C•

ELSEVIER Journal of Computational and Applied Mathematics 81 (1997) 165-179

JOURNAL OF
COMPUTATIONAL AND
APPLIED MATHEMATICS

Hermite-Sobolev and closely connected orthogonal polynomials

Andr6 Draux*, C h a r a f E1 Hami
L M I - URA CNRS 1378, INSA de Rouen, DOpartement de G~nie Mathdmatique, Place Emile Blondel, BP 08
F-76131 Mont-Saint-Aignan Cedex, France
Received 19 July 1996; received in revised form 15 February 1997

Abstract

In this paper a bilinear functional a

N

a(f, g) = ~ 2mc(m)(f (m),g(m))

m=0

which corresponds to the Hermite-Sobolev and closely connected orthogonal polynomials is studied. The domain of
N
{2m}m=l in which a is positive definite is given and the Hermite-Sobolev and closely connected formal orthogonal
polynomials are given in the nonnormal case.

Keywords: Formal orthogonal polynomials; Hermite polynomials; Herrnite-Sobolev polynomials; Definite inner product

A M S classification: 33 C 45; 42 C 05

I. Introduction

The very comprehensive bibliography [7] (more than 100 references), made by F. Marcellfin and
A. Ronveaux, shows the interest taken by numerous authors in the study of the orthogonal polynomi-
als on Sobolev spaces since more than 30 years. In the diagonal case, whether the measure is discrete
or continuous, whether the order of derivatives is one or more, these studies have always considered
only positive coefficients in the definition of the inner product. Even in their paper devoted to old
and new directions about the study of orthogonal polynomials on Sobolev spaces, Marcellfin et al.
[4] do not contemplate a study devoted to the following problem: does there exist negative values of
coefficients for which we still have a positive definite inner product? In the significant study given by
Pfrez [8] in her thesis, nothing concerns this problem; no more in the most recent papers of Marcellfin
et al. [5, 6] devoted, respectively, to an inner product with N derivatives and to Laguerre-Sobolev

* Corresponding author. E-mail: [Link]@[Link].

0377-0427/97/$17.00 (~) 1997 Elsevier Science B.V. All rights reserved

PH S 0 3 7 7 - 0 4 2 7 ( 9 7 ) 0 0 0 3 8 - 1
1'66 A. Draux, C El HamilJournal of Computational and Applied Mathematics 81 (1997) 165-179

orthogonal polynomials. In the nondiagonal Sobolev inner products, since a positive semidefinite or
definite symmetric matrix is used, some negative coefficients can be encountered - see in [1, 3] a
discrete measure and in [9] a continuous measure - but, of course, it is not the general case in
which some symmetric matrices, having some negative diagonal coefficients, could certainly be used
to define still some nondiagonal Sobolev inner products. Like in [5] we study a bilinear functional
with N derivatives which corresponds to the diagonal case, and we carefully define the set of values
of coefficients for which the bilinear functional still is a positive definite inner product. This set
corresponds to a convex domain whose the boundary is exactly given. When the boundary of the
domain of the space of coefficients is known, it seems natural to study the nonnormal case and w~
shall give the particular relations concerning the formal orthogonal polynomials.
The bilinear functional corresponds to Hermite-Sobolev orthogonal polynomials, but the same
theory can still be applied to closely connected orthogonal polynomials such that Laguerre-Sobolev
and Jacobi-Sobolev orthogonal polynomials with some measures which change with the order of the
derivatives. In a next paper we shall develop the case purely Laguerre-Sobolev and Jacobi-Sobolev
which uses some techniques completely different from the ones used here.
Let us begin to consider a bilinear functional a
a : ~ x ~---~ ~,

where ~ is the vector space of real polynomials in one real variable x E f2, open set of ~. a is
defined as follows:
N
a ( f , g ) = Z2mc(m)(f(m),o(m) ) V f and g E ~ , (1)
m=0

where c <m), m = 0,... ,N, are symmetric bilinear functionals

c(") : ~ × ~ --+ ~ for m = 0 .... ,N,

and f(m) and 9 ('n) are the derivatives of order m of f and 9 ( f ( 0 ) = f , 9(0)= 9). 2,,, m = 0,... ,N,
are N + 1 fixed real scalars with 2o = 1 and 2N ¢ 0.
Every functional c <') is defined from its moments

c(m)(xi,xJ ) = { C!m! if i >/0 and j ~> 0,

(2)
0'+j if i < 0 or j < 0 .
We look f o r the formal orthogonal polynomials with respect to a, that is to say, we look for the
polynomials Sn, n/> 0, such that

deg Sn = n,
a(S,,x i ) = O fori=0,...,n-1. (3)
If S, is written as S , ( x ) = ~j=0
~-'~n ~j
,v(n),rj
~ ' we obtain from (3) a linear system of n equations with n + 1
unknowns a~n) for j = 0,..., n.

