Journal of Theoretical Probability, Vol. 13, No.

2, 2000

A Comparison Theorem on Moment Inequalities

Between Negatively Associated and Independent
Random Variables
Qi-Man Shao 1 , 2
Received January 2, 1998; revised August 22, 1999

Let [X i , 1in] be a negatively associated sequence, and let [X i* , 1in]

be a sequence of independent random variables such that X * i and X i have
the same distribution for each i=1, 2,..., n. It is shown in this paper that
Ef ( ni=1 X i )Ef ( ni=1 X *)
i for any convex function f on R 1 and that
Ef (max 1kn  i=k X i )Ef (max 1kn  ki=1 X i*) for any increasing convex
n

function. Hence, most of the well-known inequalities, such as the Rosenthal

maximal inequality and the Kolmogorov exponential inequality, remain true for
negatively associated random variables. In particular, the comparison theorem
on moment inequalities between negatively associated and independent random
variables extends the Hoeffding inequality on the probability bounds for the
sum of a random sample without replacement from a finite population.

KEY WORDS: Negative dependence; independent random variables; com-

parison theorem; moment inequality.

1. INTRODUCTION
A finite family of random variables [X i , 1in] is said to be negatively
associated (NA) if for every pair of disjoint subsets A 1 and A 2 of
[1, 2,..., n],
Cov[ f 1(X i , i # A), f 2(X j , j # A 2 )]0 (1.1)
whenever f 1 and f 2 are coordinatewise increasing and the covariance exists.
An infinite family is negatively associated if every finite subfamily is
negatively associated. This definition was introduced by Alam and
Saxena (1) and Joag-Dev and Proschan. (15) NA is one qualitative version of

1
Department of Mathematics, University of Oregon, Eugene, Oregon 97403, E-mail: qmshao
darkwing.uoregon.edu.
2
Research supported in part by NSF Grant DMS 9802451.
343

0894-9840000400-034318.000  2000 Plenum Publishing Corporation

344 Shao

negative dependence among random variables. For other versions of

negative dependence, such as negative upper (lower) orthant dependence,
reverse regular of order two in pairs, conditionally decreasing in sequence
and negatively dependent in sequence, we refer to Lehmann, (17) Block
et al., (7) Ebrahimi and Ghosh, (10) Jogdeo and Patil, (13) and Karlin and
Rinott. (16) Among those types of negative dependence, only the NA class
enjoys the important property of being closed under formation of increas-
ing functions of disjoint sets of random variables. (15) As pointed out and
proved by Joag-Dev and Proschan, (15) a number of well known multi-
variate distributions possess the NA property, such as (a) multinomial,
(b) convolution of unlike multinomials, (c) multivariate hypergeometric,
(d) Dirichlet, (e) Dirichlet compound multinomial, (f ) negatively correlated
normal distribution, (g) permutation distribution, (h) random sampling
without replacement, and (i) joint distribution of ranks. Because of their
wide applications in multivariate statistical analysis and reliability theory,
the notions of negatively associated random variables have received more
and more attention recently. We refer to Joag-Dev and Proschan (15)
for fundamental properties, Newman (20) for the central limit theorem,
Matula (18) for the three series theorem and Barbour et al. (2) for Poisson
approximation.
The main purpose of this paper is to establish a comparison theorem
on moments between the negatively associated and independent random
variables. It is shown that the moments of the partial sum (or the maximum
partial sum) of negatively associated random variables can be bounded
by those of independent random variables. Such a comparison result is
potentially useful to obtain limit theorems, especially strong laws of large
numbers, functional central limit theorems, BerryEsseen bounds, and laws
of the iterated logarithm.

Theorem 1. Let [X i , 1in] be a negatively associated sequence,

and let [X *
i , 1in] be a sequence of independent random variables
i and X i have the same distribution for each i=1, 2,..., n. Then
such that X *
n n
Ef
\ +
: X i Ef
i=1
\ : X i*
i=1
+ (1.2)

for any convex function f on R 1, whenever the expectation on the right

hand side of (1.2) exists. If f is a non-decreasing convex function, then
k k
Ef
\ max
1kn
i=1
+
: X i Ef
\ max
1kn
: X*
i=1
i
+ (1.3)

whenever the expectation on the right hand side of (1.3) exists.

