Gauge Flow Models Overview

• Gauge Flow Models are generative flow models that embed learnable gauge fields to enforce gauge invariance and structured dynamics.
• They utilize Riemannian Flow Matching with gauge-theoretic correction terms, consistently outperforming standard flow models in synthetic experiments.
• Extensions include higher, tensor, and gauge-equivariant sampling frameworks, applicable to lattice gauge theory, Yang–Mills gradient flow, and anomaly inflow studies.

Searching arXiv for the most relevant papers on Gauge Flow Models and closely related gauge-flow frameworks. arXiv search query: "Gauge Flow Models generative flow models gauge-equivariant flow models lattice gauge theory gradient flow" Gauge Flow Models names several distinct constructions in current research rather than a single formalism. The phrase appears most narrowly as the title of a class of generative flow models that place a learnable Gauge Field inside a Flow Ordinary Differential Equation, and more broadly across lattice gauge theory, Yang–Mills gradient flow, BRST/BV canonical flow, gauge-invariant functional flow equations, and bulk gauge-field extensions used in nonperturbative chiral gauge theory. In all of these settings, the flow is not treated as an unconstrained evolution: gauge structure, gauge invariance, gauge equivariance, or gauge-parameter dependence is built into the dynamics from the outset (Strunk et al., 17 Jul 2025, Abbott et al., 2022, Lüscher et al., 2011, Quadri, 2014, Dang et al., 3 Jun 2026).

1. Terminological scope

A concise way to organize the literature is to distinguish the major uses of the term by the object that flows and the role played by gauge structure.

Usage in the literature	Representative mechanism	Representative papers
Generative Gauge Flow Models	a learnable Gauge Field within the Flow ODE	(Strunk et al., 17 Jul 2025, Strunk et al., 22 Jul 2025, Strunk et al., 18 Nov 2025)
Gauge-equivariant flow-based sampling	invertible maps between simple priors and lattice gauge-field or gauge-plus-pseudofermion distributions	(Kanwar et al., 2020, Abbott et al., 2022, Abbott et al., 2022)
Yang–Mills gradient flow	a local diffusion equation in fictitious flow time $t$	(Lüscher et al., 2011, Asakawa et al., 2013, Ramos et al., 2015, Hirakida et al., 2018)
Canonical, functional, and bulk gauge-field flows	canonical flow in gauge-parameter space, gauge-invariant FRG flow, and extra-dimensional gauge-field extension	(Quadri, 2014, Wetterich, 2017, Dang et al., 3 Jun 2026, Dang et al., 3 Jun 2026)

These usages share a common insistence that gauge redundancy is not an afterthought. In the generative and sampling papers, gauge symmetry is an inductive bias or an exact architectural constraint. In the Yang–Mills and FRG papers, flow time or RG time is used to define renormalized observables, gauge-invariant representatives, or controlled dependence on gauge fixing. In the chiral-gauge constructions, gauge fields are extended into an auxiliary dimension by a prescribed flow so that anomaly inflow and current conservation remain well defined.

2. Generative Gauge Flow Models as a named model class

The 2025 paper "Gauge Flow Models" defines Gauge Flow Models as a class of generative flow models that augment a standard flow ODE with a learnable gauge-theoretic correction term. The basic dynamics replace the standard velocity field equation

$\frac{dx(t)}{dt} = v_\theta(x(t), t)$

by

$$\hat{\nabla}_{dt} x(t): = v_{\theta}(x(t), t) - \alpha(t)\,\Pi_{M}\!\Big( A_{\mu \nu}(x(t), t)\, d^{\mu}(x(t), t)\, v^{\nu}(x(t), t) \Big).$$

The model is formulated on the geometry of principal bundles, associated bundles, and connections. The associated bundle is

$\hat{A} = P \times_G F,$

the tangent space of the total space splits as

$$T_pP = H_pP \oplus V_pP,\qquad V_pP := \ker(d\pi_p),$$

and the connection one-form satisfies

$$\omega \in \Omega^1(P)\otimes \mathfrak g,\qquad H_pP=\ker(\omega_p),$$

with the local gauge potential

$A := s^*\omega \in \Omega^1(U)\otimes\mathfrak g,$

transforming as

$A'(x)=g^{-1}(x)A(x)g(x)+g^{-1}(x)\,dg(x),$

and curvature

$F = dA + A\wedge A.$

In this formulation, the novelty is the inclusion of a gauge term absent in standard flow models, and the paper presents it as a geometric inductive bias rather than only an architectural enlargement (Strunk et al., 17 Jul 2025).

