central limit theorems for multicolor urns with dominated ... - Economia

hogheavyweightElectronics - Devices

Oct 8, 2013 (3 years and 10 months ago)

68 views

CENTRAL LIMIT THEOREMS FOR MULTICOLOR
URNS WITH DOMINATED COLORS
PATRIZIA BERTI,IRENE CRIMALDI,LUCA PRATELLI,AND PIETRO RIGO
Abstract.An urn contains balls of d ≥ 2 colors.At each time n ≥ 1,a ball is
drawn and then replaced together with a random number of balls of the same
color.Let A
n
=diag

A
n,1
,...,A
n,d

be the n-th reinforce matrix.Assuming
EA
n,j
= EA
n,1
for all n and j,a few CLT’s are available for such urns.In
real problems,however,it is more reasonable to assume
EA
n,j
= EA
n,1
whenever n ≥ 1 and 1 ≤ j ≤ d
0
,
liminf
n
EA
n,1
> limsup
n
EA
n,j
whenever j > d
0
,
for some integer 1 ≤ d
0
≤ d.Under this condition,the usual weak limit
theorems may fail,but it is still possible to prove CLT’s for some slightly
different random quantities.These random quantities are obtained neglecting
dominated colors,i.e.,colors from d
0
+1 to d,and allow the same inference
on the urn structure.The sequence (A
n
:n ≥ 1) is independent but need not
be identically distributed.Some statistical applications are given as well.
1.The problem
An urn contains a
j
> 0 balls of color j ∈ {1,...,d} where d ≥ 2.At each time
n ≥ 1,a ball is drawn and then replaced together with a random number of balls
of the same color.Say that A
n,j
≥ 0 balls of color j are added to the urn in case
X
n,j
= 1,where X
n,j
is the indicator of {ball of color j at time n}.Let
N
n,j
= a
j
+
n
￿
k=1
X
k,j
A
k,j
be the number of balls of color j in the urn at time n and
Z
n,j
=
N
n,j
￿
d
i=1
N
n,i
,M
n,j
=
￿
n
k=1
X
k,j
n
.
Fix j and let n → ∞.Then,under various conditions,Z
n,j
a.s.
−→ Z
(j)
for some
random variable Z
(j)
.This typically implies M
n,j
a.s.
−→Z
(j)
.A CLT is available as
well.Define in fact
C
n,j
=

n
￿
M
n,j
−Z
n,j
￿
and D
n,j
=

n
￿
Z
n,j
−Z
(j)
￿
.
As shown in [4],under reasonable conditions one obtains
(C
n,j
,D
n,j
) −→N(0,U
j
) ×N(0,V
j
) stably
Date:September 2,2009.
2000 Mathematics Subject Classification.60F05,60G57,60B10.
Key words and phrases.Central limit theorem – Clinical trials – Random probability measure
– Stable convergence – Urn model.
1
2 PATRIZIA BERTI,IRENE CRIMALDI,LUCA PRATELLI,AND PIETRO RIGO
for certain random variables U
j
and V
j
.A nice consequence is

n
￿
M
n,j
−Z
(j)
￿
= C
n,j
+D
n,j
−→N(0,U
j
+V
j
) stably.
Stable convergence,in the sense of Aldous and Renyi,is a strong formof convergence
in distribution.The definition is recalled in Section 3.
For (C
n,j
,D
n,j
) to converge,it is fundamental that EA
n,j
= EA
n,1
for all n and
j.In real problems,however,it is more sound to assume that
EA
n,j
= EA
n,1
whenever n ≥ 1 and 1 ≤ j ≤ d
0
,
liminf
n
EA
n,1
> limsup
n
EA
n,j
whenever j > d
0
,
for some integer 1 ≤ d
0
≤ d.Roughly speaking,when d
0
< d some colors (those
labelled from d
0
+ 1 to d) are dominated by the others.In this framework,for
j ∈ {1,...,d
0
},meaningful quantities are
C

n,j
=

n
￿
M

n,j
−Z

n,j
￿
and D

n,j
=

n
￿
Z

n,j
−Z
(j)
￿
where
M

n,j
=
￿
n
k=1
X
k,j
1 +
￿
d
0
i=1
￿
n
k=1
X
k,i
,Z

n,j
=
N
n,j
￿
d
0
i=1
N
n,i
.
If d
0
= d,then D

n,j
= D
n,j
and |C

n,j
−C
n,j
| ≤
1

n
.If d
0
< d,in a sense,dealing
with (C

n,j
,D

n,j
) amounts to neglecting dominated colors.
Our problem is to determine the limiting distribution of (C

n,j
,D

n,j
),under
reasonable conditions,when d
0
< d.
2.Motivations
Possibly,when d
0
< d,Z
n,j
and M
n,j
have a more transparent meaning than
their counterparts Z

n,j
and M

n,j
.Accordingly,a CLT for (C
n,j
,D
n,j
) is more
intriguing than a CLT for (C

n,j
,D

n,j
).So,why dealing with (C

n,j
,D

n,j
)?
The main reason is that (C
n,j
,D
n,j
) merely fails to converge in case
liminf
n
EA
n,j
>
1
2
liminf
n
EA
n,1
for some j > d
0
.(1)
Fix in fact j ≤ d
0
.Under some conditions,Z
n,j
a.s.
−→ Z
(j)
with Z
(j)
> 0 a.s.;
see Lemma 3.Furthermore,condition (1) yields

n
￿
d
i=d
0
+1
Z
n,i
a.s.
−→ ∞.(This
follows from Corollary 2 of [9] for d = 2,but it can be shown in general).Hence,
D

n,j
−D
n,j
≥ Z
n,j

n
d
￿
i=d
0
+1
Z
n,i
a.s.
−→∞.
Since D

n,j
converges stably,as proved in Theorem 4,D
n,j
fails to converge in
distribution under (1).
A CLT for D
n,j
,thus,is generally not available.A way out could be looking for
the right norming factors,that is,investigating whether
α
n

n
D
n,j
converges stably
for suitable constants α
n
.This is a reasonable solution but we discarded it.In fact,
as proved in Corollary 5,(C
n,j
,D
n,j
) converges stably whenever
limsup
n
EA
n,j
<
1
2
liminf
n
EA
n,1
for all j > d
0
.(1*)
So,the choice of α
n
depends on whether (1) or (1*) holds,and this is typi-
cally unknown in applications (think to clinical trials).In addition,dealing with
CLT FOR MULTICOLOR URNS 3
(C

n,j
,D

n,j
) looks natural (to us).Loosely speaking,as the problemoccurs because
there are some dominated colors,the trivial solution is just to neglect dominated
colors.
A next point to be discussed is the practical utility (if any) of a CLT for
(C

n,j
,D

n,j
) or (C
n,j
,D
n,j
).To fix ideas,we refer to (C

n,j
,D

n,j
) but the same
comments apply to (C
n,j
,D
n,j
) provided a CLT for the latter is available.It is
convenient to distinguish two situations.With reference to a real problem,sup-
pose the subset of non dominated colors is some J ⊂ {1,...,d} and not necessarily
{1,...,d
0
}.
If J is known,the main goal is to make inference on Z
(j)
,j ∈ J.To this
end,the limiting distribution of D

