Special Topics in Applied Mathematics

2.3. q-Hermite Polynomials

Recurrence relations, such as Equation (2.21) for Hermite polynomials,
can be generalized. In this section, we consider a generalization that is sat-
isfied by polynomials that are known as q-Hermite polynomials. We assume
that q P C is a given number satisfying |q| † 1,
Such generalizations arise from an idea that dates back several centuries
in mathematics. Recall that` in Section 1.2, ˘we replaced a derivative by a
difference quotient w1 ptq fiÑ wpt ` hq ´ wptq {h to discretize an ODE. The
main idea was to represent a line by a chain of equally spaced points: t0 `n h,
where h is a small step size.
However, the grid of points we use in a discretization do not have to be
equally spaced. An alternative discretization is to consider a chain of points
where the spacing between successive points change multiplicatively, under
multiplication by q. In this case, the difference quotient is replaced by
ypqzq ´ ypzq
Dq ypzq :“ . (2.25)
qz ´ z
To understand how this can approximate a derivative consider q “ 1 ` ✏, and
take ✏ Ñ 0, to see that
ypqzq ´ ypzq dypzq
Ñ .
qz ´ z dz
The following result shows that the ratio 1 ´ q n {1 ´ q approaches n in
such a limit. This idea is the start of a huge area called q-difference calculus
in mathematics, which is also important in mathematical physics.
Lemma 2.3.1.
1 ´ qn
lim “ n.
qÑ1 1 ´ q

Proof. The numerator on the left can be factorized to give

` ˘
p1 ´ q n q “ p1 ´ qq 1 ` q ` q 2 ` . . . ` q n´1 ,
which implies that
1 ´ qn
lim “ loooooooooomoooooooooon
1 ` 1 ` 1 ` . . . ` 1 “ n.
qÑ1 1 ´ q
n
⇤
This observation leads to q-versions of factorials and many other well
known functions.
2.3.2. Definition (q-factorial and q-Pochhammer symbols). Given q P C,
with |q| † 1, and n P N, denote
1 ´ qn
rnsq “ .
1´q
Define the q-factorial by
rns!q “ r1sq r2sq . . . rnsq
and, for a P C, the q-Pochhammer symbol by
pa; qqn “ p1 ´ aqqp1 ´ aq 2 q . . . p1 ´ aq n´1 q,
with pa; qq0 “ 1.
21 ©Nalini Joshi
 February 24, 2022 DRAFT

2.3.A. Exercise. Show that

pq; qqn
rns!q “ .
p1 ´ qqn
Correspondingly, there are q-versions of many families of orthogonal
polynomials. Consider the recurrence relation
2xPrn pxq “ Prn`1 pxq ` p1 ´ q n qPrn´1 pxq, (2.26)
r r
with initial conditions P0 “ 1, P1 “ 2x, for all n P N.
2.3.B. Exercise. Find Prj pxq, for j “ 2, 3, 4. What is Pr4 p0q?
Note that a coefficient that is linear in n in Equation (2.21) has been
replaced by a coefficient that is exponential in n in (2.26) due to Lemma
2.3.1. This brings up a natural question: is there a relationship between
Equations (2.26) and (2.21)?
To answer this question, we start by simplifying Equation (2.26) a little
bit. Consider a change of variables, involving two parameters ↵ and :
Qn pzq “ ´n{2 Prn pxq, z “ ↵x.
Substitution in Equation (2.26) gives
↵ 1{2 ↵
z Qn pzq “ Qn`1 pzq ` p1 ´ q n q 1{2
Qn´1 pzq.
2 2
So taking
´1{2
↵“2 , “ 1 ´ q,
gives
1 ´ qn
z Qn pzq “ Qn`1 pzq ` Qn´1 pzq.
1´q
Taking the limit q Ñ 1 and, using Lemma 2.3.1 above, leads to Equation
(2.21).
In other words, limqÑ1 Qn pxq corresponds to the usual Hermite polyno-
mial Hn pxq. For this reason, the polynomials Prn pxq are known as continuous
q-Hermite polynomials1. (Be careful: there are at least two other varieties
of q-Hermite polynomials.) The parameter q plays an important role and so
Prn is usually denoted by Hn px|qq in most references.
1See e.g. https://en.wikipedia.org/wiki/Continuous_q-Hermite_polynomials.

©Nalini Joshi 22
 Special Topics in Applied Mathematics

2.4. Introduction to Random Matrices

When we have a large system of interacting atomic particles, the question
of how to model individual wave functions becomes very complicated and
almost impossible. But, although we cannot predict what happens in detail
to one specific particle in such a system, we can study the macroscopic
properties of the system by following an idea introduced by Eugene Wigner.
Wigner was motivated by a need to model the atomic spectra of heavy atoms.
He suggested that one such model is given by the distribution of eigenvalues
of random matrices, that is, matrices with random entries.

Figure 2.5. Eugene Wigner in 1963, when he won the Nobel

prize for physics. By Nobel foundation, Public Domain,
commons.wikimedia.org/w/index.php?curid=6141135 .

