University of Michigan

EECS 206 Laboratory Manual

Mark A. Bartsch David L. Neuhoff Gregory H. Wakefield

January 2, 2003
ii The University of Michigan, All rights reserved
Preface

This laboratory manual was written during the first three semesters that EECS 206 was taught at the University of
Michigan. It represents an effort to provide hands-on experience with signals and systems engineering and concepts
by working with the M ATLAB mathematics environment. The specific goals are:
• To reinforce the learning of the material presented in the lecture.

• To acquaint students with a number of problems/tasks addressed by signals and systems engineering, and with
some of the approaches to these problems.

• To involve students in the design, implementation and testing of systems that address some signals and systems
engineering tasks.

• To familiarize students with the use of M ATLAB as a primary prototyping tool for signals and systems engineer-
ing.
While not a “programming” class, it is important that students be able to do things for themselves, such as im-
plement a solution to some basic signals and systems task. In signals and systems engineering, this often involves
programming in a language like M ATLAB. However, for these laboratories we have attempted to limit the amount of
“programming for the sake of programming,” which is better obtained in a true programming course. What remains
should allow students to gain facility with M ATLAB without requiring advanced programming skills.
The lab assignments presume that students have had some significant programming experience, e.g. a first course
at the freshman level, and some experience with M ATLAB, e.g. two to three weeks of coverage in a first programming
course. This prerequisite notwithstanding, the lab manual begins with a tutorial, which serves to review MATLAB and
to emphasize the constructs needed in these assignments. It has been found that students with significant programming
experience but no prior M ATLAB can also succeed in these laboratory assignments, provided they make the extra effort
to focus strongly on M ATLAB during the first couple of weeks of the course. That is, M ATLAB is readily learnable by
people familiar with another programming language.
The laboratory assignments are intended to be mostly self-contained. To this end, each contains a substantial
amount of background material. This material highlights important theoretical concepts, introductions to specific
signals and systems problems/tasks, and the specific approaches to the solution of problems to be examined in the
assignments. In some cases, the material in the background section for each lab is meant as a reference rather than as
strictly necessary to the completion of the laboratory assignment. In other cases, the background material describes an
approach that you will use in the laboratory assignment.
Each lab assignment also contains a M ATLAB section introducing commands or techniques that will be important
in this assignment, a demo section listing the demonstrations that will take place in the lab session, and an assignment
section listing exactly what must be done. Note that the bullets indicate items to be included in your lab report.
It is highly recommended that you read through each laboratory before arriving at the laboratory session in which
you will begin the lab. This will not only give you a better foundation to understand the material in the laboratories,
but it will allow you to complete the laboratory more quickly once you have begun working on it.

The University of Michigan, All rights reserved iii

Preface

Commensurate with the first listed goal, all of the laboratories are meant to reinforce key concepts of the course.
However, the presentation will often be somewhat different from that of the lecture or textbook. For instance, we
develop convolution and filtering by connecting it to the operation of correlation, which we present in the Lab 2. We
also use the idea of correlation to motivate the key concepts of spectrum and the Fourier series. In other cases, we use
the visualization capabilities of M ATLAB to help develop an intuitive sense of how systems “work.” For instance, Lab
9 uses a GUI to graphically show the effects of poles and zeros on the frequency response of a filter.
While the laboratories reinforce material from the lectures and textbook, commensurate with the second and third
goals listed earlier, they also go beyond them in numerous places. For instance, the ideas of detection and classification
form a common theme throughout the laboratories. These ideas are not commonly introduced at the undergraduate
level, but they form an important component of signals and systems engineering. As another example, Lab 5 develops
a transform based image encoder, similar to JPEG. We also focus on the two-dimensional signals (images) in Labs 5
and 6, rather than solely concentrating on one-dimensional signals.
To a great extent, the amount you will get out of these laboratories is dependent upon the amount you put into
them. There are a wide variety of topics covered in these labs. We have necessarily not examined them in great depth,
but we wish to encourage further thought and exploration into many of them. In many of the labs, you will see items
labeled “food for thought.” These are exercises that will lead you to examine other aspects of a problem, often in
greater depth than in the actual assignment. While these “food for thought” items are in no way required, we strongly
recommend that you look at them and discuss them with your lab instructors and peers. Hopefully, you will find many
ideas and applications in this course that will interest you and encourage you to explore further.
A note about the “electronic” portion of this laboratory manual. Each laboratory involves the use of M ATLAB
code, data files, and programs that must be downloaded from the course web page. These programs were developed
using M ATLAB 6 (Release 12) and a Windows 2000 platform. While most of the code should work on any version
of M ATLAB, some (most notably the GUI programs) require M ATLAB 6 or greater. Additionally, we have provided
“compiled” MEX-file versions of many of the programs that you will be writing code to complete. This allows you
to check the results provided by your code with “correct” code, and also gives you a way to continue working on the
laboratory even if you cannot get the code working. Note, however, that these programs are compiled as Windows .dll
files. As such, they will ONLY operate on a Windows-based operating system. In general, we recommend that you
use CAEN machines with the latest version of Windows.
Remember that these laboratories are covered by the College of Engineering Honor Code. In particular, it is a
violation of the Honor Code to work on these laboratories with others, unless they are members of the lab group to
which you are assigned. Further, using, or in any way deriving advantage from, solutions from previous terms is a
violation. If you have any questions about how the Honor Code applies to this class, talk to your instructors.
Finally, we would like to acknowledge all of those who helped us during the development of these laboratories. In
particular, we would like to thank Professors Stephane Lafortune and Jeffrey Fessler for their input and comments on
these laboratories. We would also like to thank the GSIs who helped us to give the labs a “trial run” during the first
three semesters: Norm Adams, Dongsook Kim, Thomas Kragh, Ben Lee, Baptiste Poupard, Charles Hsin, and Fred
Zeitz. Finally, we would like to thank the students of EECS 206 during those semesters for their patient and helpful
comments during the development and revision of the laboratories.

