Parameter choice methods using minimization schemes (Q617659)

The subject of this paper is the order optimal choice of reconstruction methods for solving linear ill-posed problems. This is done in a unified framework that includes deterministic and stochastic models for the noise on the right-hand side and the solution of the considered problem. To give some details, consider a linear operator \( A: \mathcal{X} \to \mathcal{Y} \) acting between Banach spaces \( \mathcal{X} \) and \( \mathcal{Y} \). In addition, let \( y^\delta = Ax + \delta \xi \) be noisy data, where \( x \in \mathcal{X} \) denotes the unknown solution, and the noise \( \xi \in \mathcal{Y} \) may be deterministic or stochastic. In addition, \( \delta > 0 \) represents some noise level which may be unknown, in general. The authors consider a collection of continuous reconstruction methods \( R_n: \mathcal{Y} \to \mathcal{X}\), \(n \in \mathbb{N} \), e.g., \( R_n = R(\alpha_n, \cdot) \), where \( R \) is a given regularization scheme and \( \alpha_n \) denotes a regularization parameter. The considered expected errors for the corresponding approximations \( x_n^\delta = R_n(y^\delta) \) are \( E_{P,Q}(n) := (\mathbb{E}_{P} \mathbb{E}_{Q} \| x - x_n^\delta \|^2)^{1/2}\), \(n \in \mathbb{N} \), where \( P \in \mathcal{P} \) and \( Q \in \mathcal{Q} \). Here, \( \mathcal{P} \) is a given set of Radon probabilities on \( \mathcal{X} \), and \( \mathcal{Q} \) denotes a collection of cylindrical measures on \( \mathcal{Y} \). The deterministic case for the noise is obtained if \( \mathcal{Q} \) is the set of Dirac measures on the unit ball in \( \mathcal{Y} \). Similarly, one gets the deterministic case for the solution if \( \mathcal{P} \) is the set of Dirac measures on a subset of \( \mathcal{X}\). The goal is to choose from the set \( \{x_n^\delta\}_n \) an order optimal approximation by minimizing some positive function \( f = f_{x,\xi}: \mathbb{N} \to \mathbb{R}^+ \), e.\,g., the quasi-optimality function \( f(n) = \| x_n^\delta - x_{n+1}^\delta \|^2, \;n \in \mathbb{N} \). Consider, in the general situation for \( f \), the minimizer \( n_* = n_*(x,\xi) = \text{arg\,min}_n f_{x,\xi}(n) \). For this setting, conditions are given which guarantee the existence of a minimizer \( n_* \) as well as order optimality \( E_{P,Q}(n_*) \leq c \min_{n} E_{P,Q}(n) \) uniformly for \( P \in \mathcal{P}\), \(Q \in \mathcal{Q} \), where \( c > 0 \) is some finite constant. This is done for the general case first and then again for reconstruction methods \( R_n \) that are linear. Finally, for a Bayesian setting it is shown that the results are applicable to a \( \delta \)-free residual principle for the truncated singular value decomposition.