Minimax identification of a nonlinear dynamic observation system (Q1778988)

The authors deal with the problem of estimation of unknown parameters of a nonlinear regression model with incomplete information on observation error distribution. The dynamics of a system is described by the equation \[ x'(t)=f(x(t),\theta_1),\quad x(0)=\theta_2, \] where \(\theta_1\) and \(\theta_2\) are nonrandom vectors of unknown parameters and \(x(t)\in\mathbb R^p\). Observations \[ y(t_k)=\Phi(x(t_k),\theta_3)+v_k, k=1,\dots,n, \] where \(y(t_k)\in\mathbb R^q\), \(\theta_3\) is a vector of unknown parameters, \(\{v_k\}\) is a centered vector Gaussian stationary white noise with covariance \(V\), are carried out according to a discrete plan \(\varepsilon_n=\{t_1,\dots,t_n\}\). It is assumed that the covariance matrix \(V\) is unknown, but belongs to a set \(\mathcal V\) of admissible covariances. To construct the asymptotic (as \(n\to\infty\)) minimax estimate of the unknown parameters \(\theta_1,\theta_2,\theta_3\) the authors propose an iterative algorithm based on methods of possible directions and prove its convergence. Examples illustrating the theoretical results are presented. Remarks: Harald Cramér is badly translated as Kramer. See page 293.