Uniformly Minimum Variance Unbiased Estimators

Among all the unbiased ways to estimate a parameter — estimators whose average value across repeated samples equals the true parameter exactly — some are noisier than others. A uniformly minimum variance unbiased estimator (UMVUE) is the single estimator that is unbiased and has variance at least as small as every other unbiased estimator, and — crucially — this has to hold "uniformly," meaning for every possible true value of the parameter, not just for one convenient case.

That uniformity requirement is what makes UMVUEs rare and interesting rather than automatic: an estimator might beat all rivals when the true parameter is small but lose to a different estimator when it's large, in which case neither one qualifies as the UMVUE — no single winner exists across the board. When a UMVUE does exist, the standard route to finding it is the Rao-Blackwell theorem: start with any unbiased estimator, condition it on a sufficient statistic to strictly reduce (or at least not increase) its variance, and if the sufficient statistic is also complete, the result is guaranteed to be the unique UMVUE.

For example, the sample mean is the UMVUE for the population mean of a normal distribution, and the sample variance (with the usual $n-1$ denominator) is the UMVUE for the population variance — both because they arise from conditioning on a complete sufficient statistic, not merely because they seem like the natural choice.