A family of operator splitting methods revisited (Q965043)

Let \(H\) be a real Hilbert spacee; and \(A,B\) be (multivalued) maximal monotone operators from \(H\) to itself. The variable scaling DPRV algorithm Step 0: \(x^0\in H\), \(\mu_0\in (0,\infty)\), \(k:=0\); Step 1: \(a^k\in A(x^k), (y^k,b^k)\in B\): \(y^k+\mu_kb^k=x^k-\mu_ka^k\), \(x^k=y^k\) \(\Rightarrow\) stop; Step 2: \(\gamma_k> 0\), \((x^{k+1},a^{k+1})\in A\): \(x^{k+1}+\mu_ka^{k+1}=x^k+\mu_ka^k-\gamma_k(x^k-y^k)\); Step 3: \(\tau_k\in [0,1)\), \(\mu_{k+1}\in [(1-\tau_k)\mu_k,(1+\tau_k)\mu_k]\), \(k:=k+1\), go to (Step 1) for solving the inclusion \(0\in A(x)+B(x)\) is extended to the case when the parameters may vary in a non-monotonical way from iteration to iteration. Sufficient conditions are given so as to get a linear rate of convergence.