Orthogonality and probability: beyond nearest neighbor

transitions
Yevgeniy Kovchegov

Abstract
In this article, we will explore why Karlin-McGregor method of using orthogonal
polynomials in the study of Markov processes was so successful for one dimensional
nearest neighbor processes, but failed beyond nearest neighbor transitions. We will
proceed by suggesting and testing possible xtures.
1 Introduction
This paper was inuenced by the approaches described in Deift [2] and questions con-
sidered in Gr unbaum [7].
The Karlin-McGreogor diagonalization can be used to answer recurrence/transience
questions, as well as those of probability harmonic functions, occupation times and
hitting times, and a large number of other quantities obtained by solving various re-
currence relations, in the study of Markov chains, see [9], [10], [11], [12], [8], [17], [16],
[4], [14]. However with some exceptions (see [13]) those were nearest neighbor Markov
chains on half-line. Gr unbaum [7] mentions two main drawbacks to the method as (a)
typically one cannot get either the polynomials or the measure explicitly, and (b)
the method is restricted to nearest neighbour transition probability chains that give
rise to tridiagonal matrices and thus to orthogonal polynomials. In this paper we
attempt to give possible answers to the second question of Gr unbaum [7] for general
reversible Markov chains. In addition, we will consider possible applications of the
newer methods in orthogonal polynomials such as using Riemann-Hilbert approach,
see [2], [3] and [15], and their probabilistic interpretations.
In Section 2, we will give an overview of the Karlin-McGregor method from a naive
college linear algebra perspective. In 2.3, we will give a Markov chain interpretation to
the result of Fokas , Its and Kitaev, connecting orthogonal polynomials and Riemann-
Hilbert problems. Section 3 deals with one dimensional random walks with jumps
of size m, the 2m + 1 diagonal operators. There we consider diagonalizing with
orthogonal functions. In 3.2, as an example we consider a pentadiagonal operator and
use Plemelj formula, and a two sided interval to obtain the respective diagonalization.
In Section 4, we use the constructive approach of Deift [2] to produce the Karlin-
McGregor diagonalization for all irreducible reversible Markov chains. After that, we
revisit the example from Section 3.

Department of Mathematics, Oregon State University, Corvallis, OR 97331-4605, USA

kovchegy@math.oregonstate.edu
1
Y.V.Kovchegov Orthogonality and Probability
2 Eigenvectors of probability operators
Suppose P is a tridiagonal operator of a one-dimensional Markov chain on {0, 1, . . . }
with forward probabilities p
k
and backward probabilities q
k
. Suppose is an eigenvalue
of P and q
T
() =
_
_
_
_
_
Q
0
Q
1
Q
2
.
.
.
_
_
_
_
_
is the corresponding right eigenvector such that Q
0
= 1. So
q
T
= Pq
T
generates the recurrence relation for Q
j
. Then each Q
j
() is a polynomial
of j-th degree. The Karlin-McGregor method derives the existence of a probability
distribution such that polynomials Q
j
() are orthogonal with respect to . In other
words, if is stationary with
0
= 1 and < , >

is the inner product in L

2
(d), then
< Q
i
, Q
j
>

=

i,j

j
Thus {

j
Q
j
()}
j=0,1,...
are orthonormal polynomials, where
0
= 1 and
j
=
p
0
...p
j1
q
1
...q
j
(j = 1, 2, . . . ). Also observe from the recurrence relation that the leading coecient of
Q
j
is
1
p
0
...p
j1
.
Now,
t
q
T
= P
t
q
T
implies
t
q
i
= (P
t
q
T
)
i
for each i, and
<
t
Q
i
, Q
j
>

