MULTIRATE SIGNAL PROCESSING

1. APPLICATIONS
2. THE UP-SAMPLER
3. THE DOWN-SAMPLER
4. RATE-CHANGING
5. INTERPOLATION
6. HALF-BAND FILTERS
7. NYQUIST FILTERS
8. THE NOBLE IDENTITIES
9. POLYPHASE DECOMPOSITION
10. EFFICIENT IMPLEMENTATION
11. POLYNOMIALS AND MULTIRATE FILTERING
12. INTERPOLATION OF POLYNOMIALS

I. Selesnick

EL 713 Lecture Notes

APPLICATIONS

1. Used in A/D and D/A converters.

2. Used to change the rate of a signal. When two devices that operate at different rates are to be interconnected, it is necessary
to use a rate changer between them.
3. Interpolation.
4. Some efficient implementations of single rate filters are based
on multirate methods.
5. Filter banks and wavelet transforms depend on multirate methods.

I. Selesnick

EL 713 Lecture Notes

THE UP-SAMPLER
The up-sampler, represented by the diagram,


x(n)

y(n)



is defined by the relation

(
x(n/2),
y(n) =
0,

for n even
for n odd.

(1)

The usual notation is

y(n) = [2] x(n).

(2)

The up-sampler simply inserts zeros between samples. For example,

if x(n) is the sequence
x(n) = {. . . , 3, 5, 2, 9, 6, . . . }
where the underlined number represents x(0), then y(n) is given by
y(n) = [2] x(n) = {. . . , 0, 3, 0, 5, 0, 2, 0, 9, 0, 6, 0, . . . }.
Given X(z), what is Y (z)? Using the example sequence above we
directly write
X(z) = + 3 z + 5 + 2 z 1 + 9 z 2 + 6 z 3 +

(3)

Y (z) = + 3 z 2 + 5 + 2 z 2 + 9 z 4 + 6 z 6 +

(4)

and

I. Selesnick

EL 713 Lecture Notes

It is clear that
Y (z) = Z {[2] x(n)} = X(z 2 ).

(5)

We can also derive this using the definition:

X
Y (z) =
y(n) z n

(6)

x(n/2) z n

(7)

n even

x(n) z 2 n

(8)

= X(z 2 ).

(9)

How does up-sampling affect the Fourier transform of a signal?

The discrete-time Fourier transform of y(n) is given by

Y (ej ) = X(z 2 )z=ej

(10)

= X((ej )2 )

(11)

Y (ej ) = X(ej2 ).

(12)

so we have

Or using the notation Y f () = Y (ej ), X f () = X(ej ), we have

Y f () = DTFT {[2] x(n)} = X f (2).

(13)

When sketching the Fourier transform of an up-sampled signal, it is

easy to make a mistake. When the Fourier transform is as shown in
the following figure, it is easy to incorrectly think that the Fourier
transform of y(n) is given by the second figure. This is not correct,
because the Fourier transform is 2-periodic. Even though it is
usually graphed in the range or 0 , outside
I. Selesnick

EL 713 Lecture Notes

this range it is periodic. Because X f () is a 2-periodic function

of , Y f () is a -periodic function of .
The correct graph of Y f () is the third subplot in the figure.

Xf()

1
0.5

Yf() = Xf(2)

Xf(2) WRONG!

0
1

0.5

0
/

0.5

0.5

0
/

0.5

0.5

0
/

0.5

1
0.5
0
1
1
0.5
0
1

Note that the spectrum of X f () is repeated there is an extra

copy of the spectrum. This part of the spectrum is called the
spectral image.
General case: An L-fold up-sampler, represented by the diagram,


x(n)

y(n)



is defined as
(
y(n) = [L] x(n) =
I. Selesnick

x(n/L),

when n is a multiple of L

0,

otherwise.

