Ghost-Free Renormalization Group Flows for Gauge Fields from Stochastically Gauge-Fixed Langevin Dynamics

Gauge theories such as the theory of quarks and gluons contain a built-in redundancy: many field configurations describe the same physics. Any calculation of the basic building blocks, the gluon propagator and the gluon interaction vertices, must first pick one representative from each family of equivalent configurations. This choice is called gauge fixing. The functional renormalization group (FRG) is a method that adds quantum fluctuations one momentum scale at a time and solves the resulting equations without assuming weak coupling. In its standard form it fixes the gauge with Faddeev-Popov "ghost" fields. The momentum cutoff it relies on then breaks the gauge symmetry at every intermediate scale. The repair, a tower of modified symmetry identities, is so hard to enforce that most studies solve only the transverse part of the gluon field. They also leave open how the many gauge representatives that satisfy the same gauge condition (Gribov copies) should be treated. A separate community, building lattice perturbation theory on the computer, met a different symptom of the same redundancy. Their Langevin simulations, random-walk dynamics whose long-time statistics reproduce the quantum path integral, drifted without bound along the redundant directions until noise swamped the signal. They fixed this with stochastic gauge fixing, Zwanziger's idea of adding a restoring force that pushes the field back toward a gauge condition. Because this force acts purely as a gauge transformation, it cannot change any physical answer.

This note proposes to write the FRG for the Langevin process rather than for the gauge-fixed path integral. Two things follow. First, the gauge is fixed by dynamics, so no ghost fields and no Faddeev-Popov determinant appear: with Itô time-stepping, the Jacobian of the field-theoretic rewriting of the Langevin equation is a constant. Second, Zwanziger's construction contains a free knob, the strength $\alpha$ of the restoring force. Physical observables must not depend on it. That single exact statement gives a cheap, scalar check on every truncation. It also comes with an oracle: numerical stochastic perturbation theory (NSPT) computes, order by order in the coupling, exactly the same gauge-fixed stochastic process. For the first time an FRG truncation can then be tested against high-order perturbation theory in the very gauge it uses. Lattice comparisons in Landau gauge cannot offer this, because they handle Gribov copies differently.

Core Claim

The object is Yang-Mills Langevin dynamics in an auxiliary time $\tau$, with Zwanziger's drift and an infrared regulator that acts only on transverse modes:

$$\partial_\tau A_\mu = -\frac{\delta S_{\rm YM}}{\delta A_\mu} + D_\mu v - (R_k A^{T})_\mu + \eta_\mu, \qquad v = \tfrac{1}{\alpha}\,\partial_\nu A_\nu .$$

Here $A_\mu$ is the gluon field, $S_{\rm YM}$ the Yang-Mills action, $D_\mu$ the covariant derivative, so $D_\mu v$ is an infinitesimal gauge transformation. Its sign is chosen so that it drives the field toward the Landau slice $\partial_\nu A_\nu = 0$. $R_k$ is a momentum-dependent mass that suppresses transverse modes $A^T$ below scale $k$, and $\eta$ is Gaussian white noise. We rewrite this equation as a path integral over the field and a response field, following the Martin-Siggia-Rose-de Dominicis-Janssen construction used by Self-consistent formulations for stochastic nonlinear neuronal dynamics. That paper derived Wetterich-type flows with Blaizot-Méndez-Wschebor truncations for non-gauge Langevin systems. We then derive the exact flow of the scale-dependent effective action $\Gamma_k$ for the stationary state.

The claim is that this flow, truncated at the level now standard in Landau-gauge FRG, has three properties. It reproduces the decoupling-type gluon propagator without ghosts. The longitudinal sector becomes dynamical and is controlled by $\alpha$ rather than discarded. And the residual $\alpha$-dependence of gauge-invariant observables,

$$\Delta_k(\alpha) = \partial_\alpha \langle O \rangle_k ,$$

is a quantitative measure of truncation error. Here $O$ is a gauge-invariant observable such as the plaquette or a Wilson loop. $\Delta_k$ must vanish at $k=0$ in the exact theory, because $D_\mu v$ is a gauge transformation. For $k>0$ it is driven only by the transverse regulator. This one number replaces the full set of modified symmetry identities as the first diagnostic. It does not replace them as a complete constraint, and we do not claim it does.

In the abelian case the construction reduces to covariant gauge with parameter $\xi = \alpha$, which gives an exact sanity check. In the non-abelian case the stationary distribution in gauge directions is not the Faddeev-Popov weight. Zwanziger's drift is argued to confine trajectories to the first Gribov region. The flow therefore makes a definite prediction about how Gribov copies are weighted, and that prediction can be compared with lattice data.

