Structure-Preserving Kernel Methods for Dynamics: from Poisson Manifolds to Wasserstein Flows
主讲人: Juan-Pablo Ortega(Nanyang Technological University)
活动时间: 从 2026-11-23 16:00 到 17:00
场地: 北京国际数学研究中心,镜春园78号院(怀新园)77201室
Kernel methods can be made to respect the geometry of the system that is being learned. This talk presents a unified reproducing kernel Hilbert space framework for recovering the functions that generate a dynamical system out of observational data. The key ingredient is a differential reproducing property, which yields a differential representer theorem and closed-form estimators that are structure-preserving by construction. We discuss Hamiltonian systems on symplectic vector spaces, together with their convergence rates and their link with Gaussian process regression; the global formulation on Poisson manifolds, where regularization handles Casimir degeneracy and invariant kernels yield conservation of momentum maps; the recovery of energy functionals of Wasserstein gradient and Hamiltonian flows from density observations; and the learning of solution operators of linear PDEs. Joint work with Jianyu Hu and Daiying Yin.
Bio-Sketch:
Prof. Juan-Pablo Ortega is currently Lee Soo Ying Professor of Mathematics, Head of Mathematical Sciences, and Associate Chair (Faculty) of School of Physical and Mathematical Sciences at Nanyang Technological University. Prof. Ortega studied theoretical physics at the Universidad de Zaragoza and earned his master's degree and PhD in mathematics at the University of California, Santa Cruz, followed by his habilitation at the Université de Nice. After postdoctoral research at EPFL, he became a CNRS researcher and held academic appointments in France and at the University of St Gallen before NTU. He is a foreign elected member of the Royal Academy of Sciences of Spain and received the Ferran Sunyer i Balaguer Prize to the best mathematical monograph. He was a Fulbright Fellow and Rotary Ambassadorial Scholar.
