Gauge Fields, Redundancy, and Observable Content

An ordinary gauge field is a connection, but a local connection potential is not itself an observable. Its inhomogeneous transformation law makes derivatives of charged matter covariant, while the curvature transforms homogeneously. Physical quantities arise only after one declares the admissible fields, boundary or falloff conditions, and the subgroup of gauge transformations treated as redundancy. This last qualification matters: at a boundary, a transformation that looks locally like a gauge transformation can carry electric-flux charge instead of identifying two physical states. This page develops that distinction for ordinary compact Lie gauge theory and makes it explicit in Maxwell theory on a finite region.

A background connection obeys the same local transformation law but remains a fixed source. The $A$ considered here is dynamical and belongs to a constrained phase space; Background Fields versus Dynamical Gauging develops that S1 distinction.

Required background. Bundle Connections, Curvature, Gauge Transformations, and Bianchi Identities supplies the geometric connection language, while The Free Maxwell Field and Gauge Redundancy supplies the free spin-one dynamics and constraint count used in the example.

Connections and the local gauge law

Let $G$ be a compact Lie group, let $R$ be a unitary representation on the matter fields, and fix a principal $G$-bundle $P\to M$. A gauge field is a connection on $P$. On a trivializing patch it is represented by

$$A=A_\mu^a T^a\,dx^\mu,$$

where the generators $T^a$ are Hermitian. The associated covariant derivative on a matter multiplet $\psi$ is written here for one simple or $U(1)$ gauge factor as

$$D_\mu\psi=(\partial_\mu-igA_\mu)\psi.$$

For a product group, the same formula applies factorwise with its own coupling. On the chosen trivializing patch $U\subset M$, take an active local representative $h:U\to G$ and set $\psi^h=h\psi$. Globally, a gauge transformation is a vertical bundle automorphism, equivalently a section of $\operatorname{Ad}(P)$; it is a single map $M\to G$ only after a suitable trivialization. Covariance fixes the transformation of the local potential:

$$\begin{aligned} A_\mu^h &=hA_\mu h^{-1} -\frac{i}{g}(\partial_\mu h)h^{-1}, \\ D_\mu^h(h\psi) &=hD_\mu\psi. \end{aligned}$$

The second line is the useful check. Differentiating $h\psi$ produces a term $(\partial_\mu h)\psi$; the inhomogeneous term in $A_\mu^h$ cancels it. A transformation law with the opposite sign would fail this test in the present convention.

Define the curvature through the commutator

$$[D_\mu,D_\nu]=-igF_{\mu\nu}.$$

Then

$$\begin{aligned} F_{\mu\nu} &=\partial_\mu A_\nu-\partial_\nu A_\mu -ig[A_\mu,A_\nu], \\ F_{\mu\nu}^h &=hF_{\mu\nu}h^{-1}. \end{aligned}$$

Thus the potential transforms inhomogeneously, whereas the curvature transforms homogeneously. For $h=e^{ig\epsilon}$, the infinitesimal laws are

$$\delta_\epsilon\psi=ig\epsilon\psi, \qquad \delta_\epsilon A_\mu =\partial_\mu\epsilon-ig[A_\mu,\epsilon].$$

These formulas and their normalization are developed in Schwartz 2014, § 25.2.2, pp. 490–493 and independently in Weinberg 1996, §§ 15.1–15.2, pp. 2–12. Weinberg initially absorbs coupling factors into the generators and structure constants; the displayed equations restore the site’s explicit $g$.

The formulas are local representatives of a global connection. On a nontrivial bundle, different patches have different potentials related by the same inhomogeneous law. No single $A_\mu$ need exist over all of $M$. Local Potentials and Global Gauge Configurations develops the patching data; here the bundle sector is fixed so that the physical meaning of redundancy can be isolated.

Redundancy depends on the allowed transformation group

The phrase “a gauge transformation is a redundancy” is incomplete until three choices have been made.

1. Field space. Specify the bundle sector, regularity, matter representations, and any external sources.
2. Boundary data. Specify boundary conditions or asymptotic falloffs and which transformations preserve them.
3. Zero-generator subgroup. Among the preserving transformations, identify those whose differentiable canonical generator has vanishing variation throughout the allowed constrained phase space. Any remaining state-independent constant can then be normalized to zero.

