arXiv++ Combinatorics

We study the polytope of anti-diagonal X-rays of k x k doubly stochastic matrices, the image of the Birkhoff polytope under the map that sums the entries along each anti-diagonal. Its facets other than the coordinate facets are the minimal joint tail bounds for lattice laws of a sum of two dependent uniform random variables. We give an exact facet criterion in terms of the allowed edges of a tight graph. For rays of the underlying tropical fan, total support of the tight graph suffices for a facet but is not necessary, as an example with k=19 shows. Up to affine functions and positive scaling, the potentials that define a facet form a polytope whose vertices are its rays. Facets propagate by padding and by grid refinement under explicit hypotheses. We give infinite families of facets from no step potential, facets of quadratic height, superpolynomially many facets, and a classification of three-step potentials. Total support of a step potential is characterized by a Hall-type condition on block starts; with computer assistance, total support decides facets of step potentials for k<=26. The facet counts up to k=12 are found directly for k<=3 and, for k>=4, come from an enumeration by a provably complete algorithm; an independent exact recomputation reproduces its ray lists for 5<=k<=12.

Which laws can the sum of two uniform random variables have when nothing is known about how they depend on each other? For three or more uniform summands, convex order gives the answer; for two it does not, and the question has been open. We show that the attainable laws are exactly those that satisfy a family of explicit joint bounds on sums of tail probabilities, one bound for each finite set of thresholds. For laws with a density at most one, convex order alone decides. For densities at most two, convex order implies the simplest of the joint bounds, and the constant two cannot be raised for single-threshold bounds. The number of thresholds a law needs is set by the spacing of its atoms, not by how many atoms it has.

Let $A=\{0=a_0<a_1<\dots<a_{k-1}=d\}$ be integers with greatest common divisor $1$, let $C_A$ be the projective monomial curve they define, and let $\mathrm{reg}(A)$ be the Castelnuovo-Mumford regularity of its coordinate ring. Gimenez and González-Sánchez proved $\mathrm{reg}(A)\leσ(A)+1$ for the sumsets regularity $σ(A)$; the reverse inequality $σ(A)\le\mathrm{reg}(A)$ was proposed as an open conjecture in a 2024 talk. We show that it fails and measure by how much. Let $c_1,c_2$ be the conductors of the two numerical semigroups attached to $A$ and $N=\lceil(c_1+c_2)/d\rceil$. For curves in $\mathbb{P}^3$ that are not arithmetically Cohen-Macaulay, $\mathrm{reg}(A)\ge(c_1+c_2)/d+2$, with equality exactly for the arithmetically Buchsbaum curves, and $σ(A)=\mathrm{reg}(A)$. For arithmetically Cohen-Macaulay curves in $\mathbb{P}^3$ we prove $9N\le10\,\mathrm{reg}(A)-2$, with $9/10$ best possible, and give an explicit family with $σ(A)-\mathrm{reg}(A)$ unbounded. In $\mathbb{P}^4$ a family with a band-shaped Apéry set has $\mathrm{reg}(A)/σ(A)\to(2+\sqrt2)/4$ along the solutions of a Pell equation, and gluing carries this bound to every $\mathbb{P}^r$ with $r\ge5$.