Walshness: an intrinsic neural-network representability metric for quantum states

AI summary

Overview

Research area: Quantum many-body physics, neural quantum states (NQS), and the theory of neural-network representability of quantum states.

Technical level: Advanced. The paper combines quantum many-body theory, learning theory, and numerical optimization; the central idea is accessible, but the theorems and proofs require familiarity with quantum states, neural networks, and asymptotic complexity notation.

Scope: The paper introduces “Walshness,” a basis-dependent and intrinsic complexity measure for quantum states, proves two-sided connections between low Walshness and compact neural-network representations, and demonstrates that minimizing Walshness over local basis choices improves NQS training and recovers generalized sign rules.

What This Paper Is About

Neural quantum states are powerful tools for representing quantum many-body wavefunctions, but there has been no intrinsic, quantifiable metric—analogous to entanglement for tensor networks—that predicts whether a given state is easy or hard for a neural network to represent. The problem is complicated by basis dependence: the same state can look simple or highly oscillatory depending on the local basis used to write its amplitudes. This paper proposes Walshness as a basis-sensitive complexity metric, defines an intrinsic version by minimizing over single-site unitary rotations, and shows both theoretically and numerically that low Walshness corresponds to efficient neural-network representability.

Key Contributions

1. Definition of Walshness. The authors define the fixed-basis (\alpha)-Walshness from the Walsh–Hadamard transform of the wavefunction, and an intrinsic Walshness obtained by minimizing over all tensor products of single-site (\mathrm{SU}(2)) unitaries.

2. Two-sided representability theorems. They prove that low fixed-basis Walshness implies an efficient multilayer perceptron (MLP) representation with width (n^{O(\omega)}) and depth (O(\log \omega)), while shallow MLPs can only represent states with low Walshness. An analogous result is given for restricted Boltzmann machines (RBMs) using a multiplicative version of Walshness.

3. Algorithm for finding an optimal basis. They show that minimizing Walshness over local bases is equivalent to minimizing the energy of a classical spin model whose couplings are multi-spin correlations of the quantum state, and they provide an exact single-site update algorithm for this nonconvex optimization.

4. Numerical demonstration and recovery of sign rules. Walshness-minimizing bases improve NQS training for the transverse-field Ising model and the mixed-field toric code, often reducing infidelity by orders of magnitude, and they recover and generalize the Marshall sign rule in frustrated antiferromagnets.

Main Findings

• Basic properties of Walshness: The fixed-basis (\alpha)-Walshness satisfies (0 \leq \mathrm{W}^{Z}_{\alpha} \leq n). It is invariant under rescaling and qubit permutations, additive under tensor products, and vanishes if and only if the state is a product state. The family is nondecreasing in (\alpha); as (\alpha \to 0) it reduces to the mean Hamming weight of the Walsh spectrum, and as (\alpha \to \infty) it becomes the largest Hamming weight in the spectrum’s support.

• Interpretation: The Walsh spectrum is the outcome distribution of measurements in the (X) basis, and (\alpha \mathrm{W}^{Z}_{\alpha}) is the cumulant-generating function of the number of (X=-1) spins in that distribution. Intuitively, Walshness quantifies how many discrete Fourier modes are present in the wavefunction.

• Analytic examples: For GHZ states, the fixed-basis Walshness is approximately (n g_{\alpha}), where (g_{\alpha} = \frac{1}{\alpha}\log\frac{1+e^{\alpha}}{2}), while the intrinsic Walshness is (n/2) for even (n). Haar-random states and random stabilizer states satisfy (\mathrm{W}^{Z}{\alpha} = n g{\alpha} + \tilde{o}(1)).

• Distinct from entanglement: For any bipartition (L:R), the entanglement entropy obeys (S_{L:R} \leq \frac{n}{2} h_{2}(\mathrm{W}{\alpha}/n)) whenever (\mathrm{W}{\alpha} \leq n/2). Extensive entanglement therefore requires extensive Walshness, but low-entanglement states can have extensive Walshness. GHZ states illustrate the distinction: they have constant entanglement across any cut but extensive intrinsic Walshness.

• MLP representability: If a wavefunction has fixed-basis Walshness (\omega), it admits an MLP representation of width (n^{O(\omega)}) and depth (O(\log \omega)) to any constant infidelity. For constant (\omega), (\alpha), and infidelity tolerance, the network has polynomial width and constant depth. Conversely, if (\mathrm{W}^{Z}{0} = \Omega(n)) or (\mathrm{W}^{Z}{\alpha} \geq n - O(1)), any MLP with sufficiently small constant infidelity must have depth (\Omega(\log n)) for polynomial nonlinearities, or (\Omega(\log n / \log\log n)) for standard analytic nonlinearities.

• RBM representability: A wavefunction with bounded multiplicative Walshness admits an RBM representation with (\mathrm{poly}(n)) hidden units and bounded hidden-visible connectivity to any constant infidelity. Conversely, an RBM with bounded hidden-visible connectivity has bounded multiplicative Walshness.

• Ferromagnet learning test: The states (|+\boldsymbol{x}\rangle), (|+\boldsymbol{z}\rangle), and (|-\boldsymbol{x}\rangle) have fixed-basis Walshness (0), (n g_{\alpha}), and (n), respectively, although all have zero intrinsic Walshness. In training MLP-NQS with (\tanh) nonlinearities, two hidden layers of width six, infidelity loss, and (5 \cdot 10^{4}) epochs of Adam optimization, (|+\boldsymbol{x}\rangle) was easiest, (|+\boldsymbol{z}\rangle) intermediate, and (|-\boldsymbol{x}\rangle) hardest, matching the Walshness ordering. Representing (|-\boldsymbol{x}\rangle) amounts to computing the (n)-bit parity function, which the paper describes as notoriously difficult in learning theory.