MAB, DLN, GHW

August 2002

Minor corrections by JF
Jan. 2003

iv The University of Michigan, All rights reserved

Contents

Preface iii

Contents v

List of Figures xii

An Introduction to M ATLAB 1
1 What is M ATLAB? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 M ATLAB is a mathematics environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 M ATLAB is tool for visualizing data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 M ATLAB is a prototyping language . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 M ATLAB can do more... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2 Demos for the first tutorial lab section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
3 Using M ATLAB: The basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
3.1 Starting M ATLAB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
3.2 How to get help . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
3.3 Using M ATLAB as a calculator (with variables) . . . . . . . . . . . . . . . . . . . . . . . . . 3
4 Vectors, Matrices, and Arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
4.1 Constructing arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
4.2 Concatenating arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
4.3 Transposition and “flipping” arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
4.4 Building large arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
4.5 The colon operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
5 Array Arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
6 Indexing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
6.1 Basic indexing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
6.2 Single number indexing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
6.3 Vector indexing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
6.4 Finding the size of an array . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
6.5 Vector indexing to modify arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
6.6 Conditional statements and the “find” command . . . . . . . . . . . . . . . . . . . . . . . . . 8
7 Data Visualization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
7.1 Using “plot” . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
7.2 Interpolation; line and point styles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
7.3 Axis labels and titles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
7.4 Commands related to “plot” . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
7.5 Plotting with an x-axis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

The University of Michigan, All rights reserved v

Contents

7.6 Plotting multiple vectors on the same figure . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

7.7 Legends . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
7.8 Putting several axes on one figure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
7.9 Two-dimensional arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
8 Programming in M ATLAB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
8.1 Paths and working directories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
8.2 Types of command files in M ATLAB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
M ATLAB scripts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
M ATLAB functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
Scripts versus functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
8.3 Control Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
Loops . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
Conditional statements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
8.4 Strings and string output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
9 Debugging your M ATLAB code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

List of Commonly Used M ATLAB Commands 19

10 Useful M ATLAB Facts 20

1 Signals, Signal Statistics, and Signal Detection I 21

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.1.1 “The Questions” . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.2.1 Continuous-time and discrete-time signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.2.2 Describing Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.2.3 Signal support and duration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.2.4 Periodicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.2.5 Signals in M ATLAB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.2.6 Signal Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.2.7 Measuring signal distortion and error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.2.8 Signal detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Specifying the detector’s operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Detector algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.3 Some M ATLAB commands for this lab . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
1.4 Demonstrations in the Lab Session . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
1.5 Laboratory Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

2 Signal Correlation and Detection II 37

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.1.1 “The Questions” . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.2.1 Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.2.2 Running correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.2.3 Using correlation for signal detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.2.4 Using correlation for detection of signals transmitted simultaneously with other signals . . . . 41
2.2.5 Noise, detector errors, and setting the threshold . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.2.6 An algorithm for running correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.3 Some M ATLAB commands for this lab . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

vi The University of Michigan, All rights reserved

 Contents

2.4 Demonstrations in the Lab Section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

2.5 Laboratory Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

3 Sinusoids and Sinusoidal Correlation 51

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.1.1 “The Question” . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.2.1 Complex numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.2.2 Sinusoidal and complex exponential signals in continuous time . . . . . . . . . . . . . . . . . 53
3.2.3 Finding the amplitude and phase of a sinusoid with known frequency . . . . . . . . . . . . . 53
The Amplitude and Phase Calculator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.2.4 Determining the frequency of a target sinusoid . . . . . . . . . . . . . . . . . . . . . . . . . 57
A frequency estimation algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
Estimating doppler shift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.3 Some M ATLAB commands for this lab . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.4 Demonstrations in the Lab Section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.5 Laboratory Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