Training is based on Flow Matching, specifically Riemannian Flow Matching. The Gauge Flow Model loss is

$$\mathcal{L}_{\mathrm{GFM}} = \mathbb{E}_{t\sim\mathcal U[0,1],\,x\sim p_t} \left\| \Big[ v_{\theta}(x(t), t) - \alpha(t)\Pi_M\big(A_{\mu\nu}(x(t),t)\,d^\mu(x(t),t)\,v^\nu(x(t),t)\big) \Big] - u_t(x) \right\|_{g_x}^2.$$

The experiments specialize to a trivial principal bundle $\frac{dx(t)}{dt} = v_\theta(x(t), t)$0 with $\frac{dx(t)}{dt} = v_\theta(x(t), t)$1 and $\frac{dx(t)}{dt} = v_\theta(x(t), t)$2. The training set has 15,000 samples and the test set has 5,000 samples, drawn from a mixture of $\frac{dx(t)}{dt} = v_\theta(x(t), t)$3 equally weighted isotropic Gaussian components in dimensions $\frac{dx(t)}{dt} = v_\theta(x(t), t)$4, with covariance $\frac{dx(t)}{dt} = v_\theta(x(t), t)$5 and spread parameter $\frac{dx(t)}{dt} = v_\theta(x(t), t)$6. The reported result is that both Gauge Flow variants significantly outperform the plain Flow Model across all dimensions $\frac{dx(t)}{dt} = v_\theta(x(t), t)$7, while the plain Flow Model is slightly larger in parameter count (Strunk et al., 17 Jul 2025).

Two immediate generalizations were introduced in 2025. "Higher Gauge Flow Models" replaces the ordinary gauge structure by an $\frac{dx(t)}{dt} = v_\theta(x(t), t)$8-algebra. The gauge field becomes $\frac{dx(t)}{dt} = v_\theta(x(t), t)$9-algebra valued, the auxiliary vector becomes a graded vector, and the gauge correction is mediated by the higher brackets $$\hat{\nabla}_{dt} x(t): = v_{\theta}(x(t), t) - \alpha(t)\,\Pi_{M}\!\Big( A_{\mu \nu}(x(t), t)\, d^{\mu}(x(t), t)\, v^{\nu}(x(t), t) \Big).$$0. The paper states that this lets the model integrate “higher geometry and higher symmetries associated with higher groups” into the generative modeling framework, and reports that Higher Gauge Flow Models consistently outperform both Gauge Flow Models and Plain Flow Models on synthetic Gaussian Mixture Model datasets in dimensions $$\hat{\nabla}_{dt} x(t): = v_{\theta}(x(t), t) - \alpha(t)\,\Pi_{M}\!\Big( A_{\mu \nu}(x(t), t)\, d^{\mu}(x(t), t)\, v^{\nu}(x(t), t) \Big).$$1 (Strunk et al., 22 Jul 2025). "Tensor Gauge Flow Models" generalizes both ordinary Gauge Flow Models and Higher Gauge Flow Models by replacing a 1-form-like correction with higher-order Tensor Gauge Fields $$\hat{\nabla}_{dt} x(t): = v_{\theta}(x(t), t) - \alpha(t)\,\Pi_{M}\!\Big( A_{\mu \nu}(x(t), t)\, d^{\mu}(x(t), t)\, v^{\nu}(x(t), t) \Big).$$2. In the reported Euclidean experiments on Gaussian mixtures with $$\hat{\nabla}_{dt} x(t): = v_{\theta}(x(t), t) - \alpha(t)\,\Pi_{M}\!\Big( A_{\mu \nu}(x(t), t)\, d^{\mu}(x(t), t)\, v^{\nu}(x(t), t) \Big).$$3, $$\hat{\nabla}_{dt} x(t): = v_{\theta}(x(t), t) - \alpha(t)\,\Pi_{M}\!\Big( A_{\mu \nu}(x(t), t)\, d^{\mu}(x(t), t)\, v^{\nu}(x(t), t) \Big).$$4, and $$\hat{\nabla}_{dt} x(t): = v_{\theta}(x(t), t) - \alpha(t)\,\Pi_{M}\!\Big( A_{\mu \nu}(x(t), t)\, d^{\mu}(x(t), t)\, v^{\nu}(x(t), t) \Big).$$5, Tensor Gauge Flow Models achieve the lowest training loss across all tested dimensions $$\hat{\nabla}_{dt} x(t): = v_{\theta}(x(t), t) - \alpha(t)\,\Pi_{M}\!\Big( A_{\mu \nu}(x(t), t)\, d^{\mu}(x(t), t)\, v^{\nu}(x(t), t) \Big).$$6 and the lowest test loss across all tested dimensions $$\hat{\nabla}_{dt} x(t): = v_{\theta}(x(t), t) - \alpha(t)\,\Pi_{M}\!\Big( A_{\mu \nu}(x(t), t)\, d^{\mu}(x(t), t)\, v^{\nu}(x(t), t) \Big).$$7 (Strunk et al., 18 Nov 2025).