n,j
is useful.Knowing such distribution,for
instance,asymptotic confidence intervals for Z
(j)
are easily obtained.An example
(cf.Example 6) is given in Section 4.
But in various frameworks,J is actually unknown (think to clinical trials again).
Then,the main focus is to identify J and the limiting distribution of C

n,j
can help.
If such distribution is known,the hypothesis
H
0
:J = J

can be (asymptotically) tested for any J

⊂ {1,...,d} with card(J

) ≥ 2.Details
are in Examples 7 and 8.
A last remark is that our results become trivial for d
0
= 1.On one hand,this is
certainly a gap,as d
0
= 1 is important in applications.On the other hand,d
0
= 1
is itself a trivial case.Indeed,Z
(1)
= 1 a.s.,so that no inference on Z
(1)
is required.
This paper is the natural continuation of [4].While the latter deals with d
0
= d,
the present paper focus on d
0
< d.Indeed,our results hold for d
0
≤ d,but they
are contained in Corollary 9 of [4] in the particular case d
0
= d.In addition to
[4],a few papers which inspired and affected the present one are [1] and [9].Other
related references are [2],[3],[5],[7],[8],[10],[12].
The paper is organized as follows.Section 3 recalls some basic facts on stable
convergence.Section 4 includes the main results (Theorem 4 and Corollary 5).
Precisely,conditions for
(C

n,j
,D

n,j
) −→N(0,U
j
) ×N(0,V
j
) stably and
(C
n,j
,D
n,j
) −→N(0,U
j
) ×N(0,V
j
) stably under (1*)
are given,U
j
and V
j
being the same random variables mentioned in Section 1.As
a consequence,

n
￿
M

n,j
−Z
(j)
￿
= C

n,j
+D

n,j
−→N(0,U
j
+V
j
) stably and

n
￿
M
n,j
−Z
(j)
￿
= C
n,j
+D
n,j
−→N(0,U
j
+V
j
) stably under (1*).
Also,it is worth noting that D

n,j
and D
n,j
actually converge in a certain stronger
sense.
Finally,our proofs are admittedly long.To make the paper more readable,they
have been confined in Section 5 and in a final Appendix.
3.Stable convergence
Let (Ω,A,P) be a probability space and S a metric space.A kernel on S (or a
random probability measure on S) is a measurable collection N = {N(ω):ω ∈ Ω}
4 PATRIZIA BERTI,IRENE CRIMALDI,LUCA PRATELLI,AND PIETRO RIGO
of probability measures on the Borel σ-field on S.Measurability means that
N(ω)(f) =
￿
f(x) N(ω)(dx)
is A-measurable,as a function of ω ∈ Ω,for each bounded Borel map f:S →R.
Let (Y
n
) be a sequence of S-valued random variables and N a kernel on S.Both
(Y
n
) and N are defined on (Ω,A,P).Say that Y
n
converges stably to N in case
P
￿
Y
n
∈ ∙ | H
￿
−→E
￿
N(∙) | H
￿
weakly
for all H ∈ A such that P(H) > 0.
Clearly,if Y
n
→N stably,then Y
n
converges in distribution to the probability law
E
￿
N(∙)
￿
(just let H = Ω).We refer to [5] and references therein for more on stable
convergence.Here,we mention a strong form of stable convergence,introduced in
[5].Let F = (F
n
) be any sequence of sub-σ-fields of A.Say that Y
n
converges
F-stably in strong sense to N in case
E
￿
f(Y
n
) | F
n
￿
P
−→N(f) for all bounded continuous functions f:S →R.
Finally,we give two lemmas from[4].In both,G = (G
n
) is an increasing filtration.
Given kernels M and N on S,let M ×N denote the kernel on S ×S defined as
￿
M ×N
￿
(ω) = M(ω) ×N(ω) for all ω ∈ Ω.
Lemma 1.Let Y
n
and Z
n
be S-valued random variables and M and N kernels on
S,where S is a separable metric space.Suppose σ(Y
n
) ⊂ G
n
and σ(Z
n
) ⊂ G

for
all n,where G

= σ(∪
n
G
n
).Then,
(Y
n
,Z
n
) −→M ×N stably
provided Y
n
→M stably and Z
n
→N G-stably in strong sense.
Lemma 2.Let (Y
n
) be a G-adapted sequence of real randomvariables.If
￿

n=1
EY
2
n
n
2
<
∞ and E
￿
Y
n+1
| G
n
￿
a.s.
−→Y,for some random variable Y,then
n
￿
k≥n
Y
k
k
2
a.s.
−→Y and
1
n
n
￿
k=1
Y
k
a.s.
−→Y.
4.Main results
In the sequel,X
n,j
and A
n,j
,n ≥ 1,1 ≤ j ≤ d,are real random variables on the
probability space (Ω,A,P) and G = (G
n
:n ≥ 0),where
G
0
= {∅,Ω},G
n
= σ
￿
X
k,j
,A
k,j
:1 ≤ k ≤ n,1 ≤ j ≤ d
￿
.
Let N
n,j
= a
j
+
￿
n
k=1
X
k,j
A
k,j
where a
j
> 0 is a constant.We assume that
X
n,j
∈ {0,1},
d
￿
j=1
X
n,j
= 1,0 ≤ A
n,j
≤ β for some constant β,(2)
￿
A
n,j
:1 ≤ j ≤ d
￿
independent of G
n−1
∨σ
￿
X
n,j
:1 ≤ j ≤ d
￿
,
Z
n,j
= P
￿
X
n+1,j
= 1 | G
n
￿
=
N
n,j
￿
d
i=1
N
n,i
a.s..
Given an integer 1 ≤ d
0
≤ d,let us define
λ
0
= 0 if d
0
= d and λ
0
= max
d
0
<j≤d
limsup
n
EA
n,j
if d
0
< d.
CLT FOR MULTICOLOR URNS 5
We also assume that
EA
n,j
= EA
n,1
for n ≥ 1 and 1 ≤ j ≤ d
0
,(3)
m:= lim
n
EA
n,1
,m> λ
0
,q
j
:= lim
n
EA
2
n,j
for 1 ≤ j ≤ d
0
.
A few useful consequences are collected in the following lemma.Define
S

n
=
d
0
￿
i=1
N
n,i
and S
n
=
d
￿
i=1
N
n,i
.
Lemma 3.Under conditions (2)-(3),as n →∞,
S

n
n
a.s.
−→m and
S
n
n
a.s.
−→m,
n
1−λ
d
￿
i=d
0
+1
Z
n,i
a.s.
−→0 whenever d
0
< d and λ >
λ
0
m
,
Z
n,j
a.s.
−→Z
(j)
for each 1 ≤ j ≤ d
0
,
where each Z
(j)
is a random variable such that Z
(j)
> 0 a.s..
For d = 2,Lemma 3 follows from results in [9] and [10].For arbitrary d,it
is possibly known but we do not know of any reference.Accordingly,a proof of
Lemma 3 is given in the Appendix.We also note that,apart from a few particular
cases,the probability distribution of Z
(j)
is not known (even if d
0
= d).
We aim to settle the asymptotic behavior of
C
n,j
=

