This paper was converted on www.awesomepapers.org from LaTeX by an anonymous user.
Want to know more? Visit the Converter page.

Counting Parabolic Double Cosets in Symmetric Groups

Thomas Browning
(August 2021)
Abstract

Billey, Konvalinka, Petersen, Solfstra, and Tenner recently presented a method for counting parabolic double cosets in Coxeter groups, and used it to compute pnp_{n}, the number of parabolic double cosets in SnS_{n}, for n13n\leq 13. In this paper, we derive a new formula for pnp_{n} and an efficient polynomial time algorithm for evaluating this formula. We use these results to compute pnp_{n} for n5000n\leq 5000 and to prove an asymptotic formula for pnp_{n} that was conjectured by Billey et al.

1 Introduction

For a Coxeter system (W,S)(W,S), each subset ISI\subseteq S generates a subgroup WIW_{I} of WW. These subgroups of WW are called parabolic subgroups. A parabolic double coset in the Coxeter system (W,S)(W,S) is a double coset of the form WIwWJW_{I}wW_{J} for wWw\in W and I,JSI,J\subseteq S. Parabolic double cosets have several properties that make them interesting objects to study. For example, the parabolic double cosets of a finite Coxeter group (W,S)(W,S) are rank-symmetric intervals in the Bruhat order on WW [3], and the poset of presentations of parabolic double cosets gives a boolean complex that retains many properties of the Coxeter complex [6].

The problem of counting parabolic double cosets is considered in [1]. Counting parabolic double cosets is hard because a given parabolic double coset CC might arise from multiple choices of II and JJ. To avoid this problem, Billey, Konvalinka, Petersen, Slofstra, and Tenner take the approach of counting lex-minimal presentations [1]. The lex-minimal presentation of a parabolic double coset CC is the unique presentation C=WIwWJC=W_{I}wW_{J} such that ww is the minimal element of CC in the Bruhat order and such that (|I|,|J|)(\left\lvert{I}\right\rvert,\left\lvert{J}\right\rvert) is lexicographically minimal among all presentations of CC. Theorem 4.26 of [1] gives a formula for the number of parabolic double cosets with a given minimal element wWw\in W, for any Coxeter group WW. In the case of the symmetric group, W=SnW=S_{n}, summing this formula over all n!n! elements of SnS_{n} will compute pnp_{n}, the total number of distinct parabolic double cosets in SnS_{n}. The sequence {pn}\{p_{n}\} may be found at [5, A260700]. The values of pnp_{n} for n13n\leq 13 have been computed using this method [1]. However, the summation over all n!n! elements of SnS_{n} limits the number of terms of {pn}\{p_{n}\} that can be computed with this approach.

In this paper we restrict our attention to counting parabolic double cosets in SnS_{n}. Two ways of depicting parabolic double cosets in SnS_{n} are the balls-in-boxes model considered in [1] and two-way contingency tables, which are matrices of nonnegative integers with nonzero row sums and column sums. Diaconis and Gangolli use the balls-in-boxes model to give a bijection between parabolic double cosets in the double quotient WI\Sn/WJ={WIwWJ:wSn}W_{I}\backslash S_{n}/W_{J}=\{W_{I}wW_{J}:w\in S_{n}\} and two-way contingency tables with prescribed row sums and column sums [2]. They attribute the idea behind the bijection to Nantel Bergeron. This bijection is also used in [6], where it provides an alternate construction of the two-sided Coxeter complex of SnS_{n} in terms of two-way contingency tables.

There are three main results of this paper. The first result is the explicit formula

pn=1n!m=0n[nm]c=0m/2t=2cm(mt)(k=ctc(1)k(c+k)!{tc+k}(k1kc))(j=0cfmt+c,jfmt+c,cjj!(cj)!),p_{n}=\frac{1}{n!}\sum_{m=0}^{n}\genfrac{[}{]}{0.0pt}{}{n}{m}\sum_{c=0}^{\lfloor m/2\rfloor}\sum_{t=2c}^{m}\binom{m}{t}\left(\sum_{k=c}^{t-c}(-1)^{k}(c+k)!\genfrac{\{}{\}}{0.0pt}{}{t}{c+k}\binom{k-1}{k-c}\right)\left(\sum_{j=0}^{c}\frac{f_{m-t+c,j}f_{m-t+c,c-j}}{j!\,(c-j)!}\right), (1)

where the values fn,k=l=knl!{nklk}f_{n,k}=\sum_{l=k}^{n}l!\genfrac{\{}{\}}{0.0pt}{}{n-k}{l-k} are generalizations of the Fubini numbers. The second result is an efficient polynomial time algorithm to evaluate this formula, which is used to compute the values of pnp_{n} for n5000n\leq 5000. These values of pnp_{n} are summarized in the appendix. The third result is an asymptotic formula for pnp_{n} that proves the following conjecture.

Conjecture 1.1 (Conjecture 1.4 in [1]).

There exists a constant KK so that

pnn!K(log2)2n.\frac{p_{n}}{n!}\sim\frac{K}{(\log 2)^{2n}}.
Theorem 1.2.

Conjecture 1.1 holds with constant

K=e(log2)2/24(log2)20.409223.K=\frac{e^{-(\log 2)^{2}/2}}{4(\log 2)^{2}}\approx 0.409223.

We conclude this introduction by outlining the two key ideas behind formula (1). The first idea is to associate each parabolic double coset with a canonical two-way contingency table. In particular, there is a bijection between parabolic double cosets in SnS_{n} and two-way contingency tables with sum nn that are “maximal” (as defined at the end of section 2). This reduces the problem of computing pnp_{n} to the problem of counting maximal two-way contingency tables with sum nn.

Two-way contingency tables have both a restriction on the rows (nonzero row sums) and a restriction on the columns (nonzero column sums). These two restrictions are entangled since any nonzero entry in the table both makes its row sum nonzero and its column sum nonzero. The second idea is to break this entanglement by transforming the problem from counting two-way contingency tables to counting pairs of weak orders (a weak order is a binary relation that is transitive and has no incomparable pairs of elements).

This transformation is applied in sections 3.2 and 3.3 to the problem of counting maximal two-way contingency tables with sum nn, as well as to the simpler problem of counting all two-way contingency tables with sum nn. More generally, this transformation is applicable whenever counting two-way contingency tables with a restriction on the locations of nonzero entries, but not on their specific values.

2 Background

In this section, we give an overview of parabolic double cosets in SnS_{n}, and describe two ways of depicting presentations of parabolic double cosets in SnS_{n}. We will use sis_{i} to denote the iith adjacent transposition in SnS_{n} (the transposition that swaps ii and i+1i+1).

Definition 2.1.

A parabolic subgroup WIW_{I} in the symmetric group SnS_{n} is a subgroup WI=I=s:sIW_{I}=\langle I\rangle=\langle s:s\in I\rangle of SnS_{n} generated by a collection I{s1,,sn1}I\subseteq\{s_{1},\ldots,s_{n-1}\} of adjacent transpositions in SnS_{n}.

Definition 2.2.

A parabolic double coset CC in the symmetric group SnS_{n} is a double coset of the form C=WIwWJC=W_{I}wW_{J} for parabolic subgroups WIW_{I} and WJW_{J} in SnS_{n} and an element wSnw\in S_{n}. The triple (I,w,J)(I,w,J) is called a presentation of CC. The total number of distinct parabolic double cosets in SnS_{n} is denoted by pnp_{n}.
The sequence {pn}\{p_{n}\} may be found at [5, A260700].

A parabolic double coset CC will usually have many presentations. For example, if (I,w,J)(I,w,J) is a presentation of CC, then so is (I,w,J)(I,w^{\prime},J) for any wCw^{\prime}\in C. As the next example shows, there could also be multiple choices for the collections II and JJ. In other words, CC might be an element of more than one double quotient

WI\Sn/WJ={WIwWJ:wSn}.W_{I}\backslash S_{n}/W_{J}=\{W_{I}wW_{J}:w\in S_{n}\}.
Example 2.3.

The parabolic double coset C=S2C=S_{2} in S2S_{2} has six different presentations:

({},12,{s1})(\{\},12,\{s_{1}\}) ({s1},12,{s1})(\{s_{1}\},12,\{s_{1}\}) ({s1},12,{})(\{s_{1}\},12,\{\})
({},21,{s1})(\{\},21,\{s_{1}\}) ({s1},21,{s1})(\{s_{1}\},21,\{s_{1}\}) ({s1},21,{})(\{s_{1}\},21,\{\})

For each element wSnw\in S_{n}, the ball-in-boxes picture of ww is given by placing nn balls in an n×nn\times n grid of boxes in the layout of the permutation matrix of ww. A ball is placed in column aa and row bb if and only if w(a)=bw(a)=b. We label the columns left-to-right and the rows bottom-to-top, as in the Cartesian coordinate system. Left-multiplication acts by permuting the rows of the grid and right-multiplication acts by permuting the columns of the grid.

Consider a presentation (I,w,J)(I,w,J) of a parabolic double coset CC in SnS_{n}. The group WIW_{I} acts on the balls-in-boxes picture of ww by permuting certain rows of the grid. The group WJW_{J} acts on the balls-in-boxes picture of ww by permuting certain columns of the grid. For each adjacent transposition siIs_{i}\notin I, we draw a horizontal wall between row ii and row i+1i+1. For each adjacent transposition sjJs_{j}\notin J, we draw a vertical wall between column jj and column j+1j+1. Then WIW_{I} acts on the balls-in-boxes picture of ww by permuting rows of the grid within the horizontal walls, and WJW_{J} acts on the balls-in-boxes picture of ww by permuting columns of the grid within the vertical walls. The resulting picture is called the balls-in-boxes picture with walls for the presentation (I,w,J)(I,w,J) [1].

Example 2.4 (Example 3.5 in [1]).

Let w=3512467S7w=3512467\in S_{7}, let I={s1,s3,s4}I=\{s_{1},s_{3},s_{4}\}, let J={s2,s3,s5,s6}J=\{s_{2},s_{3},s_{5},s_{6}\}, and let C=WIwWJC=W_{I}wW_{J}. Figure 1 depicts the balls-in-boxes picture with walls for the presentation (I,w,J)(I,w,J) of CC.

s1s_{1}s3s_{3}s4s_{4}s2s_{2}s3s_{3}s5s_{5}s6s_{6}
Figure 1: A balls-in-boxes picture with walls.

The walls naturally divide the n×nn\times n grid into larger rectangular cells. Notice that the number of balls in each cell remains unchanged under the actions of WIW_{I} and WJW_{J}. Thus the number of balls in a given cell is constant over all elements of CC. By counting the number of balls in each cell, we obtain a matrix of nonnegative integers such that the sum of all of the entries is nn and such that each row sum and column sum is strictly positive.

Definition 2.5.

A two-way contingency table is a matrix of nonnegative integers such that each row sum and column sum is strictly positive. Let pn,0p_{n,0} denote the number of two-way contingency tables with sum nn.
The sequence {pn,0}\{p_{n,0}\} may be found at [5, A120733].

Thus, each presentation (I,w,J)(I,w,J) of CC gives rise to a two-way contingency table with sum nn. Conversely, given a two-way contingency table with sum nn, it is possible to recover the collections II and JJ, as well as the parabolic double coset CWI\Sn/WJC\in W_{I}\backslash S_{n}/W_{J} [2, Section 3B]. As a consequence, two-way contingency tables with sum nn are in bijection with triples (C,I,J)(C,I,J) consisting of two collections I,J{s1,,sn1}I,J\subseteq\{s_{1},\ldots,s_{n-1}\} and a parabolic double coset CWI\Sn/WJC\in W_{I}\backslash S_{n}/W_{J} [6]. Another way to put this is that pn,0=IJ|WI\Sn/WJ|p_{n,0}=\sum_{I}\sum_{J}\left\lvert{W_{I}\backslash S_{n}/W_{J}}\right\rvert, because the right hand side counts triples (C,I,J)(C,I,J).

Example 2.6.

The two-way contingency table associated to the triple (C,I,J)(C,I,J) from Example 2.4 is the matrix

(001001111020).\begin{pmatrix}0&0&1\\ 0&0&1\\ 1&1&1\\ 0&2&0\end{pmatrix}.

Even though a given parabolic double coset might have multiple distinct two-way contingency tables arising from multiple choices for the collections II and JJ, the following proposition shows that it is always possible to choose II and JJ to be simultaneously maximal, resulting in a canonical two-way contingency table associated to each parabolic double coset.

Proposition 2.7.

Let CC be a parabolic double coset in SnS_{n}. Let I={si:siC=C}I=\{s_{i}:s_{i}C=C\}, let J={si:Csi=C}J=\{s_{i}:Cs_{i}=C\}, and let wCw\in C be arbitrary. Then (I,w,J)(I,w,J) is a presentation of CC and is maximal in the sense that if (I,w,J)(I^{\prime},w^{\prime},J^{\prime}) is any other presentation of CC, then III^{\prime}\subseteq I and JJJ^{\prime}\subseteq J.