• TFIM analytics: For the ground state of the transverse-field Ising model (H_{\mathrm{TFIM}} = -\sum_{i} Z_{i}Z_{i+1} - g\sum_{i}X_{i}), the authors derive an exact free-fermion expression for (\mathrm{W}^{Z}{\alpha}). As (g \to 0) or (g \to \infty), the state approaches a GHZ state or (|+\boldsymbol{x}\rangle), and the Walshness approaches the corresponding value. At the critical point (g=1), the Walshness is continuous but its derivative diverges logarithmically as (\partial{g}(\mathrm{W}^{Z}_{\alpha}/n) \simeq (2\pi\alpha)^{-1}\tanh\alpha \log|g-1|).

• Optimal TFIM bases: The intrinsic Walshness approximately equals the fixed-basis Walshness for large (g), but the two differ as (g \to 0). For (g<1), the Walshness-minimizing basis shows period-two staggering that spontaneously breaks the translation symmetry of the TFIM.

• Training advantage for TFIM: Training MLP-NQS on the ferromagnetic ground state at (g=0.2) in the Walshness-minimizing basis can lead to several orders of magnitude lower infidelity than training in the computational basis at comparable compute. The advantage is most prominent for small (g) in the ferromagnetic phase and fades as (g \to 1), where the two bases coincide.

• Marshall sign rule recovery: In an XY antiferromagnet with (H_{\mathrm{XY}} = \sum_{j}(X_{j}X_{j+1} + gY_{j}Y_{j+1})), Walshness minimization recovers the Marshall sign rule at small (g>0) with no input other than the wavefunction. At larger (g), the optimal basis departs from the Marshall rule and yields lower Walshness. In a 2D XYZ antiferromagnet on a square lattice with (K_{ij} = X_{i}X_{j} + 0.3Y_{i}Y_{j} + 0.1Z_{i}Z_{j}), the Walshness-minimizing basis recovers a checkerboard sublattice rule at (g=0.2) and a stripe sublattice rule at (g=0.8), with a crossover or transition near (g \approx 0.65); at (g=0.5) it produces a canted basis outside the standard Marshall sign rule.

• Mixed-field toric code: Numerical experiments on the mixed-field toric code again show that training in the Walshness-minimizing basis can give substantially lower infidelity than the computational basis when (\mathrm{W}{\alpha} < \mathrm{W}^{Z}{\alpha}).

Methodology in Plain English

The authors begin with the Walsh–Hadamard transform, which is the discrete Fourier transform over bit strings and is closely related to measuring a state in the (X) basis. From the transformed amplitudes, they build a probability distribution over bit strings—the Walsh spectrum—and define Walshness as a cumulant-generating function of the Hamming weight of those bit strings. This captures how spread out and how oscillatory the wavefunction is in a chosen basis.

Because this quantity depends on the basis, they define an intrinsic version by minimizing over all independent single-site unitary rotations. Rather than searching over unitary matrices directly, they prove that the minimization can be rewritten as the energy minimization of a classical spin model: each qubit contributes a three-dimensional unit vector, and the energy is an expectation of a product of local operators involving those vectors. The couplings of this classical model are the multi-spin correlations of the quantum state.

For the optimization, they use an exact single-site update: freeze all spins except one, compute an environment state and a local vector from the quantum state, and set the selected spin along that vector. Sweeping over sites until convergence gives a numerical upper bound on the intrinsic Walshness and a corresponding local basis. They report rapid convergence with a tolerance of (10^{-10}) and agreement across many random initializations, while noting that local minima are possible in this nonconvex problem.

For numerical learning experiments, they rotate the target state into the Walshness-minimizing basis, then train an MLP to fit the rotated amplitudes using infidelity as the loss. For variational calculations, they note that a basis chosen from an approximate ground state can be used to train an NQS with a rotated Hamiltonian (UHU^{\dagger}), which preserves the support of each Hamiltonian term because (U) is a product of single-site unitaries.

Why This Matters

The paper supplies a conceptual counterpart to entanglement for neural quantum states: a quantitative, intrinsic property that predicts whether a state has a compact neural representation. It also turns basis selection—long treated as a heuristic or left to physical intuition—into a concrete optimization problem with a classical spin-model interpretation. This is relevant because the sign structure and basis choice of a wavefunction are central not only to NQS but also to quantum Monte Carlo.

Potential application areas, though the paper does not report deployed real-world applications, include:

• Quantum many-body simulation: Choosing better bases and diagnosing representability could improve variational calculations for strongly correlated materials and lattice models.

• Quantum chemistry and materials discovery: NQS methods are used for molecular and materials problems; a basis-selection metric could make neural ansätze easier to train.

• Quantum computing software and algorithm design: Walshness could serve as a diagnostic for state preparation, ansatz design, and benchmarking when a quantum state is represented classically or learned from data.

• Quantum Monte Carlo: Because Walshness minimization generalizes sign-rule heuristics such as the Marshall sign rule, it may help identify bases that alleviate sign problems.

Industry relevance: The work is most directly relevant to quantum technology companies, computational chemistry and materials software developers, and groups building machine-learning tools for quantum simulation. The paper itself is theoretical and numerical; it does not report industry partnerships, commercial deployments, or production applications.

Future Directions

• Hierarchies of intrinsic Walshness: Enlarging the allowed basis transformations from single-site rotations to finite-depth local circuits would define a hierarchy of increasingly intrinsic complexity measures.

• Other neural architectures and fermions: Generalizing the representability results to autoregressive models and to fermionic degrees of freedom is a natural next step.

• Variational ground-state calculations: Extending the numerical work from quantum state learning to

Authors’ 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 and only if it has bounded Walshness 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 recovers 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.