A new discrepancy principle (Q2566123)

Consider an operator equation \(Au=f\), where \(A: H \to H\) is a bounded linear operator on a Hilbert space \(H\), and \({\overline{R(A)}} \neq R(A)\). Assume that \(f=Ay\), where \(y\) is a minimal-norm solution to the equation, and that noisy data \(f_{\delta}\) are given such that \(\| f_{\delta}-f \| \leq \delta\). Let \(\| f_{\delta} \|>C\delta\), \(C>1\), \(0<b<C^2-1\). Then there exist \(\varepsilon=\varepsilon(\delta)>0\) and an element \(u_{\delta,\varepsilon}\) with \(\| Au_{\delta,\varepsilon}-f_{\delta}\|=C\delta\) and \(F(u_{\delta,\varepsilon}) \leq m+(C^2-1-b) \delta^2\), \(F(u)=\| Au-f_{\delta} \|^2+\varepsilon \| u\|^2\), \(m=\inf_{u \in H} F(u)\). For any such element \(u_{\delta,\varepsilon}\), \(\lim_{\delta \to 0} \| u_{\delta,\varepsilon(\delta)}-y \|=0\).