A new approach to optimal designs for correlated observations (Q2403430)

Consider a linear regression model with correlated observations \(Y(t_i)= \theta^T f(t_i)+ \varepsilon (t_i)\), \(i=1,\dots,n\), where \(Y(t)\) is an observation, \(\theta=(\theta_1,\dots,\theta_m)^T\) is a vector of unknown parameters, \(f(t)=(f_1(t),\dots,f_m(t))^T\) is a vector of known functions, \(\varepsilon(t)\) is a Gaussian process with zero expected mean values and known correlation function. It is well-known that the corresponding optimal design, which includes the choice of \(t_1,\dots,t_n\), is reduced to a nonconvex discrete optimization problem. Explicit solutions to this problem are not available for nearly all the cases of practical interest. In this article, a similar continuous-time model \(Y(t)= \theta^T f(t)+ \varepsilon (t)\), \(t \in [a,b]\), is proposed for consideration. For this model, an explicit expression for the best linear unbiased estimator \(\hat{\theta}_{\text{BLUE}}\) is obtained by a stochastic integral. The corresponding optimal design for finite sample size \(t_1,\dots,t_n\) is done by minimizing the mean squared error between this stochastic integral and its discrete approximation. This approach seems to give a solution that is very close to the solution to the original problem and that is very efficient to implement.