The current empirical base for this named model class is therefore synthetic and Flow Matching centered. The papers themselves leave open questions about identifiability, universality, optimization stability, and how strongly practical performance depends on the choice of gauge group, bundle structure, tensor rank, and auxiliary fields. This suggests that the 2025 line is best read as a geometric reformulation of continuous-time generative modeling rather than a settled mature methodology.

3. Gauge-equivariant flow-based sampling in lattice field theory

A separate and earlier line of work uses normalizing flows to sample lattice gauge theories while enforcing gauge symmetry exactly. In this setting one learns an invertible map from a simple prior to a target path-integral distribution. With latent variables $$\hat{\nabla}_{dt} x(t): = v_{\theta}(x(t), t) - \alpha(t)\,\Pi_{M}\!\Big( A_{\mu \nu}(x(t), t)\, d^{\mu}(x(t), t)\, v^{\nu}(x(t), t) \Big).$$8 and a diffeomorphism $$\hat{\nabla}_{dt} x(t): = v_{\theta}(x(t), t) - \alpha(t)\,\Pi_{M}\!\Big( A_{\mu \nu}(x(t), t)\, d^{\mu}(x(t), t)\, v^{\nu}(x(t), t) \Big).$$9, the model density is

$\hat{A} = P \times_G F,$0

while the target distribution is

$\hat{A} = P \times_G F,$1

Gauge symmetry is central because the target distribution is gauge invariant, so a non-equivariant model would waste capacity learning redundant gauge copies and could fail to generalize across gauge orbits. In the 2020 $\hat{A} = P \times_G F,$2 construction, each coupling layer is made gauge equivariant, a uniform Haar prior is used, and the transformed link update is built from closed-loop or gauge-covariant combinations so that symmetry is respected by construction (Kanwar et al., 2020).

The first numerical demonstration was two-dimensional $\hat{A} = P \times_G F,$3 lattice gauge theory on $\hat{A} = P \times_G F,$4 lattices with $\hat{A} = P \times_G F,$5. The architecture used 24 gauge-equivariant coupling layers, kernels implemented as mixtures of Non-Compact Projections, and 6 mixture components per kernel. The key result was not uniform improvement over all observables, but orders-of-magnitude gains for topology-sensitive quantities: for the finest lattice/strongest critical regime tested, the flow-based sampler had $\hat{A} = P \times_G F,$6, compared with $\hat{A} = P \times_G F,$7 for Heat Bath and $\hat{A} = P \times_G F,$8 for HMC, and after accounting for per-step cost the flow-based Metropolis sampler was approximately 1500× more efficient than HMC and 200× more efficient than Heat Bath for determining topological quantities (Kanwar et al., 2020).

