Predicting Which Neural Wavefunctions Can Learn Topological States, Before Training Them

Neural networks can now represent the quantum state of many interacting electrons. Such a network, called a neural wavefunction, takes electron positions as input and outputs the amplitude and phase of the state. These models are accurate but expensive. One self-attention wavefunction needed more than 10,000 A100 GPU hours for a single molecular dimer (A Self-Attention Ansatz for Ab-initio Quantum Chemistry). The hardest targets are topological states such as fractional quantum Hall states. In these states the phase winds around each electron, and standard architectures can stall at mediocre accuracy. As a result, choosing an architecture is currently a matter of expensive trial and error. We propose a target-aware test that runs in seconds on a randomly initialized network. It measures how much of the gap between the network's current state and the target state, in both amplitude and phase, the network could close with a single linearized update. We pair this score with directly measured GPU throughput and memory, and ask whether the combination picks architectures that reach high fidelity cheaply.

The core claim is specific. Generic "zero-cost proxies" from image-classifier search look only at the network and the data. Examples are gradient signal-to-noise and the conditioning of the neural tangent kernel, which describes how outputs move when weights move. These scores are blind to the property that makes topological states hard: whether the network's reachable directions include the target's phase structure. A proxy that projects the target's phase current onto the network's tangent space at initialization should rank architectures better than generic proxies, and better than parameter count, on topological targets specifically. The advantage should shrink on easy, spatially localized states such as Wigner crystals.

What exists and what is missing

The file contains three partial bridges, and each leaves out a different piece.

WF-Bench benchmarks FermiNet and Psiformer on 31 strongly correlated target states. It does this by fitting each state's density and phase current with a second-order optimizer. It finds three things. Topological states are the hardest category. Psiformer reaches fidelity near 0.99 on laughlin_m5, where FermiNet stays below 0.90 at equal parameter budgets. Gains from more than two layers or more than about 32 determinants are small, while compute grows linearly. This paper supplies our ground truth and our main motivation. It runs no training-free estimation, no search, and no hardware measurement.

Training-Free Quantum Architecture Search ranks parameterized quantum circuits at initialization. It uses a graph path-count proxy and an expressibility proxy based on sampled fidelity, reaching Spearman correlation up to 0.85 on small spin and molecular Hamiltonians. It is the closest existing work. It differs from ours in three ways: it targets qubit circuits rather than classical neural wavefunctions, it does not use the target state's structure, and the authors note that their proxies have no trainability component.

HaeNAS combines zero-cost proxies with hardware latency tables, and ZiCo combines them with FLOP budgets. Both work only on vision networks.

From classical training-free search we reuse one method. Chen et al. estimate generalization with closed-form kernel regression on the tangent kernel at initialization. Our proxy is the wavefunction analogue of that idea, with the target supplied explicitly.

Two findings constrain the design. First, many proxies fail to beat raw parameter count: ZiCo reports Kendall τ of 0.61 versus 0.57 for #Params on NATS-Bench-TSS (ZiCo). Any proxy we propose must therefore be tested with parameter count held fixed. Second, proxy scores depend on the initialization distribution, so we average scores over random seeds.

From the hardware side, HW-NAS-Bench and HAT show two things. FLOPs track measured latency poorly, with a FLOPs–latency correlation as low as 0.18 on the Edge TPU. And the best architecture shape depends on the device. This motivates measuring cost directly rather than counting operations.

The physics work on projected quantum geometry (Low-energy optical absorption in correlated insulators) motivates the proxy in a looser sense. It shows that how the geometry of a correlated state varies, rather than its averages, governs observable behavior. Our proxy uses the geometry of the network's parameter manifold. That geometry is related to the quantum Fisher information of the network-generated state family, but it is a different object from that paper's projected position operator, and we do not claim an equivalence.

