Gradient Flow, Flow Scales, and Renormalized Gauge Observables

Gauge-field gradient flow evolves a configuration into a smoother field at positive flow time $t$ . Local gauge-invariant composites built from the flowed field are finite after the underlying theory is renormalized, and dimensionless conditions on the flowed energy density define precise reference scales. Flow remains a scale-dependent transformation: its smoothing radius, lattice discretization, integration error, finite volume, and order of limits must be controlled.

Required background. Topology and lattice index diagnostics distinguishes sector information from a particular smoothed estimator, while gauge ensembles and renormalized observables supplies covariance and continuum controls.

Helpful background. Scale setting and nonperturbative operator renormalization explain how a flow scale or flowed operator enters a physical result.

Continuum and lattice flow equations

Local convention and regulator card. Use $B_\mu(0,x)=A_\mu(x)$ , flow toward positive $t$ by $\partial_tB_\mu=D_\nu G_{\nu\mu}$ , and quote the smoothing radius as $\sqrt{8t}$ in four dimensions. State the simulation action, flow action, integrator and step, energy-density discretization, boundary conditions, $t/a^2$ , $\sqrt{8t}/L$ , and whether the continuum limit is taken at fixed positive physical $t$ before any small-flow-time limit.

For a continuum gauge field $B_\mu(t,x)$ with $B_\mu(0,x)=A_\mu(x)$ , the Yang–Mills gradient flow is

\partial_tB_\mu=D_\nu G_{\nu\mu},

up to an optional gauge-damping term that changes only the flow-time gauge. Since $[t]=\text{length}^2$ , the linearized equation is a heat equation. The field is averaged over a characteristic radius

r_{\mathrm{sm}}\simeq\sqrt{8t}

in four dimensions.

For lattice links $V_t(x,\mu)$ , a standard group-valued flow has the form

\partial_tV_t(x,\mu)V_t(x,\mu)^{-1} =-g_0^2\,\partial_{x,\mu}S_f[V_t], \qquad V_0(x,\mu)=U_\mu(x),

where $S_f$ is a chosen flow action and $\partial_{x,\mu}$ its Lie-algebra derivative. The simulation action, flow action, and discretization of a measured composite may all differ; their combination determines leading cutoff effects.

Numerically, confirm group unitarity, monotonic decrease of $S_f$ for the gradient convention, and convergence under flow-step refinement. These are necessary algorithmic checks, not substitutes for a continuum extrapolation.

The smoothing interpretation and reference-scale construction are developed in Lüscher 2010, §§ 2–3; the all-orders perturbative finiteness analysis is given in Lüscher and Weisz 2011, §§ 2–8.

Reference scales from the energy density

Let

E(t,x)=\frac14G_{\mu\nu}^a(t,x)G_{\mu\nu}^a(t,x), \qquad \mathcal E(t)=t^2\langle E(t)\rangle.

The dimensionless condition

\mathcal E(t_0)=c

defines $t_0$ , while

\left.t\frac{d}{dt}\mathcal E(t)\right|_{t=w_0^2}=c

defines $w_0$ . The conventional constant $c$ and all discretizations must be reported. These quantities are precise reference scales, not automatically physical observables; an experimentally known input or another accepted physical scale is still required to express them in conventional units.

Because $t_0/a^2$ and $w_0/a$ come from the same configurations as many target observables, their covariance should be retained in dimensionless ratios and continuum fits.

The flow window

A useful positive-flow-time observable requires scale separation,

a\ll\sqrt{8t}\ll L,

or equivalently $a^2\ll t\ll L^2$ , with stronger numerical margins determined empirically. The left inequality suppresses discretization artifacts such as $a^2/t$ ; the right controls finite-volume distortion and avoids smearing around periodic boundaries.

A representative expansion is

\mathcal E(t,a,L)=\mathcal E(t,0,\infty) +k_a\frac{a^2}{t}+k_L\frac{t^2}{L^4}+\cdots.

Improved flow and observable choices can reduce $k_a$ but do not justify including points with $t\sim a^2$ . Window stability is tested by moving both endpoints and repeating the continuum fit.