The extension to dynamical fermions introduces pseudofermions as stochastic estimators for the fermion determinant. For positive definite $\hat{A} = P \times_G F,$9,

$$T_pP = H_pP \oplus V_pP,\qquad V_pP := \ker(d\pi_p),$$0

and for $$T_pP = H_pP \oplus V_pP,\qquad V_pP := \ker(d\pi_p),$$1 the joint target distribution can be factorized as $$T_pP = H_pP \oplus V_pP,\qquad V_pP := \ker(d\pi_p),$$2. The learned model correspondingly factorizes as

$$T_pP = H_pP \oplus V_pP,\qquad V_pP := \ker(d\pi_p),$$3

The central architectural development is the parallel transport convolutional network (PTCN), which parallel transports neighboring fields to a common site before combining them, thereby making the pseudofermion flow gauge equivariant. This framework was demonstrated in the Schwinger model and in two-dimensional $$T_pP = H_pP \oplus V_pP,\qquad V_pP := \ker(d\pi_p),$$4 gauge theory with $$T_pP = H_pP \oplus V_pP,\qquad V_pP := \ker(d\pi_p),$$5 fermions, and it explicitly incorporates even/odd preconditioning, Hasenbusch factorization, and multiple pseudofermion samples per gauge field (Abbott et al., 2022).

The QCD status update then combines the marginal gauge-field model and the conditional pseudofermion model into a four-dimensional proof-of-principle algorithm. The key pieces are: a marginal flow for gauge fields $$T_pP = H_pP \oplus V_pP,\qquad V_pP := \ker(d\pi_p),$$6, a conditional flow for pseudofermions $$T_pP = H_pP \oplus V_pP,\qquad V_pP := \ker(d\pi_p),$$7 conditioned on $$T_pP = H_pP \oplus V_pP,\qquad V_pP := \ker(d\pi_p),$$8, a joint model $$T_pP = H_pP \oplus V_pP,\qquad V_pP := \ker(d\pi_p),$$9, self-training of the flow parameters using stochastic estimates of the reverse KL objective, independence Metropolis acceptance/reweighting, and stochastic estimation of the fermion determinant with multiple pseudofermion samples per gauge field. The numerical demonstration is for $$\omega \in \Omega^1(P)\otimes \mathfrak g,\qquad H_pP=\ker(\omega_p),$$0 fermion flavors on a $$\omega \in \Omega^1(P)\otimes \mathfrak g,\qquad H_pP=\ker(\omega_p),$$1 lattice at strong coupling $$\omega \in \Omega^1(P)\otimes \mathfrak g,\qquad H_pP=\ker(\omega_p),$$2 and $$\omega \in \Omega^1(P)\otimes \mathfrak g,\qquad H_pP=\ker(\omega_p),$$3. The marginal model uses a Haar-uniform prior and 48 gauge-equivariant spline coupling layers; the conditional pseudofermion model uses a Gaussian prior and 36 pseudofermion layers built from PTCNs. The trained flow model generates gauge ensembles that are statistically consistent with HMC at the chosen test point. With 512 pseudofermion draws, the effective sample size is about 0.1, compared with $$\omega \in \Omega^1(P)\otimes \mathfrak g,\qquad H_pP=\ker(\omega_p),$$4 for direct reweighting from the base distribution, and comparisons with HMC for the plaquette, Polyakov loop, pion correlation function, and topological charge agree within statistical uncertainties (Abbott et al., 2022).

The same paper is explicit that this is a proof-of-principle demonstration rather than a production-ready QCD algorithm. It states that the space of architectures and training strategies has barely been explored for QCD; there is no reliable answer yet on whether flow-based sampling will outperform HMC at realistic scale; performance depends strongly on the chosen flow architecture, training schedule, stopping criterion, and parameter regime; and real-world use will likely require hybrid algorithms and substantial engineering tightly integrated with existing lattice QCD codes on HPC systems (Abbott et al., 2022).