To explain this idea, we start by recalling some basic concepts. The

matrices of interest will depend on continuous random variables, that is,
variables that take values in a continuous interval.
2.4.1. Definition. Let X be a random variable. It has a continuous prob-
ability distribution function (PDF) ⇢ : R Ñ r0, 8q if for every interval
I “ ra, bs Ä R, the probability of X belonging to I is given by the inte-
gral of ⇢ over I:
ªb
P ra § X § bs “ ⇢pxq dx.
a
Moreover, the cumulative density function is defined by
ªx
F pxq “ P r´8 § X § xs “ ⇢pxq dx.
´8
2.4.2. Remark. Note that F pxq has many properties:
(i) F is non-decreasing.
(ii) 0 § F pxq § 1.
(iii) limxÑ´8 F pxq “ 0 and limxÑ`8 F pxq “ 1
2.4.3. Definition. Let X1 , X2 be two continuous random variables. They
have a “joint probability distribution function” (JPDF) ⇢ : R2 Ñ r0, 8q ˆ
r0, 8q if for every domain I1 ˆ I2 Ä R2 , the probability of X1 belonging to

23 ©Nalini Joshi
 February 24, 2022 DRAFT

I1 and X2 belonging to I2 is given by

ª ª
⇢px1 , x2 q dx1 dx2 .
I1 I2
If a JPDF is factorized, i.e., ⇢px1 , x2 q “ ⇢1 px1 q⇢2 px2 q, then the variables are
said to be independent. Otherwise, they are dependent.
If , X1 , X2 are independent and in addition we have ⇢1 pxq “ ⇢2 pxq, then
the random variables are called independent and identically distributed (IID).
2.4.4. Note. There are many possible probability distribution functions. A
common example is the Gaussian or normal distribution:
` ˘2
1 1
⇢pxq “ ? e´ 2 px´µq{ ,
2⇡
where µ is the mean and 2 is the variance of the distribution. See Figure
2.6 for examples of normal probability distribution functions.

0.5

0.4

μ=-1, σ=0.75
0.3
μ=0, σ=1.5
0.2 μ=1, σ=3

0.1

-5 5

Figure 2.6. Three examples of Gaussian or normal probabil-

ity distribution functions.

Random matrices, i.e., matrices with random variables as entries, orig-

inate in the study of genetics in the 1930s [Hsu]. We will focus here on
ensembles of random matrices with entries that are given by a Gaussian dis-
tribution. Note that this means each matrix H has entries given by random
variables Hkj .
There are three types of matrix ensembles:
(i) Symmetric matrices, which form the Gaussian Orthogonal Ensem-
ble (GOE);
(ii) Hermitian matrices, which form the Gaussian Unitary Ensemble
(GUE);
(iii) Hermitian quarternionic matrices, which form the Gaussian Sym-
plectic Ensemble.
Consider the first type: GOEs.
2.4.5. Example. We can generate examples of random matrices by using
symbolic algebra packages. For example, the command
RandomVariate[GaussianOrthogonalMatrixDistribution[2]]

©Nalini Joshi 24
 Special Topics in Applied Mathematics

in Mathematica leads to samples of 2 ˆ 2 GOE random matrices such as

ˆ ˙ ˆ ˙ ˆ ˙
´1.34594 1.15733 ´0.393224 ´1.00721 ´1.80379 0.426458
, , .
1.15733 ´0.891689 ´1.00721 1.36485 0.426458 ´0.69862
2.4.6. Note. Assume H is a real symmetric matix in a GOE. Then the
following statements hold.
(i) The ensemble is invariant under
Ä “ W T HW.
H ބ H
where W is any real orthogonal matrix, that is, W T W “ W W T “ I.
(ii) Since W T is the inverse of W , we have trpHq “ trpHq.Ä
(iii) The entries Hij , i § j, of H are statistically independent.
(iv) Suppose±P pHqdH is the probability that H lies in a volume element
dH “ i § jdHij . Then
Ä H
P pHqd Ä “ P pHqdH.
(v) From the above, we can also show that
π
P pHq “ fij pHij q.
i§j

Similar arguments for the remaining ensembles lead to the following re-
sult [RS1960, W28].
Theorem 2.4.7. The probability P pHqdH that a random ± matrix H (in a
GOE, GUE or GSE) will lie in a volume element dH “ i § jdHij is given
by ` ˘
P pHq “ exp ´atrH 2 ` btrH ` c ,
where a, b, c are real numbers and a ° 0.
This theorem allows us to write down a description of the joint probabil-
ity distribution function for eigenvalues of random matrices. Let x1 , . . . , xn
be the eigenvalues of an N ˆ N matrix. Then the joint probability distribu-
tion of x1 , . . . , xn is given by [Mehta, Theorem 3.3.1]
˜ ¸
ÿN π
1 2
⇢ ,N px1 , . . . , xn q “ exp ´ xj |xj ´ xk | , (2.27)
Z ,N 2 j“1 j†k
where “ 1 for GOE, “ 2 for GUE, and “ 4 for GSE.
Here Z ,N , called the “partition function”, is given by
ª ´ ÿN ¯π
Z ,N “ C ,N exp ´N x2j |xj ´ xk | dx1 . . . dxN , (2.28)
RN 2 j“1 j†k
where C ,N is chosen to ensure that ⇢ is normalized to be a probability
distribution, by requiring
ª ª
. . . ⇢ ,N dx1 . . . dxN “ 1.

