Title:SVM/SVR Kernels as Quantum Propagators

Abstract:In this work, we establish the equivalence between Support Vector Machine (SVM) kernels and quantum Green's functions. Drawing on the analogy between margin maximization in SVMs and action extremization in Lagrangian mechanics, we show that many standard kernels correspond naturally to Green's functions and that this correspondence arises from the inversion of physical operators. We further demonstrate how positive semi-definiteness, which is essential for valid SVM kernels, aligns with the spectral properties that ensure well-defined Green's functions.
We employ the Kernel Polynomial Method (KPM) to create custom kernels for cases where the commonly employed kernels don't lead to a convergence. These custom kernels approximate the desired Green's functions. We furthermore demonstrate numerically on examples taken from physical problems, such as electrical conductivity, scattering amplitudes, photonic crystals, and energy levels of anharmonic oscillators, that selecting kernel functions that mirror the mathematical form of the associated Green's function can significantly enhance the predictive accuracy of machine learning models.