4. Yang–Mills gradient flow as a gauge-flow framework

Another major meaning of gauge flow in the literature is the Yang–Mills gradient flow. In continuum form it evolves a gauge field $$\omega \in \Omega^1(P)\otimes \mathfrak g,\qquad H_pP=\ker(\omega_p),$$5 into a one-parameter family $$\omega \in \Omega^1(P)\otimes \mathfrak g,\qquad H_pP=\ker(\omega_p),$$6 by

$$\omega \in \Omega^1(P)\otimes \mathfrak g,\qquad H_pP=\ker(\omega_p),$$7

The flow is local, diffusion-like, and gauge covariant; high-momentum modes are exponentially suppressed by the heat-kernel factors in the integral representation. The central perturbative theorem is that, once the ordinary four-dimensional gauge theory is renormalized in the usual way, correlation functions of the flowed field $$\omega \in \Omega^1(P)\otimes \mathfrak g,\qquad H_pP=\ker(\omega_p),$$8 at positive flow time $$\omega \in \Omega^1(P)\otimes \mathfrak g,\qquad H_pP=\ker(\omega_p),$$9 are finite to all orders and require no additional renormalization. In the $A := s^*\omega \in \Omega^1(U)\otimes\mathfrak g,$0-dimensional formulation the retarded structure eliminates genuine bulk loops, and BRS symmetry excludes the relevant boundary counterterms (Lüscher et al., 2011).

The thermodynamic application that made this framework especially influential is the reconstruction of the renormalized energy-momentum tensor from flowed operators. For pure $A := s^*\omega \in \Omega^1(U)\otimes\mathfrak g,$1 gauge theory, the relevant gauge-invariant flowed operators are

$A := s^*\omega \in \Omega^1(U)\otimes\mathfrak g,$2

and

$A := s^*\omega \in \Omega^1(U)\otimes\mathfrak g,$3

Their small-flow-time expansion yields the renormalized EMT,

$A := s^*\omega \in \Omega^1(U)\otimes\mathfrak g,$4

This provides a direct operator-based determination of $A := s^*\omega \in \Omega^1(U)\otimes\mathfrak g,$5 and $A := s^*\omega \in \Omega^1(U)\otimes\mathfrak g,$6 and avoids the traditional integral method. The 2013 SU(3) thermodynamics study used $A := s^*\omega \in \Omega^1(U)\otimes\mathfrak g,$7 lattices, the Wilson plaquette gauge action, temperatures $A := s^*\omega \in \Omega^1(U)\otimes\mathfrak g,$8, and a double limit $A := s^*\omega \in \Omega^1(U)\otimes\mathfrak g,$9 then $A'(x)=g^{-1}(x)A(x)g(x)+g^{-1}(x)\,dg(x),$0 inside the fiducial window $A'(x)=g^{-1}(x)A(x)g(x)+g^{-1}(x)\,dg(x),$1. The final continuum results agree with earlier integral-method determinations within errors (Asakawa et al., 2013).

Because gradient flow observables are widely used for precision lattice calculations, cutoff effects became a separate subject. The Symanzik improvement analysis in the 4+1-dimensional local formulation shows that the classical nature of the flow equation allows all cutoff effects at $A'(x)=g^{-1}(x)A(x)g(x)+g^{-1}(x)\,dg(x),$2 originating from the discretized gradient flow equation or the gradient flow observable to be eliminated. The paper’s choice is the Zeuthen flow,

$A'(x)=g^{-1}(x)A(x)g(x)+g^{-1}(x)\,dg(x),$3

together with the tree-level improved observable

$A'(x)=g^{-1}(x)A(x)g(x)+g^{-1}(x)\,dg(x),$4

After improving flow and observable, the remaining $A'(x)=g^{-1}(x)A(x)g(x)+g^{-1}(x)\,dg(x),$5 effects arise from boundary counterterms localized at flow time $A'(x)=g^{-1}(x)A(x)g(x)+g^{-1}(x)\,dg(x),$6; compared to the 4-dimensional pure gauge theory only a single additional counterterm is required, corresponding to a modified initial condition for the flow equation (Ramos et al., 2015).

