7.Transform coding 651 - eecs.umich.edu · Ui's than it increases for the small Ui's. We will show:...

27
Tr-1 T RANSFORM C ODING orthogonal transformation T T -1 Q X U 2 U 1 V V Y 2 U 1 V 2 Q 1 Q k U k Scalar Quantizers . . . V k R bits 2 R bits 1 R bits k A way to use scalar quantizers to efficiently encode dependent sources. Assume, as usual, the X 1 , X 2 , is a stationary or WSS random process. Tr-2 Encoding: Instead of quantizing each X i with a scalar quantizer 1) Parse data sequence into blocks of some length k, X 1 , X 2 , = X 1 , X 2 ... where X i = (X (i-1)k+1 ,,X ik ); encode each X i as follows. 2) Find and scalar quantize the coefficients of an orthonormal expansion of X (= X i ) with respect to some orthonormal basis t 1 ,, t k . ( t i = (t i1 ,,t ik ) t ) a) Find coefficients U 1 ,...,U k such that X = U 1 t 1 + ... + U k t k b) Scalar quantize each U i with a quantizer that is optimized for it. The quantized values are V 1 = Q 1 (U 1 ), , V k = Q k (U k ) The encoder output is B = e 1 (U 1 ), , e k (U k ) Decoding: The decoder receives e 1 (U 1 ), , e k (U k ) and produces Y = V 1 t 1 + ... + V k t k

Transcript of 7.Transform coding 651 - eecs.umich.edu · Ui's than it increases for the small Ui's. We will show:...

Tr-1

TR

AN

SFO

RM

CO

DIN

G

orth

ogon

altr

ansf

orm

atio

n T

T-1

QX

U2

U

1V

VY

2

U1

V2

Q1

Qk

Uk

Sca

lar

Qua

ntiz

ers

.. .V

k

R

bits

2

R

bits

1

R

bits

k

A w

ay t

o us

e sc

alar

qua

ntiz

ers

to e

ffici

ently

enc

ode

depe

nden

t so

urce

s.

Ass

ume,

as

usua

l, th

e X

1, X

2, …

is

a s

tatio

nary

or

WS

S r

ando

mp

roce

ss.

Tr-2

En

cod

ing

: I

nste

ad o

f qu

antiz

ing

each

X

i w

ith a

sca

lar

quan

tizer

1) P

arse

dat

a se

quen

ce in

to b

lock

s of

som

e le

ngth

k,

X1,

X2,

… =

X1,

X2

...

whe

re

Xi

=

(X(i-

1)k+

1,…

,Xik

);

enco

de e

ach

Xi

as f

ollo

ws.

2) F

ind

and

scal

ar q

uant

ize

the

coef

ficie

nts

of a

n or

thon

orm

al e

xpan

sion

of

X

(= X

i) w

ith r

espe

ct t

o so

me

orth

onor

mal

bas

is

t 1,…

,t k.

(t

i =(t

i1,…

,t ik)

t )

a) F

ind

coef

ficie

nts

U1,

...,U

k s

uch

that

X =

U1t

1 +

... +

Uk

t k

b) S

cala

r qu

antiz

e ea

ch

Ui

with

a q

uant

izer

tha

t is

opt

imiz

ed f

or it

.

The

qua

ntiz

ed v

alue

s ar

e

V1

= Q

1(U

1), …

, V

k =

Qk(

Uk)

The

enc

oder

out

put

is

B =

e1(

U1)

, …, e

k(U

k)

Dec

od

ing

: T

he d

ecod

er r

ecei

ves

e1(

U1)

, …

, e k

(Uk)

an

d pr

oduc

es

Y

=

V1t

1 +

... +

Vk

t k

Tr-3

Key

Id

eas:

A:

Cho

ose

t1,

…,t k

so

tha

t U

1,...

,Uk

are

unc

orre

late

d.

(If

they

wer

e co

rrel

ated

, th

en d

ecor

rela

ting

wou

ld s

eem

to

do b

ette

r.)

B:

Cho

ose

t1,

…,t k

so

tha

t re

lativ

ely

few

U

i's

are

larg

e.

(In

this

cas

e,th

e tr

ansf

orm

is s

aid

to d

o "e

nerg

y co

mpa

ctio

n.)

T

his

mea

ns t

hat

the

sour

ce v

ecto

r X

is

des

crib

ed p

retty

wel

l with

just

a f

ew o

f th

eba

sis

func

tions

.

Use

hig

her

rate

sca

lar

quan

tizer

s on

the

U

i's

that

are

larg

er o

n th

eav

erag

e, t

han

on t

he

Ui's

th

at a

re s

mal

ler

on t

he a

vera

ge.

It re

mai

ns t

o be

see

n th

at d

isto

rtio

n is

red

uced

mor

e fo

r th

e la

rge

Ui's

th

an it

incr

ease

s fo

r th

e sm

all U

i's.

We

will

sho

w:

A

⇔ B

.

Tr-4

Co

mp

lexi

ty:

sto

rag

eo

ps/

sam

ple

Enc

odin

gk2 +

∑ i=1k M

i n

um

be

rs 2

k +

1 k ∑ i=

1k lo

g 2 M

i = 2

k+R

op

s/sa

mpl

e

Dec

odin

gk2

+ ∑ i=

1k M

i nu

mbe

rs

2k

op

s/sa

mp

le

(

tran

sfor

m +

qua

ntiz

ers)

(t

rans

form

+ q

uant

izer

s)

(Ass

umin

g tr

ansf

orm

is d

one

by m

atrix

mul

tiplic

atio

n in

k2

op

s, a

ndqu

antiz

er p

artit

ioni

ng d

one

in

log 2

Mi

op's

“bi

nary

sea

rch”

Tr-5

MA

TH

EM

AT

ICA

L D

ET

AIL

S (F

RO

M L

INE

AR

AL

GE

BR

A)

Def

init

ion

s:

•Rk

=

set

of r

eal-v

alue

d k-

dim

ensi

onal

vec

tors

of

the

form

x

= (

x 1,…

,xk)

.

We

ofte

n th

ink

of v

ecto

rs a

s co

lum

n ve

ctor

s.

•In

ner

or d

ot p

rodu

ct:

(x,

y)

= (

y, x

) =

xt y

= y

t x =

∑ i=1k x

i yi

•Le

ngth

of

x:

||x||

= √

∑ i=1k x

2 i =

(x,x

) =

√x

t x

•F

act:

||x

+ y|

|2 =

(x+

y)t (

x+y)

= |

|x||2 +

2(x

, y)

+ ||y

||2

•O

rtho

gona

lity:

x a

nd y

ar

e or

thog

onal

if

xty

= 0

.

•Li

near

ind

epen

denc

e:

x 1,…

,xm

ar

e lin

early

inde

pend

ent

if th

ere

exis

t no

coef

ficie

nts

u1,

…,u

m s

uch

that

u1

x 1 +

… +

uk x k

= 0

, ex

cept

u1=

u 2=

…=

uk=

0.

•B

asis

: A

bas

is f

or R

k is

a li

near

ly in

depe

nden

t se

t of

vec

tors

{ t

1,…

,t k}

in R

k th

at s

pan

Rk .

T

hat

is,

ever

y ve

ctor

x

∈ R

k is

a li

near

com

bina

tion

of t

he m

em-

bers

of

the

set;

i.e.

the

re a

re c

oeff’

s u

1,…

,um

su

ch th

at x

= u

1t1

+ …

+ u

k t k

•O

rtho

norm

al b

asis

for

R

k :

{t1,

…,t k

} s

uch

that

tha

t m

embe

rs o

f th

e ba

sis

are:

orth

ogon

al:

tt i tj =

0 w

hen

i ≠

j,

and

nor

mal

ized

: ||t

i|| =

1 f

or e

ach

i Tr-6

Fa

cts

:

•E

very

bas

is f

or R

k h

as e

xact

ly

k e

lem

ents

.

•T

here

exi

sts

an o

rtho

norm

al b

asis

for

R

k fo

r ev

ery

k,

e.g.

{(1

,0,…

,0),

(0,

1,0,

…,0

), …

, (0,

…,0

,1)}

•G

iven

a b

asis

{t 1

,…,t k

} a

nd a

vec

tor

x

ther

e is

exi

sts

one

and

only

one

set

of c

oeffi

cien

ts

u 1,…

,uk

suc

h th

at

x =

u1t

1 +

… u

k t k

•G

iven

an

orth

onor

mal

bas

is

t 1,…

,t k

and

a ve

ctor

x,

th

e un

ique

coe

ffici

ents

u 1,…

,uk

suc

h th

at x

= u

1t1

+ …

+ u

k t k

are

u i =

tt i x,

i = 1

,...,k

Pro

of:

If x

= u

1t1

+ ..

. + u

k t k

, th

en

tt i x =

tt i (

u 1 t 1

+...

+ u

k t k

) =

u1 tt i t

1 +

... +

uk

tt i tk

=

ui

sin

ce t

i's o

rtho

norm

al

•If

u =

(u 1

,…,u

k)

is t

he v

ecto

r of

coe

ffici

ents

of

the

expa

nsio

n of

x

with

resp

ect

to o

rtho

norm

al b

asis

{t

1,…

,t k},

th

en

||u||2

= ||x

||2

(T

his

is a

ver

sion

of

Par

seva

l’s T

heor

em)

Pro

of:

||x||2

= ||

u 1 t 1

+…+u

k t k

||2 =

(u 1

t 1+…

+uk t k

)t (u1 t 1

+…

+u k

t k)

= u

2 1 +

… +

u2 k

beca

use

tt i t

j = 1

, i=

j0

, i ≠

j

= |

|u||2

Tr-7

Rec

all

Ste

p 2

a o

f tr

ansf

orm

en

cod

ing

2) G

iven

an

orth

onor

mal

bas

is

t 1,…

,t k.

