Work

Law-determined norm inequalities do not imply complete mixability

Year: 2026
Type: preprint
Abstract: A law $F$ with mean $\mu$ is $n$-completely mixable if there are $n$ random variables with law $F$ whose sum is almost surely $n\mu$. Wang proved that such a law satisfies a family of inequalities between the norms of the positive and negative parts of $X-t$ and $X-s$, where $X\sim F$ and $s=(n\mu-t)/(n-1)$, for every threshold $t$ and every law-determined norm, and asked when these inequalities are also sufficient (Open Problem 4 in his 2015 survey). We show that for bounded $F$ the whole family is equivalent to the single convex-order condition $X-\mu\cx-(n-1)(X-\mu)$, and that the same holds for integrable $F$ if the norms are required to have the Fatou property. The inequalities do not imply $n$-complete mixability: for $n=3$ the analytic density $\frac1{10}+\frac{27}{10}(2x-1)^2$ on $[0,1]$ satisfies all of them and is not $3$-completely mixable, and for $n=3$ and $n=4$ there are counterexamples with densities bounded away from zero on $[0,1]$ and with $C^\infty$ densities, the one for $n=4$ not symmetric. A Bernoulli$(p)$ law satisfies the inequalities exactly when $1/n\le p\le1-1/n$, while it is $n$-completely mixable exactly when $np$ is an integer. For $n=2$ the inequalities are equivalent to symmetry, hence to $2$-complete mixability, for every integrable law. For $n=3$ the examples refine the symmetric non-mixable laws of Yin and Zhu, and the reduction explains why those laws satisfy the inequalities.
Language: English
Open Access status: green
OpenAlex ID: W7226019529
Cites:
Cited by:
Related to: 10