Approximate proximal point algorithms for finding zeroes of maximal monotone operators in Hilbert spaces (Q938427): Difference between revisions

Let \(T\) be a maximal monotone operator on a real Hilbert space \(H\), with \(T^{-1}\{0\}\) nonempty. In [SIAM J.~Control Optimization 14, No.\,5, 877--898 (1976; Zbl 0358.90053)], \textit{R.\,T.\thinspace Rockafellar} proved that the proximal point algorithm, \[ x_n+e_{n+1}\in x_{n+1}+{\beta}_nTx_{n+1},\;n=0,1,\dots, \] converges weakly to a zero of \(T\) for every \(x_0\in H\), provided that \({\beta}_n\geq \beta>0\), \(n=0,1,\dots\) and \((| e_n| )\in \ell^1\). The present paper deals with a similar algorithm, given by \[ x_{n+1}={\alpha}_nu + (1-{\alpha}_n)P_{\Omega}((I+{\beta}_nT)^{-1}(x_n+e_n)-e_n),\;n=0,1,\dots, \] where \(\Omega:=D(A)\) is assumed to be closed and convex and \(P_{\Omega}\) denotes the projection on \(\Omega\). The authors prove the strong convergence of this algorithm to a zero of \(T\) (nearest to \(u\)) for every \(u_0\in H\), \(u\in \Omega\), under the following additional conditions: \(\beta_n>0\), \(\beta_n \rightarrow \infty\), \(\alpha_n\in [0,1]\), \(\alpha_n \rightarrow 0\), \(\sum_0^\infty\alpha_n=\infty\), \((| e_n| )\in \ell^2\). Another result of the paper states the weak convergence to a zero of \(T\) of a similar algorithm obtained by replacing \(u\) in the above equation by \(x_n\). These results are then used to approximate minimizers of proper convex lower semicontinuous functions defined on \(H\).