Proof.

Consider the stabilizer subgroups HL={σSn:σC=C}H_{L}=\{\sigma\in S_{n}\colon\sigma C=C\} and HR={σSn:Cσ=C}H_{R}=\{\sigma\in S_{n}\colon C\sigma=C\}. Then I=HL{s1,,sn1}I=H_{L}\cap\{s_{1},\ldots,s_{n-1}\} and J=HR{s1,,sn1}J=H_{R}\cap\{s_{1},\ldots,s_{n-1}\}. In particular, IHLI\subseteq H_{L} and JHRJ\subseteq H_{R} so WI=IHLW_{I}=\langle I\rangle\subseteq H_{L} and WJ=JHRW_{J}=\langle J\rangle\subseteq H_{R}. Since wCw\in C, this shows that WIwWJCW_{I}wW_{J}\subseteq C. We will now show that CWIwWJC\subseteq W_{I}wW_{J}.

Let (I,w,J)(I^{\prime},w^{\prime},J^{\prime}) be any presentation of CC. Then C=WIwWJC=W_{I^{\prime}}w^{\prime}W_{J^{\prime}}. If σI\sigma\in I^{\prime}, then

σC=(σWI)(wWJ)=(WI)(wWJ)=C\sigma C=(\sigma W_{I^{\prime}})(w^{\prime}W_{J^{\prime}})=(W_{I^{\prime}})(w^{\prime}W_{J^{\prime}})=C

so σHL\sigma\in H_{L}. Thus, IHLI^{\prime}\subseteq H_{L} and similarly JHRJ^{\prime}\subseteq H_{R}. Intersecting with {s1,,sn1}\{s_{1},\ldots,s_{n-1}\} shows that III^{\prime}\subseteq I and JJJ^{\prime}\subseteq J. Then WIWIW_{I^{\prime}}\subseteq W_{I} and WJWJW_{J^{\prime}}\subseteq W_{J} so

C=WIwWJ=WIwWJWIwWJ.C=W_{I^{\prime}}w^{\prime}W_{J^{\prime}}=W_{I^{\prime}}wW_{J^{\prime}}\subseteq W_{I}wW_{J}.

Combining this with the earlier inclusion WIwWJCW_{I}wW_{J}\subseteq C gives the equality WIwWJ=CW_{I}wW_{J}=C. Thus, (I,w,J)(I,w,J) is a presentation of CC. Finally, (I,w,J)(I,w,J) is maximal since we already showed that if (I,w,J)(I^{\prime},w^{\prime},J^{\prime}) is any presentation of CC, then III^{\prime}\subseteq I and JJJ^{\prime}\subseteq J. ∎

For an alternative proof of Proposition 2.7, see Proposition 3.7 in [1]. Proposition 2.7 gives a bijection

{parabolic doublecosets CSn}\displaystyle\left\{\begin{subarray}{c}\text{\small{parabolic double}}\\ \text{\small{cosets $C\in S_{n}$}}\end{subarray}\right\} {Triples (C,I,J) consisting of two collectionsI,J{s1,,sn1} and a parabolicdouble coset CWI\Sn/WJ such that I={si:siC=C} and J={si:Csi=C}}.\displaystyle\longleftrightarrow\left\{\begin{subarray}{c}\text{\small{Triples $(C,I,J)$ consisting of two collections}}\\ \text{\small{$I,J\subseteq\{s_{1},\ldots,s_{n-1}\}$ and a parabolic}}\\ \text{\small{double coset $C\in W_{I}\backslash S_{n}/W_{J}$ such that }}\\ \text{\small{$I=\{s_{i}:s_{i}C=C\}$ and $J=\{s_{i}:Cs_{i}=C\}$}}\end{subarray}\right\}.

This bijection is given by

C\displaystyle C (C,{si:siC=C},{si:Csi=C}),\displaystyle\longmapsto(C,\{s_{i}:s_{i}C=C\},\{s_{i}:Cs_{i}=C\}),
C\displaystyle C (C,I,J),\displaystyle\longmapsfrom(C,I,J),

where Proposition 2.7 guarentees that the triple (C,{si:siC=C},{si:Csi=C})(C,\{s_{i}:s_{i}C=C\},\{s_{i}:Cs_{i}=C\}) satisfies the condition CWI\Sn/WJC\in W_{I}\backslash S_{n}/W_{J} required by the set on the right hand side of the bijection.

The bijection between triples (C,I,J)(C,I,J) and two-way contingency tables described after Definition 2.5 gives a bijection

{Triples (C,I,J) consisting of two collectionsI,J{s1,,sn1} and a parabolicdouble coset CWI\Sn/WJ such that I={si:siC=C} and J={si:Csi=C}}{Two-way contingency tables with sum nwhose associated triple (C,I,J) satisfiesI={si:siC=C} and J={si:Csi=C}}.\left\{\begin{subarray}{c}\text{\small{Triples $(C,I,J)$ consisting of two collections}}\\ \text{\small{$I,J\subseteq\{s_{1},\ldots,s_{n-1}\}$ and a parabolic}}\\ \text{\small{double coset $C\in W_{I}\backslash S_{n}/W_{J}$ such that }}\\ \text{\small{$I=\{s_{i}:s_{i}C=C\}$ and $J=\{s_{i}:Cs_{i}=C\}$}}\end{subarray}\right\}\longleftrightarrow\left\{\begin{subarray}{c}\text{\small{Two-way contingency tables with sum $n$}}\\ \text{\small{whose associated triple $(C,I,J)$ satisfies}}\\ \text{\small{$I=\{s_{i}:s_{i}C=C\}$ and $J=\{s_{i}:Cs_{i}=C\}$}}\end{subarray}\right\}.

This motivates the following definition.

Definition 2.8.

A two-way contingency table is said to be maximal if the associated triple (C,I,J)(C,I,J) satisfies I={si:siC=C}I=\{s_{i}:s_{i}C=C\} and J={si:Csi=C}J=\{s_{i}:Cs_{i}=C\}.

We remark that the maximal two-way contingency table associated to a given parabolic double coset has the smallest possible dimensions and the largest possible entries.

Putting this all together gives a bijection

{parabolic doublecosets CSn}{Maximal two-waycontingency tableswith sum n}.\left\{\begin{subarray}{c}\text{\small{parabolic double}}\\ \text{\small{cosets $C\in S_{n}$}}\end{subarray}\right\}\longleftrightarrow\left\{\begin{subarray}{c}\text{\small{Maximal two-way}}\\ \text{\small{contingency tables}}\\ \text{\small{with sum $n$}}\end{subarray}\right\}.
Corollary 2.9.

The sequence {pn,0}\{p_{n,0}\} counts the number of two-way contingency tables with sum nn. The sequence {pn}\{p_{n}\} counts the number of maximal two-way contingency tables with sum nn.

We will now turn our attention to counting maximal two-way contingency tables with sum nn.

3 Computation of pnp_{n}

3.1 Characterization of Maximality

The following proposition describes how to determine whether the conditions siC=Cs_{i}C=C and Csi=CCs_{i}=C are satisfied from the balls-in-boxes picture with walls. We will use the term cell to mean one of the rectangular regions formed by the walls of the balls-in-boxes picture with walls.

Proposition 3.1.

Let C=WIwWJC=W_{I}wW_{J} be a parabolic double coset in SnS_{n}, and let PP be the balls-in-boxes picture with walls for the presentation (I,w,J)(I,w,J).

  1. 1.

    Let si{s1,,sn1}Is_{i}\in\{s_{1},\ldots,s_{n-1}\}\setminus I. Then siC=Cs_{i}C=C if and only if the row of cells of PP directly below the sis_{i}-wall and the row of cells of PP directly above the sis_{i}-wall each have only one nonempty cell, the two of which are vertically adjacent.

  2. 2.

    Let sj{s1,,sn1}Js_{j}\in\{s_{1},\ldots,s_{n-1}\}\setminus J. Then Csj=CCs_{j}=C if and only if the column of cells of PP directly left the sjs_{j}-wall and the column of cells of PP directly right the sjs_{j}-wall each have only one nonempty cell, the two of which are horizontally adjacent.

Proof.

Let si{s1,,sn1}Is_{i}\in\{s_{1},\ldots,s_{n-1}\}\setminus I. First suppose that the row of cells of PP directly below the sis_{i}-wall and the row of cells of PP directly above the sis_{i}-wall each have only one nonempty cell, the two of which are vertically adjacent. Let AA be the nonempty cell of PP directly below the sis_{i}-wall and let BB be the nonempty cell of PP directly above the sis_{i}-wall. Since AA and BB are each the only nonempty cells in their row of cells, the ball in row ii must lie in AA and the ball in row i+1i+1 must lie in BB. Then the ball in column w1(i)w^{-1}(i) must lie in AA and the ball in column w1(i+1)w^{-1}(i+1) must lie in BB. Since the cells AA and BB are vertically adjacent, w1(i)w^{-1}(i) and w1(i+1)w^{-1}(i+1) lie in the same column of cells. Then the transposition w1siww^{-1}s_{i}w that swaps w1(i)w^{-1}(i) and w1(i+1)w^{-1}(i+1) is an element of WJW_{J}. Taking w1siwWJw^{-1}s_{i}w\in W_{J} and multiplying on the left by ww gives siwwWJCs_{i}w\in wW_{J}\subseteq C. Since wCw\in C was arbitrary, this shows that siCCs_{i}C\subseteq C. Since |siC|=|C|\left\lvert{s_{i}C}\right\rvert=\left\lvert{C}\right\rvert, we must have siC=Cs_{i}C=C.

For the converse, suppose that PP has a nonempty cell AA directly below the sis_{i}-wall and a nonempty cell BB directly above the sis_{i}-wall such that AA and BB are not vertically adjacent. This is just the negation of the original assumption on PP. Since AA and BB are nonempty, we can multiply ww on the left by an element xWIx\in W_{I} to make AA contain the ball in row ii and to make BB contain the ball in row i+1i+1. Then sixws_{i}xw has fewer balls in cells AA and BB than xwxw. However, the number of balls in a given cell is constant over all elements of CC. Since xwCxw\in C, this forces sixwCs_{i}xw\not\in C. In particular, siCCs_{i}C\neq C. This proves the first statement. The proof of the second statement is similar. ∎

By definition, a two-way contingency table is maximal precisely when neither of the “if and only if”s in Proposition 3.1 are ever satisfied. This gives a characterization of maximal two-way contingency tables.

Corollary 3.2.

Let TT be a two-way contingency table. Then TT is maximal if and only if TT satisfies both of the following two conditions:

  1. 1.

    There does not exist a pair of vertically adjacent nonzero entries, each of which is the only nonzero entry in its row.

  2. 2.

    There does not exist a pair of horizontally adjacent nonzero entries, each of which is the only nonzero entry in its column.

Example 3.3.

There is p1=1p_{1}=1 maximal two-way contingency table with sum 1:

11

There are p2=3p_{2}=3 maximal two-way contingency tables with sum 2:

22110110011011

There are p3=19p_{3}=19 maximal two-way contingency tables with sum 3:

110111101111111111110111101101102211022002201122011001101101111011011011001111001111001133001101101100011000111100001111000110011011000011110000110110110001100011

3.2 Sequence Transformation

Note that the conditions in Corollary 3.2 only depend on the configuration of nonzero entries of TT. Because of this, it will be helpful to ignore the specific values of the nonzero entries of TT.

Definition 3.4.

A pattern is a two-way contingency table consisting of 0’s and 1’s.

  1. 1.

    Let 𝒫m,0\mathcal{P}_{m,0} denote the collection of all patterns with sum mm.

  2. 2.

    Let 𝒫m\mathcal{P}_{m} denote the collection of all patterns with sum mm that satisfy the conditions of Corollary 3.2.

  3. 3.

    For a two-way contingency table TT, we define the pattern associated to TT to be the two-way contingency table formed by replacing all nonzero entries of TT with 1’s.

For each pattern PP with sum mm, there are exactly (n1nm)\binom{n-1}{n-m} two-way contingency tables with sum nn whose associated pattern is PP. This is because two-way contingency tables with sum nn whose associated pattern is PP are in bijection with ways to surjectively distribute nn identical balls among mm distinguishable urns.

Then combining Corollary 2.9 with Corollary 3.2 gives the formulas

pn,0=m=0n(n1nm)|𝒫m,0|andpn=m=0n(n1nm)|𝒫m|.p_{n,0}=\sum_{m=0}^{n}\binom{n-1}{n-m}\left\lvert{\mathcal{P}_{m,0}}\right\rvert\quad\text{and}\quad p_{n}=\sum_{m=0}^{n}\binom{n-1}{n-m}\left\lvert{\mathcal{P}_{m}}\right\rvert.