(t i

= (

t i1,…

,t ik)

t )

(a)

Fin

d co

effic

ient

s U

1,...

,Uk

suc

h th

at X

= U

1t1

+ ..

. + U

k t k

.

By

a pr

evio

us f

act,

Ui

= tt i X

=

∑ j=1k t i

j Xj ,

i =

1,..

.,k

Equ

ival

ently

, w

e ca

n co

mpu

te t

he c

oeffi

cien

ts v

ia m

atrix

mul

tiplic

atio

n

U =

U

1… U

k =

t 1

1 t 1

2 …

t 1k

t 21

t 22

… t 2

k... t k

1 t k

2 …

t kk

X

1X

2 ... Xk

= T

X

whe

re

T

is t

he

k×k

mat

rix

T =

--tt 1-

-

--tt 2-

-... --

tt k--

=

t 1

1 t 1

2 …

t 1k

t 21

t 22

… t 2

k... t k

1 t k

2 …

t kk

Mo

re D

efin

itio

ns:

•In

thi

s co

ntex

t th

e m

atrix

T

is

cal

led

a tr

ansf

orm

, an

d T

X

is c

alle

d a

tran

sfor

mat

ion

of

X.

U

is t

he v

ecto

r of

tra

nsfo

rm c

oeffi

cien

ts.

•A

mat

rix w

ith o

rtho

norm

al r

ows,

suc

h as

the

one

s w

e ar

e in

tere

sted

in,

is c

alle

d "o

rtho

gona

l".

("O

thon

orm

al"

wou

ld b

e a

bette

r na

me,

but

"ort

hogo

nal"

is

wha

t th

e m

athe

mat

icia

ns u

se.)

Tr-8

Fac

ts:

The

fol

low

ing

are

equi

vale

nt s

tate

men

ts

(a)

T

is a

n or

thog

onal

mat

rix,

(b)

Row

s of

T

ar

e or

thor

mal

(c)

Col

umns

of

T

are

orth

onor

mal

, (

d)

||Tx|

| = |

|x||

for

all x

(e)

T-1

= T

t , (

f) T

-1

is o

rtho

gona

l.

Pro

ofs

:(a

) ⇔

(b)

: B

y de

finiti

on o

f an

ort

hogo

nal m

atrix

(b)

⇔ (

e):

Not

ice

that

T T

t =

--tt 1-

-

--tt 2-

-... --

tt k--

| t 1 |

| t 2 | - - - | t k |

=

1

0 ...

.. 0

0 1

0 ..

0 0

.....

0 1

= I

= id

entit

ym

atr

ix

if an

d on

ly if

the

row

s of

T

are

ort

hono

rmal

. T

hus

Tt

is th

e in

vers

e of

T

iff

its

row

s ar

e or

thon

orm

al.

(N

ote:

by

defin

ition

of

the

inve

rse

T-1

T =

I =

T T

-1 .)

(c)

⇔ (

e):

By

a si

mila

r ar

gum

ent

Tt T

= I

iff

th

e co

lum

ns o

f T

ar

e or

thon

orm

al.

(b)

⇒ (

d):

Pro

ved

earli

er.

(e)

⇒ (d

):

(e)

⇒ |

|Tx|

|2 =

(T

x,T

x) =

(T

x)t T

x =

xt T

t Tx

= x

t x =

||x

||2

(d)

⇒ (

e):

Ass

umin

g ||

Tx|

| =

||x||

for

all x

, w

e ha

ve

0 =

||Tx|

|2 -

||x||2

= (

Tx,

Tx)

- (

x,x)

= (

Tx)

t Tx

- x

t x =

xt T

t Tx

- xt I

x =

xt (T

t T-I

) x

sinc

e th

is h

olds

for

all

x

it m

ust

be t

hat

Tt T

=I,

i.e.

Tt =

T-1

.

(e)

⇔ (

f):

E

lem

enta

ry.

Tr-9

Fac

t:

For

an

orth

ogon

al t

rans

form

atio

n ||

Tx-

Ty|

| = ||

x-y

||

Pro

of:

||T

x-T

y|| =

||T

(x-y

)||

= |

|x-y

||

by (

d) o

f the

pre

viou

s fa

ct.

Bec

ause

an

orth

ogon

al t

rans

form

atio

n pr

eser

ves

the

leng

th o

f ve

ctor

s an

dal

so t

he d

ista

nce

betw

een

vect

ors,

we

may

vie

w i

t as

ess

entia

lly a

rota

tion.

(T

here

mig

ht a

lso

be a

xis

flips

.)

Exa

mpl

es o

f or

thog

onal

tra

nsfo

rmat

ions

:

DF

Tdi

scre

te F

ourie

r tr

ansf

orm

DC

Tdi

scre

te c

osin

e tr

ansf

orm

DW

Tdi

scre

te w

avel

et t

rans

form

WH

TW

alsh

-Had

amar

d

Tw

o-di

men

sion

al v

ersi

ons

of t

he a

bove

KLT

K

arhu

nen-

Loev

e tr

ansf

orm

Tr-

10

EX

AM

PLE

: D

CT

BA

SIS

VE

CT

OR

S (O

NE

-DIM

EN

SIO

NA

L)

k =

8k

= 16

Tr-

11

Exa

mp

le:

(R

eal)

FF

T B

asis

Vec

tors

(o

ne-

dim

ensi

on

al)

k =

8k

= 16

Tr-

12

PE

RFO

RM

AN

CE

OF

TR

AN

SFO

RM

CO

DIN

G

orth

ogon

altr

ansf

orm

atio

n T

T-1

QX

U2

U

1V

VY

2

U1

V2

Q1

Qk

Uk

Sca

lar

Qua

ntiz

ers

.. .V

k

R

bits

2

R

bits

1

R

bits

k

Rat

e:

R =

1 k ∑ i=

1k R

(Qi)

=

aver

age

rate

of

the

scal

ar q

uant

izer

s Q

1,...

,Qk

=

rate

of

the

the

prod

uct

quan

tizer

Q

1×...

×Qk

Tra

nsfo

rm c

ode

is f

ixed

rat

e if

scal

ar q

uant

izer

s ar

e fix

ed-r

ate;

var

iabl

era

te if

sca

lar

quan

t'rs

are

varia

ble-

rate

.

Tr-

13

Dis

tort

ion

:

D=

1 k ∑ i=

1k E

(Xi-Y

i)2 =

1 k

E

∑ i=

1k (

Xi-Y

i)2 =

1 k E

||X

-Y||2

=

1 k E

||U

-V||2

sin

ce fo

r or

thog

onal

tran

sfor

m,

||X-Y

|| =

||U

-V||

=

1 k E

∑ i=

1k (

Ui-V

i)2

=

1 k ∑ i=1k E

(Ui-V

i)2

=

1 k ∑ i=

1k D

Ui(Q

i)

=

avg

dis

tort

ion

of t

he s

cala

r qu

antiz

ers

Q1,

...,Q

k

=

D U

1,...

,Uk(Q

1×...

×Qk)

=

di

stor

tion

of p

rodu

ct q

uant

izer

Q

1×...

×Qk

Thi

s is

why

we

rest

rict

to o

rtho

gona

l tr

ansf

orm

s.

Tr-

14

OPT

IMA

L D

ESI

GN

AN

D T

HE

OPT

A F

UN

CT

ION

•G

iven

a d

imen

sion

k

and

a r

ate

R,

we

wis

h to

opt

imiz

e k-

dim

ensi

onal

tran

sfor

m c

odin

g.

Tha

t is

, w

e w

ish

to f

ind

the

k×k

tr

ansf

orm

T

an

d th

esc

alar

qua

ntiz

ers

Q1,

...,

Qk

tha

t m

inim

ize

dist

ortio

n su

bjec

t to

the

rat

eco

nstr

aint

. W

e al

so w

ish

to f

ind

the

dist

ortio

n th

at r

esul

ts w

hen

the

code

par

amet

ers

are

optim

ized

. T

hat

is,

we

wis

h to

fin

d th

e O

PT

Afu

nctio

n

δ tr

(k,R

) =∆

le

ast

dist

ort’n

of

tran

sfor

m c

odes

with

dim

'n k

and

rat

e R

or

less

=

min T

m

in

Q1,

...,Q

k: 1 k

∑ i-1k R

(Qi)≤

R

1 k

∑ i=1k D

Ui(Q

i),

whe

re

U =

TX

=

min T

δU

,pr(R

),

whe

re

δ U,p

r(R

) =

O

PT

A f

unct

ion

of p

rodu

ct q

uant

izer

s fo

r U

=T

X

•W

e fir

st in

vest

igat

e δ

U,p

r(R

) t

he p

rodu

ct q

uant

izer

OP

TA

fun

ctio

n, a

ndw

e'll

lear

n ho

w t

o ch

oose

the

Q

i's.

If

all t

he

Ui's

ar

e id

entic

al r

ando

mva

riabl

es,

then

the

opt

imal

Q

i's

turn

out

to

be id

entic

al.

How

ever

, w

ew

ill s

ee t

hat

a "g

ood"

tra

nsfo

rm m

akes

the

U

i's

have

rat

her

diffe

rent

prob

abili

ty d

ensi

ties.

S

o fin

ding

the

Q

i's

is n

ot t

rivia

l.

•W

e w

ill t

hen

lear

n ho

w t

o ch

oose

T

to

min

imiz

e δ

U,p

r(R

).

•F

inal

ly,

we'

ll ex

plor

e th

e ef

fect

of

vary

ing

the

dim

ensi

on

k.