4 Fourier Series and the DFT 65

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.1.1 “The Questions” . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.2.1 Frequency-domain representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.2.2 Periodic Continuous-Time Signals — The Fourier Series . . . . . . . . . . . . . . . . . . . . 66
Partial Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
T -Second Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
Aperiodic Continuous-Time Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
Properties of the Fourier Coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.2.3 Periodic Discrete-Time Signals — The Discrete Fourier Transform . . . . . . . . . . . . . . 71
N -point DFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
Aperiodic Discrete-Time Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
Approximating Fourier series coefficients with the DFT . . . . . . . . . . . . . . . . . . . . 73
Properties of the DFT coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.2.4 Separating Signals Based on Differing Harmonic Series . . . . . . . . . . . . . . . . . . . . 76
4.3 Some M ATLAB commands for this lab . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.4 Demonstrations in the Lab Section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.5 Laboratory Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

5 Images, Compression, and Coding 83

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.1.1 “The Question” . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.2.1 Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.2.2 Signal statistics for images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.2.3 Data compression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.2.4 Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.2.5 Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.2.6 Binary coding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

The University of Michigan, All rights reserved vii

Contents

5.2.7 Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.2.8 Designing a transform coder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
5.3 Some M ATLAB commands for this lab . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.4 Demonstrations in the Lab Section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.5 Laboratory assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Postscript: JPEG Compression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

6 FIR Filtering and Image Processing 99

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.1.1 “The Question” . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
6.2.1 Implementing FIR Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
6.2.2 Edge effects and delay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
6.2.3 Noise and distortion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
Nonlinear filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.2.4 Filtering two-dimensional signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
6.2.5 Image processing with FIR filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
6.3 Some M ATLAB commands for this lab . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6.4 Demonstrations in the Lab Section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
6.5 Laboratory Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

7 Decoding DTMF: Filters in the Frequency Domain 113

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
7.1.1 “The Question” . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
7.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
7.2.1 DTMF signals and Touch ToneTM Dialing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
7.2.2 Decoding DTMF Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
Step 1: Bandpass Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
Step 2: Determining filter output strengths . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
Step 3: “Detect and Decode” . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
7.2.3 Decoder Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
7.2.4 Sidenote: Searching Parameter Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
7.3 Some M ATLAB commands for this lab . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
7.4 Demonstrations in the Lab Section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
7.5 Laboratory Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

8 Classification and Vowel Recognition 125

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
8.1.1 “The Question” . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
8.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
8.2.1 An Introduction to Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
8.2.2 A classification example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
8.2.3 A few more classification examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
8.2.4 Formalizing the feature classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
8.2.5 Measuring the performance of a classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
Data usage when designing classifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
8.2.6 Vowel Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
8.3 Some M ATLAB commands for this lab . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

viii The University of Michigan, All rights reserved

 Contents

8.4 Demonstrations in the Lab Section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

8.5 Laboratory Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

9 Filter Design, Modeling, and the z-Plane 141

9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
9.1.1 “The Question” . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
9.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
9.2.1 Filters and the z-transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
9.2.2 FIR Filters and the z-transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
9.2.3 IIR filters and rational system functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
Poles and zeros at the origin and at infinity . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
Pole-zero plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
9.2.4 Graphical interpretation of the system function . . . . . . . . . . . . . . . . . . . . . . . . . 147
9.2.5 Poles and stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
9.2.6 Filter design using manual pole-zero placement . . . . . . . . . . . . . . . . . . . . . . . . . 149
9.2.7 Design of Standard Filter Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
9.2.8 Modeling Vowel Production . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
All-pole analysis and vowel classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
9.3 Some M ATLAB commands for this lab . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
9.4 Demonstrations in the Lab Section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
9.5 Laboratory Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156

The University of Michigan, All rights reserved ix

Contents

x The University of Michigan, All rights reserved

List of Figures

1.1 Examples of continuous-time and discrete-time signals . . . . . . . . . . . . . . . . . . . . . . . . . 23

1.2 Signal value distribution and a discrete histogram approximation . . . . . . . . . . . . . . . . . . . . 28
1.3 An “overview” block diagram for a “signal present” detector. . . . . . . . . . . . . . . . . . . . . . . 30
1.4 A detailed block diagram for the “signal present” detector. . . . . . . . . . . . . . . . . . . . . . . . 30