The same EMT technology was then applied to pure $A'(x)=g^{-1}(x)A(x)g(x)+g^{-1}(x)\,dg(x),$7 thermodynamics. In that setting a reference scale is proposed by

$A'(x)=g^{-1}(x)A(x)g(x)+g^{-1}(x)\,dg(x),$8

the scale-setting function $A'(x)=g^{-1}(x)A(x)g(x)+g^{-1}(x)\,dg(x),$9 is fitted over $F = dA + A\wedge A.$0, and the equation of state is extracted from the flowed EMT on lattices with $F = dA + A\wedge A.$1. The paper describes this as the first application of the gradient-flow EMT method to SU(2) thermodynamics and reports that the equation of state is obtained with reasonably small errors (Hirakida et al., 2018).

5. Canonical, gauge-invariant, and parameter-space flows

In algebraic gauge theory, “flow” can refer neither to fictitious time smoothing nor to generative transport, but to controlled motion in the space of gauge parameters. The paper "Canonical Flow in the Space of Gauge Parameters" extends BRST symmetry by treating the gauge parameter $F = dA + A\wedge A.$2 as part of a BRST doublet,

$F = dA + A\wedge A.$3

and shows that the dependence of the 1-PI functional $F = dA + A\wedge A.$4 on $F = dA + A\wedge A.$5 is generated by a canonical transformation with respect to the BV bracket. The extended Slavnov–Taylor identity is written as

$F = dA + A\wedge A.$6

and differentiating with respect to $F = dA + A\wedge A.$7 yields

$F = dA + A\wedge A.$8

Because $F = dA + A\wedge A.$9 itself depends on $$\mathcal{L}_{\mathrm{GFM}} = \mathbb{E}_{t\sim\mathcal U[0,1],\,x\sim p_t} \left\| \Big[ v_{\theta}(x(t), t) - \alpha(t)\Pi_M\big(A_{\mu\nu}(x(t),t)\,d^\mu(x(t),t)\,v^\nu(x(t),t)\big) \Big] - u_t(x) \right\|_{g_x}^2.$$0, the full dependence is reconstructed by a Lie series rather than naive exponentiation,

$$\mathcal{L}_{\mathrm{GFM}} = \mathbb{E}_{t\sim\mathcal U[0,1],\,x\sim p_t} \left\| \Big[ v_{\theta}(x(t), t) - \alpha(t)\Pi_M\big(A_{\mu\nu}(x(t),t)\,d^\mu(x(t),t)\,v^\nu(x(t),t)\big) \Big] - u_t(x) \right\|_{g_x}^2.$$1

For the transverse gluon propagator this becomes a multiplicative evolution equation bridging the Landau gauge and a general linear covariant gauge (Quadri, 2014).

A related but distinct usage appears in the gauge-invariant functional renormalization group. The construction begins by decomposing infinitesimal gauge-field fluctuations into physical and gauge parts,

$$\mathcal{L}_{\mathrm{GFM}} = \mathbb{E}_{t\sim\mathcal U[0,1],\,x\sim p_t} \left\| \Big[ v_{\theta}(x(t), t) - \alpha(t)\Pi_M\big(A_{\mu\nu}(x(t),t)\,d^\mu(x(t),t)\,v^\nu(x(t),t)\big) \Big] - u_t(x) \right\|_{g_x}^2.$$2

with projector

$$\mathcal{L}_{\mathrm{GFM}} = \mathbb{E}_{t\sim\mathcal U[0,1],\,x\sim p_t} \left\| \Big[ v_{\theta}(x(t), t) - \alpha(t)\Pi_M\big(A_{\mu\nu}(x(t),t)\,d^\mu(x(t),t)\,v^\nu(x(t),t)\big) \Big] - u_t(x) \right\|_{g_x}^2.$$3