A Comparison Theorem on Moment Inequalities 345

Theorem 1 enables us to show that most of the well-known inequal-

ities for independent random variables, such as the Rosenthal maximal
inequality and the Kolmogorov exponential inequality, etc., remain true for
negatively associated random variables.

Theorem 2. Let p1, [X i , 1in] be a sequence of negatively

associated mean zero random variables with E |X i | p < for every 1in,
and let [X *i , 1in] be a sequence of independent random variables
i and X i have the same distribution for each i=1, 2,..., n. Then
such that X *
n p n p

} } } }
E : X i E : X *
i=1
i
i=1
(1.4)

k p n p

} } } }
E max : X 2E : X *
1kn i=1
i
i=1
i (1.5)

k p n

} }
E max : X 2
1kn
: E |X |
i=1
i
3& p

i=1
i
p
for 1<p2 (1.6)

and
k p n p2 n
E max
1in } i=1
}
: X i 2(15 p ln p) p
{\ : EX 2i
i=1
+ + : E |X i | p
i=1
= for p>2
(1.7)

Theorem 3. Let [X i , 1in] be a sequence of negatively

associated random variables with zero means and finite second moments.
Let S k = ki=1 X i and B n = ni=1 EX 2i . Then for all x>0, a>0 and
0<:<1

P( max S k x)P( max X k >a)

1kn 1kn

1 x 2: 2 ax
+
1&:
exp &
\
2(ax+B n )
} 1+ ln 1+
3 Bn { \ +=+
(1.8)
and

P( max |S k | x)2P( max |X k | >a)

1kn 1kn

2 x 2: 2 ax
+
1&:
exp &
\
2(ax+B n )
} 1+ ln 1+
3 Bn { \ +=+
(1.9)
346 Shao

In particular, we have

P( max |S k | x)
1kn
x(12a)
x2 Bn
2P( max |X k | >a)+4 exp &
1kn \ 8B n
+4
+ \
4(xa+B n ) +
(1.10)

Next we apply Theorem 1 to random sampling without replacement.

Suppose a finite population C consist of N values c 1 , c 2 ,..., c N . Let
X 1 , X 2 ,..., X n denote a random sample without replacement from C and let
Y 1 , Y 2 ,..., Y n denote a random sample with replacement from C, where
1nN. It is known that the random variables Y 1 , Y 2 ,..., Y n are inde-
pendent and identically distributed and that Y i and X i have the same dis-
tribution for every i=1,..., n. It is shown by Joag-Dev and Proschan (15)
that X 1 , X 2 ,..., X n are negatively associated, hence by Theorem 1
n n
Ef
\ : X i Ef
i=1
+ \ : Yi
i=1
+ (1.11)

for any convex function f, and

k k
Ef
\ max
1kn i=1
: X i Ef
+ \ max
1kn i=1
: Yi
+ (1.12)

for any non-decreasing convex function.

(1.11) recovers the Hoeffding (11) inequality, while (1.12) is new even in
this special case.

Remark 1.1. For the moment inequalities for (positively) associated

random variables we refer to Birkel (5) and Shao and Yu. (25) It turns
out that a negatively associated sequence possesses much nicer moment
inequalities than a positively associated sequence.

Remark 1.2. (1.6) was proved by Matula (18) when p=2.

Remark 1.3. Su and Wang (27) initially proved that

n p n p2 n

}
E : X i C p
i=1
} {\ : EX 2i
i=1
+ + : E |X i | p
i=1
=
for p2 and some constant C p depending only on p. This paper was
motivated by the desire to refine their result.
A Comparison Theorem on Moment Inequalities 347

Remark 1.4. Zhao et al. (28) derived the following result almost
simultaneously with the author using a different argument:
k p

kn } i=1
}
E max : X i C p[n p2 max(EX 2i ) p2 +n max E |X i | p ]
in in

for p2

The proofs of Theorems 13 will be given in the next section. Applica-
tions to the weak invariance principle and the complete convergence are
discussed in Section 3. Application to the law of the iterated logarithm is
considered in Shao and Su. (24)

