Gauge Cooling in Lattice Field Theory

• Gauge cooling is a technique that applies complex gauge transformations to reduce the unitarity norm, ensuring stability in complex Langevin dynamics.
• It controls excursions into the non-compact SL(N, C) space, thereby enhancing the localization of probability distributions and the reliability of gauge-invariant observables.
• Variants like adaptive and alternating descent cooling optimize the minimization process, mitigating singular-drift issues in nonabelian lattice gauge theories.

Gauge cooling is a procedure used primarily in the complex Langevin method for nonabelian lattice gauge theories with complex actions. Its central purpose is to control the exploration of the complexified configuration space by applying complexified gauge transformations in $\mathrm{SL}(N,\mathbb{C})$ that minimize a measure of distance from the original $\mathrm{SU}(N)$ manifold, typically a unitarity norm. In this role, gauge cooling is a stabilization and correctness mechanism: it suppresses large excursions into non-compact directions, improves localization of the complexified probability distribution, and can be decisive for obtaining correct gauge-invariant observables in theories with a sign problem (Bongiovanni et al., 2013, Nagata et al., 2015).

1. Definition, scope, and historical setting

In lattice gauge theory with a complex Euclidean action, the Boltzmann weight $e^{-S}$ cannot be interpreted as a probability measure. Complex Langevin dynamics replaces importance sampling by a stochastic evolution in a complexified field space. For gauge theories this means that link variables $U_{x,\mu}$, originally in $\mathrm{SU}(N)$, are extended to $\mathrm{SL}(N,\mathbb{C})$, where the non-compact directions create both formal and numerical difficulties (Bongiovanni et al., 2013).

Gauge cooling was introduced in this context as an additional deterministic step interleaved with Langevin evolution. In early applications it was used in QCD with heavy quarks and in the deconfined phase to stabilize complex Langevin simulations and to reach parameter regions inaccessible to reweighting (Seiler et al., 2012, Nagata et al., 2015). Subsequent work developed adaptive variants, generalized norms aimed at singular-drift control, and alternative solvers for the cooling minimization problem (Bongiovanni et al., 2013, Nagata et al., 2016, Dong et al., 2020).

The term has a second, older lattice usage in which “cooling” denotes gauge-field smoothing by local action minimization, often compared with the Yang–Mills gradient flow in topology studies (Bonati et al., 2014, Alexandrou et al., 2015). That usage is related in spirit but distinct in mechanism and purpose. Unless explicitly stated otherwise, gauge cooling in contemporary lattice field theory usually refers to the complex Langevin technique.

2. Geometric mechanism in the complexified gauge manifold

The basic geometric step is the complexification

$$U_{x,\mu}\in \mathrm{SU}(N)\quad \longrightarrow \quad U_{x,\mu}\in \mathrm{SL}(N,\mathbb{C}),$$

which is required because the Langevin drift becomes complex when the action is complex. In the complexified theory the links satisfy

$$UU^\dagger \neq 1, \qquad \frac{1}{N}\operatorname{Tr}\,UU^\dagger \ge 1,$$

so each Langevin step can move the system away from the unitary manifold into non-compact directions (Bongiovanni et al., 2013).

Gauge cooling exploits the fact that gauge invariance persists after complexification. A complexified gauge transformation acts as

$$U_{x,\mu}\rightarrow \Omega_x U_{x,\mu}\Omega_{x+\hat\mu}^{-1},\qquad \Omega_x\in \mathrm{SL}(N,\mathbb{C}),$$

and leaves gauge-invariant observables and the holomorphic action unchanged, while changing non-gauge-invariant measures of non-unitarity (Nagata et al., 2015).

The canonical distance measure is the unitarity norm. In one common form, for a full lattice,

$$d=\frac{1}{NV}\sum_{x,\mu}\operatorname{Tr}\bigl(U_{x,\mu}U_{x,\mu}^\dagger-1\bigr),$$

which vanishes only on $\mathrm{SU}(N)$0 configurations (Bongiovanni et al., 2013). A second form used in full QCD studies is

$\mathrm{SU}(N)$1

which is likewise zero on $\mathrm{SU}(N)$2 and positive otherwise (Nagata et al., 2016).