Tr-

15

•T

he O

PT

A f

unct

ion

of a

k-d

imen

sion

al p

rodu

ct q

uant

izer

δ U,p

r(R)

=

min

Q1,

...,Q

k: 1 k

∑ i-1k R

(Qi)

≤ R

1 k ∑ i=

1k D

Ui(Q

i),

w

here

U

= T

X

=

min

R1≥

0,...

,Rk≥

0: 1 k

∑ i-1k R

i ≤ R

min

Q1,

...,Q

k: R

(Q1)

≤ R

1,...

,R(Q

k) ≤

Rk 1 k

∑ i=1k D

Ui(Q

i)

=

min

R1≥

0,...

,Rk≥

0: 1 k

∑ i-1k R

i ≤ R

1 k

∑ i=1k

min

Qi:

R(Q

i) ≤

Ri D

Ui(Q

i)

=

min

R1≥

0,...

,Rk≥

0: 1 k

∑ i-1k R

i ≤ R

1 k

∑ i=1k δ

Ui,s

q(R

i)

•W

e ca

n no

w s

ee t

hat

to f

ind

the

OP

TA

fun

ctio

n, w

e m

ust

find

R1,

...,

Rk

that

ave

rage

to

at m

ost

R

and

that

min

imiz

e th

e av

erag

e of

the

indi

vidu

al s

cala

r qu

antiz

er O

PT

A f

unct

ions

of

U1,

...,U

k.

•T

he c

hoic

e of

R

1,...

,Rk

is c

alle

d a

bit

or r

ate

allo

catio

n,

i.e.

an a

lloca

-tio

n of

kR

bi

ts a

mon

g th

e in

divi

dual

sca

lar

quan

tizer

s.

If so

me

are

give

n ra

te la

rger

tha

n R

, t

hen

othe

rs m

ust

be g

iven

rat

e le

ss t

han

R.

Tr-

16

Rec

all o

ne o

f th

e or

igin

al id

eas

of t

rans

form

cod

ing:

so

me

coef

ficie

nts

will

be

smal

ler

on t

he a

vera

ge t

han

othe

rs,

and

thes

e w

ill b

e en

code

dat

low

er r

ates

, i.e

. al

loca

ted

few

er b

its,

than

tho

se t

hat

are

larg

er o

nth

e av

erag

e.

If th

ere

is t

o be

a n

et g

ain,

the

n it

will

dep

end

on a

succ

essf

ul b

it al

loca

tion.

•W

e w

ish

to f

ind

the

rate

allo

catio

n th

at m

inim

izes

:

δ U,p

r(R)

=

min

R1>

0,...

,Rk>

0: 1 k

∑ i-1k R

i ≤ R

1 k

∑ i=1k δ

Ui,s

q(R

i)

•F

or f

ixed

-rat

e sc

alar

qua

ntiz

atio

n, e

ach

δ U

i,sq(

Ri)

has

a s

tairc

ase

form

.F

indi

ng

δ U,p

r(R)

is a

n in

tege

r pr

ogra

mm

ing

prob

lem

. T

here

are

var

ious

itera

tive

algo

rithm

s.

•F

or v

aria

ble-

rate

sca

lar

quan

tizat

ion,

eac

h δ

Ui,s

q(R

i) is

con

tinuo

us.

We

can

find

δ U,p

r(R)

usi

ng v

aria

tiona

l, e.

g. d

eriv

ativ

e, m

etho

ds.

•F

or f

ixed

-rat

e sc

alar

qua

ntiz

atio

n, w

hen

R

is la

rge,

mos

t R

i's

are

larg

ean

d δ

Ui,s

q(R

i) c

an b

e ap

prox

imat

ed a

s co

ntin

uous

fun

ctio

n w

ith Z

ador

'sfo

rmul

a.

•F

or n

ow le

t us

ass

ume

δi(R

) is

con

tinuo

us.

Thi

s is

the

cas

e fo

rva

riabl

e-le

ngth

cod

ing

and

appr

oxim

atel

y th

e ca

se f

or f

ixed

-rat

e co

ding

whe

n R

is

larg

e.

In t

his

case

, w

e ca

n us

e va

riatio

nal m

etho

ds.

Tr-

17

A U

SEFU

L O

PTIM

IZA

TIO

N L

EM

MA

Lem

ma:

Let

g 1(R

), …

, g k

(R)

be

con

tinuo

us,

posi

tive-

valu

ed,

stric

tlyde

crea

sing

, co

ntin

uous

fun

ctio

ns d

efin

ed o

n [

0,∞

).

If R

1,…

,Rk

min

imiz

e ∑ i=1k g

i(Ri)

sub

ject

to

∑ i=1k R

i ≤ R

, R

i≥0,

i =

1,..

.,k.

then

g 'i(R

i) =

g'j(R

j) fo

r al

l i,j

s.t.

Ri >

0, R

j > 0

and

|g'i(R

i)| ≤

|g'j(R

j)| f

or a

ll i,j

s.t.

Ri =

0, R

j > 0

Tha

t is

, fo

r th

e op

timum

cho

ice

of

R1,

…,R

k,

the

slop

es o

f th

e g

ifu

nctio

ns a

re t

he s

ame.

Not

e:

The

der

ivat

ives

of

gi(R

) a

re n

egat

ive.

Tr-

18

Ri

Rj

R +

εi

R -ε j

|g'(R

)|ε

i

|g'(R

)|ε

j

Pro

of

of

Lem

ma:

B

y co

ntra

dict

ion.

Cas

e 1:

S

uppo

se

R1,

…,R

k m

inim

ize

Σk i=

1 g i

(Ri)

with

Σk i=

1 R

i ≤ R

, R

i ≥ 0

,

and

for

som

e i,

j, R

i > 0

, Rj >

0,

g 'i(R

i) ≠

g 'j(R

j).

With

out

loss

of

gene

ralit

y su

ppos

e

g 'i(R

i) <

g'j(R

j) <

0

The

n fo

r so

me

smal

l ε,

re

plac

e

Ri

by

Ri+

ε a

nd

Rj

by R

j-ε .

Thi

s ha

s no

effe

ct o

n Σ

k i=1

Ri .

Ho

we

ver,

g i(R

i) c

hang

es t

o ap

prox

imat

ely

gi(R

i+ε)

≅ g

i(Ri)+

ε g'i(R

i).

g j(R

j) c

hang

es t

o ap

prox

imat

ely

g

j(Rj-ε

) ≅

gj(R

j) -

ε-g'j(R

j).

Sin

ce g

'i(Ri)

< g

'j(Rj),

th

is r

educ

es

Σk i=

1 g i

(Ri),

bec

ause

g i(R

i+ε)

+ g

j(Rj-ε

) ≅

gi(R

i)+g j

(Rj)+

ε( g

'i(Ri)-

g 'j(R

j)) <

gi(R

i)+g j

(Rj).

(Des

pite

the

"≅"

, th

e ov

eral

l ine

qual

ity w

ill b

e st

rict

if ε

is

suf

ficie

ntly

sm

all.)

Thi

s co

ntra

dict

s th

e or

igin

al a

ssum

ptio

n th

at t

he

R1,

..,R

k m

inim

ize

Σk i=

1 g i

(Ri)

.T

here

fore

, it

is n

ot p

ossi

ble

that

R

i > 0

, R

j > 0

, g 'i

(Ri)

≠ g

'j(Rj).

Tr-

19

Ri

Rj

R +

εi

R -

εj

|g'(R

)|ε

i

|g'(R

)|ε

j

Cas

e 2:

S

uppo

se

R1,

…,R

k m

inim

ize

Σk i=

1 g i

(Ri)

with

Σk i=

1 R

i ≤ R

, R

i ≥ 0

,

and

for

som

e i,

j, R

i = 0

, R

j > 0

, an

d |g

'i(Ri)|

> |

g 'j(R

j)|.

Sin

ce t

he s

lope

s ar

ene

gativ

e,

g 'i

(Ri)|

< |g

'j(Rj)|

.

The

n fo

r so

me

smal

l ε,

re

plac

e

Ri

by

Ri+

ε a

nd

Rj

by R

j-ε .

Thi

s ha

s no

effe

ct o

n Σ

k i=1

Ri .

Ho

we

ver,

g i(R

i) c

hang

es t

o ap

prox

imat

ely

g i(R

i+ε)

≅ g

i(Ri)+

ε g'i(R

i).

g j(R

j) c

hang

es t

o ap

prox

imat

ely

g j(R

j-ε)

≅ g

j(Rj)

- ε-g

'j(Rj).

Sin

ce g

'i(Ri)

< g

'j(Rj),

th

is r

educ

es

Σk i=

1 g i

(Ri),

bec

ause

g i(R

i+ε)

+ g

j(Rj-ε

) ≅

gi(R

i)+g j

(Rj)+

ε( g

'i(Ri)-

g 'j(R

j)) <

gi(R

i)+g j

(Rj).

(Des

pite

the

"≅"

, th

e ov

eral

l ine

qual

ity w

ill b

e st

rict

if ε

is

suf

ficie

ntly

sm

all.)

Thi

s co

ntra

dict

s th

e or

igin

al a

ssum

ptio

n th

at t

he

R1,

..,R

k m

inim

ize

Σk i=

1 g i

(Ri)

.T

here

fore

, it

is n

ot p

ossi

ble

that

R

i = 0

, Rj >

0,

|g'i(R

i)| >

|g'j(R

j)|.

Tr-

20

TH

E O

PTA

FU

NC

TIO

N O

F K

-DIM

EN

SIO

NA

L P

RO

DU

CT

QU

AN

TIZ

ER

:

HIG

H-R

ESO

LU

TIO

N A

NA

LY

SIS

Rec

all:

δ U,p

r(R)

=

min

R1>

0,...

