See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/238203450

Classification of classical orthogonal polynomials

Article in Journal of the Korean Mathematical Society · January 1997

CITATIONS READS
65 2,044

2 authors, including:

Lance Littlejohn
Baylor University
143 PUBLICATIONS 2,390 CITATIONS

SEE PROFILE

All content following this page was uploaded by Lance Littlejohn on 24 June 2014.

The user has requested enhancement of the downloaded file.

J. Korean Math. Soc. 34 (1997), No. 4, pp. 973–1008

CLASSIFICATION OF CLASSICAL
ORTHOGONAL POLYNOMIALS

KIL H. KWON AND LANCE L. LITTLEJOHN

ABSTRACT. We reconsider the problem of classifying all classical orthogo-

nal polynomial sequences which are solutions to a second-order differential
equation of the form

`2 (x)y 00 (x) + `1 (x)y 0 (x) = λn y(x).

We first obtain new (algebraic) necessary and sufficient conditions on the coef-
ficients `1 (x) and `2 (x) for the above differential equation to have orthogonal
polynomial solutions. Using this result, we then obtain a complete classifi-
cation of all classical orthogonal polynomials : up to a real linear change of
variable, there are the six distinct orthogonal polynomial sets of Jacobi, Bessel,
Laguerre, Hermite, twisted Hermite, and twisted Jacobi.

1. Introduction
All polynomials in this work are assumed to be real polynomials in the
real variable x and we let P be the space of all these real polynomials.
We denote the degree of a polynomial π(x) by deg(π ) with the convention
that deg(0) = −1. By a polynomial system (PS), we mean a sequence of
polynomials {φn (x)}∞
n=0 with deg(φn ) = n, n ≥ 0. Note that a PS forms a
basis for P .
A PS {φn (x)}∞
n=0 is called orthogonal if there is a function µ : R → R of
bounded variation on the real line R such that
Z
(1.1) x n dµ(x)
R

is finite for all n = 0, 1, . . . and

Z
(1.2) φm (x)φn (x) dµ(x) = K n δmn (m and n ≥ 0),
R

Received June 25, 1997.

1991 Mathematics Subject Classification: Primary 33A65.
Key words and phrases: classical orthogonal polynomials, second-order differential equations.
974 Kil H. Kwon and Lance L. Littlejohn

where K n are non-zero real constants and δmn is the Kronecker delta function.
Furthermore, we shall say that {φn (x)}∞ n=0 is classi cal if each φn (x) (n ≥ 0)
satisfies a fixed second-order differential equation of the form

(1.3) L[y](x) = `2 (x)y 00 (x) + `1 (x)y 0 (x) = λn y(x),

where `2 (x) and `1 (x) are real-valued functions independent of n and λn is a

real constant depending only on n. The classification of classical orthogonal
polynomials is generally attributed to Bochner [3]. In fact, Bochner [3]
considered a general second-order Sturm-Liouville differential equation of
the form

(1.4) a2 (x)y 00 (x) + a1 (x)y 0 (x) + a0 (x)y(x) + λy(x) = 0,

where ai (x) (i = 0, 1, 2) are real- or complex-valued functions and λ is a

constant. He then raised and solved the problem : determine all cases such
that for each integer n ≥ 0, there is an eigenvalue λ = λn for which there
is a corresponding polynomial solution of degree n. He first observed that if
the differential equation (1.4) has polynomial solutions of degree 0, 1, and
2, then ai (x) must be a polynomial of degree ≤ i , i = 0, 1, 2. He then
considered cases according to the degree of a2 (x) and, in each case, reduced
the differential equation into a normal form by a suitable complex linear
change of variable. Then, through a detailed analysis of each case, Bochner
showed that up to a complex linear change of variable, the only PS’s that arise
as eigensolutions of the differential equation (1.4) are the following (apart
from non-zero constant factors) :
(α,β)
(a) Jacobi polynomials {Pn (x)}∞ n=0 (α, β, α + β + 1 ∈ / {−1, −2,
. . . }) ;
∞
(b) Laguerre polynomials {L (α)
n (x)}n=0 (α ∈
/ {−1, −2, . . . }) ;
∞
(c) Hermite polynomials {Hn (x)}n=0 ;
(d) {x n }∞n=0 ;
(α,β)
(e) Bessel polynomials {Bn (x)}∞ n=0 (α ∈
/ {0, −1, −2, . . . } and β 6=
0).
The orthogonality of the Jacobi polynomials for α and β > −1, Laguerre
polynomials for α > −1, and Hermite polynomials was known long before
Bochner’s work. In fact, Bochner [3] did not mention the orthogonality of
the PS’s that he found. The problem of classifying all classical orthogonal
 Classical orthogonal polynomials 975

polynomials was handled by many authors thereafter : see, for example, [1],
[5], [6], [11], [27], and [31]. The problem was settled by Lesky [27] in 1962 at
least for classical orthogonal polynomials satisfying the orthogonality relation
(1.2) in which the function µ(x) is non-decreasing. Lesky [27] showed that
the only such orthogonal polynomials are Jacobi polynomials with α and
β > −1, Laguerre polynomials with α > −1, and Hermite polynomials.
It is easy to see that the PS {x n }∞
n=0 in case (d) above cannot be orthogonal.
The orthogonality of the Bessel polynomials was first observed by H.L. Krall
[18] and later investigated in depth by Krall and Frink [19]. Bochner [3]
observed the relation between the PS in case (e) above and the half-integer
Bessel functions and it is this relation which motivates the name Bessel
polynomials in [19]. The orthogonality of the Jacobi polynomials for α or
β < −1 and Laguerre polynomials for α < −1 was recently treated by
Morton and Krall [32].
A natural question arises : are these four PS’s of Jacobi, Laguerre, Hermite,
and Bessel the only classical orthogonal polynomials? Of course, if we allow
for a complex linear change of variable, as Bochner does in [3], the answer is
yes. However, if we restrict our attention to a real linear change of variable, as
we shall do in this paper, are there any more classical orthogonal polynomials?
As far as the authors know, no previous work on this classification problem
really exhausts all possibilities.
After obtaining necessary and sufficient conditions (see Theorem 2.9) for
the differential equation (1.3) to have orthogonal polynomials of solutions
in section two, we give a complete classification of classical orthogonal
polynomials in section three. Finally, in section four, we will discuss the
integral or distributional representation of orthogonality for each classical
orthogonal polynomial system found in section three.

2. Necessary and sufficient conditions

We call any linear functional σ on P a moment functional and denote its
action on a polynomial π(x) by hσ, π i. We define the nth moment of σ by
hσ, x n i (n = 0, 1, . . . ).
We shall remind the reader in section four below that any moment func-
tional σ has a representation of the form
Z
hσ, π i = π(x) dµ(x) (π ∈ P ),
R
976 Kil H. Kwon and Lance L. Littlejohn

or
Z
hσ, π i = π(x)φ(x) d x (π ∈ P ),
R

where µ(x) is, in general, a function of bounded variation on R and where

φ(x) is a C ∞ -function of the Schwartz class. Hence, the orthogonality relation
in (1.2) can be expressed in terms of moment functionals. As we shall see,
it is very convenient and advantageous to use moment functionals instead of
using their integral representations in discussing orthogonal polynomials.
We say that a moment functional σ is quasi-definite (respectively, positive-
definite) if its moments {σn }∞n=0 satisfy the Hamburger condition

(2.1) 1n (σ ) = det[σi+ j ]ni, j =0 6= 0 (respectively, 1n (σ ) > 0)

for every n ≥ 0.
Any PS {Pn (x)}∞ n=0 determines a moment functional σ (uniquely up to
a non-zero constant multiple), called a canonical moment functional for
{Pn (x)}∞
n=0 , by the conditions

(2.2) hσ, P0 i 6= 0 and hσ, Pn i = 0, n ≥ 1.

DEFINITION 2.1. A PS {Pn (x)}∞n=0 is called a weak orthogonal polynomial

system (WOPS) if there is a non-trivial moment functional σ such that

(2.3) hσ, Pm Pn i = 0 if m 6= n (m and n ≥ 0).

If we further have

(2.4) hσ, Pm Pn i = K n δmn ,

where K n are non-zero real constants, then we call {Pn (x)}∞

n=0 an orthogonal
polynomial system (OPS). If each K n > 0, then we call {Pn (x)}∞n=0 a positive-
∞
definite OPS. In either case, we say that {Pn (x)}n=0 is a WOPS or an OPS
relative to σ and call σ an orthogonalizing moment functional of {Pn (x)}∞ n=0 .

