Exploring the Effect of Basis Rotation on NQS Performance
Sven Benjamin Kožić, V. Zlatić, F. Franchini, S. M. Giampaolo · 2025
Read Exploring the Effect of Basis Rotation on NQS Performance by Sven Benjamin Kožić, V. Zlatić, F. Franchini, S. M. Giampaolo (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-17893.txt.
Introduction
Sven Benjamin Kožić and three coauthors probe a puzzle in quantum computing: why neural network guesses for a quantum system's wave pattern work well in one frame and poorly in another. They use a simple chain model whose exact answer is known. Rotating the frame moves the target answer but leaves the training terrain unchanged.
That shift alone can push training toward bad resting spots, so a run can report low energy while its wave pattern is wrong. The failure can happen even when the moved target still fits the network's reach. The lesson is that frame choice shapes training in ways builders must plan for.
Catalog record
| CREATOR | Sven Benjamin Kožić, V. Zlatić, F. Franchini, S. M. Giampaolo |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 30 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2512-17893
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Exploring the Effect of Basis Rotation on NQS Performance Sven Benjamin Kožić, V. Zlatić, F. Franchini, S. M. Giampaolo Abstract Neural Quantum States (NQS) use neural networks to represent wavefunctions of quantum many-body systems, but their performance depends on the choice of basis, yet the underlying mechanism remains poorly understood. We use a fully solvable one-dimensional Ising model to show that local basis rotations leave the loss landscape unchanged while relocating the exact wavefunction in parameter space, effectively increasing its geometric distance from typical initializations. By sweeping a rotation angle ϕ \phi , we compute quantum Fisher information and Fubini–Study distances to quantify how the rotated wavefunction moves within the loss landscape. Shallow architectures (with focus on Restricted Boltzmann Machines (RBMs)) trained with quantum natural gradient are more likely to fall into saddle-point regions depending on the rotation angle: they achieve low energy error but fail to reproduce correct coefficient distributions. In the ferromagnetic case, near-degenerate eigenstates create high-curvature barriers that trap optimization at intermediate fidelities. We introduce a framework based on an analytically solvable rotated Ising model to investigate how relocating the target wavefunction within a fixed loss landscape exposes information-geometric barriers—such as saddle points and high-curvature regions—that hinder shallow NQS optimization, underscoring the need for landscape-aware model design in variational training. I Introduction Understanding the ground-state properties of quantum many-body systems remains a central challenge in statistical physics and quantum information theory. This difficulty originates from the exponential growth of the Hilbert space with system size, which renders exact representations of the state infeasible except for the smallest systems. While a limited number of models admit analytic solutions 4 ; 15 , most physically relevant systems do not. Consequently, a wide range of approximate approaches has been developed, including Quantum Monte Carlo methods 6 , tensor-network techniques 22 , and density functional theory 17 . In addition to these well-established approaches, neural-network-based variational approaches are emerging as a powerful framework for representing quantum many-body states in recent years. Introduced by Carleo and Troyer 9 , Neural Quantum States (NQS) encode the amplitudes of a quantum state through a parametrized neural network—most commonly a Restricted Boltzmann Machine (RBM)—whose parameters are optimized to approximate the ground state of a given Hamiltonian. This approach combines high representational flexibility with compatibility with stochastic optimization and variational Monte Carlo techniques. Unlike tensor networks, neural networks can, in principle, represent states with volume-law entanglement 13 ; 24 ; 14 , and their training naturally benefits from parallelization and advances in machine-learning software and hardware 30 . Despite these advantages, the performance of NQS is known to depend strongly on the choice of basis used to represent the quantum state. While basis dependence is an intrinsic feature of neural-network representations of a state 31 , the mechanisms through which a change of basis impacts variational optimization remain poorly understood. In particular, it is unclear whether degraded performance under basis rotations should be attributed primarily to limitations in representational expressivity or instead to changes in the geometry of the optimization landscape encountered during training. Several recent works have also noted and investigated the issues related to basis dependency of NQS 25 ; 31 ; 11 . In this work, we investigate this issue within a controlled and analytically transparent