Consistency of Landweber algorithm in an ill-posed problem with random data (Q857089)

The note deals with the linear ill-posed problem, described by the operator equation \(Ax=u\) of the first kind, in which \(A\) is an injective bounded linear operator in a separable Hilbert space \(H\) and the second member is measured with random errors [cf., e.g., \textit{A.\,M.\thinspace Fedotov}, Некорректные задачи со случајными ошибками в данных (Ill-posed problems with random errors in the data) (Nauka, Novosibirsk) (1990; Zbl 0762.65034)]. Carrying out \(n\) experiments, we obtain a sample \( \{u_1,\dots,u_n \}\) such that \(u_i=u_{ex} + \xi_i\). Here, \(u_{ex}\) represents the unknown exact value of \(u\) and \(\{\xi_i \}_{i \in N}\) is a sequence of independent and identically distributed random variables defined on a probability space \((\Omega, F)\) with values in \(H\). The proposed approach has two main steps. The first step is concerned by the estimation of \(u\). The strong law of large numbers gives the empirical mean \(\overline{u}\). Therefore we have to solve the equation \(Ax=\overline{u}\). In addition, \(AH\) is not closed, so the problem is ill-posed. This leads to the introduction of two concepts: quasi-solution and pseudo-solution [cf.\ \textit{V.\,K.\thinspace Ivanov, V.\,V.\thinspace Vasin} and \textit{V.\,P.\thinspace Tanana}, ``Theory of linear ill-posed problems and its applications'' (2nd ed.; Inverse and Ill--Posed Problems Series, VSP, Utrecht) (2002; Zbl 1037.65056)]. The existence and the uniqueness of the pseudo-solution for such a problem is shown. Then the Landweber algorithm is used to estimate the pseudo-solution. This algorithm is defined by the sequence \[ x_{m+1}=x_m - A^*Ax_m + A^* \overline{u}+ a_m \xi_m. \] Here, \(x_0\) is an arbitrary random variable, \(A^*\)is the adjoint operator of \(A\), and \(\{a_m\}\) is a sequence of real numbers such that \(m a_m\) converges to a constant when \(m \to \infty\). Following the above procedure, exponential inequalities of Bernstein--Fréchet type are established. The ``almost complete convergence'' of this algorithm to the pseudo-solution is shown and its convergence rate is estimated. Finally, a confidence domain for the pseudo-solution is build.