a(S"'xi)= Z 2mc(m)((xJ)(m)'(xi)(ra))O~5"n)' i = 0 . . . . . n - 1. (4)

j=0 m=0
 A. Draux, C. El HamilJournal of Computational and Applied Mathematics 81 (1997) 165-179 167

z (m) ~,n-- 1
Let M(nm) be the Hankel matrices (ci+j)i=j=o for m = 0 . . . . . N and Vn E ~.
Let K, be the n x n matrix such that (Kn)i+l,i = i, i = 1 , . . . , N - 1 and the other (K~)i,j are equal
to 0.
Setting Mn = (a(xJ,xi))nZjlo, it is clear that this matrix can be written as
N
Mn = Z "] [(m "~n
"~m~'n l~'[(m)(':~n
k"m'~T
: • (5)
m=0

Since degSn = n, ~(,") is different from zero. Then the linear system (4) can be written as

M-nXn = -a(~")An, (6)

where Xn is a vector o f ~" o f components ai'(n), i----0,. .., n - I, and An is a vector of ~" o f components
a(x n,x i), i ----0 , . . . , n - 1.
From (6) we immediately have the following result o f existence and uniqueness o f the solution S,.

Theorem 1. The formal orthogonal polynomial Sn ex&ts and is unique when the leading coefficient
is fixed, if and only if the matrix Mn is regular.

Definition 2. The bilinear functional a is called quasi-definite if the matrix Mn is regular Vn ~> 1.

It seems very difficult to study the property of quasi-definition o f a or the regularity o f a finite
sequence o f matrices M,, 1 ~< n ~< k. It is the reason for which we shall give some equivalent
properties in the quasi-definite case as well as in the positive definite case in the next section, in
order to study the following problem: c ('n), m = 0 , . . . ,N, being a definite inner product for a certain
space, to give the domain o f ~N in which 2 o f components 21 . . . . . 2N can be taken in order that a
is a definite inner product.

2. Orthogonal polynomials in Sobolev spaces

Let L 2 ( Q ; ]2m), m ----0 , . . . ,N, be the inner product space of square integrable real functions on the
open set f2 C ~ for the Borel measure ~,~, positive on f2.
On this space we have the classical definite inner product:

c(m)(f,g) = (f,g)(m) = fo f(x)g(x)dt'tm(X) V f , g EL 2 ((2; #m) (7)

and the norm c(m)(f , f ) = ( f , f)(m) = Ilfll~m) = fa f 2 ( x ) d#m(X).

HN((2; #0, # 1 , . . . , #N) will denote the Sobolev space such that
HN(f2; #0,#,,...,kiN) = { f l f (m) E L2(f2; #m), m = 0, 1 , . . . , N } ,
where the derivatives of order m are taken in the distribution sense. The definite inner product on
this space is
N
s ( f , g ) = y~c(m)(f("),g ~'n)) V f and gEHN(y2;#O, pl,...,#N).
m=0
168 A. Draux, C El Hami / Journal of Computational and Applied Mathematics 81 (1997) 165-179

The corresponding norm, denoted by 11. II~, is given by

: IIf(m)llm) '7'f C HN((2; #0, #1,..., #N).

m=O

Now we consider the symmetric bilinear functional a

a :HN(Q',#O,#1,... , # N ) X HN(O;#O,#I,... , # N ) ---+ ~

defined as in (1) V f and g E HN(Q; #0, #1,..., #N).

The functional a is continuous o n HN(Q; #0, #1,..., #N) X HN((2; #0, #1,..., #N). Indeed we have

la(f,g)l max
m
12m[ Ilflls I1011 , Vf, gEHN(Q;#O,#1,'",#N) • (8)
N
To see that it suffies to apply the triangular inequality on [~m=0[, Cauchy-Schwarz's inequality on
each definite inner product c (m), m = 0,... ,N, and finally Cauchy-Schwarz's inequality in •N+I for
the euclidian inner product.
Obviously, if 2i >~ 0 Vi= 1,..., N, a also is a definite inner product. Moreover, if 2i > 0 Vi= 1 , . . . , N ,
a is coercive, that is to say, there exists a constant p > 0 such that

a(f,f) >~ pllfll 2, V f EHN(f2;pO, P l , . . . , p N ) - {0}. (9)

