Many physics simulations average a quantity over configurations weighted by $e^{-S}$, where $S$ is the action, a score assigned to each configuration. When $S$ is complex, which happens at nonzero matter density or with topological terms, that weight oscillates in sign. Ordinary Monte Carlo then needs exponentially many samples. Complex Langevin (CL) avoids this. It lets the variables wander into the complex plane under a noisy gradient flow, and it estimates averages from the real, positive distribution $P(x,y)$ that the walk settles into. Its weakness is that it can fail silently. If $P$ thins out too slowly along the imaginary direction $y$, an integration-by-parts step in the correctness proof leaves leftover "boundary terms," and the simulation converges smoothly to a wrong answer. Worse, the failure differs between observables. In the 3D XY model at nonzero chemical potential, the action density is badly biased while the particle number density is fine in the same run (Controlling complex Langevin simulations of lattice models by boundary term analysis). No current tool predicts which observables will fail before each one is measured and corrected separately.
We propose that this observable dependence is not mysterious. It comes from a race between two rates. The first is how fast the CL distribution decays at large $|y|$, which is a property of the simulation. The second is how fast the observable, after the Langevin operator acts on it, grows at large $|y|$, which is a property of the observable. One measured tail exponent per simulation, together with a growth exponent computed analytically for each observable, should then say in advance which observables carry boundary terms. It should also say how fast their bias shrinks as a stabilizing regulator is turned up. That prediction can drive the expensive regularize-and-extrapolate correction (the 2R method), so that it is run only where needed and with a regressor whose form is fixed rather than guessed.
Core Claim
Write $L_c$ for the Langevin operator, the generator of the complexified dynamics, and $P_Y$ for the CL distribution restricted to configurations whose distance from the real manifold exceeds $Y$. The volume-integral estimator of Controlling complex Langevin simulations of lattice models by boundary term analysis measures the boundary term $B_O$ of an observable $O$ as the large-cutoff behaviour of $\langle L_c O \rangle$ restricted to $|y| \le Y$. Suppose the radial tail of $P$ decays like $Y^{-\alpha}$, and $|L_c O|$ grows like $Y^{g_O}$. For compact variables, where growth is exponential, replace powers by rates. The tail's contribution to the boundary term then scales as
$$B_O(Y) - B_O(\infty) \sim \int_Y^\infty y^{g_O - \alpha}\,dy \sim Y^{\,g_O + 1 - \alpha},$$
so $B_O$ vanishes only when the margin $m_O = \alpha - g_O - 1$ is positive.
The claim has three parts. First, $\alpha$ is a single, system-wide number that can be estimated from one trajectory. Second, $g_O$ follows from the observable and the drift by inspection. Third, the sign of $m_O$ sorts observables into safe and unsafe before any boundary-term measurement is made. Adding the stabilizing term $\tfrac{s}{2}\sum_k \phi_k^2$ of Regularization of the complex Langevin method raises $\alpha(s)$, which gives a second claim. The regulator value at which $m_O(s)$ crosses zero marks where each observable's regularized results become trustworthy for regression. The way the bias closes near that point fixes the form of the extrapolation back to $s \to 0$.
This is a sufficient-condition heuristic built on the integration-by-parts argument of Complex Langevin and boundary terms, not a theorem. Near drift poles, which arise from fermion determinants, a second tail appears around each pole. We treat that tail the same way but expect it to be harder.
Where the three directions sit
The boundary-term literature (Direction 1) supplies the mechanism and the measurement tools. Complex Langevin and boundary terms links slow decay to failed integration by parts in a one-dimensional model, and the 2020 lattice paper turns that link into surface and volume estimators. Both papers measure boundary terms after the fact, one observable at a time. We aim to predict them.
The sign-problem literature (Direction 3) sets the regime and the constraints. Plain reweighting dies exponentially with volume, as Kühn explains. Hamiltonian tensor-network methods such as Exploring the CP-violating Dashen phase in the Schwinger model with tensor networks are free of the sign problem but essentially restricted to one spatial dimension. CL remains the scalable option in higher dimensions, so its reliability is the bottleneck worth attacking.
