Chirality Is a Hue: Learning the Chiral Central Charge Without Regressing to Zero

Colorization networks trained with squared-error loss turn red-or-blue objects gray. When the input does not determine which of two vivid answers is right, the loss-minimizing prediction is their average. This proposal argues that the same failure will appear, in a sharper and more checkable form, when neural networks are trained to estimate a quantized topological number from measurement data on quantum states. The number in question is the chiral central charge $c_-$. It counts the net number of edge modes that circulate one way around a two-dimensional topological phase, and it flips sign under time reversal. A state and its complex conjugate have identical entanglement spectra but opposite $c_-$. A regressor fed only time-reversal-symmetric information should therefore output $c_- \approx 0$ for every chiral phase, which is the quantum analogue of predicting gray.

The fix we propose borrows directly from colorization. We predict a probability distribution over the allowed discrete values of $c_-$, rebalance classes because non-chiral phases dominate generated data the way gray pavement dominates photographs, and decode with a temperature-controlled mean. We pair this with a symmetry diagnostic that tells in advance whether a mean collapse is a fixable model failure or an unavoidable identifiability limit. A colleague could summarize it in two sentences. Learned estimators of chirality from quantum measurements will "desaturate" toward zero for the same reason colorizers go gray. Treating $c_-$ as a quantized class, and making sure the inputs carry information that flips under time reversal, fixes this, and a clean ablation shows exactly when it cannot be fixed.

Core Claim

The modular Hamiltonian of a region $A$ is $K_A = -\ln \rho_A$, the operator whose exponential gives the region's reduced density matrix. Modular commutator in gapped quantum many-body systems and Chiral Central Charge from a Single Bulk Wave Function showed that for adjacent regions $A, B, C$ of a disk in a gapped two-dimensional state,

$$ J(A,B,C) = i\,\mathrm{Tr}\big(\rho_{ABC}[K_{AB}, K_{BC}]\big) = \frac{\pi}{3}\, c_- , $$

and that $J$ is odd under time reversal. They also showed that nonzero $c_-$ forbids a real-valued wavefunction in any local product basis. That obstruction has a concrete consequence for learning. The information that distinguishes $c_-$ from $-c_-$ lives only in the imaginary parts of local correlations. In randomized single-qubit Pauli measurements, those imaginary parts are reached only through correlators containing an odd number of $Y$ factors.

We make three claims. First, an estimator whose inputs are invariant under complex conjugation provably cannot separate $\pm c_-$. Under squared-error training such an estimator converges to the class-weighted mean, so it reports $|c_-| \approx 0$ on chiral states. Inputs of this kind include Rényi entropies, which is exactly what Machine learning spatial geometry from entanglement features fits by mean squared error, and snapshots in only the $X$ and $Z$ bases. Second, a quantized classifier on the same inputs should output a calibrated near-even split between $+c_-$ and $-c_-$. That split is the honest answer, the analogue of the $P(\text{red}) \approx P(\text{blue})$ output in Colorful Image Colorization. Third, once the inputs include odd-$Y$ information, the classifier should reach the correct quantized $c_-$ at a smaller measurement budget than either squared-error regression or plug-in evaluation of $J$. Two effects drive this. Near phase transitions and at small system sizes, regression produces fractional, "desaturated" values. Plug-in evaluation requires taking a matrix logarithm of a noisy reconstructed density matrix, which is unstable.

What is new relative to the file is not "use classification for phases." It is the specific prediction, with a derivation and an ablation, that learned topological-invariant estimators collapse toward the time-reversal-symmetric mean, plus a measurement-budget comparison against the modular-commutator estimator itself. No paper in the file connects the modular commutator to learned estimation or to mean collapse.

Research Program

The data comes from two families of states with exactly known $c_-$.

The first family is free-fermion Chern insulators on square lattices. A single-particle mass parameter tunes these between Chern numbers $0$, $+1$ and $-1$, and multi-band variants reach $\pm 2$. For these Gaussian states, modular Hamiltonians, $J$ and simulated measurement snapshots are cheap to compute at large sizes.

The second family is the bosonic lattice Laughlin (semion) state with $c_- = 1$ at $N = 12$ to $26$ qubits, together with its complex conjugate at $c_- = -1$. This reproduces the regime of Modular commutator in gapped quantum many-body systems. Non-chiral gapped states with $c_- = 0$ fill out the dataset and are deliberately over-represented, so that the rebalancing question is real rather than staged.

Inputs are classical shadows, meaning collections of single-shot measurement outcomes in randomly chosen Pauli bases, at measurement budgets $M$ spanning several orders of magnitude. Each budget is restricted to an annulus-shaped $ABC$ partition so that the model sees only local data, as the invariant requires. We build three input variants. The first uses all three Pauli bases. The second uses $X$ and $Z$ only, and is therefore exactly invariant under complex conjugation in that basis. The third consists only of estimated Rényi entropies of subregions.

The model is a permutation-invariant set encoder over snapshots. It is trained with one of four output heads: a squared-error regression head on $c_-$; a mixture-density head, meaning a small mixture of Gaussians; a categorical head over the allowed set, for example $\{-2, -1, 0, 1, 2\}$; and the same categorical head with inverse-frequency rebalancing following Colorful Image Colorization. Two decoding rules for the categorical head are compared: the most probable class, and the temperature-annealed mean $q_i \propto \exp(\ln p_i / T)$ from the same paper. We expect the annealed mean to matter near transitions. There the true finite-size $J$ interpolates, which is the analogue of the jitter-versus-blur tradeoff in colorization.

