On the balancing principle for some problems of numerical analysis (Q2373403): Difference between revisions

The authors develop the problem of concordance of convergence and stability of regularization/discretization methods of solving calculating problems in functional spaces when the method depends on some scalar parameter \(\sigma\). It is assumed that a solution of the problem \(u^\dag\) can be approximated by an ideal element \(u_{\sigma}\), i.e., there exists a non-decreasing continuous function \(\varphi(\sigma)\) such that \(\varphi(0)=0\) and \(\| u^\dag-u_{\sigma}\| \leq\varphi(\sigma)\) for any \(\sigma\in(0,1)\). In view of unavoidable errors generated by data errors and/or by discretization, one can construct some element \(u_{\sigma}^{\Delta}\) instead of \(u_{\sigma}\), where \(\Delta\) is a level of unavoidable error. The stability of approximation with respect to \(\Delta\) can be described by the inequality \(\| u^{\sigma}-u_{\sigma\Delta}\| \leq\vartheta\Delta/\sigma^v\), where \(\vartheta\) and \(v\) are positive constants. The optimal parameter \(\sigma\) is defined through balancing of the right-hand parts of the mentioned estimates. The authors propose and analyze an adaptive strategy for the choice of the parameter \(\sigma\), which does not rely on using the function \(\varphi\) and the exact value of \(v\), but on some properties of \(\varphi\) and on estimations of \(v\). The strategy is detailed for severely ill-posed linear operator equations (it is realized for the sideways heat equation) and for the choice of a weight in interior point discontinuous approximation on non-matching grids for elliptic partial differential equations.