Adaptive algorithms for first principal eigenvector computation (Q557636): Difference between revisions

The paper is really devoted to applications of simple iterative algorithms for finding approximations \(w_k\) of the first principal eigenvector of an unknown matrix \(A\) when only its approximations \(A_k \) are given. These \(A_k\) are generated by a sequence of random observation vectors \(x_k\in {\mathbb R}^n\); for example, \(A_k\equiv x_kx_k^T\), \(A_k\equiv \beta A_{k-1}+k^{-1}(x_kx_k^T-\beta A_{k-1}).\) Several gradient methods of the type \(w_{k+1}=w_k+\tau_kr(w_k,A_k)\) are considered with different objective functions; some of them deal with the Rayleigh quotient \(\lambda_k\) and have \(r(w_k,A_k)\) of the form \(r(w_k,A_k)\equiv [x_kx_k^T]^m(A_kw_k-\lambda_k(w_k)w_k) \) where \(m\in \{0,-1,1\}\). As an intermediate step in analyzing stochastic convergence, the ordinary differential equation \(dw/{dt}=\lim_{k}E[r(w,A_k)]\) is used. Convergence of known methods and of two new modifications are discussed. Numerical examples deal mainly with \(n=10\), \(k=250\), \(k=500.\)