Unimodality of a refinement of Lassalle’s sequence

We study a refinement of Lassalle’s sequence introduced by Defant, Engen, and Miller [5]. Lassalle’s sequence was originally defined by Lassalle in [9] by the recurrence

with the initial condition $A_{1}=1$, and where the $C_{k}=\binom{2k}{k}/(k+1)$ are the Catalan numbers. In [9] Lassalle proved the sequence had positive terms, and the sequence has been further explored in [2, 7, 11, 12]. This sequence also has relations with noncommutative probability: $(-1)^{m-1}A_{m}$ is the $(2m)$-th classical cumulant of the standard semicircular probability distribution. More details about the connection between noncommutative probability and stack sorting can be found in [4].

We are interested in combinatorial interpretations of Lassalle’s sequence, particularly those found by Josuat-Vergès in [7] and by Defant, Engen, and Miller in [5].

Defant, Engen, and Miller’s interpretation of Lassalle’s sequence came chronologically after that of Josuat-Vergès, but we will discuss it first. The interpretation involves the stack-sorting map, which was originally defined by West [13] as a slight modification of an algorithm originally defined by Knuth [8]. Since we will not end up working directly with this map, we will not fully define the map, instead referring readers to [5] for the definition. Essentially, the stack-sorting map “partially sorts” a permutation, in such a way that any permutation can be sorted via enough applications of the stack-sorting map. If $\pi$ is a permutation, let $s(\pi)$ denote its image under stack sorting. We say that a permutation is uniquely sorted if there is a unique permutation which stack-sorts to it. That is to say, $\pi$ is uniquely sorted if $s^{-1}(\pi)$ has size 1. (Uniquely sorted permutations have also been studied further in [3, 10].) Defant, Engen, and Miller proved in [5] that $A_{k+1}$ counts the total number of uniquely sorted permutations of length $2k+1$. Furthermore, they defined the sequences $(A_{k+1}(\ell))_{1\leq\ell\leq 2k+1}$, letting $A_{k+1}(\ell)$ equal the number of uniquely sorted permutations of length $2k+1$ whose first element is $\ell$. Note that the sum of each such sequence is $A_{k+1}$, so these may be regarded as refinements of Lassalle’s sequence. They proved that each such sequence is symmetric (in $\ell$), and conjectured that these sequences are unimodal, and furthermore, log-concave. In this paper we will prove the former.

Using 3.3, which is a recursion-like identity for a generalization of these sequences, we also (with computer assistance) verify the following.

In this section we will describe Josuat-Vergès’s interpretation of the Lassalle sequence, which will be useful to us in order to prove 1.1. First we will need some definitions.

Let $\mathcal{P}(n)$ denote the set of partitions of $\{0,\dots,n-1\}$, and if $n$ is even, also let $\mathcal{M}(n)\subset\mathcal{P}(n)$ denote the set of matchings on $\{0,\dots,n-1\}$, where a matching is just a partition containing only blocks of size 2.

If $\rho\in\mathcal{P}(n)$ and $B,B^{\prime}$ are blocks in $\rho$, then we say that $B$ and $B^{\prime}$ form a crossing if there exist $i,k\in B$ and $j,\ell\in B^{\prime}$ such that either $i<j<k<\ell$ or $i>j>k>\ell$. If we put the elements of $\{0,\dots,n-1\}$ in order on a circle, and represent each block by the polygon whose vertices are its elements, then two blocks form a crossing exactly when their corresponding polygons intersect. Note that in the special case where $\rho$ is a matching, all its blocks are represented by line segments.

Now define the crossing graph $G(\rho)$ of a partition $\rho$ to be the graph whose vertices are the blocks of $\rho$ and with an edge between two blocks if and only if they form a crossing. Define an orientation $r$ of the edges of $G(\rho)$ to be root-connected with root $B$ if it is acyclic and the block $B\in\rho$ is the only source in the orientation. Equivalently, $r$ is root-connected with root $B$ if it is acyclic and there exists a path from $B$ to every other vertex in $G(\rho)$. (Note that $G(\rho)$ must be connected in order for it to have a root-connected orientation.) Greene and Zaslavsky in [6] proved that the number of root-connected orientations of $G(\rho)$ with any fixed root is $T_{G(\rho)}(1,0)$, where $T_{G(\rho)}$ is the Tutte polynomial of $G(\rho)$, defined in [1].

