Complex Langevin Algorithms

• Complex Langevin algorithms are stochastic simulation methods that extend real variables to complex space to overcome the sign problem in quantum systems.
• They employ techniques like gauge cooling and dynamic stabilisation to enhance stability and ensure proper convergence by mitigating non-unitarity and runaway behaviors.
• Applications include finite-density QCD, real-time dynamics, and supersymmetric models, with effectiveness hinging on strict convergence diagnostics and boundary term suppression.

Complex Langevin Algorithms provide a stochastic approach to simulating systems with complex actions, circumventing the sign problem that obstructs conventional importance-sampling-based Monte Carlo methods. By complexifying the configuration space and evolving the degrees of freedom according to a Langevin stochastic process with complex drift, the method aims to simulate the path integral for theories including finite-density QCD, chiral random matrix models, real-time dynamics, and supersymmetric systems. Ensuring correct convergence and avoiding spurious or runaway behavior requires careful attention to the mathematical properties and practical diagnostics of the algorithm, as well as the employment of various stabilizing modifications.

1. Formal Structure and Mathematical Justification

Given a target measure $\rho(x) = e^{-S(x)}$ with $S(x)$ complex, the Complex Langevin (CL) method promotes $x \in \mathbb{R}^n$ to $z = x + i y \in \mathbb{C}^n$ and evolves $z$ in fictitious Langevin time via

$dz = K(z)\,dt + dw, \qquad K(z) = -\nabla S(z),$

with $dw$ a real Wiener process and $K(z)$ the holomorphic drift. Observables $\mathcal{O}(z)$ are computed along this trajectory, and ergodic averages are intended to reproduce path-integral expectations: $$\langle \mathcal{O} \rangle = \int dx \, e^{-S(x)} \mathcal{O}(x) / Z.$$ The process induces a probability density $S(x)$0 on $S(x)$1 governed by a real Fokker–Planck (FP) equation,

$S(x)$2

where $S(x)$3 and $S(x)$4 are the real and imaginary parts of $S(x)$5, respectively.

A central step in the justification invokes integration by parts in complex space, linking the stationary distribution of the process to the original complex measure. This connection is valid only if all surface/boundary terms at infinity and near singularities of the drift vanish. In equilibrium, the process must satisfy the system of Schwinger-Dyson (zero-operator) conditions for holomorphic observables: $S(x)$6 The system's convergence and correctness rely crucially on the suppression of boundary terms and the analyticity of both the drift and observables (Seiler, 2017, Scherzer et al., 2018, Mandl, 14 Apr 2026).

2. Practical Implementations and Algorithmic Modifications

The Euler–Maruyama scheme is a standard discretization: $S(x)$7 For nonabelian gauge theories, the drift is implemented via left Lie derivatives on complexified gauge links, e.g., $S(x)$8 (Attanasio et al., 2016, Aarts et al., 2013). The following modifications are critical for stability and extending the algorithmic reach:

• Gauge Cooling: SL($S(x)$9,ℂ) gauge transformations minimize non-unitarity norms (e.g., $x \in \mathbb{R}^n$0), suppressing excursions into noncompact directions and mitigating fat-tailed distributions in the imaginary sector. Essential in QCD-like theories (Aarts et al., 2013, Aarts et al., 2013).
• Dynamic Stabilisation (DS): A nonholomorphic restoring force is added to the drift in gauge theories—$x \in \mathbb{R}^n$1 with $x \in \mathbb{R}^n$2, $x \in \mathbb{R}^n$3. DS suppresses the unitarity norm, keeps the process proximate to SU(3), and prevents runaways and late-time metastabilities even when gauge cooling alone fails (Attanasio et al., 2016, Attanasio et al., 2018).
• Kernel-Controlled CL: The drift and noise can be modified by a (usually constant) kernel, optimized using prior information (symmetry, Euclidean correlators) and CL-based correctness diagnostics. Properly tuned, this approach extends the reach of CL to real-time simulations with longer Schwinger-Keldysh contours (Alvestad et al., 2022, Alvestad et al., 2022).
• Regularization and Reweighting: Augmenting the action with a confining term (e.g., $x \in \mathbb{R}^n$4, or a unitarity norm in gauge theory) regularizes runaway trajectories. The bias is then removed via extrapolation (2R method) or reweighting (3R method), preferably with regression across several regulator strengths (Cai et al., 2021).
• Reweighted Complex Langevin (RCL): Trajectories are generated at parameters where CL is valid and reweighted post hoc to target parameters. This method extends CL's applicability into regimes where direct simulation is unreliable, provided the auxiliary ensemble is sufficiently close for efficient reweighting (Bloch, 2017, Bloch et al., 2017).

3. Convergence Criteria, Diagnostics, and Correctness

The main convergence criteria for Complex Langevin algorithms are:

• Decay of Probability and Drift Distributions: The probability density $x \in \mathbb{R}^n$5 and the distribution of drift magnitudes $x \in \mathbb{R}^n$6 must decay exponentially or faster for large imaginary excursions and large $x \in \mathbb{R}^n$7. Mere power-law decay is inadequate: boundary terms in the correctness proof survive and CL fails (Nagata et al., 2016, Shimasaki et al., 2016, Mandl, 14 Apr 2026).
• Vanishing of Boundary Terms: Empirical verification that all surface integrals at $x \in \mathbb{R}^n$8 and around singularities/poles vanish in the long-time limit is accomplished via drift-norm tests, marginal distribution fits, and monitoring of the Schwinger–Dyson conditions for a basis of holomorphic observables (Scherzer et al., 2018, Mandl, 14 Apr 2026).
• Pole and Cut Effects: The presence of singularities (drift poles, e.g., from zeroes of the determinant in QCD or log-determinants in chiral random matrix theory) can transiently or permanently induce boundary terms or ergodicity breaking. Monitoring the distribution of eigenvalues/argument of the determinant is essential for diagnosing improper convergence due to drift ambiguity or branch cut crossings (Mollgaard et al., 2013, Seiler, 2021).
• Consistency Conditions: In stationary distributions, expectation values of $x \in \mathbb{R}^n$9 vanish for all holomorphic $z = x + i y \in \mathbb{C}^n$0. Satisfying these up to a finite basis is a practical necessary (and under analytic control, sufficient) correctness criterion (Mandl, 14 Apr 2026).