Here p = minm(2m). Of course, in this last case, the norm ( a ( f , f ) ) 1/2 is equivalent to the norm
Ilflls.
Again we look for the formal orthogonal polynomials with respect to a, as in first section. We
shall only consider the case where c (°) corresponds to Hermite, Laguerre and Jacobi functionals.
If d is a classical positive definite linear functional, then d can be characterized by the existence
of two polynomials q) and ~, with deg ¢ ~< 2 and deg ~9= 1, such that D(~pd) = ipd, where D is the
derivative operator (we have D(dpd) : (b'd + dpDd).
Up to a linear change of variable, the only positive definite classical linear functionals are the
Hermite, Laguerre and Jacobi functionals.
Let {Pn}, the monic orthogonal polynomials sequence associated to the linear functional d (also
denoted by d(°)), then the derivatives of order m of the polynomials {P,}, are orthogonal with
respect to the (classical) linear functional d (m) = dpmd(°). In fact, if we denote by {p(m)}, the monic
orthogonal polynomial sequence associated with d (m), w e have

Dmp.(x ) = (n )mp(m) (x ), (lO)

where (n)m denotes the Pochhammer's symbol (n)m = n(n - 1 ) . . . (n - m + 1).
Moreover, if we denote by k(,m) = d(m)((P,(m))2), we obtain the usual expression
k(0)
ompn(x) : r~(m) 'n p(m) (x'] (11)

We give all the elements corresponding to the Hermite, Laguerre and Jacobi functionals in
Table 1.
 A. Draux, C. El Hamil Journal of Computational and Applied Mathematics 81 (1997) 165-179 169

Table 1

Name Hermite Laguerre Jacobi

1 x 1 -- X 2

-2x ~+ 1 -x ( f l - ~ ) - (~ + fl + 2)x
~J ]0, + ~ [ 1 - 1,1[
/lo e -x2 x% -x (1 -- x)~(1 + x) ~

k(~) v~.! F(n + ~ + m + 1)n! 2~(m)n!ro(~m))r(~(~))r(~(~)-")

2n 8(nm)(F(e(m)))2

withd~ m ) = n + ~ + m + l
/~") = n + fl + m + 1
~(.,) = d(m) + if m) _ 1
¢(:) 1 1 (n+ot+fl+m),n
Restrictions ~>--1 ~>--1, fl>--I

Of course, we have c ( ' ) ( f , 9 ) = c(m)(1,f9) in L2(~;/lm), m = 0, 1,...,N. Our study is devoted to

the case where
c(m)(f,9) = d(m)(fg), m = O, 1. . . . . N. (12)
From the orthogonality of the derivatives we deduce trivially the orthogonality with respect to the
bilinear functional a, and
N N
a(Pn,Pn) = Z 2mc(m)(Dmpn,Ompn)= ~ J'm(n)2mk~m-)m"
m=0 m=0

We can summarize all this in the following property.

Property 3.
(i) I f n # i, a(P~,P~) = O. (13)

(ii) a(P.,P,) : k(n°) (N

1 + ZC(nm)~m
m=l
) , (14)

where
C~m) = (n )2mk~_)m/k~°) = (n )m¢~m) (15)
with the convention that i f n - m + 1 <<.0 f o r an index m, then C(nm) = O, or in other words this
index does not exist in the sum.

Therefore, if the matrix Mn is regular, Pn is the (unique) monic orthogonal polynomial with respect
to the bilinear functional a (Sn = Pn if S, is monic)

Definition 4. The bilinear functional a is called positive definite o n HN(K2;/./0,Pl .... ,/[AN) × H N
(f2; ~ , #1,..., #N) (resp. on ~ × ~ ) if, Vf E HN(f2; I~o,I~1,..., #N) -- {0} (resp. ~ -- {0}), a ( f , f ) > O.
170 A. Draux, C. El Hami/Journal of Computationaland Applied Mathematics 81 (1997) 165-179

Now we want to know in which domain of ~ N /~ of components 21 ..... '~,N, must be located
in order that a is quasi-definite or positive definite on HN(O; #0, #1,..., # N ) x HN(~; ltO,l~l. . . . , #U).
First we start the study on ~.

Theorem 5. (i) The bilinear functional a is quasi-definite if and only if a(Pi,Pi) ¢ O, Vi E ~,

that is to say if and only if the coefficients 2m, m = 1,...,N, satisfy the following
conditions:
N
1-bZf~m)~m~O Vi>~I. (16)
m=l

(ii) The bilinear functional a is positive definite on ~ × ~ if and only if a(Pi,P,-)>0, ViE ~,
that is to say if and only if the coefficients 2m, m = 1,..., N, satisfy the following conditions:
N
1+ ~ c~m)l~m>O Vi~>1. (17)
m=l

Proof. (i) We have to prove that {mi}i>~l is a sequence of regular matrices if and only if a(P,-,P/) ¢ 0,
Vi E [~. Remark that a(Po, Po)=k~°) # 0 and ~t] :(c~°))=(k~°)). First let us assume that M, is regular,
and prove that M,+I regular is equivalent to a(P,,P,) ¢ O. Therefore, from (14) we get that condition
(16) is satisfied for i = n, since k,(°)>0. Indeed, let us write P, as P,(x)= ~ = 0 7~")xt with 7(,") = 1,
and let us denote by )(, the vector of R N of components Yf"(~), Y = 0,.. . , n - 1. ~t,+l is multiplied
on the right by the following (n + 1) × (n + 1) regular matrix

