How many samples are needed to reliably approximate the best linear estimator for a linear inverse problem?