It is immediate from the orthogonality (2.3) that for any WOPS {Pn (x)}∞
n=0 ,
its orthogonalizing moment functional must be a canonical moment functional
for {Pn (x)}∞n=0 so that it is unique up to a non-zero constant multiple.
 Classical orthogonal polynomials 977

It is well known (for example, see [4, Chapter 1]) that a moment functional
σ is quasi-definite (respectively, positive-definite) if and only if there is an
OPS (respectively, a positive-definite OPS) relative to σ . It is clear that
if {Pn (x)}∞ ∞
n=0 is an OPS relative to σ , then so is {C n Pn (x)}n=0 for every
sequence of non-zero constants Cn . Conversely if σ is any quasi-definite
moment functional and {Pn (x)}∞ n=0 is an OPS relative to σ , then each Pn (x)
is uniquely determined up to an arbitrary non-zero factor. In particular, for
any quasi-definite moment functional σ , there is a unique monic OPS relative
to σ given by
 
σ0 σ1 ... σn
 σ1 σ2 ... σn+1 
1  .. .. .. .. 
(2.5) Pn (x) = det 
 . . . .

 (n ≥ 0),
1n−1 (σ )  
σn−1 σn ... σ2n−1
1 x ... xn

where 1−1 (σ ) = 1 (see [4, Chapter 1]).

We shall call an OPS {Pn (x)}∞ n=0 a classi cal OP S if for each n ≥ 0,
Pn (x) satisfies the differential equation (1.3) for some eigenparameter λn . As
mentioned in the introduction, if the differential equation (1.3) has a PS of
solutions, then it is necessary that the coefficients `2 (x), `1 (x), and λn be
given by

X
i
`i (x) = `i j x j (i = 1, 2),
(2.6) j =0
λn = n(n − 1)`22 + n`11 (n ≥ 0),

where `211 + `222 6= 0.

From here on, we shall assume that the differential equation (1.3) has
coefficients given by (2.6).
In 1938, H.L. Krall [17] obtained necessary and sufficient conditions for
an OPS to satisfy a Sturm-Liouville type differential equation of any order.
In case of the second-order differential equation (1.3), Krall’s result can be
stated as :
THEOREM 2.1. A PS {Pn (x)}∞ n=0 is an OPS (respectively, a positive-definite
OPS) satisfying the differential equation (1.3) if and only if its canonical
978 Kil H. Kwon and Lance L. Littlejohn

moment functional σ is quasi-definite (respectively, positive-definite) and the

moments {σn }∞n=0 of σ satisfy the recurrence relation
(2.7)
(n`22 + `11 )σn+1 + (n`21 + `10 )σn + n`20 σn−1 = 0 (n ≥ 0 ; σ−1 = 0).

For a new and somewhat simpler proof of Theorem 2.1, see [23] ; for
another proof of the general Krall characterization theorem, see [20] and
[25].
We call the recurrence relation (2.7) the moment equati on for the dif-
ferential equation (1.3). We may use Theorem 2.1 to classify all possible
classical OPS’s. However it is very difficult, in general, to solve the moment
equation (2.7) and to see whether the corresponding moment functional is
quasi-definite or not. The disadvantage of the conditions in Theorem 2.1 is
that the equation (2.7) contains not only the coefficients of (1.3) but also the
moments of a canonical moment functional of a classical OPS of which the
existence is not known apriori.
Below, we shall first obtain a necessary condition (see Theorem 2.5) and
then necessary and sufficient conditions (see Theorem 2.9) for the differential
equation (1.3) to have an OPS of solutions. Unlike those in Theorem 2.1,
these conditions involve only the coefficients of the differential equation (1.3).
We begin with introducing some formal calculus on moment functionals.
For a moment functional σ and π ∈ P , we let σ 0 , the derivative of σ and π σ ,
multiplication of σ by a polynomial, be those moment functionals defined by

(2.8) hσ 0 , pi = −hσ, p 0 i ( p ∈ P)

and

(2.9) hπ σ, pi = hσ, π pi ( p ∈ P ).

It is then easy to obtain the following Leibnitz rule for any moment functional
σ and polynomial π(x) :

(2.10) (π σ )0 = π 0 σ + π σ 0 .

LEMMA 2.2. Let σ be a moment functional and π(x) a polynomial.

(i) Then σ = 0 if and only if σ 0 = 0.
(ii) If σ is quasi-definite, then π(x)σ = 0 if and only if π(x) = 0.
 Classical orthogonal polynomials 979

Proof. (i) If σ 0 = 0, then

1 −1
hσ, x n i = hσ, (x n+1 )0 i = hσ, x n+1 i = 0
n+1 n+1
for every n ≥ 0 so that σ = 0. The converse is trivial.
(ii) Assume σ is quasi-definite and π(x)σ = 0. Let {Pn (x)}∞
n=0 be an
OPS relative to σ . Suppose π(x) 6≡ 0 so that deg(π ) = N ≥ 0 and write
XN
π(x) = Ck Pk (x) with C N 6= 0. Then we have
k=0

X
K
0 = hπ σ, PN i = Ck hσ, Pk PN i = C N hσ, PN2 i
k=0

so that C N = 0 since hσ, PN2 i 6= 0, contradicting our assumption. The

converse is trivial.

LEMMA 2.3. If the differential equation (1.3) has a PS of solutions, then

any canonical moment functional σ of this PS satisfies the functional equation

(2.11) (`2 (x)σ )0 − `1 (x)σ = 0.

Proof. Suppose that {Pn (x)}∞n=0 is a PS of solutions of the differential

equation (1.3). Let σ be a canonical moment functional for this PS. Then we
have for each integer n ≥ 1,

0 = λn hσ, Pn i = hσ, λn Pn i = hσ, `2 Pn00 + `1 Pn0 i = h`1 σ − (`2 σ )0 , Pn0 i,

which implies (2.11) since {Pn0 (x)}∞

n=1 is also a PS.

Note that the zero in the right hand side of the equation (2.11) means the
zero moment functional. In other words, the equation (2.11) means

h(`2 σ )0 − `1 σ, x n i = 0 (n ≥ 0),

which is exactly the moment equation (2.7) when it is expressed in terms of

the moments {σn }∞n=0 of σ .
We call the equation (2.11) the weight equation for the differential equation
(1.3).
980 Kil H. Kwon and Lance L. Littlejohn

REMARK 2.1. If we view the equation (2.11) as a classical differential

equation:
(2.12) (`2 (x)s)0 − `1 (x)s = 0,
then any non-trivial solution s(x) of (2.12) is a symmetr y f actor (see [29])
of the differential expression L[·] in (1.3). In this sense, we call the equation
(2.12) the symmetr y equati on of the differential expression L[·]. For more
details on symmetry factors, symmetry equations, and their applications to
orthogonal polynomials, see [16], [25], [28], and [29].
It is natural to ask if the differential equation (1.3) always has a PS of
solutions. By direct calculation, it is easy to see that (1.3) has a unique monic
polynomial solution of degree n for each integer n ≥ 0 except possibly for
a finite number of values of n and, for those exceptional cases of n (if there
is any), there may be no polynomial solution of degree n or there will be
infinitely many monic polynomial solutions of degree n.
EXAMPLE. Consider the following second-order differential equation :
(2.13) L[y](x) = (1 + x 2 )y 00 (x) + [(1 − k)x + b]y 0 (x) = n(n − k)y(x),
where k ≥ 1 is an integer and b is a real constant. Now it is easy to see that
the equation (2.13) has a PS of solutions if and only if k is odd and b = 0.
Moreover when k = 2 j +1, j ≥ 0 and b = 0, the equation (2.13) has a unique
monic polynomial solution of degree n for n ∈ / { j + 1, j + 2, . . . , 2 j + 1}.
For n ∈ { j + 1, j + 2, . . . , 2 j + 1}, it has infinitely many monic polynomial
solutions of degree n.
DEFINITION 2.2 (Krall and Sheffer [21]). The differential expression L[·]
in (1.3) (or the differential equation (1.3) itself) is called admissible if
(2.13) λm 6= λn for m 6= n (m and n ≥ 0).
LEMMA 2.4. For the differential expression L[·] in (1.3), the following are
equivalent :
(i) L[·] is admissible ;
(ii) λn = n(n − 1)`22 + n`11 6= 0 (n ≥ 1) ;
(iii) `11 ∈
/ {−n`22 | n = 0, 1, 2, . . . } ;
(iv) The moment equation (2.7) (or equivalently the weight equation
(2.11)) has only one linearly independent solution ;
(v) For each n ≥ 0, the differential equation (1.3) has a unique monic
polynomial solution of degree n .
 Classical orthogonal polynomials 981

Proof. The proofs of (i)⇒(ii) and (ii)⇔(iii)⇔(iv) are trivial.

(ii)⇒(i) : This follows from the identity

(n + m)(λn − λm ) = (n − m)(n + m)(`22 (n + m − 1) + `11 ) = (n − m)λn+m .

(i)⇒(v) : For any integer n ≥ 1, let

X
n
Pn (x) = Ckn x k (Cnn = 1)
k=0

be a monic polynomial of degree n. Then Pn (x) satisfies (1.3) if and only if

(2.14) `20 (k + 2)(k + 1)Ck+2

n
+ (k + 1)(`21 k + `10 )Ck+1
n
+ (λk − λn )Ckn = 0

(k = 0, 1, . . . , n − 1), where Cn+1

n
= 0. If L[·] is admissible, then all
Ck (k = 0, 1, . . . , n − 1) are uniquely and successively determined by the
n

equation (2.14) and our assumption that Cnn = 1.

