pith. sign in

arxiv: 2512.11632 · v3 · submitted 2025-12-12 · 🪐 quant-ph · cond-mat.stat-mech

Basis dependence of Neural Quantum States for the Transverse Field Ising Model

Ronald Santiago Cortes , Aravindh S. Shankar , Marcello Dalmonte , Roberto Verdel , Nils Niggemann This is my paper

Pith reviewed 2026-05-16 23:09 UTC · model grok-4.3

classification 🪐 quant-ph cond-mat.stat-mech
keywords Neural Quantum StatesTransverse-Field Ising ModelBasis DependenceRestricted Boltzmann MachinesCumulant ExpansionGround State DegeneracyAmplitude UniformityMany-Body Quantum Simulation
0
0 comments X

The pith

Neural quantum state performance for the transverse-field Ising model varies with the computational basis due to ground-state degeneracies and amplitude uniformity.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper examines how restricted Boltzmann machine neural quantum states represent the ground state of rotated versions of the transverse-field Ising model. Different bases change whether the ground state is degenerate and how uniform its amplitudes and phases are, which directly affects how accurately the neural network can capture it. The authors tie these physical features to the rate at which a cluster or cumulant expansion of multi-spin operators converges. A reader cares because the work supplies a concrete way to predict which basis will work well before running expensive optimizations.

Core claim

The basis-dependence of NQS performance is linked to the convergence properties of a cluster or cumulant expansion of multi-spin operators, which in turn depends on the presence of ground state degeneracies as well as the uniformity of amplitudes and phases in the chosen basis.

What carries the argument

The cluster or cumulant expansion of multi-spin operators, whose convergence properties in a given basis determine how well an NQS can represent the corresponding ground state.

If this is right

  • Bases without ground-state degeneracies yield higher NQS accuracy for the transverse-field Ising model.
  • Uniform amplitude and phase distributions in the ground state improve convergence of the neural-network representation.
  • The cluster-expansion convergence rate provides a direct diagnostic that connects physical basis choice to numerical performance.
  • The same framework can be used to gauge whether NQS methods are likely to succeed on a new Hamiltonian before computation begins.
  • An optimal computational basis for a given model can be identified by inspecting ground-state properties and expansion behavior.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same basis-selection logic may apply to other neural-network architectures used in quantum many-body problems.
  • Practitioners could screen candidate bases by computing low-order cumulants before launching full NQS training runs.
  • The approach suggests a route to automated basis optimization that minimizes the number of required samples or parameters.
  • Similar performance variations are likely to appear in models with different symmetries or in open quantum systems.

Load-bearing premise

Observed performance differences across bases arise primarily from ground-state physical properties such as degeneracies and amplitude uniformity rather than from details of the neural-network optimization procedure.

What would settle it

A demonstration that NQS accuracy differences disappear when the same ground-state wave function is represented in two bases that differ only in how the optimizer is tuned, or conversely that the differences remain after identical hyperparameter sweeps in every basis.

Figures

Figures reproduced from arXiv: 2512.11632 by Aravindh S. Shankar, Marcello Dalmonte, Nils Niggemann, Roberto Verdel, Ronald Santiago Cortes.

Figure 1
Figure 1. Figure 1: Energy gap ratio for a finite-size TFIM with [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Flowchart summarizing whether an RBM is expected to accurately learn [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: a) Relative energy error δE and infidelity I of the wavefunction obtained by energy minimization over different basis rotations. For each angle, 10 independent realizations are represented by the markers. The dotted line connects the realizations with the lowest energy at each angle, and the solid line shows the infidelity of these selected realizations. b) Average sign of the true ground state |Ψ〉, the su… view at source ↗
Figure 4
Figure 4. Figure 4: Amplitudes Ψθ (s) = 〈s|ψRBM〉θ of the ground state wavefunction of the TFIM with λ = 1.5 for a) θ = 0 and b) θ = π, sorted by increasing absolute value |Ψ(s)|. c) Energy error for different realizations of the optimization. transformation, i.e. ΘH (x, x ′ ) = ϑH (x ′ ) − ϑH (x). Again, in this case, the problem of optimiz￾ing amplitudes and phases decouples, as the optimal phase structure can be found by fi… view at source ↗
Figure 5
Figure 5. Figure 5: a) Relative energy error (red) and infidelity (blue) of the RBM wavefunc [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Infidelity over the number of coefficients of the truncated cumulant expan [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Infidelity over the number of coefficients of the truncated cumulant expan [PITH_FULL_IMAGE:figures/full_fig_p022_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Relative error of the coefficients in the truncated cumulant expansion. The [PITH_FULL_IMAGE:figures/full_fig_p023_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Relative error of the coefficients in the truncated cumulant expansion. The [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Infidelity curves vs the number of coefficients kept in the cluster expan [PITH_FULL_IMAGE:figures/full_fig_p024_10.png] view at source ↗
read the original abstract

