Chapter 2

Discrete-Time Sequences
and Systems

Contents
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 82
2.2 Sequences . . . . . . . . . . . . . . . . . . . . . . . . 84
2.3 Systems . . . . . . . . . . . . . . . . . . . . . . . . . 90
2.4 Analysis of Sequences and Systems . . . . . . . . . 98
2.5 Multirate Sequences and Systems . . . . . . . . . . 131
2.6 Stochastic Sequences and Systems . . . . . . . . . 147
2.7 Algorithms . . . . . . . . . . . . . . . . . . . . . . . 147
2.A Elements of Complex Analysis . . . . . . . . . . . . 149
2.B Elements of Algebra: Discrete Polynomials . . . . 151
Chapter at a Glance . . . . . . . . . . . . . . . . . . . . . . 153
Historical Remarks . . . . . . . . . . . . . . . . . . . . . . . 155
Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . 155
Exercises with Solutions . . . . . . . . . . . . . . . . . . . . 156
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159

The key word in the title of this chapter is time, and the key property of time
is that it has an inherent direction; it goes from past to future. The word discrete
brings about yet another key concept—the time points are equidistant. To move
from one time point to the next, we use the shift operator. If we consider vectors
now to be sequences (the domain is discrete time), appropriate Hilbert spaces of
such sequences might be ℓ1 (Z), ℓ2 (Z) and ℓ∞ (Z), all corresponding to real physical
phenomena (finite power, finite energy, boundedness). The shift operator can be
applied repeatedly to such sequences, forming new sequences. Moreover, it generates
a whole space of operators we call discrete-time systems, which typically belong to
ℓ1 (Z) or ℓ2 (Z). As we will see later in the chapter, a shift induces convolution,
the operation of the discrete-time system on a sequence. Once these concepts are
0 Last edit: JK Mar 04 08.

81
82 Chapter 2. Discrete-Time Sequences and Systems

available (shift, sequence and system spaces, convolution), spectral theory allows us
to construct an appropriate Fourier transform.
All of the above discussion implicitly assumed that the underlying domain—
time, was infinite. While this is true, in practice, we do not observe infinite time,
and thus, typically, we only glimpse a finite portion of it. While one might think
that dealing with finite time is mathematically easier, a host of issues arises. We
discuss these throughout the chapter.
The purpose of this chapter is the study sequences in their different guises and
review methods for analyzing and processing such sequences.

2.1 Introduction
While sequences in the real world are often one-sided infinite (they start at some
initial point and then go on), for mathematical convenience, we look at them as
two-sided infinite:6
 T
x = . . . x−2 x−1 x0 x1 x2 . . . . (2.1)

For example, if you measure some physical quantity every day at some fixed time
(for example, the temperature at noon in front of your house), you obtain a sequence
starting at time 0 (say January 14th) and continuing to infinity,
h iT
x = ... 0 32◦ 29◦ 30◦ ... . (2.2)

Implicit in the index is the fact that xn corresponds to the temperature (at noon)
on the nth day. A sequence is also known under the names discrete-time signal (in
signal processing) or time series (in statistics); mathematically, these are all vectors,
most often in an infinite-dimensional Hilbert space.
In real life, most things are finite, that is, we observe only a finite portion
of an infinite sequence. Moreover, computations are always done on finite inputs.
For our temperature example, serious observations started in the 18th century and
necessarily stop at the present time. Thus, we have a finite sequence of length N ,
 T
x = x0 x1 x2 . . . xN −1 . (2.3)

Once we have a finite sequence, the questions immediately arises: where does it
come from? Is it a truly finite sequence, or, is it a piece of an infinite one? In
either case, how do we analyze such a sequence? As we will see later in the chapter,
this question immediately calls for a decision on what happens at the boundaries.
Do we assume everything outside the sequence support is zero? Do we extend it
periodically, symmetrically, ...?
While perhaps not a natural extension, a typical one is periodic; We periodize
the finite sequence and assume we are dealing with one period of a periodic sequence
6 The“boxing” of the origin is intended to serve as a reference point. When we deal with finite

vectors/matrices, this is not necessary; however, when these are infinite, knowing where the origin
is, is essential.
2.1. Introduction 83

with period N ∈ N:
x = [. . . x0 x1 . . . xN −1 x0 x1 . . .]T . (2.4)
| {z }
one period

While finite sequences in (2.3) and infinite periodic ones in (2.4) are clearly different
in nature, we will use the same types of tools to analyze them. We will assume that
both belong to some space on the domain {0, 1, . . . , N − 1} (say ℓ1 or ℓ2 ), with
circular boundary conditions, implicitly making even the first type periodic. When
we want to analyze finite sequences as windowed versions of infinite sequences, we
will do just that; we will assume such sequence are infinite in nature with only a
finite number of nonzero samples.
These considerations allow us to define two broad classes of sequences:
(i) Those with support in Z we will call infinite dimensional ; and
(ii) Those with support in {0, 1, . . . , N − 1} we will call finite dimensional.
Example 2.1 (Sequences). As an illustration:
• Infinite-Dimensional Sequences: The geometric sequence from (P1.20-1) with
x = 21 :
1 h iT
xn = ( )n , n ∈ Z, x = . . . 4 2 1 12 14 . . . , (2.5)
2
is infinite dimensional and not convergent (if we made it nonzero only for
n ≥ 0, it would be).
• Finite-Dimensional Sequences: The sequence obtained by tossing a fair coin
N times is finite dimensional:
h iT
x = h h t ... h .

A sinusoidal sequence sampled at submultiples of its period is periodic with

period N , and is another example of a finite-dimensional sequence:
 
2π
xn = sin n+θ ,
N
   
2π 2π
x = [. . . sin (θ) sin + θ . . . sin (N − 1) + θ sin (θ) . . .]T ,
N N
| {z }
one period

Given sequences (signals, vectors), one can apply operators (systems, filters). These
map input sequences into output ones, and since they involve discrete-time se-
quences, they are usually called discrete-time systems (operators). Among them,
linear ones are most common. Even more restricted is the class of linear shift-
invariant systems (filters, defined later in the chapter), an example of which is the
moving average filter:
84 Chapter 2. Discrete-Time Sequences and Systems

Example 2.2 (Moving average filter). Consider our temperature example, and
assume we want to detect seasonal trends. The day-by-day variation might be too
noisy, and thus, we can compute a local average
N
X
1
yn = xn+k , n ∈ Z, (2.6)
2N + 1
k=−N

where N is a small integer. The “local” average eliminates daily variations; this
simple filter is linear and shift invariant, since at all n, the same local averaging is
performed.

Chapter Outline
The next several sections follow the progression of topics in this brief introduction:
We start by formally defining the various types of sequences we discussed above.
Section 2.3 considers linear discrete-time systems, especially of the shift-invariant
kind, which correspond to difference equations, the discrete-time analog of differen-
tial equations. Next, in Section 2.4, we develop the tools to analyze discrete-time
sequences and systems, in particular the discrete-time Fourier transform, the dis-
crete Fourier transform and the z-transform. We discuss the fundamental theorem
relating filtering to multiplication in Fourier domain—the convolution theorem.
Section 2.5 looks into discrete-time sequences and systems that operate with dif-
ferent rates—multirate systems, which are key for filter bank development in later
chapters. Then, we present discrete-time stochastic sequences and systems in Sec-
tion 2.6, while important algorithms for discrete-time processing, such as the fast
Fourier transform, are covered in Section 2.7. Appendix 2.A lists basic elements of
complex analysis, while Appendix 2.B discusses polynomial sequences.

2.2 Sequences
2.2.1 Infinite-Dimensional Sequences
The set of sequences as in (2.1), where xn is either real or complex, together with
vector addition and scalar multiplication, forms a vector space (see Definition 1.1).
The inner product between two sequences is defined as in (1.14), and induces the
standard ℓ2 , Euclidean norm given in (1.21a), with p = 2. Other norms of interest
are the ℓ1 -norm given in (1.21a) with p = 1, and the ∞-norm given in (1.21b).
As opposed to generic infinite-dimensional spaces, where ordering of indices
does not matter in general, discrete-time sequences belong to an infinite-dimensional
space where the order of indices is important, since it represents time (or space in
two dimensions).

Sequence Spaces
Space of Finite-Energy Sequences ℓ2 (Z) The constraint of a finite square norm
is necessary for turning a simple vector space into a Hilbert space of finite-energy
sequences ℓ2 (Z).
2.2. Sequences 85

Space of Finite-Power Sequences ℓ∞ (Z) A looser constraint is finite power, which

only bounds the magnitude of the samples. This space, which contains all sequences
xn such that |xn | < ∞, is denoted ℓ∞ (Z), since the ℓ∞ -norm is finite. While
in practice many sequences are of the latter kind, the former is more convenient
because of the Euclidean geometry associated with Hilbert spaces.

Space of Absolutely-Summable Sequences ℓ1 (Z) One more important class of

sequences are absolutely-summable sequences, that is, sequences belonging to ℓ1 (Z).
We briefly touched upon those in Section 1.2.3 (remember that ℓ1 (Z) ⊂ ℓ2 (Z)). By
definition, sequences in ℓ1 (Z) have a finite ℓ1 -norm.

Example 2.3 (Sequence spaces). Revisiting the sequences seen in Example 2.1,
we see that xn = αn , this time zero for n < 0, α ∈ R, is in the following spaces:

|α| < 1 ℓ2 (Z), ℓ1 (Z), ℓ∞ (Z),

|α| = 1 ℓ∞ (Z)
|α| > 1 none.

Geometry in ℓ2 (Z) We now recall a few “geometrical” facts from Chapter 1, valid
for sequences in ℓ2 (Z):
(i) As in (1.15),
hx, yi = kxk kyk cos θ,
where θ is the angle between the two infinite sequences (vectors).
(ii) As in (1.12), if the inner product is zero,

hx, yi = 0,

then either at least one of the sequences is identically zero, or they are said to
be orthogonal to each other.
(iii) As in (1.32), given a unit-norm sequence y, kyk = 1,

x̂ = hx, yi y

is the orthogonal projection of the sequence x onto the space spanned by the
sequence y.

Special Sequences
Dirac impulse Sequence The simplest and one of the most useful sequences is
the Dirac impulse sequence at location k (it is also defined in the Quick Reference
section at the beginning of the book)

1, for n = k;
δn−k = 1 . . .]T ,
δn−k = [. . . 0 0 . . . |{z} (2.7)
0, otherwise,
n=k
86 Chapter 2. Discrete-Time Sequences and Systems

where n, k ∈ Z. The set of Dirac impulses {δn−k }k∈Z along the discrete timeline
forms an ONB for ℓ2 (N). We can easily check that as (a) it is clearly an orthonormal
set, and (b) any sequence from ℓ2 (N) can be written as a linear combination of the
basis vectors ϕkn = δn−k :
!
X X X X
xn = hxn , ϕkn in ϕkn = xn δn−k δn−k = xk δn−k . (2.8)
k∈Z k∈Z n∈Z k∈Z

This identity, while seemingly trivial, might require a moment of thought: for any
n, a single term in the infinite sum that is different from zero is the one for k = n,
when the sum equals . . . xn ! A list of properties of this sequence is given in Chater
at a Glance.

Sinc Sequence With n ∈ Z, the sinc sequence is given by:

n n
sin Tn π ej T π − e−j T π
sincT,n = n = . (2.9)
Tπ 2j( Tn π)

The zeros nk of a discrete-domain sinc function are at:

nk
π = kπ, → nk = kT, k ∈ Z, (2.10)
T
shown in Fig. 1.12(b). As opposed to the continuous-time sinc (1.65), the zeros of
the discrete-time sinc sequence will not appear at kT for all k, but just at those kT
which are integers. An example for T = 2 gives
sin n2 π h iT
sincn = n , sinc = . . . − π2 0 2
π 1 2
π 0 − π2 . . . . (2.11)
2π

This sequence is in ℓ∞ (Z) and in ℓ2 (Z), but not in ℓ1 (Z), since the sequence decays
as 1/n (see the illustration of the inclusion property (1.22) of ℓp spaces). The value
of sinc at t = 0 is obtained the same way as for the continuous-domain sinc, leading
to:
sincT,0 = 1. (2.12)

Heaviside Sequence The Heaviside sequence is defined as

 h iT
1, n ∈ N;
un = u = ... 0 1 1 ... , (2.13)
0, otherwise,

that is, a unit-step sequence at the origin. As this sequence is bounded, it belongs
to ℓ∞ (Z), but not to ℓ1 (Z) or ℓ2 (Z). The two sequences we introduced so far—Dirac
and Heaviside—are related (via “integration” or “differentiation”):
n
X
un = δk .
k=−∞
2.2. Sequences 87

Figure 2.1: A sinusoidal sequence and its windowed versions using a (a) rectangular
window and a (b) raised cosine window.

Window Sequences Most often, both because of real-world considerations and for
mathematical convenience, a finite-length sequence is seen as a glimpse of an infinite
sequence. One way to state this is by introducing a rectangular window sequence
as the indicator sequence of {0, 1, . . . , N − 1},

1, 0 ≤ n ≤ N − 1;
wn = w = [. . . 0 1 1 . . . 1 0 . . .]. (2.14)
0, otherwise, | {z }
N

Then, we can define an infinite sequence y with values given by (2.3) for n ∈
{0, 1, . . . , N − 1}, and arbitrary otherwise,

xn , 0 ≤ n ≤ N − 1;
yn = y = [. . . y−1 x0 x1 . . . xN −1 yN . . .].
arbitrary, otherwise, | {z }
N

Multiplying yn with the window wn , we obtain the projection ŷ:

ŷn = yn wn , ŷ = [. . . 0 x0 x1 . . . xN −1 0 . . .], (2.15)
| {z }
N

which equals xn for n ∈ {0, 1, . . . , N − 1} and is zero otherwise. This allows us

to study an infinite sequence ŷn that coincides with xn looking at it through the
window of interest.
How good is the window we just used? For example, if y is smooth, its
projection ŷ is not because of the abrupt boundaries of the rectangular window.
We might thus decide to use a different window to smooth the boundaries.
Example 2.4 (Windows). Consider an infinite sinusoidal sequence of frequency
ω0 and phase θ0 ,
yn = sin(ω0 n + θ0 ),
and the following two windows (in both cases N = 2K + 1):
(i) Rectangular window of length N .
(ii) A raised cosine window:
 π
cos( K (n − K)) + 1, 0 ≤ n ≤ 2K;
wn = (2.16)
0, otherwise.
The raised cosine window tapers off smoothly at the boundaries, while the rectangu-
lar one does not. The trade-off between the two windows is obvious from Figure 2.1;
The rectangular window does not modify the sequence inside the window, but leads
to abrupt changes at the boundary, while the raised cosine window has smooth
transitions, but at the price of modifying the sequence.
88 Chapter 2. Discrete-Time Sequences and Systems

2.2.2 Finite-Dimensional Sequences

As we said, finite-dimensional sequences as in (2.3) are those with the domain in
{0, 1, . . . , N − 1}, and can be either seen as truly finite or as a period of a periodic
sequence (their domain is a “discrete” circle). Once we start applying operators on
such sequences, we start worrying about whether the space they belong to is closed.
Operations such as convolution would clearly push elements out of the domain of
support, and we must consider how to extend such sequences at the boundaries. One
possible extension, and the one that is used in discrete-time signal processing, is
circular; the sequence is extended periodically, and in effect, the domain becomes the
discrete circle (operations mod N ). We can then denote the space of such sequences
as RN or CN .7 Note that we could have chosen another extension; however, this is
not typically done. More details on this topic can be found in Further Reading.
As we will see in Section 2.4.3, a natural basis for such sequences are the har-
monic complex exponential sequences, that is, N sequences8 ϕk , k ∈ {0, 1, . . . , N −
1}, given by
2πk
ϕkn = ej N n , k ∈ {0, 1, . . . , N − 1}, n ∈ Z. (2.17)

These sequences are periodic with period N , and their linear combinations span the
space of all N -periodic sequences. In Exercise 2.1, we explore a few properties of
complex exponential sequences and their periodicity.

2.2.3 Two-Dimensional and Multidimensional Sequences

Two-Dimensional Sequences
Today, one of the most widespread devices is the digital camera, and the size of
the acquired image in megapixels9 is the topic of intense marketing battles. In our
notation, a digital picture is a two-dimensional sequence, xnm , and can be seen
either as an infinite-dimensional sequences with a finite number of nonzero samples
(example of a window view of the world, that is, a finite two-dimensional sequence
being part of a larger, but invisible sequence):

xnm , n, m ∈ Z, (2.18)

or as a sequence with domain n ∈ {0, 1, . . . , N − 1}, m ∈ {0, 1, . . . , M − 1},

 
x00 x01 ... x0,M−1
 x10 x11 ... x1,M−1 
 
x =  . . .. .. , (2.19)
 .. .. . . 
xN −1,0 xN −1,1 . . . xN −1,M−1
7 Wecould denote such a space as ℓ(ZN ) to keep the awarness of the boundary conditions; we
choose to go with a simpler, RN /CN notation.
8 In the double subscript, the first, k, denotes the sequence, and the second, n, the domain of

the sequence.
9 A megapixel is either a true set of 1000 by 1000 pixels, or more often a combination of 500 000

green and 250 000 red and blue pixels.

2.2. Sequences 89

assumed to be circularly extended at the borders (similarly to the one-dimensional

case). Each element xnm is called a pixel, the total image having N M pixels. In
reality, for xnm to represent a color image, it must have more than one component;
often, red, green and blue components are used (RGB space). While perhaps such
an extension is not natural (the left side of the image appears at the right boundary),
it is the one that leads to the use of the DFT, as we will see later in the chapter.

Sequence Space Symbol Finite Norm

X
Absolutely-summable ℓ1 (Z2 ) kxk1 = |xnm |
n,m∈Z
0 11/2
X
2A
Finite-energy ℓ2 (Z2 ) kxk = @ |xnm |
n,m∈Z
Absolutely summable ℓ∞ (Z2 ) kxk∞ = max |xnm |
n,m∈Z

Table 2.1: Norms and two-dimensional sequence spaces.

Sequence Spaces The spaces we introduced in one dimension generalize to mul-

tiple dimensions; for example in two dimensions, the inner product, given two se-
quences x and y, is XX
∗
hx, yi = xnm ynm , (2.20)
n∈Z m∈Z

while the ℓ -norm and the appropriate space ℓ2 (Z2 ) are given in Table 2.1 (together
2

with other relevant norms and spaces). For example, a digital picture, having finite
size and pixel values that are bounded, clearly belongs to all three spaces defined in
Table 2.1. Infinite multidimensional sequences, on the other hand, can be trickier
to analyze.
Example 2.5 (Norms of two-dimensional sequences). Consider the sequence
1
xnm = , n, m ∈ N.
2n 3m
Its ℓ2 -norm can be evaluated as
XX 1 1 4 X 1 4 9 3
hx, xi = n m
= = = , (2.21)
4 9 3 9m 3 8 2
n∈N m∈N m∈N
p
yielding kxk = 3/2. Similarly, kxk1 = 3 and kxk∞ = 1.

Three-Dimensional Sequences
Higher-dimensional objects are also possible; we briefly consider the three-dimensional
case, video, for example. A video sequence is a sequence of digital pictures we rep-
resent as:
xnmk , n ∈ {0, 1, . . . , N − 1}, m ∈ {0, 1, . . . , M − 1}, k ∈ Z, (2.22)
90 Chapter 2. Discrete-Time Sequences and Systems

Figure 2.2: Multidimensional data sets. (a) Two-dimensional sequence (image), with the
first quadrant indicated, as well as a finite N × M digital picture. (b) Three-dimensional
sequence (video).

Figure 2.3: A discrete-time system.

where N × M is the size of the individual picture in the sequence, while k is a time
index (assumed here to be doubly infinite). The size of multidimensional sequences
grows as the product of the sizes in each dimension. For example, a video sequence
in (2.22) with k ∈ {0, 1, . . . , K − 1}, has N M K pixels, which quickly leads to
enormous amounts of data. Figure 2.2 gives examples of two- and three-dimensional
sequences.

2.3 Systems
Discrete-time systems are operators having discrete-time sequences as inputs and
outputs. Among all discrete-time systems, we will concentrate on the subclass of
linear systems, in particular, linear and shift-invariant (LSI) systems. This last
class is both important in practice and amenable to easy analysis. The moving
average filter in (2.6) is such an LSI system. After an introduction to difference
equations, which are natural descriptions of discrete-time systems, we study LSI
systems in detail. We then turn our attention to shift-varying systems and an
important subclass, linear periodically shift-varying systems.