(v)⇒(i) : Assume that the differential equation (1.3) has a unique monic
PS {Pn (x)}∞n=0 of solutions but L[·] is not admissible. Hence from (ii), we
have λ N = λ0 = 0 for some integer N ≥ 1. But then

L[PN + k P0 ] = λ N PN + kλ0 P0 = 0 = λ N (PN + k P0 )

for any constant k. Hence L[y] = λ N y has infinitely many monic polynomial
solutions of degree N , which contradicts our assumption.

REMARK 2.2. Let N ≥ 0 be the largest integer such that λ N = 0. Then

for any n > N the differential equation (1.3) can have only one linearly
independent polynomial solution of degree n.
REMARK 2.3. When `2 (x) ≡ 0, the differential equation (1.3) reduces to
the first-order equation

(`11 x + `10 )y 0 (x) = n`11 y(x),

which is admissible if and only if `11 6= 0. In this case, the corresponding

weight equation is
(`11 x + `10 )σ = 0,
982 Kil H. Kwon and Lance L. Littlejohn

of which the general solution is

σ = cδ(`11 x + `10 ),

where c is an arbitrary constant and δ(`11 x + `10 ) is the Dirac delta moment
functional defined by

hδ(`11 x + `10 ), π(x)i = π(−`10 /`11 ) (π ∈ P ).

Since σ is not quasi-definite, we can conclude that the above first-order

differential equation can never have an OPS of solutions (see Theorem 2.1).
By Remark 2.3, we may assume `2 (x) 6≡ 0 in the differential equation
(1.3). Now we are ready to give a necessary condition for the differential
equation (1.3) to have an OPS of solutions, which will be very useful in our
classification in the next section.
THEOREM 2.5. If the differential equation (1.3) has an OPS {Pn (x)}∞
n=0 of
solutions, then L[·] is admissible.
Proof. Assume that (1.3) has an OPS {Pn (x)}∞ n=0 of solutions and let σ be
an orthogonalizing moment functional of {Pn (x)}∞ n=0 . Then σ is a canonical
∞
moment functional of {Pn (x)}n=0 and, by Lemma 2.3, σ satisfies the weight
equation (2.11). If L[·] is not admissible, then, by Lemma 2.4 (ii), there is an
integer N ≥ 1 such that λ N = 0. Consequently, we have

0 = λ N PN σ = (`2 PN00 + `1 PN0 )σ

= (`2 PN0 σ )0 − PN0 (`2 σ )0 + PN0 (`1 σ ) = (`2 PN0 σ )0 .

Hence, by Lemma 2.2, we have `2 PN0 ≡ 0. However, `2 (x) 6≡ 0 (see

Remark 2.3) so that PN0 (x) ≡ 0, which implies N = 0 contradicting the fact
that N ≥ 1.

Theorem 2.5 was first proved by Lesky [27] only for positive-definite
classical OPS’s. However his method of proof cannot be extended to general
classical OPS’s since he used the following fact which holds only for positive-
definite OPS’s : for any positive-definite OPS {Pn (x)}∞
n=0 , the zeros of Pn (x),
n ≥ 1, are real and distinct and no two polynomials from {Pn (x)}∞ n=0 can
have common zeros (see Chihara [4]).
 Classical orthogonal polynomials 983

The converse of Theorem 2.5 does not hold in general. For example, the
PS {x n }∞
n=0 satisfies the admissible differential equation

x 2 y 00 (x) + x y 0 (x) = n 2 y(x)

but {x n }∞
n=0 is not an OPS. However, we have the following partial converse
of Theorem 2.5.
THEOREM 2.6. If the differential operator L[·] in (1.3) is admissible, then
any PS {Pn (x)}∞
n=0 of solutions to the differential equation (1.3) is a WOPS.

Proof. By Lemma 2.4, we may assume that {Pn (x)}∞ n=0 is the unique
monic PS of solutions to (1.3). Let σ be a canonical moment functional of
{Pn (x)}∞
n=0 . Then σ 6= 0 by definition and, by Lemma 2.3, σ satisfies the
weight equation (2.11). Then we have for m and n ≥ 0

(λm − λn )Pm Pn = `2 (Pm0 Pn − Pm Pn0 )0 + `1 (Pm0 Pn − Pm Pn0 )

0
= `2 Wm,n + `1 Wm,n ,

where Wm,n = Pm0 Pn − Pm Pn0 is the Wronskian of Pm and Pn . Hence, by

(2.11), we have
0
(λm − λn )hσ, Pm Pn i = hσ, `2 Wm,n + `1 Wm,n i
= h`1 σ − (`2 σ )0 , Wm,n i = 0.

Consequently, hσ, Pm Pn i = 0 for m 6= n if L[·] is admissible.

REMARK 2.4. In fact we can prove, by the same reasoning as in the proof
of Theorem 2.6, something more than Theorem 2.6. If L[ p] = λp and
L[q] = µq for some polynomials p(x) and q(x) and λ 6= µ, then hσ, pqi = 0
for any moment functional solution σ of the weight equation (2.11). Here we
do not need to assume L[·] is admissible.
We now seek a criterion for when a WOPS {Pn (x)}∞ n=0 is an OPS, which
does not involve a canonical moment functional of {Pn (x)}∞
n=0 .
For any monic PS {Pn (x)}n=0 , there are constants {αn }∞
∞ ∞
n=1 and {βn }n=1
such that

(2.15) Pn+1 (x) − (x − αn )Pn (x) + βn Pn−1 (x) (n ≥ 1)

984 Kil H. Kwon and Lance L. Littlejohn

Pn
is a polynomial of degree ≤ n − 2. In fact if Pn (x) = k=0 Ckn x k (Cnn = 1 ;
n ≥ 1), then

(2.16)
αn = Cn−1
n
− Cnn+1

and

(2.17)
βn = Cn−2
n
− (Cn−1
n
− Cnn+1 )Cn−1
n
− Cn−1
n+1
(C−1
1
= 0).

At this moment, we need to recall Favard’s theorem (see [10]) which asserts
that a monic PS {Pn (x)}∞ n=0 is an OPS (respectively, a positive-definite OPS)
∞
if and only if {Pn (x)}n=0 satisfies a three term recurrence relation

(2.18) Pn+1 (x) = (x − αn )Pn (x) − βn Pn−1 (x) (n ≥ 1),

where each βn 6= 0 (respectively, βn > 0).

In the case of WOPS’s, Favard’s theorem can be improved as follows.

PROPOSITION 2.7 (Krall and Sheffer [21]). A monic WOPS {Pn (x)}∞
n=0 is an
OPS (respectively, a positive-definite OPS) if and only if

(2.19) βn 6= 0 (respectively, βn > 0)

for n ≥ 1, where βn is the constant given in (2.15).

Proof. See Lemma 1.1 in [21].

Once we know a PS {Pn (x)}∞ ∞

n=0 is a WOPS (it is so if {Pn (x)}n=0 satisfies
an admissible equation (1.3) : see Theorem 2.6), the advantage of applying
Proposition 2.7 over Favard’s theorem is evident. In order to check condition
(2.19), we only need to know the coefficients of x n−1 and x n−2 of each Pn (x)
from a monic WOPS {Pn (x)}∞ n=0 . More precisely, we have the following
result from Proposition 2.7 and equation (2.17).
 Classical orthogonal polynomials 985

∞
X
COROLLARY 2.8. Let {Pn (x)}∞
n=0 be a WOPS and Pn (x) = Ckn x k
k=0
(Cnn = 1) for n ≥ 0. Then {Pn (x)}∞ n=0 is an OPS (respectively, a positive-
definite OPS) if and only if
(2.20)
βn = Cn−2
n
− (Cn−1
n
− Cnn+1 )Cn−1
n
− Cn−1
n+1
6= 0 (respectively, βn > 0)

for n ≥ 1, where C−1

1
= 0.
Now combining Theorem 2.5, Theorem 2.6, and Corollary 2.8, we can
obtain necessary and sufficient conditions for the differential equation (1.3)
to have an OPS of solutions in terms of only the coefficients of the differential
expression L[·].
THEOREM 2.9. The differential equation (1.3) has an OPS (respect- ively,
a positive-definite OPS) of solutions if and only if
(i) `11 ∈
/ {−n`22 | n = 0, 1, 2, . . . }
and
(ii) the condition (2.20) holds; i.e. βn 6= 0 (r especti vely, βn > 0),
where
n[`10 + `21 (n − 1)]
(2.21) n
Cn−1 =
`11 + 2`22 (n − 1)
and
n(n − 1)[`20 (`11 + 2`22 (n − 1)) + (`10 + `21 (n − 2))(`10 + `21 (n − 1)]
(2.22) n
Cn−2 =
2[`11 + 2`22 (n − 1)][`11 + `22 (2n − 3)]

(n ≥ 1 ; C−1
1
= 0).
Proof. By Lemma 2.4, the above condition (i) is just the admissibility of
L[·] which is also equivalent to the fact that the differential equation (1.3) has a
Xn
∞
unique monic PS {Pn (x)}n=0 of solutions. If we set Pn (x) = Ckn x k (Cnn =
k=0
1 ; n ≥ 0), then Cn−1
n n
and Cn−2 are given by (2.21) and (2.22), respectively, by
solving the equation (2.14) for k = n − 1 and k = n − 2. Hence, Theorem 2.9
follows from Theorem 2.5, Theorem 2.6, and Corollary 2.8.

