The averaged Robbins-Monro method for linear problems in a Banach space (Q2433968): Difference between revisions

Let \(B\) be a separable real Banach space, \(A: B\to B\) be a bounded linear operator with spectrum in \(]0,\infty[\), \(V\) is a vector in \(B\). Consider the problem \(Ax= V\) with unique solution, but assume that \(A\) and \(V\) are observable with some random error only. The authors use the recursion \[ X(n +1)= X(n)- a(n)(A(n)X(n)- V(n)),\quad n\in\mathbb{N}, \] where \(A(n)\) and \(V(n)\) are noisy observations of \(A\) and \(V\), respectively, and where \((a(n))\) is a gain sequence. They conclude that ``under fairly weak conditions on the noise process given by \(((A(n)- A, V(n)- V))\), almost sure and distributional invariance principles are shown'', by using the averaged sequence \((1/n)\Sigma X(i)\), summing for \(i\) between 1 and \(n\). For the exact conditions, we have to refer to the paper that has the following structure: Section 1 (Introduction) and Section 2 (Statement of the result) contain together the description of the basis problem with references to the literature, the statement of the conditions that will be used, the statement of the theorem, and two examples; Section 3 (Lemmas) contains an extended description of the notations that are used, together with the statements of 8 lemmas; in Section 4 (Proofs), the proofs of those lemmas are given, together with the concluding part of the proof of the theorem.