Asymptotic behaviour of the minimum bound method for choosing the regularization parameter (Q4381827)

We consider a linear ill-posed operator equation \[ Kf(x)= g(x),\quad x\in [0,1]\tag{1} \] with discrete noisy data \(y_i= g(x_i)+ \varepsilon_i\), \(i= 1,\dots, n\). The errors \(\varepsilon_i\) are modelled as uncorrelated random variables, each with mean \(0\) and variance \(\sigma^2\). It is assumed that \(K: L^2(0, 1)\to L^2(0, 1)\) and the solution \(f_0\in L^2(0, 1)\) of (1) exists and is unique. Let \(f_{n\lambda}\) be the regularized solution defined as the solution of the regularization problem \[ \text{min}_{f\in W} n^{-1} \sum^n_{i= 1} [Kf(x_i)- y_i]^2+ \lambda\| f\|^2_W,\tag{2} \] where \(W\) is a Hilbert space of smooth functions, for example the Sobolev space \(W^{m,2}[0, 1]\). It is well known that the accuracy of the regularized solution depends greatly on the choice of the regularization parameter \(\lambda> 0\). The author considers a parameter choice method (called the minimum bound method) for regularization of linear ill-posed problems that was developed by \textit{T. Raus} [Uch. Zap. Tartu Gos. Univ. 672, 16-26 (1984; Zbl 0578.65050); ibid. 715, 12-20 (1985; Zbl 0594.65037)] and \textit{H. Gfrerer} [Math. Comput. 49, 507-522 (1987; Zbl 0631.65056)] for the case with continuous, deterministic data. The method is adapted and analyzed in a discrete, stochastic framework. It is shown that asymptotically, as the number of data points approaches infinity, the method (with a constant set to 2) behaves like an unbiased error method, which selects the parameter by minimizing a certain unbiased estimate of the expected squared error in the regularized solution. The method is also shown to be weakly asymptotically optimal, in that the `expected' estimate achieves the optimal rate of convergence with respect to the expected squared error criterion and it has the optimal rate of decay.