arXiv:2610.00505
Walshness: an intrinsic neural-network representability metric for quantum states
Nisarga Paul, Kehuang Chen, Jessica K. Jiang, Haimeng Zhao, Di Luo
Abstract
Neural quantum states (NQS) have emerged as powerful representations of quantum states with rapidly expanding applications across quantum many-body physics. Yet our understanding of when neural networks can efficiently represent physical quantum states remains limited, in part due to the nonlinear parameterization of NQS, the intricate sign structure of quantum states, and the sensitive dependence on basis choice. We introduce Walshness, a complexity metric of a quantum state which quantifies whether it admits a compact neural-network representation. We rigorously prove that a general quantum many-body state admits an efficient NQS representation if it has bounded Walshness, and vice versa, for various NQS architectures. We provide an algorithm for finding the optimal basis by minimizing the energy of a corresponding classical spin model. This allows us to empirically improve the learnability of various physical quantum states, including ground states of the transverse-field Ising model and mixed-field toric code, often reducing infidelity by orders of magnitude, and generalizes the basis choice given by the well-known Marshall sign rule in frustrated antiferromagnets. Our results provide an intrinsic complexity metric for quantum states with provable connections to their neural-network representability.
AI Summary
The paper asks whether a basis-dependent property of quantum states can characterize their neural-network representability and guide local basis selection. It defines fixed-basis Walshness as $W_\alpha^Z=\alpha^{-1}\log\mathbb E_{r\sim p_\psi}e^{\alpha|r|}$, where $p_\psi$ is the normalized squared Walsh–Hadamard spectrum, and intrinsic Walshness by minimizing over single-qubit unitaries. The latter is additive, vanishes exactly for product states, and bounds bipartite entanglement while remaining distinct from it. For fixed $\alpha$ and constant target infidelity, fixed-basis Walshness $\omega$ implies an MLP representation with width $n^{O(\omega)}$ and depth $O(\log\omega)$. The construction truncates the spectrum at degree $q=\lceil\omega+\alpha^{-1}\log(1/\varepsilon)\rceil$ and represents the retained parity functions. Converse results impose depth lower bounds—not general parameter-count lower bounds—under specified high-Walshness conditions: $\Omega(\log n)$ for polynomial activations and $\Omega(\log n/\log\log n)$ for standard analytic activations. A separate multiplicative metric based on $\log\psi$ characterizes polynomial-size, bounded-connectivity RBMs. Local-basis minimization is rewritten as a classical spin-energy optimization involving quantum correlators, then approached through single-site updates. Supervised fits to Ising and mixed-field toric-code ground states show infidelity reductions of several orders of magnitude in some regimes. Optimized bases also recover Marshall sign patterns and yield canted alternatives in frustrated examples. The numerical optimization can encounter local minima; evaluating its nonlocal objective at large sizes remains challenging. Representability guarantees do not establish efficient training or sampling. Scalable basis selection and systematic variational ground-state applications remain open.
Central claim may not hold: The MLP converse supplies depth lower bounds compatible with efficient polynomial-size networks, rather than proving that efficient NQS representations require bounded Walshness. The stated tanh-network depth lower bound is contradicted by a one-hidden-neuron approximation to |0^n> with vanishing infidelity, despite that state's limiting fixed-basis Walshness being n/2. The RBM characterization uses a different, logarithm-dependent multiplicative Walshness and restricted connectivity, so it does not establish the abstract's general converse for the original metric. The basis-search algorithm is a local optimization heuristic without a global-optimality guarantee, contrary to the abstract's unqualified promise of finding the optimal basis.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 9+36 pages; v2 submitted version