Here we use the conventions (10)=1\binom{-1}{0}=1 and also (k1k)=0\binom{k-1}{k}=0 for k1k\geq 1, since (n1nm)\binom{n-1}{n-m} is counting ways to surjectively distribute nn identical balls among mm distinguishable urns. We now invoke the identity

(n1nm)=m!n!k=mn[nk]{km},\binom{n-1}{n-m}=\frac{m!}{n!}\sum_{k=m}^{n}\genfrac{[}{]}{0.0pt}{}{n}{k}\genfrac{\{}{\}}{0.0pt}{}{k}{m},

where []\genfrac{[}{]}{0.0pt}{}{\cdot}{\cdot} denotes unsigned Stirling numbers of the first kind and {}\genfrac{\{}{\}}{0.0pt}{}{\cdot}{\cdot} denotes Stirling numbers of the second kind (see the exercise on Lah numbers in [8]). If we define

qk,0=m=0km!{km}|𝒫m,0|andqk=m=0km!{km}|𝒫m|,q_{k,0}=\sum_{m=0}^{k}m!\genfrac{\{}{\}}{0.0pt}{}{k}{m}\left\lvert{\mathcal{P}_{m,0}}\right\rvert\quad\text{and}\quad q_{k}=\sum_{m=0}^{k}m!\genfrac{\{}{\}}{0.0pt}{}{k}{m}\left\lvert{\mathcal{P}_{m}}\right\rvert,

then we obtain the identities

pn,0=1n!k=0n[nk]qk,0andpn=1n!k=0n[nk]qk.p_{n,0}=\frac{1}{n!}\sum_{k=0}^{n}\genfrac{[}{]}{0.0pt}{}{n}{k}q_{k,0}\quad\text{and}\quad p_{n}=\frac{1}{n!}\sum_{k=0}^{n}\genfrac{[}{]}{0.0pt}{}{n}{k}q_{k}.

This reduces the computation of the sequences {pn,0}\{p_{n,0}\} and {pn}\{p_{n}\} to the computation of the sequences {qn,0}\{q_{n,0}\} and {qn}\{q_{n}\}. The following proposition gives a combinatorial interpretation of the sequences {qn,0}\{q_{n,0}\} and {qn}\{q_{n}\}.

Proposition 3.5.

The sequence {qn,0}\{q_{n,0}\} counts the number of functions from nn distinguishable balls to a rectangular matrix of boxes, such that each row and column is nonempty. The sequence {qn}\{q_{n}\} counts the number of such functions that satisfy the following two conditions:

  1. 1.

    There does not exist a pair of vertically adjacent nonempty boxes, each of which is the only nonempty box in its row.

  2. 2.

    There does not exist a pair of horizontally adjacent nonempty boxes, each of which is the only nonempty box in its column.

Proof.

Relabeling indices (from qk,0q_{k,0} to qn,0q_{n,0} and from qkq_{k} to qnq_{n}) gives

qn,0=k=0nk!{nk}|𝒫k,0|andqn=k=0nk!{nk}|𝒫k|.q_{n,0}=\sum_{k=0}^{n}k!\genfrac{\{}{\}}{0.0pt}{}{n}{k}\left\lvert{\mathcal{P}_{k,0}}\right\rvert\quad\text{and}\quad q_{n}=\sum_{k=0}^{n}k!\genfrac{\{}{\}}{0.0pt}{}{n}{k}\left\lvert{\mathcal{P}_{k}}\right\rvert.

Then the proposition follows from the fact that k!{nk}k!\genfrac{\{}{\}}{0.0pt}{}{n}{k} counts the number of surjective functions from nn distinguishable balls to the kk nonempty boxes prescribed by the pattern in 𝒫k,0\mathcal{P}_{k,0} or 𝒫k\mathcal{P}_{k}. ∎

3.3 Pairs of Weak Orders

Definition 3.6.

A weak order on a set SS is a binary relation \leq on SS that is transitive and has no incomparable pairs of elements (so for all s,tSs,t\in S, sts\leq t or tst\leq s). The total number of weak orders on {1,,n}\{1,\ldots,n\} is denoted by fnf_{n}. The sequence {fn}\{f_{n}\} is known as the Fubini numbers or the ordered Bell numbers, and may be found at [5, A000670].

Given a weak order on a set SS, it is possible for two elements s,tSs,t\in S to be tied with each other, in the sense that both sts\leq t and tst\leq s. The “tied” relation is an equivalence relation and partitions the set SS into tied subsets. Furthermore, the weak order on SS gives a total order on this partition. In other words, we obtain an ordered set-partition S=S1SkS=S_{1}\cup\cdots\cup S_{k}. Conversely, an ordered set-partition S=S1SkS=S_{1}\cup\cdots\cup S_{k} determines a weak order on SS by setting sts\leq t whenever sSis\in S_{i} and tSjt\in S_{j} and iji\leq j. Thus, there is a bijection between weak orders on SS and ordered set-partitions on SS. In what follows, we will freely translate back and forth between weak orders and ordered set-partitions.

In the special case where S={1,,n}S=\{1,\ldots,n\}, weak orders on {1,,n}\{1,\ldots,n\} correspond to ordered set-partitions {1,,n}=S1Sk\{1,\ldots,n\}=S_{1}\cup\cdots\cup S_{k}. Summing over the possible values for kk gives the formula fn=k=0nk!{nk}f_{n}=\sum_{k=0}^{n}k!\genfrac{\{}{\}}{0.0pt}{}{n}{k}.

Definition 3.7.

Let ((S1,,Sk),(T1,,Tk))((S_{1},\ldots,S_{k}),(T_{1},\ldots,T_{k^{\prime}})) be a pair of ordered set-partitions of {1,,n}\{1,\ldots,n\}. A left consecutive embedding is a consecutive pair (i,i+1)l(i,i+1)_{l} such that SiSi+1TjS_{i}\cup S_{i+1}\subseteq T_{j} for some (necessarily unique) TjT_{j}. Similarly, a right consecutive embedding is a consecutive pair (i,i+1)r(i,i+1)_{r} such that TiTi+1SjT_{i}\cup T_{i+1}\subseteq S_{j} for some (necessarily unique) SjS_{j}. The phrase consecutive embedding will refer to either a left consecutive embedding or a right consecutive embedding.

Since weak orders can be viewed as ordered set-partitions, it makes sense to talk about consecutive embeddings in a pair of weak orders.

Proposition 3.8.

The sequence {qn,0}\{q_{n,0}\} counts the total number of pairs of weak orders on {1,,n}\{1,\ldots,n\}, so qn,0=fn2q_{n,0}=f_{n}^{2}. The sequence {qn}\{q_{n}\} counts the number of such pairs that have no consecutive embeddings.

Proof.

Consider one of the qn,0q_{n,0} functions from nn distinguishable balls to a rectangular matrix of boxes, such that each row and column is nonempty. Considering the row of each ball gives a weak order on {1,,n}\{1,\ldots,n\}. Similarly, considering the column of each ball gives a weak order on {1,,n}\{1,\ldots,n\}. Conversely, given this pair of weak orders on {1,,n}\{1,\ldots,n\}, we can recover the dimensions of the matrix and the position of each ball. In particular, the dimensions of the matrix are equal to the number of parts of the two ordered set-partitions, and the position of the iith ball in the matrix is given by the location of ii within the two ordered set-partitions. This gives a bijection

{Functions from n distinguishable ballsto a rectangular matrix of boxes suchthat each row and column is nonempty}{Pairs of weak orderson {1,,n}}.\left\{\begin{subarray}{c}\text{\small{Functions from $n$ distinguishable balls}}\\ \text{\small{to a rectangular matrix of boxes such}}\\ \text{\small{that each row and column is nonempty}}\end{subarray}\right\}\longleftrightarrow\left\{\begin{subarray}{c}\text{\small{Pairs of weak orders}}\\ \text{\small{on $\{1,\ldots,n\}$}}\end{subarray}\right\}.

In particular, the sequence {qn,0}\{q_{n,0}\} counts the total number of pairs of weak orders on {1,,n}\{1,\ldots,n\}.

As for {qn}\{q_{n}\}, observe that this bijection takes failures of conditions 1 and 2 of Corollary 3.2 to left and right consecutive embeddings. ∎

Example 3.9.

There is q1,0=1q_{1,0}=1 pair of weak orders on {1}\{1\}:

1
1

This pair of weak orders has no consecutive embeddings, so we also have q1=1q_{1}=1. Here we write the first weak order in the first column, and the second weak order in the second column. Tied elements of a weak order are written in the same row as each other.

There are q2,0=9q_{2,0}=9 pairs of weak orders on {1,2}\{1,2\}:

1
2
1
2
   
1
2
2
1
   
2
1
1
2
   
2
1
2
1
   
1,2
1,2
1,2
1
2
   
1,2
2
1
   
1
2
1,2
   
2
1
1,2

Of these, the five in the first row have no consecutive embeddings, whereas the first two in the second row each have a right consecutive embedding, and the last two in the second row each have a left consecutive embedding. Thus, q2=5q_{2}=5. Then

p2,0=12!([21]q1,0+[22]q2,0)=12(1+9)=5andp2=12!([21]q1+[21]q2)=12(1+5)=3.p_{2,0}=\frac{1}{2!}\left(\genfrac{[}{]}{0.0pt}{}{2}{1}q_{1,0}+\genfrac{[}{]}{0.0pt}{}{2}{2}q_{2,0}\right)=\frac{1}{2}(1+9)=5\qquad\text{and}\qquad p_{2}=\frac{1}{2!}\left(\genfrac{[}{]}{0.0pt}{}{2}{1}q_{1}+\genfrac{[}{]}{0.0pt}{}{2}{1}q_{2}\right)=\frac{1}{2}(1+5)=3.

3.4 Inclusion-Exclusion

Recall that we have reduced the computation of the sequences {pn,0}\{p_{n,0}\} and {pn}\{p_{n}\} to the computation of the sequences {qn,0}\{q_{n,0}\} and {qn}\{q_{n}\}. The formula qn,0=fn2q_{n,0}=f_{n}^{2} from Proposition 3.8 gives a fast polynomial-time algorithm for computing the sequence {qn,0}\{q_{n,0}\}. We now restrict our attention to computing the sequence {qn}\{q_{n}\}. This subsection will give an inclusion-exclusion formula for qnq_{n}.

Definition 3.10.
  1. 1.

    A pair of weak orders on {1,,n}\{1,\ldots,n\} with distinguished consecutive embeddings consists of a pair of weak orders (,)(\leq,\leq^{\prime}) on {1,,n}\{1,\ldots,n\}, together with set 𝒞\mathcal{C} of consecutive embeddings in (,)(\leq,\leq^{\prime}). In other words, 𝒞\mathcal{C} is a subset of the set of all consecutive embeddings in (,)(\leq,\leq^{\prime}).

  2. 2.

    A pairs of weak orders on {1,,n}\{1,\ldots,n\} with kk distinguished consecutive embeddings is a pair of weak orders on {1,,n}\{1,\ldots,n\} with distinguished consecutive embeddings whose set 𝒞\mathcal{C} has cardinality kk.

  3. 3.

    Let qn,kq_{n,k} count the number of pairs of weak orders on {1,,n}\{1,\ldots,n\} with kk distinguished consecutive embeddings.

The notation qn,kq_{n,k} does not conflict with the notation qn,0q_{n,0} because pairs of weak orders with 0 distinguished consecutive embeddings are in bijection with (ordinary) pairs of weak orders.

Example 3.11.

We will circle distinguished consecutive embeddings. There are q2,0=9q_{2,0}=9 pairs of weak orders on {1,2}\{1,2\} with 0 distinguished consecutive embeddings:

1
2
1
2
   
1
2
2
1
   
2
1
1
2
   
2
1
2
1
   
1,2
1,2
1,2
1
2
   
1,2
2
1
   
1
2
1,2
   
2
1
1,2

There are q2,1=4q_{2,1}=4 pairs of weak orders on {1,2}\{1,2\} with 1 distinguished consecutive embedding:

1,2
1
2
   
1,2
2
1
   
1
2
1,2
   
2
1
1,2
Example 3.12.

Figure 2 depicts a more complicated example which we will return to later. The example in Figure 2 is a pair of weak orders on {1,2,3,4,5,6,7,8,9}\{1,2,3,4,5,6,7,8,9\} with 5 total consecutive embeddings, 4 of which are distinguished:

2
7
5
1,3,4,6,8,9
8
6
1,4
3
2,5,7
9
Figure 2: A pair of weak orders with distinguished consecutive embeddings.

