Estimation of random fields by piecewise constant estimators (Q1965873): Difference between revisions

The problem of optimal design for estimation of random fields by piecewise constant estimates is studied. The random field is considered on the continuous square \(D= [0,1]^2\) with an unknown mean function \(m(t)\) and with a covariance structure which is satisfied e.g. by Ornstein-Uhlenbeck or by Gaussian random fields. The random field in \(D\) has to be estimated by a collection of observed values at some points, each of them representing the field in a whole subregion. The estimation error is measured by the integrated mean-square error. The optimal procedure is proved to be the following. The region \(D\) is partioned into a number of subregions and each of them is further tessalated into regular diamonds (or hexagons -- in dependence on the covariance structure) with size depending on the properties of the particular subregions (in general different). Then the centers of these diamonds or hexagons are asymptotically optimal sampling points. The asymptotics is understood in the sense of growing number of sampling points.