We end this section by the following remark.

986 Kil H. Kwon and Lance L. Littlejohn

REMARK 2.5. If we assume that the differential equation (1.3) has a monic
PS {Pn (x)}∞ ∞
n=0 of solutions, then {Pn (x)}n=0 is an OPS if and only if the
condition (2.20) holds. For a proof of this statement, see [24, Proposition 3.7].
Note here that apriori we do not assume that {Pn (x)}∞ n=0 is a WOPS (as in
Proposition 2.7) or L[·] is admissible (as in Theorem 2.9). Furthermore, only
condition (2.2) must be checked but not with the conditions given in (2.21)
and (2.22). In general, these latter two equations are not well defined unless
the expression L[·] is admissible.

3. Classification
We say that any two OPS’s are equivalent if either one differs from the
other by non-zero constant factors or one is obtained from the other by a real
linear change of variable.
In this section, we will classify all classical OPS’s up to equivalence classes
using Theorem 2.9.
In the following, we let N be the set of all positive integers and use the
notation
   
α α α(α − 1) · · · (α − k + 1)
= 1 and =
0 k k!
for any complex number α and any integer k in N. As with Bochner, we
divide the cases according to the roots of the leading coefficient `2 (x) of the
differential expression L[·] in (1.3).
Cases 1: Jacobi polynomials
We assume `22 6= 0 and `221 − 4`22 `20 > 0. Then, by a real linear change
of variable, the equation (1.3) can be transformed into
L[y](x) = (1 − x 2 )y 00 (x) + [(β − α) − (α + β + 2)x]y 0 (x)
(3.1)
= − n(n + α + β + 1)y(x).
We assume −(α + β + 1) ∈ / N so that L[·] in (3.1) is admissible. Then the
(α,β)
equation (3.1) has a unique monic PS {Pn (x)}∞ n=0 , called the Jacobi PS, of
solutions :
(3.2)
  n   
(α,β) 2n + α + β −1 X n + α n + β
Pn (x) = (x − 1)n−k (x + 1)k
n k=0
k n−k
(n ≥ 0).
 Classical orthogonal polynomials 987

(α,β)
PROPOSITION 3.1. The Jacobi PS {Pn (x)}∞ n=0 is
(i) a WOPS if −(α + β + 1) ∈ /N;
(ii) an OPS if and only if −α , −β , and −(α + β + 1) ∈/ N:
(iii) a positive-definite OPS if and only if α and β > −1.
Proof. The proof of (i) follows from Theorem 2.6. Now we assume
−(α + β + 1) ∈/ N. We then have, from (2.20), (2.21), (2.22), and (3.1),
(3.3)
4n(α + β + n)(α + n)(β + n)
βn = (n ≥ 1).
(α + β + 2n − 1)(α + β + 2n)2 (α + β + 2n + 1)
Hence, βn 6= 0 for n ≥ 1 if and only if α + n 6= 0 and β + n 6= 0 for n ≥ 1 so
that (ii) follows from Theorem 2.9. To prove (iii), it suffices to show βn > 0
for n ≥ 1 if and only if α and β > −1. If α and β > −1, then every factor
in (3.3) is positive so that βn > 0 for n ≥ 1. Conversely, assume βn > 0
for n ≥ 1 but α < −1 (when β < −1, the proof is essentially the same).
Then, from β1 > 0, we have (β + 1)(α + β + 3) < 0. If β + 1 < 0 and
α + β + 3 > 0, then α + β + 2 < 0 and 0 < α + 2, β + 2 < 1 so that β2 < 0,
which is a contradiction. If β + 1 > 0 and α + β + 3 < 0, then α < −2.
Then, from β2 > 0, we have α + β + 5 < 0 and so α < −4. Continuing
the same process, we have that α < −2k for any integer k ≥ 1, which is
impossible.
(α,β)
The explicit orthogonality of the Jacobi PS {Pn (x)}∞
n=0 for α or β < −1
(but −α and −β ∈ / N) has been treated by Morton and Krall [32].
Case 2: Bessel polynomials
We assume `22 6= 0 and `221 − 4`22 `20 = 0. Then, by a real linear change
of variable, the equation (1.3) can be transformed into
(3.4) L[y](x) = x 2 y 00 (x) + (αx + β)y 0 (x) = n(n + α − 1)y(x).
We assume −(α −1) ∈ / N so that L[·] in (3.4) is admissible. Then the equation
(α,β)
(3.4) has a unique monic PS {Bn (x)}∞ n=0 of solutions :
 n
 x if β = 0
  
(3.5) Bn(α,β) (x) = 1 X
n
n! 0(α + n + k − 1) x k

 β n 0(α + 2n − 1) if β 6= 0
(n − k)! k! β
k=0

(α,β)
(n ≥ 0). When β 6= 0, we call {Bn (x)}∞ n ∞
n=0 the Bessel PS. The PS {x }n=0
is a WOPS by Theorem 2.6 but it cannot be an OPS.
988 Kil H. Kwon and Lance L. Littlejohn

(α,β)
PROPOSITION 3.2. The Bessel PS {Bn (x)}∞ n=0 is an OPS (but not a
positive-definite OPS) if and only if −(α − 1) ∈
/ N and β 6= 0.

Proof. We assume −(α − 1) ∈

/ N . We then have, from (2.20), (2.21),
(2.22), and (3.4),

−nβ 2 (α + n − 2)
(3.6) βn = (n ≥ 1).
(α + 2n − 3)(α + 2n − 2)2 (α + 2n − 1)

Hence βn 6= 0 for n ≥ 0 if and only if β 6= 0 and βn < 0 for n large enough.

Therefore, we have the proposition by Theorem 2.9.

The Bessel PS, as an OPS, was first observed by H.L. Krall [18]. Earlier
these polynomials were discussed by Romanovski [33] and Bochner [3]. In
[19], Krall and Frink studied the Bessel polynomials in detail and found,
explicitly, their complex orthogonality.

Case 3: Laguerre polynomials

We assume `22 = 0 and `21 6= 0. Then by a real linear change of variable,
the equation (1.3) can be transformed into

(3.7) L[y](x) = x y 00 (x) + (α + 1 − x)y 0 (x) = −ny(x).

The differential expression L[·] in (3.7) is admissible and so the equation

(3.7) has a unique monic PS {L (α) ∞
n }n=0 , called the Laguerre polynomials, of
solutions :

n 
X 
n + α (−x)k
(3.8) L (α)
n (x) = (−1) n!
n
(n ≥ 0).
k=0
n−k k!

∞
PROPOSITION 3.3. The Laguerre PS {L (α)
n (x)}n=0 is
(i) a WOPS for every α ;
(ii) an OPS if and only if −α ∈ /N;
(iii) a positive-definite OPS if and only if α > −1.
 Classical orthogonal polynomials 989

Proof. (i) follows from Theorem 2.6. We have from (2.20), (2.21), (2.22),
and (3.7)

(3.9) βn = n(α + n) (n ≥ 1).

Hence βn 6= 0 (respectively, βn > 0) for n ≥ 1 if and only if α + n 6= 0 for

n ≥ 1 (respectively, α > −1) so that (ii) and (iii) follow from Theorem 2.9.

The case α = 0 is the one originally studied by Laguerre [26]. The case
α > −1 is due to Sonine [34] and the generalized Laguerre PS for α < −1
and −α ∈/ N has been recently studied by Morton and Krall [32].
Case 4: Hermite polynomials
We assume `22 = `21 = 0, `20 6= 0, and `11 < 0. Then, by a real linear
change of variable, the equation (1.3) can be transformed into

(3.10) L[y](x) = y 00 (x) − 2x y 0 (x) = −2ny(x).

The differential expression L[·] in (3.10) is admissible and so the equation

(3.10) has a unique monic PS {Hn (x)}∞ n=0 of solutions called the Hermite
polynomials :

X
[n/2]
(−1)k x n−2k
(3.11) Hn (x) = n! (n ≥ 0),
k=0
k! (n − 2k)! 4k

where [x] is the integer part of x.

PROPOSITION 3.4. The Hermite PS {Hn (x)}∞
n=0 is a positive-definite OPS.

Proof. We have from (2.20), (2.21), (2.22), and (3.10)

n
(3.12) βn = (n ≥ 0).
2
Hence, the proposition follows from Theorem 2.9.

Case 5: Twisted Hermite polynomials

Assume `22 = `21 = 0, `20 6= 0, and `11 > 0. Then, by a real linear
change of variable, the equation (1.3) can be transformed into

(3.13) L[y](x) = y 00 (x) + 2x y 0 (x) = 2ny(x).