,, en),
1

where/~ is the n x n identity matrix and On the zero vector of ~". Using relations a(P,,x i) = 0 for
i < n deduced from (13) we get

Ar
On)
a(Pn,xn) '

where An is the vector of ~" of components a(xJ,xn), j = 0,. :_:~n- 1. Since a(P,,x")=a(P,,P,), the
result a(Pn,P,)¢ 0 holds as equivalence to the regularity of Mn+l.
As a consequence of that, -¢M. l i J~,+1
i = l is a sequence of regular matrices if and only if a(P,-, P~) ¢ 0 for
i----0,... ,n, that is to say if and only if (16) is satisfied for i--0,... ,n. In particular a is quasi-definite
if and only if (16) holds Vi>~ 1.
(ii) Thanks to (14) we can see that it is necessary to have (17) in order that a(P,,P,)>O.
Any non-zero polynomial of degree n can be written as

p(x) = ~ fl,Pi(x) with fin # 0.

i=O
 A. Draux, C. El Hami / Journal of Computational and Applied Mathematics 81 (1997) 165-179 171

Using relation (14), we get

a(p, p ) = fl~k~°) 1 + ~ c)m) l~m .

i=0 m=l /

Therefore, if relations (17) are satisfied, then a(p,p)>O, VpC ~ - {0}. []

If relations (17) are satisfied, then ~ is an inner product space for the definite inner product a.
There exists a sequence of orthogonal polynomials that we shall call Hermite (resp. Laguerre or
Jacobi) Sobolev orthogonal polynomials.
Now, we recall that a sequence {w~}~>0 is closed in a separable inner product space ~ ( O )
which has definite inner product denoted by (.,.) if and only if, for f E 9f'(~2), there exists a
sequence {'Wn).>~O (~J. = ~___~j=oflj,nWj) such that l i m , - . ~ ( ~ n - f , ~ , - f ) = 0. This is equivalent to
(f, wi) = 0, Vi~>0, for f c oW(f2) if and only if f = 0 (see [2]). Then Vf E A~(f2), f can be written
as f = ~_,i>~ofWi.
On the other hand, L2(f2; #m) is a separable inner product space, as well as Hw(Y2; P0, #1,..., #U).
Moreover {P,}i~>0 is closed in L2(O; ]Am)- We generalize this property to H N ( o ; Igo, # l , . . . , fiN) with
the positive definite inner product a.

Theorem 6. If relations (16) are satisfied, then

(i) the bilinear functional a is positive definite on HN(O; #0,121. . . . , / A N ) X I-IN ( ~'~;~0, #1,..., ]AN).
(ii) the sequence {P,}n~>0 is closed in H S ( f 2 ; # 0 , p t , . . . , # s ) with the inner product a.

Proof. Since {Dmpn}n>~m is closed in L2(O;/~m), m = 0, 1,...,N, then Vf E n N ( o , flo, t/1,...,~N) and
for a fixed integer r, l<~r<<.N, f(O can be written as f(r)=~>.r~Drp,, and ~>.rf2C(~)(DrP~,D~P~)
<+oc.
In ~'(f2), space of distributions, we have a primitive of f(r) given by
y(r--1) = Z f~D r--1 P~,
i>>-r

where the primitives are taken in the distribution sense, f ( r - 1 ) C L 2 ( ~ 2 ; #r--l). Indeed

Z *'~2r(r-1)CF)r-lp. /Tr--lp.'l 2 ~(r--1) 2 (k°) 2

Ji ~ ~,~ "t,~ "tl = Z£ ( i ) /,.(r--I)
i>~r i>~r 'Vi--r+l

~ ZJi \ i J k~r)r < "-~(X),

i>~r

since ~Ir-1)/~l ") ~<1 and k~_~/k[r-~+~<<.1Vi>~r.

f(r-1) which also is a primitive of f(r), belongs to L : ( ~ ; # r - 1 ) and can be written as ~J>~r-1 f
Dr-I Pi.
Since f ( r - l ) _ f(r--1) is a constant which can be written in a unique way as gr_lDr-lP~_l, we get
f=-)~, Vi>~r and gr--1 = f - - 1 .
172 A. Draux. C. El Hamil Journal of Computationaland Applied Mathematics 81 (1997) 165-179

Therefore f(m) = ~j>>.O )~o(m)ej in L2(I2;/Am), m = 1,... ,N.

