Topological phase transitions, such as the confinement of a toric-code-like phase or the gapping of Dirac fermions without symmetry breaking, cannot be seen in local measurements. Physicists detect them with non-local probes instead. These include string or membrane operators (products of spin flips along a whole line or region) and entanglement-based quantities. Inside a Monte Carlo simulation, such a probe is a ratio of two partition functions, a twisted system over an untwisted one. Its value shrinks exponentially with the size of the region. Estimating it is therefore a rare-event problem: the configurations that dominate the twisted system almost never appear in an ordinary simulation.
The standard fix is to split the ratio into a chain of small steps and estimate each step by reweighting. Near a critical point, the distribution of those per-step reweighting factors develops a long right tail, so a handful of samples dominate. Ordinary error bars then become overconfident. Universal exponents fitted to these numbers inherit the problem without anyone noticing. We propose to treat the step chain as an adaptive rare-event estimator. Step placement would be chosen and certified by tail diagnostics borrowed from heavy-tailed importance sampling, and the steps would be combined by a multistate estimator. The test is whether this yields confidence intervals with correct coverage, and better accuracy per sample, on transitions where the exact answer is known through duality.
Core Claim
Our claim is about the estimator, not the physics. Near criticality, the error in non-local order-parameter estimates is controlled by the tail of the per-stage reweighting factors, not by their variance as naively estimated. A stage schedule chosen so that every stage passes a tail check should do two things at the same total sample budget. First, it should give near-nominal coverage of 95% intervals in cases where uniform schedules under-cover. Second, it should give lower error than uniform schedules.
The rare-event direction supplies the estimator structure. The heavy-tailed direction supplies the diagnostic and the safeguard. Topological transitions supply the application where the failure matters and where exact answers exist for checking.
The relevant literature motivates this in three ways. He et al. showed that at the symmetric-mass-generation critical point every local order parameter (valence-bond, charge-density and spin-density orders) vanishes, so only gaps and non-local quantities carry the transition (Quantum critical point of Dirac fermion mass generation without spontaneous symmetry breaking). Chen et al. proved that local data change smoothly across non-local level crossings and that conditional mutual information, a multi-region entanglement quantity, is the robust detector (Discontinuity of maximum entropy inference and quantum phase transitions). They also noted that it is costly to compute on large geometries. Wu et al. argued that classifying phases with non-local order parameters is classically hard in general (Quantum Phase Recognition via Quantum Kernel Methods). We do not dispute that hardness. We restrict to sign-free models, where sampling is efficient and the bottleneck is purely statistical.
Research Program
Write the non-local probe for a region of size $\ell$ as $O(\ell)=Z_1/Z_0$. Here $Z_0$ is the ordinary partition function and $Z_1$ is the one with the string or membrane twist inserted. Introduce intermediate ensembles $Z_{\lambda_k}$, with $0=\lambda_0<\dots<\lambda_K=1$, that switch the twist on gradually. Each stage ratio is $r_k = \mathbb{E}_{\lambda_k}[w_k]$, where $w_k$ is the reweighting factor from stage $k$ to stage $k+1$.
The factors $w_k$ are bounded on any finite lattice, so their variance is technically finite. When stages are too far apart, however, their empirical distribution looks power-law over the range reachable with practical sample sizes. The sample variance then underestimates the true error.
The method has three parts.
Diagnosing each stage. For every stage we compute a Hill-type tail-index estimate on the upper order statistics of $w_k$. A tail index is the exponent of the power-law decay. We also compute the gap between the ordinary effective sample size and its α-generalized version from AHTIS. AHTIS (Adaptive importance sampling for heavy-tailed distributions via α-divergence minimization) adapts Student-$t$ proposals by matching moments of a power-sharpened "escort" target and tunes the proposal through α-ESS. We reuse only its α-ESS diagnostic and its escort idea, on a discrete lattice ensemble rather than a continuous unimodal target.
The constraint on this part comes from Aliakbarpour et al. (Aliakbarpour et al.), who showed that light and heavy tails cannot be told apart from samples without structural assumptions. We therefore never treat the Hill estimate as a proof. We exploit known structure instead: the maximum possible $w_k$ for a given step is computable from the lattice. A stage is accepted only if the estimated tail index exceeds 2 with margin and the observed maximum weight is a small fraction of the summed weights.
Placing stages adaptively. Stages are inserted by bisection wherever a stage fails the check. This is the same logic as the adaptive levels of the cross-entropy method (de Boer et al.), except the level criterion is tail-based rather than quantile-based. As an alternative to the linear path in $\lambda$, we test an escort path. This path raises the twist's Boltzmann weight to a power, mirroring $p_\alpha \propto p^\alpha$, and squeezes the tails within each step.
Combining stages. All stages are pooled with the Multistate Bennett Acceptance Ratio (MBAR), a method that combines samples from many ensembles into one minimum-variance estimate of their free-energy differences. Dorman et al. imported this tool, together with transition path sampling, to estimate language-model tails (Rare Event Analysis of Large Language Models). We move it back to lattice physics, where it originated, and add a tail-certified error bar computed by block bootstrap over certified stages.