The first few values of qn,kq_{n,k} are given in Table 1. There are two main observations from this table that we will prove below. First, qn,k=0q_{n,k}=0 for k>nk>n (Corollary 3.16). Second, k=0n(1)kqn,k=qn\sum_{k=0}^{n}(-1)^{k}q_{n,k}=q_{n} (Lemma 3.17).

nn qnq_{n} qn,0q_{n,0} qn,1q_{n,1} qn,2q_{n,2} qn,3q_{n,3} qn,4q_{n,4}
0 1 1 0 0 0 0
11 1 1 0 0 0 0
22 5 9 4 0 0 0
33 97 169 84 12 0 0
44 3365 5625 2812 600 48 0
55 177601 292681 145380 34380 4320 240
Table 1: Small Values of qnq_{n} and qn,kq_{n,k}

We will need the notion of a chain in a pair of weak orders with distinguished consecutive embeddings.

Definition 3.13.

In a pair of weak orders on {1,,n}\{1,\ldots,n\} with distinguished consecutive embeddings, a chain is a maximal sequence of overlapping distinguished consecutive embeddings.

The example in Figure 2 has one chain on the left side and two chains on the right side. The following lemma is a key property of chains.

Lemma 3.14.

In a pair of weak orders on {1,,n}\{1,\ldots,n\} with distinguished consecutive embeddings, any two chains have no elements of {1,,n}\{1,\ldots,n\} in common, even if the two chains are on different sides.

Proof.

It is clear that two chains on the same side have no elements of {1,,n}\{1,\ldots,n\} in common. It remains to show that any two consecutive embeddings on different sides have no elements of {1,,n}\{1,\ldots,n\} in common. Let (i,i+1)l(i,i+1)_{l} be a left consecutive embedding and let (i,i+1)r(i^{\prime},i^{\prime}+1)_{r} be a right consecutive embedding. Then SiSi+1TjS_{i}\cup S_{i+1}\subseteq T_{j} for some (necessarily unique) TjT_{j} and TiTi+1SjT_{i^{\prime}}\cup T_{i^{\prime}+1}\subseteq S_{j^{\prime}} for some (necessarily unique) SjS_{j^{\prime}}. If these two consecutive embeddings have an element of {1,,n}\{1,\ldots,n\} in common, then TjT_{j} must be either TiT_{i^{\prime}} or Ti+1T_{i^{\prime}+1}, and SjS_{j^{\prime}} must be either SiS_{i} or Si+1S_{i+1}. Then

|SiSi+1||Tj|<|TiTi+1||Sj|<|SiSi+1|\left\lvert{S_{i}\cup S_{i+1}}\right\rvert\leq\left\lvert{T_{j}}\right\rvert<\left\lvert{T_{i^{\prime}}\cup T_{i^{\prime}+1}}\right\rvert\leq\left\lvert{S_{j^{\prime}}}\right\rvert<\left\lvert{S_{i}\cup S_{i+1}}\right\rvert

which is a contradiction. ∎

We can now determine qn,kq_{n,k} for kn1k\geq n-1.

Proposition 3.15.

If knk\geq n and k1k\geq 1, then qn,k=0q_{n,k}=0. If k=n1k=n-1 and k1k\geq 1, then qn,k=2n!q_{n,k}=2\cdot n!.

Proof.

Consider a pair of weak orders on {1,,n}\{1,\ldots,n\} with k1k\geq 1 distinguished consecutive embeddings. Let c1c\geq 1 be the number of chains. Note that c+kc+k counts the number of parts that are contained in the chains, because each chain has one more part than distinguished consecutive embedding. Note that these c+kc+k parts have no elements in common, since any two parts within a chain have no elements in common, and Lemma 3.14 states that distinct chains have no elements in common. This gives the inequality c+knc+k\leq n.

If knk\geq n, then this is impossible, which shows that qn,k=0q_{n,k}=0. Now suppose that k=n1k=n-1. Then c=1c=1, so there is one chain consisting of n1n-1 overlapping distinguished consecutive embeddings. Thus, one side (either left or right) has nn parts in one of the n!n! possible permutations, and the other side has just one part consisting of the unordered set {1,,n}\{1,\ldots,n\}. This gives 2n!2\cdot n! possibilities. ∎

Corollary 3.16.

If k>nk>n, then qn,k=0q_{n,k}=0.

Proof.

If k>nk>n, then k1k\geq 1 so Proposition 3.15 gives qn,k=0q_{n,k}=0. ∎

We now give the inclusion-exclusion formula for qnq_{n} in terms of qn,kq_{n,k}.

Lemma 3.17.

We have the inclusion-exclusion formula

qn=k=0n(1)kqn,k.q_{n}=\sum_{k=0}^{n}(-1)^{k}q_{n,k}.
Proof.

If a pair of weak orders on {1,,n}\{1,\ldots,n\} has mm total consecutive embeddings, then mnm\leq n by Corollary 3.16, so that pair of weak orders is counted k=0m(1)k(mk)\sum_{k=0}^{m}(-1)^{k}\binom{m}{k} times by the sum. This is equal to 0 if m1m\geq 1 and equal to 1 if m=0m=0. Thus, the sum counts pairs of weak orders with no consecutive embeddings. ∎

3.5 The Final Count

It remains to count pairs of weak orders on {1,,n}\{1,\ldots,n\} with kk distinguished consecutive embeddings. To deal with the fact that consecutive embeddings can overlap, we will work with chains. The first step of our formula will be to sum over the number of chains, which we will call cc, and the total number of elements from {1,,n}\{1,\ldots,n\} to put inside these chains, which we will call tt. Keep in mind Lemma 3.14, which states that distinct chains have no elements in common, even if they are on different sides. In other words, each of the tt elements will end up in exactly one of the cc chains.

There are (nt)\binom{n}{t} ways to choose which tt elements from {1,,n}\{1,\ldots,n\} to put inside the cc chains. In the proof of Proposition 3.15 we saw that the cc chains contain c+kc+k total parts, so tc+kt\geq c+k. Then there are (c+k)!{tc+k}(c+k)!\genfrac{\{}{\}}{0.0pt}{}{t}{c+k} ways to partition these tt elements into an ordered sequence of c+kc+k parts.

Now we cut the ordered sequence of c+kc+k parts into cc chains, each of which must consist at least 2 parts. Since the parts are already ordered, this just requires choosing how many parts to put into each chain. The number of ways to do this equals the number of ways to distribute c+kc+k identical balls among cc distinguishable urns, each of which must get at least 2 balls. This is the same as the number of ways to surjectively distribute kk identical balls among cc distinguishable urns. There are (k1kc)\binom{k-1}{k-c} such ways.

At this point, we have an ordered sequence of cc chains, along with ntn-t non-chain elements. We can minimize the distinction between the cc chains and the ntn-t elements by collapsing each chain into a single element, as depicted in Figure 3. For example, consider the top right chain in Figure 3. Both the left and right side of this chain get collapsed to a single element (which is labeled c2c_{2} in Figure 3). The key difference between the left c2c_{2} and the right c2c_{2} is that the right c2c_{2} must be in a part by itself. Thus, the circling in the collapsed form indicates which side of the collapsed chain must be in a part by itself.

2
7
5
1,3,4,6,8,9
8
6
1,4
3
2,5,7
9

\quad\longrightarrow\quad c1c_{1} c2c_{2},c3c_{3},9 c2c_{2} c3c_{3} c1c_{1} 9

Figure 3: Collapsing chains.

We will need a generalization of the Fubini numbers that allows for the possibility of elements that are required to be placed in a part by themselves. For 0kn0\leq k\leq n, let fn,kf_{n,k} count the number of weak orders on {1,,n}\{1,\ldots,n\} where each of the elements 1,,k1,\ldots,k must be placed in a part by itself. These generalized Fubini numbers have the formula fn,k=l=knl!{nklk}f_{n,k}=\sum_{l=k}^{n}l!\genfrac{\{}{\}}{0.0pt}{}{n-k}{l-k}, because there are {nklk}\genfrac{\{}{\}}{0.0pt}{}{n-k}{l-k} ways to partition the set {k+1,,n}\{k+1,\ldots,n\} into lkl-k parts, and l!l! ways to order these lkl-k parts along with the remaining kk elements {1,,k}\{1,\ldots,k\}. The ordinary Fubini numbers are the special case when k=0k=0.

Returning to the situation prior to Figure 3, we now sum over the number of chains on the left side, which we will call jj. This forces the number of chains on the right side to be cjc-j. We already have an ordering on the cc chains, so we will take the first jj chains to be the chains on the left side, and the last cjc-j chains to be the chains on the right side. After collapsing each of the cc chains into a single element on each side, we have nt+cn-t+c total elements to distribute. On the left side, jj of these elements are marked as having to be placed in a part by themselves. On the right side, cjc-j of these elements are marked as having to be placed in a part by themselves. Then there are fnt+c,jfnt+c,cjf_{n-t+c,j}f_{n-t+c,c-j} ways to produce a valid pair of weak orders. However, since we already have an ordering for the cc chains, we must divide by j!j! and (cj)!(c-j)!. This gives the formula

qn,k=c=0kt=c+kn(nt)choose telements(c+k)!{tc+k}partition the telements into anordered sequenceof c+k parts(k1kc)split thec+k partsinto c chainsj=0cfnt+c,jfnt+c,cjcollapsing each chain intoa single element on each sidemakes n become nt+cj!(cj)!the chains are alreadyordered at this point,so the numerator isovercounting by this factor.q_{n,k}=\sum_{c=0}^{k}\sum_{t=c+k}^{n}\underbrace{\binom{n}{t}}_{\begin{subarray}{c}\text{choose $t$}\\ \text{elements}\end{subarray}}\underbrace{(c+k)!\genfrac{\{}{\}}{0.0pt}{}{t}{c+k}}_{\begin{subarray}{c}\text{partition the $t$}\\ \text{elements into an}\\ \text{ordered sequence}\\ \text{of $c+k$ parts}\end{subarray}}\underbrace{\binom{k-1}{k-c}}_{\begin{subarray}{c}\text{split the}\\ \text{$c+k$ parts}\\ \text{into $c$ chains}\end{subarray}}\sum_{j=0}^{c}\frac{\overbrace{f_{n-t+c,j}f_{n-t+c,c-j}}^{\begin{subarray}{c}\text{collapsing each chain into}\\ \text{a single element on each side}\\ \text{makes $n$ become $n-t+c$}\end{subarray}}}{\underbrace{j!\,(c-j)!}_{\begin{subarray}{c}\text{the chains are already}\\ \text{ordered at this point,}\\ \text{so the numerator is}\\ \text{overcounting by this factor}\end{subarray}}}.

We should point out that the second sum is zero for c>nkc>n-k. Plugging this into Lemma 3.17 and rearranging the summations gives the formula

qn=c=0n/2t=2cn(nt)(k=ctc(1)k(c+k)!{tc+k}(k1kc))(j=0cfnt+c,jfnt+c,cjj!(cj)!).q_{n}=\sum_{c=0}^{\lfloor n/2\rfloor}\sum_{t=2c}^{n}\binom{n}{t}\left(\sum_{k=c}^{t-c}(-1)^{k}(c+k)!\genfrac{\{}{\}}{0.0pt}{}{t}{c+k}\binom{k-1}{k-c}\right)\left(\sum_{j=0}^{c}\frac{f_{n-t+c,j}f_{n-t+c,c-j}}{j!\,(c-j)!}\right). (2)

This proves our formula for pnp_{n} (equation 1 from the introduction).

Theorem 3.18.

For each nonnegative integer nn, we have the formula

pn=1n!m=0n[nm]c=0m/2t=2cm(mt)(k=ctc(1)k(c+k)!{tc+k}(k1kc))(j=0cfmt+c,jfmt+c,cjj!(cj)!),p_{n}=\frac{1}{n!}\sum_{m=0}^{n}\genfrac{[}{]}{0.0pt}{}{n}{m}\sum_{c=0}^{\lfloor m/2\rfloor}\sum_{t=2c}^{m}\binom{m}{t}\left(\sum_{k=c}^{t-c}(-1)^{k}(c+k)!\genfrac{\{}{\}}{0.0pt}{}{t}{c+k}\binom{k-1}{k-c}\right)\left(\sum_{j=0}^{c}\frac{f_{m-t+c,j}f_{m-t+c,c-j}}{j!\,(c-j)!}\right),

where the values fn,k=l=knl!{nklk}f_{n,k}=\sum_{l=k}^{n}l!\genfrac{\{}{\}}{0.0pt}{}{n-k}{l-k} are generalizations of the Fubini numbers.

Equation 2 has been written so that there are several independent components inside the formula. The next subsection shows how we can precompute these components to speed up computation.