2.1 Examples of positively correlated, uncorrelated, and anticorrelated signals. . . . . . . . . . . . . . . . 38

2.2 Signals in a radar detection system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.3 A generalized block diagram for a correlation-based detection system. . . . . . . . . . . . . . . . . . 40
2.4 Code signals for a simultaneous communication system . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.5 Example of several overlapping communications signals . . . . . . . . . . . . . . . . . . . . . . . . 42
2.6 Example of a M ATLAB figure with subplots. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

3.1 Three-dimensional plot of a complex exponential signal. . . . . . . . . . . . . . . . . . . . . . . . . 54

3.2 System diagram for the “amplitude and phase calculator.” . . . . . . . . . . . . . . . . . . . . . . . . 56
3.3 System diagram for the “frequency, amplitude, and phase estimator.” . . . . . . . . . . . . . . . . . . 59

4.1 Time-domain and frequency-domain representations . . . . . . . . . . . . . . . . . . . . . . . . . . 66

4.2 A comparison of the Fourier Series and DFT coefficients . . . . . . . . . . . . . . . . . . . . . . . . 75
4.3 The DFT of a harmonic series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

5.1 A block diagram of a general data compression system. . . . . . . . . . . . . . . . . . . . . . . . . . 83

5.2 A block diagram of a general image encoder/compressor. . . . . . . . . . . . . . . . . . . . . . . . . 86
5.3 A block diagram of a general image decoder/decompressor. . . . . . . . . . . . . . . . . . . . . . . . 87
5.4 Block diagram of a direct quantizer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.5 Plot of the tradeoff between coding rate and distortion . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.6 A block diagram of a transform coder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

6.1 A graphical illustration of filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

6.2 Input and output of a 5-point moving average filter. . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
6.3 A block diagram of a noise-reduction system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.4 The coefficients of a two-dimensional moving average filter . . . . . . . . . . . . . . . . . . . . . . . 105
6.5 The coefficients for g smooth filters with varying widths. . . . . . . . . . . . . . . . . . . . . . . . 108

7.1 A spectrogram of a DTMF signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115

7.2 A block diagram of the DTMF decoder system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
7.3 A comparison of half-wave and full-wave rectification . . . . . . . . . . . . . . . . . . . . . . . . . . 117
7.4 An illustration of the DTMF detector subsystem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

The University of Michigan, All rights reserved xi

List of Figures

8.1 Block diagram of a general classifier system. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126

8.2 A histogram for one-feature classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
8.3 A case where two-feature classification is necessary . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
8.4 An example where two features are not as clearly separated. . . . . . . . . . . . . . . . . . . . . . . 129
8.5 Various classification examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
8.6 Block diagram of a distance-based feature classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
8.7 The magnitude spectrum (in decibels) of four vowel signals . . . . . . . . . . . . . . . . . . . . . . . 133

9.1 A pole-zero plot of an IIR filter. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146

9.2 System function, frequency response, and pole-zero plot of an FIR filter . . . . . . . . . . . . . . . . 148
9.3 System function, frequency response, and pole-zero plot of an all-pole filter . . . . . . . . . . . . . . 148
9.4 System function, frequency response, and pole-zero plot of a general IIR filter . . . . . . . . . . . . . 148
9.5 An illustration of the various bands of a lowpass filter. . . . . . . . . . . . . . . . . . . . . . . . . . . 150
9.6 A diagram showing the larynx and vocal tract. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
9.7 A block diagram of the source-filter model of speech production. . . . . . . . . . . . . . . . . . . . 151
9.8 Plot of the spectral evolution of a vowel signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
9.9 The GUI window for Pole-Zero Place 3-D. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154

xii The University of Michigan, All rights reserved

Lab 0: An Introduction to M ATLAB

1 What is M ATLAB?
The Mathworks, Inc., makers of M ATLAB, claims that M ATLAB is “the language of technical computing.” By and
large, they are right. M ATLAB is widely used in a great number of scientific fields. For those who work with signals
and systems, M ATLAB is a de facto standard. Engineers from a wide array of disciplines, in both academia and in
industry, use M ATLAB on a regular basis. As such, a knowledge of M ATLAB will not only be useful for this course,
but for future courses and in your career as a whole. One of the main reasons for MATLAB’s popularity arises from its
wide array of uses. So what is M ATLAB?