Gauge cooling consists of choosing $\mathrm{SU}(N)$3 so as to reduce such a norm along the complexified gauge orbit. Because only gauge transformations are used, the procedure is not gauge fixing and requires no Faddeev–Popov determinant (Bongiovanni et al., 2013).

3. Core algorithms and major variants

The standard cooling step follows the gradient of the unitarity norm along the gauge orbit. A local transformation can be written as

$\mathrm{SU}(N)$4

where $\mathrm{SU}(N)$5 is of the order of the Langevin step size, $\mathrm{SU}(N)$6 is a tunable cooling parameter, and $\mathrm{SU}(N)$7 is the cooling force. In the formulation of adaptive gauge cooling,

$\mathrm{SU}(N)$8

with the sum over directions $\mathrm{SU}(N)$9 understood (Bongiovanni et al., 2013). A Langevin step is then followed by one or more cooling sweeps.

The one-link analysis already shows the essential behavior. For $e^{-S}$0, if a link is gauge-equivalent to an $e^{-S}$1 matrix, the cooled distance decays exponentially to zero,

$e^{-S}$2

where $e^{-S}$3. If it is not gauge-equivalent to an $e^{-S}$4 matrix, cooling converges to a finite minimal distance (Bongiovanni et al., 2013).

Adaptive gauge cooling was introduced because fixed $e^{-S}$5 becomes inefficient on larger systems. The adaptive parameter is chosen as

$e^{-S}$6

with representative choices

$e^{-S}$7

In the $e^{-S}$8 Polyakov chain, adaptive cooling with $e^{-S}$9 yields a much faster approach to the fixed point, with an asymptotic decay $U_{x,\mu}$0, and reaches the minimum distance several orders of magnitude faster than the non-adaptive case (Bongiovanni et al., 2013).

A distinct algorithmic development is the alternating descent method, which treats gauge cooling explicitly as a minimization problem over the complexified gauge orbit. The method performs exact local minimizations on even and odd sublattices, is parameter-free at the iteration level, and shows better performance than classical gradient descent especially when the lattice size is large (Dong et al., 2020). In the one-dimensional periodic setting, exact minimizers of the Frobenius norm can be constructed analytically, clarifying how optimal cooling reduces the effective dynamics to eigenvalue evolution and suppresses non-compact wandering (Cai et al., 2019).

4. Correctness theory, localization, and singular-drift control

The formal justification of complex Langevin requires more than numerical stability. The drift must be holomorphic in the sampled region, the distribution in the complexified variables must be sufficiently localized, and boundary terms must vanish in the integrations by parts used to derive the complex Fokker–Planck relation (Nagata et al., 2015, Nagata et al., 2016).

Gauge cooling modifies the probability distribution $U_{x,\mu}$1 of the complexified variables, but for gauge-invariant observables it does not change the Fokker–Planck equation for the complex weight corresponding to the original path integral. This is because the cooling contribution can be written as an extra deterministic drift generated by a complexified gauge transformation parameter $U_{x,\mu}$2, and its action on holomorphic gauge-invariant observables vanishes identically (Nagata et al., 2015). In that precise sense, gauge cooling can improve convergence conditions without changing the target gauge-invariant physics.

A practical correctness diagnostic is the probability distribution of the drift norm,

$U_{x,\mu}$3

Correct convergence requires $U_{x,\mu}$4 to fall off exponentially or faster at large $U_{x,\mu}$5; power-law tails signal failure (Nagata et al., 2016).

The excursion problem motivated the original unitarity-norm cooling, but later work generalized gauge cooling to the singular-drift problem caused by small eigenvalues of $U_{x,\mu}$6. In chiral Random Matrix Theory, three norms were emphasized: $U_{x,\mu}$7 which controls deviation from the original Hermitian relation;

$U_{x,\mu}$8

which measures violation of anti-Hermiticity of the Dirac operator; and

$U_{x,\mu}$9

where $\mathrm{SU}(N)$0 are the smallest eigenvalues of $\mathrm{SU}(N)$1 (Nagata et al., 2015, Nagata et al., 2016). Cooling with $\mathrm{SU}(N)$2 narrows the Dirac spectrum toward the imaginary axis, while $\mathrm{SU}(N)$3 directly repels eigenvalues of $\mathrm{SU}(N)$4 from the origin. In both cases the objective is to suppress the regions where the drift becomes singular.

