A sieve estimator for the covariance of a Gaussian process (Q1116251)

Let \(\{X_ t,t\in T\}\) be a Gaussian process with mean zero and unknown covariance R. R is parameterized as \[ R(s,t)=R_ 0(s,t)+\sum_{k}a_ kg_ k(s)g_ k(t) \] where \(R_ 0\) is a fixed covariance, \(\{g_ k(t)\}\) is an orthogonal sequence in the reproducing kernel Hilbert space generated by \(R_ 0\), \(\{a_ k\}\) is an \(\ell^ 2\)-sequence with inf \(a_ k>-1\). The sieve estimator for R, based on independent observations (realizations of the process) of size n, is \[ \hat R(s,t)=R_ 0(s,t)+\sum^{m}_{k=1}\hat a_ kg_ k(s)g_ k(t), \] where \(\hat a_ k\), \(k=1,....,m\), are the maximum likelihood estimators under the restriction that \(a_ k=0\) for \(k>m\). Th mean and covariance of \(\hat R\) are calculated, and asymptotic unbiasedness follows immediately. It is shown that pointwise weak and mean-square consistency is valid when \(m=O(n)\), and strong consistency in \(l^ 2\) is valid when \(m=o(n)\).