=< (P
t
q
T
)
i
, Q
j
>

=
p
t
(i, j)

j
Therefore
p
t
(i, j) =
j
<
t
Q
i
, Q
j
>

Since the spectrum of P lies entirely inside [1, 1] interval, then so is the support
of . Hence, for |z| > 1, the generating function
G
i,j
(z) =
+

t=0
z
t
p
t
(i, j) = z
j
<
Q
i
z
, Q
j
>

= z
j
_
Q
i
()Q
j
()
z
d()
2.1 Converting to a Jacobi operator
Let b
k
=
_

k

k+1
p
k
, then b
k
=
_

k+1

k
q
k+1
due to reversibility condition. Thus the
recurrence relation for q,

k
Q
k
= q
k

k
Q
k1
+ (1 q
k
p
k
)

k
Q
k
+ p
k

k
Q
k+1
,
can be rewritten as

k
Q
k
= b
k1

k
Q
k1
+ a
k

k
Q
k
+ b
k

k
Q
k+1
,
where a
k
= 1 q
k
p
k
. Therefore q = (

0
Q
0
,

1
Q
1
, . . . ) solves

P q = q, where

P =
_
_
_
_
_
_
_
a
0
b
0
0 . . .
b
0
a
1
b
1
.
.
.
0 b
1
a
2
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
_
_
_
_
_
_
_
2
Y.V.Kovchegov Orthogonality and Probability
is a Jacobi (symmetric triangular with b
k
> 0) operator. Observe that

P is self-adjoint.
The above approach extends to all reversible Markov chains. Thus every reversible
Markov operator is equivalent to a self-adjoint operator, and therefore has an all real
spectrum.
2.2 Karlin-McGregor: a simple picture
It is a basic fact from linear algebra that if
1
, . . . ,
n
are distinct real eigenvalues of
an nn matrix A, and if u
1
, . . . , u
n
and v
1
, . . . , v
n
are the corresponding left and right
eigenvectors. Then A diagonalizes as follows
A
t
=

t
v
T
j
u
j
u
j
v
T
j
=
_
(A)

t
v
T
()u()d() ,
where u(
j
) = u
j
, v(
j
) = v
j
, spectrum (A) = {
1
, . . . ,
n
}, and
() =

j
1
u()v
T
()

j
() =
n
u()v
T
()
U
(A)
()
Here U
(A)
() is the uniform distribution over the spectrum (A).
It is important to observe that the above integral representation is only possible
if u() and v() are well dened - each eigenvalue has multiplicity one, i.e. all distinct
real eigenvalues. As we will see later, this will become crucial for Karlin-McGregor
diagonalization of reversible Markov chains. The operator for a reversible Markov
chain is bounded and is equivalent to a self-adjoint operator, and as such has a real
bounded spectrum. However the eigenvalue multiplicity will determine whether the
operators diagonalization can be expressed in a form of a spectral integral.
Since the spectrums (P) = (P

), we will extend the above diagonalization

identity to the operator P in the separable Hilbert space l
2
(R). First, observe that
u() = (
0
Q
0
,
1
Q
1
, . . . ) satises
uP = P
due to reversibility. Hence, extending from a nite case to an innite dimensional space
l
2
(R), obtain
P
t
=
_

t
q
T
()u()d() =
_

t
_
_
_

0
Q
0
Q
0

1
Q
0
Q
1

0
Q
1
Q
0

1
Q
1
Q
1

.
.
.
.
.
.
.
.
.
_
_
_d() ,
where
() = lim
n+

n
()
The above is the weak limit of

n
() =
n
u()q
T
()
U
(An)
() ,
3
Y.V.Kovchegov Orthogonality and Probability
where A
n
is the restriction of P to the rst n coordinates, < e
0
, . . . , e
n1
>
A
n
=
_
_
_
_
_
_
_
_
_
1 p
0
p
0
0 0
q
1
1 q
1
p
1
p
1
.
.
.
.
.
.
0 q
2
1 q
2
.
.
. 0
.
.
.
.
.
.
.
.
.
.
.
. p
n2
0 0 q
n1
1 q
n1
p
n1
_
_
_
_
_
_
_
_
_
Observe that if Q
n
() = 0 then (Q
0
(), . . . , Q
n1
())
T
is the corresponding right
eigenvector of A
n
. Thus the spectrum of (A
n
) is the roots of
Q
n
() = 0
So