2.3.1 Discrete-Time Systems and Their Properties

A discrete-time system maps an input sequence xn ∈ H into an output sequence
y ∈ H through an operator T (see Figure 2.3)

yn = Tn x. (2.23)

The subscript n in Tn indicates that the operator may be changing over time. While
in the above, Tn operates only on x, in general, it can operate both on x and y.
While this may seem self-referential, it can be done as long as yn is computable.
As a discrete-time system Tn is an operator, everything we learned in Chap-
ter 1 about operators (in particular, linear operators) holds here. In general, oper-
ators map a sequence from H1 to H2 , while here, we stay in the same Hilbert space
H, space of sequences.
2.3. Systems 91

Basic Systems
Shift The shift-by-k operator is defined as:

yn = T x = xn−k , (2.24a)
   
. . . x−1 x0 x1 . . . → . . . xk−1 xk xk+1 ... . (2.24b)

which simply delays xn by k samples. While this is one of the simplest discrete-time
systems, it is also the most important, as the whole concept of time processing is
based on this simple operator. The shift-by-one is usually termed a delay.

Modulation While the shift we just saw is the shift in time, modulation is shift
in frequency (as we will see later in this chapter). A modulation by a sinusoid of
frequency ω0 , is given by (we use the cosine here)

yn = T x = cos(ω0 n) xn . (2.25)

For those already familiar with Fourier analysis, (2.25) shifts the spectrum of x to a
position ω0 in frequency (and to a corresponding negative frequency −ω0 ). Imagine
xn = α, for all n. Assume now ω0 = π. The corresponding cos(ω0 n) becomes
cos(πn) = (−1)n . Thus, a constant sequence xn turns into a fast-varying sequence
xn = (−1)n α (shifted in frequency to higher frequency):
   
... α α α α α ... → . . . α −α α −α α . . . .

Hard Limiter Given a bound ∆, the hard limiter computes the following output:

 ∆, xn ≥ ∆;
yn = T x = −∆, xn ≤ −∆; (2.26)

xn , otherwise,
h iT
[. . . x−2 x−1 x0 x1 x2 . . .]T → . . . x−2 ∆ ∆ x1 −∆ . . . .
|{z} |{z} |{z}
≥∆ ≥∆ ≤∆

Accumulator The output of the accumulator is the “integral” of the input:

n
X
yn = T x = xk .
k=−∞

Linear Systems
Definition 2.1 (Linear system). A discrete-time system is linear if Tn is linear,
that is, if it satisfies Definition 1.16.

For the examples above, it is easy to verify that except for the hard limiter, the
systems are linear.
92 Chapter 2. Discrete-Time Sequences and Systems

Memoryless Systems
Definition 2.2 (Memoryless system). A system y = Tn x is memoryless, if its
output yn at time n depends only on the input xn at time n:

yn = T xn . (2.27)

In the examples seen previously, the modulation and the hard limiter are memory-
less, while the two others are clearly not.

Shift-Invariant Systems
Definition 2.3 (Shift-invariant system). Given an input xn and an output
yn , a system is shift invariant if a shifted input produces a shifted output:

yn = T xn ⇒ yn−k = T xn−k . (2.28)

The above implies that the operator Tn in (2.23) cannot be dependent on time,
and thus Tn = T . Shift invariance is often a very desirable property. For example,
listening to a piece of music on an MP3 player should produce the same music the
first as well as the seventh time. Moreover, shift-invariant systems have desirable
mathematical properties, especially when they are also linear. Among the examples
studied earlier, only modulation is not shift invariant.

Causal Systems
Definition 2.4 (Causal system). A causal system is a system whose output at
time n depends only on the present and past inputs:

yn = Tn xk , k ≤ n. (2.29)

We will see once we define the impulse response of a system, that causality implies
that the impulse response hn satisfies hn = 0 for all n < 0. If the domain of our
function is time, then any real-world system will necessarily be causal. There are
no real systems where the output can depend on future inputs, unless the input is
completely deterministic and known in advance (for example, it is a constant). Note
that causality is a property of discrete-time systems and not of sequences. However,
by abuse of language, we often refer to sequences which are nonzero only for positive
n as causal sequences, meaning they could be impulse responses of causal systems.
Often times this can lead to confusion since we will try to interpret the statement
of causality as “there is no future” and thus will think that the past in the sequence
is expressed on the right side of the origin. Of course, this is not true, as a sample
at (n + 1) definitely comes after the sample at n.

Stable Systems
While there are various definitions of stability for systems, we consider bounded
input bounded output (BIBO) stability exclusively.
2.3. Systems 93

Definition 2.5 (BIBO stability). A system is BIBO stable if, for every bounded
input x ∈ ℓ∞ (Z), the input y = Tn x is bounded as well:

x ∈ ℓ∞ (Z) ⇒ y ∈ ℓ∞ (Z). (2.30)

In the four example systems seen at the start of the section, the first three are BIBO
stable, while the accumulator is not, since, for example, a constant input xn = 1
will lead to an unbounded output.

2.3.2 Difference Equations

An important class of discrete-time systems can be described by linear difference
equations that relate the input sequence and past outputs to the current output,10

X ∞
X
(n) (n)
yn = ak xn−k + bk yn−k .
k∈Z k=1

(n) (n)
If we require shift invariance, then the coefficients ak and bk are constant, and
we get a linear, constant-coefficient difference equation. To make the system causal,
we restrict the inputs xn to the current and past inputs, leading to
∞
X ∞
X
yn = ak xn−k + bk yn−k . (2.31)
k=0 k=1

Realizable systems will have a finite number of coefficients ak , k ∈ {0, 1, . . . , M }

and bk , k ∈ {1, 2, . . . , N }, leading to
M
X N
X
yn = ak xn−k − bk yn−k . (2.32)
k=0 k=1

To find the solution to such a difference equation, we follow a similar path as for
ordinary differential equations.

• Homogeneous solution: First, we find a solution to the homogeneous equation,

N
X
yn(h) = bk yn−k , (2.33a)
k=1

by setting the input x in (2.32) to zero. The solution is of the form

N
X
yn(h) = βk λnk , (2.33b)
k=1

10 In this discussion, we concentrate on sequences where the index n corresponds to time, and

one solves the system forward in time, as is natural in causal systems. A general treatment of
difference equations involves also anti-causal solutions, obtained by solving backward in time.
94 Chapter 2. Discrete-Time Sequences and Systems

where λnk are obtained by solving the characteristic equation of the system
N
X
bk λN −k = 0. (2.33c)
k=1

(p)
• Particular solution: Then, any particular solution to (2.32), yn , is found
(h) (h)
(independent of yn ). This is typically done by assuming that yn is of the
same form as xn , possibly scaled.

• Complete solution: By superposition, the complete solution yn is the sum of

(h) (p)
the homogeneous solution yn and the particular solution yn :
N
X
yn = yn(h) + yn(p) = βk λnk + yn(p) . (2.34)
k=1

(h)
We determine the coefficients βk in yn by specifying initial conditions for yn
and then solving the system.

Example 2.6. As an example, consider the accumulator seen previously,

n
X n−1
X
yn = xk = xn + xk = xn + yn−1 ,
k=−∞ k=−∞

which is of the form (2.32), with a0 = 1, b1 = 1. The infinite sum has been turned
into a recursive formula, which now requires initial conditions. For example, assume
the system starts at n = 0, with initial condition y−1 given:
n
X
yn = y−1 + xk , n ∈ N.
k=0

From the above example, we see that unless the initial conditions are zero, the
system is not linear. Similarly, the system is shift invariant only if the initial
conditions are zero. These properties are fundamental and hold beyond the case
of the accumulator: difference equations as in (2.32) are linear and shift invariant
if and only if initial conditions are zero. This also means that the homogeneous
solution is necessarily zero (the proof is left as Exercise 2.2).

2.3.3 Linear Shift-Invariant Systems

Impulse Response
Consider the linear shift-invariant (LSI) difference equation (2.31) with zero initial
conditions. Assume a single nonzero input at the origin of time, δn , solve yn for
2.3. Systems 95

n ∈ N, and call this solution hn :

∞
X ∞
X ∞
X
hn = ak δn−k + bk hn−k = an + bk hn−k , n ∈ N. (2.35)
k=0 k=1 k=1

This is called the impulse response of the LSI difference equation and can be seen
as the output (response) of the system to the Dirac impulse input.

Convolution
If the input is an impulse at time ℓ, or δn−ℓ , then the output is delayed by ℓ:
∞
X ∞
X ∞
X
yn = ak δn−ℓ−k + bk yn−k = an−ℓ + bk yn−k , = hn−ℓ ,
k=0 k=1 k=1

for n => ℓ, since, by shift invariance, it is the solution of the difference equation,
but starting at time ℓ. We can use the superposition principle, since the system is
linear, and use the decomposition of the input sequence seen in (2.8) but now for
two-sided sequences: X
xn = xk δn−k ,
k∈Z

to write the output of the difference equation as a superposition of impulse re-

sponses, properly weighted and delayed, as
X X
yn = xk hn−k = xn−k hk = (h ∗ x) = T x. (2.36)
k∈Z k∈Z

This is the convolution 11 of the input sequence xn with the impulse response hn of
the system, denoted by (h ∗ x) (symmetric in xn and hn ). When not obvious, we
will use subscript to denote theP argument over which we perform the convolution
(2.36). For example, the sum k xk hn−k−ℓ will be (xk ∗ hn−k−ℓ )k .
Example 2.7 (Solution of LSI difference equations). Consider the differ-
ence equation
yn = a0 xn + b1 yn−1 , n ∈ N,
with zero initial conditions, y−1 = 0. The impulse response, or solution for the
input x = δn , is
 T
h = . . . 0 a0 b1 a0 b21 a0 b31 a0 . . . .
If the input is delayed by k samples, x = δn−k , the output is

y = [. . . 0 0 0 . . . 0 a0 b1 a0 b21 a0 . . .],
| {z }
k

or, hn−k . Using superposition, we get the convolution formula as in (2.36).

11 Convolution is sometimes called linear convolution to distinguish it from the circular convo-

lution of finite sequences.

96 Chapter 2. Discrete-Time Sequences and Systems

Filters The impulse response plays a key role (2.35); an LSI difference equation
(with zero initial conditions) is entirely specified by its impulse response. The
impulse response is often called a filter and the convolution is called filtering. Here
are some basic classes of filters:
• Causal filters: are such that hn = 0, n < 0. This corresponds to a causal
solution of a difference equation.
• Anticausal filters: are such that hn = 0, n > 0.
• Two-sided filters: are neither causal nor anticausal.
• Finite impulse response filters (FIR): have only a finite number of coefficients
hn different from zero.
• Infinite impulse response filters (IIR): have an infinite number of nonzero
terms. For example, the impulse response in Example 2.7 is causal and IIR.

Matrix View of the Convolution Operator As we have shown in Appendix 1.C,

any linear operator can be expressed in matrix form. While this is not often done
when dealing with convolution, (2.36) can be expressed as:
   .. .. .. .. ..  
.. .. .. ..
. . . . . . . .
    . 
 y−2   . . . h0 h−1 h−2 h−3 h−4 . . .  x−2 
    
 y−1   . . . h1 h0 h−1 h−2 h−3 . . .  x−1 
    
 y 0  =  . . . h2 h 1 h0 h−1 h−2 . . .  
     x0  . (2.37)
 y1  ... h h h h h . . .   x 
   3 2 1 0 −1  1 
 y2  ... h h0 . . .  
   4 h3 h2 h1   x2 
.. .. .. .. .. .. .. .. .
..
. . . . . . . .
This again shows that linear discrete-time system, linear operator, filter and matrix
are all synonyms. The key elements in (2.37) are the time reversal of the impulse
response (in each row of the matrix, the impulse response goes from right to left),
and the Toeplitz structure of the matrix (each row is a shifted version of the previous
row, the matrix is constant along diagonals, see (1.87)). In Figure 2.4, an example
convolution is computed graphically, emphasizing the time reversal.

Stability We now discuss stability of LSI systems.

Proposition 2.1. An LSI system is BIBO stable if and only if its impulse response
is absolutely summable.

Proof. To prove sufficiency, consider an absolutely-summable impulse response h ∈

ℓ1 (Z), so khk1 < ∞. Consider a bounded input x ∈ ℓ∞ (Z), so kxk∞ < ∞. The
norm of the samples at the output can be bounded as follows:
 
  (b) X
(a) X  (c) X (d)
|yn | =  hk xn−k  ≤ |hk | |xn−k | ≤ kxk∞ |hk | = kxk∞ khk1 < ∞,
 
k∈Z k∈Z k∈Z
2.3. Systems 97

Figure 2.4: Example of the convolution between a sequence and a filter. (a) Sequence xn .
(b) Impulse response hn . (c)–(e) Various time-reversed or shifted versions of the impulse
response. (f) Output y = (h ∗ x).

where (a) follows from (2.36); (b) from the triangle inequality (Definition 1.7); (c)
from x ∈ ℓ∞ (Z); and (d) from h ∈ ℓ1 (Z), proving that y is bounded.
We prove necessity (BIBO stability implies absolute summability) by contra-
diction. For any hn that is not absolutely summable we choose a particular input
xn (that depends on hn ), to create an unbounded output. Consider a real impulse
98 Chapter 2. Discrete-Time Sequences and Systems

response12 hn , and define the input sequence to be


 −1, c < 0;
xn = sgn(h−n ), where sgn(c) = 0, c = 0;

1, c > 0,
is the sign function. Now, compute the convolution of xn with hn at time 0:
X X
y0 = hk x−k = |hk | = khk1 , (2.38)
k∈Z k∈Z\{0}

which is unbounded when hn is not in ℓ1 (Z).

Among examples we have seen so far, the accumulator does not belong to
ℓ1 (Z), and can lead to unbounded outputs. Example 2.7, for |b1 | < 1, leads to an
absolutely-summable geometric series, and thus a BIBO-stable filter. For |b1 | > 1,
any nonzero input leads to an unstable system. Lastly, for |b1 | = 1 (which means
+1 or −1 for b1 real), the system is not BIBO stable either.
Among spaces we considered for our filters, two are most often used: (a) ℓ1 (Z)
as it ensures BIBO stability and (b) CN (or RN ) when sequences on which we
operate are truly finite.

2.4 Analysis of Sequences and Systems

In this section, we introduce various ways to analyze sequences and discrete-time
systems. They range from the analytical to the computational, but they all are
variations of the Fourier transform. Why this prominent role of Fourier methods?
Quite simply because the Fourier sequences are eigensequences of LSI systems (con-
volution operator), and operating in these eigenspaces is a natural representation
for such systems. This leads immediately to the convolution theorem, which states
that the convolution operator is diagonalized by the Fourier transform.
The above is a part of a recipe (see Table in Chater at a Glance:
(i) Start with a given time shift δn−1 ;
(ii) Induce an appropriate convolution operator T x = (h ∗ x);
2πk
(iii) Find the eigensequences xn of T (ejωn for infinite time and ej N n for finite
time);
(iv) Identify the frequency response as the eigenvalue corresponding to the above
eigensequence (H(ejωn ) for infinite time, Hk for finite time);
(v) Find the appropriate Fourier transform by projecting on the eigenspaces spanned
by eigensequences identified in (iii) (discrete-time Fourier transform for infinite
time, discrete Fourier transform for finite time).
We start with the discrete-time Fourier transform (DTFT)—the Fourier trans-
form for infinite sequences (discrete-time sequences). The DTFT of a sequence xn is
12 For a complex-valued impulse response, a slight modification, using x = h∗ / |h | for |h | 6= 0,
n n n n
0, otherwise, leads to the same result.
2.4. Analysis of Sequences and Systems 99

a 2π-periodic function X(ejω ), with the argument, ejω , a complex number of mag-
nitude 1. The extension to more general complex arguments is the z-transform (ZT)
X(z), and we discuss when such a generalization makes sense. We then discuss the
discrete Fourier transform (DFT)—the Fourier transform for both infinite periodic
sequences as well as finite sequences with integer period N (both of these can be
viewed as existing on a “discrete” circle of length N ). The sequence is thus specified
by one of its periods; this also allows us to analyze the system as if it were finite
dimensional of size N (with appropriate boundary considerations). The transform
is then finite dimensional of the same size, and denoted as Xk . Both DTFT as well
as DFT will then have a corresponding matrix/vector product (infinite for DTFT
and finite for DFT). We end the section by discussing some real-world issues, in
particular, when to use which transform.

2.4.1 Fourier Transform of Infinite Sequences—DTFT

We start with and demonstrate a fundamental property of LSI systems: complex
exponential sequences v are eigensequences of the convolution operator T (see Def-
inition 1.17), that is
T vλ = (h ∗ vλ ) = λ vλ , (2.39)
with the convolution operator as in (2.36). To prove this, we must find these vλ and
the corresponding eigenvalues λ. Thus, consider a complex exponential sequence of
frequency ω,
vω,n = ejωn . (2.40)
(This sequence is bounded since |xn | = 1. If h ∈ ℓ1 (Z), according to Proposition 2.1,
(h ∗ vλ ) is bounded as well.)
This sequence generates an entire space:

Hω = {αejωn | α ∈ C, ω ∈ R}. (2.41)

Let us now check what happens if we convolve h with vω :

X X X
(T vω )n = (h ∗ vω )n = vω,n−k hk = ejω(n−k) hk = jωn
hk e−jωk e|{z} .
k∈Z k∈Z k∈Z vω
| {z }
λω
(2.42)
Indeed, applying the convolution operator to the complex exponential sequence vω
results in the same sequence, albeit scaled by the corresponding eigenvalue λω . We
call that eigenvalue the frequency response of the system:
X
H(ω) = λω = hk e−jωk . (2.43)
k∈Z

We can thus rewrite (2.42) as

T ejωn = h ∗ ejωn = H(ω)ejωn . (2.44)
100 Chapter 2. Discrete-Time Sequences and Systems

The above is true for any vω ∈ Hω , and thus, that space does not change under the
operation of convolution, making it an eigenspace.
Finding the appropriate Fourier transform of x now amounts to projecting x
onto each of the eigenspaces Hω :

Definition 2.6 (Discrete-time Fourier transform). Given a sequence x, its

discrete-time Fourier transform X is given by
X
X(ejω ) = hx, vω i = xn e−jωn , ω ∈ R. (2.45a)
n∈Z

X(ejω ) is a 2π-periodic function called the spectrum of the sequence.13

The inverse discrete-time Fourier transform (IDTFT) is given by
Z π Z π
1 1
xn = hx, vω i ejωn dω = X(ejω )ejωn dω. (2.45b)
2π −π 2π −π

DTFT Vs Frequency Response The DTFT is defined for sequences and we use
spectrum to denote their DTFTs. The frequency response is defined for filters
(systems), and thus, strictly speaking, “DTFT of a filter” or the “spectrum of a
filter” are not correct. However, by abuse of language, as the form of the DTFT
of a sequence (2.45a) and the frequency response of a filter (2.43) are the same, we
often interchange the two; this is fine as long as we keep in mind the fundamental
difference between them.

Magnitude and Phase Response To understand the frequency response of a filter

or the DTFT of a sequence, we often write the magnitude and phase separately:
  jω
H(ejω ) = H(ejω ) ej arg(H(e )) ,
 
where H(ejω ) is a 2π-periodic real positive function—magnitude response, while
arg(H(ejω )) is a 2π-periodic, real function between 0 and 2π—phase response.14 A
filter is said to possess linear phase if the phase response is linear in ω.

Existence and Convergence of the DTFT

The existence of the DTFT and its inverse depends on the sequence x. Consider a
sequence x that is nonzero over a finite number of indices around the origin, xn = 0
for |n| > N . Then the DTFT is a trigonometric polynomial with at most (2N + 1)
terms,
N
X
XN (ejω ) = xn e−jωn , (2.46)
n=−N

13 We follow a somewhat redundant practice in signal processing of explicitly showing the peri-

odicity of X(ejω ) through the appropriate use of the argument ejω .

14 While the argument of the complex number H(ejω ) can be defined to be on [0, 2π) or [−π, π),

in either case the function arg(H(ejω )) is discontinuous at interval boundaries.

