发表机构
Universitat de València; BBVA Corporate & Investment Banking; Universidad Autónoma de Madrid; BBVA Quantum（瓦伦西亚大学; 西班牙对外银行企业与投资银行部; 马德里自治大学; 西班牙对外银行量子计算部门）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
研究针对信用估值调整问题，开发端到端噪声感知量子工作流程，引入对比感知贝叶斯迭代量子幅度估计，经硬件校准实验验证其有效性，能更有效利用设备放大能力，降低经典后处理运行时间，并分解CVA误差。
AI 中文摘要
信用估值调整（CVA）需要重复进行风险中性期望估计，这使其成为量子幅度估计的天然测试平台，其相干放大原则上可降低蒙特卡罗采样成本。但在现实金融编码和有噪声的硬件条件下该优势是否存在尚不明朗。我们为CVA开发了一个端到端、噪声感知的量子工作流程，涵盖市场校准、离散化、预言机构建、硬件执行和误差预算分析。该模型结合了相关双资产风险敞口与贴现和违约因素，通过基于量子电路玻尔兹曼机的联合时间 - 市场分布和受控收益旋转进行编码。我们引入了对比感知贝叶斯迭代量子幅度估计（CABIQAE），将实验校准的格罗弗对比度损失纳入贝叶斯推理和电路深度选择。硬件校准实验表明，CABIQAE比无噪声感知的替代方法更有效地利用了当前设备上有限的放大能力，并且比噪声感知的BAE基线实现了低得多的经典后处理运行时间。分析还将总CVA误差分解为统计、编码、离散化和硬件贡献。完整的CVA预言机仍受电路深度和离散化分辨率的限制。
英文摘要
Credit Valuation Adjustment (CVA) requires repeated risk-neutral expectation estimation, making it a natural test bed for quantum amplitude estimation, whose coherent amplification can in principle reduce Monte Carlo sampling cost. Whether this advantage survives realistic financial encoding and noisy hardware remains open. We develop an end-to-end, noise-aware quantum workflow for CVA, covering market calibration, discretisation, oracle construction, hardware execution and error-budget analysis. The model combines a correlated two-asset exposure with discount and default factors, encoded through a QCBM-based joint time-market distribution and controlled payoff rotations. We introduce contrast-aware Bayesian iterative quantum amplitude estimation (CABIQAE), which incorporates experimentally calibrated Grover-contrast loss into Bayesian inference and circuit-depth selection. Hardware-calibrated experiments show that CABIQAE exploits the limited amplification available on current devices more effectively than noise-agnostic alternatives and achieves a much lower classical post-processing runtime than the noise-aware BAE baseline. The analysis further decomposes the total CVA error into statistical, encoding, discretisation and hardware contributions. The full CVA oracle remains limited by circuit depth and discretisation resolution.
Comments56 pages, 14 figures, 18 tables. Includes supplementary appendices with implementation details, extended results, calibration data, and quantum-resource analysis
