We consider the problem of invariant estimation of a discrete distribu
tion function F under the Cramer-von Mises loss. It is proved that the
best invariant estimator is admissible. This extends a result of Brow
n (1988) and settles an open question (Brown (1988)). The idea used in
the proof of admissibility is a new refinement of the standard Bayes
argument, which is different from the step-wise Bayes approach and Bly
th's (1951) Lemma.