Download as pdf or txt
Download as pdf or txt
You are on page 1of 71

ASP ( <367)

-
Multirate signal processing
>
-
Normal LT
systems

x (3 +un -
y]


underlying assumption-sampling
rate is
same
for up Op ,

& impulse response

ultivate systems multiple rate


sampling
-

Broadcast Audio -
32KUE

CD -
44 . 1 K4E
DAT- 48 KHE

CD - Broadcast Audio
44 . 1 KZ 32kME
I

I
*
--> Broadcast
-
-D

Data
>
Thing >
-
Broadcast
Audio


advantageous
.
port
Basic operations
-

Time
sealing
n (n] =
2 2
,
3 , 4 ,
5
,
4, 3, 2
, 13

F [4] ,
=
↑52]

..
000
o
O O

O O 8
·
-

0 12345678
usn]

O 2
M = I

27 = 02 4

un] 50 12] U]

-
>

2 345 % ]
↑ ,

down
sampling/compression
xpansion

2n(y
n = 0 3 ...
,

82(2]
,

Ow

& [0] = [O]

72[B] = (1]

y2 [b] =
a [2]

22 [9] =
n(3]
0
i . jijj =j
n[9]
72 (12)
=

: 27 =
[E] -
(1) zeros
between
samples .

Expansion loss
of information
>
no
-

compression Poss loss


information
of
>
-

.
sampler/Expander
x
- XE []

Increase in
sampling rate
>
-

by am
Integer factor

E
IL , 12L,
[E] if
0
↑ =
x ,
n =
,
...

O ,
otherwise .

. (2) =
Xin] zt
n = - 0

n = mult .

of L

n = k L
*

XSE
X [R]

1
(2)
x (E) =
Ft
(e) LeSWL)
.

x =
x

set periodic
· in w with

Period = 2 A

ecol--period
=

T ↑
-
A/5
2
-

45

↓ upsampling 5
As

- -
In

(L-1) additional
# >
- =>

copies of
spectrum
I
-
-

-
>Interpolation

* - -

To
Linearity & Dini Invariance
- -
.

Sampler i 2x(E) a
x() = ,
n =

- O
,
Otherwise

XE(z) = x (z))

(L-1) Copies
of spectrum

23, 23 (3 ,
1 0 0
,
1 0
,
0
. , , ,

23 ! 23 23
0, 0 1 0 0 2 ...
.
, , , ,

Linear but Dime Variant

"delayed i/p does not result into

delayed off.
sampler/Decimation
own

>
-
>
-

x i>
x
D] = x [n]

C
results in

Stretchinseeat(
Derivation
-

* (2) =
(1 2

n = - 0

= n[n]z
#
= -

Intermed , sea
mupa
f+
2 []
.
n =

[n] =

O otherwise-

* Enn) -
+ [M)

x(z)
= 7 ,]
En
Mr =
*M
x
-

XD(z) =
, [k]z
k= - N

- (z) (E
&+, =
X ,

C
(zY )
*

Xp(z) =
y,

xp (ejw) =
x, (esw()
↑: 3 Work w =

/b

-
Comb sea.
x
,
[n] x (n)
mu
)
212
:
=

·
Otherwise
X , [M] = X [] []

ee
Clus
= !
. . .
- ↓ I
29
- j2n
DFT
COCYV
e
.

Ind =
Enj = ne

X
, [n] = X [n] [L]

spectrum

zu
x (7)
2 x , 2)
=
,

n =
-
*
(n)
(mm) z

x
I
↓ I
scaling ies Freq shifted

est
je
Lejwe C
-

= Zeilw-2%)
x

K = 0
,
. - -
+ -
1

B
--


-

↑p =
3
edit
the

can see
k 2
= 0 k =
1 k =

I E

I Her

-
-
Expression &
doteinsampung This is simply
ap .

scoring re-scaling
re-scaling
stift

operation on

to Frequency

-
illustration with an ple.
- -
-

of
Sampling a Nyquist
in
& 34
Nyquist
over sampling

......
>
-

27
~
...... X-,


-

5/3 *3

Now if

Nat
Now look at the scenario where the
original
Spectrum is not sampled at
Hyquist .