The deliverable is open code wrapped around a sign-free classical or quantum Monte Carlo sampler. It outputs $O(\ell)$ with intervals, and fitted critical scaling of $O(\ell)$ with uncertainties propagated from those intervals.
Baselines
The naive reference is direct measurement: average the twist operator in the untwisted simulation. For values of $O(\ell)$ well below $1/\sqrt{N}$ at $N$ samples, it should return noise or zero. This mirrors the failure Dorman et al. report, where brute-force sampling observed no completions in the tail of the readability score.
The strongest competitor is the current practical method: a uniform schedule of $K$ stages combined by MBAR, with $K$ tuned by grid search. Its error bars are the usual sample-variance or binning ones. It is given the same sampler, the same lattices and the same total number of sweeps, and its $K$ is chosen by the best grid value in hindsight against the exact answer. That choice deliberately advantages the baseline. No reported result was found on Lacuna for disorder-operator or membrane-operator estimation with any of these methods. For context only, outside this setting, activation-steering importance sampling estimated harm probabilities down to $10^{-4}$ with 500 samples, 10–20× fewer than Monte Carlo (Estimating Tail Risks in Language Model Output Distributions).
Beating the tuned uniform baseline means two things. Our intervals must stay within a few points of 95% coverage at criticality where the baseline's coverage falls clearly below. We must also match or reduce mean squared error at equal sweeps, without the hindsight tuning.
Three ablations isolate the key ingredient. In the first, the adaptive schedule is kept but standard error bars are reported, which separates certification from placement. In the second, stages are placed by a plain effective-sample-size threshold instead of the tail check, which asks whether tails add anything beyond ESS. In the third, the escort path is swapped for the linear path.
Why not just…?
Why not just use more uniform stages? This is the real threat. With enough stages, every factor becomes well behaved. The issue is that the number needed depends sharply on distance to criticality and on $\ell$, and a practitioner has no way to know it in advance. If the hindsight-tuned uniform baseline already achieves nominal coverage at equal cost across our grid, the contribution shrinks to a cheap diagnostic, and we will report it that way. The coverage-versus-$\ell$ experiment below settles this.
Why not just compute the quantity exactly with tensor networks or transfer matrices? In one dimension, and for the 2D classical Ising model, one can. That is exactly why we use these cases as ground truth. For two-dimensional quantum models such as the SU(4) honeycomb model of He et al., studied on lattices up to $L=18$, only Monte Carlo reaches relevant sizes. The method targets that regime.
Why not just learn a proposal with a diffusion model or normalizing flow, as DiFS does for failures? DiFS iteratively retrains a conditional diffusion model toward failure regions and avoids mode collapse (Diffusion-Based Failure Sampling for Evaluating Safety-Critical Autonomous Systems). Its runtime is hours rather than minutes, on continuous disturbances. Lattice configurations are discrete and already sampled efficiently by cluster or loop updates. In addition, Pachebat shows that Lipschitz networks fed Gaussian noise cannot produce power-law tails without a coordinate transform (Tail Annealing for Heavy-Tailed Flow Matching). The trouble here sits in the weights, not the coordinates, so learned proposals do not address the failure mode.
Proposed Experiments
Two-dimensional classical Ising model. Its disorder operator equals the dual-temperature spin correlator exactly (Kramers–Wannier duality). This model also describes the Kitaev-chain topological transition through the quantum–classical mapping. We sweep $\ell$ and temperature across the critical point at several lattice sizes and run about 200 independent replicas per setting. We report interval coverage, relative error at equal sweeps, and the fitted scaling dimension against its exact dual value.
Three-dimensional classical Ising model. Through Wegner duality, this model describes the confinement transition of a Z2 gauge theory, the canonical topologically ordered phase. The membrane disorder operator follows perimeter scaling with corner corrections. Here the check is internal consistency: estimates must agree across independent schedules and replicas, and must be compatible with established 3D Ising exponents.
The SU(4) honeycomb model of He et al. is a stretch goal and depends on access to a sign-free projector QMC code. We would estimate a non-local disorder-type operator across $V/t$ and test whether its scaling locates the transition consistently with the reported $V_c/t = 2.00 \pm 0.05$.
Decisive check. The decisive check is the 2D coverage curve as a function of $\ell$ at criticality. If tuned uniform schedules under-cover there and the tail-certified schedule does not, at equal cost, the claim holds. If both cover, it fails, and we will report that negative result.
References
- Quantum critical point of Dirac fermion mass generation without spontaneous symmetry breaking
- Discontinuity of maximum entropy inference and quantum phase transitions
- Quantum Phase Recognition via Quantum Kernel Methods
- Adaptive importance sampling for heavy-tailed distributions via α-divergence minimization
- Testing Tail Weight of a Distribution Via Hazard Rate
- A Tutorial on the Cross-Entropy Method
- Rare Event Analysis of Large Language Models
- Estimating Tail Risks in Language Model Output Distributions
- Diffusion-Based Failure Sampling for Evaluating Safety-Critical Autonomous Systems
- Tail Annealing for Heavy-Tailed Flow Matching