Explicit numerical algorithms for stochastic differential equations and their applications in data science and optimization
In this thesis, we focus on the analysis of a new generation of explicit numerical algorithms for (non-linear) random systems and their applications in computational statistics, non-convex optimization, data science, and finance. In the first part, a tamed (explicit) order 1.5 algorithm is proposed to approximate solutions of stochastic differential equations (SDEs) with super-linear coefficients. Under certain conditions on the coefficients, a convergence result in L 2 is obtained. Then, in the second part, a new higher order Langevin Monte Carlo (LMC) algorithm is constructed, which is based on the application of an order 1.5 scheme to the Langevin SDE. The proposed LMC algorithm is considered in the context of sampling from a target distribution π that has a density πˆ on R d known up to a normalizing constant. To obtain the convergence results, − log ˆπ is assumed to have a locally Lipschitz gradient and its third derivative is locally Hölder continuous with exponent α ∈ (0, 1]. Non-asymptotic bounds are established for the convergence to stationarity of the new sampling method with convergence rate 1 + α/2 in Wasserstein distance, while it is shown that the rate is 1 in total variation even in the absence of convexity. In particular, in the case where − log ˆπ is strongly convex and its gradient is Lipschitz continuous, explicit constants are provided. Finally, in the last part of the thesis, the stochastic gradient Langevin dynamics (SGLD) algorithm is considered, which can be viewed as an extension of the unadjusted Langevin algorithm (ULA). We focus on a stochastic optimization problem via the SGLD algorithm. Crucially, the aim is to obtain non-asymptotic error bounds of the SGLD algorithm with discontinuous gradient H(θ, x). Theoretical guarantees in Wasserstein distances are provided under such an assumption for both convex and non-convex objective functions. Moreover, explicit upper estimates are obtained for the expected excess risk associated with the approximation of global minimizers of these objective functions. Numerical results for key examples in statistics and finance are presented which support the main findings.