2.4. Analysis of Sequences and Systems 101

where the subscript N indicates the boundaries of the summation. This can be seen
as a partial sum, and the DTFT as the limit of XN as N → ∞, assuming the limit
exists.
• If x ∈ ℓ1 (Z), then X(ejω ) is bounded, since
 
  X  (a) X  
X(ejω ) =  
xn e−jωn  ≤ |xn | ejωn  = kxk1 < ∞,
 
n∈Z n∈Z

where (a) follows from the triangle inequality. To verify the inversion formula,
note that from the definitions (2.45a) and (2.45b), we have
Z π X X Z π
1 1
xn = xk e−jωk ejωn dω = xk ejω(n−k) dω.
2π −π 2π −π
k∈Z k∈Z

When n 6= k, the integral vanishes,

Z π π
1 
ejω(n−k) dω = ejω(n−k)  = 0,
−π j2π(n − k) −π

and when n = k, the integral is 2π, thus proving the inversion.

The partial sum in (2.46) converges uniformly to its limit X(ejω ), which is
bounded as we have just seen.
• If x√∈ ℓ2 (Z), the inversion
√ formula holds as well, and follows from {ϕω,n =
(1/ 2π)vω,n = (1/ 2π)ejωn }n∈Z being an ONB for the interval [−π, π] (we
will see that when discussing Fourier series in the next chapter).
The partial sum in (2.46) converges to a function X(ejω ) which is square
integrable over [0, 2π], as seen in Parseval’s equality shortly (2.56a). However,
this convergence might be nonuniform as we will show.
• If x belongs to other spaces, the DTFT can be found under certain conditions
and using distributions. For example, by allowing a transform with Dirac
functions, one can handle sinusoidal sequences. We will treat these cases
more formally in the next chapter.
The frequency response of a filter is typically used to design a filter with
specific properties. We now explore both the design questions as well as point out
some convergence issues.
Example 2.8 (Design of an ideal filter and Gibbs’ phenomenon). Filtering
passes/attentuates/blocks certain frequencies. For example, an ideal lowpass filter
passes frequencies below some cut-off frequency ω0 (passband [−ω0 , ω0 ]) and blocks
the others (stopband [ω0 , 2π − ω0 ]). Figure 2.5(b) illustrates an ideal lowpass filter.
To find the impulse response of such an ideal filter, we start with the desired

magnitude response (we assume a real impulse response, H(ejω ) = H(e−jω )):

  1, |ω| ≤ ω0 ;
H(ejω ) = (2.47)
0, otherwise.
102 Chapter 2. Discrete-Time Sequences and Systems

Figure 2.5: Ideal lowpass filter. (a) Impulse response of an ideal lowpass filter. (b)
Magnitude response of an ideal lowpass filter. (c) Impulse response of an ideal highpass
filter. (d) Magnitude response of an ideal highpass filter. (e) Impulse response of an idea
bandpass filter. (f) Magnitude response of an ideal bandpass filter.

Figure 2.6: DTFT of the truncated impulse response of the ideal lowpass filter, illustrat-
ing the Gibbs phenomenon. Shown are HN (ejω ) for N = 10, 100 and 1000.

For this design example, we set the phase to zero. Applying the inverse DTFT15
given in (2.45b), we obtain the impulse response as
Z ω0 ω
1 1 jωn  0 1 ejω0 n − e−jω0 n 1 sin(ω0 n)
hn = ejωn dω = e  = = .
2π −ω0 2πjn −ω0 π 2jn π n

A particular case of interest is the halfband filter, which passes through

 half of the
√
spectrum, from −π/2 to π/2. We normalize it so khk = 1, and thus H(ejω ) = 2
(see (2.56a)). Then, its impulse response is
1 sin(π/2n)
hn = √ , (2.48)
2 π/2n
as shown in Figure 2.5(a). The impulse responses of these ideal filters have a slow
decay of order 1/n, and are thus not absolutely summable. This leads to convergence
problems, because of lack of uniform convergence.
This is illustrated in Figure 2.6, where the trigonometric polynomial
N
1 X sin(π/2n) −jωn
HN (ejω ) = √ e , (2.49)
2 n=−N π/2n

which corresponds to the frequency response of (2.48), is truncated to n ∈ {−N, −N +

1, . . . , N − 1, N }, and this for various values of N . As is clear from the figure, the
convergence is nonuniform; the result oscillates around the points of discontinuity
and becomes narrower but does not decrease in size. This over- and undershoot is
of the order of 9%, and is called the Gibbs’ phenomenon (see also Figure 0.4).

In the above example, we saw the lack of uniform convergence, and the poten-
tially disturbing oscillations around points of discontinuity due to the Gibbs’ phe-
nomenon. Fortunately, there is mean-square convergence; Given a 2π-periodic func-
tion H(ejω ), square integrable on a period, and its coefficients hn from (2.45b), the
truncated DTFT is
XN
HN (ejω ) = hn e−jωn .
n=−N
15 The inverse frequency response is equal to the inverse IDTFT.
2.4. Analysis of Sequences and Systems 103

The limit of HN (ejω ) as N → ∞ is equal to H(ejω ) in L2 -norm,

lim kH(ejω ) − HN (ejω )k2 = 0.

N →∞

While the limit of HN (ejω ) is not necessarily equal to H(ejω ) everywhere, it is equal
almost everywhere.

Properties of the DTFT

Convolution Given a sequence x and filter h, in the DTFT domain, their convo-
lution maps to the product of the spectrum of the sequence and frequency response
of the filter, the convolution theorem states that:
DT F T
(h ∗ x)n ←→ H(ejω )X(ejω ). (2.50)

This result is a direct consequence of the eigensequence property of complex expo-

nential sequences vω from (2.40): since each spectral component is the projection
of the sequence x onto the appropriate eigenspace, the DTFT diagonalizes the con-
volution operator. Assume that both x and h are in ℓ1 (Z). Then, (h ∗ x) is also in
ℓ1 (Z), since
 
X X 
 XX X X
 xk hn−k  ≤ |xk | |hn−k | = |xk | |hn−k |
 
n k n k k n
X
= |xk | khk1 = kxk1 khk1 < ∞.
k

The spectrum Y (ejω ) of the output sequence y = (h ∗ x) can be written as

!
X X X
Y (ejω ) = yn e−jωn = xk hn−k e−jωn
n n k
XX
−jωk
= xk e hn−k e−jω(n−k)
n k
(a) X X
= xk e−jωk hn−k e−jω(n−k) = X(ejω )H(ejω ),
k n

where in (a) we interchanged the order of summation, an allowed operation since

xk hn−k is absolutely summable over k and n.
While the convolution is clearly commutative, it is associative only under
certain conditions. Given sequences a, b, c ∈ ℓ1 (Z),

(a ∗ (b ∗ c)) = ((a ∗ b) ∗ c) = (a ∗ b ∗ c),

verified by using the DTFT:

   
A(ejω ) B(ejω )C(ejω ) = A(ejω )B(ejω ) C(ejω ) = A(ejω )B(ejω )C(ejω ).
104 Chapter 2. Discrete-Time Sequences and Systems

Example 2.9. When sequences are not in ℓ1 (Z) or ℓ2 (Z), one needs to be careful
about associativity, as the following counterexample shows: Take the Heaviside
sequence from (2.13) an = un . As the second sequence, choose the first-order
differencing sequence bn = δn − δn−1 . Finally, choose as cn = 1, the constant
sequence. Note that neither an nor cn is in ℓ1 (Z). Now,
(a ∗ (b ∗ c)) = (a ∗ 0) = 0,
while
((a ∗ b) ∗ c) = (δn ∗ cn )n = 1.
Thus, the convolution of these particular sequences is not associative.

Modulation Given sequences x and h, their product maps into the convolution of
their spectra in the DTFT domain,
1
DT F T
xn hn ←→
(X ∗ H)ω , (2.51)
2π
where the convolution of the 2π-periodic spectra is defined as
Z π
(X ∗ H)ω = X(ejθ )H(ej(ω−θ) ) dθ. (2.52)
−π

Modulation is dual to convolution (2.50), (the proof is left as Exercise 2.3).

Autocorrelation Given a sequence x, its autocorrelation a16 and the corresponding

DTFT pair are given by
X DT F T  2
an = xk xk+n ←→ A(ejω ) = X(ejω ) . (2.53)
k∈Z

The autocorrelation measures

P the similarity of a sequence with respect to shifts of
itself. Note that a0 = k |xk |2 = kxk2 . To show (2.53), introduce the time-reversed
version of xn , x−n , and its DTFT,
DT F T
x−n ←→ X ∗ (ejω ). (2.54)
Then, another way to write a is simply
an = (xn ∗ x−n )n .
Using the convolution theorem (2.50), we obtain (2.53).

Crosscorrelation Given two sequences x and h, their crosscorrelation c and the

corresponding DTFT are given by
X DT F T
cn = xk hk+n ←→ C(ejω ) = X(ejω )H ∗ (ejω ). (2.55)
k∈Z

Further properties of autocorrelation and crosscorrelation sequences and their trans-

forms are explored in Exercise 2.4.
16 This is the deterministic autocorrelation of a sequence. The stochastic case will be seen later.
2.4. Analysis of Sequences and Systems 105

Parseval’s Equality Given xn ∈ ℓ2 (Z), its energy is conserved by the DTFT,

X Z π
1  
kxkℓ2 = 2
|xn | = X(ejω )2 dω = kXkL2 . (2.56a)
2π −π
n∈Z

To verify this formula, recall that the left-hand side is equal to the central term a0 in
the autocorrelation (2.53). Inverting the spectrum and using the inversion formula
(2.45b) leads to the desired result. Similarly, generalized Parseval’s relation can be
written as Z π
X 1
xn h∗n = X(ejω )H ∗ (ejω) dω, (2.56b)
n
2π −π

the proof of which is left as Exercise 2.5.

Both formulas can be rewritten in terms of inner products defined on the
appropriate spaces. For example, the following inner product
Z π
1
hX, Hiω = X(ejω )H ∗ (ejω) ) dω, (2.57)
2π −π

is defined on L2 ([−π, π]), with an appropriate L2 -norm:

1
kX(ejω )k2 = hX, Xiω2 . (2.58)
Then (2.56b) is
hx, hin = hX, Hiω ,
while (2.56a) is the special case with hn = xn . Parseval’s equality (norm conserva-
tion) indicates that the DTFT is a unitary map.

Frequency Response of Filters

So far, we have mostly seen ideal filters, which have rather simple frequency re-
sponses and are not realizable. Filters that are realizable in practice have more
subtle frequency responses. We now explore a few of these.
Example 2.10 (Frequency response of a moving average filter). The im-
pulse response of the moving average filter in (2.6) is:
 1
hn = 2N +1 , |n| ≤ N ; (2.59a)
0, otherwise,
while its frequency response is
N
X 2N
X
1 1
H(ejω ) = e−jωn = ejωN e−jωn
2N + 1 2N + 1 n=0
n=−N
1 1
(a) 1 1 − e−jω(2N +1) 1 e−jω/2 (ejω(N + 2 ) − e−jω(N + 2 ) )
= ejωN =
2N + 1 1 − e−jω 2N + 1 e−jω/2 (ejω/2 − e−jω/2 )
1 sin ω(N + 21 )
= , (2.59b)
2N + 1 sin ω/2
106 Chapter 2. Discrete-Time Sequences and Systems

Figure 2.7: Moving average filter (2.6). (a) Impulse response (2.59a). (b) Frequency
response (2.59b).

where (a) follows from (P1.20-6). Figure 2.7 shows the impulse response and fre-
quency response of this filter for N = 7.

Example 2.11 (Allpass filters and shift-invariant ONBs). Consider a sim-

ple shift-by-k filter given in (2.24a) with the impulse response hn = δn−k . If xn
is a complex exponential sequence as in (2.40), the filter does not influence the
magnitude of xn ; only the phase is changed:
X
(h ∗ x) = δℓ−k ejω(n−ℓ) = e−jωk ejωn = H(ejω )xn ,
ℓ

from which  
H(ejω ) = 1, arg(H(ejω )) = ωk mod 2π.
Since all frequencies are going through without change of magnitude, this is an
example of an allpass filter. In addition, it has linear phase with a slope k given by
the delay.
An example of both a more sophisticated allpass filter, as well as the fact
that while key properties are sometime not plainly visible in the time domain, they
become obvious in the frequency domain, is
h iT
g n = an u n , g = ... 0 1 a a2 a3 ... ,

where a is real or complex with |a| < 1, and un is the Heaviside sequence from
(2.13). Now, construct
hn = −a∗ gn + gn−1 .
We will show that filtering a sequence x with h will not change its magnitude, and
that, moreover, h is of norm 1 and orthogonal to all its shifts
  DT F T
H(ejω ) = 1 ←→ hhn , hn−k in = δk .

To show this, find the frequency response of gn ,

∞
X 1
G(ejω ) = an e−jωn = ,
1 − ae−jω
n∈Z

where we have used (P1.20-1). Then, H(ejω ) follows as

ejω − a∗
H(ejω ) = −a∗ G(ejω ) + e−jω G(ejω ) = . (2.60)
1 − ae−jω
2.4. Analysis of Sequences and Systems 107

The magnitude of H(ejω ) is

  (e−jω − a∗ )(ejω − a)
H(ejω )2 = H(ejω )H ∗ (ejω ) = = 1,
(1 − ae−jω )(1 − a∗ ejω )
 
and thus H(ejω ) = 1 for all ω. By Parseval’s equality, khk = 1. In frequency
domain, the result of (h ∗ x) is
      
H(ejω )X(ejω ) = H(ejω ) X(ejω ) = X(ejω ) .

The magnitude remains unchanged, from which the name—allpass filter—is derived.
As to be expected, this also corresponds to energy conservation, since using
Parseval’s relation (2.56a), we have
Z π
1 1  
jω
kH(e )X(e )k =jω 2 H(ejω )X(ejω )2 dω
2π 2π −π
Z π
1  
= X(ejω )2 dω = kxk2 .
2π −π
 
Finally, let us show that H(ejω ) = 1 implies that the impulse response hn
is orthogonal to all its shifts. Using the generalized Parseval’s relation (2.56b) and
DT F T
hn−k ←→ e−jωk H(ejω ):
X Z π
(a) 1
∗
hhn , hn−k in = hn h∗n−k = H(ejω ) e−jωk H(ejω ) dω
n
2π −π
Z π Z π
1 (b) 1
= ejωk H(ejω )H ∗ (ejω ) dω = ejωk |H(ejω )|2 dω
2π −π 2π −π | {z }
1
Z π jkπ −jkπ
1 e −e (c) sin kπ (d)
= ejωk dω = = = δk ,
2π −π 2jkπ kπ
where (a) follows from the generalized Parseval’s relation (2.56b); (b) from our
assumption; (c) from Euler’s formula (2.163); and (d) from the properties of the
sinc function (1.68). Therefore, the set {ϕkn = hn−k }k∈Z forms an orthonormal set
in ℓ2 (Z). To check whether it is ONB, we check if we can write any xn as
X X
xn = αk ϕkn = αk hn−k ,
k∈Z k∈Z

with αk = hxn , hn−k in , for which, it is sufficient to verify that kαk = kxk. Write
α as X X
αk = xn hn−k = xn h−(k−n) = (xn ∗ h−n )n .
n∈Z n∈Z
We now prove kαk = kxk:
(a) (b) (c)
kαk2 = kX(ejω )H ∗ (ejω )k2 = kX(ejω )k2 = kxk2 ,
where (a) follows from the convolution theorem (2.50), Parseval’s equality (2.56a)
and (2.54); (b) from h being an allpass filter; and (c) from Parseval’s equality again.
Figure 2.8 shows the phase of H(ejω ) given in (2.60).
108 Chapter 2. Discrete-Time Sequences and Systems

Figure 2.8: Phase of a first-order allpass filter, with a = 21 .

The above example contains a piece of good news—there exist shift-invariant ONBs
for ℓ2 (Z), as well as a piece of bad news—these bases have no frequency selectivity
(they are allpass sequences). This is one of the main reasons to search for more
general ONBs for ℓ2 (Z), as will be done in Part II of the book.

2.4.2 The z-transform

While the DTFT has many nice properties, we discussed it mostly for sequences in
ℓ1 (Z) and ℓ2 (Z), and extending it further requires care, limiting its applicability.
Take the Heaviside sequence from (2.13), which is neither in ℓ1 (Z) nor ℓ2 (Z) (nor
any other ℓp (Z) space except ℓ∞ (Z)). If we were to multiply it by an exponential
sequence rn , |r| < 1 yielding xn = rn un , we could take the DTFT of xn , as
it became an absolutely-summable sequence. What would this do to some of the
key properties of the DTFT? The way we derived the DTFT relied on the fact
that complex exponential sequences are eigensequences of the convolution operator.
However, this is true for any exponential sequence vr,n = rn , since
X X
(hn ∗ rn )n = hn rk−n = rk hn r−n = rk H(r),
n∈Z n∈Z

where X
H(r) = hn r−n (2.61)
n∈Z

is now the counterpart of the frequency response in (2.43), and is well defined for
values of r for which hn r−n is absolutely summable.
This sets the stage for an extension of the DTFT to a larger class, those based
on exponential sequences of the form

vr,ω,n = (rejω )n , (2.62)

for r ∈ R+ , ω ∈ R, n ∈ Z. Except for r = 1, these exponential sequences are

unbounded, and thus, convergence issues of the resulting transform are even more
critical than for the DTFT. For values r for which (2.61) is well defined, any
complex exponential sequence as in (2.62) will lead to a bounded eigenvalue H(r).
Thus, it is intuitive that we will have access to a convolution property similar
to the one for the DTFT, but for more general sequences. Indeed, as we will see
shortly, convolution of finite-length sequences becomes polynomial multiplication
in the transform domain. This is the essential motivation behind extending the
analysis which uses unit-norm complex exponential sequences vω,n = ejωn as in
(2.40), to more general complex exponential sequences vr,ω,n = (rejω )n as in (2.62).
For historical reasons, the complex number rejω is denoted by

z = rejω ,
2.4. Analysis of Sequences and Systems 109

where r ≥ 0 is the magnitude, ω the phase, and the associated transform is called
the z-transform:
Definition 2.7 (z-transform). Given a sequence x, the power series
X
X(z) = xn z −n , z = rejω , r ∈ R+ , ω ∈ [−π, π), (2.63)
n∈Z

is called the z-transform of xn .

Associated to this transform is a region of convergence (ROC), which is the set of
values of z for which (2.63) converges:
ROC = {z | |X(z)| < ∞}, (2.64)
and without specifying the ROC, the z-transform makes little sense.Let us explore
a few features of the ROC.
For the z-transform to exist, the following must be true:
 
X  X   X    X 
 −n 
|X(z)| =  xn z  ≤ |xn | z −n  = |xn | r−n  e−jωn  = xn r−n  < ∞.
 
n∈Z n∈Z n∈Z n∈Z

Thus, absolute summability of xn r−n is a sufficient condition for convergence. If

z1 ∈ ROC, than any z2 such that |z2 | = |z1 | also belongs to the ROC, and thus the
ROC is a ring of the form
ROC = {z | 0 ≤ α1 < |z| < α2 ≤ ∞}. (2.65)
Exercise 2.9 explores a number of properties of the ROC, summarized in Table 2.2.

Sequence ROC

General 0 ≤ α1 < |z| < α2 ≤ ∞

Finite length All z, expect possibly 0 and/or ∞
Right sided Outward from largest pole
Left sided Inward from smallest pole
Stable system Contains unit circle

Table 2.2: Some properties of the ROC of the z-transform. A right-sided sequence is
zero for some n < N , and a left-sided sequence is zero for some n > N , with N arbitrary
but finite.

Example 2.12. To develop intuition, we look at a few examples:

(i) We start with the shift-by-k sequence, xn = δn−k . Then

 |z| > 0, k > 0;
X(z) = z −k , ROC = all z, k = 0;

|z| < ∞, k < 0.
110 Chapter 2. Discrete-Time Sequences and Systems

Figure 2.9: ROC for Example 2.12. (a) The right-sided sequence, α < 1. (b) The
left-sided sequence, α < 1.