,Rk>

0: 1 k

∑ i-1k R

i ≤ R

1 k

∑ i=1k δ

Ui,s

q(R

i)

Ass

ume

R

is s

o la

rge

that

the

opt

imal

R

i's

are

so la

rge

that

we

may

use

the

appr

oxim

atio

ns

δ sq,U

i(Ri)

≅≅≅≅

1 12

σ2 iα i

2-2

Ri

,

whe

re

σ2 i =

var

ianc

e of

Ui

, an

d fo

r fix

ed-r

ate

codi

ng

α i =

β U

i =

1 σ2 i

∫ f1/

3U

i(u

) du

3

and

for

varia

ble-

rate

cod

ing

α i =

η U

i =

1 σ2 i 2

2h(U

i)

,

whe

re

h(U

i) =

-

∫ -∞∞ f Ui(u

) lo

g 2 f U

i(u)

du

We

now

hav

e

δ U,p

r(R)

=

min

R1>

0,...

,Rk>

0: 1 k

∑ i-1k R

i ≤ R

1 k

∑ i=1k 1 1

2 σ2 i

α i 2

-2R

i

Tr-

21

Ap

ply

ing

th

e O

pti

miz

atio

n L

emm

a

The

lem

ma

show

s th

at if

R

1,…

,Rk

> 0

m

inim

ize

1 k ∑ i=

1k

1 12

σ2 iα i

2-2

Ri

su

bjec

t to

1 k

∑ i=1k R

i = R

,

then

the

re is

a c

onst

ant

c

suc

h th

at f

or e

ach

i,

c =

d dR

i

1 12

σ2 iα i

2-2

Ri

=

1 12

σ2 iα i

2-2

Ri

(

-2 ln

2)

Sol

ving

the

abo

ve y

ield

s

Ri

= 1 2

log 2

σ2 iα i

- 1 2

log 2

(-6

cln

2)

.

Sin

ce t

he

Ri's

av

erag

e to

R

, e

quat

ing

the

avg.

of

the

RH

S t

erm

s t

o R

gi

ves

- 1 2

log 2

(-6c

ln 2

) =

R -

1 k ∑ j=1k 1 2 lo

g 2 σ

2 jα j

=

R -

1 2 log 2

∏ j=

1k σ

2 jα j

1/k

The

refo

re t

he o

ptim

al r

ate

allo

catio

n is

Ri

=

1 2 log 2

σ2 iα i

+ R

- 1 2 lo

g 2

∏ j=

1k σ

2 jα j

1/k =

R +

1 2 log

2 σ2 iα

i

∏ j=1k σ

2 jαj1/

k

Sub

stitu

ting

the

optim

al a

lloca

tion

show

n ab

ove

into

the

exp

ress

ion

for

dist

ortio

n yi

elds

δ U,p

r(R)

≅ 1 1

2

∏ j=

1k σ

2 jα j

1/k 2

-2R

Tr-

22

Not

ice

also

tha

t su

bstit

utin

g th

e op

timal

R

i in

to

δ sq,U

i(Ri)

≅≅≅≅ 1 1

2 σ

2 iα i

2-2

Ri

give

s

δ sq,U

i(Ri)

≅≅≅≅

1 12

∏ j=

1k σ

2 jα j

1/k 2

-2R

≅≅≅≅

δ U,p

r(R

)

Thu

s, w

ith t

he o

ptim

al b

it al

loca

tion,

eac

h U

i is

qua

ntiz

ed w

ith t

he s

ame

dist

or-

tion.

Is

thi

s w

hat

you

expe

cted

? I

s th

is w

hat

JPE

G d

oes?

If

not,

why

not

?

Sum

mar

y:

Giv

en a

k-d

imen

sion

al r

ando

m v

ecto

r U

a

nd la

rge

R,

the

optim

al r

ate

allo

catio

n is

Ri

= R

+ 1 2 l

og2

σ2 iαi

Γ U ,

( *)

the

min

imal

dis

tort

ion,

i.e

the

OP

TA

fun

ctio

n, is

δ U,p

r(R)

≅≅≅≅

1 12

Γ U 2

-2R

,

( **)

and

with

the

opt

imal

rat

e al

loca

tion

all c

oeffi

cien

ts a

re q

uant

ized

with

app

roxi

mat

ely

the

sam

e di

stor

tion,

w

here

Γ U =

∏ j=

1k σ

2 jα j

1/k

=

geom

etric

ave

rage

of

σ2 jα

j's

Tr-

23

No

tes

:

•T

he U

i's

with

larg

er

σ2 iαi

are

quan

tized

with

hig

her

rate

tha

n th

ose

with

smal

ler

σ2 iα

i. T

hose

with

σ2 i

α i >

ΓU

are

quan

tized

with

rat

e R

i > R

,an

d th

ose

with

σ2 i

α i <

ΓU

are

quan

tized

with

rat

e R

i < R

. B

ut a

ll ar

equ

antiz

ed w

ith a

ppro

xim

atel

y th

e sa

me

dist

ortio

n.

•C

oeffi

cien

t co

rrel

atio

ns d

o no

t en

ter

into

for

mul

a fo

r δ

U,p

r(R).

T

hus,

ther

e is

no

dire

ct n

eed

for

the

coef

ficie

nts

to b

e un

corr

elat

ed.

•S

uppo

se in

stea

d of

the

opt

imal

bit

allo

catio

n, w

e as

sign

equ

al n

umbe

rsof

bits

to

each

U

i. T

hat

is,

supp

ose

Ri =

R,

all

i. T

hen

D ≅

1 k

∑ i=1k δ

Ui,s

q(R

i) =

1 k

∑ i=1k δ

Ui,s

q(R

) ≅

1 k

∑ i=1k

1 12

σ2 iα i

2-2

R

=

1 12

1 k

∑ i=1k σ2 i

α i 2

-2R

Sin

ce a

n ar

ithm

etic

ave

rage

is la

rger

tha

n a

geom

etric

ave

rage

(un

less

the

term

s be

ing

aver

aged

are

iden

tical

),

1 k ∑ i=

1k σ

2 iαi

≥ Γ

U.

The

refo

re,

the

SN

R g

ain

of t

he o

ptim

al b

it al

loca

tion

over

a n

aive

all-

equa

l bit

allo

catio

n is

10 lo

g 10

1 k Σk i=

1sσ2 i

α iΓ U

Tr-

24

•If

σ2 iα i

is

the

sam

e fo

r al

l i,

the

n th

e op

timal

bit

allo

catio

n is

the

all-

equa

l bit

allo

catio

n, a

nd

1 k

Σk i=1σ

2 iα i

=

ΓU.

•W

hat

to d

o if

(*)

yie

lds

a sm

all o

r ne

gativ

e R

i fo

r so

me

i?

Thi

sha

ppen

s w

hen

som

e σ

2 iαi's

ar

e ve

ry s

mal

l.

(Typ

ical

ly,

we

need

eac

hR

i ≥~ 3

fo

r th

is h

igh-

reso

lutio

n an

alys

is t

o be

acc

urat

e.)

If

only

a s

mal

lfr

actio

n of

the

coe

ffici

ents

hav

e sm

all o

r ne

gativ

e R

i, t

hen

the

anal

ysis

will

be

fairl

y ac

cura

te.

If

not,

we

may

hav

e to

res

ort

to a

n in

tege

rpr

ogra

mm

ing

appr

oach

to

findi

ng a

goo

d bi

t al

loca

tion.

Tr-

25

TH

E O

PTA

FU

NC

TIO

N O

F T

RA

NSF

OR

M C

OD

ING

WIT

H T

RA

NSF

OR

M T

•A

pply

ing

wha

t w

e ju

st le

arne

d ab

out

prod

uct

code

s, w

e fin

d th

at f

ortr

ansf

orm

cod

ing

with

a s

peci

fic t

rans

form

T

,

-th

e op

timal

bit

allo

catio

n is

Ri

= R

+ 1 2 l

og2

σ2 iαi

Γ TX

( *

)

-th

e O

PT

A f

unct

ion

of t

rans

form

cod

ing

with

k-d

imen

sion

al t

rans

form

T

is

δ tr,

T(k

,R)

= δ

TX

,pr(R

) ≅≅≅≅

1 1

2 Γ T

X 2

-2R

( *

*)

-w

ith t

he o

ptim

al r

ate

allo

catio

n, a

ll co

effic

ient

s ar

e qu

antiz

ed w

ith

appr

oxim

atel

y th

e sa

me

dist

ortio

n,

-Γ T

X =

∏ j=

1k σ

2 jα j

1/k

=

geom

etric

ave

rage

of

σ2 jα

j's

•W

e se

e th

at a

goo

d tr

ansf

orm

is o

ne t

hat

mak

es

Γ TX =

()

∏k j=

1 σ2 j

α j1/

k

smal

l.

•C

ompa

re

( **)

to

the

hig

h-re

solu

tion

appr

oxim

atio

n to

the

OP

TA

of

dire

ct s

cala

r qu

antiz

atio

n ap

plie

d to

X

δ sq,

X(R

) ≅≅≅≅

1 12

σ2 X α

X 2

-2R

Tr-

26

The

SN

R g

ain

of t

rans

form

cod

ing

with

tra

nsfo

rm

T

over

dire

ct s

cala

rqu

antiz

atio

n of

X

is

10 lo

g 10

δ sq,

X(R

)δ T

X,p

r(R)

≅≅≅≅ 1

0 lo

g 10

σ2 Xα X

∏ j=

1k σ

2 jα j

1/k

= 1

0 lo

g 10

σ2 Xα X

Γ TX

•W

e w

on't

com

pare

to

k-di

men

sion

al V

Q u

ntil

we

optim

ize

T.