The tonal-to-non-tonal speech literature (Direction 2) contributes no mechanism, and the research file finds no work linking it to the other two. It contributes one design principle and one evaluation design. The principle comes from intonation modelling. A Statistical Phrase/Accent Model for Intonation Modeling splits a pitch contour into a global phrase layer and local accent excursions, and that explicit split is what made the model controllable. Our split of the boundary-term signal into a global $\alpha$ and a per-observable $g_O$ is the same move.
The evaluation design comes from cross-lingual transfer. Cross-lingual emotion transfer has so far been tested only between two tonal languages (Zero-Shot Emotion Transfer For Cross-Lingual Speech Synthesis) or only within English (RSET). The hard and unanswered test is transfer across a change in what a feature means: pitch is a free expressive channel in English but part of word identity in Mandarin. We apply the same standard. A diagnostic calibrated on Abelian scalar models has to be tested on non-Abelian gauge models, where non-unitarity, the drift structure, and gauge cooling change what large $|y|$ means, rather than only on more models of the same kind.
Research Program
The first stage builds the estimator and checks it where exact answers exist. We run CL on the one-dimensional complex-$\beta$ model of Complex Langevin and boundary terms and on the U(1) one-link model. We estimate $\alpha$ by fitting the radial tail of the sampled imaginary parts, using a standard tail-index fitting approach with bootstrap errors. We derive $g_O$ for a family of observables, including polynomial moments of increasing order and plaquette-type exponentials. We then check that $\operatorname{sign}(m_O)$ matches the boundary terms measured by the volume estimator, and that the measured $B_O(Y)$ follows the predicted $Y^{m_O}$ approach.
The second stage is the decisive transfer test inside one system: the 3D XY model at nonzero chemical potential. The method must predict, from one run per parameter point, that the action density has boundary terms and the number density does not, as the 2020 lattice paper reports. It must also predict where in chemical potential that split begins.
The third stage moves to non-Abelian systems, the SU(3) Polyakov chain and heavy-dense QCD (HDQCD). There, distance from the real manifold is measured by a unitarity norm. Gauge cooling is known to suppress boundary terms in these models (Controlling complex Langevin simulations of lattice models by boundary term analysis), so a successful diagnostic must show the margins turning positive under cooling.
The fourth stage uses the margins to drive regularization, giving "margin-guided 2R." For each observable we compute $\alpha(s)$ along a coarse sweep of the regulator, place regression points only where $m_O(s) > 0$, and fit an extrapolant whose leading behaviour near the crossover is fixed by the tail law. The aim is to fix the two weaknesses Regularization of the complex Langevin method reports: sensitivity to the choice of regression points in the XY model, and extreme sensitivity to the reweighting anchor $s_0$ for the Polyakov chain.
Baselines
The naive reference point is unregularized CL with only a check that trajectories stay bounded. Its failure mode is documented: it can look stable and still be wrong. In the one-dimensional model at $\beta = 0.1$, unregularized CL disagrees with the exact integral, while $s = 0.1$ restores agreement (Complex Langevin and boundary terms). We run it with the same step size, Langevin time and seeds as every other method. Any useful diagnostic must flag every case where this baseline is biased.
The strongest diagnostic competitor is direct volume-integral boundary-term measurement for each observable, as in Controlling complex Langevin simulations of lattice models by boundary term analysis. It is the accepted way to detect and subtract boundary terms. It recovers exact U(1) one-plaquette results, though no raw numerical value for that result is reported in the file. We run it on the same trajectories with the same cutoff schedule, measuring every observable we test. Beating it means matching its safe/unsafe classification on all benchmarks while needing only one tail fit per simulation instead of one cutoff study per observable. For any observable where measured $B_O$ is small but clearly nonzero, it also means predicting the sign of $B_O$ before it is measured.