2. PROOFS
We first list some basic properties of the negatively associated random
variables and convex functions and two inequalities.
(H1) Increasing functions defined on disjoint subsets of a set of
negatively associated random variables are negatively
associated (see Ref. 15).
(H2) For any convex function f on R 1, the right derivative f $+ exists
and is increasing. Moreover for all real numbers a, b (see
Ref. 22)

b
f (b)& f (a)= | a
f $+(t) dt

(H3) For any &<x< and 1<p2 (see Lemma 2.3 of Ref. 9),

|1+x| p 1+ px+2 2& p |x| p

(H4) For any x0,

x x2 2
ln(1+x) +
1+x 2(1+x) 2
1+ ln(1+x)
3 \ +
In what follows, random variables marked with asterisks mean that
they are independent and every asterisked random variable and the
original one have the same distribution. For example, X *1 , X*
2 ,..., X *
n are
independent, and X *i and X i have the same distribution for each i.
348 Shao

Proof of Theorem 1. We prove (1.2), by induction on n. Let (Y 1 , Y 2 )

be an independent copy of (X 1 , X 2 ). It follows from (H2) that

f (X 1 +X 2 )+ f (Y 1 +Y 2 )& f (X 1 +Y 2 )& f (Y 1 +X 2 )
Y2
= | X2
( f $+(Y 1 +t)& f $+(X 1 +t)) dt


= | &
( f $+(Y 1 +t)& f $+(X 1 +t))(I [Y2 >t] &I [X2 >t] ) dt

Since f $+(x+t) and I [x>t] are increasing functions of x for each t,

f $+(X 1 +t) and I [X2 >t] are negatively associated. By applying Fubini's
theorem, we conclude that

2(Ef (X 1 +X 2 )&Ef (X *+X

1 *))
2

=E( f (X 1 +X 2 )+ f (Y 1 +Y 2 )& f (X 1 +Y 2 )& f (Y 1 +X 2 ))


= | &
Cov( f $+(X 1 +t), I [X2 >t] ) dt0 (2.1)

which proves (1.2) for n=2.

Assume that (1.2) holds for n=1. Let

g(x)=Ef (x+S n&1 )

By the induction hypothesis,

n&1

\
g(x)Ef x+ : X i*
i=1
+ (2.2)

It follows from (H1) that  n&1

i=1 X i and X n are also negatively associated.
Thus,
n n&1
Ef
\ i=1
+
: X i Ef (X n*+S* *)Ef X *+
n&1 )=Eg(X n n : X i*
\ i=1
+
by (2.2). This proves that (1.2) remains true for n.
To prove (1.3), write
k k
Sk = : X i , S k = : X *
i , M k = max S i , and M k = max S i
1ik 1ik
i=1 i=1
A Comparison Theorem on Moment Inequalities 349

It suffices to show that for each n and for any real numbers a and b,

Ef (max(a, b+max(0, M n )))Ef (max(a, b+max(0, M n ))) (2.3)

In fact, if (2.3) holds, letting

g 1(x)=Ef (max(x, x+max[0, max (S k &X 1 )]))

2in

yields

g 1(x)Ef (max(x, x+max[0, max (S k &X *)]))

1 (2.4)
2in

It follows from (H1) that X 1 and max[0, max 2kn(S k &X 1 )] are
negatively associated. Hence,

i
Ef
\ max : X j
1in j=1 +
=Ef (X 1 +max[0, max (S k &X 1 )])
2in

Ef (X *+[max[0,
1 max (S k &X 1 )]]*) [by (1.2)]
2in

=E(E[ f (X *+[
1 max (S k &X 1 )]*) | X *])
1
1in

=Eg 1(X *)
1

1 , X *+max[0,
Ef (max(X * 1 max (S k &X 1*)])) [by (2.4)]
2in

i
=Ef
\ max : X *
1in
j
j=1
+
as desired.
We now turn to the proof of (2.3), by induction on n. It is easy to see
that M 2 =X 1 +max(0, X 2 ), X 1 and max(0, X 2 ) are negatively associated,
and

h(x) :=f (max(a, b+max(0, x)))= f(max(a, b, b+x))

=max( f (max(a, b)), f (b+x))

350 Shao

is convex on R 1. Applying (1.2) to h, we obtain

Ef (max(a, b+max(0, M 1 )))=Eh(X 1 +max(0, X 2 ))

Eh(X *+(max(0,
1 X 2 ))*)
=Eh(X *+max(0,
1 X *))
2

=Ef (max(a, b+max(0, M 2 )))

This proves (2.3) for n=2.

