Discovery-Grade Error Bars for Thermodynamic-Limit Extrapolation

Simulations of quantum materials run on finite lattices. The quantity physicists care about, such as whether a frustrated magnet keeps a small amount of magnetic order when the lattice becomes infinite, is obtained by fitting the finite-size results and reading off the value at infinite size. The error bar on that value is usually a least-squares error that reflects only Monte Carlo sampling noise. It leaves out two larger sources of doubt. The first is the choice of fitting curve. The second is the systematic bias of an approximate wave function, and that bias grows with lattice size. We propose to treat the extrapolated order parameter the way particle physicists treat a new-particle claim. We would build a likelihood that carries these doubts as explicit "nuisance" parameters, test "zero order" against "nonzero order" with a profile likelihood ratio, and choose in advance which lattice sizes and sample counts give the most discovery significance per unit of compute.

The case that motivates this is a single number. Approaching the Thermodynamic Limit with Neural-Network Quantum States trains neural-network wave functions on triangular lattices up to 42×42 sites and extrapolates the magnetization to $M_0 = 0.148(1)$. That paper also reports that earlier, less accurate variational states systematically overestimated $M_0$. So the quoted error of one part in the third decimal cannot be the whole uncertainty. The dominant error is bias that depends on wave-function quality, and the current fitting procedure does not see it. Our claim is that a calibrated, bias-aware significance can be built from quantities already computed in such runs, chiefly the energy variance at each size. We further claim this changes both the stated confidence and the choice of which simulations to run.

Core Claim

A neural-network quantum state is a variational wave function written as a neural network and evaluated by Monte Carlo sampling. At lattice size $L$, it reports an observable that we model as

$$\hat M(L) = M_0 + a_1 L^{-1} + a_2 L^{-2} + \kappa\, \sigma_E^2(L) + \epsilon_L .$$

Here $M_0$ is the thermodynamic-limit value and $a_1, a_2$ are finite-size correction coefficients. $\sigma_E^2(L)$ is the measured energy variance per site, which is zero for an exact eigenstate. $\kappa$ is an unknown slope linking wave-function error to observable bias, and $\epsilon_L$ is Monte Carlo noise with autocorrelation-corrected variance. We treat $a_1, a_2, \kappa$ as nuisance parameters and profile over them. The test statistic is $q_0 = -2\ln[\mathcal{L}(M_0{=}0)/\mathcal{L}(\hat M_0)]$, with its null distribution calibrated by simulation rather than assumed chi-squared, because $M_0 \ge 0$ sits on a boundary. We also profile over a small discrete family of correction forms, including logarithmic terms. The resulting trials penalty is the finite-size-scaling version of the look-elsewhere effect, the correction made in the Telescope Array hotspot analysis when a local excess is judged against all the places one could have looked.

The design half of the claim goes as follows. Before any production run, we fit the likelihood to "expected" data, meaning data generated at the best-guess parameters with no noise (particle physicists call this an Asimov dataset). That gives the expected significance as a function of the lattice sizes $\{L_i\}$ and the per-size budgets. The budget at each size is split between optimization steps, which reduce $\sigma_E^2$ and therefore bias, and sampling, which reduces $\epsilon_L$. We then maximize expected $q_0$ subject to a total cost $C = \sum_i c(L_i)\,[n^{\rm opt}_i + n^{\rm samp}_i]$.

Relation to Prior Work

The work that motivates this proposal is the neural-network scaling line. Approaching the Thermodynamic Limit with Neural-Network Quantum States supplies the triangular-lattice data and extrapolation practice, using polynomial fits in $1/L$ and $1/L^3$ after the simulations are finished. Message-passing neural quantum states for the homogeneous electron gas extends the approach to 128 electrons in continuous space. WF-Bench shows that global infidelity grows as a power law in particle number. That result is exactly why we tie bias to a size-dependent quality measure rather than assuming it is constant. None of these papers propagates bias into the extrapolated value or plans the size ladder.

The work that constrains the proposal comes from two directions. Finite-size scaling of random k-SAT and the spiked Wigner analysis connect system size to detection limits, but only in settings with exact or analytically known statistics and no variational bias. They are nonetheless useful as sandboxes with exact answers. Fidelity mapping shows how badly unmodeled finite-size effects can mislead: it places the XXZ transition at Δ≈1.5 when the infinite-size value is 1.0. It serves as a test of whether our look-elsewhere-corrected fits flag such shifts rather than reporting them as confident results.

The statistical machinery comes from rare-event physics. We reuse unbinned extended likelihoods with nuisance profiling, as used in the EXO-200 search, and apply them to a regression setting instead of event counting.

Research Program

The program has three parts. First, we build the likelihood described above. It includes blocking-based autocorrelation estimates for $\epsilon_L$, a menu of correction forms (power laws, power laws plus logarithms, and forms that include $\kappa$), and a simulation-calibrated null distribution. Second, we check calibration on systems where the answer is independently known. Third, we apply the design procedure to the triangular lattice under a fixed GPU budget and report the significance of $M_0>0$ achieved by designed and undesigned campaigns.