n
() =
n
u()q
T
()
U
Qn=0
() =
n

n1
k=0

k
Q
2
k
()
U
Qn=0
() .
The orthogonality follows if we plug in t = 0. Since
0
Q
0
Q
0
= 1, should integrate
to one.
Example. Simple random walk and Chebyshev polynomials. The Chebyshev polyno-
mials of the rst kind are the ones characterizing a one dimensional simple random
walk on half line, i.e. the ones with generator
P
ch
=
_
_
_
_
_
_
_
_
_
0 1 0 0
1
2
0
1
2
0
0
1
2
0
1
2
.
.
.
0 0
1
2
0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
_
_
_
_
_
_
_
_
_
So, T
0
() = 1, T
1
() = and T
k+1
() = 2T
k
() T
k1
() for k = 2, 3, . . . . The
Chebyshev polynomials satisfy the following trigonometric identity:
T
k
() = cos(k cos
1
())
Now,

n
() =
n

n1
k=0

k
T
2
k
()
U
{cos(ncos
1
())=0}
() ,
where (0) = 1 and (1) = (2) = = 2. Here
U
{cos(ncos
1
())=0}
() = U
{cos
1
())=

2n
+
k
n
, k=0,1,...,n1}
()
Thus if X
n
U
{cos(ncos
1
())=0}
, then Y
n
= cos
1
(X
n
) U
{

2n
+
k
n
, k=0,1,...,n1}
and Y
n
converges weakly to Y U
[0,]
. Hence X
n
converges weakly to
X = cos(Y )
1

1
2

[1,1]
()d ,
4
Y.V.Kovchegov Orthogonality and Probability
i.e.
U
{cos(ncos
1
())=0}
()
1

1
2

[1,1]
()d
Also observe that if x = cos(), then
n1

k=0

k
T
2
k
() = 1 + 2
n1

k=0
cos
2
(kx) = n
1
2
+
sin((2n 1)x)
2 sin(x)
Thus
d
n
() d() =
1

1
2

[1,1]
()d
2.3 Riemann-Hilbert problem and a generating function
of p
t
(i, j)
Let us write

j
Q
j
() = k
j
P
j
(), where k
j
=
1

p
0
...p
j1

q
1
...q
j
is the leading coecient
of

j
Q
j
(), and P
j
() is therefore a monic polynomial.
In preparation for the next step, let w() be the probability density function as-
sociated with the spectral measure : d() = w()d on the compact support,
supp() [1, 1] = . Also let
C(f)(z) =
1
2i
_

f()
z
d()
denote the Cauchy transform w.r.t. measure .
First let us quote the following theorem.
Theorem. [Fokas, Its and Kitaev, 1990] Let
v(z) =
_
1 w(z)
0 1
_
be the jump matrix. Then, for any n {0, 1, 2, . . . },
m
(n)
(z) =
_
P
n
(z) C(P
n
w)(z)
2ik
2
n1
P
n1
(z) 2ik
2
n1
C(P
n1
w)(z)
_
, for all z C \ ,
is the unique solution to the Riemann-Hilbert problem with the above jump matrix v(x)
and that satises the following condition
m
(n)
(z)
_
z
n
0
0 z
n
_
I as z . (1)
The Riemann-Hilbert problem, for an oriented smooth curve , is the problem of
nding m(z), analytic in C \ such that
m
+
(z) = m

(z)v(z), for all z ,

where m
+
and m

denote respectively the limit from the left and the limit from the
right of function m as the increment approaches a point on .
5
Y.V.Kovchegov Orthogonality and Probability
Suppose we are given the weight function w() for the Karlin-McGregor orthogonal
polynomials q. If m
(n)
(z) is the solution of the Riemann-Hilbert problem as in the
above theorem, then for |z| > 1,
m
(n)
(z) =
_
1
kn

