English

Convergence Rates and Decoupling in Linear Stochastic Approximation Algorithms

Statistics Theory 2015-01-13 v1 Statistics Theory

Abstract

Almost sure convergence rates for linear algorithms hk+1=hk+1kχ(bkAkhk)h_{k+1} = h_k +\frac{1}{k^\chi} (b_k-A_kh_k) are studied, where χ(0,1)\chi\in(0,1), {Ak}k=1\{A_{k}\}_{k=1}^\infty are symmetric, positive semidefinite random matrices and {bk}k=1\{b_{k}\}_{k=1}^\infty are random vectors. It is shown that hnA1b=o(nγ)|h_n- A^{-1}b|=o(n^{-\gamma}) a.s. for the γ[0,χ)\gamma\in[0,\chi), positive definite AA and vector bb such that 1nχγk=1n(AkA)0\frac{1}{n^{\chi-\gamma}}\sum\limits_{k=1}^n (A_{k}- A)\to 0 and 1nχγk=1n(bkb)0\frac{1}{n^{\chi-\gamma}}\sum\limits_{k=1}^n (b_k-b)\to 0 a.s. When χγ(12,1)\chi-\gamma\in\left(\frac12,1\right), these assumptions are implied by the Marcinkiewicz strong law of large numbers, which allows the {Ak}\{A_k\} and {bk}\{b_k\} to have heavy-tails, long-range dependence or both. Finally, corroborating experimental outcomes and decreasing-gain design considerations are provided.

Keywords

Cite

@article{arxiv.1501.02414,
  title  = {Convergence Rates and Decoupling in Linear Stochastic Approximation Algorithms},
  author = {Michael A. Kouritzin and Samira Sadeghi},
  journal= {arXiv preprint arXiv:1501.02414},
  year   = {2015}
}

Comments

27 pages

R2 v1 2026-06-22T07:57:27.373Z