中文

多期鞅最优传输:经典理论、神经加速与金融应用

计算金融 2026-04-21 v2 数理金融 证券定价

摘要

本文开发了多期鞅最优传输(MMOT)的计算框架,探讨了收敛速率、算法效率与金融校准。我们的贡献包括:(1)理论分析:通过 Donsker 原理确立了 O(Δtlog(1/Δt))O(\sqrt{\Delta t} \log(1/\Delta t)) 的离散收敛速率,以及 (1κ)2/3(1-\kappa)^{2/3} 的线性算法收敛性;(2)算法改进:引入了增量更新(O(M2)O(M^2) 复杂度)与自适应稀疏网格;(3)数值实现:提出了一种混合神经投影求解器,结合了基于 Transformer 的热启动与 Newton-Raphson 投影。一旦训练完成,纯神经求解器实现了 1,597×1{,}597\times 的在线推理加速(4.74.7s 2.9\to 2.9ms),适用于实时应用,同时混合求解器确保鞅约束精度达到 10610^{-6}。在 12,000 个合成实例(GBM, Merton, Heston)和 120 个真实市场场景上进行了验证。

关键词

引用

@article{arxiv.2601.05290,
  title  = {Multi-Period Martingale Optimal Transport: Classical Theory, Neural Acceleration, and Financial Applications},
  author = {Sri Sairam Gautam B},
  journal= {arXiv preprint arXiv:2601.05290},
  year   = {2026}
}

备注

This preprint is being withdrawn by the authors. We identified errors in the reference list, including incorrect attribution of works to authors -- references and were cited inaccurately with wrong author arrangements and publication details. We are withdrawing the manuscript to correct these errors before any further dissemination. We apologize for the oversight