EL 713 Lecture Notes

(14)
The L-fold up-sampler simply inserts L 1 zeros between samples.
For example, if the sequence x(n)
x(n) = {. . . , 3, 5, 2, 9, 6, . . . }
is up-sampled by a factor L = 4, the result is the following sequence
y(n) = [4] x(n)
= {. . . , 0, 3, 0, 0, 0, 5, 0, 0, 0, 2, 0, 0, 0, 9, 0, 0, 0, 6, 0, . . . }.
Similarly, we have
Y (z) = Z {[L] x(n)} = X(z L ),

(15)

Y (ej ) = X(ejL ),

(16)

Y f () = DTFT {[L] x(n)} = X f (L ).

(17)

The L-fold up-sampler will create L 1 spectral images. For example, when a signal is up-sampled by 4, there are 3 spectral images
as shown in the following figure.

Xf()

1
0.5

Yf() = Xf(L)

0
1

0
/

0.5

0.5

0
/

0.5

1
0.5
0
1

I. Selesnick

0.5

EL 713 Lecture Notes

Remarks
1. No information is lost when a signal is up-sampled.
2. The up-sampler is a linear but not a time-invariant system.
3. The up-sampler introduces spectral images.

I. Selesnick

EL 713 Lecture Notes

THE DOWN-SAMPLER
The down-sampler, represented by the following diagram,


x(n)

y(n)



is defined as
y(n) = x(2 n).

(18)

The usual notation is

y(n) = [2] x(n).

(19)

The down-sampler simply keeps every second sample, and discards

the others. For example, if x(n) is the sequence
x(n) = {. . . , 7, 3, 5, 2, 9, 6, 4, . . . }
where the underlined number represents x(0), then y(n) is given by
y(n) = [2] x(n) = {. . . , 7, 5, 9, 4, . . . }.
Given X(z), what is Y (z)? This is not as simple as it is for the
up-sampler. Using the example sequence above we directly write
X(z) = +7 z 2 +3 z +5+2 z 1 +9 z 2 +6 z 3 +4 z 4 + (20)
and
Y (z) = + 7 z + 5 + 9 z 1 + 4 z 2 +

(21)

How can we express Y (z) in terms of X(z)? Consider the sum of

X(z) and X(z). Note that X(z) is given by
X(z) = +7 z 2 3 z+52 z 1 +9 z 2 6 z 3 +4 z 4 + . (22)
I. Selesnick

EL 713 Lecture Notes

The odd terms are negated. Then

X(z)+X(z) = 2 + 7 z 2 + 5 + 9 z 2 + 4 z 4 +

(23)

or
X(z) + X(z) = 2 Y (z 2 )

(24)

or
1

X(z 2 ) + X(z 2 )
Y (z) = Z {[2] x(n)} =
2

(25)

How does down-sampling affect the Fourier transform of a signal?

The discrete-time Fourier transform of y(n) is given by

1
1 
2
2
X(z ) + X(z ) 
Y (ej ) =

 j
2
z=e

1
= X(ej 2 ) + X(ej 2 )
2

1
= X(ej 2 ) + X(ej ej 2 )
2 

1
j 2
j( 2 )
)
= X(e ) + X(e
2 

1
f
f 2
= X ( )+X (
)
2
2
2

(26)
(27)
(28)
(29)
(30)



1

2
Y f () = DTFT {[2] x(n)} = X f ( ) + X f (
)
2
2
2
(31)
where we have used the notation Y f () = Y (ej ), X f () =
X(ej ).
Note that because X f () is periodic with a period of 2, the functions X f ( 2 ) and X f ( 2
2 ) are each periodic with a period of 4.
I. Selesnick

EL 713 Lecture Notes

But as Y f () is the Fourier transform of a signal, it must be 2periodic. What does Y f () look like? It is best illustrated with an
example.
1.2
1
Xf()

0.8
0.6
0.4
0.2
0
3

0
/

0
/

0
/

0
/

1.2

0.5 * Xf(/2)

1
0.8
0.6
0.4
0.2
0
3

0.5 * Xf((2)/2)

1.2
1
0.8
0.6
0.4
0.2
0
3
1.2
1
Yf()

0.8
0.6
0.4
0.2
0
3

Notice that while the two terms X f ( 2 ) and X f ( 2

2 ) are 4periodic, because one is shifted by 2, their sum is 2-periodic, as
a Fourier transform must be.
Notice that when a signal x(n) is down-sampled, the spectrum
X f () may overlap with adjacent copies, depending on the specific
shape of X f (). This overlapping is called aliasing. When aliasing
occurs, the signal x(n) can not in general be recovered after it
I. Selesnick

EL 713 Lecture Notes

10

is down-sampled. In this case, information is lost by the downsampling. If the spectrum X f () were zero for /2 || ,
then no overlapping would occur, and it would be possible to recover
x(n) after it is down-sampled.
General case: An M -fold down-sampler, represented by the diagram,


x(n)

y(n)



is defined as
y(n) = x(M n).