990 Kil H. Kwon and Lance L. Littlejohn

The differential expression L[·] in (3.13) is admissible and so the equation

(3.13) has a unique monic PS { Ȟn (x)}∞ ∞
n=0 of solutions. We call { Ȟn (x)}n=0
the twi sted Her mi te PS. In order to find Ȟ√n (x) explicitly, we set x = i t
and Ȟn (x) = Ȟn (i t) = i Z n (t) with i = −1. Then Z n (t) is a monic
n

polynomial of degree n and satisfies the Hermite differential equation (3.10)

so that Z n (t) = Hn (t). Hence, we have

X
[n/2]
1 x n−2k
(3.14) Ȟn (x) = i Hn (−i x) = n!
n
(n ≥ 0).
k=0
k! (n − 2k)! 4k

PROPOSITION 3.5. The twisted Hermite PS { Ȟn (x)}∞

n=0 is an OPS but not
a positive-definite OPS.
Proof. We have from (2.20), (2.21), (2.22), and (3.13)

−n
(3.15) βn = (n ≥ 1).
2
Hence, the proposition follows from Theorem 2.9.

Case 6: Twisted Jacobi polynomials

We assume `22 6= 0 and `221 − 4`22 `20 < 0. Then, by a real linear change
of variable, the equation (1.3) can be transformed into

(3.16) L[y](x) = (1 + x 2 )y 00 (x) + (d x + e)y 0 (x) = n(n + d − 1)y(x).

We assume −(d − 1) ∈ / N so that L[·] in (3.16) is admissible. Then the

equation (3.16) has a unique monic PS { P̌n (x; d, e)}∞ n=0 of solutions. We call
∞
{ P̌n (x; d, e)}n=0 the twi sted Jacobi PS. In order to find P̌n (x; d, e) explicitly,
we set x = i t and P̌n (x; d, e) = P̌n (i t; d, e) = i n Z n (t). Then Z n (t) is a
monic polynomial of degree n and satisfies

(1 − t 2 )y 00 (t) + (i e − dt)y 0 (t) = −n(n + d − 1)y(t),

which is the Jacobi differential equation (3.1) when

ie = β − α and d = α + β + 2.
 Classical orthogonal polynomials 991

(α,β) (α,β) (α,β)

Hence we have Z n (t) = Pn (t) and P̌n (x) = i n Pn (−i x) so that
(3.17)
  n   
(α,β) 2n + α + β −1 X n + α n + β
P̌n (x) = (x − i )n−k (x + i )k
n k=0
k n − k

(α,β)
(n ≥ 0), where P̌n (x; d, e) = P̌n (x; α + β + 2, i (α − β)) = P̌n (x). Note
(α,β) (α,β)
that even though the expression for P̌n (x) in (3.17) involves i , P̌n (x)
is a real polynomial of degree n since β = ᾱ.
(α,β)
PROPOSITION 3.6. The twisted Jacobi PS { P̌n (x)}∞ n=0 is an OPS (but
not a positive-definite OPS) if and only if −(α + β + 1) ∈
/ N.
Proof. We have, from (2.20), (2.21), (2.22), and (3.16),
(3.18)
−4n(α + β + n)(α + n)(β + n)
βn = (n ≥ 1).
(α + β + 2n − 1)(α + β + 2n)2 (α + β + 2n + 1)

Since β = ᾱ, βn 6= 0 for n ≥ 1 if and only if α + β + n 6= 0 for n ≥ 2 and

βn < 0 for n large enough. Hence, the proposition follows from Theorem 2.9.

The twisted Jacobi PS first appeared in the paper [33] of Romanovski as a

PS satisfying the differential equation

(x 2 + a 2 )y 00 (x) + [2(1 − m)x − va]y 0 (x) − n(n + 1 − 2m)y(x) = 0,

where a, m, v > 0. He provided identities for the twisted Jacobi PS includ-

ing the three term recurrence relation, the differentiation formula, and the
orthogonality (with an incorrect weight function ; see section four).
As discussed in the introduction, Bochner [3] classified the so-called
Sturm-Liouville polynomial systems that can arise as eigenfunctions of the
differential equation (1.3). His analysis allowed a complex linear change of
variable in his classification. Consequently, he identified the Hermite PS with
the twisted Hermite PS, and the Jacobi PS with the twisted Jacobi PS.
Later, Cryer [5] found the twisted Jacobi PS as the Jacobi PS with com-
plex parameters in his characterization of the classical OPS’s through the
Rodrigues’ type formula.
Lastly in this section, we discuss briefly the problem of finding moments
of the classical OPS’s.
992 Kil H. Kwon and Lance L. Littlejohn

For each classical OPS, we can compute the moments {σn }∞ n=0 of its canoni-
cal moment functional σ by solving the corresponding moment equation (2.7)
successively starting from any non-zero value for σ0 . However, the moment
equation is, in general, a three term recurrence relation, which is not easy to
solve. Morton and Krall [32] introduced an idea by which we can always
reduce a three term recurrence relation to a two term recurrence relation. Let
t = x − x0 , where x0 is a constant, possibly complex, that will be chosen
later. Then, in terms of the new variable t, the differential equation (1.3) and
the corresponding moment equation (2.7) become
[`22 t 2 + (2`22 x0 + `21 )t + `22 x02 + `21 x0 + `20 ]y 00 (t)
(3.19)
+ [`11 t + (`11 x0 + `10 )]y 0 (t) = λn y(t),
and
(`11 + n`22 )σn+1 (x0 ) + [`11 x0 + `10 + n(2`22 x0 + `21 )]σn (x0 )
(3.20)
+ n(`22 x02 + `21 x0 + `20 )σn−1 (x0 ) = 0 (n ≥ 0),
where σn (x0 ) = hσ, (x − x0 )n i is the nth moment of σ about x0 . If we choose
x0 so that `22 x02 + `21 x0 + `20 = 0, then the equation (3.20) becomes a two
term recurrence relation and we have
(3.21)
Xn  
n n−k
σn = hσ, x i = hσ, [(x − x0 ) + x0 ] i =
n n
x σk (x0 ) (n ≥ 0).
k=0
k 0
We illustrate the above procedure for the twisted Jacobi polynomials; see
Morton and Krall [32] for a similar discussion of the moments for the other
classical OPS’s, except the twisted Hermite PS. The moment equation for the
twisted Hermite PS is a two-term recurrence relation, which can be solved
easily.
Let σ̌ = σ̌ (α,β) be the canonical moment functional of the twisted Jacobi PS
(α,β)
{ P̌n (x)}∞ n=0 with σ̌0 = hσ̌ , 1i = 1. The corresponding moment equation
is
(3.22) (α + β + n + 2)σ̌n+1 + i (α − β)σ̌n + n σ̌n−1 = 0 (n ≥ 0),
which is a three-term recurrence relation unless α = β. If we choose x0 to be
i and let {σ̌n (i )}∞ ∞
n=0 be the moments of σ̌ about i , then {σ̌n (i )}n=0 satisfies a
two-term recurrence relation
(α + β + n + 2)σ̌n+1 (i ) + 2i (α + n + 1)σ̌n (i ) = 0 (n ≥ 0),
 Classical orthogonal polynomials 993

from which it follows that

(−1)n (2i )n (α + 1)n
σ̌n (i ) = (n ≥ 0),
(α + β + 2)n
where (α)0 = 1 and (α)k = α(α + 1) · · · (α + k − 1) for any complex number
α and integer k ≥ 1. We now obtain, from (3.21),


X n n
j
(−2) j (α + 1)j
(3.23) σ̌n = i n (n ≥ 0).
j =0
(α + β + 2) j

Similarly, if we use σ̌n (−i ) instead of σ̌n (i ), we then obtain

n

X n
j
(−2) j (β + 1)j
(3.24) σ̌n = (−i ) n
(n ≥ 0).
j =0
(α + β + 2) j

Note that all σ̌n are real since the complex conjugate of σ̌n (recall β = ᾱ) in
(3.23) is exactly σ̌n in (3.24).