n
Q
n
(z)
1
2ikn

nz
n+1
G
0,n
2i
k
n1

n1
Q
n1
(z)
k
n1

n1
z
n
G
0,n1
(z)
_
=
_
q
1
. . . q
n
Q
n
(z)
q
1
...qn
2iz
n+1
G
0,n
2i
p
0
...p
n2
Q
n1
(z)
1
p
0
...p
n2
z
n
G
0,n1
(z)
_
3 Beyond nearest neighbor transitions
Observe that the Chebyshev polynomials were used to diagonalize a simple one di-
mensional random walk reecting at the origin. Let us consider a random walk where
jumps of sizes one and two are equiprobable
P =
_
_
_
_
_
_
_
_
_
_
_
_
_
_
0
1
2
1
2
0 0 0 . . .
1
4
1
4
1
4
1
4
0 0 . . .
1
4
1
4
0
1
4
1
4
0 . . .
0
1
4
1
4
0
1
4
1
4
.
.
.
0 0
1
4
1
4
0
1
4
.
.
.
0 0 0
1
4
1
4
0
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
_
_
_
_
_
_
_
_
_
_
_
_
_
_
The above random walk with the reector at the origin is reversible with
0
= 1 and

1
=
2
= = 2. The Karlin-McGregor representation with orthogonal polynomials
will not automatically extend to this case. However this does not rule out obtaining a
Karlin-McGregor diagonalization with orthogonal functions.
In the case of the above pentadiagonal Chebyshev operator, some eigenvalues will
be of geometric multiplicity two as
P = P
2
ch
+
1
2
P
ch

1
2
I ,
where P
ch
is the original tridiagonal Chebyshev operator.
3.1 2m + 1 diagonal operators
Consider a 2m + 1 diagonal reversible probability operator P. Suppose it is Karlin-
McGregor diagonalizable. Then for a given (P), let q
T
() =
_
_
_
_
_
Q
0
Q
1
Q
2
.
.
.
_
_
_
_
_
once again
denote the corresponding right eigenvector such that Q
0
= 1. Since the operator is
more than tridiagonal, we encounter the problem of nding the next m1 functions,
Q
1
() =
1
(), Q
2
() =
2
(), . . . , Q
m1
() =
m1
().
6
Y.V.Kovchegov Orthogonality and Probability
Observe that q = q
0
+ q
1

1
+ + q
m1

m1
, where each q
T
j
() =
_
_
_
_
_
Q
0,j
Q
1,j
Q
2,j
.
.
.
_
_
_
_
_
solves Pq
T
j
= q
T
j
recurrence relation with the initial conditions
Q
0,j
() = 0, . . . , Q
j1,j
() = 0, Q
j,j
() = 1, Q
j+1,j
() = 0, . . . , Q
m1,j
() = 0
In other words, q
T
() = Q()
T
, where Q() =
_

_
| | |
q
T
0
q
T
1
q
T
m1
| | |
_

_
and

T
=
_
_
_
_
_
1

1
()
.
.
.

m1
()
_
_
_
_
_
is such that q() l
2
(R) for each (P).
Let A
n
denote the restriction of P to the rst n coordinates, < e
0
, . . . , e
n1
>.
Observe that if Q
n
() = = Q
n+m1
() = 0 then (Q
0
(), . . . , Q
n1
())
T
is the
corresponding right eigenvector of A
n
. Thus the spectrum of (A
n
) consists of the
roots of
det
_
_
_
_
_
Q
n,0
() Q
n,1
() Q
n,m1
()
Q
n+1,0
() Q
n+1,1
() Q
n+1,m1
()
.
.
.
.
.
.
.
.
.
Q
n+m1,0
() Q
n+m1,1
() Q
n+m1,m1
()
_
_
_
_
_
= 0
3.2 Chebyshev operators
Let us now return to the example generalizing the simple random walk reecting at
the origin. There one step and two step jumps were equally likely. The characteristic
equation z
4
+ z
3
4z
3
+ z
2
+ z = 0 for the recurrence relation
c
n+2
+ c
n+1
4c
n
+ c
n1
+ c
n2
= 0
can be easily solved by observing that if z is a solution then so are z and
1
z
. The
solution in radicals is expressed as z
1,2
= r
1
i
_
1 r
2
1
and z
3,4
= d
2
i
_
1 r
2
2
,
where r
1
=
1+

