a
r
X
i
v
:
m
a
t
h
.
C
A
/
0
5
1
2
4
2
4
v
1
1
8
D
e
c
2
0
0
5
Orthogonal Polynomials
Vilmos Totik
11/11/05
Abstract
In this survey, diﬀerent aspects of the theory of orthogonal polynomials
of one (real or complex) variable are reviewed. Orthogonal polynomials
on the unit circle are not discussed.
Contents
1 Introduction 71
2 Orthogonal polynomials 73
Orthogonal polynomials with respect to measures . . . . . . . . . 73
The Riemann–Hilbert approach . . . . . . . . . . . . . . . . . . . 74
Orthogonal polynomials with respect to inner products . . . . . . 76
Varying weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
Matrix orthogonal polynomials . . . . . . . . . . . . . . . . . . . 77
3 The L
2
extremal problem 77
4 Orthogonal polynomials on the real line 79
5 Classical orthogonal polynomials 80
6 Where do orthogonal polynomials come from? 82
Continued fractions . . . . . . . . . . . . . . . . . . . . . . . . . . 82
Pad´e approximation and rational interpolation . . . . . . . . . . 83
Moment problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
Jacobi matrices and spectral theory of selfadjoint operators . . . 85
Quadrature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
Random matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
Surveys in Approximation Theory
Volume 1, 2005. pp. 70–125.
Copyright c 2005 Surveys in Approximation Theory.
ISSN 1555578X
All rights of reproduction in any form reserved.
70
Orthogonal Polynomials 71
7 Some questions leading to classical orthogonal polynomials 88
Electrostatics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
Polynomial solutions of eigenvalue problems . . . . . . . . . . . . 89
Harmonic analysis on spheres and balls . . . . . . . . . . . . . . 89
Approximation theory . . . . . . . . . . . . . . . . . . . . . . . . 90
8 Heuristics 91
9 General orthogonal polynomials 93
Lower and upper bounds . . . . . . . . . . . . . . . . . . . . . . . 94
Zeros . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Regularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
10 Strong, ratio and weak asymptotics 98
Strong asymptotics . . . . . . . . . . . . . . . . . . . . . . . . . . 98
Ratio asymptotics . . . . . . . . . . . . . . . . . . . . . . . . . . 99
Weak and relative asymptotics . . . . . . . . . . . . . . . . . . . 100
Widom’s theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
Asymptotics for Christoﬀel functions . . . . . . . . . . . . . . . . 102
11 Recurrence coeﬃcients and spectral measures 102
12 Exponential and Freud weights 104
13 Sobolev orthogonality 107
14 NonHermitian orthogonality 109
15 Multiple orthogonality 110
Types and normality . . . . . . . . . . . . . . . . . . . . . . . . . 111
Recurrence formulae . . . . . . . . . . . . . . . . . . . . . . . . . 112
The Riemann–Hilbert problem . . . . . . . . . . . . . . . . . . . 113
16 Matrix orthogonal polynomials 114
Matrix orthogonal polynomials . . . . . . . . . . . . . . . . . . . 114
Threeterm recurrence and quadrature . . . . . . . . . . . . . . . 115
Families of orthogonal polynomials . . . . . . . . . . . . . . . . . 116
Connection with higher order scalar recurrence . . . . . . . . . . 117
1 Introduction
The theory of orthogonal polynomials can be divided into two main but only
loosely related parts. The two parts have many things in common, and the
division line is quite blurred, it is more or less along algebra vs. analysis. One
of the parts is the algebraic aspect of the theory, which has close connections
V. Totik 72
with special functions, combinatorics and algebra, and it is mainly devoted
to concrete orthogonal systems or hierarchies of systems such as the Jacobi,
Hahn, AskeyWilson, . . . polynomials. All the discrete polynomials and the q
analogues of classical ones belong to this theory. We will not treat this part;
the interested reader can consult the three recent excellent monographs [39] by
M. E. H. Ismail, [28] by W. Gautschi and [6] by G. E. Andrews, R. Askey and
R. Roy. Much of the present state of the theory of orthogonal polynomials of
several variables lies also close to this algebraic part of the theory. To discuss
them would take us too far from our main direction; rather we refer the reader
to the recent book [24] by C. F. Dunkl and Y. Xu.
The other part is the analytical aspect of the theory. Its methods are ana
lytical, and it deals with questions that are typical in analysis, or questions that
have emerged in and related to other parts of mathematical analysis. General
properties ﬁll a smaller part of the analytic theory, and the greater part falls
into two main and extremely rich branches: orthogonal polynomials on the real
line and on the circle. The richness is due to some special features of the real
line and the circle. Classical real orthogonal polynomials, sometimes in other
forms like continued fractions, can be traced back to the 18th century, but their
rapid development occurred in the 19th and early 20th century. Orthogonal
polynomials on the unit circle are much younger, and their existence is largely
due to Szeg˝ o and Geronimus in the ﬁrst half of the 20th century. B. Simon’s
recent treatise [80, 81] summarizes and greatly extends what has happened since
then.
The connection of orthogonal polynomials with other branches of mathe
matics is truly impressive. Without even trying to be complete, we mention
continued fractions, operator theory (Jacobi operators), moment problems, an
alytic functions (Bieberbach’s conjecture), interpolation, Pad´e approximation,
quadrature, approximation theory, numerical analysis, electrostatics, statistical
quantum mechanics, special functions, number theory (irrationality and tran
scendence), graph theory (matching numbers), combinatorics, random matrices,
stochastic processes (birth and death processes; prediction theory), data sorting
and compression, Radon transform and computer tomography.
This work is a survey on orthogonal polynomials that do not lie on the unit
circle. Orthogonal polynomials on the unit circle—both the classical theory and
recent contributions—will be hopefully dealt with in a companion article.
This work is meant for nonexperts, and it therefore contains introductory
materials. We have tried to list most of the actively researched ﬁelds not directly
connected with orthogonal polynomials on the unit circle, but because of space
limitation we have only one or two pages on areas where dozens of papers
and several books had been published. As a result, our account is necessarily
incomplete. Also, the author’s personal taste and interest is reﬂected in the
survey, and the omission of a particular direction or a set of results reﬂects in
no way on the importance or quality of the omitted works.
For further backgound on orthogonal polynomials, the reader can consult
Orthogonal Polynomials 73
the books Szeg˝ o [91], Simon [80][81], Freud [27], Geronimus [34], Gautschi [28],
Chicara [18], Ismail [39].
This is a largely extended version of the ﬁrst part of the article Golinskii–
Totik [36].
Acknowledgement. Research was supported by NSF grant DMS040650 and
OTKA T049448, TS44782, and was carried out within the Analysis Research
Group of the Hungarian Academy of Sciences. The author participates in the
project INTAS 03516637 that supported a visit by Leonid Golinskii to Szeged,
during which the outline of this paper was laid out.
2 Orthogonal polynomials
Orthogonal polynomials with respect to measures
Let µ be a positive Borel measure on the complex plane, with an inﬁnite number
of points in its support, for which
[z[
m
dµ(z) < ∞
for all m > 0. There are unique polynomials
p
n
(z) = p
n
(µ, z) = κ
n
z
n
+ , κ
n
> 0, n = 0, 1, . . .
that form an orthonormal system in L
2
(µ), i.e.
p
m
p
n
dµ =
0 if m = n
1 if m = n.
These p
n
’s are called the orthonormal polynomials corresponding to µ. κ
n
is
the leading coeﬃcient, and p
n
(z)/κ
n
= z
n
+ is called the monic orthog
onal polynomial. The leading coeﬃcients play a special and important role in
the theory, many properties depend on their behavior. When dµ(x) = w(x)dx
on some interval, say, then we talk about orthogonal polynomials with respect
to the weight function w.
The p
n
’s can be easily generated: all we have to do is to make sure that
p
n
(z)
κ
n
z
k
dµ(z) = 0, k = 0, 1, . . . , n −1, (2.1)
which is an nn system of equations for the nonleading coeﬃcients of p
n
(z)/κ
n
with matrix (σ
i,j
)
n−1
i,j=0
, where
σ
i,j
=
z
i
z
j
dµ(z)
V. Totik 74
are the complex moments of µ. This matrix is nonsingular: if some linear
combination with coeﬃcients c
0
, . . . , c
n−1
of the rows is zero, then the polyno
mial P
n−1
(z) = c
0
+ +c
n−1
z
n−1
is orthogonal to every z
j
, j < n, and hence
it is orthogonal to itself, i.e.,
[P
n−1
[
2
dµ =
P
n−1
P
n−1
dµ = 0,
which implies P
n−1
(z) ≡ 0. Thus, c
0
= = c
n−1
= 0, which shows the
nonsingularity of (σ
i,j
). Therefore, the system (2.1) has a unique solution for
the nonleading coeﬃcients of p
n
(z)/κ
n
(note that the leading coeﬃcient is 1),
and ﬁnally κ
n
comes from normalization.
In particular, the complex moments already determine the polynomials. In
terms of them one can write up explicit determinant formulae:
p
n
(z) =
1
D
n−1
D
n
σ
0,0
σ
0,1
σ
0,n−1
1
σ
1,0
σ
1,1
σ
1,n−1
z
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
σ
n−1,0
σ
n−1,1
σ
n−1,n−1
z
n−1
σ
n,0
σ
n,1
σ
n,n−1
z
n
(2.2)
where
D
n
= [σ
i,j
[
n
i,j=0
(2.3)
are the so called Gram determinants.
Note that if µ is supported on the real line then
σ
i,j
=
x
i+j
dµ(x) =: α
i+j
,
so D
n
= [α
i+j
[
n
i,j=0
is a Hankel determinant, while if µ is supported on the unit
circle then
σ
i,j
=
z
i−j
dµ(z) =: β
i−j
,
so D
n
= [β
i−j
[
n
i,j=0
is a Toeplitz determinant. In these two important cases the
orthogonal polynomials have many special properties that are missing in the
general theory.
The Riemann–Hilbert approach
Let µ be supported on the real line, and suppose that it is of the form dµ(t) =
w(t)dt with some smooth function w. A new approach to generating orthogonal
polynomials that has turned out to be of great importance was given in the early
1990’s by Fokas, Its and Kitaev [26]. Consider 2 2 matrices
Y (z) =
Y
11
(z) Y
12
(z)
Y
21
(z) Y
22
(z)
Orthogonal Polynomials 75
where the Y
ij
are analytic functions on C` R, and solve for such matrices the
following matrixvalued Riemann–Hilbert problem:
1. for all x ∈ R
Y
+
(x) = Y
−
(x)
1 w(x)
0 1
where Y
+
, resp. Y
−
, is the limit of Y (z) as z tends to x from the upper, resp.
lower half plane, and
2.
Y (z) =
I +O
1
z
z
n
0
0 z
−n
at inﬁnity, where I denotes the identity matrix.
Thus, one is looking for 4 functions Y
11
, ..., Y
22
analytic on C`R, such that if
Y
±
ij
(x) denote the boundary limits of these functions at x ∈ R from the upper,
resp. lower half plane, then
Y
−
11
(x) = Y
+
11
(x), Y
−
21
(x) = Y
+
21
(x) (2.4)
and
Y
−
12
(x) = Y
+
11
(x)w(x) +Y
+
12
(x), Y
−
22
(x) = Y
+
21
(x)w(x) +Y
+
22
(x). (2.5)
These connect the functions on the upper and lower half planes only very mildly,
and what puts the problem into a rigid form is the second condition, namely it
is required that for large z uniformly on the plane we have
Y
11
(z) = z
n
+O([z[
n−1
), Y
21
(z) = O([z[
n−1
) (2.6)
and
Y
12
(z) = O(z
−n−1
), Y
22
(z) = z
−n
+O([z[
−n−1
). (2.7)
It can be shown that there is a unique solution Y (z). The relevance of this to
orthogonal polynomials is that the entry Y
11
(z) is precisely the monic polyno
mial p
n
(µ, z)/κ
n
. Indeed, (2.4) shows that Y
11
and Y
12
are analytic everywhere,
and if an entire function is O([z[
m
) as z →∞, then it is a polynomial of degree
at most m. Thus, we get from (2.6) that Y
11
(z) = z
n
+ is a monic polynomial
of degree n, and Y
21
(z) is a polynomial of degree at most n − 1. The relation
(2.7) gives that the integral of z
k
Y
12
(z) over the circle [z[ = R is O(R
k−n
) for
all k < n and hence it tends to 0 as R → ∞. By analyticity, the integral over
the upper part of the circle can be deformed into an integral from R to −R on
the upper part of R, i.e., into
−R
R
x
k
Y
+
12
(x)dx,
and similarly the integral over the lower part of the circle can be deformed into
an integral from −R to R on the lower part of R, i.e., into
R
−R
x
k
Y
−
12
(x)dx.
V. Totik 76
The ﬁrst relation in (2.5) implies
x
k
Y
−
12
(x) −x
k
Y
+
12
(x) = x
k
Y
11
(x)w(x),
therefore for k = 0, 1, . . . , n −1 we have
R
−R
x
k
Y
11
(x)w(x)dx = O(R
k−n
) = O(R
−1
) →0,
which implies
∞
−∞
x
k
Y
11
(x)w(x)dx = 0.
Thus, Y
11
is indeed the monic nth orthogonal polynomial with respect to w.
The other entries can also be explicitly written in terms of the orthogonal
polynomials p
n
and p
n−1
: Y
21
is a constant multiple of p
n−1
,
Y
12
(z) =
1
2iπκ
n
p
n
(x)w(x)
x −z
dx
is the Cauchy transform of p
n
(x)w(x)/κ
n
, and Y
22
is the Cauchy transform of
Y
21
(= const p
n−1
). Furthermore, κ
n
and the recurrence coeﬃcients a
n
, b
n
(see
Section 4) can be expressed in terms of the entries of Y
1
, where Y
1
is the matrix
deﬁned by
Y (z)
z
−n
0
0 z
n
=: I +z
−1
Y
1
+O
1
z
2
.
For details on this Riemann–Hilbert approach, see Deift [20].
Orthogonal polynomials with respect to inner products
Sometimes one talks about orthogonal polynomials with respect to an inner
product ', ` which is deﬁned on some linear space containing all polynomials,
and orthogonality means 'p
n
, p
m
` = 0 for m = n. In this case the aforemen
tioned orthogonalization process can be used, and with σ
i,j
= 'x
i
, x
j
`, the
determinantal formula (2.2) is still valid.
Sometimes one has an ', ` with the standard inner product properties,
except that positive deﬁniteness may not hold (as an example consider non
Hermitian orthogonality from Section 14). Then the orthogonalization process
and the determinantal formulae can still be used provided the Gram determi
nants (2.3) are diﬀerent from zero. If this is not so, then we write
p
n
(z) = γ
n
z
n
+ γ
n−1
z
n−1
+ ,
and make sure that p
n
is orthogonal to all powers z
k
, 0 ≤ k < n, i.e., solve the
homogeneous system of equations
n
¸
j=0
γ
j
σ
j,k
= 0, k = 0, . . . , n −1,
Orthogonal Polynomials 77
for γ
0
, γ
1
, . . . , γ
n
. Since the number of unknowns is bigger than the number of
equations, there is always a nontrivial solution, which gives rise to nontrivial
orthogonal polynomials. However, then we cannot assert any more γ
n
= 0, so
the degree of p
n
may be smaller than n, and there may be several choices for
p
n
. Still, in applications where nonHermitian orthogonality is used, these p
n
play the role of orthogonal polynomials.
Varying weights
In the last 25 years orthogonal polynomials with respect to varying measures
have played a signiﬁcant role in several problems, see e.g. the sections on expo
nential and Freud weights or on random matrices in Section 4. In forming them
one has a sequence of measures µ
n
(generally with some particular behavior),
and for each n one forms the orthogonal system ¦p
k
(µ
n
, z)¦
∞
k=0
. In most cases
one needs the behavior of p
n
(µ
n
, z) or that of p
n±k
(µ
n
, z) with some ﬁxed k.
We mention three examples.
The ﬁrst example is that of Freud weights: W(x) = e
−x
λ
, λ > 0. If one
substitutes x = n
1/λ
y, then with P
n
(y) = p
n
(W, x) orthogonality takes the form
P
n
(y)P
m
(y)e
−ny
λ
dy = 0, n = m,
and it turns out that this is just the right scaling, e.g. the zeros of P
n
have
an asymptotic zero distribution (while those of p
n
(W, z) are spreading out to
inﬁnity). Thus, studying orthogonal polynomials with respect to Freud weights
W is equivalent to studying orthogonal polynomials with respect to the varying
weights w
n
(x) = W(x)
n
, and actually, working with w
n
turns out to be very
natural.
For the second and third examples see multipoint Pad´e approximation and
random matrix theory in Section 6.
Matrix orthogonal polynomials
Orthogonality of matrix polynomials (i.e., when the entries of the ﬁxed size ma
trix are polynomials of degree n = 0, 1, . . . and orthogonality is with respect to
a matrix measure) is a very active area which shows extreme richness compared
to the scalar case. See Section 16 for a short discussion.
3 The L
2
extremal problem
One of the most useful tools in the study of orthogonal polynomials is the fact
that they solve the following extremal problem: minimize the L
2
(µ) norm for
all monic polynomials P
n
(z) = z
n
+ of degree n. The minimum turns out
V. Totik 78
to be 1/κ
2
n
, i.e., the nth monic orthogonal polynomial is the (unique) extremal
polynomial in
inf
Pn(z)=z
n
+···
[P
n
[
2
dµ =
1
κ
2
n
. (3.1)
Indeed, any P
n
is a linear combination
¸
n
k=1
c
k
p
k
with the orthonormal polyno
mials p
k
, and here c
n
= 1/κ
n
because P
n
is monic, i.e., it has leading coeﬃcient
1. Orthogonality gives
[P
n
[
2
dµ =
n
¸
k=0
[c
k
[
2
,
from which we can see that this is always ≥ [c
n
[
2
= 1/κ
2
n
, and equality occurs
if and only if all the other c
k
’s are 0.
A related extremum problem leads to the so called Christoﬀel functions
associated with µ. They are deﬁned as
λ
n
(µ, z) = inf
Pn(z)=1, deg(Pn)≤n
[P
n
[
2
dµ. (3.2)
If we write again P
n
=
¸
n
k=0
c
k
p
k
(µ, ), then P
n
(z) = 1 means
n
¸
k=0
c
k
p
k
(µ, z) = 1,
and hence by Cauchy’s inequality
1 ≤ (
n
¸
k=0
[c
k
[
2
)(
n
¸
k=0
[p
k
(µ, z)[
2
).
Therefore,
[P
n
[
2
dµ =
n
¸
k=0
[c
k
[
2
≥ (
n
¸
k=0
[p
k
(µ, z)[
2
)
−1
with equality if and only if
c
k
=
p
k
(µ, z)
¸
n
k=0
[p
k
(µ, z)[
2
.
Thus, we have arrived at the formula
λ
n
(µ, z)
−1
=
n
¸
k=0
[p
k
(µ, z)[
2
(3.3)
for all z ∈ C for the Christoﬀel function λ
n
(µ, z).
For example, for measures µ lying on the real line it is easy to see from
this formula that µ has a point mass at x
0
, i.e., µ(¦x
0
¦) > 0 if and only if
¸
k
p
k
(µ, x
0
)
2
< ∞, and then
µ(¦x
0
¦) = (
∞
¸
k=0
p
k
(µ, x
0
)
2
)
−1
.
Orthogonal Polynomials 79
4 Orthogonal polynomials on the real line
Let µ be supported on the real line. In this case orthogonalization leads to real
polynomials (i.e., all the coeﬃcients are real). The most remarkable property
of this real case is that the p
n
’s obey a threeterm recurrence formula
xp
n
(x) = a
n
p
n+1
(x) +b
n
p
n
(x) +a
n−1
p
n−1
(x), (4.1)
where
a
n
=
κ
n
κ
n+1
> 0, b
n
=
xp
2
n
(x)dµ(x) (4.2)
are the so called recurrence coeﬃcients. Indeed, if we write xp
n
(x) as a
linear combination
¸
n+1
k=0
c
k
p
k
(z) with
c
k
:=
xp
n
(x)p
k
(x)dµ(x),
then all the c
k
’s for k < n −1 vanish by orthogonality:
c
k
=
xp
n
(x)p
k
(x)dµ(x) =
(xp
k
(x))p
n
(x)dµ(x) = 0
because xp
k
(x) is a polynomial of degree smaller than n. Comparison of the
leading coeﬃcients on both sides gives that c
n+1
= κ
n
/κ
n+1
, but since c
n+1
is
also the integral of xp
n
(x)p
n+1
(x) against µ, we get that
c
n−1
=
xp
n
(x)p
n−1
(x)dµ(x) =
κ
n−1
κ
n
.
Finally, c
n
is the integral given in (4.2).
We emphasize that the threeterm recurrence is a very special property of
real orthogonal polynomials, and it is due to the fact that in this case the
polynomials are real, hence
xp
n
(x)p
k
(x)dµ(x) =
p
n
(x)(xp
k
(x))dµ(x) = 0
for k < n − 1. In the nonreal case the two sides here are totally diﬀerent.
The threeterm recurrence is missing in the general case, and it is replaced by
a diﬀerent recurrence for polynomials on the circle.
Conversely, any system of polynomials satisfying (4.1) with real a
n
> 0, b
n
is an orthonormal system with respect to a (not necessarily unique) measure on
the real line (Favard’s theorem). The unicity of the measure in question is the
same as the determinacy of the moment problem, which in turn is again closely
related to the behavior of orthogonal polynomials; see Section 6.
In the real case the zeros of p
n
are real and simple and the zeros of p
n
and p
n+1
interlace, i.e., in between any two zeros of p
n+1
there is a zero of
V. Totik 80
p
n
. In fact, p
n
must have n sign changes, for if it had only m < n, say at the
points y
1
, . . . , y
n
∈ R, then it could not be orthogonal to the polynomial q(x) =
¸
m
j=1
(x−y
j
) of degree m < n, for then q(x)p
n
(x) would be of constant sign. Let
now x
n
< x
n−1
< . . . < x
1
be the zeros of p
n
, and suppose that we already know
that the zeros of p
n
and p
n−1
interlace, which implies sign(p
n−1
(x
k
)) = (−1)
k−1
.
If we substitute x
k
into the recurrence formula (4.1) then a
n
> 0 gives that
p
n+1
(x
k
) and p
n−1
(x
k
) are of opposite signs at x
k
, i.e., sign(p
n+1
(x
k
)) = (−1)
k
,
and this gives that the zeros of p
n
and p
n+1
also interlace. Thus, the interlacing
property follows by induction.
The threeterm recurrence implies for the so called reproducing kernel
the ChristoﬀelDarboux formula
n
¸
k=0
p
k
(x)p
k
(t) =
κ
n
κ
n+1
p
n+1
(x)p
n
(t) −p
n
(x)p
n+1
(t)
x −t
. (4.3)
Indeed, use the recurrence formula for p
n+1
on the right and a
n
= κ
n
/κ
n+1
;
then induction gives (4.3). The special case
λ
n
(µ, x)
−1
=
n
¸
k=0
p
k
(x)
2
=
κ
n
κ
n+1
p
′
n+1
(x)p
n
(x) −p
′
n
(x)p
n+1
(x)
(4.4)
is worth mentioning.
The starting values of the recurrence (4.1) are p
−1
≡ 0, p
0
= (µ(C)))
−1/2
. If
one starts from q
−1
= −1, q
0
≡ 0 and uses the same recurrence (with a
−1
= 1)
xq
n
(x) = a
n
q
n+1
(x) +b
n
q
n
(x) +a
n−1
q
n−1
(x), (4.5)
then q
n
is of degree n − 1, and by Favard’s theorem the diﬀerent q
n
’s are or
thogonal with respect to some measure. The q
n
’s are called orthogonal poly
nomials of the second kind (sometimes for p
n
we say that they are of the
ﬁrst kind). They can also be written in the form
q
n
(z) = (µ(C))
−1/2
p
n
(z) −p
n
(x)
z −x
dµ(x).
5 Classical orthogonal polynomials
These are
• Jacobi polynomials P
(α,β)
n
, α, β > −1, orthogonal with respect to the
weight (1 −x)
α
(1 +x)
β
on [−1, 1],
• Laguerre polynomials L
(α)
n
, α > −1, with orthogonality weight x
α
e
−x
on [0, ∞),
Orthogonal Polynomials 81
• Hermite polynomials H
n
orthogonal with respect to e
−x
2
on the real
line (−∞, ∞).
In the literature various normalizations are used for them.
They are very special, for they possess many properties that no other or
thogonal polynomial system does. In particular,
• they have derivatives which form again an orthogonal polynomial system,
e.g. the derivative of P
(α,β)
n
is a constant multiple of P
(α+1,β+1)
n−1
:
(P
(α,β)
n
)
′
(x) =
1
2
(n +α +β + 1)P
(α+1,β+1)
n−1
(x),
• they all possess a Rodrigue’s type formula
P
n
(x) =
1
d
n
w(x)
d
n
dx
n
¦w(x)σ(x)
n
¦,
where w is the weight function and σ is a polynomial that is independent
of n, for example,
L
(α)
n
(x) = e
x
x
−α
1
n!
d
n
dx
n
e
−x
x
n+α
,
• they satisfy a diﬀerentialdiﬀerence relation of the form
π(x)P
′
n
(x) = (α
n
x +β
n
)P
n
(x) +γ
n
P
n−1
(x),
e.g.
x(L
(α)
n
)
′
(x) = nL
(α)
n
(x) −(n +α)L
(α)
n−1
(x),
• they satisfy a nonlinear equation of the form
σ(x) (P
n
(x)P
n−1
(x))
′
= (α
n
x +β
n
)P
n
(x)P
n−1
(x)
+γ
n
P
2
n
(x) +δ
n
P
2
n−1
(x),
with some constants α
n
, β
n
, γ
n
, δ
n
, and σ a polynomial of degree at most
2, e.g.
(H
n
(x)H
n−1
(x))
′
= 2xH
n
(x)H
n−1
(x) −H
2
n
(x) + 2nH
2
n−1
(x).
Every one of these properties has a converse, namely if a system of orthogonal
polynomials possesses any of these properties, then it is (up to a change of
variables) one of the classical systems, see AlSalam [3]. See also Bochner’s
result in the next section claiming that the classical orthogonal polynomials are
essentially the only polynomial (not just orthogonal polynomial) systems that
satisfy a certain second order diﬀerential equation.
Classical orthogonal polynomials are also special in the sense that they pos
sess a relatively simple
V. Totik 82
• second order diﬀerential equation, e.g.
xy
′′
+ (α + 1 −x)y
′
+ny = 0
for L
(α)
n
,
• generating function, e.g.
¸
n
H
n
(x)
n!
w
n
= exp(2xw −w
2
),
• integral representation, e.g.
(1 −x)
α
(1 +x)
β
P
(α,β)
n
(x) =
(−1)
n
2
n+1
πi
(1 −t)
n+α
(1 +t)
n+β
(t −x)
−n−1
dt
over an appropriate contour,
and these are powerful tools to study their behavior.
For all these results see Szeg˝ o [91].
6 Where do orthogonal polynomials come from?
In this section we mention a few selected areas where orthogonal polynomials
naturally arise.
Continued fractions
Continued fractions played an extremely important role in the development of
several branches of mathematics, but their signiﬁcance has unjustly diminished
in modern mathematics. A continued fraction is of the form
B
1
A
1
+
B2
A2+···
,
and its nth convergent is
S
n
R
n
=
B
1
A
1
+
B2
A2+···
Bn
An
, n = 1, 2, . . . .
The value of the continued fraction is the limit of its convergents. The denom
inators and numerators of the convergents satisfy the threeterm recurrence
relations
R
n
= A
n
R
n−1
+B
n
R
n−2
, R
0
≡ 1, R
−1
≡ 0
S
n
= A
n
S
n−1
+B
n
S
n−2
, S
0
≡ 0, S
−1
≡ 1,
Orthogonal Polynomials 83
which immediately connects continued fractions with threeterm recurrences
and hence with orthogonal polynomials.
But the connection is deeper than just this formal observation. Many el
ementary functions (like z −
√
z
2
−1) have a continued fraction development
where the B
n
’s are constants while the A
n
’s are linear functions, in which case
the convergents are ratios of some orthogonal polynomials of the second and
ﬁrst kind. An important example is that of Cauchy transforms of measures µ
with compact support on the real line (so called Markov functions):
f(z) =
dµ(x)
x −z
= −
α
0
z
−
α
1
z
2
−. . . . (6.1)
The coeﬃcients α
j
in the development of (6.1) are the moments
α
j
=
x
j
dµ(x), j = 0, 1, . . .
of the measure µ. The continued fraction development
f(z) ∼
B
1
z −A
1
+
B2
z−A2+···
of f at inﬁnity converges locally uniformly outside the smallest interval that
contains the support of µ (A. Markov’s theorem).
As has been mentioned, the numerators S
n
(z) and the denominators R
n
(z)
of the nth convergents
S
n
(z)
R
n
(z)
=
B
1
z −A
1
+
B2
z−A2+···
Bn
z−An
, n = 1, 2, . . .
satisfy the recurrence relations
R
n
(z) = (z −A
n
)R
n−1
(z) +B
n
R
n−2
(z), R
0
≡ 1, R
−1
≡ 0 (6.2)
S
n
(z) = (z −A
n
)S
n−1
(z) +B
n
S
n−2
(z), S
0
≡ 0, S
−1
≡ 1.
These are precisely the recurrence formulae for the monic orthogonal polyno
mials of the ﬁrst and second kind with respect to µ, hence the nth convergent
is cq
n
(z)/p
n
(z) with c = µ(C)
1/2
.
See Szeg˝ o [91, pp. 54–57] as well as Kruschev [42] and the numerous refer
ences there.
Pad´e approximation and rational interpolation
One easily gets from the recurrence relations (6.2) that
S
m
(z)
R
m
(z)
−
S
m+1
(z)
R
m+1
(z)
= (−1)
n
B
1
B
2
B
n+1
R
n
(z)R
n+1
(z)
,
V. Totik 84
and summation of these for m = n, n + 1, . . . yields that
S
n
(z)
R
n
(z)
=
2n
¸
k=0
−α
k
z
k+1
+O(z
−2n−1
),
i.e., with the preceding notation the rational function
S
n
(z)/R
n
(z) = cq
n
(z)/p
n
(z) with c = µ(C)
1/2
of numerator degree n − 1 and of denominator degree n interpolates f(z) at
inﬁnity to order 2n. This is the analogue (called [n − 1/n] Pad´e approxi
mant) of the nth Taylor polynomial (which interpolates the function to order
n) for rational functions. The advantage of Pad´e approximation over Taylor
polynomials lies in the fact that the poles of Pad´e approximants may imitate
the singularities of the function in question, while Taylor polynomials are good
only up to the ﬁrst singularity. The error in [n −1/n] Pad´e approximation has
the form
f(z) −c
q
n
(z)
p
n
(z)
=
1
p
2
n
(z)
p
2
n
(x)
x −z
dµ(x).
Orthogonal polynomials appear in more general rational interpolation (called
multipoint Pad´e approximation) to Markov functions, see e.g. Stahl–Totik
[87, Sec. 6.1]. For every n select a set A
n
= ¦x
0,n
, . . . , x
2n,n
¦ of 2n+1 interpola
tion points from C`I where I is the smallest interval that contains the support
of µ. The points need not be distinct, but we assume that A
n
is symmetric with
respect to the real line. Put
ω
n
(z) :=
2n
¸
j=0
xjn=∞
(z −x
jn
).