A gauge-invariant representative $$\mathcal{L}_{\mathrm{GFM}} = \mathbb{E}_{t\sim\mathcal U[0,1],\,x\sim p_t} \left\| \Big[ v_{\theta}(x(t), t) - \alpha(t)\Pi_M\big(A_{\mu\nu}(x(t),t)\,d^\mu(x(t),t)\,v^\nu(x(t),t)\big) \Big] - u_t(x) \right\|_{g_x}^2.$$4 is then associated to every gauge field $$\mathcal{L}_{\mathrm{GFM}} = \mathbb{E}_{t\sim\mathcal U[0,1],\,x\sim p_t} \left\| \Big[ v_{\theta}(x(t), t) - \alpha(t)\Pi_M\big(A_{\mu\nu}(x(t),t)\,d^\mu(x(t),t)\,v^\nu(x(t),t)\big) \Big] - u_t(x) \right\|_{g_x}^2.$$5, and the effective average action is made to depend on $$\mathcal{L}_{\mathrm{GFM}} = \mathbb{E}_{t\sim\mathcal U[0,1],\,x\sim p_t} \left\| \Big[ v_{\theta}(x(t), t) - \alpha(t)\Pi_M\big(A_{\mu\nu}(x(t),t)\,d^\mu(x(t),t)\,v^\nu(x(t),t)\big) \Big] - u_t(x) \right\|_{g_x}^2.$$6 only through $$\mathcal{L}_{\mathrm{GFM}} = \mathbb{E}_{t\sim\mathcal U[0,1],\,x\sim p_t} \left\| \Big[ v_{\theta}(x(t), t) - \alpha(t)\Pi_M\big(A_{\mu\nu}(x(t),t)\,d^\mu(x(t),t)\,v^\nu(x(t),t)\big) \Big] - u_t(x) \right\|_{g_x}^2.$$7. The resulting exact gauge-invariant flow equation is

$$\mathcal{L}_{\mathrm{GFM}} = \mathbb{E}_{t\sim\mathcal U[0,1],\,x\sim p_t} \left\| \Big[ v_{\theta}(x(t), t) - \alpha(t)\Pi_M\big(A_{\mu\nu}(x(t),t)\,d^\mu(x(t),t)\,v^\nu(x(t),t)\big) \Big] - u_t(x) \right\|_{g_x}^2.$$8

In the simplest pure $$\mathcal{L}_{\mathrm{GFM}} = \mathbb{E}_{t\sim\mathcal U[0,1],\,x\sim p_t} \left\| \Big[ v_{\theta}(x(t), t) - \alpha(t)\Pi_M\big(A_{\mu\nu}(x(t),t)\,d^\mu(x(t),t)\,v^\nu(x(t),t)\big) \Big] - u_t(x) \right\|_{g_x}^2.$$9 Yang–Mills truncation, the flow reproduces one-loop asymptotic freedom and yields a beta function whose first term is the standard one-loop result. The same anomalous dimension $\frac{dx(t)}{dt} = v_\theta(x(t), t)$00 controls both the running of the coupling and the running of the physical gluon propagator in this formulation (Wetterich, 2017).

These algebraic and functional constructions show that “gauge flow” can denote a structural control problem: how gauge dependence is generated, projected, or removed. The flow is then a mathematically organized transport through gauge choices or theory space, not a sampling algorithm.

6. Boundary extensions, anomaly flow, and topological flow geometry

In nonperturbative proposals for chiral gauge theories, gauge fields are extended into an extra dimension by a prescribed flow. On the slab, the physical wall is at $\frac{dx(t)}{dt} = v_\theta(x(t), t)$01, the anti-wall is at $\frac{dx(t)}{dt} = v_\theta(x(t), t)$02, and the boundary condition is

$\frac{dx(t)}{dt} = v_\theta(x(t), t)$03

The original prescription uses gradient flow,

$\frac{dx(t)}{dt} = v_\theta(x(t), t)$04

while the alternative EOM flow requires the field away from the wall to satisfy the higher-dimensional equation of motion. In the $\frac{dx(t)}{dt} = v_\theta(x(t), t)$05-dimensional case the modified Maxwell action is

$\frac{dx(t)}{dt} = v_\theta(x(t), t)$06

leading to

$\frac{dx(t)}{dt} = v_\theta(x(t), t)$07

The lattice implementation shows that both gradient flow and EOM flow suppress the gauge field away from the wall, decouple the mirror sector, and realize anomaly inflow and current conservation on the lattice (Dang et al., 3 Jun 2026).