3.6 The Algorithm

We now give an algorithm to compute the values p1,,pNp_{1},\ldots,p_{N} in O(N2)O(N^{2}) memory and O(N3)O(N^{3}) time. The statement “O(N2)O(N^{2}) memory and O(N3)O(N^{3}) time” assumes that storing a number requires constant memory and that arithmetic operations require constant time. This assumption is not true in practice since the computation involves numbers that grow super-exponentially. However, the number of digits of these numbers grows polynomially. Then storing a number requires polynomial memory and arithmetic operations require polynomial time. Thus, even with these practical considerations, the algorithm uses polynomial memory and polynomial time. Empirically, the runtime for the algorithm grows like N4.75N^{4.75}.

Here is the algorithm:

  1. 1.

    Compute the values (nk),[nk],{nk}\displaystyle\binom{n}{k},\genfrac{[}{]}{0.0pt}{}{n}{k},\genfrac{\{}{\}}{0.0pt}{}{n}{k}, for 0knN0\leq k\leq n\leq N using the standard recurrences.

  2. 2.

    Compute the values fn,k=l=knl!{nklk}f_{n,k}=\displaystyle\sum_{l=k}^{n}l!\genfrac{\{}{\}}{0.0pt}{}{n-k}{l-k} for 0knN0\leq k\leq n\leq N.

  3. 3.

    Compute the values gn,c=j=0cfn,jfn,cjj!(cj)!g_{n,c}=\displaystyle\sum_{j=0}^{c}\frac{f_{n,j}f_{n,c-j}}{j!\,(c-j)!} for 0cnN0\leq c\leq n\leq N.

  4. 4.

    Compute the values ht,c=k=ctc(1)k(c+k)!{tc+k}(k1kc)h_{t,c}=\displaystyle\sum_{k=c}^{t-c}(-1)^{k}(c+k)!\genfrac{\{}{\}}{0.0pt}{}{t}{c+k}\binom{k-1}{k-c} for 02ctn0\leq 2c\leq t\leq n.

    In this calculation, note that (10)=1\binom{-1}{0}=1 and also (k1k)=0\binom{k-1}{k}=0 for k1k\geq 1. We will give a faster and simpler algorithm for ht,ch_{t,c} in the next subsection.

  5. 5.

    Compute the values qn=c=0n/2t=2cn(nt)ht,cgnt+c,cq_{n}=\displaystyle\sum_{c=0}^{\lfloor n/2\rfloor}\sum_{t=2c}^{n}\binom{n}{t}h_{t,c}g_{n-t+c,c} for 0nN0\leq n\leq N.

  6. 6.

    Compute the values pn=1n!k=0n[nk]qkp_{n}=\displaystyle\frac{1}{n!}\sum_{k=0}^{n}\genfrac{[}{]}{0.0pt}{}{n}{k}q_{k} for 0nN0\leq n\leq N.

I have written two versions of this algorithm in Java (see Web Resources at the end of this paper).

Version 1 is a straightforward single-threaded implementation of this algorithm, with the faster and simpler step 4 given in the next subsection. Version 1 can compute the values p1,,p1000p_{1},\ldots,p_{1000} in 40 minutes on a laptop with a 2.4 GHz processor using 2 GB of memory. Version 1 is limited by memory.

Version 2 is a memory-optimized multi-threaded implementation of the algorithm that was used to compute the values p1,,p5000p_{1},\ldots,p_{5000}. The computation took 9.4 days on a server with 20 CPU cores, each running at 2.4 GHz. Version 2 achieves O(N)O(N) memory and O(N3)O(N^{3}) time (with the same caveat as before). It achieves O(N)O(N) memory by not storing precomputed two-dimensional arrays and instead recomputing values on the fly. Version 2 maintains the O(N3)O(N^{3}) time of Version 1 by rewriting formula (2) for qnq_{n} so that the outside sum is a sum over the values of nt+cn-t+c, which avoids expensive recomputation of the values gnt+c,cg_{n-t+c,c}. Version 2 is limited by time, rather than by memory.

3.7 Combinatorial Interpretation of ht,ch_{t,c}

We conclude this section by giving a combinatorial interpretation of ht,ch_{t,c}. This will give a fast and simple algorithm for ht,ch_{t,c}, as well as a bound on ht,ch_{t,c} that will be used when proving asymptotic formulas.

Proposition 3.19.

For all integers 02ct0\leq 2c\leq t, the quantity (1)cht,c/2c(-1)^{c}h_{t,c}/2^{c} counts the number of ordered set-partitions of {1,,t}\{1,\ldots,t\} into cc parts of even cardinality. In particular, if tt is odd, then ht,c=0h_{t,c}=0.

Proof.

First, we rewrite ht,ch_{t,c} as

ht,c=(1)ck=2ct(1)kk!{tk}(kc1k2c).h_{t,c}=(-1)^{c}\sum_{k=2c}^{t}(-1)^{k}k!\genfrac{\{}{\}}{0.0pt}{}{t}{k}\binom{k-c-1}{k-2c}.

Recall that the quantity

(c+k)!{tc+k}(k1kc)(c+k)!\genfrac{\{}{\}}{0.0pt}{}{t}{c+k}\binom{k-1}{k-c}

from equation (2) counts the number of ways to choose an ordered set-partition of {1,,t}\{1,\ldots,t\} into c+kc+k parts, which are further grouped into cc chains of at least two parts each. Then the quantity

k!{tk}(kc1k2c)k!\genfrac{\{}{\}}{0.0pt}{}{t}{k}\binom{k-c-1}{k-2c}

counts the number of ways to choose an ordered set-partition of {1,,t}\{1,\ldots,t\} into kk parts, which are further grouped into cc consecutive blocks of at least two parts each. Here’s example with c=3c=3, k=7k=7, and t=10t=10:

3 5
4 8,10 2
1,7,9 6

Considering the cc blocks gives an ordered set-partition of {1,,t}\{1,\ldots,t\} into cc parts. If we let Ω\Omega denote the collection of all ordered set-partitions of {1,,t}\{1,\ldots,t\} into cc parts, then we can write

k!{tk}(kc1k2c)=πΩaπ,k!\genfrac{\{}{\}}{0.0pt}{}{t}{k}\binom{k-c-1}{k-2c}=\sum_{\pi\in\Omega}a_{\pi},

where aπa_{\pi} counts the number of ways that πΩ\pi\in\Omega arises as the cc blocks of an ordered set-partition of {1,,t}\{1,\ldots,t\} into kk parts, which are further grouped into cc consecutive blocks of at least two parts each. If π\pi is the ordered set-partition {1,,t}=π1πc\{1,\ldots,t\}=\pi_{1}\cup\cdots\cup\pi_{c}, then we can write aπa_{\pi} as

aπ=k=k1++kcki2i=1cki!{|πi|ki}.a_{\pi}=\sum_{\begin{subarray}{c}k=k_{1}+\cdots+k_{c}\\ k_{i}\geq 2\end{subarray}}\prod_{i=1}^{c}k_{i}!\genfrac{\{}{\}}{0.0pt}{}{\left\lvert{\pi_{i}}\right\rvert}{k_{i}}.

Thus,

k!{tk}(kc1k2c)=πΩk=k1++kcki2i=1cki!{|πi|ki}k!\genfrac{\{}{\}}{0.0pt}{}{t}{k}\binom{k-c-1}{k-2c}=\sum_{\pi\in\Omega}\sum_{\begin{subarray}{c}k=k_{1}+\cdots+k_{c}\\ k_{i}\geq 2\end{subarray}}\prod_{i=1}^{c}k_{i}!\genfrac{\{}{\}}{0.0pt}{}{\left\lvert{\pi_{i}}\right\rvert}{k_{i}}

Putting this all together gives

ht,c\displaystyle h_{t,c} =(1)ck=2ct(1)kπΩk=k1++kcki2i=1cki!{|πi|ki}\displaystyle=(-1)^{c}\sum_{k=2c}^{t}(-1)^{k}\sum_{\pi\in\Omega}\sum_{\begin{subarray}{c}k=k_{1}+\cdots+k_{c}\\ k_{i}\geq 2\end{subarray}}\prod_{i=1}^{c}k_{i}!\genfrac{\{}{\}}{0.0pt}{}{\left\lvert{\pi_{i}}\right\rvert}{k_{i}}
=(1)cπΩk=2ctk=k1++kcki2i=1c(1)kiki!{|πi|ki}\displaystyle=(-1)^{c}\sum_{\pi\in\Omega}\sum_{k=2c}^{t}\sum_{\begin{subarray}{c}k=k_{1}+\cdots+k_{c}\\ k_{i}\geq 2\end{subarray}}\prod_{i=1}^{c}(-1)^{k_{i}}k_{i}!\genfrac{\{}{\}}{0.0pt}{}{\left\lvert{\pi_{i}}\right\rvert}{k_{i}}
=(1)cπΩk1,,kc2ki|πi|i=1c(1)kiki!{|πi|ki}\displaystyle=(-1)^{c}\sum_{\pi\in\Omega}\sum_{\begin{subarray}{c}k_{1},\ldots,k_{c}\\ 2\leq k_{i}\leq\left\lvert{\pi_{i}}\right\rvert\end{subarray}}\prod_{i=1}^{c}(-1)^{k_{i}}k_{i}!\genfrac{\{}{\}}{0.0pt}{}{\left\lvert{\pi_{i}}\right\rvert}{k_{i}}
=(1)cπΩi=1cki=2|πi|(1)kiki!{|πi|ki}.\displaystyle=(-1)^{c}\sum_{\pi\in\Omega}\prod_{i=1}^{c}\sum_{k_{i}=2}^{\left\lvert{\pi_{i}}\right\rvert}(-1)^{k_{i}}k_{i}!\genfrac{\{}{\}}{0.0pt}{}{\left\lvert{\pi_{i}}\right\rvert}{k_{i}}.

We now invoke the identity

k=2n(1)kk!{nk}={2if n is even,0if n is odd\sum_{k=2}^{n}(-1)^{k}k!\genfrac{\{}{\}}{0.0pt}{}{n}{k}=\begin{cases}2&\text{if }n\text{ is even},\\ 0&\text{if }n\text{ is odd}\end{cases}

which follows from setting x=1x=-1 in the identity

k=0n{nk}(x)k=xn.\sum_{k=0}^{n}\genfrac{\{}{\}}{0.0pt}{}{n}{k}(x)_{k}=x^{n}.

where (x)k=x(x1)(xk+1)(x)_{k}=x(x-1)\cdots(x-k+1) denotes the falling factorial. Thus, if each πi\pi_{i} has even cardinality, then π\pi contributes (1)c2c(-1)^{c}2^{c} to ht,ch_{t,c}, and otherwise π\pi contributes nothing to ht,ch_{t,c}. The result follows. ∎

Corollary 3.20.

For all integers 02ct0\leq 2c\leq t, we have the inequality |ht,c|2cct\left\lvert{h_{t,c}}\right\rvert\leq 2^{c}c^{t}.

Proof.

Proposition 3.19 states that |ht,c|/2c\left\lvert{h_{t,c}}\right\rvert/2^{c} counts the number of ordered set-partitions of {1,,t}\{1,\ldots,t\} into cc parts of even cardinality. This is bounded above by ctc^{t} (the number of functions {1,,t}{1,,c}\{1,\ldots,t\}\to\{1,\ldots,c\}). ∎

The combinatorial interpretation of ht,ch_{t,c} given in Proposition 3.19 also gives a recurrence for ht,ch_{t,c}. It will be convenient to introduce an auxiliary sequence.

Definition 3.21.

We define integers T(n,k)T(n,k) for 0kn0\leq k\leq n by

  • T(n,0)=0T(n,0)=0 for n1n\geq 1,

  • T(n,n)=1T(n,n)=1 for n0n\geq 0,

  • T(n,k)=T(n1,k1)+k2T(n1,k)T(n,k)=T(n-1,k-1)+k^{2}T(n-1,k) for 1kn11\leq k\leq n-1.

The values T(n,k)T(n,k) maybe found at [5, A036969].

Proposition 3.22.

For all integers 0kn0\leq k\leq n, we have h2n,k=(1)k(2k)!T(n,k)h_{2n,k}=(-1)^{k}(2k)!\,T(n,k).

The significance of Proposition 3.22 is that it reduces the computation of ht,ch_{t,c} in step 4 of the algorithm from cubic time to quadratic time. Let S(n,k)=(1)kh2n,k/2kS(n,k)=(-1)^{k}h_{2n,k}/2^{k}, which by Proposition 3.19 counts the number of ordered set-partitions of {1,,2n}\{1,\ldots,2n\} into kk parts of even cardinality. The proof of Proposition 3.22 boils down to showing S(n,k)S(n,k) satisfies the recurrence S(n,k)=k2S(n1,k)+k(2k1)S(n1,k1)S(n,k)=k^{2}S(n-1,k)+k(2k-1)S(n-1,k-1). This recurrence is known [5, A241171]. We include a proof of Proposition 3.22 for completeness.