n
￿
M
n,j
−Z
n,j
￿
,D
n,j
=

n
￿
Z
n,j
−Z
(j)
￿
,
C

n,j
=

n
￿
M

n,j
−Z

n,j
￿
,D

n,j
=

n
￿
Z

n,j
−Z
(j)
￿
,
where j ∈ {1,...,d
0
} and
M
n,j
=
￿
n
k=1
X
k,j
n
,M

n,j
=
￿
n
k=1
X
k,j
1 +
￿
n
k=1
￿
d
0
i=1
X
k,i
,Z

n,j
=
N
n,j
￿
d
0
i=1
N
n,i
.
Let N(a,b) denote the one-dimensional Gaussian law with mean a and variance
b ≥ 0 (where N(a,0) = δ
a
).Note that N(0,L) is a kernel on R for each real non
negative random variable L.We are in a position to state our main result.
Theorem 4.If conditions (2)-(3) hold,then
C

n,j
−→N(0,U
j
) stably and
D

n,j
−→N(0,V
j
) G-stably in strong sense
for each j ∈ {1,...,d
0
},where U
j
= V
j
−Z
(j)
(1 −Z
(j)
)
and V
j
=
Z
(j)
m
2
￿
q
j
(1 −Z
(j)
)
2
+ Z
(j)
￿
i≤d
0
,i￿=j
q
i
Z
(i)
￿
.
In particular (by Lemma 1),
(C

n,j
,D

n,j
) −→N(0,U
j
) ×N(0,V
j
) stably.
As noted in Section 2,Theorem 4 has been thought for the case d
0
< d,and
it reduces to Corollary 9 of [4] in the particular case d
0
= d.We also remark
that some assumptions can be stated in a different form.In particular,under
6 PATRIZIA BERTI,IRENE CRIMALDI,LUCA PRATELLI,AND PIETRO RIGO
suitable extra conditions,Theorem 4 works even if (A
n,1
,...,A
n,d
) independent of
G
n−1
∨σ(X
n,1
,...,X
n,d
) is weakened into
(A
n,1
,...,A
n,d
) conditionally independent of (X
n,1
,...,X
n,d
) given G
n−1
;
see Remark 8 of [4].
The proof of Theorem 4 is deferred to Section 5.Here,we stress a few of its
consequences.
We already know (from Section 2) that (C
n,j
,D
n,j
) may fail to converge when
d
0
< d.There is a remarkable exception,however.
Corollary 5.Under conditions (2)-(3),if 2λ
0
< m (that is,(1*) holds) then
C
n,j
−→N(0,U
j
) stably and D
n,j
−→N(0,V
j
) G-stably in strong sense
for each j ∈ {1,...,d
0
}.In particular (by Lemma 1),
(C
n,j
,D
n,j
) −→N(0,U
j
) ×N(0,V
j
) stably.
Proof.By Theorem4,it is enough to prove D

n,j
−D
n,j
P
−→0 and C

n,j
−C
n,j
P
−→0.
It can be assumed d
0
< d.Note that
￿
￿
￿
D

n,j
−D
n,j
￿
￿
￿
=

nZ
n,j
￿
S
n
S

n
−1
￿

S
n
S

n

n
d
￿
i=d
0
+1
Z
n,i
,
C

n,j
−C
n,j
= D
n,j
−D

n,j
+ M
n,j

n
￿
d
i=d
0
+1
M
n,i

1
n
1
n
+
￿
d
0
i=1
M
n,i
.
By Lemma 3 and 2λ
0
< m,there is α >
1
2
such that n
α
￿
d
i=d
0
+1
Z
n,i
a.s.
−→0.Thus,
it remains only to see that

nM
n,i
a.s.
−→ 0 for each i > d
0
.Fix i > d
0
and define
L
n,i
=
￿
n
k=1
X
k,i
−Z
k−1,i

k
.Since (L
n,i
:n ≥ 1) is a G-martingale and
￿
n
E
￿
(L
n+1,i
−L
n,i
)
2
| G
n
￿
=
￿
n
Z
n,i
(1 −Z
n,i
)
n +1

￿
n
n
α
Z
n,i
n
1+α
< ∞ a.s.,
then L
n,i
converges a.s..By Kronecker lemma,
1

n
n
￿
k=1
(X
k,i
−Z
k−1,i
) =
1

n
n
￿
k=1

k
X
k,i
−Z
k−1,i

k
a.s.
−→ 0.
Since
1

n
￿
n
k=1
k
−α
−→0 and Z
k,i
=o(k
−α
) a.s.,it follows that

nM
n,i
=
1

n
n
￿
k=1
(X
k,i
−Z
k−1,i
) +
1

n
n−1
￿
k=0
Z
k,i
a.s.
−→ 0.
￿
Theorem 4 has some statistical implications as well.
Example 6.(A statistical use of D

n,j
).Suppose d
0
> 1,conditions (2)-(3)
hold,and fix j ≤ d
0
.Let (V
n,j
:n ≥ 1) be a sequence of consistent estimators of
V
j
,that is,V
n,j
P
−→V
j
and σ(V
n,j
) ⊂ D
n
for each n where
D
n
= σ
￿
X
k,i
A
k,i
,X
k,i
:1 ≤ k ≤ n,1 ≤ i ≤ d
￿
CLT FOR MULTICOLOR URNS 7
is the σ-field corresponding to the ”available data”.Since (V
n,j
) is G-adapted,
Theorem 4 yields
(D

n,j
,V
n,j
) −→N(0,V
j
) ×δ
V
j
G-stably in strong sense.
Since d
0
> 1,then 0 < Z
(j)
< 1 a.s.,or equivalently V
j
> 0 a.s..Hence,
I
{V
n,j
>0}
D

n,j
￿
V
n,j
−→N(0,1) G-stably in strong sense.
For large n,this fact allows to make inference on Z
(j)
.For instance,
Z

n,j
±
u
α

n
￿
V
n,j
provides an asymptotic confidence interval for Z
(j)
with (approximate) level 1 −α,
where u
α
is such that N(0,1)(u
α
,∞) =
α
2
.
An obvious consistent estimator of V
j
is
V
n,j
=
1
m
2
n
￿
Q
n,j
(1 −Z
n,j
)
2
+ Z
2
n,j
￿
i≤d
0
,i￿=j
Q
n,i
￿
where
m
n
=
￿
n
k=1
￿
d
i=1
X
k,i
A
k,i
n
and Q
n,i
=
￿
n
k=1
X
k,i
A
2
k,i
n
.
In fact,E(X
n+1,i
A
2
n+1,i
| G
n
) = Z
n,i
EA
2
n+,i
a.s.
−→ Z
(i)
q
i
for all i ≤ d
0
,so that
Lemma 2 implies Q
n,i
a.s.
−→Z
(i)
q
i
.Similarly,m
n
a.s.
−→m.Therefore,V
n,j
a.s.
−→V
j
.
Finally,Theorem4 also implies

n
￿
M

n,j
−Z
(j)
￿
= C

n,j
+D

n,j
−→N(0,U
j
+V
j
)
stably.So,another asymptotic confidence interval for Z
(j)
is M

n,j
±
u
α

n
￿
G
n,j
where G
n,j
is a consistent estimator of U
j
+V
j
.One merit of the latter interval is
that it does not depend on the initial composition a
i
,i = 1,...,d
0
(provided this
is true for G
n,j
as well).
Example 7.(A statistical use of C

n,j
).Suppose
EA
n,j
= µ
j
and var(A
n,j
) = σ
2
j
> 0 for all n ≥ 1 and 1 ≤ j ≤ d.
Suppose also that conditions (2)-(3) hold with some J ⊂ {1,...,d} in the place of
{1,...,d
0
},where card(J) > 1,that is
µ
r
= m> µ
s
whenever r ∈ J and s/∈ J.
Both J and card(J) are unknown,and we aim to test the hypothesis H
0
:J = J

where J

⊂ {1,...,d} and card(J

) > 1.Note that U
j
can be written as
U
j
=
Z
(j)
m
2
￿
(1 −Z
(j)
)
2
σ
2
j
+Z
(j)
￿
i∈J,i￿=j
Z
(i)
σ
2
i
￿
,j ∈ J.
Fix j ∈ J

