Towards a Combinatorial Theory of Multiple Orthogonal Polynomials

A THESIS
SUBMITTED TO THE FACULTY OF THE GRADUATE SCHOOL
OF THE UNIVERSITY OF MINNESOTA
BY

Daniel Allen Drake

IN PARTIAL FULFILLMENT OF THE REQUIREMENTS

FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY

Dennis Stanton, Advisor

August 2006

c Daniel Allen Drake, August 2006
 Acknowledgments

The combinatorics group here at the University of Minnesota has been supporting and helpful
throughout my time here. There is a genuine community of students, postdocs, and faculty, and I
have benefited greatly from it. My advisor, Dennis Stanton, has always skillfully moved between
letting me struggle and figure things out for myself, and pointing—some might say shoving—me in
the right direction.
I am grateful to Professor Ravi Janardan for agreeing to serve on my reviewing committee on
very short notice.
I probably wouldn’t have starved without them, but the Village Wok restaurant and item 164
in the vending machine in the basement of Vincent Hall were always welcome when I got hungry. At
times, I ate at the Wok so frequently I could distinguish different cooks based on the chicken hot &
sour soup, which I still recommend, along with their eggplant and the squid with black bean sauce.

i
 Contents

Acknowledgments i

List of Figures iv

Chapter 1. Introduction 1
1.1. Notation and conventions 2

Chapter 2. General multiple orthogonal polynomials 3

2.1. Usual OP and Viennot’s general combinatorial theory 3
2.2. Multiple orthogonal polynomials 7
2.3. Four-term recurrences and nearly-diagonal polynomials 9
2.4. New results 11

Chapter 3. Multiple Hermite polynomials 19

3.1. Matchings and usual Hermite polynomials 19
3.2. Definition 21
3.3. A combinatorial model 21
3.4. Identities and other results 25
3.5. Multiple Laguerre polynomials 30

Chapter 4. Nearly-diagonal Chebyshev polynomials 34

4.1. Definition 34
4.2. Equivalence of moments of nearly-diagonal and x-axis Chebyshevs 38
4.3. Orthogonality and integrals with respect to second moments 43
4.4. Generating functions and other identities 45
4.5. The corresponding MOP 48
4.6. Unanswered questions and future directions 55

Chapter 5. Associated Hermite polynomials 57

5.1. Definition and orthogonality 57
5.2. Identities 61
5.3. Associated Hermite moments and oscillating tableaux 65
5.4. Associated Hermite moments and rooted maps 67
5.5. Unanswered questions and future directions 73

Bibliography 75

Index 79
ii
 List of Figures

2.1.1 A paving of [n] of weight b1 λ3 λ6 b8 x3 . 5

2.1.2 A Motzkin path of weight b0 b1 λ2 λ21 6
2.1.3 The domino/NE-SE steps case of the orthogonality sign-reversing involution. 7
2.1.4 The monomino/E step case of the orthogonality sign-reversing involution. 7
2.4.1 Possible recurrence shapes for Pn,m . 17
2.4.2 The recurrence shapes from Theorem 2.4.6 and Corollary 2.4.7. 18

3.1.1 The bijection from labeled Dyck paths to complete matchings. 20

c,d
3.3.1 A matching that contributes cd2 x2 to H5,6 (x). Edges all have weight −1. 23
3.3.2 A paired matching in the integral of H6,3 (x)H4,7 (x) with respect to the first weight. 24
3.4.1 A pair of matchings that are part of H4,7 (x)H4,7 (y). 27
3.5.1 The four ways to add a new vertex to a partial injective function from Lemma 3.5.1. 32
3.5.2 A directed graph representation of (A, B, f, g) for n = 6 and m = 7. 33

4.1.1 A matching of weight c2 d2 that is part of L1 (x10 ). 37

4.2.1 Four-term matchings of weight ±c3 d2 illustrating the core idea of the CD involution. 39
4 2
4.2.2 The CD involution and parity flip fix applied to matchings of weight ±c d . 41
4.2.3 The CD involution and parity flip fix applied to matchings of weight ±cd2 . 41
4.2.4 Two matchings of weight c4 that correspond to each other via the parity flip fix. 43
4.5.1 The iterative cancel-and-correct process of (4.5.5). 50
4.5.2 The basic idea of the sign-reversing involution applied to matchings on the left side. 54

5.1.1 A matching on 8 points using the rightmost-choice weighting. 58

5.1.2 A complete matching on 6 points under the leftmost-choice weighting for the moments. 59
5.1.3 The same complete matching under the rightmost-choice weighting. 59
5.1.4 A paired matching for n = 5 and m = 3. 60
5.1.5 The permutation 3142 as a matching. 61
5.2.1 A matching on 6 vertices of the type described by Lemma 5.2.2. 62
5.3.1 A complete matching and the corresponding oscillating tableau. 67
iii
 LIST OF FIGURES iv

5.3.2 An oscillating tableaux that corresponds to the complete matching (13)(26)(48)(57). 67

5.4.1 A rooted map to which we’ll apply the bijection to connected matchings. 69
5.4.2 The connected matching which corresponds to Figure 5.4.1. 69
5.4.3 The tail-swapping bijection applied to the connected matching of Figure 5.4.2. 70
5.4.4 A matching on 9 vertices that contributes cx to H7 (x; c). 72
 CHAPTER 1

Introduction

In this work we present the beginnings of a combinatorial theory of multiple orthogonal poly-
nomials, and also the first combinatorial treatment of associated Hermite polynomials. The goal is
a generalization of Viennot’s combinatorial theory of general orthogonal polynomials and interpre-
tations of some classical multiple orthogonal polynomials.
This thesis is a complement of sorts to Emmanuel Roblet’s thesis [53], in which he developed a
combinatorial interpretation of Padé approximants. Multiple orthogonal polynomials are intimately
related to Padé approximations for systems of functions, and so-called multicontinued fractions, or
branching continued fractions, all of which Roblet addressed. Moreover, Roblet built upon Viennot’s
work, just as we do here.
Multiple orthogonal polynomials are also known as polyorthogonal polynomials, vector orthog-
onal polynomials, and polynomials of simultaneous orthogonality. They are polynomials of a single
variable orthogonal to several measures. In chapter 2 we will discuss the definition of these poly-
nomials, the basic theorems, and present our partial results on developing a combinatorial theory.
These include a combinatorial proof of the orthogonality for P1,m (x) for any set of multiple orthog-
onal polynomials (Theorem 2.4.4); the relationship between the moments of Pn,0 (x), P0,m (x), and
the two sets of moments for the nearly-diagonal polynomials (Theorem 2.4.1); and a classification
of some of the possible recurrence relations a set of multiple orthogonal polynomials must satisfy
(Theorem 2.4.6 and Corollary 2.4.7).
Multiple Hermite polynomials have proven quite tractable, and in chapter 3 a combinatorial
interpretation of those polynomials is presented, along with a number of generating functions and
interesting identities. We derive the exponential generating function for multiple Hermite polynomi-
als with respect to an arbitrary number of weights (Theorem 3.3.1, Theorem 3.4.5) and find several
formulas for the multiple Hermites as sums of usual Hermite polynomials (equations (3.3.3), (3.3.4),
and analogues in Section 3.4.1). We state the Mehler formula for those polynomials (Theorem 3.4.2)
and provide a proof analogous to that given by Foata for usual Hermite polynomials [32]. Multiple
Laguerre polynomials are briefly addressed at the end of the chapter since we have scant results on
the combinatorics of those polynomials.
The nearly-diagonal Chebyshev polynomials are a sequence of polynomials satisfying a four-term
recurrence; they are not multiple orthogonal polynomials, but are closely related to them—they are
the “nearly-diagonal polynomials” for a set of multiple orthogonal polynomials—and we discuss them
in chapter 4. The major result of that chapter is a sign-reversing involution and weight-preserving
bijection that shows the first set of moments of the nearly-diagonal Chebyshevs are the same as the

1
 1.1. NOTATION AND CONVENTIONS 2

moments of a shifted set of usual Chebyshev polynomials (Theorem 4.2.1). We prove not only the 2-
orthogonality of those polynomials (Theorem 4.1.4) with respect to the first set of moments, but find
a formula for the integral of any two nearly-diagonal Chebyshev polynomials (Theorem 4.1.6); this
sort of result is not necessarily expected for a set of polynomials satisfying a four-term recurrence.
Similar results are also found with respect to the second moments (Theorem 4.3.2, Theorem 4.5.5,
and we also find generating functions for the polynomials and their moments.
In the last chapter, we switch gears to usual orthogonal polynomials and study the associated
Hermite polynomials. We can write the associated Hermites in terms of usual Hermites (Theo-
rem 5.2.1) and have a mixed linearization formula (Theorem 5.2.5). Much of the chapter is devoted
to weight-preserving bijections between a wide variety combinatorial objects enumerating the mo-
ments of the associated Hermites, all the way from oscillating tableaux (Theorem 5.3.1) to rooted
maps and connected matchings (Theorem 5.4.2 and Theorem 5.4.3).

1.1. Notation and conventions

We will frequently abbreviate “orthogonal polynomials” and minor variants thereof as “OP”
and “multiple orthogonal polynomials” as “MOP”.
In this work we will find it convenient to use both rising and falling factorials, which sadly are
represented by the same notation. We shall use (a)k for the rising factorial a(a + 1) · · · (a + k − 1),
and use ak for the falling factorial a(a − 1) · · · (a − k + 1).
We will use the bracket notation of Knuth for a characteristic function: [P ] equals 1 if the logical
statement P is true, and 0 if P is false. The reader may be familiar with χ(P ) which means the
same as our [P ]. This notation also generalizes, and makes easier to read, the Kronecker delta.
The traditional combinatorial notation [n] for the set of integers {1, 2, . . . , n} will also appear;
context will always make it easy to distinguish between this notation and the Knuth notation above.
We will also use [n] t [m] to mean the disjoint union of the two sets.
n


We shall interpret empty sums as zero, empty products as one, and binomial coefficients k as
zero when k is negative and n nonnegative.
The continued fraction
b0
a0 +
b1
a1 +
b2
a2 +
a3 + · · ·
will be written using the easier-to-read notation
b0 b1 b2
a0 + ···
a1 + a2 + a3 +
Finally, the figures in this document will often have black and green elements (usually edges in
graphs). The PDF version of this document will have the colors, of course, along with extensive
hyperlinking, but in a black and white print version, the green will appear gray. In this work we
will refer to, for instance, black edges and green edges—the reader should understand that green
elements of figures appear gray if reading the print version of this thesis.
 CHAPTER 2

General multiple orthogonal polynomials

In a Euclidean vector space, two vectors are orthogonal if their dot product is zero. Let’s think
of the dot product of two n-dimensional vectors this way: first, multiply them together—multiply
corresponding entries—and then “integrate” the resulting vector—add up the entries.
The subject of orthogonal polynomials arises from copying that idea over to a vector space of
polynomials. We define the inner product of two polynomials by multiplying them together and then
integrating them over some interval with respect to some measure or weight function. For example,
over the interval [−1, 1] with respect to a unit weight, the polynomials 5x3 − 3x and 3x2 − 1 are
orthogonal to each other. If we want to generate a set of polynomials which are mutually orthogonal,
we can start with an arbitrary basis for the space of polynomials, typically the monomials xk , and
perform Gram-Schmidt orthogonalization. The result is a set of polynomials {Pn (x)}∞
n=0 , with Pn (x)
of degree n, which are orthogonal.
The classic reference on orthogonal polynomials is Szegő [60]. Other good references are Chi-
hara [17], chapter 2 of Wilf [67], Ismail [41], Askey [8], and many chapters in [2].
Orthogonal polynomials arose in the 18th century in the study of celestial mechanics, and their
study continued in the 19th and most of the 20th century as creatures of real and complex analysis—
they are, for instance, very efficient at approximating a function. By the 1970’s, it was becoming
obvious that many common sets of OP, the so-called classical OP, could be thought of as generating
functions for combinatorial objects, and a number of combinatorial proofs of OP identities were
developed. This phenomenon was transformed from a curiosity to a coherent theory by Gérard
Viennot in [66, 65], where he established a combinatorial theory that encompassed any set of
orthogonal polynomials. We’ll begin by describing that theory, then define multiple orthogonal
polynomials, which generalize usual orthogonal polynomials.

2.1. Usual OP and Viennot’s general combinatorial theory

Before defining multiple orthogonal polynomials, we will in this section present a number of
basic definitions and theorems on usual orthogonal polynomials and describe the basic idea of the
combinatorial theory of general orthogonal polynomials.
Above we defined an inner product by integration against a measure or weight function, which
we will always require have total mass 1: the integral of the weight function itself should be 1.
(This is sometimes expressed by saying that we use a probability measure.) For our combinatorial
purposes, since we will only be integrating polynomials, it will suffice to work with a linear functional
L on the space of polynomials; such a functional is defined by the quantities µn := L(xn ) which are
3
 2.1. USUAL OP AND VIENNOT’S GENERAL COMBINATORIAL THEORY 4

called the moments of L. We will freely switch back and forth between notations such as
Z
f (x)g(x) dµ and L(f (x)g(x));

in this work, the integral notation is a synonym for the linear functional notation. (In the analytic
theory of OP, linear functionals are “too coarse”; they are insufficient for the types of operations and
theorems analysts wish to investigate because two distinct measures can have the same sequence of
moments.)
The fact that starts the combinatorial theory of orthogonal polynomials is that a sequence of
orthogonal polynomials always satisfies a three-term recurrence:

Theorem 2.1.1. Given L and the corresponding sequence of orthogonal polynomials, normalized
so that each polynomial is monic, there exist sequences {bn }∞ ∞
n=0 and {λn }n=1 of scalars such that,
for n ≥ 0,

(2.1.1) Pn+1 (x) = (x − bn )Pn (x) − λn Pn−1 (x),

with P0 (x) = 1 and P−1 (x) = 0.

The statement of the theorem assumes that a sequence of OP exists for any L; this certainly
isn’t true. A necessary and sufficient condition for existence is the nonvanishing of the Hankel
determinants
 
µ0 µ1 ··· µn
 µ1 µ2 · · · µn+1 
 
(2.1.2) ∆n := det   .. .. .. .. 
.

 . . . 
µn µn+1 · · · µ2n
for n ≥ 0 [17, Theorem 3.1], but in this work we shall ignore questions of existence. These determi-
nants also allow one to express the nth orthogonal polynomial as
 
µ0 µ1 · · · µn
µ1 µ2 · · · µn+1 
 
Pn (x) = det 
 .. .. .. .
.. 
. . . . 
1 x ··· xn

In analogy to real analysis, the integral of the square of a polynomial is the L2 norm of that
polynomial. The above theorem allows us to easily calculate the L2 norms of the polynomials: under
the above hypotheses (dropping the “(x)”),
2
L(xPn Pn+1 ) = L((Pn+1 + bn Pn + λn Pn−1 )Pn+1 ) = L(Pn+1 )

by using orthogonality and the three-term recurrence. On the other hand, we can associate the x
with Pn+1 , which yields

L(Pn xPn+1 ) = L(Pn (Pn+2 + bn+1 Pn+1 + λn+1 Pn )) = λn+1 L(Pn2 ).

 2.1. USUAL OP AND VIENNOT’S GENERAL COMBINATORIAL THEORY 5

By induction, we have proved that L(Pn2 ) = λ1 λ2 · · · λn L(P02 ), and since P0 = 1 and total mass is
1, we see that

(2.1.3) L(Pn2 ) = λ1 λ2 · · · λn .

The converse to Theorem 2.1.1 is true:

Theorem 2.1.2. If {Pn (x)}∞

n=0 is a sequence of polynomials satisfying a three-term recurrence
relation as in (2.1.1), with P0 (x) = 1 and P−1 (x) = 0, then there exists a unique linear functional
L with total mass 1 such that
L(Pn (x)Pm (x)) = Kn [n = m].
Furthermore, Kn is always nonzero iff all λn are nonzero.

Proof. For all n > 0, Pn (x) must be orthogonal to P0 (x) = 1, so it is necessary that L(Pn (x)) =
0 for n > 0. This condition and total mass 1 is sufficient to inductively determine the moments,
so L is uniquely determined. That L(Pn (x)Pm (x)) = 0 if 0 < m < n follows by induction from
Theorem 2.1.1 or alternatively, from Viennot’s combinatorial proof (Section 2.1.1).
Above we calculated that the L2 norm of Pn is λ1 λ2 · · · λn , which shows that Kn is nonzero for
all n iff λn is likewise nonzero. 

The early development of the combinatorics of orthogonal polynomials was done by Gérard
Viennot [65, 66]. He defined several weighted combinatorial objects: the first is a paving of [n],
which is a set of nonoverlapping dominos and monominos that cover a subset of [n]. A domino
starting at k has weight −λk ; a monomino at k + 1 has weight −bk ; an “empty” point has weight x;
and finally, the weight of an entire paving is the product of the weights of its dominos, monominos,
and fixed points. See Figure 2.1.1. By convention, the empty paving will have weight 1.

x −b1 −λ3 x −λ6 x −b8

Figure 2.1.1. A paving of [n] of weight b1 λ3 λ6 b8 x3 .

Any set of orthogonal polynomials may be described as the generating function of pavings:

Theorem 2.1.3. If {Pn (x)} is an OPS with recursion coefficients {bn } and {λn }, then
X
(2.1.4) Pn (x) = wt(π).
π a paving of [n]

The proof is obvious from the three-term recurrence (2.1.1).

Figure 2.1.2 shows a Motzkin path, which is a lattice path consisting of steps (1, 1), (1, 0), and
(1, −1), called northeast, east, and southeast respectively. We weight northeast steps with weight
1, east steps at ordinate (y-coordinate) k with weight bk , and southeast steps that start at ordinate
k with weight λk . We require Motzkin paths to start at height zero, to stay at or above the x-axis,
and to end at height zero except when working with polynomials satisfying higher-order recurrences
(see Section 2.3.1).
 2.1. USUAL OP AND VIENNOT’S GENERAL COMBINATORIAL THEORY 6

1 b1 1 λ2 λ1 b0 1 λ1

Figure 2.1.2. A Motzkin path of weight b0 b1 λ2 λ21

.

One can define a linear functional by setting the nth moment equal to the generating function
for weighted Motzkin paths of length n. The remarkable fact is that these moments are precisely
the moments for the unique linear functional corresponding to the sequence of OP with the same
recursion coefficients.

Theorem 2.1.4. If {µn } are the moments of the linear functional for the sequence of OP defined
by recursion coefficients {bn } and {λn }, then
X
(2.1.5) µn = wt(ω).
ω Motzkin path
of length n

This theorem is a consequence of Theorem 2.1.5 below. We will show that the correspond-
ing orthogonal polynomials really are orthogonal to the moments defined by the right-hand side
of (2.1.5).

2.1.1. A combinatorial proof of orthogonality. Here we shall present Viennot’s proof that
a sequence of polynomials defined by a three-term recurrence (2.1.1) are orthogonal to moments
defined by (2.1.5). This theorem will finish the proofs of Theorems 2.1.2 and 2.1.4.

Theorem 2.1.5 (Viennot). If {Pn } is defined as in (2.1.1) and L is defined by the moments
in (2.1.5), then
L(Pn (x)Pm (x)) = λ1 λ2 · · · λn [n = m].

The proof amounts to a simple sign reversing involution.

Proof. Figures 2.1.3 and 2.1.4 present all the important ideas. See [65, p. I-15] or [66, p. 143]
for a more thorough version of this proof.
Applying L to Pn Pm yields a sum over weighted triples (π1 , π2 , ω) in which π1 (resp. π2 ) is
a paving of [n] (resp. [m]) in which fixed points have weight 1, and ω a Motzkin path of length
f (π1 ) + f (π2 ), where f denotes the number of fixed points in a paving. The weight of such a triple
is of course the product of the weights of the objects that make it up.
The involution works as follows. Find the leftmost nonempty point in the paving or non-NE
step in the Motzkin path and exchange that domino for NE-SE steps, or vice versa (Figure 2.1.3).
If we have a monomino and an E step, exchange those (Figure 2.1.4). If there’s a tie, move from
 2.2. MULTIPLE ORTHOGONAL POLYNOMIALS 7

the paving to the path, since we can always make a Motzkin path longer, but can’t overlap pieces
in the paving.

−λk k

1 +λk

k−1 k k+1 k+2

Figure 2.1.3. The domino/NE-SE steps case of the orthogonality sign-reversing involution.

−bk−1
+bk−1
k−1

k−1 k k+1

Figure 2.1.4. The monomino/E step case of the orthogonality sign-reversing involution.

Let m ≤ n. Fixed points of this involution occur when the first paving is empty and the Motzkin
path starts with n NE steps. If m is strictly smaller than n, the path must go from ordinate n to
the x-axis in at most m steps, an impossibility: the sum is zero and Pn (x) is orthogonal to Pm (x).
If m = n, then there is precisely one such triple: π1 and π2 empty and ω consisting of n NE steps,
then n SE steps. This triple has weight λ1 λ2 · · · λn , which agrees with (2.1.3). 

Because Viennot’s work in [66, 65] is so central to this dissertation, we will not usually cite it;
we will use phrases like “Viennot’s theory of OP” or “Viennot’s general theory”.

2.2. Multiple orthogonal polynomials

Instead of a singly-indexed set of polynomials orthogonal to one measure, what about a multiply-
indexed set of polynomials orthogonal to many measures?
That question is the beginning of multiple orthogonal polynomials. In this section we present
the basics of MOP. As in the previous section, we will not address questions of existence.
There are several slightly different ways to define multiple orthogonal polynomials. The def-
inition we shall use is for type II multiple orthogonal polynomials, described by Van Assche and
Coussement in [63]. It is also the definition used by Aptekarev in [4] and de Bruin in [19]. Discrete
measures and some of our basic theorems are discussed in [7]. A good general introduction to MOP
is chapter 23 of [41].
We start with r weight functions w1 (x), . . . , wr (x), where each wi (x) has total mass 1, or equiv-
(i) (i)
alently, r sequences of moments {µn } for n ≥ 0 and 1 ≤ i ≤ r, with µ0 = 1 for all i. We will
work with r-vectors n = (n1 , . . . , nr ) ∈ Nr and define |n| := n1 + · · · + nr . With the exception of
Section 3.4.1, we will always use r = 2.
 2.2. MULTIPLE ORTHOGONAL POLYNOMIALS 8

The multiple orthogonal polynomials {P (x)n }n∈Nr corresponding to the weights wi (x) satisfy
Z
(2.2.1) Pn (x)xk wi (x) dx = 0
Si

and
Z
(2.2.2) Pn (x)xni wi (x) dx 6= 0
Si

where Pn (x) has degree |n|, 1 ≤ i ≤ r, and 0 ≤ k < ni . Although it is a slight abuse of terminology,
we will refer to the second condition (2.2.2) as the L2 norm with respect to weight wi , or just the
L2 norm if the weight is clear from the context. For usual OP, the L2 norm is the “first time” you
get a nonzero integral, which is why we use the same terminology here.
We will ignore questions of existence; for MOP there is a Hankel determinant condition similar
to (2.1.2), see [41, p. 608].
For example, with respect to the weight 1 + sin x on [−π/2, π/2] and 1 on [−1, 1], the polynomial

3
p(x) = 48π 3 (π 2 − 10) − (π 8 − 4π 6 − 2160π 4 + 35520π 2 − 145920)x
2
− 144π 3 (π 2 − 10)x2 + 10(π 6 − 4π 4 − 432π 2 + 3648)x3

is a candidate for P2,1 (x), since

Z π/2 Z 1
p(x)x0 (1 + sin x) dx = 0, p(x)x0 dx = 0,
−π/2 −1
Z π/2 Z 1
p(x)x1 (1 + sin x) dx = 0, p(x)x1 dx 6= 0.
−π/2 −1
Z π/2
p(x)x2 (1 + sin x) dx 6= 0,
−π/2

We say that the polynomial is a candidate because we could scale the polynomial by a nonzero
constant—just as the length of a vector, so long as it is nonzero, does not affect orthogonality
relations, so does the leading coefficient of an orthogonal polynomial not affect its orthogonality
relations. In this work we shall always use monic polynomials, which are unique for a given set of
weight functions or moment sequences.
(i)
Write µj for the jth moment of wi (x). As before, we will interchangeably use integration
against weight functions and linear functionals; for MOP we will frequently use L1 and L2 for our
functionals.
Usual OP satisfy a three-term recurrence; MOP with two weights satisfy a four-term recurrence:

Theorem 2.2.1. For any set {Pn,m (x)} of MOP orthogonal to weights w1 and w2 , there exist
three sets of constants {bn,m }, {cn,m }, and {dn,m } such that

(2.2.3) Pn+1,m (x) = (x − bn,m )Pn,m (x) − cn,m Pn,m−1 (x) − dn,m Pn−1,m−1 (x).

Proof. We show that we can always solve for the coefficients bn,m , cn,m , and dn,m so that the
right-hand side of (2.2.3) satisfies the same orthogonality relations as Pn+1,m (x); then, because a
 2.3. FOUR-TERM RECURRENCES AND NEARLY-DIAGONAL POLYNOMIALS 9

set of monic OP or MOP with respect to measures of total mass 1 is unique, the two sides are equal.
Writing Qn+1,m (x) for the right-hand side of (2.2.3), we need to show that
Z
(2.2.4) Qn+1,m (x)xj w1 (x) dx = 0, 0 ≤ j ≤ n,
Z
(2.2.5) Qn+1,m (x)xj w2 (x) dx = 0, 0 ≤ j ≤ m − 1.

