Multi-parameter regularization and its numerical realization (Q537875): Difference between revisions

The authors propose and analyze a choice of parameters in the multi-parameter regularization of Tikhonov-type: Let \(A\) be a bounded linear operator between Hilbert spaces \(X\), \(Y\). The equation \(Ax= y\) is often ill-posed. Instead of \(y\) only a noisy data \(y_\delta\in Y\), \(\| y-y_\delta\|\leq\delta\) is known. In the two-parameter Tikhonov regularization a solution \(x^\delta(\alpha, \beta)\) is defined as the minimizer of the functional \[ \Phi(\alpha,\beta, x)=\| Ax- y_\delta\|^2+\alpha\| Bx\|^2+\beta\| x\|^2, \] where \(B\) is assumed to be a densely defined unbounded self-adjoint strictly positive operator on \(X\). The goal is to find an a posteriori strategy for choosing the parameter pair \((\alpha,\beta)\) so that the so-called discrepancy principle is fulfilled. The definitions are generalized to multi-parameter \(\alpha_1,\dots, \alpha_e\), \(\beta\) in Chapter 4. A numerical realization of the multi-parameter discrepancy principle is proposed. For a serie of examples the two-parameter regularization is tested against the single-parameter one. Experiments with 3-parameter regularizations are compared with single-parameter regularizations. First kind Fredholm integral equations are studied. The discrepancy curves are given. The discrepancy principle and the model function approximation are applied to the problem of learning from labeled and unlabeled data.