Le. Sjoberg, THE BEST QUADRATIC MINIMUM BIAST NONNEGATIVE DEFINITE ESTIMATOR FOR AN ADDITIVE 2 VARIANCE COMPONENT MODEL, Manuscripta geodaetica, 20(2), 1995, pp. 139-143
We derive the Best Quadratic Minimum Bias Non-negative Definite Estima
tors of the variance components sigma(1)(2) and sigma(2)(2) of the var
iance-covariance matrix model Q = sigma(1)(2)I + sigma(2)(2)F for the
observation vector L in the linear model AX = E{L}. I is the unit matr
ix and F is a positive diagonal matrix. The result is [GRAPHICS] and [
GRAPHICS] Here f(min) and f(max) are the minimum and maximum elements
of F, and <(epsilon)over cap>(i) and <(epsilon)over cap>(j) the corres
ponding residuals obtained in the least squares adjustment. In additio
n to the design matrix A we introduce nu = sigma(2)/sigma(1) and A(O)
= A(A(T)Q(O)(-1)A)(-1) A(T) Q(O)(-1), where Q(O) is the estimate for Q
given by a priori presented components sigma(1) and sigma(2). As each
estimator is determined from merely one residual, it turns out that t
heir variances are poor.