- A/6



ot
multiple comes
cu my---

/t
re-scalump
-

-
/6

He
/6
He
C I

20s t

Freq
of .

-
How if have sigual two form
of
we a

--

T Copies

L In
-

the He in s
ins
↓ re .
scaling

Alessing
-
2-3
e ↓ 24
3/2

-
Amp .

scaling
Downsampling >
-
shifted copils

>
Crescaling of freq

- rescaled crosses = n
If spectrum
A
leads to thes any
Inpect Aliasing will
crosses
Spectrum
occur

To Avoid Alcasing
6

-
Filter

↓ ing
Antialias
M wor
-

F-
observation >
-

spectrum is Band limited


up signal
-

If
/ Aliasing after
<
( Wal > No

down sampling
bog
- A
Sufficient condition -

Necessary condition -X

for B signals

However ,
if we can make an
statement
- For
Baseband/low pass symals it is
condition
necessary & Sufficient

Example :-

Arsen
speech
0
,
50004z

Fs = 7500Hz

What portion of the


ilp ↑
signal
is free
Spectrum
of aliasing 25047
>
-

& What isone


highest freq
Component
·

-ten
So 500

C
sample .
w
-

(4) of
2
x mult L
lat if no
,
-

=
X [4] - xe(us
O ,
otherwise .

Freq
- domain-usier/z
domain

x(z) = X (z) )

1.
(L-1) additional copies of
of servations .

spectrum in [0 25]
,

2 Freq
Scaling of .
.

.
3 No .
overlap of the copies

of Spectrum
4 No. loss of information
Les
.

↑ D +

I As
3

--
=
Downsamplel
-
-

+(2) =
x [m]
x [n]
Scaling
Freq -

Freq
domain

x, (z) x (n
= ↳
Spectral
I
T
copies

scaling additional
are
shifted
Copies

observations
-
. ↓ possibility of Overlap of
spectral copies
.
2 Potential for loss of information
↑ spectral distortion

upsample +
filtering
-
& terpolation

X En]

-pot iosspetrum

ne
·
-

Is / Ins line

- 2 A

ownsamplian Filtering
with

Antialiasing Filter-alias-free downsampling

·

·
↑O
band-kind
she
-
set

Sufficient
condition
to
avoid
&

alcasung
.
↳ Essence
t
-
O sampling
SR
SR Lk

-
.

X 2 [n]

E ES , s
TS

xote
I =>
[n]

2

We have
Ts
achieved

xci) -L -
x 2]


Ts/L
----

A Filter

X -
E
& -Invariance #
*nearity
.


↳+

- [un]
X
X. [m] >
-
,

t
x2[n] >
- X 2 [n]
Also
ax , [n] + bx2(n7 >
- a X
, [rn) + bx2 [M]
② - ↓
Hence Linear

anc -x - Xp[] = X (2)

of
Suitted
[M(n 1)]
18 x , (y x [n 1]
2 xy (n 1) x
= - - = -

X
[rn-
y
I
-

+ +
↑ -
=

XX ②
x , [Hn] = x [rn-1]

① Ot same

x (n -
1 Xp[n-1]
Produce

~ .
Downsample is not time invalian

and nea -Valiant


also
Sundar reasoning
is
with upsampler

Nime-variant

Example:
X [15
-
D +2 -

x b [n]

[n]
x -

20, .
. ..
0 ,
1 S
0 2
,
0 ,
3 ,
0
,
4 ,
0 5
,
0
,
0
,
.
0
]

x ] =

20 ,
0 ..

