How Many Counts Does It Take to Certify a Topological Invariant?

Topological materials are built around a whole-number "twist" in the interior of the material, called a bulk invariant, such as a winding number or a Chern number. Bulk-boundary correspondence is the rule that this twist forces robust states to appear at the material's edges or corners. Experiments rarely measure the invariant directly. They infer it from finite, noisy, often sparse measurements: port-to-port responses of a circuit, or photon counts from a photonic lattice. Physics has a mature language for when such inference is limited by too few events. In counting experiments, relative error falls only as one over the square root of the number of events, and a "discovery" conventionally requires the 5σ threshold, a false-alarm probability near $2.87\times10^{-7}$. Topological physics has never used this language. No one has stated how many measurements are needed to claim, at a stated confidence, that a sample has winding number 1 rather than 0.

This note proposes to answer that question. We will derive and test a sample-complexity law for certifying a topological invariant from counting-limited data. We will also build a test that reaches that limit. The central observation is simple. An integer invariant turns estimation into a choice between discrete hypotheses, so the error should fall exponentially in the number of counts rather than as one over its square root. The rate of that exponential is set by the hardest-to-distinguish pair of samples on opposite sides of a phase transition, and that pair sits where the bulk gap closes. The result is a concrete prediction that can be falsified: the counts needed for 5σ certification should grow as the inverse square of the bulk gap, with a prefactor we can compute.

Core Claim

Consider a composite hypothesis: all disordered samples sharing invariant $\nu = 0$, against all sharing $\nu = 1$, where "disordered" means any disorder that keeps the bulk gap at least $\Delta$. The minimum number of counts $N^\ast$ needed to decide between them with false-certification probability $\alpha$ obeys

$$N^\ast(\alpha,\Delta) \approx \frac{\ln(1/\alpha)}{C(\Delta)}, \qquad C(\Delta) \approx \kappa\,\Delta^{2},$$

Here $C(\Delta)$ is the Chernoff information, the best achievable per-count exponential rate for telling apart the two least-distinguishable samples, one from each class. The constant $\kappa$ depends on the measurement scheme: which ports or sites are read out, and the noise model. The $\Delta^2$ form has a simple origin. Near a gap-closing transition, the expected measurement signals of the closest pair differ by an amount linear in $\Delta$. Chernoff information between two nearby noise distributions is quadratic in that difference. For 5σ, $\ln(1/\alpha)\approx 15$, so the claim fixes absolute count budgets, not just scaling.

Topology enters through the classes themselves. Bulk-boundary correspondence guarantees that every sample in a class shares the same quantized response far from the transition. Robustness to disorder then becomes a statistical statement: $C(\Delta)$ stays bounded away from zero for any disorder that keeps the gap open. This is what we will test.

Relation to Prior Work

Two papers motivate the gap directly. The Universal Spectral Moment Theorem proves that averages of powers of eigenvalues are boundary-independent in non-Hermitian systems, but only in the infinite-size limit. It says nothing about the variance of those quantities when eigenvalues are inferred from finite, noisy data. Experimental identification of the second-order non-Hermitian skin effect with physics-graph-informed machine learning measured a non-Bloch winding number $v_{2D}=1$ on a 10×10 topolectrical circuit, a board of discrete electronic components wired to mimic a lattice. It worked around the measurement cost, which grows as the square of the port count, with a method called PGIML. PGIML groups subcircuits that are wired identically, using graph isomorphism and K-means clustering, measures only a few ports per group, and fills in the full matrix with a teacher-student network. That paper supplies our main testbed and our strongest baseline. Its reliance on exact graph symmetry is exactly what disorder breaks, and it reports no confidence statement on the inferred invariant.

Two machine-learning papers constrain the design. Symbolic Learning of Topological Bands in Photonic Crystals reaches about 90% classification accuracy from 64 training samples using noise-free dielectric Fourier coefficients. Its "Rule of Three" (three Fourier coefficients control the low-band topology) suggests that the information distinguishing the classes lives in a low-dimensional statistic, which our test exploits. It does not address photon-noise-limited observations. Gradient-based search of quantum phases likewise assumes exact spectra.

On the statistics side, the EXO-200 analysis in Search for Neutrinoless Double-β Decay with the Complete EXO-200 Dataset shows that exploiting event structure can be worth years of runtime. Its neural network on detector track shapes raised signal efficiency from 80.8% to 96.4±3.0% and sensitivity by about 25%. Its "topology" means the spatial clustering of charge deposits, not a quantized invariant, so it is an analogy for the payoff rather than a method we reuse. Work on the Telescope Array cosmic-ray cross section (proton-air cross section), where statistical error (±70.5 mb) exceeds systematic error, illustrates the counting-limited regime we borrow the vocabulary from.