Assume that (2.3) holds for n&1. Let

M 1(n&1)= max (S i+1 &X 1 )

1in&1

M 1(n&1)= max (S i+1 &X *)

1
1in&1

u(x)=Ef (max[a, b+max(0, x+max[0, M 1(n&1)])])

Note that

max[a, b+max(0, x+max(0, M 1(n&1)))]

=max[a, b, b+x+max(0, M 1(n&1))]
=max[max(a, b), b+x+max(0, M 1(n&1))]

By the induction hypothesis,

u(x)Ef (max[a, b+max(0, x+max[0, M 1(n&1)])])

It follows from (H1) that X 1 and M 1(n&1) are negatively associated.

Thus,

Ef (max(a, b+max(0, M n )))

=Ef (max[a, b+max(0, X 1 +max[0, M 1(n&1)])])
Ef (max[a, b+max(0, X *+max[0,
1 M 1(n&1)*])]) [by (1.2)]
=Eu(X *)Ef
1 (max[a, b+max(0, X *+max(0,
1 M 1(n&1)))])
=Ef (max(a, b+max(0, M n )))

This proves that (2.3) remains valid for n. The proof is complete. g
A Comparison Theorem on Moment Inequalities 351

Proof of Theorem 2. With f (x)= |x| p, (1.4) is an immediate conse-

quence of (1.2). Let f (x)=max(0, x) p. It follows from (1.3) that
k p k p

\ {
E max 0, max
1kn
: Xi
i=1
=+ \ {
E max 0, max
=+
1kn
: X i*
i=1
n p

\ {
2E max 0, : X *
=+
i=1
i (2.5)

where the last inequality comes from Theorem 2.3 of Choi and Klass. (8)
Similarly,
k p n p

\ {
E max 0, max
1kn
i=1
: (&X i )
=+ \ {
2E max 0, & : X *
i
i=1
=+ (2.6)

This proves (1.5), by (2.5) and (2.6).

From (H3), it is easy to show that
n p n n

}
E : X*
i
i=1
}
2 2& p : E |X * p
i | =2
2& p
: E |X i | p
i=1 i=1

for 1p2, which together with (1.5), implies (1.6). (1.7) follows from
(1.5) and Theorem 4.1 of Ref. 14. g
The following lemma may be known, but we could not locate a
reference.

Lemma 1. If [T i , 1in] is a non-negative supermartingale, then

for any 0<:<1,

E max T :i (ET 1 ) :(1&:) (2.7)

1in

Proof. From the proof of Corollary 5.4.1 in Ref. 26, one has

P( max T i >x)ET 1 x
1in

for every x>0. Thus


E max T :i =:
1in
| 0
x :&1P( max T i >x) dx
1in

ET1 
: | 0
x :&1 dx+:ET 1 | ET1
x :&2 dx=(ET 1 ) :(1&:)

as desired.
352 Shao

Proof of Theorem 3. (1.9) is a direct consequence of (1.8), while

(1.10) follows from (1.9) easily, by taking :=12 and considering whether
B n xa or B b >xa. So, we need only to prove (1.8). Let
i
t=a &1 ln(1+xaB n ), S i = : min(X j , a)
j=1
i
Ui = : Yj , i=1, 2,..., n (2.8)
j=1

where Y 1 , Y 2 ,..., Y n are independent random variables such that Y i and

min(X i , a) have the same distribution for each i=1, 2,..., n. It is easy to
show that [min(X i , a), 1in] is a negatively associated sequence.
Applying (1.3) yields

P( max S i x)P( max X i >a)+P( max S i x)

1in 1in 1in

P( max X i >a)+e &t:xE exp(t: max S i )

1in 1in

&t:x
P( max X i >a)+e E exp(t: max U i ) (2.9)
1in 1in

Put

J 1 =e &t:xE exp(t: max U i ) and T i =exp(tU i &(e ta &1&ta) a &2B i )