.Under H
0
,a consistent estimator of U
j
is
U
n,j
=
Z
n,j
￿m
2
n
￿￿
i∈J

Z
n,i
￿
4
￿
(1 −Z
n,j
)
2
￿σ
2
n,j
+Z
n,j
￿
i∈J

,i￿=j
Z
n,i
￿σ
2
n,i
￿
where
￿m
n
=
1
card(J

)
￿
i∈J

￿m
n,i
,￿m
n,i
=
￿
n
k=1
X
k,i
A
k,i
￿
n
k=1
X
k,i
,￿σ
2
n,i
=
￿
n
k=1
X
k,i
￿
A
k,i
− ￿m
n,i
)
2
￿
n
k=1
X
k,i
.
8 PATRIZIA BERTI,IRENE CRIMALDI,LUCA PRATELLI,AND PIETRO RIGO
A couple of remarks are in order.First,
F
n
:=
￿
i∈J

Z
n,i
a.s.
−→1 under H
0
.
Indeed,the factor F
−4
n
has been inserted into the definition of U
n,j
in order that
K
n,j
fails to converge in distribution to N(0,1) when H
0
is false,where K
n,j
is
defined a few lines below.Second,
￿
n
k=1
X
k,i
> 0,eventually a.s.,so that ￿m
n,i
and
￿σ
2
n,i
are well defined.Similarly,￿m
n
> 0 eventually a.s..
Next,defining C

n,j
in the obvious way (i.e.,with J

in the place of {1,...,d
0
}),
Theorem 4 implies
K
n,j
:= I
{U
n,j
>0}
C

n,j
￿
U
n,j
−→N(0,1) stably under H
0
.
The converse is true as well,i.e.,K
n,j
fails to converge in distribution to N(0,1)
when H
0
is false.(This can be proved arguing as in Remark 10;we omit a formal
proof).Thus,an asymptotic critical region for H
0
,with approximate level α,is
￿
|K
n,j
| ≥ u
α
￿
with u
α
satisfying N(0,1)(u
α
,∞) =
α
2
.In real problems,some-
times,it is known in advance that j
0
∈ J for some j
0
∈ J

.Then,j = j
0
is a
natural choice in the previous test.Otherwise,an alternative option is a critical
region of the type
￿
i∈J

￿
|K
n,i
| ≥ u
i
￿
for suitable u
i
.This results in a more pow-
erful test but requires the joint limit distribution of
￿
K
n,i
:i ∈ J

￿
under H
0
.Such
a distribution is given in [4] when J

= {1,...,d},and can be easily obtained for
arbitrary J

using the techniques of this paper.
Example 8.(Another statistical use of C

n,j
).As in Example 7 (and under
the same assumptions),we aim to test H
0
:J = J

.Contrary to Example 7,
however,we are given observations A
k,j
,1 ≤ k ≤ n,1 ≤ j ≤ d,but no urn is
explicitly assigned.This is a main problem in statistical inference,usually faced
by the ANOVA techniques and their very many ramifications.A solution to this
problem is using C

n,j
,as in Example 7,after simulating the X
n,j
.The simulation
is not hard.Take in fact an i.i.d.sequence (Y
n
:n ≥ 0),independent of the A
k,j
,
with Y
0
uniformly distributed on (0,1).Let a
i
= 1,Z
0,i
=
1
d
for i = 1,...,d,and
X
1,j
= I
{F
0,j−1
<Y
0
≤F
0,j
}
where F
0,j
=
j
￿
i=1
Z
0,i
and F
0,0
= 0.
By induction,for each n ≥ 1,
X
n+1,j
= I
{F
n,j−1
<Y
n
≤F
n,j
}
where F
n,j
=
j
￿
i=1
Z
n,i
,
F
n,0
= 0 and Z
n,i
=
1 +
￿
n
k=1
X
k,i
A
k,i
d +
￿
d
r=1
￿
n
k=1
X
k,r
A
k,r
.
Now,H
0
can be asymptotically tested as in Example 7.In addition,since A
k,i
is
actually observed (unlike Example 7,where only X
k,i
A
k,i
is observed),￿m
n,i
and
￿σ
2
n,i
can be taken as
￿m
n,i
=
￿
n
k=1
A
k,i
n
and ￿σ
2
n,i
=
￿
n
k=1
￿
A
k,i
− ￿m
n,i
)
2
n
.
CLT FOR MULTICOLOR URNS 9
Clearly,this procedure needs to be much developed and investigated.By now,
however,it looks (to us) potentially fruitful.
5.Proof of Theorem 4
Next result,of possible independent interest,is inspired by ideas in [4] and [5].
Proposition 9.Let F = (F
n
) be an increasing filtration and (Y
n
) an F-adapted
sequence of real integrable random variables.Suppose Y
n
a.s.
−→Y for some random
variable Y and H
n
∈ F
n
are events satisfying P(H
c
n
i.o.) = 0.Then,

n(Y
n
−Y ) −→N(0,V ) F-stably in strong sense,
for some random variable V,whenever
E
￿
I
H
n
￿
E(Y
n+1
| F
n
) −Y
n
￿
2
￿
= o(n
−3
),(4)

nE
￿
I
H
n
sup
k≥n
|E(Y
k+1
| F
k
) −Y
k+1
|
￿
−→0,(5)
n
￿
k≥n
(Y
k
−Y
k+1
)
2
P
−→V.(6)
Proof.We base on the following result,which is a consequence of Corollary 7 of [5].
Let (L
n
) be an F-martingale such that L
n
a.s.
−→L.Then,

n(L
n
−L) −→N(0,V )
F-stably in strong sense whenever
(i) lim
n

nE
￿
I
H
n
sup
k≥n
|L
k
−L
k+1
|
￿
= 0;(ii) n
￿
k≥n
(L
k
−L
k+1
)
2
P
−→V.
Next,define the F-martingale
L
0
= Y
0
,L
n
= Y
n

n−1
￿
k=0
E
￿
Y
k+1
−Y
k
| F
k
￿
.
Define also T
n
= E
￿
Y
n+1
−Y
n
| F
n
￿
.By (4),