Tr-

27

•If

X

is

Gau

ssia

n, t

hen

each

U

i is

Gau

ssia

n,

and

α 1=

... =

αk

= α

X =

αG =

32.

6 fo

r F

RC

and

17.

08 fo

r V

RC

.

The

refo

re,

the

α's

ca

ncel

out

of

the

rate

allo

catio

n fo

rmul

a, w

hich

be

com

es

Ri

= R

+ 1 2 l

og2

σ2 i

Σ2 TX

wh

ere

Σ2 TX =

∏ j=

1k σ

2 j1/

k =

ge

omet

ric m

ean

of t

he c

oeffi

cien

t va

rianc

es

The

α'

s a

lso

canc

el o

ut o

f th

e th

e ga

in o

ver

scal

ar q

uant

izat

ion

form

ula,

whi

ch b

ecom

es

10 lo

g 10

σ2 X

Σ2 TX

=

10 lo

g 10

ratio

of v

aria

nce

to g

eom

etric

mea

n of

coe

f va

rian

ces

Sin

ce

σ2 X =

1 k E

||X||2

= 1 k

E||U

||2 =

1 k ∑ j=

1k σ

2 j ,

th

e ga

in is

the

rat

io o

f th

e

arith

met

ic a

vera

ge t

o th

e ge

omet

ric a

vera

ge o

f th

e co

effic

ient

var

ianc

es.

And

the

OP

TA

fun

ctio

n fo

r tr

ansf

orm

cod

ing

with

tra

nsfo

rm

T

bec

omes

δ X,tr

,T(R

) ≅

1 12

α G Σ

2 TX 2

-2R

Tr-

28

•C

oncl

usio

ns f

or t

he G

auss

ian

case

:

-A

goo

d tr

ansf

orm

is

one

that

min

imiz

es t

he g

eom

etric

ave

rage

of

the

coef

ficie

nt v

aria

nces

Σ2 TX =

∏ j=

1k σ

2 j1/

k ,

Sin

ce t

he a

rithm

etic

ave

rage

1 k

∑ j=1k σ

2 j

is u

naffe

cted

by

the

tran

sfor

m,

this

is d

one

by m

akin

g so

me

σ2 j 's

as

sm

all

as p

ossi

ble

(thi

s is

wha

t m

akes

the

aver

age

smal

l),

but

sinc

e th

e ar

ithm

etic

ave

rage

is u

ncha

nged

, s

ome

σ2 j's

mus

t al

so b

e la

rge.

In

som

e se

nse,

a g

ood

tran

sfor

m m

akes

the

σ2 j 's

as d

iffer

ent

as p

ossi

ble.

Tr-

29

•M

ore

conc

lusi

ons

for

the

Gau

ssia

n ca

se

-If

inst

ead

of t

he o

ptim

al b

it al

loca

tion,

one

use

d th

e eq

ual-b

it-al

loca

tion,

the

n th

e tr

ansf

orm

cod

e w

ould

per

form

no

bette

r th

andi

rect

sca

lar

quan

tizat

ion

appl

ied

to

X.

Tha

t is

, th

e di

stor

tion

wou

ldb

e

D ≅

1 k

∑ i=1k δ

Ui,s

q(R

i) =

1 k

∑ i=1k δ

Ui,s

q(R

) ≅

1 k

∑ i=1k

1 12

σ2 iα G

2-2

R

=

1 12

αG

1 k

∑ i=1k σ2 i

2-2

R

=

1 12

αG σ

2 X 2

-2R

≅ δ

X,s

q(R

) .

Thi

s co

nclu

sion

app

lies

to a

ll tr

ansf

orm

s.

We

conc

lude

tha

t th

e pr

oper

bit

allo

catio

n is

cru

cial

.

-If

the

tran

sfor

m w

ere

to p

rodu

ce c

oeffi

cien

ts a

ll of

whi

ch h

ad t

he s

ame

varia

nce

σ2 i,

th

en t

he o

ptim

al r

ate

allo

catio

n w

ould

be

the

equa

l-bit

allo

catio

n, a

nd t

he t

rans

form

cod

e w

ould

per

form

no

bette

r th

an d

irect

scal

ar q

uant

izat

ion

appl

ied

to

X.

•If

X

is n

ot G

auss

ian,

the

re's

no

sim

ple

rela

tion

amon

g th

e α

i's.

Nev

erth

eles

s, m

inim

izin

g Σ

2 TX

is u

sual

ly a

goo

d st

rate

gy t

o m

akin

g Γ

TX

sma

ll.

Tr-

30

TH

E O

PTIM

AL

TR

AN

SFO

RM

We

now

dis

cuss

how

to

choo

se t

he o

rtho

gona

l tr

ansf

orm

T

to

min

imiz

e

Σ2 =∆

∏ j=

1k σ

2 j1/

k

Thi

s w

ill b

e th

e op

timal

tra

nsfo

rm f

or t

he G

auss

ian

case

, an

d a

"goo

d"tr

ansf

orm

mor

e ge

nera

lly.

Mai

n F

act:

A k

-dim

ensi

onal

tra

nsfo

rm

T

min

imiz

es

Σ2 if

and

onl

y if

it is

aK

arhu

nen-

Loev

e T

rans

form

(K

LT)

for

the

k×k

co

varia

nce

mat

rix

KX o

fX

.

A K

LT i

s an

y tr

ansf

orm

who

se r

ows

are

an o

rtho

norm

al s

et o

fei

genv

ecto

rs o

f K

X.

For

a K

LT,

the

coef

ficie

nt v

aria

nces

σ

2 i a

re t

he e

igen

valu

es o

f K

X,

deno

ted

λ1,

…,λ

k a

nd

Σ2 =

∏ j=

1k λ

j1/

k

= |

KX|1/

k

whe

re

|KX|

den

otes

the

det

erm

inan

t of

K

X.

Tr-

31

MO

RE

FA

CT

S FR

OM

LIN

EA

R A

LG

EB

RA

The

pro

of o

f th

e M

ain

Fac

t is

bas

ed o

n m

ore

fact

s fr

om li

near

alg

ebra

.

Def

init

ion

: S

uppo

se

K

is a

k×k

m

atrix

, λ

is

a r

eal o

r co

mpl

ex n

umbe

r an

dv

is a

k-d

imen

sion

al v

ecto

r su

ch t

hat

Kv

= λ

v.

The

n λ

is

sai

d to

be

an e

igen

valu

e of

K

, v

is s

aid

to b

e an

eig

enve

ctor

of

K,

and

(λ,

v) a

re a

n ei

genp

air

for

K.

Fac

t 1:

E

very

k×k

mat

rix h

as

k e

igen

valu

es,

thou

gh t

hey

need

not

be

dist

inct

.

Fac

t 2:

T

he e

igen

valu

es o

f a

k×k

diag

onal

mat

rix

K

are

the

diag

onal

ele

me

nts

.

Pro

of:

One

may

dire

ctly

ver

ify t

hat

the

(K

(i,i),

v)

is a

n ei

genp

air,

whe

re

v =

(0

... 0

1 0

... 0

) w

ith a

1 in

the

ith p

lace

.

Fac

t 3:

R

eal

sym

met

ric m

atric

ies

have

rea

l (n

ot c

ompl

ex)

eige

nval

ues.

The

eig

enve

ctor

s as

soci

ated

with

dis

tinct

eig

enva

lues

are

ort

hogo

nal.

The

re i

s an

ort

hono

rmal

set

of

eige

nvec

tors

.

If th

e m

atrix

is a

lso

posi

tive

(res

pect

ivel

y, n

onne

gativ

e) d

efin

ite,

i.e.

if x

t K x

> 0

fo

r al

l x,

(

resp

ectiv

ely,

xt K

x ≥

0),

th

en t

heei

genv

alue

s ar

e po

sitiv

e (r

espe

ctiv

ely,

non

nega

tive)

.

Tr-

32

Fac

t 4:

T

he d

eter

min

ant

of a

squ

are

mat

rix is

the

pro

duct

of

itsei

genv

alue

s. i.

e.

if a

k×k

m

atrix

has

eig

enva

lues

λ 1

,…,λ

k,

then

|K|

= ∏ j=

1k λ

j

Fac

t 5:

The

det

erm

inan

t of

an

orth

ogon

al m

atrix

T

is

|T

| = ±

1.

Pro

of:

Sup

pose

λ

is a

n ei

genv

alue

of

T

and

v

is a

cor

resp

ondi

ngei

genv

ecto

r, i.

e.

Tv

= λ

v.

The

n

|| v||

= ||T

v|| =

||λv

|| =

|λ|

||v|

|

whi

ch im

plie

s |

λ| =

1.

Sin

ce a

ll th

e ei

genv

alue

s ha

ve m

agni

tude

one

,F

act 4

impl

ies

|T| =

±1.

Tr-

33

FA

CT

S A

BO

UT

CO

VA

RIA

NC

E M

AT

RIC

ES

Fac

t 6:

F

or a

ny k

×k c

ovar

ianc

e m

atrix

K

, ∏ i=

1k K

i,i ≥

|K

|, e

qual

ity if

f K

is

diag

'l.

Pro

of:

See

Ger

sho

and

Gra

y, p

p. 2

41-2

42

Fac

t 7:

If U

= T

X,

then

K

U =

T K

X T

t

Pro

of:

KU

= E

U U

t =

E T

X (

TX

)t = E

TX

Xt T

t = T

EX

Xt T

t = T

KX T

t .