Write $\mathcal C_{\mathrm{adm}}$ for the resulting admissible field histories and $\mathcal G_{\mathrm{adm}}$ for its admissible gauge transformations. The subgroup that is actually treated as redundancy will be denoted $\mathcal G_0$. At the history level, its orbits are

$$\mathcal C_{\mathrm{adm}}/\mathcal G_0.$$

In a canonical description, let $\mathcal P$ denote phase space and $\mathcal P_{\mathrm{Gauss}}$ the Gauss-constraint surface. The corresponding reduction is

$$\mathcal P_{\mathrm{phys}} = \frac{\mathcal P_{\mathrm{Gauss}}}{\mathcal G_0}.$$

This schematic formula does not by itself prove that every orbit is regular or that all global components have been identified. One should not silently replace $\mathcal G_0$ by every local representative of a bundle automorphism.

For transformations connected to the identity, the Gauss constraint supplies the infinitesimal candidate directions. Three further tests remain:

• a transformation can fail to preserve the chosen bundle, boundary condition, or falloff;
• a transformation can preserve the field space but have a nonzero boundary or asymptotic generator;
• a disconnected or topologically nontrivial transformation is not classified by an infinitesimal Gauss generator, so its role requires separate global input.

The first two tests are explicit in the Maxwell example below. Gauge Orbits, Gauss Constraints, and Stabilizers supplies the full constraint reduction, presymplectic interpretation, and special configurations with nontrivial stabilizers. Large Gauge Transformations and Topological Sectors treats the third test.

Observable content descends to the quotient

A classical observable must assign the same value to every representative of a $\mathcal G_0$-orbit:

$$\mathcal O[A^h,h\psi]=\mathcal O[A,\psi], \qquad h\in\mathcal G_0.$$

Covariant objects remain indispensable, but covariance is not invariance. The basic transformation test gives the following first inventory.

Object	Transformation	Status before further completion
$A_\mu$	Inhomogeneous	Local representative, not an observable
$\psi$	$\psi^h=h\psi$	Charged covariant field
$F_{\mu\nu}$	$F_{\mu\nu}^h=hF_{\mu\nu}h^{-1}$	Covariant; invariant only for Abelian $G$
$\operatorname{tr}(F_{\mu\nu}F^{\mu\nu})$	Unchanged	Local invariant contraction
$\psi^\dagger\psi$	Unchanged	Neutral local composite
Open parallel transporter	Endpoint covariance	Needs endpoint data or dressing
Character of closed holonomy	Unchanged	Extended gauge-invariant observable

For a path $\gamma$ from $x$ to $y$, define parallel transport consistently with $D_\mu=\partial_\mu-igA_\mu$ by

$$U_\gamma(y,x) = \mathcal P\exp\!\left(ig\int_\gamma A\right).$$

It transforms at both endpoints:

$$U_\gamma^h(y,x) =h(y)U_\gamma(y,x)h(x)^{-1}.$$

An open transporter is therefore not gauge invariant by itself. If $\gamma$ is a closed curve $C$ based at $x$, the matrix holonomy still transforms by conjugation. Only an invariant function of that matrix, such as the Wilson loop in representation $R$,

$$W_R(C)=\operatorname{tr}_R U_C(x,x),$$

is invariant; cyclicity of the trace removes the base-point conjugation. Schwartz gives the parallel-transport and finite-transformation calculation at Schwartz 2014, §§ 25.2.1–25.2.2, pp. 488–493.

This criterion is necessary, not sufficient, for a quantum observable. Composite operators still require renormalization. A proposed line must have allowed endpoint and global data to be genuine. Gauge invariance alone does not prove locality, finiteness, or completeness, and curvature alone need not separate flat connections with different holonomy. Gauge-Invariant and Dressed Observables develops charged dressings and the distinction between local, extended, and asymptotic observables.

Maxwell theory on a finite region

Take four-dimensional compact $U(1)$ Maxwell theory on $M=\mathbb R\times\Sigma$, where the spatial region $\Sigma$ has smooth boundary $\partial\Sigma$. Work in a fixed bundle sector and impose the Dirichlet condition used in the example below: the pullback of $A$ is fixed on the timelike boundary $\Gamma=\mathbb R\times\partial\Sigma$.

The local transformation is

$$A_\mu\longmapsto A_\mu+\partial_\mu\lambda, \qquad F_{\mu\nu}\longmapsto F_{\mu\nu}.$$

Preserving the boundary condition requires the tangential pullback $\iota_\Gamma^*(d\lambda)=0$, so $\lambda$ is locally constant on each connected component of $\Gamma$. Let

$$\begin{aligned} \Pi^i&=F^{i0}, \\ \{A_i(\mathbf x),\Pi^j(\mathbf y)\} &=\delta_i^{\ j}\delta^{(3)}(\mathbf x-\mathbf y). \end{aligned}$$

The differentiable generator suggested by the canonical bracket is

$$\begin{aligned} G[\lambda] &=\int_\Sigma d^3x\,\Pi^i\partial_i\lambda \\ &=-\int_\Sigma d^3x\, \lambda\,\partial_i\Pi^i +\int_{\partial\Sigma} d^2S\, \lambda\Pi^\perp, \end{aligned}$$

where $\Pi^\perp=n_i\Pi^i$ for the outward normal. Indeed, $\{A_i,G[\lambda]\}=\partial_i\lambda$. On the Gauss constraint surface, $\partial_i\Pi^i=0$ in the source-free region, the bulk term vanishes and the generator reduces to the surface electric flux.