0 ,
2
,
3 ,
4
,
5
,
0
. -
0

(n-
if x it - x ]
x
(2 1
-
=

20 ,
0 . . -
0
,
1
,
0
,
2
,
0
,
3
,
0
,
4 ,
05
,
o]

x() =

20 , 0. - -
0 .
0
.
0 . . .
o]
Downsampling process

x (n]
T
-↳ >
-
x [2]
We
-
can Obtain M
Possible Sequence
.

x' [+ L]
[n -
>
-

x [n 2) -
> X" [M2]
-

=
multiple of N
I

2
n =

Otherwise
&

C
Comb
Sequence

0 ...
In
,
-00
I
24

Ch-1
= 2 newe
I I I
of
&

4 -

1 = Bull 19

H . 1 =
KM
.....
- n =
k +x +
1
↑+ +

&

C (n -2) =

!
Cop (n -
+ ]
sample
& -

x
[] =

21 ,
2 ,
3
,
4
,
5 ,
6 ,
7, 8, 9 ,
0 . . .

3
(()x 2) =

21 ,
0 ,
0
,
4 ,
0
,
0
, 7, 0 .
0
,
0 --)

⑪14 . 73 D

Ch -

1 x [] =

2 0
,
2
, 0
,
0
,
5
,
0
,
0
,
8, 0 . . . )

-I -

2.2 .
5, 8 ,
0
...

43[n - 2] x [2]
↳ -
13 2 ->
3
,
6
,
9, .

-
&
mins--as

x )
-

E - () (n- 15

< 2]
-I (n] -

-
-

1 + n(n]((n 1) -


-1
-
<>
= x
3 (n 2)

[n]
-
↑ ractional Rate change
x [2 & 12 HE -
X & 16 E

'[n]
XI
Case
-J > X
-
, 2) -

12
1 4 164z
x =
x

Xi [] = x [34]

gues of
3 xi Y] it
x [n] = ,
n = y
,

otherwise.
o ,

- i nemunda W
,
x2 [n]

-
:

B -
x2

16K4E
[L] - Cars

12knz

a
xesn) =

2 xle7 , it ismustot
Y2[n] = X2 [3n]

y
I
e

i ine
,

4
② - -

① ② are same
,
as
304
are both

Prime
numbers
.

x , (n] = X2 [n]

↳ Is it true in General ?
General as a

X2 [L]

]
I tiny - -
y (n] I X =-
[n]
J-F2[n]
I
>

=
X [n]

special case 1. L =
M

Intuitively
of information
-

N orloss

I
usof information .
~

(if (n) not BL


C )
x +

to
/up x z(z) = X (z

x
,
() =
x (
*
wi ↑ 2(z)
= (

Y(z) X 2zY)
(*
-
=
,

)
*

xq27 (z
=
X,
1=

x()(zy- e
+

d
-

Ix
+
=
z

ene I

safred = x(z)
Spectrum

'
to
[] = x [n] C [n] add ↑ times

(y)
divide
4o times

= x (2)
2


ye [n] =
x (2]
2 L

I
x[n]
7
-L
↓ N -
X ,
[]

,]
2)
- -I
X Eng
-
Y]

xc(z) (z)
=
=
X
x, (z)
Y(z) = + 2 (*)
Y (z) = x , (EV)
(
,

(
=

=
-

NisL
-
Q
j2π K
WH
-

= e

⑦ &Q same
are
only
if L are
relatively
Prime .
=
]
Earlier ,
we saw that im case of interchanging
upsampling & down sampling rescued in same

L
o

not
spectrally they may
so
but
correct

ineyt can be Spectrally correct

if we insert
filtering
Let's see how

Objective - -
Fractional
sampling
-
xCn]
↑ at
16kHz ot output
z change
I want

# *
at
op

A A 2A S is

-I
Case 1

T+T
- I↳
Interpol
L - ↑ 4 6
Filter
-

&
S

-es
Aa
- -
/3 5/s
~
silver
Interpolation
-

-
to
& 10

t
s

↓ rate
sampling i -

No
aliasing
No distortion
I

&
C
T
-
↑4
-

-
-

- u
A

48-2A

/
125
6
,y
-
- -

48

/it 4 2n

⑪ Filter of from / ,
/s (
i/p signal


- 2n

Filter is redundant
In a
way ,
An
Here

-
- A
**
~
A

we
signal warned
-
Review

rate with
↓ sampling conversion

filling
2
.
Always perform upsampling/Interpolations
first
③ +7
Hing -

# T
- -

x(L]
x [] -
[]
x
CGy
can cel each other
- important
powerful ↓ leads to

application observation

- -

I)--
-

- ent
- -
>
MA D

Denny - High rate stream >


-
Bultiple lower
-
-
rate stream .

