Supplementary Notes for ELEN 4810 Lecture 12

More Z-transform
John Wright
Columbia University
November 13, 2017

Disclaimer: These notes are intended to be an accessible introduction to the subject, with no pretense at com-
pleteness. In general, you can find more thorough discussions in Oppenheim’s book. Please let me know if
you find any typos.
Reading suggestions: Oppenheim and Schafer Sections 3.3 - 3.6

In this lecture, we continue our discussion of the Z-transform, focusing on issues that arise when
we try to invert the transform to recovery x[n].

1 Inverting the Z-transform

In this section, we describe approaches to computing the inverse Z transform. The most conceptu-
ally straightforward approach is to use the following inversion formula, which asserts that x[n] can
be obtained by integrating X(z) over any closed contour C in the ROC:
I
1
x[n] = X(z)z n−1 dz. (1.1)
2πj C

In this formula, C can be any closed contour which (i) lies entirely in the ROC, (ii) encircles 0, and
(iii) is oriented in a counterclockwise direction.
This formula looks intimidating. To make it more comfortable, consider the case π in which the
 jω jω 
ROC contains the unit circle e | ω ∈ R . Taking the particular contour C = e  , and noting
−π
that for z = ejω , dz = jejω dω, we can express the contour integral as
Z π
1
x[n] = X(ejω )ejωn dω, (1.2)
2π −π

which is just the familiar expression for the inverse DTFT. When the DTFT does not exist (i.e., the
unit circle is not in the ROC), we can still invert the Z-transform by computing a similar integral
around a different circle which does lie in the ROC.

1
 It may seem strange that in the inversion formula (1.1), C can be any closed contour that encloses
zero and lies inside the ROC. The fact that the particular choice of contour does not matter follows
from a remarkable result in complex analysis known as the Cauchy integral theorem. Proving this is
beyond our scope. However, we will briefly justify (1.1) based on the related Cauchy integral formula,
which asserts that I (
1 1 k = 1,
z −k dz = (1.3)
2πj C 0 else.
Based on this, consider the right hand side of (1.1):
I I X
1 1
X(z)z n−1 dz = x[k]z n−k−1 dz
2πj C 2πj C
k
I
X 1
= x[k] z n−k−1 dz
2πj C
k
X
= x[k]δ[n − k]
k
= x[n]. (1.4)

Above, the interchange of integration and summation is justified by the fact that C lies within the
ROC, and hence the summation converges absolutely.

Easier ways to deduce the inverse. In practice, contour integration is complicated, and it is desir-
able to avoid it when possible. The most obvious way around it, called the “inspection method” in the
text, is to use a table of known Z transforms in conjunction with the properties of the Z-transform
to try to guess the inverse.
A more structured approach is to use partial fraction expansion to write X(z) as a superposition
of simpler functions, and then invert each of these simpler functions individually. For our applica-
tions, we are almost exclusively interested in Z-transforms with a rational functional form. In this
situation, we can use polynomial division and partial fraction expansion to express Z as a super-
position of simpler functions. Since the Z transform is linear, we can then invert it by summing the
inverses of these simpler functions. To make this more concrete, let us assume that
PM −k
k=0 bk z
X(z) = PN (1.5)
−`
`=0 a` z
QM
b0 k=1 (1 − ζk z −1 )
= QN (1.6)
a0 `=1 (1 − ρ` z −1 )
−1
P (z )
= . (1.7)
Q(z −1 )

We have chosen here to express X(z) as a quotient of polynomials in z −1 , rather than polynomials in
z. This is for consistency with the text, which mostly describes rational Z-transforms as polynomials
in z −1 . However, everything we will say about this situation translates into equivalent statements
about polynomials in z, with very minor modifications.

2
 In (1.6), the ζk are the zeros of X(z), while the ρ` are the poles. P is a polynomial of degree M ;
Q is a polynomial of degree N . We first show how to express X(z) as a sum of simpler terms when
M < N and the poles ρ1 , . . . , ρN are distinct:
Proposition 1.1 (Partial fraction expansion, distinct poles). Suppose that X(z) has the form (1.6), with
M < N and ρ1 , . . . , ρN distinct. Then
N
X A`
X(z) = . (1.8)
1 − ρ` z −1
`=1

with 
A` = (1 − ρ` z −1 )X(z) 

. (1.9)
z=ρ`

Proof. By placing everyone the over a common denominator, we obtain

n PN
− ρk z −1 )
Q
X A` i=1 Ai k6=i (1
= Qn . (1.10)
1 − ρ` z −1 `=1 (1 − ρ` z −1 )
`=1

Qnthe above expression consist of polynomials of degree at most N − 1 divided by the

Both X(z) and
polynomial `=1 (1 − ρ` z −1 ). It is a basic fact in algebra that if two polynomials of degree at most
N − 1 agree at N distinct points, they are equal everywhere.1 Our choice of coefficients A` ensures
PN
that the two expressions agree at the N points ρ1 , . . . , ρ` , and hence that X(z) = `=1 1−ρA``z−1 .