We already know that (2.2.4) is true for 0 ≤ j ≤ n − 2 and (2.2.5) is true for 0 ≤ j ≤ m − 2. We
need three more orthogonality relations, which translates into a linear homogeneous system in three
variables, which is nonsingular if the L2 norms are nonzero—i.e., if
Z
Pn,m (x)xn[k=1]+m[k=2] wk (x) dx 6= 0

for all n, m ≥ 0 and k = 1 or 2. The necessary conditions are similar to the determinant condition
mentioned above. If we let P−1,m (x) = 0 for all m ≥ 0, then (2.2.3) is valid for all n, m ≥ 1. The
orthogonality conditions imply that along the n = 0 and m = 0 axes we get ordinary orthogonal
polynomials and so they will satisfy their own three-term recurrences. 

The shape of the points (n + 1, m), (n, m), (n, m − 1), and (n − 1, m − 1) in the plane resembles
the S-shaped piece from the game Tetris, so we will call this the Tetris S-piece recurrence or simply
the Tetris recurrence. A set of MOP satisfies other types of recurrences; see Section 2.4.1.
The polynomials Pn,0 (x) and P0,m (x) will be called the x- and y-axis polynomials, respectively,
(1) (2)
and we will also call the moments µn and µm the x- and y-axis moments.

2.3. Four-term recurrences and nearly-diagonal polynomials

Definition 2.3.1. Given a set {Pn (x)} of MOP with respect to r weights, the corresponding
nearly-diagonal polynomials are are those Pn (x) with indices

(n, . . . , n) → (n + 1, n, . . . , n) → (n + 1, n + 1, n, . . . , n) → · · ·

They are a singly-indexed sequence of polynomials denoted Pen (x).

We will use the tilde notation to denote the nearly-diagonal polynomials corresponding to a
set of MOP. As a mnemonic, a tilde goes over ‘n’ in Spanish, so a tilde denotes nearly-diagonal
polynomials. Because these polynomials are singly-indexed, they are simpler to study; indeed, many
papers on MOP only concern the nearly-diagonal polynomials.

Theorem 2.3.2. The nearly-diagonal polynomials for a set of MOP with respect to r weights
satisfy a recurrence of order r + 2.

Proof. Since we are almost exclusively concerned with the r = 2 case, we’ll indicate a proof
for that case only. We wish to show that {Pen (x)} satisfies a four-term recurrence:

(2.3.1) Pen+1 (x) = (x − bn )Pen (x) − cn Pen−1 (x) − dn Pen−2 (x).

For Pe2k (x), the above recurrence follows directly from the more general recurrence (2.2.3) with
(n, m) = (k, k). For Pe2k+1 , (2.3.1) follows from the general recurrence when (n, m) = (k + 1, k) and
 2.3. FOUR-TERM RECURRENCES AND NEARLY-DIAGONAL POLYNOMIALS 10

the roles of w1 and w2 have been reversed, since Pn,m with respect to (w1 , w2 ) equals Pm,n with
respect to (w2 , w1 ). 

This fact is very intriguing. It suggests that polynomial or integer sequences that naturally
involve recurrences of order higher than three can be attacked using multiple OP—for example,
in [29] the number of vertically symmetric alternating sign matrices is connected to a four-term
sequence of polynomials.
Hereafter, whenever referring to multiple orthogonal polynomials or nearly-diagonal polynomi-
als, we will mean there are two weights (r = 2) unless indicated otherwise.
Because higher-order recurrences are central to the theory of MOP, it is appropriate that we
briefly discuss d-orthogonality.

2.3.1. d-orthogonality. The notion of d-orthogonality or orthogonality of dimension d is a

generalization of usual OP and the three-term recurrence (which is a recurrence relation of order 2).
It was introduced by Van Iseghem in her thesis [64] and has been studied by a number of Francophone
researchers [15, 26, 27, 28, 50, 18, 13].
We shall say that a sequence of polynomials satisfies a genuine recurrence relation of order d or
are genuinely d-orthogonal if the coefficient of Pn−d (x) in determining Pn (x) is always nonzero for
n ≥ d. Another way to express this: if we write
d+1
X
(2.3.2) Pn+1 (x) = (x − λn,1 )Pn (x) − λn,i Pn−i+1 (x),
i=2

the polynomials are genuinely d-orthogonal if λn,d+1 is nonzero for all appropriate n.
The primary result relevant to our purposes is

Theorem 2.3.3 ([50], Theorem 2.1). Given a sequence {Pn (x)}n≥0 of monic polynomials with
Pn (x) of degree n, the following two statements are equivalent:
• The polynomials satisfy a genuine recurrence relation of order d + 1, as in (2.3.2), with
suitable initial conditions. (Typically P0 (x) = 1 and polynomials with negative indices are
zero.)
• There exist d linear functionals Li , 1 ≤ i ≤ d, satisfying
Li (Pn (x)Pm (x)) = 0 n > md + i − 1,
(2.3.3)
Li (Pmd+i−1 (x)Pm (x)) 6= 0 m ≥ 0.
Similar to MOP, we will call the second line of (2.3.3) the L2 norm with respect to Li . We
will refer to the moments with respect to L1 and L2 as the first moments and second moments,
respectively. When d = 1 we recover the usual notion of OP and Theorems 2.1.1 and 2.1.2. The
nearly-diagonal polynomials are 2-orthogonal since they satisfy a recurrence relation of order 3:

Corollary 2.3.4. For any set of MOP, there are two unique linear functionals L1 and L2 such
that, for the corresponding nearly diagonal polynomials Pn (x),
L1 (Pn (x)Pm (x)) = 0, n > 2m, L2 (Pn (x)Pm (x)) = 0, n > 2m + 1,
L1 (P2m (x)Pm (x)) 6= 0, m ≥ 0, L2 (P2m+1 (x)Pm (x)) 6= 0, m ≥ 0.
 2.4. NEW RESULTS 11

There is a close relationship between L1 , L2 , and the moments of the weight functions for the
full set of MOP; see Theorem 2.4.1.
It is not difficult to extend Viennot’s theory to d-orthogonal sets of polynomials; roughly speak-
ing, we need longer dominos in the pavings and steeper steps in the Motzkin path. This setup was
described by Viennot [66, p. 149] and further studied by Roblet [53, p. 138]. For polynomials
satisfying a recurrence of order d, one uses pavings with “dominos” of length 1, 2, . . . , d + 1, and
lattice paths with NE and E steps, as well as steps (1, −i) for 1 ≤ i ≤ d − 1. Viennot and Roblet call
these paths Lukasiewicz paths or d-Lukasiewicz paths; here we shall call them generalized Motzkin
paths. We will call (1, −2) steps double-down steps or SSE steps since they go down (or south) twice.
The different sets of moments require Motzkin paths that end at different heights: the functional
Li of (2.3.3) is the generating function for generalized Motzkin paths that end at height i − 1.
The combinatorial proof of Theorem 2.3.3 mimics that of Theorem 2.1.5 and Figures 2.1.4 and
2.1.3: one exchanges a paving piece of length i with the leftmost sequence of lattice path steps that
consist of i − 1 NE steps followed by a (1, −i + 1) step leaving from the appropriate height.
In addition to expressing the moments for d-orthogonal polynomials as generating functions of
generalized Motzkin paths, we can also express them with a matrix product: for usual OP (d = 1),
let M be the infinite tridiagonal matrix, with entries indexed by nonnegative integers, with 1’s on
the superdiagonal, bn on the diagonal, and λn on the subdiagonal. Then the (0, 0) entry of M n
is µn . (Compare with [65, equation (29)], and (70) for arbitrary d.) For d = 2, one can put the
θn ’s on the subsubdiagonal of M —the entries (2, 0), (3, 1), and so on—and the (0, 0) entry of M n
(1)
is again the first moment µn . The reason is that the (i, j) entry of M n is the generating function
for generalized Motzkin paths of length n that start at height i and end at height j. The proof is a
simple use of induction, and the obvious recurrence for such paths: the (i, j) entry of M n equals
X
(the (i, k) entry of M n−1 ) · (the (k, j) entry of M ),
k≥0

which is another way of saying that such a path of length n can be uniquely described as a path
of length n − 1 starting from height i and ending at height k, and a path of length 1 starting from
height k and ending at height j. This implies that

Li (xn ) = (0, i) entry of M n ,

where M has the appropriate recurrence coefficients below the main diagonal.

2.4. New results

So far this chapter has consisted of a summary of the basic definitions and results for multiple
orthogonal polynomials. Here and in subsequent chapters we present new results.
Since we were just discussing d-orthogonality, let’s begin with the relationship between the
moments for a set of MOP and the moments for the corresponding nearly-diagonal polynomials.

Theorem 2.4.1. Let Pn,m (x) be a set of MOP with respect to weight functions w1 and w2 of total
(1) (2)
mass 1, with moments µn and µn , and let Pen (x) be the corresponding nearly-diagonal polynomials
with linear functionals L1 and L2 as in Corollary 2.3.4. If the total mass of L1 and the two weight
 2.4. NEW RESULTS 12

functions are 1, and L2 (Pe1 (x)) = 1, then

L1 (xn ) = µ(1)
n

(2.4.1) (1) (2)

µn − µn
L2 (xn ) = (1) (2)
.
µ1 − µ1
That is: the first moments for the nearly-diagonal polynomials coincide with the x-axis moments,
and the second moments for the nearly-diagonals are a linear combination of x- and y-axis moments.

This is quite surprising. The x-axis polynomials Pn,0 (x) are a set of usual OP, satisfying a three-
term recurrence, and have moments described by Motzkin paths; the nearly-diagonal polynomials
satisfy a four-term recurrence and their first set of moments are generating functions for generalized
Motzkin paths with double-down steps—but they are exactly the same. Let us see why this is so.

Proof. We need only check the corresponding linear functionals in (2.4.1) satisfy the same or-
thogonality conditions: the functionals for the nearly-diagonal polynomials are defined by L1 (Pen (x)) =
[n = 0] and L2 (Pen (x)) = [n = 1]. With respect to the first weight,
Z
Pen (x)w1 dx = [n = 0],

since Pe0 (x) = P0,0 (x) and all the other nearly-diagonal polynomials have a first coordinate greater
than zero, so the first weight sends them to zero—this is just a restatement of the orthogonality
condition for L1 .
For L2 , we posit that it is a linear combination of the first and second weights:
Z Z
L2 (f ) = A f w1 dx + B f w2 dx.

Plug in Pe0 (x) = P0,0 (x) = 1 to see that B = −A. Now plug in Pe1 (x) = P1,0 (x), which must equal
(1)
x − µ1 : Z Z 
1=A P1,0 w1 dx − P1,0 w2 dx .
(1) (2)
The first integral is zero by orthogonality. The second equals µ1 − µ2 , which establishes necessary
conditions on A and B; to see the these conditions are sufficient as well, one checks that the right
side of the second line of (2.4.1) satisfies the orthogonality condition that defines L2 : it sends Pen (x)
to zero for all nonnegative n 6= 1, and sends Pe1 (x) to 1. 

This theorem has two important consequences:

Corollary 2.4.2. A set of monic polynomials satisfying a genuine recurrence relation of order
3 determines, up to the choice of a single constant, a set of MOP for which those polynomials are
the nearly-diagonal polynomials.

Proof. The polynomials determine two linear functionals as in Theorem 2.3.2; the first func-
tional determines the moments for the first weight function, and the second functional determines
(2)
the moments for the second weight function, except for µ1 . Choosing a value for that moment
determines the set of MOP. 
 2.4. NEW RESULTS 13

In (2.4.1), if the weight functions w1 and w2 are the same, then the moments for L2 are undefined;
they are all 0/0. This cannot even be circumvented with a limit, and we highlight this fact with the
following remark.

Remark 2.4.3. The first and second weight functions for a set of MOP cannot be the same.

The reason is the orthogonality relations for the polynomials with unequal subscripts are con-
tradictory if w1 and w2 are identical: we want xP2,1 (x) to be zero with respect to the first weight,
but nonzero with respect to the second weight, which cannot happen if the weights are the same.
This reasoning also shows that, for an arbitrary number of weights, no two of them can be the same.
Having just discussed the relationship between the moments of the x- and y-axis polynomials and
the moments of the nearly-diagonal polynomials, it is natural to explore the relationship between the
recurrence coefficients of those three sets of polynomials. The x- and y-axis polynomials determine
the MOP, and hence the nearly-diagonal polynomials, so the recurrence coefficients for the nearly-
diagonals must be functions of the x- and y-axis recurrence coefficients. It would be desirable to
have a combinatorial description of this function.
Say the axis polynomials satisfy the recurrence

Pn+1 (x) = (x − bn(i) )Pn (x) − λ(i)

n Pn−1 (x)

with i = 1 for the x-axis and i = 2 for the y-axis, and let the nearly-diagonal polynomials satisfy
the recurrence
Pen+1 (x) = (x − Bn )Pen (x) − Λn Pen−1 (x) − θn Pen−2 (x).
What are Bn , Λn , and θn as functions of the b’s and λ’s for the x- and y-axis?
By considering the orthogonality relations that Pen must satisfy, it is easy to find the first two
Bn ’s:
(1)
B0 = b0
(1) (2)
(2) λ1 − λ1
B1 = b0 + (1) (2)
.
b0 − b0
B2 is also a rational function in the b’s and λ’s; its numerator has 142 terms, its denominator 56.
For Λn the prospects are no better: we have

Λ0 = 0,
(1)
Λ 1 = λ1 ,

and Λ2 has 354 and 79 terms in its numerator and denominator, respectively. Unsurprisingly, θn is
similar: θ0 = θ1 = 0 and θ2 has a similarly unwieldy expression as a rational function, although it
 2.4. NEW RESULTS 14

is worth noting that it has the same denominator as B2 :

 
(1) (2) (1) (1) (2) (1) (1) (2) (1) (2) (2) (1) (2) (2) (1) (1) (2) (1)
b0 − b0 b0 3 b1 b0 2 − b0 2 b1 b0 3 − b0 3 b0 2 b1 + b0 2 b0 3 b1 − b0 b1 b0 λ 1
(1) (2) 2 (1) (1) (2) 2 (1) (1) (1) (2) 2 (1) (1) 2 (2) 2 (1) (2) 3 (1) (1) (2) 3 (1)
+ b0 b0 λ1 + b1 b0 λ1 + 2b0 b1 b0 λ1 + b1 b0 λ 1 − b0 λ1 − b1 b0 λ1
(1) (2) 2 (2) (1) (1) (2) (2) (1) (2) (2) (1) (2) (1) (1) (2) (1) (1) (2) (1) (2) (2)
− 2b0 b0 b1 λ 1 − b1 b0 2 b1 λ 1 + b0 3 b1 λ 1 + b0 λ 1 2 − b0 3 λ 1 + b0 3 b1 λ 1 + b0 2 b0 λ 1
(1) 2 (1) (2) (2) (1) (2) (2) (1) (2) (2) (1) (1) (2) (2) (1) (2) (2) (2) (1) (2) (2) (2)
− 2b0 b1 b0 λ1 + b0 2 b1 λ1 − b0 3 b1 λ1 − b0 2 b1 b1 λ1 − b0 b0 b1 λ1 + 2b0 2 b0 b1 λ1
(1) 2 (2) 2 (2) (1) (1) (2) (1) (1) (1) (2) (1) 2 (1) (2) (2) (1) (2) (1) (2) (1) (2)
+ b0 b1 λ 1 − b0 λ1 λ1 + 2b0 b1 λ1 λ1 + b1 λ1 λ1 − b0 λ1 λ1 − 2b1 b0 λ1 λ1

(1) (2) (1) (2) (1) (2) (1) (2) (2) (2) (1) (2) (2) 2 (1) (2) (1) (2) 2
−2b0 b1 λ1 λ1 − 2b1 b1 λ1 λ1 + 2b0 b1 λ1 λ1 + b1 λ1 λ1 + b0 λ 1

A computer program could not find B3 , Λ3 , or θ3 in a reasonable amount of time.

All this seems to imply that is is extremely difficult to work with very general sets of MOP
because the coefficients are too complicated. However, we do have the following partial result on
the orthogonality of P1,m (x) using Viennot’s general theory.

Theorem 2.4.4. Given a set of MOP with weight functions w1 and w2 , assume the y-axis
polynomials P0,m (x) satisfy the recurrence

(2.4.2) P0,m+1 (x) = (x − dn )P0,m (x) − en P0,m−1 (x).

Then the polynomials P1,m (x) satisfy the recurrence relation

(2.4.3) P1,m (x) = (x − b0,m )P0,m (x) − c0,m P0,m−1 , m ≥ 1,

where
Z  ! Z 

b0,m = P0,m+1 (x)w1 dx P0,m (x)w1 dx ,
dm →0

c0,m = em .

The notation in the numerator of b0,m means that one should think of the integral of P0,m+1 (x)
as a polynomial in the d’s, e’s, and moments of w1 , and send dm to zero. The coefficients in (2.4.3)
follow the notation of Theorem 2.2.1.

Proof. We need only show that the orthogonality conditions are satisfied; by uniqueness we
must have the right polynomials.
First, we need to show that the first weight sends each polynomial to zero, so apply it to the
right-hand side of (2.4.3):
Z Z  ! Z

xP0,m w1 dx − P0,m+1 w1 dx
 − em P0,m−1 w1 dx.
dm →0

Use the recurrence relation (2.4.2) for the y-axis polynomials on xP0,m to rewrite the first term of
that expression, which becomes
Z Z  ! Z

(P0,m+1 + dm P0,m + em P0,m−1 )w1 dx − P0,m+1 w1 dx − em P0,m−1 w1 dx.
dm →0
 2.4. NEW RESULTS 15

The em terms cancel, and we will have proved orthogonality if we can show that
Z Z Z  !

(P0,m+1 w1 dx − (−dm ) P0,m w1 dx − P0,m+1 w1 dx = 0.
dm →0

(The double negative in front of dm is intentional.) This is easy to do using Viennot’s theory: the
y-axis polynomials of (2.4.2) are generating functions for pavings with dominos and monominos
weighted by en and dn . The first term on the left side of the equation is a generating function
for path-paving pairs in which the paths have length m + 1 and empty points have weight 1; the
corresponding Motzkin paths are weighted by the recurrence coefficients for w1 . (This is the set
E0,m+1,0 from [65, p. I-16].)
The third term is very similar, but in this case, we have removed all path-paving pairs in which
the paving ends with a monomino on vertex m + 1. The middle term exactly compensates for that:
the integral is the generating function for path-paving pairs with a paving of length m; multiplying
it by −dm yields exactly the missing part of the third term. The second and third terms, then, add
up to the first—so their difference is zero. This means P1,m so defined has the correct orthogonality
relation with w1 .
Now to w2 . We require P1,m to be orthogonal to xk for 1 ≤ k ≤ m − 1. Again we show that the
right-hand side of (2.4.3) has the correct orthogonality.
That right-hand side already is orthogonal to xk for 1 ≤ k ≤ m − 2, so what about xm−1 P1,m ?
Integrate that with respect to w2 :
Z Z Z
m m−1
x P0,m w2 dx − b0,m x P0,m w2 dx − em xm−1 P0,m−1 w2 dx.

The middle term is zero by orthogonality, the first term is the product e1 e2 · · · em , and the rightmost
integral is e1 e2 · · · em−1 ; after multiplying that by em , the difference is clearly zero. 

This is the first combinatorial proof of any orthogonality relation for general MOP. Note that
by switching the weights and the roles of the recurrence coefficients appropriately, one obtains a
recurrence for Pn,1 of the form

Pn,1 = (x − bn,0 )Pn,0 − cn,0 Pn−1,0 ;

this is not the Tetris recurrence of (2.2.3)—the two polynomials on the right would need to be Pn−1,1
and Pn−1,0 to fit that recurrence “shape”. We will address the issue of different kinds of recurrence
relations in Section 2.4.1; first let’s discuss the recurrence coefficients for the Tetris recurrence.
The above theorem has an explicit expression for some of those coefficients, and the natural
thing to do is try and get similar expressions for more of the coefficients. We have no results in this
direction, but computer evidence indicates that these coefficients become very, very complicated:
we have calculated the general coefficients for the polynomials with degree at most 6, not including
P3,3 , and it takes 91 megabytes to store the expressions for those coefficients.
It does appear that the “integrate with the wrong weight” phenomenon seen in Theorem 2.4.4
continues, and we conjecture the following:
 2.4. NEW RESULTS 16

Conjecture 2.4.5. Assume the hypotheses of Theorem 2.4.4, and also assume that the x-axis
polynomials have recurrence coefficients {bn } and {cn }. Using the notation of Theorem 2.2.1, the
recurrence coefficients for Pn,1 are
Z  Z  Z −2
(2.4.4) cn−1,1 = cn−1 Pn−2,0 w2 dx Pn,0 w2 dx Pn−1,0 w2 dx
Z  Z −1
+ xPn−1,1 w2 dx Pn−1,0 w2 dx
Z  Z 
(2.4.5) dn−1,1 = −cn−1 Pn,0 w2 dx Pn−1,0 w2 dx

We are abusing notation slightly; cn−1,1 is a coefficient for Pn,1 , and cn−1 is a coefficient for
Pn,0 .
Note that some factors of the numerator of cn−1,1 become factors in the denominator of cn,1 . This
behavior has also been observed with cn,2 and it seems likely that those coefficients can be expressed
similarly. The numerator-becomes-a-denominator phenomenon is reminiscent of the convergents of
the moment generating function [65, Section V.3]: for a usual set of OP with recurrence coefficients
bn and λn , the moment generating function can be written
X 1 λ 1 t2 λ 2 t2 λ 3 t2
µn tn = ···
1 − b0 t − 1 − b1 t − 1 − b2 t − 1 − b3 t −
n≥0

Viennot showed that the nth convergent of that continued fraction is δPn∗ (t)/Pn+1
∗
(t), where the
∗ n
star superscript indicates the reciprocal polynomial: Pn (t) = t Pn (1/t). The δ notation in this
context means δPn (t) is the generating function for pavings on the vertices {2, 3, . . . , n + 1}, with the
monominos and dominos the same weights as usual. Perhaps the recurrence coefficients (2.4.4) can
be described as convergents of some continued fraction. This is an intriguing avenue of investigation.

2.4.1. Recurrence relation shapes. In this section we investigate the sorts of recurrence
relations that a set of MOP can satisfy. In Theorem 2.2.1 we saw that any set of MOP satisfies the
Tetris recurrence; in the next chapter we will see that the multiple Hermite polynomials naturally
satisfy a recurrence involving (n + 1, m), (n, m), (n − 1, m − 1) and (n − 1, m), which we might call
a Tetris T-piece recurrence. The results of this section are a generalization of [41, Theorem 23.1.7]
for the r = 2 case.
Let’s restrict ourselves to recurrence relations of the following form: the polynomial Pn,m equals
(x − α) times Pn−1,m , plus a linear combination of two more polynomials of degree n + m − 2 or
n + m − 3, both of which have a first coordinate at most n and second coordinate at most m.
Figure 2.4.1 shows which polynomials we are considering.
The reasoning we will use is best illustrated by an example. Let’s prove the Tetris S-piece
recurrence of Theorem 2.2.1. We assume that the polynomials exist and satisfy the orthogonality
conditions of (2.2.1) and (2.2.2). We wish to show that we can solve for α, β, and γ in

Pn+1,m = xPn,m − αPn,m − βPn,m−1 − γPn−1,m−1 .

 2.4. NEW RESULTS 17

(n − 1, m)
(n, m)
a g

b f

de
gre
e
n
e

+
n
c

m
+

−
m

1
−
2
d

n
+
m
−
3
Figure 2.4.1. We wish to express Pn,m as (x−α)Pn−1,m plus a linear combination
of two of the seven polynomials of degree n + m − 2 and n + m − 3.

We need orthogonality to xk for 0 ≤ k ≤ n with respect to the first measure, and to xk for
0 ≤ k ≤ m − 1 with respect to the second. We need only show orthogonality for n − 1, n, and m − 1;
those three conditions yield a 3 × 3 linear system, which we may write

0 = N + 0α + 0β + N γ wrt xn−1 , weight 1

0 = ∗ + N α + N β + ∗γ wrt xn , weight 1
0 = N + 0α + N β + N γ wrt xm−1 , weight 2

where N means a nonzero number and ∗ means any number. We can rewrite the system as
    
0 0 N α N
N N ∗  β  =  ∗  ,
    

0 N N γ N
which has a nonzero solution since the matrix must be invertible—its determinant in the product of
the nonzero entries in positions (1, 3), (2, 1), and (3, 2).
The hypotheses stated above mean that we have 72 = 21 possible recurrence shapes, corre-

sponding to unordered pairs in {a, b, . . . , g} in Figure 2.4.1. Using the above reasoning to test those
21 cases, we get

Theorem 2.4.6. Under the above hypotheses:

• The recurrence shapes that work are {a, g}, {b, g}, {b, f }, {c, f }, and {f, g}.
• {c, e} yields a matrix which is not necessarily invertible.
• {e, g} yields an invertible matrix, but the coefficient of the polynomial for e is zero.
• The remaining 14 possibilities each produce more than three orthogonality conditions that
need to be satisfied.