2.4.8. Note. Be careful with conventions for terminology! There is a va-

riety of different conventions used in the literature, which may change the
constants given in such formulas in different references.

25 ©Nalini Joshi
 February 24, 2022 DRAFT

We will focus on Mehta’s integral:

ª ´ ÿN ¯π
r
Z ,N “ exp ´N x2j |xj ´ xk | dx1 . . . dxN . (2.29)
RN 2 j“1 j†k
It is related to the famous Selberg integral, whose explicit calculation was
of interest to those pursuing elementary methods in number theory. (See
Figure 2.7 for a photograph of Atle Selberg – I was sorry not to be able to
find a good picture of ML Mehta, who was a very influential figure in random
matrix theory.)

Figure 2.7. Atle Selberg who was was awarded the Fields
medal in 1950 for his work in number theory. By
Konrad Jacobs, Erlangen - Mathematisches Institut Oberwolfach,
https://opc.mfo.de/detail?photoID=3792,CCBY-SA2.0de.

Now we come to a major observation, which ties random matrix theory

closely to the theory
± of orthogonal polynomials. This link is built on the fact
that the product j†k |xj ´ xk | in ⇢ ,N and Z ,N can be expressed as the
determinant of a Vandermonde matrix.
2.4.9. Definition. A Vandermonde matrix is a matrix with a geometric
progression in each row (or column).
2.4.A. Exercise. Consider an N ˆ N Vandermonde matrix V .
(i) Suppose N “ 3 so that
¨ ˛
1 1 1
V “ ˝x1 x2 x3 ‚.
x21 x22 x23
Show that π
detpV q “ pxj ´ xk q.
1§j†k§3
(ii) Show that
π
detpV q “ pxj ´ xk q.
1§j†k§N

©Nalini Joshi 26
 Special Topics in Applied Mathematics

By using row (or column) operations, we can show that determinants of

Vandermonde matrices can be re-expressed in terms of orthogonal polyno-
mials.
2.4.B. Exercise. Consider the Vandermonde matrix in Exercise 2.4.A(i).
(i) Show that
¨ ˛
P0 px1 q P0 px2 q P0 px3 q
detpV q “ det ˝P1 px1 q P1 px2 q P1 px3 q‚,
P2 px1 q P2 px2 q P2 px3 q
for Hermite polynomials tPn pxqu8n“0 .
(ii) Relate detV to the determinant of
¨ ˛
L0 px1 q L0 px2 q L0 px3 q
˝L1 px1 q L1 px2 q L1 px3 q‚,
L2 px1 q L2 px2 q L2 px3 q
where Lj are given by the first three Laguerre polynomials:
L0 pxq “ 1, L1 pxq “ ´x ` 1, L2 pxq “ x2 {2 ´ 2x ` 1.
The expression of Vandermonde determinants in terms of orthogonal
polynomials enables us to calculate Mehta’s integral explicitly.
2.4.C. Exercise. Consider Zr2,N defined with “ 2 in Equation (2.29).
(i) Consider the case N “ 2. Show by using Hermite polynomials
P0 pxq, P1 pxq and their squared norms h0 , h1 that Zr2,2 can be ex-
pressed in terms of the product of h0 h1 .
(ii) Can you extend the above calculation to get Zr2,N for any positive
integer N ?
A calculation of Zr1,N is also possible, but it is more complicated
± than
the one presented in in Exercise 2.4.C because the product j†k |xj ´ xk |
has an absolute value sign.
2.4.D. Exercise. In [Mehta], a calculation of Zr1,N for even N is carried
out by transforming to a slightly different perspective, given by defining 'n
as follows:
x2 {4
Pn pxq “ h1{2
n 'n pxqe ,
where tPn pxqu are Hermite polynomials.
(i) Show that ª
8
'n pxq'm pxq “ nm .
´8
(ii) By using the 3-term recurrence relation for tPn pxqu and the relation
between Pn1 and Pn´1 , show that
? ?
2'1n “ n'n´1 ´ n ` 1'n`1 .
(iii) Replace '2j`1 by '12j by using the relation in the above exercise,
and show that for even N ,
˜ ¸
ÿN π ` ˘
exp ´ x2i {2 pxi ´ xj q “ c det '2j´2 pxi q, '12j´2 pxi q ,
i“1 j†i

where c is a constant and 1 § i § N , 1 § j § N {2.

27 ©Nalini Joshi