0% found this document useful (0 votes)
9 views13 pages

Homework 2

The document discusses the decomposition of bias and variance in the context of estimating a function using kernel methods. It outlines the use of Taylor expansion to derive bias and variance expressions, and emphasizes the importance of integrating by parts in the calculations. Additionally, it addresses the conditions under which certain kernels can be used and the implications of these conditions on the properties of the estimators.

Uploaded by

zwu363
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
9 views13 pages

Homework 2

The document discusses the decomposition of bias and variance in the context of estimating a function using kernel methods. It outlines the use of Taylor expansion to derive bias and variance expressions, and emphasizes the importance of integrating by parts in the calculations. Additionally, it addresses the conditions under which certain kernels can be used and the implications of these conditions on the properties of the estimators.

Uploaded by

zwu363
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 13

Decompose USE into bins and variance

MSE l f ,
f) :
ET If (X0) -
f'(x) ( 2
] = LET Y'CXOC ] -

f(x) 5 + var >


FCXOLC
First , find the of f'nbias (x0)

2 f(x)/8X
F'(x) = Wor t . No Not Xi

i IiE Xi
10
-

=
, K'(

#[f' (xO] =
ii ZE , Elk') *** >]

= Elk ( * X017 ind

= Sk( ** > f(x)dx


=
i SkCusf(un Xo) du + Mo Y (x . -

X0)

Here , we
may not use the mean value theorem to
replace f(uh + Xo) Since

the integration part contains K(u) instead of K(u).

f(uh+) - orderot ,
B
Instead ,
use taylor's expansion to expand .
p

f f(8-1(xo \uhyB-
fi tub)
+

f (uh + xo) =
f(x) +
(an) +
nee
(wh +.

(B- 1) :
contain B-3 terms -
remainder term

where Xo += Uh [xo ,
Xo + Uh]
Plug EIf'(x01]
this into

#[f(xoS] in Sk (n) =
fixo) du + in 9 k(n) fixo) wh du

-
[
in Sicus faces chdu +.+
(
Skcu) frotun cubs da

Contain B-3 terms

in /k' (n) du i f(x)k(u)


The first term is f(xo) = =
0 by 11

second calculated
The term can be
by integration by part

Sk'cn)udu =
uk(u) -

Sklu) do W= W do = dn v k(a) du=


=
k'(r) du

=
UK(u) -
1


f'(x) SK (w)ndn = -

f'(xo)Uk(u) +
f' (x) =
f'(x0) by C

This term cancels out with -f'(x) in bias calculation

The B-3 calculated


remaining terms can be
by integration by part.

Let at [2 , B-2] , GER.

& k'lasnadu = n9km)- aSkcusuda = nakcu) by 22

i Skicusfaxes Cancadu = " SKCuda = 0


by 2

Thus ,
every other terms cancel out except the remainder term in the bias

IEI f(xOC] -
f'(x)

I Skcu) fettu (B d
1We
=

o + tuh) -
f(x)
n
=

<Bhs : Sk(u) fiB (0) \nhB"du

·
Skcus [f- (o +
enh) -

fB-P(xos] Lunc
+
du

"
=
<is SkcnsfP(oteah) f8-P(e) - Cut do
& i als : J1k)) IfBr" (xo + EUS -
-18 (x) & \nhs9-du
= ha : SIkcus) < Inh ChP"du by Holder's
property
LI
bo"supu /K Cas /S us de
(B-) !

=
Che where c = supulk'cus/Sudde so
Next , find the variance of Fr(XO
van)f(x) = var (wh Z k'( *** 1)
= whe var (1 ( *** I ind
= nE(k ( *** )(
= n Sk( *** ) f(x.) dx
.

has SK (n) f(ub + x) do


* Xo
-

=
n= h
I

· supuK (n) Sf(ub + X du

= G was wherecn =
suyuk Las If (un +
Xo) do

We know that

(f f) C: rB Ch
supf USE ,
=
+
-

2
haB wishs
+ 1

Choose h so that ,
then hi n

-18-2)/
RB+
3/28 +
Thus supfMSE (f f) C C wil
= +
, .

- 2(B 1)/(B + 1)
-

= (C , + G) n

-
(B x/kB +
1)
-

0(n
=
① G holds
tricially by the indicator function .