The above proposition gives a very simple way of expanding certain rational functions X(z) as
sums of simpler functions, which are then easily inverted. However, it requires that in the expression
X(z) = P (z −1 )/Q(z −1 ), deg(P ) < deg(Q). If this condition does not hold, we can use polynomial
long division to write X(z) as a sum of a polynomial in z −1 and a rational function R(z −1 )/Q(z −1 )
with deg(R) < deg(Q):
−1
P (z )
Proposition 1.2 (Polynomial division). Suppose X(z) = Q(z −1 ) , with P and Q polynomials. Then there

is a unique expression
R(z −1 )
X(z) = D(z −1 ) + , (1.11)
Q(z −1 )
with R, D polynomials and deg(R) < deg(Q).
Proof. To show that such a representation exists, we give an algorithm for constructing it. For an
Pd
arbitrary polynomial H(z −1 ) = `=0 h` z
−`
, let lead(H) = hd z −d denote the leading term. The
algorithm is as follows:
Set D = 0, R = P .

while deg(R) ≥ deg(Q),

lead(R)
Set D = D + lead(Q) .

1 This can be deduced from the fact that the Vandermonde matrix constructed from N distinct points has full rank N .

3
 lead(R)
Set R = R − lead(Q) Q.

end while
To show that this algorithm produces a representation of the desired form, note that because at
each iteration deg(R) ≥ deg(Q), the term lead(R)/lead(Q) is always a polynomial in z −1 . Hence,
D remains a polynomial in z −1 , as does R. Moreover, the degree of R decreases by at least one in
each iteration. Hence, after a certain number of iterations, the algorithm terminates with D and R
polynomials in z −1 and deg(R) < deg(Q).
To show that this representation is unique, note that if X(z) = D0 (z −1 ) + R0 (z −1 )/Q(z −1 ) is
another factorization of the same form, then

Q(z −1 )(D − D0 )(z −1 ) = (R0 − R)(z −1 ). (1.12)

The left hand side is either zero (if D0 = D) or has degree at least deg(Q). The right hand side has
degree strictly smaller than deg(Q), and so the only possibility is that D0 = D, which then implies
that R0 = R.
The proof of the above proposition gives a very practical way for computing D and R, which
you may recognize as polynomial long division. If we are confronted with a function X(z) of the
form QM
b0 i=1 (1 − ζi z −1 )
X(z) = QN , (1.13)
a0 `=1 (1 − ρ` z −1 )
with ρ1 , . . . , ρN distinct, but M ≥ N , we can use Proposition 1.2 and Proposition 1.1 to express X(z)
as
R(z −1 )
X(z) = D(z −1 ) + QN (1.14)
`=1 (1 − ρ` z −1 )
N
X A`
= D(z −1 ) + , (1.15)
1 − ρ` z −1
`=1

which can be inverted piece-by-piece.

The final generalization needed is to consider rational X(z) in which there may be repeated
poles. In this case, the form of the partial fraction expansion becomes slightly more complicated:
Proposition 1.3. Suppose that

P (z −1 ) P (z −1 )
X(z) = −1
= QU , (1.16)
Q(z ) `=1 (1 − ρ` z
−1 )d`

where deg(P ) < deg(Q), U is the number of unique (distinct) roots of Q, and ρ1 , . . . , ρU are distinct. Then
d
U X̀
X A`,k
X(z) = , (1.17)
(1 − ρ` z −1 )k
`=1 k=1

with
d(d` −k)
 
1 d` −1 
A`,k = (1 − ξ` w) X(w )  (1.18)
(d` − k)!(−ρ` )d` −k dwd` −k w=ρ−1
`

4
An example. The following example is taken from the text. We consider

z −2 + 2z −1 + 1
X(z) = 1 −2 , |z| > 1. (1.19)
2z − 32 z −1 + 1

Factoring the numerator and denominator, we obtain

(1 + z −1 )2
X(z) = . (1.20)
(1 − 12 z −1 )(1 − z −1 )

Thus, there are no repeated poles. However, the numerator and denominator have the same degree,
and so we cannot directly apply partial fraction expansion. Using polynomial long division, we
obtain that
5z −1 − 1
X(z) = 2 + (1.21)
(1 − 12 z −1 )(1 − z −1 )
We then express the second term as

A1 A2
1 −1 + 1 − z −1 (1.22)
1 − 2z

where
5z −1 − 1 
 
A1 = = −9 (1.23)
1 − z −1 z=1/2

 −1 
5z − 1 
A2 = = 8, (1.24)
1 − 12 z −1 z=1


and so
9 8
X(z) = 2 − + . (1.25)
1 − 21 z −1 1 − z −1
Since
Z
2δ[n] ←
→ 2 (1.26)
 n
1 Z 1
u[n] ←
→ , |z| > 1/2 (1.27)
2 1 − 12 z −1
Z 1
u[n] ←
→ , |z| > 1, (1.28)
1 − z −1
we conclude that  n
1
x[n] = 2δ[n] − 9 u[n] + 8u[n]. (1.29)
2

Power series expansion. The last trick for avoiding heavy calculation in the inverse P∞Z-transform
is to use power series expansion. Namely, we seek to express X(z) as a sum X(z) = n=−∞ αn z −n ,
and then read of the values x[n] = αn . This approach is perhaps best illustrated through an example.

5
 Let us suppose that X(z) = log(1 + az −1 ), with ROC |z| > |a|. We know that log(1 + s) has a
convergent Taylor series for |s| ≤ 1, namely,
∞
X (−1)n+1 sn
log(1 + s) = . (1.30)
n=1
n

Setting s = az −1 , we obtain
∞
X (−1)n+1 an z −n
log(1 + az −1 ) = , (1.31)
n=1
n

from which we read off

(−1)n+1 an
(
n n≥1
x[n] = . (1.32)
0 else