EVENT DOSSIER
The researchers proposed a new method for Unbalanced Dynamic Optimal Transport that does not require simulation. This method introduces a universal growth penalty term, aiming to address the limitations of traditional methods when dealing with data distributions that vary significantly or contain noise. Through this innovation, the model can more effectively handle unbalanced data scenarios without relying on large-scale simulation calculations, thereby enhancing the robustness and efficiency of the algorithm under complex distributions.
arXiv:2609.04710v1 提出 Simulation-free Unbalanced Dynamic Optimal Transport with General Growth Penalty (SUDO) 框架，旨在解决非二次凸增长惩罚下的无平衡动态最优传输求解效率问题。现有 UDOT 求解器依赖昂贵的 NeuralODE 模拟或仅支持二次惩罚的解析解，限制了其通用性。SUDO 通过证明凹性增长惩罚导致退化解，引入无需模拟的条件路径学习与半耦合问题求解机制，并利用不平衡流匹配技术实现高效计算。在 WFR 基准测试中，SUDO 精度与解析算法相当且速度更快；此外，该框架支持非对称惩罚以编码增殖主导先验，在合成及单细胞数据集上生成更合理的轨迹与生长估计。