The degree of ω
n
is equal the number of ﬁnite points in A
n
. By r
n
(z) =
u
n
(z)/Q
n
(z) we denote a rational function of numerator and denominator de
gree at most n that interpolates the function f at the 2n + 1 points of the set
A
n
= ¦x
0,n
, . . . , x
2n,n
¦ in the sense that
f(z) −r
n
(f, A
n
; z)
ω
n
(z)
= O(z
−(2n+1)
) as [z[ →∞;
the expression on the left is bounded at every ﬁnite point of A
n
, and at inﬁnity
it has the indicated behavior. Now for Markov functions this rational inter
polant uniquely exists, Q
n
is the nth orthogonal polynomial with respect to
the varying weight dµ(x)/ω
n
(x), and the remainder term of the interpolation
has the representation
(f −r
n
(f, A
n
; ))(z) =
ω
n
(z)
Q
2
n
(z)
Q
2
n
(x)
ω
n
(x)(x −z)
dµ(x)
Orthogonal Polynomials 85
for all z outside the support of µ. Thus, the rate of convergence of the ra
tional interpolants is intimately connected with the behavior of the orthogonal
polynomials with respect to the varying weight dµ(x)/ω
n
(x).
Moment problem
The moments of a measure µ, µ(C) = 1, supported on the real line, are
α
n
=
x
n
dµ(x), n = 0, 1, . . . .
The Hamburger moment problem is to determine if a sequence ¦α
n
¦ (with nor
malization α
0
= 1) of real numbers is the moment sequence of a measure with
inﬁnite support, and if this measure is unique (the Stieltjes moment problem
asks the same, but for measures on [0, ∞)). The existence is easy: ¦α
n
¦ are
the moments of some measure supported on R if and only if all the Hankel
determinants [α
i+j
[
m
i,j=0
, m = 0, 1, . . ., are positive. The unicity (usually called
determinacy) depends on the behavior of the orthogonal polynomials (2.2) de
ﬁned from the moments σ
i,j
= α
i+j
. In fact, there are diﬀerent measures with
the same moments α
j
if and only if there is a nonreal z
0
with
¸
n
[p
n
(z
0
)[
2
< ∞,
which in turn is equivalent to
¸
n
[p
n
(z)[
2
< ∞ for all z ∈ C. Furthermore, the
Cauchy transforms of all solutions ν of the moment problem have the parametric
form
dν(x)
z −x
=
C(z)F(z) +A(z)
D(z)F(z) +B(z)
,
where F is an arbitrary analytic function (the parameter) mapping the upper
half plane C
+
into C
+
∪¦∞¦, and A, B, C and D have explicit representations
in terms of the ﬁrst and second kind orthogonal polynomials p
n
and q
n
:
A(z) = z
¸
n
q
n
(0)q
n
(z); B(z) = −1 +z
¸
n
q
n
(0)p
n
(z);
C(z) = 1 +z
¸
n
p
n
(0)q
n
(z); D(z) = z
¸
n
p
n
(0)p
n
(z).
For all these results see Akhiezer [2], and for an operator theoretic approach
to the moment problem see Simon [78] (in particular, Theorems 3 and 4.14).
Jacobi matrices and spectral theory of selfadjoint operators
Tridiagonal, so called Jacobi matrices
J =
¸
¸
¸
¸
¸
¸
b
0
a
0
0 0
a
0
b
1
a
1
0
0 a
1
b
2
a
2
0 0 a
2
b
2
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
¸
V. Totik 86
with bounded a
n
> 0 and bounded real b
n
deﬁne a bounded selfadjoint operator
J in l
2
, a so called Jacobi operator. These are the discrete analogues of second
order linear diﬀerential operators of Schr¨odinger type on the half line. Every
bounded self adjoint operator with a cyclic vector is a Jacobi operator in an
appropriate base.
The formal eigenequation Jπ = λπ is equivalent to the threeterm recur
rence
a
n−1
π
n−1
+b
n
π
n
+a
n
π
n+1
= λπ
n
, n = 1, 2, . . . ,
b
0
π
0
+ a
0
π
1
= λπ
0
, π
0
= 1.
Thus, π
n
(λ) is of degree n in λ.
By the spectral theorem, J, as a selfadjoint operator having a cyclic vector
((1, 0, 0, . . .)), is unitarily equivalent to multiplication by x in some L
2
(µ) with
some probability measure µ having compact support on the real line. This µ
is called the spectral measure associated with J (and with its spectrum). More
precisely, if p
n
(x) = p
n
(µ, x) are the orthonormal polynomials with respect
to µ, and U maps the unit vector e
n
= (0, . . . , 0, 1, 0, . . .) into p
n
, then U
can be extended to a unitary operator from l
2
onto L
2
(µ), and if Sf(x) =
xf(x) is the multiplication operator by x in L
2
(µ), then J = U
−1
SU. The
recurrence coeﬃcients for p
n
(µ, x) are precisely the a
n
’s and b
n
’s from the Jacobi
matrix, i.e., p
n
(x) = cπ
n
(x) with some ﬁxed constant c. These show that Jacobi
operators are equivalent to multiplication by x in L
2
(µ) spaces if the particular
basis ¦p
n
(µ, )¦ is used (see e.g. Deift [20, Ch. 2]).
The truncated n n matrix
J
n
=
¸
¸
¸
¸
¸
¸
b
0
a
0
0 0
a
0
b
1
a
1
0
0 a
1
b
2
a
2
.
.
.
.
.
.
.
.
.
.
.
.
0 0 0 a
n−2
b
n−1
¸
has n real and distinct eigenvalues, which turn out to be the zeros of p
n
, i.e.,
the monic polynomial p
n
(z)/κ
n
is the characteristic polynomial of J
n
.
Quadrature
For a measure µ, an npoint quadrature (rule) is a sequence of n points
x
n,1
, . . . , x
n,n
and a sequence of associated numbers λ
n,1
, . . . , λ
n,n
. It is ex
pected that
fdµ ∼
n
¸
k=1
λ
n,k
f(x
n,k
)
in some sense for as large a class of functions as possible. Often the accuracy of
the quadrature is measured by its exactness, which is deﬁned as the largest m
Orthogonal Polynomials 87
such that the quadrature is exact for all polynomials of degree at most m, i.e.,
m is the largest number with the property that
x
j
dµ(x) =
n
¸
k=1
λ
n,k
x
j
n,k
for all 0 ≤ j ≤ m.
For µ with support on the real line and for quadrature based on n points this
exactness m cannot be larger than 2n − 1, and this optimal value 2n − 1 is
attained if and only if x
n,1
, . . . , x
n,n
are the zeros of the orthonormal polynomial
p
n
(µ, x) corresponding to µ, and the so called Cotes numbers λ
n,k
are chosen
to be
λ
n,k
= λ
n
(µ, x
n,k
) =
¸
n
¸
j=0
p
j
(µ, x
n,k
)
2
¸
−1
,
where λ
n
(µ, z) is the Christoﬀel function (3.2) associated with µ.
See Szeg˝ o [91, Ch. XV].
Random matrices
Some statisticalmechanical models in quantum systems use random matrices.
Let H
n
be the set of all nn Hermitian matrices M = (m
i,j
)
n
i,j=1
, and let there
be given a probability distribution on H
n
of the form
P
n
(M)dM = D
−1
n
exp(−nTr¦V (M)¦)dM,
where V (λ), λ ∈ R, is a realvalued function that increases suﬃciently fast at
inﬁnity (typically an even polynomial in quantum ﬁeld theory applications),
Tr¦H¦ denotes the trace of the matrix H,
dM =
n
¸
k=1
dm
k,k
¸
k<j
d ℜm
k,j
d ℑm
k,j
is the “Lebesgue” measure for Hermitian matrices, and D
n
is a normalizing
constant so that the total integral of P
n
(M)dM is one.
Every matrix M ∈ H
n
has n real eigenvalues which carry physical informa
tion on the system when it is in the state described by M. The quantity
N
n
(D) =
#¦eigenvalues in D¦
n
is the random variable that equals the normalized number of eigenvalues in the
interval D. This model is known as the unitary ensemble associated with V .
Let p
j
(w
n
, x) be the orthonormal polynomials with respect to the varying
weight w
n
(x), w(x) = exp(−V (x)). Then the joint probability density of the
eigenvalues can be written in the form
d
n
p
i−1
(w
n
, λ
j
)w
n/2
(λ
j
)
2
1≤i,j≤n
,
V. Totik 88
where d
n
is a normalizing constant built up from the leading coeﬃcients of the
p
j
(w
n
, ). With the so called weighted reproducing kernel
K
n
(t, s) =
n−1
¸
j=0
p
j
(w
n
, t)w
n/2
(t) p
j
(w
n
, s)w
n/2
(s),
it can also be written in the form
1
n!
[K
n
(λ
i
, λ
j
)[
1≤i,j≤n
.
In particular, for the expected number of eigenvalues in an interval D we have
EN
n
(D) =
D
K
n
(λ, λ)
n
dλ,
where 1/K
n
(λ, λ) is known in the theory of orthogonal polynomials as the nth
(weighted) Christoﬀel function associated with the weight w
n
, while the limit
of the left hand side (as n →∞) is known as the density of states.
See, e.g., Mehta [60] and Pastur–Figotin [68].
7 Some questions leading to classical orthogonal
polynomials
There are almost an inﬁnite number of problems where classical orthogonal
polynomials emerge. Let us just mention a few.
Electrostatics
Put at 1 and −1 two positive charges p and q, and with these ﬁxed charges
put n positive unit charges on [−1, 1] at the points x
1
, . . . , x
n
. On the plane
the Coulomb force is proportional with the reciprocal of the distance, and so a
charge generates a logarithmic potential ﬁeld. Therefore, the mutual energy of
all these charges is
I(x
1
, . . . , x
n
) = p
n
¸
j=1
log
1
[1 −x
j
[
+q
n
¸
j=1
log
1
[1 +x
j
[
+
¸
i<j
log
1
[x
i
−x
j
[
,
and the equilibrium problem asks for ﬁnding x
1
, . . . , x
n
for which this energy is
minimal. The unique minimum occurs (see Szeg˝ o [91, Section 6.7]) for the zeros
of the Jacobi polynomial P
(2p−1,2q−1)
n
orthogonal with respect to the weight
(1 −x)
2p−1
(1 +x)
2q−1
.
There is a similar characterization of the zeros of Laguerre and Hermite
polynomials, and even of more general orthogonal polynomials (for the latter
see Ismail [39, Section 3.5]).
Orthogonal Polynomials 89
Polynomial solutions of eigenvalue problems
Consider the eigenvalue problem
f(x)
d
2
dx
2
y(x) +g(x)
d
dx
y(x) +h(x)y(x) = λy(x),
where f, g, h are ﬁxed polynomials and λ is a free constant, and it is required that
this have a polynomial solution of exact degree n for all n = 0, 1, . . ., for which
the corresponding λ and y(x) will be denoted by λ
n
and y
n
(x), respectively.
Bochner’s theorem from [16] states that, except for some trivial solutions of the
form y(x) = ax
n
+ bx
m
and for some polynomials related to Bessel functions,
the only solutions are (in all of them we can take h(x) = 0)
• Jacobi polynomials P
(α,β)
n
(f(x) = 1 − x
2
, g(x) = β − α − x(α + β + 2),
λ
n
= −n(n +α +β + 1))
• Laguerre polynomials L
(α)
n
(f(x) = x, g(x) = 1 +α −x, λ
n
= −n) and
• Hermite polynomials H
n
(x) (f(x) = 1, g(x) = −2x, λ
n
= −2n).
Harmonic analysis on spheres and balls
Harmonic analysis on spheres and balls in R
d
is based on spherical harmonics,
i.e., harmonic homogeneous polynomials. In this theory, special Jacobi polyno
mials, the so called ultraspherical or Gegenbauer polynomials P
(α)
n
, play a
fundamental role – they are orthogonal with respect to the weight (1−x
2
)
α−1/2
.
Let S
d−1
be the unit sphere in R
d
and let H
d
n
be the restriction to S
d−1
of
all harmonic polynomials Q(x
1
, . . . , x
n
) of d variables that are homogeneous of
degree n, i.e.,
n
¸
k=1
∂
2
∂x
2
k
Q = 0, Q(λx
1
, . . . , λx
n
) = λ
n
Q(x
1
, . . . , x
n
), λ > 0.
The dimension of H
d
n
is
n +d −1
d −1
−
n +d −3
d −1
,
and an orthogonal basis in it can be produced as follows. With ρ = x
2
d−1
+ x
2
d
let g
s,0
= ρ
s
P
(0)
s
(x
d−1
/ρ) and g
s,1
= x
d
ρ
s
P
(1)
s
(x
d−1
/ρ). With n
d
= 0 or n
d
= 1
consider all multiindices n = (n
1
, n
2
, . . . , n
d
) such that n
1
+ + n
d
= n, and
if for such a multiindex we deﬁne the function Y
n
(x
1
, . . . , x
d
) as
g
n
d−1
,n
d
d−2
¸
j=1
(x
2
j
+ +x
2
d
)
nj
P
(λj )
nj
(x
j
(x
2
j
+ +x
2
d
)
−1/2
)
,
V. Totik 90
then these Y
n
constitute an orthogonal basis in H
d
n
(see e.g. Dunkl–Xu [24, p.
35]).
If x = (x
1
, . . . , x
n
) and 'x, y` =
¸
k
x
k
y
k
is the inner product in R
d
, then
the reproducing kernel for these spherical polynomials is P
(d−2)/2
n
('x, y`) in the
sense that for all Q ∈ H
d
n
and for all x ∈ S
d−1
we have
c
n,d
P
(d−2)/2
n
('x, y`)Q(y)dσ(y) = Q(x),
where integration is with respect to surface area, and c
n,d
is an explicit normal
izing constant (see e.g. Dunkl–Xu [24, p. 37]).
As a result, Gegenbauer polynomials are all over the theory of spherical
harmonics, as well as in the corresponding theory for the unit ball.
Approximation theory
In the literature, expansions of functions into classical orthogonal polynomial
series are second only to trigonometric expansions, and numerous works have
been devoted to their convergence and approximation properties, see e.g. Szeg˝ o
[91, Ch. XIII].
The Chebyshev polynomials cos(narccos x) are orthogonal on [−1, 1]
with respect to the weight w(x) = (1 − x
2
)
−1/2
. These directly correspond
to trigonometric functions, and expansions into them have virtually the same
properties as trigonometric Fourier expansions. But there are many other as
pects of approximation where Chebyshev polynomials appear. If one considers,
for example, the best approximation on [−1, 1] of x
n
in the uniform norm by
polynomials P
n−1
(x) of smaller degree then the smallest error appears when
x
n
− P
n−1
(x) = 2
1−n
cos(narccos x) is the monic nth Chebyshev polynomial.
Actually, monic Chebyshev polynomials minimize all L
p
(w), p > 0, norms
among monic polynomials of a given degree.
As we have seen in (3.1), the monic orthogonal polynomials p
n
(µ)/κ
n
are
the solutions to the extremal problem
[P
n
[
2
dµ →min, (7.3)
where the minimum is taken for all monic polynomials of degree n. This ex
tremal property makes orthogonal polynomials, in particular Chebyshev poly
nomials, indispensable tools in approximation theory.
Lagrange interpolation and its various generalizations like HermiteFej´er or
Hermite interpolation etc. is mostly done on the zeros of some orthogonal poly
nomials. In fact, these nodes are often close to optimal in the sense that the
Lebesgue constant increases at the optimal rate. In many cases interpolation on
zeros of orthogonal polynomials has special properties due to explicitly calcula
ble expressions. Recall e.g. Fej´er’s result that if P
2n−1
is the unique polynomial
Orthogonal Polynomials 91
of degree at most 2n−1 that interpolates a continuous function f at the nodes
of the nth Chebyshev polynomial and that has zero derivative at each of these
nodes, then P
2n−1
uniformly converges to f on [−1, 1] as n →∞. For the role
of orthogonal polynomials in interpolation see the books Szabados–V´ertesi [90]
and Mastroianni–Milovanovic [55].
8 Heuristics
In this section we do not state precise results. We just want to indicate some
heuristics on the behavior of orthogonal polynomials. For the concepts below,
as well as for a more precise form of some of the heuristics see the following
sections, in particular Section 9.
As we have seen, the monic orthogonal polynomials p
n
(µ)/κ
n
minimize the
L
2
(µ) norm in (7.3). Therefore, these polynomials try to be small where the
measure is large, e.g. one expects the zeros to cluster at the support S(µ) of
µ. The example of arc measure on the unit circle, for which the orthogonal
polynomials are z
n
, shows however, that this is not true (due to the fact that
the complement of the support is not connected). The statement is true when
the support lies on R or on some systems of arcs, and also in the general case
when instead of the support one considers the polynomial convex hull of the
support of µ (for the deﬁnition, see the next section): on any compact set
outside the polynomial convex hull there can only be a ﬁxed number of zeros of
p
n
(µ) for every n. When the complement of S(µ) is connected and S(µ) has no
interior, then the distribution of the zeros shows a remarkable universality and
indiﬀerence to the size of µ. In many situations the distribution of the zeros
is the equilibrium distribution of the support S(µ). When S(µ) = [−1, 1], this
means that under very weak assumptions the zero distribution is always the
arcsine distribution dx/π
√
1 −x
2
.
The L
2
(µ) minimality of p
n
(µ)/κ
n
in the sense of (7.3) is something like
minimality in the L
∞
norm on S(µ). Therefore, p
n
(µ)/κ
n
should behave like
the monic polynomial T
n
minimizing the L
∞
norm on S(µ) (so called Chebyshev
polynomials for S(µ)). Since
1
n
log [T
n
(z)[ =
log [z −t[dν
n
(t)
where ν
n
has mass 1/n at each zero of T
n
, in the limit the behavior should be
like
U
ν
(z) =
log [z −t[dν(t), (8.1)
where ν is the probability measure on S(µ) for which the maximum of U
ν
on
S(µ) is as small as possible (this is the so called equilibrium measure of S(µ)).
More generally, if dν = dν
n
= w
n
(x)dx is a varying weight in the speciﬁed way,
then the same reasoning leads to a behavior like (8.1), but now ν is a measure
V. Totik 92
for which the supremum of U
ν
(z) + log w(z) is as small as possible (weighted
equilibrium measure).
Universal behavior can also be seen for the polynomials themselves. Usually
they obey
1
n
log [p
n
(µ, z)[ →g
C\S(µ)
(z, ∞), z ∈ S(µ), (8.2)
where g
C\S(µ)
(z, ∞) is the Green function with pole at inﬁnity associated with
the complement of the support. When the unbounded component of the com
plement of S(µ) is simply connected, then in that component often there is a
ﬁner asymptotic behavior of p
n
(µ) of the form
p
n
(z) ∼ d
n
g
µ
(z)Φ(z)
n
, z ∈ S(µ), (8.3)
where Φ is the mapping function that maps C ` S(µ) conformally onto the
outside of the unit disk, and g
µ
is a function (might be called generalized Szeg˝ o
function) that depends on µ. Such a ﬁne asymptotic is restricted to the simply
connected case, see e.g. Section 10.
Asymptotics of orthogonal polynomials have a hierarchy, and the diﬀerent
types of asymptotics usually require the measure to be suﬃciently strong with
diﬀerent degree on its support. Consider ﬁrst the case of compact support
S(µ). The weakest is nth root asymptotics stating the behavior (8.2) for
[p
n
(µ, z)[
1/n
outside the support of the measure. It is mostly equivalent to a
corresponding distribution of the zeros, as well as asymptotical minimal behav
ior of κ
1/n
n
. It holds under very weak assumptions on the measure, roughly
stating that the logarithmic capacity of the points where µ
′
> 0 (derivative
with respect to equilibrium measure), be the same as the capacity of S(µ). Ra
tio asymptotics, i.e., asymptotic behavior of p
n+1
(µ, z)/p
n
(µ, z), is stronger,
and is equivalent with asymptotics for the ratio κ
n+1
/κ
n
of consecutive leading
coeﬃcients. It can only hold when C`S(µ) (more precisely its unbounded com
ponent) is simply connected, and in this case it is enough that µ
′
> 0 almost
everywhere with respect to the equilibrium measure of the support of µ (see
Section 10). Finally, strong asymptotics of the form (8.3) needs roughly that
log µ
′
be integrable (Szeg˝o condition, see Section 10).
All these are outside the support. On the support the orthogonal polyno
mials are of oscillatory behavior, and in the real case under smoothness as
sumptions on the measure often a so called PlancherelRotach type asymptotic
formula
p
n
(µ, x) ∼ d
n
g(x) sin(nh(x) + H(x))
holds, where g, h, H are ﬁxed functions. Here h(x) is directly linked with the
zeros, h
′
/π is precisely the distribution of the zeros. When S(µ) = [−1, 1] and
the measure is smooth, then h(x) = arccos x.
When S(µ) is not of compact support (like Laguerre, Hermite or Freud
weights), then usually the zeros are spreading out, and one has to rescale them
to [−1, 1] (or to [0, 1]) to get a distribution, which is mostly NOT the arcsine
Orthogonal Polynomials 93
distribution. In a similar fashion, various asymptotics hold for the polynomials
only after a corresponding rescaling.
The Christoﬀel function (3.2) or, what is the same, the square sum
1
λ
n
(µ, z)
=
n
¸
k=0
[p
k
(µ, z)[
2
,
behaves much more regularly than the orthogonal polynomials. Outside the
support the behavior of λ
n
is what one gets from the heuristics above on the
polynomials (after square summation). On S(µ) the typical behavior of λ
n
(µ, z)
is like µ(D
n
(z)), where D
n
(z) is the disk about z with equilibrium measure
1/n (equilibrium measure of the support S(µ)). In particular, nλ
n
(µ, z) tends
pointwise to the RadonNikodym derivative of µ with respect to the equilibrium
measure. As a rule of thumb the estimate [p
n
(µ, x)[
2
≤ C/nλ
n
(µ, x) holds in
many cases.
If f ∈ L
2
(µ), its Fourier expansion into ¦p
n
(µ, )¦ is
f(x) ∼
∞
¸
k=0
c
k
p
k
(x), c
k
=
fp
k
dµ.
The nth partial sum has the closed form
f(t)K
n
(x, t)dµ(t), K
n
(x, t) =
n
¸
k=0
p
k
(x)p
k
(t).
In the real case for the reproducing kernel K
n
(x, t) we have the Christoﬀel
Darboux formula
K
n
(x, t) =
κ
n
κ
n+1
p
n+1
(x)p
n
(t) −p
n
(x)p
n+1
(t)
x −t
,
which suggests a singular integraltype behavior for the partial sums. In general,
Fourier expansions into orthogonal polynomials are sensitive to the weight (re
call e.g. Pollard’s theorem that Legendre expansions are bounded in L
p
[−1, 1]
only for 4/3 < p < 4), but sometimes convergence properties are equivalent to
those of a related trigonometric Fourier series (so called transplantation theo
rems, see e.g. Askey [11]).
9 General orthogonal polynomials
In this section µ always has compact support S(µ). For all the results below
see Stahl–Totik [87] and the references therein.
V. Totik 94
Lower and upper bounds
The energy V (K) of a compact set K is deﬁned as the inﬁmum of
I(ν) =
log
1
[x −t[
dν(x)dν(t) (9.1)
where the inﬁmum is taken for all positive Borel measures on K with total
mass 1. The logarithmic capacity is then cap(K) = e
−V (K)
. For the leading
coeﬃcients κ
n
of the orthonormal polynomials p
n
(µ) we have
1
cap(S(µ))
≤ liminf
n→∞
κ
1/n
n
. (9.2)
To get an upper bound we need the concept of carrier: a Borel set E is a
carrier for µ if µ(C` E) = 0. The capacity of a Borel set is the supremum of
the capacities of its compact subsets, and the minimal carrier capacity c
µ
associated with µ is the inﬁmum of the capacities of all carriers. With this
limsup
n→∞
κ
1/n
n
≤
1
c
µ
. (9.3)
When cap(K) is positive, then there is a unique measure ν = ω
K
minimizing
the energy in (9.1), and this measure is called the equilibrium measure of
K. Green’s function g
C\K
(z, ∞) with pole at inﬁnity of C` K can then be
deﬁned as
g
C\K
(z, ∞) = log
1
cap(K)
−
log
1
[z −t[
dω
K
(t). (9.4)
We have for all µ (with cap(S(µ)) > 0) the estimate
liminf
n→∞
1
n
log [p
n
(µ, z)[
1/n
≥ g
C\S(µ)
(z, ∞) (9.5)
locally uniformly outside the convex hull of S(µ), while in the convex hull but
outside the so called polynomial convex hull Pc(S(µ)) (for the deﬁnition see
below) (9.5) is true quasieverywhere (i.e., with the exception of a set of zero
capacity). The same is true on the outer boundary of S(µ), which is deﬁned
as the boundary ∂Ω of the unbounded component Ω of the complement C`S(µ),
namely for quasievery z ∈ ∂Ω
liminf
n→∞
[p
n
(µ, z)[
1/n
≥ 1.
The minimal carrier Green function g
µ
(z, ∞) is the supremum for all carri
ers E of the Green function of C` E, where the latter is deﬁned as the inﬁmum
of g
C\K
for all compact subsets K of E. With this,
limsup
n→∞
1
n
log [p
n
(µ, z)[
1/n
≤ g
µ
(z, ∞) (9.6)
Orthogonal Polynomials 95
locally uniformly on the whole plane.
All these estimates are sharp.
When the bounds in (9.2) and (9.3) coincide we have convergence for κ
1/n
n
,
and these bounds coincide precisely when the bounds in (9.5) and (9.6) do so.
Zeros
The zeros of p
n
(µ) always lie in the convex hull of the support S(µ) of the
measure µ. This is a consequence of the L
2
extremal property (3.1) of orthogonal
polynomials. In fact, if there was a zero z
0
of p
n
(µ, z) outside the convex hull
of the support, then we could move that zero towards the convex hull (along
a line segment that is perpendicular to a line separating z
0
and S(µ)). During
this move the absolute value of the polynomial decreases at all points of S(µ)
and hence so does its L
2
(µ) norm, but that is impossible by (3.1).
To say somewhat more on the location of zeros we need the concept of the
polynomial convex hull. When Ω is the unbounded component of the com
plement C ` S(µ), then Pc(S(µ)) = C ` Ω is called the polynomial convex
hull of S(µ) (it is the union of S(µ) with all the “holes” in it, i.e., with the
bounded components of C ` S(µ)). Now the zeros cluster on Pc(S(µ)) in the
sense that for any compact subset K of Ω there is a number N
K
independent
of n, such that p
n
(µ) can have at most N
K
zeros in K. The proof of this is
based on the following lemma: Let V, S ⊆ C be two compact sets. If V and
Pc(S) are disjoint, then there exist a < 1 and m ∈ N such that for arbitrary
m points x
1
, . . . , x
m
∈ V there exist m points y
1
, . . . , y
m
∈ C for which the
rational function
r
m
(z) :=
m
¸
j=1
z −y
j
z −x
j
(9.7)
has on S a supnorm satisfying
r
m

S
≤ a. (9.8)
Taking this for granted, assume that V is a compact set contained in Ω. We
apply the lemma with S = S(µ), and let a < 1 and m ∈ N be the numbers
in the lemma. Let us assume that p
n
(µ; z) has at least m zeros x
1
, . . . , x
m
on
V. By the lemma there exist m points y
1
, . . . , y
m
∈ C such that the rational
function r
m
deﬁned as in (9.7) by the points x
1
, . . . , x
m
and y
1
, . . . , y
m
satisﬁes
the inequality (9.8). With r
m
we deﬁne the modiﬁed monic polynomial
q
n
(z) := r
m
(z)p
n
(µ; z) = z
n
+. . . ,
For the L
2
(µ) norm of this polynomial we have the estimate
q
n

L
2
(µ)
≤ r
m

S(µ)
p
n
(µ; )
L
2
(µ)
< p
n
(µ; )
L
2
(µ)
,
V. Totik 96
which contradicts the minimality (3.1) of the monic orthogonal polynomial
p
n
(µ; z). Hence, we have proved that p
n
(µ; z) has at most m−1 zeros on V , as
was stated.
What we have said about the zeros can be sharpened for measures on the
real line. For example, if µ is supported on the real line, then Pc(S(µ)) = S(µ),
and if K is a closed interval disjoint from the support, then there is at most
one zero in K. It was shown in Denison–Simon [23] that if x
0
∈ R is not in
the support, then for some δ > 0 and all n either p
n
or p
n+1
has no zero in
(x
0
−δ, x
0
+δ). Note that if µ is a symmetric measure on [−1, −1/2] ∪[1/2, 1],
then p
2n+1
(0) = 0 for all n, so the result is sharp.
Any isolated point in the support that lies on the outer boundary attracts
precisely one zero. Let z
0
be an isolated point of S(µ), such that its distance
from the convex hull of S(µ) ` ¦z
0
¦ is δ > 0. Then p
n
has at most one zero in
the disk ¦[z −z
0
[ < δ/3¦ (Simon [80, Section 8.1]). It is also clear that for any
symmetric measure µ with S(µ) = [−1, −1/2] ∪ ¦0¦ ∪ [1/2, 1] the polynomials
p
2n
(µ) have 2 zeros near 0, so the result is sharp (in this case δ = 0). Moreover,
if µ lies on the unit circle, then there exist two positive constants C and a and
a zero z
n
of p
n
such that [z
n
−z
0
[ ≤ Ce
−an
.
In general, each component of the polynomial convex hull consisting of more
than one point attracts inﬁnitely many zeros: if γ is a Jordan curve in Ω such
that S(µ) ∩ γ is inﬁnite, then the number of zeros of p
n
that lie inside γ tends
to inﬁnity (Saﬀ–Totik [76]). Mass points of µ do not necessarily attract zeros
(above we have mentioned that they do if they lie on the outer boundary).
In fact, it was shown in Saﬀ–Totik [76] that if ρ is the measure on the unit
circle given by the density function sin
2
(θ/2), then for any measure σ that is
supported in the open unit disk there is a λ > 0 such that all zeros of the nth
orthogonal polynomials with respect to µ = ρ + λσ tend to the unit circle as
n →∞.
Next put a unit mass at every zero of p
n
(µ) (counting multiplicity). This
gives the so called counting measure ν
pn(µ)
on the zero set. Zero distri
bution amounts to ﬁnding the limit behavior of
1
n
ν
pn(µ)
. The normalized arc
measure on the unit circle (for which p
n
(µ, z) = z
n
) shows that if the interior
of the polynomial convex hull Pc(S(µ)) is not empty, then the zeros may be
far away from the outer boundary ∂Ω, where the equilibrium measure ω
S(µ)
is
supported. Thus, assume that Pc(S(µ)) has empty interior and also that there
is no Borel set of capacity zero and full µmeasure, i.e., the minimal carrier
capacity c
µ
is positive (the c
µ
= 0 case is rather pathological, almost anything
can happen with the zeros then). In this case
lim
n→∞
κ
1/n
n
= log
1
cap(S(µ))
(9.9)
if and only if
lim
1
n
ν
pn(µ)
= ω
S(µ)
Orthogonal Polynomials 97
in weak
∗
sense, i.e., asymptotically minimal behavior of κ
1/n
n
(see (9.2)) is equiv
alent to the fact that the zero distribution is the equilibrium distribution. In a
similar way, asymptotic maximal behavior (see (9.3)), i.e.,
lim
n→∞
κ
1/n
n
=
1
c
µ
(9.10)
holds precisely when
lim
n→∞
1
n
ν
pn(µ)
= ω
µ
,
where ω
µ
is the so called minimal carrier equilibrium measure, for which a
representation like (9.4) is true, but for the minimal carrier Green function g
µ
.
