Simulated annealing for stochastic semilinear equations on Hilbert spaces (Q1374621)

A semilinear stochastic evolution equation \[ dX_{t} = -AX_{t} dt + b(X_{t}) dt +\varepsilon (t) dW_{t} \tag{1} \] in a real separable Hilbert space \(H\) is considered, where \(A\) is a positive self-adjoint operator, \(b :H\to H\) is a bounded Lipschitz continuous mapping, and \(W\) is a standard cylindrical Wiener process in \(H\). It is supposed that the operator \(A^{-1+\delta}\) is nuclear for a \(\delta \in \left ]0,1\right [\) and that \(b\) has a potential, i.e. \(b=\nabla V\) for a bounded real function \(V\) on \(H\). The author investigates the simulated annealing problem for the equation (1), which consists in describing the long-time behaviour of the law \(\mathcal L(X_{t})\) of the solution \(X_t\) to (1) if \(\varepsilon\) is a function slowly vanishing at infinity. More precisely, she proves the following theorem: Denote by \(\mu ^{V}_{\alpha}\) the invariant measure of the equation (1) with a constant function \(\varepsilon (t)\equiv \alpha \), that is, \( d\mu ^{V}_{\alpha}= C_{\alpha}\exp (2\alpha ^{-2}V) dn_{\alpha}\), \(n_\alpha\) standing for the Gaussian measure on \(H\) with zero mean and covariance operator \(\frac 12\alpha ^{2}A^{-1}\). Choose \(\varepsilon (t) = c(\log (e^{1/c^{2}}+t))^{-1/2}\) with an arbitrary but sufficiently large constant \(c>0\). Then \(\operatorname {Var}\bigl (\mathcal L(X_{t})-\mu ^{V}_{\varepsilon (t)} \bigr)\to 0\) as \(t\to \infty \), whenever the law \(\nu\) of \(X_0\) satisfies \(\int _{H}\exp (\beta |y|) d\nu (y)<\infty\) for a \(\beta >0\). The obtained result is shown to apply e.g. to stochastic evolutions of Ginzburg-Landau type.