Neural Quantum States (NQS) are powerful tools used to represent complex quantum many-body states in an increasingly wide range of applications. However, despite their popularity, at present only a rudimentary understanding of their limitations exists. In this work, we investigate the dependence of NQS on the choice of the computational basis, focusing on restricted Boltzmann machines. Considering a family of rotated Hamiltonians corresponding to the paradigmatic transverse-field Ising model, we discuss the properties of ground states responsible for the dependence of NQS performance, namely the presence of ground state degeneracies as well as the uniformity of amplitudes and phases, carefully examining their interplay. We identify that the basis-dependence of the performance is linked to the convergence properties of a cluster or cumulant expansion of multi-spin operators -- providing a framework to directly connect physical, basis-dependent properties, to performance itself. Our results provide insights that may be used to gauge the applicability of NQS to new problems and to identify the optimal basis for numerical computations.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper investigates the basis dependence of restricted Boltzmann machine (RBM) Neural Quantum States (NQS) performance for the transverse-field Ising model (TFIM) under a family of rotated Hamiltonians. It attributes observed variations in NQS accuracy to ground-state properties including degeneracies, amplitude and phase uniformity, and links these to the convergence rate of cluster/cumulant expansions of multi-spin operators, providing a framework to connect physical basis-dependent features directly to representational performance.

Significance. If the central claim holds after addressing controls, the work offers a concrete way to diagnose NQS applicability via physical properties of the target state and to select optimal bases, which is useful for variational quantum many-body simulations beyond the TFIM.

major comments (2)
  1. [§4] §4 and Figure 3: The correlation between cumulant-expansion convergence and NQS performance is reported, but the protocol varies both the physical basis and the RBM training schedule (learning rate, sweeps, regularization) simultaneously. Without an ablation that fixes optimizer hyperparameters while sweeping only the basis, it remains possible that performance ordering partly reflects basis-dependent loss-landscape conditioning rather than truncation error of the expansion.
  2. [Abstract] Abstract and §3: No error bars, training-convergence diagnostics, or controls that isolate optimizer artifacts from the claimed physical mechanism (degeneracies and amplitude uniformity) are provided, which is load-bearing for the quantitative claim that basis dependence is primarily physical rather than numerical.
minor comments (2)
  1. [Figures] Figure captions should explicitly list the rotation angles and the precise definition of the cumulant expansion used for each basis.
  2. [§2] Notation for the RBM hidden-unit density and the multi-spin operator expansion should be introduced once with a clear equation reference.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their careful reading and constructive comments, which have helped us strengthen the manuscript. We address each major comment below and have performed additional numerical controls to isolate the physical mechanism from optimizer effects.

read point-by-point responses

  1. Referee: [§4] §4 and Figure 3: The correlation between cumulant-expansion convergence and NQS performance is reported, but the protocol varies both the physical basis and the RBM training schedule (learning rate, sweeps, regularization) simultaneously. Without an ablation that fixes optimizer hyperparameters while sweeping only the basis, it remains possible that performance ordering partly reflects basis-dependent loss-landscape conditioning rather than truncation error of the expansion.