1.1 M ATLAB is a mathematics environment that can easily handle vectors and matrices
M ATLAB was originally written to provide an easy-to-use interface to the mathematical subroutines included in LIN-
PACK and EISPACK. These two packages are sets of subroutines written in FORTRAN for a wide variety of linear
algebra operations. M ATLAB’s original focus on linear algebra means that it has very well developed capabilities for
handling vectors and matrices1 . In fact, M ATLAB is short for “Matrix Laboratory.” For our purposes, both vectors
and matrices are examples of signals – a mathematical environment that can easily handle vectors and matrices makes
working with signals just as easy.
Let’s look at an example to see exactly what this buys us. Suppose that we have two signals, x and y, each of
which is simply an array with 100 elements. How would we add these signals in a language like C++? The easiest
way probably involves the following fragment of code:

double z[100];
for(int i = 0; i < 100; i++)
{
z[i] = x[i] + y[i];
}

This is a simple enough piece of code, but it is not as clear as it could be. In M ATLAB we can simply do the following:

z = x + y;

Simply adding two signals (vectors or matrices) with the same size automatically performs an element-by-element
sum. Which of these two is easier to understand? Using this M ATLAB syntax, we can see immediately what is
happening. M ATLAB takes care of any necessary looping and variable declarations for us. This is a very common
feature in M ATLAB; many operations that you would normally need to perform explicitly in another programming
language can be performed implicitly in M ATLAB.
1 Vectors and matrices are simply one- and two-dimensional arrays, respectively

The University of Michigan, All rights reserved 1

An Introduction to M ATLAB

1.2 M ATLAB is tool for visualizing data

You are probably very familiar with how much easier it is to interpret a graph than a table of numbers or a formula.
By producing a plot of the relevant data or formula, you can gain a visual sense of what is going on that otherwise
might be lacking. This is one of the motivations behind the use of graphing calculators in high school math. Simply
put, M ATLAB is one of the best tools for visualizing data that is currently available.
You will find that these capabilities very useful in your study of signals and systems. By looking at a signal, you
can often gain some insight into how it behaves. The same applies to systems. Certain systems are said to “smooth”
signals because of the visual appearance of the resulting signal. In certain cases (like image processing), the visual
result of a system is the primary reason for its use.

1.3 M ATLAB is a prototyping language

In many respects, M ATLAB is like a UNIX shell. It has the same sort of interactive interface for normal usage, but it
also has most of the standard programming language constructs like loops and conditional statements. You can put
commands into a file and call it as a script. Alternatively, you can write functions with input and output parameters.
The main difference between M ATLAB and programming languages like C++ is the ease with which you can
implement algorithms (especially mathematical algorithms). This is because M ATLAB operates at a higher level
than many other programming languages. It is also usually easier to understand M ATLAB code than code in other
programming languages. The sum-of-vectors example given above is a prime example of this. All of this makes
M ATLAB a very good prototyping language. It is easy to whip up a “proof of concept” program in M ATLAB to
make sure that your algorithm actually works. Then, you can code a “development” version using a more traditional
compiled programming language.

1.4 M ATLAB can do more...

One of the key rules of thumb to remember about M ATLAB is that it can perform almost any mathematical task you
could want. Often, there will be a built-in function to do what you want. If it’s not a part of the main M ATLAB
distribution, it is probably available as part of an add-on called a toolbox. Some toolboxes can be purchased from the
Mathworks, while others are developed and distributed for free by third party developers.
In this course, we will be focusing on the core M ATLAB distribution and the Signal Processing Toolbox. (We
will also be doing some image processing, but you will not need the Image Processing Toolbox for this course.) We
recommend that you consider purchasing a version of MATLAB and the Signal Processing Toolbox; you find it to be
useful throughout your academic career.

2 Demos for the first tutorial lab section

1. Recording, displaying, and manipulating signals in M ATLAB
2. Image Compression via JPEG
3. DTMF (Touch-tone) telephone tones

3 Using M ATLAB: The basics

3.1 Starting M ATLAB
The first step to using M ATLAB is to bring up the program on your computer system. This series of laboratories was
designed for Windows-based computers, so we recommend using these machines if possible. Starting M ATLAB on a

2 The University of Michigan, All rights reserved

 3.2 How to get help

Windows machine or a Macintosh usually requires finding the appropriate icon either on the desktop or in the Start
menu2 . At a UNIX system, simply typing “matlab” should be sufficient. Note that you can run M ATLAB remotely on
UNIX servers through telnet or ssh, but M ATLAB versions 6 and higher generally require an X-windows connection to
run3 . When M ATLAB is finished loading, you’ll see the M ATLAB program window, possibly with several subwindows.
The most important window is the command window, which contains a command prompt that looks something like
this:

>>

3.2 How to get help

So now what do you do? Well, the first step is to make use of M ATLAB’s single most useful command:

>> help

See that list of categories? You can call help on any of these categories to get an organized list of commands with
brief discussions. Then, you can call help on any of the commands for a complete description of that command.
The description also includes a “see also” line near the bottom which suggests other commands that may be related
to the one you’re looking at. Select a category that looks interesting and call help on it. Do the same for whichever
command strikes your fancy. For instance:

>> help elfun

>> help abs

Most often you’ll use help in this last capacity. Note that help abs lists commands related to the absolute value
function as well.
Unfortunately the traditional help system isn’t so helpful if you don’t know the name of the command you’re
looking for. One way around this is to use the lookfor command. For instance, if you know you’re looking for a
function that deals with time, you can try:

>> lookfor time

This searches the first line of the every help description for the word “time.” This can take a while, though (depending
upon your system’s configuration). You should get into the habit of reading the help on every new command that
you run across. So call help on both help and lookfor. There’s some useful information there.
Another very good source of help is the M ATLAB helpdesk. It may or may not be available on your system; to
find out, simply try:

>> helpdesk

If it is available, you will see a help window. The M ATLAB helpdesk contains all of the help pages that you can
find using help or lookfor, along with many other useful documents. The helpdesk is also easily searchable
(and often much faster than lookfor), so you would benefit from becoming familiar with this tool.

3.3 Using M ATLAB as a calculator (with variables)

Not surprisingly, you can use M ATLAB to do arithmetic. It operates very much like you might expect, employing infix
arithmetic like that used on standard calculators. M ATLAB can evaluate simple expressions or arbitrarily complicated
ones with parentheses used to enforce a particular order of operations.
2 CAEN machines may have multiple versions installed; you should try to locate the most recent version of M ATLAB.
3 Versionsof M ATLAB prior to 6.x run by default in a terminal window, without an X-windows connection.

The University of Michigan, All rights reserved 3

An Introduction to M ATLAB

>> 6 * 7
>> (((12 + 5) * 62/22.83) - 5)ˆ2.4
(The ˆ operator performs exponentiation.) Notice that when you execute these commands, MATLAB indicates that
ans = 7.4977e+003 (or whatever the answer is). This indicates that the result has been stored in a variable called
ans. We can then refer to this quantity like this:
>> 0 * ans
>> ans + 1
It is important to note that each of these commands overwrites ans. If we want to save an answer, we can simply
perform assignment, like this:
>> my_variable = 42
This is the only declaration of my_variable that is needed, and we can use this variable later just as we could with
ans. Further, my_variable will retain its value until we explicitly assign something else to it.
We can also remove variables with the command clear. Typing who or whos will list what variables we have
in our workspace.
Using variables, then, is straightforward.
>> x = 5.4
>> y = 2
>> z = (my_variable*y)ˆx
Note that sometimes you don’t need or want to see what M ATLAB returns in response to a particular command. To
suppress the output, simply add a semicolon, ;, after the command. Try any of the above commands with and without
the semicolon to see what this does.
We also have access to a wealth of standard mathematical functions. Thus, we can if we want to calculate the sine
of the square-root of two and store it in a variable called var, we simply type:
>> var = sin(sqrt(2))
Type help elfun to see how to call most of the elementary mathematical functions like these.
There are also a number of constants built into M ATLAB that are√very useful. The number π is referred to as
pi (note that M ATLAB is case sensitive!). Both i and j default to −1, but you can still use either (or both) as
variable names if you like. You should glance at help i so that you can see the various options for building complex
numbers. Note that you can overwrite variables like pi, i, and j, but then you will not be able to use their special
properties. The special variables (and matrices) built-in to M ATLAB are listed under help elmat.

4 Vectors, Matrices, and Arrays

So far, we’ve been using M ATLAB to deal with scalar numbers. The real power of M ATLAB, though, comes from
its ability to handle vectors and matrices. In M ATLAB, vectors and matrices are simply one-dimensional and two-
dimensional arrays, respectively. An array is simply a collection of numbers, each of which is indexed by some
ordered set of numbers. For instance, a vector is indexed by a single integer, while a matrix is indexed by an ordered
pair. The number of indices is equal to the dimension of the array. For instance, consider the following vector and
matrix:  
1  
 2  1 2 3
v=  3 
 M = 4 5 6  (1)
7 8 9
4

4 The University of Michigan, All rights reserved

 4.1 Constructing arrays

To access the 3 from vector v, we simply need to know that it is in the third row. (In M ATLAB, we use v(3) to access
this element.) Thus the vector is one-dimensional. To access the 6 in the matrix M , though, we need to know that it
is in the second row and the third column. We index the 6 using the pair (2,3), and so matrix is two-dimensional. (In
M ATLAB, we use M(2,3) to access this element.) M ATLAB arrays can have any number of dimensions. In practice,
though, we will only need vectors and matrices.