We can conclude that {P,},~o is closed in HN(f2; #o, #1,..., #N) with the positive definite inner
product s. Indeed, for fEHN(f2;#o,#I .... ,#N) there exists a sequence {Pn}n>_-o (fin = ~'=oJ}Pj)
such that l i m n ~ II f(,n)]In ----0, m : 0, 1 , . . . , N . Therefore l i m n ~ II ?. - f II~ = ZT,=o
II Dmfi, - f(m) IIm ----0,
Since a is continuous over H N ( f 2 ; # o , # I , . . . , # N ) × HN(I2; #o, #I .... ,#N), l i m , ~ a(Pn- f , P n -
f ) : 0. Moreover,

l i r n [a(Pn,Pn)- a(f, f ) l = lijno~ [a(Pn - f , Pn " f ) + 2a(Pn - f , f ) [ : o

A A

thanks to relation (8). Thus a(f,f)>>.O as limit of a(P,,Pn)>~O.

Using Qn:~i"=o fP,, we have limn__.~ a(Qn, Q n ) : a ( f , f ) and also l i m n ~ a(Qn, Qn)=limn--.~ ~i"---o
f/2a(P/,Pi) = ~i~>0 f/2a(P/,P/) •
Thus a(f, f ) = ~i~>0 f/2a(P/, P,) > 0, V f E HN((2; #0, #1,.-., #N) -- {0}.
Therefore the results (i) and (ii) hold. []

3. Domain of positive definition and nonnormal case

Let 2 be an element of ~N of components 2m, m : I , . . . , N .

Let Jf~, i i> 1, be the hyperplanes of R iv

~i = 2 E ~ N l+~-~c~m)2m=O
m=l

and 9i, i~> 1, the open half-spaces of ~N

m=l

Define 9 by 9 = Ni>119i.
9 is non-empty, since V2 such that 2i > 0, i = 1,..., N, 2 E 9 . Moreover, V2 E 9 , 2N > 0. Indeed
the limit open half-space 9 ~ corresponds to

lim 1 + Z 2m + 2N •
i---~ cx~
m=l

Thanks to relations (15) it is clear that l i m ~ l / C ~ N ) = lim~_~c~m)/c~U)= 0 for m = 1 , . . . , U - 1.

Since 2N is assumed to be different from zero, then 2u >0.
From Theorem 5 the following result is obvious.

Corollary 7. (i) The bilinear functional a is positive definite if and only if 2 6 9.

(ii) The bilinear functional a is quasi-definite if and only if 2 c Ru\Ui>>.l~ii.
 A. Draux, C. El Hami/Journal of Computational and Applied Mathematics 81 (1997) 165-179 173

Obviously, ~ is a convex domain, since, if 2 and 2 E 9 , then p2 + (1 - p ) 2 E N with 0~<p ~<1.

Indeed if we denote by a (resp. ~) the bilinear functional (1) associated with 2 (resp. 2), then

V f EHN(f2;#o,I~I, .... #N)-- {0}, ~(f,f)>0a(f'f)>0 }¢=~

VfEHN(f2;I~o,I~,...,#N) - {0} and Vpc[0,1],

p a ( f , f ) + (1 - p)'d(f , f ) > O .

Definition 8.
The case where 2 E Ui~>l~" will be called the non-normal case•

Now, let us prove that (N + 1) different hyperplanes o~i have no common point•

Property 9. L e t L be a set of N + 1 distinct inteoers f l , . . . , ( u + l (1 ~E1 < f 2 < "'" <#N+I).

Then AicL ~ i = dp.

ProoL If the matrix A(fl,..., ~N+I) is regular

l1 C dl
°) C (N)
(l

1 C (1) C (N)
d2 d2
A(~I,...,{N+I)~--- (18)
• . •

1 C dN+~
(1) C (N)
~N+I

then the linear system A(f~,...,fN+l)(lx) = 0 has no solution.

If the determinant of this matrix is developed with respect to its last row, we obtain a polynomial
function in one variable YN+~. Its degree is equal to N and its roots are f~,... ,(N in the Hermite
and Laguerre cases; its degree is equal to 2N and its roots are El,..., EN,--Yl --6 .... , - - f N - 5, where
6 = e + fl + 1, in the Jacobi case.
If we denote by A((l,..., YN) the submatrix o f A ( f l , . . . , fN+l) which corresponds to its N first rows
and columns, we get the following:

Hermite and Laguerre cases:

N
detA((1,..., ~U+l) = detA((l,..., fN) II(fu+l -- ~)
j=l

= II (t,-6).
l<~j<i<.N+l
174 A. Draux, C. El Hami l Journal of Computational and Applied Mathematics 81 (1997) 165-179

Jacobi case:
N
d e t A ( E 1 , . . . , flN+l ) = d e t A ( f l . . . . , fiN ) II(gaN+l -- 6 )(ffN+l '~ 4 "~- 6)
j=l

= ]-I
l <~j < i < ~ N + l

Since the integers d,- are distinct, the result holds.