Regularity
(9.9) is called regular limit behavior, and in this case we write µ ∈ Reg.
Thus, the important class Reg is deﬁned by the property (9.9). µ ∈ Reg is
equivalent to either of
• lim
n→∞
[p
n
(µ, z)[
1/n
= exp(g
C\S(µ)
(z, ∞)), z ∈ Con(S(µ))
• limsup
n→∞
[p
n
(µ, z)[
1/n
= 1 for quasievery z ∈ ∂Ω.
If Ω is a regular set with respect to the Dirichlet problem, then µ ∈ Reg is
equivalent to either of
• lim
n→∞
p
n
(µ)
1/n
sup,S(µ)
= 1
• For any sequence ¦P
n
¦ of polynomials of degree n = 1, 2, . . .
lim
n→∞
P
n

sup,S(µ)
P
n

L
2
(µ)
1/n
= 1.
The last statement expresses the fact that in the nth root sense the L
2
(µ) and
L
∞
norms (on S(µ)) are asymptotically the same.
All equivalent formulations of µ ∈ Reg point to a certain “thickness” of
µ on its support. Regularity is an important property, and it is desirable to
know “thickness” conditions under which it holds. Several regularity criteria
are known, e.g. either of the conditions
• all Borel sets B ⊆ S(µ) with full measure (i.e with µ(B) = µ(S(µ))) have
capacity cap(B) = cap(S(µ)), i.e., c
µ
= cap(S(µ)) or
• dµ/dω
S(µ)
> 0 (RadonNikodym derivative) ω
S(µ)
almost everywhere
V. Totik 98
is suﬃcient for µ ∈ Reg. Regularity holds under fairly weak assumptions on
the measure, e.g. if S(µ) = [0, 1], and
liminf
r→0
r log µ([x −r, x +r]) ≥ 0
for almost every x ∈ [0, 1] (i.e., if µ is not exponentially small around almost
every point), then µ ∈ Reg.
No necessary and suﬃcient condition for regularity in terms of the size of
the measure µ is known. The only existing necessary condition is for the case
S(µ) = [0, 1], and it reads that for every η > 0
lim
n→∞
cap
¸
x µ([x −1/n, x + 1/n]) ≥ e
−ηn
¸
=
1
4
(here 1/4 is the capacity of [0, 1]).
10 Strong, ratio and weak asymptotics
Strong asymptotics
Let µ be supported on [−1, 1] and suppose that the so called Szeg˝ o condition
1
−1
log µ
′
(t)
√
1 −t
2
dt > −∞ (10.1)
holds, where µ
′
is the RadonNikodym derivative of µ with respect to linear
Lebesgue measure. Note that this condition means that the integral is ﬁnite,
for it cannot be ∞. It expresses a certain denseness of µ, and under this con
dition G. Szeg˝ o proved several asymptotics for the corresponding orthonormal
polynomials p
n
(µ). This theory was developed on the unit circle and then was
translated into the real line. The Szeg˝ o function associated with µ is
D
µ
(z) := exp
z
2
−1
1
2π
1
−1
log µ
′
(t)
z −t
dt
√
1 −t
2
(10.2)
and it is the outer function in the Hardy space on C ` [−1, 1] with boundary
values [D
µ
(x)[
2
= µ
′
(x). Outside [−1, 1] the asymptotic formula
p
n
(µ, z) = (1 +o(1))
1
√
2π
(z +
z
2
−1)
n
D
µ
(z)
−1
(10.3)
holds locally uniformly. In particular, the leading coeﬃcient κ
n
of p
n
(µ) is of
the form
κ
n
= (1 +o(1))
2
n
√
2π
exp
−1
2π
1
−1
log µ
′
(t)
√
1 −t
2
dt
. (10.4)
Orthogonal Polynomials 99
If dµ(x) = w(x)dx and h(t) = w(cos t) sin t satisﬁes a DiniLipshitz condition
[h(t +δ) −h(t)[ ≤
C
[ log δ[
1+ε
, ε > 0,
then with
Γ
w
(x) :=
1
2π
1
−1
log w(ξ) −log w(x)
ξ −x
1 −x
2
1 −ξ
2
1/2
dξ,
we have uniformly on [−1, 1]
(1 −x
2
)
1/4
w(x)
1/2
p
n
(x) =
2
π
1/2
cos
(n +
1
2
) arccos x + Γ
w
(x) −
π
4
+ O((log n)
−ε
).
For all these results see Szeg˝ o [91], Chapter 6. The Szeg˝ o condition is also
necessary for these results, e.g. an asymptotic formula like (10.3) and (10.4) is
equivalent to (10.1).
Ratio asymptotics
If one assumes weaker conditions then necessarily weaker results will follow. A
large and important class of measures is the Nevai class M(b, a) (see Nevai
[62]), for which the coeﬃcients in the threeterm recurrence
xp
n
(x) = a
n
p
n+1
(x) +b
n
p
n
(x) +a
n−1
p
n−1
(x)
satisfy a
n
→a, b
n
→b. This is equivalent to ratio asymptotics
lim
n→∞
p
n+1
(z)
p
n
(z)
=
z −b +
(z −b)
2
−4a
2
2
for large z (actually, away from the support of µ), and the monograph Nevai
[62] contains a very detailed treatment of orthogonal polynomials in this class.
It is also true that if the limit of p
n+1
(z)/p
n
(z) exists at a single nonreal z,
then µ ∈ M(b, a) for some a, b (Simon [79]).
The classes M(b, a) are scaled versions of each other, and the most im
portant condition ensuring M(0, 1/2) is given in Rakhmanov’s theorem from
[75]: if µ is supported in [−1, 1] and µ
′
> 0 almost everywhere on [−1, 1],
then µ ∈ M(0, 1/2). Conversely, Blumenthal’s theorem from [15] states that
µ ∈ M(0, 1/2) implies that the support of µ is [−1, 1] plus at most count
ably many points that converge to ±1. Thus, in this respect the extension of
Rakhmanov’s theorem given in [22] by Denisov is of importance: if µ
′
> 0 almost
everywhere on [−1, 1] and outside [−1, 1] the measure µ has at most countably
many mass points converging to ±1, then µ ∈ M(0, 1/2). However, M(0, 1/2)
V. Totik 100
contains many other measures not just those that are in these theorems, e.g. in
Delyon–Simon–Souillard [21] a continuous singular measure in the Nevai class
was exhibited, and the result in Totik [92] shows that the Nevai class contains
practically all types of measures allowed by Blumenthal’s theorem.
Weak and relative asymptotics
Under Rahmanov’s condition supp(µ) = [−1, 1], µ
′
> 0 a.e., some parts of
Szeg˝ o’s theory can be proven in a weaker form (see e.g. M´ at´e–Nevai–Totik
[57, 58]). In these the Tur´an determinants
T
n
(x) := p
2
n
(x) −p
n−1
(x)p
n+1
(x)
play a signiﬁcant role. In fact, then given any interval D ⊂ (−1, 1) the Tur´an
determinant T
n
is positive on D for all large n, and T
n
(x)
−1
dx converges in the
weak
∗
sense to dµ on D. Furthermore, the absolutely continuous part µ
′
can
be also separately recovered from T
n
:
lim
n→∞
T
n
(x)µ
′
(x) −
2
π
(1 −x
2
)
1/2
dx = 0.
Under Rahmanov’s condition we also have weak convergence, for example,
lim
n→∞
f(x)p
2
n
(x)µ
′
(x)dx =
1
π
1
−1
f(x)
√
1 −x
2
dx (10.5)
for any continuous function f. Pointwise we only know a highly oscillatory
behavior: for almost all x ∈ [−1, 1]
limsup
n→∞
p
n
(x) ≥
2
π
(µ
′
(x))
−1/2
(1 −x
2
)
−1/4
,
liminf
n→∞
p
n
(x) ≤ −
2
π
(µ
′
(x))
−1/2
(1 −x
2
)
−1/4
,
and if E
n
(ε) is the set of points x ∈ [−1, 1] where
[p
n
(x)[ ≥ (1 +ε)
2
π
(µ
′
(x))
−1/2
(1 −x
2
)
−1/4
,
then [E
n
(ε)[ → 0 for all ε > 0. However, it is not true that the sequence
¦p
n
(µ, x)¦ is pointwise bounded, since for every ε > 0 there is a weight function
w > 1 on [−1, 1] such that p
n
(0)/n
1/2−ε
is unbounded (see Rakhmanov [74]).
Simon [79] extended (10.5) by showing that if the recurrence coeﬃcients
satisfy b
n
→b, a
2n+1
→a
′
and a
2n
→a
′′
, then there is an explicitly calculated
measure ρ depending only on b, a
′
, a
′′
such that
lim
n→∞
f(x)p
2
n
(x)µ
′
(x)dx =
1
−1
f(x)dρ(x) (10.6)
Orthogonal Polynomials 101
for any continuous function f, and conversely, if (10.6) exists for f(x) = x, x
2
, x
4
,
then b
n
→b, a
2n+1
→a
′
and a
2n
→a
′′
with some b, a
′
, a
′′
.
For measures in Nevai’s class, part of Szeg˝ o’s theory can be extended to
relative asymptotics, i.e., when sequences of orthogonal polynomials corre
sponding to two measures are compared. Here is a sample theorem: let α be
supported in [−1, 1] and in Nevai’s class M(0, 1/2), and let dβ = gdα, where g
is a function such that for some polynomial R both Rg and R/g are Riemann
integrable. Then
lim
n→∞
p
n
(β, z)
p
n
(α, z)
= D
g
(z)
−1
uniformly on C away from [−1, 1], where D
g
is Szeg˝ o’s function with respect to
the measure g(x)dx.
Widom’s theory
Szeg˝ o’s theory can be extended to measures lying on a single Jordan curve
or arc J (see Kaliaguine [40] where also additional outside lying mass points
are allowed), in which case the role of z +
√
z
2
−1 in (10.3) is played by the
conformal map Φ of C ` J onto the exterior of the unit disk, and the role of
2
n
in (10.4) is played by the reciprocal of the logarithmic capacity of J (see
Section 9). Things change considerably if the measure is supported on a set J
consisting of two or more smooth curve or arc components J
1
, . . . , J
m
. A general
feature of this case is that κ
n
cap(J)
n
does not have a limit, its limit points ﬁll
a whole interval (though if some associated harmonic measures are all rational
then the limit points may form a ﬁnite set). The polynomials themselves have
asymptotic form
p
n
(z)
κ
n
= cap(J)
n
Φ(z)
n
(F
n
(z) +o(1))
uniformly away fromJ, where Φ is the (multivalued) complex Green function of
the complement C` J, and where F
n
is the solution of an L
2
extremal problem
involving analytic functions belonging to some class Γ
n
. The functions F in
Γ
n
are determined by an H
2
condition plus an argument condition, namely
if the change of the argument of Φ as we go around J
k
is γ
k
2π modulo 2π,
then in Γ
n
we consider functions whose change of the argument around J
k
is
−nγ
k
2π modulo 2π. Now the point is that these function classes Γ
n
change
with n, and hence so does F
n
, and this is the reason that a single asymptotic
formula like (10.4) or (10.3) does not hold. The fundamentals of the theory
were laid out in H. Widom’s paper [97]; and since then many results have been
obtained by F. Peherstorfer and his collaborators, as well as A. I. Aptekarev, J.
Geronimo, S. P. Suetin and W. Van Assche. The theory has deep connections
with function theory, the theory of Abelian integrals and the theory of elliptic
functions. We refer the reader to the papers Aptekarev [7], Geronimo–Van
Assche [31], Peherstorfer [69]–[72] and Suetin [88]–[89].
V. Totik 102
Asymptotics for Christoﬀel functions
The Christoﬀel functions
λ
n
(µ, x)
−1
=
n
¸
k=0
p
k
(µ, x)
2
behave somewhat more regularly than the orthogonal polynomials. In M´ at´e–
Nevai–Totik [59] it was shown that if µ is supported on [−1, 1], it belongs to
the Reg class there (see Section 9) and log µ
′
is integrable over an interval
I ⊂ [−1, 1], then for almost all x ∈ I
lim
n→∞
nλ
n
(µ, x) = π
1 −x
2
µ
′
(x).
This result is true (see Totik [93]) in the form
lim
n→∞
nλ
n
(µ, x) =
dµ(x)
dω
supp(µ)
(x)
, a.e. x ∈ I
when the support is a general compact subset of R, µ ∈ Reg and log µ
′
∈ L
1
(I).
Often only a rough estimate is needed for Christoﬀel functions, and such a
one is provided in Mastroianni–Totik [56]: if µ is supported on [−1, 1] and it is
a doubling measure, i.e.,
µ(2I) ≤ Lµ(I)
for all I ⊂ [−1, 1], where 2I is the twice enlarged I, then uniformly on [−1, 1]
λ
n
(µ, x) ∼ µ(∆
n
(x)) ; ∆
n
(x) =
√
1 −x
2
n
+
1
n
2
.
11 Recurrence coeﬃcients and spectral measures
Let µ be a unit measure of compact support on the real line, and
xp
n
(x) = a
n
p
n+1
(x) +b
n
p
n
(x) +a
n−1
p
n−1
(x),
the recurrence relation for the corresponding orthogonal polynomials. We have
already mentioned in Section 6 that µ is the spectral measure for the Jacobi
matrix
J =
¸
¸
¸
¸
¸
¸
b
0
a
0
0 0
a
0
b
1
a
1
0
0 a
1
b
2
a
2
0 0 a
2
b
2
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
¸
,
and this gives a onetoone correspondence between unit measures with compact
(and inﬁnite) support on the real line and Jacobi operators with bounded en
tries. Every such Jacobi operator is a bounded selfadjoint operator on l
2
, hence
Orthogonal Polynomials 103
operator theory and orthogonal polynomials meet at this point, and techniques
and questions from both areas are relevant. Information between the measure
µ and the recurrence coeﬃcients might be called the spectral analysis of or
thogonal polynomials. In this, a special role is played by the Chebyshev case
a
n
= 1/2 and b
n
= 0, where the Jacobi matrix is denoted by J
0
. If the sequences
¦a
n
¦ and ¦b
n
¦ have limits then we may assume a
n
→ 1/2, b
n
→0 (this is just
rescaling, and this was the Nevai class M(0, 1/2)), hence in this case the Jacobi
operator is a compact perturbation of J
0
, and one of the main questions of the
theory is how properties of J −J
0
are reﬂected in the spectral measure µ.
We have already mentioned in Section 10 that µ ∈ M(0, 1/2) implies that
the support of µ is [−1, 1] plus some additional mass points converging to ±1
(called Blumenthal’s theorem in orthogonal polynomials; it is a special case of
Weyl’s theorem in operator theory on the invariance of the essential spectrum
under compact perturbation). Conversely, if the support of µ is [−1, 1] plus
some additional mass points converging to ±1 and µ
′
(x) > 0 for almost all
x ∈ [−1, 1], then µ ∈ M(a, b) (Denisov [22], Rakhmanov [75], Nevai–Totik [63]).
No spectral characterization of µ ∈ M(0, 1/2) is known; this important class
seems to contain all sorts of measures. For example, if ν is any measure with
support [−1, 1] then there is a µ ∈ M(0, 1/2) which is absolutely continuous with
respect to ν. In particular, M(0, 1/2) contains discrete measures, continuously
singular measures or measures that are given by a continuous density which is
positive on a set of measure < ε.
Strengthening the condition µ ∈ M(0, 1/2) can be done in several ways.
After numerous works in the subject by Szeg˝ o, Shohat, Geronimus, Krein, Kol
mogorov and others, a complete characterization for J − J
0
being a Hilbert
Schmidt operator was given in [43] by R. Killip and B. Simon (note that µ is
assumed to have total mass 1):
¸
n
(a
n
−1/2)
2
+
¸
n
b
2
n
< ∞ (11.1)
if and only if the following conditions hold:
(i) the support of µ is [−1, 1] plus some additional mass points E
±
j
converging
to ±1,
(ii) if µ
′
is the absolutely continuous part of µ on [−1, 1], then
1
−1
(log µ
′
(t))
1 −t
2
dt > −∞,
(iii) for the mass points E
±
j
lying outside [−1, 1] we have
¸
j
[E
+
j
−1[
3/2
+
¸
j
[E
−
j
+ 1[
3/2
< ∞.
V. Totik 104
It was also shown in Killip–Simon [43] that if J −J
0
is trace class, i.e.,
¸
n
[a
n
−1/2[ +
¸
n
b
n
< ∞,
then Szeg˝ o’s condition
1
−1
log µ
′
(t)
√
1 −t
2
dt > −∞ (11.2)
holds. The conclusion is also true if µ ∈ M(0, 1/2), for the mass points E
±
j
lying outside [−1, 1] we have
¸
j
[E
+
j
−1[
1/2
+
¸
j
[E
−
j
+ 1[
1/2
< ∞,
and
limsup
n
(2
n
a
1
a
n
) > 0. (11.3)
If the support of µ is contained in [−1, 1], then Szeg˝ o’s condition automatically
holds if (11.3) is true. Actually, when supp(µ) = [−1, 1] then Szeg˝ o’s condition
(11.2) is equivalent to (11.1) and to the (conditional) convergence of the series
¸
n
(a
n
−1/2) and
¸
n
b
n
.
There is also an extended theory of orthogonal polynomials with several dif
ferent applications when the recurrence coeﬃcients do not converge, but they are
asymptotically periodic in the sense that for some k all the sequences (a
kn+j
)
∞
n=1
and (b
kn+j
)
∞
n=1
, j = 1, . . . , k converge. These are related to so called sieved or
thogonal polynomials and to orthogonal polynomials generated by polynomial
mappings. In this case the essential support of the spectral measure lies on
several intervals. There are numerous papers on this subject by M. E. H. Is
mail, N. A. AlSalam, J. A. Charris, J. Wimp, J. Bustoz, J. Geronimo, W. Van
Assche, F. Peherstorfer, R. Steinbauer, N. I. Akhiezer, B. P. Osilenker and oth
ers; see e.g. Charris–Ismail [17], Geronimo–Van Assche [31], Peherstorfer [70],
Peherstorfer–Steinbauer [73], Akhiezer [1] for details and for further references.
12 Exponential and Freud weights
These are weight functions of the form e
−2Q(x)
, where x is on the real line or
on some subinterval thereof. For simplicity we shall ﬁrst assume that Q is even.
We get Freud weights when Q(x) = [x[
α
, α > 0, x ∈ R, and Erd˝os weights
if Q tends to inﬁnity faster than any polynomial as [x[ →∞. G. Freud started
to investigate these weights in the sixties and seventies, but they independently
appeared also in the Russian literature and in statistical physics. One can
safely say that some of Freud’s problems and the work of P. Nevai and E. A.
Rahmanov were the primary cause of the sudden revitalization of the theory of
orthogonal polynomials since the early 1980’s. In the last 20 years D. Lubinsky
Orthogonal Polynomials 105
with coauthors have conducted systematic studies on exponential weights, see
e.g. Levin–Lubinsky [46, 45], Lubinsky [49], Lubinsky–Saﬀ [50], Van Assche
[94]; we should mention the names E. Levin, E. B. Saﬀ, W. Van Assche, E. A.
Rahmanov and H. N. Mhaskar. In the mid 1990’s a new stimulus came from
the Riemann–Hilbert approach that was used together with the steepest descent
method by P. Deift, T. Kriecherbauer, K. T.R. McLaughlin, S. Venakides and
X. Zhou ([20]) to give complete asymptotics when Q is analytic.
One can roughly say that because of the fast vanishing of the weight around
inﬁnity, things happen on a ﬁnite subinterval [−a
n
, a
n
] (depending on the de
gree of the polynomials), and on [−a
n
, a
n
] techniques developed for [−1, 1] are
applied. For Freud weights one can also make the substitution x → n
1/λ
x and
go to orthogonality with respect to the varying weight e
−nx
λ
, in which case
things are automatically reduced to a ﬁnite interval which is the support of a
weighted energy problem.
The a
n
are the so called MhaskarRahmanovSaﬀ numbers deﬁned by
n =
2
π
1
0
a
n
tQ
′
(a
n
t)
√
1 −t
2
dt. (12.1)
The zeros of p
n
(w
2
), w(x) = exp(−Q(x)) are spreading out and the largest zero
is very close to a
n
, which tends to ∞.
To describe the distribution of the zeros and the behavior of the polynomials
one has to make appropriate contractions. Let us consider ﬁrst the case of Freud
weight w(x) = exp(−[x[
α
), and let p
n
be the nth orthogonal polynomial with
respect to w
2
(on (−∞, ∞)). In this case
a
n
= n
1/α
γ
α
, γ
α
α
:= Γ
α
2
Γ
1
2
2Γ
α
2
+
1
2
.
Thus, for the largest zero x
n,n
we have x
n,n
/n
1/α
→ γ
α
as n → ∞, and to
describe zero distribution we divide (contract) all zeros x
n,i
by n
1/α
γ
α
. These
contracted zeros asymptotically have the Ullman distribution
dµ
w
(t)
dt
:=
α
π
1
t
u
α−1
√
u
2
−t
2
du, t ∈ [−1, 1]. (12.2)
This measure µ
w
minimizes the weighted energy
log
1
[x −t[
dµ(x)dµ(t) + 2
Qdµ (12.3)
among all probability measures compactly supported on R. It is a general
feature of exponential weights that the behavior of zeros of the polynomials is
governed by the solution of a weighted energy problem (weighted equilibrium
measures, see Saﬀ–Totik [77]). If κ
n
is the leading coeﬃcient of p
n
, i.e., p
n
(z) =
κ
n
z
n
+ , then (Lubinsky–Saﬀ [50])
lim
n→∞
κ
n
π
1/2
2
−n
e
−n/α
n
(n+1/2)/α
= 1,
V. Totik 106
and we have
lim
n→∞
[p
n
(n
1/α
γ
α
z)[
1/n
= exp
log [z +
z
2
−1[ + Re
1
0
zu
α−1
√
z
2
−u
2
du
locally uniformly outside [−1, 1]. This latter is so called nth root asymptotics,
while the former is strong asymptotics. Strong asymptotics for p
n
(z) on diﬀerent
parts of the complex plane was given using the Riemann–Hilbert approach, see
Deift [20] and Kriecherbauer–McLaughlin [44] and the references there. On the
real line we have a Plancherel–Rotach type formula
n
1/2α
p
n
(w
α
; n
1/α
γ
α
x) exp(−nγ
α
α
[x[
α
)−
−
2
π
1
4
√
1 −x
2
cos
1
2
arccos x +nπµ
w
([x, 1]) −
π
4
→0
uniformly on any subinterval of (−1, 1).
Things become more complicated for nonFreud weights, but the correspond
ing results are of the same ﬂavor. In this case the weight is not necessarily
symmetric, but under some conditions (like Q being convex or xQ
′
(x) being in
creasing for x > 0 and an analogous condition for x < 0) the relevant weighted
equilibrium measure’s support is an interval, and the deﬁnition of the Mhaskar–
Rahmanov–Saﬀ numbers a
±n
is
n =
1
π
an
a−n
xQ
′
(x)
(x −a
−n
)(a
n
−x)
dx,
0 =
1
π
an
a−n
Q
′
(x)
(x −a
−n
)(a
n
−x)
dx.
Now one solves the weighted equilibrium problem (12.3) for all measures µ with
total mass n, and if µ
n
is the solution then [a
−n
, a
n
] is the support of µ
n
and
µ
n
/n will play the role of the measure µ
w
from (12.2) above.
The weight does not even have to be deﬁned on all R, e.g. in [46] a theory
was developed by Levin and Lubinsky that simultaneously includes far reaching
generalizations of nonsymmetric Freud, Erd˝ os and Pollaczek weights such as
(a) nonsymmetric Freudtype weights
Q(x) =
[x[
α
, x ∈ [0, ∞)
[x[
β
, x ∈ (−∞, 0),
(b) nonsymmetric Erd˝ os weights such as
Q(x) =
exp
l
([x[
α
) −exp
l
(0), x ∈ [0, ∞)
exp
k
([x[
β
) −exp
k
(0), x ∈ (−∞, 0)
Orthogonal Polynomials 107
with exp
l
is the ltimes iterated exponential function, or
(c) nonsymmetric Pollaczek type weights that vanish fast at ±1 such as
Q(x) =
exp
l
((1 −x)
−α
) −exp
l
(1), x ∈ [0, 1)
exp
k
((1 −x)
−β
) −exp
k
(0), x ∈ (−1, 0).
In all cases the interval [a
−n
, a
n
] is where things happen, e.g. this is the
shortest interval on which the supremum norm of a weighted polynomial is
attained:
wP
n

sup
= wP
n

sup,[−a−n,an]
for all polynomials of degree at most n. These numbers a
±n
are everywhere in
the theory, e.g.
sup
x
[p
n
(x)[w(x)[x −a
n
[
1/4
[x −a
−n
[
1/4
∼ 1.
13 Sobolev orthogonality
In Sobolev orthogonality we consider orthogonality with respect to an inner
product
(f, g) =
r
¸
k=0
f
(k)
g
(k)
dµ
k
, (13.1)
where µ
k
are given positive measures. There are several motivations for this
kind of orthogonality. Perhaps the most natural one is smooth data ﬁtting. The
Spanish school around F. Marcell´ an, G. Lopez and A. MartinezFinkelshtein
has been particularly active in developing this area (see the surveys Marcell´ an–
Alfaro–Rezola [51] and MartinezFinkelshtein [53, 52] and the references therein).
In this section let Q
n
(z) = z
n
+ denote the monic orthogonal polyno
mial with respect to the Sobolev inner product (13.1), and q
n
(µ
k
) the monic
orthogonal polynomials with respect to the measure µ
k
.
Most arguments for the standard theory fail in this case, e.g. it is no longer
true that the zeros lie in the convex hull of the support of the measures µ
k
,
k = 0, 1, . . . , r. It is not even known if the zeros are bounded if all the measures
µ
k
have compact support. Nonetheless, for the case r = 1, and µ
0
, µ
1
∈ Reg
(see Section 9) it was shown in Gautschi–Kuijlaars [29] that the asymptotic
distribution of the zeros of the derivative Q
′
n
is the equilibrium measure ω
E0∪E1
,
where E
i
is the support of µ
i
, i = 0, 1 (which also have to be assumed to be
regular). Furthermore, if, in addition, E
0
⊆ E
1
, then the asymptotic zero
distribution of Q
n
is ω
E0
.
In general, both the algebraic and the asymptotic/analytic situation is quite
complicated, and there are essentially two important cases which have been
understood to a satisfactory degree.
Case I: The discrete case. In this case µ
0
is some “strong” measure, e.g. from the
Nevai class M(b, a) (see Section 10), and µ
1
, . . . , µ
k
are ﬁnite discrete measures.
V. Totik 108
It turns out that then the situation is similar to adding these discrete measures
to µ
0
(the new measure will also be in the same Nevai class), and considering
standard orthogonality with respect to this new measure. For ecample, if r = 1,
then
lim
n→∞
Q
n
(z)
q
n
(µ
0
+µ
1
, z)
= 1
holds uniformly on compact subsets of C ` supp(µ
0
+ µ
1
). Thus, the Sobolev
orthogonal polynomials diﬀer from those of the measure µ
0
, but not more than
what happens when adding mass points to µ
0
.
In this discrete case the Q
n
’s satisfy a higher order recurrence relation, hence
this case is also related to matrix orthogonality (see the end of the Section 16).
Case II: The Szeg˝ o case. Suppose now that µ
0
, . . . , µ
k
are all supported on
the same smooth curve or arc J, and they satisfy Szeg˝ o’s condition there (see
Section 10). In this case the kth derivative of Q
n
satisﬁes, locally uniformly in
the complement of J, the asymptotic formula
lim
n→∞
Q
(k)
n
(z)
n
k
q
n−k
(µ
r
, z)
=
1
[Φ
′
(z)]
m−k
,
where Φ is the conformal map that maps C ` J onto the complement of the
unit disk. That is, in this case the measures µ
0
, . . . , µ
r−1
do not appear in
the asymptotic formula, only µ
r
matters. The reason for this is the following:
Q = Q
n
minimizes
(Q, Q) =
r
¸
k=0
[Q
(k)
[
2
dµ
k
(13.2)
among all monic polynomials of degree n, while q = q
n−k
(µ
k
) minimizes
[q[
2
dµ
k
among all monic polynomials of degree n − k. But the polynomial Q
(k)
n
(t) =
n(n −1) (n −k + 1)t
n−k
+ is a monic polynomial times the factor n(n −
1) (n −k +1) ∼ n
k
, and this factor is dominant for k = r, so everything else
will be negligible. There are results for compensation of this n
k
factor which
lead to Sobolev orthogonality with respect to varying measures.
Under the much less restrictive assumption that µ
0
∈ Reg (see Section 9)
and the other measures µ
k
are supported in the support E of µ
0
it is true
(L´opez–PijeiraCabrera–Izquierdo [47]) that the asymptotic zero distribution of
Q
(k)
n
is the equilibrium measure ω
E
for all k,
lim
n→∞
Q
(k)
n

1/n
sup,E
= cap(E),
Orthogonal Polynomials 109
and hence, away from the zeros in the unbounded component of the complement
of E, we have
lim
n→∞
[Q
k
n
(z)[
1/n
= e
g
C\E
(z)
where g
C\E
is the Green function for this unbounded component.
The techniques developed for exponential weights and for Sobolev orthog
onality were combined in Geronimo–Lubinsky–Marcellan [30] to prove strong
asymptotics for Sobolev orthogonal polynomials when r = 1 and µ
0
= µ
1
are
exponential weights.
14 NonHermitian orthogonality
We refer to nonHermitian orthogonality in either of these cases:
• the measure µ is nonpositive or even complexvalued and we consider p
n
with
p
n
(z)z
k
dµ = 0, k = 0, 1, . . . , n −1, (14.1)
• µ is again nonpositive or complexvalued, or positive but lies on a complex
curve or arc and orthogonality is considered without complex conjugation,
i.e.,
p
n
(z)z
k
dµ = 0, k = 0, 1, . . . , n −1. (14.2)
More generally, one could consider nonpositive inner products, but we shall
restrict our attention to complex measures and orthogonality (14.2).