The one-dimensional $\mathrm{SU}(N)$5 analysis makes this stabilizing mechanism explicit. After optimal cooling, the effective stochastic equation for the angular variable $\mathrm{SU}(N)$6 takes the form

$\mathrm{SU}(N)$7

so gauge cooling adds a $\mathrm{SU}(N)$8 drift term that confines the imaginary part of $\mathrm{SU}(N)$9 to a strip and helps localize the process near the real axis (Cai et al., 2019).

5. Principal applications in lattice gauge theory

The first extensive application was heavy-dense QCD. In that approximation the fermion determinant reduces to a product over Polyakov-loop factors,

$\mathrm{SL}(N,\mathbb{C})$0

with $\mathrm{SL}(N,\mathbb{C})$1 and $\mathrm{SL}(N,\mathbb{C})$2. Gauge cooling was essential for stabilizing the complex Langevin evolution, and the resulting observables agreed with reweighting within the estimated errors while extending simulations to previously inaccessible high densities (Seiler et al., 2012).

Adaptive gauge cooling was then tested in an $\mathrm{SL}(N,\mathbb{C})$3 Polyakov chain and in $\mathrm{SL}(N,\mathbb{C})$4 Yang–Mills theory with a $\mathrm{SL}(N,\mathbb{C})$5-term. In the Polyakov chain it improved localization of the distribution of the imaginary part of the action, kept the unitarity norm several orders of magnitude smaller than in the uncooled case, and restored agreement of observables such as $\mathrm{SL}(N,\mathbb{C})$6 with exact analytic results when sufficient cooling was used (Bongiovanni et al., 2013). In $\mathrm{SL}(N,\mathbb{C})$7 Yang–Mills with lattice action

$\mathrm{SL}(N,\mathbb{C})$8

simulations on a $\mathrm{SL}(N,\mathbb{C})$9 lattice at $$U_{x,\mu}\in \mathrm{SU}(N)\quad \longrightarrow \quad U_{x,\mu}\in \mathrm{SL}(N,\mathbb{C}),$$0 showed that for $$U_{x,\mu}\in \mathrm{SU}(N)\quad \longrightarrow \quad U_{x,\mu}\in \mathrm{SL}(N,\mathbb{C}),$$1 the plaquette agreed with Hybrid Monte Carlo, while for real $$U_{x,\mu}\in \mathrm{SU}(N)\quad \longrightarrow \quad U_{x,\mu}\in \mathrm{SL}(N,\mathbb{C}),$$2 cooling controlled the distribution of the complex action more effectively at larger $$U_{x,\mu}\in \mathrm{SU}(N)\quad \longrightarrow \quad U_{x,\mu}\in \mathrm{SL}(N,\mathbb{C}),$$3 than at smaller $$U_{x,\mu}\in \mathrm{SU}(N)\quad \longrightarrow \quad U_{x,\mu}\in \mathrm{SL}(N,\mathbb{C}),$$4 (Bongiovanni et al., 2013).

Full QCD at finite density and low temperature provided a more stringent test. On a $$U_{x,\mu}\in \mathrm{SU}(N)\quad \longrightarrow \quad U_{x,\mu}\in \mathrm{SL}(N,\mathbb{C}),$$5 lattice with $$U_{x,\mu}\in \mathrm{SU}(N)\quad \longrightarrow \quad U_{x,\mu}\in \mathrm{SL}(N,\mathbb{C}),$$6 and $$U_{x,\mu}\in \mathrm{SU}(N)\quad \longrightarrow \quad U_{x,\mu}\in \mathrm{SL}(N,\mathbb{C}),$$7, the singular-drift problem turned out to be mild in the explored region, so gauge cooling was used only to control the unitarity norm. The norm remained bounded throughout the simulations, the drift distribution $$U_{x,\mu}\in \mathrm{SU}(N)\quad \longrightarrow \quad U_{x,\mu}\in \mathrm{SL}(N,\mathbb{C}),$$8 was exponentially suppressed for $$U_{x,\mu}\in \mathrm{SU}(N)\quad \longrightarrow \quad U_{x,\mu}\in \mathrm{SL}(N,\mathbb{C}),$$9, and the onset of the baryon number density appeared at larger $$UU^\dagger \neq 1, \qquad \frac{1}{N}\operatorname{Tr}\,UU^\dagger \ge 1,$$0 than in phase-quenched QCD (Nagata et al., 2016). This study also made clear that on larger lattices and at stronger density the singular-drift problem is expected to become more severe, motivating the more sophisticated cooling norms developed in Random Matrix Theory (Nagata et al., 2016).

