跳转至

文章背景与核心概要

在近期量子硬件(NISQ)上,估计玻恩概率(Born probability)需要通过重复执行量子电路(即采样/shots)。对于受限于固定总测量预算 \(B\) 的量子学习实验,研究人员面临着一个根本性的权衡:应该使用多少个不同的训练态(\(n\)),以及为每个状态分配多少次采样(\(S\),满足 \(B = nS\))。

本文针对使用固定或独立选择的测量算子 \(M\) 的二元量子分类器深入探讨了这一权衡,其中理想得分由 \(\text{Tr}(M\rho)\) 给出。作者推导出了清晰分离有限样本与有限采样贡献的无分布泛化界,并通过最小化保守的闭式替代边界,导出了最优分配规则。PennyLane 仿真实验验证了该理论界的有效性,为实验前规划和有限采样评估提供了稳健的统计学指导。


Measurement-Budget Allocation in Quantum Learning with Finite-Shot Generalization Guarantees

Authors: Ferhat Ozgur Catak
Submitted on: June 1, 2026
Primary Subject: Artificial Intelligence (cs.AI)
arXiv ID: 2608.24891 | DOI: 10.48550/arXiv.2608.24891

Authors: Ferhat Ozgur Catak
Submitted on: June 1, 2026
Primary Subject: Artificial Intelligence (cs.AI)
arXiv ID: 2608.24891 | DOI: 10.48550/arXiv.2608.24891


Abstract Summary

On near-term quantum hardware, estimating a Born probability requires repeated circuit executions (shots). For a quantum learning experiment constrained by a fixed total measurement budget \(B\), researchers face a fundamental tradeoff: how many distinct training states (\(n\)) should be used versus how many shots (\(S\)) should be allocated to each state (\(B = nS\)).

This paper investigates this tradeoff for binary quantum classifiers using fixed or independently selected measurement operators \(M\), where the ideal score is given by \(\text{Tr}(M\rho)\).

Key Contributions:

  • Generalization Bounds: The author proves a distribution-free generalization bound that clearly separates finite-sample and finite-shot contributions:
  • Sample term: Scales as \(\sqrt{d/n}\)
  • Shot term: Scales as \(\sqrt{(\log n)/S}\)
  • Optimal Allocation Rule: By minimizing a conservative closed-form surrogate of the bound, the optimal allocation rules are derived as: $\(\begin{aligned} n^\star &= 2\sqrt{\frac{2dB}{\log(2B/\delta)}} \\ S^\star &= \frac{B}{n^\star} \end{aligned}\)$ This surrogate matches the asymptotic scaling of the exact minimizer, yielding a worst-case convergence rate of \(B^{-1/4}\).
  • Simulations: PennyLane simulations using 2-qubit and 4-qubit variational quantum circuits across nine synthetic binary classification benchmarks demonstrate that the one-sided empirical generalization gap consistently remains below the theoretical bound.

Note: While this framework offers a conservative statistical guideline for pre-experimental planning and finite-shot evaluation, extending these guarantees to fully adaptive shot-noisy training remains an open problem.

在近期量子硬件上,估计玻恩概率需要重复执行量子电路(采样)。对于受限于固定总测量预算 \(B\) 的量子学习实验,研究人员面临着一个根本性的权衡:应该使用多少个不同的训练态(\(n\)),以及为每个状态分配多少次采样(\(S\))(满足 \(B = nS\))。

本文针对使用固定或独立选择的测量算子 \(M\) 的二元量子分类器研究了这一权衡,其中理想得分由 \(\text{Tr}(M\rho)\) 给出。

核心贡献:

  • 泛化界(Generalization Bounds): 作者证明了一个无分布(distribution-free)的泛化界,清晰地分离了有限样本和有限采样的贡献:
  • 样本项: 尺度为 \(\sqrt{d/n}\)
  • 采样项: 尺度为 \(\sqrt{(\log n)/S}\)
  • 最优分配规则(Optimal Allocation Rule): 通过最小化该边界的一个保守闭式替代(surrogate),推导出的最优分配规则为: $\(\begin{aligned} n^\star &= 2\sqrt{\frac{2dB}{\log(2B/\delta)}} \\ S^\star &= \frac{B}{n^\star} \end{aligned}\)$ 该替代方案与精确极小值的渐近尺度相匹配,产生了 \(B^{-1/4}\) 的最坏情况收敛率。
  • 仿真实验(Simulations): 在九个合成二元分类基准测试中,使用 2 量子比特和 4 量子比特变分量子电路进行 PennyLane 仿真,结果表明单侧经验泛化间隙始终保持在理论界之下。

注:虽然该框架为实验前规划和有限采样评估提供了保守的统计指导,但将这些保证扩展到完全自适应的采样噪声训练仍然是一个开放性问题。


Article Metadata & Links

Article Metadata & Links