Let $\widetilde{\mathcal{P}}(n)$ denote the set of pairs $(\rho,r)$, where $\rho\in\mathcal{P}(n)$ and $r$ is a root-connected orientation of $G(\rho)$, where the root is the block containing 0. Similarly let $\widetilde{\mathcal{M}}(n)$ denote the subset of elements $(\rho,r)$ of $\widetilde{\mathcal{P}}(n)$ such that $\rho\in\mathcal{M}(n)$. Josuat-Vergès proved that Lassalle’s sequence $A_{k+1}$ counts the number of elements of $\widetilde{\mathcal{M}}(2k+2)$. Furthermore, Defant, Engen, and Miller proved by bijection that the refinement $A_{k+1}(\ell)$ can also be counted by root-connected orientations of matchings:

This is the interpretation of the sequence $A_{k+1}(\ell)$ which we will use to prove unimodality.

We will now prove 1.1. First we will need to define a slight generalization of $\widetilde{\mathcal{M}}(n)$, where we allow the root to be a set of any size.

For a set $S\subset\{0,\dots,n-1\}$, let $\mathcal{M}_{S}(n)$ denote the set of partitions of $\{0,\dots,n-1\}$ in which one of the blocks is $S$ and all other blocks have 2 elements. Let $\widetilde{\mathcal{M}}_{S}(n)$ denote the set of $(\rho,r)$ such that $\rho\in\mathcal{M}_{S}(n)$ and $r$ is a root-connected orientation of $G(\rho)$ with root $S$. Then, let $A_{k+1}(S)$ be the size of the set $\widetilde{\mathcal{M}}_{S}(|S|+2k)$. We then have $A_{k+1}(\ell)=A_{k+1}(\{0,\ell\})$.

Let $n=|S|+2k$. We first note some basic properties of the function $A_{k+1}(S)$. Each such property will be accompanied by a visual depiction of it, where the elements of $\{0,\dots,n-1\}$ are placed (equally spaced and clockwise) on a circle and $S$ is represented by a polygon whose vertices are its elements. First, the property of rotation states that we can rotate the elements of $S$ on the circle without changing $A_{k+1}(S)$.

We will generally use the property of rotation implicitly throughout this proof.

The property of merging states that if $S$ contains two consecutive elements then we can merge these two points on the circle into one point, reducing the size of $S$ by 1 and also reducing $n$ by 1. Note that this doesn’t affect any crossings. To state this, we will let $S=\{a_{1},\dots,a_{m}\}$, where $m\geq 1$ and $a_{1}<\dots<a_{m}$. We allow indices to “wrap around” the circle, so that $a_{m+1}=a_{1}+n$, and so on.

Now we show a sort of recursion for $A_{k+1}(S)$, which will be the main tool that we will use.

We now use 3.3 to prove one final property, which is that we can reflect $a_{j}$ over the midpoint of $a_{j-1},a_{j+1}$. As a consequence, $A_{k+1}(S)$ depends only on the (unordered) multiset of differences $\{a_{2}-a_{1},a_{3}-a_{2},\dots,a_{m+1}-a_{m}\}$ (recall that the last difference here equals $(a_{1}+n)-a_{m}$).

We now show the main theorem. In fact we will show the following which is a generalization of 1.1.

The main remaining open question is Defant, Engen, and Miller’s conjecture of log-concavity:

It turns out that the recursion 3.3 actually allows us to more efficiently compute the sequence elements $A_{k+1}(\ell)$. Using this recursion, we used a computer program to verify that 4.1 is true for all $k\leq 53$. However, the problem remains open for large $k$.

This research was conducted at the University of Minnesota, Duluth REU and was supported by NSF-DMS grant 1949884 and NSA grant H98230-20-1-0009. Thanks to Joe Gallian for running the REU program and providing helpful comments. Thanks also to Colin Defant for suggesting the problem and for useful comments.