A smaller second regime tests whether the same logic explains a bias in learned emulators of Page curves. The Page curve is the rise-and-fall of the entanglement entropy of a black hole's emitted radiation. The test uses the closed-form two-branch curves of evaporating-black-hole models with end-of-the-world branes, as in Page curve of AdS-Vaidya model for evaporating black holes and Page curve of effective Hawking radiation. In those models the radiation entropy is the minimum of a rising surface contribution and a flat island contribution.

The emulator sees time and an observable mass parameter, but not the brane parameter, so the Page time varies across instances. A squared-error emulator must then output the instance-averaged curve, which is smooth and has no kink. We will label this output the "annealed" ensemble average. That object resembles the disorder-averaged quantity whose physical status is debated in Eigenstate Thermalization and Disorder Averaging in Gravity. We have not established that the two are the same, and this part of the program will not test the resemblance; the label is offered for interpretation only. The comparison emulator is a classifier over which branch dominates, combined with branch-wise regression. It should recover single-instance kinks. This regime is deliberately secondary. It checks that the diagnostic transfers beyond chirality; it is not a claim about resolving the information paradox.

Baselines

Plug-in modular commutator. The strongest non-learned competitor reconstructs $\rho_{ABC}$ from the same shadows, evaluates $J$ directly, divides by $\pi/3$, and rounds to the nearest allowed value. With exact wavefunctions this method is very accurate. Modular commutator in gapped quantum many-body systems reports $1.054 \pm 0.013$ against the theoretical $\pi/3 \approx 1.047$ for the semion model after finite-size extrapolation, an error of about 0.7%. The same paper reports raw small-system errors of around 10%. We run it on identical snapshot budgets, give it the same partition geometry, and tune its density-matrix regularization on held-out states with a search comparable to the learned models' hyperparameter budget. Beating it means reaching a fixed accuracy on quantized $c_-$, for example 95% correct, at a smaller $M$, or reaching higher accuracy at equal $M$. We also report the budget at which the two methods converge, because the plug-in estimator should eventually win or tie.

Squared-error regression. This is the naive reference point and the closest analogue of entanglement-feature learning, which fits entropies by squared error in Machine learning spatial geometry from entanglement features. It uses the same encoder, data, steps and tuning; only the loss differs. Our prediction is quantitative. On conjugation-invariant inputs, its mean $|\hat c_-|$ on chiral states should approach zero.

Mixture-density regression. This is the strongest learned competitor. It handles multimodality without discretization, so it isolates whether the gain comes from quantization or merely from a multimodal output. It is trained identically.

Ablations. Rebalanced and unrebalanced categorical heads isolate the role of class weighting. The three input variants isolate the role of time-reversal-odd information, and this is the key ingredient. A further check augments training with complex-conjugated states and flipped labels. This should be harmless when odd-$Y$ data is present and should force exact even splits when it is absent.

Why not just…?

Why not just compute the modular commutator directly and round it? This is the most serious threat. The plug-in baseline exists precisely to test it. With full wavefunctions, direct evaluation already works, at 0.7% error after extrapolation. The open question is the measurement-limited regime, where the logarithm of a noisy estimated density matrix amplifies sampling error. If the plug-in estimator matches the classifier at every budget, we will report that. The contribution then narrows to the symmetry diagnostic and the documented failure of regression estimators.

Why not just use a mixture-density network instead of discrete bins? It may work as well. Nothing in the file shows it failing. Because $c_-$ has known discrete support, we expect bins to give better calibration at transitions and simpler rebalancing. The mixture-density baseline decides this question. If it ties, the paper's claim becomes "multimodal heads are required" rather than "quantization is required."

Why not just feed the network entanglement entropies, as geometry-learning work does? Entropies are spectra of reduced density matrices, and complex conjugation leaves spectra unchanged. So no estimator can recover the sign of $c_-$ from entropies alone. This follows from the time-reversal oddness established in Chiral Central Charge from a Single Bulk Wave Function. The entropy-only arm demonstrates the limit empirically; it is not a strawman.

Why not just train a standard phase classifier and skip the analysis? That classifier is one arm of our study. What a plain phase classifier would not provide is a prediction of when its outputs must be split, and of how the outputs relate to the physically meaningful invariant $J$. The annealed-mean decoder supplies the second link: it maps classifier outputs back onto values comparable with finite-size $J$.

Decisive Experiments

The cleanest test is the $X$/$Z$-only arm. The predicted outcomes are mean $|\hat c_-| \to 0$ for squared-error regression and a posterior split between $\pm c_-$ within sampling error of 50/50 for the classifier, holding on both the Chern insulators and the semion pair. Turning on $Y$-basis snapshots should restore correct signs. If regression nonetheless recovers signs in the $X$/$Z$-only arm, either our basis-invariance argument is wrong or the simulation leaks information, and we would investigate before proceeding.

The second test is the accuracy-versus-$M$ curve against the plug-in commutator on semion systems from $N = 12$ to $26$. The question is whether learning buys budget efficiency at small sizes, where raw $J$ errors sit around 10%.

The third test sweeps the Chern-insulator mass across transitions. We compare calibration, measured as expected calibration error, and a "chiral saturation" score, defined as the mean $|\hat c_-|$ on chiral test states and playing the role of colorfulness, across the regression, mixture-density and annealed-categorical heads. We also test generalization to system sizes held out from training.

The fourth test, in the Page-curve regime, checks whether squared-error emulators erase the kink when Page times vary across instances. If they do, ML emulation of entanglement dynamics with hidden parameters returns ensemble averages, and should be reported as such.