Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

Statistics and Probability Letters 81 (2011) 19401944

Contents lists available at SciVerse ScienceDirect

Statistics and Probability Letters


journal homepage: www.elsevier.com/locate/stapro

Deviations of discrete distributions and a question of Mri


Kenneth S. Berenhaut , John V. Baxley, Robert G. Lyday
Wake Forest University, Department of Mathematics, Winston-Salem, NC 27109, United States

article info abstract


Article history: In this note, we consider a question of Mri regarding estimating the deviation of the kth
Received 24 October 2010 terms of two discrete probability distributions in terms of the supremum distance between
Received in revised form 27 June 2011 their generating functions over the interval [0, 1]. An optimal bound for distributions on
Accepted 29 June 2011
finite support is obtained. Properties of Chebyshev polynomials are employed.
Available online 19 July 2011
2011 Published by Elsevier B.V.
MSC:
60E05
60E10

Keywords:
Probability generating functions
Discrete distributions
Chebyshev polynomials

1. Introduction

In this paper, we consider a question of Mri regarding estimating the deviation of the kth terms of two discrete
probability distributions in terms of the supremum distance between their generating functions over the interval [0, 1].
In particular, suppose p = (p0 , p1 , . . .) and q = (q0 , q1 , . . .) are discrete probability distributions with generating functions
gp (t ) and gq (t ), respectively, and set

= p,q = max |gp (t ) gq (t )|. (1)


0t 1

We will assume throughout that p = q, so that = 0. In Mri (1996), it was asked whether for fixed k 0, the value
|pk qk |/p,q is bounded independent of p and q, i.e. whether there exist positive constants ck , k 0 such that for all p
and q and k 0
|pk qk | ck . (2)
The following weaker result was proven in Mri (1996).

Theorem 1 (Mri, 1996). For every real 0 < 1 and non-negative integer k, there exists a finite constant ck () not depending
on p or q such that

|pk qk | ck () . (3)

Recently, Mri (2009) proved the following theorem, which shows that = 1 in (3) is not possible, for sufficiently
small .

Corresponding author. Tel.: +1 336 758 5922; fax: +1 336 758 7190.
E-mail addresses: berenhks@wfu.edu (K.S. Berenhaut), baxley@wfu.edu (J.V. Baxley), lydarg6@wfu.edu (R.G. Lyday).

0167-7152/$ see front matter 2011 Published by Elsevier B.V.


doi:10.1016/j.spl.2011.06.019
K.S. Berenhaut et al. / Statistics and Probability Letters 81 (2011) 19401944 1941

Theorem 2 (Mri, 2009). Let k be an arbitrary positive integer and C a positive constant satisfying

1
C < Ck = 2k . (4)
2(2k)! log( 2 + 1)

Then, for every sufficiently small > 0 there exists a p and q such that max0t 1 |gp (t ) gq (t )| = and
2k
1
|pk qk | C log . (5)

In proving Theorem 2, the author employs a construction based on the Chebyshev polynomials of the first kind, {Tn }, and
in particular the spread polynomials {Sn } given by

1 Tn (1 2t )
Sn (t ) = . (6)
2
In Mri (2009), it is remarked that Chebyshev polynomials are characterized by their heavy oscillation with small
amplitude, and that the oscillation could somehow be necessary for obtaining large deviations in coefficients.
In what follows, we will consider further, the quantity

|pk qk |
, (7)
p,q
and show how Chebyshev polynomials arise in an optimal
sense in this setting. In particular, let F be the set of non-zero
real power series f satisfying f (t ) = k=0 ak (f )t with
k
k=0 ak (f ) = 0. In addition, let



F =
f F : |ak (f )| 2 , (8)
k =0

Pn = { F : is a polynomial and deg( ) n} (9)

and for f F ,

= f = max |f (t )|. (10)


0t 1

Note that we will use the notation ak (f ) = [xk ]f (x), to denote the coefficient of xk in f (x).
Considering = gp gq Pn , where p and q have finite support, for fixed k 0, we are interested in bounds on

sup , (11)
Pn

where

|ak ( )|
= (k) =
def
. (12)

Note that for c = 0,

c = (13)

and hence is scale invariant. Thus, when considering the value in (11), we may restrict attention to for in the subset
Pn Pn given by

Pn = Pn : s = 1 ,

(14)

where for Pn ,

n

s = |aj ( )|. (15)
j =1

We will prove the following theorem.


1942 K.S. Berenhaut et al. / Statistics and Probability Letters 81 (2011) 19401944


Theorem 3. For n 1, set n = 1 + cos 2n and define Tn via

Tn (x) = Tn (n x 1) = An,n xn + An,n1 xn1 + + An,0 . (16)

Then

Tn
Pn (17)
sTn

and

(2n)2k
sup = sup = Tn /s = |An,k | . (18)
Pn Pn
Tn (2k)!

Note that {Tn } as defined in (16) differs from the standard shifted Chebyshev polynomials {Tn }, wherein a shift from the
interval [1, 1] to the interval [0, 1] is employed via

Tn (x) = Tn (2x 1) (19)


for 0 x 1 (see for instance Mason and Handscomb, 2003).
Theorem 3 leads to the following result regarding deviations which implies that for all k 0 and > 0, a bound such
as in (2) is not possible.

Theorem 4. Suppose n 1 and k 0 are fixed and define {An,j } as in (16). Let Rn be the set of discrete distributions supported
on {0, 1, . . . , n}, i.e.