Remark that the previous proof also is valid for integers ~- such that 1 ~<li ~ N . Indeed, in this
case, the corresponding row of A(g~,.. ,~N+I) is: 1 C (1) C (+') 0 . ,0 (the row is ended by a
group of N - ~ zero elements). Expanding the determinant of A(:~,..., :N+I) with respect to its last
row, when #N+I is replaced by ~, the N - f/ last terms are well equal to zero since they have a
factor EN+~-- (,. []

Note that with the choice (1 ~f~ < g ° 2 < "'" < ~ N + I ) , detA(~l,...,fu+l)>0.
Corollary 10. 2 E ~N can only belon9 to N hyperplanes ~ , at most.

Now we can give a property about the part played by the hyperplanes gg,. or by the open half-
spaces ~;, i >/1, which are all necessary for defining ~ .

Property 11. l f N >>.2, each open half-space ~i contributes to restrict the domain in which a &
positive definite, that is to say ~ is inside AJj~] ~ j strictly: 32 E (]~] ~ j and 2 f~ ~.

Proof. The domain ~ can also be defined by the following linear system (2;) of inequalities:

(2;) A2>A0, (19)

where A is a (cx~ × N ) matrix; the jth row contains the elements C~1),..., C~N). Ao is a vector in
infinite dimension of components all equal to - 1 .
We consider the linear system (Si), deduced from (S), such that the ith inequality of (2;) is
replaced by
N
- )--~ c~'n)2,, ~> 1,
m=l
where i is a fixed integer, i >/. 1. To prove that each open half-space ~ i contributes to restrict the
domain in which a is positive definite comes to prove that the system (2;i) has a solution 2 such
that 2N>0.

In fact it suffies to look for solutions 2 with components 2~,0,...,0,2;,0 . . . . . 0,2:: if i < N , or
21,0,...,0,2N if i>~N.
Let Ej(2) be the equation
/~(1)] c(r)]
Ej(2) = 1 + ,..j ~1 -'~ "-'j -°r, j >i 1, (20)
 A. Draux, C. El Hamil Journal of Computationaland Applied Mathematics 81 (1997) 165-179 175

with r = i if i < N and r = N if i>~N. I f j < r , then CF

~ j ) = 0 . Let g~, k = 1,2,3, be the three distinct
integers with ~k/> 1. Consider the following three equations:

Ee~ = 0 , k = 1,2,3. (21)

Thanks to a similar proof as in Property 9, it is possible to show that the matrix B(f~, g2, ~3) of the
system (21) is regular.

1 C~I1) c(r)el)
B(~I, ~2, d3) = 1 C(e~) C ~2
(r) " (22)
1 C~ ) C (~)~

If i>~2, we consider the system (21) with E1 = i - 1, E2=i, E3= i + 1, and the three points F~,F2 and
F3 obtained as solution of two equations taken among the three previous ones (F1 with E~ and Ei+l,
F2 with Ei-i and E~+~, and F3 with Ei_l and E~). Let (¢k)
k"~l , ' ;(k)5
~ r I be the coordinates of Fk, k = 1,2,3.
In any case 1-3, Eq. (20) shows us that 2]k)<0 and 2~k)>0, k = 1,2,3.
Let us prove that the three points F~,F2 and F3 are the vertices of a triangle (i.e., they are not
aligned). This property is satisfied if and only if the following matrix is regular:

1 1 1]
A=

But it is clear that

Gl 0 0 "~
B ( i - 1,i,i + 1) A = 0 G2 0 ) .
0 0 G3

Since B ( i - l , i , i + l ) is regular, Gk # 0, Vk=l,2,3, and A is regular. Since the Eqs. (21) correspond
to straight lines in the plane (21, 2r), it is obvious to see that Ei(F2)<0. Therefore the points inside
the triangle FI F2 F3 are in ~i-x and ~+1, and in NN\~..
Now let us prove that Ej(Fk)= Gj,k>0, Vk = 1,2,3, and Vj~>I such that j ~ { i - 1,i,i + 1}.
Remark that, in the particular case where i<N, if Ej(Fk)>0 and 2N>0, then Ej(F~)+ Cj(.N)2N>'0. '
Thus, if Ej(Fk)>0, since a triangle is convex, the closed triangle F1 F2 F3 is contained within all
the open half-spaces ~j., j/> 1, j ~ {i - 1, i, i + 1}. Therefore there exists, inside this triangle, a point
(21,2r) such that the corresponding point (21,0,...,0,2~,0 ..... 0,2N) with 2N arbitrary and 2N > 0
(resp. (21,0,...,0,2N)) if i < N (resp. if i>~N), is in all the open half-spaces Nj-, j>~l, j ¢ i, and
in NN\~/, and the property holds.
Consider the Eqs. (21) with gl and ~2 (fl <~2) belonging to the set { i - 1,i,i + 1} and with
~3 = j q~{i - 1, i, i + 1}. To Ee, and Et2 corresponds the point Fk (k E { 1,2, 3}). Then, for this point,
176 A. Draux, C. E1 HamiIJournal of Computational and Applied Mathematics 81 (1997) 165-179

the corresponding system (21) can be written as

B(E1, vP2,j) 2] k) = 00 . (23)