n
￿
k≥n
E|I
H
k
T
k
| ≤

n
￿
k≥n
￿
E(I
H
k
T
2
k
) =

n
￿
k≥n
o(k
−3/2
) −→0.(7)
In particular,
￿

k=0
E|I
H
k
T
k
| < ∞ so that
￿
n−1
k=0
I
H
k
T
k
converges a.s..Since Y
n
converges a.s.and P(I
H
n
￿= 1 i.o.) = 0,
L
n
= Y
n

n−1
￿
k=0
T
k
a.s.
−→L for some random variable L.
Next,write
(L
n
−L) −(Y
n
−Y ) =
￿
k≥n
(L
k
−L
k+1
) −
￿
k≥n
(Y
k
−Y
k+1
) =
￿
k≥n
T
k
.
Recalling

n
￿
k≥n
|I
H
k
T
k
|
P
−→0 (thanks to (7)),one obtains
￿
￿
￿

n(L
n
−L) −

n(Y
n
−Y )
￿
￿
￿
=

n
￿
￿
￿
￿
k≥n
T
k
￿
￿
￿


n
￿
k≥n
|I
H
k
T
k
| +

n
￿
k≥n
|(1 −I
H
k
) T
k
|
P
−→0.
10 PATRIZIA BERTI,IRENE CRIMALDI,LUCA PRATELLI,AND PIETRO RIGO
Thus,it suffices to prove

n(L
n
−L) −→N(0,V ) F-stably in strong sense,that
is,to prove conditions (i) and (ii).Condition (i) reduces to (5) after noting that
L
k
−L
k+1
= E(Y
k+1
| F
k
) −Y
k+1
.
As to (ii),since L
k
−L
k+1
= Y
k
−Y
k+1
+T
k
,condition (6) yields
n
￿
k≥n
(L
k
−L
k+1
)
2
= V + n
￿
k≥n
￿
T
2
k
+2T
k
(Y
k
−Y
k+1
)
￿
+o
P
(1).
By (4),E
￿
n
￿
k≥n
I
H
k
T
2
k
￿
= n
￿
k≥n
o(k
−3
) −→ 0.Since P(I
H
n
￿= 1 i.o.) = 0,
then n
￿
k≥n
T
2
k
P
−→0.Because of (6),this also implies
￿
n
￿
k≥n
T
k
(Y
k
−Y
k+1
)
￿
2
≤ n
￿
k≥n
T
2
k
∙ n
￿
k≥n
(Y
k
−Y
k+1
)
2
P
−→0.
Therefore,condition (ii) holds and this concludes the proof.￿
We next turn to Theorem 4.From now on,it is assumed d
0
< d (the case d
0
= d
has been settled in [4]).Recall the notations S

n
=
￿
d
0
i=1
N
n,i
and S
n
=
￿
d
i=1
N
n,i
.
Note also that,by a straightforward calculation,
Z

n+1,j
−Z

n,j
=
X
n+1,j
A
n+1,j
S

n
+A
n+1,j
− Z

n,j
d
0
￿
i=1
X
n+1,i
A
n+1,i
S

n
+A
n+1,i
.
Proof of Theorem 4.The proof is split into two steps.
(i) D

n,j
−→N(0,V
j
) G-stably in strong sense.
By Lemma 3,Z

n,j
=
Z
n,j

d
0
i=1
Z
n,i
a.s.
−→ Z
(j)
.Further,P(2S

n
< nm i.o.) = 0 since
S

n
n
a.s.
−→m.Hence,by Proposition 9,it suffices to prove conditions (4)-(5)-(6) with
F
n
= G
n
,Y
n
= Z

n,j
,Y = Z
(j)
,H
n
= {2S

n
≥ nm},V = V
j
.
Conditions (4) and (5) trivially hold.As to (4),note that
Z

n,j
d
0
￿
i=1
Z
n,i
= Z
n,j
d
0
￿
i=1
Z

n,i
= Z
n,j
.
Therefore,
E
￿
Z

n+1,j
−Z

n,j
| G
n
￿
= Z
n,j
E
￿
A
n+1,j
S

n
+A
n+1,j
| G
n
￿
− Z

n,j
d
0
￿
i=1
Z
n,i
E
￿
A
n+1,i
S

n
+A
n+1,i
| G
n
￿
= −Z
n,j
E
￿ A
2
n+1,j
S

n
(S

n
+A
n+1,j
)
| G
n
￿
+ Z

n,j
d
0
￿
i=1
Z
n,i
E
￿ A
2
n+1,i
S

n
(S

n
+A
n+1,i
)
| G
n
￿
,
so that I
H
n
￿
￿
￿ E
￿
Z

n+1,j
−Z

n,j
| G
n
￿
￿
￿
￿ ≤ I
H
n
d
0
β
2
(S

n
)
2

4 d
0
β
2
m
2
1
n
2
.
CLT FOR MULTICOLOR URNS 11
As to (5),
￿
￿
￿ E
￿
Z

k+1,j
| G
k
￿
−Z

k+1,j
￿
￿
￿ ≤

S

k
+ N
k,j
￿
￿
￿E
￿
1
S

k+1
| G
k
￿

1
S

k+1
￿
￿
￿

2 β
S

k
+ N
k,j
￿
1
S

k

1
S

k

￿


S

k
,
so that I
H
n
sup
k≥n
|E(Z

k+1,j
| G
k
) −Z

k+1,j
| ≤ I
H
n
3 β
S

n

6 β
m
1
n
.
Finally,let us turn to (6).For every i ∈ {1,...,d
0
},
n
2
E
￿
A
2
n+1,i
(S

n
+A
n+1,i
)
2
| G
n
￿
≤ n
2
EA
2
n+1,i
(S

n
)
2
a.s.
−→
q
i
m
2
and
n
2
E
￿ A
2
n+1,i
(S

n
+A
n+1,i
)
2
| G
n
￿
≥ n
2
EA
2
n+1,i
(S

n
+β)
2
a.s.
−→
q
i
m
2
.
Since X
n+1,r
X
n+1,s
= 0 for r ￿= s,it follows that
n
2
E
￿
(Z

n+1,j
−Z

n,j
)
2
| G
n
￿
= n
2
Z
n,j
(1 −Z

n,j
)
2
E
￿
A
2
n+1,j
(S

n
+A
n+1,j
)
2
| G
n
￿
+
+n
2
(Z

n,j
)
2
￿
i≤d
0
,i￿=j
Z
n,i
E
￿ A
2
n+1,i
(S

n
+A
n+1,i
)
2
| G
n
￿
a.s.
−→Z
(j)
(1 −Z
(j)
)
2
q
j
m
2
+ Z
2
(j)
￿
i≤d
0
,i￿=j
Z
(i)
q
i
m
2
= V
j
.
Let R
n+1
= (n +1)
2
I
H
n
(Z

n+1,j
−Z

n,j
)
2
.Since H
n
∈ G
n
and P(I
H
n
￿= 1 i.o.) = 0,
then E(R
n+1
| G
n
)
a.s.
−→V
j
.On noting that |Z

n+1,j
−Z

n,j
| ≤
d
0
β
S

n
,
ER
2
n
n
2
≤ (d
0
β)
4
n
2
E
￿
I
H
n−1
(S

n−1
)
4
￿

￿
2d
0
β
m
￿
4
n
2
(n −1)
4
.
By Lemma 2 (applied with Y
n
= R
n
),
n
￿
k≥n
I
H
k
(Z

