{"id":"arxiv/2610.17835","url":"https://arxiv.org/abs/2610.17835","tags":["math.NA","math.SG"],"score":0,"title":"Learning Hamiltonians from Trajectory Data with Kernel Methods","domain":"math.NA","source":"arxiv","authors":["Jianyu Hu","David Martin de Diego","Juan-Pablo Ortega","Miguel Vaquero"],"reviews":0,"subject":"math.NA","summary":"We study the recovery of a Hamiltonian from discrete trajectory observations using structure-preserving kernel methods. When exact-flow data are fitted through a symplectic one-step scheme, the fixed-step learning target is generally not the physical Hamiltonian itself, but a scheme-dependent inverse modified Hamiltonian. For the implicit midpoint and $p$-implicit symplectic Euler methods, we establish the existence of analytic inverse targets and show that their discrepancies from the physical Hamiltonian are second and first order in the observation step, respectively. Within an RKHS…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-10-13","reactions":{}}