This linear system can be transformed in an equivalent system having a upper triangular matrix
thanks to Gauss algorithm:

(i x
)1) (0)
2~k) 0 .
0 det B(t~, f2,j) Gj, k

Since 2~k) > 0, Gj, k and detB(Yl, E2,j) have the same sign. If j < ~1, the last row of B(fl, f2,j) can be
placed in the first row without change of sign of detB(fl,f2,j). In any case, a similar proof as this
one of Property 9 can be used for obtaining the same remark which follows this property: det B > 0.
Thus Gj, k > 0 and the result holds.
If i = 1, using the expression of C~1~ and C~2) in the cases 1-3, and taking r = 2 in the defini-
tion (20) of Ei(2 ), it is easy to see that { E j ( 2 ) = 0}j~>2 is a sequence of straight lines in (21,22)
whose the negative slope increases strictly with j and the coordinate 22 of the intersection point
of E1(2) = 0 decreases strictly with j. Thus the open cone, in the plane (21,22), of vertex F1,
defined by E 1 ( 2 ) < 0 and E z ( 2 ) > 0 , is strictly inside every open half-spaces E f t 2 ) > 0 , Vj~>3, in
A A

the plane (21,22). Let (21,22) be an arbitrary point of this cone. Using an arbitrary 2N, 2N>0,
we get that (21,22,0,...,0,~N) is in @, Vj>~2, and in ~ U \ ~ l , which concludes the
proof. []
This property is not true in the case where N = 1 as it can be seen in Section 4.
Remark that the proof of Property 11 shows us that ~ contains a part of ~U corresponding to some
2~s<0, 1 <~i<~N- 1. To illustrate this fact thoroughly, let us show in Property 12, that a non-empty
domain of ~ can be included in the region of ~U defined by {2E RN I 2i-..<0 for i = 1 , . . . , N - 1
and 2N > 0}.
We shall use ~N-1 as normed space for the norm I1.111 defined by
N--1
ilu111 = E 1nil
i=l
where u is a vector of ~N--1 of components ui, i = 1 , . . . , N - 1.

Property 12. Let 2 be the vector of ~N-1 of components 21,...,AN_l, and B the open ball in
RN-1 of center 0 and of radius min(2N, 1/c'(N-1)~-/-~N-I
j with respect to the norm I1-111. Then, ifN>~2,
A

M = {2E~Nl2i<<.O for i = l , . . . , N - - 1 and 2N>O, 2 E B } C ~ .

Proof. Let us prove that if 2 C ~ which is obviously non-empty, then 2 E ~, that is to say that
relations (17) are satisfied.
 A. Draux, C. El Hami/Journal of Computational and Applied Mathematics 81 (1997) 165-179 177

If 1 <~i<~N- 1, remark that

1 + ~"=1 C} m)2" > 0 S E i : l cSm)lA"l < 1
2i<<.0, i = 1 , . . . , N - 1 2i-~<0, i : 1 , . . . , N - 1.
Now
i i N--I
c}m)lA" I ~ C(NU~a)~-~ 12"[ ~ C(NNC1) ~ lAm[< 1
m=l m=l m=l
if 2 E ~. Thus 2 E 9.
If i/> N, since "-'it"~(m)~<" ~ic'(N) and 2,, ~<0, m = 1, ... , N - 1, we have 1 + EN=I C}") 2 m ~ ] JI-c}N)~Nm=1 2rn"
~. N N
If 2 E , ~ " = 1 2m > 0, and therefore 1 + ~ " = l c}m)2m> 1 and 2 E ~. []

Therefore ~ is not only, like studied until now, the positive domain of ~N.
Note that, with respect to the norm II. I1,, while 2N<I/C~U_-~ ~), ~ is a cone whose the vertex is 0
and the axis is this one of AN. When 2N/> 1ICON_-1~), we have a cylinder whose the axis is this one
of AN, the generating line is parallel to this axis and the section is B.
Let J be a set of p distinct integers j l , . . . , j p (p<~N) such that 1 ~<jl < j 2 < ' " "<jp.
Assume that 2 E ~N, 2 E 9~j, Vj E J.
We look for the formal Sobolev orthogonal polynomials Sn, like in (3), in the nonnormal case.