Proof of Proposition 3.22.

Let S(n,k)S(n,k) be defined as above. It suffices to show that S(n,k)=(2k)!T(n,k)/2kS(n,k)=(2k)!\,T(n,k)/2^{k}, because then h2n,k=(1)k2kS(n,k)=(1)k(2k)!T(n,k)h_{2n,k}=(-1)^{k}2^{k}S(n,k)=(-1)^{k}(2k)!\,T(n,k). We need to prove the following three properties:

  1. 1.

    S(n,0)=0S(n,0)=0 for n1n\geq 1.

  2. 2.

    S(n,n)=(2n)!/2nS(n,n)=(2n)!/2^{n} for n0n\geq 0.

  3. 3.

    S(n,k)=k2S(n1,k)+k(2k1)S(n1,k1)S(n,k)=k^{2}S(n-1,k)+k(2k-1)S(n-1,k-1) for 1kn11\leq k\leq n-1.

The first property follows from the observation that if n1n\geq 1, then there are no ordered set-partitions of {1,,2n}\{1,\ldots,2n\} into 0 parts. For the second property, note that an ordered set-partition of {1,,2n}\{1,\ldots,2n\} into nn parts of even cardinality consists of pairing up {1,,2n}\{1,\ldots,2n\} into an ordered sequence of nn unordered pairs. There are (2n)!/2n(2n)!/2^{n} such pairings. It remains to show the third property (the recurrence). Consider the following two ways of constructing an ordered set-partition of {1,,2n}\{1,\ldots,2n\} into kk parts of even cardinality:

  • Start with an ordered set-partition {1,,2n2}=S1Sk\{1,\ldots,2n-2\}=S_{1}\cup\cdots\cup S_{k} into kk parts of even cardinality, and choose two (possibly equal) indices i,j{1,,k}i,j\in\{1,\ldots,k\}. Consider the smallest element of the union SiSjS_{i}\cup S_{j} and flip its position (if it was in SiS_{i}, then move it to SjS_{j}, and vice versa). Finally, add 2n12n-1 to SiS_{i} and add 2n2n to SjS_{j}.

  • Start with an ordered set-partition {1,,2n2}=S1Sk1\{1,\ldots,2n-2\}=S_{1}\cup\cdots\cup S_{k-1} into k1k-1 parts of even cardinality, and choose an index k0{1,,k}k_{0}\in\{1,\ldots,k\}. Inserting the empty set at position k0k_{0} and relabeling gives {1,,2n2}=S1Sk\{1,\ldots,2n-2\}=S_{1}\cup\cdots\cup S_{k}, which would be an ordered set-partition except that Sk0S_{k_{0}} is empty. Now choose two (possibly equal) indices i,j{1,,k}i,j\in\{1,\ldots,k\}, at least one of which is equal to k0k_{0}. Consider the smallest element of the union SiSjS_{i}\cup S_{j} and flip its position (if i=j=k0i=j=k_{0}, then do nothing). Finally, add 2n12n-1 to SiS_{i} and add 2n2n to SjS_{j}.

We claim that every ordered set-partition of {1,,2n}\{1,\ldots,2n\} into kk parts of even cardinality arises uniquely from one of these two constructions. This is because the process can be reversed:

  • Start with an ordered set-partition of {1,,2n}=S1Sk\{1,\ldots,2n\}=S_{1}\cup\cdots\cup S_{k} into kk parts of even cardinality. Let SiS_{i} be the set containing 2n12n-1 and let SjS_{j} be the set containing 2n2n. Remove 2n12n-1 from SiS_{i} and remove 2n2n from SjS_{j}. Consider the smallest element of the union SiSjS_{i}\cup S_{j} and flip its position (if Si=Sj=S_{i}=S_{j}=\varnothing, then do nothing). If SiS_{i} and SjS_{j} are still nonempty, then we are in the first case. If either SiS_{i} or SjS_{j} is now empty, then we are in the second case.

In the first case, there are S(n1,k)S(n-1,k) ways of choosing the initial ordered set-partition, and k2k^{2} ways of choosing the indices i,ji,j. In the second case, there are S(n1,k1)S(n-1,k-1) ways of choosing the initial ordered set-partition, kk ways of choosing the index k0k_{0}, and 2k12k-1 ways of choosing the indices i,ji,j. This proves the recurrence S(n,k)=k2S(n1,k)+k(2k1)S(n1,k1)S(n,k)=k^{2}S(n-1,k)+k(2k-1)S(n-1,k-1). ∎

4 Asymptotics

4.1 Asymptotics of fn,kf_{n,k}

We first give a formula for the generalized Fubini numbers fn,kf_{n,k} in terms of the ordinary Fubini numbers fnf_{n}.

Proposition 4.1.

If k0k\geq 0 and n0n\geq 0 are integers, then

fn+k,k=k!n0++nk=nn!n0!nk!fn0fnk,f_{n+k,k}=k!\sum_{n_{0}+\cdots+n_{k}=n}\frac{n!}{n_{0}!\cdots n_{k}!}f_{n_{0}}\cdots f_{n_{k}},

where each nin_{i} is a nonnegative integer.

Proof.

Recall that fn+k,kf_{n+k,k} counts the number of weak orders on {1,,n+k}\{1,\ldots,n+k\} where each of the elements 1,,k1,\ldots,k must be placed in a part by itself. We first choose the permutation of the elements 1,,k1,\ldots,k. This leaves k+1k+1 regions between the elements 1,,k1,\ldots,k where we must place the remaining nn elements k+1,,n+kk+1,\ldots,n+k. We then choose how many elements nin_{i} should go into each of the k+1k+1 regions. There are n!n0!nk!\frac{n!}{n_{0}!\cdots n_{k}!} ways to choose how to distribute the nn elements k+1,,n+kk+1,\ldots,n+k into these k+1k+1 regions. Finally, there are fnif_{n_{i}} choices for the weak order on the nin_{i} elements within each of the k+1k+1 regions. ∎

We remark that Proposition 4.1 gives a generating function identity

n=0fn+k,kn!xn=k!(n=0fnn!xn)k+1=k!(2ex)k+1.\sum_{n=0}^{\infty}\frac{f_{n+k,k}}{n!}x^{n}=k!\left(\sum_{n=0}^{\infty}\frac{f_{n}}{n!}x^{n}\right)^{k+1}=\frac{k!}{(2-e^{x})^{k+1}}.

It is possible to prove Lemma 4.3 below using complex-analytic generating function techniques by looking at the poles of k!/(2ex)k+1k!/(2-e^{x})^{k+1}. However, we will prove Lemma 4.3 using more direct real-analytic techniques. We will need the following technical lemma.

Lemma 4.2.

Let {an}n=0\{a_{n}\}_{n=0}^{\infty} be a sequence of real numbers and let k0k\geq 0 be an integer. If an1a_{n}\to 1, then

1(n+kk)n0++nk=nall ni0an0ank1\frac{1}{\binom{n+k}{k}}\sum_{\begin{subarray}{c}n_{0}+\cdots+n_{k}=n\\ \text{all }n_{i}\geq 0\end{subarray}}a_{n_{0}}\cdots a_{n_{k}}\to 1

as nn\to\infty.

Proof.

We first remark that (n+kk)\binom{n+k}{k} is the number of terms in the sum, since the number of solutions of n0++nk=nn_{0}+\cdots+n_{k}=n equals the number of ways to distribute nn identical balls into k+1k+1 distinguishable urns. Let ε>0\varepsilon>0. Then there exists an m0m\geq 0 such that an[1ε,1+ε]a_{n}\in[1-\varepsilon,1+\varepsilon] for nmn\geq m. Let C=max|an|C=\max\left\lvert{a_{n}}\right\rvert. For n(k+1)mn\geq(k+1)m, we will split up the sum as

n0++nk=nan0ank=n0++nk=nall niman0ank+n0++nk=nsome ni<man0ank.\sum_{n_{0}+\cdots+n_{k}=n}a_{n_{0}}\cdots a_{n_{k}}=\sum_{\begin{subarray}{c}n_{0}+\cdots+n_{k}=n\\ \text{all }n_{i}\geq m\end{subarray}}a_{n_{0}}\cdots a_{n_{k}}+\sum_{\begin{subarray}{c}n_{0}+\cdots+n_{k}=n\\ \text{some }n_{i}<m\end{subarray}}a_{n_{0}}\cdots a_{n_{k}}.

The number of terms of the first sum equals the number of solutions to n0++nk=nn_{0}+\cdots+n_{k}=n with all nimn_{i}\geq m, which equals the number of solutions to n0++nk=n(k+1)mn_{0}+\cdots+n_{k}=n-(k+1)m. Then the first sum has (n(k+1)m+kk)\binom{n-(k+1)m+k}{k} terms (by the same argument as at the start of the proof), each of which lies in the interval [(1ε)k+1,(1+ε)k+1][(1-\varepsilon)^{k+1},(1+\varepsilon)^{k+1}]. The second sum has (n+kk)(n(k+1)m+kk)\binom{n+k}{k}-\binom{n-(k+1)m+k}{k} terms, each of which lies in the interval [Ck+1,Ck+1][-C^{k+1},C^{k+1}]. Thus,

(1ε)k+1(n(k+1)m+kk)Ck+1((n+kk)(n(k+1)m+kk))n0++nk=nan0ank(1-\varepsilon)^{k+1}\binom{n-(k+1)m+k}{k}-C^{k+1}\left(\binom{n+k}{k}-\binom{n-(k+1)m+k}{k}\right)\leq\sum_{n_{0}+\cdots+n_{k}=n}a_{n_{0}}\cdots a_{n_{k}}

and

n0++nk=nan0ank(1+ε)k+1(n(k+1)m+kk)+Ck+1((n+kk)(n(k+1)m+kk)).\sum_{n_{0}+\cdots+n_{k}=n}a_{n_{0}}\cdots a_{n_{k}}\leq(1+\varepsilon)^{k+1}\binom{n-(k+1)m+k}{k}+C^{k+1}\left(\binom{n+k}{k}-\binom{n-(k+1)m+k}{k}\right).

Dividing through by (n+kk)\binom{n+k}{k} gives the inequalities

(1ε)k+1(n(k+1)m+kk)(n+kk)Ck+1(1(n(k+1)m+kk)(n+kk))1(n+kk)n0++nk=nan0ank(1-\varepsilon)^{k+1}\frac{\binom{n-(k+1)m+k}{k}}{\binom{n+k}{k}}-C^{k+1}\left(1-\frac{\binom{n-(k+1)m+k}{k}}{\binom{n+k}{k}}\right)\leq\frac{1}{\binom{n+k}{k}}\sum_{n_{0}+\cdots+n_{k}=n}a_{n_{0}}\cdots a_{n_{k}}

and

1(n+kk)n0++nk=nan0ank(1+ε)k+1(n(k+1)m+kk)(n+kk)+Ck+1(1(n(k+1)m+kk)(n+kk)).\frac{1}{\binom{n+k}{k}}\sum_{n_{0}+\cdots+n_{k}=n}a_{n_{0}}\cdots a_{n_{k}}\leq(1+\varepsilon)^{k+1}\frac{\binom{n-(k+1)m+k}{k}}{\binom{n+k}{k}}+C^{k+1}\left(1-\frac{\binom{n-(k+1)m+k}{k}}{\binom{n+k}{k}}\right).

Note that

(n(k+1)m+kk)(n+kk)=(n(k+1)m+1)(n(k+1)m+k)(n+1)(n+k)1 as n.\frac{\binom{n-(k+1)m+k}{k}}{\binom{n+k}{k}}=\frac{(n-(k+1)m+1)\cdots(n-(k+1)m+k)}{(n+1)\cdots(n+k)}\to 1\text{ as }n\to\infty.

Then taking nn\to\infty gives

(1ε)k+1lim infn1(n+kk)n0++nk=nan0anklim supn1(n+kk)n0++nk=nan0ank(1+ε)k+1.\displaystyle(1-\varepsilon)^{k+1}\leq\liminf_{n\to\infty}\frac{1}{\binom{n+k}{k}}\sum_{n_{0}+\cdots+n_{k}=n}a_{n_{0}}\cdots a_{n_{k}}\leq\limsup_{n\to\infty}\frac{1}{\binom{n+k}{k}}\sum_{n_{0}+\cdots+n_{k}=n}a_{n_{0}}\cdots a_{n_{k}}\leq(1+\varepsilon)^{k+1}.

The result follows from taking ε0\varepsilon\to 0. ∎

We can now give an asymptotic formula for fn,kf_{n,k} as nn\to\infty. Recall that f(n)g(n)f(n)\sim g(n) means limnf(n)g(n)=1\lim\limits_{n\to\infty}\frac{f(n)}{g(n)}=1.