9+16
4
and r
2
=
1

9+16
4
.
Observe that r
1
and r
2
are the two roots of s(x) = , where s(x) = x
2
+
1
2
x
1
2
is
the polynomial for which
P = s(P
ch
)
In general, the following is true for all operators P that represent symmetric random
walks reecting at the origin, and that allow jumps of up to m ights: there is a
polynomial s(x) such that P = s(P
ch
) and the roots z
j
of the characteristic relation
in c = Pc will lie on a unit circle with their real parts Re(z
j
) solving s(x) = . The
reason for the latter is the symmetry of the corresponding characteristic equation of
7
Y.V.Kovchegov Orthogonality and Probability
order 2m, implying
1
z
j
= z
j
, and therefore the characteristic equation for c = Pc can
be rewritten as
s
_
1
2
_
z +
1
z
__
= ,
where
1
2
_
z +
1
z

is the Zhukovskiy function.

In our case, s(x) =
_
x +
1
4
_
2

9
16
, and for
_

9
16
, 0

, there will be two candidates

for
1
(),

+
() = r
1
=
1 +

9 + 16
4
and

() = r
2
=
1

9 + 16
4
Taking 0 arg z < 2 branch of the logarithm log z, and applying Plemelj formula,
one would obtain

1
(z) =
1
4
+ z
1
2
exp
_
1
2
_
0

9
16
ds
s z
_
,
where
+
() = lim
z, Im(z)>0

1
(z) and

() = lim
z, Im(z)<0

1
(z).
Now, as we dened
1
(z), we can propose the limits of integration to be a contour
in C consisting of
_

9
16
, 0
_
+
= lim
0
_
z = x + i : x
_

9
16
, 0
__
, and
_

9
16
, 0
_

=
lim
0
_
z = x i : x
_

9
16
, 0
__
, and the [0, 1] segment. Then
P
t
=
_
[
9
16
,0)

[
9
16
,0)
+
[0.1]

t
q
T
()u()d(),
where u() is dened as before, and
d() =
1
2
_
+
9
16
_
_
_
_

[
9
16
,0)

()
_
1
__
+
9
16
+
1
4
_
2
+

[
9
16
,0)
+
() +
[0,1]
()
_
1
__
+
9
16

1
4
_
2
_
_
_
_
d
Let us summarize this section as follows. If the structure of the spectrum does
not allow Karlin-McGregor diagonalization with orthogonal functions over [1, 1], say
when there are two values of
T
() for some , then one may use Plemelj formula to
obtain an integral diagonalization of P over the corresponding two sided interval.
4 Spectral Theorem and why orthogonal poly-
nomials work
The constructive proofs in the second chapter of Deift [2] suggest the reason why
Karlin-McGregor theory of diagonalizing with orthogonal polynomials works for all
time reversible Markov chains. The approach goes back to the works of M.Krein and
N.I.Akhiezer, see [1] and [6]. Using the same logical steps as in [2], we can construct
a map M which assigns a probability measure d to a reversible transition operator
P on a countable state space {0, 1, 2, . . . }. W.l.o.g. we can assume P is symmetric as
8
Y.V.Kovchegov Orthogonality and Probability
one can instead consider
_
_
_
_

0
0
0

1
.
.
.
.
.
.
.
.
.
.
.
.
_
_
_
_
P
_
_
_
_
1

0
0
0
1

1
.
.
.
.
.
.
.
.
.
.
.
.
_
_
_
_
which is symmetric, and its spectrum coinciding with spectrum (P) [1, 1].
Now, for z C \ R let G(z) = (e
0
, (P zI)
1
e
0
). Then
ImG(z) =
1
2i
_
(e
0
, (P zI)
1
e
0
) (e
0
, (P zI)
1
e
0
)

= (Im(z))|(P zI)
1
e
0
|
2
and therefore G(z) is a Herglotz function, i.e. G(z) is an analytic map from
{Im(z) > 0} into {Im(z) > 0}, and as all such functions, it can be represented
as
G(z) = az + b +
_
+