Xo X3 . . .
I can also be though of
Paralel
-

s a serial to
x4 ...
conversion

T
-
7 X2
a
X5 - . -
T

things
-

x o X
Similar happened in
x2 X


, > X4X5 down samplers case .

x
o x . *
-xoxs +

E
+

·
T
x x +
X
.
,

T
-

I -
x2x5 o

-engii

-

x .
2, + ...
~
still demultiplayer ,
only difference
is instead dockwise
of going , denay is
going dockwise


stack with the upper branch
swings
=
,

around to the lower branch


.

# An
interpretation
Important
is
application ·

of
the OV-

useful in application. F
data communication & signal
analysis in

processing. where
you need to perform FFTs

Let's have a data


of
the
following form
Yo X
, . . .
-1 --
* Yay, . . .
Xu ,
,

--- -
BLOCK BLOK 2
O Block I

Blocks
of

creation
Blocking" of
data
.
-


Blocking can on done
efficiently
using the hardware but ,

Bulirate
by
concept

e
7

L
!
Multiplexes X

- -

a
-

swith
i
-
by rotating the ,
wo

offanfictiplay 1 ower data


rates higher data rates
-

to

xoxs -
-
- x0x ,23 +5 --

x . = E
!
&

B
+ x5
-
-

x04 -

13- X


000X 3

D X 00 x3

--⑰
0 , ,

For
0
Xy I
I
.


yes -
00 Xo ,
0 ,
0 , 43

- -
*24 , o .
oox
&X ,

-
x 200 X 5

Not o
sam
& X2X / X05 XyX3 *

-
-

Interchanging
- of build
Blocks .

1
x
- J

ideadical x 2) -
#
J
-

to

&
downsampling
Since addition linear
are

-- =--

scaling ↳ should be
preferred a

becaus throwin away the samples

muhlication
doing the is
&
resulti officiency doing
more than the

Ell also Bultiplication first


will for ,
&

word au g
ups
** - -

+ -

4 -

1 +
-

x2[] >
-

7 +


More
officient .

All these are


equivalent the we can think

wasteful computious
,

,
in terms
of whether we are
doing any
If so how we can avai& it
dealt with delay
& We have not

element
; because when
they come in Picture

have
We to be weny Careful .

- -
-
-

Summary
- -
> Denny -
Blocking of data

S/p(r)
serial-parallel
(1) -

(44)
I

one
chammed chammer

-
- MOX >
Unblocking of data
P(s (v)

Paralel-Series
(1) -
C

2AUX
-

I xoxs +
* T
x
o x .

Xo X3 . . .

-
-

IX4 ...
-
x
-
7
X5
- , x +

I
X2 . . .

T a
-

x o X x2 X
, > X4X5
# x2x5 10

-

data
Blocking
of clockwise
direction
/P (1 =
3) with
-
(p( + )
YA X
-

-
XOx x

Yo
- -
,

&
a
>
-

-
-
a

i
I
z
-
1

b
&

B
+ x5
-
-
- I

P(s(L 3) =

With anticlockwise

-
commutator

as(e Shorthand
-

notation
clockwise in one direction (at Demus (
director
& Antideckwise in occor
(at +x
So that data will come out in

correct Sequence .

>
-
All these are simple examples ,
if we

pieces et wa will
Dut all there
gett to ,

interesting
-

applications .

Dasfeel
Se

of festively
-
-
! giviot a
- -

i
a D
-

Jani of
unly
M

Lif
considere
as
a
black
boy
-
......
=
XM - [T -- >
-
X -,
[n]

What is
Relation W
X [] & x[n]

- Y0[]-D-VOC] =
XoC]

[u) Xi Sn]
x: -

D -
<n] = X,

!
X . . -
D - X ,
[] =
X M - / []

-
- efficient
L

-- -D-
-

- Vs
-

1. Mimize "wasteful computations


19 . computations to Obtain samples
that will be thrown away .

