Comparisons of parameter choice methods for regularization with discrete noisy data (Q4381828)

We consider a linear ill-posed operator equation \[ Kf(x)= g(x), \quad x\in [0,1] \tag{1} \] where \(g(x)\) is known only as discrete, noisy data \(y_i= g(x_i)+ \varepsilon_i\), \(i=1, \dots,n\). Assume that the operator \(K:L^2(0,1)\to L^2 (0,1)\) is known without error and the solution \(f_0\in L^2 (0,1)\) of (1) exists and is unique. The prototype example is the first-kind integral equation \[ Kf(x)= \int^1_0 k(x,t)f(t)dt =g(x). \] To obtain an approximate solution to (1), it is used the method of regularization in the form \[ \min_{f\in W} n^{-1} \sum^n_{i=1} \bigl(Kf(x_i) -y_i\bigr)^2 +\lambda \| f\|^2_W \tag{2} \] 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 \(f_{n \lambda}\) depends critically on the regularization parameter \(\lambda>0\). If \(\lambda\) is too small, then \(f_{n \lambda}\) contains noise of large magnitude and if \(\lambda\) is too large, then the features of the true solution are smoothed out. Therefore, it is important to have a reliable method for choosing a good regularization parameter. The author develops several methods for the crucial selection of the parameter in regularization of linear ill-posed problems with discrete, noisy data. The discrepancy principle (DP), minimum bound (MB) method and generalized cross-validation (GCV) are known to be at least waeakly asymptotically optimal with respect to appropriate loss functions as the number \(n\) of data points approaches infinity. The author compares these methods in three other ways. First, \(n\) is taken to be fixed and, using a discrete Picard condition, upper and lower bounds on the `expected' DP and MB estimates are derived in terms of the optimal parameters with respect to the risk and expected error. Next, a simple measure of the variability of a practical estimate is defined and, for each of the five methods, its asymptotic behaviour is determined. The results are that the asymptotic stability of GCV is the same as for the unbiased risk method and is superior to that of DP, which is better than for MB and an unbiased error method. Finally, the results of numerical simulations of the five methods (on a first-kind integral equation with kernel \(\kappa (x,t)= | x-t |^{1/2}\) and solution \(f_0(t)=t\), using 100 data points) demonstrate that the theoretical conclusions hold in practice.