No label defined (Q3973447)

Let \((X_1,\ldots,X_n)\) denote a random vector with \(N(a,R)\) distribution. Suppose that the rank of \(R\) is equal to \(m\) and denote \(\mu_1(R)>\ldots>\mu_m(R)\) the eigenvalues of \(R\) and \(h_1(R),\ldots,h_m(R)\) the eigenvectors. \textit{T. W. Anderson} [An introduction to multivariate statistical analysis. New York: John Wiley (1958; Zbl 0083.14601)] introduced the maximum likelihood estimates of the parameters \(\mu_1(\hat R),\ldots,\mu_m(\hat R)\), \(h_1(\hat R),\ldots h_m(\hat R)\) where \[ \hat R=(n-1)^{-1}\sum_{k=1}^n (X_k-\overline{X})(X_k-\overline{X})'\text{ and }\overline{X}=n^{-1}\sum_{k=1}^n X_k. \] For large \(m\) the bias of these estimates can be important. The author introduces a new consistent estimate for \(\mu_1(R)\) using his method of \(G\)-analysis [see the author, Theory Probab. Appl. 32, No. 2, 229---242 (1987); translation from Teor. Veroyat. Primen. 32, No. 2, 252--265 (1987; Zbl 0653.62043)]. This method introduces some nonrandom function \(g\) such that \[ p-\lim_{n\to \infty}(\mu_ 1(\hat R)-g(\mu_ 1(R),R))=0 \] and the estimate of \(\mu_ 1\) is chosen to be the solution of the equation \(g(x,R)=\mu_ 1(\hat R)\). As \(R\) is unknown, the author proposes to replace this equation by an approximation \(\varphi(x,\hat R)=\mu_ 1(\hat R)\). The derivation of the result is a very technical one. It uses a ``\(G_{26}\)-estimate'' the definition of which cannot be summarized in a few lines. For the general definition of \(G\)-estimates, the author refers to his previous publications. Along the derivation, many references are made to the works of the author and often, for some intermediate proofs, references are only given. The only references of the paper are the publications of the author and the Anderson's book. [For the entire collection see Zbl 0725.00013.]