    Authors: We acknowledge that the original protocol optimized hyperparameters separately for each basis, which is common but does not fully isolate the basis effect. To address this concern, we have performed a new ablation study in which the optimizer hyperparameters are held fixed (learning rate 0.01, 2000 sweeps, L2 regularization 10^{-4}) while only the computational basis is varied. The resulting performance ordering remains consistent with the cumulant-expansion convergence rates reported in the original Figure 3. These fixed-hyperparameter results have been added as a new panel in the revised Figure 3 and are discussed in §4. The correlation with the physical properties (degeneracies and amplitude/phase uniformity) is preserved, indicating that the basis dependence is not an artifact of loss-landscape conditioning. revision: yes

  2. Referee: [Abstract] Abstract and §3: No error bars, training-convergence diagnostics, or controls that isolate optimizer artifacts from the claimed physical mechanism (degeneracies and amplitude uniformity) are provided, which is load-bearing for the quantitative claim that basis dependence is primarily physical rather than numerical.

    Authors: We agree that quantitative claims require statistical controls. In the revised manuscript we have added error bars to all NQS accuracy metrics in §3 and Figure 3, obtained from ten independent training runs with different random seeds for each basis. We have also included training-convergence diagnostics (variational energy versus sweep number) for representative bases in a new supplementary figure, showing that all trainings reach stable plateaus within the allotted sweeps. These additions demonstrate that the observed performance differences track the physical properties rather than incomplete optimization. The abstract has been updated to note the inclusion of these statistical controls. revision: yes

Circularity Check

0 steps flagged

No significant circularity; central claims rest on direct numerical optimization without reduction to inputs by construction.

full rationale

The paper computes NQS performance for the TFIM ground states in rotated bases via direct RBM optimization and correlates it with the convergence rate of a cluster/cumulant expansion of multi-spin operators. Neither quantity is defined in terms of the other, no parameters are fitted to a subset and then relabeled as predictions, and no self-citation chain or uniqueness theorem is invoked to force the result. The observed correlation is obtained from independent numerical experiments on known Hamiltonians; the derivation chain therefore remains self-contained and externally falsifiable.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The analysis rests on the standard variational principle for ground-state approximation and the assumption that RBM expressivity is sufficient to reveal basis-dependent convergence differences; no new entities are postulated.

free parameters (1)
  • RBM hidden-unit density
    Hyperparameter controlling network capacity, chosen to achieve convergence for each basis.
axioms (2)
  • domain assumption Variational Monte Carlo minimization yields a faithful approximation to the ground state when the ansatz is sufficiently expressive.
    Invoked when interpreting NQS energy minimization results as reflecting true ground-state properties.
  • standard math The transverse-field Ising model ground states are exactly solvable or well-characterized in the rotated bases.
    Used as benchmark for measuring NQS performance deviations.

pith-pipeline@v0.9.0 · 5484 in / 1292 out tokens · 44556 ms · 2026-05-16T23:09:51.631855+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Enhancing Neural-Network Variational Monte Carlo through Basis Transformation

    cond-mat.str-el 2026-04 unverdicted novelty 6.0

    A learnable Gaussian basis transformation lowers variational energies in neural-network variational Monte Carlo for the three-dimensional homogeneous electron gas.

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages · cited by 1 Pith paper

  1. [1]

    Carleo and M

    [1]G. Carleo and M. Troyer,Solving the quantum many-body problem with artificial neural networks, Science355(6325), 602 (2017), doi:10.1126/science.aag2302. [2]Y. Nomura, A. S. Darmawan, Y. Yamaji and M. Imada,Restricted boltzmann machine learning for solving strongly correlated quantum systems, Phys. Rev. B96, 205152 (2017), 16 REFERENCES REFERENCES doi:...

    work page doi:10.1126/science.aag2302 2017

  2. [2]

    can neural quantum states learn volume-law ground states?

    10705. [14]Y. Gu, W . Li, H. Lin, B. Zhan, R. Li, Y. Huang, D. He, Y. Wu, T . Xiang, M. Qinet al.,Solving the hubbard model with neural quantum states, arXiv preprint arXiv:2507.02644 (2025). [15]D.-L. Deng, X. Li and S. Das Sarma,Quantum entanglement in neural network states, Physical Review X7(2), 021021 (2017), doi:10.1103/PhysRevX.7.021021. [16]Z. Den...

    work page doi:10.1103/physrevx.7.021021 2025