_
1
s z

s
s
2
+ 1
_
d(s), Im(z) > 0
In the above representation a 0 and b are real constants and d is a Borel measure
such that
_
+

1
s
2
+ 1
d(s) <
Deift [2] uses G(z) = (e
0
, (P zI)
1
e
0
) =
1
z
+O(z
2
) to show a = 0 in our case, and
b =
_
+

s
s
2
+ 1
d(s)
as well as the uniqueness of d. Hence
G(z) =
_
d(s)
s z
, Im(z) > 0
The point of all these is to construct the spectral map
M: {reversible Markov operators P} {probability measures on [1, 1] with compact supp()}
The asymptotic evaluation of both sides in
(e
0
, (P zI)
1
e
0
) =
_
d(s)
s z
, Im(z) > 0
implies
(e
0
, P
k
e
0
) =
_
s
k
d(s)
Until now we were reapplying the logical steps in Deift [2] for the case of reversible
Markov chains. However, in the original, the second chapter of Deift [2] gives a con-
structive proof of the following spectral theorem, that summarizes as
U : {bounded Jacobi operators on l
2
} {probability measures on R with compact supp()},
where U is one-to-one onto.
9
Y.V.Kovchegov Orthogonality and Probability
Theorem 1. [Spectral Theorem] For every bounded Jacobi operator A there exists
a unique probability measure with compact support such that
G(z) =
_
e
0
, (AzI)
1
e
0
_
=
_
+

d(x)
x z
The spectral map U : A d is one-to-one onto, and for every f L
2
(d),
(UAU
1
f)(s) = sf(s)
in the following sense
(e
0
, Af(A)e
0
) =
_
sf(s)d(s)
So suppose P is a reversible Markov chain, then
M: P d and U
1
: d P

,
where P

is a unique Jacobi operator such that

(e
0
, P
k
e
0
) =
_
s
k
d(s) = (e
0
, P
k

e
0
)
Now, if Q
j
() are the orthogonal polynomials w.r.t. d associated with P

, then
Q
j
(P

)e
0
= e
j
and

i,j
= (e
i
, e
j
) = (Q
i
(P

)e
0
, Q
j
(P

)e
0
) = (Q
i
(P)e
0
, Q
j
(P)e
0
)
Thus, if P is irreducible, then f
j
= Q
j
(P)e
0
is an orthonormal basis for Karlin-
McGregor diagonalization. If we let F =
_
_
| |
f
0
f
1

| |
_
_
, then
P
t
=
_
_
(P
t
e
i
, e
j
)
_
_
= F
_
_
_
1
1
s
t
Q
i
(s)Q
j
(s)d(s)
_
_
F
T
,
where F
T
= F
1
. Also Deift [2] provides a way for constructing
U
1
M: P P

Since P

is a Jacobi operator, it can be represented as

P

=
_
_
_
_
_
_
_
a
0
b
0
0
b
0
a
1
b
1
.
.
.
0 b
1
a
2
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
_
_
_
_
_
_
_
b
j
> 0
Now,
(e
0
, Pe
0
) = (e
0
, P

e
0
) = a
0
, (e
0
, P
2
e
0
) = (e
0
, P
2

e
0
) = a
2
0
+ b
2
0
10
Y.V.Kovchegov Orthogonality and Probability
(e
0
, P
3
e
0
) = (e
0
, P
3

e
0
) = (a
2
0
+ b
2
0
)a
0
+ (a
0
+ a
1
)b
2
0
and (e
0
, P
4
e
0
) = (e
0
, P
4

e
0
) = (a
2
0
+ b
2
0
)
2
+ (a
0
+ a
1
)
2
b
2
0
+ b
2
0
b
2
1
thus providing us with the coecients of the Jacobi operator, a
0
, b
0
, a
1
, . . . , and
therefore the orthogonal polynomials Q
j
.
Example. Pentadiagonal Chebyshev operator. For the pentadiagonal P that repre-
sents the symmetric random walk with equiprobable jumps of sizes one and two,
(e
0
, Pe
0
) = 0, (e
0
, P
2
e
0
) =
1
4
, (e
0
, P
3
e
0
) =
3
32
, (e
0
, P
4
e
0
) =
9
64
, . . .
Thus
a
0
= 0, b
0
=
1
2
, a
1
=
3
8
, b
1
=