We see that a shift-by-one maps to z −1 , which is why z −1 is often called a

delay operator.
(ii) The right-sided geometric series sequence has the following z-transform pair:
ZT
X 1
xn = αn un ←→ X(z) = (αz −1 )n = , (2.66)
1 − αz −1
n∈N

with the ROC = {z | |z| > |α|}. Now, z = α is a pole (zero of the denominator)
of the complex function X(z), and we see that the ROC is bounded from
inside by a circle containing the pole. This is a general property, since the
ROC cannot contain a singularity where X(z) does not exist.
(iii) The left-sided geometric series has the following z-transform pair:
ZT
X 1
xn = −u−n−1 αn ←→ X(z) = (αz −1 )n = , (2.67)
1 − αz −1
n∈N

with the ROC = {z | |z| < |α|}. The expression for X(z) is exactly as in
the previous case; the only difference is in the ROC. Had we been given only
the this X(z) without the associated ROC, we would not be have been able
to tell whether it originated from x in (2.66) or (2.67). This shows why the
z-transform and its ROC form a pair that should not be broken.
A standard way of showing the ROC is a picture of the complex plane,
see Figure 2.9. Note the importance of the unit circle z = ejω .
A number of useful z-transforms are given in Table 2.3.

z-transform and the DTFT Given a sequence x and its z-transform X(z), with
an ROC that includes the unit circle |z| = 1, the z-transform evaluated on the unit
circle is equal to the DTFT of the same sequence:

X(z)|z=ejω = X(ejω ). (2.68)

Conversely, given a sequence x scaled by rn and its DTFT X(rejω ), for all
values of r for which it is absolutely summable, the DTFT of that scaled sequence
is equal to the z-transform of the same sequence evaluated on the circle z = rejω :

X(rejω ) = X(z)|z=rejω . (2.69)

Rational z-transforms
An important class of z-transforms are those that are rational functions, since z-
transforms of most realizable systems (systems that can be built and used in prac-
2.4. Analysis of Sequences and Systems 111

Sequence z-transform ROC

δn 1 all z
δn−k z −k all z except possibly
z = 0 (for k > 0)
or z = ∞ (for k < 0)

an un 1/(1 − az −1 ) |z| > |a|

−an u−n−1 1/(1 − az −1 ) |z| < |a|

nan un (az −1 )/(1 − az −1 )2 |z| > |a|

−nan u−n−1 (az −1 )/(1 − az −1 )2 |z| < |a|

cos(ωc n)un (1 − cos(ωc )z −1 )/(1 − 2 cos(ωc )z −1 + z −2 ) |z| > 1

sin(ωc n)un (1 − sin(ωc )z −1 )/(1 − 2 cos(ωc )z −1 + z −2 ) |z| > 1
(
an , 0 ≤ n ≤ N − 1;
(1 − aN z −N )/(1 − az −1 ) |z| > 0
0, otherwise.

Table 2.3: Commonly encountered sequences and their z-transforms with the associated
ROCs.

tice) are rational. Such z-transforms are of the form

A(z)
H(z) = , (2.70)
B(z)

where A(z) and B(z) are polynomials in z −1 , of degree M and N , respectively.

The degrees satisfy M ≤ N , otherwise, polynomial division would lead to a sum
of a polynomial and a rational function satisfying this constraint. It is a standard
fact of complex analysis that a rational function has the same number of poles and
zeros, max(M, N ).
 T
Consider a finite-length sequence h = h0 h1 . . . hM . Then H(z) =
PM −k
k=0 hk z , which has M poles at z = 0 and M zeros at the roots {zk }Mk=1 of the
polynomial H(z).17 Therefore, H(z) can be written as
M
Y
H(z) = h0 (1 − zk z −1 ) |z| > 0, (2.71)
k=1

where the form of the factorization shows explicitly both the roots as well as the
multiplicative factor h0 .
The rational z-transform in (2.70) can thus also be written as
Q
a0 M k=1 (1 − zk z
−1
)
H(z) = QN , (2.72)
b0 k=1 (1 − pk z ) −1

17 The fundamental theorem of algebra (see Theorem 1.8) states that a degree-M polynomial

has M complex roots.

112 Chapter 2. Discrete-Time Sequences and Systems

Figure 2.10: Simple discrete-time system, where z −1 stands for a unit delay.

where {zk }M N
k=1 are the zeros and {pk }k=1 are the poles of the rational function.
(Without loss of generality, we can set b0 = 1.) The ROC cannot contain any
poles, and is thus, assuming a right-sided sequence, outside of the pole largest in
magnitude. If M is smaller than N , then H(z) has additional (N − M ) zeros at
0. This can be best seen on our previous example (2.66), which can be written as
1/(1 − αz −1 ) = z/(z − α) and has thus a pole at z = α and a zero at z = 0.

Difference Equations with Finite Number of Coefficients Where do such rational

z-transforms come from? For the sake of simplicity and unless otherwise specified,
we consider right-sided sequences (causal systems). Thus, consider a causal solution
of a difference equation with a finite number of terms as in (2.32) with zero initial
conditions. Assuming x and y have well-defined z-transforms X(z) and Y (z), and
ZT
using xn−k ←→ z −k X(z), we can rewrite (2.32) as
M
! N
!
X X
−k −k
Y (z) = ak z X(z) − bk z Y (z).
k=0 k=1

Using (2.72) with b0 = 1, the transfer function is given by

PM −k
k=0 ak z
Y (z) = PN X(z) = H(z)X(z). (2.73)
1 + k=1 bk z −k
| {z }
H(z)

In other words, the impulse response of the difference equation (2.32) (when xn = δn
or X(z) = 1), is given by a rational H(z) in the ZT domain. That is, the impulse
response of a linear discrete-time system with a finite number of coefficients has a
rational z-transform.

Example 2.13 (Transfer function). Consider the simple system in Figure 2.10,
where the z −1 operators stand for unit delays. The output yn and the associated
transfer function are:
1 + a1 z −1
yn = xn + a1 xn−1 − b1 yn−1 − b2 yn−2 ⇒ H(z) = .
1 + b1 z −1 + b2 z −2

The above discussion leads to the following important result on stability:

Proposition 2.2 (Stability). A causal discrete-time linear system with a finite
number of coefficients is BIBO stable, if and only if the poles of its (reduced) z-
transform are inside the unit circle.
2.4. Analysis of Sequences and Systems 113

Proof. By what we have just seen, the z-transform of the transfer function of such
a system is a rational function H(z) = A(z)/B(z). We assume that any pole-zero
cancellation has been done (reduced system),18 as well as that N > N . Using
partial fraction expansion, we can write H(z) as a sum of first-order terms
N
X αk
H(z) =
1 − pk z −1
k=1

when all poles are distinct. (In the case of multiple poles, powers of (1 − pk z −1 ) are
involved.) Each pole pk leads to a geometric series (pk )n , and thus
N
X
hn = αk pnk .
k=1

Each term in the above sum is absolutely summable if |pi | < 1. Since hn is a
finite sum of such terms, the sum is absolutely summable as well, and according to
Proposition 2.1, the system is BIBO stable.

Inverse z-transform
Given a z-transform and its ROC, what is the time-domain sequence, how do we
invert the z-transform? The question is all the more important as we have seen
that only when the z-transform is accompanied by it ROC can we say which se-
quence it originated from without ambiguity. The general inversion formula for
the z-transform involves contour integration, a standard topic of complex analysis.
However, most z-transforms encountered in practice can be inverted using simpler
methods which we now discuss; for a more detailed treatment inverse z-transform,
Further Reading gives pointers.

Inversion by Inspection The first method is just a way of recognizing certain

z-transform pairs. For example, by looking at Table 2.3, we would see that the
following z-transform:
1
X(z) = . (2.74a)
1 − 14 z −1
has the same form as 1/(1 − az −1), with a = 1/4. From the table, we can then read
the appropriate sequence that generated it as one of the following two:


1 n 1

1 n 1
4 un , |z| > , or − 4 u−n−1 , |z| < . (2.74b)
4 4

Inversion Using Partial Fraction Expansion This method works when the z-
transform is given as ratio of two polynomials as in (2.70), which can also be
expressed as in (2.72).
18 While in real systems, pole-zero cancellation can be a delicate issue, we do not discuss it here.
114 Chapter 2. Discrete-Time Sequences and Systems

• When M < N , we have (N − M ) zeros. If, moreover, all the poles are first
order, we can express X(z) as
N
X Ak 
X(z) = , Ak = (1 − dk z −1 )X(z)z=d . (2.75)
1 − dk z −1 k
k=1

If we are able to write our X(z) as above, then

N
X
xn = xkn , (2.76)
k=1

where, depending on the ROC we are given, each xk will be either (dk )n un or
−(dk )n u−n−1 .
• When M ≥ N , then we can write X(z) as
M−N
X N
X Ak
X(z) = Bk z −k + . (2.77)
1 − dk z −1
k=0 k=1

If there is a pole di of order s, then

M−N
X N
X s
X
Ak Ck
X(z) = Bk z −k + + . (2.78)
1 − dk z −1 (1 − di z −1 )k
k=0 k=1,k6=i k=1

Example 2.14 (Inversion using partial fraction expansion). Given is

1 − z −1 1 − z −1
X(z) = = , (2.79)
1 − 5z −1 + 6z −2 (1 − 2z −1 )(1 − 3z −1 )
with poles at z = 2 and z = 3. We compute the coefficients as in (2.75):
 
1 − z −1  1 − z −1 
A1 = = 2, A2 = = −1, (2.80)
1 − 2z −1 z=3 1 − 3z −1 z=2
yielding
2 1
X(z) = −1
− . (2.81)
1 − 2z 1 − 3z −1
Again, the original sequence will depend on the ROC. (If the ROC is given as the
ring 2 < |z| < 3, what is xn ?)

Inversion Using Power-Series Expansion This method is most useful for finite-
length sequences. For example, given X(z) = (1 − z −1 )(1 − 2z −1 ), we can expand
it in its power-series form as
X(z) = 1 − 3z −1 + 2z −2 . (2.82)
Knowing that each of the elements in this power series corresponds to a delayed
Dirac impulse, we can read directly that the sequence is
xn = δn − 3δn−1 + 2δn−2 . (2.83)
2.4. Analysis of Sequences and Systems 115

Example 2.15 (Inversion using power-series expansion). Suppose now

X(z) = log(1 + 2z −1 ), |z| > 2. (2.84)

To invert this z-transform, we use its power-series

 expansion
 from Table P1.20-1.
Substituting x = 2z −1 , we confirm that |x| = 2z −1 < 2( 21 ) = 1, and thus X(z)
and its z-transform pair are:
∞
X 
2n −n ZT (−1)n+1 2n n−1 , n ≥ 1;
log(1+2z −1 ) = (−1)n+1 z ←→ xn =
n 0, otherwise.
n=1

Properties of the z-transform

As expected, the z-transform has the same properties as the DTFT but for a larger
class of sequences. As an example, we will compute the convolution of two sequences
that may not have proper DTFTs by moving to a region of the z-transform plane
where the sequences are absolutely summable, and then use the DTFT convolution
property. A summary of z-transform properties can be found in Table 2.4.

Sequence z-transform ROC

xn real X ∗ (z) X(z ∗ )

= ROCx
x−n X(z −1 ) 1/ROCx
a1 x1n + a2 x2n a1 X1 (z) + a2 X2 (z) contains ROCx1 ∩ ROCx2
xn−k z −k X(z) ROCx
αn xn X(z/α) α ROCx1
nxn −z(dX(z)/dz) ROCx
x∗n X ∗ (z ∗ ) ROCx
ℜ(xn ) [X(z) + X ∗ (z ∗ )]/2 contains ROCx
ℑ(xn ) [X(z) − X ∗ (z ∗ )]/(2j) contains ROCx
x∗−n X ∗ (1/z ∗ ) 1/ROCx
x1n ∗ x2n X1 (z)X2 (z) contains ROCx1 ∩ ROCx2
P
xNn
( (1/N ) N−1 k=0 X(WN Z
k 1/N ) (ROC)1/N
xn/N , n = kN ;
X(z N ) (ROC)N
0, n 6= kN

Table 2.4: Properties of the z-transform.

 T
Convolution Given a length-N sequence x = x0 x1 . . . xN −1 and a length-
 T
M impulse response h = h0 h1 . . . hM−1 , their convolution is given by
116 Chapter 2. Discrete-Time Sequences and Systems

(2.36). The z-transforms of x and h are19

M−1
X N
X −1
H(z) = hn z −n , X(z) = xn z −n .
n=0 n=0

The product polynomial H(z)X(z) has powers of z −1 going from 0 to M + N − 2,

and its nth coefficient is obtained from the coefficients in H(z) and X(z) that have
powers summing to n, that is, the convolution (h ∗ x) given in (2.36). Therefore:
given a sequence x and an impulse response h, with z-transforms X(z) and H(z),
and having regions of convergence ROCX and ROCH with a nonempty intersection:
ZT
(h ∗ x) ←→ H(z)X(z), (2.85)

where ROC(h∗x) contains ROCh ∩ ROCx .

Example 2.16 (DTFT does not work, z-transform works). Given are

xn = un , hn = αn un , 0 < α < ∞,

where un is the Heaviside sequence from (2.13). We would like to compute the
convolution y = (h ∗ x), where x is not absolutely summable, and for α ≥ 1, neither
is h. While using the DTFT does not seem to be a viable option, for |z| > max{α, 1},
the z-transforms of both x and h are well defined:
1 1
X(z) = , |z| > 1, H(z) = , |z| > α,
1 − z −1 1 − αz −1
and thus
1
Y (z) = , |z| > max{α, 1}.
(1 − αz −1 )(1 − z −1 )
By partial fraction expansion, we can rewrite Y (z) as
A B
Y (z) = + ,
(1 − αz −1 ) (1 − z −1 )

with A = α/(α − 1) and B = 1/(1 − α). Inverting each term leads to

1 − αn+1
yn = Aαn un + Bun = un .
1−α
As a sanity check, we can compute the time-domain convolution directly:
X ∞
X n
X 1 − αn+1
yn = xk hn−k = hn−k = αn−k = , n ≥ 0,
1−α
k∈Z k=0 k=0

and is 0 otherwise. When α > 1, yn is unbounded as well, but the z-transform Y (z)
exists for |z| > α.
19 Again, for historical reasons, the polynomial is in z −1 , but one can think of x = z −1 .
2.4. Analysis of Sequences and Systems 117

Example 2.17 (DTFT does not work, z-transform does not work either).
Here is an example where even the z-transform does not help:

xn = 1, n ∈ Z, hn = αn un , 0 < α < 1.

In sequence domain,
X X 1
yn = (h ∗ x)n = hn xk−n = αn = .
1−α
n∈N n∈N

However, because xn is two-sided and not absolutely summable on either side, there
is no z for which the z-transform converges, thus leading to an empty ROC. This
prohibits the use of the z-transform for the computation of this convolution.

Autocorrelation Given a real sequence x, its autocorrelation is given in (2.53),

and by extension of the argument used for the DTFT, we have
X ZT
an = xk xk+n ←→ A(z) = X(z)X(z −1 ), (2.86)
k∈Z

with an ROC that is the intersection of ROCx and 1/ROCx .20

It is easy to verify that on the unit circle, the autocorrelation is the square
 2
magnitude of the spectrum X(ejω ) as in (2.53). This “quadratic form,” when
extended to the z-plane, leads to a particular symmetry of poles and zeros when
X(z), and thus A(z) as well, are rational functions.

Proposition 2.3 (Rational autocorrelation of a stable and real sequence).

A rational z-transform A(z) is the autocorrelation of a stable real sequence xn , if
and only if:
(i) Its complex poles and zeros appear in quadruples:

{zi , zi∗ , zi−1 , (zi−1 )∗ }, {pi , p∗i , p−1 −1 ∗

i , (pi ) }; (2.87a)

(ii) Its real poles and zeros appear in pairs:

{zi , zi−1 }, {pi , p−1

i }; (2.87b)

(iii) Its zeros on the unit circle are double zeros:

{zi , zi∗ , zi−1 , (zi−1 )∗ } = {ejωi , e−jωi , ejωi , e−jωi }, (2.87c)

with the special case of possibly double zeros at z = ±1. There are no poles
on the unit circle.

Proof.
The proof follows from the following two facts:
20 By 1/ROCx , we denote the set {z|z −1 ∈ ROCx }.
118 Chapter 2. Discrete-Time Sequences and Systems

• an is real, since xn is real. From Table 2.4, this means that:

pi pole ⇒ p∗i pole

A∗ (z) = A(z ∗ ) ⇒ (2.88a)
zi zero ⇒ zi∗ zero

• an is symmetric, since a−n = an . From Table 2.4, this means that:

pi pole ⇒ p−1
i pole
A(z −1 ) = A(z) ⇒ (2.88b)
zi zero ⇒ zi−1 zero

(i) From (2.88a)-(2.88b), we have that

p∗i pole
pi pole ⇒ ⇒ (p∗i )−1 pole,
p−1
i pole

similarly for zeros, and we obtain the pole/zero quadruples in (2.87a).

(ii) If a zero/pole is real, it is its own conjugate, and thus, quadruples from (2.87a)
become pairs from (2.87b).
(iii) xn stable, real ⇒ zeros double on the unit circle: Since x is stable, there are
no poles. Since x is real, X ∗ (z) = X(z ∗ ). Thus, a rational A(z) has only zeros
from X(z) and X(z −1 ).

zi−1 zero of X(z −1 )

zi zero of X(z) ⇒
zi∗ zero of X(z)
⇒ (zi∗ )−1 zero of X(z −1 ) ⇒ zi = (zi∗ )−1 zero of X(z).

Thus, both X(z) and X(z −1 ) have zi as a zero, leading to double zeros on the
unit circle.
zeros double on the unit circle ⇒ xn stable, real: TBD.

Spectral Factorization The particular pattern of poles and zeros which character-
ize a rational autocorrelation lead to a key procedure called spectral factorization.
It is essentially taking the square root of A(ejω ), and by extension, of A(z), factoring
it into rational factors X(z) and X(z −1 ).21

Theorem 2.4 (Spectral factorization). A rational z-transform A(z) is the

autocorrelation of a stable real sequence xn if and only if it can be factored as
A(z) = X(z)X(z −1).

Instead of proving the theorem, we discuss the actual procedure, based on the results
of Proposition 2.3. As A(z) is rational, it is of the form (2.72), with poles and zeros
p
21 Note that since A(ejω ) ≥ 0 and real, one could write X(ejω ) = A(ejω ). However, such a
spectral root will in general not be rational.
2.4. Analysis of Sequences and Systems 119

Figure 2.11: Pole/zero locations of rational autocorrelations. (a) Zero locations, with
double zeros on the unit circle. (b) Pole locations.

being poles and zeros of X(z). Thus, spectral factorization amounts to assigning
poles and zeros from quadruples and pairs (2.87a)-(2.87c), to X(z) and X(z −1 ).
For the poles, there is a unique rule: take all poles inside the unit circle and
assign them to X(z). This is because x stable requires X(z) to have only poles
inside the unit circle (Proposition 2.2), while x real requires the conjugate pairs be
kept together.
For the zeros, there is a choice, since we are not forced to only assign zeros
inside the unit circle to X(z). Doing so, however, creates a unique solution called
the minimum-phase solution.22 It is now clear why it is important that the zeros
on the unit circle appear in pairs: it allows for the assignment of one each to X(z)
and X(z −1 ).

Example 2.18 (Spectral factorization). We now illustrate the procedure (TBD,

see Figure 2.11):

(i) A finite symmetric sequence that has spectral factors, showing zero locations.
(ii) An infinite symmetric sequence that has spectral factors, showing pole loca-
tions.
(iii) A symmetric, finite-length sequence that is not always positive on the unit
circle, and is therefore not an autocorrelation, showing the necessity of positive
semidefinitiveness.

Crosscorrelation Given a real sequence x, its crosscorrelation is given in (2.55),

and by extension of the argument used for the DTFT, we have
X ZT
cn = xk hk+n ←→ C(z) = X(z −1 )H(z), (2.89)
k∈Z

with ROC containing 1/ROCx ∩ ROCh . The proof of this is left as an exercise.

Rational Filters and Filter Design

A major application of the z-transform is in the analysis and design of filters.
As pointed out previously, there is a one-to-one relationship between a rational
z-transform and realizable difference equations (the ones with finite numbers of
coefficients). Therefore, designing a desirable filter is essentially the problem of
strategically placing poles and zeros in the z-plane.
22 The name stems from the fact that among the various solutions, this one will create a minimal

delay, or that the sequence is most concentrated towards the origin of time.
120 Chapter 2. Discrete-Time Sequences and Systems

As simple as it sounds, this is a rather sophisticated problem, and has led

to a vast literature and numerous numerical procedures. For example, a standard
way to design FIR filters is to use a numerical optimization procedure such as the
Parks-McClellan algorithm, which iteratively modifies coefficients so as to approach
a desired frequency response. Rather than embarking on a tour of filter design, we
study properties of certain classes of filters; pointers to filter design techniques are
given in Further Reading.