Recent systematic comparisons confirm the predictive power of drift-norm tail diagnostics and a small set of Schwinger–Dyson checks as key practical correctness tools (Mandl, 14 Apr 2026).

4. Algorithmic Stabilization Strategies

A spectrum of practical stabilization techniques has been developed and validated:

• Gauge Cooling and Drift-Norm Cooling: Iterative SL($z = x + i y \in \mathbb{C}^n$1,ℂ) transformations minimize non-unitarity or local drift norms between Langevin steps, directly controlling the magnitude of the drift and ensuring exponential suppression in the critical regions of configuration space (Nagata et al., 2016, Aarts et al., 2013).
• Dynamic Stabilisation: The DS force with tunable strength $z = x + i y \in \mathbb{C}^n$2 is added to the drift in gauge theories. The force is proportional to a high power of the deviation from unitarity and thus vanishes in the SU(3) limit, preserving the continuum theory while enforcing proximity to SU(3) numerically. Tuning $z = x + i y \in \mathbb{C}^n$3 such that $z = x + i y \in \mathbb{C}^n$4 for step size $z = x + i y \in \mathbb{C}^n$5 is typically sufficient to restore agreement with benchmark results (Attanasio et al., 2016, Attanasio et al., 2018).
• Kernel Learning and Optimization: For non-equilibrium and real-time settings, field-independent kernels can stabilize CL for longer contour lengths, with further improvement possible using parametrized, field-dependent kernels optimized by stochastic gradient descent against physics-based cost functions (Alvestad et al., 2022).
• Regularization/Reweighting/Extrapolation: When regularization is introduced to prevent runaways, bias is systematically removed using multicomponent regression. RCL uses auxiliary ensembles with known validity and applies postprocessing reweighting to project onto target parameter sets while monitoring overlap and sign problems (Bloch, 2017, Bloch et al., 2017, Cai et al., 2021).

5. Applications and Benchmarks

Complex Langevin schemes with these controls have delivered quantitatively valid results in multiple regimes:

• Finite-Density QCD and QCD-like Models: For heavy-dense QCD and QCD with staggered fermions, CL with gauge cooling is reliable for $z = x + i y \in \mathbb{C}^n$6; DS further extends convergence into coarser couplings (smaller $z = x + i y \in \mathbb{C}^n$7). Reweighting allows extrapolation into low-mass or high-density regimes where direct CL fails (Attanasio et al., 2016, Aarts et al., 2013, Bloch et al., 2017).
• Random Matrix Theory and One-Link Models: CL matches analytic expectations when determinant trajectories avoid the origin but fails otherwise; these models provide analytically controlled settings to test convergence diagnostics, boundary term phenomena, and the efficacy of stabilization (Mollgaard et al., 2013).
• Real-Time (Schwinger-Keldysh) Dynamics: Kernel-optimized and implicit-solver-based CL approaches enable simulations up to $z = x + i y \in \mathbb{C}^n$8 in strongly coupled quantum anharmonic oscillators—significantly beyond naive CL—with real-time correlators matching Schrödinger or Monte Carlo data (Alvestad et al., 2022, Alvestad et al., 2021).
• Supersymmetric and PT-Symmetric Systems: CL can resolve dynamical supersymmetry breaking in 1D models, provided regulators are tuned to avoid singular drift, and drift-distribution/FP operator consistency is monitored (Joseph et al., 2020).

6. Limitations, Open Problems, and Practical Guidelines

Key limitations include the absence of comprehensive existence theorems for non-selfadjoint FP operators, complications from poles and branch cuts, and unsolved issues with ergodicity breaking and metastability. Empirical stabilization techniques, while effective, are not yet justified by a unified theoretical framework. The survey (Mandl, 14 Apr 2026) recommends reliable practical guidelines:

• Always monitor drift-norm distributions and ensure exponential decay.
• Accumulate low-order Schwinger–Dyson consistency observables.
• Test for invariance under holomorphic kernel (or gauge-cooling parameter) modifications.
• Halt simulations and seek further stabilization or cooling if drift-tails fatten or Schwinger–Dyson tests fail.
• Document diagnostic histograms and correctness checks alongside observable measurements for reproducibility.

A plausible implication is that systematic correctness checking—combining drift-norm, zero-operator, and kernel-independence diagnostics—forms the backbone of modern CL validation protocols in realistic lattice field theories.

---

References:

(Attanasio et al., 2016, Attanasio et al., 2018, Aarts et al., 2013, Aarts et al., 2013, Bloch et al., 2017, Bloch, 2017, Nagata et al., 2016, Shimasaki et al., 2016, Alvestad et al., 2022, Alvestad et al., 2022, Cai et al., 2021, Alvestad et al., 2021, Mandl, 14 Apr 2026, Mollgaard et al., 2013, Scherzer et al., 2018, Seiler, 2021, Joseph et al., 2020, Seiler, 2017, Aarts et al., 2017).