framework, designed to disentangle physical effects from optimization artifacts. We employ shallow neural quantum state architectures—namely Restricted Boltzmann Machines and small feedforward networks—trained via stochastic reconfiguration (quantum natural gradient) 27 . Our analysis focuses on the one-dimensional transverse-field Ising model with periodic boundary conditions, considering both ferromagnetic ( J=-1 ) and antiferromagnetic ( J=+1 ) couplings. A site-resolved basis rotation, parameterized by an angle \phi , is applied to the exact ground state, thereby modifying its representation in Hilbert space without affecting the Hamiltonian spectrum, entanglement properties, or underlying physical content. Therefore, within this framework, basis rotations act solely to reposition the exact ground state inside an otherwise unchanged loss landscape. As a consequence, the resulting optimization difficulties originate only from geometric effects. It depends on where the target state resides relative to typical initial variational states, rather than on any modification of the Hamiltonian or the variational objective. By sweeping the rotation angle \phi , we explicitly follow the motion of the target state in parameter space, characterizing it through information-geometric measures such as the quantum Fisher information and the Fubini–Study distance, which quantify its displacement from the initial variational manifold. Our results show that shallow NQS architectures trained with quantum natural gradient are particularly sensitive to this relocation. Depending on the rotation angle, optimization trajectories can be diverted toward saddle-point regions or high-curvature barriers of the loss manifold. In these cases, the network often achieves a small relative energy error while failing to reproduce the correct structure of the ground-state. To diagnose this mismatch, we supplement energy-based metrics with fidelity and coefficient Shannon entropy (quantum coherence) 18 . The distinction between ferromagnetic and antiferromagnetic regimes further highlights the role of spectral structure in shaping optimization dynamics. In the ferromagnetic case, near-degenerate low-energy eigenstates generate high-curvature regions in parameter space that can trap optimization at intermediate fidelities, whereas in the antiferromagnetic case, the closing of the excitation gap leads to different failure modes. Importantly, these effects arise even though the applied basis rotations do not alter entanglement measures derived from reduced density matrices, underscoring that entanglement alone is insufficient to characterize variational difficulty. Beyond quantum many-body physics, our analysis connects naturally to broader questions in modern machine learning concerning optimization landscapes and information geometry 7 ; 26 . In contrast to typical supervised learning tasks that emphasize generalization, the NQS ground-state search is a high-precision regression problem on a fixed target functional, where convergence is governed by landscape curvature rather than overfitting behavior 10 . By leveraging analytically solvable models, we obtain direct insight into how representational constraints and information-geometric structure jointly influence optimization, complementing recent studies of variational quantum algorithms and neural-network ansätze 16 ; 12 . The paper is organized as follows. In Sec. II we introduce the basis-rotation framework and discuss relevant physical properties of the Ising ground state. In Sec. III we summarize the NQS architectures, loss functions, and optimization methods employed. In Sec. IV we analyze information-geometric considerations and potential pitfalls affecting optimization. Section V presents our numerical results, and Sec. VI discusses the broader implications of our findings for landscape-aware design of variational quantum algorithms. II Basis Rotation To isolate how basis choice affects NQS optimization without altering the underlying physics, we study the transverse-field Ising model on a ring: H\;=\;J\sum_{i=1}^{N}\sigma^{z}_{i}\,\sigma^{z}_{i+1}\;+\;h\sum_{i=1}^{N}\sigma^{x}_{i}, (1) where \sigma^{x} and \sigma^{z} are the usual Pauli matrices and periodic boundary conditions identify site N+1\equiv 1 . We then apply a local rotation around the y –axis at each site. The single-spin rotation operator is U_{y}(\phi)\;=\;e^{\,i\,\phi\,\sigma^{y}}\;=\;\cos\phi\,I\;+\;i\,\sin\phi\,\sigma^{y}, and the full N -spin rotation is U_{y}^{(N)}(\phi)\;=\;\bigl[\,U_{y}(\phi)\bigr]^{\otimes N}. Because U_{y} commutes with \sigma^{y} and rotates \sigma^{x}\leftrightarrow\sigma^{z} , it leaves the spectrum of H unchanged while transforming each eigenvector. In particular, if \lvert\psi_{0}\rangle is the ground state in the computational ( \sigma^{z} ) basis, then \lvert\psi_{\phi}\rangle\;=\;U_{y}^{(N)}(\phi)\,\lvert\psi_{0}\rangle is the corresponding ground state in the rotated basis. III Methods In the following section we shortly summarize the standard NQS optimization procedure. We denote