(32)

The M -fold down-sampler keeps only every M th sample. For example, if the sequence x(n)
x(n) = {. . . , 8, 7, 3, 5, 2, 9, 6, 4, 2, 1, . . . }
is down-sampled by a factor M = 3, the result is the following
sequence
y(n) = [3] x(n) = {. . . , 8, 5, 6, 1, . . . }.
Similarly, we have
M 1
1
1 X
Y (z) = Z {[M ] x(n)} =
X(W k z M )
M

(33)

k=0

where
2

W = ej M ,
I. Selesnick

(34)
EL 713 Lecture Notes

11

and


M
1
X
1

2
k
Y f () = DTFT {[M ] x(n)} =
Xf
.
M
M

(35)

k=0

Remarks
1. In general, information is lost when a signal is down-sampled.
2. The down-sampler is a linear but not a time-invariant system.
3. In general, the down-sampler causes aliasing.

I. Selesnick

EL 713 Lecture Notes

12

RATE-CHANGING
The up-sampler and down-sampler are usually used in combination
with filters, not by themselves. For example, to change the rate of
a signal, it is necessary to employ low-pass filters in addition to the
up-sampler and down-sampler.
The following system is used for interpolation.


x(n)

H(z)

y(n)



The combined up-sampling and filtering can be written as

X
y(n) = ([L] x(n)) h(n) =
x(k) h(n L k).

(36)

The filter fills in the zeros that are introduced by the up-sampler.
Equivalently, it is designed to remove the spectral images. It should
be a low-pass filter with a cut-off frequency o = /L. In this
context, the low-pass filter is often called an interpolation filter.
The following system is used for decimation.


x(n)

H(z)

y(n)



The combined filtering and down-sampling can be written as

X
y(n) = [M ] (x(n) h(n)) =
x(k) h(M n k).
(37)
k

The filter is designed to avoid aliasing. It should be a low-pass filter

with a cut-off frequency o = /M . In this context, the low-pass
filter is often called an anti-aliasing filter.
I. Selesnick

EL 713 Lecture Notes

13

A rate changer for a fractional change (like 2/3) can be obtained

by cascading an interpolation system with a decimation system.
Then, instead of implementing two separate filters in cascade, one
can implement a single filter. Structure for rational rate changer:


x(n)



H(z)



y(n)



The filter is designed to both eliminate spectral images and to avoid

aliasing. The cascade of two ideal low-pass filters is again a lowpass filter with a cut-off frequency that is the minimum of the two
cut-off frequencies. So, in this case, the cut-off frequency should
be
n o
o = min
,
.
L M

I. Selesnick

EL 713 Lecture Notes

(38)

14

INTERPOLATION EXAMPLE 1
In this example, we interpolate a signal x(n) by a factor of 4,
using the interpolation system described above. We use a linearphase Type I FIR lowpass filter of length 21 to follow the 4-fold upsampler. Note that because the filter is causal, a delay in introduced
by the interpolation system. y(n) could be aligned with x(n) by
shifting it.
5
1

0.6

|Hf()|

h(n)

0.8
0.4
0.2
0

3
2
1

0.2
0

10
n

15

20

0.2

0.4

0.6

0.8

1.2
1

x(n)

0.8
0.6
0.4
0.2
0

10

12

14

16

18

20

n
1.2
1

x2(n)

0.8
0.6
0.4
0.2
0

10

20

30

40
n

50

60

70

80

1.2
1

y(n)

