arXiv:2610.13907
A scalable sample-based quantum diagonalization solver for general Pauli-sum Hamiltonians
Yutaro Iiyama
Abstract
A scalable ground-state solver to be used in sample-based quantum diagonalization (SQD) is presented. The solver is built on JAX API, and is intended to be run on GPU platforms and possibly many-GPU high-performance computing facilities. It is capable of subspace diagonalization in up to $2^{30}$-dimensional subspace, with possible upgrades to enable even larger-scale operations. The two main components, a general-purpose eigensolver and an SQD-specific matrix-vector operation function, are described in detail. Efficiency of the eigensolver with respect to the GPU memory is demonstrated using a simple physics model. Furthermore, good strong scaling of the eigensolver in many-GPU environment is observed.
AI Summary
The paper addresses the classical diagonalization bottleneck in sample-based quantum diagonalization (SQD), where a Hamiltonian is projected onto computational-basis states selected through quantum sampling and classical postprocessing. It presents an open-source, Python/JAX solver for general Pauli-sum Hamiltonians, combining a matrix-free eigensolver with an SQD-specific Hamiltonian–vector operation and supporting distributed GPU execution. The eigensolver uses the established locally optimal conjugate-gradient method, retaining as few as five full-size vectors through memory reuse. For $D=10^9$, these vectors require approximately $80$ GB in complex128 precision, excluding additional matrix–vector workspace. The SQD operation represents Pauli strings by binary signatures, finds connections between selected basis states using stable sorting, and groups terms sharing an X signature. Optional caching trades memory against repeated computation; caching source indices requires $4LN$ bytes for $N$ basis states and $L$ distinct X signatures. Demonstrations concern the standalone eigensolver: the two lowest energies of a 30-site transverse-field Ising chain are computed on one 96-GB NVIDIA H100, and a 32-site ground-state calculation is distributed across 8–128 GPUs. Identical energies and good strong scaling are reported, without numerical timing values in the supplied text. The current SQD implementation is limited to $N\leq 2^{30}-1$ by JAX’s stable-sort size restriction; block-wise sorting is proposed to remove this limit. Matrix–vector workspace can increase memory requirements, and the optional sequential approach to excited states is acknowledged to be numerically unstable.
Error found: The reported benchmarks validate only the standalone eigensolver, leaving the complete SQD projection/matvec capacity and distributed performance unbenchmarked. The documented sorting restriction permits at most 2^30−1 SQD basis states, inconsistent with the introduction's 2^31 limit and slightly below the abstract's stated 2^30. For the valid Pauli Hamiltonian A=I, the published eigensolver pseudocode treats a zero residual as a Rayleigh–Ritz basis vector and subsequently divides a zero update vector by its norm. For positive Rayleigh quotients, the published stopping criterion rejects sufficiently small nonzero residuals and fails to recognize an exact positive-energy eigenpair.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 10 pages, 1 figure