Core Claim and the Proxy

Let $\psi_\theta$ be the network at random initialization, with $N_e$ electron positions $x$. WF-Bench fits phase currents rather than raw phases because phase windings make the log-phase multivalued. Following that choice, we work with the complex gradient field $g_\theta(x) = \nabla_x \log \psi_\theta(x)$. Its real part is the amplitude gradient and its imaginary part is the phase current. Let $g_\star$ be the same field for the target. The proxy asks what fraction of the field residual a single linearized parameter step can remove:

$$ R^2_{c} = 1 - \frac{\min_{\delta}\ \mathbb{E}_{x}\,\big\| \mathrm{c}\!\left[g_\star(x) - g_\theta(x) - J_\theta(x)\,\delta\right]\big\|^2 + \lambda\|\delta\|^2}{\mathbb{E}_{x}\,\big\|\mathrm{c}\!\left[g_\star(x) - g_\theta(x)\right]\big\|^2}. $$

Here $J_\theta(x) = \partial g_\theta(x)/\partial\theta$ is the Jacobian of the field with respect to the parameters. The operator $\mathrm{c}$ takes either the real part (giving $R^2_{\rm amp}$) or the imaginary part (giving $R^2_{\rm phase}$). The scalar $\lambda$ is a fixed ridge regularizer, and $x$ is sampled from a mixture of $|\psi_\theta|^2$ and $|\psi_\star|^2$.

Computing the score costs a few thousand samples and one regularized least-squares solve in the kernel form, which has the size of the sample count. No optimization step is taken. The score is target-aware, it separates amplitude from phase, and the ridge term penalizes ill-conditioned directions. The ridge term therefore folds in the trainability component that TF-QAS lacked.

Our prediction is as follows. On topological targets, $R^2_{\rm phase}$ ranks fully trained fidelity better than any generic proxy. It does so even within groups of architectures with similar parameter counts. On Wigner crystals, where WF-Bench finds fitting easy, the amplitude term and the generic proxies should do about as well.

Research Program

We first build a search space around the two WF-Bench architectures. It varies backbone type (multi-stream MLP versus self-attention), depth (1 to 4), width, number of attention heads, number of determinants (1 to 64), and envelope and Jastrow-factor choices. This yields a few thousand configurations.

For a stratified sample of roughly 150 to 200 architectures, we run the full WF-Bench fidelity protocol to obtain ground-truth fidelity. The targets are laughlin_m5, a Moore-Read state, one d-wave BCS state, and one Wigner crystal, each at two small electron numbers. The protocol uses the same loss, optimizer, and step budget as WF-Bench. This ground truth is the dominant cost of the project. It is affordable because fidelity matching at small $N_e$ is far cheaper than variational energy minimization.

For every configuration we compute the proxy over five initialization seeds, along with all baseline proxies. On an A100 and on one smaller consumer GPU, we also measure wavefunction-plus-Laplacian evaluations per second and peak memory, since that is the inner loop of Monte Carlo sampling. These networks are small, so direct measurement replaces lookup tables. This also sidesteps the failure of additive lookup tables noted by HW-NAS-Bench.

Finally, we run a constrained evolutionary search over the full space. It maximizes the proxy subject to a hinge-style throughput floor and memory cap, as in DUCCIO (Enhancing NAS with Multiple Hardware Constraints). We then fully train the top candidates.

Baselines

The naive reference point is to adopt WF-Bench's own recommendation without search: a two-layer Psiformer with 32 determinants, trained with the standard protocol. Its fidelity on laughlin_m5 is reported as near 0.99 (WF-Bench). The searched model beats this baseline if it matches that fidelity at measurably higher throughput or lower memory on a given GPU, or if it exceeds that fidelity at equal throughput. The threshold for "matching" is set by the seed-to-seed spread we measure.

A second naive baseline ranks architectures by parameter count or by FLOPs. It is the bar many proxies fail to clear. ZiCo reports #Params at Kendall τ 0.57 on NATS-Bench-TSS (ZiCo). No reported result was found on Lacuna for parameter count as a predictor of neural-wavefunction fidelity, so we measure it ourselves on the identical architecture sample.