Ib .
computations on
Zero-values
Samples
.

Let's back take look


go and a at

Filtering
with
4/h

-E1-conflicts previously
as

S
We sech thof any
peration after upsampling
·

is in-efficient

- 1 In efficient

How do we address this


conflicting situation I
& observation

, my

x
, (z) =
x ( *
Y
,
(E) = X , (E) H(ET

=
()
[x( I


Now , Let's take another structure .

, [n]
X

T -eius
2

+L
*

x 23 () H

-
If4(7) Xe(z) H(z) (z)

E
is = x

9 LPF ; H(z)
is a BPF at
Ye() =

= x *
I
multiple freq.

#
[! Ein
.

·
2(z) = ()

&

Y , (E) = Yz(E) >
Ye[n] -
Y, [n]
↓ efficient

-
↓ if we can

represent Ast Futre


in this form ,
we

can achive
efficient
Implementation by
using this identity
.

2 observation

x as -, <]

X3(z) = H(z) x (z)


↑3(z) =
x z(zY)
=
H(zY) X (E)
another
structure.

x Ya

Ya(z) = x
(zY)H(EY)
↑> (E) -

Yn(z)
W

-
=
14
L
-
-- >

efficient .

--

I
S
*
Hobe Identifies &
9 H2z ) Can't

W lettbe
BPF

=I to

- -

* identifies
must be defined
Blocks
& realizabe
after applying noble
.

--

HOW
,
since in
upsampling & downsampling ,

the LPF is used ,

can we represent LPF using


above form for officient implementation ?

Let ,

h[n) -
LPF FIR .

z
With Hz) =

Zh(u]
2 3 Y

= no + 2
,
z
+
+ hez +
hyz + 2nz -
5
6
hyz
-

+
25 +
regrouping - Polyphase decompo
sition
-

41
Ho

(no
7
2jz)
2

n(z) =
+ hez +
24z +

↑ 5
z" hzz3
-

h , + +
n z

-
H
,
+

H(z) =
to(z) + z H , (z) (60(0 =

z"(
- 2

, + 4yz

Y)
-

+
h5z
now Let's
say
we want to do a

downconversion factor of 2
by a


J
M(E)
not
efficient

C- = >
-

-
L

Es J
I
-
I J
L
[In↑
-
- ·
+ 2 -

-
following linearity ,
We
I can

more the downsampler inside


How
invoking Note identities
.

Epox
a ...
-

-
I
-

I
12 -
7 ⑦> -

:
I

-
I
' we made it
& -. . .

C , YzX5
officient
X

it
...

3 as
Demux .

Can

power Bulbirate. be visualize


of
>
- can .

identifies
any filtering
We note to
apply
>
- can

operation to make it efficient


very
.

We Con

use the

Following
not alon
-

as well

for addition operator

-
I
>
-
2x
-
- -
-
-

III
using notta

identities -

T- Py -

My
-
-

15+) -

# Time domain description of strate


Filtering

"T
XI

x J
(E) -
>
- y , 2)

x, [n] =

= Wu

Y , [n] = X
, []

]W]
-us
I
3 xCE),
L
Cs ↳ =
Bult.
of
xe[] =

otherwise

Ye [] = <][n-y
= -

Ye [n] = = x 2(k) h[n -


k]
0
k
-
=

k = Bull. of L
- K = ML

=
x2(L] w(n -
mL]
m = -

x (m) whm

t ↓ Y [n]

O
Y] X2 []
-

x 22] =
x (myhen] - 2

② in Q

x
I S-my
(m) h
Deposition
Phase
(Bellanger 1976)
Gwy polyphase termi
-what is
magnitude
element
response
·f this

to
unity-all pass filter
but it has a phase

- also all
pass-only Phase

Basically ,
he wants to look for mechanism
by
which
M

a
fillor can be represented using
Crese
safilters with these elements (, = )
--


Wear co
poly phase

4(z) n z
=

nz hi e
-
I - E Cze]
grouping terms
,

all
even
&eval form Pophase Decomposition
↳ % time
used 90
most
frequently 9.
- .

