Large Sample Theory

Ferguson

Exercises, Section 5, Central Limit Theorems.

1. (a) Using a Chebyshev’s-Inequality-like argument, show that (assuming the expec-

tations exist) E|X|2+α ≥ tα E[X 2 I(|X| ≥ t)] for all α > 0 and t > 0.
(b) Using part (a) and Lindeberg, prove Liapounov’s Theorem: Let Xn1 , Xn2 , . . . , Xnn
be independent with EXnj = 0 and E|Xnj |2+α < ∞ for some α > 0 and all n and j. Let

n
n L
Zn = j=1 Xnj and Bn2 = VarZn = j=1 VarXnj . Then Zn /Bn −→ N (0, 1), provided
1 
n
E|Xnj |2+α → 0 as n → ∞.
Bn2+α j=1

2. Let X1 , X2 , . . . be independent exponential random variables with

means β1 , β2 , . . .
n
respectively, and let Zn = X1 + · · · + Xn . Show that if max1≤j≤n βj2 / j=1 βj2 → 0 as
√ L
n → ∞, then (Zn − EZn )/ VarZn −→ N (0, 1). (Use Liapounov’s Theorem with α = 2.)
3. (a) Let X1 , X2 , . . . be independent Poisson random variables with means λ1 , λ2 , . . .
√ L
respectively,
and let Zn = X1 + · · · + Xn . Show that (Zn − EZn )/ VarZn −→ N (0, 1) if
n
and only if 1 λj → ∞.
(b) Show that this can provide an example to show you can get asymptotic normality
without the Lindeberg condition being satisfied.
4. As an illustration of the use of Kendall’s tau, here is a famous little example taken
from M. G. Kendall’s 1948 book, Rank Correlation Methods. Suppose a number of boys
are ranked according to their ability in mathematics and music. Such a pair of rankings
for ten boys, denoted by the letters A to J , might be as follows :

Boy : A B C D E F G H I J
Maths. : 7 4 3 10 6 2 9 8 1 5
Music : 5 7 3 10 1 9 6 2 8 4

Compute Tn , the number of discrepencies, and τn , Kendall’s rank correlation coefficient.

According to Kendall’s tables, P (Tn ≥ 32) = .054 for n = 10. Compare this probabilty
with the normal approximation, using the correction for continuity.
5. Let X be a Poisson random variable with mean λ = 10.
(a) Find the exact probability, P (X ≤ 10). (You may use the calculators found on
the web page http://www.math.ucla.edu/ tom/distributions/CONTENTS.html)
(b) Find the normal approximation to P (X ≤ 10).
(c) Find the first Edgeworth approximation to P (X ≤ 10).
(d) Find the second Edgeworth approximation to P (X ≤ 10). (Please make the
corrections for continuity in all these approximations.)

n
6. (a) Let X1 , X2 , . . . be i.i.d. with EXi = 0 and VarXi = 1. Let Sn = j=1 anj Xj

n
n
and Tn = j=1 bnj Xj , where anj and bnj are constants, normalized so that j=1 a2nj =

n
n
j=1 anj bnj . Assume that ρn → ρ, maxj≤n anj → 0 and
2 2
j=1 bnj = 1. Let ρn =
maxj≤n b2nj → 0 as n → ∞. Show that
 
L 1 ρ
(Sn , Tn ) −→ N ((0, 0), ).
ρ 1

n
n
(b) Apply the above to find the asymptotic joint distribution of 1 Xj and 1 jXj .
7. Let X1 , X2 , . . . be independent random variables with Xn having a uniform distri-
bution over the interval [−n, n].
P
(a) Does Xn −→ 0 as n → ∞?
√ L
(b) Does nXn −→ N (0, σ 2 ) for some number σ 2 ? If not, what can you say about
the large sample distribution of Xn ? (Maybe you should answer (a) after (b).)
8. The Coupon Collector’s Problem. Coupons are drawn at random with replacement
from among N distinct coupons until exactly n distinct coupons are observed. Let Sn
denote the total number of coupons drawn. Then Sn = Y1 + · · · + Yn , where Yj is the
number of coupons drawn after observing j − 1 distinct coupons until the jth distinct
coupon is drawn. Then Y1 , . . . , Yn are independent geometric random variables with means,
EYj = N/(N −j+1), and variances, Var(Yj ) = N (j−1)/(N −j+1)2 . Let n = Nr for some
√ L
fixed r ∈ (0, 1), and let N , and hence n, tend to ∞. Show n((Sn /n) − m) −→ N (0, σ 2 )
and find m and σ 2 as functions of r.
9. (a) Assume for a triangular array of independent variables that the Xnj are

n bounded, say |Xn,j | < A for all n and j and some fixed constant A. Let
uniformly
Sn = j=1 Xnj . Show that

Sn − E(Sn ) L
 −→ N (0, 1) provided Var(Sn ) → ∞.
Var(Sn )

(b) Apply this to the binomial random√variable, Yn ∈ B(n, pn ), (which is a sum of

independent Bernoullis) in the case pn = 1/ n to show that
 
√ Yn L
4
n √ − 1 −→ N (0, 1).
n

10. Let X1 , X2 , . . . , Xn be a sample from a distribution with distribution function

F (x) and density f(x). A simple estimate of the density at a point x is given by fˆn (x) =
F̂n (x + bn ) − F̂n (x − bn )
, where F̂n (x) is the sample distribution function. Here, bn is a
2bn
sequence of constants tending to zero at an appropriate rate. Note that Zn = n(F̂n (x +
bn ) − F̂n (x − bn )) has a binomial distribution, B(n, pn ) where pn = F (x + bn ) − F (x − bn ).
Assume that x is a point of continuity of f and that f(x) > 0.
√ L
(a) Using the preceding exercise, show that 2nbn (fˆn (x) − Efˆn (x)) −→ N (0, f(x)),
provided bn → 0 and nbn → ∞.
 (b) Assuming that f(x) is differentiable a suitable number of times, find extra condi-
√ L
tions on bn such that is it true that 2nbn (fˆn (x) − f(x)) −→ N (0, f(x)).
11. Let X1 , X2 , . . . be independent with P(Xn = n) = P(Xn = −n) = pn /2 and
P(Xn = 0) = 1 − pn , and let Zn = X1 + · · · + Xn . Take pn = 1/n2 .
(a) Show, using the converse to the Lindeberg-Feller Theorem, that Zn /Bn is not
asymptotically normal.
(b) What can you say about the asymptotic distribution of Zn or Zn /Bn?
12. Show that the Lindeberg condition implies the uniformly asymptotically negligi-
2
blity (UAN) condition: maxj≤n σnj /Bn2 → 0 as n → ∞.