The {b, f } shape is the Tetris S-piece recurrence. Theorem 23.1.7 in [41] shows that the {b, g}
shape works.
 2.4. NEW RESULTS 18

Theorem 23.1.7 also shows that two valid recurrence shapes involve writing Pn,m as (x−α) times
Pn,m−1 ; we can, so to speak, flip Theorem 2.4.6 about the line y = x, and by symmetry immediately
conclude

Corollary 2.4.7. If we wish to write Pn,m as (x − α) times Pn,m−1 plus a linear combination
of two more polynomials, and assume the hypotheses stated at the beginning of this section, then the
recurrence shapes which work are {d, e}, {c, e}, {c, f }, {b, f }, and {e, f }. The other 16 possibilities
do not work.

The shapes {c, f } and {c, e} are the other two predicted by Theorem 23.1.7. Figure 2.4.2 shows
what these ten recurrence shapes look like; it is amusing to note that all the shapes are Tetris pieces,
and only the square Tetris piece is missing (that shape yields a noninvertible matrix, sadly).

Figure 2.4.2. The recurrence shapes from Theorem 2.4.6 and Corollary 2.4.7.

There are other possible recurrence shapes that will work. For instance, the same reasoning as
above shows that one can write

Pn,m = (x − A)Pn−2,m+1 − BPn−1,m − CPn−1,m−1

for some A, B, and C.

 CHAPTER 3

Multiple Hermite polynomials

In this chapter we will study the simplest set of multiple orthogonal polynomials: the multiple
Hermite polynomials. Let’s begin with some definitions and terminology related to matchings and
a brief review of the combinatorics of the usual Hermite polynomials, and matchings, which occur
frequently in this chapter and others.

3.1. Matchings and usual Hermite polynomials

A matching of [n] is a set of disjoint pairs of numbers in [n]. A matching is frequently represented
by a graph with vertices labeled 1 to n with edges connecting paired-up numbers, and we require
that each vertex be adjacent to at most one edge. A complete matching is one in which each number
is paired with another number. A matching can be viewed as an involution of [n], so we will refer to
non-paired-up numbers (equivalently, isolated vertices in the graph) as fixed points of the matching,
and paired-up numbers as transpositions. A matching of [n] is sometimes (for instance, in [25])
thought of as a matching of the complete graph on n vertices. In that context, a matching is a
subset of edges of a graph with the property that no two edges in the subset are adjacent to a
common vertex.
The reader should be aware of a possible source of confusion: we will, in this chapter and others,
be performing involutions on sets of matchings (which are themselves involutions), and that there
are fixed points of the matching and fixed points of the involution on the set of matchings.
The usual Hermite polynomials are orthogonal polynomials which may be defined by the recur-
rence relation

(3.1.1) Hn+1 (x) = xHn (x) − nHn−1 (x),

with H0 (x) = 1 and H1 (x) = x. In the normalization we are using here, they are orthogonal to the
2
weight function e−x /2
on the real line:
Z +∞
1 2
(3.1.2) √ Hn (x)Hm (x) e−x /2 dx = [n = m]n!.
2π −∞
The corresponding moments are
Z +∞
1 2
(3.1.3) µn = √ xn e−x /2
dx = [n even](n − 1)!!.
2π −∞

The combinatorial interpretations of these polynomials and moments are well-known: the nth poly-
nomial is the generating function for matchings or involutions of [n] in which fixed points have weight
x and transpositions have weight −1, and the nth moment is the number of complete matchings
19
 3.1. MATCHINGS AND USUAL HERMITE POLYNOMIALS 20

of [n]. See, for example, the papers [12, 32, 37, 25] for the classical combinatorics of Hermite
polynomials.
We note briefly that, by the exponential formula, the Hermite polynomials have the following
generating function:
tn
 2 
X t
(3.1.4) Hn (x) = exp − + xt .
n! 2
n≥0

This fact will be used to derive a formula for the multiple Hermites in Section 3.3.
The combinatorial interpretation can be derived from the recurrence relation as follows: the
vertex n + 1 may be fixed, and we multiply all matchings on [n] by x; or we may connect vertex n + 1
to any of the n vertices to its right, and then multiply −1 by all matchings on the n − 1 remaining
vertices. This sort of interpretation of the recurrence relation, and variations on it, occur frequently
in Chapter 5.
The moments count the number of complete matchings, which may be seen from the integral
(3.1.3), or from Viennot’s general theory: the nth moment is the generating function for Dyck paths1
in which SE steps leaving from height k have weight k—or, to use the language of “histories” from
Viennot’s general theory, we have labeled Dyck paths in which SE steps leaving from height k may
have any label 1, 2, . . . , k. Here is a bijection from this set to the set of complete matchings: read
the path from left to right. For each NE step, place an isolated vertex. For a SE step labeled i
leaving from height k, there must be k isolated vertices in the graph—connect a new vertex to the
ith available vertex, starting from the left. Figure 3.1.1 has an example of this bijection. There are,
of course, a number of small variations on this bijection—one can build the complete matching from
right to left, attach an edge to the ith available vertex reading from the right, et cetera—and some
of these variations will be used in Chapter 5.

3
2 1
1 1 2 3 4 5 6

Figure 3.1.1. The bijection from labeled Dyck paths to complete matchings. We
are considering the 6th step in the path (the green edge), which corresponds to
matching vertex 6 (the green vertex) to another vertex. The label 3 tells us to
connect vertex 6 to the third unmatched vertex, counting from the left, so we will
connect to vertex 5. The final complete matching corresponding to the Dyck path
is (17)(23)(48)(56).

For our purposes, integrating a product of Hermite polynomials consists of taking pairs of
incomplete matchings, weighted as described above, and replacing the fixed points with complete
matchings with edges weighted +1. This setup will occur often in this work, so let us make a
definition:
1The moments are generating functions for Motzkin paths, but in the recurrence relation (3.1.1), b = 0, so any
n
Motzkin path with a horizontal edge has weight zero. Thus we can use Dyck paths.
 3.3. A COMBINATORIAL MODEL 21

Definition 3.1.1. The set of paired matchings on [n] t [m] is the set of complete matchings on
[n] t [m] with green and black edges. Black edges are homogeneous (they stay within [n] or [m]) and
have weight −1. Green edges may connect anywhere and have weight 1.

This definition is not standard terminology, it is only for our convenience in this work.
The orthogonality relation may then be phrased as “the generating function for paired matchings
is zero if n 6= m, and is n! if n = m”. A sign reversing involution that proves this orthogonality is:
Arrange the vertices so that [n] is to the left of [m]. Find the leftmost homoge-
neous edge and flip its color.
If n 6= m, there must be at least one homogeneous edge, and hence there are no fixed points in that
case. If n = m, every edge must go from [n] to [m]; this obviously describes a permutation of [n].

3.2. Definition

Walter van Assche and his coauthors define the multiple Hermite polynomials for two weights
in [63] and for an arbitrary number of weights in [5]. We will mostly be concerned with the two-
weight case here, although many of our results have obvious generalizations to an arbitrary number
c,d
of weights. (See Section 3.4.1.) The multiple Hermite polynomial Hn,m (x) is the monic polynomial
of degree n + m satisfying
Z +∞
1 2
√ c,d
Hn,m (x) xk e−(x−c) /2 dx = [k < n]n!(c − d)m ,
2π −∞
(3.2.1) Z +∞
1 2
√ c,d
Hn,m (x) xk e−(x−d) /2 dx = [k < m]m!(d − c)n .
2π −∞
In the sequel we will use dµ(1) and dµ(2) for the first and second measures (c and d, respectively),
and also often omit the c, d superscript on the polynomials.
Van Assche et al. also derive the recurrence relation:

c,d
Theorem 3.2.1. Hn,m (x) satisfies the following recurrence relation for n, m ≥ 1:

(3.2.2) Hn+1,m (x) = (x − c)Hn,m (x) − (n + m)Hn,m−1 (x) − n(c − d)Hn−1,m−1 (x),

with Hn,0 (x) = Hn (x − c) and H0,m = Hm (x − d).

This recurrence is given in [5, p. 3890]—use δ = −1, their β1 and β2 are our c and d, respectively,
and their polynomials differ from ours by a factor of (−1)n+m . We will also offer a bijective proof
of this recurrence in Theorem 3.4.1 after we’ve described a combinatorial model for the multiple
Hermites.

3.3. A combinatorial model

In this section we’ll describe the combinatorial model of the multiple Hermites. We could pull
the model out of thin air and show that it is correct by proving the orthogonality relations and
appealing to uniqueness, but we’ll take a slightly more roundabout route which explains how this
model was derived, and also yields some nice results along the way. We’ll start with the Rodrigues
 3.3. A COMBINATORIAL MODEL 22

formula for multiple Hermites, use that to find the exponential generating function, and from the
generating function, express the multiple Hermites as sums of usual Hermites.
The Rodrigues formula for the multiple Hermite polynomials [5, eq. (5.2), use δ = −1] is
dn dm 
 
2 2
(3.3.1) Hn,m (x) = (−1)n+m ex /2−cx n e(c−d)x m e−x /2+dx .
dx dx
That formula implies the exponential generating function:

Theorem 3.3.1. The multiple Hermites have the exponential generating function
∞ X ∞
sn tm s2 t2
X  
(3.3.2) Hn,m (x) = exp s(x − c) + t(x − d) − st − − .
n=0 m=0
n! m! 2 2

Proof. Use (3.3.1) and two applications of Taylor’s formula in the form
X xn
f (x + a) = f (n) (a)
n!
n≥0

and the result follows immediately. This generating function is also a consequence of the exponential
formula and the model which will be described shortly. 

The exponential in (3.3.2) is similar to the exponential generating function for the usual Hermite
polynomials (3.1.4) and allows us to prove

Theorem 3.3.2. The multiple Hermites may be written as a sum of usual Hermite polynomials:
n X m   
X n m
(3.3.3) Hn,m (x) = Hi+j (x)(−c)n−i (−d)m−j .
i=0 j=0
i j

Proof. Write (3.3.2) as

(s + t)2
 
exp − + x(s + t) exp(−cs) exp(−dt);
2
the first exponential is the generating function for usual Hermite polynomials with t replaced by
s + t, so the above expression equals
   
X Hi+j (x) X (−cs)n X (−dt)n
 si tj    .
i!j! n! n!
i,j≥0 n≥0 n≥0

n m
The coefficient of s t /(n!m!) in the above expression is Hn,m (x), and it equals
n X
m
X Hi+j (x) (−c)n−i (−d)m−j
n!m! ,
i=0 j=0
i!j! (n − i)! (m − j)!

which equals the expression given in (3.3.3). 

Equation 3.3.3 motivates the following combinatorial definition. Its r-weight version is equation
(3.4.8).
c,d
Theorem 3.3.3. The multiple Hermite polynomial Hn,m (x) is the generating function for match-
ings on [n] t [m] in which edges have weight −1 and fixed points in [n] (resp. [m]) may have weight
x or −c (resp. x or −d).
 3.3. A COMBINATORIAL MODEL 23

Proof. This is clear from Theorem 3.3.2: choose i vertices in [n], j vertices in [m], and put an
arbitrary matching with the Hermite weighting on those i + j vertices. Make the remaining n − i
(resp. m − j) vertices in [n] (resp. [m]) fixed points of weight −c (resp. −d). 

One such matching is shown in Figure 3.3.1. This interpretation offers an alternative proof
of the generating function (3.3.2): using the exponential formula, we have the following connected
components:
• a fixed point in [n] (resp. [m]), weight sx (resp. tx);
• a fixed point in [n] ([m]), weight −sc (−td);
• an edge staying within [n] ([m]), weight −s2 /2 (−t2 /2);
• an edge going from [n] to [m], weight −st.
The exponential of the sum of those weights coincides with (3.3.2).

−c x −d x −d

[n] [m]

c,d
Figure 3.3.1. A matching that contributes cd2 x2 to H5,6 (x). Edges all have weight −1.

Our model immediately implies another exact formula for the multiple Hermites: since fixed
points in, say, [n] may have weight x or −c (equivalently, weight x−c), we can take care of all the fixed
points and homogeneous edges with shifted Hermite polynomials, and express the inhomogeneous
edges separately:
min(m,n)   
X n m
(3.3.4) Hn,m (x) = k!Hn−k (x − c)Hm−k (x − d).
k k
k=0

Both versions of Proposition 3.4.6 are r-weight versions of this expression.

3.3.1. Combinatorial proof of orthogonality. In this section we present a combinatorial

proof of the orthogonality relations for the multiple Hermite polynomials which generalizes the
combinatorial orthogonality proof for usual Hermite polynomials found in the introduction to this
chapter. We will prove the relations (3.2.1) in a slightly different but equivalent form:

Theorem 3.3.4. The following orthogonality relations hold for the multiple Hermite polynomials:
Z
Hn1 ,m1 (x)Hn2 ,m2 (x) dµ(1) = 0 if n2 + m2 < n1 ,
(3.3.5) Z
Hn1 ,m1 (x)Hn2 ,m2 (x) dµ(2) = 0 if n2 + m2 < m1 ,
 3.3. A COMBINATORIAL MODEL 24

and the L2 norms

Z
Hn1 ,m1 (x)Hk,n1 −k (x) dµ(1) = n1 !(c − d)m1 0 ≤ k ≤ n1 , and
(3.3.6) Z
Hn1 ,m1 (x)Hk,m1 −k (x) dµ(2) = m1 !(c − d)n1 0 ≤ k ≤ m1 .

The rest of this section is devoted to a proof of this theorem. Let’s begin by noting that the
(1)
weight function exp(−x2 /2 − cx) has moments µn that are generating functions for incomplete
matchings of [n] with edges weighted +1 and fixed points weighted c; the reason, essentially, is that
we have take the weight function for the usual Hermites and sent x to x + c and we may choose
whether a vertex is adjacent to an edge or is fixed and has weight c. See [39] and [40] for more on
this phenomenon in which “the moments are the polynomials”.
Integrating Hn1 ,m1 (x)Hn2 ,m2 (x) with respect to one of the weights gives us paired matchings
of a sort, but our definition of paired matching needs to stretch to fit the current scenario: we now
have a pair of pairs of matchings (!), on [n1 ] t [m1 ] and [n2 ] t [m2 ]. We’ll refer to the “left pair”
(n1 and m1 ) and “right pair” (n2 and m2 ) of vertices. On these four sets of vertices, we have black
edges which have weight −1 and must stay within the left or right pair, and green +1 edges which
may go anywhere. There are black fixed points of weight −c (resp. −d) if inside [n1 ] or [n2 ] (resp.
[m1 ] or [m2 ]). And finally, if integrating with respect to the first (second) weight, we have green
fixed points of weight +c (+d) which may be anywhere.
Such a paired matching is shown in Figure 3.3.2.

−c +c

−d +c

Figure 3.3.2. A paired matching that is part of the integral of H6,3 (x)H4,7 (x) with
respect to the first weight. Here H6,3 is on the left and H4,7 is on the right. The
orthogonality involution would flip the color of the leftmost edge which connects
vertices 1 and 3 in [6], the upper left set of vertices—although these two multiple
Hermites have no special orthogonality relation; the integral of their product with
respect to the first weight is (7620480 + 4233600(c − d)2 + 529200(c − d)4 + 15120(c −
d)6 )(c − d)2 .

Assume that n2 + m2 < n1 , and integrate with respect to the first weight. Apply the following
involution to the set of paired matchings described above:
 3.4. IDENTITIES AND OTHER RESULTS 25

Arrange the sets of vertices in the following order: n1 , m1 , n2 , m2 . Find the

leftmost edge internal to the left or right pair and flip its sign. If there is no such
edge, find the leftmost fixed point in [n1 ] or [n2 ] and flip its sign.
Paired matchings left alone by this involution are ones in which all edges have weight +1 and go
from the left pair to the right pair, and neither [n1 ] nor [n2 ] have any fixed points. If n2 + m2 < n1 ,
there can be no paired matchings left alone, and we have proved orthogonality with respect to the
first weight. The proof of orthogonality with respect to the second weight is exactly similar.
If n1 = n2 + m2 and we use the first weight, then after applying the above involution, we are
left with paired matchings in which all the edges from [n1 ] attach to [n2 ] or [m2 ]; there are n1 ! ways
to arrange this. The vertices in [m1 ] must all be fixed points, and since they can have weight c or
−d, our L2 norm is n1 !(c − d)m1 . The proof for the L2 norm with respect to the second weight is
again exactly similar. Theorem 3.3.4 is proved. 

3.4. Identities and other results

Our first result in this section is to reconcile the Tetris-piece recurrence of Theorem 3.2.1 and
the natural recurrence suggested by the combinatorial interpretation in Theorem 3.3.3, which works
as follows: to find Hn+1,m (x), consider vertex n + 1. That vertex could be fixed of weight x − c,
and we can fill in the remaining points with a multiple Hermite matching on [n] t [m], contributing
(x − c)Hn,m (x), or it could connect to another vertex in [n + 1] and yield −nHn−1,m (x), or it could
connect to any vertex in [m] and contribute −mHn,m−1 (x).

Theorem 3.4.1. The multiple Hermite polynomials satisfy both the Tetris S-piece recurrence of
Theorem 3.2.1:

(3.2.2) Hn+1,m (x) = (x − c)Hn,m (x) − (n + m)Hn,m−1 (x) − n(c − d)Hn−1,m−1 (x),

and the natural recurrence, described above, which is a T-piece recurrence in Theorem 2.4.6:

(3.4.1) Hn+1,m (x) = (x − c)Hn,m (x) − nHn−1,m (x) − mHn,m−1 (x).

Proof. Ignore the common terms on the right sides of both recurrences; we will be done if we
can prove that
Hn,m−1 (x) + (c − d)Hn−1,m−1 (x) = Hn−1,m (x).
We’ll provide a bijective proof of this. The idea is that a matching on [n] t [m − 1] is very nearly the
same as a matching on [n−1]t[m]. One may turn a weighted matching on the first set into a weighted
matching on the second set by moving vertex n into the [m − 1] set of vertices. This operation will
be weight-preserving unless that vertex is a fixed point of weight −c, so, to the generating function
for matchings on [n] t [m − 1] (i.e., Hn,m−1 (x)) we add the generating function for matchings on the
same set in which vertex n has weight +c; that’s cHn−1,m−1 (x). That cancels those matchings for
which moving vertex n would not be a weight-preserving operation. Then, we add in the generating
function for matchings on [n]t[m−1] in which vertex n has weight −d: −dHn−1,m−1 (x). Altogether
we’ve described exactly the identity stated above. 
 3.4. IDENTITIES AND OTHER RESULTS 26

Our first major result in this section is a version of the Mehler formula. The Mehler formula is
a bilinear generating function for the usual Hermite polynomials:
∞
tn 2txy + t2 (x2 + y 2 )
 
X 1
(3.4.2) Hn (x)Hn (y) = √ exp .
n=0
n! 1 − t2 2(1 − t2 )

We will state a generalization of this formula and provide a proof analogous to that given by Foata
in [32].

Theorem 3.4.2. The Mehler formula for multiple Hermite polynomials with two weights is
X sn tm
(3.4.3) Hn,m (x)Hn,m (y) = exp(A(s, t) + · · · + G(s, t)),
n! m!
n,m≥0

where
1
A(s, t) = − log(1 − (s + t)2 ),
2
−st
B(s, t) = ((x − c)(x − d) + (y − c)(y − d)) ,
1 − (s + t)2
−s2
C(s, t) = ((x − c)2 + (y − c)2 ) ,
2(1 − (s + t)2 )
−t2
D(s, t) = ((x − d)2 + (y − d)2 ) ,
2(1 − (s + t)2 )
s2 t + st2
E(s, t) = ((x − d)(y − c) + (x − c)(y − d)) ,
1 − (s + t)2
s3 + s2 t
 
F (s, t) = (x − c)(y − c) s + ,
1 − (s + t)2
t3 + t2 s
 
G(s, t) = (x − d)(y − d) t + .
1 − (s + t)2

Proof. As in [32], we just need to find the connected components. Think of Hn,m (x)Hn,m (y)
as pairs of weighted matchings on [n] t [m], with the “x matching” having black edges and the “y
matchings” having green edges. See Figure 3.4.1 for an example of this setup. Observe that these
are not the paired matchings of Definition 3.1.1.
There are seven types of connected components:

Case A: An even cycle, weight 1. Since the total number of points must be even, n and m
can be both even (n, m ≥ 0, except for (0, 0)) or both odd (n, m ≥ 1).
Case B: An odd path, with one endpoint in [n] and one in [m], of weight −(x − d)(x − d) −
(y − c)(y − d) (the endpoints can either be in the black matching or green matching). Odd
paths have an even number of vertices, so n and m can be both even (n, m ≥ 2) or both
odd (n, m ≥ 1).
Case C: An odd path, with both endpoints in [n], weight −(x − c)2 − (y − c)2 (similarly,
endpoints can be in the black or green matching).
Case D: Case C with n and m switched (you must also switch c and d).
 3.4. IDENTITIES AND OTHER RESULTS 27

Case E: An even path, with one endpoint in [n] and one in [m], of weight (x − c)(y − d) +
(x − d)(y − c). Even paths have an odd number of vertices, so we may have n even and
≥ 2, m odd and ≥ 1 or vice versa.
Case F: An even path, with both endpoints in [n], of weight (x − c)(y − c). We include the
special case of (n, m) = (1, 0).
Case G: Case F with n and m switched (you must also switch c and d).
We’ll do case A as an example; the others are similar.

x−c x−d x−d

y−d

Figure 3.4.1. A pair of matchings that are part of H4,7 (x)H4,7 (y). The matching
for H4,7 (x) is in black; H4,7 (x) is in green. All edges have weight −1. There are
four connected components, which from left to right are in case C, A, A, and G.

On n + m vertices there are (n + m − 1)! cycles, so the generating function—ignoring the parity
requirements for a moment; we’ll fix that by sieving—is
X sn tm
(3.4.4) (n + m − 1)! .
n! m!
n,m≥0
(n,m)6=(0,0)

Split off the n = 0 sum and write the above power series as
X sn X m + n − 1 X tm
tm + (m − 1)! ;
n m m!
n≥1 m≥0 m≥1

the second summand is a logarithm, and the m sum of the first summand can be found using the
elementary identity
X n + k  1
xn = ,
n (1 − x)k+1
n≥0

so that (3.4.4) is
X 1  s n
− log(1 − t) = − log(1 − s − t).
n 1−t
n≥1
We need to sieve this power series to get only terms in which n and m are both even, or both odd.
If f (s, t) is a power series in s and t, then
f (s, t) + f (−s, −t)
(3.4.5)
2
is a power series of terms of the form s2n t2m or s2n+1 t2m+1 . The correct case A generating function
is therefore
1
A(s, t) = − log(1 − (s + t)2 ).
2
 3.4. IDENTITIES AND OTHER RESULTS 28

The other cases are found using similar techniques; if you need to sieve to get “odd-even” terms,
you just subtract in (3.4.5). Call those generating functions B(s, t), . . . , G(s, t), and we have proved
the multiple Mehler formula. 

Note that if we set s = 0 (or t = 0) in the multiple Mehler formula, we obtain the usual Mehler
formula for x − c and y − c (or x − d and y − d).
Our final result of this section is another proof of the orthogonality relation for the multiple
Hermites, by integrating the product of two generating functions. For usual Hermites, starting from
(3.1.4) we have, on the one hand,
Z  2   2 
1 t s
√ exp − + xt exp − + xs exp(−x2 /2) dx = exp(st),
2π 2 2
which means when integrating the product of the series, the integral of any pair Hn (x)Hm (x) where
n 6= m must be zero: on the other hand, that equation is
  
Z n n X sn tn
1 X t X s
√  Hn (x)   Hn (x)  exp(−x2 /2) dx = n! .
2π n! n! n! n!
n≥0 n≥0 n≥0

This proves orthogonality and the L2 norm simultaneously. We can do the same thing for the
multiple Hermites, although naturally it is more complicated.

Theorem 3.4.3. The orthogonality relation for multiple Hermite polynomials may be obtained
by integrating the product of two generating functions. Let f (s, t) be the exponential generating
function for the multiple Hermites (3.3.2):
∞ X
∞
sn tm s2 t2
X  
α,β
f (s, t) = Hn,m (x) = exp s(x − α) + t(x − β) − st − − .
n=0 m=0
n! m! 2 2

The orthogonality relation with respect to the first weight is implied by the equation
Z
(3.4.6) f (s1 , t1 )f (s2 , t2 ) dµ(1) = exp ((α − β)(t1 + t2 ) + (s1 + t1 )(s2 + t2 )) .
√
Here dµ(1) = exp(−(x − α)2 /2)/ 2π.

