A simple numerical method based simultaneous stochastic perturbation for estimation of high dimensional matrices

Abstract

We describe a simple algorithm for estimating the elements of a matrix as well as its decomposition under the condition that only the product of this matrix with a vector is accessible. The algorithm is based on application of the stochastic simultaneous perturbation method. Theoretical results on the convergence of the proposed algorithm are proven. Numerical experiments are presented to show the efficiency of the proposed algorithm.

Appendix A

The following definitions are given in Chicone (2006).

Definition A1

Equilibria for the system (2) (s.t $w_k = 0$) is the point X such that $\phi (X) = X$.

Let ||X|| denote the Euclidean norm.

Definition A2

Let X be a critical point of the system (2).

• (i) The critical point X is stable if, for any $\epsilon > 0$, there is a $\delta > 0$, such that if a solution $x = \psi _k$ satisfies $||\psi _0 - X|| < \delta$, then $||\psi _k - X|| < \epsilon , \forall k$.

• (ii) The critical point X is unstable if it is not stable as defined above.

• (iii) The critical point X is asymptotically stable if there exists a $\delta > 0$ such that if a solution $X = \psi _k$ satisfies $||\psi _0 - x|| < \delta$, then $\text{ lim }_{k \rightarrow \infty } \psi _k = X$.

Lemma A

(Theorem 1.25 in Chicone (2006)). Assume $\phi (x)$ is a continuously differentiable function, and X an equilibrium solution of $x_{k+1} = \phi (x_k)$ (see (2)). If all eigenvalues of $\varPhi = \partial {\phi (X)}/{\partial {x}}$ are stable and there exist $c > 0$ such that $||\mu _k (X)|| \le c ||X||$ (for definition of $\mu _k$ see below) then the system $x_{k+1} = \phi (x_k)$ is asymptotically stable.

Proof of Theorem 4.1

By the conditions, $\phi (x)$ is continuous differentiable function.

Introduce $e_{k} := \hat{x}_k - x_k$. Subtracting Eq. (2) from Eq. (23) yields

(31)

Using the Taylor expansion $\delta \varphi = \varPhi e_k + O(||e_k||^2)$, $\delta h = H{\varPhi } e_k + O(||e_k||^2)$, the Eq. (31) can be represented as

(32)

It is seen that $e_k = 0$ is an equilibrium point for F(x).

Consider the nonlinear system (32). If we linearize the system (32) around $e_k = 0$ we have $L_k$ as its transition matrix.

From Lemma A it implies that the filter is a.s if the linear system $e_{k+1}^l = L_k e_k^l$ is a.s.

Using the results in Hoang et al. (2001) for the linear filtering problem, it is established that under the conditions of the Theorem, all the eigenvalues of $L_k$ are stable hencet the present, he works as a research engineer on applied mathematics at the DOPS/HOM/REC of the Service Hydrographique et Océanographique de la Marine, France. His research interests lie in numerical methods for hyperbolic partial differential equations, oceanographic modelling, and data assimilation.

The linear system $e_{k+1}^l = L_k e_k^l$ is a.s. It implies in its turn, by Lemma A, that the nonlinear filter (23) (24) (25) is a.s. $\square$