4.1 Constructing arrays

There are many different ways to build up and manipulate arrays in M ATLAB. For instance, consider (and execute) the
following commands:

>> a = [1 2 3 4 5 6 7]
>> b = [1, 2, 3, 4, 5, 6, 7]
>> c = [1; 2; 3; 4; 5; 6; 7]
>> d = [1 2 3 4; 5 6 7 8; 9 10 11 12]

The first two commands both build up the same vector, a 1×7 row-vector4 . The third command builds up a 7×1
column-vector with the same elements. The fourth command builds a 3×4 matrix.

4.2 Concatenating arrays

The comma (or the space) within the square brackets concatenates horizontally and the semicolon concatenates verti-
cally. The elements being concatenated do not need to be scalars, either:

>> e = [a b]
>> f = [a; b]
>> g = [c d]

Oops! That last command produced an error. When concatenating arrays, the concatenated arrays must have sizes
such that the resulting array is rectangular.

4.3 Transposition and “flipping” arrays

The single apostrophe, ', is M ATLAB’s transposition operator. It will turn a row-vector into a column-vector and vice
versa. Similarly, it will make an n × m matrix into an m × n matrix. To see how this works, type d' and look at
the results. (Warning: ' is actually a complex conjugate transpose, so complex numbers will have the sign of their
imaginary parts changed. To perform a straight transposition, use the .' operator. For real arrays, both operators are
identical.) Other useful commands for matrix manipulation include flipud and fliplr, which mirror matrices
top-to-bottom and left-to-right, respectively. Look at help elmat for other useful functions.

4.4 Building large arrays

Building small arrays by hand is fine, but it can become very tedious for larger arrays. There are a number of commands
to facilitate this. The ones and zeros commands build matrices that are populated entirely with ones or zeros. The
eye command builds identity matrices. repmat is especially useful for making matrices out of vectors. diag
builds diagonal matrices from vectors, or returns the diagonal (vector) of a matrix. Check the help for all of these
commands. For an example, try these:
4 In M ATLAB, indices are given as row × column.

The University of Michigan, All rights reserved 5

An Introduction to M ATLAB

>> ones(5,3)
>> zeros(3,4)
>> zeros(5)
>> eye(4)

4.5 The colon operator

The colon operator is one way of creating long vectors that are useful for indexing (see the next section). Execute the
following commands:

>> 1:7
>> 1:2:13
>> 0.1:0.01:2.4

Each of these commands defines a row-vector. With only two arguments, as in the first command, the colon operator
produces a row vector starting with the first argument and incrementing by one until the second argument has been
reached. The optional middle argument (seen in the second two commands) provides a different increment amount.
The colon operator is extremely useful, so it is recommended that you check out help colon for more details. Play
with some other combinations of parameters to familiarize yourself with the behavior of this operator.

5 Array Arithmetic
M ATLAB allows you to perform mixed arithmetic between scalars and arrays as well as two different types of arith-
metic on arrays. Mixed scalar/array arithmetic is the most straightforward. Adding, subtracting, multiplying or divid-
ing a scalar from an array is equivalent to performing the operation on every element of the array. For instance,

>> [5 10 15 20]/5

returns the vector [1 2 3 4].

It is also useful to note that most of the provided mathematical functions (like sqrt and sin) operate in a similar
element-by-element fashion. Thus, the commands

>> t = 0:.1:pi;
>> sin(t)

return a 32-element vector (the same size as t) containing the sine of each element of t.
If we have two arrays, addition and subtraction is also straightforward. Provided that the arrays are the same size,
adding and subtracting them performs the operation on an element-by-element basis. Thus, the (3,4) element in the
output (for instance) is the result of the operation being performed on the (3,4) elements in the input arrays. If the
arrays are not the same size, M ATLAB will generate an error message.
For multiplication, division, exponentiation, and a few other operations, there are two different ways of performing
the operation in question. The first involves matrix arithmetic, which you may have studied previously. You may recall
that the product of two matrices is only defined if the “inner dimensions” are the same; that is, we can multiply an mxn
matrix with an nxp matrix to yield an mxp matrix, but we cannot reverse the order of the matrices. Then, the (p,q)
element of the result is equal to the sum of the element-by-element product of the pth row of the first matrix and the q th
column of the second. Division and exponentiation are defined with respect to this matrix product. It is not imperative
that you recall matrix multiplication here (most likely you will see it in a linear algebra course in the future); however,
it is important that you note that in M ATLAB the standard mathematical operators (*, /, and ˆ) default to these forms
of the operations.