1in

Noting that (e x &1&x)x 2 is a non-decreasing function of x on R 1 and

that EY i 0, tY i ta, we have

(e tYi &1&tY i ) 2
Ee tYi =1+tEY i +E
\ Y 2i
Yi
+
1+(e ta &1&ta) a &2EY 2i 1+(e ta &1&ta) a &2EX 2i
exp((e ta &1&ta) a &2EX 2i )

Hence, [T i , 1in] is a supermartingale. Note that, by (H4)

(xa+B n ) a &2 ln(1+xaB n )&xa

xa (xa) 2 2
(xa+B n ) a &2
\ +
xa+B n 2(xa+B n ) 2
\
1+ ln(1+xaB n )
3 ++ &xa
x2 2
=
2(xa+B n ) \
1+ ln(1+xaB n )
3 +
A Comparison Theorem on Moment Inequalities 353

Thus, by Lemma 1,
ta &1&ta) a &2B
J 1 =e &t:xE[ max T i e (e i ]:
1in

exp(&t:x+:(e ta &1&ta) a &2B n ) E[ max T i ] :

1in

(1&:) &1 exp(&t:x+:(e &a &1&ta) a &2B n )

=(1&:) &1 exp(&:[a &1x ln(1+xaB n )
&(xaB n &ln(1+xaB n )) a &2B n ]) [by (2.8)]
=(1&:) &1 exp(&:[(xa+B n ) a &2 ln(1+xaB n )&xa])
:x 2 2
(1&:) &1 exp &
\ 2(xa+B n ) \
1+ ln(1+xaB n )
3 ++ (2.10)

This proves (1.8), by (2.9) and (2.10). g

3. APPLICATIONS
Throughout this section let [X i , i1] be a negatively associated
sequence and let S n = ni=1 X i . The maximal inequalities in Section 1
enable us to establish easily the weak invariance principle and the complete
convergence for negatively associated random variables.

3.1. The weak invariance principle

Let [X i , i1] be a strictly stationary negatively associated sequence
with EX i =0, EX 2i < and _ 2 :=EX 21 +2   i=2 E(X 1 X i )>0. Let

W n(t)=(_ 2n) &12 S [nt] for 0t1

The central limit theorem for negatively associated sequence is due to

Newman (19) (see also Ref. 20). The following theorem gives the weak
invariance principle, which solves an open problem of Newman. (20)

Theorem 4. The sequence of processes [W n(t), 0t1] converges

in distribution to the standard Wiener process.
Proof. It suffices to verify the tightness of W n (see Ref. 21 and
Theorem 15.1 of Ref. 4). By Theorem 8.4 of Ref. 4, we only need to show
354 Shao

that for any =>0, there exist a * and an integer n 0 such that for every
nn 0

P(max |S i | *n 12 )=* &2 (3.1)

in

Applying (1.10) with x=*n 12 and a=*n 1248, we obtain

P(max |S i | *n 12 )

in

2P( max |X i | >*n 1248)

1in

*2 EX 21 4
+4 exp &
\ 8EX 21+ \
+4
4(EX 21 +* 248) +
2(48) 2 * &2EX 21 I [ |X1 | >*n1248]
*2
+4 exp &
\ 8EX 21++4(12EX 21 * 2 ) 4

=* &2,

provided that * is sufficiently large. This proves (3.1). g

3.2. The complete convergence

Complete convergence gives a convergence rate with respect to the
strong law of large numbers. One can refer to Hsu and Robbins (12) and
Baum and Katz (3) for details. Applying the maximal inequality (1.7) or
(1.10), one can get the following result easily.