4. Integral representation of orthogonality

Although using moment functionals to introduce orthogonality has many
advantages as we have seen in previous sections, it is still desirable to express
the orthogonality as an integral with respect to a suitable measure. Such an
integral representation of orthogonality is always possible due to the following
classical results on the moment problem .
Given any sequence {σn }∞ n=0 of real numbers,
(i) (Boas [2]) there is a function µ: R → R of bounded variation on R
such that
Z
(4.1) σn = x n dµ(x) (n ≥ 0);
R

(ii) (Duran [7]) there is a C ∞ -function φ: R → R in the Schwartz space

S such that
Z
(4.2) σn = x n φ(x) d x (n ≥ 0).
R
994 Kil H. Kwon and Lance L. Littlejohn

Hence for any moment functional σ , there is a distribution wσ (x) on R (for

example, we may take wσ (x) to be φ(x) in (4.2)) such that
(4.3) hσ, π i = hwσ , π i (π ∈ P ),
where hwσ , π i is the action of the distribution wσ (x) on the test function
π(x). In particular, if σ is an orthogonalizing moment functional of an OPS
{Pn (x)}∞ ∞
n=0 , we call wσ (x) in (4.3) an orthogonalizing weight for {Pn (x)}n=0 .
Recently, there have been several attempts of effectively finding orthogo-
nalizing weights for various classes of OPS’s. Morton and Krall [32] intro-
duced a formal δ-series expansion of a moment functional σ :
∞
X
σ ≈ (−1)n σn δ (n) (x)/n!
n=0

and found, via the Fourier transform, orthogonalizing weights for the Jacobi,
Laguerre, and Hermite PS’s. This formal δ-series expansion was also used in
Kim and Kwon [14] to produce an orthogonalizing hyperfunctional weight
for the Bessel PS {Bn(2,2)(x)}∞
n=0 .
In case of a classical OPS {Pn (x)}∞ n=0 satisfying the differential equation
(1.3), we may use the corresponding weight equation (2.11) to find an or-
thogonalizing weight for {Pn (x)}∞ n=0 . To do this, however, we must interpret
(2.11) as a classical differential equation with the right-hand side of (2.11)
replaced by a function (not necessarily identically zero) having zero moments.
To be precise we have the following Theorem, which is a special case
of Theorem 2.3 in [22] for second-order differential equations (see also [28,
Theorem 5.6]).
THEOREM 4.1. Let {Pn (x)}∞ n=0 be a classical OPS satisfying the differ-
ential equation (1.3). If w(x) is an orthogonalizing weight distribution for
{Pn (x)}∞
n=0 , then w(x) satisfies the distributional differential equation

(4.4) (`2 (x)w(x))0 − `1 (x)w(x) = g(x),

where g(x) is a distribution having zero moments; that is,
(4.5) hg(x), x n i = 0 (n ≥ 0).
Conversely, if w(x) is a distribution such that
(i) w(x) decays rapidly at infinity so that < w, x n > exists and is finite
for all n ≥ 0 ;
(ii) w(x) is a solution to equation (4.4) on R distributionally ;
 Classical orthogonal polynomials 995

and
(iii) w(x) is non-trivial as a moment functional,
then w(x) is an orthogonalizing weight distribution for {Pn (x)}∞
n=0 .

Condition (iii) in the above Theorem 4.1 means that hw, x n i 6= 0 for some
n ≥ 0. For any classical OPS, there always exists a distributional orthogo-
nalizing weight w(x) satisfying the conditions (i), (ii), (iii) in Theorem 4.1.
In fact, it is enough to take w(x) to be φ(x) in (4.2) where {σn }∞ n=0 are the
moments of any canonical moment functional σ of the given classical OPS.
We call the equation (4.4) the non-homogeneous weight equation for the
differential equation (1.3). When g(x) ≡ 0, the homogeneous weight equa-
tion

(4.6) (`2 (x)w(x))0 − `1 (x)w(x) = 0

is exactly the symmetry equation (2.12) of (1.3) (see Remark 2.1).

Although it turns out that it is enough to solve classically the homogeneous
weight equation (4.6) for an orthogonalizing weight for any positive-definite
classical OPS (as shown by Lesky [27]), we must, in general, consider the
non-homogeneous weight equation (4.4) in the space of distributions; see,
for example, Kwon, Kim, and Han [22] for the case of the Bessel PS and
Littlejohn [28] and Krall and Littlejohn [16] for other classical OPS’s as well
as non-classical OPS’s satisfying higher order differential equations.
There are several examples of non-trivial continuous functions having zero
moments available. For example, the function g(x) given by

0 if x ≤ 0
(4.7) g(x) = 1 1
exp(−x 4 ) sin(x 4 ) if x > 0
is continuous on R and has zero moments. This function was found by
Stieltjes [35]. For more such examples, we refer to Hardy [12] and Maroni
[30].
Once an orthogonalizing weight w(x) (or any orthogonalizing moment
functional σ ) of an OPS {Pn (x)}∞
n=0 is chosen, the squared norms hw, Pn i
2

can be computed most easily from the three-term recurrence relation (2.18).
In fact, we have (see [4, Theorem 4.2 in Chap. 1])
Y
n
(4.8) hw(x), Pn2 i = βj ,
j =0
996 Kil H. Kwon and Lance L. Littlejohn

where β0 = hw, P02 i = hw, 1i and βn (n ≥ 1) are the constants in (2.18).

We shall now construct an orthogonalizing weight for each classical OPS
found in section three. We always assume that the parameters involved in
each PS are restricted so that the PS is an OPS.

Case 1: Jacobi polynomials

In this case, the homogeneous weight equation corresponding to the Jacobi
differential equation (3.1) is

(4.9) (1 − x 2 )w0 (x) + [(α + β)x − (β − α)]w(x) = 0,

which is equivalent to

(4.10) (1 − x 2 )[(1 − x)−α (1 + x)−β w(x)]0 = 0

for x 6= ±1. Then the general distributional solution of (4.10) for x 6= ±1 is

w(x) = [c1 H (1 − x) + c2 H (1 + x) + c3 ](1 − x)α (1 + x)β ,

where ci (i = 1, 2, 3) is an arbitrary constant and H (x) is the Heaviside

function. If we choose c1 = −1, c2 = +1, and c3 = 0, then this w(x)
extends to a distribution on R (see Remark 4.1 below) :

β
(4.11) w(α,β) (x) = (1 − x)α+ (1 + x)+ ,

which is a non-trivial distributional solution to (4.9) on R with compact

support [−1, 1]. Since w(α,β) (x) satisfies the conditions (i), (ii), (iii) in
(α,β)
Theorem 4.1, w(α,β) (x) is an orthogonalizing weight for {Pn (x)}∞ n=0 . We
then have, from (3.3) and (4.8),
(4.12)
hw(α,β) (x), [Pn(α,β) (x)]2 i
22n+α+β+1 0(n + α + 1)0(n + β + 1)0(n + α + β + 1)n!
=
0(2n + α + β + 1)0(2n + α + β + 2)

2α+β+1 0(α+1)0(β+1)
for n ≥ 0, since hw(α,β) (x), 1i = 0(α+β+2)
.
 Classical orthogonal polynomials 997

REMARK 4.1. For any complex number a, consider the function f a : R →

C defined by 
0 if x ≤ 0
f a (x) = ,
x a if x > 0
where we take log x to be real for x > 0 so that x a is defined uniquely
for x > 0. The function f a (x) always extends to a distribution x+a on R
with support in [0, ∞). For Re a > −1, f a (x) is locally integrable on R so
that x+a = f a (x) and for Re a ≤ −1, x+a is obtained from f a (x) by analytic
continuation and regularization. For details on the distribution x+a , we refer to
Hörmander [13, Chap. 3.3.2]; see also Morton and Krall [32] for an explicit
integral representation of the distribution w(α,β) (x) in (4.11).

Case 2: Bessel polynomials

βx
In this case, it is more convenient to replace x by 2
and α by α + 2 so
that the equation (3.4) becomes

(4.13) L[y](x) = x 2 y 00 (x) + [(α + 2)x + 2]y 0 (x) = n(n + α + 1)y(x),

/ N. We then denote Bn(α+2,2) (x) by Bn(α) (x). Now, the

where −(α + 1) ∈
homogeneous weight equation corresponding to (4.13) is

(4.14) x 2 w0 (x) − (αx + 2)w(x) = 0,

of which the only one linearly independent distributional solution with support
in [0, ∞) is

0 if x ≤ 0
(4.15) w0 (x) =
x α exp(−2/x) if x > 0.

Romanovski [33] used w0 (x) as an orthogonalizing weight for Bessel PS, but
w0 (x) cannot be an orthogonalizing weight since it does not decay rapidly at
infinity. In fact, we have

lim x n w0 (x) = ∞
x→∞

for n + α > 0. We now consider the non-homogeneous weight equation

(4.16) x 2 w0 (x) − (αx + 2)w(x) = g(x),

998 Kil H. Kwon and Lance L. Littlejohn

where g(x) is a function with zero moments. For x 6= 0, the general solution
of (4.16) is

c1 (−x)α e−2/x if x < 0
w(x) = Rx
x α e−2/x 0 e2/t t −2−α g(t) dt + c2 x α e−2/x if x > 0,

where c1 and c2 are arbitrary constants. With concern for the

R ∞ boundary
condition (i) in Theorem 4.1, we choose c1 = 0 and c2 = − 0 e2/t t −2−α
g(t) dt to obtain

(α) 0 if x ≤ 0
(4.17) w (x) = α −2/x
R∞ 2/t −2−α
−x e x e t g(t) dt if x > 0.