As an example, consider the diagonal Pad´e approximant to the Cauchy
transform
f(z) =
dµ(t)
z −t
of a signed or complexvalued measure, i.e., consider polynomials p
n
and q
n
of
degree at most n such that
f(z)p
n
(z) −q
n
(z) = O(z
−n−1
)
at inﬁnity. Then p
n
satisﬁes the nonHermitian orthogonality relation
p
n
(x)x
j
dµ(x) = 0, j = 0, 1, . . . , n −1. (14.3)
In this nonHermitian case even the GramSchmidt orthogonalization pro
cess may fail, and then p
n
is deﬁned as the solution of the orthogonality condi
tion (14.1), resp. (14.2), which give a system of homogeneous equations for the
V. Totik 110
coeﬃcients of p
n
. Thus, p
n
may have smaller degree than n, and things can get
pretty wild with this kind of orthogonality. For example, in the simple case
dµ(x) = (x −cos πα
1
)(x −cos πα
2
)(1 −x
2
)
−1/2
dx, x ∈ [−1, 1],
with 0 < α
1
< α
2
< 1 rationally independent algebraic numbers, the zeros
of p
n
from (14.3) are dense on the whole complex plane (compare this with
the fact that for positive µ all zeros lie in [−1, 1]). In Stahl [86] it was shown
that it is possible to construct a complex measure µ on [−1, 1], such that for an
arbitrary prescribed asymptotic behavior some subsequence ¦p
n
k
¦ will have this
zero behavior. Nonetheless, the asymptotic distribution of the zeros is again the
equilibrium distribution of the support of µ under regularity conditions on µ.
For example, this is the case if
• [µ[ belongs to the Reg class (see Section 9), and the argument of µ, i.e.,
dµ(t)/d[µ[(t), is of bounded variation (Baratchart–K¨ ustner–Totik [12]), or
• dµ(x) = g(arccos x)(1 − x
2
)
−1/2
dx, x ∈ [−1, 1], g is bounded away from
zero and inﬁnity, and satisﬁes [g(θ + δ) −g(θ)[ ≤ K[ log δ[
−1−δ
, or
• µ is supported on ﬁnitely many intervals, the argument of µ is uniformly
continuous and for a(δ) = inf
x∈supp(µ)
[µ[([x − δ, x + δ]) the property
lim
δ→0
log a(δ) = 0 holds (Stahl [86]).
In [84]–[86] H. Stahl obtained asymptotics for nonHermitian orthogonal poly
nomials even for varying measures and gave several applications of them to Pad´e
approximation. When the measure µ is of the form dµ(x) = g(x)(1−x
2
)
−1/2
dx,
x ∈ [−1, 1], with an analytic g, for z ∈ C` [−1, 1], a strong asymptotic formula
of the form
p
n
(z)
κ
n
= (1 +o(1))
(z +
√
z
2
−1)
n
2
n
D
µ
(z)
−1
exp
1
2π
1
−1
log µ
′
(t)
√
1 −t
2
dt
(with D
µ
the Szeg˝ o function (10.2)) was proved by J. Nuttall [66], [67], A. A.
Gonchar and S. P. Suetin [38]. For a recent Riemann–Hilbert approach see
the paper [10] by A. I. Aptekarev and W. Van Assche. A similar result holds
on the support of the measure, as well as for the case of varying weights, see
Aptekarev–Van Assche [10].
15 Multiple orthogonality
Multiple orthogonality comes from simultaneous Pad´e approximation. It is a
relatively new area where we have to mention the names of E. M. Nikishin,
V. N. Sorokin, A. A. Gonchar and E. A. Rahmanov, A. I. Aptekarev, A. B.
J. Kuijlaars, J. Geronimo and W. Van Assche (see the survey [95] by W. Van
Assche and the references there and the paper Gonchar–Rakhmanov [37]). The
Orthogonal Polynomials 111
analogues of many classical concepts and properties have been found, and also
the analogues of the classical orthogonal polynomials are known, e.g. in the
multiple Hermite case the measures are dµ(x) = e
−x
2
+cjx
dx.
Asymptotic behavior of multiple orthogonal polynomials is not fully under
stood yet due to the interaction of the diﬀerent measures. For the existing
results see Aptekarev [8], Van Assche [95], Van Assche’s Chapter 23 in [39] and
the references therein.
Types and normality
On R let there be given r measures µ
1
, . . . , µ
r
with ﬁnite moments and inﬁnite
support, and consider multiindices n = (n
1
, . . . , n
r
) of nonnegative integers
with norm [n[ = n
1
+ + n
r
. There are two types of multiple orthogonality
corresponding to the appropriate HermitePad´e approximation.
In type I multiple orthogonality we are looking for polynomials Q
n,j
of
degree n
j
−1 for each j = 1, . . . , r, such that
r
¸
j=1
x
k
Q
n,j
(x)dµ
j
(x) = 0, k = 0, 1, . . . , [n[ −2.
These orthogonality relations give [n[ −1 homogeneous linear equations for the
[n[ coeﬃcients of the r polynomials Q
n,j
, so there is a nontrivial solution. If the
rank of the system is [n[ −1, then the solution is unique up to a multiplicative
factor, in which case the index n is called normal. This happens precisely if
each Q
n,j
is of exact degree n
j
−1.
In type II multiple orthogonality we are looking for a single polynomial
P
n
of degree [n[ such that
x
k
P
n
(x)dµ
1
(x) = 0, k = 1, . . . , n
1
−1
.
.
.
x
k
P
n
(x)dµ
r
(x) = 0, k = 1, . . . , n
r
−1.
These are [n[ homogeneous linear equations for the [n[ + 1 coeﬃcients of P
n
,
and again if the solution is unique up to a multiplicative constant, then n is
called normal. This is again equivalent to P
n
being of exact degree n.
n is normal for type I orthogonality precisely when it is normal for type
II, so we just speak of normality. This is the case, for example, if the µ
j
’s
are supported on intervals [a
j
, b
j
] that are disjoint except perhaps for their
endpoints; in fact, in this case P
n
has n
j
simple zeros on (a
j
, b
j
). Normality
also holds if dµ
j
= w
j
dµ with a common µ supported on some interval [a, b],
and for all m
j
≤ n
j
, j = 1, . . . , r, every nontrivial linear combination of the
functions
w
1
(x), xw
1
(x), . . . , x
m1−1
w
1
(x), w
2
(x), xw
2
(x), . . . , x
mr−1
w
r
(x)
V. Totik 112
has at most m
1
+ + m
r
− 1 zeros on [a, b] (this means that these functions
form a so called Chebyshev system there). In this case P
n
has [n[ −1 zeros on
[a, b].
Recurrence formulae
To describe recurrence formulae, let e
j
= (0, . . . , 1, . . . , 0) where the single 1
entry is at position j. Under the normality assumption if P
n
is the monic
orthogonal polynomial, then for any k
xP
n
(x) = P
n+e
k
(x) +a
n,0
P
n
(x) +
r
¸
j=1
a
n,j
P
n−ε
j
(x).
Another recurrence formula is
xP
n
(x) = P
n+e
k
(x) +b
n,0
P
n
(x) +
r
¸
j=1
b
n,j
P
n−ε
π(1)
−···−e
π(j)
(x),
where π(1), . . . , π(r) is an arbitrary, but ﬁxed, permutation of 1, 2, . . . , r. The
orthogonal polynomials with diﬀerent indices are strongly related to one an
other, e.g. P
n+e
k
(x) −P
n+e
l
(x) is a constant multiple of P
n
(x).
If dµ
j
= w
j
dµ, then similar recurrence relations hold in case of type I or
thogonality for
Q
n
(x) =
r
¸
j=1
Q
n,j
(x)w
j
(x).
Also, type I and type II are related by a biorthogonality property:
P
n
Q
m
dµ = 0
except for the case when m = n+e
k
for some k, and then the previous integral
is not zero (under the normality condition).
To describe an analogue of the ChristoﬀelDarboux formula let ¦m
j
¦ be
a sequence of multiindices such that m
0
is the identically 0 multiindex, and
m
j+1
coincides with m
j
except for one component which is 1 larger than the
corresponding component of m
j
. Set P
j
= P
m
j
, Q
j
= Q
m
j+1
and with m = m
n
h
(j)
m
:=
P
m
(x)x
(m)j
dµ
j
(x),
where (m)
j
denotes the jth component of the multiindex m. Then (see Daems–
Kuijlaars [19]), again with m = m
n
,
(x −y)
n−1
¸
k=0
P
k
(x)Q
k
(y) = P
m
(x)Q
m
(y) −
r
¸
j=1
h
(j)
m
h
(j)
m−e
j
P
m−e
j
(x)Q
m+e
j
(y).
Thus, the left hand side depends only on m = m
n
and not on the particular
choice of the sequence m
j
leading to it.
Orthogonal Polynomials 113
The Riemann–Hilbert problem
There is an approach (see Van Assche–Geronimo–Kuijlaars [96]) to both types
of multiple orthogonality in terms of matrixvalued Riemann–Hilbert problem
for (r + 1) (r + 1) matrices Y = (Y
ij
(z))
r
i,j=0
.
If dµ
j
(x) = w
j
dx, then one requires that
• Y is analytic on C` R,
• if Y
±
(x) denote the limit of Y (z) as z → x ∈ R from the upper, respec
tively the lower, half plane, then we have Y
+
(x) = Y
−
(x)S(x), where
S(x) :=
1 w
1
(x) w
2
(x) w
r
(x)
0 1 0 0
0 0 1 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 0 1
¸
¸
¸
¸
¸
¸
¸
,
• as z →∞
Y (z) =
I +O
1
z
z
n
0 0 0
0 z
−n1
0 0
0 0 z
−n2
0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 0 z
−nr
¸
¸
¸
¸
¸
¸
¸
.
The ﬁrst entry Y
11
(z) is precisely the orthogonal polynomial P
n
of type II, and
the other entries are also explicit in terms of the P
n
’s and w
j
’s (all other entries
are either a constant multiple of P
n−e
k
or a Cauchy transform of its multiple
with w
j
). For type I orthogonality the transfer matrix is
S(x) :=
1 0 0 0
−w
1
(x) 1 0 0
−w
2
(x) 0 1 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
−w
r
(x) 0 0 1
¸
¸
¸
¸
¸
¸
¸
,
the behavior at inﬁnity is of the form
Y (z) =
I +O
1
z
z
−n
0 0 0
0 z
n1
0 0
0 0 z
n2
0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 0 z
nr
¸
¸
¸
¸
¸
¸
¸
,
and the multiple orthogonal polynomials Q
n,j
are Y
1,j+1
/2πi.
V. Totik 114
16 Matrix orthogonal polynomials
In the last 20 years the fundamentals of matrix orthogonal polynomials have
been developed mainly by A. Dur´ an and his coauthors (see also the work [9] by
A. I. Aptekarev and E. M. Nikishin). The theory shows many similarities with
the scalar case, but there is an unexpected richness which is still to be explored.
For all the results in this section see L´ opezRodriguez–Dur´an [48] and Dur´ an–
Gr¨ unbaum [25] and the numerous references there.
Matrix orthogonal polynomials
An N N matrix
P(t) =
¸
¸
p
11
(t) p
1N
(t)
.
.
.
.
.
.
.
.
.
p
N1
(t) p
NN
(t)
¸
with polynomial entries p
ij
(t) of degree at most n is called a matrix polyno
mial of degree at most n. Alternatively, one can write
P(t) = C
n
t
n
+ +C
0
with numerical matrices C
n
, . . . , C
0
of size N N.
The number t = a is called a zero of P if P(a) is singular, and the mul
tiplicity of a is the multiplicity of a as a zero of detP(a). When the leading
coeﬃcient matrix C
n
is nonsingular, then P has nN zeros counting multi
plicity.
From now on we ﬁx the dimension to be N, but the degree n can be any
natural number. I will denote the N N unit matrix and 0 stands for all kinds
of zeros (numerical or matrix).
A matrix
W(t) =
¸
¸
µ
11
(t) µ
1N
(t)
.
.
.
.
.
.
.
.
.
µ
N1
(t) µ
NN
(t)
¸
of complex measures deﬁned on (or part of) the real line is positive deﬁnite
if for any Borel set E the numerical matrix W(E) is positive semideﬁnite. We
assume that all moments of W are ﬁnite. With such a matrix we can deﬁne a
matrix inner product on the space of N N matrix polynomials via
(P, Q) =
P(t)dW(t)Q
∗
(t),
and if (P, P) is nonsingular for any P with nonsingular leading coeﬃcient, then
just as in the scalar case one can generate a sequence ¦P
n
¦
∞
n=0
of matrix poly
nomials of degree n = 0, 1, . . . which are orthonormal with respect to W:
P
n
(t)dW(t)P
∗
m
(t) =
0 if n = m
I if n = m,
Orthogonal Polynomials 115
and here P
n
has nonsingular leading coeﬃcient matrix. The sequence ¦P
n
¦ is
determined only up to left multiplication by unitary matrices, i.e., if U
n
are
unitary matrices, then the polynomials U
n
P
n
also form an orthonormal system
with respect to W.
Threeterm recurrence and quadrature
Just as in the scalar case, these orthogonal polynomials satisfy a threeterm
recurrence relation
tP
n
(t) = A
n+1
P
n+1
(t) +B
n
P
n
(t) +A
∗
n
P
n−1
(t), n ≥ 0, (16.1)
where A
n
are nonsingular matrices, and B
n
are Hermitian. Conversely, the
analogue of Favard’s theorem is also true: if a sequence of matrix polynomials
¦P
n
¦ of corresponding degree n = 0, 1, 2, . . ., satisfy (16.1) with nonsingular A
n
and Hermitian B
n
, then there is a positive deﬁnite measure matrix W such that
the P
n
are orthonormal with respect to W.
The threeterm recurrence formula easily yields the ChristoﬀelDarboux for
mula:
(w −z)
n−1
¸
k=0
P
∗
k
(z)P
k
(w) = P
∗
n−1
(z)A
n
P
n
(w) −P
∗
n
(z)A
∗
n
P
n−1
(w),
from which for example it follows that
P
∗
n−1
(z)A
n
P
n
(z) −P
∗
n
(z)A
∗
n
P
n−1
(z) = 0,
n−1
¸
k=0
P
∗
k
(z)P
k
(z) = P
∗
n−1
(z)A
n
P
′
n
(z) −P
∗
n
(z)A
∗
n
P
′
n−1
(z).
The orthogonal polynomials Q
n
of the second kind
Q
n
(t) =
P
n
(t) −P
n
(x)
t −x
dW(x), n = 1, 2, . . . ,
also satisfy the same recurrence and are orthogonal with respect to some other
matrix measure. For them we have
P
∗
n−1
(t)A
n
Q
n
(t) −P
∗
n
(z)A
∗
n
Q
n−1
(t) ≡ I,
and
Q
n
(t)P
∗
n−1
(t) −P
n
(t)Q
∗
n−1
(t) ≡ A
−1
n
.
With the recurrence coeﬃcient matrices A
n
, B
n
one can form the block
Jacobi matrix
J =
¸
¸
¸
¸
¸
¸
B
0
A
0
0 0
A
∗
0
B
1
A
1
0
0 A
∗
1
B
2
A
2
0 0 A
∗
2
B
2
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
¸
.
V. Totik 116
The zeros of P
n
are real and they are the eigenvalues (with the same multiplicity)
of the Ntruncated block Jacobi matrix (which is of size nN). If a is a zero
then its multiplicity p is at most N, the rank of P
n
(a) is N − p, and the space
of those vectors v for which P
n
(a)v = 0 is of dimension p. If we write x
n,k
,
1 ≤ k ≤ m, for the diﬀerent zeros of P
n
, and l
k
is the multiplicity of x
n,k
, then
the matrices
Γ
k
=
1
(det(P
n
(t)))
(l
k
)
(x
n,k
)
(Adj(P
n
(t)))
(l
k
−1)
(x
n,k
)Q
n
(x
n,k
), 1 ≤ k ≤ m
are positive semideﬁnite of rank l
k
, and with them the matrix quadrature
formula
P(t)dW(t) =
m
¸
k=1
P(x
n,k
)Γ
n,k
holds for every matrix polynomial P of degree at most 2n −1.
If we assume that A
n
→A, B
n
→B where A is nonsingular, then
P
n
(z)P
−1
n−1
(z)A
−1
n
→
dW
A,B
(t)
z −t
locally uniformly outside the cluster set of the zeros, where W
A,B
(t) is the mea
sure matrix of orthogonality for the sequence of matrix orthogonal polynomials
S
n
with recurrence coeﬃcients A, B for all n, i.e., which satisfy the threeterm
recurrence
tS
n
(t) = A
∗
S
n+1
(t) +BS
n
(t) +AS
n−1
(t).
The distribution of the zeros themselves will be 1/Ntimes the trace of the ma
trix measure of orthogonality for another sequence of matrix orthogonal poly
nomials R
n
satisfying
tR
n
(t) = AR
n+1
(t) +BR
n
(t) +AR
n−1
(t), n ≥ 2,
with appropriate modiﬁcations for n < 2.
Families of orthogonal polynomials
If the matrix of orthogonality is diagonal (or similar to a diagonal matrix) with
diagonal entries µ
i
, then the orthogonal matrix polynomials are also diagonal
with ith entry equal to p
n
(µ
i
), the nth orthogonal polynomial with respect
to µ
i
. Many matrix orthogonal polynomials in the literature can be reduced to
this scalar case. Recently however, some remarkably rich nonreducible families
have been obtained by A. Duran and F. Gr¨ unbaum (see [25] and the references
therein), which may play the role of the classical orthogonal polynomials in
higher dimension. They found families of matrix orthogonal polynomials that
satisfy second order (matrix) diﬀerential equations just like the classical orthog
onal polynomials. Their starting point was a symmetry property between the
Orthogonal Polynomials 117
orthogonality measure matrix and a second order diﬀerential operator. They
worked out several explicit examples. Here is one of them: N = 2, the measure
matrix (more precisely its density) is
H(t) := e
−t
2
1 +[a[
2
t
4
at
2
at
2
1
, t ∈ R,
where a ∈ C` ¦0¦ is a free parameter. The corresponding P
n
(t) satisﬁes
P
′′
N
(t) +P
′
n
(t)
−2t 4at
0 −2t
+ P
n
(t)
−4 2a
0 0
=
−2n −4 2a(2n + 1)
0 −2n
P
n
(t).
There is an explicit Rodrigues’ type representation for the polynomials them
selves, and the threeterm recurrence (16.1) holds with B
n
= 0,
A
n+1
:=
n + 1
2
γ
n+3
/γ
n+2
aγ
n+2
γ
n+1
0 γ
n
/γ
n+1
,
where
γ
2
n
:= 1 +
[a[
2
2
n
2
.
Connection with higher order scalar recurrence
Matrix orthogonality is closely connected to (2N + 1)term recurrences for
scalar polynomials. To describe this we need the following operators on poly
nomials p: if p(t) =
¸
k
a
k
t
k
, then
R
N,m
(p) =
¸
s
a
sN+m
t
s
,
i.e., from a polynomial the operator R
N,m
takes those powers where the expo
nent is congruent to m modulo N, removes the common factor t
m
and changes
t
N
to t.
Now suppose that ¦p
n
¦
∞
n=0
is a sequence of scalar polynomials of correspond
ing degree n = 0, 1, . . ., and suppose that this sequence satisﬁes a (2N+1)term
recurrence relation
t
N
p
n
(t) = c
n,0
p
n
(t) +
N
¸
k=1
(c
n,k
p
n−k
(t) +c
n+k,k
p
n+k
(t)),
V. Totik 118
where c
n,0
is real, c
n,N
= 0 (and p
k
(t) ≡ 0 for k < 0). Then
P
n
(t) =
¸
¸
¸
¸
R
N,0
(p
nN
) R
N,N−1
(p
nN
)
R
N,0
(p
nN+1
) R
N,N−1
(p
nN
+ 1)
.
.
.
.
.
.
.
.
.
R
N,0
(p
nN+N−1
) R
N,N−1
(p
nN
+N −1)
¸
is a sequence of matrix orthogonal polynomials with respect to a positive deﬁnite
measure matrix. Conversely, if P
n
= (P
n,m,j
)
N−1
m,j=0
is a sequence of orthonormal
matrix polynomials, then the scalar polynomials
p
nN+m
(t) =
N−1
¸
j=0
t
j
P
n,m,j
(t
N
), 0 ≤ m < N, n = 0, 1, 2, . . . ,
satisfy a (2N + 1)recurrence relation of the above form.
References
[1] N. I. Ahiezer, Orthogonal polynomials on several intervals, Dokl. Akad.
Nauk SSSR, 134(1960), 9–12. (Russian); translated as Soviet Math. Dokl.,
1(1960), 989–992.
[2] N. I. Akhiezer, The Classical Moment Problem and Some Related Questions
in Analysis, Oliver and Boyd, Edinburgh, 1965
[3] W. A. AlSalam, Characterization theorems for orthogonal polynomials,
in: Orthogonal Polynomials: Theory and Practice, M. Ismail and P. Nevai
(eds), NATO ASI Series C, 294, Kluwer, Dordrecht, 1990.
[4] M. P. Alfaro, M. Bello, J. M. Montaner and J. L. Varona, Some asymptotic
properties for orthogonal polynomials with respect to varying measures, J.
Approx. Theory, 135(2005), 22–34.
[5] M. P. Alfaro and L. Vigil, Solution of a problem of P. Tur´an, J. Approx.
Theory, 53(1988), 195–197.
[6] G. E. Andrews, R. Askey and R. Roy, Special Functions, Encyclopedia of
Mathematics and its Applications, Cambridge University Press, New York,
1999.
[7] A. I. Aptekarev, Asymptotic properties of polynomials orthogonal on a
system of contours, and periodic motions of Toda chains. (Russian) Mat.
Sb. (N.S.), 125(167)(1984), 231–258.
[8] A. I. Aptekarev, Multiple orthogonal polynomials, J. Comput. Appl. Math.,
99(1988), 423–447.
Orthogonal Polynomials 119
[9] A. I. Aptekarev and E. M. Nikishin, The scattering problem for a discrete
SturmLiouville operator, Mat. Sb. (N.S.), 121(1983), 327–358. (Russian)
[10] A. I. Aptekarev and W. Van Assche, Scalar and matrix Riemann–Hilbert
approach to the strong asymptotics of Pad´e approximants and complex or
thogonal polynomials with varying weights, J. Approx. Theory, 129(2004),
129–166.
[11] R. Askey, A transplantation theorem for Jacobi series. Illinois J. Math.,
13(1969), 583–590.
[12] L. Baratchart, R. K¨ ustner and V. Totik, Zero distribution via orthogonality,
Annales de l’Institut Fourier, 55(2005), 14551499.
[13] D. Barrios and G. L´ opez, Ratio asymptotics for polynomials orthogonal on
arcs of the unit circle, Constr. Approx., 15(1999), 1–31.
[14] M. Bello and G. L´ opez, Ratio and relative asymptotics of polynomials
orthogonal on an arc of the unit circle, J. Approx. Theory, 92(1998), 216–
244.
[15] O. Blumenthal,
¨
Uber die Entwicklung einer willk¨ urlichen Funktion nach
den Nennern des Kettenbruches f¨ ur
0
−∞
[φ(ξ)/(z − ξ)] dξ, Inaugural Dis
sertation, G¨ottingen, 1898.
[16] S. Bochner,
¨
Uber SturmLiouvillesche Polynomsysteme, Math. Zeitschrift,
29(1929), 730–736.
[17] J. A. Charris and M. E. H. Ismail, Sieved orthogonal polynomials. VII.
Generalized polynomial mappings, Trans. Amer. Math. Soc., 340(1993),
71–93.
[18] T. S. Chihara, An introduction to orthogonal polynomials. Mathematics
and its Applications, Vol. 13, Gordon and Breach Science Publishers, New
YorkLondonParis, 1978.
[19] E. Daems and A. B. J. Kuijlaars, A ChristoﬀelDarboux formula for mul
tiple orthogonal polynomials, J. Approx. Theory, 130(2004), 190–202.
[20] P. Deift, Orthogonal Polynomials and Random Matrices: A Riemann–
Hilbert Approach, Courant Lecture Notes, 1999.
[21] F. Delyon, B. Simon, and B. Souillard, From power pure point to con
tinuous spectrum in disordered systems, Annales Inst. Henri Poincar´e,
42(1985), 283–309.
[22] S. Denisov, On Rakhmanov’s theorem for Jacobi matrices, Proc. Amer.
Math. Soc., 132(2004), 847–852.
V. Totik 120
[23] S. Denisov and B. Simon, Zeros of orthogonal polynomials on the real line,
J. Approx. Theory, 121(2003), 357–364.
[24] C. F. Dunkl and Y. Xu, Orthogonal polynomials of several variables, Ency
clopedia of Mathematics and its Applciations, Cambridge University Press,
New York, 2001.
[25] A. J. Dur´ an and F. A. Gr¨ unbaum, A survey on orthogonal matrix poly
nomials satisfying second order diﬀerential equations, J. Comput. Appl.
Math., 178(2005), 169190.
[26] A. S. Fokas, A. R. Its and A. V. Kitaev, The isomonodromy approach to
matrix models in 2D quantum gravity, Comm. Math. Phys., 147(1992),
395–430.
[27] G. Freud, Orthogonal Polynomials. Akad´emiai Kiad´ o/Pergamon Press, Bu
dapest, 1971.
[28] W. Gautschi, Orthogonal Polynomials: Computation and Approximation,
Numerical Mathematics and Scientiﬁc Computation, Oxford University
Press, New York, 2004.
[29] W. Gautschi and A. B. J. Kuijlaars, Zeros and critical points of Sobolev
orthogonal polynomials, J. Approx. Theory, 91(1997), 117–137.
[30] J. Geronimo, D. S. Lubinsky and F. Marcellan, Asymptotics for Sobolev
orthogonal polynomials for exponential weights, to appear in Constructive
Approx.
[31] J. S. Geronimo and W. Van Assche, Orthogonal polynomials on several
intervals via a polynomial mapping, Trans. Amer. Math. Soc., 308(1988),
559–581.
[32] Ya. L. Geronimus, On the character of the solution of the moment problem
in the case of the periodic in the limit associated fraction, Bull. Acad. Sci.
USSR Math., 5(1941), 203–210. [Russian]
[33] Ya. L. Geronimus, On polynomials orthogonal on the circle, on trigono
metric moment problem, and on allied Carath´eodory and Schur functions,
Mat. Sb., 15(1944), 99–130. [Russian]
[34] Ya. L. Geronimus, On the trigonometric moment problem, Ann. of Math.,
47(1946), 742–761.
[35] Ya. L. Geronimus, Orthogonal polynomials on the unit circle and their
applications, Amer. Math. Soc. Translation, 1954, Vol. 104.
[36] L. Golinskii and V. Totik, Orthogonal polynomials from Stieltjes to Simon,
Proc. Conf. honoring B. Simon, AMS Conference Series (to appear).
Orthogonal Polynomials 121
[37] A. A. Gonchar and E. A. Rakhmanov, On the convergence of simultane
ous Pad´e approximants for systems of functions of Markov type, Num
ber theory, mathematical analysis and their applications. Trudy Mat. Inst.
Steklov., 157(1981), 31–48. (Russian)
[38] A. A. Gonchar and S. P. Suetin, On Pad´e approximation of Markov type
meromorphic functions, Current Problems in Mathematics, 5., Rossiiskaya
Akademiya Nauk, Matematicheskii Institut im. V. A. Steklova, Moscow,
2004. 68 pp. (electronic). (Russian)
[39] M. E. H. Ismail, Classical and quantum orthogonal polynomials in one vari
able, Encyclopedia in Mathematics, Cambridge University Press, 2005.
[40] V. Kaliaguine, A note on the asymptotics of orthogonal polynomials on
a complex arc: the case of a measure with a discrete part. , J. Approx.
Theory, 80(1995), 138–145.
[41] S. Khrushchev, A singular Riesz product in the Nevai class and inner func
tions with the Schur parameters in ∩
p>2
ℓ
p
, J. Approx. Theory, 108(2001),
249–255.
[42] S. Khrushchev, Continued fractions and orthogonal polynomials on the
unit circle, Proc. Conf. OPSFA, Copenhagen, 2003.
[43] R. Killip and B. Simon, Sum rules for Jacobi matrices and their applications
to spectral theory, Annals of Math., 158(2003), 253321
[44] T. Kriecherbauer and K. TR. McLaughlin, Strong asymptotics of polyno
mials orthogonal with respect to a Freud weight, Internat. Math. Research
Notices, 6(1999), 299–333.
[45] A. Levin and D. S. Lubinsky, Christoﬀel Functions and Orthogonal Poly
nomials for Exponential Weights on [−1, 1], Memoirs Amer. Math. Soc. ,
535, (1994)
[46] A. Levin and D. S. Lubinsky, Orthogonal polynomials for exponential
weights, CMS Books in Mathematics/Ouvrages de Mathmatiques de la
SMC, 4, SpringerVerlag, New York, 2001.
[47] G. L´ opez, H. PijeiraCabrera and I. P. Izquierdo, Sobolev orthogonal poly
nomials in the complex plane, J. Comput. Appl. Math., 127(2001), 219230.
[48] P. L´ opezRodriguez and A. J. Dur´ an, Orthogonal matrix polynomials,
Laredo Lecture Notes (Lecture notes of the 1st SIAG Summer School on
Orthogonal Polynomials and Special Functions, Laredo, Spain, July 2429
2000). Nova Science Publisher (2003). R.
´
AlvarezNodarse, F. Marcell´ an
and W. Van Assche (Eds. )
V. Totik 122
[49] D. S. Lubinsky, Strong Asymptotics for Extremal Errors Associated with
Erd¨ ostype Weights, Pitman Res. Notes, Vol. 202, Longmans, Nurn Mill,
Essex, 1989.
[50] D. S. Lubinsky and E. B. Saﬀ, Strong Asymptotics for Extremal Polyno
mials Associated with Weights on (−∞, ∞), Springer Lecture Notes, Vol.
1305, Springer Verlag, Berlin, 1988.
[51] F. Marcell´ an, M. Alfaro and M. L. Rezola, Orthogonal polynomials on
Sobolev spaces: old and new directions, J. Comput. Appl. Math., 48(1993),
113–131.
[52] A. MartinezFinkelshtein, Analytic aspects of Sobolev orthogonal polyno
mials, J. Comput. Appl. Math., 99(1998), 491–510.
[53] A. MartinezFinkelshtein, Analytic properties of Sobolev orthogonal poly
nomials revisited, J. Comput. Appl. Math., 127(2001), 255266.
[54] A. MartinezFinkelshtein, Szeg˝ o polynomials: a view from the Riemann–
Hilbert window, preprint arXiv:math. CA/0508117, 2005.
[55] G. Mastroianni and G. Milovanovic, Approximation of Functions and Func
tionals with Application, World Scientiﬁc, to appear.
[56] G. Mastroianni and V. Totik, Polynomial inequalities with doubling and
A
∞
weights, Constructive Approx., 16(2000), 37–71.
[57] A. M´ at´e, P. Nevai and V. Totik, Extensions of Szeg˝ o’s theory of orthogonal
polynomials, Constructive Approx., 3(1987), 51–72.
[58] A. M´ at´e, P. Nevai and V. Totik, Strong and weak convergence of orthogonal
polynomials, Amer. J. Math., 109(1987), 239–282.
[59] A. M´ at´e, P. Nevai and V. Totik, Szeg˝ o’s extremum problem on the unit
circle, Annals of Math., 134(1991), 433–453.
[60] M. Mehta, Random matrices, second ed., Academic Press, Inc. , Boston,
1991.
[61] H. N. Mhaskar and E. B. Saﬀ, On the distribution of zeros of polynomials
orthogonal on the unit circle, J. Approx. Theory, 63(1990), 30–38.
