arXiv:2610.31515
Problem-aware sampling for bias and variance reduction in randomised Hamiltonian simulation
Brian Coyle, Jonathan H. Fetherolf, Antonio Márquez Romero, Stefano Scali, Giuseppe Buonaiuto, Josh Kirsopp
Abstract
Randomised quantum simulation algorithms such as qDRIFT reduce circuit depth by replacing deterministic product formulas with stochastic compilations, but achieving high accuracy can require many steps. In vanilla qDRIFT, Hamiltonian terms are sampled with probabilities proportional to their coefficient magnitudes. We propose StateDRIFT, a variant of qDRIFT which optimises its sampling distribution to estimate time-evolved quantities using knowledge of the initial state and target observable, with a provable truncation bound that keeps qDRIFT's guarantees. The distribution at each depth is chosen by solving an optimisation problem targeting bias and/or variance. StateDRIFT can significantly reduce the bias and variance of estimators, and hence circuit depth, which we test on short-time dynamics of lattice models and molecules under Jordan-Wigner up to 18 qubits. For example, with BeH2 StateDRIFT can reduce the number of two-qubit gates required to achieve a given accuracy from 11,201 to 983. StateDRIFT involves measuring observables and commutators on the initial state, so we propose two practical ways to estimate these: classical surrogates such as tensor networks, and classical shadows, which recover most or all of the saving from exact oracles in the systems tested.
AI Summary
StateDRIFT asks whether qDRIFT sampling can be improved for estimating a particular time-evolved expectation value, rather than approximating evolution uniformly over states and observables. It replaces coefficient-proportional sampling with a distribution optimised using the initial state, target observable, and Hamiltonian. Fourth-order bias and trajectory-variance expansions supply the objectives, using initial-state disturbances and nested commutators. The distribution is fixed before each simulation, without additional ancillas. Single-step analysis gives closed-form variance-optimal allocations and, under sign and ordering assumptions, proves that some allocation is asymptotically no worse than vanilla qDRIFT in both bias and variance. Multistep allocations are optimised numerically. Mixing with vanilla sampling enforces $q_j\ge\tau |c_j|/\lambda$, bounding relative reweighting by $1/\tau$. A bias-truncation bound has leading behaviour $O(\|\delta_2\|\lambda^3T^5\|O\|/N)$. Classical simulations cover lattice models and molecular quenches on 6–18 qubits. Molecular examples require 1.7–26 times fewer steps and 1.2–11 times fewer two-qubit gates at the tested resolution thresholds; BeH$_2$ decreases from 11,201 to 983 gates. Approximate tensor-network references and classical shadows are tested as alternatives to exact initial-state inputs, including matchgate-shadow estimation requiring at most the 5-RDM for the full fourth-order objective. Limitations include short-time truncation accuracy, which does not improve merely by increasing depth at fixed $T$, loose rigorous bounds, and a heuristic validity diagnostic. Water fails at $T=1$ and is benchmarked at $T=1/2$. Poor classical surrogates can perform worse than vanilla qDRIFT. The variance analysed is compilation-trajectory variance, not finite-shot observable-measurement noise; demonstrations remain classically tractable.
The authors used large language model tools (Claude, Anthropic; Codex, OpenAI) to assist with software development, code review and derivations, and editing of the manuscript text and its \LaTeX{} structure.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 15 pages main text, 14 figures