A generalization of the maximal entropy method for solving ill-posed problems (Q5932602): Difference between revisions

Let \(Z(T)\) be a space of functions \(z(x)\), \(x\in\mathbb R^N\), defined on a closed bounded domain. The space \(Z(T)\) is furnished with some sequential convergence topology \(\tau\). Let \(A\) be an operator from \(Z(T)\) into a normed space \(U\). A nonempty set \(D\subset Z(T)\) is fixed and the following operator equation \[ Az = u,\quad z = z(x),\;u\in U, \] is considered on the set \(D\). It is assumed that the set \(Z^* = \arg\inf\{ \|Az - u\|_U: z\in D\}\) of quasisolutions to this equation is nonempty. For finding special quasisolutions to the equation, some functional \(\Omega(z)\) on \(D\) is used. This functional is selected by the so-called \(\Omega\)-optimal quasisolution to the equation, i.e., functions \(\overline z(x)\in Z^*\) for which \[ \Omega(\overline z) = \inf\{\Omega(z): z\in Z^*\}. \] Denote the set of such functions by \(\overline{Z}\). The above-mentioned equation represents an ill-posed problem in which some approximations \(\{A_h, u_{\delta}\}\) with accuracy \(\eta = \{h, \delta\}\) are known instead of the exact data \(\{A, u\}\). In this case, given \(\{A_h, u_{\delta}, h, \delta\}\), it is required to construct a stable approximate solution to the problem, i.e., a function \(z_{\eta}\in D\) such that \(z_{\eta}(x)\overset\tau\rightarrow\overline{Z}\) as \(\eta\to 0\). These are the specific properties of the regularizing functional \(\Omega(z)\) on \(D\) that are well known to guarantee solvability of the problem and possibility of finding stable approximate solutions \(z_{\eta}(x)\) to the inverse problem. In this case the problem under consideration with \(\Omega(z) = \Omega_E(z)\), \(\Omega_E(z) = \int_Tz(x)\ln z(x) dx\) is equivalent to maximizing the entropy functional \(E(z) = -\Omega_E(z)\) on the set of all nonnegative quasisolutions to the equation. The properties of the functional \(\Omega_E\) were studied by \textit{U.~Amato} and \textit{W.~Hughes} in [Inver. Prob. 7, No. 6, 793-808 (1991; Zbl 0752.65091)] and by \textit{P.~P. B. Eggermont} [SIAM J. Math. Anal. 24, No. 6, 1557-1576 (1993; Zbl 0791.65099)] in connection with solving ill-posed linear problems by the residual method and the Tikhonov regularization method. The main result of these articles is the theorem on strong convergence of the approximate solutions \(z_{\eta}(x)\) obtained by these methods to a unique exact solution to the equation with maximal entropy in the space \(L_1(T)\) as \(\eta\to 0\). It turns out that such a result can be obtained for functionals different from \(\Omega_E\) in the problem. Convergence of approximate solutions in \(L_1(T)\) to \(\Omega\)-optimal quasisolutions in this space can be also obtained for a rather wide class of integral functionals \(\Omega(z)\) and for the general case of nonlinear operator equations which are studied by the author. In the article under review, the author introduces and studies such functionals \(\Omega(z)\). The corresponding problem (see the above-mentioned problem) represents a generalization of the maximal entropy method.