Research Program

The theory part computes $C(\Delta)$ for two canonical settings. The first is a non-Hermitian two-dimensional circuit model matching the 10×10 board, where invariant information is read out as noisy admittance entries. The second is a $C_2$-symmetric photonic crystal read out by photon-counting detectors with Poisson noise. In each setting we find the least-favorable pair by optimizing over disorder realizations constrained to a gap of at least $\Delta$. A gradient search in the spirit of Fonseca et al. does this, with the target loss replaced by negative Chernoff information. The output is $\kappa$ and a check of the $\Delta^2$ law, including where it fails. Failure is expected for corner modes of higher-order phases, which can be fragile to symmetry breaking, as discussed in Lee et al..

The method part builds a certifying test. It computes a local marker, a site-resolved version of the invariant averaged over interior sites, directly from the sparse noisy measurements. It then applies a sequential likelihood-ratio test between the integer hypotheses, using the least-favorable pair as the alternative. The test stops as soon as the evidence crosses the threshold set by $\alpha$, so its count budget adapts to the sample. Because the hypotheses are composite and disordered, we calibrate the threshold by simulation over the disorder ensemble rather than trusting the nominal error rate.

A secondary question tests the claim that boundary hypersensitivity in skin-effect systems makes better sensors. Under counting noise, we compute whether the amplified response to a small perturbation raises Chernoff information, or whether the readout noise is amplified by the same factor.

Baselines

The naive reference is full measurement with a plug-in estimate. We measure every port pair (or record a fixed photon budget on every site), compute the winding number from the reconstructed matrix, and repeat with bootstrap resampling to attach a confidence. It is run on the same simulated noise model and the same total count budget. Our method beats it if it reaches a given false-certification rate with substantially fewer counts.

The strongest existing method is PGIML from the second-order skin-effect paper. It is the only published approach that certifies a higher-order non-Hermitian invariant from sparse port data. We reimplement its sampling mask and teacher-student reconstruction, give it the same count budget and comparable tuning time, and attach a bootstrap confidence to its plug-in winding number. Its reported outcome is $v_{2D}=1$ on the 10×10 board. No reported result was found on Lacuna for its measurement count versus confidence, or for its behavior under disorder. Beating it means two things. In the clean, symmetric regime we should match it while also supplying a calibrated confidence. Under disorder, where its isomorphism clusters fragment, we should hold the nominal false-certification rate at lower counts.

A learned baseline is a classifier trained on simulated noisy measurements to output the invariant. We use a KAN following Ghorashi et al. and a plain neural network, with the same simulation budget. Their reported ~90% accuracy at 64 samples comes from noise-free inputs, so it is not directly comparable.

The ablation isolates the key ingredient. We run our test with the alternative hypothesis set to a typical $\nu=1$ sample instead of the least-favorable one, and with the marker averaged over all sites including the boundary. If either variant certifies as reliably, the least-favorable-pair and bulk-marker design is not what matters.

Why not just…?

Why not just collect more counts until the plug-in estimate stabilizes? That works, but it leaves the cost unknown and the confidence uncalibrated. The core claim says the required budget is a computable quantity. The full-measurement baseline will show directly how many counts it wastes relative to $N^\ast$. If it lands within a small factor of $N^\ast$, the practical contribution shrinks to the bound itself, and we would report that.

Why not just train a classifier on simulated noisy data? This is the real threat. An amortized classifier trained on enough simulations can in principle approach the best possible test. The proposal does not depend on beating it. The bound tells us whether it is already optimal, and the experiments compare its error-versus-counts curve to $\ln(1/\alpha)/C(\Delta)$. What a classifier does not give is a guaranteed false-certification rate at $10^{-7}$. Reaching that level would require on the order of $10^{8}$ held-out simulations per operating point, whereas a likelihood-ratio threshold can be set analytically and checked at moderate α.

Why not just apply generic hypothesis testing with no topology? Generic testing supplies the exponential rate, but not the structure of the hypotheses. Topology determines which samples count as equivalent and places the least-favorable pair at gap closing. That placement is what yields the $\Delta^2$ law and the disorder robustness. The ablation with a typical alternative tests whether this matters.

Proposed Experiments

All primary experiments use simulations of the 10×10 non-Hermitian circuit model from the PGIML paper and of $C_2$ photonic crystals parameterized as in Ghorashi et al.. Disorder strength and gap are swept independently. We will request the measured board data where available.

The decisive check plots measured $N^\ast$ against $\Delta$ on log-log axes for a fixed α. The theory passes if the slope is near −2 and the absolute counts fall within a factor of two of $\ln(1/\alpha)/(\kappa\Delta^2)$. It fails if the slope is shallower and disorder-dependent, which would mean robustness does not translate into a statistical guarantee.

The second check measures the realized false-certification rate of each method as disorder grows at fixed gap. PGIML's rate is expected to degrade as its clusters fragment. Ours should stay at nominal.

The third check, for skin-effect sensing, compares Chernoff information per count for open- versus periodic-boundary readouts of a small perturbation. The amplification claim survives only if the open-boundary readout gains information per count rather than just signal amplitude.