Fac

t 8:

If T

is

ort

hogo

nal a

nd

U =

TX

, th

en

(a)

KX

and

KU

have

the

sam

e ei

genv

alue

s.

(b)

|KU|

= |K

X|

Pro

of:(

a) L

et (

λ,v)

be

an e

igen

pair

for

KX.

The

n K

UT

v =

TK

XT

t Tv

= T

λv =

λT

v.H

ence

, (

λ,T

v)

is a

n ei

genp

air

for

KU.

Thu

s ev

ery

eige

nval

ue o

f K

X

is a

lso

an e

igen

valu

e of

K

U.

The

sam

e ar

gum

ent

appl

ied

to K

U a

nd T

-1 s

how

s th

at e

very

eig

enva

l. of

K

U

is a

lso

an e

igen

valu

e of

K

X.

Thu

s K

X

and

KU

have

the

sam

e ei

genv

alue

s.

(b)

Sin

ce

KX

and

KU

have

the

sam

e ei

genv

alue

s, F

act

4 im

plie

s th

ey h

ave

the

sam

e de

term

inan

ts.

Fac

t 9:

For

any

cov

aria

nce

mat

rix t

here

is a

set

of

k o

rtho

norm

al e

igen

vect

ors.

Pro

of:

B

ecau

se c

ovar

ianc

e m

atric

es a

re r

eal a

nd s

ymm

etric

, an

d F

act

3.

Tr-

34

PR

OO

F O

F O

PTIM

AL

ITY

OF

KL

T

Lem

ma:

If

T

is

ort

hogo

nal a

nd m

akes

K

U

is d

iago

nal,

then

for

any

oth

eror

thog

onal

mat

rix

~ T,

~ Σ2 ≥

Σ2 ,

with

equ

ality

if a

nd o

nly

if K

~ U

is d

iago

nal.

Pro

of:

For

sim

plic

ity c

onsi

der

the

kth

pow

er o

f ~ Σ2 :

~ Σ2k =

∏ i=

1k ~ σ2 i

by

def

initi

on o

f ~ Σ2 ,

w

here

~ σ2 i

= E

~ U2 i

&

~ U=~ T

X.

≥ |

K ~ U|

F

act 6

fact

that

~ σ2 i's

are

dia

g el

emen

ts o

f K

~ U

= |

KX|

F

act 8

(b)

= |

KU|

Fac

t 8(b

)

=

∏ i=1k σ

2 i

Fac

t 6 &

fact

that

σ2 i '

s ar

e di

ag e

lem

ents

of

KU

=

Σ2k

for

T

By

Fac

t 6, e

qual

ity h

olds

iff

K ~ U

is d

iago

nal.

We

see

from

thi

s le

mm

a th

at o

ne c

an d

o no

bet

ter

than

to

mak

e th

e tr

ansf

orm

coef

ficie

nts

have

a d

iago

nal c

ovar

ianc

e m

atrix

.

Whe

n th

is is

don

e, t

he c

oeff.

var

ianc

es a

re t

he d

iago

nal e

lem

ents

, w

hich

by

Fac

t 2,

ar

e th

e ei

genv

alue

s of

K

U.

By

Fac

t 8a

, th

ese

are

also

the

eig

enva

lues

of

KX

as c

laim

ed in

the

Mai

n F

act.

Tr-

35

It re

mai

ns o

nly

to s

how

tha

t th

ere

is a

cho

ice

of T

tha

t m

akes

KU d

iago

nal.

Acc

ordi

ngly

, le

t T

be

the

typ

e of

mat

rix m

entio

ned

in t

he M

ain

Fac

t, i.e

. its

row

s ar

e an

ort

hono

rmal

set

of

eige

nvec

tors

t 1

,...,

t k

The

n, b

y F

act

5, a

nd t

he f

act

that

t i

is a

n ei

genv

ecto

r w

ith e

igen

valu

e λ

i

KU =

T K

X T

t =

--tt 1-

-

--tt 2-

-... --

tt k--

KX

| t 1 |

| t 2 | - - - | t k |

=

--tt 1-

-

--tt 2-

-... --

tt k--

| λ 1

t 1|

| λ 2t 2

| - - - | λ k

t k|

=

λ 1

0 ..

... 0

0 λ 2

0 ..

0 0

.....

0 λ k

.

So

as w

e ho

ped,

K

U

is d

iago

nal.

Thi

s fin

ishe

s th

e pr

oof

of t

he m

ain

fact

, na

mel

y, t

hat

Σ2

is m

inim

ized

by

a K

LTan

d th

at t

he r

esul

ting

min

imum

val

ue is

Σ2 =

∏ j=

1k λ

j1/

k

= |

KX|1/

k

.

Tr-

36

We

can

now

fin

d th

e O

PT

A f

unct

ion

for

tran

sfor

m c

odin

g a

Gau

ssia

n so

urce

.R

ecal

l tha

t fo

r a

Gau

ssia

n so

urce

δ X,tr

,T(R

) ≅

1 12

α G

Σ2 T

X 2

-2R

With

the

KLT

, σ

2 1,...

,σ2 k

= λ

1,...

,λk

and

Σ2 T

X =

∏ j=

1k σ

2 k1/

k

=

∏ j=

1k λ

j1/

k

= |

KX|1/

k

.

The

refo

re.

Th

e O

PT

A F

un

ctio

n f

or

k-d

imen

sio

nal

Tra

nsf

orm

Co

din

g a

pp

lied

to

aS

tati

on

ary,

Gau

ssia

n S

ou

rce:

For

larg

e R

,

δ tr(

k,R

) ≅≅≅≅

1 12

|KX|1/

k

αG

2-2

R

,

wh

ere α G

= 2

π 3

3/2

= 3

2.6

fo

r FR

C2

π e

= 1

7.0

8 fo

r VR

C

Tr-

37

Mo

reo

ver

•T

he b

est

tran

sfor

m is

the

KLT

, i.

e. r

ows

are

orth

onor

mal

eig

enve

ctor

s fo

rK

X.

•T

he r

esul

ting

coef

ficie

nts

U1,

…,U

k a

re u

ncor

rela

ted

(inde

ed,

inde

pend

ent)

.

•T

heir

varia

nces

σ2 1,

...,σ

2 k e

qual

the

eig

enva

lues

λ 1

,…,λ

k o

f K

X.

•T

he r

ate

allo

cate

d to

the

ith

coef

ficie

nt is

:

Ri

=

R +

1 2 log 2

λ i

|KX|1/

k

•T

he r

esul

ting

coef

ficie

nt d

isto

rtio

ns a

re a

ll th

e sa

me

and

equa

l to

δtr(k

,R).

Tr-

38

CO

MPA

RIS

ON

S FO

R T

HE

GA

USS

IAN

CA

SE

Opt

imal

k-D

imen

sion

al T

rans

form

cod

ing

δ tr(

k,R

) ≅≅≅≅

1 12

|KX|1/

k

αG

2-2

R

,

α G

= 2

π 3

3/2

= 3

2.6

fo

r FR

C2

π e

= 1

7.0

8 fo

r VR

C

Opt

imal

Sca

lar

Qua

ntiz

atio

n

δ sq(

R)

≅≅≅≅

1 12

σ2

α G 2

-2R

SN

R G

ain

over

sca

lar

quan

tizat

ion.

10 lo

g 10

δ sq(

R)

δ tr(

k,R

) ≅≅≅≅

10

log 1

0 σ2 X

|KX|1/

k

Opt

imal

k-d

imen

sion

al V

Q:

δ vq(

k,R

) ≅

m

* k σ2 X

αG

,k 2

-2R

,

α G

,k =

(k+2 k)(k

+2)

/2 |K

|1/k

1 σ2 X

for

FR

C

2π e

|K|1

/k 1 σ2 X

fo

r VR

C

SN

R G

ain

of O

ptim

al k

-dim

VQ

ove

r O

pt k

-dim

'l T

rans

form

Cod

ing

10 lo

g 10

δ tr(

k,R

)δ v

q(k,

R)

≅≅≅≅ 1

0 lo

g 10

1/12 m* k

|K

X|1/

k

αG

σ2 X α

G,k

=

10

log 1

0 1/

12 m* k

×

33/2

(k+2 k)(k

+2)

/2 f

or F

RC

1 f

or V

RC

Tr-

39

SN

R G

ain

of O

ptim

al h

igh

dim

'l V

Q o

ver

high

dim

'l T

rans

form

Cod

ing

in d

B

10 lo

g 10

δ tr(

∞,R

)δ v

q(∞

,R)

≅≅≅≅ 1

0 lo

g 10

1/12

1/2π

e +

10

log 1

0 3

3/2 e

for

FR

C1 f

or V

RC

=

1.53

+ 2

.81

fo

r FR

C0 f

or V

RC

=

4.35

fo

r FR

C1

.53 f

or V

RC

The

se a

re t

he s

ame

gain

s as

opt

imal

VQ

ove

r op

timal

SQ

for

IID

Gau

ssia

nso

urc

e.

Why

?

Whe

n op

timiz

ed,

tran

sfor

m c

odin

g su

ffers

no

mem

ory

loss

,bu

t it

suffe

rs t

he s

ame

cubi

c, o

blon

gitis

and

poi

nt d

ensi

ty lo

sses

as

optim

ized

sca

lar/

prod

uct

quan

tizat

ion

for

the

IID c

ase.

Fix

ed-r

ate

codi

ng:

One

can

sho

w t

hat

tran

sfor

m c

odin

g co

uld

be d

esig

ned

(a)

To

have

opt

im'a

l poi

nt d

ensi

ty.

In t

his

case

it w

ould

hav

e hi

ghob

long

itis

loss

.