Proof. Let’s find the coefficient of sa1 tb1 sc2 td2 in the right-hand side of the equation above. Note
that the parameters of the polynomials are α and β, not c and d—those are nonnegative integers in
this proof.
Assume that a > c + d and write the exponential as exp(s1 s2 + s1 t2 + s2 t1 + t1 t2 + (α − β)t1 +
(α − β)t2 ) and label those six terms A, B, C, D, E, and F , respectively. We may split up the
exponential into a product of 6 exponentials, and we just need to pick out the correct terms from
each of them to get sa1 tb1 sc2 td2 .
In A, we can get sk1 sk2 , where 0 ≤ k ≤ min(a, c). Then in B we are forced to take sa−k
1 ta−k
2 to
a
get s1 in the product. (We are ignoring the coefficients here; they are not necessary at the moment.)
In C we must take sb−k
2 tb−k
1 to get the right exponent of s2 . At this point, we have sa1 sb2 tb−k
1 ta−k
2
and we just need to get the exponents of the ti ’s correct.
 3.4. IDENTITIES AND OTHER RESULTS 29

In D, we can take tj1 tj2 , where 0 ≤ j ≤ min(c − (b − k), d − (a − k)), and after that we can use
E and F as needed to get the correct exponents of t1 and t2 . The crucial step, then, is in D. Note
that j must satisfy
j ≤ d − a + k,
which can be rewritten as
a + j ≤ k + d.
Since k ≤ c,
a + j ≤ c + d,
but if a > c + d, no such nonnegative j can exist: the coefficient of sa1 tb1 sc2 td2 must be zero.
To derive the L2 norm, assume now that a = c + d. Following the same strategy, after choosing
terms from B, we have s1c+d sk2 t2c+d−k , but since we want td2 in the end, this means k must equal c.
In C, we use the constant term (s2 t1 )0 since we already have sc2 ; for similar reasons we take the
constant term in D and F . In E, we need tb1 , so we take b, which gives us a coefficient of (α − β)b .
α,β α,β
Equating the other coefficients that piled up along the way, we see that the integral of Hc+d,b Hc,d
is (c + d)!(α − β)b . 

Orthogonality with respect to the second weight is, of course, exactly similar.

3.4.1. For an arbitrary number of weights. The combinatorial model presented in this
chapter generalizes easily to an arbitrary number of weights:

Theorem 3.4.4. Let n = (n1 , n2 , . . . , nr ) and c = (c1 , c2 , . . . , cr ) be r-tuples of nonnegative

integers and arbitrary parameters respectively. Then Hnc (x) is the generating function for matchings
F
on i [ni ] in which edges have weight −1 and can go anywhere, and fixed points inside of [ni ] may
have weight x or −ci .

The exponential formula allows us to find the generating function for such polynomials:

Theorem 3.4.5. The exponential generating function for the multiple Hermites with respect to
r weights is
r
!
n
X t p2 (t) X
(3.4.7) Hnc (x) = exp −e2 (t) + + ti (x − ci ) ,
n
n! 2 i=1

where the sum extends over all r-tuples n = (n1 , . . . , nr ), tn denotes the product tn1 1 . . . tnr r , n!
denotes n1 ! . . . nr !, and p2 and e2 are the power sum and elementary symmetric functions of degree
2, respectively.

Proof. The connected components are fixed points, in any one of the r sets of vertices, which
have weight ti x or −ti ci ; inhomogeneous edges which have weight −ti tj ; and homogeneous edges
which have weight −t2i /2. The sum accounts for fixed points, the e2 symmetric function accounts
for inhomogeneous edges, but over-accounts for the homogeneous edges: it includes −t2i terms, so
we add half of p2 (t) to get the correct weights. 
 3.5. MULTIPLE LAGUERRE POLYNOMIALS 30

Equations (3.3.3) and (3.3.4) also extend to r weights; the analogue of (3.3.3) is
r  
X Y ni
(3.4.8) Hnc (x) = H|a| (x) (−ci )ni −ai ,
a i=1
ai

in which the sum runs over r-tuples a for which 0 ≤ (a1 , . . . , ar ) ≤ n in componentwise order. All
inequalities in this chapter involving vectors are with componentwise order.
Our first expression of the analogue of (3.3.4) is rather complicated; instead of having only one
choice for inhomogeneous edges (they must go from [n] to [m]), we now have 2r choices for where

inhomogeneous edges can go.

Let A(n) be the set of weak compositions of 0, 1, . . . , mini {ni } into 2r parts, in which the parts

are indexed by unordered pairs of distinct integers between 1 and r. Given a composition in A(n),
let
Xr
f (i) = ni − a{i,j} .
j=1,j6=i
Then we have

Proposition 3.4.6. The r-weight analogue of Equation 3.3.4 may be written as

 
r  
X Y Y ni
(3.4.9) Hnc (x) =  a{i,j} ! Hf (i) (x − ci ).
i=1
a{i,1} · · · a[
{i,i} · · · a{i,r}
a∈A(n) 1≤i<j≤r

The notation a[ {i,i} indicates an omitted term in the multinomial coefficient, and polynomials with
negative indicies are zero.

The proof is left to the reader; all the difficulty lies in expressing where the inhomogeneous
edges go. There is a much simpler way to express this, which relies on [25]: de Sainte-Catherine
and Viennot show that the integral of r Hermite polynomials enumerates perfect matchings with
no homogeneous edges in which edges have weight 1. This much more compactly expresses the
inhomogeneous edges portion of the above expression, and is especially fitting for an expression of
a multiple Hermite polynomial, so we may rephrase the above proposition as

Proposition 3.4.6 (version 2). For r weights, the analogue of Equation 3.3.4 is
r
! r  
X Z Y Y ni
0 c |a|/2
(3.4.9 ) Hn (x) = Hai (x) dµ (−1) Hni −ai (x − ci ).
a i=1 i=1
ai

in which the sum runs over r-tuples a for which 0 ≤ (a1 , . . . , ar ) ≤ n. In the integral, dµ is the
√
usual Hermite measure: exp(−x2 /2)/ 2π.

3.5. Multiple Laguerre polynomials

It may seem strange for a chapter on multiple Hermite polynomials to feature a section on
multiple Laguerre polynomials, but our results are so limited that this is the natural place to discuss
them.
The usual Laguerre polynomials are, in some sense, the second simplest classical orthogonal
polynomials. They have one parameter and sit immediately above the Hermite polynomials in the
 3.5. MULTIPLE LAGUERRE POLYNOMIALS 31

Askey-Wilson table [46]. They are orthogonal to the weight function xα e−x on the interval [0, ∞)
and satisfy the recurrence relation

(3.5.1) Lα α α
n+1 (x) = (x − (2n + α + 1))Ln (x) − n(n + α)Ln−1 (x).

Laguerre polynomials have a combinatorial interpretation as “partial injective functions” and their
moments count permutations by left-to-right maxima [65, 48, 34]. See page 60 for discussion of
left-to-right maxima in the context of associated Hermite polynomials.
These polynomials commonly appear in combinatorial problems related to permutations [10,
9, 30, 55, 54, 35, 36], but also can be thought of as matching polynomials for complete bipartite
graphs [25].
Van Assche and others have defined multiple Laguerre polynomials of the first and second kind
[63, 5]; we are using the first kind here. For two weights, the polynomials Lα,β
n,m (x) are orthogonal
to xα e−x and xβ e−x . The multiple Laguerres have the explicit formula [5, p. 3896]
n X m 
n + α n + m + β − i (−x)n+m−i−j
X  
Lα,β
n,m (x) = n!m!
i=0 j=0
i j (n − i)!(m − j)!
(3.5.2) n X m   
X n m
= (α + 1 + n − i)i (β + n − i + 1 + m − j)j (−x)n−i (−x)m−j .
i=0 j=0
i j

Before we state the combinatorial interpretation of the above formula, we need a lemma of Foata
and Strehl about the rising factorials:

Lemma 3.5.1 ([34], Lemma 2.1). The rising factorial (α + 1 + k)n−k is the generating function
for injective functions from A, a fixed subset of [n] of size n − k, to [n], in which each cycle of f is
weighted α + 1.
In particular, (α + 1)n is the generating function for permutations of [n] weighted either by
number of cycles or by left-to-right maxima.

Proof. Fix k and consider values of n ≥ k: for n = k or n = k + 1 the result is clear. If the
result is true for a particular n: take such a function f from A (of size n − k) to [n]. Think of f as
a directed graph on n vertices, with the vertices of A numbered 1 to n − k. There are four different
ways we can add a new vertex 0 to the directed graph:
We can make 0 its own cycle, thus multiplying the weight of f by α + 1. Or, for each of the
n − k edges whose source is in A, we can split the edge with vertex 0. This leaves the number of
cycles unchanged. For each of the k vertices outside of A, if the vertex is at the end of a path, we
can add vertex 0 to the beginning of that path, and if the vertex is not at the end of a path, we can
add 0 to A and draw an edge from 0 to the vertex. This too leaves the number of cycles unchanged.
See Figure 3.5.1 for pictures of these four operations.
Each such function for n + 1 and k can be uniquely obtained by one of the above options,
which collectively multiply the previous generating function by α + 1 + n, which yields exactly
(α + 1 + k)n+1−k .
 3.5. MULTIPLE LAGUERRE POLYNOMIALS 32

The final claim, that weighting permutations by cycles or left-to-right maxima gives the same
generating function, follows from the Foata-Schützenberger fundamental transform ([57, Prop. 1.3.1]
or [33, pp. 13-15]). 

(nothing) i, not in A
i, not in A

0
0
0 i, not in A
i, not in A
0

Figure 3.5.1. The four ways to add a new vertex to a partial injective function
described in Lemma 3.5.1. The top row is “before”, the bottom row “after”. Adding
the cycle multiplies the weight of the function by α + 1; the other three operations
leave the weight unchanged, and there are n total ways to do them.

With that lemma in hand, we have the following combinatorial interpretation:

Theorem 3.5.2. The multiple Laguerre polynomials of the first kind Lα,β n,m (x) may be expressed
as the generating function for weighted 4-tuples (A, B, f, g), in which A (respectively B) is a subset
of [n] (resp. [m]) and f (resp. g) is an injective function from A to [n] (resp. B to [m]). Let
“ cyc(f )” denote the number of cycles of such a function. Then
X
(3.5.3) Lα,β
n,m (x) = (α + 1)cyc(f ) (β + n − |A| + 1)cyc(g) (−x)n−|A| (−x)m−|B| .
(A,B,f,g)

Proof. This interpretation is clear from Lemma 3.5.1 and the second line of equation (3.5.2).


An example of such a 4-tuple is shown in Figure 3.5.2.

While this interpretation generalizes that of the usual Laguerre polynomials, unfortunately at
this time we cannot prove the orthogonality relation in a combinatorial way. The main difficulty is
that the weight of cycles of g depend on the size of A, so if we integrate with respect to the first
measure, which effectively replaces vertices weighted x with permutations with cycles weighted α+1,
we cannot change the sign of a cycle inside of [n] without changing the weight of the cycles inside
[m]. On the other hand, if we integrate with respect to the second measure—its effect is similar to
the first, but with β + 1—we have cycles inside [m] whose weight may or may not depend on the
size of A, so a simple sign flipping involution will not work.
A combinatorial proof of orthogonality will either involve a much more complicated involution,
or perhaps will use a model based on complete bipartite graphs.
 3.5. MULTIPLE LAGUERRE POLYNOMIALS 33

[n] [m]
1 2
7 4
5 6 4 6
2
5 1 3
A 3 B

Figure 3.5.2. A directed graph representation of (A, B, f, g) for n = 6 and m = 7.

Here A = {1, 2, 3, 5, 6} and B = {1, 4, 5, 7}. The two cycles of f contribute weight
(α + 1)2 , vertices 4 in [n] and 2, 3, and 6 in [m] have weight x, and g’s single cycle
has weight β + 2. This configuration contributes (α + 1)2 (β + 2)x4 to Lα,β 6,7 (x).
 CHAPTER 4

Nearly-diagonal Chebyshev polynomials

4.1. Definition

The nearly-diagonal Chebyshev polynomials U

en (x) are defined by the four-term recurrence re-
lation

(4.1.1) en+1 (x) = (x − bn )U

U en − U
en−1 (x) − θn U
en−2 (x),

where  
c n even c − d n even
bn = and θn =
d n odd 0 n odd
and we continue our convention that polynomials with negative indices are zero and U
e0 (x) = 1.
These polynomials arose in a peculiar way. So far it seems that one decides to study a particular
set of multiple orthogonal polynomials, and then considers the corresponding set of nearly-diagonal
polynomials—this is the typical approach often seen in van Assche’s work, for example. In this case
we began with the nearly-diagonal polynomials and have worked back to the full set of MOP. We
know from Remark 2.4.3 that one cannot use two copies of the same measure to define a set of MOP,
so an obvious thing to try is shifting the measure: given a measure µ(x), try using µ(x − c) and
µ(x − d) to define a set of MOP. We may think of c and d as either distinct real numbers or formal
parameters. We started with the Chebyshev polynomials (c = 0 in (4.1.2) below), shifted x to x − c,
and then worked out a four-term recurrence that seemed to have the same moments as those shifted
Chebyshev polynomials.
Those four-term polynomials are the nearly-diagonal Chebyshevs described above. We began
with those polynomials, and have worked backwards to find a corresponding set of MOP.
These nearly-diagonal polynomials have the same moments as shifted Chebyshev polynomials
of the second kind:

(4.1.2) Un+1 (x) = (x − c)Un (x) − Un−1 (x).

When c = 0, these polynomials are obtained by plugging x/2 into the traditional normalization of
Chebyshev polynomials of the second kind, which are defined by Un (cos θ) = sin ((n + 1)θ) / sin θ.
The combinatorial model of these polynomials for the c = 0 case was developed in [25]: Un (x) is
the generating function for matchings on [n] in which each vertex is either fixed and has weight x, or
is connected by an edge of weight −1 to an adjacent vertex. Vertices can have degree at most 1. This
model may also be viewed as the matching polynomial of a path with n vertices. These polynomials

34
 4.1. DEFINITION 35

are also called Fibonacci polynomials since one obtains the Fibonacci numbers if x = 1 and the
weight of the edges is changed to +1.
The moments for these polynomials when c = 0 are Catalan numbers: µ2n = 2n


n /(n + 1),
µ2n+1 = 0, and we interpret them as noncrossing matchings. Integrating a product of two poly-
nomials yields paired matchings with no crossings; the orthogonality involution finds the leftmost
edge connecting two adjacent vertices and swaps its color. (In a nonempty complete noncrossing
matching, every edge either connects two adjacent vertices, or nests such an edge.) When we allow
c to be a parameter, the polynomials (4.1.2), their moments, and orthogonality satisfy the following
the model:

Theorem 4.1.1. The polynomials (4.1.2) are the generating function for matchings of [n] in
which each vertex is either fixed and has weight x or −c, or is connected by an edge of weight −1 to
an adjacent vertex. Every vertex has degree at most 1.
The corresponding moments are generating functions for noncrossing incomplete matchings in
which fixed points have weight c and edges weight 1.

Proof. That the “x or −c” model describes the polynomials is clear; we have just shifted the
traditional three-term recurrence by c. The orthogonality involution proceeds as above, except that
we must find the leftmost object that is either a fixed point, or an edge connecting adjacent vertices,
and flip its color. 

Let’s see what the model is for the nearly-diagonal Chebyshevs of (4.1.1). We will also think
of these polynomials as the generating functions for matchings, but we will need to stretch our
definition of what a matching is. First we need two definitions.

Definition 4.1.2. A double edge in a matching on [n] is two edges, one connecting i to j, the
other connecting j to k, with i < j < k. We think of this configuration as one edge adjacent to
three vertices. We will refer to the left arch and right arch of a double edge. We will also refer to
usual edges as single edges if necessary to remove ambiguity. In this chapter, “edge” will refer to
both single and double edges.

Definition 4.1.3. The nesting parity of an edge or fixed point in a matching is the parity of
the number of edges nesting the object, in which double edges contribute 1 if the object if it is under
the left arch of the double edge, and contribute 2 if it is under the right arch.

For example, the nesting parity of the fixed point (vertex 4) in the matching 16|235|4 is odd,
and the nesting parity of the fixed point in the matching 16|245|3 is even. One could also think of
these objects as weighted set partitions into blocks of size 1, 2, or 3, but we prefer to use a loose
definition of matching to emphasize the connections with usual Chebyshev polynomials.

Theorem 4.1.4. The polynomials U

en (x) are the generating function for noncrossing matchings
consisting of fixed points, single edges connecting two adjacent vertices, and double edges that connect
three adjacent vertices. They are weighted as follows:
• fixed points may have weight x, or weight −c if the vertex is odd, or weight −d if the vertex
is even;
 4.1. DEFINITION 36

• single edges have weight −1;

• double edges have weight d or −c if their middle vertex is even, otherwise have weight 0.
We will interpret this as meaning double edges centered on odd vertices are forbidden.
The first moments L1 (xn ) are noncrossing matchings with fixed points, single and double edges
such that:
• fixed points have weight c if they have even nesting parity and weight d if they have odd
nesting parity,
• regular edges have weight 1, and
• double edges must have even nesting parity and may have weight c or −d.

Before we begin the proof, let us make clear some notation about the four sets of moments in this
chapter. As in Corollary 2.3.4, there are two sets of moments for the nearly-diagonal polynomials;
we’ll call those the first moments and second moments and denote them L1 (xn ) and L2 (xn ). The
shifted Chebyshev polynomials of (4.1.2) will be the x-axis polynomials Un,0 (x) in the full set of
(1)
MOP, and we’ll refer to their moments as the x-axis moments and denote them µn . In Section 4.5
(2)
we’ll describe the y-axis polynomials and the corresponding y-axis moments denoted µn .
In Theorem 4.3.1 and Theorem 4.5.5 we will address orthogonality with respect to the second
moments.

Proof. That the polynomials are described by the above combinatorial model is clear from the
en+1 (x), consider vertex n + 1. It can be a fixed point of weight x, −c, or −d,
recurrence: to build U
according to whether n + 1 is odd or even; it can connect with a single edge to vertex n; or, if n + 1
is odd, we can have a double edge connecting the previous two available vertices. Then one may fill
in the remaining vertices with such a matching on n, n − 1, or n − 2 vertices, as appropriate.
The model for the moments arises from Viennot’s lattice path model for the moments of orthog-
onal polynomials. In Section 2.3 we learned that for a four-term recurrence, we need to consider
generalized Motzkin paths, in which double-down or south-southeast edges that go (1, −2) are al-
lowed. Here, the nth moment is the generating function for weighted lattice paths with such edges
that go from the origin to (n, 0), in which NE edges have weight 1; E edges have weight c or d if
they are at even or odd height, respectively; SE edges have weight 1; and SSE edges have weight
c − d or 0 if they start at even or odd height, respectively.
The usual bijection from Dyck paths to noncrossing matchings extends easily to a weight-
preserving bijection from the above lattice paths to matchings weighted as described in the statement
of the theorem. Double-down edges reduce the height of the path by two, so we must somehow
connect to the previous two vertices. This is what motivated the definition of double edge. See
Figure 4.1.1 for a picture of one of the matchings for the moment L1 (x10 ). 

Now that we’ve described the combinatorial models for the polynomials and their moments, let’s
consider the orthogonality relation.

Theorem 4.1.5. The polynomials (4.1.1) are 2-orthogonal to the above moments:
 
(4.1.3) L1 Uen (x)U
em (x) = 0 if n > 2m.
 4.1. DEFINITION 37

d d c

Figure 4.1.1. A matching of weight c2 d2 that is part of L1 (x10 ); it corresponds to

the path NE, E, S, NE, E, NE, S, NE, E, SSE with the SSE edge given weight c.

and
 
(4.1.4) L1 U em (x) = (c − d)m .
e2m (x)U

Proof. In light of the fact that the moments of the nearly-diagonal polynomials coincide with
the moments of the x-axis polynomials (Theorem 2.4.1), we will show 2-orthogonality with respect
to the moments of Theorem 4.1.1. A combinatorial proof that the moments of Theorem 4.1.4 and
Theorem 4.1.1 coincide is rather involved and is presented in Section 4.2. Assuming the result of
that section, we can easily show (4.1.3).
Integrating the product U en (x)U
em (x) yields paired matchings with homogeneous black edges
from the polynomials and green edges from the moments. Apply the following involution to the set
of paired matchings:
Find the leftmost single edge that connects adjacent vertices, and flip its sign. If
there is no such edge, find the leftmost fixed point on an odd vertex and flip its
sign. Configurations not canceled by those operations are those with:
• inhomogeneous green edges;
• no odd fixed points, and even fixed points of weight c or −d;
• green edges of weight 1 connecting nonadjacent vertices, so because of the
above fact, they connect consecutive odd vertices;
• fixed points of weight c or −d on even vertices;
• double edges of weight d or −c connecting adjacent vertices whose middle is
on an even vertex.
Find the leftmost fixed point or double edge, and, if it’s a fixed point, pull down
the single edge nesting it and make it a double edge of the opposite weight; or, if
it’s a double edge, pop up the middle and give the new fixed point the opposite
weight.
This involution cancels any paired matching with a homogeneous edge, and since if n > 2m there
must always be a homogeneous edge in [n] or a fixed point nested by a homogeneous edge in [n], the
above involution will cancel all the paired matchings in such cases, which proves the orthogonality
relation (4.1.3).
The L2 norm of (4.1.4) follows from the above involution; the paired matchings not canceled
by the involution are those in which vertex i ∈ [m] connects to vertex 2m − 2i + 1 in [2m], and the
even vertices in [2m] may have weight c or −d. 
 4.2. EQUIVALENCE OF MOMENTS OF NEARLY-DIAGONAL AND X-AXIS CHEBYSHEVS 38

Note that the L2 norm of (c − d)m is exactly what we’d expect from the generalization of
Viennot’s paths-and-pavings involution discussed in Section 2.3.1: the only term left over in the
integral of U
e2m (x)U
em (x) has two empty pavings and a path of length 3m that begins with 2m NE
steps and ends with m SSE steps, each of which has weight c − d.
In general, for a four-term sequence of polynomials, there is very little we can say about the
integral of Pn (x)Pm (x) when m ≤ n < 2m, but with our nearly-diagonal Chebyshevs, we can
describe the integral of any pair of those polynomials:
Theorem 4.1.6. For 0 ≤ k ≤ n, with respect to the first moments we have
  bk/2c
X bn/2cn − dk/2e
(4.1.5) L Ue2n−k (x)U
en (x) = (c − d)n−k+2j .
j=0
j bk/2c − j

Note that k = 0 is the L2 norm, and that if we interpret an empty sum as zero, we can actually
use any integer k ≤ n. We needn’t consider any k greater than n by symmetry.

Proof. Think of U e2n−k (x) as on the left and U

en (x) on the right. The orthogonality involution
applied to the integral of U
e2n−k (x)Uen (x) cancels any paired matching with a homogeneous edge or
fixed point on an odd vertex, so every odd vertex must be incident to a single green inhomogeneous
edge. Since 2n − k ≥ n, there can be at most n edges going from right to left, and since we must
hit all of the n − bk/2c odd vertices on the left, there must be at least that many edges going from
right to left. That explains the bounds on the sum.
Note that if we choose sets of vertices on the right and left that will be incident to edges going
between them, there is only one way to connect those two sets of vertices because of the noncrossing
condition.
On the right, choose j of the bn/2c even vertices to remain fixed points, so n − j edges go to
the left. All the odd vertices must be hit by an edge, which leaves n − j − (n − bk/2c) = bk/2c − j
even vertices that must get hit. We choose that many even vertices out of the n − dk/2e total even
vertices on the left, and match the left and right sets up. There are 3n − k vertices, 2(n − j) of
which are taken up by edges, leaving n − k + 2j even vertices which may have weight c or −d. This
accounts for the three factors in each term of the sum. 

This theorem may also be proved using (4.5.2) and the equivalence of the first moments and
x-axis moments: the left side of (4.1.5) becomes
  
Z X  X bn/2c
n − dk/2e
 (c − d)j U2n−k−j (x)  (c − d)j Un−j (x) dµ
j j
j≥0 j≥0

Using orthogonality and the L2 norm of the Un ’s (which is 1), we immediately get the right side of
(4.1.5). See Theorem 4.5.5 for the analogue of this theorem for the second moments.
When k = n, one obtains a polynomial in c − d whose coefficients are the squares of the entries
of row bn/2c of Pascal’s triangle.

4.2. Equivalence of moments of nearly-diagonal and x-axis Chebyshevs

The aim of this section is a combinatorial proof of

 4.2. EQUIVALENCE OF MOMENTS OF NEARLY-DIAGONAL AND X-AXIS CHEBYSHEVS 39

−d c

Figure 4.2.1. Two four-term matchings of weight ±c3 d2 that illustrate the core
idea of the CD involution. Recall that fixed points beneath the left arch of a double
edge have weight d, and those under the right arch have weight c. If the double
edge has weight c, we move its center to the left one vertex and change its weight
to −d. This creates a new fixed point of weight c, so the weight is preserved but
sign is reversed. For a double edge of weight −d, we move the center to the right
one vertex.

Theorem 4.2.1. The generating functions for the following two kinds of noncrossing matchings
of [n] are equal:
• Incomplete matchings with single edges and fixed points of weight c allowed. In this sec-
tion, we’ll call those “three-term matchings” since they come from the shifted Chebyshev
polynomials that satisfy a three-term recurrence.
• Incomplete matchings as in Theorem 4.1.4, with fixed points and double edges, with fixed
points of weight c or d if they have even or odd nesting parity, respectively; double edges of
weight c or −d with even nesting parity. We’ll call such matchings “four-term matchings”.
(1)
That is: µn for the shifted Chebyshevs (4.1.2) equals L1 (xn ) for the nearly-diagonal Chebyshevs
(4.1.1).

In Section 4.5, we shall see that the shifted Chebyshevs are the x-axis polynomials for the nearly-
diagonal Chebyshevs; given that result, Theorem 2.4.1 implies that these two generating functions
coincide, but in this section we will prove this combinatorially. The bulk of our work will be in
describing a sign-reversing involution on the four-term matchings that shows that the generating
function for those moments is a polynomial in c with nonnegative integer coefficients—that is, all
the −d edges and d fixed points cancel. Hence we’ll call this the CD involution for “cancel d’s”.
Then we will describe a weight-preserving bijection from the four-term matchings whose weight
is a power of c to the 3-term matchings.

