Necessary and sufficient conditions for the Robbins-Monro method (Q792727)
From MaRDI portal
scientific article
Language | Label | Description | Also known as |
---|---|---|---|
English | Necessary and sufficient conditions for the Robbins-Monro method |
scientific article |
Statements
Necessary and sufficient conditions for the Robbins-Monro method (English)
0 references
1984
0 references
Consider the Robbins-Monro method \(X_{n+1}=X_ n-a_ nf(X_ n)+a_ n\xi_ n\), \(a_ n\cong c/n\), \(c>0\), \(n=0,1,...\). This paper shows that the law of large numbers \(\lim_{n\to \infty}a_ n\sum^{n- 1}_{j=1}\xi_ j=0\) is necessary and sufficient for the convergence of \(X_ n\) to \(\theta\) of the unknown f, almost surely or in \(L_ p\), \(p\geq 2\), and the weak law of large numbers \(\lim_{n\to \infty}a_ n\sum^{n-1}_{j=0}\xi_ j=0\) is neither necessary nor sufficient for the convergence of \(X_ n\) to \(\theta\) in probability. The proof is carried out by embedding the R-M method in a continuous- parameter RM differential equation \(X(t)=- a(t)f(X(t))+a(t)\xi(t)+a(t)\zeta(t)\), where a(t)\(\zeta\) (t) is the error due to embedding and tends to zero in some probabilistic sense as \(t\to \infty\). Counter-examples for non-convergence in probability are presented.
0 references
differential inequalities
0 references
embedded differential equation
0 references
Lyapunov function
0 references
Robbins-Monro method
0 references
law of large numbers
0 references
weak law of large numbers
0 references
Counter-examples for non-convergence in probability
0 references