Distribution

Convergence in

Motivation
Outline
{ Convergence
properties
in distribution

Motivation
Part 2 :
:
Consider the estimation
problem :

Xy Xy -

y Xn
Iid U f O -

I , It { ) ,
where 0EUR is the

parameter
ѵm
.

Let's focus on the Method of moment estimator .

It 'z
die
since ETH -

So I
z = 0

yay
-

dahm =
In = a-
'

§,
✗ i. sample mean .

WLCN last lecture )

We know that by the ( ,

dam I>
E- Ex) = 0 as a-in
 i. e. farm is a consisent estimator for d.

•
How about confidence interval ?

Find a (1) and but ) ,

such that

Pf Ñmm -

alk ) c- de Finn -1b¥ ) ) =

0.95 Confidence level!

Ipl -

bk ) e Ñmm -
O e acts ) =
0.95
To find appropriate alk ) , bk ) , we need to

flan XT

n €
know the distn of =
.

From the
graph , we can see that
the dist 'n
of 8mm becomes more and
more concentrated around I.
( i. e. ѵm is consistent )
.

However to construct the confidence interval

x-p
, ,
pdf of
 we need to know the more refined information about the

around
shape of the
pdf 0
by zooming into a local
region around d.
Let's inflate the difference 1dam -0 ) by factor

ya
•

of in which gives

€☒,
,

Yn
-
-

rn Comm -01 .

atttibp
It turns out that the dish of
tends to increase
Yn converge as n .
,
f yn
normal distn
Its pdf approaches to a

NCO ,
¥) .

will
In the language of probability theory ,
we
say
disks
that Yn converges in to NCO , E) ,
 or write it as Yu d→ No , E)
Yn=rnlѵm
.

•
This means that the dish of -
O )
is very close to Nlo %) when
, n is large In > 20 ) .

or distln of Imm is very close to

Nld ,
In ) ( Emm =
Ot Fn Yn )
Therefore ,
to make

Rfb IX ) e doin -

Of alk ) ) = 0.95 ,

we
may simply choose alk ) =

blIl=Y¥n .

This forms an approximate 95% -

confidence interval
for 0 as
[ imm -

;¥rn
I
,
Ñmm+t¥rn ] .
 To make this aymptotic analysis rigorous ,
we need
notion disks
to formally define the
of convergence in
.

Part 2 :
Definition
distribution )
•
Deft convergence in

let Xi Xi
, ,
- '

ixn r.is and ✗

be a
sequence of
be a random variable Let Fxn and Fx be .

respectively the cdfs of Xn and ✗ Let CCF✗ )

, .

denote the set of all points where Fx is continuous .

We in distribution to ✗ or Fx if
say ✗ n converges ,

A Fxn a) =
Fxcx ) for all X C- CCF×)
,
A) is

→dF×
.

This is writer as Xn d→ ✗ or Xn .

•
connection with statistics : when we analyze an estimator
in ,
we first want to show that in is consistent

for 0 , i. e. ☒ n
-13 0 "
.
Then we want to
determine how close is On to 0 by finding
an increasing sequence { an :
n -4,2 ,
- -
-

} ,
St .

anton -
d) →d -1-64 .

where an is called the rate of convergence

and Fix) is called limiting distln ( cdf )
the .

e.
g. in our
motivating example ,

rnl Imm o)
d→ NCO ¥ )
-

in in
an r FIX )
This means 0mm approaches 0 with error

I ѵm -

01 of order
tan =

trn which is
,
 the order of the length of the confidence interval
Edam -

tiffin ,
0mm -1
"¥rn ]
The form of Ttx) can be used to construct the

confidence interval

Remark only
why requiring Fxncx ) → Fxlx ) for
• :

those x c- cCF× ) ?
Consider the following example :

1171 Xn = Tn ) =
1
,
Pf ✗ =
07--1 .
C both deterministic
since Xn → ✗ as a deterministic sequence ,

it makes sense to have ✗n ✗ as well .

 {
¥-0m
Fix / =
0 ✗ CO cred

7- 21>-0
curve
)

the blue
{?
✗'
Final
¥É
-

Carve )
✗ 25
As see in this
example CCF✗ ) C- pros)\{ 01
you can , .
=
.

and for all 21 C- CCFX) ,

or ✗ to ,

0 NO
Filk )
{
V-x-to.lt
e- Fxnck ) = =

> is

2 ✗> 0

However ,
¥ ,
Fxn 103--0=1 2 =
7×10)

This exclude
explains why we want to
points not
 in CCFX ) in the definition of convergence in distribution .

Let Xi Xn be iid from 410,0 ) 0 is

e-
g.
-

y
-

, .

the parameter with value 0>0 .

We know that the MLE fer o is

ѵhE Yn
= = Max ( Xi Xz
, ,
- .

;Xn ) .

We proved in last lecture that Rane is a assistant

estimator for d.

Let's consider the random variable

Zn =
n .

( d- Ya)

We know that pdf of Yn is

" -1
fyn 4) =
nfxlii ) [ F×M]
 =
night o c
y co
p-n
.

So Fznlt ) =
1171 Zn Et ) =p ( n IO -
Yn ) et )

=
Alyn all -

th )
o

go.in MINI
in
dy -

Fenton
fyncy )
=
I -

CI -

they
Use the calculus formula that
"
"

to Clt F) = e for all a HR .

Heh
"
We obtain Fzn A) = I -

C l -

)
 "0
I -

e- [ 1- ex )
Cx= -
to )
I
cdf of Exp ( t )
Zn →d Exp ( t ) ( exponential dish with rate

parameter to ,

toe 10.x )
✗ -

> 0

or nlÑme -

O ) →d -

Explte ) .

an ¥
Part } :

Properties
Xn→dX
.

I) .
Theorem 5.2.1 : If Xn ✗ ,
then .
i. e.
convergence in
pwbalibity is stronger than
in distribution
convergence .

ii ) Theorem 5.2.2 :
Zf Xn b ,
b is a anstant ,
then Xn -13 b.
i ) is
i. e. the converse of property partly true :

when the limit X is constant ( random

)
only a non -

iii ) Theorem 5.2.3 -

If Xn -1 ✗ and Yn -40 ,

Xntyn →d
.

then ✗
In general if Yn -94 then Xntyn may ,

not →d ✗ + Y
.
[ different from convergence is
prob )
.

Yn
.

e-
g. ✗n =
-

n N°11 )
→d Yn -9
then Xn Nun )
Hz
,
NIEL )
but Xn 1- Yn =D 22 = Not 41 .
iv ) Theorem 5.2.4 ( continuous
mapping theorem )

If Xn ✗ and g is a continuous function ,

-4
then glxn ) gcx)
v ) . Theorem 5.2.5 ( Slutsky's theorem )

Lf Xn d→ ✗ Ah a
,
→ constants
Bn -9
.

b →

then Ant Bnxn →d at BX .

[ generalizes property iii)

vi ) Theorem :
If Mxn It ) → Mxlt ) , Itch

mgtf of Xn Fgf of ✗
-9
,

then Xu ✗ .
 Question How show the distn
to convergence in
• :

of a
sequence of r-v.is in statistics
problem .

Next lecture !