[62] P. Nevai, Orthogonal Polynomials, Memoirs Amer. Math. Soc., 213(1979)
[63] P. Nevai and V. Totik, Orthogonal polynomials and their zeros, Acta Sci.
Math.(Szeged), 53(1989), 99–104.
[64] P. Nevai and V. Totik, Denisov’s theorem on recurrent coeﬃcients, J. Ap
prox. Theory, 127(2004), 240–245.
Orthogonal Polynomials 123
[65] E. M. Nikishin, An estimate for orthogonal polynomials, Acta Sci.
Math.(Szeged), 48(1985), 395–399. [Russian]
[66] J. Nuttall, Asymptotics for diagonal HermitePad´e approximants, J. Ap
prox. Theory, 42(1984), 299–386.
[67] J. Nuttall, Pad´e polynomial asymptotics from a singular integral equation,
Constructive Approx., 6(1990), 157–166.
[68] L. Pastur and A. Figotin, Spectra of Random and AlmostPeriodic Opera
tors, Grundlehren der mathematischen Wissenschaften, Vol. 297, Springer
Verlag, Berlin, 1992.
[69] F. Peherstorfer, On BernsteinSzeg˝ o orthogonal polynomials on several in
tervals, SIAM J. Math. Anal., 21(1990), 461–482.
[70] F. Peherstorfer, On BernsteinSzeg˝ o orthogonal polynomials on several in
tervals. II. Orthogonal polynomials with periodic recurrence coeﬃcients, J.
Approx. Theory, 64(1991), 123–161.
[71] F. Peherstorfer, Orthogonal and extremal polynomials on several intervals,
J. Comput. Appl. Math., 48(1993), 187–205.
[72] F. Peherstorfer, Zeros of polynomials orthogonal on several intervals. Int.
Math. Res. Notes, 7(2003), 361–385.
[73] F. Peherstorfer and R. Steinbauer, On polynomials orthogonal on several
intervals, Special functions (Torino, 1993), Ann. Numer. Math., 2(1995),
353–370.
[74] E. A. Rakhmanov, On V. A. Steklov’s problem in the theory of orthogonal
polynomials, (Russian) Dokl. Akad. Nauk SSSR, 254(1980), 802–806.
[75] E. A. Rakhmanov, On the asymptotics of the ratio of orthogonal polyno
mials, Math. USSR Sbornik, 46(1983), 105–117; Russian original, Mat. Sb.,
118(1982), 104–117.
[76] E. B. Saﬀ and V. Totik, What parts of a measure’s support attract zeros
of the corresponding orthogonal polynomials? Proc. Amer. Math. Soc.,
114(1992), 185–190.
[77] E. B. Saﬀ and V. Totik, Logarithmic Potentials with External Fields,
Grundlehren der mathematischen Wissenschaften, 316, SpringerVerlag,
Berlin, 1997.
[78] B. Simon, The classical moment problem as a selfadjoint ﬁnite diﬀerence
operator, Advances in Math., 137(1998), 82–203.
V. Totik 124
[79] B. Simon, Ratio asymptotics and weak asymptotic measures for orthogonal
polynomials on the real line, J. Approx. Theory, 126(2004), 198–217.
[80] B. Simon, Orthogonal Polynomials on the Unit Circle, V. 1: Classical The
ory, AMS Colloquium Series, American Mathematical Society, Providence,
RI, 2005.
[81] B. Simon, Orthogonal Polynomials on the Unit Circle, V. 2: Spectral The
ory, AMS Colloquium Series, American Mathematical Society, Providence,
RI, 2005.
[82] B. Simon and A. Zlatoˇs, Sum rules and the Szeg˝ o condition for orthogonal
polynomials on the real line, Comm. Math. Phys., 242(2003), 393–423.
[83] B. Simon and A. Zlatoˇs, Higherorder Szeg˝ o theorems with two singular
points, J. Approx. Theory, 134(2005), 114–129.
[84] H. Stahl, Divergence of diagonal Pad´e approximants and the asymptotic
behavior of orthogonal polynomials associated with nonpositive measures,
Constr. Approx., 1(1985), 249–270.
[85] H. Stahl, Orthogonal polynomials with complexvalued weight function. I,
II. Constr. Approx., 2(1986), 225–240, 241–251.
[86] H. Stahl, Orthogonal polynomials with respect to complexvalued mea
sures. Orthogonal polynomials and their applications (Erice, 1990), IMACS
Ann. Comput. Appl. Math., 9(1991), 139–154.
[87] H. Stahl and V. Totik, General Orthogonal Polynomials, Encyclopedia of
Mathematics and its Applications, 43, Cambridge University Press, New
York, 1992.
[88] S. P. Suetin, On the uniform convergence of diagonal Pad´e approximants for
hyperelliptic functions, Mat. Sb., 191(2000), 81–114; (Russian), translation
in Sb. Math., 191(2000), 1339–1373.
[89] S. P. Suetin, Pad´e approximants and the eﬀective analytic continuation of a
power series, Uspekhi Mat. Nauk, 57(2002), 45–142; (Russian), translation
in Russian Math. Surveys, 57(2002), 43–141.
[90] J. Szabados and P. V´ertesi, Interpolation of Functions, World Scientiﬁc,
Singapore, 1990.
[91] G. Szeg˝ o, Orthogonal Polynomials, Coll. Publ. , XXIII, Amer. Math. Soc.
, Providence, 1975.
[92] V. Totik, Orthogonal polynomials with ratio asymptotics, Proc. Amer.
Math. Soc., 114(1992), 491–495.
Orthogonal Polynomials 125
[93] V. Totik, Asymptotics for Christoﬀel functions for general measures on the
real line, J. D’Analyse Math., 81(2000), 283–303.
[94] W. Van Assche, Asymptotics for Orthogonal Polynomials, Spriger Lecture
Notes, Vol. 1265, Springer Verlag, Berlin, 1987.
[95] W. Van Assche, HermitePad´e approximation and multiple orthogonal
polynomials, preprint.
[96] W. Van Assche, J. S. Geronimo, A. B. J. Kuijlaars, Riemann–Hilbert prob
lems for multiple orthogonal polynomials, in: J. Bustoz et al. (eds), Special
Functions 2000: Current Perspectives and Future Directions, Kluwer, Dor
drecht, 2001, 23–59.
[97] H. Widom, Extremal polynomials associated with a system of curves in the
complex plane, Adv. Math., 3(1969), 127–232.
Bolyai Institute, University of Szeged
Szeged
Aradi v. tere 1, 6720
Hungary, and
Department of Mathematics
University of South Florida
Tampa, FL, 33620
USA
Orthogonal Polynomials 7 Some questions leading to classical orthogonal Electrostatics . . . . . . . . . . . . . . . . . . . Polynomial solutions of eigenvalue problems . . Harmonic analysis on spheres and balls . . . . Approximation theory . . . . . . . . . . . . . . 8 Heuristics polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
71 88 88 89 89 90 91
. . . .
9 General orthogonal polynomials 93 Lower and upper bounds . . . . . . . . . . . . . . . . . . . . . . . 94 Zeros . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95 Regularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97 10 Strong, ratio and weak asymptotics Strong asymptotics . . . . . . . . . . Ratio asymptotics . . . . . . . . . . Weak and relative asymptotics . . . Widom’s theory . . . . . . . . . . . . Asymptotics for Christoﬀel functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98 . 98 . 99 . 100 . 101 . 102 102 104 107 109
11 Recurrence coeﬃcients and spectral measures 12 Exponential and Freud weights 13 Sobolev orthogonality 14 NonHermitian orthogonality
15 Multiple orthogonality 110 Types and normality . . . . . . . . . . . . . . . . . . . . . . . . . 111 Recurrence formulae . . . . . . . . . . . . . . . . . . . . . . . . . 112 The Riemann–Hilbert problem . . . . . . . . . . . . . . . . . . . 113 16 Matrix orthogonal polynomials Matrix orthogonal polynomials . . . . . . . . . Threeterm recurrence and quadrature . . . . . Families of orthogonal polynomials . . . . . . . Connection with higher order scalar recurrence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 114 115 116 117
1
Introduction
The theory of orthogonal polynomials can be divided into two main but only loosely related parts. The two parts have many things in common, and the division line is quite blurred, it is more or less along algebra vs. analysis. One of the parts is the algebraic aspect of the theory, which has close connections
V. Totik
72
with special functions, combinatorics and algebra, and it is mainly devoted to concrete orthogonal systems or hierarchies of systems such as the Jacobi, Hahn, AskeyWilson, . . . polynomials. All the discrete polynomials and the qanalogues of classical ones belong to this theory. We will not treat this part; the interested reader can consult the three recent excellent monographs [39] by M. E. H. Ismail, [28] by W. Gautschi and [6] by G. E. Andrews, R. Askey and R. Roy. Much of the present state of the theory of orthogonal polynomials of several variables lies also close to this algebraic part of the theory. To discuss them would take us too far from our main direction; rather we refer the reader to the recent book [24] by C. F. Dunkl and Y. Xu. The other part is the analytical aspect of the theory. Its methods are analytical, and it deals with questions that are typical in analysis, or questions that have emerged in and related to other parts of mathematical analysis. General properties ﬁll a smaller part of the analytic theory, and the greater part falls into two main and extremely rich branches: orthogonal polynomials on the real line and on the circle. The richness is due to some special features of the real line and the circle. Classical real orthogonal polynomials, sometimes in other forms like continued fractions, can be traced back to the 18th century, but their rapid development occurred in the 19th and early 20th century. Orthogonal polynomials on the unit circle are much younger, and their existence is largely due to Szeg˝ and Geronimus in the ﬁrst half of the 20th century. B. Simon’s o recent treatise [80, 81] summarizes and greatly extends what has happened since then. The connection of orthogonal polynomials with other branches of mathematics is truly impressive. Without even trying to be complete, we mention continued fractions, operator theory (Jacobi operators), moment problems, analytic functions (Bieberbach’s conjecture), interpolation, Pad´ approximation, e quadrature, approximation theory, numerical analysis, electrostatics, statistical quantum mechanics, special functions, number theory (irrationality and transcendence), graph theory (matching numbers), combinatorics, random matrices, stochastic processes (birth and death processes; prediction theory), data sorting and compression, Radon transform and computer tomography. This work is a survey on orthogonal polynomials that do not lie on the unit circle. Orthogonal polynomials on the unit circle—both the classical theory and recent contributions—will be hopefully dealt with in a companion article. This work is meant for nonexperts, and it therefore contains introductory materials. We have tried to list most of the actively researched ﬁelds not directly connected with orthogonal polynomials on the unit circle, but because of space limitation we have only one or two pages on areas where dozens of papers and several books had been published. As a result, our account is necessarily incomplete. Also, the author’s personal taste and interest is reﬂected in the survey, and the omission of a particular direction or a set of results reﬂects in no way on the importance or quality of the omitted works. For further backgound on orthogonal polynomials, the reader can consult
Orthogonal Polynomials
73
the books Szeg˝ [91], Simon [80][81], Freud [27], Geronimus [34], Gautschi [28], o Chicara [18], Ismail [39]. This is a largely extended version of the ﬁrst part of the article Golinskii– Totik [36]. Acknowledgement. Research was supported by NSF grant DMS040650 and OTKA T049448, TS44782, and was carried out within the Analysis Research Group of the Hungarian Academy of Sciences. The author participates in the project INTAS 03516637 that supported a visit by Leonid Golinskii to Szeged, during which the outline of this paper was laid out.
2
Orthogonal polynomials
Orthogonal polynomials with respect to measures Let µ be a positive Borel measure on the complex plane, with an inﬁnite number of points in its support, for which zm dµ(z) < ∞ for all m > 0. There are unique polynomials pn (z) = pn (µ, z) = κn z n + · · · , that form an orthonormal system in L2 (µ), i.e. pm pn dµ = 0 1 if m = n if m = n. κn > 0, n = 0, 1, . . .
These pn ’s are called the orthonormal polynomials corresponding to µ. κn is the leading coeﬃcient, and pn (z)/κn = z n + · · · is called the monic orthogonal polynomial. The leading coeﬃcients play a special and important role in the theory, many properties depend on their behavior. When dµ(x) = w(x)dx on some interval, say, then we talk about orthogonal polynomials with respect to the weight function w. The pn ’s can be easily generated: all we have to do is to make sure that pn (z) k z dµ(z) = 0, κn k = 0, 1, . . . , n − 1, (2.1)
which is an n×n system of equations for the nonleading coeﬃcients of pn (z)/κn n−1 with matrix (σi,j )i,j=0 , where σi,j = z i z j dµ(z)
V. Totik
74
are the complex moments of µ. This matrix is nonsingular: if some linear combination with coeﬃcients c0 , . . . , cn−1 of the rows is zero, then the polynomial Pn−1 (z) = c0 + · · · + cn−1 z n−1 is orthogonal to every z j , j < n, and hence it is orthogonal to itself, i.e., Pn−1 2 dµ = Pn−1 Pn−1 dµ = 0,
which implies Pn−1 (z) ≡ 0. Thus, c0 = · · · = cn−1 = 0, which shows the nonsingularity of (σi,j ). Therefore, the system (2.1) has a unique solution for the nonleading coeﬃcients of pn (z)/κn (note that the leading coeﬃcient is 1), and ﬁnally κn comes from normalization. In particular, the complex moments already determine the polynomials. In terms of them one can write up explicit determinant formulae: σ0,0 σ1,0 . . . σn−1,0 σn,0 where are the so called Gram determinants. Note that if µ is supported on the real line then σi,j = xi+j dµ(x) =: αi+j , Dn = σi,j n i,j=0 (2.3) σ0,1 σ1,1 . . . σn−1,1 σn,1 ··· σ0,n−1 ··· σ1,n−1 . .. . . . · · · σn−1,n−1 σn,n−1 1 z . . . z n−1 zn
pn (z) =
1 Dn−1 Dn
(2.2)
···
so Dn = αi+j n i,j=0 is a Hankel determinant, while if µ is supported on the unit circle then σi,j = z i−j dµ(z) =: βi−j ,
n so Dn = βi−j i,j=0 is a Toeplitz determinant. In these two important cases the orthogonal polynomials have many special properties that are missing in the general theory.
The Riemann–Hilbert approach Let µ be supported on the real line, and suppose that it is of the form dµ(t) = w(t)dt with some smooth function w. A new approach to generating orthogonal polynomials that has turned out to be of great importance was given in the early 1990’s by Fokas, Its and Kitaev [26]. Consider 2 × 2 matrices Y (z) = Y11 (z) Y12 (z) Y21 (z) Y22 (z)
i.e.7) It can be shown that there is a unique solution Y (z).5)
These connect the functions on the upper and lower half planes only very mildly. resp. and Y21 (z) is a polynomial of degree at most n − 1. (2. The relation (2. and solve for such matrices the following matrixvalued Riemann–Hilbert problem: 1. lower half plane. and 2. resp.Orthogonal Polynomials
75
where the Yij are analytic functions on C \ R. is the limit of Y (z) as z tends to x from the upper. Thus.
Y21 (z) = O(zn−1 )
(2. − + + Y22 (x) = Y21 (x)w(x) + Y22 (x).. namely it is required that for large z uniformly on the plane we have Y11 (z) = z n + O(zn−1 ). the integral over the upper part of the circle can be deformed into an integral from R to −R on the upper part of R. then it is a polynomial of degree at most m. and if an entire function is O(zm ) as z → ∞. resp. one is looking for 4 functions Y11 . The relevance of this to orthogonal polynomials is that the entry Y11 (z) is precisely the monic polynomial pn (µ.6) that Y11 (z) = z n +· · · is a monic polynomial of degree n. into
−R R + xk Y12 (x)dx. where I denotes the identity matrix. such that if ± Yij (x) denote the boundary limits of these functions at x ∈ R from the upper.. then
− + Y11 (x) = Y11 (x). for all x ∈ R 1 w(x) Y + (x) = Y − (x) 0 1 where Y + .
. i.4)
and
− + + Y12 (x) = Y11 (x)w(x) + Y12 (x).7) gives that the integral of z k Y12 (z) over the circle z = R is O(Rk−n ) for all k < n and hence it tends to 0 as R → ∞.. and what puts the problem into a rigid form is the second condition. Indeed. Thus. By analyticity.
(2.e.4) shows that Y11 and Y12 are analytic everywhere.. lower half plane. into
R −R − xk Y12 (x)dx. (2. and Y12 (z) = O(z −n−1 ). z)/κn . Y22 analytic on C\R.. Y22 (z) = z −n + O(z−n−1 ).6)
and similarly the integral over the lower part of the circle can be deformed into an integral from −R to R on the lower part of R. Y − . . − + Y21 (x) = Y21 (x)
(2. 1 zn 0 Y (z) = I + O 0 z −n z at inﬁnity. we get from (2.
and orthogonality means pn . Y12 (z) = 1 2iπκn pn (x)w(x) dx x−z
is the Cauchy transform of pn (x)w(x)/κn .k = 0. Y11 is indeed the monic nth orthogonal polynomial with respect to w. pm = 0 for m = n.e.
76
therefore for k = 0. In this case the aforementioned orthogonalization process can be used. .
j=0
k = 0. · with the standard inner product properties. n − 1. Then the orthogonalization process and the determinantal formulae can still be used provided the Gram determinants (2. κn and the recurrence coeﬃcients an . . Sometimes one has an ·.. where Y1 is the matrix deﬁned by 1 z −n 0 Y (z) =: I + z −1 Y1 + O . and Y22 is the Cauchy transform of Y21 (= const · pn−1 ). .2) is still valid. 0 zn z2 For details on this Riemann–Hilbert approach. If this is not so. · which is deﬁned on some linear space containing all polynomials. . the determinantal formula (2. 0 ≤ k < n. xj . . then we write pn (z) = γn z n + γn−1 z n−1 + · · · . and with σi. and make sure that pn is orthogonal to all powers z k . The other entries can also be explicitly written in terms of the orthogonal polynomials pn and pn−1 : Y21 is a constant multiple of pn−1 .
∞ −∞
which implies
xk Y11 (x)w(x)dx = 0. .j = xi . Orthogonal polynomials with respect to inner products Sometimes one talks about orthogonal polynomials with respect to an inner product ·. bn (see Section 4) can be expressed in terms of the entries of Y1 .
. . n − 1 we have
R −R
xk Y11 (x)w(x)dx = O(Rk−n ) = O(R−1 ) → 0. .3) are diﬀerent from zero. i.5) implies
− + xk Y12 (x) − xk Y12 (x) = xk Y11 (x)w(x). solve the homogeneous system of equations
n
γj σj.
Thus. except that positive deﬁniteness may not hold (as an example consider nonHermitian orthogonality from Section 14). Totik The ﬁrst relation in (2.V. 1. Furthermore. see Deift [20].
in applications where nonHermitian orthogonality is used. For the second and third examples see multipoint Pad´ approximation and e random matrix theory in Section 6. . and orthogonality is with respect to a matrix measure) is a very active area which shows extreme richness compared to the scalar case..Orthogonal Polynomials
77
for γ0 . . . and for each n one forms the orthogonal system {pk (µn .g. λ The ﬁrst example is that of Freud weights: W (x) = e−x . z) are spreading out to inﬁnity). Varying weights In the last 25 years orthogonal polynomials with respect to varying measures have played a signiﬁcant role in several problems. studying orthogonal polynomials with respect to Freud weights W is equivalent to studying orthogonal polynomials with respect to the varying weights wn (x) = W (x)n . γ1 . If one 1/λ substitutes x = n y. Thus. . then we cannot assert any more γn = 0. then with Pn (y) = pn (W. when the entries of the ﬁxed size matrix are polynomials of degree n = 0. Still. working with wn turns out to be very natural. In most cases k=0 one needs the behavior of pn (µn . which gives rise to nontrivial orthogonal polynomials.
λ
n = m. Since the number of unknowns is bigger than the number of equations. and actually. . e. there is always a nontrivial solution. z) with some ﬁxed k. 1. In forming them one has a sequence of measures µn (generally with some particular behavior). the sections on exponential and Freud weights or on random matrices in Section 4. However. Matrix orthogonal polynomials Orthogonality of matrix polynomials (i. z) or that of pn±k (µn . See Section 16 for a short discussion. The minimum turns out
. γn .e. x) orthogonality takes the form Pn (y)Pm (y)e−ny dy = 0. and there may be several choices for pn . these pn play the role of orthogonal polynomials.
and it turns out that this is just the right scaling. We mention three examples.
3
The L2 extremal problem
One of the most useful tools in the study of orthogonal polynomials is the fact that they solve the following extremal problem: minimize the L2 (µ) norm for all monic polynomials Pn (z) = z n + · · · of degree n. . z)}∞ . so the degree of pn may be smaller than n. the zeros of Pn have an asymptotic zero distribution (while those of pn (W.g. see e. λ > 0. .
z) = If we write again Pn =
Pn (z)=1. i. and then
∞
µ({x0 }) = (
pk (µ..e. For example. n
then Pn (z) = 1 means
ck pk (µ.3)
for all z ∈ C for the Christoﬀel function λn (µ. z).
k=0
and hence by Cauchy’s inequality
n n
1≤( Therefore. z)2 )−1
ck =
Thus. z) = 1.
k=0
ck 2 )(
k=0
pk (µ. any Pn is a linear combination k=1 ck pk with the orthonormal polynomials pk . the nth monic orthogonal polynomial is the (unique) extremal polynomial in 1 (3. Totik
78
2 to be 1/κn . deg(Pn )≤n
inf
Pn 2 dµ.
k=0
. z)−1 =
k=0
pk (µ. µ({x0 }) > 0 if and only if 2 k pk (µ. They are deﬁned as λn (µ.1) inf Pn 2 dµ = 2 . x0 )2 )−1 .
(3. and equality occurs n if and only if all the other ck ’s are 0. for measures µ lying on the real line it is easy to see from this formula that µ has a point mass at x0 .
from which we can see that this is always ≥ cn 2 = 1/κ2 .. z)
n
λn (µ. i. we have arrived at the formula
pk (µ.V.
n
n
Pn 2 dµ = with equality if and only if
k=0
ck 2 ≥ (
k=0
pk (µ. A related extremum problem leads to the so called Christoﬀel functions associated with µ. z) .2)
n k=0 ck pk (µ.e. z)2 ). Orthogonality gives n
Pn 2 dµ =
k=0
ck 2 .. z)2
(3. ·).e. it has leading coeﬃcient 1. n 2 k=0 pk (µ. x0 ) < ∞. κn Pn (z)=z n +··· n Indeed. i. and here cn = 1/κn because Pn is monic.
see Section 6. which in turn is again closely related to the behavior of orthogonal polynomials. if we write xpn (x) as a linear combination n+1 ck pk (z) with k=0 ck := xpn (x)pk (x)dµ(x).. κn+1 xp2 (x)dµ(x) n (4. cn is the integral given in (4.1)
bn =
(4.2)
are the so called recurrence coeﬃcients. In the real case the zeros of pn are real and simple and the zeros of pn and pn+1 interlace. In the nonreal case the two sides here are totally diﬀerent.
then all the ck ’s for k < n − 1 vanish by orthogonality: ck = xpn (x)pk (x)dµ(x) = (xpk (x))pn (x)dµ(x) = 0
because xpk (x) is a polynomial of degree smaller than n. but since cn+1 is also the integral of xpn (x)pn+1 (x) against µ.e. Indeed.1) with real an > 0. Conversely. where an = κn > 0.2). hence xpn (x)pk (x)dµ(x) = pn (x)(xpk (x))dµ(x) = 0
for k < n − 1. The most remarkable property of this real case is that the pn ’s obey a threeterm recurrence formula xpn (x) = an pn+1 (x) + bn pn (x) + an−1 pn−1 (x). κn
Finally.Orthogonal Polynomials
79
4
Orthogonal polynomials on the real line
Let µ be supported on the real line. Comparison of the leading coeﬃcients on both sides gives that cn+1 = κn /κn+1 . in between any two zeros of pn+1 there is a zero of
. all the coeﬃcients are real).. we get that cn−1 = xpn (x)pn−1 (x)dµ(x) = κn−1 . bn is an orthonormal system with respect to a (not necessarily unique) measure on the real line (Favard’s theorem). i. and it is replaced by a diﬀerent recurrence for polynomials on the circle. In this case orthogonalization leads to real polynomials (i. and it is due to the fact that in this case the polynomials are real. We emphasize that the threeterm recurrence is a very special property of real orthogonal polynomials. any system of polynomials satisfying (4. The unicity of the measure in question is the same as the determinacy of the moment problem.e. The threeterm recurrence is missing in the general case.
< x1 be the zeros of pn .V. q0 ≡ 0 and uses the same recurrence (with a−1 = 1) xqn (x) = an qn+1 (x) + bn qn (x) + an−1 qn−1 (x). and suppose that we already know that the zeros of pn and pn−1 interlace.e. for if it had only m < n. i. sign(pn+1 (xk )) = (−1)k . and by Favard’s theorem the diﬀerent qn ’s are orthogonal with respect to some measure. α > −1. which implies sign(pn−1 (xk )) = (−1)k−1 .
(α) (α. In fact. If we substitute xk into the recurrence formula (4. κn+1 x−t
(4.. then it could not be orthogonal to the polynomial q(x) = m j=1 (x−yj ) of degree m < n. . The threeterm recurrence implies for the so called reproducing kernel the ChristoﬀelDarboux formula
n
pk (x)pk (t) =
k=0
κn pn+1 (x)pn (t) − pn (x)pn+1 (t) .1) are p−1 ≡ 0. . Thus.4)
is worth mentioning.5)
then qn is of degree n − 1.3)
Indeed. for then q(x)pn (x) would be of constant sign. use the recurrence formula for pn+1 on the right and an = κn /κn+1 . Totik
80
pn . z−x
5
Classical orthogonal polynomials
• Jacobi polynomials Pn . . . orthogonal with respect to the weight (1 − x)α (1 + x)β on [−1. yn ∈ R.1) then an > 0 gives that pn+1 (xk ) and pn−1 (xk ) are of opposite signs at xk . x)−1 =
k=0
pk (x)2 =
κn p′ (x)pn (x) − p′ (x)pn+1 (x) n κn+1 n+1
(4. with orthogonality weight xα e−x on [0. . say at the points y1 . They can also be written in the form qn (z) = (µ(C))
−1/2
pn (z) − pn (x) dµ(x). pn must have n sign changes. Let now xn < xn−1 < . • Laguerre polynomials Ln .3). The qn ’s are called orthogonal polynomials of the second kind (sometimes for pn we say that they are of the ﬁrst kind). α. and this gives that the zeros of pn and pn+1 also interlace. then induction gives (4. (4. . ∞). the interlacing property follows by induction. If one starts from q−1 = −1. β > −1. p0 = (µ(C)))−1/2 .β)
These are
. The starting values of the recurrence (4. The special case
n
λn (µ. 1].
β) ′ ) (x) = (Pn
1 (α+1.
2 2 (Hn (x)Hn−1 (x))′ = 2xHn (x)Hn−1 (x) − Hn (x) + 2nHn−1 (x). for they possess many properties that no other orthogonal polynomial system does. 2
• they all possess a Rodrigue’s type formula Pn (x) =
dn 1 {w(x)σ(x)n }. dn w(x) dxn
where w is the weight function and σ is a polynomial that is independent of n. Classical orthogonal polynomials are also special in the sense that they possess a relatively simple
. (α.
e. γn .
with some constants αn .g. See also Bochner’s result in the next section claiming that the classical orthogonal polynomials are essentially the only polynomial (not just orthogonal polynomial) systems that satisfy a certain second order diﬀerential equation. βn .g.β+1) (n + α + β + 1)Pn−1 (x). and σ a polynomial of degree at most 2.β) (α+1. In the literature various normalizations are used for them. In particular.
Every one of these properties has a converse. the derivative of Pn is a constant multiple of Pn−1 :
(α.β+1) e. see AlSalam [3]. for example. then it is (up to a change of variables) one of the classical systems. e. δn . x(L(α) )′ (x) = nL(α) (x) − (n + α)Ln−1 (x). They are very special.Orthogonal Polynomials
2
81
• Hermite polynomials Hn orthogonal with respect to e−x on the real line (−∞.
(α) Ln (x) = ex x−α
1 dn e−x xn+α .g. n! dxn
• they satisfy a diﬀerentialdiﬀerence relation of the form
′ π(x)Pn (x) = (αn x + βn )Pn (x) + γn Pn−1 (x). ∞). n n • they satisfy a nonlinear equation of the form σ(x) (Pn (x)Pn−1 (x))
′ (α)
=
(αn x + βn )Pn (x)Pn−1 (x) 2 2 +γn Pn (x) + δn Pn−1 (x). namely if a system of orthogonal polynomials possesses any of these properties. • they have derivatives which form again an orthogonal polynomial system.
β) (1 − x)α (1 + x)β Pn (x) =
(−1)n 2n+1 πi
(1 − t)n+α (1 + t)n+β (t − x)−n−1 dt
over an appropriate contour. For all these results see Szeg˝ [91]. . o
6
Where do orthogonal polynomials come from?
In this section we mention a few selected areas where orthogonal polynomials naturally arise. 2.V. B2 A1 + A2 +··· and its nth convergent is Sn B1 . e. An Sn−1 + Bn Sn−2 . . n!
• integral representation. e. Continued fractions Continued fractions played an extremely important role in the development of several branches of mathematics.
The value of the continued fraction is the limit of its convergents. A continued fraction is of the form B1 .
(α. and these are powerful tools to study their behavior. = B2 Rn A1 + A +··· Bn
2 An
n = 1. . Totik • second order diﬀerential equation.g.g. The denominators and numerators of the convergents satisfy the threeterm recurrence relations Rn Sn = = An Rn−1 + Bn Rn−2 . • generating function.g.
n (α)
82
Hn (x) n w = exp(2xw − w2 ). . xy ′′ + (α + 1 − x)y ′ + ny = 0 for Ln . e.
. R−1 ≡ 0
S0 ≡ 0. R0 ≡ 1. S−1 ≡ 1. but their signiﬁcance has unjustly diminished in modern mathematics.
the numerators Sn (z) and the denominators Rn (z) of the nth convergents Sn (z) B1 = Rn (z) z − A1 + z−A satisfy the recurrence relations Rn (z) = Sn (z) = (z − An )Rn−1 (z) + Bn Rn−2 (z).2)
B2 Bn 2 +··· z−An
. pp.
n = 1.. . . An important example is that of Cauchy transforms of measures µ with compact support on the real line (so called Markov functions): f (z) = α0 α1 dµ(x) =− − 2 − . Rm (z) Rm+1 (z) Rn (z)Rn+1 (z)
. 54–57] as well as Kruschev [42] and the numerous refero ences there. in which case the convergents are ratios of some orthogonal polynomials of the second and ﬁrst kind.
of the measure µ.. 1.
