可复用连续酉变换的编译一次操作图
Compile-once operation graphs for reusable continuous unitary transformations
- Pacific Northwest National Laboratory（太平洋西北国家实验室）
机构由 AI 辅助整理，请以论文原文为准。
AI总结：
本文提出将连续酉变换中的算子代数编译为可复用操作图，通过预分配存储和直接写入，在超3.5亿贡献的压力测试中内存仅增7%，实现CPU/GPU高效复用。
AI中文摘要：
连续酉变换在哈密顿量及其可观测量系数演化过程中，会重复评估相同的算子代数。我们将这些代数关系一次性编译为数值操作图，该图可保存并复用于新的系数集、额外算子、导数以及CPU和GPU上的正向与反向计算。通过将同一编译图应用于简化分子模型的氢和氘参数化，且无需重新生成算子方程，展示了其复用性。主要计算挑战在于构建阶段：在数值传播开始前，生成的多体代数可能包含数亿个贡献项。我们通过预先确定所需存储空间，并将贡献直接写入最终的磁盘支持表示来解决此问题。在一个刻意设计的大型诊断启用压力测试中，包含超过3.5亿个生成贡献，峰值主机内存仅比最终存储大小高出约7%。交叉检查表明，CPU和GPU实现、正向与反向操作以及重复加载编译图，均能以数值精度复现所测试的操作。这种编译一次表示将昂贵的算子代数生成与复用它进行数值变换分离开来。
英文摘要：
Continuous unitary transformations repeatedly evaluate the same operator algebra as the coefficients of a Hamiltonian and its observables evolve. We compile these algebraic relationships once into a numerical operation graph that can be saved and reused for new coefficient sets, additional operators, derivatives, and forward and reverse calculations on CPUs and GPUs. Reuse is demonstrated by applying the same compiled graph to hydrogen and deuterium parameterizations of a reduced molecular model without regenerating the operator equations. The main computational challenge is construction: generated many-body algebra can contain hundreds of millions of contributions before numerical propagation begins. We address this by determining the required storage in advance and writing contributions directly into the final disk-backed representation. In a deliberately large diagnostic-enabled stress test containing more than 350 million generated contributions, peak host memory remains only about 7\% above the final stored size. Cross-checks show that CPU and GPU implementations, forward and reverse operations, and repeated loading of the compiled graph reproduce the tested operations to numerical precision. This compile-once representation separates expensive operator-algebra generation from the numerical transformations that reuse it.