k+1,j
−Z

k,j
)
2
=
n
n +1
(n +1)
￿
k≥n+1
R
k
k
2
a.s.
−→ V
j
.
Since P(I
H
n
￿= 1 i.o.) = 0 then n
￿
k≥n
(Z

k+1,j
−Z

k,j
)
2
a.s.
−→ V
j
,that is,condition
(6) holds.
(ii) C

n,j
−→N(0,U
j
) stably.
Define T
n,i
=
￿
n
k=1
X
k,i
,T
0,i
= 0,and note that
C

n,j
= −

nZ

n,j
1 +
￿
d
0
i=1
T
n,i
+
n
1 +
￿
d
0
i=1
T
n,i
T
n,j
−Z

n,j
￿
d
0
i=1
T
n,i

n
and
T
n,j
−Z

n,j
d
0
￿
i=1
T
n,i
=
n
￿
k=1
￿
X
k,j
−Z

k,j
d
0
￿
i=1
T
k,i
+Z

k−1,j
d
0
￿
i=1
T
k−1,i
￿
=
n
￿
k=1
￿
X
k,j
−Z

k−1,j
d
0
￿
i=1
X
k,i

d
0
￿
i=1
T
k,i
(Z

k,j
−Z

k−1,j
)
￿
.
12 PATRIZIA BERTI,IRENE CRIMALDI,LUCA PRATELLI,AND PIETRO RIGO
Define also H
n
= {2S

n
≥ nm} and
C
∗∗
n,j
=
1

n
n
￿
k=1
I
H
k−1
￿
X
k,j
−Z

k−1,j
d
0
￿
i=1
X
k,i
+
d
0
￿
i=1
T
k−1,i
￿
E(Z

k,j
| G
k−1
) −Z

k,j
￿￿
.
Recalling (from point (i)) that P(I
H
n
￿= 1 i.o.) = 0,lim
n

d
0
i=1
T
n,i
n
= 1 a.s.,and
I
H
k−1
￿
￿
￿
E
￿
Z

k,j
−Z

k−1,j
| G
k−1
￿
￿
￿
￿

c
(k−1)
2
a.s.for some constant c,it is not hard to
see that C

n,j
−→N stably if and only if C
∗∗
n,j
−→N stably for any kernel N.
We next prove C
∗∗
n,j
−→N(0,U
j
) stably.For k = 1,...,n,let F
n,k
= G
k
and
Y
n,k
=
I
H
k−1
￿
X
k,j
− Z

k−1,j
￿
d
0
i=1
X
k,i
+
￿
d
0
i=1
T
k−1,i
(E(Z

k,j
| G
k−1
) −Z

k,j
)
￿

n
.
Since E(Y
n,k
| F
n,k−1
) = 0 a.s.,the martingale CLT (see Theorem 3.2 of [6])
applies.As a consequence,C
∗∗
n,j
=
￿
n
k=1
Y
n,k
−→N(0,U
j
) stably provided
sup
n
E
￿
max
1≤k≤n
Y
2
n,k
￿
< ∞;max
1≤k≤n
|Y
n,k
|
P
−→0;
n
￿
k=1
Y
2
n,k
P
−→U
j
.
As shown in point (i),I
H
k−1
￿
￿
￿E(Z

k,j
| G
k−1
) − Z

k,j
￿
￿
￿ ≤
d
k−1
a.s.for a suitable
constant d.Hence,the first two conditions follow from
Y
2
n,k

2
n
+
2
n
I
H
k−1
(k −1)
2
￿
E(Z

k,j
| G
k−1
) −Z

k,j
￿
2

2 (1 +d
2
)
n
a.s..
To conclude the proof,it remains to see that
￿
n
k=1
Y
2
n,k
P
−→ U
j
.After some
(long) algebra,the latter condition is shown equivalent to
1
n
n
￿
k=1
I
H
k−1
￿
X
k,j
−Z

k−1,j
+k
￿
Z

k−1,j
−Z

k,j
￿￿
2
P
−→U
j
.(8)
Let R
n+1
= (n +1)
2
I
H
n
(Z

n+1,j
−Z

n,j
)
2
.Since E(R
n+1
| G
n
)
a.s.
−→V
j
,as shown in
point (i),Lemma 2 implies
1
n
n
￿
k=1
I
H
k−1
k
2
￿
Z

k−1,j
−Z

k,j
￿
2 a.s.
−→V
j
.
A direct calculation shows that
1
n
n
￿
k=1
I
H
k−1
(X
k,j
−Z

k−1,j
)
2
a.s.
−→Z
(j)
(1 −Z
(j)
).
Finally,observe the following facts
￿
Z

n,j
−Z

n+1,j
￿
(X
n+1,j
−Z

n,j
) = −(1 −Z

n,j
)
X
n+1,j
A
n+1,j
S

n
+A
n+1,j
−Z

n,j
(Z

n,j
−Z

n+1,j
),
(n +1) Z

n,j
I
H
n
￿
￿
￿E
￿
Z

n,j
−Z

n+1,j
| G
n
￿
￿
￿
￿ ≤
c (n +1)
n
2
a.s.
−→0,
(n +1) E
￿
X
n+1,j
A
n+1,j
S

n
+A
n+1,j
| G
n
￿

n +1
S

n
Z
n,j
EA
n+1,j
a.s.
−→Z
(j)
,
(n +1) E
￿
X
n+1,j
A
n+1,j
S

n
+A
n+1,j
| G
n
￿

n +1
S

n

Z
n,j
EA
n+1,j
a.s.
−→Z
(j)
.
CLT FOR MULTICOLOR URNS 13
Therefore,
(n +1) I
H
n
E
￿
(Z

n,j
−Z

n+1,j
) (X
n+1,j
−Z

n,j
) | G
n
￿
a.s.
−→−Z
(j)
(1 −Z
(j)
)
and Lemma 2 again implies
2
n
n
￿
k=1
I
H
k−1
k (Z

k−1,j
−Z

k,j
) (X
k,j
−Z

k−1,j
)
a.s.
−→−2Z
(j)
(1 −Z
(j)
).
Thus condition (8) holds,and this concludes the proof.
￿
Remark 10.Point (ii) admits a simpler proof in case EA
k,j
= mfor all k ≥ 1 and
1 ≤ j ≤ d
0
.This happens,in particular,if the sequence (A
n,1
,...,A
n,d
) is i.i.d..
Given the real numbers b
1
,...,b
d
0
,define
Y
n,k
=
1

n
d
0
￿
j=1
b
j
X
k,j
(A
k,j
−EA
k,j
),F
n,k
= G
k
,k = 1,...,n.
By Lemma 2,
￿
n
k=1
Y
2
n,k
a.s.
−→
￿
d
0
j=1
b
2
j
(q
j
− m
2
) Z
(j)
:= L.Thus,the martingale
CLT implies
￿
n
k=1
Y
n,k
−→ N(0,L) stably.Since b
1
,...,b
d
0
are arbitrary con-
stants,
￿
￿
n
k=1
X
k,j
(A
k,j
−EA
k,j
)