If we further take g(x) to be the function given in (4.7), then w(α) in (4.17) is a
continuous function on R satisfying the conditions (i) and (ii) in Theorem 4.1
(see [9], [22], and [30]). Hence, w(α) (x) in (4.17) (with g(x) in (4.7)) is an
orthogonalizing weight for Bessel PS {Bn(α)(x)}∞ n=0 if and only if

Z ∞ Z ∞ 
(α) α −2/x 2/t −2−α
(4.18) hw (x), 1i = − x e e t g(t) dt d x 6= 0.
0 x

Condition (4.18) was first proved in [22] for α = 0 and, recently, Maroni [30]
proved (4.18) for all α ≥ 12( π2 )4 − 2.
If we let σ (α) be the canonical moment functional of {Bn(α)(x)}∞ n=0 with
(α)
σ0 = 1, then we have from (3.6) and (4.8)

(α) (−4)n n!0(α + 2)0(α + n + 1)

(4.19) hσ , [Bn(α)(x)]2 i = (n ≥ 0).
0(α + 2n + 1)0(α + 2n + 2)

REMARK 4.2. Krall and Frink [19] found the complex orthogonality (now
called the Bessel orthogonality) of the Bessel PS through the contour integral
along the unit circle in the complex plane. Although the homogeneous weight
equation (4.14) cannot yield a distributional orthogonalizing weight for the
Bessel PS, it has a non-trivial hyperfunctional solution with support at {0}
with respect to which the Bessel PS is orthogonal (see [9], [14], and [15]).
Later in this section, we will discuss again real orthogonalizing weights
for {Bn(α)(x)}∞
n=0 for any α with −(α + 1) ∈/ N.
 Classical orthogonal polynomials 999

Case 3: Laguerre polynomials

In this case, the homogeneous weight equation corresponding to the La-
guerre differential equation (3.7) is

(4.20) xw0 (x) + (x − α)w(x) = 0.

If we set v(x) = e x w(x), then v(x) satisfies the Euler equation

xv 0 (x) − αv(x) = 0,

of which the general distributional solution is

v(x) = c1 x+α + c2 x−α ,

where c1 and c2 are arbitrary constants and x−α is the distribution on R with
support in (−∞, 0] (defined similarly as x+α ; see Remark 4.1 and Hörmander
[13, Chap. 3.3.2]). Hence, the general solution of (4.20) is

w(x) = c1 x+α e−x + c2 x−α e−x .

For this w(x) to vanish at infinity, c2 must be zero. Then by taking c1 = 1,

we obtain

(4.21) w(α) (x) = x+α e−x .

Since w(α) (x) in (4.21) satisfies the conditions (i), (ii), (iii) in Theorem 4.1,
∞
w(α) (x) is an orthogonalizing weight for {L (α)
n (x)}n=0 . Since

hx+α e−x , 1i = 0(α + 1)

we have, from (3.9) and (4.8),

(4.22) hw(α) (x), [L (α)

n (x)] i = n! 0(n + α + 1).
2

Case 4: Hermite polynomials

In this case, the homogeneous weight equation corresponding to the Her-
mite differential equation (3.10) is

(4.23) w0 (x) + 2xw(x) = 0,

1000 Kil H. Kwon and Lance L. Littlejohn

of which the only one linearly independent distributional solution is

(4.24) w(x) = exp(−x 2 ).

Since w(x) in (4.24) satisfies the conditions (i), (ii), (iii) in Theorem 4.1,
w(x) is an orthogonalizing weight for {Hn (x)}∞ n=0 . We then have, from (3.12)
and (4.8),
(4.25) Z ∞
√
hw(x), Hn (x)i =
2
Hn2 (x) exp(−x 2 ) d x = πn! 2−n (n ≥ 0).
−∞

Case 5: Twisted Hermite polynomials

In this case, the homogeneous weight equation corresponding to the twisted
Hermite differential equation (3.13) is

(4.26) w0 (x) − 2xw(x) = 0,

of which the only one linearly independent distributional solution is

w0 (x) = exp(x 2 ),

which cannot be an orthogonalizing weight. However, from (3.14) and (4.25),

we can obtain the complex orthogonality :
(4.27)
Z i∞
√
Ȟm (x) Ȟn (x) exp(x 2 ) d x = (−1)n π n! 2−n i δmn (m and n ≥ 0).
−i∞

Let us now consider the non-homogeneous weight equation

(4.28) w0 (x) − 2xw(x) = g(x),

where g(x) is a non-trivial continuous function on R with zero moments and

support in [0, ∞). Then the general solution of (4.28) is
Z x
x2 x2
e−t g(t) dt,
2
w(x) = ce + e
0
 Classical orthogonal polynomials 1001

where c is an arbitrary constant. For this w(x) to vanish at infinity, c must be

zero and
Z ∞
e−x g(x) d x = 0.
2
(4.29)
0

Then we have

0 if x ≤ 0
(4.30) w(x) = x2
Rx −t 2
e 0 e g(t) dt if x > 0.
Note that w(x) in (4.30) is a classical solution to (4.26) on R. If we further
assume

(4.31) lim x n g(x) = 0 (n ≥ 0),

x→∞

then it is easy to see that

lim x n w(x) = 0 (n ≥ 0),

x→∞

and so w(x) satisfies the conditions (i), (ii) in Theorem 4.1. Consequently,
w(x) in (4.30) is a real orthogonalizing weight for { Ȟn (x)}∞
n=0 if and only if
Z ∞ Z x 
x2 −t 2
(4.32) hw(x), 1i = e e g(t) dt d x 6= 0.
0 0

The existence of a weight w(x) for the twisted Hermite PS, of the form
given in (4.30) and satisfying (4.29), is discussed below in Remark 4.3.
∞
√ let σ be the orthogonalizing moment functional for { Ȟn (x)}n=0 with
If we
σ0 = π , then we have, from (3.15) and (4.8),
√
(4.33) hσ, [ Ȟn (x)]2 i = (−1)n π n!2−n (n ≥ 0).

REMARK 4.3. We can easily see that there is a non-trivial function g(x)
with zero moments, which also satisfies the condition (4.29). Choose any
two linearly independent continuous functions g1 (x) and g2 (x) with zero
moments and support in [0, ∞). Set
Z ∞
e−x gi (x) d x
2
Ai = (i = 1, 2).
0
1002 Kil H. Kwon and Lance L. Littlejohn

If Ai 6= 0 (i = 1, 2), then

g(x) = A2 g1 (x) − A1 g2 (x)

satisfies the condition (4.29) and has zero moments.

Case 6: Twisted Jacobi polynomials

In this case, the homogeneous weight equation corresponding to the twisted
Jacobi differential equation (3.16) is

(4.34) (1 + x 2 )w0 (x) + [(d − 2)x + e]w(x) = 0,

of which the only linearly independent distributional solution is

2−d
f (x) = (1 + x 2 ) 2 exp(−e arctan x).

Romanovski [33] used f (x) as an orthogonalizing weight for the twisted

Jacobi PS, but f (x) cannot be an orthogonalizing weight since it does not
decay rapidly at infinity. In fact, we have

lim x n f (x) = ∞
x→∞

for n + 2 − d > 0. However, from (3.17) and (4.12), we can obtain the
complex orthogonality
(4.35)
β
h(1 − x)α+ (1 + x)+ , P̌m(α,β) (i x) P̌n(α,β) (i x)i
(−1)n 22n+α+β+1 0(n + α + 1)0(n + β + 1)0(n + α + β + 1)n!
= δmn
0(2n + α + β + 1)0(2n + α + β + 2)
(m and n ≥ 0), where i e = β − α and d = α + β + 2.
Let us now consider the non-homogeneous weight equation

(4.36) (1 + x 2 )w0 (x) + [(d − 2)x + e]w(x) = g(x),

where g(x) is a non-trivial continuous function on R with zero moments and

support in [0, ∞). Then the general solution of (4.36) is
Z x
f (x)
w(x) = e [c + e− f (t) (1 + t 2 )−1 g(t) dt],
0
 Classical orthogonal polynomials 1003

where c is an arbitrary constant. For this w(x) to vanish at infinity, c must be

zero and
Z ∞
(4.37) e− f (t) (1 + t 2 )−1 g(t) dt = 0.
0

Then we have

0 if x ≤ 0
(4.38) w̌(x) = f (x)
Rx − f (t) 2 −1
e 0 e (1 + t ) g(t) dt if x > 0.
Note that w̌(x) in (4.38) is a classical solution to (4.36). If g(x) satisfies
the condition (4.31), then w̌(x) satisfies the conditions (i) and (ii) in Theo-
rem 4.1. Consequently, w̌(x) in (4.38) is a real orthogonalizing weight for
(α,β)
{ P̌n (x)}∞n=0 if and only if
Z ∞ Z x 
f (x) − f (t) 2 −1
(4.39) hw̌(x), 1i = e e (1 + t ) g(t) dt d x 6= 0.
0 0

(α,β)
If we let σ̌ be the orthogonalizing moment functional of { P̌n (x)}∞
n=0
with
2α+β+1 0(α + 1)0(β + 1)
σ̌0 = ,
0(α + β + 2)
then we have, from (3.18) and (4.8),
(4.40)
hσ̌ ,[ P̌n(α,β) (x)]2 i
(−1)n 22n+α+β+1 0(n + α + 1)0(n + β + 1)0(n + α + β + 1)n!
= .
0(2n + α + β + 1)0(2n + α + β + 2)
(n ≥ 0)

