This thesis studies nonparametric estimation of drift functions in causal Lyapunov models from samples of a stationary distribution. The underlying dynamics are modeled by a stochastic differential equation whose drift is structured by a directed acyclic graph with self-loops, and each drift component is estimated in a reproducing kernel Hilbert space according to its parent set. For the Gaussian kernel and Tikhonov regularization, concentration bounds are derived for the estimator, first when parent drifts are known and then in the recursive setting where they are estimated from the data. The analysis also includes an explicit convergence rate for a one-dimensional drift consisting of a linear term plus an RKHS function. The computational part compares Tikhonov, Spectral Cutoff, and Landweber regularization, implements Nyström approximations to reduce computational cost, and evaluates the method on synthetic examples and the Sachs protein-signaling data.
«
This thesis studies nonparametric estimation of drift functions in causal Lyapunov models from samples of a stationary distribution. The underlying dynamics are modeled by a stochastic differential equation whose drift is structured by a directed acyclic graph with self-loops, and each drift component is estimated in a reproducing kernel Hilbert space according to its parent set. For the Gaussian kernel and Tikhonov regularization, concentration bounds are derived for the estimator, first when p...
»