Theorem 5. Let 1r>12, pr1, and let [X i , i1] be a negatively

associated sequence of identically distributed random variables with
EX i =0 and E |X i | p <. Then

\=>0, : n pr&2P(max |S i | >=n r )< (3.2)
in
n=1

ACKNOWLEDGMENT
The author thanks Su and Wang for sending their manuscript to him
prior to its publication, Victor de la Pena and a referee for helpful comments.
A Comparison Theorem on Moment Inequalities 355

REFERENCES

1. Alam, K., and Saxena, K. M. L. Positive dependence in multivariate distributions. Comm.

Statist. A 10, 11831196.
2. Barbour, A. D., Holst, L., and Janson, S. (1992). Poisson Approximation, Oxford University
Press.
3. Baum, L. E., and Katz, M. (1965). Convergence rate in the law of large numbers. Trans.
Amer. Math. Soc. 120, 108123.
4. Billingsley, P. (1968). Convergence of Probability Measures, John Wiley, New York.
5. Birkel, T. (1988). Moment bounds for associated sequences. Ann. Probab. 16, 11841193.
6. Block, H. W., Savits, T. H., and Shaked, M. (1982). Some concepts of negative
dependence. Ann. Probab. 10, 765772.
7. Block, H. W., Savits, T. H., and Shaked, M. (1985). A concept of negative dependence
using stochastic ordering. Statist. Probab. Lett. 3, 8186.
8. Choi, K. P., and Klass, M. J. (1997). Some best possible prophet inequalities for convex
functions of sums of independent variates and unordered martingale difference sequences.
Ann. Probab. 25, 803811.
9. Csorgo , M., Horvath, L., and Shao, Q. M. (1993). Convergence of integrals of uniform
empirical and quantile processes. Stochastic Process. Appl. 45, 283294.
10. Ebrahimi, N., and Ghosh, M. (1981). Multivariate negative dependence. Comm. Statist. A
10, 307337.
11. Hoeffding, W. (1963). Probability inequalities for sums of bounded random variables.
J. Amer. Statist. Assoc. 58, 1330.
12. Hsu, P. L., and Robbins, H. (1947). Complete convergence and the law of large numbers.
Proc. Nat. Acad. Sci. U.S.A. 33, 2531.
13. Jogdeo, K., and Patil, G. P. (1975). Probability inequalities for certain multivariate
discrete distributions. Sankhya B 37, 158164.
14. Johnson, W. B., Schechtman, G., and Zinn, J. (1985). Best constants in moment
inequalities for linear combinations of independent and exchangeable random variables.
Ann. Probab. 13, 234253.
15. Joag-Dev, K., and Proschan, F. (1983). Negative association of random variables with
applications. Ann. Statist. 11, 286295.
16. Karlin, S., and Rinott, Y. (1980). Classes of orderings of measures and related correlation
inequalities. Il. Multivariate reverse rule distributions. J. Multivar. Anal. 10, 499516.
17. Lehmann, E. L. (1966). Some concepts of dependence. Ann. Math. Statist. 43, 1137
1153.
18. Matula, P. (1992). A note on the almost sure convergence of sums of negatively dependent
random variables. Statist. Probab. Lett. 15, 209213.
19. Newman, C. M. (1980). Normal fluctuations and the FKG systems. Comm. Math. Phys.
74, 119128.
20. Newman, C. M. (1984). Asymptotic independence and limit theorems for positively and
negatively dependent random variables. In Y. L. Tong (ed.), Inequalities in Statistics and
Probability, Institute of Mathematical Statistics, Hayward, CA, pp. 127140.
21. Newman, C. M., and Wright, A. L. (1981). An invariance principle for certain dependent
sequences. Ann. Probab. 9, 671675.
22. Roberts, A. W., and Verberg, D. E. (1973). Convex Functions, Academic Press, New York.
23. Rosenthal, H. P. (1970). On the subspaces of L p( p>2) spanned by sequence of indepen-
dent random variables. Israel J. Math. 8, 273303.
24. Shao, Q. M., and Su, C. (1999). The law of the iterated logarithm for negatively
associated random variables. Stoch. Process. 83, 139148.
356 Shao

25. Shao, Q. M., and Yu, H. (1996). Weak convergence for weighted empirical processes of
dependent sequences. Ann. Probab. 24, 20982127.
26. Stout, W. F. (1974). Almost Sure Convergence, Academic Press, New York.
27. Su, C., and Wang, Y. B. (1995). A moment inequality of negatively associated sequence
with its applications, Manuscript.
28. Su, C., Zhao, L. C., and Wang, Y. B. (1997). Moment inequalities and weak convergence
for negatively associated sequences, Science in China (A) 40, 172182.