Rn = {p : pi = 0 for i > n}. (20)

Then, for all > 0 there exist p, q Rn , satisfying

|pk qk |
= |An,k |, (21)
p ,q
and furthermore,

|pk qk | (2n)2k
sup = |An,k | . (22)
p,qRn p,q (2k)!

Proof. Let = Tn /(sTn ) Pn and note that by (18),

= Tn = |An,k |. (23)

Set

p = (|a0 ( )| + a0 ( ), |a1 ( )| + a1 ( ), . . . , |an ( )| + an ( ), 0, 0, . . .) (24)

and

q = (|a0 ( )|, |a1 ( )|, . . . , |an ( )|, 0, 0, . . .). (25)

Then p and q are legitimate distributions on the set {0, 1, . . . , n}. To see this, note that, since Pn

n
n

|aj ( )| = (|aj ( )| + aj ( )) = 1. (26)
j =0 j=0

The result now follows from (18), since for all p and q, supported on {0, 1, . . . , n}, = gp gq Pn . 

Now, for {An,k } as in (16), define Pn,k Pn via

Pn,k = { Pn : ak ( ) = An,k }. (27)


Theorem 3 will follow directly from the following lemma.

Lemma 1. For k 0, n 1 and Pn,k ,

1. (28)
K.S. Berenhaut et al. / Statistics and Probability Letters 81 (2011) 19401944 1943

Before proving Lemma 1, we recall some standard facts regarding the Chebyshev polynomials, {Tn } (see for instance Davis,
1975).

Lemma 2. Suppose n 1. Then,


(a) Tn is a polynomial in x of degree n.
(b) |Tn (x)| 1 for x [1, 1].
(c) The zeros of Tn are given by

2i 1
xi = cos , i = 1, 2, . . . , n. (29)
2n

(d) Tn (ti ) = (1)i for i = 1, 2, . . . , n, where

i

ti = cos . (30)
n

Lemma 2 leads to the following regarding {Tn } as given in (16).

Lemma 3. Suppose n 1. Then,

(a) Tn is apolynomial in x of degree n.


(b) Tn (x) 1 for x [0, 1].

(c) Tn (1) = Tn (cos( /(2n))) = Tn (x1 ) = 0, and hence Tn Pn,k .


(d) For {xi } as in (29), define


2i1
1 + xi 1 + cos
x i = = 2n , i = 1, 2, . . . , n. (31)
n 1 + cos 2n

Then, the zeros of Tn are given by

0 < x n < x n1 < < x 1 = 1. (32)


(e) For {ti } as in (30), define
i
1 + ti 1 + cos
ti = = . n
(33)
n 1 + cos 2n
Then

Tn (ti ) = (1)i , i = 1, 2, . . . , n, (34)


and
0 = tn < tn1 < < t1 < 1. (35)

We are now in a position to prove Lemma 1.

Proof of Lemma 1. Suppose that there exists an Pn,k such that

< 1 (36)

and define the polynomial via

(x) = Tn (x) (x). (37)

Note that sign((ti )) = (1)i for 1 i n and since Tn , Pn , (1) = 0. Thus

has n zeros on (0, 1]. (38)

If k = 0, then (0) = 0, is of degree at most n with n + 1 zeros and Tn which is a contradiction to (36). If k = 1,
then has n 1 zeros on (0, 1) and (0) = 0. Thus is of degree at most n 1 with n zeros on [0, 1] and hence 0.
Thus, by (38), 0, and Tn , which is a contradiction. In general, if k = r, with 1 r n, then the rth derivative (r )
has n r zeros on (0, 1) and (r ) (0) = 0. Thus (r ) is of degree at most n r with n r + 1 zeros on [0, 1]. Thus (r ) 0
and has degree at most r 1 < n. Thus, Tn , which is a contradiction and the lemma is proven. 
1944 K.S. Berenhaut et al. / Statistics and Probability Letters 81 (2011) 19401944

For further optimality properties of the polynomials {Tn }, see for instance Davis (1975) or Mason and Handscomb (2003).
We now turn to a proof of Theorem 3.
Proof of Theorem 3. Suppose Pn . Note that if ak ( ) = 0, = 0. Hence, assume ak ( ) = 0 and define

An,k
= ,k = Pn,k . (39)
ak ( )
Noting that = , and employing Lemma 1 gives

|An,k |
= |An,k |. (40)
inf
Pn,k

The equalities in (18) follow upon noting that

Tn
Pn (41)
sTn

and
Tn /s = Tn = |An,k |. (42)
Tn

Finally, we have (see for instance Gautschi, 1979), for k 0,

(1)nk n(n + k 1)!2k 1 + cos k nn2k1 4k



|An,k | = 2n
, (43)

(n k)!(2k)! (2k)!
and the theorem is proven. 

Acknowledgments

We are very thankful to a referee for the comments and insights that improved this manuscript.
The first and third authors acknowledge support through NSF grant DMS-063664. The first author also acknowledges
financial support from a Z. Smith Reynolds Foundation Faculty Fellowship.

References

Davis, P.J., 1975. Interpolation and Approximation. Dover Publications, Inc., New York.
Gautschi, W., 1979. The condition of polynomials in power form. Math. Comp. 33 (145), 343352.
Mason, J.C., Handscomb, D.C., 2003. Chebyshev Polynomials. Chapman & Hall, CRC, Boca Raton, FL.
Mri, T.F., 1996. Bonferroni inequalities and deviations of discrete distributions. J. Appl. Probab. 33 (1), 115121.
Mri, T.F., 2009. Deviation of discrete distributionspositive and negative results. Statist. Probab. Lett. 79 (8), 10891096.

You might also like