Small-flow-time expansion

For a flowed renormalized composite $O_t(x)$ ,

O_t(x)=\sum_i c_i(t,\mu)O_i^R(x;\mu)+O(t)

as $t\to0$ , ordered by operator dimension and symmetries. The coefficients are a matching problem. A common controlled order is:

take $a\to0$ at fixed positive physical $t$ ;
control $L\to\infty$ or bound finite-volume effects;
apply the matching relation and, if required, take $t\to0$ within a stable window.

Taking $t\to0$ at fixed $a$ collapses the smoothing radius onto the cutoff and invalidates the expansion. Identifying $1/\sqrt{8t}$ with a renormalization scale without computing the matching coefficients is likewise incomplete.

Flowed topology

At positive flow time, an improved field-strength charge often clusters near integers and definitions agree more closely. This makes flow a valuable diagnostic. Yet two questions remain separate:

Does the charge definition approach the continuum topological observable?
Did the Markov chain sample sectors with their equilibrium weights?

Flow may improve the first and cannot repair the second. Track charge histories before and after flow, state $t$ in physical units, and test definition and spacing dependence.

A complete flow result

Report the ensemble action, flow action, integrator and step, energy or operator discretization, $c$ , $t/a^2$ , $\sqrt{8t}/L$ , scale input, autocorrelation, and continuum model. The regime map emphasizes that the path from smoothing to a renormalized claim passes through both a flow window and a matching or reference-scale definition.

Gauge configurations branch into gauge-invariant loops, gauge-fixed correlators, strong-coupling series, topology, and gradient-flow observables, each with a distinct validity test. — The positive-flow branch requires a scale-separation window, integration refinement, and matching before joining the final continuum claim. Smoothness alone is not sufficient. The diagram is schematic and not to scale.

Adversarial failure: a smooth diagonal path with unsuppressed cutoff error

Take a sequence with $t/a^2=2$ at every spacing and observe that $t^2\langle E(t)\rangle$ changes little. The configurations look smoother as the nominal $a$ decreases, yet

\frac{a^2}{t}=\frac12, \qquad \sqrt{8t}=4a\longrightarrow0

in physical units. The leading $a^2/t$ artifact never becomes small, so this diagonal path cannot establish the positive- $t$ continuum observable or its small-flow-time expansion. Fixed physical $t$ at several $a$ values, followed by a separate $t$ -window analysis, is the test that this apparently stable sequence fails.

Observable-level validation checklist

Verify $V_{t=0}=U$ , link unitarity, gauge covariance, and monotonic flow-action decrease with the stated sign convention.
Refine the integration step and require $t^2\langle E(t)\rangle$ , $t_0/a^2$ , and $w_0/a$ to converge within the numerical error.
Compare at least two flow or energy discretizations and fit their $a^2/t$ dependence at fixed positive physical $t$ .
Move both ends of the window $a\ll\sqrt{8t}\ll L$ and repeat at another volume.
Carry flow-scale covariance into the target ratio, and test matching-scale and small- $t$ stability only after the fixed- $t$ continuum limit.

Common pitfalls

Calling flow “just cooling.” Gradient flow has a precise differential equation and supports renormalized positive- $t$ composites. That precision is lost if the action and integrator are unspecified.

Treating $t$ as harmless. Flow time is a physical resolution scale. Compare at fixed physical $t$ , not merely at equal $t/a^2$ .

Removing the cutoff and flow limits together without analysis. Terms such as $a^2/t$ make diagonal limit paths dangerous. Demonstrate a stable ordered or joint extrapolation.

Learning outcomes

Given a lattice flow prescription, compute $t_0$ or $w_0$ , verify integration and scale-separation conditions, and propagate the resulting reference-scale covariance into an observable.
Given data over $a$ , $L$ , and $t$ , distinguish a finite positive-flow-time observable from arbitrary smoothing and design ordered continuum, volume, matching, and small-flow-time tests.

Exercises

If $t/a^2=2$ is held fixed while $a\to0$ , what happens to $r_{\mathrm{sm}}$ in physical units and to $a^2/t$ ?

Solution

$r_{\mathrm{sm}}=\sqrt{16}\,a=4a\to0$ , while $a^2/t=1/2$ remains finite. Cutoff effects are therefore not suppressed.

A measured $t_0/a^2=4$ gives $\sqrt{8t_0}=\sqrt{32}\,a$ . Is this automatically inside the flow window?