The strongest existing training-free methods come from the vision literature, which has stronger proxies than the quantum literature, and we port them to wavefunctions. ZiCo is computed on the WF-Bench pre-training loss. It reports Kendall τ 0.61 on NATS-Bench-TSS. The tangent-kernel condition number and kernel-regression score of Chen et al. are also ported, but the file gives no correlation figure for them. TF-QAS's sampled expressibility proxy is adapted by measuring the spread of fidelities between pairs of randomly initialized networks. It reports Spearman correlation up to 0.85, on qubit circuits (TF-QAS). All of these receive the same samples, the same seeds, and the same wall-clock budget per architecture as our proxy. No reported result was found on Lacuna for any zero-cost proxy applied to neural wavefunctions.

The key ablation isolates target-awareness. It replaces $g_\star$ with a random target field of matched norm, which keeps the kernel's conditioning information but removes the target's phase structure. A second variant drops the phase term and keeps only $R^2_{\rm amp}$. If the full proxy does not beat both variants on topological states, the claim fails.

Our proxy beats the field if its rank correlation with trained fidelity exceeds that of every baseline on the topological targets. The confidence intervals come from bootstrapping over architectures. The advantage must also survive within groups of similar parameter count.

Why not just…?

Why not just rank by parameter count or FLOPs? WF-Bench already shows parameter count is unreliable here. At equal parameters, the two backbones differ by roughly 0.09 in fidelity on laughlin_m5, and extra determinants add cost without improving accuracy. HW-NAS-Bench shows FLOPs can track measured latency poorly. Both orderings are still included as baselines, and the partial-correlation analysis controlling for parameter count settles the question directly.

Why not just train each candidate briefly and rank by early fidelity? This is the serious threat. A few hundred optimizer steps may rank architectures well, especially since WF-Bench documents self-trapping on phase-winding states that might appear early. We compare head-to-head at equal wall-clock time per architecture. If brief training ranks as well, the proxy's value reduces to speed, and we report that honestly. In that case we would reposition the proxy as a pre-filter in front of short training rather than a replacement for it. The cost-matched comparison decides which outcome holds.

Why not just use Psiformer at WF-Bench's recommended depth and determinant count? For a single GPU and a single target, that may be nearly optimal, and it is our naive baseline. The case for search depends on two outcomes. One is that the best shape changes across devices, as HAT found for transformers in translation. The other is that cheaper architectures match its fidelity. If neither happens, the hardware part of this program is a negative result worth publishing.

Proposed Experiments and Decisive Checks

The first decisive check is the ranking study. We report Kendall τ and Spearman correlation between each proxy and trained fidelity, per target category, with bootstrap intervals and with parameter count controlled. The hypothesis is falsified if, on laughlin_m5 and Moore-Read, $R^2_{\rm phase}$ does not exceed both ported ZiCo and the random-target ablation. It is also falsified if our proxy's advantage over generic proxies is as large on Wigner crystals as on topological states, which would mean it measures something other than phase reachability.

The second check is transfer across electron number. We test whether rankings computed at the smaller $N_e$ predict trained fidelity at the larger $N_e$, by checking whether proxy-selected architectures sit on better curves of the WF-Bench form $F = 1-\alpha(N_e-2)^\beta$. A training-free proxy that fails to transfer across $N_e$ would have little practical value.

The third check is the constrained search. We report fidelity against measured throughput for the selected architectures on both GPUs, compared against the naive baseline and against search driven by ported ZiCo under identical constraints. We also test whether the selected shapes differ between devices.

Lattice models, such as the sign-problem-free Ising-mediated Dirac fermions of Dynamical generation of topological masses, are a natural follow-up. That would require target fields from exact or quantum Monte Carlo data, which are outside this study's scope.