Theorem 13. I f 2 E ~U is such that 2 E ~ j Vj E J = { j l , . . . , j p } with p < . N and 1 <<.j~<j2 < " " <
jp, then
(i) I f n <<.jl, Sn=~(nn)Pn, where ~") is a non-zero constant determined by the wanted normalization
Of Sn.
(ii) Let Jr be a subset o f J :Jr ---- { j l , . . . , . ~ ) , r<.p. F o r r = 1 , . . . , p - 1, and n such that
j~ < n <<.j~ + 1, we have
an : a(n)pn "~- Z ~(n)pj,
jEJ~
where "8~n), j E Jr, are r arbitrary constants, and "8(nn) a non-zero constant.
(iii) I f n > ~, then S, =~,(")P, + [Link]
~ c~n)~,
~j 1-j., where =(n)
O;j , j E J, are p arbitrary constants, and "8(")
a non-zero constant.

Proof. We can still write Sn as

n
~(n)n
Sn = ~-'~ (zj f j.
j=0
Thus thanks to a(Sn,Pi)= 0, i = 0 , . . . , n - 1 which is equivalent to (3), we get

a(Sn,Pi) = ~ "~(n'a(Pj, Pi) = ~n'a(Pi,Pi).

j=0

Therefore if i E J , then ~") is an arbitrary constant since a(P/,P~) = O, and if i ~ J , a(P/,P,) ~ 0

implies that 8 ~ ) = O. []
178 A. Draux, C. E1 Hami/Journal of Computational and Applied Mathematics 81 (1997) 165-179

Remark that Vn E ~ we always have

a(S,,S,) = (~"))2a(P,,P~)
and therefore a(Sn, Sn) -- 0 if n E J and a(S,, S,) ¢ 0 if n ~ J.

4. Some examples

Now let us present two examples (N = 1 and N = 2) which well illustrate our purpose.
Case N = 1:
~i = {21E n I 1 + C~[)21 > 0 } Vi>~I.

Since 21 ¢ 0, the only positive definite inner products (1) correspond to 21 > 0. It always is the case
studied until now.
The set

A---{ 2~i) E R' 2~0 - C;1) Vi>~ 1} =u°~/,>~1

is the one of values for which we are in the nonnormal case.

Case N =- 2:
~,-={2ER211 -~-C}1)21--~-C}2)22>0}Vi>~I.
is a polyhedric domain for which each Ni is a face and we give a picture of N in the case 2
(Laguerre-Sobolev orthogonal polynomials).
We have drawn the Jct~/s for i = 1 to 7 and the limit straight line (i = c~).

\
½J
Domain ~D

0 ,~a
-1
...-°/-'.o,-°.. tff.
\
 A. Draux, C. El Hami l Journal of Computational and Applied Mathematics 81 (1997) 165-179 179

We see that ~ contains a part of ~2 corresponding to 2~ < O.

References

[1] M. Alfaro, F. Marcellfin, M.L. Rezola, A. Ronveaux, Sobolev-type orthogonal polynomials: the nondiagonal case,
J. Approx. Theory 83 (1995) 266-287.
[2] R. Dautray, J.-L. Lions, Analyse math6matique et calcul num~rique pour les sciences et les techniques, vol. 3:
Transformations, Sobolev, op6rateurs, Masson, Paris, 1987; vol. 4: M6thodes variationnelles, Masson, Paris, 1988;
vol. 5: Spectre des op6rateurs, Masson, Paris, 1988.
[3] K.H. Kwon, L.L. Littlejohn, The orthogonality of the Laguerre polynomials {L~-k)(x)} for positive integers k, Ann.
Numer. Anal. 2 (1995) 1-4, 289-303.
[4] F. Marcellfin, M. Alfaro, M.L. Rezola, Orthogonal polynomials on Sobolev spaces: old and new directions, J. Comput.
Appl. Math. 48 (1993) 113-131.
[5] F. Marcell~in, T.E. P6rez, M.A. Pifiar, A. Ronveaux, General Sobolev orthogonal polynomials, J. Math. Anal. Appl.
200 (1996) 614-634.
[6] F. Marcellfin, T.E. P6rez, M.A. Pifiar, Laguerre-Sobolev orthogonal polynomials, J. Comput. Appl. Math. 71 (1996)
245 -265.
[7] F. Marcellfin, A. Ronveaux, Orthogonal polynomials and Sobolev inner products: a bibliography, Internal report to
the Universities of Madrid and Namur, November 1996.
[8] T.E. P6rez, Polinomios orthogonales respecto a productos de Sobolev: el caso continuo, Doctoral dissertation,
University of Granada, 1994.
[9] T.E. P6rez, M.A. Pifiar, On Sobolev orthogonality for the generalized Laguerre polynomials, J. Approx. Theory 86
(1996) 278-285.
[10] G. Szeg6, Orthogonal polynomials, A.M.S. Colloquium publications, vol. XXIII, Providence, RI, 1939.