11
8
, etc.
So
P

=
_
_
_
_
_
_
_
0
1
2
0
1
2
3
8

11
8
.
.
.
0

11
8
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
_
_
_
_
_
_
_
and
Q
0
() = 1, Q
1
() = 2, Q
2
() =
32

11

11

4

11
, . . .
Then applying classical Fourier analysis, one would obtain
_
e
0
, (P zI)
1
e
0
_
=
1
2
_
2
0
d
1
2
[cos() + cos(2)] z
=
_
1

9
16
d(s)
s z
,
where
d(s) =
1
2
_
s +
9
16
_
_
_
_

[
9
16
,1]
(s)
_
1
__
s +
9
16

1
4
_
2
+

[
9
16
,0)
(s)
_
1
__
s +
9
16
+
1
4
_
2
_
_
_
_
ds
To obtain the above expression for d we used the fact that
_
e
0
, (P zI)
1
e
0
_
would
be the same if there were no reector at zero.
4.1 Applications of Karlin-McGregor diagonalization
Let us list some of the possible applications of the diagonalization. First, the diagonal-
ization can be used to extract a sharp rate of convergence to a stationary probability
distribution, if there is one, see Diaconis et al [4]. If we consider a nearest-neighbor
Markov chain on a half-line with nite , i.e.
=

k
< ,
11
Y.V.Kovchegov Orthogonality and Probability
then the spectral measure will contain a point mass

1
()
P

k=0

k
Q
2
k
()
=

1
()

at 1 (and
naturally no point mass at 1).
In order to measure the rate of convergence to a stationary distribution, the following
distance is used.
Denition 1. If and are two probability distributions over a sample space , then
the total variation distance is

TV
=
1
2

x
|(x) (x)| = sup
A
|(A) (A)|
Observe that the total variation distance measures the coincidence between the distri-
butions on a scale from zero to one.
In our case, =
1

is the stationary probability distribution. Now, suppose the process

commences at site
0
= i, then the total variation distance between the distribution

t
=
0
P
t
and is given by

t

TV
=
1
2

j
_
(1,1]

t
Q
i
()Q
j
()d()

=
1
2

_
(1,1)

t
Q
i
()Q
j
()d()

The rates of convergence are quantied via mixing times.

Denition 2. Suppose P is an irreducible and aperiodic Markov chain with stationary
probability distribution . Given an > 0, the mixing time t
mix
() is dened as
t
mix
() = min{t :
t

TV
}
Thus, in the case of a nearest-neighbor process commencing at
0
= i, the correspond-
ing mixing time has the following simple expression
t
mix
() = min
_
_
_
t :

_
(1,1)

t
Q
i
()Q
j
()d()

2
_
_
_
Observe that the above expression is simplied when
0
= 0. In general, if P is an
irreducible aperiodic reversible Markov chain with a stationary distribution , that
commences at
0
= 0, then its mixing time is given by
t
mix
() = min
_

_
t :
_
_
_
_
_
_
e
i
F
_
_
_
(1,1)
s
t
Q
j
(s)Q
k
(s)d(s)
_
_
F
T
_
_
_
_
_
_

1
2
_

_
,
where Q
j
, F and are as described earlier in the section.
12
Y.V.Kovchegov Orthogonality and Probability
Now, if
0
= 0, then e
0
F = ((e
0
, f
0
), (e
0
, f
1
), . . . ) = e
0
by construction, and there-
fore
t
mix
() = min
_
t :
_
_
_
_
_
_
_
(1,1)
s
t
Q
0
(s)d(s),
_
(1,1)
s
t
Q
1
(s)d(s), . . .
_
F
T
_
_
_
_
_