& We know that 90j] is a


complete orthonormal basis on L(E-1 13) , ee
i

Si Uj (a)
&j(u)du
= 1

Since [rj]EL (Ey , 1) ,


then Oj(m) =
[Andlu) where Sass -.. 3 is

square-summable sequence infinite sequence that the of


.

a i e . an sum
,

each elements
squared converge .
For i t 90 , 1. . .
-

B c)
-

↓ i km) du =
J Ijeojcos0j(u) 1914 *
1] de by def of Los
=
Si n jajlos UJ Lus d

=
Injjlos Ep Ulus da

=
I 10) I a S, wicklunde by dominate
convergence theorem

:
=

o by or the normality
9 kusdu
:

When := 0 ,
= 0 =

02
When i so , gui kln)dn = = 0

③ Since K is a polynomial and the domain is on the compact set [ I


-
. 17 .

as' also holds


c) is not pdf -

There is no Kernel
satisfying 4 , Cr, C and is also pdf .

Prove by contradiction .

Assume I
following C .. C, G is pdf
By Lemma 2 2
. . 5 , K( ) > .
0

By C2 , Suklu) du =

uk(n) 0
=
a. .
e

k(u) = 0

Then Siku) duzo this contradicts with C.


First , we want to find the
expression for
the bias and variance .

Here , we consider the case BC2 , then we need to


apply the taylor expansion
bias = ETFn (o)] -

for exe

=
S If(xauh) -

fLxo3] klms de

f" f(B-(xo subs"- f(x1] Klasdn


fil
(X0) + = wb)
=
/[f(x) +
(ab) +
nee
2 ! (wh +.

(B- 1) :
contain B-3 terms -
remainder term

where Xo += Uh [xo ,
Xo + Uh]

=
Sf'(o)whk(n) + & (h) k(a) +

=
JIjf(o) Lub Klus/51 dn +
&
fig-(o +
Enhsh"klws da.
(B - 13 !

= Jac fis cub Klas/jhdeagus18't"Gottuns-f'd" KLusdm (0) .

=
I f(x295 3
+hi gu gcotuh-fookLud
is even
, . .

Next evaluate the second term in the bias .

B gu , gottuk-fokLud . .

% gue glcotuk-fokLud a
>has "
-

Jiah) Kasdn by Holder's condition

-h guSt
kn Enh = uh
Then ,
we can write Bias asf" (0) + 0 (h)
Next , we find the format of the variance
+
Var (In(t) = Vardns[k ( 2**
=
nh var(k( *** () ind

= is [Elk[xix1] -
EEKLY 13]
= ↓ ( f()Skuldn O(1)-Ek( *** <]) plug dias
+
in

- (
=

f(x)(k(u)dn O() [f(x0) -"(x) 0(4452 +


-
+
+

= (In f(x)
·

f(x) + O(k)

= nh f(x) / +
0(t) f(x)S kcustu =I

suppose hn = 15

By Lindberg -
Feller theorem ,

↑ nn (X0) -
Elfhm(x0)] In n ) - Ek(
Var
[frn (XO)J' La [Var(Xi n))
-
Xo

= Ik(Xi]x) -
ETk(x1]
& [ var (E , >
= N(0 1).

Since Var(fucos) = win f(x)/ + Oct .


Then whWar(fn(x) >
-

f(x)/ as new .

By Slutsky's theorem
EnexoS El Enxos]
(fn(x0) ElEnkos]) Var(Enixos)]'k
-

dnd -
=
Inh
Var
(fn(0)1
=> N(0 , f(x) (2)

Since bias = f"(x) +


Olh) ,
taking h = vills,
- 2/5
n

bias = 6 f" (x0) +


0 (1215s

By Slutsky's theorem ,

dnb [f(0) -
f(x0)] Es N2f"(xo) /6 , f(x) ((

This distribution is not mean-zero .

From the previous part , we know that

↓ n4 (fn(x0) E[Enko(T)
-
N20 f(x) (2) = ,

Sinc bias = f"(x) Olh) by Slutcky's


+
, Lemma

duh (fn(x0) -
f(x)) => N10, f(x) (2)
By the bing expression
(Eanxo)) (fnkol
Fankos EnKos
-
Bias -
Bias
# y =

-
=

(h(z -
f "(x0) + o(h)

= bf"(x0) + O(b))

= Bias (fn(x)) + 0 (4)

By rearranging ,
we
get Bias (fn(x) 5 (fan()- fnx))
-
=
OCn

Next , evaluate the variance

Var (fu(x0) 5(frn(x0) fn(xo) (


- -

: var > Fr(x)) + ↑ var(Fan(x)) - Cor(Fn(x0) ,


fan(0)
/ wh
=

f(x)k + inf(x)k -

&Cor(fne) · fan(e)) + O 1)
= ni fixol -
En Cor(fuko) · Frn(x)( + 0 (t)

Take a look at the covariance term.