0.8
0.6
0.4
0.2
0

I. Selesnick

10

20

30

40

50
n

60

70

EL 713 Lecture Notes

80

90

100

15

INTERPOLATION EXAMPLE 2
This time we use a filter of length 7,
1
h(n) = {1, 2, 3, 4, 3, 2, 1}.
(39)
4
Note that this filter has the effect of implementing linear interpolation between the existing samples x(n). The result is rather poor
the signal y(n) is not very smooth. Similarly, quadratic interpolation can be implemented by using an appropriate filter h(n).
5
1

0.6

|Hf()|

h(n)

0.8
0.4
0.2
0

3
2
1

0.2
0

0.2

0.4

0.6

0.8

1.2
1

x(n)

0.8
0.6
0.4
0.2
0

10

12

14

16

18

20

n
1.2
1

x2(n)

0.8
0.6
0.4
0.2
0

10

20

30

40
n

50

60

70

80

1.2
1

y(n)

0.8
0.6
0.4
0.2
0

10

20

30

40

50

60

70

80

I. Selesnick

EL 713 Lecture Notes

16

HALF-BAND FILTERS
When interpolating a signal x(n), the interpolation filter h(n) will
in general change the samples of x(n) in addition to filling in the
zeros. It is natural to ask if the interpolation filter can be designed
so as to preserve the original samples x(n).
To be precise, if
y(n) = h(n) [2] x(n)
then can we design h(n) so that y(2n) = x(n), or more generally,
so that y(2n + no ) = x(n) ?
It turns out that this is possible. When interpolating by a factor of
2, if h(n) is a half-band, then it will not change the samples x(n).
A no -centered half-band filter h(n) is a filter that satisfies
(
1,
for n = no
h(n) =
0,
for n = no 2, 4, 6, . . .

(40)

That means, every second value of h(n) is zero, except for one such
value, as shown in the figure.
A HALFBAND FILTER
1.2
1

h(n)

0.8
0.6
0.4
0.2
0
0.2

10
n

12

14

16

18

20

In the figure, the center point is no = 10. The definition of a

half-band filter can be written more compactly using the Kronecker
delta function (n). Half-band filter:
h(2 n + no ) = (n),
I. Selesnick

EL 713 Lecture Notes

(41)
17

when no = 0, we get simply:

h(2 n) = (n).

(42)

Note that the transfer function of a half-band filter (centered at

no = 0) can be written as
H(z) = 1 + z 1 H1 (z 2 ).

(43)

Here H1 (z) contains the odd samples of h(n).

NYQUIST FILTERS

When interpolating a signal x(n) by a factor L, the original samples

of x(n) are preserved if the interpolation filter h(n) is a Nyquist-L
filter. A Nyquist-L filter simply generalizes the notion of the halfband filter to L > 2. A (0-centered) Nyquist-L filter h(n) is one
for which
h(L n) = (n).

(44)

A Nyquist-4 filter is shown in the following figure.

A NYQUIST4 FILTER
1.2
1

h(n)

0.8
0.6
0.4
0.2
0
0.2
0

I. Selesnick

10

15
n

EL 713 Lecture Notes

20

25

30

18

THE NOBLE IDENTITY FOR THE UP-SAMPLER

The two following equivalences are sometimes called the noble identities.
Can you reverse the order of an up-sampler and a filter?
Yes and no it depends. There are two cases.
1. If the up-sampler comes after the filter, then you can reverse
the order of the filter and the up-sampler, but the filter needs
to be modified as shown in the figure.
2. If the up-sampler comes before the filter, then you can not
reverse their order unless the filter is of the special form H(z L ).
This can be summarized by the following figure.


x(n)

H(z)

y(n)



m


x(n)



H(z L )

y(n)

Equivalently:
[L] (h(n) x(n)) = [L] h(n) [L] x(n)

I. Selesnick

EL 713 Lecture Notes

19

THE NOBLE IDENTITY FOR THE UP-SAMPLER

This identity is most easily derived using the Z-transform and equation (15). In the following figure the intermediate signal v(n) is
shown.
x(n)

v(n)

H(z)


-

y(n)



Then, using the Z-transform, we have

V (z) = H(z) X(z)