4.2.1. The CD involution. The basic idea of this involution is very simple: we need to cancel
(almost all) double edges, so to cancel such an edge, we’ll move its center to the left or right and
change its weight. Figure 4.2.1 demonstrates this process.
We need two definitions before we begin the formal description of the involution.

Definition 4.2.2. An edge e in a matching immediately nests an edge or fixed point if it nests
that edge or fixed point, and no other edge nested by e does so.

Definition 4.2.3. An edge in a four-term matching has property I if the edge does not nest
any edge of weight −d and does not nest any edge that nests a fixed point of weight d.

We use “I” as a mnemonic for “innermost” since such edges do not nest certain other edges; they
are innermost with respect to the described conditions. With this definition in hand, we can define
our first involution: given a four-term matching, find the leftmost edge that satisfies property I and
falls into one or both of the following categories:
 4.2. EQUIVALENCE OF MOMENTS OF NEARLY-DIAGONAL AND X-AXIS CHEBYSHEVS 40

A: the edge has weight −d;

B: the edge does not have weight −d, and it nests a fixed point of weight d.

Any matching that has no such edge is left fixed by this involution. In particular, note that any
matching that has no edge or fixed point of weight d will be fixed by the involution.
Assume M is a four-term matching in case A. Call the selected edge E. (It is a double edge,
since it has weight −d.) The CD involution works as follows on M : if there is a fixed point of weight
c immediately nested by E, move the center of E to the leftmost such fixed point, change the weight
of E to c, and make the old center of E a fixed point, which must have weight d.
When we move the center, we may hop over some edges, and we will then flip their nesting
parity. We must account for this, and we do so with the “parity flip fix”, a sign- and weight-
preserving involution described in Section 4.2.2.
We have preserved the weight and the reversed the sign of the matching since we removed the
−d weight from edge E but added a new fixed point with weight d, and removed a fixed point of
weight c but added weight c to E. Observe that since E no longer has weight −d and now nests a
fixed point of weight d, it falls into case B.
It may be the case that E does not nest any fixed point of weight c. In this case, simply pop
up the middle of E—replace it with a single edge with the same endpoints, and make the old center
of E a fixed point, which must have weight d because of nesting parity requirements. Do the parity
flip fix on all edges that were previously nested by the right arch of E. This also preserves weight,
reverses sign, and results in a matching in case B.
Now assume that M is a four-term matching in case B and again let E be the selected edge,
which could be a single or double edge.
If E has weight c—i.e., it’s a double edge—it must immediately nest a fixed point of weight
d, so do the opposite of what we did for the first part of case A above: move the center leftwards
to the rightmost such fixed point, change its weight to −d, and do the parity flip fix on any edges
that were originally nested by the right arch of E but are now nested by the left arch. Weight is
preserved, sign is reversed, and the result is a matching in case A.
If E has weight 1—i.e., it’s a single edge—replace E with a double edge of weight −d with the
same endpoints and a center at the rightmost fixed point of weight d nested by E. Do the parity
flip fix on any edges now nested by the right arch of E. We replaced an edge of weight 1 and a fixed
point of weight d with a double edge of weight −d, and now have a matching in case A.
Figures 4.2.2 and 4.2.3 show two examples of the CD involution. In both figures, the top
matching is in case A, the bottom matching in case B. In Figure 4.2.2, we move the center of the
double edge between vertices 5 and 11, and perform the parity flip fix (described in the next section)
on vertices 6 through 10. Figure 4.2.3 shows the “pop up the center” operation; there we perform
the parity flip fix on vertices 9 through 12, although in that case the submatching is unchanged
when the parity flip fix is applied.
This process takes matchings in case A and produces matchings in case B, and vice versa, but a
careful reader may wonder if the same edge will be chosen after performing the CD map described
above. However, because the above process preserves property I, and we always select the leftmost
 4.2. EQUIVALENCE OF MOMENTS OF NEARLY-DIAGONAL AND X-AXIS CHEBYSHEVS 41

−d c

Figure 4.2.2. The CD involution and parity flip fix applied to matchings of weight ±c4 d2 .

−d

Figure 4.2.3. The CD involution and parity flip fix applied to matchings of weight
±cd2 . The center of the double edge on vertices 1, 8, 13 cannot be moved any farther
right, so we pop up the middle.

edge with property I, we will in fact choose the same edge. That is, applying CD twice to a matching
will always yield the same matching.
Since every four-term matching whose weight is a multiple of d will have at least one edge in case
A or case B, we have proved, modulo the parity flip fix of the next section, the following statement.

Theorem 4.2.4. The generating function for four-term matchings on 2n vertices, weighted as
in Theorem 4.1.4, is a polynomial in c with nonnegative coefficients. Equivalently, we may say that
the first moments of the nearly-diagonal Chebyshev polynomials are noncrossing matchings with fixed
points, and single and double edges such that
• fixed points and double edges have weight c and even nesting parity, and
• regular edges have weight 1.

In Section 4.2.3 we will show a weight-preserving bijection between the matchings described
above and the 3-term matchings.
 4.2. EQUIVALENCE OF MOMENTS OF NEARLY-DIAGONAL AND X-AXIS CHEBYSHEVS 42

4.2.2. The parity flip fix. It’s the dance sensation that’s sweeping the nation. Actually, no,
it’s a weight- and sign-preserving involution on matchings that accounts for a flip in nesting parity
when the CD involution moves the center of a double edge. The need for this is demonstrated with
the matching 127|345|6, with the outer double edge of weight −d. The fixed point has weight c,
so the CD involution will change that matching into 167|2|345. But this matching is not a legal
four-term matching, since the double edge connecting vertices 3, 4, and 5 has odd nesting parity.
The parity flip fix will be applied to edges such as the 345 edge, and yield a new configuration whose
weight and sign is the same as before.
The parity flip fix (abbreviated PFF) is a recursive procedure applied to each edge E and
everything nested by it when the CD involution changes its nesting parity. The PFF does nothing
to a single edge that nests only single edges, since the weight and sign of such a configuration does
not depend on its nesting parity; we’ll describe what it does to an edge that is a double edge, nests
a double edge, or nests a fixed point. Note that, because of property I, all the double edges and
fixed points we will consider for the PFF have weight c.
Assume that E is such an edge, and that it is a double edge of weight c. Its nesting parity was
originally even, and now it is odd. Replace E with a single edge with the same endpoints, and make
the old center of E a fixed point, which under the new nesting parity will have weight c. Apply the
PFF to any edges that were previously immediately nested by the left arch of E. This preserves
weight and sign of the configuration formed by E and everything it nests.
If E is a single edge that immediately nests a fixed point, do the opposite of the above. The
nesting parity of E must have originally been odd (since it nests a fixed point of weight c), and now
it is even. Replace E with a double edge of weight c by adding a center at the leftmost fixed point
immediately nested by E—this too preserves the weight of the edge. Then apply the PFF to every
edge under the new right arch of the new double edge.
If E is a single edge that immediately nests no fixed points, apply the PFF to each edge
immediately nested by E.
In all cases, we preserve the weight of the configuration, and since we go back and forth between
a double edge of weight c and a single edge immediately nesting a fixed point of weight c, we have
an involution. Figure 4.2.4 shows an example of the PFF.

4.2.3. Proof of Theorem 4.2.1. Now we may complete the proof of Theorem 4.2.1. Begin
with four-term matchings on n vertices: these have fixed points and double edges satisfying
• fixed points have weight c if they have even nesting parity and weight d if they have odd
nesting parity,
• regular edges have weight 1, and
• double edges must have even nesting parity and may have weight c or −d.
The CD involution leaves us with matchings whose weight is a power of c. We need only describe
a weight-preserving bijection f between these matchings and the three-term matchings of Theo-
rem 4.1.1. We define f as follows: it maps any four-term matching with no double edges to itself,
and if there are double edges, it pops up the middle (as in the parity flip fix) and gives all fixed
points beneath the new edge weight c.
 4.3. ORTHOGONALITY AND INTEGRALS WITH RESPECT TO SECOND MOMENTS 43

Figure 4.2.4. Two matchings of weight c4 that correspond to each other via the
parity flip fix. The top matching has even nesting parity, the bottom has odd
nesting parity. All fixed points and double edges have weight c.

One may easily show that f is injective by considering two distinct four-term matchings and
considering the leftmost point where they differ: after applying f , one matching will have a fixed
point at that vertex, and the other an edge; or both matchings will have a single edge at that vertex,
but one will nest a fixed point and the other won’t. This map is surjective since, given a three-term
matching, any single edge with even nesting parity that immediately nests a fixed point is the image
of a double edge with the same endpoints and a middle point attached to the leftmost fixed point
immediately nested by the single edge.
Starting with the four-term matchings of Theorem 4.1.4, the CD involution shows that all
matchings with double edges of weight −d and fixed points of weight d cancel. The f bijection
above showed that the resulting four-term matchings have the same generating function as the
three-term matchings of Theorem 4.1.1. This completes the proof of Theorem 4.2.1. 

4.3. Orthogonality and integrals with respect to second moments

As we know from Theorem 2.3.3, there is a second set of moments for polynomials satisfying a
four-term recurrence. The lattice path model for these moments and the natural generalization of
Viennot’s theory gives us generalized Motzkin paths that end at height 1. In this section we shall
show the orthogonality relation and analogue of Theorem 4.1.6. We postponed these theorems until
now because we need the CD involution. First, we describe the second moments.

Theorem 4.3.1. The nth second moment L2 (xn ) for the nearly-diagonal Chebyshev polynomials
is the generating function for noncrossing matchings on n+1 points, with single edges and fixed points
(but no double edges), where vertex n + 1 is the right endpoint of a single edge. They are weighted
as follows: single edges have weight 1 and all fixed points have weight c except those immediately
nested by the edge incident to vertex n + 1; those fixed points have weight d.

Proof. We know that the moments are the generating function for generalized Motzkin paths
of length n that end at height 1, weighted as described at the beginning of the chapter. By adding
a final SE edge to each such path, we have a bijection between matchings on n + 1 vertices with
fixed points, single edges, and double edges, weighted as in Theorem 4.1.4, where, because of the
 4.3. ORTHOGONALITY AND INTEGRALS WITH RESPECT TO SECOND MOMENTS 44

final SE edge, vertex n + 1 is the right endpoint of a single edge. Call that single edge incident to
vertex n + 1 the “fake edge”.
For any such matching on n + 1 vertices, consider the submatching on the vertices strictly to
the left of the fake edge. That submatching is exactly the kind of matching that contributes to the
first moments, and we may apply the CD involution. Therefore, to the right of the fake edge, we
have a submatching with single edges and fixed points weighted c.
Consider the submatching nested by the fake edge. Fixed points immediately nested by the fake
edge have odd nesting parity and must have weight d; there is nothing we can do to cancel such fixed
points.1 However, underneath each edge immediately nested by the fake edge, we have a matching
which again is exactly the type of matching contributing to the first moments, and we can apply the
CD involution. Hence underneath each such edge, we have a matching with only single edges and
fixed points weighted c. 

Now that we have a convenient description for the second moments, let’s address the orthogo-
nality relation:
Theorem 4.3.2. The orthogonality relation for the nearly-diagonal Chebyshevs with respect to
the second moments is
 
(4.3.1) L2 Uen (x)U
em (x) = 0

if n > 2m + 1.
Proof. When integrating the product U
en (x)U
em (x), we replace k fixed points with a matching
on k + 1 vertices—to fix this discrepancy, add a new vertex to the right of [m] and use that as the
rightmost vertex of the matching from the moments. For the purposes of this proof, that vertex will
be neither odd nor even, and the fake edge will not be considered homogeneous, even if it connects
to a vertex in [m].
For n > 2m + 1, the orthogonality follows from an involution similar to the one in the proof
of Theorem 4.1.4: our paired matchings consist of green fixed points which have weight c if not
immediately nested by the green fake edge, otherwise they have weight d. There are black fixed
points of weight −c or −d on odd or even vertices respectively. Black single edges must connect
adjacent vertices and have weight −1; black double edges must connect three adjacent vertices, be
centered on an even vertex, and have weight d or −c. Green edges may connect anywhere and always
have weight 1.
Apply the following involution:
Find the leftmost single edge connecting adjacent vertices and flip its color.
If there’s no such edge, find the leftmost black double edge or homogeneous
green edge connecting consecutive odd vertices. If it’s a double edge, pop up the
middle and make the nested fixed point weight c or −d according to the weight of
the double edge. If it’s a green edge, pull down the edge to make a double edge
and give the edge the appropriate weight.
1Consider L (x2 ) = c + d; the c term corresponds to the matching 1|23 and the d term corresponds to the matching
2
13|2. There’s nothing of weight −d nesting such fixed points; we can’t cancel them.
 4.4. GENERATING FUNCTIONS AND OTHER IDENTITIES 45

Note that the above two operations will cancel all paired matchings with a black edge. Now find the
leftmost fixed point which is on odd vertex and not immediately nested by the fake edge, or which
is on an even vertex, immediately nested by the fake edge, and flip the color of that vertex.
Are there any paired matchings with homogeneous green edges that are not canceled by the
above operations? No: any homogeneous green edge in a paired matching not canceled by the above
involution must connect adjacent vertices, connect consecutive odd vertices, connect consecutive
even vertices, or nest both even and odd vertices—or, it must nest such an edge. All of those
possibilities have been canceled, so if n > 2m + 1, there are no paired matchings left alone by the
involution, and we have proved orthogonality. 

What remains are paired matchings with inhomogeneous green edges in which the fake edge
connects to 1 ∈ [n], to m ∈ [m], or if m is even, to m − 1 ∈ [m]. All the fixed points are either on
an even vertex with weight c − d if not immediately nested by the fake edge, or on an odd vertex
with weight d − c if immediately nested by the fake edge. These polynomials don’t satisfy a genuine
recurrence relation of order 3.
The L2 norm of the polynomials with respect to this functional is zero; similar to the comment at
the end of the proof of Theorem 4.1.4, when we apply L2 to Ue2m+1 (x)U em (x), the Viennot involution
cancels everything but two empty pavings and a path that begins with 2m + 1 NE steps and ends
with m SSE steps—but each of those steps leaves from an odd height, and so those steps have weight
zero.

4.4. Generating functions and other identities

In this section we collect a number of results about the generating functions and identities
related to the nearly-diagonal Chebyshev polynomials.

Theorem 4.4.1. The nearly-diagonal Chebyshev polynomials of (4.1.1) have the generating
function
X
en (x)tn = 1 + (x − c)t + t2 + (d − c)t3
(4.4.1) U .
1 + (2 + dx − x2 + c(x − d))t2 + (1 + d2 − dx + c(x − d))t4
n≥0

Proof. This result follows from a relatively straightforward use of the recurrence relation and
solving a system of functional equations. Let p(t) denote the sum on the left side of (4.4.1). Since
bn and θn depend on whether n is even or odd, we will need to use sieving. (See the proof of
P e 2n
U2n+1 t2n+1 , where each sum runs over
Pe
Theorem 3.4.2.) Let peven (t) = U2n t and podd (t) =
nonnegative n, so that p(t) = peven (t) + podd (t). The recurrence relation for U
e2n is

U2n (x) = (x − d)U2n−1 (x) − U2n−2 (x)

so that if we multiply both sides of the above equation by t2n and sum over nonnegative n, we have

peven (t) − 1 = t(x − d)podd (t) − t2 peven (t).

 4.4. GENERATING FUNCTIONS AND OTHER IDENTITIES 46

(We subtract one from peven (t) because the recurrence doesn’t work for n = 0.) A similar process
with U
e2n+1 (x) yields

podd (t) = t(x − c)peven (t) − t2 podd (t) − t3 (c − d)peven (t).

We have two equations, linear in peven (t) and podd (t), so that it is easy to find
−1 − t2
peven (t) =
(−2 + c(d − x) − dx + x2 )t2 − (1 + d2 − dx + c(x − d))t4 − 1
(c − x)t + (c − d)t3
podd (t) = .
(−2 + c(d − x) − dx + x2 )t2 − (1 + d2 − dx + c(x − d))t4 − 1
Adding and simplifying the fractions yields p(t). 

Let’s now consider the generating function for the first set of moments: m(t) = L1 (xn )tn . By
P

Viennot’s general theory, the generating function for the moments is given by a continued fraction
expansion:
1 λ 1 t2 λ 2 t2 λ 3 t2
(4.4.2) ···
1 − b0 t− 1 − b1 t− 1 − b2 t− 1 − b3 t−
That continued fraction can be compactly described with the “δ” notation used by Viennot: m(t)
depends on the infinite sequence of values bn , λn , and θn , and he defines

δm(t, b0 , b1 , λ1 , b2 , λ2 , θ2 , b3 , . . . ) = m(t, b1 , b2 , λ2 , b3 , λ3 , θ3 , b4 , . . . ),

with δ k m(t) defined the k-fold composition: δ(δ · · · δ(m(t))). One should think of δ k m(t) as the
generating function for generalized Motzkin paths that begin, end, and stay at or above height k.
The moment generating function is then
1
m(t) = .
1 − b0 t − λ1 t2 δm(t) − θ2 t3 (δm(t))(δ 2 m(t))
The proof of the above equation amounts to thinking of the right side as a geometric series; see [53,
sections 1.2, 4.2.2]. In our case, the periodicity of bn and θn (and because λn = 1 for all n) allow us
to think of the moment generating function as m(t, beven , bodd , θeven , θodd ), and for the first moments
of the nearly-diagonal Chebyshev polynomials, we have
X
(4.4.3) L1 (xn )tn = m(t, c, d, c − d, 0) =
n≥0
1
.
1 − ct − t2 m(t, d, c, 0, c − d) − (c − d)t3 m(t, d, c, 0, c − d)m(t, c, d, c − d, 0)
That expression for the moment generating function is not very explicit, but because of Theo-
rem 4.2.1, we can also write (abbreviating m(t, c, d, c − d, 0) with m(t))
1
m(t) = .
1 − ct − t2 m(t)
One may solve this functional equation just as one does to derive the generating function of the
Catalan numbers, and we have
 4.4. GENERATING FUNCTIONS AND OTHER IDENTITIES 47

Theorem 4.4.2. The moment generating function for the nearly-diagonal Chebyshev polynomi-
als is
√
X
n n 1 − ct + 1 − 2ct − 4t2 + c2 t2
(4.4.4) L1 (x )t = m(t) = .
2t2
n≥0

Interestingly, the inverse of the generating function is very easy to describe:

Theorem 4.4.3. Let m(t) be the moment generating function for the nearly-diagonal Chebyshev
polynomials. The compositional inverse of t · m(t) is
t
(4.4.5) .
1 + ct + t2
Proof. By Theorem 4.2.1, the moments are the generating function for incomplete noncrossing
matchings with fixed points weighted c, or, equivalently, Motzkin paths in which NE and SE edges
have weight 1 and E edges have weight c. The Lagrange Inversion Formula [58, second proof of
Theorem 5.4.2] immediately implies the result.
One may also verify that replacing t by t · m(t) in (4.4.5) yields the identity function. 

This idea can be generalized to arbitrary rooted trees with a constraint on the degree of each
vertex: see, for example, sequence A036765 in [56].
The generating function for the second moments can also be found without too much difficulty,
since their recurrence coefficients are the same as those for the first moments except for b0 .

Theorem 4.4.4. The generating function for the second moments of the nearly-diagonal Cheby-
shev polynomials is
X 1
(4.4.6) m2 (t) = L2 (xn )tn = m(t) · t · ,
1 − dt − t2 m(t)
n≥1

where m(t) is the generating function for the first moments.

Proof. To the left of the fake edge, we have a first moments-style matching; that explains the
m(t) factor. We multiply by t to account for the left vertex of the fake edge (we artificially added
the right vertex of the fake edge, so that doesn’t get weight t). The generating function for the
matching underneath the fake edge is, according to (4.4.2),
1 t2 t2
···
1 − dt− 1 − ct− 1 − ct−
After the dt, we have a copy of m(t), so the generating function for those matchings is
1
.
1 − dt − t2 m(t)
Each matching for the second moments can be uniquely “factored” into a first-moment-style match-
ing (generating function m(t)), the left vertex of the fake edge (generating function t), and a match-
ing with nonnested fixed points weighted d and others weighted c (generating function (1 − dt −
t2 m(t))−1 ), so we multiply those three generating functions to find m2 (t). 
 4.5. THE CORRESPONDING MOP 48

4.5. The corresponding MOP

In this section we’ll discuss the MOP that give rise to the nearly-diagonal Chebyshev polynomi-
als. More precisely, we’ll describe the moments of the corresponding y-axis polynomials, and since
we already know the moments of the x-axis polynomials, that will determine the set of MOP.
Of course, in light of Corollary 2.4.2, it is somewhat disingenuous to talk about “the” multiple
(2)
OP corresponding to a set of nearly-diagonal polynomials, since we can choose µ1 to be anything
(2)
but c for our nearly-diagonal Chebyshevs. In this case, though, it is most natural to choose µ1 , and
hence b0 for the y-axis polynomials, to be d. That parameter is already part of our nearly-diagonal
polynomials, and because of that we’ll continue to refer to “the” MOP and y-axis polynomials.
Let Vn (x) be a set of OP defined by Vn+1 (x) = (x−bn )Vn (x)−Vn−1 (x), where b0 = d and bn = c
for n > 0. Note the close similarity to the shifted Chebyshev polynomials (4.1.2) at the beginning
of this chapter; these polynomials are identical to those except for b0 (where they necessarily differ).
These polynomials are the y-axis polynomials for the nearly-diagonal Chebyshevs:

Theorem 4.5.1. The polynomials Vn (x) defined by

(4.5.1) Vn+1 (x) = (x − bn )Vn (x) − Vn−1 (x),

where bn = c except for b0 = d, are the y-axis polynomials for the nearly-diagonal Chebyshev
polynomials.

More precisely: the nearly-diagonal polynomials for the set of MOP with Vn (x) as the y-axis
polynomials and Un (x) of (4.1.2) as the x-axis polynomials are exactly the nearly-diagonal Cheby-
shev polynomials we’ve studied in this chapter.
Our strategy for the proof is to first prove two identities for U
en (x) involving Un (x) and Vn (x),
and then use those identities to show that Uen (x) satisfies the correct orthogonality relations for the
corresponding moments.
The two identities express Uen (x) as a sum of the U ’s and V ’s and are interesting in their own
right:

Theorem 4.5.2. The nearly-diagonal Chebyshev polynomials may be written

X bn/2c
(4.5.2) Un (x) =
e (c − d)i Un−i (x)
i
i≥0

and
X bn/2c − 1
(4.5.3) U
en (x) = (c − d)i Vn−i (x).
i
i≥0

The first sum is valid for all nonnegative n, the second for n ≥ 2.

Proof. A simple but tedious way to prove these identities is to show that the sums on the right
satisfy the same recurrence relation (4.1.1) as the polynomials on the left. Let S(n) be the sum on
the right side of (4.5.2) and consider

(x − c)S(2n) − S(2n − 1) − (c − d)S(2n − 2).

 4.5. THE CORRESPONDING MOP 49

One needs only some reindexing, Pascal’s relation, and the recurrence for Un (x) to show that the
above expression equals S(2n + 1). The S(2n − 1) case is similar, as is the sum for Vn (x).
Let’s prove these identities using the combinatorial models for these polynomials. The model
for Vn (x) is identical to that of Un (x), except that vertex 1, if it is a fixed point, has weight x or −d.
For the purposes of this proof, let’s think of the double edges in U en (x), which have weight d or
−c and must be centered on an even vertex and connect three consecutive vertices, as a single edge
of weight −1 connecting two consecutive odd vertices, with the nested fixed point of weight c or
−d. That arrangement obviously has the same weight and because all the other edges must connect
consecutive vertices, it’s clear that these interpretation gives the same polynomials.
The binomial coefficients and powers of c − d in the sum on the right side of (4.5.2) suggest that
the sum may be interpreted as the generating function for matchings of [n] with edges connecting
consecutive vertices in which some even vertices “don’t count” for the purpose of consecutivity. More
precisely, we have matchings of [n] satisfying:

• odd vertices are colored black and even vertices may be colored black or green;
• black vertices have weight −c, green vertices may have weight c or −d;
• edges have weight −1 and must connect consecutive black vertices.

An obvious sign-reversing involution (find the leftmost nonnested even fixed point, and if it’s green
of weight c, change it to black, or vice versa) cancels any such matching with a nonnested green
fixed point of weight c or a black even fixed point. The matchings that remain are exactly those
en (x): ignore the color, and even fixed points have weight −c, odd fixed points have weight −d,
for U
single edges connect consecutive vertices and have weight −1, and double edges have weight c or
−d—here we’re thinking of them as a single edge which nests an even fixed point.
The second sum is slightly trickier. We first note that matchings for U
en (x) that begin with a
fixed point and a single edge (which together have weight c) cancel matchings that begin with a
double edge of weight −c. So it changes nothing to restrict U
en (x) to those matchings that begin
with two fixed points, a single edge, or a double edge of weight d—equivalently, an edge of weight
−1 connecting vertices 1 and 3 nesting a fixed point of weight −d. In this situation, we may just as
well think of a fixed point on vertices 1 and 2 as having weight −d and −c, respectively.
We interpret the sum as before: we have matchings on [n] in which some of the even vertices—
except vertex 2—have been colored green. We get matchings weighted as described above, apply the
same involution, and the uncanceled matchings are just as above except for the weights of vertices
1 and 2 as fixed points.
The remaining matchings are exactly those for Uen (x) (with the conventions described above)
except for those matchings that begin with a fixed point of weight −d and a single edge, but those
matchings clearly correspond with those that begin with a double edge of weight d. Altogether we
have a weight-preserving bijection between the two sets of matchings. 

