A global structure-preserving kernel method for the learning of Poisson systems (Q6961695)

In this paper, a novel kernel ridge regression method is proposed that allows the recovery of globally defined nonlinear Hamiltonian functions on Poisson manifolds out of datasets made of noisy observations of Hamiltonian vector fields. The kernel method exhibits various key advantages. The differential equation constraints necessary for structure preservation are incorporated in the kernel ridge regression framework due to the the differential reproducing property. For kernel regression, ill-posedness is inherent in recovering the Hamiltonian function in the presence of Poisson degeneracy due to the presence of Casimir functions. The Tikhonov regularized kernel regression ensures a unique solution due to strict convexity. An operator-theoretic framework and a kernel estimator representation is presented that enables a rigorous analysis of estimation and approximation errors. The estimator converges to a true Hamiltonian function with high probability, as the number of random samples increases. The structure-preserving kernel ridge regression is implemented in Python and extensive numerical experiments are conducted on Hamiltonian systems on symplectic manifolds: the rigid body underwater vehicle dynamics, and recovery of two-vortex dynamics.