FIR Filters An FIR filter has a z-transform that is a polynomial in z −1 ,

L−1
X
H(z) = hn z −n ,
n=0

and is given in its factored form as (2.71).

Linear Phase Filters An important subclass of FIR filters are linear phase filters,
obtained when filters (of length L) possess certain symmetries:

symmetric antisymmetric ZT
←→ H(z) = ±z −L+1H(z −1 ), (2.90)
hn = hL−1−n hn = −hL−1−n

where ± indicates symmetry/antisymmetry, respectively. It is easy to see why the

above is true; H(z −1 ) reverses the filter, z −L+1 makes it causal again, and ± then
determines the type of symmetry. Figure 2.12 illustrates the above concepts for L
even and odd.
We now show an even-length, symmetric filter has linear phase; other cases
follow similarly. Given (2.90), the DTFT of the filter is:

L−1 L/2−1  
X X
jω −jωn
H(e ) = hn e = hn e−jωn + e−jω(L−1−n)
n=0 n=0
L/2−1  
X
= hn e−jω(L−1)/2 ejω(n−(L−1)/2) + e−jω(n−(L−1)/2)
n=0
L/2−1
X
L−1
=2 hn cos ω(n − 12 (L − 1)) e−jω 2 = rejaω ;
n=0

because of symmetry, the sum is real with a single phase factor making it linear.

Allpass Filters The basic single-zero/single-pole allpass building block has the z-
transform was given in (2.60) in Example 2.11:

z −1 − a∗
H(z) = , (2.91)
1 − az −1
2.4. Analysis of Sequences and Systems 121

Figure 2.12: Filters with symmetries. (a) Symmetric filter of even length. (b) Symmetric
filter of odd length. (c) Antisymmetric filter of even length. (d) Antisymmetric filter of
odd length.

Figure 2.13: Magnitude response of a 17-tap, symmetric filter designed with the Parks-
McClellan algorithm, and approximating a lowpass filter.

with the zero (a∗ )−1 and pole a. For stability, |a| < 1 is required. A more general
allpass filter is formed by cascading these elementary building blocks as

YN
z −1 − a∗i −N A∗ (z
−1
)
H(z) = −1
= z , (2.92)
i=1
1 − a i z A(z)

(recall A∗ denotes conjugation of coefficients but not of z). The autocorrelation of

such an allpass filter is given by

YN N
−1 z −1 − a∗i Y z − ai
P (z) = H(z)H∗ (z ) = = 1.,
i=1
1 − ai z −1 i=1 1 − a∗i z

and thus, an allpass filter has an autocorrelation function with pm = δm and P (z) =
P (ejω ) = 1. Poles and zeros appear in pairs as {a, (a∗ )−1 } = {r0 ejω0 , 1/r0 ejω0 } for
some real r0 < 1 and angle ω0 . They appear across  the unit
 circle at reciprocal
magnitudes, and is the reason why the magnitude H(ejω ) is not influenced, but
phase is, as shown in Figure 2.14.
122 Chapter 2. Discrete-Time Sequences and Systems

Figure 2.14: Phase behavior of an allpass filter.

2.4.3 Fourier Transform of Finite Sequences—DFT

The DFT is a computational tool; unlike the DTFT, it does not arise naturally
from physical systems described by difference equations. This is true for the circular
convolution as well, as it is the convolution operator that gives rise to the DFT. As
such, it is questionable how to present the DFT. There are a couple of possibilities:
(1) We can assume that sequences live on a discrete circle, define the appropriate
convolution which will turn out to be circular, and then mimic the approach used
for the DTFT by finding eigensequences of the convolution operator and defining
the DFT based on that. In this approach, implicitly, we have chosen how our finite
sequence behaves at the boundaries; it has been periodized. While one might argue
whether this is natural or not, the approach is valid and leads to the DFT. (2) The
second view asks the following question: Given that we have a finite sequence of
unknown origin, how can we analyze it? One option is to discretize the existing
tool—DTFT, by sampling. Thus, we start with the DFT from which we can deduce
the circular convolution. The resulting sequence space will be the same as in the
first case, discrete circle. As there does not seem to be a consensus whether the
first or the second approach are the right ones, we present both. We leave it up to
you to decide what makes more sense. The main point to remember is that when
one deals with finite sequences (always in practice), one has to keep in mind the
boundary issues; a simple cutout of a signal followed by the DFT might produce
disastrous results if not handled with care.

Periodic Assumption Leads to the DFT

Periodizing Finite Sequences We thus start with a finite sequence. As discussed
before, we must decide what happens at the boundaries, and choose to assume a
circular extension, making the sequence xn periodic with an integer period N , that
is
xn+kN = xn , k ∈ Z.

As such a sequence cannot be in ℓ2 (Z) (unless it is the zero sequence), we just

require it to be bounded xn ∈ ℓ∞ (Z). This also means that one period has finite
ℓ2 -norm. Since a period is all we need to specify the sequence, we will define an
N -dimensional vector xn describing a period, or
 T
x = x0 x1 . . . xN −1 . (2.93)

Circular Convolution Take now an LSI filter with impulse response h ∈ ℓ2 (Z).
The convolution of a periodic input xn with hn leads to a periodic output of the
2.4. Analysis of Sequences and Systems 123

same period, since

X X
yn+ℓN = xn+ℓN −k hk = xn−k hk = yn ,
k k

and, according to Proposition 2.1, the output is bounded. Therefore, the convolu-
tion y = (h ∗ x) when xn is periodic of period N is a map from RN (or CN ) to RN
(or CN )).
We want to describe this map based on the impulse response hn . While hn is
not periodic, each subsequent piece of length N sees an identical piece of the input.
More precisely, breaking the time index into pieces of length N , or k = ℓN + m
where m ∈ {0, 1, . . . , N − 1},

X (a) X NX
−1
yn = xn−k hk = xn−ℓN −m hℓN +m
k∈Z ℓ∈Z m=0
N
X −1 X N
X −1
(b) (c)
= xn−m hℓN +m = xn−m gm (2.94)
m=0 ℓ∈Z m=0

where in (a) we split the sum over k = ℓN + m into pieces of length N ; in (b) we
used periodicity of xn ; and in (c) we introduced
X
gm = hℓN +m , m ∈ {0, 1, . . . , N − 1}, (2.95)
ℓ∈Z
 T P P P T
g = g0 g1 . . . gN −1 = ℓ hℓN ℓ hℓN +1 ... ℓ hℓN +N −1 .

As we can see, the elements of this sequence are sums of the impulse response hn
at multiples of N shifted by m, and is bounded since hn is absolutely summable by
assumption. Note that gm is itself periodic with period N . Because both xn and gn
are periodic, their indices can be taken mod N , leading to the periodic or circular
convolution as
N
X −1 N
X −1
yn = x[(n−m) mod N ] gm = xm g[(n−m) mod N ] = (g ∗ x) . (2.96)
m=0 m=0

We use the same notation for circular convolution as for the convolution we defined
before without risking confusion; depending on the domain of the sequence, we will
use one or the other.
In matrix format,
    
y0 g0 gN −1 gN −2 . . . g1 x0
 y1   g1 g0 gN −1 . . . g2   
     x1 
 y2   g2 g1 g0 . . . g3   x2 
 
  =   = Gx, (2.97)
 ..   .. .. .. .. ..   .. 
 .   . . . . .  . 
yN −1 gN −1 gN −2 gN −3 . . . g0 xN −1
124 Chapter 2. Discrete-Time Sequences and Systems

where G is a circulant matrix containing g (one could say it is a transpose of a

circulant matrix defined in (1.86)). Therefore, the convolution of a periodic input
with a filter hn results in a linear operator from RN (or CN ) to RN (or CN ) which
is circulant, and the entries are given by a periodized impulse response gn .

Eigensequences of the Convolution Operator Now that we have the appropriate

convolution operator defined, we could just mimic what we did for the DTFT, start-
ing with identifying the eigensequences of the convolution operator (eigenvectors of
a circulant matrix). We could easily guess what these eigensequences are, but in
this case, we have a better method. Since we can identify the space of our sequences
with periodic sequences, to find them, take vω,n = ejωn from (2.40). Since vω,n is
periodic with period N :
2π
vω,n+N = ejω(n+N ) = vω,n ⇒ ejωN = 1 ⇒ ω= k, (2.98)
N
for k ∈ Z. Since k and k + ℓN lead to the same complex exponential sequence, we
have N complex exponential sequences of period N , indexed by k ∈ {0, 1, . . . , N −1}
instead of ω:
2π
h iT
vk,n = ej( N k)n = WN−kn , vk = 1 WN−k . . . WN−(N −1)k . (2.99)

Thus, instead of infinitely many eigensequences, we now have N of them. Let

us check that these are indeed eigensequences of the convolution operator. We
can check that easily by applying the circular convolution operator G to vk . We
compute the nth element of the resulting vector:
N
X −1 N
X −1 N
X −1
[(n−i) mod N ]k (a) (n−i)k (b)
(G vk )n = gi WN = gi WN = gi WN−ik WNnk ,
| {z }
i=0 i=0 i=0
| {z } vk,n
Gk
(2.100)
where (a) follows from the fact that if (n − i) = ℓN + p, then (n − i) mod N = p and
[(n−i) mod N ] (n−i) (ℓN +p)
WN = WNp , but also WN = WN = WNp ; and (b) from the fact
nk
that WN does not depend on i and can be pulled out in front of the sum. Thus
indeed, the vector vk,n is the eigenvector of the convolution operator G, with the
associated eigenvalue
N
X −1
Gk = gi WNik , k ∈ {0, 1, . . . , N − 1}. (2.101)
i=0

the frequency response.

Finding the appropriate Fourier transform of x now amounts to projecting
x onto each of the eigenspaces Hk each spanned by one of the eigenvectors vk ,
k ∈ {0, 1, . . . , N − 1}.
2.4. Analysis of Sequences and Systems 125

Definition 2.8 (Discrete Fourier transform). Given a finite sequence x, its

discrete Fourier transform (DFT) X is given by
N
X −1
Xk (ejω ) = hx, vk i = xn WNnk , (2.102a)
n=0

k ∈ {0, 1, . . . , N −1}. Xk is a length-N sequence called the spectrum of the sequence.

The inverse discrete Fourier transform (IDFT) is given by
N −1
1 X
xn = Xk WN−nk . (2.102b)
N
k=0

n ∈ {0, 1, . . . , N − 1}.

Sampling the DTFT to Obtain DFT

When given a finite sequence of unknown origin to analyze, we might first turn to
what we already have—the DTFT. To turn it into a tool for analyzing length-N
sequences, we can simply sample it. As we need only N points, we can choose them
anywhere; let us choose N uniformly spaced points in frequency, which are actually
given by (2.98).
Thus, we evaluate the DTFT at k ∈ {0, 1, . . . , N − 1} as follows:
 (a) 2π (b) X 2π
X(ejω ) 2π k = X(ej N k ) = xn e−j N kn
N
n∈Z
N
X −1 X N
X −1
(c) 2π (d) 2π
= xℓN +i e−j N k(ℓN +i) = xi e−j N ki = XK ,
i=0 ℓ∈Z i=0

where (a) follows from sampling the DTFT uniformly at ω = 2π N k; (b) from the
expression for the DTFT (2.45a); (c) from breaking the sum into subsequences
modulo N ; (d) from x being finite of lenght N , and thus ℓ = 0. The final expression
is the one for the DFT we have seen in (2.102a). Thus, sampling the DTFT results
in the DFT.
We can now easily find the appropriate convolution operator corresponding to
the DFT (we know what it is already, but we reverse engineer it here). The way to
do it is to ask ourselves which operator will have as its eigensequences the columns
of the DFT matrix F ∗ (since then, that operator is diagonalized by the DFT). In
other words, which operator T will satisfy the following:
1 ∗
T vk = λk vk ⇒ T F ∗ = F ∗Λ ⇒ T = F ΛF.
N
Even though we do not know Λ, we know it is a diagonal matrix of eigenvalues λk .
Using the expressions (2.103)-(2.104) for F and F ∗ , we can find the expression for
the element Tim as
NX−1
ℓ(m−i)
Tim = λℓ WN ,
ℓ=0
126 Chapter 2. Discrete-Time Sequences and Systems

which shows T to be a circulant matrix (circular convolution). We leave the details

of the derivation as exercise.
What we have seen in this short account is how to reverse engineer the circular
convolution given a finite sequence and the DTFT.

Properties of the DFT

Matrix View of the DFT Introduce the DFT matrix F , formed from the N
eigensequences vk :
 
1 1 1 ... 1
 1 W W 2
... WNN −1 
 N N 
 1 WN 2
WN 4
. . . WN
2(N −1) 
 
 . . .. .. .. 
F =  .. .. . . .
.
 (2.103)
 (N −1) 2(N −1) (N −1)2 
 1 WN WN . . . WN 
 
 
|{z} |{z} |{z} |{z}
v0 v1 v2 vN −1

This matrix has the following inverse:

 
1 1 1 ... 1
1 WN−1 WN−2 . . . WN
−(N −1) 
 
1 ∗ 1 
1 WN−2 WN−4
−2(N −1) 
. . . WN ,
F −1 = F = (2.104)
N N 
 .. .. .. .. .. 

. . . . . 
−(N −1) −2(N −1) −(N −1)2
1 WN WN . . . WN

which shows that F is a unitary matrix up to a scaling factor.23 To prove this, the
product of the ℓth row of F with the kth column of F ∗ is of the form
N −1 N −1
1 X −kn ℓn 1 X n(k−ℓ)
WN WN = W .
N n=0 N n=0 N

Using the orthogonality of the roots of unity from (2.170c), this is equal to 1 when
k = ℓ, and otherwise it is 0.

The DFT and the Circular Convolution Equipped with the DFT matrix and its
inverse, we are ready to state the various properties seen so far.

Proposition 2.5 (DFT and circular convolution). Given are a periodic se-
quence xn , or a length-N sequence periodically extended with period N , and a filter
hn with absolutely-summable
P impulse response h ∈ ℓ1 (Z). The periodized impulse
response is gn = ℓ∈Z hℓN +n . Then,
23 A
√
normalized version uses a 1/ N on both F and its inverse.
2.4. Analysis of Sequences and Systems 127

(i) The convolution (h ∗ x) is equivalent to the circular convolution (g ∗ x), y =

Gx, where y and x are vectors in RN (or CN ) representing one period, and G
is a circulant matrix having g as its first column.
(ii) The circulant matrix G has eigenvalues Gk , k ∈ {0, 1, . . . , N − 1}, called
frequency response, given by (2.101) and
   
G0 go
 G1   g1 
   
 ..  = F  ..  , (2.105)
 .   . 
GN −1 gN −1
that is, the vector of these eigenvalues is the DFT of the periodized impulse
response.
(iii) The corresponding eigenvectors are the columns of F ∗ as in (2.101):
GF ∗ = F ∗ Λ (2.106)
where Λ = diag(G0 , G1 , . . . , GN −1 ) is the diagonal matrix of eigenvalues
(frequency response).
(iv) The circular convolution matrix G can be diagonalized as
1 ∗
G = F ΛF, (2.107)
N
that is, the DFT diagonalizes the circular convolution.
(v) Circular convolution can be computed in DFT domain as
DF T
(g ∗ x)n ←→ Gk Xk , (2.108)
where Gk and Xk are the DFT coefficients of gn and xn , respectively.

Proof. Parts (i)-(iii) follow readily from the developments above, and the details
are the subject of Exercise 2.6. Part (iv) follows from Part (iii) by right multiplying
with F , using F ∗ F = N I and dividing by N . For Part (v) write
(a) 1 ∗
Gx = F ΛF x, (2.109)
N  
G0 X0
(b) 1

(e)  G1 X1 
(c) (d) 
F (Gx) = (F F ∗ )ΛF x = ΛF x = ΛX =  .. , (2.110)
N  . 
GN −1 XN −1
where the left side of (a) describes circular convolution as in (2.97), the equality
in (a) follows from Part (iv); (b) follows by left multiplying (2.109) by F ; (c) from
(F F ∗ ) = N I; (d) from X being the DFT of x, X = F x; and (e) from evaluating
GX.

The relationships between the convolution of a periodic sequence by hn and the

circular convolution by gn are shown schematically in Figure 2.15 and explored
further in the context of circulant matrices in solved Exercise 2.2.
128 Chapter 2. Discrete-Time Sequences and Systems

Figure 2.15: Convolution of periodic sequence. (a) The filter hn , convolved with the
sequence xn , leads to a periodic output. (b) The equivalent, periodized filter gn , circularly
convolved with xn , leads to the same output.

The DFT as an ONB That the set of vectors vk , k ∈ {0, 1, . . . , N − 1} is an

orthogonal set was shown when proving that the inverse of F is given by 1/NF ∗ as
in (2.104). Normalizing vk to have norm 1, it is easy to create an ONB for RN (or
CN ),
1 1 h (N −1)k
iT
ϕk = √ vk = √ 1 WNk WN2k . . . WN , (2.111)
N N
for k ∈ {0, 1, . . . , N − 1}, which satisfies
hϕk , ϕℓ i = δk−ℓ . (2.112)
By the eigenvalue/eigenvector arguments developed in Proposition 2.5, each of the
basis vectors spans an eigenspace of the circular convolution operator. Finally, by
periodization, we have a natural representation of periodic sequences in terms of
linear combinations of N complex exponential sequences as in (2.99), that is, given
either a period of a periodic sequence, or a finite sequence periodically extended,
 T
x0 x1 . . . xN −1 , we can write it as
N
X −1
xn = Xk ϕkn , (2.113a)
k=0

where
N
X −1
Xk = hx, ϕk i = xn ϕ∗kn , (2.113b)
n=0
which is an ONB representation of periodic sequences. The coefficients Xk are the
expansion coefficients, and the representation is the DFT.
Example 2.19. Consider the “square wave” given by

1, 0 ≤ n < N ;
xn =
0, K ≤ n < N,
periodic with period N . The expansion coefficients are
K−1
X 1 − WNKk
Xk = WN−nk = .
n=0
1 − WNk

Note that XN −k is related to Xk ,

K−1
X K−1
X
−n(N −k)
XN −k = WN = WNnk = Xk∗ ,
n=0 n−0
2.4. Analysis of Sequences and Systems 129

which is true in general for real sequences. Table 2.5 summarizes the important
properties of the DFT, while Exercise 2.7 works out some details.

Time domain DFT domain Property

N−1
1 X −nk
xn = Xk WN Xk Inverse DFT
N k=0
N−1
X
nk
xn Xk = xn W N Forward DFT
n=0

yn = x−n Yk = Xk∗ Time reversal

xn real ∗
Xk = XN−k Hermitian symmetry
xn = x−n Xk real Real, symmetric sequence
xn = −x−n Xk imaginary Real, antisymmetric sequence

(g ∗ x) Gk Xk Circular convolution
xg (G ∗ X) Modulation

Table 2.5: Properties of the DFT.

Relation between Linear and Circular Convolutions We have seen how a linear
convolution of a periodic or periodized sequence with h, can be computed as a
circular convolution with the periodized impulse response g. When the input x
is not periodic and has a finite number of nonzero samples M , with the impulse
response of hn with a finite number of nonzero samples L,
 T
x = ... 0 x0 x1 . . . xM−1 0 ... ,
h iT
h = ... 0 h0 h1 . . . hL−1 0 ... ,

the result of the convolution has also a finite number of nonzero samples, since
L−1
X
yn = (h ∗ x) = hk xn−k (2.114)
k=0

is nonzero when (n − k) is between 0 and M − 1, or, equivalently, it is nonzero

when 0 ≤ n ≤ L + M − 2, since k is between 0 and L − 1. Thus, the result of the
convolution is of length L + M − 1:
h iT
y = ... 0 y0 y1 . . . yL+M−1 0 ... .

We have chosen to write the vectors as infinite sequences, as their finite supports are
all different. We could have written each as a finite-length vector with appropriate
length; however, as they are all different, one would have to choose a common period
130 Chapter 2. Discrete-Time Sequences and Systems

Figure 2.16: Equivalence of circular and linear convolutions: (a) Linear convolution of
two sequences of length M and L, resulting in a sequence of length L + M − 1. (b) Periodic
convolution with a period N ≥ L + M − 1 leads to the same result as in (a).