Cor(fn(x) · Frn(xo) ) = [Kn(X 1


-

Xo) Kan (x -
x0)] + O(1

= Skn(x . -0) Kah(x , -xo) f(x1)dx .


+
0(1)
=
zinSkla)k(2) f(xo + Uh) da + Pla
= 2 f(xosSkLas Kluk) da +OL

= an +
P(
1

Var (fn(x0) -
51 fan (x0) -

Eu(x)
=
isin f(x) -
- In f(x) + 0 (i)
= win f(x) + 0 (t)

By Lindeberg -
Feller CLT and Slutsky's Lemma ,

dub [fn(x - 5 (fan (10)


-

Enke)] = N20 , f(x)


Before need
proceeding , we to preve the
following lemma .

Chemma) 25 X, Y are iid , then var(Y) = El (X-45]

# :
I E[(x -
45] : El Et (x -
%5143] law of total expectation

= E[E[X -
2xy + y 1 % 3]

= I El EIX] -

2YE(X(Y) +
Y]
= ELEIY]-2YE[Y] + /23 by id
= I Elvarly) + (EYS-2YETY) +
Y 2]
= IE[ var() +
(EY -
Y( ]
*

= var (4) I

Denote : (X ., X Xi-1 Xits Xa)


Xi
of
:
·

and Xi' is the


ghost sample
-

Xi
-
=
-

,
,

This construction of %: gives (Y: 1X) IXj :


j+i·
Then ,
we can calculate

var (f(x) (Xj :


j + i) = ET (f(X) -

f(Y)(Xj j + i] :
by independence and the lemma

EEEEIIX-Y /5 1Xj : :
jti] by Lipchitz property
~
[E[/X : -
X: 'R] X and % :
Only have one
different element

= oar (Xi) by lemma


varif(x)) =
I :E
E[(f(x) ETf(x) (Xj jti])) by Efron-Steve inequality
- :

=
IE
El El(f(x) E[f(x) (Xj jti])"(Xj j law of total expectation
- : : + :])

= Zivar (i)

before proceeding . first following Lemma


prove the

Clemma 2) For a
differentiable convex functions , .
OXY in the domain of g
.
g(y)
-

g(x))yg(x) ,
y
-
x]

Of Fix X , y in the domain of g


.

By convexity . EtE Lo , K ,

(k +(x)
g(x +
+(y -
x)) =
g(ty + =
+g(y) + ( -

t)g(x) .

definition of derivative also have


By the ,
he

g(x +
+ (y x)) -
=
g(x) + + >
Og(x) ,
y
-
x3 + O(t)

We
may
conclude that g(y) -g(x)> <
g(x) ,
yex > + "

We know that El f(x) -E(f(x))xj jfi)] : =


infEICE Elf(x)-Zig2
Let Y : =
infxietoy & IX , x2)"Xi-1 , Xis Lit --
Xr)

Since is differentiable , and thus it is continuous .

Since the domain is a compact set , f attains its minimum .

Let Wi =
min
argmitous & (X ,
X ·
Xin Xi , Xit , ... Xn)

By continuity of f , Y := infxEtOBlQf(X ., X-Xiv . Xis Xit ...


Xm) .

Defince Fas r(X ., X


... Xin , Xit ...
Xn)

For x :E to 1]eR .
, fLXX-Xiv is XitssXn) is fil-many
Then Yi is also Fil -
meas
We know &Wi <
w] = Socw3n[f(X ,
:
Xin , W, xiti .... Xn >
Yiz ,
thus W: is also

↓is -
measurable·

By Lemma

# f(x) -
El f(x) /X :
jti]] =
EEfWs-Yi]
=
EEf(x) f(X -

,
...
Xin , Wis Xinc"--Xn)]

>
-
EI (x))(x: -
Wil]
·
El(2
By E from -Stein inequality,
var (f(x))
= [ Et f(x) -

E(f(x) (j :
jei]

= In EI((x()
=
E[l0f(x)12]
Let f(x) = Zi Xi

Then for X , xEIR,

|f *
. (x) -

f(x))) =
1 [(X :
-

Xic)
= I 1x -X: 1 by Triangle's inequality
=
/IX-XII ,

This shows that fo *


E F,

var (f * (x() =
var (ii) = : Var(x:

Let &* (X) =


c where ceR is constant.

It's clear that fat e F2

E[ fc (x)112]
*
var (f(x)) =
0 = 110

You might also like