If $\lambda|_{\partial\Sigma}=0$, then $G[\lambda]$ equals the smeared Gauss constraint and, after the usual normalization, vanishes throughout the constraint surface. The transformation is a redundant direction. More generally, write $\partial\Sigma=\bigsqcup_a\partial\Sigma_a$ and let $\lambda|_{\partial\Sigma_a}=c_a$. On the Gauss surface the same local transformation can instead have generator

$$Q[\lambda] = \sum_a c_a \int_{\partial\Sigma_a}d^2S\,\Pi^\perp.$$

If this surface functional is nonzero and is allowed to vary, the transformation acts nontrivially on phase space and must not be divided out as an element of $\mathcal G_0$. Harlow and Wu derive this conclusion for the stated finite-boundary Dirichlet system at Harlow and Wu 2020, § 1, pp. 3–4, and § 3.3, pp. 22–23, JHEP PDF.

The qualifications are physical, not cosmetic. A different boundary condition changes which $\lambda$ are allowed and which surface variation makes the generator differentiable. Integrated Gauss law can also relate the fluxes of different boundary components; in a source-free region with a single connected boundary it may force the total flux to vanish. The correct statement is therefore that an admissible nonzero boundary value can carry charge after the boundary phase space is specified, not that every nonzero boundary gauge parameter is automatically a symmetry. Proper and Improper Gauge Transformations develops the general charge distinction.

Gauge fixing selects calculational representatives

The Maxwell action has a degenerate kinetic operator because $A_\mu$ and $A_\mu+\partial_\mu\lambda$ describe the same interior physics when $\lambda\in\mathcal G_0$. In free Abelian theory, add the covariant gauge-fixing term

$$S_{\mathrm{gf}} = -\frac{1}{2\xi} \int d^4x\,(\partial_\mu A^\mu)^2$$

For finite nonzero $\xi$, and after compatible boundary conditions and a Green-function or $i0$ prescription are chosen, this term removes the longitudinal algebraic degeneracy and yields the standard gauge-fixed propagator. The bulk term alone does not remove residual harmonic transformations or boundary zero modes. This is a choice of calculational representative and weighting, not a change in the observable content. In this free Abelian setting, physical gauge-invariant results are required to be independent of $\xi$ when the calculation and boundary conditions are treated consistently.

In Lorenz gauge,

$$\partial_\mu A^\mu=0,$$

a transformed representative remains in the same gauge only if

$$\Box\lambda=0.$$

These are residual transformations, and they remain subject to the same boundary restrictions as the original field space. Gauge fixing has therefore not turned every residual solution into a redundancy or proved that one global slice intersects every orbit once. Schwartz develops the free Maxwell gauge equivalence and covariant gauge-fixing calculation at Schwartz 2014, § 8.2.3, pp. 118–120, § 8.4.2 and §§ 8.5–8.6, pp. 126–132.

This bounded example does not establish the corresponding claims for an interacting non-Abelian theory. The Faddeev–Popov determinant, ghosts, BRST cohomology, and Gribov obstruction enter later, beginning with The Faddeev–Popov Construction.

Four descriptions of the same Maxwell content

The finite-region example can now be read in four compatible languages.

Description	What is imposed or removed	What remains physical
Orbit	Identify $A$ with $A+d\lambda$ for $\lambda\in\mathcal G_0$	The orbit $[A]$, not a chosen potential
Constraint	Impose $\partial_i\Pi^i=0$ and quotient its identity-connected zero-generator flow	Transverse bulk data and allowed boundary data
Gauge-fixed	Select representatives such as $\partial\cdot A=0$	Gauge-invariant results, with residual transformations checked separately
Observable	Use $F$, fluxes, neutral composites, and closed holonomies	Quantities well defined on the physical quotient

No row defines a different theory. The orbit description states the quotient, the constraint description generates its identity-connected infinitesimal part, gauge fixing coordinates it for a calculation, and observables are functions on it. Disconnected identifications require the separate global input already signposted. Agreement among the four descriptions is the central consistency check.

