Correctness criteria for complex Langevin

This paper systematically compares the applicability, ease of use, and predictive power of prominent diagnostic tools for assessing the correctness of complex Langevin simulations by applying them to four simple but nontrivial models.

The Gist

The Big Problem: The "Ghostly" Sign Problem

Imagine you are trying to calculate the average height of people in a crowded room. Usually, you just ask everyone, add up their heights, and divide by the number of people. This is easy because everyone's height is a positive number.

In physics, specifically in Quantum Field Theory, scientists try to do the same thing but with "fields" (invisible forces that fill the universe). However, in certain situations—like when particles are moving very fast or interacting with specific forces—the math produces numbers that are negative or even imaginary (like $\sqrt{-1}$).

This is called the "Sign Problem."

• The Analogy: Imagine trying to calculate the average temperature of a room, but half the thermometers are broken and say "100 degrees" while the other half say "-100 degrees." If you just add them up, they cancel each other out to zero, giving you a completely wrong answer. The "noise" of the positive and negative numbers drowns out the real signal.

The Solution: The Complex Langevin Method

To fix this, physicists invented a clever trick called the Complex Langevin Method.

• The Analogy: Instead of trying to measure the temperature in the "real" room (where the broken thermometers are), they imagine a parallel universe where the rules are slightly different. In this parallel universe, the numbers are all positive and behave nicely. They run a simulation there, get a result, and hope that this result magically translates back to the real world correctly.

It's like trying to find the shortest path through a foggy, confusing maze (the real world). Instead of walking through the fog, you fly above it in a helicopter (the parallel universe) where you can see the whole map clearly. You hope that the path you see from above is the same as the path you'd have to walk on the ground.

The Danger: The "Wrong Turn" Trap

Here is the catch: Sometimes the helicopter path looks perfect, but it leads you to the wrong destination.

The Complex Langevin method is powerful, but it has a flaw. Sometimes, the simulation gets stuck in a "wrong" version of the parallel universe. It converges (settles down) to a stable answer, but that answer is completely wrong. It's like the helicopter landing in a fake city that looks exactly like the real one, but the streets are in the wrong order.

If a physicist uses this wrong answer to design a nuclear reactor or predict particle collisions, the results could be disastrous.

The Paper's Mission: Building a "Lie Detector"

Since we can't always check the answer against the "real world" (because in complex physics, we often don't know the real answer!), we need a way to tell if our simulation is telling the truth or lying.

This paper is a comparative review of different "Lie Detectors" (called Correctness Criteria). The author, Michael Mandl, tested eight different tools on four simple toy models to see which ones work best.

Think of these tools as different ways to check if a car engine is running correctly:

1. Listening to the engine (Dyson-Schwinger equations): Does it make the right sounds?
2. Checking the oil (Histograms): Is the distribution of particles smooth and healthy?
3. Looking for leaks (Boundary Terms): Is the simulation spilling out of the box?
4. Checking the speedometer (Drift Criterion): Is the engine revving too high or too low?
5. Measuring the heat (Configurational Temperature): Is the engine too hot?

The Results: Which Lie Detector Wins?

Mandl tested these tools and found some surprising results:

1. The "Drift Criterion" is the Champion:

   • The Analogy: This tool checks how "hard" the simulation is pushing against the walls of the imaginary world. If the push is too weak or decays too slowly, the simulation is likely lying.
   • Verdict: This was the most reliable tool. It was like a smoke detector that almost always went off when there was a fire. It's cheap to use and very sensitive.

2. The "Histogram" is a Good Runner-Up:

   • The Analogy: This looks at the shape of the data cloud. If the cloud is too spread out or has weird tails, something is wrong.
   • Verdict: Very reliable, but sometimes it gets fooled by "ghosts" (unwanted mathematical loops) that don't actually affect the final answer.

3. The "Configurational Temperature" is Unreliable:

   • The Analogy: This tries to measure the "heat" of the simulation.
   • Verdict: It gave false alarms. Sometimes it said "Fire!" when the engine was fine, and sometimes it stayed silent when the engine was smoking. It's too sensitive to the size of the system.