The strongest correction competitor is 2R and 3R from Regularization of the complex Langevin method. The file reports that 2R recovers HDQCD observables where standard CL fails, but gives no numerical error table: no reported result found on Lacuna for 2R/3R error magnitudes on U(1), XY, Polyakov chain or HDQCD. We reimplement both with their published regressor families. We give them the same total Langevin time as margin-guided 2R and tune their regression points and $s_0$ with the same search budget we spend on our coarse $\alpha(s)$ sweep. Beating them means lower error against exact or high-precision reference values at equal cost. It also means smaller spread in the extrapolated value when the regression points are moved, since point sensitivity is their stated weakness. Where gauge cooling applies, it is run as an extra baseline.
The key ablation keeps margin-guided 2R but replaces the per-observable $g_O$ with one global stability score that ignores the observable. If this does as well, the observable-specific half of the decomposition is not doing the work, and the central claim fails.
Why not just…?
Why not just measure boundary terms directly for every observable? This is the real threat. If the volume estimator is cheap enough per observable, prediction adds little. The proposal's value then rests on two things. One is cost: a single tail fit against a cutoff study for each observable, which matters in lattices with many observables and dynamical fermions, where the 2020 lattice paper notes direct subtraction is hard. The other is guidance: telling 2R where to place points, which direct measurement does not do. We report wall-clock cost at matched accuracy for both. If direct measurement wins on cost and margin-guided 2R does not beat tuned 2R, we will say the prediction is only an interpretive aid.
Why not just run 2R with more regulator values and a better regressor? Spending more compute does reduce variance. But Regularization of the complex Langevin method reports that in the XY model the trouble is possibly infinite variance at some regression points, which more samples do not fix. Choosing points by the margin targets exactly that. The equal-budget comparison and the point-sensitivity test settle this question.
Why not just reweight, or move to a Hamiltonian method? Phase reweighting needs a number of samples growing like $e^{2\beta V \Delta f}$, where $\beta$ is the inverse temperature, $V$ the spatial volume and $\Delta f$ a free-energy-density difference (Kühn). Matrix-product-state methods such as Fermion Discretization Effects in the Two-Flavor Lattice Schwinger Model are exact in one dimension but face growing entanglement costs beyond it. We use reweighting only as a small-volume, small-density reference check.
Why not just trust CL when trajectories look stable? The one-dimensional and XY results above show stable-looking runs that converge to wrong values, so trajectory inspection is exactly what a diagnostic must improve on.
Decisive Checks
The proposal stands or falls on four outcomes. In the one-dimensional and U(1) models, the sign of $m_O$ must agree with measured boundary terms for every observable, and the measured bias must approach its limit with the predicted exponent within error. In the 3D XY model, a single run per point must reproduce the reported split between action density and number density. In the SU(3) Polyakov chain and HDQCD, the margins must turn positive under gauge cooling, which tests whether the diagnostic carries over from Abelian scalar models to non-Abelian gauge models. Margin-guided 2R must match or beat tuned 2R/3R at equal Langevin time, with visibly reduced sensitivity to regression points and $s_0$.
Failure of the third check while the first two pass would be informative in its own right. It would show where a diagnostic built on scalar models stops carrying over to gauge models, which is the same kind of limit the speech literature has yet to measure for tonal and non-tonal languages.
References
- Controlling complex Langevin simulations of lattice models by boundary term analysis
- Regularization of the complex Langevin method
- Complex Langevin and boundary terms
- Quantum and classical simulation of High Energy Physics
- Exploring the <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mi>C</mml:mi><mml:mi>P</mml:mi></mml:mrow></mml:math>-violating Dashen phase in the Schwinger model with tensor networks
- A Statistical Phrase/Accent Model for Intonation Modeling
- Zero-Shot Emotion Transfer for Cross-Lingual Speech Synthesis
- RSET: Remapping-Based Sorting Method for Emotion Transfer Speech Synthesis
- Fermion Discretization Effects in the Two-Flavor Lattice Schwinger Model: A Study with Matrix Product States