N for both the input and the impulse response. Choosing this common period has
implications in computing convolutions, as is shown next.
Observe that if a sequence x has M nonzero coefficients, then its periodized
version, with a period N ≥ M , conserves the M coefficients of the sequence. This,
of course, is not the case when N < M (as the copies of the base period will overlap).
Proposition 2.6 (Equivalence of circular and linear convolutions). For
two finite-length sequences x and h of lengths M and L, respectively, linear and
circular convolution lead to the same result, if N , the period of the circular convo-
lution, satisfies
N ≥ M + L − 1. (2.115)
 T  T
Proof. Take x = x0 x1 . . . xM−1 and h = h0 h1 . . . hL−1 . Assum-
ing x and h are of infinite support with just a finite number of nonzero samples,
the linear convolution y (lin) is given by (2.114), while the circular convolution (see
(2.96)) is
N
X −1 n
X L−1
X
yn(circ) = hk x[(n−k) mod N ] = hk xn−k + hk xn+N −k ,
k=0 k=0 k=n+1

for n ∈ {0, 1, . . . , N − 1}. In the above, we broke the sum into positive indices of
xn , while the negative ones, mod N , are made positive by adding N . Also, k goes
from 0 to (N − 1), but stops at (L − 1) since hk is zero after that. Now it suffices
to show that the second sum is zero when N ≥ L + M − 1. When the index k goes
from (N + 1) to (L − 1), xn goes from xN −1 to xN −L+1 . But since N ≥ L + M − 1,
and xn is zero for n ≥ M , xn−N −k is zero, and the right sum vanishes, proving that
(lin) (circ)
yn = yn .

Figure 2.16 depicts this equivalence, and the following example shows it in matrix
notation in a particular case.
Example 2.20 (Equivalence of circular and linear convolutions). We now
look at a length-3 filter convolved with a length-4 sequence. The result of the linear
convolution is of length 6, and is given by
   
y0 h0 0 0 0  
y1   h1 h0 0 0
    x0
y2  h2 h1 h0 0  x1 
  =   
y3   0 h2 h1 h0  x2  .
   
y4  0 0 h2 h1  x3
y5 0 0 0 h2
2.5. Multirate Sequences and Systems 131

Strictly speaking, we should have used infinite matrices above; however, since yn is
nonzero only for n ∈ {0, 1, . . . , 5}, we can use finite matrices only.
To calculate circular convolution, we choose N = M + L − 1 = 6, and form a
6×6 circulant matrix G by using h as its first column. Then the circular convolution
leads to the same result as before:
    
y0 h0 0 0 0 h2 h1 x0
y1   h1 h0 0 0 0 h 2
 x1 
    
y2   0  
  =  h2 h1 h0 0 0  x2  .
y3   0 h2 h1 h0 0 0  
   x3 
y4  0 0 h2 h1 h0 0   0 
y5 0 0 0 h2 h1 h0 0
Note that if the period N had been chosen smaller (for example, N = 5), then the
equivalence would not hold anymore.
This example also shows that to compute the linear convolution, we can compute
the circular convolution instead by choosing the appropriate period N ≥ M + L − 1.
Then, the circular convolution can be computed using the DFT of size N (see
Exercise 2.8).

2.5 Multirate Sequences and Systems

So far, we considered sequences in time indexed by integers, and the time index was
assumed to be the same for all sequences. Physically, this is as if we had observed
a physical process, and each term in the sequence corresponded to a sample of the
process taken at regular intervals (for example, every second).24
In multirate sequence processing, different sequences may have different time
scales. Thus, the index n of the sequence may refer to different “physical” times
for different sequences. The question is both why do that and how to go between
these different scales. Let us look at a simple example. Start with a sequence xn
and derive a downsampled sequence yn by dropping every other sample
 T
yn = x2n → . . . x−2 x0 x2 . . . , (2.116)
Clearly, if xn is the sample of a physical process taken at t = n, then yn is a sample
taken at time t = 2n. In other words, yn has a timeline with intervals of 2 seconds
if xn has a timeline with intervals of 1 second; the clock of the process is twice as
slow. The process above is called downsampling by 2, and while simple, it has a
number properties we will study in detail. For example, is is irreversible; once we
remove samples, we cannot go back from yn to xn , and is shift-varying, requiring
more complicated analysis.
The dual operation to downsampling is upsampling. For example, upsampling
a sequence xn by 2 results in a new sequence yn by inserting zeros between every
two samples as

xn/2 , n even;  T
yn = → . . . x−1 0 x0 0 x1 0 . . . (2.117)
0, n odd,
24 The issue of when this sampling can be done, and how, is treated in detail in the next chapter.
132 Chapter 2. Discrete-Time Sequences and Systems

The index of yn corresponds to a time that is half of that for xn . For example, if xn
has an interval of 1 second between samples, then yn has intervals of 0.5 seconds;
the clock of the process is twice as fast.
What we just saw for rate changes by 2 can be done for any integer as well
as rational rate changes (the latter ones by combining upsampling by N and down-
sampling by M ). In addition, to smooth the sequence before dropping samples
downsampling is preceded by lowpass filtering, while to fill in the zeros upsampling
is followed by interpolation, thus combining filtering with sampling rate changes.
The multirate operations are used in any number of today’s physical systems, from
MP3 players, to JPEG, MPEG, to name a few.
The purpose of this section is to study these various multirate operations
and their consequences on the resulting sequences and their spectra. While the
periodically shift-varying nature of multirate systems does complicate analysis, we
use a relatively simple and powerful tool called polyphase analysis to mediate the
problem. The outline of the section follows naturally the above introduction, moving
from down- and upsampling together with filtering to polyphase analysis. Multirate
processing, while not standard DSP material, is central to filter bank and wavelet
constructions.

2.5.1 Downsampling
Downsampling by 2 Downsampling25 by 2, as introduced in (2.116) and shown in
Figure 2.17(a), is clearly not shift invariant. If the input is xn = δn , the output is
yn = δn ; however, if the input is xn = δn−1 , the output is zero! It is instructive to
look at (2.116) in matrix notation:
 
..
    .  
.. .. .. .. .. .. .. .. .. ..   ..
. .  x−2 
 .   . . . . . . .    . 
y−1  . . . 1 0 0 0 0 0 0 . . .    x−2 
     x−1   
 y0  . . . . . .    x0 
   0 0 1 0 0 0 0   x0   
 y1  = . . . . . .    
   0 0 0 0 1 0 0   x1  =  x2  , (2.118a)
 y2  . . .  
. . .  x2    
   0 0 0 0 0 0 1  x4 
 y3  . . . 0 0 0 0 0 0 0  
. . .  x3    
    x6 
.. .. .. .. .. .. .. .. .. 
. .  x4   ..
. . . . . . . . . . . .
| {z } ..
D2

y = D2 x, (2.118b)
where D2 stands for the operator describing downsampling by 2. Inspection of D2
shows that it is similar to an identity matrix, but with the odd rows taken out.
Intuitively, it is a ”rectangular” operator, with the output space being a subspace
of the input space (one would like to say of half the size, but both are infinite
dimensional). It is a linear operator and has no inverse.
25 Other possible terms are subsampling or decimation.
2.5. Multirate Sequences and Systems 133

Figure 2.17: Downsampling by 2. (a) The block diagram of the downsampling operation
by 2. (b) The spectrum of a sequence and (c) its downsampled version.

Given an input sequence xn , the output yn keeps only every second sample,
given by (2.116). To find the z-transform Y (z), the downsampled sequence may be
seen as a sequence of even samples with ones set to zero x0n , followed by contraction
of the sequence by removing those zeros. Thus,

xn , n even;
x0n =
0, n odd.
Its z-transform is
1
X0 (z) = [X(z) + X(−z)]
2
 
= (· · · + x0 + x1 z −1 + x2 z −2 + · · · ) + (· · · x0 − x1 z −1 + x2 z −2 + · · · )
1
2
X
= (· · · + x−2 z 2 + x0 + x2 z −2 + · · · ) = x2n z −2n ,
n∈Z

canceling the odd powers of z cancel and keeping the even ones. We now get Y (z)
by contracting X0 (z) as:
X 1
h 1 1
i
Y (z) = x2n z −n = X0 (z 2 ) = 12 X(z 2 ) + X(−z 2 ) . (2.119)
n∈Z

To find its DTFT, we simply evaluate Y (z) at z = ejω :

h ω ω−2π

Y (ejω ) = 12 X(ej 2 ) + X(ej 2 ] , (2.120)
134 Chapter 2. Discrete-Time Sequences and Systems

where −ejω/2 can be written as ej(ω−2π)/2 since e−jπ = −1. With the help of
Figure 2.17, we now analyze this formula. X(ejω/2 ) is a stretched version of X(ejω )
(by a factor of 2) and is 4π periodic (since downsampling contracts time, it is natural
that frequency expands accordingly). This is shown as solid line in Figure 2.17(c).
X(ej(ω−2π)/2 ) is not only a stretched version of X(ejω ), but also shifted by 2π,
shown as dashed line in Figure 2.17(c). The sum is again 2π-periodic, since Y (ejω ) =
Y (ej(ω−k2π) ). Both stretching and shifting create “new frequencies”, an unusual
occurrence, since in LSI processing, no new frequencies ever appear. The shifted
version X(ej(ω−2π)/2 ) is called the aliased version of the original (a ghost image!).

Example 2.21 (Dowsampling of sequences). To see how downsampling can

change the nature of the sequence, take xn = (−1)n = cos πn—the highest-frequency
discrete sequence. The result, its downsampled version, is yn = x2n = cos 2πn = 1—
the lowest-frequency discrete sequence (a constant):
h iT 2↓
h iT
. . . 1 −1 1 −1 1 ... −→ ... 1 1 1 1 1 ... .

Consider now xn = αn un , |α| < 1. The z-transform of this right-sided geometric

series is
1
X(z) = ,
1 − α2 z −1
and its downsampled version is

yn = x2n = α2n u2n = β n un ,

with β = α2 . The z-transform of the downsampled sequence is

1
Y (z) = .
1 − α2 z −1
We can obtain the same expression using (2.119):
  1 1
1 1 1 (1 + αz 2 ) + (1 − αz − 2 ) 1
1 + 1 = 1 1 = .
2 1 − αz 2 1+ αz − 2 (1 − αz − 2 )(1 + αz − 2 ) 1 − α2 z −1

Downsampling by N The formulas for downsampling by 2 can be generalized to

downsampling by N . Using the usual shorthand WN from (2.169), we can write the
downsampled-by-N sequence yn ,

yn = xnN , (2.121)

In the ZT domain as
N −1
1 X 1
Y (z) = X(WNk z N ). (2.122)
N
k=0
2.5. Multirate Sequences and Systems 135

The DTFT of yn is obtained by evaluating the above on the unit circle and using

1  2π ω (ω−2πk)
WNk z N  jω = e−j N k ej N = ej N ,
z=e

leading to
N −1
jω 1 X (ω−2πk)
Y (e ) = X(ej N ). (2.123)
N
k=0
The proof is an extension of the N = 2 case, and we leave it as an exercise.

2.5.2 Upsampling
Upsampling by 2 Upsampling by 2 as introduced in (2.117) and shown in Fig-
ure 2.18(a), also creates new frequencies, as it “stretches” time by a factor of 2. In
matrix notation, similarly to (2.118a), we have
     
.. .. .. .. .. .. . . ..
. .  
 .   . . . .  ..  . 
y−2  . . . 1 0 0 0 . . . .  x−1 
      
y−1  . . . 0 0 0 0 . . . x−1   0 
      
 y0  . . . 0 1 0 0 . . .  x0   x0 
      
 y1  = . . . 0 0 0 0 . . .  x1  =  0  , (2.124a)
      
 y2  . . . 0 0 1 0 . . .  x2   x1 
      
 y3  . . . 0 0 0 0 . . .  x3   0 
      
 y4  . . . 0 0 0 1 . . . ..  x2 
    .  
.. .. .. .. .. .. . . ..
. . . . . . . .
| {z }
U2

y = U2 x, (2.124b)
where U2 stands for the upsampling-by-2 operator. The matrix U2 looks like an
identity matrix with rows of zeros in between every two rows. Another way to look
at it is as an identity matrix with every other column removed. It is a “rectangular”
matrix in the sense that is has twice as many rows as columns—but when we say
that, remember that twice infinity is infinity. The downsampling and upsampling
operators are transposes of each other, D2 removes every other row, while U2 adds
a row in between every two rows:
U2 = D2T . (2.125)
In the ZT domain, the expression for upsampling by 2 is
X X
Y (z) = yn z −n = xn z −2n = X(z 2 ), (2.126)
n∈Z n∈Z

since all odd terms of yn are zero, while the even ones are y2n = xn following (2.117).
In frequency domain,
Y (ejω ) = X(ej2ω ), (2.127)
a contraction by a factor of 2 as shown in Figure 2.18(b) and (c).
136 Chapter 2. Discrete-Time Sequences and Systems

Figure 2.18: Upsampling by 2. (a) The block diagram of the upsampling operation. (b)
The spectrum of a sequence and (c) its upsampled version.

Example 2.22. Take the constant sequence xn = 1. Its upsampled version is

h iT
yn = . . . 1 0 1 0 1 . . . , and can be written as

yn = 1
2 [1 + (−1)n ] = 1
2 (cos 2πn + cos πn),

indicating that it contains both the original frequency (constant, DC value at the
origin) and a new high frequency (at ω = π, since (−1)n = ejπn = cos πn).

Relations between Upsampling and Downsampling What about combinations of

upsampling and downsampling? Clearly, upsampling by 2 followed by downsam-
pling by 2 results in the identity:

D2 U2 = I, (2.128)

since the zeros added by upsampling are at odd-indexed locations and are subse-
quently eliminated by downsampling. The operations in the reverse order is more
interesting; Downsampling by 2 followed by upsampling by 2 results in a sequence
where all odd-indexed samples have been replaced by zeros, or
 T 2↓,2↑  T
. . . x−1 x0 x1 x2 ... −→ . . . 0 x0 0 x2 ...

This operator, P = U2 D2 , is an orthogonal projection operator onto the subspace

of all even-indexed samples. To verify this, we check idempotency (for the operator
2.5. Multirate Sequences and Systems 137

to be a projection, see Definition 1.20),

P 2 = (U2 D2 )(U2 D2 ) = U2 (D2 U2 )D2 = U2 D2 = P,

using (2.128), as well as self-adjointness (for the operator to be an orthogonal pro-

jection, see Definition 1.21),

P ∗ = (U2 D2 )T = D2T U2T = U2 D2 = P,

using (2.125).

Upsampling by N Finally, generalize the formulas to upsampling by N . A se-

quence upsampled by N is of the form

xn/N , n = ℓN ;
yn = (2.129)
0, otherwise,

that is, (N − 1) zeros are inserted between every two samples. By the same argu-
ments used for (2.126), the z-transform of yn is simply

Y (z) = X(z N ) (2.130)

or, on the unit circle for the DTFT,

Y (ejω ) = X(ejN ω ). (2.131)

2.5.3 Filtering and Interpolation

We have seen above that for both downsampling and upsampling, new frequencies
appear, and is thus often desirable to filter out such frequencies. While lowpass
filtering is usually applied before downsampling to avoid aliasing, interpolation (also
form of filtering) is usually applied to fill in the zeros between the samples of the
original sequence to smooth the output sequence. We now consider these two cases
in more detail.

Filtering Followed by Downsampling Consider filtering followed by downsampling

by 2, as illustrated in Figure 2.19. For simplicity, we look at a causal FIR filter
 T
of length L, h = . . . 0 h0 h1 . . . hL−1 0 . . . . We denote by H the
band Toeplitz convolution operator corresponding to hn (see (1.87) and (1.88). An
example for L = 4 is
 
.. .. .. .. .. .. .. . .
 . . . . . . . . 
 . . . h3 h 2 h 1 h 0 0 0 . . .
 
H =  . . . 0 h3 h2 h1 h0 0 . . . .

. . . 0 0 h3 h2 h1 h0 . . .
 
.. . .. .. .. .. .. . .
. .. . . . . . .
138 Chapter 2. Discrete-Time Sequences and Systems

Convolution with the filter hn followed by downsampling by 2 can then be written

as y = D2 H,
   . .. .. .. .. ..  
.. .. .. ..
. . .. . . . . . . .
    
 y−1   . . . h2 h 1 h 0 0 0 0 . . . x−1 
    
 y0   . . .  
  =  . . . h3 h 2 h 1 h 0 0 0   x0  = D2 Hx; (2.132)
 y1  . . . 0 h h h1 h0 0 . . .  x1 
 
   3 2 
 y2  . . . 0 . . .  
   0 h3 h2 h1 h0   x2 
.. .. . .. .. .. .. .. .. .
..
. . .. . . . . . .
| {z }
D2 H

nothing else but the convolution operator H with the odd rows removed. From
the above matrix-vector product, we can also write the filtered and downsampled
sequence using inner products as
L−1
X L−1
X
yn = hk x2n−k = hx2n−k , hk ik = xk h2n−k = hh2n−k , xk ik . (2.133)
k=0 k=0

Note how the convolution and the inner product operators are related by time
reversal:
(h ∗ x)n = hh2n−k , xk ik . (2.134)
in the ZT domain, apply (2.119) to H(z)X(z),
h 1 1 1 1
i
Y (z) = 12 H(z 2 )X(z 2 ) + H(−z 2 )X(−z 2 ) , (2.135)

and, on the unit circle for the DTFT,

h ω ω ω−2π ω−2π
i
Y (ejω ) = 12 H(ej 2 )X(ej 2 ) + H(ej 2 )X(ej 2 ) . (2.136)

Figure 2.19 shows an input spectrum and its downsampled version. Figure 2.17(c)
showed the spectrum without filtering, while Figure 2.19(c) shows the spectrum
with filtering. As can be seen, when no filtering is used, then aliasing perturbs
the spectrum. When ideal lowpass filtering is used (Figure 2.19(b)), the spectrum
from −π/2 to π/2 is conserved, the rest is put to zero, then no aliasing occurs, and
the central lowpass part of the spectrum is conserved in the downsampled version
(Figure 2.19(c)).

Example 2.23 (Filtering before downsampling). Consider the 2-point aver-

age filter hn = 12 (δn + δn−1 ), whose output, downsampled by 2, y = D2 Hx, is
1
yn = 2 (x2n + x2n−1 ).

Because of filtering, all samples of the input influence the output, as opposed to
downsampling without filtering, where the odd-indexed samples had no impact.
2.5. Multirate Sequences and Systems 139

Figure 2.19: Filtering and downsampling. (a) Block diagram. (b) Original spectrum
with and without filtering. (c) Downsampled spectrum when filtering is applied first.

Upsampling Followed by Filtering Consider now filtering after upsampling, as

shown in Figure 2.20. Using the matrix-vector notation, we can write the output
as the product GU2 , where G is a band Toeplitz operator just like H:
 
.. . .. .. .. . .
  . .. . . . .  
..   ..
 
 .   . . . g2 g0 0 0 . . .  . 
 y−1   . . . g3 g1 0 0 . . .  
    x−1 
 y0   . . . 0 g2 g0  
0 . . .  x0 
    = GU2 x;
 y1  =  . . . 0 g3 g1 0 . . .   (2.137)
     x1 
 y2  . . . 0 0 g2 g0 . . .  
     x2 
..   ..
. . . . 0 0 g3 g1 . . .
.
.. .. .. .. .. . .
. . . . . .
| {z }
GU2

nothing else but the convolution operator H with the odd columns removed. Using
inner products, we can express yn as
X
yn = gn−2k xk = hxk , gn−2k ik . (2.138)
k

Another way to look at (2.137)–(2.138) is to see that each input sample xk generates
an impulse response gn delayed by 2k samples and weighted by xk . In the ZT
140 Chapter 2. Discrete-Time Sequences and Systems

Figure 2.20: Upsampling by 2, followed by filtering or interpolation with a filter having

impulse response gn .

domain, the output of filtering followed by downsampling is

Y (z) = G(z)X(z 2 ), (2.139)

while it is
Y (ejω ) = G(ejω )X(ej2ω ) (2.140)
in the DTFT domain.
Figure 2.21 shows a spectrum, its upsampled version, and finally, its ideally
filtered or interpolated version. As can be seen, the “ghost” spectrum at ωπ is
removed, and only the base spectrum around the origin remains.