The central scientific question is whether $\kappa\,\sigma_E^2$ captures most of the variational bias. We test this directly. For several deliberately under-trained checkpoints at each size, we regress the observable against $\sigma_E^2$ and look for linear collapse. If there is no collapse, the bias term needs a richer covariate, such as the variance of the local observable estimator, and we report that as a negative finding.

Baselines

The naive reference is the current practice: an ordinary least-squares polynomial fit in $1/L$, whose error comes from the fit covariance, applied to whatever sizes were run. On the triangular lattice, this procedure yields the reported $M_0 = 0.148(1)$ and a ground-state energy of $-0.55168(2)$ per site (Approaching the Thermodynamic Limit with Neural-Network Quantum States).

The strongest existing competitor, as practiced by careful groups, is inverse-variance weighted least squares with bootstrap resampling of Monte Carlo samples. It already handles sampling noise and is honest about it. We run it on identical checkpoints and samples. A second strong competitor is Bayesian model averaging over fit forms, which handles fit-form ambiguity but not variational bias. We give it the same form menu and the same data.

The key ablation is our likelihood with $\kappa$ fixed to zero. This isolates whether the energy-variance bias term is what improves calibration. A second ablation keeps our likelihood but replaces the designed size ladder with the paper's ladder at equal compute, which isolates the value of the design step.

Beating the baselines means two things on the calibration systems. The first is empirical coverage of nominal 68% and 95% intervals within a few percent of nominal where the baselines under-cover. The second is a false-discovery rate at a $3\sigma$ threshold near its nominal value on the null system. On the triangular lattice, beating them means reaching a stated significance for $M_0>0$ at lower compute, or showing that the baselines' significance was overstated. No reported result was found on Lacuna for coverage or budget-allocation performance of any extrapolation method, so these comparisons are new measurements.

Why not just…?

Why not just use inverse-variance weighted fits with bootstrap errors? For pure sampling noise this is nearly optimal, and we concede that. The proposal adds nothing there. The threat it misses is bias that grows with $L$, which bootstrap resampling of one fixed wave function cannot reveal. The observation in Approaching the Thermodynamic Limit with Neural-Network Quantum States that weaker wave functions overestimated $M_0$ is exactly this effect. The κ=0 ablation versus the full model on under-trained checkpoints settles it: if weighted least squares covers the truth as well as our likelihood, the bias term is unnecessary and we will say so.

Why not just run larger lattices? Larger $L$ shrinks the finite-size terms, but WF-Bench indicates that wave-function error grows with size. Larger lattices can therefore trade finite-size error for bias, while cost rises steeply. The design experiment compares a "largest affordable lattice" allocation with the optimized ladder at equal compute.

Why not just average over fit forms (Bayesian model averaging)? This is a real threat on fit-form ambiguity, and it may match us on exact-state sandboxes. It does not model the bias of approximate states, so the decisive comparison is on the under-trained-checkpoint experiments. If averaging with $\sigma_E^2$ added as an extra regressor performs as well as profiling, we will adopt the simpler Bayesian presentation. In that case the contribution becomes the bias covariate and the design procedure.

Why not just extrapolate the observable to zero energy variance? Variance extrapolation is a single-covariate special case of our model at fixed $L$. It ignores the fact that finite-size and bias corrections are correlated across sizes. The joint fit versus sequential "variance-then-size" extrapolation is run as an additional ablation.

Proposed Experiments

The calibration systems are chosen to have independently known answers. The first is the 1D Heisenberg chain, a true null ($M_0=0$) with logarithmic corrections that tempt $1/L$ fits into spurious nonzero intercepts. The second is the unfrustrated square-lattice Heisenberg model, an ordered state with sign-free quantum Monte Carlo references. The third is the XXZ chain, where we re-run the fidelity-map boundary with sampled rather than exact states and test whether the corrected fit moves from Δ≈1.5 toward 1.0 or at least widens its interval to include it. The fourth is random k-SAT, a noise-only sandbox where both instance count and $N$ can be scanned cheaply.

On each system we train neural-network wave functions to several quality levels and repeat the entire pipeline over independent seeds. We measure interval coverage, false-discovery rate at $3\sigma$ and $5\sigma$ thresholds, and the error in the extrapolated value.

The decisive checks are as follows. On the chain, weighted least squares should under-cover and produce false $M_0>0$ claims with under-trained states, while our method should not. The energy-variance collapse should hold within each size. On the triangular lattice, under a budget matched to the reported campaign, we report whether $M_0>0$ survives at $5\sigma$ once bias and fit-form nuisances are profiled, and how the resulting interval compares with $0.148(1)$. The $J_1$-$J_2$ model on the 20×20 square lattice from the same study and prior neural-network work serves as an out-of-sample application. There, the question of order versus no order is contested, and a calibrated "not yet significant" would itself be a result.