Even in Abelian theory, $F$ is not always a complete coordinate on the orbit space. On a spacetime with noncontractible cycles, flat connections can have different holonomies. Conversely, at a boundary a flux that is invariant under $A\mapsto A+d\lambda$ can also generate a genuine boundary symmetry. Local covariance, global configuration data, and the physical transformation group must therefore be specified together.

Common pitfalls

Calling the non-Abelian curvature gauge invariant. It transforms by conjugation. Invariant contractions such as $\operatorname{tr}(F_{\mu\nu}F^{\mu\nu})$ descend, whereas the matrix $F_{\mu\nu}$ is only covariant.

Quotienting by every formal map to $G$. A transformation must preserve the field space, and it is a redundancy only when its physical generator vanishes. Boundary conditions and disconnected components can change the answer.

Treating a closed holonomy matrix as an observable. It still transforms by conjugation at its base point. A trace or character is invariant; an open line needs endpoint matter, a dressing, or boundary data.

Assuming gauge fixing removes the gauge structure. A gauge condition selects representatives and can leave residual transformations. Globally it can also fail to provide a unique slice.

Assuming curvature exhausts the physics. Flat connections can have nontrivial holonomy, and boundary fluxes can carry charges. Both effects are invisible in a purely pointwise list of curvature components.

Check your understanding

1. Starting from the finite transformation laws, verify $D_\mu^h(h\psi)=hD_\mu\psi$ and $F_{\mu\nu}^h=hF_{\mu\nu}h^{-1}$.
2. Derive the endpoint transformation of $U_\gamma(y,x)$ and explain why a trace makes only a closed transporter invariant.
3. In the Maxwell example, show that $G[\lambda]$ generates $\delta A_i=\partial_i\lambda$. State separately the conditions for it to be a redundant transformation and for it to carry boundary charge. What equation characterizes residual transformations in Lorenz gauge?

Solution

Substitute

$$A_\mu^h=hA_\mu h^{-1} -\frac{i}{g}(\partial_\mu h)h^{-1}$$

into $D_\mu^h(h\psi)$. The derivative term $(\partial_\mu h)\psi$ cancels the inhomogeneous connection term, leaving $h(\partial_\mu-igA_\mu)\psi$. Conjugating the commutator then gives

$$[D_\mu^h,D_\nu^h] =h[D_\mu,D_\nu]h^{-1},$$

so $F_{\mu\nu}^h=hF_{\mu\nu}h^{-1}$.

Parallel transport maps a vector in the fiber at $x$ to one at $y$. Changing the local frame at the two endpoints therefore gives

$$U_\gamma^h(y,x) =h(y)U_\gamma(y,x)h(x)^{-1}.$$

For a closed path, $x=y$, this is conjugation by one matrix, whose trace is unchanged. For an open path the two endpoint matrices cannot be removed by a trace because they act in different endpoint fibers.

Finally, the canonical bracket gives

$$\{A_i(\mathbf x),G[\lambda]\} =\partial_i\lambda(\mathbf x).$$

After integrating by parts, $G[\lambda]$ is the Gauss constraint plus the surface term. It is redundant when $\lambda$ preserves the boundary data and the complete generator has vanishing variation throughout the allowed constraint surface, with any fixed constant normalized away. For the stated Dirichlet example, this includes $\lambda|_{\partial\Sigma}=0$. If flux variations are allowed and admissible boundary constants produce a nonzero, variable surface functional, it is instead a charge. Lorenz gauge leaves $\Box\lambda=0$, together with the original boundary restrictions.

What to carry forward

Gauge Orbits, Gauss Constraints, and Stabilizers turns the quotient into a systematic constraint reduction. Local Potentials and Global Gauge Configurations restores bundle patching and topological sectors, while Gauge-Invariant and Dressed Observables develops charged and extended observables. Which Wilson representations are allowed and genuine depends on Global Form, Matter Representations, and the Faithful Gauge Group together with Genuine Line Spectra, Discrete Theta Data, and Theory Specification. Interacting dynamics and matter content belong to Dynamical Gauge Fields and Matter. For a shorter conceptual bridge from the free vector field, continue with Vector Fields and Gauge Redundancy.

References

• Harlow, Daniel, and Jie-qiang Wu. “Covariant Phase Space with Boundaries.” Journal of High Energy Physics 10 (2020): 146. DOI. Open PDF
• Schwartz, Matthew D. Quantum Field Theory and the Standard Model. First edition. Cambridge: Cambridge University Press, 2014. DOI
• Weinberg, Steven. The Quantum Theory of Fields, Volume II: Modern Applications. First edition; 2005 Cambridge paperback printing consulted. Cambridge: Cambridge University Press, 1996. DOI