Y (z) = V (z L )

and

and therefore,
Y (z) = H(z L ) X(z L ).
Now consider the system that we claim to be equivalent. In the
following figure the intermediate signal w(n) is shown.


x(n)

w(n)



H(z L )

y(n)

Then, using the Z-transform, we have

W (z) = X(z L )

and

Y (z) = H(z L ) W (z)

and therefore,
Y (z) = H(z L ) X(z L ).
This shows that the systems are equivalent.
I. Selesnick

EL 713 Lecture Notes

20

THE NOBLE IDENTITY FOR THE DOWN-SAMPLER

Can you reverse the order of an down-sampler and a filter?
Yes and no it depends. There are two cases.
1. If the down-sampler comes before the filter, then you can reverse the order of the filter and the down-sampler, but the
filter needs to be modified as shown in the figure.
2. If the down-sampler comes after the filter, then you can not reverse their order unless the filter is of the special form H(z M ).
This can be summarized by the following figure.


x(n)

H(z)

y(n)



m


x(n)

H(z )

y(n)



Equivalently:
h(n) [M ] x(n) = [M ] ([M ] h(n) x(n))

I. Selesnick

EL 713 Lecture Notes

21

THE NOBLE IDENTITY FOR THE DOWN-SAMPLER

For convenience, we prove it just for M = 2. This identity is most
easily derived using the Z-transform and equation (25). In the
following figure the intermediate signal v(n) is shown.


x(n)

v(n)

H(z)

y(n)



Then, using the Z-transform, we have

1
1
1
1
V (z) = X(z 2 ) + X(z 2 )
2
2

and

Y (z) = H(z) V (z)

and therefore,
1
1
1
1
Y (z) = H(z) X(z 2 ) + H(z) X(z 2 )
2
2

Now consider the system that we claim to be equivalent. In the

following figure the intermediate signal w(n) is shown.
x(n)

H(z 2 )

w(n)


-

y(n)



Then, using the Z-transform, we have

W (z) = H(z 2 ) X(z)

and

1
1
1
1
Y (z) = W (z 2 ) + W (z 2 )
2
2

and therefore,
1
1
1
1
Y (z) = H(z) X(z 2 ) + H(z) X(z 2 ).
2
2

This shows that the systems are equivalent.

I. Selesnick

EL 713 Lecture Notes

22

POLYPHASE DECOMPOSITION
The polyphase decomposition of a signal is simply the even and odd
samples,
x0 (n) = x(2 n)

(45)

x1 (n) = x(2 n + 1).

(46)

Then the Z-transform X(z) is given by

X(z) = X0 (z 2 ) + z 1 X1 (z 2 )

(47)

where X0 (z) and X1 (z) are the Z-transforms of x0 (n) and x1 (n).
For example, if x(n) is:
x(n) = {3, 1, 5, 6, 2, 4, 3, 7}
then the polyphase components are
x0 (n) = {3, 5, 2, 3}

(48)

x1 (n) = {1, 6, 4, 7}.

(49)

The Z-transforms for this example are given by

X(z) = 3 + z 1 + 5 z 2 + 6 z 3 + 2 z 4 + 4 z 5 3 z 6 + 7 z 7
X0 (z) = 3 + 5 z 1 + 2 z 2 3 z 3
X1 (z) = 1 + 6 z 1 + 4 z 2 + 7 z 3 .
In general, X0 (z) and X1 (z) can be obtained from X(z) as,
1
X0 (z 2 ) = (X(z) + X(z))
(50)
2
z
X1 (z 2 ) = (X(z) X(z)) .
(51)
2
The polyphase components x0 (n), x1 (n) can be obtained with the
following structure.
I. Selesnick

EL 713 Lecture Notes

23

x(n)


2 

x0 (n)


2 

x1 (n)

General case: An M -component polyphase decomposition of x(n)

is given by
x0 (n) = x(M n)

(52)

x1 (n) = x(M n + 1)
..
.

(53)

xM 1 (n) = x(M n + M 1).