Example 2.24 (Upsampling and interpolation). Consider the piecewise con-

stant interpolator, that is, a filter gn with impulse response gn = δn + δn−1 . The
sequence xn , upsampled by 2 and interpolated with gn leads to
 T
yn = . . . x−1 x−1 x0 x0 x1 x1 . . . , (2.141)

a “staircase” sequence, with stairs of height xn and length 2. A smoother interpo-

lation is obtained with a linear interpolator:
1
gn = 2 δn−1 + δn + 21 δn+1 .

From (2.137) or (2.138), we can see that even-indexed outputs are equal to input
samples (at half the index), while odd-indexed outputs are averages of two input
2.5. Multirate Sequences and Systems 141

Figure 2.21: Upsampling and interpolation. (a) Block diagram. (b) Original spectrum
(c) Upsampled spectrum after filtering with an ideal filter which keeps only frequencies
between [−π/2, π/2].

Figure 2.22: Upsampling and interpolation. (a) Example signal. (b) Piecewise constant
interpolation. (c) Linear interpolation.

samples,

xn/2
 ,  n even;
yn = 1 (2.142)
2 x(n+1)/2 + x(n−1)/2 , n odd.
 1 1
T
y = . . . x−1 2 (x−1 + x0 ) x0 2 (x0 + x1 ) x1 ... .

Compare (2.142) with (2.141) to see why (2.142) is a smoother interpolation, and
see Figure 2.22 for an example.

Upsampling, Downsampling and Filtering Earlier, we noted the duality of down-

sampling and upsampling, made explicit in the transposition relation (2.125). What
happens when filtering is involved, when is H T equal to G? By inspection, this holds
142 Chapter 2. Discrete-Time Sequences and Systems

when hn = g−n , since

   
.. .. .. .. .. . .. .. .. ..
 . . . . .  . . . . . .
. . . h1 h 2 h3 . . . . . . g−1 g−2 g−3 . . .
   
. . . h0 h1 h2 . . .  . . .
T
H =   = . . . g0 g−1 g−2  = G. (2.143)
. . . h−1 h0 h1 . . . 
 . . . g1 g0 g−1 . . .
. . . h−2 h−1 h0 . . .  . . .
  . . . g2 g1 g0 
.. .. .. .. .. .. .. .. .. ..
. . . . . . . . . .

Then, it follows that hn = g−n followed by downsampling by 2 is the transpose of

upsampling by 2 followed by interpolation with gn :

(D2 H)T = H T D2T = GU2 , (2.144)

where we used (2.143). The above “transposition equals time-reversal” will appear
prominently in our analysis of orthogonal filter banks in Chapter 6. In particular,
we will prove that when the impulse response of the filter gn is orthogonal to its even
shifts as in (2.145), and hn = g−n , then the operation of filtering, downsampling by
2, upsampling by 2 and interpolation as in (6.16) is an orthogonal projection onto
the subspace spanned by gn and its even shifts.

2.5.4 Multirate Identities

Orthogonality of Filter’s Impulse Response to its Even Shifts Filters that have
impulse responses orthogonal to their even shifts

hgn , gn−2k i = δk , (2.145)

will play an important role in the analysis of filter banks. Geometrically, (2.145)
means that the columns of GU2 in (2.137) are orthonormal to each other (similarly
for the rows of D2 H). Moreover, we can see (2.145) as the autocorrelation of g
sampled by 2. Write the autocorrelation as

ak = hgn , gn−k in ,

and note that it has a single nonzero even term, g0 = 1,

a2k = δk . (2.146)

In the ZT domain, A(z) = G(z)G(z −1 ) using (2.86). Keeping only the even terms
can be accomplished by adding A(z) and A(−z) and dividing by 2. Therefore,
(2.146) can be expressed as

A(z) + A(−z) = G(z)G(z −1 ) + G(−z)G(−z −1 ) = 2, (2.147)

which on the unit circle leads to

   2
G(ejω )2 + G(ej(ω+π) ) = 2. (2.148)
2.5. Multirate Sequences and Systems 143

(a) N H(z) H(zN) N

(b) H(z) N N H(zN)

Figure 2.23: Interchange of multirate operations and filtering. (a) Filtering and down-
sampling. (b) Upsampling and filtering. (c) Commutativity of up- and downsampling
(TBD).

This quadrature formula, also called power complementarity, will be central in the
design of orthonormal filter banks in Chapter 6. In the above we have assumed that
gn is real, and used both
  2
G(z)G(z −1 )z=ejω = G(ejω )G(e−jω ) = G(ejω )G∗ (ejω ) = G(ejω ) ,
as well as  2
 
G(−z)G(−z −1 ) = G(ej(ω+π) ) .
In summary, a filter satisfying any of the conditions below is called orthogonal :
ZT
←→ G(z)G(z −1 ) + G(−z)G(−z −1 ) = 2
hgn , gn−2k i = δk     (2.149)
DT F T
←→ G(ejω )2 + G(ej(ω+π) )2 = 2

Noble Identities These identities relate to the interchange of multirate operations

and filtering. The first states that downsampling by 2 followed by filtering with
H(z) is equivalent to filtering with H(z 2 ) followed by downsampling by 2, as shown
in Figure 2.23(a).
The second states that filtering with G(z) followed by upsampling by 2 is
equivalent to upsampling by 2 followed by filtering with G(z 2 ), shown in Fig-
ure 2.23(b). The proof of the Noble identities is left as Exercise 2.10, and both
results generalize to sampling rate changes by N .

Commutativity of Upsampling and Downsampling Up/downsampling by the

same integer does not commute, since, as we have seen, U2 followed by D2 is the
identity, while D2 followed by U2 is a projection onto the subspace of even-indexed
samples. Interestingly, upsampling by N and downsampling by M commute when
N and M are relatively prime (that is, they have no common factor). This is shown
in Figure 2.23(c), and the proof is the topic of Exercise 2.11. A couple of illustrative
examples are given as the solved Exercise 2.3.
Example 2.25 (Commutativity of upsampling and downsampling). We look
at upsampling by 2 and downsampling by 3. If we apply U2 to xn
 T
U2 x = . . . x0 0 x1 0 x2 0 x3 0 . . . ,
144 Chapter 2. Discrete-Time Sequences and Systems

followed by D3 , we get
 T
D3 U2 x = . . . x−3 0 x0 0 x3 0 x6 ... .

Applying D3 first
 T
D3 x = . . . x−3 x0 x3 x6 ... ,

followed by U2 leads to the same result, U2 D3 x = D3 U2 x.

2.5.5 Polyphase Representation

Polyphase Representation of Sequences A major twist multirate processing brings
to signal processing is that shift invariance is replaced by periodic shift variance. A
convenient way to express this periodic shift variance is to split the input xn into
its even- and odd-indexed parts. Call x0 and x1 the following sequences
 T ZT
X
x0 = . . . x−2 x0 x2 x4 . . . ←→ X0 (z) = x2n z −n ,
n∈Z
 T ZT
X
x1 = . . . x−1 x1 x3 x5 ... ←→ X1 (z) = x2n+1 z −n .
n∈Z

Then, we can express the z-transform of xn as

X(z) = X0 (z 2 ) + z −1 X1 (z 2 ), (2.150)

which clearly shows the even and odd parts of X(z). This very simple transform
that separates X(z) into X0 (z) and X1 (z) is called the polyphase transform, is
shown in Figure 2.24(a). The even and odd parts X0 (z) and X1 (z) are called the
polyphase components of X(z). The inverse transform simply upsamples the two
polyphase components and puts them back together by interleaving, as in (2.150),
shown in Figure 2.24(b).
For example, in polyphase transform domain, the effect of the operator Y =
U2 D2 in is very simple: X0 (z) is kept while X1 (z) is zeroed out:

Y (z) = X0 (z 2 ).

Polyphase Representation of Filters We now look into how filters can be ex-
pressed in the polyphase transform domain.
We start with filtering followed by downsampling by 2, as we discussed previ-
ously. We need to find the polyphase components of H(z)X(z), given the polyphase
components of X(z):

H(z)X(z) = H0 (z 2 )X0 (z 2 )+H1 (z 2 )X1 (z 2 )+z −1 (H0 (z 2 )X1 (z 2 )+z 2 H1 (z 2 )X0 (z 2 )).

Therefore, H(z) can be decomposed as

H(z) = H0 (z 2 ) + zH1 (z 2 ) (2.151)

2.5. Multirate Sequences and Systems 145

Figure 2.24: Forward and inverse polyphase transform for (a)-(b) rate changes by 2, and
(c)-(d) rate changes by N .

Figure 2.25: Polyphase representation of filters. (a) Filtering and downsampling. (b)
Upsampling and interpolation.

(note the z instead of the z −1 ). In particular, filtering followed by downsampling

by 2 can now be compactly written as (see Figure 2.25(a)):
 
  X0 (z)
Y (Z) = H0 (z) H1 (z) . (2.152)
X1 (z)

Upsampling by 2 followed by interpolation with G(z) leads to an output Y (z),

which when expressed in its polyphase components,

Y (z) = Y0 (z 2 ) + z −1 Y1 (z 2 ),

uses the polyphase decomposition of G(z)

G(z) = G0 (z 2 ) + z −1 G1 (z 2 ), (2.153)

since, in terms of the input X(z), Y (z) equals

Y (z) = (G0 (z 2 ) + z −1 G1 (z 2 ))X(z 2 ),

or, in terms of polyphase components of Y (z),

   
Y0 (z) G0 (z)
= X(z), (2.154)
Y1 (z) G1 (z)
146 Chapter 2. Discrete-Time Sequences and Systems

Figure 2.26: Filtering, downsampling by 3, upsampling by 3, and interpolation. (a)

Time domain. (b) Polyphase domain.

shown in Figure 2.25(b). Note that the duality between (filtering, downsampling)
and (upsampling, interpolation) we have seen earlier, shows through polyphase
decomposition as well, (2.151) and (2.153). This duality, including the change from
z −1 to z, is related to the transposition and time reversal seen in (2.143).

Polyphase Representation with Rate Changes by N The polyphase transform of

size N decomposes a sequence into N phases

 T
xk = . . . xN (n−1)+k xN n+k xN (n+1)+k ... , (2.155)

from which we can write X(z) as

N
X −1
X(z) = z −k Xk (z N ). (2.156)
k=0

A filter preceding downsampling by N is decomposed in the reverse order as

N
X −1
H(z) = z k Hk (z N ), (2.157)
k=0

so that the output of filtering and downsampling by N can be written as

N
X −1
Y (z) = Hk (z)Xk (z). (2.158)
k=0

Conversely, upsampling by N followed by interpolation by G(z) uses a decomposi-

tion
N
X −1
G(z) = z −k Gk (z N ).
k=0

To illustrate these, in Figure 2.26 we show a complete system for N = 3. Solved

Exercise 2.4 illustrates our discussion on polyphase transform.
2.6. Stochastic Sequences and Systems 147

2.6 Stochastic Sequences and Systems

2.6.1 Stationary and Wide-Sense Stationary Processes
2.6.2 Linear Shift-Invariant Processing
2.6.3 Multirate Processing
2.6.4 Wiener Filtering
2.7 Algorithms
2.7.1 Fast Fourier Transforms
We have thus far studied the DFT as the natural analysis tool for periodic signals.
The use of DFTs extends beyond the scope of periodic signals in part because there
are fast algorithms for computing them. These algorithms are called fast Fourier
transform (FFT) algorithms.
The DFT, as defined in (2.102a), appears as the sum of N complex terms.
Each term is the product of a complex constant (a power of WN ) and a compo-
nent of the input signal x. Thus, each DFT coefficient can be computed with N
multiplications and (N − 1) additions.26 The full DFT can be computed with N 2
multiplications and N (N − 1) additions. Writing the DFT as a matrix multiplica-
tion by FN does not change the operation count—in fact, N 2 multiplications and
N (N − 1) additions is the cost of a direct multiplication of an N × N matrix by an
N × 1 vector.
Special structure of the DFT allows its computation with far fewer operations.
For simplicity, we will consider only the case of N = 2γ with γ ∈ Z+ , though fast
algorithms exist for all values of N . Many classes of FFT algorithms are based on
decomposing a DFT computation into smaller DFTs and a few simple operations
to combine results. We develop one such algorithm here.
Starting with the definition of the DFT (2.102a), write
N
X −1
Xk = xn WNkn
n=0
N
X −1 N
X −1
(a) k(2n) k(2n+1)
= x2n WN + x2n+1 WN
n=0 n=0
N
X −1 N
X −1
(b) kn
= x2n WN/2 + WNk kn
x2n+1 WN/2 ,
n=0 n=0

where (a) uses separates the summation over odd- and even-numbered terms; and
(b) follows from WN2 = WN/2 . Recognize the first sum as the length-(N/2) DFT
 T
of the sequence x0 x2 . . . xN −2 , and the second sum as the length-(N/2)
26 We are counting complex multiplications and complex additions. It is customary to not count

multiplications by (−1) and thus lump together additions and subtractions.

148 Chapter 2. Discrete-Time Sequences and Systems

 T
DFT of the sequence x1 x3 . . . xN −1 . It is now apparent that the length-
N DFT computation can make use of FN/2 D2 x and FN/2 C2 x, where D2 is the
downsampling-by-2 operator defined in (2.118b), and C2 is a similar operator, except
that it keeps the odd-indexed values. Since the length-(N/2) DFT expression is
(N/2)-periodic in k, the length-(N/2) DFT is useful even for k ∈ {N/2, N/2 +
1, . . . , N − 1}.
To help get a compact matrix representation, we define a diagonal matrix to
encapsulate the WNk factor:
h i
AN/2 = diag( 1 WN WN2 . . . WN(N/2)−1 ).

We can now express the length-N DFT vector X as

   
IN/2 AN/2 FN/2 0 D2
X = x, (2.159)
IN/2 −AN/2 0 FN/2 C2
| {z }
FN

k−N/2
where the final twist is to realize that WNk = −WN . If this turns out to be a
useful factorization, then we can repeat it to represent FN/2 using FN/4 , etc. So let
us count computations in the factored form.
Let µN and νN denote the number of multiplications and additions in comput-
ing a length-N DFT. The factorization (2.159) shows that a length-N DFT can be
accomplished with two length-(N/2) DFTs, N/2 multiplications and N additions,
yielding the following recursions:
N N N N
µN = 2µN/2 + = µ2 + (log N − 1) = (log N − 1) ,
2 2 2 2
N
νN = 2νN/2 + N = ν2 + (log N − 1)N = (log N − 1)N,
2
solved with the initial conditions µ2 = 0 and ν2 = 0. Thus, recursive application of
(2.159) has allowed us to reduce the number of multiplications from N 2 to (log N −
1)N/2.
Other well-known FFT algorithms are split-radix which reduces the number
of multiplications; Good-Thomas for N = N1 N2 where N1 and N2 are coprime;
Winograd, typically used for small factors; Rader’s for prime size N ; and many
others. Further Reading gives pointers to texts discussing this wealth of FFTs.

2.7.2 Convolution
TBD
2.A. Elements of Complex Analysis 149

Appendix
2.A Elements of Complex Analysis
A complex number z ∈ C, is a number of the form:

z = a + jb, a, b ∈ R, (2.160)

and j is by definition
√ the positive root of −1. (Mathematicians and physicists
typically use i for −1, while j is more common in engineering.)
In (2.160), a is called the real part while b is called the imaginary part. The
complex conjugate of z is denoted by z ∗ and is by definition:

z ∗ = a − jb. (2.161)

Any complex number can be represented in its polar form as well:

z = rejθ , (2.162)

where r is called the modulus or magnitude and θ is the argument. Using the Euler’s
formula:
ejθ = cos θ + j sin θ, (2.163)
we can express a complex number further as

z = rejθ = r(cos θ + j sin θ). (2.164)

Euler’s formula is particularly useful, since then we can easily find a power of a
complex number as

(cos θ + j sin θ)n = (ejθ )n = ejnθ = cos nθ + j sin nθ. (2.165)

Two other useful relations using Euler’s formula are:

ejθ + e−jθ ejθ − e−jθ

cos θ = , sin θ = . (2.166)
2 2j
We can easily go between standard and polar forms as follows:
√
r = a2 + b2 θ = arctan b/a,
(2.167)
a = r cos θ b = r sin θ.

Just as a reminder, the basic operations on complex numbers are as follows:

z1 + z2 = (a1 + a2 ) + j(b1 + b2 ),
z1 − z2 = (a1 − a2 ) + j(b1 − b2 ),
z1 · z2 = (a1 a2 − b1 b2 ) + j(b1 a2 + a1 b2 ),
z1 a1 a2 + b 1 b 2 a2 b 1 − a1 b 2
= 2 2 + j .
z2 a2 + b 2 a22 + b22
150 Chapter 2. Discrete-Time Sequences and Systems

Im z Im z
jθ 1
z =r e ( W3)
b
r
( W3)0
θ
a Rez 1 Rez

( W3) 2

(a) (b)
Figure 2.27: (a) A pictorial representation of a complex number in a complex plane.
(b) Third roots of unity.

Complex numbers are typically shown in the complex plane (see Fig. 2.27(a)).
Similarly to the triangle inequality we have seen in Definition 1.7, we have the
triangle inequality for complex numbers
 
X  X
 
 zi  ≤ |zi | . (2.168)
 
i i

Roots of Unity The same way j was defined as the second root of unity, we can
define the principal N th root of unity as
1
WN = (1) N = e−j2π/N . (2.169)

You can easily check that WNk , for k ∈ {1, 2, . . . , N − 1}, are also roots of unity. If
we drew all N roots of unity in the complex plane, we would see that they slice up
the unit circle by equal angles. An example with N = 3 is given in Fig. 2.27(b),
while for N = 4 they are 1, −j, −1, j. Here are some useful identities involving the
roots of unity:

WNN = 1, (2.170a)
WNkN +n = WNn , with k, n ∈ Z, (2.170b)
N
X −1 
N, n = ℓN, ℓ ∈ Z;
WNnk = (2.170c)
0, otherwise.
k=0

The last relation is often referred to as orthogonality of the roots of unity. Its proof
is given next.
2.B. Elements of Algebra: Discrete Polynomials 151

Proof. We use the finite sum from (P1.20-6):

N
X −1
1 − WNnN
(WNn )k = . (2.171a)
1 − WNn
k=0

If n = ℓN + p, this reduces to
N
X −1
WNkn = 0. (2.171b)
k=0

For n = ℓN , substitute n directly into (2.170c).

2.B Elements of Algebra: Discrete Polynomials

A polynomial sequence is a sequence whose nth element is a finite sum of the fol-
lowing form:
N
X
pn = ak n k . (2.172)
k=0
For example, a constant polynomial sequence is of the form pn = a, a linear poly-
nomial sequence is of the form pn = a0 + a1 n, and a quadratic polynomial sequence
is of the form pn = a0 + a1 n + a2 n2 . The z-transform of such a sequence is:
N
!
X X X
P (z) = pn z −n = ak nk z −n .
n∈Z n∈Z k=0

When we study wavelets and filter banks, we will be concerned with the moment
anihilating/preserving properties of such systems. In that, the following fact will
be of use: Convolution of the polynomial sequence with a differencing filter dn =
(δn − δn−1 ), or, multiplication of P (z) by D(z) = (1 − z −1 ), reduces the degree of
the polynomial by 1, as in
N
!
X X X
−1 −n −1
D(z)P (z) = (1 − z ) pn z = (1 − z ) k
ak n z −n
n n k=0
N
XX N
XX
= ak nk z −n − ak nk z −(n+1)
n k=0 n k=0
N
XX N
XX
= ak nk z −n − ak (n − 1)k z −n
n k=0 n k=0
N
XX
= ak (nk − (n − 1)k )z −n
n k=0
N −1
!
(a) X X X
k
= bk n z −n = rn z −n ,
n k=0 n
152 Chapter 2. Discrete-Time Sequences and Systems

where rn is a polynomial of degree (N − 1), and (a) follows from (nN − (n − 1)N )
being a polynomial of degree (N − 1). The above process can be seen as applying a
differencing filter with a zero at z = 1. Extending the above argument, we see that
by repeatedly applying the differencing filter, we will kill the constant term, then
the linear, then the quadratic, and so on.
Chapter at a Glance 153

Chapter at a Glance

We now summarize the main concepts and results seen in this chapter in a tabular form.