(b)

To

have

cub

ic c

ells

. I

n th

is c

ase

it su

ffers

larg

e po

int

dens

ity lo

ss

The

opt

imal

is

a co

mpr

omis

e th

at c

ause

s sa

me

loss

es a

s in

the

IID

cas

e.

Var

iabl

e-ra

te c

odin

g:

One

can

des

ign

the

tran

sfor

m c

ode

to h

ave

the

optim

al p

oint

den

sity

(w

hich

is u

nifo

rm)

and

cubi

c ce

lls.

So

it su

ffers

onl

yth

e cu

bic

loss

.

Tr-

40

TH

E E

FFE

CT

OF

DIM

EN

SIO

N k

ON

|K

(k)

|

Let

K(k

)

be

the

k×k

co

varia

nce

mat

rix o

f X

w

ith e

igen

valu

es

λ(k)

1,…

,λ(k

)k

.

Fac

t 10:

|K

(k)

|

= M

k-1

|K(k

-1)

|

= σ

2 X ∏ i=

1

k-1 M

i

whe

re

Mk

is t

he M

SE

of

the

best

line

ar p

redi

ctor

for

X

i fr

om

Xi-k

, …,X

i-1.

Pro

of:

G

iven

in t

he D

PC

M n

otes

.

Fac

t 11:

|K

(k+1

)

|1/(k

+1)

|K

(k)

|1/

k

Thi

s im

plie

s th

at i

ncre

asin

g di

men

sion

will

not

dec

reas

e pe

rfor

man

ce.

Pro

of:

By

Fac

t 10,

|K(k

)

|1/k

=

geo

met

ric a

vera

ge o

f σ

2 X, M

1, …

, Mk

|K(k

+1)

|1/

(k+1

)

=

g

eom

etric

ave

rage

of

σ2 X, M

1, …

, Mk,

Mk+

1.

Obs

erva

tion:

M

k+1

≤ M

k,

beca

use

the

best

(k+

1)th

ord

er p

redi

ctor

mus

t be

at

leas

t as

goo

d as

the

bes

t kt

h or

der

pred

icto

r.

Obs

erva

tion:

th

e se

cond

geo

met

ric a

vera

ge is

like

the

firs

t ex

cept

it h

ason

e ad

ditio

nal t

erm

tha

t is

no

larg

er t

han

all t

he o

ther

s.

The

refo

re,

the

seco

nd g

eom

etric

ave

rage

is n

o la

rger

tha

n th

e fir

st.

Tr-

41

Not

e:

(Opt

iona

l Rea

ding

)

Fac

ts 1

0 an

d 11

are

rem

inis

cent

of

h(X

1,…

,Xk)

= h

(X1,

…,X

k-1)

+ h

(Xk|

X1,

…,X

k-1)

and

1 k h

(X1,

…,X

k) ≤

1 k-1

h(X

1,…

,Xk-

1)

Inde

ed,

they

can

be

prov

ed b

y us

ing

the

abov

e re

latio

ns f

or d

iffer

entia

len

trop

y.

Tr-

42

Fac

t 12:

lim k→∞ |K

(k)

|1/

k

= Q

(sta

ted

earli

er w

ithou

t pro

of)

wh

ere Q

=

exp

{ 1 2π

∫ -ππ ln S

(ω)

dω }

=

"one

-ste

p pr

edic

tion

erro

r"

=

MS

E o

f opt

imum

line

ar p

redi

ctor

for

Xi

base

d on

X

i-1, X

i-2, …

S(ω

) =

∑ n=-∞∞

RX(n

) e-jn

ω

=

pow

er s

pect

ral d

ensi

ty o

f ra

ndom

pro

cess

X

Pro

of:

lim k→

∞ |K

(k)

|1/

k

= l

im k→∞

(∏ i=

1k λ

(k)

i)1/

k

=

l

im k→∞

exp

{ ln

(∏ i=

1k λ

(k)

i)1/

k

}

= l

im k→∞

exp

{ 1 k

∑ i=1k ln

λ(k

)i

}

=

exp{

lim k→∞

1 k ∑ i=

1k ln

λ(k

)i

}

To

com

plet

e th

e pr

oof

we

need

to

show

lim k→∞

1 k ∑ i=

1k ln

λ(k

)i

=

1 2π

∫ -ππ ln S

(ω)

Tr-

43

Fac

t 13

: S

zeg

o's

Eig

enva

lue

Dis

trib

uti

on

Th

eore

m

Let

{X

i} b

e w

ide-

sens

e st

atio

nary

ran

dom

pro

cess

with

pow

er

spec

tral

den

sity

S

(ω)

and

with

k-d

imen

sion

al c

ovar

ianc

e m

atrix

K

(k)

ha

ving

eig

enva

lues

λ(k

)1

,…,λ

(k)

.

The

n fo

r an

y pi

ecew

ise

cont

inuo

us

func

tion

g

lim k→∞ 1 k

∑ i=1k g

(λ(k

)i

) =

1 2π

∫ -ππ g(S

(ω))

Ref

eren

ces:

U.

Gre

nand

er a

nd G

. S

zego

, T

oepl

itz F

orm

s an

d T

heir

App

licat

ions

(bo

ok)

R.M

. G

ray,

"T

oepl

itz a

nd C

ircul

ant

Mat

rices

: A

Rev

iew

" ,

(pap

er)

http

://w

ww

-ee.

stan

ford

.edu

/~gr

ay/t

oepl

itz.h

tml

Inte

rpre

tati

on

: T

his

theo

rem

det

erm

ines

the

asy

mpt

otic

"dis

trib

utio

n" o

f th

e ei

genv

alue

s of

the

cov

aria

nce

mat

rices

of

{X

i}.F

or e

xam

ple,

whe

n k

is

larg

e,

it de

term

ines

wha

t fr

actio

n of

the

eige

nval

ues

lie b

etw

een

a

and

b

for

any

a,b

.

See

nex

t pa

ge.

Tr-

44

Con

side

r th

e fu

nctio

n

g λ(s)

= 1

, s

≤ λ

0 ,

els

e

F(λ

) =

lim k→

∞ 1 k

∑ i=1k g

λ(λ(k

)i

) =

a

sym

pt fr

ac o

f e.v

.'s ≤

λ

=

dis

trib

utio

n fu

nctio

n fo

r th

e ei

genv

alue

s

The

G-S

The

orem

im

plie

s

F(λ

) =

1 2π

∫ -ππ g λ(

S(ω

)) d

ω

=

leng

th o

f {ω

: S(ω

) ≤ λ

}2π

λ

ω1

ω2

ω3

ω4

ω5

ω6

S(ω

)

ω

π-π

{ ω

: S

(ω)

≤ λ}

The

refo

re t

he "

dens

ity"

of e

igen

valu

es w

ith v

alue

s ne

ar

λ is

f(λ)

=

d dλ F

(λ)

≅≅≅≅

1 2π

1

|S'(ω

1)| +

1

|S'(ω

2)| +

whe

re

ω1,

ω

2, …

ar

e th

e fr

eque

ncie

s su

ch t

hat

S(ω

1) =

λ.

We

see

ther

ear

e m

any

eige

nval

ues

whe

re s

lope

is

flat

and

few

whe

re s

lope

is

stee

p.

Tr-

45

Com

plet

ion

of P

roof

of

Fac

t 12

:

Let

g(s)

= ln

(s).

The

n by

Fac

t 13

(th

e ei

genv

alue

dis

trib

utio

n th

eore

m)

lim k→∞ |K

(k)

|1/

k

=

lim k→

∏ i=1k λ

(k)

i1/

k

=

ex

p{ lim k→

1 k ∑ i=

1k ln

λ(k

)i

}

=

exp

{ 1 2π

∫ -ππ ln S

(ω) d

ω}

=

Q

whi

ch i

s th

e de

sire

d re

sult.

Tr-

46

Th

e O

PT

A F

un

ctio

n f

or

k-d

imen

sio

nal

Tra

nsf

orm

Co

din

g a

pp

lied

to

aS

tati

on

ary,

Gau

ssia

n S

ou

rce:

For

larg

e R

,

δ tr(

k,R

) ≅≅≅≅

1 12

|KX|1/

k

αG

2-2

R

δ tr(

R)

≅≅≅≅

1 12

Q α

G 2

-2R

as

k →

wh

ere α G

= 2

π 3

3/2

= 3

2.6

fo

r FR

C2

π e

= 1

7.0

8 fo

r VR

C

Q =

e

xp{

1 2π

∫ -ππ ln S

(ω)

dω }

In c

ompa

rison

, fo

r op

timal

VQ

and

Gau

ssia

n

δ vq(

k,R

) ≅

m

* k |K

X|1/

k

αG 2

-2R

δ tr(

R)

≅≅≅≅

1 2πe

Q α

G 2

-2R

a

s k

→ ∞

Tr-

47

EX

AM

PLE

: FI

RST

-OR

DE

R A

R, G

AU

SSIA

N S

OU

RC

E:

•X

i =

ρ X

i-1 +

Zi

whe

re

Zi's

ar

e IID

Gau

ssia

n, z

ero

mea

n w

ith v

aria

nces

σ2 Z

, an

d Z

i is

inde

pend

ent

of p

ast

X's

. I

n th

is c

ase,

•B

est

linea

r pr

edic

tor

for

Xi

from

X

i-1,…

,Xi-k

is

~ Xi

= ρ

Xi-1

. Its

MS

E is

σ2 Z.