(54)
(55)

The Z-transform X(z) is then given by

X(z) = X0 (z M ) + z 1 X1 (z M ) + + z (M 1) XM 1 (z M )
(56)
where Xi (z) is the Z-transform of xi (n). The polyphase component
Xi (z) can be found from X(z) with
M 1
z i X ik
Xi (z ) =
W X(W k z)
M
M

(57)

k=0

where
2

W = ej M .

I. Selesnick

(58)

EL 713 Lecture Notes

24

EFFICIENT IMPLEMENTATION
The noble identities and the polyphase decomposition can be used
together to obtain efficient structures. Consider again the system
for interpolation: an up-sampler is followed by a filter. In this
system, the up-sampler inserts zeros between the samples x(n).
There are two disadvantages.
1. Half the samples of the input to the filter are zero. That means
the filter is doing unnecessary computations (multiplications by
zero, adding zeros).
2. The filter operates at the higher rate.
A more efficient implementation can be obtained by writing the
filter in polyphase form, and then using the noble identities. This
is done through the following transformation of the block diagram.
x(n)

x(n)


2 


2 - H0 (z 2 )




x(n)



y(n)

+ z 1 H1 (z 2 )

H0 (z 2 )

-+
l-

y(n)

y(n)

I. Selesnick

H(z)

z 1 H1 (z 2 )

EL 713 Lecture Notes

25

EFFICIENT IMPLEMENTATION

x(n)

x(n)


2 

H0 (z 2 )


2 

z 1 H1 (z 2 )

-+
l-

y(n)


2 

H0 (z 2 )

-+
l-

y(n)

z 1

- 2


x(n)

H0 (z)

6
2

H1 (z )


2 - +l 
6

y(n)

z 1
 6
-

H1 (z)



Note that in the last block diagram, the filters operate at the slower
rate, and the filter inputs are not zero. Also note that the filters
h0 (n), h1 (n) are each half the length of the original filter h(n). The
adding node in the last diagram does not incur any actual additions
it implements an interleaving of the two branches.

I. Selesnick

EL 713 Lecture Notes

26

HALF-BAND CASE

If h(n) is a half-band filter, then the polyphase component H0 (z)

is 1 (assuming the half-band filter is centered at no = 0). In this
case, the block diagram becomes more simple as shown.

x(n)


2 - +l 
6

y(n)

z 1
 6
-

I. Selesnick

H1 (z)



EL 713 Lecture Notes

27

POLYNOMIAL SIGNALS
A (discrete-time) polynomial signal x(n) is a signal of the form
x(n) = c0 + c1 n + c2 n2 + + cd nd .
The degree is d. The set of polynomial signals of degree d or less
is denoted by Pd .
Consider a system described by the rule
y(n) = x(n) x(n 1).
This system gives the first difference of the signal x(n). It has the
impulse response
h(n) = (n) (n 1),
and the transfer function
H(z) = 1 z 1
and so we can write
y(n) = h(n) x(n)
or


Y (z) = 1 z 1 X(z).
Clearly if x(n) is a constant signal (x(n) = c, so we can write
x(n) P0 ), then the first difference of x(n) is identically zero,


Y (z) = 1 z 1 X(z) = 0 for x(n) P0 .
Moreover, the first difference Y (z) is identically zero only if x(n)
is a constant signal.
I. Selesnick

EL 713 Lecture Notes

28

POLYNOMIAL SIGNALS
Similarly, if x(n) is a ramp signal (x(n) = c0 + c1 n, so we can
write x(n) P1 ), then the first difference is a constant signal.
Therefore the second difference, (defined as the first difference of
the first difference), must be identically zero. Writing this using
the Z-transform gives
1 z 1

2

X(z) = 0 for x(n) P1 .