These are precisely the recurrence formulae for the monic orthogonal polynomials of the ﬁrst and second kind with respect to µ. . Markov’s theorem). 2. . j = 0. (6. Many el√ ementary functions (like z − z 2 − 1) have a continued fraction development where the Bn ’s are constants while the An ’s are linear functions. The continued fraction development f (z) ∼ B1 B z − A1 + z−A22+···
of f at inﬁnity converges locally uniformly outside the smallest interval that contains the support of µ (A. But the connection is deeper than just this formal observation. As has been mentioned.Orthogonal Polynomials
83
which immediately connects continued fractions with threeterm recurrences and hence with orthogonal polynomials. R−1 ≡ 0 (z − An )Sn−1 (z) + Bn Sn−2 (z). x−z z z (6. Pad´ approximation and rational interpolation e One easily gets from the recurrence relations (6. S0 ≡ 0.2) that B1 B2 · · · Bn+1 Sm+1 (z) Sm (z) − = (−1)n . hence the nth convergent is cqn (z)/pn (z) with c = µ(C)1/2 .1) are the moments αj = xj dµ(x).1)
The coeﬃcients αj in the development of (6. . See Szeg˝ [91. . S−1 ≡ 1. R0 ≡ 1..
Totik and summation of these for m = n. Put
2n
ωn (z) :=
j=0 xjn =∞
(z − xjn ). The advantage of Pad´ approximation over Taylor e polynomials lies in the fact that the poles of Pad´ approximants may imitate e the singularities of the function in question. but we assume that An is symmetric with respect to the real line. The error in [n − 1/n] Pad´ approximation has e the form qn (z) p2 (x) 1 n f (z) − c = 2 dµ(x). . . pn (z) pn (z) x−z
Orthogonal polynomials appear in more general rational interpolation (called multipoint Pad´ approximation) to Markov functions.g. Stahl–Totik e [87. see e. This is the analogue (called [n − 1/n] Pad´ approxie mant) of the nth Taylor polynomial (which interpolates the function to order n) for rational functions. The points need not be distinct. An . . . x2n. z k+1
i. n + 1. Sec. .1].n . . .
The degree of ωn is equal the number of ﬁnite points in An . .V. Qn is the nth orthogonal polynomial with respect to the varying weight dµ(x)/ωn (x). . with the preceding notation the rational function Sn (z)/Rn (z) = cqn (z)/pn (z) with c = µ(C)1/2 of numerator degree n − 1 and of denominator degree n interpolates f (z) at inﬁnity to order 2n.n } of 2n+1 interpolation points from C \ I where I is the smallest interval that contains the support of µ.e. x2n. Now for Markov functions this rational interpolant uniquely exists. . and at inﬁnity it has the indicated behavior. . and the remainder term of the interpolation has the representation (f − rn (f.n } in the sense that f (z) − rn (f. For every n select a set An = {x0. ·))(z) = ωn (z) Q2 (z) n Q2 (x) n dµ(x) ωn (x)(x − z)
. An .. By rn (z) = un (z)/Qn (z) we denote a rational function of numerator and denominator degree at most n that interpolates the function f at the 2n + 1 points of the set An = {x0. 6. yields that Sn (z) = Rn (z)
2n k=0
84
−αk + O(z −2n−1 ). while Taylor polynomials are good only up to the ﬁrst singularity.
the expression on the left is bounded at every ﬁnite point of An .n . z) = O(z −(2n+1) ) ωn (z) as z → ∞.
In fact. . B. D(z) = z
pn (0)pn (z). n = 0. are positive. and A.j = αi+j . . 1. Thus. which in turn is equivalent to n pn (z)2 < ∞ for all z ∈ C. supported on the real line. C and D have explicit representations in terms of the ﬁrst and second kind orthogonal polynomials pn and qn : A(z) = z
n qn (0)qn (z). . . the Cauchy transforms of all solutions ν of the moment problem have the parametric form C(z)F (z) + A(z) dν(x) = . . . .
For all these results see Akhiezer [2].2) deﬁned from the moments σi. Furthermore. and for an operator theoretic approach to the moment problem see Simon [78] (in particular.. . Jacobi matrices and spectral theory of selfadjoint operators Tridiagonal. The unicity (usually called i.
The Hamburger moment problem is to determine if a sequence {αn } (with normalization α0 = 1) of real numbers is the moment sequence of a measure with inﬁnite support. there are diﬀerent measures with the same moments αj if and only if there is a nonreal z0 with n pn (z0 )2 < ∞. µ(C) = 1. ∞)). .
··· ··· ··· ··· .j=0 determinacy) depends on the behavior of the orthogonal polynomials (2. n
B(z) = −1 + z
n
n qn (0)pn (z). .
C(z) = 1 + z
pn (0)qn (z). . m = 0. . .
. the rate of convergence of the rational interpolants is intimately connected with the behavior of the orthogonal polynomials with respect to the varying weight dµ(x)/ωn (x). z−x D(z)F (z) + B(z) where F is an arbitrary analytic function (the parameter) mapping the upper half plane C+ into C+ ∪ {∞}. but for measures on [0. .
0 0 a2 b2 .14).. . 1. Moment problem The moments of a measure µ. and if this measure is unique (the Stieltjes moment problem asks the same. . Theorems 3 and 4. are αn = xn dµ(x). so called Jacobi matrices b 0 a0 0 a0 b 1 a1 J = 0 a1 b 2 0 0 a2 . The existence is easy: {αn } are the moments of some measure supported on R if and only if all the Hankel determinants αi+j m . .Orthogonal Polynomials
85
for all z outside the support of µ. .
2]).
Thus.. 2. . By the spectral theorem. . J. These show that Jacobi operators are equivalent to multiplication by x in L2 (µ) spaces if the particular basis {pn (µ. . .k )
k=1
in some sense for as large a class of functions as possible. . The recurrence coeﬃcients for pn (µ. Ch. . an npoint quadrature (rule) is a sequence of n points xn. x) are precisely the an ’s and bn ’s from the Jacobi matrix. i. the monic polynomial pn (z)/κn is the characteristic polynomial of Jn . which turn out to be the zeros of pn . . . . . π0 = 1.V. ··· .g. 1.)). ·)} is used (see e.1 .1 . Every o bounded self adjoint operator with a cyclic vector is a Jacobi operator in an appropriate base.) into pn . Quadrature For a measure µ. a so called Jacobi operator. . if pn (x) = pn (µ.k f (xn. This µ is called the spectral measure associated with J (and with its spectrum). 0. . λn. . Totik
86
with bounded an > 0 and bounded real bn deﬁne a bounded selfadjoint operator J in l2 . πn (λ) is of degree n in λ. . 0. then U can be extended to a unitary operator from l2 onto L2 (µ).. It is expected that
n
f dµ ∼
λn. n = 1. Deift [20. and if Sf (x) = xf (x) is the multiplication operator by x in L2 (µ). More precisely. .e. . . . xn. The formal eigenequation Jπ = λπ is equivalent to the threeterm recurrence an−1 πn−1 + bn πn + an πn+1 = λπn . i. 0. pn (x) = cπn (x) with some ﬁxed constant c. 0. . is unitarily equivalent to multiplication by x in some L2 (µ) with some probability measure µ having compact support on the real line. Often the accuracy of the quadrature is measured by its exactness. and U maps the unit vector en = (0.e. as a selfadjoint operator having a cyclic vector ((1. .n and a sequence of associated numbers λn. . 0 0 0 an−2 bn−1 has n real and distinct eigenvalues. which is deﬁned as the largest m
. . . . These are the discrete analogues of second order linear diﬀerential operators of Schr¨dinger type on the half line. . then J = U −1 SU .n .. . The truncated n × n matrix b 0 a0 0 0 ··· a0 b 1 a1 0 ··· 0 a1 b 2 a2 ··· Jn = . . . x) are the orthonormal polynomials with respect to µ. . . b0 π0 + a0 π1 = λπ0 . .
k
k<j
d ℜmk. xn.2) associated with µ.1 .k ) =
pj (µ.Orthogonal Polynomials
87
such that the quadrature is exact for all polynomials of degree at most m.j d ℑmk. o Random matrices Some statisticalmechanical models in quantum systems use random matrices. Every matrix M ∈ Hn has n real eigenvalues which carry physical information on the system when it is in the state described by M .
where V (λ). XV]. This model is known as the unitary ensemble associated with V .k = λn (µ.k are chosen to be
n −1
where λn (µ. i. . x) be the orthonormal polynomials with respect to the varying weight wn (x).n are the zeros of the orthonormal polynomial pn (µ. and this optimal value 2n − 1 is attained if and only if xn. Ch. xn.. and Dn is a normalizing constant so that the total integral of Pn (M )dM is one. w(x) = exp(−V (x)). and let there i.k )
j=0
2
.
λn.j )n .j=1 be given a probability distribution on Hn of the form
−1 Pn (M )dM = Dn exp(−nTr{V (M )})dM.e. .j≤n
.k xj n. Then the joint probability density of the eigenvalues can be written in the form Nn (D) = dn pi−1 (wn . is a realvalued function that increases suﬃciently fast at inﬁnity (typically an even polynomial in quantum ﬁeld theory applications). m is the largest number with the property that
n
xj dµ(x) =
k=1
λn.k
for all 0 ≤ j ≤ m. Tr{H} denotes the trace of the matrix H. Let pj (wn .
. See Szeg˝ [91.
n
dM =
k=1
dmk. Let Hn be the set of all n × n Hermitian matrices M = (mi. . The quantity #{eigenvalues in D} n is the random variable that equals the normalized number of eigenvalues in the interval D. λj )wn/2 (λj )
2 1≤i. xn. x) corresponding to µ.j
is the “Lebesgue” measure for Hermitian matrices. z) is the Christoﬀel function (3. λ ∈ R. . and the so called Cotes numbers λn.
For µ with support on the real line and for quadrature based on n points this exactness m cannot be larger than 2n − 1.
and with these ﬁxed charges put n positive unit charges on [−1. n! In particular.
it can also be written in the form 1 Kn (λi . for the expected number of eigenvalues in an interval D we have ENn (D) =
D
Kn (λ. n
where 1/Kn (λ. while the limit of the left hand side (as n → ∞) is known as the density of states. . xi − xj 
and the equilibrium problem asks for ﬁnding x1 .j≤n . . . There is a similar characterization of the zeros of Laguerre and Hermite polynomials.V. . With the so called weighted reproducing kernel
n−1
Kn (t.
. . . Electrostatics Put at 1 and −1 two positive charges p and q.. Let us just mention a few. λ) dλ.5]).2q−1) of the Jacobi polynomial Pn orthogonal with respect to the weight (1 − x)2p−1 (1 + x)2q−1 . xn ) = p
j=1
log
1 +q 1 − xj 
n
log
j=1
1 + 1 + xj 
log
i<j
1 . xn . and even of more general orthogonal polynomials (for the latter see Ismail [39. . s) =
j=0
pj (wn . e.7]) for the zeros o (2p−1. . Totik
88
where dn is a normalizing constant built up from the leading coeﬃcients of the pj (wn .
7
Some questions leading to classical orthogonal polynomials
There are almost an inﬁnite number of problems where classical orthogonal polynomials emerge. ·). The unique minimum occurs (see Szeg˝ [91. xn for which this energy is minimal. Therefore. t)wn/2 (t) pj (wn . . 1] at the points x1 . Section 3. the mutual energy of all these charges is
n
I(x1 . On the plane the Coulomb force is proportional with the reciprocal of the distance. . Section 6. . . s)wn/2 (s). Mehta [60] and Pastur–Figotin [68]. λ) is known in the theory of orthogonal polynomials as the nth (weighted) Christoﬀel function associated with the weight wn . λj )1≤i.g. and so a charge generates a logarithmic potential ﬁeld. See.
the so called ultraspherical or Gegenbauer polynomials Pn . harmonic homogeneous polynomials. . nd ) such that n1 + · · · + nd = n. Bochner’s theorem from [16] states that. . .
d The dimension of Hn is
n+d−1 n+d−3 − . . . .nd
j=1
(λ (x2 + · · · + x2 )nj Pnj j ) (xj (x2 + · · · + x2 )−1/2 ) . Harmonic analysis on spheres and balls Harmonic analysis on spheres and balls in Rd is based on spherical harmonics. g.e. . With ρ = xd−1 + xd
let gs. λn = −2n). In this theory. d Let S d−1 be the unit sphere in Rd and let Hn be the restriction to S d−1 of all harmonic polynomials Q(x1 .1 = xd ρs Ps (xd−1 /ρ). i. i.. . With nd = 0 or nd = 1 consider all multiindices n = (n1 . dx2 dx
89
where f. xd ) as
d−2
(0)
(1)
gnd−1 .. respectively.Orthogonal Polynomials Polynomial solutions of eigenvalue problems Consider the eigenvalue problem f (x) d d2 y(x) + g(x) y(x) + h(x)y(x) = λy(x).
n
k=1
∂2 Q = 0. . xn ). λn = −n) and
• Hermite polynomials Hn (x) (f (x) = 1. • Jacobi polynomials Pn λn = −n(n + α + β + 1)) • Laguerre polynomials Ln
(α) (α. . d−1 d−1
2 2 and an orthogonal basis in it can be produced as follows.e. h are ﬁxed polynomials and λ is a free constant. n2 . play a fundamental role – they are orthogonal with respect to the weight (1−x2 )α−1/2 . . . the only solutions are (in all of them we can take h(x) = 0) (f (x) = 1 − x2 . λxn ) = λn Q(x1 .. xn ) of d variables that are homogeneous of degree n. and it is required that this have a polynomial solution of exact degree n for all n = 0. g(x) = β − α − x(α + β + 2). g(x) = 1 + α − x. . . special Jacobi polyno(α) mials. .β)
(f (x) = x. for which the corresponding λ and y(x) will be denoted by λn and yn (x). . ∂x2 k
Q(λx1 . .0 = ρs Ps (xd−1 /ρ) and gs.
λ > 0. j d j d
. . . . . except for some trivial solutions of the form y(x) = axn + bxm and for some polynomials related to Bessel functions. and if for such a multiindex we deﬁne the function Yn (x1 . . g(x) = −2x. 1. .
If one considers. . indispensable tools in approximation theory. y = k xk yk is the inner product in Rd . then
the reproducing kernel for these spherical polynomials is Pn d sense that for all Q ∈ Hn and for all x ∈ S d−1 we have cn. But there are many other aspects of approximation where Chebyshev polynomials appear. Szeg˝ o [91. Totik
90
d then these Yn constitute an orthogonal basis in Hn (see e.d is an explicit normalizing constant (see e. as well as in the corresponding theory for the unit ball. In fact. As we have seen in (3. Lagrange interpolation and its various generalizations like HermiteFej´r or e Hermite interpolation etc. . In many cases interpolation on zeros of orthogonal polynomials has special properties due to explicitly calculable expressions. these nodes are often close to optimal in the sense that the Lebesgue constant increases at the optimal rate. . norms among monic polynomials of a given degree. 1] with respect to the weight w(x) = (1 − x2 )−1/2 . monic Chebyshev polynomials minimize all Lp (w). p. y ) in the
(d−2)/2 Pn ( x. Dunkl–Xu [24. 35]). The Chebyshev polynomials cos(n arccos x) are orthogonal on [−1. p > 0.
where integration is with respect to surface area. .3)
where the minimum is taken for all monic polynomials of degree n. Ch. the best approximation on [−1. Dunkl–Xu [24. in particular Chebyshev polynomials. 37]).1). Gegenbauer polynomials are all over the theory of spherical harmonics. Fej´r’s result that if P2n−1 is the unique polynomial e
. and numerous works have been devoted to their convergence and approximation properties. 1] of xn in the uniform norm by polynomials Pn−1 (x) of smaller degree then the smallest error appears when xn − Pn−1 (x) = 21−n cos(n arccos x) is the monic nth Chebyshev polynomial.d
(d−2)/2
( x. Recall e.g.V.g. for example. is mostly done on the zeros of some orthogonal polynomials. the monic orthogonal polynomials pn (µ)/κn are the solutions to the extremal problem Pn 2 dµ → min. p.g. expansions of functions into classical orthogonal polynomial series are second only to trigonometric expansions.g. This extremal property makes orthogonal polynomials. Actually. If x = (x1 . As a result. and expansions into them have virtually the same properties as trigonometric Fourier expansions. and cn.
Approximation theory In the literature. XIII]. see e. xn ) and x. y )Q(y)dσ(y) = Q(x). These directly correspond to trigonometric functions. (7.
3) is something like minimality in the L∞ norm on S(µ). 1] as n → ∞. for which the orthogonal polynomials are z n . see the next section): on any compact set outside the polynomial convex hull there can only be a ﬁxed number of zeros of pn (µ) for every n. then the same reasoning leads to a behavior like (8. in particular Section 9.1). The example of arc measure on the unit circle. as well as for a more precise form of some of the heuristics see the following sections. 1]. Since 1 log Tn (z) = n log z − tdνn (t)
where νn has mass 1/n at each zero of Tn . e.
8
Heuristics
In this section we do not state precise results. Therefore. As we have seen. For the role of orthogonal polynomials in interpolation see the books Szabados–V´rtesi [90] e and Mastroianni–Milovanovic [55]. We just want to indicate some heuristics on the behavior of orthogonal polynomials. When S(µ) = [−1. More generally. if dν = dνn = wn (x)dx is a varying weight in the speciﬁed way. The L2 (µ) minimality of pn (µ)/κn in the sense of (7. The statement is true when the support lies on R or on some systems of arcs. one expects the zeros to cluster at the support S(µ) of µ.g. shows however. In many situations the distribution of the zeros is the equilibrium distribution of the support S(µ). but now ν is a measure
. then the distribution of the zeros shows a remarkable universality and indiﬀerence to the size of µ.Orthogonal Polynomials
91
of degree at most 2n − 1 that interpolates a continuous function f at the nodes of the nth Chebyshev polynomial and that has zero derivative at each of these nodes. in the limit the behavior should be like U ν (z) = log z − tdν(t). Therefore.1) where ν is the probability measure on S(µ) for which the maximum of U ν on S(µ) is as small as possible (this is the so called equilibrium measure of S(µ)). these polynomials try to be small where the measure is large. that this is not true (due to the fact that the complement of the support is not connected). and also in the general case when instead of the support one considers the polynomial convex hull of the support of µ (for the deﬁnition. pn (µ)/κn should behave like the monic polynomial Tn minimizing the L∞ norm on S(µ) (so called Chebyshev polynomials for S(µ)). the monic orthogonal polynomials pn (µ)/κn minimize the L2 (µ) norm in (7.3). (8. When the complement of S(µ) is connected and S(µ) has no interior. then P2n−1 uniformly converges to f on [−1. this means that under very weak assumptions the zero distribution is always the √ arcsine distribution dx/π 1 − x2 . For the concepts below.
z ∈ S(µ). and the diﬀerent types of asymptotics usually require the measure to be suﬃciently strong with diﬀerent degree on its support. Hermite or Freud weights).V. Finally. ∞). and one has to rescale them to [−1. x) ∼ dn g(x) sin(nh(x) + H(x)) holds. (8. Such a ﬁne asymptotic is restricted to the simply connected case. which is mostly NOT the arcsine
. z). z)/pn (µ.. 1]) to get a distribution. Totik
92
for which the supremum of U ν (z) + log w(z) is as small as possible (weighted equilibrium measure). strong asymptotics of the form (8.e. h. Asymptotics of orthogonal polynomials have a hierarchy. z ∈ S(µ). Here h(x) is directly linked with the zeros. When the unbounded component of the complement of S(µ) is simply connected. It can only hold when C \ S(µ) (more precisely its unbounded component) is simply connected. z) → gC\S(µ) (z. It is mostly equivalent to a corresponding distribution of the zeros. Section 10. 1] (or to [0. where g. be the same as the capacity of S(µ). Consider ﬁrst the case of compact support S(µ). then in that component often there is a ﬁner asymptotic behavior of pn (µ) of the form pn (z) ∼ dn gµ (z)Φ(z)n . and in this case it is enough that µ′ > 0 almost everywhere with respect to the equilibrium measure of the support of µ (see Section 10). z)1/n outside the support of the measure. When S(µ) = [−1. The weakest is nth root asymptotics stating the behavior (8. see Section 10). as well as asymptotical minimal behav1/n ior of κn . h′ /π is precisely the distribution of the zeros.3) needs roughly that log µ′ be integrable (Szeg˝ condition. When S(µ) is not of compact support (like Laguerre. and in the real case under smoothness assumptions on the measure often a so called PlancherelRotach type asymptotic formula pn (µ. then h(x) = arccos x. is stronger. see e. o All these are outside the support.2) for pn (µ. It holds under very weak assumptions on the measure.3)
where Φ is the mapping function that maps C \ S(µ) conformally onto the outside of the unit disk. Usually they obey 1 log pn (µ. On the support the orthogonal polynomials are of oscillatory behavior. i. ∞) is the Green function with pole at inﬁnity associated with the complement of the support. asymptotic behavior of pn+1 (µ.g. and gµ is a function (might be called generalized Szeg˝ o function) that depends on µ. Ratio asymptotics. then usually the zeros are spreading out. roughly stating that the logarithmic capacity of the points where µ′ > 0 (derivative with respect to equilibrium measure). 1] and the measure is smooth. (8. Universal behavior can also be seen for the polynomials themselves.2) n where gC\S(µ) (z. and is equivalent with asymptotics for the ratio κn+1 /κn of consecutive leading coeﬃcients. H are ﬁxed functions.
For all the results below see Stahl–Totik [87] and the references therein. t)dµ(t). its Fourier expansion into {pn (µ. Pollard’s theorem that Legendre expansions are bounded in Lp [−1. In general. In a similar fashion.g. ·)} is
∞
f (x) ∼
ck pk (x). z) tends pointwise to the RadonNikodym derivative of µ with respect to the equilibrium measure.2) or. κn+1 x−t
which suggests a singular integraltype behavior for the partial sums.
In the real case for the reproducing kernel Kn (x. z) is like µ(Dn (z)). t) we have the ChristoﬀelDarboux formula Kn (x. x) holds in many cases. z)2 . see e. t) = κn pn+1 (x)pn (t) − pn (x)pn+1 (t) . Fourier expansions into orthogonal polynomials are sensitive to the weight (recall e.Orthogonal Polynomials
93
distribution.
The nth partial sum has the closed form
n
f (t)Kn (x. but sometimes convergence properties are equivalent to those of a related trigonometric Fourier series (so called transplantation theorems.
k=0
ck =
f pk dµ.g. 1] only for 4/3 < p < 4).
. nλn (µ. the square sum 1 = λn (µ.
Kn (x. The Christoﬀel function (3.
9
General orthogonal polynomials
In this section µ always has compact support S(µ). If f ∈ L2 (µ).
behaves much more regularly than the orthogonal polynomials. In particular. what is the same. As a rule of thumb the estimate pn (µ. z)
n k=0
pk (µ. On S(µ) the typical behavior of λn (µ. t) =
k=0
pk (x)pk (t). various asymptotics hold for the polynomials only after a corresponding rescaling. Askey [11]). Outside the support the behavior of λn is what one gets from the heuristics above on the polynomials (after square summation). where Dn (z) is the disk about z with equilibrium measure 1/n (equilibrium measure of the support S(µ)). x)2 ≤ C/nλn (µ.
Green’s function gC\K (z.. then there is a unique measure ν = ωK minimizing the energy in (9. z)1/n ≥ gC\S(µ) (z. ∞) n
(9. The capacity of a Borel set is the supremum of the capacities of its compact subsets. Totik Lower and upper bounds The energy V (K) of a compact set K is deﬁned as the inﬁmum of I(ν) = log 1 dν(x)dν(t) x − t
94
(9. while in the convex hull but outside the so called polynomial convex hull Pc(S(µ)) (for the deﬁnition see below) (9.4) cap(K) z − t We have for all µ (with cap(S(µ)) > 0) the estimate lim inf
n→∞
1 log pn (µ. The logarithmic capacity is then cap(K) = e−V (K) . cµ
(9. With this.1)
where the inﬁmum is taken for all positive Borel measures on K with total mass 1. z)1/n ≥ 1. which is deﬁned as the boundary ∂Ω of the unbounded component Ω of the complement C\S(µ). where the latter is deﬁned as the inﬁmum of gC\K for all compact subsets K of E. and the minimal carrier capacity cµ associated with µ is the inﬁmum of the capacities of all carriers.e.5) is true quasieverywhere (i. ∞) n
(9. and this measure is called the equilibrium measure of K. For the leading coeﬃcients κn of the orthonormal polynomials pn (µ) we have 1 ≤ lim inf κ1/n . (9.1). lim sup
n→∞
1 log pn (µ. namely for quasievery z ∈ ∂Ω lim inf pn (µ.V.6)
. ∞) = log − log dωK (t). With this lim sup κ1/n ≤ n
n→∞
1 .2)
To get an upper bound we need the concept of carrier: a Borel set E is a carrier for µ if µ(C \ E) = 0. ∞) with pole at inﬁnity of C \ K can then be deﬁned as 1 1 gC\K (z.3)
When cap(K) is positive. n n→∞ cap(S(µ)) (9. with the exception of a set of zero capacity). ∞) is the supremum for all carriers E of the Green function of C \ E. z)1/n ≤ gµ (z.
n→∞
The minimal carrier Green function gµ (z.5)
locally uniformly outside the convex hull of S(µ). The same is true on the outer boundary of S(µ).
. .8)
Taking this for granted. xm ∈ V there exist m points y1 .5) and (9. then there exist a < 1 and m ∈ N such that for arbitrary m points x1 .e. z) has at least m zeros x1 . .
.7) by the points x1 . and let a < 1 and m ∈ N be the numbers in the lemma. ym ∈ C such that the rational function rm deﬁned as in (9.7) rm (z) := z − xj j=1 has on S a supnorm satisfying rm
S
≤ a. . . z) = z n + . .8). . ym ∈ C for which the rational function m z − yj (9. . When Ω is the unbounded component of the complement C \ S(µ). All these estimates are sharp. xm on V. . To say somewhat more on the location of zeros we need the concept of the polynomial convex hull. ·)
L2 (µ) . . This is a consequence of the L2 extremal property (3. S ⊆ C be two compact sets.3) coincide we have convergence for κn .1). The proof of this is based on the following lemma: Let V. with the bounded components of C \ S(µ)). but that is impossible by (3. .1) of orthogonal polynomials. . We apply the lemma with S = S(µ). If V and Pc(S) are disjoint. 1/n When the bounds in (9. assume that V is a compact set contained in Ω.
(9.2) and (9. i. . . . Now the zeros cluster on Pc(S(µ)) in the sense that for any compact subset K of Ω there is a number NK independent of n. and these bounds coincide precisely when the bounds in (9. .Orthogonal Polynomials
95
locally uniformly on the whole plane. . then we could move that zero towards the convex hull (along a line segment that is perpendicular to a line separating z0 and S(µ)). For the L2 (µ) norm of this polynomial we have the estimate qn
L2 (µ)
≤ rm
S(µ)
pn (µ. . if there was a zero z0 of pn (µ. xm and y1 . . Zeros The zeros of pn (µ) always lie in the convex hull of the support S(µ) of the measure µ. such that pn (µ) can have at most NK zeros in K. . z) outside the convex hull of the support. In fact. .6) do so. then Pc(S(µ)) = C \ Ω is called the polynomial convex hull of S(µ) (it is the union of S(µ) with all the “holes” in it. . By the lemma there exist m points y1 . . .. . ·)
L2 (µ)
< pn (µ. With rm we deﬁne the modiﬁed monic polynomial qn (z) := rm (z)pn (µ. ym satisﬁes the inequality (9. . Let us assume that pn (µ. . During this move the absolute value of the polynomial decreases at all points of S(µ) and hence so does its L2 (µ) norm.
In fact. Let z0 be an isolated point of S(µ). so the result is sharp. Zero distri1 bution amounts to ﬁnding the limit behavior of n νpn (µ) . i. we have proved that pn (µ. it was shown in Saﬀ–Totik [76] that if ρ is the measure on the unit circle given by the density function sin2 (θ/2). then the number of zeros of pn that lie inside γ tends to inﬁnity (Saﬀ–Totik [76]). where the equilibrium measure ωS(µ) is supported. such that its distance from the convex hull of S(µ) \ {z0 } is δ > 0. Mass points of µ do not necessarily attract zeros (above we have mentioned that they do if they lie on the outer boundary).1]). so the result is sharp (in this case δ = 0).. then for some δ > 0 and all n either pn or pn+1 has no zero in (x0 − δ.1) of the monic orthogonal polynomial pn (µ. then there is at most one zero in K. then Pc(S(µ)) = S(µ). Next put a unit mass at every zero of pn (µ) (counting multiplicity). almost anything can happen with the zeros then). Moreover. Thus. Totik
96
which contradicts the minimality (3. What we have said about the zeros can be sharpened for measures on the real line. if µ is supported on the real line. z). In this case
n→∞
lim κ1/n = log n
1 cap(S(µ))
(9. The normalized arc measure on the unit circle (for which pn (µ. It was shown in Denison–Simon [23] that if x0 ∈ R is not in the support. Section 8. This gives the so called counting measure νpn (µ) on the zero set. assume that Pc(S(µ)) has empty interior and also that there is no Borel set of capacity zero and full µmeasure. the minimal carrier capacity cµ is positive (the cµ = 0 case is rather pathological. 1]. −1/2] ∪ [1/2. 1] the polynomials p2n (µ) have 2 zeros near 0. then for any measure σ that is supported in the open unit disk there is a λ > 0 such that all zeros of the nth orthogonal polynomials with respect to µ = ρ + λσ tend to the unit circle as n → ∞. then there exist two positive constants C and a and a zero zn of pn such that zn − z0  ≤ Ce−an . In general. For example. and if K is a closed interval disjoint from the support. Then pn has at most one zero in the disk {z − z0  < δ/3} (Simon [80. Hence.V. z) = z n ) shows that if the interior of the polynomial convex hull Pc(S(µ)) is not empty. It is also clear that for any symmetric measure µ with S(µ) = [−1. Note that if µ is a symmetric measure on [−1. as was stated.e. z) has at most m − 1 zeros on V . then p2n+1 (0) = 0 for all n. each component of the polynomial convex hull consisting of more than one point attracts inﬁnitely many zeros: if γ is a Jordan curve in Ω such that S(µ) ∩ γ is inﬁnite. x0 + δ).9)
if and only if lim
1 νp (µ) = ωS(µ) n n
. then the zeros may be far away from the outer boundary ∂Ω. if µ lies on the unit circle. Any isolated point in the support that lies on the outer boundary attracts precisely one zero. −1/2] ∪ {0} ∪ [1/2.
10)
n→∞
holds precisely when
1 νp (µ) = ωµ . the important class Reg is deﬁned by the property (9. i.9) is called regular limit behavior. i. cµ = cap(S(µ)) or • dµ/dωS(µ) > 0 (RadonNikodym derivative) ωS(µ) almost everywhere
. lim κ1/n = n 1 cµ (9. then µ ∈ Reg is equivalent to either of • limn→∞ pn (µ)
1/n sup. z ∈ Con(S(µ)) • lim supn→∞ pn (µ.e.e. for which a representation like (9. but for the minimal carrier Green function gµ . Thus..2)) is equivalent to the fact that the zero distribution is the equilibrium distribution. All equivalent formulations of µ ∈ Reg point to a certain “thickness” of µ on its support. z)1/n = 1 for quasievery z ∈ ∂Ω. In a similar way.4) is true. e. asymptotic maximal behavior (see (9. either of the conditions • all Borel sets B ⊆ S(µ) with full measure (i. z)1/n = exp(gC\S(µ) (z. Several regularity criteria are known.
n→∞
lim
Regularity (9. If Ω is a regular set with respect to the Dirichlet problem. Regularity is an important property.e. n n where ωµ is the so called minimal carrier equilibrium measure. 2. . asymptotically minimal behavior of κn (see (9.S(µ) Pn L2 (µ)
1/n
= 1.S(µ)
=1
• For any sequence {Pn } of polynomials of degree n = 1. . . ∞)). and in this case we write µ ∈ Reg..
n→∞
lim
Pn sup.g. µ ∈ Reg is equivalent to either of • limn→∞ pn (µ.Orthogonal Polynomials
1/n
97
in weak∗ sense.. and it is desirable to know “thickness” conditions under which it holds. i.