n
:j = 1,...,d
0
￿
−→N
d
0
(0,Σ) stably
where Σ is the diagonal matrix with σ
j,j
= (q
j
−m
2
)Z
(j)
.Let T
n,j
=
￿
n
k=1
X
k,j
.
Since EA
k,j
= m and
T
n,j
n
a.s.
−→Z
(j)
> 0 for all j ≤ d
0
,one also obtains
￿

n
￿
￿
n
k=1
X
k,j
A
k,j
T
n,j
−m
￿
:j = 1,...,d
0
￿
−→N
d
0
(0,Γ) stably
where Γ is diagonal with γ
j,j
=
(q
j
−m
2
)
Z
(j)
.Next,write
￿
C
n,j
:=

n
￿
T
n,j
￿
d
0
i=1
T
n,i

￿
n
k=1
X
k,j
A
k,j
￿
d
0
i=1
￿
n
k=1
X
k,i
A
k,i
￿
=
T
n,j
￿
d
0
i=1
￿
n
k=1
X
k,i
A
k,i
￿
i≤d
0
,i￿=j
T
n,i
￿
d
0
i=1
T
n,i

n
￿
m−
￿
n
k=1
X
k,j
A
k,j
T
n,j
￿
+
+
T
n,j
￿
d
0
i=1
￿
n
k=1
X
k,i
A
k,i
1
￿
d
0
i=1
T
n,i
￿
i≤d
0
,i￿=j
T
n,i

n
￿
￿
n
k=1
X
k,i
A
k,i
T
n,i
−m
￿
.
Clearly,C

n,j

￿
C
n,j
a.s.
−→ 0.To conclude the proof,it suffices noting that
￿
C
n,j
converges stably to the Gaussian kernel with mean 0 and variance
￿
Z
(j)
(1 −Z
(j)
)
m
￿
2
q
j
−m
2
Z
(j)
+
Z
2
(j)
m
2
￿
i≤d
0
,i￿=j
Z
2
(i)
q
i
−m
2
Z
(i)
= U
j
.
APPENDIX
14 PATRIZIA BERTI,IRENE CRIMALDI,LUCA PRATELLI,AND PIETRO RIGO
Proof of Lemma 3.We first note that N
n,j
a.s.
−→ ∞ for each j ≤ d
0
.Arguing
as in the proof of Proposition 2.3 of [9],in fact,
￿

n=1
X
n,j
= ∞ a.s..Hence,
￿
n
k=1
X
k,j
EA
k,j
a.s
−→∞,and N
n,j
a.s.
−→∞follows from
L
n
= N
n,j

￿
a
j
+
n
￿
k=1
X
k,j
EA
k,j
￿
=
n
￿
k=1
X
k,j
￿
A
k,j
−EA
k,j
￿
is a G-martingale such that |L
n+1
−L
n
| ≤ β for all n.
We also need the following fact.
CLAIM:τ
n,j
=
N
n,j
(S

n
)
λ
converges a.s.for all j > d
0
and λ ∈ (
λ
0
m
,1).
On noting that (1 −x)
λ
≤ 1 −λx for 0 ≤ x ≤ 1 and
￿
d
0
i=1
Z
n,i
=
S

n
S
n
,one can
estimate as follows
E
￿
τ
n+1,j
τ
n,j
−1 | G
n
￿
= E
￿
N
n,j
+X
n+1,j
A
n+1,j
N
n,j
(
S

n
S

n+1
)
λ
| G
n
￿
−1

Z
n,j
EA
n+1,j
N
n,j
+ E
￿
(
S

n
S

n+1
)
λ
−1 | G
n
￿

EA
n+1,j
S
n
− λ
d
0
￿
i=1
E
￿
X
n+1,i
A
n+1,i
S

n+1
| G
n
￿

EA
n+1,j
S
n
− λ
d
0
￿
i=1
Z
n,i
EA
n+1,i
S

n

=
EA
n+1,j
S
n
− λEA
n+1,1
S

n
S
n
(S

n
+β)
=
1
S
n
￿
EA
n+1,j
− λEA
n+1,1
S

n
S

n

￿
a.s..
Since limsup
n
￿
EA
n+1,j
−λEA
n+1,1
￿
≤ λ
0
−λm< 0,there are ￿ > 0 and n
0
≥ 1
such that EA
n+1,j
−λEA
n+1,1
≤ −￿ whenever n ≥ n
0
.Thus,
E
￿
τ
n+1,j
−τ
n,j
| G
n
￿
= τ
n,j
E
￿
τ
n+1,j
τ
n,j
−1 | G
n
￿
≤ 0 a.s.whenever n ≥ n
0
and S

n
≥ c
for a suitable constant c.Since S

n
≥ N
n,1
a.s.
−→∞,thus,(τ
n,j
) is eventually a non
negative G-super-martingale.Hence,τ
n,j
converges a.s..
Let λ ∈ (
λ
0
m
,1).A first consequence of the Claim is that Z
n,j

τ
n,j
S
1−λ
n
a.s.
−→0 for
each j > d
0
.Letting Y
n
=
￿
d
0
i=1
X
n,i
A
n,i
,this implies
E(Y
n+1
| G
n
) =
d
0
￿
i=1
Z
n,i
EA
n+1,i
= EA
n+1,1
(1 −
d
￿
i=d
0
+1
Z
n,i
)
a.s.
−→m.
Thus,Lemma 2 yields
S

n
n
a.s.
−→m.Similarly,
S
n
n
a.s.
−→m.Applying the Claim again,
n
1−λ
Z
n,j
= (
n
S
n
)
1−λ
(
S

n
S
n
)
λ
τ
n,j
converges a.s.for each j > d
0
.
Since j > d
0
and λ ∈ (
λ
0
m
,1) are arbitrary,it follows that n
1−λ
￿
d
j=d
0
+1
Z
n,j
a.s.
−→0
for each λ >
λ
0
m
.
CLT FOR MULTICOLOR URNS 15
Next,fix j ≤ d
0
.For Z
n,j
to converge a.s.,it suffices that
￿
n
E
￿
Z
n+1,j
−Z
n,j
| G
n
￿
and
￿
n
E
￿
(Z
n+1,j
−Z
n,j
)
2
| G
n
￿
converge a.s.;
see Lemma 3.2 of [11].Since
Z
n+1,j
−Z
n,j
=
X
n+1,j
A
n+1,j
S
n
+A
n+1,j
− Z
n,j
d
￿
i=1
X
n+1,i
A
n+1,i
S
n
+A
n+1,i
,
then |Z
n+1,j
−Z
n,j
| ≤