Lemma 4.3.

For each fixed nonnegative integer kk, we have

fn,kn!2k+1(log2)n+1f_{n,k}\sim\frac{n!}{2^{k+1}(\log 2)^{n+1}}

as nn\to\infty.

Proof.

Proposition 4.1 gives the combinatorial identity

fn+k,k=k!n0++nk=nn!n0!nk!fn0fnkf_{n+k,k}=k!\sum_{n_{0}+\cdots+n_{k}=n}\frac{n!}{n_{0}!\cdots n_{k}!}f_{n_{0}}\cdots f_{n_{k}}

which we can rewrite as

fn+k,k((n+k)!2k+1(log2)n+k+1)=1(n+kk)n0++nk=nfn0(n0!2(log2)n0+1)fnk(nk!2(log2)nk+1).\frac{f_{n+k,k}}{\displaystyle\left(\frac{(n+k)!}{2^{k+1}(\log 2)^{n+k+1}}\right)}=\frac{1}{\displaystyle\binom{n+k}{k}}\sum_{n_{0}+\cdots+n_{k}=n}\frac{f_{n_{0}}}{\displaystyle\left(\frac{n_{0}!}{2(\log 2)^{n_{0}+1}}\right)}\cdots\frac{f_{n_{k}}}{\displaystyle\left(\frac{n_{k}!}{2(\log 2)^{n_{k}+1}}\right)}.

Then the result follows from Lemma 4.2, along with the asymptotic formula fnn!2(log2)n+1f_{n}\,{\sim}\,\frac{n!}{2(\log 2)^{n+1}} [9, p.175-176]. ∎

4.2 Asymptotics of qnq_{n}

We will need a short technical lemma.

Lemma 4.4.

For all integers 02ctn0\leq 2c\leq t\leq n,

(nt)((nt+c)!n!)21t!.\binom{n}{t}\left(\frac{(n-t+c)!}{n!}\right)^{2}\leq\frac{1}{t!}.
Proof.

We have

(nt)((nt+c)!n!)2=1t!n!(nt)!((nt+c)!n!)2=1t!(nt+c)!2(nt)!n!,\binom{n}{t}\left(\frac{(n-t+c)!}{n!}\right)^{2}=\frac{1}{t!}\frac{n!}{(n-t)!}\left(\frac{(n-t+c)!}{n!}\right)^{2}=\frac{1}{t!}\frac{(n-t+c)!^{2}}{(n-t)!\,n!},

where

(nt+c)!2(nt)!n!(nt+c)!(nt)!(nt+c)!(nt+2c)!=nt+1nt+c+1nt+cnt+2c1.\frac{(n-t+c)!^{2}}{(n-t)!\,n!}\leq\frac{(n-t+c)!}{(n-t)!}\frac{(n-t+c)!}{(n-t+2c)!}=\frac{n-t+1}{n-t+c+1}\cdots\frac{n-t+c}{n-t+2c}\leq 1.\qed
Theorem 4.5.

We have

qne(log2)2qn,0e(log2)2n!24(log2)2n+2.q_{n}\sim e^{-(\log 2)^{2}}q_{n,0}\sim e^{-(\log 2)^{2}}\frac{n!^{2}}{4\,(\log 2)^{2n+2}}.

Thus, approximately e(log2)2e^{-(\log 2)^{2}} (around 62%62\%) of pairs of weak orders have no consecutive embeddings.

Proof.

Taking the formula qn=ct(nt)ht,cgnt+c,cq_{n}=\sum_{c}\sum_{t}\binom{n}{t}h_{t,c}g_{n-t+c,c} in equation (2), unfolding the definition of gnt+c,cg_{n-t+c,c}, and dividing through by fn2f_{n}^{2} gives

qnfn2=c=0n/2t=2cnj=0c(nt)ht,cj!(cj)!fnt+c,jfnt+c,cjfn2.\frac{q_{n}}{f_{n}^{2}}=\sum_{c=0}^{\lfloor n/2\rfloor}\sum_{t=2c}^{n}\sum_{j=0}^{c}\binom{n}{t}\frac{h_{t,c}}{j!\,(c-j)!}\frac{f_{n-t+c,j}f_{n-t+c,c-j}}{f_{n}^{2}}.

Now fix c0c\geq 0, t2ct\geq 2c, and 0jc0\leq j\leq c, and consider the summand

(nt)ht,cj!(cj)!fnt+c,jfnt+c,cjfn2\binom{n}{t}\frac{h_{t,c}}{j!\,(c-j)!}\frac{f_{n-t+c,j}f_{n-t+c,c-j}}{f_{n}^{2}}

for ntn\geq t as nn\to\infty. By the asymptotic formula for fn,kf_{n,k} in Lemma 4.3 and the fact that fn=fn,0f_{n}=f_{n,0} we have

(nt)ht,cj!(cj)!fnt+c,jfnt+c,cjfn2\displaystyle\binom{n}{t}\frac{h_{t,c}}{j!\,(c-j)!}\frac{f_{n-t+c,j}f_{n-t+c,c-j}}{f_{n}^{2}} ntt!ht,cj!(cj)!((nt+c)!2j+1(log2)nt+c+1)((nt+c)!2cj+1(log2)nt+c+1)(n!2(log2)n+1)2\displaystyle\sim\frac{n^{t}}{t!}\frac{h_{t,c}}{j!\,(c-j)!}\frac{\displaystyle\left(\frac{(n-t+c)!}{2^{j+1}(\log 2)^{n-t+c+1}}\right)\left(\frac{(n-t+c)!}{2^{c-j+1}(\log 2)^{n-t+c+1}}\right)}{\displaystyle\left(\frac{n!}{2(\log 2)^{n+1}}\right)^{2}}
ntt!ht,cj!(cj)!(log2)2t2c2cn2t2cht,cj!(cj)!(log2)2t2c2ct!1nt2c.\displaystyle\sim\frac{n^{t}}{t!}\frac{h_{t,c}}{j!\,(c-j)!}\frac{(\log 2)^{2t-2c}}{2^{c}n^{2t-2c}}\sim\frac{h_{t,c}}{j!\,(c-j)!}\frac{(\log 2)^{2t-2c}}{2^{c}t!}\frac{1}{n^{t-2c}}.

If t>2ct>2c, then the summand converges to 0. If t=2ct=2c, then ht,c=(1)ct!h_{t,c}=(-1)^{c}t! (this follows from the definition of ht,ch_{t,c}, as well as from Proposition 3.22) so the summand converges to

(1)c1j!(cj)!(log2)2c2c.(-1)^{c}\frac{1}{j!\,(c-j)!}\frac{(\log 2)^{2c}}{2^{c}}.

After applying dominated convergence theorem (justified below), the t>2ct>2c terms contribute nothing so we can drop the sum over tt and get

qnfn2c=0j=0c(1)c1j!(cj)!(log2)2c2c=c=0(1)c(log2)2cc!=e(log2)2.\frac{q_{n}}{f_{n}^{2}}\to\sum_{c=0}^{\infty}\sum_{j=0}^{c}(-1)^{c}\frac{1}{j!\,(c-j)!}\frac{(\log 2)^{2c}}{2^{c}}=\sum_{c=0}^{\infty}(-1)^{c}\frac{(\log 2)^{2c}}{c!}=e^{-(\log 2)^{2}}.

The result follows from the identity qn,0=fn2q_{n,0}=f_{n}^{2} and the asymptotic formula fnn!2(log2)n+1f_{n}\sim\frac{n!}{2(\log 2)^{n+1}} (Lemma 4.3).

It remains to justify this application of the dominated convergence theorem. Note that the asymptotic formula fnn!2(log2)n+1f_{n}\sim\frac{n!}{2(\log 2)^{n+1}} gives a constant CC (not depending on cc, tt, jj) such that we have

|(nt)ht,cj!(cj)!fnt+c,jfnt+c,cjfn2|\displaystyle\left\lvert\binom{n}{t}\frac{h_{t,c}}{j!\,(c-j)!}\frac{f_{n-t+c,j}f_{n-t+c,c-j}}{f_{n}^{2}}\right\rvert =|ht,c|j!(cj)!(nt)fnt+c,jfnt+c,cjfn2\displaystyle=\frac{\left\lvert{h_{t,c}}\right\rvert}{j!\,(c-j)!}\binom{n}{t}\frac{f_{n-t+c,j}f_{n-t+c,c-j}}{f_{n}^{2}}
|ht,c|j!(cj)!(nt)(fnt+cfn)2\displaystyle\leq\frac{\left\lvert{h_{t,c}}\right\rvert}{j!\,(c-j)!}\binom{n}{t}\left(\frac{f_{n-t+c}}{f_{n}}\right)^{2}
C|ht,c|j!(cj)!(nt)((nt+c)!(log2)nt+cn!(log2)n)2\displaystyle\leq C\frac{\left\lvert{h_{t,c}}\right\rvert}{j!\,(c-j)!}\binom{n}{t}\left(\frac{\displaystyle\frac{(n-t+c)!}{(\log 2)^{n-t+c}}}{\displaystyle\frac{n!}{(\log 2)^{n}}}\right)^{2}
C|ht,c|j!(cj)!(nt)((nt+c)!n!)2\displaystyle\leq C\frac{\left\lvert{h_{t,c}}\right\rvert}{j!\,(c-j)!}\binom{n}{t}\left(\frac{(n-t+c)!}{n!}\right)^{2}
C|ht,c|j!(cj)!1t!,\displaystyle\leq C\frac{\left\lvert{h_{t,c}}\right\rvert}{j!\,(c-j)!}\frac{1}{t!},

where the last inequality uses Lemma 4.4. It remains to show that the sum

c=0t=2cj=0c|ht,c|j!(cj)!1t!\sum_{c=0}^{\infty}\sum_{t=2c}^{\infty}\sum_{j=0}^{c}\frac{\left\lvert{h_{t,c}}\right\rvert}{j!\,(c-j)!}\frac{1}{t!}

is finite. Applying Corollary 3.20 and the inequality 2ct2c\leq t from the bounds of summation gives

c=0t=2cj=0c|ht,c|j!(cj)!1t!\displaystyle\sum_{c=0}^{\infty}\sum_{t=2c}^{\infty}\sum_{j=0}^{c}\frac{\left\lvert{h_{t,c}}\right\rvert}{j!\,(c-j)!}\frac{1}{t!} =c=0t=2c2cc!t!|ht,c|c=0t=2c22cctc!t!c=0t=2c(2c)tc!t!c=0t=0(2c)tc!t!=c=0e2cc!\displaystyle=\sum_{c=0}^{\infty}\sum_{t=2c}^{\infty}\frac{2^{c}}{c!\,t!}\left\lvert{h_{t,c}}\right\rvert\leq\sum_{c=0}^{\infty}\sum_{t=2c}^{\infty}\frac{2^{2c}c^{t}}{c!\,t!}\leq\sum_{c=0}^{\infty}\sum_{t=2c}^{\infty}\frac{(2c)^{t}}{c!\,t!}\leq\sum_{c=0}^{\infty}\sum_{t=0}^{\infty}\frac{(2c)^{t}}{c!\,t!}=\sum_{c=0}^{\infty}\frac{e^{2c}}{c!}

which is finite by the ratio test. ∎

4.3 Asymptotics of pnp_{n}

We will need the following technical lemma.

Lemma 4.6.

For each fixed nonnegative integer kk, we have

[nnk]n2k2kk!\genfrac{[}{]}{0.0pt}{}{n}{n-k}\sim\frac{n^{2k}}{2^{k}k!}

as nn\to\infty. We also have

[nnk]n2k2kk!\genfrac{[}{]}{0.0pt}{}{n}{n-k}\leq\frac{n^{2k}}{2^{k}k!}

for all nkn\geq k.

Proof.

The first statement follows from equation 1.6 of [4]. For the second statement, we will use the recurrence for the Stirling numbers of the first kind. If k=0k=0 or k=nk=n, then the inequality is clear. Now suppose that 1kn11\leq k\leq n-1, and inductively assume that the inequality holds for smaller values of nn. Then

[nnk]\displaystyle\genfrac{[}{]}{0.0pt}{}{n}{n-k} =(n1)[n1nk]+[n1nk1]\displaystyle=(n-1)\genfrac{[}{]}{0.0pt}{}{n-1}{n-k}+\genfrac{[}{]}{0.0pt}{}{n-1}{n-k-1}
=(n1)[n1(n1)(k1)]+[n1(n1)k]\displaystyle=(n-1)\genfrac{[}{]}{0.0pt}{}{n-1}{(n-1)-(k-1)}+\genfrac{[}{]}{0.0pt}{}{n-1}{(n-1)-k}
(n1)(n1)2(k1)2k1(k1)!+(n1)2k2kk!\displaystyle\leq(n-1)\frac{(n-1)^{2(k-1)}}{2^{k-1}(k-1)!}+\frac{(n-1)^{2k}}{2^{k}k!}
=12kk!(2k+n1)(n1)2k1\displaystyle=\frac{1}{2^{k}k!}(2k+n-1)(n-1)^{2k-1}
12kk!((2k+n1)+(2k1)(n1)2k)2k\displaystyle\leq\frac{1}{2^{k}k!}\left(\frac{(2k+n-1)+(2k-1)(n-1)}{2k}\right)^{2k}
=n2k2kk!,\displaystyle=\frac{n^{2k}}{2^{k}k!},

where the last inequality (the penultimate step) uses the AM-GM inequality. ∎

We can now prove the asymptotic formula for pnp_{n} (Theorem 1.2 from the introduction).