M1

= M

2 =

M3

= …

= σ

2 Z =

σ2 X (

1-ρ2

) =

Q =

exp

{ 1 2π

∫ -ππ ln S

(ω) d

ω }

•F

act 1

0 ⇒

|K

(k)

|

= M

k-1

|K(k

-1)

|

= σ

2 X ∏ i=

1

k-1 M

i =

σ2k X

(1-

ρ2)k-

1

⇒ |

K(k

)

|1/k

=

σ2 X (

1-ρ2 )

(k-1

)/k

(

stat

ed e

arlie

r w

/o p

roof

in Z

ador

not

es)

σ2 X (

1-ρ2 )

as

k →

∞.

•O

PT

A,

k-di

m.

Tra

nsf.

Cod

ing

of S

tat'r

y, G

auss

, A

R S

ourc

e w

ith c

orr.

coe

f. ρ:

For

larg

e R

,

δ tr(

k,R

) ≅≅≅≅

1 12

σ2 X (

1-ρ2 )

(k-1

)/k

α

G 2

-2R

,

αG =

33

/2 =

32

.6 f

or F

RC

e =

17

.08 fo

r VR

C

1 12

σ2 X (

1-ρ2 )

αG

2-2

R

as

k →

Tr-

48

SNR

FO

R F

IXE

D-R

AT

E T

RA

NSF

OR

M C

OD

ING

Gau

ss A

R S

ourc

e --

cor

r. c

oeff.

ρ

= .9

, R

= 3

dim

ensi

on k

SNR, dB

101214161820222426

12345678

1216244896

192256

1024

Lob

SN

R*k

(3)

SN

R*(

3)

SNR

tr,k

(3)

Lcu

Lpt

(Ign

ore

the

loss

es.

In

this

plo

t, th

ey a

re d

efin

ed in

a d

iffer

ent

way

tha

n be

fore

.)

Tr-

49

Co

mp

aris

on

of

Tra

nsf

orm

Co

din

g a

nd

DP

CM

Con

side

r a

stat

iona

ry,

Gau

ssia

n so

urce

and

larg

e R

.F

or k

-dim

ensi

onal

tra

nsfo

rm c

odin

g:

δ tr,

k(R

) ≅

1 12

|K(k

)

|1/k

α

G 2

-2R

For

DP

CM

with

kth

-ord

er li

near

pre

dict

ion

δ dpc

m,k

(R)

1 12

Mk

α G 2

-2R

whe

re

Mk

is M

SE

of

optim

al k

th-o

rder

line

ar p

redi

ctor

for

X

i, i.

e.

from

Xi-k

,…,X

i-1.

Fac

t 14:

M

k+1

≤ M

k

Pro

of:

Sin

ce k

th o

rder

pre

dict

ion

is a

spe

cial

cas

e of

(k+

1)th

ord

erpr

edic

tion,

the

bes

t (k

+1)

th o

rder

pre

dict

or m

ust

be a

t le

ast

as g

ood

asth

e be

st k

th o

rder

pre

dict

or;

i.e.

Mk+

1 ≤

Mk.

Tr-

50

Fac

t 15:

|K

(k)

|1/

k

≥ M

k

Pro

of:

Fro

m F

acts

10

and

14

|K(k

)

|1/k

=

(σ2 X

∏ i=1

k-1 M

i)1/k

=

(σ2 X

∏ i=1

k-1 M

k-1)

1/k

Mk-

1 ≥

Mk

Fac

t 16:

lim k→∞ M

k =

lim k→

∞ |K

(k)

|1/

k

= e

xp{

1 2π

∫ -ππ ln S

(ω) d

ω }

Pro

of:

Firs

t w

e no

te t

hat

sinc

e th

e M

k's

are

non

nega

tive

and

noni

ncre

asin

g,th

ey c

onve

rge

to a

lim

it. S

econ

dly,

fro

m F

act

10

|K(k

)

|1/k

=

(σ2 X

∏ i=1

k-1 M

i)1/k

and

the

latte

r ex

pres

sion

con

verg

es t

o l

im k→

∞ M

k b

ecau

se w

hen

one

take

sth

e ge

omet

ric a

vera

ge o

f th

e fir

st

k t

erm

s of

a c

onve

rgen

t se

quen

ce (

the

Mk'

s),

that

geo

met

ric a

vera

ge c

onve

rges

to

the

sam

e va

lue

as t

hese

quen

ce.

Hen

ce,

lim k→∞ |K

(k)

|1/

k

=

lim k→∞ M

k

•B

y co

nsid

erin

g F

acts

11,

14,

15

and

16,

we

see

that

the

per

form

ance

s of

both

DP

CM

and

tra

nsfo

rm c

odin

g im

prov

e m

onot

onic

ally

with

k

to

the

sam

elim

it.

Mor

eove

r, t

he p

erfo

rman

ce o

f D

PC

M c

onve

rges

mor

e ra

pidl

y to

the

limit

than

doe

s th

e pe

rfor

man

ce o

f tr

ansf

orm

cod

ing.

Tr-

51

Exa

mp

le:

Firs

t-or

der

AR

, G

auss

ian

sour

ce:

Xi

= a

Xi-1

+ Z

i

whe

re

Zi's

ar

e IID

Gau

ssia

n, z

ero

mea

n w

ith v

aria

nces

σ2 Z

, an

d Z

i is

inde

pend

ent

of p

ast

X's

. I

n th

is c

ase,

M1

= M

2 =

M3

= …

= σ

2 Z =

σ2 X (

1-a2

) =

exp

{ 1 2π

∫ -ππ ln S

(ω) d

ω },

whi

ch m

eans

tha

t D

PC

M w

ith f

irst-

orde

r lin

ear

pred

ictio

n is

as

good

as

kth

-or

der

linea

r pr

edic

tion

for

any

k.

Der

ivat

ion:

F

or a

ny

k ≥

1, t

he b

est

linea

r pr

edic

tor

for

Xi

from

X

i-1,…

,Xi-k

is

~ Xi

= a

Xi-1

. T

his

can

be v

erifi

ed v

ia t

he o

rtho

gona

lity

prin

cipl

e, i.

e. b

y ch

ecki

ngth

at t

he e

rror

due

to

this

pre

dict

ion

is o

rtho

gona

l to

each

X

j, i-

k≤j≤

i-1:

E (

Xi-a

Xi-1

)Xj

= E

Zi X

j =

0

beca

use

Zi

is in

depe

nden

t of

X

j an

d ha

s ze

ro m

ean.

T

he r

esul

ting

MS

E is

:

Mk

= E

(Xi-~ X

i)2 =

E (

Xi-a

Xi-1

)2 = E

Z2 i

= σ

2 Z =

σ2 X (

1-a2

)

For

k-d

imen

sion

al t

rans

form

cod

ing

|K(k

)

|1/k

=

(σ2 X

(∏ i=

1

k-1 M

i))1/

k

= (

σ2 X (

∏ i=1

k-1 σ

2 Z))

1/k

= (σ

2 X (σ

2 X(1

-a2 )

)k-1

)1/

k

=

σ

2 X (

1-a2 )

(k-1

)/k

Thi

s is

larg

er t

han

M1

= σ

2 X (

1-a2

),

but

conv

erge

s to

it a

s k

→ ∞

.

Tr-

52

The

SN

R g

ain

in d

B o

f D

PC

M w

ith k

th-o

rder

line

ar p

redi

ctio

n ov

erk-

dim

ensi

onal

tra

nsfo

rm c

odin

g is

10 lo

g 10

δ tr,

k(R

)δ d

pcm

,k =

10

log 1

0 (1

-a2 )

(k-1

)/k

1-a2

= -

10 k lo

g 10

(1-a

2 )

For

a

= .

9,

this

gai

n is

plo

tted

belo

w.

05

1015

2025

3035

40012345678

k=tr

ansf

orm

cod

e di

men

., or

der

of D

PC

M li

near

pre

dict

or

dB

gain

of D

PC

M o

ver

tran

sfor

m c

odin

g

Tr-

53

Oth

er t

rans

form

s ar

e of

ten

used

:

DC

T,

FF

T a

nd w

avel

ets

The

y ar

e

fast

er,

sig

nal i

ndep

ende

nt a

nd h

ence

mor

e ro

bust

, p

erce

ptua

lly r

elev

ant,

appr

oxim

atio

ns t

o th

e K

LT

Rec

onsi

der

JPE

G

uses

DC

T in

stea

d of

KLT

, f

or a

ll of

the

abo

ve r

easo

ns

uses

US

Q w

ith V

L co

ding

, be

caus

e th

is is

the

bes

t ki

nd o

f sc

alar

quan

tizat

ion.

it

is a

lso

sim

ple

the

VLC

is n

ot s

impl

y H

uffm

an c

odin

g be

caus

e m

any

of t

he c

oeffi

cien

tsha

ve

H(U

i) <

1

and

in s

uch

case

s H

uffm

an c

odin

g yi

elds

giv

es p

oor

rate

.so

it u

ses

a m

ore

soph

istic

ated

sch

eme

in o

rder

to

get

Ri ≅

H(U

i)

JPE

G is

not

opt

imiz

ed f

or M

SE

, if

it w

ere

it w

ould

sho

ot f

or s

omet

hing

like

equa

l dis

tort

ion

of a

ll co

effic

ient

s.

Wha

t bl

ock

size

to

use?

If da

ta is

sta

tiona

ry G

auss

ian,

the

n pe

rfor

man

ce im

prov

es w

ith d

imen

sion

,to

a li

mit.

How

ever

, fo

r re

al d

ata

this

is n

ot a

lway

s th

e ca

se.

Ins

tead

the

dat

a is

mul

timod

al,

inho

mog

eneo

us,

blo

bby,

m

ixtu

res

of G

auss

ian,

non

stat

iona

ry

In t

his

case

, ch

oosi

ng t

oo la

rge

a bl

ockl

engt

h w

ill h

urt.