The last statement expresses the fact that in the nth root sense the L2 (µ) and L∞ norms (on S(µ)) are asymptotically the same.e with µ(B) = µ(S(µ))) have capacity cap(B) = cap(S(µ)).3)).9).
2)
and it is the outer function in the Hardy space on C \ [−1. z) = (1 + o(1)) √ (z + 2π z 2 − 1)n Dµ (z)−1 (10. No necessary and suﬃcient condition for regularity in terms of the size of the measure µ is known. This theory was developed on the unit circle and then was translated into the real line. Regularity holds under fairly weak assumptions on the measure. (10. 1] (i. and it reads that for every η > 0 lim cap x µ([x − 1/n.3)
holds locally uniformly. e.V. 1]). 1] with boundary values Dµ (x)2 = µ′ (x). for it cannot be ∞. x + 1/n]) ≥ e−ηn = 1 4
n→∞
(here 1/4 is the capacity of [0. Outside [−1. if µ is not exponentially small around almost every point). 1] and suppose that the so called Szeg˝ condition o
1 −1
log µ′ (t) √ dt > −∞ 1 − t2
(10.1)
holds.e.4) κn = (1 + o(1)) √ exp 2π −1 1 − t2 2π
. x + r]) ≥ 0
r→0
for almost every x ∈ [0. The Szeg˝ function associated with µ is o Dµ (z) := exp z2 − 1 1 2π
1 −1
log µ′ (t) dt √ z−t 1 − t2
(10. ratio and weak asymptotics
Strong asymptotics Let µ be supported on [−1. Totik
98
is suﬃcient for µ ∈ Reg. where µ′ is the RadonNikodym derivative of µ with respect to linear Lebesgue measure. the leading coeﬃcient κn of pn (µ) is of the form −1 1 log µ′ (t) 2n √ dt . if S(µ) = [0. and lim inf r log µ([x − r. 1] the asymptotic formula 1 pn (µ.
10
Strong. In particular. It expresses a certain denseness of µ. The only existing necessary condition is for the case S(µ) = [0. Szeg˝ proved several asymptotics for the corresponding orthonormal o polynomials pn (µ). Note that this condition means that the integral is ﬁnite.g. then µ ∈ Reg. and under this condition G.. 1]. 1].
away from the support of µ). Ratio asymptotics If one assumes weaker conditions then necessarily weaker results will follow.
we have uniformly on [−1. The Szeg˝ condition is also o o necessary for these results. an asymptotic formula like (10. 1].g. in this respect the extension of Rakhmanov’s theorem given in [22] by Denisov is of importance: if µ′ > 0 almost everywhere on [−1.1). Chapter 6.
log w(ξ) − log w(x) ξ−x
1/2
1 − x2 1 − ξ2
1/2
dξ. Blumenthal’s theorem from [15] states that µ ∈ M (0. A large and important class of measures is the Nevai class M (b. 1/2). 1/2) implies that the support of µ is [−1. 1] and µ′ > 0 almost everywhere on [−1. b (Simon [79]). This is equivalent to ratio asymptotics z−b+ pn+1 (z) = n→∞ pn (z) lim (z − b)2 − 4a2 2
for large z (actually. 1/2) is given in Rakhmanov’s theorem from [75]: if µ is supported in [−1.3) and (10. bn → b.4) is equivalent to (10. M (0. 1] the measure µ has at most countably many mass points converging to ±1. Conversely. However. then µ ∈ M (0. then µ ∈ M (b. and the most important condition ensuring M (0. Thus. The classes M (b. 1] plus at most countably many points that converge to ±1. 1/2).Orthogonal Polynomials
99
If dµ(x) = w(x)dx and h(t) = w(cos t) sin t satisﬁes a DiniLipshitz condition h(t + δ) − h(t) ≤ then with Γw (x) := 1 2π
1 −1
C . and the monograph Nevai [62] contains a very detailed treatment of orthogonal polynomials in this class. 1] and outside [−1. a) (see Nevai [62]). 1] (1 − x2 )1/4 w(x)1/2 pn (x) = + 2 π 1 π cos (n + ) arccos x + Γw (x) − 2 4
O((log n)−ε ). for which the coeﬃcients in the threeterm recurrence xpn (x) = an pn+1 (x) + bn pn (x) + an−1 pn−1 (x) satisfy an → a.  log δ1+ε
ε > 0. e. It is also true that if the limit of pn+1 (z)/pn (z) exists at a single nonreal z. then µ ∈ M (0. a) are scaled versions of each other.
For all these results see Szeg˝ [91]. 1/2)
. a) for some a.
e. 1].g. π
2 lim inf pn (x) ≤ − (µ′ (x))−1/2 (1 − x2 )−1/4 . 1] where 2 pn (x) ≥ (1 + ε) (µ′ (x))−1/2 (1 − x2 )−1/4 ..g. x)} is pointwise bounded. n→∞ π and if En (ε) is the set of points x ∈ [−1. a′ . some parts of Szeg˝’s theory can be proven in a weaker form (see e. 1] lim sup pn (x) ≥
n→∞
2 ′ (µ (x))−1/2 (1 − x2 )−1/4 . a′′ such that
1 n→∞
lim
f (x)p2 (x)µ′ (x)dx = n
f (x)dρ(x)
−1
(10. However. it is not true that the sequence {pn (µ. µ′ > 0 a. then there is an explicitly calculated measure ρ depending only on b. the absolutely continuous part µ′ can be also separately recovered from Tn :
n→∞
lim
Tn (x)µ′ (x) −
2 (1 − x2 )1/2 dx = 0. π then En (ε) → 0 for all ε > 0. and Tn (x)−1 dx converges in the weak∗ sense to dµ on D.e. a2n+1 → a′ and a2n → a′′ . for example.5)
for any continuous function f .V. then given any interval D ⊂ (−1. Simon [79] extended (10. In fact. 1] such that pn (0)/n1/2−ε is unbounded (see Rakhmanov [74]). In these the Tur´n determinants a Tn (x) := p2 (x) − pn−1 (x)pn+1 (x) n play a signiﬁcant role. 58]).6)
. in Delyon–Simon–Souillard [21] a continuous singular measure in the Nevai class was exhibited. 1) the Tur´n a determinant Tn is positive on D for all large n. Furthermore. Totik
100
contains many other measures not just those that are in these theorems. Weak and relative asymptotics Under Rahmanov’s condition supp(µ) = [−1. π
Under Rahmanov’s condition we also have weak convergence. Pointwise we only know a highly oscillatory behavior: for almost all x ∈ [−1. M´t´–Nevai–Totik o a e [57. since for every ε > 0 there is a weight function w > 1 on [−1.5) by showing that if the recurrence coeﬃcients satisfy bn → b. lim f (x)p2 (x)µ′ (x)dx = n 1 π
1 −1
n→∞
f (x) √ dx 1 − x2
(10. and the result in Totik [92] shows that the Nevai class contains practically all types of measures allowed by Blumenthal’s theorem.
.
. . Geronimo–Van Assche [31]. Then pn (β.4) or (10. namely if the change of the argument of Φ as we go around Jk is γk 2π modulo 2π. part of Szeg˝’s theory can be extended to o relative asymptotics.e. We refer the reader to the papers Aptekarev [7]. A general feature of this case is that κn cap(J)n does not have a limit. . z) = Dg (z)−1 lim n→∞ pn (α. Now the point is that these function classes Γn change with n. Suetin and W.4) is played by the reciprocal of the logarithmic capacity of J (see Section 9).Orthogonal Polynomials
101
for any continuous function f . where Φ is the (multivalued) complex Green function of the complement C \ J.3) is played by the conformal map Φ of C \ J onto the exterior of the unit disk. I. then in Γn we consider functions whose change of the argument around Jk is −nγk 2π modulo 2π. . when sequences of orthogonal polynomials corresponding to two measures are compared. The polynomials themselves have asymptotic form pn (z) = cap(J)n Φ(z)n (Fn (z) + o(1)) κn uniformly away from J. and conversely. and where Fn is the solution of an L2 extremal problem involving analytic functions belonging to some class Γn . P. Widom’s theory Szeg˝’s theory can be extended to measures lying on a single Jordan curve o or arc J (see Kaliaguine [40] where also additional outside lying mass points √ are allowed). a2n+1 → a′ and a2n → a′′ with some b. where g is a function such that for some polynomial R both Rg and R/g are Riemann integrable. The fundamentals of the theory were laid out in H. x2 . 1]. Peherstorfer [69]–[72] and Suetin [88]–[89].6) exists for f (x) = x.. then bn → b. J. The functions F in Γn are determined by an H 2 condition plus an argument condition. 1/2). where Dg is Szeg˝’s function with respect to the measure g(x)dx. the theory of Abelian integrals and the theory of elliptic functions. Here is a sample theorem: let α be supported in [−1. x4 . and this is the reason that a single asymptotic formula like (10. S. and the role of 2n in (10. z) o uniformly on C away from [−1. Aptekarev.3) does not hold. 1] and in Nevai’s class M (0. a′′ . i. if (10. Peherstorfer and his collaborators. and since then many results have been obtained by F. and hence so does Fn . and let dβ = gdα. Widom’s paper [97]. Things change considerably if the measure is supported on a set J consisting of two or more smooth curve or arc components J1 . Geronimo. as well as A. For measures in Nevai’s class. Van Assche. in which case the role of z + z 2 − 1 in (10. Jm . The theory has deep connections with function theory. a′ . its limit points ﬁll a whole interval (though if some associated harmonic measures are all rational then the limit points may form a ﬁnite set).
Often only a rough estimate is needed for Christoﬀel functions. and
the recurrence relation for the corresponding orthogonal polynomials. and this gives a onetoone correspondence between unit measures with compact (and inﬁnite) support on the real line and Jacobi operators with bounded entries. it belongs to the Reg class there (see Section 9) and log µ′ is integrable over an interval I ⊂ [−1.
This result is true (see Totik [93]) in the form
n→∞
lim nλn (µ.e. n n
11
Recurrence coeﬃcients and spectral measures
xpn (x) = an pn+1 (x) + bn pn (x) + an−1 pn−1 (x). i. . x) =
dµ(x) . In M´t´– a e Nevai–Totik [59] it was shown that if µ is supported on [−1. . 0 0 a2 b 2 · · · .. x)2
behave somewhat more regularly than the orthogonal polynomials.
Let µ be a unit measure of compact support on the real line. 1] and it is a doubling measure. 1] √ 1 − x2 1 λn (µ. 1]. . µ(2I) ≤ Lµ(I) for all I ⊂ [−1. . . x) ∼ µ (∆n (x)) . . and such a one is provided in Mastroianni–Totik [56]: if µ is supported on [−1. x ∈ I
when the support is a general compact subset of R. . Every such Jacobi operator is a bounded selfadjoint operator on l2 . hence
. . µ ∈ Reg and log µ′ ∈ L1 (I).e. ∆n (x) = + 2. . Totik Asymptotics for Christoﬀel functions The Christoﬀel functions
n
102
λn (µ. then for almost all x ∈ I
n→∞
lim nλn (µ. . then uniformly on [−1. x) = π
1 − x2 µ′ (x). .V. 1].. 1]. . . dωsupp(µ) (x)
a. x)−1 =
k=0
pk (µ. We have already mentioned in Section 6 that µ is the spectral measure for the Jacobi matrix b 0 a0 0 0 · · · a0 b 1 a1 0 · · · J = 0 a1 b 2 a2 · · · . where 2I is the twice enlarged I.
hence in this case the Jacobi operator is a compact perturbation of J0 . 1] plus some additional mass points converging to ±1 and µ′ (x) > 0 for almost all x ∈ [−1. and techniques and questions from both areas are relevant. 1] then there is a µ ∈ M (0. where the Jacobi matrix is denoted by J0 . and this was the Nevai class M (0. Kolo mogorov and others.
(ii) if µ′ is the absolutely continuous part of µ on [−1. bn → 0 (this is just rescaling. If the sequences {an } and {bn } have limits then we may assume an → 1/2. 1] plus some additional mass points Ej converging to ±1. 1] plus some additional mass points converging to ±1 (called Blumenthal’s theorem in orthogonal polynomials. b) (Denisov [22]. Nevai–Totik [63]). continuously singular measures or measures that are given by a continuous density which is positive on a set of measure < ε. it is a special case of Weyl’s theorem in operator theory on the invariance of the essential spectrum under compact perturbation).Orthogonal Polynomials
103
operator theory and orthogonal polynomials meet at this point. After numerous works in the subject by Szeg˝. M (0. if ν is any measure with support [−1. a special role is played by the Chebyshev case an = 1/2 and bn = 0. then
1 −1
(log µ′ (t)) 1 − t2 dt > −∞.1)
n
n
if and only if the following conditions hold:
± (i) the support of µ is [−1. Shohat.
± (iii) for the mass points Ej lying outside [−1. 1/2) implies that the support of µ is [−1. 1]. 1/2) contains discrete measures. and one of the main questions of the theory is how properties of J − J0 are reﬂected in the spectral measure µ. Krein. Simon (note that µ is assumed to have total mass 1): (an − 1/2)2 + b2 < ∞ n (11. 1/2) is known. For example. We have already mentioned in Section 10 that µ ∈ M (0. this important class seems to contain all sorts of measures. Strengthening the condition µ ∈ M (0. 1/2)). a complete characterization for J − J0 being a HilbertSchmidt operator was given in [43] by R. if the support of µ is [−1. 1/2) which is absolutely continuous with respect to ν. In particular. 1] we have + Ej − 13/2 + − Ej + 13/2 < ∞. Information between the measure µ and the recurrence coeﬃcients might be called the spectral analysis of orthogonal polynomials. No spectral characterization of µ ∈ M (0. Killip and B.
j
j
. Conversely. 1]. 1/2) can be done in several ways. In this. then µ ∈ M (a. Rakhmanov [75]. Geronimus.
Nevai and E. when supp(µ) = [−1. J.V.1) and to the (conditional) convergence of the series n bn . I.
log µ′ (t) √ dt > −∞ 1 − t2
(11. Geronimo–Van Assche [31]. Osilenker and others. x ∈ R. Freud started to investigate these weights in the sixties and seventies.
then Szeg˝’s condition o
± holds. Wimp. 1].g. 1] we have + Ej − 11/2 + − Ej + 11/2 < ∞. 1] then Szeg˝’s condition o (11. Rahmanov were the primary cause of the sudden revitalization of the theory of orthogonal polynomials since the early 1980’s. B. Peherstorfer–Steinbauer [73]. for the mass points Ej lying outside [−1. A. but they are ∞ asymptotically periodic in the sense that for some k all the sequences (akn+j )n=1 ∞ and (bkn+j )n=1 . α > 0. N. Totik It was also shown in Killip–Simon [43] that if J − J0 is trace class. j = 1. In this case the essential support of the spectral measure lies on several intervals. Bustoz.2)
j
j
and lim sup(2n a1 · · · an ) > 0. Charris–Ismail [17]. P. J. J. . Ismail. but they independently appeared also in the Russian literature and in statistical physics. Peherstorfer [70]. Akhiezer.2) is equivalent to (11. Charris. then Szeg˝’s condition automatically o holds if (11.3) is true. There are numerous papers on this subject by M. i. 1/2). In the last 20 years D. J. F. A. . We get Freud weights when Q(x) = xα . Akhiezer [1] for details and for further references. E. . Van Assche. A. For simplicity we shall ﬁrst assume that Q is even. N. W. One can safely say that some of Freud’s problems and the work of P.. Peherstorfer. R. AlSalam.
n
104
an − 1/2 +
1 −1
n
bn < ∞.
n
(11. Actually. Geronimo. n (an − 1/2) and There is also an extended theory of orthogonal polynomials with several different applications when the recurrence coeﬃcients do not converge. . G. These are related to so called sieved orthogonal polynomials and to orthogonal polynomials generated by polynomial mappings. k converge.
12
Exponential and Freud weights
These are weight functions of the form e−2Q(x) . The conclusion is also true if µ ∈ M (0. and Erd˝s weights o if Q tends to inﬁnity faster than any polynomial as x → ∞. Steinbauer. where x is on the real line or on some subinterval thereof. Lubinsky
.3)
If the support of µ is contained in [−1. H. see e.e.
McLaughlin.
. Van Assche. in which case things are automatically reduced to a ﬁnite interval which is the support of a weighted energy problem. A. W. For Freud weights one can also make the substitution x → n1/λ x and λ go to orthogonality with respect to the varying weight e−nx . Rahmanov and H. and let pn be the nth orthogonal polynomial with respect to w2 (on (−∞. Lubinsky [49].1)
α Γ 2
1 2
2Γ
α 1 + 2 2
. S.3)
among all probability measures compactly supported on R. E. These contracted zeros asymptotically have the Ullman distribution dµw (t) α := dt π
1 t
uα−1 √ du. 45]. E. see e. w(x) = exp(−Q(x)) are spreading out and the largest zero is very close to an .n we have xn. things happen on a ﬁnite subinterval [−an .2)
This measure µw minimizes the weighted energy log 1 dµ(x)dµ(t) + 2 x − t Qdµ (12. Saﬀ. an ] techniques developed for [−1. Venakides and X. In this case an = n1/α γα . and to describe zero distribution we divide (contract) all zeros xn. Kriecherbauer. Zhou ([20]) to give complete asymptotics when Q is analytic. pn (z) = κn z n + · · ·. then (Lubinsky–Saﬀ [50])
n→∞
lim κn π 1/2 2−n e−n/α n(n+1/2)/α = 1. To describe the distribution of the zeros and the behavior of the polynomials one has to make appropriate contractions. Let us consider ﬁrst the case of Freud weight w(x) = exp(−xα ). u 2 − t2
t ∈ [−1. see Saﬀ–Totik [77]).. One can roughly say that because of the fast vanishing of the weight around inﬁnity. and on [−an .R. K. If κn is the leading coeﬃcient of pn . Van Assche [94]. 1 − t2
(12. for the largest zero xn. Levin–Lubinsky [46. which tends to ∞. 1].Orthogonal Polynomials
105
with coauthors have conducted systematic studies on exponential weights. T.
α γα := Γ
an tQ′ (an t) √ dt. In the mid 1990’s a new stimulus came from the Riemann–Hilbert approach that was used together with the steepest descent method by P. ∞)).g. Levin. N. T.i by n1/α γα . B.n /n1/α → γα as n → ∞.e. The an are the so called MhaskarRahmanovSaﬀ numbers deﬁned by n= 2 π
1 0
The zeros of pn (w2 ). an ] (depending on the degree of the polynomials). Deift. 1] are applied. we should mention the names E. i.
Thus. It is a general feature of exponential weights that the behavior of zeros of the polynomials is governed by the solution of a weighted energy problem (weighted equilibrium measures. Mhaskar. Lubinsky–Saﬀ [50].
(12.
(x − a−n )(an − x) Q′ (x) dx. e. The weight does not even have to be deﬁned on all R. x ∈ (−∞. 1]. and the deﬁnition of the Mhaskar– Rahmanov–Saﬀ numbers a±n is n = 1 π 1 π
an a−n an a−n
2 1 √ cos 4 π 1 − x2
π 1 arccos x + nπµw ([x. Things become more complicated for nonFreud weights. x ∈ [0. 1). n1/α γα x) exp(−nγα xα )−
−
uniformly on any subinterval of (−1. ∞) expk (xβ ) − expk (0). Erd˝s and Pollaczek weights such as o (a) nonsymmetric Freudtype weights Q(x) = xα . an ] is the support of µn and µn /n will play the role of the measure µw from (12. 0)
. in [46] a theory was developed by Levin and Lubinsky that simultaneously includes far reaching generalizations of nonsymmetric Freud.g. On the real line we have a Plancherel–Rotach type formula
α n1/2α pn (wα . 0).V.
(b) nonsymmetric Erd˝s weights such as o Q(x) = expl (xα ) − expl (0). In this case the weight is not necessarily symmetric. but the corresponding results are of the same ﬂavor. This latter is so called nth root asymptotics.2) above.3) for all measures µ with total mass n. x ∈ [0. ∞) xβ . x ∈ (−∞. Strong asymptotics for pn (z) on diﬀerent parts of the complex plane was given using the Riemann–Hilbert approach. and if µn is the solution then [a−n . but under some conditions (like Q being convex or xQ′ (x) being increasing for x > 0 and an analogous condition for x < 0) the relevant weighted equilibrium measure’s support is an interval. (x − a−n )(an − x)
Now one solves the weighted equilibrium problem (12. Totik and we have
n→∞
106
lim pn (n1/α γα z)1/n
1
= exp log z +
z 2 − 1 + Re
0
zuα−1 √ du z 2 − u2
locally uniformly outside [−1. see Deift [20] and Kriecherbauer–McLaughlin [44] and the references there. while the former is strong asymptotics. 1]) − 2 4
→0
0 =
xQ′ (x) dx.
Most arguments for the standard theory fail in this case. Case I: The discrete case. this is the shortest interval on which the supremum norm of a weighted polynomial is attained: wPn sup = wPn sup.[−a−n .
107
In all cases the interval [a−n . x ∈ [0.g. e. r. . . Perhaps the most natural one is smooth data ﬁtting. 52] and the references therein).g. In this case µ0 is some “strong” measure. a) (see Section 10). e. G. k = 0. in addition. . . Lopez and A. e.
. There are several motivations for this kind of orthogonality.
(13.g. for the case r = 1. 0). . from the Nevai class M (b. e. Nonetheless. then the asymptotic zero distribution of Qn is ωE0 .1). and µ0 . an ] is where things happen.g. g) =
k=0
f (k) g (k) dµk . Marcell´n. µk are ﬁnite discrete measures. MartinezFinkelshtein a has been particularly active in developing this area (see the surveys Marcell´n– a Alfaro–Rezola [51] and MartinezFinkelshtein [53. These numbers a±n are everywhere in the theory. 1. µ1 ∈ Reg (see Section 9) it was shown in Gautschi–Kuijlaars [29] that the asymptotic distribution of the zeros of the derivative Q′ is the equilibrium measure ωE0 ∪E1 . . . 1) x ∈ (−1. E0 ⊆ E1 . The Spanish school around F. In general. n where Ei is the support of µi .1)
where µk are given positive measures. 1 (which also have to be assumed to be regular). and µ1 . and there are essentially two important cases which have been understood to a satisfactory degree. . i = 0. In this section let Qn (z) = z n + · · · denote the monic orthogonal polynomial with respect to the Sobolev inner product (13. It is not even known if the zeros are bounded if all the measures µk have compact support. it is no longer true that the zeros lie in the convex hull of the support of the measures µk . sup pn (x)w(x)x − an 1/4 x − a−n 1/4 ∼ 1. Furthermore.
x
13
Sobolev orthogonality
r
In Sobolev orthogonality we consider orthogonality with respect to an inner product (f. if. and qn (µk ) the monic orthogonal polynomials with respect to the measure µk .Orthogonal Polynomials with expl is the ltimes iterated exponential function. expk ((1 − x)−β ) − expk (0).an ] for all polynomials of degree at most n. both the algebraic and the asymptotic/analytic situation is quite complicated. or (c) nonsymmetric Pollaczek type weights that vanish fast at ±1 such as Q(x) = expl ((1 − x)−α ) − expl (1).
in this case the measures µ0 . µk are all supported on o the same smooth curve or arc J. . In this case the kth derivative of Qn satisﬁes.V. Totik
108
It turns out that then the situation is similar to adding these discrete measures to µ0 (the new measure will also be in the same Nevai class). locally uniformly in the complement of J. Case II: The Szeg˝ case. Suppose now that µ0 . For ecample. That is. . . and considering standard orthogonality with respect to this new measure.
n→∞ (k)
lim
Q(k) n
1/n sup. But the polynomial Qn (t) = n(n − 1) · · · (n − k + 1)tn−k + · · · is a monic polynomial times the factor n(n − 1) · · · (n − k + 1) ∼ nk . but not more than what happens when adding mass points to µ0 .2)
among all monic polynomials of degree n. z) [Φ (z)]m−k lim where Φ is the conformal map that maps C \ J onto the complement of the unit disk. if r = 1. so everything else will be negligible. Under the much less restrictive assumption that µ0 ∈ Reg (see Section 9) and the other measures µk are supported in the support E of µ0 it is true (L´pez–PijeiraCabrera–Izquierdo [47]) that the asymptotic zero distribution of o (k) Qn is the equilibrium measure ωE for all k. There are results for compensation of this nk factor which lead to Sobolev orthogonality with respect to varying measures.E
= cap(E). then Qn (z) =1 lim n→∞ qn (µ0 + µ1 . . . the asymptotic formula 1 Qn (z) . µr−1 do not appear in the asymptotic formula. The reason for this is the following: Q = Qn minimizes
r (k)
(Q. In this discrete case the Qn ’s satisfy a higher order recurrence relation. . . . Q) =
k=0
Q(k) 2 dµk
(13. = ′ n→∞ nk qn−k (µr . and they satisfy Szeg˝’s condition there (see o Section 10). the Sobolev orthogonal polynomials diﬀer from those of the measure µ0 . hence this case is also related to matrix orthogonality (see the end of the Section 16).
. and this factor is dominant for k = r. Thus. only µr matters. z) holds uniformly on compact subsets of C \ supp(µ0 + µ1 ). while q = qn−k (µk ) minimizes q2 dµk among all monic polynomials of degree n − k.
or positive but lies on a complex curve or arc and orthogonality is considered without complex conjugation. The techniques developed for exponential weights and for Sobolev orthogonality were combined in Geronimo–Lubinsky–Marcellan [30] to prove strong asymptotics for Sobolev orthogonal polynomials when r = 1 and µ0 = µ1 are exponential weights. Then pn satisﬁes the nonHermitian orthogonality relation pn (x)xj dµ(x) = 0.3)
In this nonHermitian case even the GramSchmidt orthogonalization process may fail. .. 1. we have lim Qk (z)1/n = egC\E (z) n
n→∞
where gC\E is the Green function for this unbounded component. consider the diagonal Pad´ approximant to the Cauchy e transform dµ(t) f (z) = z−t
(14. . . and then pn is deﬁned as the solution of the orthogonality condition (14. k = 0. • µ is again nonpositive or complexvalued. . resp.Orthogonal Polynomials
109
and hence.2). away from the zeros in the unbounded component of the complement of E. consider polynomials pn and qn of degree at most n such that f (z)pn (z) − qn (z) = O(z −n−1 ) at inﬁnity.
More generally.e. 1. n − 1. . i. .2).2)
of a signed or complexvalued measure. . (14. .e. . but we shall restrict our attention to complex measures and orthogonality (14. pn (z)z k dµ = 0.1). n − 1.. . As an example. 1. (14. . j = 0. one could consider nonpositive inner products. i. which give a system of homogeneous equations for the
. n − 1. (14.
14
NonHermitian orthogonality
We refer to nonHermitian orthogonality in either of these cases: • the measure µ is nonpositive or even complexvalued and we consider pn with k = 0. .1) pn (z)z k dµ = 0.
P. Sorokin. the zeros of pn from (14. [67]. Nonetheless. A similar result holds on the support of the measure. B. and the argument of µ. dµ(t)/dµ(t). A. Stahl obtained asymptotics for nonHermitian orthogonal polynomials even for varying measures and gave several applications of them to Pad´ e approximation. Nikishin. the asymptotic distribution of the zeros is again the equilibrium distribution of the support of µ under regularity conditions on µ. Rahmanov. is of bounded variation (Baratchart–K¨ stner–Totik [12]). in the simple case dµ(x) = (x − cos πα1 )(x − cos πα2 )(1 − x2 )−1/2 dx. A. Totik
110
coeﬃcients of pn . with an analytic g. x + δ]) the property limδ→0 log a(δ) = 0 holds (Stahl [86]). The
. x ∈ [−1. A. In [84]–[86] H. for z ∈ C \ [−1.. 1]. A.3) are dense on the whole complex plane (compare this with the fact that for positive µ all zeros lie in [−1. 1]. It is a e relatively new area where we have to mention the names of E. such that for an arbitrary prescribed asymptotic behavior some subsequence {pnk } will have this zero behavior. 1]. J. i. For example. Geronimo and W. pn may have smaller degree than n. this is the case if • µ belongs to the Reg class (see Section 9). see Aptekarev–Van Assche [10]. Van Assche (see the survey [95] by W. I. Gonchar and E. M. Aptekarev. Van Assche and the references there and the paper Gonchar–Rakhmanov [37]). A. J. or u • dµ(x) = g(arccos x)(1 − x2 )−1/2 dx. V. 1]). Suetin [38].e. and things can get pretty wild with this kind of orthogonality. A. For example. In Stahl [86] it was shown that it is possible to construct a complex measure µ on [−1. For a recent Riemann–Hilbert approach see the paper [10] by A. the argument of µ is uniformly continuous and for a(δ) = inf x∈supp(µ) µ([x − δ.
with 0 < α1 < α2 < 1 rationally independent algebraic numbers. When the measure µ is of the form dµ(x) = g(x)(1 − x2 )−1/2 dx.2)) was proved by J. 1]. x ∈ [−1. o Gonchar and S. A. 1]. Thus. Kuijlaars. Van Assche. as well as for the case of varying weights. I. x ∈ [−1. and satisﬁes g(θ + δ) − g(θ) ≤ K log δ−1−δ . a strong asymptotic formula of the form √ 1 (z + z 2 − 1)n 1 log µ′ (t) pn (z) √ = (1 + o(1)) Dµ (z)−1 exp dt κn 2n 2π −1 1 − t2 (with Dµ the Szeg˝ function (10. or • µ is supported on ﬁnitely many intervals.
15
Multiple orthogonality
Multiple orthogonality comes from simultaneous Pad´ approximation.V. Nuttall [66]. N. g is bounded away from zero and inﬁnity. Aptekarev and W.
. . . .Orthogonal Polynomials
111
analogues of many classical concepts and properties have been found. . Van Assche [95]. . . Types and normality On R let there be given r measures µ1 . xm1 −1 w1 (x). . in the 2 multiple Hermite case the measures are dµ(x) = e−x +cj x dx.g. . e In type I multiple orthogonality we are looking for polynomials Qn. . This happens precisely if each Qn. n − 2. . so there is a nontrivial solution.j is of exact degree nj − 1. Asymptotic behavior of multiple orthogonal polynomials is not fully understood yet due to the interaction of the diﬀerent measures. . r. e. k = 1. . 1. µr with ﬁnite moments and inﬁnite support. j = 1. . . bj ] that are disjoint except perhaps for their endpoints. . . . If the rank of the system is n − 1. . . . .
j=1
k = 0. r. Normality also holds if dµj = wj dµ with a common µ supported on some interval [a. . then the solution is unique up to a multiplicative factor. 0. bj ). xmr −1 wr (x)
. if the µj ’s are supported on intervals [aj . every nontrivial linear combination of the functions w1 (x). xw1 (x). In type II multiple orthogonality we are looking for a single polynomial Pn of degree n such that xk Pn (x)dµ1 (x) = . . xw2 (x). Van Assche’s Chapter 23 in [39] and the references therein. . This is again equivalent to Pn being of exact degree n. n1 − 1
These are n homogeneous linear equations for the n + 1 coeﬃcients of Pn . so we just speak of normality. in which case the index n is called normal. . . . . and for all mj ≤ nj .j (x)dµj (x) = 0. and also the analogues of the classical orthogonal polynomials are known. For the existing results see Aptekarev [8]. w2 (x).