= he
e
↳ (z) T

z tez-..... - (4-1)
↑ H(z) = z Er(
M
ut-polynomials
.

I
S
-
- E

eszu
·er ( +

n =
- 0

J

polyphase components .

apphostin
I see =

Em-1-e(z)
R(z) =
- >

-
T =

I - 7

: IT
&

↳ -
-

# Computational advantage
Agai consider a
simpler case .

--It L
>
-


Filler with
length -
meaning there

wil sa

Direct form (2 1)
Polyphases
-

structure s .

deloy
- -

ee ruts

&
2 N

multipliers .
I ↳
Block ,
Block

[2

I

S
M
2n-

IN

+
# computations

2n
1 >
-

-
/Op

1)

Block
A]
S
computations
& then
sample

multiplies

and des

&

I
2
-

two
.

↓ 2 will
throw away
Samples .

of
)
=
had we done
using pophase structure ,
what
would have been the implementation .

length
O -
-

computations/op

Even sample

E
↳ .

Stream

[NM +
(x -

1A] 2 + 1A

Strea +7 ↳E , (z)
>
-
&

forlor
of savings
in
~
by 2
in hardware
prominent
advantage of polyphase
is
So
Implementation .

#banks -> set


of filters .


Graphic Equalizer
.

Et
-
x [n] E ....
I
T
-
↑2 -
↑ [n] >
-
-
Q
Non-zero samples
are Odd
↳ Apply
cng

note identines
samples of
[n]
.

T-
--i

T IT
I
-

I
-
undefined
Block .

to

AT &
-
-

2
-
Y
1, non-zero
[n] ↳
samples are

even samples
of x my

# We have used a undefined block in the process .

visualize actually
If we
① they
,
are not

This we did sin


wrongx
means
same :

undefined
ok
optional to discuss
Err
-
-
-

n(- -
> 50 S to < n]

E
As
·

I 2(3n(M + 1] e , >

I
>
-

he
4(z)
=

-
>
-

his =
Eo(z) + zE (E)
structure
,

such
+ - -

is
at when Eo() =
Ih[n]z
th
food
Y
on

eei :
an in
=

ze , (e)

Polyphose
gives
componenti
ord P
#- h(-13 .
12-15 -

ear >

frodondrat
-

-
change
-
x, (n] x 2 (n] ↳
x - both
nothing
X [n]
- -I Y[n]

> + >
-
happanc
-

desin
A
· fal want
Y() - xe( *) -
same
When
yousampling
karp Rat
to
unchanged
)
*

rede
-
x2(z) = H(z) x (2 -

do easing
some
& ①

(e)
= (*

Y (1) =

) (
#
Enni S
-

=- -
this

down sampling
is

&
f 4 (n)
·

i
(n not Perfect Reconstruction
if y() = < x -

is possible .

Le
CONSA .

is
no .
-

Instagr ·
IP
sected delayed
version of p

Sp . Case
.
-

Pedet
C = 1 Reconstructio

U. = 0 system
-

Y (n] =
[]

Lets
apply
this condition to above expression ,

We want


* I
-

RMs in is
basically down sampling of
n [n]
In this case, which polyphase component have

We picked up I

n [M]

->
-

&

Zusteszu
E

-t
n =
- 0

↓ zeworth
polyphase components .

Polyphase
component
So
for ③ to hold ,
we can
say that
,
Oh component has to be coust
polyphase

Let's earlier observation


relate with our
Now ,
this

during sampling rate


change .

For Interpolation did filtering


,
we
after
up sampling or
PR .

We gol

#
LPF .

e
>
-

!

downsampled

Dome domain response


#

= -

M 1
-
-
--

think
Now
of this inpulse response ,
passing
through
on
polyphase component
-
only samples at O is non-zero ,
all other

samples multiple
of of 49 are zeros .

Both observations are consistent

Filter
banks

collection
of digital filters
analysisers .
with a common
i/p/op

O

S
↓lay
x

f x <] sub-band signals


I
-
> ,

Swanal
ce
,
#

↑ refer to
alysisafter
as bank .