Moreover, the second difference of x(n) is identically zero only

if x(n) is of the form c0 + c1 n. Therefore, the set of first degree
polynomial signals P1 is exactly the set of signals that is annihilated
by (1 z 1 )2 .
Similarly, if x(n) is a polynomial signal of degree d, then
Y (z) = 1 z 1

d+1

X(z) = 0 for x(n) Pd .

or equivalently,
y(n) = h(n) h(n) h(n) x(n) = 0 for x(n) Pd .
{z
}
|
d + 1 terms

Moreover, y(n) = 0 only if x(n) has the form x(n) = c0 + c1 n +

c2 n2 + + cd nd .
Therefore we have the following result.
x(n) Pd

I. Selesnick

1 z 1

EL 713 Lecture Notes

d+1

X(z) = 0

29

INTERPOLATION OF POLYNOMIALS
We saw before that the interpolation of discrete-time signals can
be carried out by using an upsampler together with a filter. For
interpolation by a factor of two (2X interpolation) we have the
following diagram.
x(n)


2 

H(z)

y(n)

Suppose x(n) is a polynomial signal of degree d. Then it is natural

to ask that y(n) also be a polynomial signal of degree d. But
for just any filter h(n) that will not be the case. What condition
must h(n) satisfy, to ensure that y(n) is also a polynomial signal
of degree d?
It turns out that if (1 + z 1 )d+1 is a factor of H(z),
H(z) = Q(z) (1 + z 1 )d+1
then y(n) Pd whenever x(n) Pd . This can be verified using the
boxed result on the previous page together with the noble identity,
as we will now show.
We may determine if y(n) Pd by filtering y(n) with the transfer
function (1 z 1 )d+1 and checking that the result is zero. If the
signal v(n) in the following figure is zero, then we know that y(n)
Pd as explained earlier.

x(n)

I. Selesnick


- 2


Q(z) (1 + z 1 )d+1

y(n)

EL 713 Lecture Notes

- (1 z 1 )d+1

v(n)

30

We can rearrange the order of the filters to get the following diagram
which is an equivalent structure (end-to-end).

- 2


x(n)

(1 z 1 )d+1 (1 + z 1 )d+1

Q(z)

v(n)

Recognizing that (1 z 1 ) (1 + z 1 ) = 1 z 2 we get the following

diagram.

- 2
- (1 z 2 )d+1


x(n)

Q(z)

v(n)

Observing that the transfer function (1 z 2 )d+1 is a function of

z 2 , and using the noble identity for upsampling, we get the final
diagram.
- (1 z 1 )d+1

x(n)


w(n) 2


Q(z)

v(n)

As explained above, if x(n) Pd , then w(n) = 0 and therefore

v(n) = 0. Because v(n) = 0, we know that y(n) Pd .
In other words, if H(z) = Q(z) (1+z 1 )d+1 then when it is used for
the 2X interpolation, it preserves Pd , the set of polynomial signals
of degree d.
H(z) = Q(z) (1 + z 1 )d+1

h(n) [2] x(n) preserves Pd

It should be said that in the interpolation structure above, even if H(z) is chosen
so that y(n) is ensured to be a polynomial signal of degree d like x(n) is, it does
not mean that y(2n) = x(n). That is only true when the filter H(z) is in addition
a half-band filter, as discussed above.
I. Selesnick

EL 713 Lecture Notes

31

POLYNOMIAL INTERPOLATION BY L
How should the condition above be modified if we are interpolating
by a factor L rather than just by a factor of 2? If we guess that
H(z) should be of the form H(z) = Q(z) R(z) and follow the same
procedure used above, we will see that we will want the product
R(z) (1 z 1 )d+1 to be equal to (1 z L )d+1 . For in that case,
we could again exchange the order of the (L-fold) upsampler and
this term. This gives
R(z) (1 z 1 )d+1 = (1 z L )d+1
or
id+1
(1 z L )d+1 h
1
2
(L1)
= 1 + z + z + + z
R(z) =
(1 z 1 )d+1
where we have used
(1 z) (1 + z + z 2 + + z L1 ) = 1 z L .

H(z) = Q(z) 1 + z

+z

+ + z

(L1)

id+1

h(n) [L] x(n) preserves Pd

It should be said that in the LX interpolation structure, even if

H(z) is chosen so that y(n) is ensured to be a polynomial signal of
degree d like x(n) is, it does not mean that y(Ln) = x(n). That
is only true when the filter H(z) is in addition a Nyquist-L filter.

I. Selesnick

EL 713 Lecture Notes

32