The sums (4.5.2) and (4.5.3) make Theorem 4.5.1 easy to prove:
 4.5. THE CORRESPONDING MOP 50

Proof of Theorem 4.5.1. To prove that the Vn polynomials are the correct y-axis polyno-
mials, we need to show that U
en (x) has the appropriate orthogonality relation with respect to the
(2)
corresponding moments µn .
Both U
e2n (x) and U
e2n+1 (x), as polynomials for the full set of MOP, have second coordinate equal
to n, so we need to show that those polynomials are both orthogonal to xk for 0 ≤ k ≤ n − 1 with
respect to the y-axis moments. But the sum (4.5.3) makes that obvious, since both polynomials are
a linear combination of Vi (x)’s with i at least n + 1. 

In a similar manner, the sum (4.5.2) provides an alternate proof of the orthogonality of U
en (x)
with respect to the x-axis moments.
Having written the Ue ’s in terms of the U ’s and V ’s, it is nice to write the U ’s and V ’s in terms
of each other.

Theorem 4.5.3. For n ≥ 0,

(4.5.4) Vn (x) = Un (x) + (c − d)Un−1 (x),

X
(4.5.5) Un (x) = (d − c)i Vn−i (x).
i≥0

Proof. The first sum is much like Theorem 3.4.1: Un (x) includes matchings with the first vertex
weight −c, which are canceled by subtracting −c times the generating function for the matchings
on vertices 2 through n, and then we need to add in a similar term so that we have matchings in
which vertex 1 is fixed and has weight −d.
The second sum can be proved by showing both sides satisfy the recurrence for Un (x), or by
an iterated process like the one described in the previous paragraph: we start with Vn (x), but need
to cancel the fixed point of weight −d on vertex 1, so we add in d − c times Vn−1 (x)—but that
effectively adds in a fixed point of weight −d on vertex 2. Figure 4.5.1 shows this idea. 

Vn (x)

Vn−1 (x)

Vn−2 (x)

Vn−3 (x)

etc

Figure 4.5.1. The iterative cancel-and-correct process of (4.5.5). Each vertex

above has weight d − c; first we cancel and correct the weight of vertex 1 with
(d − c)Vn−1 (x), but then we need to cancel the resulting matchings with weight −d
on vertex 2, so we add (d − c)2 Vn−2 , and so on.
 4.5. THE CORRESPONDING MOP 51

Finally, to be extremely thorough, we can write the U ’s and V ’s in terms of the nearly-diagonal
polynomials and obtain the inverse of Theorem 4.5.2.

Theorem 4.5.4. For n ≥ 1,

n  
X n − 1 − bk/2c
(4.5.6) Un (x) = (−1)n−k (c − d)n−k U
ek (x),
b(k − 1)/2c
k=1

and for n ≥ 3,
n  
X n − 3 − bk/2 − 1c
(4.5.7) Vn (x) = (−1)n−k (c − d)n−k U
ek (x).
b(k − 3)/2c
k=1

e1 (x) + (c − d)U
The zeroth polynomials are all 1, V1 (x) = U e0 (x), and V2 (x) = U
e2 (x).

Proof. Consider the sums connecting U

en (x) and Un (x). We can phrase (4.5.2) with a matrix
equation:
   
U
e0 U0
U1  U1 
e   
  = M  ,
U
e2  U2 
   
.. ..
. .
where M is an infinite matrix with coefficients from (4.5.2), with rows and columns indexed starting
from 0. The matrix equation is well-defined because each row contains only finitely many nonzero
entries. The sum (4.5.6) can be expressed with a similar matrix equation, but with Uk ’s on the left
and Uek ’s on the right. The indexing in that matrix, since (4.5.6) is only valid for n ≥ 1 and begins
the sum with k = 1, is not quite the same as for M ; to fix that, add a new first row and column
of all zeros except for the (0, 0) position, which should be 1. Call that matrix in that equation M 0 .
If we can show that M M 0 equals the infinite identity matrix, we are done. We will find the (n, m)
entry of M M 0 and use the Chu-Vandermonde sum to see that it must equal 0 if m 6= n and 1 if
m = n.
The nth row of the M matrix is
  
bn/2c
(c − d)n−k
k − dn/2e k≥0

and by considering n as the summation index in (4.5.6), the mth column of M 0 is

  
k − bm/2c − 1
(d − c)k−m .
dm/2e − 1 k≥0

In the matrix product, the (n, m) entry is

X  bn/2c k − bm/2c − 1
n−m
(c − d) (−1)k−m .
k − dn/2e dm/2e − 1
k≥0

Observe that if n = m, the only nonzero term in the sum is when k = n and the result is 1. In
general, by using the Chu-Vandermonde summation, the above sum equals
 
m n−m dn/2e − dm/2e − 1
(−1) (c − d) .
−bn/2c + dm/2e − 1
 4.5. THE CORRESPONDING MOP 52

If n < m, the binomial coefficient is zero because the “numerator” is less than zero; if n > m, we
get zero because of the denominator. This means the product M M 0 is the identity matrix.
The proof for (4.5.7) is similar. The matrix M has entries from (4.5.3) and its nth row is
  
bn/2c − 1
(c − d)n−k
k − dn/2e − 1 k≥0

when n ≥ 2; the matrix M 0 , with coefficients from (4.5.7), has mth column
  
k − bm/2c − 2 k−m
(d − c)
dm/2e − 2 k≥0

for m ≥ 3. After dealing manually with the first rows and columns, the Chu-Vandermonde sum
again shows that M M 0 is the identity matrix. 

Those sums are unlikely to be very useful, but the sum decompositions for U
en (x) in terms of
U ’s and V ’s allow us to prove the second-moments analogue of Theorem 4.1.6:

Theorem 4.5.5. For 0 ≤ k ≤ n, with respect to the second moments we have

 
(4.5.8) L2 Ue2n−k (x)U
en (x) =

bk/2c      
X bn/2c n − dk/2e − 1 bn/2c − 1 n − dk/2e − 1
+ (c − d)n−k−1+2j
j=0
j bk/2c − j j−1 bk/2c − 1 − j

Proof. Our only proof at the moment is algebraic; a proof using an involution, analogous to
the proof of Theorem 4.1.6 is desirable but unknown.
We know that
Z Z 
  1 (1) (2)
L2 U2n−k (x)Un (x) =
e e U2n−k (x)Un (x) dµ − U2n−k (x)Un (x) dµ
e e e e ,
c−d
so use (4.5.2) and (4.5.3) to expand the nearly-diagonal polynomials into Un ’s inside the dµ(1)
integral, and into Vn ’s inside the dµ(2) integral. The first integral is
Z X  ! 
b(2n − k)/2c X bn/2c
(c − d)i U2n−k−i (x)  (c − d)j Un−j (x) dµ(1)
i
i j
j

which because of the orthogonality of the Un ’s, and because the L2 norm of Un is always 1, equals
X b(2n − k)/2c bn/2c 
(c − d)k−n+2j .
j
j k − n + j

Reindex that sum by changing j into n − k + j, use the fact that b(2n − k)/2c = n − dk/2e, and
rewrite the binomial coefficients to turn the above sum into
X n − dk/2ebn/2c
(4.5.9) (c − d)n−k+2j .
j
bk/2c − j j

An exactly similar process with the V ’s and the dµ(2) integral yields
X n − dk/2e − 1bn/2c − 1
(4.5.10) (c − d)n−k+2j .
j
bk/2c − j − 1 j
 4.5. THE CORRESPONDING MOP 53

Our goal is to show that the difference of (4.5.9) and (4.5.10), divided by c − d, equals (4.5.8). The
above sums combine easily, and after we divide by c − d, we have
X n − dk/2ebn/2c n − dk/2e − 1bn/2c − 1
(4.5.11) − (c − d)n−k−1+2j .
j
bk/2c − j j bk/2c − j − 1 j

We’ll use several applications of Pascal’s relation to coerce the binomial coefficients above into the
correct form. The coefficient of (c − d)n−k−1+2j in the above sum is
          
n − dk/2e bn/2c n − dk/2e n − dk/2e − 1 bn/2c bn/2c − 1
− − − ;
bk/2c − j j bk/2c − j bk/2c − j j j−1
the first term cancels with one term from the binomials, yielding
        
n − dk/2e bn/2c − 1 bn/2c n − dk/2e − 1 n − dk/2e − 1 bn/2c − 1
+ −
bk/2c − j j−1 j bk/2c − j bk/2c − j j−1

and after rewriting the leftmost binomial coefficient as n−dk/2e−1 + n−dk/2e−1

bk/2c−j−1 bk/2c−j
, we see that the
sum (4.5.11) is exactly the same as the sum (4.5.8), which completes the proof. 

Theorem 4.5.1 proved orthogonality with respect to the moments of Vn ; those moments are very
(2)
similar to those for Un (x): µn is the generating function for noncrossing incomplete matchings of
[n] in which edges have weight 1, nonnested fixed points have weight d, and nested fixed points have
weight c.
Those matchings appear in Theorem 4.3.1, and we would expect a connection between them
and the second moments for the nearly diagonal Chebyshev polynomials because of Theorem 2.4.1.
Using the CD involution and parity flip fix, we produced a combinatorial proof of the first line of
(2.4.1) in Theorem 4.2.1; now we may do the same for the second line of that equation.
Theorem 4.5.6. Let L2 (xn ) denote the second moments for the nearly-diagonal Chebyshev poly-
(1) (2)
nomials, and µn and µn denote the x- and y-axis moments described above. There is a sign-
reversing involution and a weight-preserving bijection that proves

(4.5.12) (c − d)L2 (xn ) = µ(1) (2)

n − µn ,

which is equivalent to the second line of (2.4.1).

Proof. On the right side of the above equation, we have matchings in which all fixed points
have weight c, and also matchings in which all nonnested fixed points have weight d and nested fixed
points have weight c. The latter sort of matchings also have an overall negative sign.
Observe that on the right side of the equation, matchings with no nonnested fixed points will
(1) (2)
cancel; any such matching will come from µn and also from µn with a negative sign.
Now to the left side. Those moments were described in Theorem 4.3.1; the nth such moment is
the generating function for matchings of [n + 1] with a fake edge—vertex n + 1 must be part of an
edge. We interpret the c − d factor as meaning that fake edge may have weight c or −d.
Much like the proof of Theorem 4.2.1, we will first use a sign-reversing involution to cancel extra
matchings, and then show the remaining matchings are in bijection with the matchings described
for the right side. The sign-reversing involution is somewhat similar to the CD involution, but is far
less complicated. Figure 4.5.2 shows the basic idea.
 4.5. THE CORRESPONDING MOP 54

−d: move left c: move right

x-axis style matching y-axis style matching

Figure 4.5.2. The basic idea of the sign-reversing involution applied to matchings
on the left side.

To the left of the fake edge’s left endpoint, we have a matching weighted like the x-axis mo-
ments, so nonnested fixed points will have weight c. To the right of the fake edge’s left endpoint
(underneath the fake edge), we have a matching weighted like the y-axis moments, and there fixed
points immediately nested by the fake edge will have weight d. The involution is:
If the fake edge has weight −d, move the left endpoint of the fake edge to the
left, connect it to the rightmost nonnested fixed point, and change the weight of
the fake edge to c. If there’s no such fixed point to the left of the fake edge, do
nothing.
If the fake edge has weight c, do the opposite: move the left endpoint of
the fake edge to the right and connect it to the leftmost fixed point immediately
nested by the fake edge, then change the weight of the fake edge to −d. If there’s
no such fixed point immediately nested by the fake edge, do nothing.
In each case, we will create a new fixed point which, together with the new weight of the fake edge,
will preserve the weight and reverse the sign of the matching.
The matchings not canceled by that involution are those in which the fake edge has weight c
and there are no fixed points immediately nested by the fake edge, and those with the fake edge of
weight −d and no nonnested fixed points to the left of the fake edge. In the first case, the entire
matching is a power of c, and in the second case, the only place where fixed points of weight d occur
is immediately underneath the fake edge.
Now we show that those matchings are in (weight-preserving) bijection with the uncanceled
matchings on the right side. All we do is remove the fake edge and replace its left endpoint with a
fixed point of the same weight as the fake edge—one can think of the left endpoint swallowing up
the rest of the fake edge.
First we should verify that the result of this process is a matching found on the right side. If the
fake edge has weight c, then the result is a matching of [n] with at least one nonnested fixed point,
using the x-axis weighting. If the fake edge has weight −d, then the process yields a matching of [n],
again with at least one nonnested fixed point, with the y-axis weighting. In the latter case, note that
since there were no nonnested fixed points to the left of the fake edge, we will obtain a matching in
which every nonnested fixed point does have weight d, except for the leftmost one, which will have
weight −d. In both cases, we get a matching of the type found on the right side.
It is not difficult to see why this map is injective; consider two distinct matchings and look at
the leftmost vertex where they differ. They will still differ after the above process. For surjectivity,
 4.6. UNANSWERED QUESTIONS AND FUTURE DIRECTIONS 55

any x-axis matching on the right side is the image of a matching in which vertex n + 1 is added and
connected to the rightmost nonnested fixed point in the matching, and the new fake edge is given
weight c. To show surjectivity for the y-axis matchings on the right side, attach the left endpoint of
the new fake edge to the leftmost nonnested fixed point. 

4.6. Unanswered questions and future directions

What three-term polynomials have the second moments for the nearly-diagonal Chebyshevs as
their moments? We showed that the first moments of the four-term polynomials coincide with the
moments of the shifted Chebyshevs; it seems appropriate to try the same thing with the second
moments. A small shift would be required, since L(1) = 0; the nth moment for such a set of OP
would equal L(xn−1 ).
Explicit formulas for these polynomials are desirable, but not very simple to derive because the
edges need to connect adjacent vertices. For the usual Hermite polynomials, for example, it’s very
easy: choose the fixed points, and put a complete matching on the rest of the vertices:
n  
X n i
Hn (x) = x (−1)n−i (n − i)!!.
i=0
i

One may think of these polynomials as generating functions for weighted compositions of n into 1,
2, and 3; is there a nice expression for that? Can we use the sums of Theorem 4.5.2 to derive an
explicit formula? Those sums resemble (3.3.3) and (3.3.4) and perhaps we could use the proofs of
those formulas for inspiration here.
In the next chapter we will connect the moments of associated Hermite polynomials, which are
complete matchings weighted in a special way, to oscillating tableaux and rooted maps. In this
chapter we dealt with noncrossing complete matchings, and it would be interesting to find if there
are tableaux or rooted maps that enumerate the matchings of this chapter. Pylyavskyy has proposed
[52] a notion of noncrossing tableaux; will that definition yield anything interesting for polynomials
studied in this chapter?

4.6.1. Nearly-diagonal Hermite polynomials. The moments of Chebyshev polynomials are

generating functions for noncrossing matchings; the moments of Hermite polynomials are arbitrary
complete matchings. The nearly-diagonal Hermite polynomials are in some sense opposite to the
nearly-diagonal Chebyshevs: we began with the multiple Hermites, and can study the corresponding
nearly-diagonal polynomials. With the Chebyshevs, we began with the nearly-diagonal polynomials
and have worked backwards to the set of MOP. In both cases the state of knowledge is unsatisfactory.
From Theorem 3.2.1, the recurrence for general multiple Hermite polynomials, we know that
the nearly-diagonal Hermite polynomials H
e n (x) are defined by the four-term recurrence relation

(4.6.1) e n+1 (x) = (x − bn )H

H e n − nH
e n−1 (x) − θn H
e n−2 (x),

where 
c n even jnk
bn = and θn = (−1)n (c − d)
d n odd 2
 4.6. UNANSWERED QUESTIONS AND FUTURE DIRECTIONS 56

We can use the combinatorial model for the multiple Hermites, but we would like a model derived
directly from the above recurrence. This has proved more complicated than one might expect. The
weight of a fixed point seems to depend on whether the vertex is even or odd, as well as some sort
of nesting parity. It is also unclear what properties the double edges must have; none of the several
obvious ways to place a double edge so that there are bn/2c ways to do it seem easy to work with,
since the weight seems to depend on the crossing parity of the double edge. No workable definition
of “crossing parity” is known for these polynomials.
 CHAPTER 5

Associated Hermite polynomials

In this chapter we will step away from multiple orthogonal polynomials and consider some usual
orthogonal polynomials: the associated Hermite polynomials. We will present the first combinatorial
treatment of these polynomials. The reader may wish to refer to Section 3.1 for the combinatorics
of the ordinary Hermite polynomials.
The associated Hermite polynomials were considered by Askey and Wimp in [11], who analyti-
cally derived a number of results about these polynomials. They are also treated in [41, Section 5.6].
In Section 5.1 we provide a combinatorial interpretation of these polynomials, their moments, and
describe an involution that proves the orthogonality and L2 norms of the polynomials with respect to
those moments. Then in Section 5.2 we shall describe several identities and interesting linearization
formulas involving associated Hermite polynomials. We finish the chapter with weight-preserving
bijections between a number of classes of combinatorial objects whose generating functions all yield
the moments of the associated Hermites.

5.1. Definition and orthogonality

The associated Hermite polynomials are defined by a small shift of the recurrence relation for
the ordinary Hermite polynomials (3.1.1):

(5.1.1) Hn+1 (x; c) = xHn (x; c) − (n + c)Hn−1 (x; c),

with the same initial conditions. Askey and Wimp use a different normalization than we do; one
√ √
n
obtains our normalization from plugging x/ 2 into their associated Hermites and dividing by 2 .
Build the matchings as in Section 3.1; we’ll think of the parameter c as meaning that one special
choice for the edge from n + 1 has weight −1 or −c. Two natural choices are to make the special
choice be the leftmost available vertex, or the rightmost available vertex. Choosing the rightmost
available vertex happens to make the orthogonality involution easy to prove, and yields the following
result:

Theorem 5.1.1. The nth associated Hermite polynomial is the sum over certain weighted match-
ings:
X
(5.1.2) Hn (x; c) = wt(M ),
matchings M of [n]

in which fixed points have weight x, edges that nest no fixed points or edges and have no left crossings
may have weight −1 or −c, and all other edges have weight −1.
57
 5.1. DEFINITION AND ORTHOGONALITY 58

Proof. We build the matching from right to left, and if at some point we add an edge and do
not choose the rightmost available vertex, then that edge will nest a vertex, and when we come to
that vertex, we will either leave it fixed (resulting in a fixed point underneath that edge), connect
to another vertex underneath the edge (resulting in an edge nested by the original edge), or connect
to a vertex to the left of the edge, resulting in a left crossing for the original edge. Any of these
possibilities indicate that the rightmost vertex was not chosen, so edges for which none of those
happen are “eligible” for weight −c. 

An example of such a weighted matching is shown in Figure 5.1.1.

−1 −1 or −c −1

Figure 5.1.1. A matching on 8 points using the rightmost-choice weighting. This

weighting is used throughout this chapter.

With nothing more than this model, we can easily explain an “unexpected” limit that Askey
and Wimp derive [11, eq. (5.9)]. Using our normalizations, the limit is
√
(5.1.3) lim c−n/2 Hn (x c; c) = Un (x),
c→∞

where Un (x) is the c = 0 version of (4.1.2). (In their paper, there is a small typo: it should
√
be Hn (x 2c; c).) Using the combinatorial models of (4.1.2) and Theorem 5.1.1, there is nothing
√
unexpected about this limit: take Hn (x c; c) and give each vertex, whether fixed or incident to
√ √
an edge, weight 1/ c, so that c−n/2 Hn (x c; c) is the generating function for incomplete matchings
with fixed points weighted x, and all edges weighted −1/c except those which nest no fixed points
or edges, and have no left crossings—such edges have weight −1. As c goes to infinity, we effectively
restrict the generating function to matchings in which no edge has weight −1/c; i.e., every edge nests
no fixed points or edges, and has no left crossings, so every edge must connect adjacent vertices.
This is exactly the c = 0 case of (4.1.2)
We need to describe a linear functional Lc with respect to which the associated Hermite poly-
nomials are orthogonal. (This notation should not be confused with the functionals in Section 2.3.1;
all polynomials in this chapter are ordinary OP.) The sequence of polynomials uniquely determines
the sequence of moments, the first few of which are

µ0 = 1, µ4 = 3 + 5c + 2c2 ,
µ2 = 1 + c, µ6 = 15 + 32c + 22c2 + 5c3 .

The moments µ2n+1 are all zero; we shall see why in a moment. Using Viennot’s theory, the moments
are a sum over weighted Dyck paths in which a northeast edge has weight 1 and a southeast edge
leaving from height j has weight j + c. There no Dyck paths of odd length, so the odd moments are
zero.
 5.1. DEFINITION AND ORTHOGONALITY 59

Building matchings as in Figure 3.1.1, we have two combinatorial interpretations for the mo-
ments:

Theorem 5.1.2. If the special choice is the leftmost available vertex, then the nth moment µn (c)
is the generating function for complete matchings in which edges which are not nested by any other
edge may have weight 1 or c, and all other edges have weight 1.
If the special choice is the rightmost available vertex, then the nth moment µn (c) is the generating
function for complete matchings in which edges with no right crossings may have weight 1 or c and
all other edges have weight 1.

These interpretations also explain why the odd moments are zero, since there are no complete
matchings on an odd number of vertices. For the proof of orthogonality, we shall use the rightmost
weighting; later we shall use the leftmost weighting. Figures 5.1.2 and 5.1.3 show a matching using
the two weightings.
1 or c
1 or c
1

Figure 5.1.2. A complete matching on 6 points under the leftmost-choice weight-

ing for the moments.

1
1 or c
1 or c

Figure 5.1.3. The same complete matching under the rightmost-choice weighting.
This is used in the orthogonality proof.

5.1.1. Proof of orthogonality. We wish to prove the following theorem in a combinatorial

manner:

Theorem 5.1.3. The associated Hermite polynomials Hn (x; c) are orthogonal with respect to
the linear function Lc with the above moments. They satisfy

(5.1.4) Lc (Hn (x; c)Hm (x; c)) = [n = m](c + 1)n .

Proof. The proof proceeds very similarly to the proof of orthogonality for usual Hermite poly-
nomials (page 21). The product Hn (x; c)Hm (x; c) is the generating function for pairs of matchings
with, say, solid black edges, using the rightmost weighting. Integrating has the effect of putting
a complete matching with the rightmost weighting with green edges on the fixed points. Using
Theorems 5.1.1 and 5.1.2, the integral (5.1.4) is the generating function for paired matchings, where
black edges may have weight −1 or −c if they nest no edges, have no green crossings and no left
 5.1. DEFINITION AND ORTHOGONALITY 60

black crossing; otherwise black edges have weight −1. Green edges may have weight 1 or c if they
have no right green crossing and weight 1 otherwise. See Figure 5.1.4 for an example of such an
object for n = 5 and m = 3.

c
−1
−1
1

Figure 5.1.4. A paired matching for n = 5 and m = 3.

We need an involution that shows the generating function for paired matchings equals zero
when n 6= m, and equals (c + 1)n otherwise. Assume that n ≥ m and put [n] to the left of [m].
The involution is the very similar to that used in the combinatorial proof of orthogonality for usual
Hermite polynomials:
Find the leftmost homogeneous edge in [n] that nests no other edges and change
its color.
For example, in Figure 5.1.4, one would change the color of the leftmost green edge that connects
vertices 2 and 3 to black. This operation is evidently an involution and will certainly change the
sign; we need to verify that the weight of no other edge is affected by this change, and that if we
change the color of an edge weighted ±c, the new edge is eligible for a weight of ∓c.
We begin with the following observation: the leftmost homogeneous edge in [n] that nests no
edges can have no left crossing.
All green and black edges are eligible for weight +1 and −1 respectively, so switching the color
of such an edge will always preserve weight and reverse sign. The two remaining possibilities for the
weight and color of this edge are:
• Edge is black, weight −c: to get weight −c, the edge must in particular have no green
crossing, and therefore as green , it will be eligible for weight c.
• Edge is green, weight c: the edge nests no edges by assumption, and has no right green
crossing. By our observation above, it has no left crossings, hence will be eligible for weight
−c as a black edge.
Thus the weight of the edge is preserved and the sign is reversed. We leave it to the reader to check
that the weight and sign of no other edge is affected by this operation.
If n > m, there must be a homogeneous edge in [n]; in that case, the above involution has no
fixed points, and we have proved that Hn (x; c) is orthogonal to Hm (x; c).
Now we shall prove that the L2 norm of the associated Hermites is (c + 1)n by interpreting
the paired matchings as something whose generating function is known to be (c + 1)n : namely,
permutations weighted by left-to-right maxima. See [20, 34] for proofs of this fact in the context of
Laguerre polynomials. (“Left-to-right maxima” is “éléments saillants inférieurs gauches” in French.)
Also see Section 3.5 for some information on multiple Laguerre polynomials where rising factorials
 5.2. IDENTITIES 61

occur. This bijection naturally generalizes the usual combinatorial proof that the L2 norm of the
Hermite polynomials is n!.
First, apply the above involution to paired matchings with n = m; that involution will cancel
all matchings with a homogeneous edge. To set up the bijection, begin with a matching on [n] t [n]
with no homogeneous edges. Number the edges starting from the right as shown in Figure 5.1.5.
Thinking of the right side as the domain, and the left side as the range, we see that we can interpret
these objects are permutations, and that all permutations of n can be obtained in this way. We
shall show that edges that get weight 1 or c (here we shall think of that as having weight 1 + c)
correspond exactly to digits in the permutation that are left-to-right maxima.