S
n
.Hence,
￿
n
E
￿
(Z
n+1,j
−Z
n,j
)
2
| G
n
￿
≤ d
2
β
2
￿
n
1
n
2
(
n
S
n
)
2
< ∞ a.s..
Moreover,
E
￿
Z
n+1,j
−Z
n,j
| G
n
￿
= Z
n,j
E
￿
A
n+1,j
S
n
+A
n+1,j
| G
n
￿
− Z
n,j
d
￿
i=1
Z
n,i
E
￿
A
n+1,i
S
n
+A
n+1,i
| G
n
￿
= −Z
n,j
E
￿ A
2
n+1,j
S
n
(S
n
+A
n+1,j
)
| G
n
￿
+ Z
n,j
d
￿
i=1
Z
n,i
E
￿ A
2
n+1,i
S
n
(S
n
+A
n+1,i
)
| G
n
￿
+
+Z
n,j
EA
n+1,j
S
n
− Z
n,j
d
￿
i=1
Z
n,i
EA
n+1,i
S
n
a.s.,and
EA
n+1,j

d
￿
i=1
Z
n,i
EA
n+1,i
= EA
n+1,1
d
￿
i=d
0
+1
Z
n,i

d
￿
i=d
0
+1
Z
n,i
EA
n+1,i
.
Therefore,
￿
n
E
￿
Z
n+1,j
−Z
n,j
| G
n
￿
converges a.s.since
￿
￿
￿E
￿
Z
n+1,j
−Z
n,j
| G
n
￿
￿
￿
￿ ≤

2
S
2
n
+2β
￿
d
i=d
0
+1
Z
n,i
S
n
= o(n
λ−2
) a.s.for each λ ∈ (
λ
0
m
,1).
Thus,Z
n,j
a.s.
−→ Z
(j)
for some random variable Z
(j)
.To conclude the proof,we
let Y
n,i
= log
Z
n,i
Z
n,1
and prove that
￿
n
E
￿
Y
n+1,i
−Y
n,i
| G
n
￿
and
￿
n
E
￿
(Y
n+1,i
−Y
n,i
)
2
| G
n
￿
converge a.s.whenever i ≤ d
0
.
In this case,in fact,log
Z
n,i
Z
n,1
converges a.s.for each i ≤ d
0
and this implies Z
(i)
> 0
a.s.for each i ≤ d
0
.
Since Y
n+1,i
−Y
n,i
= X
n+1,i
log
￿
1 +
A
n+1,i
N
n,i
￿
−X
n+1,1
log
￿
1 +
A
n+1,1
N
n,1
￿
,then
E
￿
Y
n+1,i
−Y
n,i
| G
n
￿
= Z
n,i
E
￿
log
￿
1+
A
n+1,i
N
n,i
￿
| G
n
￿
−Z
n,1
E
￿
log
￿
1+
A
n+1,1
N
n,1
￿
| G
n
￿
a.s..
Since EA
n+1,i
= EA
n+1,1
,a second order Taylor expansion of x ￿→log(1+x) yields
￿
￿
￿
E
￿
Y
n+1,i
−Y
n,i
| G
n
￿
￿
￿
￿

β
2
S
n
￿
1
N
n,i
+
1
N
n,1
￿
a.s..
A quite similar estimate holds for E
￿
(Y
n+1,i
−Y
n,i
)
2
| G
n
￿
.Thus,it suffices to see
￿
n
1
S
n
N
n,i
< ∞ a.s.for each i ≤ d
0
.
16 PATRIZIA BERTI,IRENE CRIMALDI,LUCA PRATELLI,AND PIETRO RIGO
Define R
n,i
=
(S

n
)
u
N
n,i
where u ∈ (0,1) and i ≤ d
0
.Since (1+x)
u
≤ 1+ux for x ≥ 0,
one can estimate as
E
￿
R
n+1,i
R
n,i
−1 | G
n
￿
= E
￿
(
S

n+1
S

n
)
u
−1 | G
n
￿
−E
￿
(
S

n+1
S

n
)
u
X
n+1,i
A
n+1,i
N
n,i
+A
n+1,i
| G
n
￿
≤ uE
￿
S

n+1
−S

n
S

n
| G
n
￿
−E
￿
X
n+1,i
A
n+1,i
N
n,i

| G
n
￿
=
u
S

n
d
0
￿
p=1
Z
n,p
EA
n+1,p

Z
n,i
EA
n+1,i
N
n,i

=
EA
n+1,1
S
n
￿
u −
N
n,i
N
n,i

￿
a.s..
As in the proof of the Claim,thus,
E
￿
R
n+1,i
−R
n,i
| G
n
￿
= R
n,i
E
￿
R
n+1,i
R
n,i
−1 | G
n
￿
≤ 0 a.s.whenever N
n,i
≥ c
for a suitable constant c.Since N
n,i
a.s.
−→∞,then (R
n,i
) is eventually a non negative
G-super-martingale,so that R
n,i
converges a.s..Hence,
￿
n
1
S
n
N
n,i
=
￿
n
R
n,i
S
n
(S

n
)
u
=
￿
n
R
n,i
n
S
n
(
n
S

n
)
u
1
n
1+u
< ∞ a.s..
This concludes the proof.
￿
References
[1] Aletti G.,May C.and Secchi P.(2008) A central limit theorem,and related
results,for a two-color randomly reinforced urn,Preprint,currently available
at:ArXiv:math.PR/0811.2097v1
[2] Bay Z.D.and Hu F.(2005) Asymptotics in randomized urn models,Ann.Appl.
Probab.,15,914-940.
[3] Berti P.,Pratelli L.and Rigo P.(2004) Limit theorems for a class of identically
distributed random variables,Ann.Probab.,32,2029-2052.
[4] Berti P.,Crimaldi I.,Pratelli L.and Rigo P.(2009) Acentral limit theoremand
its applications to multicolor randomly reinforced urns,submitted,currently
available at:http://arxiv.org/abs/0904.0932
[5] Crimaldi I.,Letta G.and Pratelli L.(2007) A strong form of stable conver-
gence,Sem.de Probab.XL,LNM,1899,203-225.
[6] Hall P.and Heyde C.C.(1980) Martingale limit theory and its applications,
Academic Press.
[7] Janson S.(2004) Functional limit theorems for multitype branching processes
and generalized Polya urns,Stoch.Proc.Appl.,110,177-245.
[8] Janson S.(2005) Limit theorems for triangular urn schemes,Probab.Theo.
Rel.Fields,134,417-452.
[9] May C.and Flournoy N.(2009) Asymptotics in response-adaptive designs
generated by a two-color,randomly reinforced urn,Ann.Statist.,37,1058-
1078.
[10] Muliere P.,Paganoni A.M.and Secchi P.(2006) A randomly reinforced urn,
J.Statist.Plann.Inference,136,1853-1874.
[11] Pemantle R.and Volkov S.(1999) Vertex-reinforced random walk on Z has
finite range,Ann.Probab.,27,1368-1388.
CLT FOR MULTICOLOR URNS 17
[12] Pemantle R.(2007) A survey of random processes with reinforcement,Probab.
Surveys,4,1-79.
Patrizia Berti,Dipartimento di Matematica Pura ed Applicata ”G.Vitali”,Univer-
sita’ di Modena e Reggio-Emilia,via Campi 213/B,41100 Modena,Italy
E-mail address:patrizia.berti@unimore.it
Irene Crimaldi,Dipartimento di Matematica,Universita’ di Bologna,Piazza di Porta
San Donato 5,40126 Bologna,Italy
E-mail address:crimaldi@dm.unibo.it
Luca Pratelli,Accademia Navale,viale Italia 72,57100 Livorno,Italy
E-mail address:pratel@mail.dm.unipi.it
Pietro Rigo (corresponding author),Dipartimento di Economia Politica e Metodi
Quantitativi,Universita’ di Pavia,via S.Felice 5,27100 Pavia,Italy
E-mail address:prigo@eco.unipv.it