Theorem 4.7.

We have

pne(log2)2/2fn2n!e(log2)2/24(log2)2n!(log2)2n.p_{n}\sim e^{-(\log 2)^{2}/2}\frac{f_{n}^{2}}{n!}\sim\frac{e^{-(\log 2)^{2}/2}}{4(\log 2)^{2}}\frac{n!}{(\log 2)^{2n}}.
Proof.

Applying the formula pn=1n!k=0n[nk]qk\displaystyle p_{n}=\frac{1}{n!}\sum\limits_{k=0}^{n}\genfrac{[}{]}{0.0pt}{}{n}{k}q_{k} and replacing kk with nkn-k gives

n!pnfn2=k=0n[nk]qkfn2=k=0n[nnk]qnkfn2.\frac{n!\,p_{n}}{f_{n}^{2}}=\sum_{k=0}^{n}\genfrac{[}{]}{0.0pt}{}{n}{k}\frac{q_{k}}{f_{n}^{2}}=\sum_{k=0}^{n}\genfrac{[}{]}{0.0pt}{}{n}{n-k}\frac{q_{n-k}}{f_{n}^{2}}.

If kk is fixed, then applying Lemma 4.3, Theorem 4.5, and Lemma 4.6 gives

[nnk]qnkfn2n2k2kk!e(log2)2(nk)!24(log2)2(nk)+2(n!2(log2)n+1)2e(log2)2(log2)2k2kk!.\genfrac{[}{]}{0.0pt}{}{n}{n-k}\frac{q_{n-k}}{f_{n}^{2}}\sim\frac{n^{2k}}{2^{k}k!}\frac{\displaystyle e^{-(\log 2)^{2}}\frac{(n-k)!^{2}}{4(\log 2)^{2(n-k)+2}}}{\displaystyle\left(\frac{n!}{2(\log 2)^{n+1}}\right)^{2}}\sim\frac{e^{-(\log 2)^{2}}(\log 2)^{2k}}{2^{k}k!}.

By the dominated convergence theorem (justified below),

n!pnfn2k=0e(log2)2(log2)2k2kk!=e(log2)2e(log2)2/2=e(log2)2/2.\frac{n!\,p_{n}}{f_{n}^{2}}\to\sum_{k=0}^{\infty}\frac{e^{-(\log 2)^{2}}(\log 2)^{2k}}{2^{k}k!}=e^{-(\log 2)^{2}}e^{(\log 2)^{2}/2}=e^{-(\log 2)^{2}/2}.

Then the result follows from the asymptotic formula fnn!2(log2)n+1f_{n}\sim\frac{n!}{2(\log 2)^{n+1}}.

It remains to justify this application of the dominated convergence theorem. Note that the asymptotics fnn!2(log2)n+1f_{n}\sim\frac{n!}{2(\log 2)^{n+1}} and qne(log2)2n!4(log2)2n+2q_{n}\sim e^{-(\log 2)^{2}}\frac{n!}{4(\log 2)^{2n+2}} give a constant CC (not depending on nn or kk) such that we have

[nnk]qnkfn2Cn2k2kk!(nk)!2(log2)2(nk)(n!(log2)n)2Cn2k2kk!(nk)!2n!2\genfrac{[}{]}{0.0pt}{}{n}{n-k}\frac{q_{n-k}}{f_{n}^{2}}\leq C\frac{n^{2k}}{2^{k}k!}\frac{\displaystyle\frac{(n-k)!^{2}}{(\log 2)^{2(n-k)}}}{\displaystyle\left(\frac{n!}{(\log 2)^{n}}\right)^{2}}\leq C\frac{n^{2k}}{2^{k}k!}\frac{(n-k)!^{2}}{n!^{2}}

for nkn\geq k. Now view the expression

n2k(nk)!2n!2=((nnk+1)(nnk+2)(nn))2n^{2k}\frac{(n-k)!^{2}}{n!^{2}}=\left(\left(\frac{n}{n-k+1}\right)\left(\frac{n}{n-k+2}\right)\cdots\left(\frac{n}{n}\right)\right)^{2}

as a function of nkn\geq k for fixed kk. This function has a fixed number of terms, each of which is decreasing as a function of nn. In particular, this function is maximized at n=kn=k with value k2k/k!2k^{2k}/k!^{2}. This shows that

[nnk]qnkfn2Cn2k2kk!(nk)!2n!2Ck2k2kk!3\genfrac{[}{]}{0.0pt}{}{n}{n-k}\frac{q_{n-k}}{f_{n}^{2}}\leq C\frac{n^{2k}}{2^{k}k!}\frac{(n-k)!^{2}}{n!^{2}}\leq C\frac{k^{2k}}{2^{k}k!^{3}}

for nkn\geq k. Finally, the sum

k=0k2k2kk!3\sum_{k=0}^{\infty}\frac{k^{2k}}{2^{k}k!^{3}}

is finite by the ratio test. ∎

5 Further Directions

5.1 Higher Order Asymptotics

Let KK denote the constant

K=e(log2)2/24(log2)20.409223.K=\frac{e^{-(\log 2)^{2}/2}}{4(\log 2)^{2}}\approx 0.409223.

Theorem 4.7 states that

pn(log2)2nn!K.\frac{p_{n}(\log 2)^{2n}}{n!}\to K.

Figure 4 is a graph of y=log(Kpn(log2)2nn!)y=\log\big{(}K-\frac{p_{n}(\log 2)^{2n}}{n!}\big{)} against x=lognx=\log n for the values of nn given in the appendix.

Refer to caption
Figure 4: Graph of y=log(Kpn(log2)2nn!)y=\log\big{(}K-\frac{p_{n}(\log 2)^{2n}}{n!}\big{)} against x=lognx=\log n

For large nn, the points appear to approach a line with slope 1-1 and yy-intercept b2.23b\approx-2.23. In other words,

log(Kpn(log2)2nn!)blogn.\log\left(K-\frac{p_{n}(\log 2)^{2n}}{n!}\right)\approx b-\log n.

Exponentiating and rearranging terms gives

pn(log2)2nn!Kebn.\frac{p_{n}(\log 2)^{2n}}{n!}\approx K-\frac{e^{b}}{n}.

This suggests the following conjecture.

Conjecture 5.1.

There exists a constant c>0c>0 such that

pn(log2)2nn!=Kcn+O(1n2).\frac{p_{n}(\log 2)^{2n}}{n!}=K-\frac{c}{n}+O\left(\frac{1}{n^{2}}\right).

5.2 Congruence Conjecture

Theorem 2 of [7] states that if pp is prime and nmn\geq m, then

fn+φ(pm)fn(modpm),f_{n+\varphi(p^{m})}\equiv f_{n}\pmod{p^{m}},

where φ\varphi is Euler’s totient function. In other words, the Fubini numbers are periodic modulo pmp^{m}. Squaring both sides gives the congruence

qn+φ(pm),0qn,0(modpm)q_{n+\varphi(p^{m}),0}\equiv q_{n,0}\pmod{p^{m}}

for pp prime and nmn\geq m. Replacing qn,0q_{n,0} with qnq_{n} suggests the following conjecture, which is supported by the computed values of qnq_{n} for n5000n\leq 5000.

Conjecture 5.2.

If pp is prime and nmn\geq m, then

qn+φ(pm)qn(modpm).q_{n+\varphi(p^{m})}\equiv q_{n}\pmod{p^{m}}.

Unfortunately, this conjecture does not appear to imply anything about the behavior of the sequence {pn}\{p_{n}\} modulo pmp^{m}, because of the division by n!n! when converting from {qn}\{q_{n}\} to {pn}\{p_{n}\}.

6 Acknowledgements

Many thanks to Sara Billey for proposing the problem and for helpful discussions, and to the WXML (Washington Experimental Mathematics Lab) for providing the project that led to this paper. Also thanks to Sara Billey and the anonymous referees for their comments and suggestions on this paper.

7 Web Resources

The basic and memory-optimized implementations of the algorithm described in section 3.6 and the values of pnp_{n} for n5000n\leq 5000 can be found at

https://github.com/tb65536/ParabolicDoubleCosets

8 Appendix: Values of pnp_{n}

nn pnp_{n} (pnlog2n2)/n!(p_{n}\log^{2n}2)/n!
1 1 (1 digits) 0.480453
2 3 (1 digits) 0.346253
3 19 (2 digits) 0.3512
4 167 (3 digits) 0.370774
5 1791 (4 digits) 0.382093
6 22715 (5 digits) 0.388048
7 334031 (6 digits) 0.391663
8 5597524 (7 digits) 0.394169
9 105351108 (9 digits) 0.396036
10 2200768698 (10 digits) 0.397485
11 50533675542 (11 digits) 0.398644
12 1265155704413 (13 digits) 0.399593
13 34300156146805 (14 digits) 0.400385
14 1001152439025205 (16 digits) 0.401056
15 31301382564128969 (17 digits) 0.401631
16 1043692244938401836 (19 digits) 0.402131
17 36969440518414369896 (20 digits) 0.402569
18 1386377072447199902576 (22 digits) 0.402955
19 54872494774746771827248 (23 digits) 0.403299
20 2285943548113541477123970 (25 digits) 0.403608
30 382079126820…882950534546 (42 digits) 0.405528
40 179736290098…532574927537 (61 digits) 0.406469
50 102365379120…338473199289 (81 digits) 0.407028
60 427699505826…027450945465 (101 digits) 0.407398
70 940027093836…926979570377 (122 digits) 0.407662
80 857360695445…439742054481 (144 digits) 0.407859
90 271659624624…300501685746 (167 digits) 0.408011
100 260443549181…383464403196 (190 digits) 0.408133
200 150691150471…390138470043 (439 digits) 0.40868
300 400039289653…047576602840 (710 digits) 0.408862
400 572423854465…686938545249 (996 digits) 0.408952
500 745894661762…526127432358 (1293 digits) 0.409006
600 529056570650…070570426529 (1599 digits) 0.409042
700 692359539273…658799872850 (1912 digits) 0.409068
800 150717237472…313160125048 (2232 digits) 0.409088
900 902565318506…968550812571 (2556 digits) 0.409103
1000 367762337807…336792083803 (2886 digits) 0.409115
1500 657393993927…489306609387 (4592 digits) 0.409151
2000 677187561025…781759174668 (6372 digits) 0.409169
2500 497164609537…894142980291 (8207 digits) 0.40918
3000 189293873430…434167136044 (10086 digits) 0.409187
3500 163043229993…353705274487 (12001 digits) 0.409192
4000 186384435725…985721119395 (13947 digits) 0.409196
4500 346443781530…440739293425 (15920 digits) 0.409199
5000 962766473267…951984139754 (17917 digits) 0.409201

References

  • [1] S. Billey, M. Konvalinka, T. K. Peterson, W. Slofstra, and B. E. Tenner. Parabolic Double Cosets in Coxeter Groups. The Electronic Journal of Combinatorics, 25(1), 2018.
  • [2] P. Diaconis and A. Gangolli. Rectangular Arrays with Fixed Margins. Discrete Probability and Algorithms, pages 15-41. The IMA Volumes in Mathematics and its Applications, vol 72, 1995. Springer, New York.
  • [3] M. Kobayashi. Two-Sided Structure of Double Cosets in Coxeter Groups, June 2011. Accessed online October 2018.
  • [4] L. Moser and M. Wyman. Asymptotic Development of the Stirling Numbers of the First Kind. Journal of the London Mathematical Society, 33(2):131-146, 1958.
  • [5] The On-Line Encyclopedia of Integer Sequences, published electronically at https://oeis.org, October 2018.
  • [6] T. K. Petersen. A two-sided analogue of the Coxeter complex. The Electronic Journal of Combinatorics, 25(4), 2018.
  • [7] B. Poonen. Periodicity of a Combinatorial Sequence. The Fibonacci Quarterly, 26(1), 1988.
  • [8] J. Riordan. An Introduction to Combinatorial Analysis. Princeton University Press, 1978.
  • [9] H. S. Wilf. generatingfunctionology. Accessed online on August 2020.