1
2
_
= min
_
_
_
t :
_
_
_
_
_
_
e
0
_
_

j
_
_
(1,1)
s
t
Q
j
(s)d(s)
_
Q
j
(P)
_
_
_
_
_
_
_
_

1
2
_
_
_
in this case. For example, if we estimate the rate with which

_
(1,1)
s
t
Q
j
(s)d(s)

e
0
Q
j
(P)

1
converges to zero, if it does, we would obtain an upper bound on the mixing time.
We conclude by listing some other important applications of the method.
The generator
G(z) =
_
_
G
i,j
(z)
_
_
= F
_
_
z
_
1
1
Q
i
()Q
j
()
z
d()
_
_
F
T
One can use the Fokas, Its and Kitaev results, and benet from the connection
between orthogonal polynomials and Riemann-Hilbert problems.
One can interpret random walks in random environment as a random spectral
measure.
References
[1] N.I.Aheizer and M.Krein, Some Questions in the Theory of Moments. Translations
of Math Monographs, Vol.2, Amer. Math. Soc., Providance, RI, (1962) (transla-
tion of the 1938 Russian edition)
[2] P.Deift, Orthogonal Polynomials and Random Matrices: A Riemann-Hilbert Ap-
proach. Amer. Math. Soc., Providance, RI, (2000)
[3] P.Deift, Riemann-Hilbert Methods in the Theory of Orthogonal Polynomials Spec-
tral Theory and Mathematical Physics, Vol. 76, Amer. Math. Soc., Providance,
RI, (2006) pp.715-740
[4] P.Diaconis, K.Khare and L.Salo-Coste Gibbs Sampling, Exponential Families and
Orthogonal Polynomials Statistical Science, Vol. 23, No.2, (2008), pp.151-178
[5] H.Dym and H.P.McKean, Gaussian processes, function theory, and the inverse
spectral problem Probability and Mathematical Statistics, 31, Academic, New
York - London (1976)
[6] M.L.Gorbachuk and V.I.Gorbachuk, M.G.Kreins Lectures on Entire Operators
Birkhauser Verlag (1997)
[7] F.A. Gr unbaum, Random walks and orthogonal polynomials: some challenges
Probability, Geometry and Integrable Systems - MSRI Publications, Vol. 55,
(2007), pp.241-260.
13
Y.V.Kovchegov Orthogonality and Probability
[8] S.Karlin, Total Positivity Stanford University Press, Stanford, CA (1968)
[9] S.Karlin and J.L.McGregor, The dierential equations of birth and death processes,
and the Stieltjes moment problem Transactions of AMS, 85, (1957), pp.489-546.
[10] S.Karlin and J.L.McGregor, The classication of birth and death processes Trans-
actions of AMS, 86, (1957), pp.366-400.
[11] S.Karlin and J.L.McGregor, Random Walks Illinois Journal of Math., 3, No. 1,
(1959), pp.417-431.
[12] S.Karlin and J.L.McGregor, Occupation time laws for birth and death processes
Proc. 4th Berkeley Symp. Math. Statist. Prob., 2, (1962), pp.249-272.
[13] S.Karlin and J.L.McGregor, Linear Growth Models with Many Types and Multi-
dimensional Hahn Polynomials In: R.A. Askey, Editor, Theory and Applications
of Special Functions, Academic Press, New York (1975), pp. 261288.
[14] Y.Kovchegov, N.Meredith and E.Nir Occupation times via Bessel functions
preprint
[15] A.B.J.Kuijlaarsr, Riemann-Hilbert Analysis for Orthogonal Polynomials Orthog-
onal Polynomials and Special Functions (Springer-Verlag), Vol. 1817, (2003)
[16] W.Schoutens, Stochastic Processes and Orthogonal Polynomials. Lecture notes in
statistics (Springer-Verlag), Vol. 146, (2000)
[17] G.Szego, Orthogonal Polynomials. Fourth edition. AMS Colloquium Publications,
Vol. 23, (1975)
14