1 2 3 4 4 3 2 1

Figure 5.1.5. The permutation 3142 as a matching. The right side is the domain,
the left the range. The digits 3 and 4 are left-to-right maxima in the permutation,
and indeed the green edges connecting to 3 and 4 on the left have weight 1 + c under
the associated Hermite moment weighting.

The base case of n = 1 clearly works; assume that the edges correspond so for some n. Any
permutation of n + 1 may be obtained by the following procedure: increment the label on each
vertex in the matching by 1, and add a new vertex with label 1 at the far right. We may connect
this vertex to a new vertex on the left side (i.e., choose what position 1 takes in the permutation) in
any one of n + 1 ways. If we put 1 in the first position, we will put an edge over the entire matching.
That edge will have no crossings, and hence have weight 1 + c; this corresponds to putting 1 at the
beginning of the permutation, where it will be a right-to-left maximum and contribute 1 + c to the
weight of the permutation. If we connect edge 1 to any other place, we must cross another edge, and
the new edge will have a right crossing and weight 1. This corresponds to putting 1 in the 2nd, 3rd,
etc, position in the permutation, where it cannot be a left-to-right maximum and will contribute 1
to the weight of the permutation.
Since this bijection from the fixed points of the involution to permutations preserves weight, we
see that the L2 norms of the associated Hermite polynomials are (c + 1)n . This completes the proof
of Theorem 5.1.3. 

The L2 norm of the associated Hermites also follows from (2.1.3). By Lemma 3.5.1, the L2 norm
can also be interpreted as the generating function for permutations with cycles weighted by c + 1.

5.2. Identities

Our first result is a combinatorial proof of the following identity, found in Askey and Wimp [11,
equation (4.18)]:
 5.2. IDENTITIES 62

−1
−c
−c

Figure 5.2.1. A matching on 6 vertices of the type described by Lemma 5.2.2. If

the new edge on the right is to have weight −1 and satisfy the conditions, it must
connect to a new vertex in one of the three available slots, indicated by the solid
arrows.

Theorem 5.2.1. The associated Hermites may be written as a sum of usual Hermite polynomials:
 
X
k n−k
(5.2.1) Hn (x; c) = (−1) (c)k Hn−2k (x).
k
k≥0

We will need two lemmas to prove Theorem 5.2.1.

Lemma 5.2.2. (−1)k (c)k is the generating function for complete matchings on 2k vertices, with
the associated Hermite polynomial weighting, such that all edges of weight −1 have a left crossing
by an edge of weight −c. Furthermore, in such matchings there are exactly k “slots” available
underneath the edges weighted −c where one could place the endpoint of a new edge.

Proof. The proof goes by induction. For k = 0 or 1, it’s obvious. If true for k, given any
configuration for that k, we can either:
• add a new edge connecting vertices 2k + 1 and 2k + 2 which has weight −c, and hence we
multiply the generating function for 2k vertices by −c and add a new slot, or
• add a new edge from the rightmost vertex and put its left endpoint in any one of the k
“slots” underneath one of the −c edges. Such an edge must have weight −1, and there are
k ways to place this edge, hence we effectively multiply the generating function by k, and
since we put a new edge into one of the k slots, there are now k + 1 slots available below
edges weighted −c.
See Figure 5.2.1 for an example of case 2. Altogether we’ve multiplied (−1)k (c)k , the generating
function for 2k vertices, by −(c + k), so the lemma is true by induction. 

Lemma 5.2.3. For such a configuration on 2k vertices as described in Lemma 5.2.2, there are
k + 1 places where one or more green edges of weight 1 could be placed without affecting the weight
of the configuration.

Proof. Induction again. The green edges cannot “pierce” the −c edges. For example, in
Figure 5.2.1, there are four places where one could place such an edge, indicated by the dotted
arrows. 
 5.2. IDENTITIES 63

Proof of Theorem 5.2.1. Since Hn (x; c) is an even or odd polynomial if n is even or odd,
respectively, we can certainly write
X
(5.2.2) Hn (x; c) = ank Hn−2k (x)
k≥0

for some coefficients ank . We shall show that those coefficients equal (−1)k (c)k n−k


k . Fix k between
0 and n/2, multiply both sides by Hn−2k (x), and apply the usual Hermite linear functional L0 . On
the right side, we use orthogonality and equation (5.2.2) becomes

L0 (Hn (x; c)Hn−2k (x)) = ank (n − 2k)!.

Thinking of the left side as paired matchings on [n] and [n − 2k] with black edges of −1 and −c
as appropriate, and green edges all of weight 1, we may apply the following involution: find the
leftmost homogeneous edge of weight ±1 in [n] or [n − 2k] and flip its color, unless that edge has
has a left crossing with an edge of weight −c. Swapping the colors on such edges does not preserve
the weight of the paired matching.
Lemma 5.2.2 tells us exactly what configurations of edges will remain in [n] after applying the
involution; Lemma 5.2.3 tells us that such configurations may be thought of as consisting of k
“chunks” of vertices. Placing the green edges into those chunks is equivalent to forming a weak
composition of k into n − 2k + 1 parts; there are n−k


k such compositions, and having chosen where
the n − 2k edges in [n] start, we can choose their endpoints in [n − 2k] in (n − 2k)! ways. Together
we have  
k n−k
(−1) (c)k (n − 2k)! = L0 (Hn (x; c)Hn−2k (x)) = ank (n − 2k)!
k
which proves the identity of Theorem 5.2.1. 

5.2.1. Linearization formulas. In [49, theorem 3.1], Markett shows that the linearization
coefficients in
min(N,M )
X
(5.2.3) HN (x; c)HM (x; c) = f (N, M, j) HN +M −2j (x; c)
j=0

are
!
j−N j−M −j
(5.2.4) f (N, M, j) = (N + M − 2j + c + 1)j 3 F2 ,
j−N −M −c 1
where the 3 F2 notation indicates a hypergeometric function evaluated at x = 1. We wish to prove

Theorem 5.2.4. The linearization coefficients f (N, M, j) of equation (5.2.4) are polynomials in
c with nonnegative integer coefficients.

Proof. Take the rising factorial in front and reverse the order of multiplication: it becomes
(−1)j (j − N − M − c)j . We have two k! factors in the denominator of the 3 F2 ; combine them with
the (j − N )k and (j − M )k in the numerator to get (−1)k Nk−j and (−1)k Mk−j . The (−1)k factors

cancel. Finally rewrite (−j)k = (j − k + 1)k (−1)k .

Put (−1)j (j − N − M − c)j inside the sum. There is a factor of (j − N − M − c)k in the
denominator; those cancel and yield (−1)j (j − N − M − c + k)j−k in the numerator of the sum.
 5.2. IDENTITIES 64

Reverse the order again and it turns into (−1)k (N + M − 2j + c + 1)j−k . This (−1)k cancels with
the earlier one from the (−j)k .
The sum is now
X N − j M − j 
(5.2.5) (j − k + 1)k (N + M − 2j + c + 1)j−k .
k k
k≥0

This is clearly a polynomial in c with nonnegative coefficients. 

Note that when c = 0, the 3 F2 of (5.2.4) sums by the Pfaff-Saalschütz identity to

(N + 1 − j)j (M + 1 − j)j
(5.2.6) ,
j!
and we recover the linearization coefficients for usual Hermite polynomials; the expression above,
after you multiply by (N + M − 2j)!, counts inhomogeneous matchings on [n] t [m] t [n + m − 2j],
as shown by de Sainte-Catherine and Viennot in [25].
A combinatorial interpretation of the coefficients (5.2.4) is not known, but the best starting
points seem to be [1, 51, 3]; the first two papers concern the usual Pfaff-Saalschütz identity, the
third features a combinatorial proof of the q-Pfaff-Saalschütz identity.

5.2.2. A mixed linearization formula. Recall that ak denotes the falling factorial a(a −
1) · · · (a − k + 1). Rising factorials are still denoted (a)k . In this section we shall prove

Theorem 5.2.5. If n ≥ m − 1, then

X n + cm
(5.2.7) Hn (x; c)Hm (x) = k!Hn+m−2k (x; c),
k k
k

where the sum runs from 0 to min(m, b(n + m)/2c).

Proof. Fix n; we’ll induct on m. For m = 0 and m = 1 the formula is a tautology and the
recurrence relation, respectively. Assume that the formula works for some m ≤ n; multiply both
sides of the formula by x and use the recurrence:

Hn (x; c)(Hm+1 (x) + mHm−1 (x)) =

X n + cm
k!(Hn+m+1−2k (x; c) + (n + m − 2k + c)Hn+m−1−2k (x; c)).
k k
k

If we move the mHn (x; c)Hm−1 (x) term over and use the induction hypothesis, we find that the
coefficient of Hn+m+1−2k (x; c) on the left side is
        
n+c m n+c m n+c m−1
k! + (n + m − 2k + c) (k − 1)! − m (k − 1)!
k k k−1 k−1 k−1 k−1
which simplifies to   
n+c m+1
k!,
k k
exactly the coefficient we want. 
 5.3. ASSOCIATED HERMITE MOMENTS AND OSCILLATING TABLEAUX 65

One must be careful with that recurrence, though. If k gets too large the recurrence fails,
because

(5.2.8) xH−1 (x; c) = H0 (x; c) + (−1 + c)H−2 (x; c)

is false. The induction argument works to go from m = n to m = n + 1 because xH0 = H1 + cH−1 ,

as long as one assumes polynomials with negative indices are zero.
Let’s try to go further. We’ll use
n   
X n+c n+1
(5.2.9) Hn (x; c)Hn+1 (x) = k!H2n+1−2k (x; c)
k k
k=0

and use the recurrence to move up to Hn (x; c)Hn+2 (x). We only need to worry about the constant
coefficient: since the sum for Hn (x; c)Hn+1 (x) runs up to k = n and the sum for Hn (x; c)Hn+2 (x)
goes to k = n + 1, if we try to use the recurrence in (5.2.9) for k = n + 1, we’re using the false
recurrence relation (5.2.8).
The correct coefficient of H0 (x; c) in the sum for Hn (x; c)Hn+2 (x) is (n + 1)(c)n+1 . Multiply
both sides of (5.2.9) by x; our induction hypothesis will work except when k = n. On the sum side
of that equation, since pq = pq /q!, the xH1 contributes (n + 1)(n + c)n (1 + c) to the coefficient of

H0 (x; c). On the product side of the equation, we write

xHn (x; c)Hn+1 (x) = Hn (x; c)(Hn+2 (x) − (n + 1)Hn (x))

and when we move over the (n + 1)Hn (x; c)Hn (x) and use the induction hypothesis, we subtract
(n + 1)(n + c)n from (n + 1)(n + c)n (1 + c) to get (n + 1)(n + c)n c = (n + 1)(c)n+1 .
The sum formula (5.2.7) predicts (n+2)(c)n+1 for the coefficient of H0 (x; c) in Hn (x; c)Hn+2 (x),
so we conclude that
n+1
X    
n+c n+2
(5.2.10) Hn (x; c)Hn+2 (x) = k! H2n+2−2k (x; c) − (c)n+1 H0 (x; c).
k k
k=0

Using that equation and more manual fiddling, we can prove that
n+1
X n + cn + 3 
(5.2.11) Hn (x; c)Hn+3 (x) = k! H2n+3−2k (x; c) − (c)n+1 H1 (x; c)
k k
k=0

and
n+2
X    
n+c n+4
(5.2.12) Hn (x; c)Hn+4 (x) = k! H2n+4−2k (x; c)
k k
k=0

− (c)n+1 H2 (x; c) + (c)n+1 ((n + 4)c − 2n − 5)H0 (x; c).

5.3. Associated Hermite moments and oscillating tableaux

In this section we will describe a statistic on oscillating tableaux , also known as up-down
tableaux, and a bijection from these tableaux to complete matchings which is weight-preserving
when using the weight for associated Hermite moments. Oscillating tableaux were described by
 5.3. ASSOCIATED HERMITE MOMENTS AND OSCILLATING TABLEAUX 66

Sundaram [59]; see section 5 of [16] for discussion of their origins. Briefly, an oscillating tableau
is a path in the Hasse diagram of the Young lattice in which at each point one either moves up to
a partition that covers the current partition, or moves down to a partition covered by the current
partition. For our purposes, the path will always begin and end with the empty shape. The length
of an oscillating tableau is the number of edges in the path. Figure 5.3.2 has an example of an
oscillating tableau of length 8.
In this section, we use Theorem 5.1.2’s “leftmost-available” weighting of complete matchings,
in which edges that are not nested by other edges may have weight 1 or c, and all other edges have
weight 1.
Roughly speaking, the bijection from complete matchings to oscillating tableaux works by RSK-
inserting numbers when edges start, and deleting them when edges end. More precisely, given a
complete matching, number the edges from right to left as in Figure 5.3.1. (Equivalently, write
the matching as a double occurrence word; see Definition 5.4.1.) We will map this matching to a
sequence of Ferrers shapes. Begin with the empty Ferrers shape and read the matching left to right.
When edge j starts, RSK-insert a j into the tableaux; when edge j ends, delete the box containing j.
Figure 5.3.1 has an example.
There is a possible point of confusion here. A tableau, in this context, is a path in the Hasse
diagram of the Young lattice—a sequence of Ferrers shapes. A traditional Young tableau is a path
that continually moves up, and therefore it is simple to record the path with a single Ferrers shape
filled with numbers that strictly increase in rows and columns. In Figure 5.3.1, the Ferrers shapes
are written as Young tableaux, which is only for our convenience. The actual image of the complete
matching is the same sequence without the numbers in the shapes. The reason for this is that RSK
is a bijection, and one can unbump numbers.
Figure 5.3.2 describes the inverse map from tableaux to matchings. We read the sequence of
Ferrers shapes from right to left. Because of how we number of the edges, the first box must have
a 1 in it. In general, when the shape gets larger, we put the next-largest number into the new box,
because we’ve started a new edge. The third shape from the right is 13 2 , and the shape to its left
must be 13 , because unbumping the 2 is the only way to produce the second shape. This oscillating
tableaux corresponds to the matching 43412321, using the vertex-numbering scheme described above.
Let us weight oscillating tableaux with the following statistic: numbers that appear in the first
column have weight 1 or c, and all other numbers have weight 1. That statistic is exactly what we
need to prove the following theorem.

Theorem 5.3.1. There is a weight-preserving bijection between oscillating tableaux of length

2n weighted with the above statistic and complete matchings weighted with the leftmost-available
associated Hermite weighting.

We will use several preliminary results to prove this theorem.

Lemma 5.3.2. In an oscillating tableau, when a number is added to a shape, the corresponding
edge is nested by all edges whose number is smaller, and has a left crossing from all edges whose
number is bigger. Edges that never appear together in a shape neither nest nor cross one another.
 5.4. ASSOCIATED HERMITE MOMENTS AND ROOTED MAPS 67

2 4 3 4 3 2 1 1

∅ 2 2 4 2 3 2 3 2 ∅ 1 ∅
4

Figure 5.3.1. A complete matching and the corresponding oscillating tableau.

The numbers in the Ferrers shapes are not strictly part of the oscillating tableau;
they are only used in the bijection from the matching to the tableau.

∅ ∅

Figure 5.3.2. An oscillating tableaux that corresponds to the complete matching

(13)(26)(48)(57).

For example, when we move from 2 4 to 24 3 in Figure 5.3.1, edge 3 is nested by edge 2 and has
a left crossing from edge 4. The proof of this is left to the reader; it follows from the way the edges
are numbered and in what order we add numbers to the tableau.
The above lemma implies the following facts:

Proposition 5.3.3. In an oscillating tableau, edges that get nested by other edges are exactly
those whose number appears in the 2nd, 3rd, etc, column of a shape. Edges that have a right crossing
are exactly those whose number appears in the 2nd, 3rd, etc row of a shape.

Proof of Theorem 5.3.1. The bijection between complete matchings and oscillating tableaux
clearly preserves weight: edges that do not get nested by another edge must appear in the first col-
umn only. Note also that we could have used the rightmost-available weighting from Theorem 5.1.2;
in that case, we would have needed to make our statistic “entries that appear in the first row and
stay there get weight 1 or c”. 

5.4. Associated Hermite moments and rooted maps

In addition to the weight-preserving bijection between associated Hermite moments and oscil-
lating tableaux of Section 5.3, there is a weight-preserving bijection between associated Hermite
moments and rooted maps. See [62, 42] for introductions to maps. A map may be thought of as a
graph along with an embedding into a surface. A rooted map is a map in which one edge has been
oriented. There is an axiomatic construction of maps that makes it natural to think of the edges in
a map as pairs of half-edges or edge ends and we will speak of edge ends in this section.
This connection was motivated by setting c = 1; if we have a polynomial that we suspect is the
generating function for some objects, setting c to 1 gives us a count of how many objects there are,
 5.4. ASSOCIATED HERMITE MOMENTS AND ROOTED MAPS 68

which facilitates searching. Doing so yields

1, 1, 2, 10, 74, 706, 8162, 110410, 1708394, . . .

which is sequence A000698 in [56]. This sequence likely first appeared in [61]; it counts connected
matchings (see below).
Table 1 of [6, page 10] essentially lists associated Hermite moments: row n is µ2n (c), and the
columns are powers of c. We will weight each vertex in such a map by c except the vertex at the head
of the root edge, and use the bijection between rooted maps in orientable surfaces and connected
matchings found in the work of de Mendez and Rosenstiehl [23, 21].

Definition 5.4.1. A connected matching on 2n vertices is one in which all vertices except 1
and 2n are nested by an edge. Equivalently, one can write a matching as a double occurrence word
in the letters 1, 2, . . . , n where each letter appears exactly twice; then a matching is connected if
the corresponding double occurrence word cannot be written as the concatenation of two double
occurrence words w1 and w2 .

A double occurrence word is the same thing as the vertex-numbering scheme used in Section 5.3.
We shall weight connected matchings by giving weight c to all nonnested edges except the edge
containing vertex 1. Then we have

Theorem 5.4.2. The function given in [23] and [21] is a weight-preserving bijection from rooted
maps in orientable surfaces with k vertices and n edges to connected matchings on 2n + 2 vertices
of weight ck−1 .

Proof. We summarize the bijection below. The bijection is weight-preserving because when
deciding the next vertex to visit, the algorithm chooses the vertex in the rooted map corresponding
to the leftmost unattached vertex in the partially-constructed matching. As we add edge ends to
the list, we will add a new edge to the matching that contains that leftmost unattached vertex. No
edge can then nest the newly created edge, so every visit to a new vertex in the rooted map results
in exactly one nonnested edge in the matching. 

The bijection works as follows:

(1) Add a new loop immediately clockwise of the root edge, oriented so that its head is next
to the old root.
(2) Beginning with the head of the new root, go around the adjacent vertex counterclockwise
and make a list of the edge ends encountered.
(3) After coming back around to the new root, read through the list and find the first edge
label that has not had both ends visited. Move to that edge end and, starting with the
immediate counterclockwise edge end, read counterclockwise around the vertex and add
the encountered edge ends to the list.
(4) Take the new list and repeat the above step until all edge ends have been visited.
In Figure 5.4.1 there is an example. We start at the head of edge α and read counterclockwise
around vertex a; our initial list is
α 1 2 3 α.
 5.4. ASSOCIATED HERMITE MOMENTS AND ROOTED MAPS 69

We have visited both ends of α, so we move to the unvisited end of edge 1, and go around vertex b
and add 4 4 5 2 5 1 to the list, which is now

α 1 2 3 α 4 4 5 2 5 1.

Now move to the unvisited end of edge 3 and do the same thing; we just append 3 to the list. We
end up with:
α 1 2 3 α 4 4 5 2 5 1 3,
which corresponds to the matching shown in Figure 5.4.2.
5 b
4
2

c 1
a
3

Figure 5.4.1. A rooted map to which we’ll apply the bijection to connected match-
ings. Green vertices have weight c. We have already added the extra edge, labeled
α; the original root was the end of edge 1 incident with vertex a.

Figure 5.4.2. The connected matching which corresponds to Figure 5.4.1. Edges
which get weight c are in green. The right vertices of the two green edges (reading
left to right) correspond to visiting vertices b and c respectively in the rooted map.

Now we need another weight-preserving bijection, this time from weighted connected matchings
to one of our original definitions for the moments of associated Hermite polynomials. We will
demonstrate such a bijection to the moments weighted with the leftmost-available weighting of
Theorem 5.1.2, in which nonnested edges are eligible for weight c. Call the edge containing vertex 1
the “fake edge”.1
The bijection works as follows: If the fake edge has no crossings, remove it; the remaining
matching on 2n vertices, of weight 1, is the result of the bijection. Otherwise, swap the tails of the
fake edge and that edge crossing the fake edge which has the leftmost endpoint. That crossing edge
must have weight c; give the new edge, which is now nested by the fake edge, weight c also. Continue
this tail-swapping process with the fake edge until the fake edge has no crossings, then remove it.
An example is shown in Figure 5.4.3.
1This notion is similar to the fake edge used for the second moments of the nearly-diagonal Chebyshev polynomials in
Section 4.3, but we don’t believe there is any special connection between those moments and these; the extra vertex
comes from different sources—for the Chebyshevs, the Motzkin paths end at height 1; here we have a parameter.
 5.4. ASSOCIATED HERMITE MOMENTS AND ROOTED MAPS 70

Figure 5.4.3. The steps of the tail-swapping bijection applied to the connected
matching of Figure 5.4.2; the result is a complete matching (the lower right match-
ing) in which nonnested edges are eligible for weight c.

This map is a bijection because it can be reversed: given such a weighted matching on 2n vertices,
add a new edge that nests the entire matchings, and swap tails with the green edges from right to
left. Observe that the green edges in the connected matching—which are nonnested—will end up
nonnested after the tail-swapping bijection, and vice versa. So this bijection is weight-preserving.
Note that in the example of Figure 5.4.3 and Table 5.4.1, the connected matching corresponded to a
complete matching which was also connected. Of course this does not always happen: the connected
matching (1, 5)(2, 4)(3, 8)(6, 7) corresponds to the unconnected complete matching (1, 3)(2, 4)(5, 6)
under this bijection.
Theorem 5.4.2 established that the generating functions for rooted maps and connected match-
ings are the same; that theorem, together with the bijection between connected matchings and
arbitrary complete matchings, provides a proof of the following theorem.

Theorem 5.4.3. The generating functions for rooted maps with n edges, connected matchings
on 2n + 2 vertices, and complete matchings on 2n vertices all equal the moment µ2n of the associated
Hermite polynomials.

In the last two sections, we’ve shown bijections between the moments of the associated Hermites,
connected matchings, rooted maps and oscillating tableaux. We summarize these correspondences
by going all the way from a rooted map, to a connected matching, to a regular complete matching,
to an oscillating tableaux in Table 5.4.1.

5.4.1. A second model for associated Hermite polynomials. The above discussion of
connected matchings meshes nicely with a second combinatorial model of the associated Hermites,
which is motivated by identity (5.2.1):
 
X n−k
Hn (x; c) = (−1)k (c)k Hn−2k (x).
k
k≥0

The key features of this second model are very similar to those of the connected matching model
for the moments: there are no choices for the weights of parts of the matching, and the resulting
matching is connected.
 5.4. ASSOCIATED HERMITE MOMENTS AND ROOTED MAPS 71

Table 5.4.1. A rooted map, a connected matching, a complete matching, and an

oscillating tableau, all of weight c5 , that correspond to each other using the weight-
preserving bijections of this chapter. In the tableau, we have only colored the first
box that corresponds to a number which gets weight c.

Object What gets weight c

Vertices not adjacent

to root edge.

Non-nested edges ex-

cept edge containing
vertex 1.

(1, 6)(2, 8)(3, 12)(4, 17)(5, 15)(7, 20)(9, 22)(10, 11)(13, 18)(14, 16)(19, 21)

Non-nested edges may

have weight 1 or c.

(1, 5)(2, 7)(3, 11)(4, 14)(6, 16)(8, 19)(9, 10)(12, 17)(13, 15)(18, 20)

∅
Numbers that appear
in first column may
have weight 1 or c.
∅

Theorem 5.4.4. The associated Hermite polynomial Hn (x; c) is the generating function for
certain connected incomplete matchings on n + 2 vertices with the following weights:
• The edge containing vertex 1 has weight 1. Call this edge the “fake edge”.
• Fixed points have weight x.
• Non-nested edges (except the fake edge) have weight −c.
• Nested edges have weight −1.
In such matchings, fixed points must be nested by the fake edge. All edges other than the fake edge
must either cross or be nested by the fake edge.

An example of such a matching for H7 (x; c) is shown in Figure 5.4.4. It is clear that the
requirement for nesting and crossing the fake edge yields a connected matching. Note that the
connected matching moments of Section 5.4 also have a fake edge.
 5.4. ASSOCIATED HERMITE MOMENTS AND ROOTED MAPS 72

+1
−c

−1 −1

Figure 5.4.4. A matching on 9 vertices that contributes cx to H7 (x; c) using the

combinatorial interpretation of Theorem 5.4.4. Note the “fake edge” of weight +1.