Properties of the Dirac Impulse Sequence

X
δn = 1
n∈Z
X X
xn0 −n δn = δn0 −n xn = xn0
n∈Z n∈Z
xn ∗ δn−n0 = xn+n0
xn δn = x0 δn
DT F T
δn ←→ 1
1 DT F T
←→ δ(ω)
2π

Concepts in Discrete-Time Processing

Concept Notation Infinite Time Finite Time Comments

shift linear circular

sequence xn n∈Z n ∈ {0, 1, . . . , N − 1} signal

vector

LSI system hn n∈Z n ∈ {0, 1, . . . , N − 1} filter

operator
impulse response

X N−1
X
convolution (h ∗ x) xk hn−k xk hn−k
k∈Z k=0
2πk
eigensequence v ejωn ej N n
satisfies (h ∗ vλ ) = λvλ (h ∗ vω ) = H(ejω ) vω (h ∗ vk ) = Hk vk
eigenspace Hλ Hω = {αejωn } Hk = {αejωn }
α ∈ C, ω ∈ R α ∈ C, k ∈ Z

frequency response λ λω = H(ejω ) λk = Hk eigenvalue

X∞ N−1
X 2πk
hn e−jωn hn e−j N n corresponding to v
n=−∞ n=0

Fourier transform hx, vi DTFT DFT spectrum

X(ejω ) = Xk =
∞
X N−1
X 2πk
xn e−jωn xn e−j N n
n=−∞ n=0
154 Chapter 2. Discrete-Time Sequences and Systems

Multirate, factor 2 Expression Reference

Input xn , X(z), X(ejω )

Downsampling by 2 yn = x2n (2.116)

ˆ ˜
Y (z) = 21 X(z 1/2 ) + X(−z 1/2 ) (2.119)
ˆ ˜
Y (ejω ) = 12 X(ejω/2 ) + X(ej(ω−2π)/2 ) (2.120)
(
xn/2 , n even;
Upsampling by 2 yn = (2.117)
0, otherwise.
Y (z) = X(z 2 ) (2.126)
Y (ejω ) = X(ej2ω ) (2.127)
ˆ ˜
Filtering by hn Y (z) = 12 H(z 1/2 )X(z 1/2 ) + H(−z 1/2 )X(−z 1/2 ) (2.135)
jω 1
ˆ jω/2 jω/2 j(ω−2π)/2
˜
& downsampling by 2 Y (e ) = 2 H(e )X(e ) + H(e )X(ej(ω−2π)/2 ) (2.136)

Upsampling by 2 Y (z) = G(z)X(z 2 ) (2.139)

& filtering by gn Y (ejω ) = G(ejω )X(ej2ω ) (2.140)

Multirate, factor N Expression Reference

Input xn , X(z), X(ejω )

Downsampling by N yn = xnN (2.121)

1 PN−1 k 1/N )
Y (z) = N k=0 X(ωN z (2.122)
jω 1 PN−1 j(ω−2πk/N) )
Y (e ) = N k=0 X(e (2.123)
(
xn/N , n = lN ;
Upsampling by N yn = (2.129)
0, otherwise.
Y (z) = X(z N ) (2.130)
Y (ejω ) = X(ejNω ) (2.131)
Historical Remarks 155

Historical Remarks
Cooley-Tukey FFT The impact signal processing has had in
practical terms is perhaps due in large part to the advent of the
fast Fourier transform algorithms, spurred by the paper of Cooley
and Tukey in 1965 [13]. It breaks the computation of the DFT of
size N = N1 N2 into a recursive computation of smaller DFTs of
sizes N1 and N2 , respectively (see Section 2.7.1). Unbeknownst
to them, a similar algorithm was published by Gauss some 150
years earlier, in his attempt to track asteroid trajectories.

MP3, JPEG and MPEG These already household names, are

all algorithms for coding and compressing various types of data
(audio, images and video). The basic ideas in all three all stem from transform-based work
in signal processing, called subband coding, closely-related to wavelets as we will see in
later chapters.

Further Reading
Books and Textbooks A standard book on discrete-time processing is the one Oppen-
heim, Schafer and Buck [45]. For statistical signal processing, see the text by Porat [49].
An early account of multirate signal processing was given by Crochiere and Rabiner [14], a
more recent one is by Vaidyanathan [60]. Dugeon and Mersereau cover multidimensional
signal processing in [22]. Blahut in [5] discusses fast algorithms for DSP, and in particular,
various classes of FFTs.

Inverse z-transform Via Contour Integration The formal inversion process for the
z-transform is given by contour integration using Cauchy’s integral formula when X(z) is
a rational function of z. When X(z) is not rational, inversion can be quite difficult. A
short account of inversion using contour integration is given in [39]; more details can be
found in [45].

Filter Design Numerous filter design techniques exist. They all try to approximate
the desired specifications of the system/filter by a realizable discrete-time system/filter.
For IIR filters, one of the standard methods is to design the discrete-time filters from
continuous-time ones using bilinear transformation. For FIR filters, windowing is often
used to approximate the desired response by truncating it with a window, a topic we
touched upon in Example 2.4. Then, linear phase is often incorporated as a design re-
quirement. Kaiser window design method is a standard method for FIR filter design
using windowing. Another commonly used design method is called Parks-McClellan. An
excellent overview of filter design techniques is given in [45].

Algebraic Theory of Signal Processing Algebraic theory of signal processing is a

recent development whose foundations can be found in [50]. It provides the framework for
signal processing in algebraic terms, and allows for the development of other models but
those based on time. For example, by introducing extensions different from the circular
(periodic) one we discussed in Section 2.4.3, in particular the symmetric one, one can show
that the appropriate transform is the well-known DCT. Moreover, it provides a recipe
156 Chapter 2. Discrete-Time Sequences and Systems

for building a signal model bases on sequence and filter spaces, appropriate convolutions
and appropriate Fourier transforms. As such, the existence and form of fast algorithms
for computing such transforms (among which are the well-known trigonometric ones) are
automatically given. Some of the observations in this chapter were inspired by the algebraic
framework.

Exercises with Solutions

2.1. Filtering as a Projection
Given is a filter with impulse response gn , n ∈ Z. For the filters below, check whether they
are orthogonal projections or not:
(Hint: Use the DTFT G(ejω ) of the filter’s impulse response.)
(i) gn = δn−k , k ∈ Z;
(ii) gn = 21 δn+1 + δn + 21 δn−1 ;
sin(nπ/2)
(iii) gn = √1 .
2 nπ/2

Consider now the class of real-valued discrete-time filters that perform an orthogonal pro-
jection. Give a precise characterization of this class (be more specific than just repeating
the conditions for an operator being an orthogonal projection.)

Solution:
We can represent our system as y = Gx, with the corresponding matrix notation as in
(2.37). To check that a filter (operator) is an orthogonal projection, we must check that it
is idempotent and self-adjoint as in Definitions 1.20 and 1.21.
Checking idempotency is easier in the Fourier domain since G2 = G has as its DTFT
pair G(ejω ) = G2 (ejω ). Checking self-adjointness is equivalent to checking that the oper-
ator matrix G is Hermitian.
(i) We have that G(ejω ) = e−jωk and G2 (ejω ) = e−jω2k . Thus, G2 (ejω ) 6= G(ejω ),
unless k = 0, and the filter operator is not a projection.
(ii) Similarly, we have
1 1
G(ejω ) = ejω + 1 + e−jω ,
2 2
and
„ «2
1 jω 1 1 3 1
G2 (ejω ) = e + 1 + e−jω = ej2ω + ejω + + e−jω + e−j2ω .
2 2 4 2 4

Thus, G2 (ejω ) 6= G(ejω ), and this filter is not a projection operator either.
(iii) We can rewrite gn as
Z π/2 √
1 sin nπ/2 1
gn = √ = 2 ejωn dω,
2 nπ/2 2π −π/2

yielding G(ejω ) as
 √
2, for ω ∈ [−π/2, π/2];S
G(ejω ) =
0, for ω ∈ [−π, −π/2] [π/2, π],

as well as 
2, for ω ∈ [−π/2, π/2];S
G2 (ejω ) =
0, for ω ∈ [−π, −π/2] [π/2, π].
Again, G2 (ejω ) 6= G(ejω ), and this filter is not a projection operator either. Note
sin nπ/2
that for gn = nπ/2 we would have had a projection.
Exercises with Solutions 157

In the most general case, again trying to satisfy idempotency, we see that for G2 (ejω ) =
G(ejω ), G(ejω ) can only be 1 or 0:
 S S S
1, for ω ∈ R1 R2 ... Rn ;
G(ejω ) =
0, otherwise,
where R1 , R2 , . . . , Rn are disjoint regions in [−π, π].
Self-adjointness is satisfied if gn = g−n ∗ . In Fourier domain:

X X X
G(ejω ) = gn e−jωn = ∗
g−n e−jωn = ) = G∗ (ejω ).
∗ −jωn ∗
gn (e
n n n

Since, from the first condition, G(ejω ) is real, we conclude that self-adjointness is auto-
matically satisfied.
2.2. Circulant Matrices
Given is an N × N circulant matrix C as in (1.86).
(i) Give a simple test for the singularity of C.
(ii) Give a formula for det(C).
(iii) Prove that C −1 is circulant.
(iv) Given two circulant matrices C1 and C2 , show that they commute, C1 C2 = C2 C1
and that the result is circulant as well.

Solution:
Most of the developments in this problem follow from Proposition 2.5.
(i) Consider circularly convolving two length-N sequences cn and gn :
(a) (b) 1 ∗ (c) 1 (d) 1 (e) 1
c ∗ g = Cg = F (C G) = F ∗ (C F g) = F ∗ (ΛF g) = F ∗ ΛF g,
N N N |N {z }
C
ˆ ˜T
where F is the DFT matrix given in (2.103), C1 = C0 C1 . . . CN−1 is the
DFT of c, and Λ = diag C1 . In the above, (a) is obtained by applying the circular
convolution operator of c to g; (b) follows from taking the IDFT of the DFT pair of
Cg which is C G (convolution in time corresponds to muliplication in frequency, also
F −1 = (1/N )F ∗ ); in (c) we write G as the DFT of g; in (d) we write the pointwise
multiplication in frequency as a diagonal matrix Λ multiplying F g; and finally in (e)
we recognize that C = F ∗ ΛF . In other words, we have rederived (2.107), stating
that any circulant matrix C can be diagonalized using the DFT.
This formulation implies that the columns of F ∗ are the eigenvectors of C, with
Ck being the associated eigenvalues. So the matrix C is nonsingular if and only if
none of the DFT coefficients Ck are zero.
(ii) From the diagonalization of C, and from (1.84):
N−1
Y
det(C) = Ck .
k=0

(iii) Since
1 ∗ 1 −1 −1 ∗ −1 1 1 ∗ −1 1 ∗ −1
C −1 = ( F ΛF )−1 = F Λ (F ) = ( F )Λ (N F ) = F Λ F,
N N N N N
it has the same form as C in (2.107), and is thus circulant as well.
(iv) Writing C1 and C2 as C1 = F ∗ Λ1 F and C2 = F ∗ Λ2 F , we see that
C = C1 C2 = (F ∗ Λ1 F )(F ∗ Λ2 F ) = F ∗ Λ1 Λ2 F = F ∗ ΛF,
where Λ = diag{C10 C20 , C11 C21 , . . . , C1N −1 C2N −1 }. So C = C1 C2 is a circulant
matrix with eigenvalues C1k C2k , k ∈ {0, 1, . . . , N − 1.
Since Λ1 Λ2 = Λ2 Λ1 , we have C1 C2 = C2 C1 .
158 Chapter 2. Discrete-Time Sequences and Systems

2.3. Combinations of Upsampling and Downsampling

Using matrix notation, compare:
(i) U3 D2 x to D2 U3 x;
(ii) U4 D2 x to D2 U4 x.
Explain the outcome of these comparisons.

Solution:
We will solve the problem using matrix notation for U3 , U4 and D2 :
2 3
2 3 . . . . ..
.. . . . . 6 .. .. .. .. .7
. . . . 6 7
6 . . . . .7 6· · · 1 0 0 · · ·7
6 7 6 7 2 3
6· · · 1 0 0 · · ·7 6· · · 0 0 0 · · ·7 .. .. .. .. .. .. ..
6 7 6 7
6· · · 0 0 0 · · ·7 6· · · 0 0 0 · · ·7 6 . . . . . . .7
6 7 6 7 6 7
6· · · 0 0 0 · · ·7 6· · · 0 0 0 · · ·7 6· · · 1 0 0 0 0 · · ·7
6 7 6 7 6 7
U3 = 6· · · 0 1 0 · · ·7 U4 = 6· · · 0 1 0 · · ·7 D2 = 6· · · 0 0 1 0 0 · · ·7 .
6 7 6 7 6 7
6· · · 0 0 0 · · ·7 6· · · 0 0 0 · · ·7 6· · · 0 0 0 0 1 · · ·7
6 7 6 7 4 5
6· · · 0 0 0 · · ·7 6· · · 0 0 0 · · ·7 .. .. .. .. .. .. ..
6· · · 0 0 1 · · ·7 6· · · 0 0 0 · · ·77 . . . . . . .
6 7 6
4 5 6· · · 0 0 1 · · ·7
.. .. .. .. . . 6 7
. . . . . 4 .. .. .. 5
.. ..
. . . . .

(i) For the first comparison, compute

2 3 2 3
.. . . . . . .. .. . . . . . ..
.. .. .. .. .. .. .. .. .. ..
6 . .7 6 . .7
6 7 6 7
6· · · 1 0 0 0 0 · · ·7 6· · · 1 0 0 0 0 · · ·7
6 7 6 7
6· · · 0 0 0 0 0 · · ·7 6· · · 0 0 0 0 0 · · ·7
6 7 6 7
6· · · 0 0 0 0 0 · · ·7 6· · · 0 0 0 0 0 · · ·7
6 7 6 7
U3 D2 = 6· · · 0 0 1 0 0 · · ·7 D2 U3 = 6· · · 0 0 1 0 0 · · ·7 .
6 7 6 7
6· · · 0 0 0 0 0 · · ·7 6· · · 0 0 0 0 0 · · ·7
6 7 6 7
6· · · 0 0 0 0 0 · · ·7 6· · · 0 0 0 0 0 · · ·7
6· · · 0 0 0 0 1 · · ·7 6· · · 0 0 0 0 1 · · ·7
6 7 6 7
4 .. .. .. .. .. 5 4 .. .. .. .. .. 5
.. .. .. ..
. . . . . . . . . . . . . .

Thus U3 D2 x and D2 U3 x are identical since U3 D2 = D2 U3 .

(ii) For the second comparison, compute
2 3
. . . . . . .
6 .. .. .. .. .. .. . .7
6 7 2 3
6· · · 1 0 0 0 0 · · ·7 .. .. .. .. .. .. ..
6 7
6· · · 0 0 0 0 0 · · ·7 6 . . . . . . .7
6 7 6 7
6· · · 0 0 0 0 0 · · ·7 6· · · 0 1 0 0 0 · · ·7
6 7 6 7
6· · · 0 0 0 0 0 · · ·7 6· · · 0 0 0 0 0 · · ·7
6 7 6 7
U4 D2 = 6· · · 0 0 1 0 0 · · ·7 D2 U4 = 6· · · 0 0 1 0 0 · · ·7 ,
6 7 6 7
6· · · 0 0 0 0 0 · · ·7 6· · · 0 0 0 0 0 · · ·7
6 7 6 7
6· · · 0 0 0 0 0 · · ·7 6· · · 0 0 0 1 0 · · ·7
6· · · 0 0 0 0 0 · · ·7 4 5
6 7 .. .. .. .. .. .. ..
6· · · 0 0 0 0 1 · · ·7 . . . . . . .
6 7
4 .. .. .. .. .. . . 5
..
. . . . . . .

which are evidently not the same, and thus U4 D2 x 6= D2 U4 x.

The results of the two comparisons are different because in the first case, 3 and 2 are
coprime, while in the second 4 and 2 are not (see Exercise 2.11).
Exercises 159

2.4. Periodically Shift-Varying Systems

Show that a linear, periodically shift-varying system of period N , can be implemented with
a polyphase transform followed by upsampling by N , N filter operations and a summation.

Solution:
A linear, shift-varying system is characterized by a two-variable impulse response hkn ,
the response
P of the system to the input xn = Pδn−k . Since the input can be written as
xn = k xk δn−k , the output is given by yn = k xk hkn .
When hkn is periodic in k with period N , we define the polyphase components xk as
in (2.155), to yield
h iT
xkn = xnN+k , xk = . . . x−2N+k x−N+k xk xN+k x2N+k . . . ,

(N)
for k ∈ {0, 1, . . . , N − 1}. Denote the upsampled version of xkn by xkn , so
 
(N) xk,n/N , n/N ∈ Z; xn+k , n/N ∈ Z;
xkn = =
0, otherwise; 0, otherwise.
(N)
xk = [. . . 0 xk 0 0 . . . 0 xN+k 0 . . .]T .
| {z }
N−1

If we take the above upsampled components, delay each by k, and sum them up, we get x
back:
N−1
X (N)
xn = xk,n−k .
k=0
As this is a linear system, we can find the output as

X N−1
X N−1
XX N−1
X X
(N) (N) (a) (N)
yn = hkn xi,n−i = hkn xi,n−i = hkn xi,n−i
k i=0 i=0 k i=0 k−i ∈Z
N

N−1
X X N−1
X X
(b) (N) (c) (N)
= hin xi,n−i = hkn xi,n−i ,
i=0 k−i ∈Z i=0 k
N

where (a) follows from the excluded terms being zero; (b) from the periodicity of h; and
(c) because the added terms are zero. The final expression shows the output as a sum
of N terms. The ith term is the ith polyphase component, upsampled and filtered by an
i-sample delayed version of hkn .

Exercises
2.1. Properties of Periodic Sequences
Given are complex exponential sequences of the form
xn = ejω0 n .
(i) Show that if α = ω0 /2π is a rational number, α = p/q, with p and q coprime integers,
then xn is periodic with period q.
(ii) Show that if α = ω0 /2π is irrational, then xn is not periodic
(iii) Show that if x and y are two periodic sequences with periods N and M respectively,
then x + y is periodic with period N M (in general).
2.2. Linear and Shift-Invariant Difference Equations
Consider the difference equation (2.32).
160 Chapter 2. Discrete-Time Sequences and Systems

(i) Show, possibly using a simple example, that if the initial conditions are nonzero,
then the system is not linear and not shift-invariant.
(ii) Show that if the initial conditions are zero, then (a) the homogeneous solution is
zero, (b) the system is linear and (c) the system is shift invariant.
2.3. Modulation Property of the DTFT
Given are two sequences x and h, both in ℓ1 (Z). Verify that
DT F T 1
xh ←→ X ∗ H.
2π
2.4. Auto- and Crosscorrelation
Consider autocorrelation and crosscorrelation sequences and their DTFTs as defined in
(2.53) and (2.55). Show that:
(i) an = a−n and |an | ≤ a0 ;
(ii) cn is in general not symmetric and C(ejω ) = X(ejω )H ∗ (ejω );
(iii) The generalized Parseval’s formula (2.56b) holds.
2.5. Generalized Parseval’s Relation for the DTFT
Prove the generalized Parseval’s relation for the DTFT given in (2.56b).
2.6. Circulant Matrices and the DFT
Given is the circulant matrix from Proposition 2.5:
2 3
g0 gN−1 gN−2 ... g1
6 g1 g0 gN−1 ... g2 7
6 7
G = 6 . . .. .. .. 7
4 .. .. . . .5
gN−1 gN−2 gN−3 ... g0
(i) Prove that the eigenvalues are given by (2.105) and that the right eigenvectors are
the columns of F ∗ .
(ii) Show that the left eigenvectors are the rows of F .
2.7. DFT Properties
Verify the following DFT pairs from Table 2.5:
(i) Time reversal yn = x−n ;
(ii) Real symmetric xn = x−n and real antisymmetric xn = −x−n sequences;
(iii) Modulation property.
(Hint: Start with the convolution in frequency domain.)
2.8. Computing Linear Convolution with the DFT
Prove that the linear convolution of two sequences of length M and L, can be computed
using DFTs of size N ≥ M + L − 1, and show how to do it.
2.9. ROC of z-Transform
Verify the properties of the ROCs listed in Table 2.3.
2.10. Interchange of Filtering and Sampling Rate Change
(i) Prove that downsampling by 2 followed by filtering with H(z) is equivalent to filter-
ing with H(z 2 ) followed by downsampling by 2.
(ii) Prove that filtering with G(z) followed by upsampling by 2 is equivalent to upsam-
pling by 2 followed by filtering with G(z 2 ).
2.11. Commutativity of Up- and Downsampling
Prove that downsampling by M and upsampling by N commute if and only if M and N
are coprime.