6. Related meanings, smoothing theory, and terminological extensions

The phrase “cooling” also has a long and technically important meaning in lattice gauge theory as gauge-field smoothing. In that setting, standard cooling iteratively minimizes the gauge action by local link replacements, while the Yang–Mills gradient flow evolves gauge fields in a continuous fictitious time. In $$UU^\dagger \neq 1, \qquad \frac{1}{N}\operatorname{Tr}\,UU^\dagger \ge 1,$$1 pure gauge theory these two procedures were shown to be equivalent for topology-related observables when the flow time and the number of cooling sweeps are matched according to

$$UU^\dagger \neq 1, \qquad \frac{1}{N}\operatorname{Tr}\,UU^\dagger \ge 1,$$2

with agreement not only in ensemble averages but also configuration by configuration for sufficiently fine lattices (Bonati et al., 2014). For gauge actions with rectangular terms, the mapping generalizes to

$$UU^\dagger \neq 1, \qquad \frac{1}{N}\operatorname{Tr}\,UU^\dagger \ge 1,$$3

where $$UU^\dagger \neq 1, \qquad \frac{1}{N}\operatorname{Tr}\,UU^\dagger \ge 1,$$4 is the Symanzik coefficient multiplying the rectangular term; this was verified for the Wilson, Symanzik tree-level improved, and Iwasaki actions in $$UU^\dagger \neq 1, \qquad \frac{1}{N}\operatorname{Tr}\,UU^\dagger \ge 1,$$5 twisted-mass QCD (Alexandrou et al., 2015). In pure $$UU^\dagger \neq 1, \qquad \frac{1}{N}\operatorname{Tr}\,UU^\dagger \ge 1,$$6 lattice gauge theory, cooling-flow scales were likewise found to provide scale setting with no significant loss of accuracy relative to gradient flow, while being much cheaper computationally (Berg et al., 2016).

Cooling in the topology literature also predates gradient flow as a way to reveal instanton-like structures. Creutz showed that under Wilson-action cooling, rough Monte Carlo configurations develop near-integer topological plateaus, but instantons tend to shrink and eventually “fall through the lattice,” and the resulting topological susceptibility depends sensitively on the early details of the cooling algorithm (Creutz, 2010). This older smoothing usage is conceptually separate from complex Langevin gauge cooling, even though both are deterministic procedures applied between dynamical updates.

The terminology has also been extended outside classical lattice Monte Carlo. In quantum-simulator work, “gauged cooling” denotes a state-preparation protocol in which Ising spins are coupled to a $$UU^\dagger \neq 1, \qquad \frac{1}{N}\operatorname{Tr}\,UU^\dagger \ge 1,$$7 gauge field that acts as a reservoir for removing domain-wall excitations and naturally extends to fermionic systems (Kishony et al., 2023). In digital quantum simulations of $$UU^\dagger \neq 1, \qquad \frac{1}{N}\operatorname{Tr}\,UU^\dagger \ge 1,$$8 lattice gauge theory, “gauge cooling” has been used for an active syndrome-based protocol that detects local Gauss-law violations through a group quantum Fourier transform and applies syndrome-conditional recovery operations to map the state back toward the gauge-invariant subspace (Bradshaw, 26 Mar 2026). In optomechanics, the phrase has appeared again in a different sense, referring to cooling mechanisms mediated by a synthetic phononic gauge field (Jiang et al., 2020). These usages preserve the general idea of removing unwanted excitations through gauge-structured control, but they are not the same method as complex Langevin gauge cooling.

Taken together, the literature gives “gauge cooling” a sharply defined core meaning in complex Langevin theory and a wider family of related meanings in smoothing, topology, and quantum simulation. In the narrow complex Langevin sense, however, its defining content is fixed: complexified gauge transformations are used to minimize a norm that is invisible to gauge-invariant observables but decisive for localization, stability, and correct convergence in the complexified configuration space (Bongiovanni et al., 2013, Nagata et al., 2015).