arXiv Science⌕ Search

arXiv subjects

Diancong Jin

Publications and source records attributed to Diancong Jin.

22 records · Page 2Linked to original sources

Numerically asymptotical preservation of the large deviations principles for invariant measures of Langevin equations

In this paper, we focus on two kinds of large deviations principles (LDPs) of the invariant measures of Langevin equations and their numerical methods, as the noise intensity $ε\to 0$ and the dissipation intensity $ν\to\infty$ respectively. First, by proving the weak LDP and the exponential tightness, we conclude that the invariant measure $\{μ_{ν,ε}\}$ of the exact solution satisfies the LDPs as $ε\to0$ and $ν\to\infty$ respectively. Then, we study whether there exist numerical methods asymptotically preserving these two LDPs of $\{μ_{ν,ε}\}$ in the sense that the rate functions of invariant measures of numerical methods converge pointwise to the rate function of $\{μ_{ν,ε}\}$ as the step-size tends to zero. The answer is positive for the linear Langevin equation. For the small noise case, we show that a large class of numerical methods can asymptotically preserve the LDP of $\{μ_{ν,ε}\}_{ε>0}$ as $ε\to0$. For the strong dissipation case, we study the stochastic $θ$-method ($θ\in[1/2,1]$) and show that only the midpoint scheme ($θ=1/2$) can asymptotically preserve the LDP of $\{μ_{ν,ε}\}_{ν>0}$ as $ν\to\infty$. These results indicate that in the linear case, the LDP as $ε\to0$ and the LDP as $ν\to\infty$ for the invariant measures of numerical methods have intrinsic differences: the common numerical methods can asymptotically preserve the LDP of $\{μ_{ν,ε}\}_{ε>0}$ as $ε\to0$ while the asymptotical preservation of numerical methods for the LDP of $\{μ_{ν,ε}\}_{ν>0}$ as $ν\to\infty$ depends on the choice of numerical methods. To the best of our knowledge, this is the first result of investigating the relationship between the LDPs of invariant measures of stochastic differential equations and those of their numerical methods.

math.NA↗

Large deviations principles for symplectic discretizations of stochastic linear Schrödinger Equation

In this paper, we consider the large deviations principles (LDPs) for the stochastic linear Schrödinger equation and its symplectic discretizations. These numerical discretizations are the spatial semi-discretization based on spectral Galerkin method, and the further full discretizations with symplectic schemes in temporal direction. First, by means of the abstract Gärtner--Ellis theorem, we prove that the observable $B_T=\frac{u(T)}{T}$, $T>0$ of the exact solution $u$ is exponentially tight and satisfies an LDP on $L^2(0, π; \mathbb C)$. Then, we present the LDPs for both $\{B^M_T\}_{T>0}$ of the spatial discretization $\{u^M\}_{M\in\mathbb N}$ and $\{B^M_N\}_{N\in \mathbb N}$ of the full discretization $\{u^M_N\}_{M,N\in\mathbb N}$, where $B^M_T=\frac{u^M(T)}{T}$ and $B^M_N=\frac{u^M_N}{Nτ}$ are the discrete approximations of $B_T$. Further, we show that both the semi-discretization $\{u^M\}_{M\in \mathbb N}$ and the full discretization $\{u^M_N\}_{M,N\in \mathbb N}$ based on temporal symplectic schemes can weakly asymptotically preserve the LDP of $\{B_T\}_{T>0}$. These results show the ability of symplectic discretizations to preserve the LDP of the stochastic linear \xde equation, and first provide an effective approach to approximating the LDP rate function in infinite dimensional space based on the numerical discretizations.

math.NA↗

The probabilistic superiority of stochastic symplectic methods via large deviations principles

It is well known that symplectic methods have been rigorously shown to be superior to non-symplectic ones especially in long-time computation, when applied to deterministic Hamiltonian systems. In this paper, we attempt to study the probabilistic superiority of stochastic symplectic methods by means of the theory of large deviations. We propose the concept of asymptotical preservation of numerical methods for large deviations principles associated with the exact solutions of the general stochastic Hamiltonian systems. Concerning that the linear stochastic oscillator is one of the typical stochastic Hamiltonian systems, we take it as the test equation in this paper to obtain precise results about the rate functions of large deviations principles for both exact and numerical solutions. Based on the Gärtner--Ellis theorem, we first study the large deviations principles of the mean position and the mean velocity for both the exact solution and its numerical approximations. Then, we prove that stochastic symplectic methods asymptotically preserve these two large deviations principles, but non-symplectic ones do not. This indicates that stochastic symplectic methods are able to approximate well the exponential decay speed of the "hitting probability" of the mean position and mean velocity of the stochastic oscillator. To the best of our knowledge, this is the first result about using large deviations principle to show the superiority of stochastic symplectic methods compared with non-symplectic ones in the existing literature.

math.NA↗

Modified averaged vector field methods preserving multiple invariants for conservative stochastic differential equations

A novel class of conservative numerical methods for general conservative Stratonovich stochastic differential equations with multiple invariants is proposed and analyzed. These methods, which are called modified averaged vector field methods, are constructed by modifying the averaged vector field methods to preserve multiple invariants simultaneously. Based on the prior estimate for high order moments of the modification coefficient, the mean square convergence order $1$ of proposed methods is proved in the case of commutative noises. In addition, the effect of quadrature formula on the mean square convergence order and the preservation of invariants for the modified averaged vector field methods is considered. Numerical experiments are performed to verify the theoretical analyses and to show the superiority of the proposed methods in long time simulation.

math.NA↗