Minimax estimators of parameters of linear model uncertainty in Hilbert space with stochastic regression coefficients (Q4516146)

The problem of minimax parameter estimation for a Hilbert space linear regression model \(y=Xb+\varepsilon\) is considered. Here \(y\) and \(\varepsilon\) are random elements in a Hilbert space \(H_1\), \(\beta\) is an unknown parameter from a Hilbert space \(H_2\), \(X\) is a random linear operator from \(H_2\) to \(H_1\). The author considers linear estimators of the form \(\hat b=Uy+c\), where \(U\) is a nonrandom linear operator from \(H_1\) to \(H_2\) and \(c\) is a constant in \(L_2\). The minimax estimator minimizes the criterion \[ \sup_{b\in G} E (b-\hat b)^TV(b-\hat b), \] \(V\) being some fixed symmetric nonnegative operator, and \(G\) being a unit ball in \(H_2\). A system of equations is obtained for minimax \(U\) (\(c=0\)) in terms of eigenvectors of the covariance operator of \(\varepsilon\). An analogous problem of minimax estimation of \(x\) is considered for the observations \[ y(\omega,t)=C(\omega,t)x(\omega,t)+\xi_2(\omega,t),\quad 0\leq t\leq T, \] where \(x\) is an (unobserved) solution of the equation \[ d x(\omega,t)/dt=A(\omega,t)x(\omega,t)+\xi_1(t) \] (\(\omega\) denotes dependence on chance).