REMARK 4.4. In the formula (4.12), the parameters α and β are real
numbers with −(α + β + 1), −α, and −β ∈ / N . However, by analytic
continuation, the same formula holds for complex parameters α and β as long
as − Re(α + β + 1), − Re α, and − Re β ∈ / −N. This fact is used in (4.35),
where β = ᾱ, Re α = Re β = 2 , and −(d − 1) = −(α + β + 1) ∈
d−2
/ N.
Constructing explicit real orthogonalizing weights for classical OPS’s by
solving the non-homogeneous weight equation (4.4) has been successful ex-
cept, at the moment, for the Bessel PS {Bn(α) (x)}∞
n=0 when 0 6= α < 12( π ) −2,
2 4
1004 Kil H. Kwon and Lance L. Littlejohn

(α,β)
the twisted Hermite PS { Ȟn (x)}∞
n=0 , and the twisted Jacobi PS { P̌n (x)}∞
n=0 .
For any OPS (classical or not), its real orthogonalizing weight can be explic-
itly constructed by the following remarkable result on the general moment
problem.
THEOREM 4.2 (Duran [8]). For any sequence of real or complex numbers
{σn }∞n=0 , define a function w(x) by

0 if x ≤ 0
(4.41) w(x) = 1 R ∞ P∞ √
2 0
( n=0 σn cn t n h(λn t))J0 ( xt) dt if x > 0,
(−1)n P
where cn = 22n+1 (n!)2
, λn = n + nk=0 cn , J0 (x) is the Bessel function of the
first kind, and h(x) is a C ∞ -function on R with compact support satisfying
h(0) = 1 and h (n) (0) = 0 (n ≥ 1). Then, w(x) is a function in the Schwartz
space S and satisfies
Z ∞ Z ∞
x w(x) d x =
n
x n w(x) d x = σn (n ≥ 0).
−∞ 0
In particular, if we take {σn }∞ in Theorem 4.2 to be the moments of a
n=0
canonical moment functional of any classical OPS, then the function w(x)
in (4.41) is a real orthogonalizing weight for the OPS. Moreover, by Theo-
rem 4.1, the function
g(x) = (`2 (x)w(x))0 − `1 (x)w(x)
is a function, in the Schwartz space S with zero moments and support in
[0, ∞), satisfying the condition (4.31). In the case of the twisted Hermite or
the twisted Jacobi polynomials, this g(x) also satisfies (4.29) and (4.32) or
(4.37) and (4.39) respectively.
REMARK 4.5. In the case of the Bessel, the twisted Hermite, and the twisted
Jacobi polynomials, their complex orthogonality seems more natural than
their real orthogonality. In fact, through the hyperfunctional representations
of orthogonalizing weights, we can see that any OPS has both real and
complex orthogonality : see, for example, [14].
ACKNOWLEDGEMENTS. The first author (KHK) thanks Professor L. Duane
Loveland of the Department of Mathematics and Statistics at Utah State
University for the opportunity to visit Utah State during the 1992-93 academic
year. In addition, he thanks the Korea Science and Engineering Foundation
(95-0701-02-01-3), GARC, and Korea Ministry of Education(BSRI 1420) for
their research support.
 Classical orthogonal polynomials 1005

References
[1] J. Aczél, Eine Bemerkung über die Characterisierung der klassischen orthogonale Poly-
nome, Acta Math. Sci. Hung. 4 (1953), 315–321.
[2] R. P. Boas, The Stieltjes moment problem for functions of bounded variation, Bull. Amer.
Math. Soc. 45 (1939), 399–404.
[3] S. Bochner, Über Sturm-Liouvillesche Polynomsysteme, Math. Z. 29 (1929), 730–736.
[4] T. S. Chihara, An introduction to orthogonal polynomials, Gordon and Breach, New York,
1977.
[5] C. W. Cryer, Rodrigues’ formulas and the classical orthogonal polynomials, Boll. Unione
Mat. Ital. 25 (1970), 1–11.
[6] Á. Császár, Sur les polynômes orthogonaux classiques, Annales Univ. Sci. Budapest Sec.
Math. 1 (1958), 33–39.
[7] A. J. Duran, The Stieltjes moment problem for rapidly decreasing functions, Proc. Amer.
Math. Soc. 107(3) (1989), 731–741.
[8] , Functions with given moments and weight functions for orthogonal polynomials,
Rocky Mountain J. Math. 23(1) (1993), 87–104.
[9] W. D. Evans, W. N. Everitt, K. H. Kwon, and L. L. Littlejohn, Real orthogonalizing weights
for Bessel polynomials, J. Comp. Appl. Math. 49 (1993), 51-57.
[10] J. Favard, Sur les polynomes de Tchebicheff, C.R. Acad. Sci. Paris 200 (1935), 2052–2053.
[11] J. Feldmann, On a characterization of classical orthogonal polynomials, Acta. Sc. Math.
17 (1956), 129–133.
[12] G. H. Hardy, On Stieltjes’ “Probleme des Moments", Messenger of Math. 4 (1917), 175–182.
[13] L. Hörmander, The analysis of linear partial differential operators I, Springer-Verlag, New
York, 1983.
[14] S. S. Kim and K. H. Kwon, Hyperfunctional weights for orthogonal polynomials, Results
in Math. 18 (1990), 273–281.
[15] , Generalized weights for orthogonal polynomials, Diff. and Int. Equations 4 (1991),
601–608.
[16] A. M. Krall and L. L. Littlejohn, On the classification of differential equations having
orthogonal polynomial solutions II, Ann. Mat. Pura Appl. 4 (1987), 77–102.
[17] H. L. Krall, Certain differential equations for Tchebychev polynomials, Duke Math. J. 4
(1938), 705–719.
[18] , On derivatives of orthogonal polynomials II, Bull. Amer. Math. Soc. 47 (1941),
261–264.
[19] H. L. Krall and O. Frink, A new class of orthogonal polynomials : The Bessel polynomials,
Trans. Amer. Math. Soc. 65 (1949), 100–115.
[20] H. L. Krall and I. M. Sheffer, A characterization of orthogonal polynomials, J. Math. Anal.
Appl. 8 (1964), 232–244.
[21] , Differential equations of infinite order for orthogonal polynomials, Ann. Mat. Pura
Appl. 4 (1966), 135–172.
[22] K. H. Kwon, S. S. Kim, and S. S. Han, Orthogonalizing weights for Tchebychev sets of
polynomials, Bull. London Math. Soc. 24 (1992), 361–367.
[23] K. H. Kwon, J. K. Lee, and B. H. Yoo, Characterizations of classical orthogonal polynomi-
als, Results in Math. 24 (1993), 119-128.
 1006 Kil H. Kwon and Lance L. Littlejohn

[24] K. H. Kwon and L. L. Littlejohn, Sobolev orthogonal polynomials and second-order differ-
ential equations, Rocky Mt. J. Math. (to appear).
[25] K. H. Kwon, L. L. Littlejohn, and B. H. Yoo, Characterizations of orthogonal polynomials
satisfying differential equations,R∞ SIAM J. Math. Anal. 25(3) (1994), 976–990.
[26] E. N. Laguerre, Sur l’integrale x x −1 e−x d x, Bull. de la Socíeté Math. de France 7 (1879),
72–81.
[27] P. Lesky, Die Charakterisierung der klassischen orthogonalen Polynome durch Sturm-
Liouvillesche Differentialgleichungen, Arch. Rat. Mech. Anal. 10 (1962), 341–352.
[28] L. L. Littlejohn, On the classification of differential equations having orthogonal polynomial
solutions, Ann. Mat. Pura Appl. 4 (1984), 35–53.
[29] L. L. Littlejohn and D. Race, Symmetric and symmetrisable ordinary differential expressions,
Proc. London Math. Soc. 1 (1990), 334–356.
[30] P. Maroni, An integral representation for the Bessel form, J. Comp. Appl. Math. 57 (1995),
251–260.
[31] M. Mikolás, Common characterization of the Jacobi, Laguerre and Hermite-like polynomi-
als (in Hungarian), Mate. Lapok 7 (1956), 238–248.
[32] R. D. Morton and A. M. Krall, Distributional weight functions for orthogonal polynomials,
SIAM J. Math. Anal. 9(4) (1978), 604–626.
[33] M. V. Romanovski, Sur quelques classes nouvelle de polynomes orthogonaux, C.R. Acad.
Sci. Paris 188 (1929), 1023–1025.
[34] N. J. Sonine, Recherches sur les fonctions cylindriques et le développment des fonctions
continues en series, Math. Ann. 16 (1880), 1–80.
[35] T. J. Stieltjes, Recherches sur les fractions continues, Ann. de la Faculté des Sci. de Toulouse
8 (1894), J1–122; 9 (1895), A1–47; Oeuvres 2, 398–566.

K. H. Kwon
Department of Mathematics
KAIST
Taejon 305-701, Korea
E-mail: khkwon@ jacobi.kaist.ac.kr

L. L. Littlejohn
Department of Mathematics and Statistics
Utah State University
Logan, Utah, 84322-3900
E-mail: lance@sunfs.math.usu.edu

View publication stats