How This Relates to Prior Work

Landau-gauge FRG supplies the truncation technology and the main competitor. Landau gauge Yang-Mills correlation functions computes gluon and ghost propagators and three- and four-gluon vertices self-consistently. It explicitly leaves the longitudinal sector and Gribov copies as open limitations. Nonperturbative finite-temperature Yang-Mills theory extends this to finite temperature. It reaches 10% agreement with lattice data for the magnetic gluon propagator and flags basis artifacts and missing backgrounds as limits. Nonperturbative quark, gluon, and meson correlators of unquenched QCD adds quarks. All three fix the gauge with ghosts. We keep their vertex truncations and replace the gauge-fixing mechanism.

The stochastic gauge-fixing papers supply both the mechanism and the oracle. Lattice perturbation theory on the computer showed that Zwanziger's drift turns exponentially growing noise into a flat plateau. That made the fourth-order Wilson-loop coefficient measurable. Developments and new applications of numerical stochastic perturbation theory extended NSPT to gauge-fixed correlators and checked internal consistency with Dyson-Schwinger equations. New issues for numerical stochastic perturbation theory solved the inverse Faddeev-Popov operator recursively. It also reported that background-field (Schrödinger functional) observables were too noisy. That result is a constraint: we stay at zero background in this proposal. NSPT is perturbative and has no scale-dependent effective action. We use it only as an order-by-order reference for our flow, not as a competitor in the strong-coupling regime.

The neural-dynamics FRG paper supplies the dynamical formalism. Its systems have no gauge redundancy, which is precisely the ingredient we add. The machine-learning samplers in the file sidestep gauge fixing through equivariance, for example Scaling of stochastic normalizing flows in SU(3) lattice gauge theory. They target gauge-invariant sampling and do not address gauge-fixed correlation functions, which are our subject.

Research Program

The first stage is the abelian test bed: compact scalar QED and the free and weakly interacting U(1) theory, both of which NSPT has already handled to order $g^3$. We derive the stochastic flow and check two things analytically. The stationary longitudinal propagator should equal the covariant-gauge one at $\xi=\alpha$, and $\Delta_k(\alpha)$ should vanish at $k=0$ to the accuracy of the truncation. This stage settles which time-discretization (Itô) and which regulator placement, transverse only or including the response field, keep the flow ghost-free and stable.

The second stage is the non-abelian derivation for SU(2) and SU(3). Starting from the response-field action, we derive flows for the transverse and longitudinal gluon two-point functions and the three- and four-gluon vertices, in both their static and frequency-dependent parts. The symbolic derivation will reuse the DoFun/FormTracer toolchain behind the fQCD codes. The extra cost relative to standard FRG is one frequency integral per loop. Stationarity lets us close it analytically in the lowest-order truncation and numerically beyond that.

The third stage solves the flow both on a finite periodic lattice-momentum grid ($8^4$, $12^4$), matching NSPT, and in the continuum at infinite volume. We do this for several values of $\alpha$, and we track $\Delta_k(\alpha)$ along the flow as a function of truncation level.

Baselines

The naive reference point is lowest-order perturbation theory in stochastic Landau gauge. This means the tree-level gluon propagator and the one-loop plaquette coefficient. It fixes the floor: any flow that cannot beat one-loop accuracy on the second-order Wilson-loop coefficient on an $8^4$ lattice is broken.

The perturbative ground truth is NSPT with stochastic gauge fixing, run by us on the same lattices ($8^4$ and $12^4$), the same gauge group and the same $\alpha$ values as the flow. We take the gauge-fixed gluon propagator coefficients through order $g^4$ and the Wilson-loop coefficients $c_1$ through $c_4$. The anchoring number is the one reported in Lattice perturbation theory on the computer: $c_4 = 1.211 \pm 0.003$ for the $1\times1$ Wilson loop on an $8^4$ lattice. Our NSPT runs must reproduce this before we use them. The comparison is fair because both methods target exactly the same stochastic process. The flow, expanded order by order in $g$, must reproduce $c_1$ and $c_2$ exactly. Its deviation from $c_3$ and $c_4$ measures truncation error. To count as a success, that deviation must be no larger than that of the standard ghost-based FRG truncation of equal vertex content, expanded the same way.

The strongest non-perturbative competitor is standard Landau-gauge FRG with Faddeev-Popov ghosts, as in Landau gauge Yang-Mills correlation functions and Nonperturbative finite-temperature Yang-Mills theory. The latter reports agreement with lattice data within 10% for the magnetic gluon propagator. For the zero-temperature SU(3) gluon propagator we found no reported result on Lacuna giving a numerical deviation from lattice data. We will therefore re-run the ghost-based truncation ourselves with identical vertex content, regulator shape, momentum grid and solver tolerances. Both will be compared against the non-perturbative reference below. Beating it means a smaller maximum relative deviation of the transverse propagator across momenta, at equal truncation and comparable compute.

