arXiv:2610.29779
Mamba Quantum Eigensolver (MQE): A Selective State Space Model for Scalable Ground State Simulations in Quantum Chemistry
G. X. A. Petronilo, M. R. Araújo, Erico S. Teixeira, Clebson Cruz
Abstract
Predicting molecular electronic structure using Quantum Computing approaches is a challenge in Computational Chemistry. Traditional Variational Quantum Eigensolvers (VQEs) are typically hindered by optimization challenges. The alternative Generative Quantum Eigensolver (GQE) mitigates this by reformulating ground-state search as discrete sequence generation, but its Transformer backbone suffers from quadratic computational complexity, bottlenecking scalability for larger molecules. To address this, we introduce the Mamba Quantum Eigensolver (Mamba-QE), which replaces the Transformer with a Selective State Space Model (SSM) to process operator sequences with linear $\mathcal{O}(L)$ complexity. Additionally, we integrate a distance-conditioned mechanism enabling the direct prediction of complete potential energy surfaces (PES). Evaluated on the $H_2$ molecule, Mamba-QE reproduces Full Configuration Interaction (FCI) reference energies well within chemical accuracy, achieving an energy error below $3\times10^{-6}$ Ha across 16 geometries with zero variational violations. Mamba-QE maintains high expressive power while establishing a robust, scalable foundation for quantum simulations of complex molecular systems.
AI Summary
The paper asks whether replacing the Transformer backbone of a Generative Quantum Eigensolver (GQE) with a Mamba selective state-space model can reduce classical sequence-processing costs while retaining accurate molecular ground-state circuit generation. The proposed Mamba-QE processes operator sequences with nominal $O(L)$ rather than attention’s $O(L^2)$ complexity and conditions token embeddings on internuclear distance so that one model can serve multiple geometries. The implementation uses six Mamba blocks with hidden dimension 128. Training begins with randomly sampled circuits whose intermediate energies are evaluated on a quantum simulator. A supervised mean-squared-error objective fits cumulative selected-token logits to these energies. Iterative refinement then adds generated circuits to a replay buffer using an energy threshold annealed from 50 to 5 mHa relative to the best energy observed at each geometry. This is a self-training procedure, not a policy-gradient reinforcement-learning update. Tests use $\mathrm{H}_2$ in STO-3G with CAS(2,2), four-operator sequences, 10 training geometries, and 16 evaluation geometries. Reported energies achieve chemical accuracy at all 16 evaluation points. The minimum is $-1.136186$ Ha versus the FCI value $-1.136189$ Ha, approximately $3\times10^{-6}$ Ha apart; no variational violations are reported. The paper states that this small benchmark cannot expose runtime-scaling advantages. Larger-system comparisons, hardware circuit evaluations, conditioning on multidimensional geometries, and transfer learning remain future work; the demonstrated results do not quantify an end-to-end speedup over Transformer-based GQE.
Central claim may not hold: The abstract claims errors below 3×10⁻⁶ Ha at all 16 geometries, but the supplied numerical results support that precision only near the minimum and report only chemical accuracy across the full sampled surface. The known linear sequence-processing cost of Mamba and an H₂ benchmark with four-operator sequences do not establish practical scalability of the complete eigensolver. The Transformer–Mamba comparison changes multiple training and circuit-search choices simultaneously, so it cannot attribute preserved accuracy specifically to the backbone replacement. The Methods do not reconcile their raw-logit, log-probability, and hidden-state descriptions of energy prediction or clearly distinguish the reported PES energies from independently evaluated circuit expectations.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.