hidden inside the regular dimension formula.

Boundary terms decide whether a direction is gauge. If a smeared generator is not differentiable, it must be improved or its boundary conditions changed. A transformation carrying a nonzero charge is not automatically a redundancy to quotient.

Exercises

1. Separate the two classifications. In the first finite model, how is $p_2$ classified by origin and by Poisson geometry? What calculation fixes each label?

Solution

It is primary because it follows directly from the Legendre map: $p_2=\partial L/\partial\dot q_2=0$. It is first class because it has weakly vanishing Poisson bracket with the full constraint set, which here contains only $p_2$. The two answers use different tests.

2. Diagnose the weak-equality trap. Suppose $q\approx0$ and $p\approx0$. Why can one not set them to zero before classifying the pair?

Solution

Weak equality is restriction after the ordinary bracket is evaluated. Here

$$\Delta = \begin{pmatrix} \{q,q\}&\{q,p\}\\ \{p,q\}&\{p,p\} \end{pmatrix} = \begin{pmatrix} 0&1\\ -1&0 \end{pmatrix}.$$

It is invertible, so $(q,p)$ is second class. Substituting zero first would erase the information that proves this.

3. Check the sphere bracket. Derive $\{q_i,p_j\}_D$ from $\chi_1$ and $\chi_2$, and test that the result removes the radial direction.

Solution

The needed brackets are

$$\{q_i,\chi_1\}=0, \qquad \{q_i,\chi_2\}=q_i, \qquad \{\chi_1,p_j\}=q_j.$$

Using $\Delta^{21}=q^{-2}$ gives

$$\{q_i,p_j\}_D =\delta_{ij}-\frac{q_iq_j}{q^2}.$$

Multiplication by $q_j$ gives zero on $q^2=R^2$, so the radial momentum direction has been projected out.

4. Test a reduction hypothesis. Let $S^1$ act on $\mathbb C^2$ by

$$e^{i\theta}\cdot(z_1,z_2) =(e^{i\theta}z_1,e^{2i\theta}z_2),$$

and restrict to the regular ellipsoid $|z_1|^2+2|z_2|^2=c>0$. Which theorem hypothesis fails at points with $z_1=0$, and what conclusion must be weakened?

Solution

The action is not free there because rotation by $\theta=\pi$ fixes $(0,z_2)$. The natural quotient is a weighted-projective orbifold with a $\mathbb Z_2$ point, and the level-to-quotient map is not a principal $S^1$-bundle. The regular free-action theorem’s manifold-and-bundle conclusion must therefore be replaced by an orbifold statement.

5. Transfer the sign to Maxwell theory. Starting from $\mathcal G=\partial_iE^i$, determine the sign of the smeared generator that gives $\delta A_i=+\partial_i\epsilon$. Then verify that the momentum projector is transverse.

Solution

Because

$$\left\{ A_i(\mathbf x), \int\mathrm d^3y\,\epsilon\,\partial_jE^j \right\} =-\partial_i\epsilon(\mathbf x),$$

the desired generator is $G[\epsilon]=-\int\epsilon\,\partial_iE^i$. For

$$P_i{}^j(\mathbf k) =\delta_i{}^j-\frac{k_ik^j}{\mathbf k^2},$$

one has

$$k^iP_i{}^j=0, \qquad P_i{}^\ell P_\ell{}^j=P_i{}^j.$$

Thus the bracket retains precisely the two directions transverse to $\mathbf k\ne0$.

Synthesis and continuation

The constraint matrix and the pulled-back two-form encode the same decision. An invertible constraint matrix gives a symplectic submanifold, and the Dirac bracket computes its intrinsic Poisson structure. A kernel instead marks first-class characteristic directions; if their leaf space is smooth, quotienting gives the unique form $\pi^*\omega_{\mathrm{red}}=\iota^*\omega$. Mixed systems require these operations in that order, followed by a projectability check for dynamics and observables.

The finite models show that gauge fixing and quotienting agree when a global slice really exists, while Maxwell theory shows why support, boundary, and zero-mode assumptions cannot be omitted. The next physical step is the linked treatment of gauge orbits and Gauss constraints; singular quotients, boundary charges, and covariant phase-space ambiguities require their own hypotheses rather than an extrapolation of the regular formulas used here.

References

• Ana Cannas da Silva, Lectures on Symplectic Geometry — Open PDF, revised January 2006, Lecture Notes in Mathematics 1764, Springer, doi:10.1007/978-3-540-45330-7, Chapters 23–24, pp. 141–150. This is the structural source for the reduced two-form, regular moment-map reduction, reduction at other levels, and orbifold warnings.
• Ghanashyam Date, Lectures on Constrained Systems — Open PDF, arXiv:1010.2062 [gr-qc], 2010, Chapters 3–4, pp. 14–22. This is the teaching source for stabilization, weak equality, first- and second-class constraints, the Dirac bracket, and gauge-equivalence classes.
• David Tong, Lectures on Quantum Field Theory — Open PDF, Cambridge Part III lecture notes, 2006–2007, §§ 6.2 and 6.2.1, pp. 127–130. This source supports the controlled Maxwell application: $A_0$ as a multiplier, Gauss’s law, Coulomb gauge, and the transverse projector. The page stops before Tong’s quantization step.