First proof. Consider the kth term in the sum (5.2.1):

 
k n−k
(−1) (c)k Hn−2k (x).
k
Begin with the fake edge and put k vertices to the right of it. Put the remaining n − k vertices
underneath the fake edge and choose k of them to connect with the edges that will come from the k
vertices on the right of the fake edge; that accounts for the binomial coefficient. On the remaining
n − 2k vertices underneath the fake edge, we put a regular Hermite-style matching; all the edges will
have weight −1 since they are nested by the fake edge.
The last thing to do is account for the k edges that come from the right of the fake edge and
show that they contribute (−1)k (c)k . According to Lemma 5.4.5, the generating function for such a
configuration with the edges of weight +1 and +c is (c + 1)k−1 , but in our subset, the leftmost edge
also gets weight c, so the correct factor is (c + 1)k−1 · c = (c)k . Also, we must correct for the signs:
our edges have weight −1 and −c, so we multiply by (−1)k . 

Second proof. Verify that the generating function described in the theorem satisfies the three-
term recurrence for the associated Hermites (5.1.1). We proceed very much like the usual combi-
natorial proof of the recurrence relation for Hermite polynomials: any such restricted matching on
n + 3 vertices may be obtained by placing the fake edge and considering the rightmost vertex nested
by the fake edge. There are three possibilities: one, we can leave that vertex fixed, and fill in the
remaining n + 2 vertices with any restricted matching; two, we can add an edge from that vertex
to the very rightmost vertex, and fill in the remaining n + 1 vertices with any restricted matching;
three, we can attach that vertex to any vertex except the rightmost vertex and fill in the remaining
n + 1 vertices as before. The first case contributes x times the generating function for n + 2 vertices.
The second cases contributes −c times the generating function for n + 1 vertices, since that new edge
cannot be nested, and it will not nest any of the other edges. In the third case, there are n vertices
to choose from and all of them will result in a nested edge of weight −1, so we add −n times the
generating function for n + 1 vertices. This exposition is simply another way of stating (5.1.1):

Hn+1 (x; c) = xHn (x; c) − (n + c)Hn−1 (x; c). 

The following lemma was used in the first proof of Theorem 5.4.4. It may be proved by induction,
similar to Lemma 5.2.2 and Theorem 5.1.3.
 5.5. UNANSWERED QUESTIONS AND FUTURE DIRECTIONS 73

Lemma 5.4.5. The generating function for complete matchings on 2n vertices in which all edges
go from the “left n” vertices to the “right n” vertices , with all nonnested edges having weight c
except the edge containing the leftmost vertex, is (c + 1)n−1 .
There is a weight-preserving bijection between such matchings and permutations π of [n] weighted
by c LRM(π)−1 where LRM(π) is the number of left-to-right-maxima of the permutation.

At this point, we have a combinatorial interpretation for both the associated Hermite polyno-
mials (Theorem 5.4.4) and their moments (Theorem 5.4.3) in terms of connected matchings with
a fake edge; the natural thing to do is combine these to get another proof of orthogonality. This
will be quite difficult because it is not at all obvious how to combine a pair of matchings for the
polynomials and a matching for the moments to get a paired matching; one would have two fake
edges from the polynomials and would need to somehow incorporate the fake edge from the moments
into that configuration. However, it is interesting to note that the above theorem tells us how we
would derive the L2 norm using such a setup: Hn (x; c)2 would be a pair of matchings on 2n + 4
vertices, but because of the extra fake edge mentioned above, after canceling all homogeneous edges
we would effectively get complete matchings on 2n + 2 vertices in which all the edges go from the
left n + 1 vertices to the right n + 1. The generating function for such a configuration, according to
Lemma 5.4.5, is (c + 1)n , which agrees with the known L2 norm for the associated Hermites.

5.5. Unanswered questions and future directions

We have taken the basic combinatorial model in Section 5.1 for associated Hermite polynomials
and their moments and gone in two directions: to oscillating tableaux, and to rooted maps. The
appeal of oscillating tableaux is in the recent flurry of work on k-crossings and k-nestings in matchings
and set partitions; see [16, 47, 24, 45, 44, 43]. The moments of Charlier polynomials are generating
functions for set partitions and it seems likely that some of this work could be used to treat the
associated Charlier polynomials. Also, a notion of noncrossing tableaux has been put forth [52]
which, surprisingly, suggests a connection between oscillating tableaux, which count matchings,
and the moments of the nearly-diagonal Chebyshev polynomials, which count certain noncrossing
matchings. See (5.1.3) for a limit relationship between the associated Hermites and Chebyshevs.
One disadvantage of the original combinatorial model and oscillating tableaux is that certain
parts of the corresponding combinatorial objects may have weight 1 or c. This makes proofs of
linearization formulas and other identities more difficult because our proofs are generally “geometric”
in nature, in the sense that they depend on the pictorial representation of the objects; we use words
like “leftmost” and make other spatial references. In such proofs it is desirable to have a 1-1
correspondence between terms that contribute to the generating function and geometric objects.
This makes the sign-reversing involutions we need to find much easier.
Sending c to c − 1 seems like a solution to this problem, but it would really just shuffle it around:
as it stands, the moment weighting for complete matchings on 2n points has “1 or c edges”, but the
weighting for connected complete matchings on 2n + 2 points only has edges of weight c. Shifting c
to c − 1 would fix the problem for matchings on 2n vertices, but would give our connected matchings
edges of weight c or −1, which is undesirable for the moments.
 5.5. UNANSWERED QUESTIONS AND FUTURE DIRECTIONS 74

Observe that in the connected matchings, the rooted maps, and in the second combinatorial
model for the associate Hermite polynomials of Theorem 5.4.4, each model has some sort of “fake
edge”. Combining the models for the moments and polynomials which both involve connected match-
ings would be fruitful and eliminate the problems with multiple choices for the weights discussed
above, but this has not yet shown promise. A major problem is that each incomplete matching
for the polynomial is weighted by x to the number of fixed points—say there are 2k fixed points—
but the corresponding matchings are matchings on 2k + 2 vertices. It is not clear how to combine
these two objects in a geometric or graph-theoretical way that allows a natural and easy proof of
orthogonality.
Using rooted maps holds promise, though: de Mendez and Rosenstiehl have generalized the
bijection between connected matchings and rooted maps to a bijection between permutations and
hypermaps [22, 21]. This suggests an intriguing connection to Laguerre polynomials since hyper-
maps are built out of permutations in the same way that maps are built out of complete matchings.
It may also be possible to bring the topological data that accompanies maps back to the matchings,
and see if that tells us anything interesting or new. One might also consider different families of
maps or hypermaps and try to work backwards to a new (or old!) set of orthogonal polynomials.

5.5.1. Associated Laguerre polynomials. The paper of Askey and Wimp [11] which in-
spired this chapter devotes much more attention to the associated Laguerres than to Hermites—
about two-thirds of the article. It is natural, then, to work out a corresponding combinatorial treat-
ment of those polynomials, especially given the connections between rooted maps and hypermaps
mentioned above. There is also the work of Ismail et al [38] who work with the associated Laguerres
as birth and death processes—there has been work on birth and death processes and lattice paths
[31] which suggests another avenue for a combinatorial theory of those polynomials.
 Bibliography

[1] George E. Andrews, Identities in combinatorics I: On sorting two ordered sets, Discrete Math. 11 (1975), 97–106,
doi:10.1016/0012-365X(75)90001-1. MR 0389609. pages 64
[2] George E. Andrews, Richard Askey, and Ranjan Roy, Special functions, Cambridge University Press, February
2001. ISBN 0521789885. pages 3
[3] George E. Andrews and David M. Bressoud, Identities in combinatorics III: Further aspects of ordered set sorting,
Discrete Math. 49 (1984), no. 3, 223–236, doi:10.1016/0012-365X(84)90159-6. MR 0743793. pages 64
[4] A. Aptekarev, Multiple orthogonal polynomials, Journal of Computational and Applied Mathematics 99 (1998),
no. 1-2, 423–447, doi:10.1016/S0377-0427(98)00175-7. MR 1662713. pages 7
[5] A. I. Aptekarev, A. Branquinho, and W. van Assche, Multiple orthogonal polynomials for classical weights, Trans.
Amer. Math. Soc. 355 (2003), no. 10, 3887–3914, doi:10.1090/S0002-9947-03-03330-0. MR 1990569. pages 21, 22,
31
[6] Didier Arquès and Jean-François Béraud, Rooted maps on orientable surfaces, Riccati’s equation and continued
fractions, Discrete Math. 215 (2000), no. 1-3, 1–12, doi:10.1016/S0012-365X(99)00197-1. MR 1746444. pages 68
[7] J. Arvesú, J. Coussement, and W. van Assche, Some discrete multiple orthogonal polynomials, Proceedings of
the Sixth International Symposium on Orthogonal Polynomials, Special Functions and their Applications (Rome,
2001), vol. 153, 2003, pp. 19–45, doi:10.1016/S0377-0427(02)00597-6. MR 1985676. pages 7
[8] Richard Askey, Orthogonal polynomials and special functions, Society for Industrial and Applied Mathematics,
Philadelphia, Pa., 1975. MR 481145. pages 3
[9] Richard Askey and Mourad E. Ismail, Permutation problems and special functions, Canad. J. Math. 28 (1976),
no. 4, 853–874. MR 406808. pages 31
[10] Richard Askey, Mourad E. Ismail, and Tom Koornwinder, Weighted permutation problems and Laguerre poly-
nomials, J. Combin. Theory Ser. A 25 (1978), no. 3, 277–287, doi:10.1016/0097-3165(78)90020-1. MR 0514623.
pages 31
[11] Richard Askey and Jet Wimp, Associated Laguerre and Hermite polynomials, Proc. Roy. Soc. Edinburgh Sect.
A 96 (1984), no. 1-2, 15–37. MR 741641. pages 57, 58, 61, 74
[12] Ruth Azor, J. Gillis, and J. D. Victor, Combinatorial applications of Hermite polynomials, SIAM J. Math. Anal.
13 (1982), no. 5, 879–890, doi:10.1137/0513062. MR 0668329. pages 20
[13] Yousséf Ben Cheikh and Ali Zaghouani, Some discrete d-orthogonal polynomial sets, J. Comput. Appl. Math.
156 (2003), no. 2, 253–263, doi:10.1016/S0377-0427(02)00914-7. pages 10
[14] C. Brezinski, A. Draux, A. P. Magnus, P. Maroni, and A. Ronveaux (eds.), Polynômes orthogonaux et applica-
tions, Lecture Notes in Mathematics, vol. 1171, Berlin, Springer-Verlag, 1985. MR 838964. ISBN 3-540-16059-0.
pages 76, 78
[15] Youssèf B. Cheikh and Khalfa Douak, On the classical d-orthogonal polynomials defined by certain generating
functions. I, Bull. Belg. Math. Soc. Simon Stevin 7 (2000), no. 1, 107–124. MR 1741750. pages 10
[16] William Y. C. Chen, Eva Y. P. Deng, Rosena R. X. Du, Richard P. Stanley, and Catherine H. Yan, Crossings
and nestings of matchings and partitions, Apr 2005, to appear in Transactions of the AMS, arXiv:math/0501230.
pages 66, 73
[17] Theodore S. Chihara, An introduction to orthogonal polynomials, Gordon & Breach, 1978. ISBN 0677041500.
pages 3, 4

75
 BIBLIOGRAPHY 76

[18] Zélia da Rocha, Shohat-Favard and Chebyshev’s methods in d-orthogonality, Numer. Algorithms 20 (1999),
no. 2-3, 139–164, doi:10.1023/A:1019151817161. MR 1709558. pages 10
[19] Marcel G. de Bruin, Simultaneous Padé approximation and orthogonality, pp. 74–83, vol. 1171 of Brezinski et al.
[14], 1985. MR 838972. ISBN 3-540-16059-0. pages 7
[20] Anne de Médicis and Xavier G. Viennot, Moments des q-polynômes de Laguerre et la bijection de Foata-
Zeilberger, Adv. in Appl. Math. 15 (1994), no. 3, 262–304, doi:10.1006/aama.1994.1010. MR 1288802. pages
60
[21] Patrice O. De Mendez and Pierre Rosenstiehl, Connected permutations and hypermaps, Tech. Report 183, Centre
d’Analyse et de Mathématique Sociales, July 1999, http://citeseer.ist.psu.edu/demendez99connected.html.
pages 68, 74
[22] , Transitivity and connectivity of permutations, Combinatorica 24 (2004), no. 3, 487–501,
doi:10.1007/s00493-004-0029-4. MR 2085369. pages 74
[23] , Encoding pointed maps by double occurrence words, August 2005, http://hal.ccsd.cnrs.fr/
ccsd-00007477. pages 68
[24] Anna de Mier, k-noncrossing and k-nonnesting graphs and fillings of Ferrers diagrams, Feb 2006,
arXiv:math.CO/0602195. pages 73
[25] Myriam de Sainte-Catherine and Gérard Viennot, Combinatorial interpretation of integrals of products of Her-
mite, Laguerre and Tchebycheff polynomials, in Brezinski et al. [14], pp. 120–128. MR 0838977. ISBN 3-540-
16059-0. pages 19, 20, 30, 31, 34, 64
[26] K. Douak and P. Maroni, Une caractérisation des polynômes d-orthogonaux “classiques”, Journal of Approxi-
mation Theory 82 (1995), no. 2, 177–204, doi:10.1006/jath.1995.1074. MR 1343833. pages 10
[27] , On d-orthogonal Tchebychev polynomials. I, Appl. Numer. Math. 24 (1997), no. 1, 23–53,
doi:10.1016/S0168-9274(97)00006-8. MR 1454707. pages 10
[28] , On d-orthogonal Tchebychev polynomials. II, Methods Appl. Anal. 4 (1997), no. 4, 404–429. MR 1605626.
pages 10
[29] Ömer Eğecioğlu, Timothy Redmond, and Charles Ryavec, From a polynomial Riemann hypothesis to alternating
sign matrices, Electron. J. Combin. 8 (2001), no. 1, Research Paper 36, 51 pp. (electronic). MR 1877655. pages
10
[30] S. Even and J. Gillis, Derangements and Laguerre polynomials, Math. Proc. Cambridge Philos. Soc. 79 (1976),
no. 1, 135–143. MR 0392590. pages 31
[31] Philippe Flajolet and Fabrice Guillemin, The formal theory of birth-and-death processes, lattice path combina-
torics and continued fractions, Adv. in Appl. Probab. 32 (2000), no. 3, 750–778, doi:10.1239/aap/1013540243.
MR 1788094. pages 74
[32] Dominique Foata, A combinatorial proof of the Mehler formula, J. Combinatorial Theory Ser. A 24 (1978), no. 3,
367–376, doi:10.1016/0097-3165(78)90066-3. MR 498167. pages 1, 20, 26
[33] Dominique Foata and Marcel-P Schützenberger, Théorie géométrique des polynômes eulériens, Lecture Notes in
Mathematics, Vol. 138, Springer-Verlag, Berlin, 1970. MR 272642. pages 32
[34] Dominique Foata and Volker Strehl, Combinatorics of Laguerre polynomials, pp. 123–140, Academic Press,
Toronto, ON, 1984. MR 782311. pages 31, 60
[35] Dominique Foata and Doron Zeilberger, Laguerre polynomials, weighted derangements, and positivity, SIAM J.
Discrete Math. 1 (1988), no. 4, 425–433, doi:10.1137/0401043. MR 0968850. pages 31
[36] A. M. Garsia and J. Remmel, A combinatorial interpretation of q-derangement and q-Laguerre numbers, Euro-
pean J. Combin. 1 (1980), no. 1, 47–59. MR 576766. pages 31
[37] C. D. Godsil, Hermite polynomials and a duality relation for matching polynomials, Combinatorica 1 (1981),
no. 3, 257–262. MR 637830. pages 20
[38] Mourad E. Ismail, Jean Letessier, and Galliano Valent, Linear birth and death models and associated Laguerre
and Meixner polynomials, J. Approx. Theory 55 (1988), no. 3, 337–348, doi:10.1016/0021-9045(88)90100-1.
MR 0968940. pages 74
 BIBLIOGRAPHY 77

[39] Mourad E. Ismail and Dennis Stanton, Classical orthogonal polynomials as moments, Canad. J. Math. 49 (1997),
no. 3, 520–542. MR 1451259. pages 24
[40] , More orthogonal polynomials as moments, Progr. Math., vol. 161, pp. 377–396, Birkhäuser Boston,
Boston, MA, 1998. MR 1627382. pages 24
[41] Mourad E. H. Ismail, Classical and quantum orthogonal polynomials in one variable, Encyclopedia of Mathe-
matics and its Applications, Cambridge University Press, November 2005. ISBN 0521782015. pages 3, 7, 8, 16,
17, 57
[42] David Jackson and Terry Visentin, An atlas of the smaller maps in orientable and nonorientable surfaces,
Chapman & Hall/CRC, September 2000. ISBN 1584882077. pages 67
[43] Vı́t Jelı́nek, Dyck paths and pattern-avoiding matchings, Mar 2005, arXiv:math/0503327. pages 73
[44] Anisse Kasraoui and Jiang Zeng, Distribution of crossings, nestings and alignments of two edges in matchings
and partitions, Jan 2006, arXiv:math/0601081. pages 73
[45] Martin Klazar, On identities concerning the numbers of crossings and nestings of two edges in matchings, Mar
2005, arXiv:math.CO/0503012. pages 73
[46] Roelof Koekoek and Rene F. Swarttouw, The Askey-scheme of hypergeometric orthogonal polynomials and its
q-analogue, Tech. Report 98-17, Delft University of Technology, 1998, http://aw.twi.tudelft.nl/~koekoek/
askey.html. pages 31
[47] Christian Krattenthaler, Growth diagrams, and increasing and decreasing chains in fillings of Ferrers shapes,
Oct 2005, arXiv:math.CO/0510676. pages 73
[48] Jacques Labelle and Yeong N. Yeh, The combinatorics of Laguerre, Charlier, and Hermite polynomials, Stud.
Appl. Math. 80 (1989), no. 1, 25–36. MR 1002302. pages 31
[49] Clemens Markett, Linearization of the product of symmetric orthogonal polynomials, Constr. Approx. 10 (1994),
no. 3, 317–338, doi:10.1007/BF01212564. MR 1291053. pages 63
[50] Pascal Maroni, L’orthogonalité et les récurrences de polynômes d’ordre supérieur à deux, Ann. Fac. Sci. Toulouse
Math. (5) 10 (1989), no. 1, 105–139. MR 1425747. pages 10
[51] T. S. Nanjundiah, Remark on a note of P. Turán, Amer. Math. Monthly 65 (1958), 354. MR 98042. pages 64
[52] Pavlo Pylyavskyy, Non-crossing tableaux, Jul 2006, arXiv:math.CO/0607211. pages 55, 73
[53] Emmanuel Roblet, Une interprétation combinatoire des approximants de Padé, Ph.D. thesis, L’Université Bor-
deaux I, 1994. ISBN 2892761336. pages 1, 11, 46
[54] R. Simion and D. Stanton, Specializations of generalized Laguerre polynomials, SIAM J. Math. Anal. 25 (1994),
no. 2, 712–719, doi:10.1137/S003614109322854X. MR 1266585. pages 31
[55] , Octabasic Laguerre polynomials and permutation statistics, Journal of Computational and Applied
Mathematics 68 (1996), no. 1-2, 297–329, doi:10.1016/0377-0427(95)00250-2. MR 1418763. pages 31
[56] N. J. A. Sloane, The on-line encyclopedia of integer sequences, http://www.research.att.com/~njas/
sequences/. pages 47, 68
[57] Richard P. Stanley, Enumerative combinatorics, vol. 1, Cambridge University Press, June 2000. ISBN 0521663512.
pages 32
[58] , Enumerative combinatorics, vol. 2, Cambridge University Press, February 2001. ISBN 0521789877. pages
47
[59] Sheila Sundaram, The Cauchy identity for Sp(2n), J. Combin. Theory Ser. A 53 (1990), no. 2, 209–238,
doi:10.1016/0097-3165(90)90058-5. MR 1041446. pages 66
[60] Gabor Szegő, Orthogonal polynomials, 4th ed., American Mathematical Society, 1975. ISBN 0821810235. pages
3
[61] Jacques Touchard, Sur un problème de configurations et sur les fractions continues, Canadian J. Math. 4 (1952),
2–25. MR 46325. pages 68
[62] William T. Tutte, What is a map?, New directions in the theory of graphs (Proc. Third Ann Arbor Conf., Univ.
Michigan, Ann Arbor, Mich., 1971), Academic Press, New York, 1973, pp. 309–325. MR 0376413. pages 67
[63] Walter van Assche and Els Coussement, Some classical multiple orthogonal polynomials, J. Comput. Appl. Math.
127 (2001), no. 1-2, 317–347, doi:10.1016/S0377-0427(00)00503-3. MR 1808581. pages 7, 21, 31
 BIBLIOGRAPHY 78

[64] Jeannette van Iseghem, Approximants de Padé vectoriels, Ph.D. thesis, Université des Sciences et Technologies
de Lille, Lille, France, 1987. pages 10
[65] Gérard Viennot, Une théorie combinatoire des pôlynomes othogonaux generaux, Sep-Oct 1983, Notes from a
conference at the Université du Québec à Montréal. pages 3, 5, 6, 7, 11, 15, 16, 31
[66] , A combinatorial theory for general orthogonal polynomials with extensions and applications, pp. 139–157,
vol. 1171 of Brezinski et al. [14], 1985. MR 838979. ISBN 3-540-16059-0. pages 3, 5, 6, 7, 11
[67] Herbert S. Wilf, Mathematics for the physical sciences, Dover Publications, February 2006. ISBN 0486450384.
pages 3
 Index

δ notation, 16, 46 first moments, see moments, first

four-term recurrence, see recurrence, four term
Askey-Wilson table, 31
associated Hermite polynomials generating function
as sum of usual Hermites, 62 Hermite polynomials, 20
L2 norm, 60 multiple Hermite polynomials, 22, 29
linearization, 63–65 nearly-diagonal Chebyshev polynomials, 45
moments genuine recurrence relation, 10
as oscillating tableaux, 66
orthogonality, 59 Hermite polynomials
associated Laguerre polynomials, 74 multiple, 21
nearly-diagonal, 55
Charlier polynomials, 73
Chebyshev polynomials L2 norm, 4
shifted, 34 for d-orthogonality, 10
Chu-Vandermonde identity, 51 multiple Hermite polynomials, 25
combinatorial model nearly-diagonal Chebyshev polynomials, 36
associated Hermite moments, 59 of multiple orthogonal polynomials, 8
associated Hermite polynomials, 57, 71 lattice paths
Chebyshev polynomials, 35 with double-down steps, 11
Hermite polynomials, 19 left-to-right maxima, 31, 60, 73
multiple Hermite polynomials, 22, 29 Lukasiewicz paths, see Motzkin paths, generalized
nearly-diagonal Chebyshev polynomials, 35
matching, 19
connected matching, 68
as double occurrence word, 66
continued fraction, 2, 46
complete, 19
moment generating function, 16
connected, 68

d-Lukasiewicz paths, see Motzkin paths, generalized double edge, 35

d-orthogonality, 10 fixed points, 19

nearly-diagonal Chebyshev polynomials, 36 nesting parity, 35

dominos, 5 matchings, 35

of arbitrary length, 11 Mehler formula, 26

Dyck paths, 36, 58 moments, 4

x-axis, 12
exponential formula, 20, 23 y-axis, 12
Chebyshev polynomials, 35
fake edge, 44, 71 first, 10
falling factorial, 2, 64 nearly-diagonal Chebyshev polynomials, 36

79
 INDEX 80

generating function, 16 Tetris recurrence, 17, 25

second, 10 Tetris S-piece recurrence, 9
x-axis, 9 three-term recurrence, see recurrence, three term
y-axis, 9 type II multiple orthogonal polynomials, see multiple
monominos, 5 orthogonal polynomials, type II
Motzkin paths, 5
generalized, 11, 43, 46 Vladivostok telephone directory, 25

multiple Laguerre polynomials, 30–32

x-axis polynomials, 9
multiple orthogonal polynomials
type II, 7 y-axis polynomials, 9

nearly-diagonal polynomials, 9
nearly-diagonal Chebyshev polynomials
CD involution, 39–43
first moments
generating function, 46
generating function, 45
integral with respect to first weight, 38
integral with respect to second weight, 52
L2 norm, 37, 45
orthogonality with respect to second moments, 44
parity flip fix, 42
second moments, 43–45
generating function, 47
y-axis polynomials, 48
noncrossing tableaux, 55, 73

oscillating tableaux, 65

P1,m (x) polynomials, 14

paired matchings, 21, 24, 44
Chebyshev polynomials, 35
nearly-diagonal Chebyshev polynomials, 37
paving, 5

recurrence
for P1,m (x), 14
for multiple Hermite polynomials, 21
four term, 8
four-term
nearly-diagonal Chebyshev polynomials, 34
shapes for multiple orthogonal polynomials, 16–18
three-term, 4
rising factorial, 2, 31, 62, 65, 72
rooted maps, 67–70

second moments, see moments, second

sieving a generating function, 27, 45
sign-reversing involution, 21
SSE steps, 11, see lattice paths, double-down steps