The disk-boundary version adapts the same idea to a $\frac{dx(t)}{dt} = v_\theta(x(t), t)$08-dimensional disk manifold. In the $\frac{dx(t)}{dt} = v_\theta(x(t), t)$09 $\frac{dx(t)}{dt} = v_\theta(x(t), t)$10 case the gauge field is fixed on the boundary, extended into the interior by an equation-of-motion flow, and numerically implemented on a square lattice by minimizing the plaquette action through an auxiliary flow time $\frac{dx(t)}{dt} = v_\theta(x(t), t)$11. Coupling the flowed gauge field to lattice fermions shows that the boundary anomaly is canceled by bulk inflow; the reported anomaly ratio is $\frac{dx(t)}{dt} = v_\theta(x(t), t)$12, close to the continuum expectation $\frac{dx(t)}{dt} = v_\theta(x(t), t)$13 (Dang et al., 3 Jun 2026).

A different anomaly-related use of flow occurs in orbifold gauge theory and gauge-Higgs unification, where the chiral gauge anomaly depends on the Aharonov-Bohm phase $\frac{dx(t)}{dt} = v_\theta(x(t), t)$14 in the fifth dimension. In flat space the anomaly coefficients jump discontinuously at KK level crossings as $\frac{dx(t)}{dt} = v_\theta(x(t), t)$15 varies, whereas in Randall–Sundrum space there is no level crossing and the anomaly varies smoothly with $\frac{dx(t)}{dt} = v_\theta(x(t), t)$16. After summing over KK modes, the anomaly coefficients reduce to a boundary formula depending only on the wave-function values at the UV and IR branes and on fermion parity data,

$\frac{dx(t)}{dt} = v_\theta(x(t), t)$17

The anomaly cancellation conditions are therefore independent of $\frac{dx(t)}{dt} = v_\theta(x(t), t)$18 (Hosotani, 2023).

Finally, Yang–Mills flow can be studied as a dynamical system on the space of gauge fields modulo gauge transformations. In this setting the flow acts pointwise on loops $\frac{dx(t)}{dt} = v_\theta(x(t), t)$19, and a topologically nontrivial loop of $\frac{dx(t)}{dt} = v_\theta(x(t), t)$20 gauge fields on $\frac{dx(t)}{dt} = v_\theta(x(t), t)$21 can be constructed from a zero-size instanton at one pole and a zero-size anti-instanton at the other, glued by a nontrivial $\frac{dx(t)}{dt} = v_\theta(x(t), t)$22 gauge transformation representing the generator of $\frac{dx(t)}{dt} = v_\theta(x(t), t)$23. On the slow manifold near the twisted pairs, the leading-order flow equations are

$\frac{dx(t)}{dt} = v_\theta(x(t), t)$24

showing local stability of the loop under the Yang–Mills flow and providing evidence for stable higher-dimensional manifolds of gauge fields as well (Friedan, 2010).

Taken together, these lines of work show that Gauge Flow Models are best understood as a family resemblance rather than a single doctrine. In the most recent generative papers the term denotes gauge-structured neural ODEs trained by Flow Matching. In lattice sampling it denotes gauge-equivariant normalizing flows for gauge and pseudofermion fields. In Yang–Mills theory it denotes gradient flow and its improved lattice implementations. In BRST/BV, FRG, and extra-dimensional constructions it denotes canonical, gauge-invariant, or bulk-extension flows that control gauge dependence, anomaly inflow, or topological motion in field space.