4. The "Observable Bounds" are the Gold Standard (but hard to use):

   • The Analogy: This is a rigorous mathematical proof that says, "If the answer is bigger than X, it's a lie."
   • Verdict: It is 100% accurate in theory, but in practice, it's like trying to solve a 1,000-piece puzzle blindfolded. It's too difficult to calculate for complex problems.

5. The "Unitarity Norm" is a Heuristic Guide:

   • The Analogy: This checks how far the simulation has wandered from the "real" path.
   • Verdict: It's a good rule of thumb. If the simulation wanders too far, it's probably lying. But there's no hard line where "a little wandering" becomes "too much."

The Takeaway for the Real World

The paper concludes that while there is no single "magic bullet" that works 100% of the time, the Drift Criterion is currently the best tool for the job.

• Why it matters: As physicists try to simulate the early universe, black holes, or new materials, they rely on these computer simulations. If they use the wrong tool to check their work, they might publish a "discovery" that is actually just a mathematical glitch.
• The Advice: Don't rely on just one check. Use the Drift Criterion as your primary alarm, but keep an eye on the Histograms and Boundary Terms to be sure.

In short: The paper is a user manual for physicists, teaching them how to spot a fake simulation so they don't get fooled by the ghosts in the machine.

Technical Summary

1. Problem Statement

The sign problem is a fundamental obstacle in lattice quantum field theory (QFT) and condensed matter physics, preventing the use of standard Monte Carlo importance sampling methods when the path integral weight $\rho$ is complex (e.g., in Minkowski space or at finite chemical potential).

The Complex Langevin (CL) method is a prominent stochastic approach to bypass the sign problem by evolving complexified field variables in an auxiliary time dimension. However, CL suffers from the wrong-convergence problem, where the stochastic process converges to an incorrect limit or fails to converge to the correct expectation values of holomorphic observables.

While various correctness criteria (diagnostic tools) have been developed over the years to detect these failures, there has been no systematic comparison of their reliability, predictive power, and practical applicability. This paper aims to fill that gap by rigorously testing the most prominent criteria against models with known exact solutions.

2. Methodology

The author employs a comparative analysis framework using four distinct toy models with known analytical solutions. These models are chosen to represent different failure modes of the CL method (e.g., slow decay, poles in the drift term, unwanted integration cycles).

The Models:

1. One-dimensional Quartic Model: $S(z) = \frac{\lambda}{4}z^4$. Used to study the effect of constant kernels and integration cycles.
2. One-pole Model: $\rho(z) \propto (z-z_0)^{n_p} e^{-\beta z^2}$. Used to investigate drift-term poles and the onset of incorrect convergence as a function of parameters.
3. One-site Hubbard Model: A fermionic model with infinitely many poles in the drift term. Used to test the method with complex noise and reweighting scenarios.
4. Quartic Model on a Complex Time-Contour: A discretized 1D quantum field theory (4 degrees of freedom) on a Schwinger-Keldysh contour. Used to test multi-variable systems and real-time evolution limits.

The Criteria Tested:
The paper evaluates eight specific diagnostic tools:

1. Dyson–Schwinger (DS) Equations: Checking if correlation functions satisfy exact relations.
2. Histograms: Analyzing the decay of the probability distribution $P$ in the complex plane.
3. Boundary Terms: Direct estimation of boundary terms at infinity using cutoffs.
4. Convergence Conditions: Checking if $\langle L_c O \rangle = 0$ in the equilibrium limit.
5. Drift Criterion: Analyzing the decay of the drift term magnitude distribution.
6. Observable Bounds: A rigorous necessary and sufficient condition based on norm inequalities.
7. Unitarity Norm: Monitoring the distance of the trajectory from the original real manifold.
8. Configurational Temperature: A thermodynamic consistency check comparing geometric and input temperatures.

Simulation Setup:
Simulations were performed using GPU-accelerated parallelization (CUDA) with both Euler–Maruyama and improved update schemes. Large ensembles ($N_{sim} = 2^{13}$) were generated to ensure statistical precision, allowing for the detection of subtle discrepancies.