These orthogonality relations give n − 1 homogeneous linear equations for the n coeﬃcients of the r polynomials Qn.j . n is normal for type I orthogonality precisely when it is normal for type II. and again if the solution is unique up to a multiplicative constant. for example. in this case Pn has nj simple zeros on (aj . b]. . . . There are two types of multiple orthogonality corresponding to the appropriate HermitePad´ approximation. . nr ) of nonnegative integers with norm n = n1 + · · · + nr .j of degree nj − 1 for each j = 1. . . k = 1. then n is called normal. in fact. . such that
r
xk Qn. . This is the case. . and consider multiindices n = (n1 . nr − 1. xk Pn (x)dµr (x) = 0.
r. In this case Pn has n − 1 zeros on [a. .j (x)wj (x). then similar recurrence relations hold in case of type I orthogonality for
r
Qn (x) =
j=1
Qn. permutation of 1. b]. the left hand side depends only on m = mn and not on the particular choice of the sequence mj leading to it.g. To describe an analogue of the ChristoﬀelDarboux formula let {mj } be a sequence of multiindices such that m0 is the identically 0 multiindex. . . let ej = (0. Under the normality assumption if Pn is the monic orthogonal polynomial.
n−1 r
(x − y)
k=0
Pk (x)Qk (y) = Pm (x)Qm (y) −
hm
(j)
(j)
j=1
hm−ej
Pm−ej (x)Qm+ej (y). . π(r) is an arbitrary.0 Pn (x) +
j=1
an. and then the previous integral is not zero (under the normality condition). e.V. . then for any k
r
xPn (x) = Pn+ek (x) + an. The orthogonal polynomials with diﬀerent indices are strongly related to one another. b] (this means that these functions form a so called Chebyshev system there). 0) where the single 1 entry is at position j. . type I and type II are related by a biorthogonality property: Pn Qm dµ = 0 except for the case when m = n + ek for some k.
Thus.
where (m)j denotes the jth component of the multiindex m. Totik
112
has at most m1 + · · · + mr − 1 zeros on [a. Set Pj = Pmj . .
. . Pn+ek (x) − Pn+el (x) is a constant multiple of Pn (x).j Pn−επ(1) −···−eπ(j) (x). Recurrence formulae To describe recurrence formulae. again with m = mn . . .
Also.
Another recurrence formula is
r
xPn (x) = Pn+ek (x) + bn. . but ﬁxed.0 Pn (x) +
j=1
bn. If dµj = wj dµ. . 1. Then (see Daems– Kuijlaars [19]). .
where π(1). . .j Pn−εj (x). 2. . and mj+1 coincides with mj except for one component which is 1 larger than the corresponding component of mj . Qj = Qmj+1 and with m = mn h(j) := m Pm (x)x(m)j dµj (x).
. . . . . half plane. 0
. . −wr (x) 0 0 ··· 1 0 z n1 0 . . . .
· · · z nr
. 0 0 0 z −n2 . z .
0 0 0 .. . . 0 z 0 0
n2
the behavior at inﬁnity is of the form −n z 0 1 0 Y (z) = I + O . . . .. then one requires that • Y is analytic on C \ R. 0 ··· ··· ··· . . . then we have Y + (x) = Y − (x)S(x).
Y (z) =
I +O
· · · z −nr
. For type I orthogonality the transfer matrix is 1 0 0 ··· 0 −w1 (x) 1 0 · · · 0 S(x) := −w2 (x) 0 1 · · · 0 . 0 0 z −n1 0 . . . . . .j+1 /2πi.
. . .Orthogonal Polynomials The Riemann–Hilbert problem
113
There is an approach (see Van Assche–Geronimo–Kuijlaars [96]) to both types of multiple orthogonality in terms of matrixvalued Riemann–Hilbert problem for (r + 1) × (r + 1) matrices Y = (Yij (z))r i. . . . . . .. and the other entries are also explicit in terms of the Pn ’s and wj ’s (all other entries are either a constant multiple of Pn−ek or a Cauchy transform of its multiple with wj ). 0
··· ··· ··· . . .
and the multiple orthogonal polynomials Qn. . 0 0 0 . . . . . . • if Y ± (x) denote the limit of Y (z) as z → x ∈ R from the upper. .. . If dµj (x) = wj dx. respectively the lower. . . . . where 1 w1 (x) w2 (x) · · · wr (x) 0 1 0 ··· 0 0 0 1 ··· 0 S(x) := .j are Y1. 0 0 0 ··· 1 • as z → ∞ 1 z z n 0 0 .
The ﬁrst entry Y11 (z) is precisely the orthogonal polynomial Pn of type II.j=0 . . .
I. .. Aptekarev and E. M. 1. When the leading coeﬃcient matrix Cn is nonsingular. From now on we ﬁx the dimension to be N . P (t) = . one can write P (t) = Cn tn + · · · + C0 with numerical matrices Cn . We assume that all moments of W are ﬁnite. · · · pN N (t)
with polynomial entries pij (t) of degree at most n is called a matrix polynomial of degree at most n. Dur´n and his coauthors (see also the work [9] by a A. . With such a matrix we can deﬁne a matrix inner product on the space of N × N matrix polynomials via (P. . µN 1 (t) · · · µN N (t)
of complex measures deﬁned on (or part of) the real line is positive deﬁnite if for any Borel set E the numerical matrix W (E) is positive semideﬁnite. A matrix µ11 (t) · · · µ1N (t) .
and if (P. but there is an unexpected richness which is still to be explored.V. Q) = P (t)dW (t)Q∗ (t). . . . P ) is nonsingular for any P with nonsingular leading coeﬃcient. For all the results in this section see L´pezRodriguez–Dur´n [48] and Dur´n– o a a Gr¨ nbaum [25] and the numerous references there. The theory shows many similarities with the scalar case. C0 of size N × N . pN 1 (t) · · · p1N (t) . . . which are orthonormal with respect to W :
∗ Pn (t)dW (t)Pm (t) =
0 I
if n = m if n = m. Nikishin). . .
. I will denote the N × N unit matrix and 0 stands for all kinds of zeros (numerical or matrix). . Alternatively. . The number t = a is called a zero of P if P (a) is singular. and the multiplicity of a is the multiplicity of a as a zero of detP (a). . . . then P has nN zeros counting multiplicity. u Matrix orthogonal polynomials An N × N matrix p11 (t) . W (t) = . . but the degree n can be any natural number. Totik
114
16
Matrix orthogonal polynomials
In the last 20 years the fundamentals of matrix orthogonal polynomials have been developed mainly by A.. . then just as in the scalar case one can generate a sequence {Pn }∞ of matrix polyn=0 nomials of degree n = 0. .
. . ··· . i. 1. The threeterm recurrence formula easily yields the ChristoﬀelDarboux formula:
n−1
(w − z)
k=0
∗ ∗ ∗ Pk (z)Pk (w) = Pn−1 (z)An Pn (w) − Pn (z)A∗ Pn−1 (w). . n
and With the recurrence coeﬃcient matrices An .. The sequence {Pn } is determined only up to left multiplication by unitary matrices. then the polynomials Un Pn also form an orthonormal system with respect to W . .
also satisfy the same recurrence and are orthogonal with respect to some other matrix measure. satisfy (16. . the analogue of Favard’s theorem is also true: if a sequence of matrix polynomials {Pn } of corresponding degree n = 0. . . Conversely.Orthogonal Polynomials
115
and here Pn has nonsingular leading coeﬃcient matrix. . .e.
. Jacobi matrix B0 A0 0 0 A∗ B1 A1 0 0 ∗ 2 J = 0 A1 B∗ A2 0 0 A2 B2 . . (16. 2.1)
where An are nonsingular matrices. . . n
The orthogonal polynomials Qn of the second kind Qn (t) = Pn (t) − Pn (x) dW (x). .1) with nonsingular An and Hermitian Bn . . n−1 n
Bn one can form the block ··· ··· ··· . . 2. t−x n = 1. . n n ≥ 0.
∗ Qn (t)Pn−1 (t) − Pn (t)Q∗ (t) ≡ A−1 . . these orthogonal polynomials satisfy a threeterm recurrence relation tPn (t) = An+1 Pn+1 (t) + Bn Pn (t) + A∗ Pn−1 (t). n n−1 k=0 ∗ ∗ ′ ∗ ′ Pk (z)Pk (z) = Pn−1 (z)An Pn (z) − Pn (z)A∗ Pn−1 (z). Threeterm recurrence and quadrature Just as in the scalar case. For them we have
∗ ∗ Pn−1 (t)An Qn (t) − Pn (z)A∗ Qn−1 (t) ≡ I. and Bn are Hermitian. . if Un are unitary matrices.. then there is a positive deﬁnite measure matrix W such that the Pn are orthonormal with respect to W .. . n
from which for example it follows that
∗ ∗ Pn−1 (z)An Pn (z) − Pn (z)A∗ Pn−1 (z) = 0.
Totik
116
The zeros of Pn are real and they are the eigenvalues (with the same multiplicity) of the N truncated block Jacobi matrix (which is of size nN ).k
holds for every matrix polynomial P of degree at most 2n − 1. some remarkably rich nonreducible families have been obtained by A. the nth orthogonal polynomial with respect to µi . The distribution of the zeros themselves will be 1/N times the trace of the matrix measure of orthogonality for another sequence of matrix orthogonal polynomials Rn satisfying tRn (t) = ARn+1 (t) + BRn (t) + ARn−1 (t). B for all n. Many matrix orthogonal polynomials in the literature can be reduced to this scalar case. They found families of matrix orthogonal polynomials that satisfy second order (matrix) diﬀerential equations just like the classical orthogonal polynomials.k .k . and lk is the multiplicity of xn.V.
. with appropriate modiﬁcations for n < 2. Recently however. i.k ) 1≤k≤m
are positive semideﬁnite of rank lk ..e.B (t) z−t
locally uniformly outside the cluster set of the zeros. If we write xn. Duran and F. for the diﬀerent zeros of Pn .k )Γn.k )Qn (xn. then the matrices Γk = 1 (l −1) (Adj(Pn (t))) k (xn. Families of orthogonal polynomials If the matrix of orthogonality is diagonal (or similar to a diagonal matrix) with diagonal entries µi . and with them the matrix quadrature formula
m
P (t)dW (t) =
k=1
P (xn. then the orthogonal matrix polynomials are also diagonal with ith entry equal to pn (µi ). If a is a zero then its multiplicity p is at most N . (det(Pn (t)))(lk ) (xn. then
−1 Pn (z)Pn−1 (z)A−1 → n
dWA. Gr¨ nbaum (see [25] and the references u therein). the rank of Pn (a) is N − p. If we assume that An → A. 1 ≤ k ≤ m. where WA. which may play the role of the classical orthogonal polynomials in higher dimension.B (t) is the measure matrix of orthogonality for the sequence of matrix orthogonal polynomials Sn with recurrence coeﬃcients A. Their starting point was a symmetry property between the n ≥ 2. which satisfy the threeterm recurrence tSn (t) = A∗ Sn+1 (t) + BSn (t) + ASn−1 (t). Bn → B where A is nonsingular. and the space of those vectors v for which Pn (a)v = 0 is of dimension p.k ).
..0 pn (t) +
k=1
(cn.
i. . .k pn+k (t)).Orthogonal Polynomials
117
orthogonality measure matrix and a second order diﬀerential operator.m (p) =
s
asN +m ts .m takes those powers where the exponent is congruent to m modulo N . 1.
a2 n . Now suppose that {pn }∞ is a sequence of scalar polynomials of correspondn=0 ing degree n = 0. and the threeterm recurrence (16. To describe this we need the following operators on polynomials p: if p(t) = k ak tk . Here is one of them: N = 2. An+1 := where
2 γn := 1 +
n+1 2
γn+3 /γn+2 0
aγn+2 γn+1 γn /γn+1
..1) holds with Bn = 0.
=
−2n − 4 2a(2n + 1) 0 −2n
There is an explicit Rodrigues’ type representation for the polynomials themselves. 2 2
Connection with higher order scalar recurrence Matrix orthogonality is closely connected to (2N + 1)term recurrences for scalar polynomials. the measure matrix (more precisely its density) is H(t) := e−t
2
1 + a2 t4 at2
at2 1
. and suppose that this sequence satisﬁes a (2N + 1)term recurrence relation
N
tN pn (t) = cn. They worked out several explicit examples. from a polynomial the operator RN. then RN. removes the common factor tm and changes tN to t.
.
t ∈ R. The corresponding Pn (t) satisﬁes
′′ ′ PN (t) + Pn (t)
−2t 4at 0 −2t
+
Pn (t)
−4 2a 0 0 Pn (t).k pn−k (t) + cn+k.e.
where a ∈ C \ {0} is a free parameter.
[5] M. and periodic motions of Toda chains. 989–992. (Russian) Mat. Some asymptotic properties for orthogonal polynomials with respect to varying measures. Alfaro and L. . Orthogonal polynomials on several intervals. .0 (pnN ) ··· RN.N −1 (pnN + 1) Pn (t) = . (N. Comput. cn. L. 1(1960). 53(1988). . Then RN.0 (pnN +1 ) ··· RN. P. n = 0. in: Orthogonal Polynomials: Theory and Practice. M. Vigil. I. 1965 [3] W.
satisfy a (2N + 1)recurrence relation of the above form. Aptekarev. J. Theory. Approx. Kluwer. 1. 195–197.m. 294. Ahiezer. Appl. Special Functions. Dokl. Andrews.0 (pnN +N −1 )
118
· · · RN. Akhiezer. Sb.N −1 (pnN ) RN.. Tur´n. Multiple orthogonal polynomials. . [2] N. then the scalar polynomials
N −1
pnN +m (t) =
j=0
tj Pn. . Cambridge University Press. Askey and R. Nevai (eds). AlSalam.. a Theory.N = 0 (and pk (t) ≡ 0 for k < 0). Ismail and P. Alfaro. . P. I.j=0 matrix polynomials.0 is real. 423–447. NATO ASI Series C. Oliver and Boyd. Montaner and J. [8] A.
References
[1] N. RN. J. Roy. M. Math. Asymptotic properties of polynomials orthogonal on a system of contours.. E. 125(167)(1984). Characterization theorems for orthogonal polynomials. J.
0 ≤ m < N. if Pn = (Pn.N −1 (pnN + N − 1)
is a sequence of matrix orthogonal polynomials with respect to a positive deﬁnite measure matrix. I. Edinburgh. .m. . Nauk SSSR.j )N −1 is a sequence of orthonormal m. (Russian).
. Varona. 2. R. Dordrecht. J. 9–12. . [6] G. 1999. . Conversely. 134(1960). New York. Bello. [4] M. A. Encyclopedia of Mathematics and its Applications.j (tN ). Totik where cn. Solution of a problem of P. M. 22–34. I. . 231–258.V. Approx. The Classical Moment Problem and Some Related Questions in Analysis.S. translated as Soviet Math. 135(2005). Aptekarev. Akad.). Dokl. [7] A. 1990. 99(1988).
I. R. Generalized polynomial mappings. Ratio and relative asymptotics of polynomials o orthogonal on an arc of the unit circle. Mat. Denisov. 283–309. 1898. S. Souillard. Bochner. An introduction to orthogonal polynomials. ¨ [15] O. Annales Inst. M. 13(1969). 340(1993). Delyon. [18] T. 1978. Inaugural Disu sertation. 55(2005). A ChristoﬀelDarboux formula for multiple orthogonal polynomials. Math. 847–852. Ismail. L´pez. J. 327–358. Orthogonal Polynomials and Random Matrices: A Riemann– Hilbert Approach. [14] M. Math. Approx. 216– 244. Soc. (Russian) [10] A. B. Scalar and matrix Riemann–Hilbert approach to the strong asymptotics of Pad´ approximants and complex ore thogonal polynomials with varying weights. Ratio asymptotics for polynomials orthogonal on o arcs of the unit circle. [11] R. On Rakhmanov’s theorem for Jacobi matrices.. Approx. Charris and M. Vol.. Sb. Zeitschrift. VII. Gordon and Breach Science Publishers. H. Baratchart. A. [12] L. Approx. [13] D. B. e 42(1985). L´pez. Uber die Entwicklung einer willk¨rlichen Funktion nach u 0 den Nennern des Kettenbruches f¨r −∞ [φ(ξ)/(z − ξ)] dξ. I.). Theory. 15(1999). and B. Zero distribution via orthogonality. 71–93.S. Henri Poincar´. [20] P. Bello and G. Aptekarev and E. Amer. A transplantation theorem for Jacobi series. 1–31. 583–590. (N. [21] F. K¨ stner and V. [22] S. u Annales de l’Institut Fourier. 121(1983). Trans. Math. 14551499. E. 190–202.
. o ¨ [16] S. 29(1929). Aptekarev and W. Chihara. Constr. [17] J. Blumenthal. Amer. Proc. 92(1998). 129–166. Simon. 1999. Sieved orthogonal polynomials. New YorkLondonParis. J. Deift. 13. Kuijlaars.. Totik. Barrios and G. [19] E.Orthogonal Polynomials
119
[9] A. Illinois J. J. Approx. 129(2004). 730–736. Math.. Courant Lecture Notes. G¨ttingen. Mathematics and its Applications. 132(2004). Nikishin. From power pure point to continuous spectrum in disordered systems. Daems and A. Van Assche. 130(2004). The scattering problem for a discrete SturmLiouville operator. J. Theory. Theory. Uber SturmLiouvillesche Polynomsysteme. Soc. Askey.
Kuijlaars. Conf. Approx. Sb. Math. Bue o dapest.. Dur´n and F. Acad. 1971. Comm. 2004. 169190. Geronimus. [35] Ya. On polynomials orthogonal on the circle. Cambridge University Press. Orthogonal Polynomials. Sci. F. 5(1941). Its and A. Amer. L. J.. Appl. On the trigonometric moment problem. Kitaev. of Math. and on allied Carath´odory and Schur functions. [30] J. Geronimus. [25] A. 99–130. Orthogonal polynomials on several intervals via a polynomial mapping. Theory.. Totik. Math. Simon. Vol. 15(1944). Math. 308(1988). Zeros and critical points of Sobolev orthogonal polynomials. Asymptotics for Sobolev orthogonal polynomials for exponential weights. e Mat. Dunkl and Y. S. The isomonodromy approach to matrix models in 2D quantum gravity. 742–761. [36] L. Zeros of orthogonal polynomials on the real line. Oxford University Press. Translation. on trigonometric moment problem. [26] A. Orthogonal polynomials from Stieltjes to Simon. 395–430.. [29] W. [31] J. Gautschi. 121(2003)... Lubinsky and F. Van Assche. USSR Math. Soc. 91(1997). Gautschi and A. L. B. Marcellan. Geronimus. [28] W. Freud. L. Ann. S. Theory. Xu. A survey on orthogonal matrix polya u nomials satisfying second order diﬀerential equations. Comput. Simon. 357–364. [Russian] [34] Ya. Amer. New York. J. 1954.
. V. Golinskii and V. J. New York. Geronimus. L. 147(1992). A. Geronimo and W. Gr¨ nbaum. [32] Ya.V. Geronimo. 203–210. J. Bull. On the character of the solution of the moment problem in the case of the periodic in the limit associated fraction. Trans. [24] C. Totik
120
[23] S. Orthogonal Polynomials: Computation and Approximation. honoring B. Math. Fokas. Numerical Mathematics and Scientiﬁc Computation. to appear in Constructive Approx. Proc. Encyclopedia of Mathematics and its Applciations. Orthogonal polynomials on the unit circle and their applications. [27] G. A. Akad´miai Kiad´/Pergamon Press. 559–581. 2001. Soc. Orthogonal polynomials of several variables. 178(2005). J. AMS Conference Series (to appear). R. 117–137. Denisov and B. 47(1946). D. Approx. 104. [Russian] [33] Ya. Phys. S.
Matematicheskii Institut im. Proc. Nova Science Publisher (2003).. 535. Lubinsky. New York.. H. Internat. Ismail. Sum rules for Jacobi matrices and their applications to spectral theory. On Pad´ approximation of Markov type e meromorphic functions. Continued fractions and orthogonal polynomials on the unit circle. 108(2001). S. Classical and quantum orthogonal polynomials in one variable. Simon. Nume ber theory. 138–145. Rakhmanov. PijeiraCabrera and I. July 2429 ´ 2000). Memoirs Amer. [45] A. SpringerVerlag. Math. 157(1981). 2005. A note on the asymptotics of orthogonal polynomials on a complex arc: the case of a measure with a discrete part. 2003. F. Appl. P. 219230. AlvarezNodarse. Rossiiskaya Akademiya Nauk. [48] P. 158(2003). 253321 [44] T. (Russian) [38] A. Sobolev orthogonal polyo nomials in the complex plane. A. Steklova. 299–333. Orthogonal polynomials for exponential weights. [42] S. Annals of Math. McLaughlin. Dur´n. P. R. Comput. 6(1999). Izquierdo. [40] V. TR. 127(2001). mathematical analysis and their applications. 68 pp. Moscow. Approx. A. Math. Cambridge University Press. Gonchar and E. Suetin. Theory. 2004. Conf. Spain. 5. 31–48. . Killip and B. Gonchar and S. Levin and D. 4. E. OPSFA. Lubinsky. Steklov.. [41] S. Van Assche (Eds. Kriecherbauer and K. A. Kaliaguine. )
. Marcell´n a and W. . 2001. CMS Books in Mathematics/Ouvrages de Mathmatiques de la SMC. On the convergence of simultaneous Pad´ approximants for systems of functions of Markov type. Christoﬀel Functions and Orthogonal Polynomials for Exponential Weights on [−1. Inst. Research Notices. Approx. Encyclopedia in Mathematics. J. (Russian) [39] M. Laredo. 80(1995). J. Math. Copenhagen. Levin and D. A singular Riesz product in the Nevai class and inner functions with the Schur parameters in ∩p>2 ℓp .. Current Problems in Mathematics.Orthogonal Polynomials
121
[37] A. Khrushchev. Strong asymptotics of polynomials orthogonal with respect to a Freud weight. L´pezRodriguez and A. V. J. 1]. Soc. Theory. (electronic). o a Laredo Lecture Notes (Lecture notes of the 1st SIAG Summer School on Orthogonal Polynomials and Special Functions. H. S. Orthogonal matrix polynomials. L´pez. [47] G. [43] R. 249–255. J. Khrushchev. (1994) [46] A. Trudy Mat. A.
Totik. Nevai. [60] M. B.. Pitman Res. P. Saﬀ. Totik.. Approximation of Functions and Functionals with Application. Soc. CA/0508117. [52] A. Math. Strong Asymptotics for Extremal Polynomials Associated with Weights on (−∞.. Constructive Approx. J. Comput. 213(1979) [63] P. Mastroianni and V. Nevai and V. Strong Asymptotics for Extremal Errors Associated with Erd¨stype Weights. 202. 491–510. N. [51] F. 1305. [53] A. Random matrices. J. Math. [50] D. Math. Longmans... Comput. o Essex. Totik. On the distribution of zeros of polynomials orthogonal on the unit circle. 255266. 63(1990).V. M´t´. Denisov’s theorem on recurrent coeﬃcients. [59] A. Lubinsky and E. J. Szeg˝’s extremum problem on the unit a e o circle. Theory. second ed. 1988. ∞). S. Vol. Math. Rezola. J. 99–104. 240–245. Appl. Theory. Totik
122
[49] D. 127(2004). Math. Polynomial inequalities with doubling and A∞ weights. Orthogonal polynomials on a Sobolev spaces: old and new directions. Vol. [61] H. 109(1987).
. MartinezFinkelshtein. Annals of Math. MartinezFinkelshtein. MartinezFinkelshtein. Lubinsky. 48(1993). Constructive Approx.. Notes. Totik. Springer Verlag. Saﬀ. Appl. [57] A. Orthogonal polynomials and their zeros. 239–282. to appear. L. S. P. 99(1998). 30–38. 37–71. [55] G. [54] A. Amer. Approx. Analytic aspects of Sobolev orthogonal polynomials. 2005. M´t´. Appl. Berlin. Szeg˝ polynomials: a view from the Riemann– o Hilbert window. Orthogonal Polynomials.. Nevai and V. Nevai and V. Extensions of Szeg˝’s theory of orthogonal a e o polynomials. Boston. [64] P. Comput. 1989. Analytic properties of Sobolev orthogonal polynomials revisited. Milovanovic. 127(2001). 53(1989). 134(1991). 16(2000).. Nurn Mill. Inc. Academic Press. . Springer Lecture Notes. P. [56] G. Approx. 3(1987). Mhaskar and E. 51–72. [62] P. Acta Sci. 1991. M. Nevai and V. Totik. Mastroianni and G. World Scientiﬁc. Nevai and V. M´t´. Alfaro and M. Memoirs Amer. preprint arXiv:math. B. Strong and weak convergence of orthogonal a e polynomials. 433–453. J.(Szeged). Marcell´n. J. Math. Mehta. Totik.. [58] A. 113–131.
46(1983). Grundlehren der mathematischen Wissenschaften. [72] F. Peherstorfer. 123–161. Pad´ polynomial asymptotics from a singular integral equation. An estimate for orthogonal polynomials. Asymptotics for diagonal HermitePad´ approximants. Peherstorfer and R. Approx. Totik. [Russian] [66] J. B.. What parts of a measure’s support attract zeros of the corresponding orthogonal polynomials? Proc. SpringerVerlag. Ape prox. A. (Russian) Dokl. The classical moment problem as a selfadjoint ﬁnite diﬀerence operator. Berlin. On BernsteinSzeg˝ orthogonal polynomials on several ino tervals. Nuttall. Advances in Math. 1997. On BernsteinSzeg˝ orthogonal polynomials on several ino tervals. Peherstorfer. 104–117. On V. Steinbauer. [71] F. 21(1990). 254(1980). Math. 64(1991). Special functions (Torino. Springer Verlag. Appl.. Vol. Steklov’s problem in the theory of orthogonal polynomials. Soc.(Szeged). J. 6(1990). Theory. A. 185–190. Saﬀ and V. Notes. 42(1984). Math. Amer. Theory. [76] E. Math. Zeros of polynomials orthogonal on several intervals. Rakhmanov. 7(2003). Akad. USSR Sbornik. Pastur and A. 1992. On the asymptotics of the ratio of orthogonal polynomials. Peherstorfer. J. [69] F. Peherstorfer. 105–117. 187–205. 299–386.. Simon. 1993). J. [70] F. [68] L. Res. Logarithmic Potentials with External Fields. Saﬀ and V. 353–370. Figotin. Anal.Orthogonal Polynomials
123
[65] E. 137(1998). 157–166. Comput.. Spectra of Random and AlmostPeriodic Operators. Grundlehren der mathematischen Wissenschaften. Mat. [75] E. [67] J. 316. Acta Sci. Math. Int. Nikishin. Math. II... 48(1985). 114(1992). Orthogonal polynomials with periodic recurrence coeﬃcients. Nauk SSSR. Nuttall. 48(1993). [78] B. 82–203. [73] F. Math. Russian original. Rakhmanov.
.. Numer. 2(1995). A. Totik. 802–806. 395–399. Ann. 297. [74] E. B. Berlin. 461–482. 361–385. e Constructive Approx. Orthogonal and extremal polynomials on several intervals. Math. 118(1982). SIAM J. Sb. M. [77] E. On polynomials orthogonal on several intervals.
Sb. 242(2003). Orthogonal Polynomials. [90] J. I. 225–240. o . On the uniform convergence of diagonal Pad´ approximants for e hyperelliptic functions. Nauk. New York. 393–423. 134(2005).. Sum rules and the Szeg˝ condition for orthogonal s o polynomials on the real line. Math. II. V. Math. Interpolation of Functions. Orthogonal Polynomials on the Unit Circle. Theory. 1990. Zlatoˇ. AMS Colloquium Series. (Russian). Mat. General Orthogonal Polynomials. 1339–1373. 139–154. Orthogonal Polynomials on the Unit Circle.. Simon. Stahl. Orthogonal polynomials with ratio asymptotics. Totik. Comput.. Amer.. IMACS Ann.V. [83] B. [86] H. 114–129. Theory. World Scientiﬁc. . 57(2002). Coll. 1992. [84] H. Soc. 249–270. Providence. V´rtesi. Orthogonal polynomials with complexvalued weight function. J. 1: Classical Theory. 191(2000). Simon. [82] B. 57(2002). 2005. 1(1985). Appl. Szeg˝. 1990). Comm. [80] B. translation in Sb. 1975. RI. 114(1992). Totik
124
[79] B. Approx. Ratio asymptotics and weak asymptotic measures for orthogonal polynomials on the real line. [91] G. [87] H. Phys. Cambridge University Press. Amer. Simon and A. [92] V. 2(1986). Constr. Divergence of diagonal Pad´ approximants and the asymptotic e behavior of orthogonal polynomials associated with nonpositive measures. Providence. Approx. 198–217. American Mathematical Society. Encyclopedia of Mathematics and its Applications. (Russian). XXIII. Constr. Orthogonal polynomials with respect to complexvalued measures. Uspekhi Mat. J. 43–141. AMS Colloquium Series. Soc. [88] S. [89] S. 81–114. 43. 2005. Totik. Math. Orthogonal polynomials and their applications (Erice.. e Singapore. Stahl. Math. Approx. RI. Simon and A. Higherorder Szeg˝ theorems with two singular s o points. Suetin.
. Stahl. 9(1991). Approx.. Proc. [81] B. translation in Russian Math. Stahl and V. Publ. 491–495. [85] H. Providence. Surveys. P. 2: Spectral Theory. Pad´ approximants and the eﬀective analytic continuation of a e power series. Math. Suetin. 126(2004). Szabados and P.. American Mathematical Society. Simon. 241–251. 45–142. Zlatoˇ. P. 191(2000). V.
B. [97] H. [95] W. Bolyai Institute. Adv. Vol. J. Van Assche. Special Functions 2000: Current Perspectives and Future Directions. Widom. Spriger Lecture Notes. Van Assche. Math. 33620 USA
. J. Kuijlaars. 127–232. 1265. J. University of Szeged Szeged Aradi v. 2001. Dordrecht. tere 1. in: J. Asymptotics for Christoﬀel functions for general measures on the real line. 6720 Hungary. Extremal polynomials associated with a system of curves in the complex plane. FL. Van Assche. Asymptotics for Orthogonal Polynomials... (eds). Riemann–Hilbert problems for multiple orthogonal polynomials. 3(1969). 23–59. Kluwer. 1987. 283–303. [94] W. Geronimo. S. D’Analyse Math. Berlin. Bustoz et al. [96] W. Totik. preprint. and Department of Mathematics University of South Florida Tampa.Orthogonal Polynomials
125
[93] V. Springer Verlag. 81(2000). A. HermitePad´ approximation and multiple orthogonal e polynomials.