AI 中文总结
该研究提出无需训练的元启发式优化框架，在Azure Quantum Resource Estimator上实现自适应误差预算分配，可将433个基准电路的时空成本平均降低33%以上，为容错量子计算资源估计提供优化层。
AI 中文摘要
系统级资源估计是容错量子计算（FTQC）工具链的关键组成部分，其效率取决于如何将全局容错能力分配到逻辑操作、T态蒸馏和旋转合成中，以最小化物理资源开销。常用的均匀分配策略忽略了电路特定结构，可能会为非活跃或欠关键子系统过度配置资源，导致时空估计值膨胀。先前的研究尝试使用在离线生成的数据集上训练的监督模型解决这一局限，但该方法会产生额外的数据生成成本，并限制部署灵活性。为克服这些缺陷，我们提出一种无需训练的优化框架，该框架直接在Azure Quantum Resource Estimator（AQRE）上执行无导数搜索，可为先前未见过的电路实现实例特定的误差预算分配，而无需离线训练数据。为评估对优化器选择的鲁棒性，我们用两种结构不同的元启发式算法实例化该框架：模拟退火和量子粒子群优化。我们在MQT Bench套件中来自31个家族、量子比特数为2至91的433个电路上评估了该框架。在整个基准套件中，两种方法均将时空成本平均降低了33%以上，且一致性在1.34个百分点以内，表明在不同元启发式搜索策略中，收益是稳定的。我们的分析进一步发现，优化收益主要由误差分布的不对称性而非电路规模驱动，且优化分配的基尼系数这一指标为预期改进提供了可解释的诊断。总之，这些结果表明，自适应误差预算分配是FTQC资源估计流程的系统软件优化层。
英文摘要
System-level resource estimation is a key component of fault-tolerant quantum computing (FTQC) toolchains. Its efficiency depends on how global error tolerance is allocated across logical operations, T-state distillation, and rotation synthesis to minimize physical resource overhead. The commonly used uniform-allocation strategy ignores circuit-specific structure and can overprovision inactive or less critical subsystems, leading to inflated space-time estimates. Prior work aims to address this limitation using supervised models trained on offline-generated datasets. However, this approach incurs additional data-generation costs and limits deployment flexibility. To overcome these drawbacks, we propose a training-free optimization framework that performs derivative-free search directly on the Azure Quantum Resource Estimator (AQRE), enabling instance-specific error budget allocation for previously unseen circuits without requiring offline training data. To evaluate robustness to optimizer choice, we instantiate the framework with two structurally distinct metaheuristics, simulated annealing and quantum particle swarm optimization. We evaluate our framework across 433 circuits spanning 2 to 91 qubits from 31 families in the MQT Bench suite. Across the benchmark suite, both methods reduce space-time cost by more than 33\% on average and agree within 1.34\% points, indicating that the gains are stable across different metaheuristic search strategies. Our analysis further finds that the optimization benefit is driven primarily by error-profile asymmetry rather than circuit scale, and the metric, Gini coefficient of optimized allocation, provides an interpretable diagnostic of expected improvement. Together, these results position adaptive error budget allocation as a system-software optimization layer for FTQC resource estimation pipeline.