- LF
Ex be low freg sign of
Ho can to allow

&

:
> Y ~
a ,

4 ,
-could be a BF to allow another
Freq
range
↑y could be DF
C
- a

Im Music
-
application in Busi system ,
Graphic Equalizes ,

the music
When
you want to playback .

G -Speaker E

↳J
ass,
vow ,
Gre

Hi - # -
mid free

↳-
.

** -
High frequency -

)
·
~

chess
bank

is
Synthes
- filers
10
--
7
-

Y - 7

est !
Far() , pa

utiple &
are
signals to form one sigmal
combined
More often we only focus
On
Analysis Filter

Danks

Uniform Spectral alysis



Signal is passed through filters Which
are
uniform in Bandwidth

#
+
In
S
the
Consty of Filter
An
--
bank these fulters are
,

related to each other

ros J wo)
· u ,
(z) = o(ze-
=
Mo(ew =Two
=

Ho(e)(w-wo)
/
[0 25)
of
our unterest is
frequencies ,

Now if we want to analyze if using M


filters
(of uniform size)
.

↳ each filter have


~2
must

#X --
4 , ( 40(ze
=
i)
a

-
--

- /na Wha
·

42) =

MOLE

- passband I - Stoband
Transition
band
> usually a don't care band
&
signat of intensedont the
#overseasin
a
band

# overlapping passband .

Ideal sectual Analysis malio

#. -

My na Yo

S
Ho LPF-symmetric may response > noth us

real

, 42 -
M 1
BPF- asymmetric magi> W , [n] is

response .

Complex .
results
usual
=> --

C
K 0 1 2
If
=
, ,

= -
D - k = 0

O
=
k = 1 2
- ,

↑General
-
.

2)
-
-
- -

III
1 = 0 1, . . . M - 1
,

3 -

=
D -
k

k
=

=
0

1 ,
2
,
--
-

M - 1

↳ K

muriple
of u

2 -

D-
O
Otherwise
a

applicationfeaborers

-Y-
-

Let 4(z) =
Ro(z3) + z" , (3)
+ z 22(3)

-
...... Dype - R

I polywase
+
103 - =>
J -
.
E ↑
as
I

7

I
,
I

-I
11

III
>

-
I
Sampling change
Fraction
of

> T
-

-
-
- if polyphase decomposition

this paci

on ,
we

can
get complexity
24 reduction
of
Complexity reduction

can we built a structure that can

give by complexity reduction 1


-

!.
combine with upsampling
↳ for upsampling
↳ LYPe - R is used .

- J+
7
7 7

T· &
M -

was to
decimator
y
the move
,

an the
left -
7 I R (2) -
- 7] 7 ↑

27

T
.

Di
& stepe

-E-X

-
Con
be interchange
relatively prima

&
-
Step 3

7 I R (2) -
T J
-

-
27
7 ↑

T
.

E
&
>
+ MB -

>
-
-E- -

-
>
z=3

--B-I-7
1
z
-

chine
> remamides
-

6 Theorem
Channel Eiler bark

#, (z) = 40 (z(x)

n
~
j2q
I = = & FX

(w-2)
>
= n .
(e

4 (z) () E (
=
,
=
4 .

n (z) = Ne (zze(*)

J
,

0
e =

Mc(z) =
40 (z(3)
.
**

4
+ .
(z) =
4022 Y

k = 1.. -
+ -
/
4x(z) =
40(Z)
-

) Er(((
Z
= -ke (zre(
*

)
l= 0

No
-

X x(z) = 4 , ( >
=
X (z)

S
dagger
I


+ -
'

Ets

[1 y
+
*
which band limited
signal
to
When ,
we have a is
2/
we can downsample by up without aliasing

e '

- Cam be moved
were .
As DFT
has multiplication
*
Very efficient implementation & addition
a peech on

do
hoble
Again
identities
can or
.
applied
S
DF + is also

Fast to implement
using FFT
downsample
How what's the role
of

-
x (z)

Lo >
Saul
fro9
NO

'

Fi
-

- I No
or
e
de

No

-
-

esirable for

Storage Purposes
-
n of necessary
.

You might also like