The non-perturbative reference is a full, non-expanded Langevin lattice simulation of SU(2) and SU(3) with Zwanziger's drift on $16^4$ lattices at several couplings. This is standard lattice Langevin with the restoring force, not NSPT. It gives the gluon propagator in exactly the stochastic gauge the flow computes. It also gives an $\alpha$-scan of the plaquette that should be flat within statistics.

The key ablation removes the stochastic formulation but keeps everything else. We run the standard ghost-based flow in linear covariant gauge at $\xi = \alpha$ with identical truncation, and evaluate its gauge-parameter dependence of the plaquette in place of $\Delta_k(\alpha)$. If our $\Delta_k$ is no more informative about truncation error than this $\xi$-dependence, the stochastic ingredient is not pulling its weight.

Why not just…?

Why not just simulate the Zwanziger-fixed Langevin process on the lattice and read off the propagator? For the regime this proposal can test, that simulation is the ground truth, and we run it as a baseline. It is a real threat as a source of numbers. The flow earns its place only through what simulation cannot easily give: infinite volume and the deep infrared without finite-size extrapolation, an analytic handle on how the longitudinal sector and the Gribov-region restriction shape the propagator, and a route to settings where lattice sampling degrades. The file documents that $n_{\rm step} \sim (L/a)^4$ for state-of-the-art SU(3) samplers in Scaling of stochastic normalizing flows in SU(3) lattice gauge theory. If the flow does not match the simulation at $16^4$, the proposal fails, and we will say so.

Why not just keep Faddeev-Popov ghosts and check gauge-parameter dependence in covariant gauges? This is the closest alternative, and it is our key ablation. The difference is that the $\xi$-dependence of ghost-based flows mixes truncation error with the ghost sector and with the regulator's breaking of the symmetry between ghosts and gluons. Our $\alpha$-dependence involves no ghosts and is driven by the transverse regulator alone. Whether that makes it the cleaner diagnostic is an empirical question. The test is to correlate each diagnostic, across truncation levels, with the true deviation from NSPT's $c_3$ and $c_4$ and from the Langevin propagator, and to see which tracks the error better.

Why not just use a simpler equal-time (Fokker-Planck) formulation instead of the dynamical response-field one? The stationary distribution of the Zwanziger process has no closed form in the non-abelian case. That is exactly why Gribov-region effects enter, and an equal-time flow would need that unknown weight as input. The dynamical formulation avoids this at the cost of a frequency integral. In the abelian stage we will run both and quantify the overhead.

Why not just use mean-field or one-loop self-consistent approximations? The file documents that freezing fluctuations produces spurious phase structure, in Quark-meson-diquark model for two-color QCD, and that first-order expansions fail near critical points, in Transitions between asynchronous and synchronous states: a theory of correlations in small neural circuits. The one-loop baseline above makes the size of this failure explicit for the gluon propagator.

Proposed Experiments

The first decisive check is abelian exactness. In U(1) and compact scalar QED, the stochastic flow must reproduce the covariant-gauge longitudinal propagator at $\xi=\alpha$, and must match NSPT coefficients through order $g^3$ at the statistical precision of our runs. Failure here means the formulation or discretization is wrong.

The second check is order-by-order agreement with NSPT for SU(3) on $8^4$. We reproduce $c_4 = 1.211 \pm 0.003$ with our NSPT runs, then expand the truncated stochastic flow and the truncated ghost-based flow to fourth order. We compare both to $c_1$ through $c_4$ and to the gauge-fixed propagator coefficients. The stochastic flow must be exact at orders one and two and at least as accurate as the ghost-based truncation at orders three and four.

The third check is non-perturbative agreement and the infrared. On $16^4$ SU(2) and SU(3), we compare the flow's transverse and longitudinal propagators with the full Zwanziger-Langevin simulation at three or more values of $\alpha$. We then extrapolate the flow to infinite volume to determine whether the deep-infrared solution is of decoupling type and how its gluon mass scale depends on $\alpha$. In the exact theory the transverse propagator may depend on $\alpha$, but gauge-invariant quantities may not.

The fourth check is diagnostic validity. Across a ladder of truncations, we test whether $\Delta_k(\alpha)$ at $k=0$, evaluated for the plaquette, tracks the measured deviation from the Langevin and NSPT references better than the ghost-based flow's $\xi$-dependence does. The proposal is supported if the stochastic flow matches the references with no more error than the ghost-based flow, while dropping ghosts and solving the longitudinal sector, and if $\Delta_k$ proves a reliable error estimate. It is refuted if the frequency integrals make the truncation less accurate at equal cost, or if $\Delta_k$ stays small while the propagator is wrong.