Methods for Identifying Minimal Sufficient Statistics

Let $f_θ$ be a joint density of a random sample $X$. A frequently used criterion asserts that a statistic $T(X)$ is minimal sufficient if, for any sample points $x$ and $y$, $T(x)=T(y)$ exactly when there exists a finite constant $h_{xy}>0$, independent of $θ$, such that $f_θ(y)=f_θ(x)h_{xy}$ for every $θ$. Although valid under regularity assumptions identified by Lehmann (1950), this statement is false as stated in general: we give a counterexample based on the choice of versions of Radon--Nikodym derivatives. We then propose a version-robust criterion that can be applied once sufficiency is established, and we use it to extend the method of Sato (1996) beyond Euclidean settings to analytic Borel sample spaces and standard Borel statistic spaces. While Lehmann--Scheffé--type regularity conditions and Sato's approach can be difficult to verify in practice, our method is straightforward to check when the statistic is known to be sufficient. Finally, we analyze a distinct criterion for identifying minimal sufficient statistics developed by (Pfanzagl,1994, Pfanzagl, 2017) and provide a discrete counterexample showing that it also requires additional assumptions.