3. Key Contributions

• Systematic Benchmarking: This is the first large-scale, systematic comparison of all major CL correctness criteria across diverse models and failure modes.
• Identification of Limitations: The paper demonstrates that many criteria previously thought to be robust are either insufficient, heuristic, or sensitive to specific model features (e.g., integration cycles vs. boundary terms).
• Practical Guidelines: The author provides a hierarchy of criteria based on their predictive power and computational cost, offering a practical guide for researchers.
• Resolution of Discrepancies: The study clarifies the role of finite-step-size effects in boundary term calculations and distinguishes between statistical anomalies and genuine convergence failures.

4. Key Results

A. Performance of Specific Criteria

• Drift Criterion (Most Powerful): The analysis of the drift term magnitude distribution (checking for exponential decay) emerged as the most reliable and robust indicator of overall correctness. It successfully detected incorrect convergence in almost all cases, including those caused by unwanted integration cycles (with the exception of the specific 1D quartic model where it failed to distinguish between rotated cycles).
• Histograms: Generally reliable for detecting slow decay (boundary terms) but can be insensitive to contributions from unwanted integration cycles if the decay remains fast.
• Boundary Terms & Convergence Conditions:
  • Boundary Terms: Effective for detecting non-vanishing boundary terms at infinity. However, they are highly sensitive to finite step-size effects and can yield false positives if the step size is not sufficiently small.
  • Convergence Conditions: Found to be unreliable. They often hold even when the simulation is incorrect (e.g., when unwanted integration cycles contribute), as they essentially measure equilibration rather than correctness.
• Dyson–Schwinger Equations: A necessary but not sufficient condition. They are satisfied even when the CL evolution samples unwanted integration cycles (linear combinations of different contours).
• Observable Bounds: Theoretically rigorous (necessary and sufficient), but practically difficult. Finding a control observable that violates the bound for a specific incorrect case is a non-trivial, often heuristic task.
• Unitarity Norm: A useful heuristic indicator. Large values correlate with incorrectness, but no strict threshold exists. It failed in the Hubbard model where correct results were obtained for one observable despite large norms.
• Configurational Temperature: Unreliable for small systems. It produced false positives and negatives and diverged in models with poles, likely due to the violation of the thermodynamic limit assumption.

B. Model-Specific Findings

• Quartic Model: Demonstrated that different kernels can lead to correct results, incorrect results due to boundary terms, or incorrect results due to sampling the wrong integration cycle. The drift criterion failed to distinguish between the "correct" cycle and a "wrong" cycle that was simply a 90-degree rotation of the correct one.
• One-pole Model: Incorrect convergence for small $\beta$ was caused by unwanted integration cycles. The drift and histogram criteria successfully detected this, whereas boundary terms and DS equations did not.
• Hubbard Model: A scenario was found where tuning imaginary noise yielded the correct fermion number density but incorrect higher moments. The drift criterion correctly flagged the overall simulation as incorrect, while boundary terms were ambiguous due to step-size dependence.
• Complex Time-Contour: In the multi-variable model, incorrect convergence for large real-time extents was detected by the drift and histogram criteria. Notably, the boundary term analysis showed a complex behavior (multiple plateaus) that was more conclusive than the convergence conditions.

5. Significance and Conclusions

The paper concludes that no single criterion is perfect, but the Drift Criterion stands out as the most effective general-purpose tool for detecting incorrect convergence in complex Langevin simulations.

• Hierarchy of Reliability:

  1. Drift Criterion: Best for overall detection of failure.
  2. Histograms & Boundary Terms: Good for specific diagnostics, but require careful handling of step sizes and interpretation.
  3. Dyson–Schwinger Equations: Necessary but insufficient; useful for ruling out correctness but not for confirming it.
  4. Configurational Temperature & Unitarity Norm: Heuristic and prone to failure in small or specific systems.
  5. Observable Bounds: Theoretically sound but practically difficult to apply.

• Implications for Realistic Theories: The findings suggest that for realistic theories (like QCD), relying on the drift criterion and histogram analysis is the most pragmatic approach. The study highlights that the "wrong-convergence" problem often stems from the sampling of unwanted integration cycles, a failure mode that many standard diagnostics (like DS equations) fail to detect.

The work provides a crucial roadmap for practitioners, emphasizing that while the CL method is powerful, its results must be validated using a combination of diagnostics, with the drift criterion serving as the primary filter.