Totally non-negativity of a family of change-of-basis matrices

TL;DR

This paper characterizes when a family of change-of-basis matrices, encoding combinatorial sequences, is totally non-negative, providing a practical check and a planar network representation for such matrices.

Contribution

It extends previous results by giving a complete, efficient criterion for total non-negativity of these matrices for arbitrary real sequences.

Findings

Provides an $O(n^2)$ algorithm to determine total non-negativity.

Characterizes total non-negativity via planar network representations.

Identifies explicit negative minors when matrices are not totally non-negative.

Abstract

Let ${\bf a}=(a_1, a_2, \ldots, a_n)$ and ${\bf e}=(e_1, e_2, \ldots, e_n)$ be real sequences. Denote by $M_{{\bf e}\rightarrow {\bf a}}$ the $(n+1)\times(n+1)$ matrix whose $(m,k)$ entry ($m, k \in \{0,\ldots, n\}$) is the coefficient of the polynomial $(x-a_1)\cdots(x-a_k)$ in the expansion of $(x-e_1)\cdots(x-e_m)$ as a linear combination of the polynomials $1, x-a_1, \ldots, (x-a_1)\cdots(x-a_m)$. By appropriate choice of ${\bf a}$ and ${\bf e}$ the matrix $M_{{\bf e}\rightarrow {\bf a}}$ can encode many familiar doubly-indexed combinatorial sequences, such as binomial coefficients, Stirling numbers of both kinds, Lah numbers and central factorial numbers. In all four of these examples, $M_{{\bf e}\rightarrow {\bf a}}$ enjoys the property of total non-negativity -- the determinants of all its square submatrices are non-negative. This leads to a natural question: when, in general, is $M_{{\bf e}\rightarrow {\bf a}}$ totally non-negative? Galvin and Pacurar found a simple condition on ${\bf e}$ that characterizes total non-negativity of $M_{{\bf e}\rightarrow {\bf a}}$ when ${\bf a}$ is non-decreasing. Here we fully extend this result. For arbitrary real sequences ${\bf a}$ and ${\bf e}$, we give a condition that can be checked in $O(n^2)$ time that determines whether $M_{{\bf e}\rightarrow {\bf a}}$ is totally non-negative. When $M_{{\bf e}\rightarrow {\bf a}}$ is totally non-negative, we witness this with a planar network whose weights are non-negative and whose path matrix is $M_{{\bf e}\rightarrow {\bf a}}$. When it is not, we witness this with an explicit negative minor.