6 The University of Michigan, All rights reserved

 6 Indexing

A form of multiplication, division, and exponentiation for arrays that is more useful for our purposes is the element-
by-element variety. To use this form, we must use the “dot” forms of the respective operators, .*, ./, and .ˆ). Once
again, the arrays must have the same dimensions or M ATLAB will return an error. Thus, the commands

>> [1 2 3 4].*[9 8 7 6]
>> [7; 1; 4]./(1:3)'
>> [5 6 7].ˆ[2 3 4]
>> 2.ˆ[1 2 3 4 5 6]

perform element-by-element multiplication, division, and two slightly different forms of exponentiation. Note that the
.ˆ form is necessary even for scalar-to-array exponentiation operations.
The array arithmetic capabilities of M ATLAB contribute greatly to its power as a programming language. Using
these operators, we can perform mathematical operations on hundreds or thousands of numbers with a single com-
mand. This also has the side effect of simplifying M ATLAB code, making it shorter and easier to read (usually).

6 Indexing
6.1 Basic indexing
To make arrays truly useful, we need to be able to access the elements in those arrays. First, let’s fill a couple of arrays:

>> a = 5:5:60
>> d = [9, 8, 7, 6 ; 5, 4, 3, 2]

Now, let’s access elements in them:

>> a(6)
>> a(3) = 12
>> d(2,3)

The first command retrieves the sixth element from the vector a. The second assigns a number to the third element of
the same vector. For the third command, the order of the dimensions is important. In M ATLAB, the first dimension
is always rows and the second dimension is always columns. Note particularly that this is the opposite of (x, y)
indexing. Thus, the third command retrieves the element from row two, column three.

6.2 Single number indexing

We can also index into matrices using single numbers. In this case, the numbers count down the columns. This is
called “column-major” and is the opposite of array indexing in C or C++. For instance, notice what happens when you
use the following commands:

>> d(2)
>> d(3)
>> d(7)

6.3 Vector indexing

It is not necessary to index arrays only with scalars. One of the most powerful features of M ATLAB is the ability to
use one array to index into another one. For instance, consider the following commands:

The University of Michigan, All rights reserved 7

An Introduction to M ATLAB

>> a([1 4 6])

>> b(3:7)
>> c(2:2:end)

These commands return a subset of the appropriate vector, as determined by the indexing vector. For instance, the
first command returns the first, fourth, and sixth elements from the vector a. Notice the use of the end keyword in
the third command. In an indexing context, end is interpreted as the length of the currently indexed dimension. This
is particularly useful because M ATLAB will return an error if you try to access the eighth element of a seven-element
vector, for instance. In general, indices must be strictly positive integers less than the length of the dimension being
indexed. Thus, unlike C or C++, the indices begin at one rather than at zero.
Using multiple indices into multi-dimensional arrays is more complicated than doing so with vectors, but in some
cases it can be extremely useful. Consider the following commands:

>> d([1 3],2)

>> d([2 3],[1 4])
>> d(2,:)

The first command, as you might expect, returns the first and third elements of the second column. The second
command returns the second and third rows from the first and fourth columns. Note particularly that this command
does not return the individual elements at (2,1) and (1,4). (To index individual elements in this manner, we need to use
single-index method along with the sub2ind command). The colon operator in the second command is a shortcut
for 1:end; thus, the third command returns all of the second row.

6.4 Finding the size of an array

Two very useful commands that can be used to facilitate indexing are size and length. size returns a vector
containing the length of each dimension of an array. Alternately, size can be used to request the length of a single
dimension. length is primarily useful for vectors when you’re not sure about their orientation. length returns the
length of the longest dimension. Thus, length(v) is the same whether v is a row-vector or a column-vector, but
size(v,1) will only properly return the length of a column-vector.

6.5 Vector indexing to modify arrays

It is important to note that all of these indexing techniques are used not only to retrieve many elements from an array
but also to set them. When performing array assignment, you must be careful to make sure the array being assigned
has the same size as the array to which it is being assigned. For instance, consider the following command:

>> d([1 3],[2 4]) = [9 8; 7 6]

Note that both of the matrixes on the left and right of the equal sign are 2 × 2, so the assignment is valid. Look at the
results of this command and make sure you understand what it does and why.

6.6 Conditional statements and the “find” command

One last command that is extremely useful in context of indexing in M ATLAB is find. find will return a vector
containing the indices of any nonzero elements in an array. Note that find uses the single-index indexing scheme that
was mentioned earlier. At first glance, this has relatively few uses; however, it is in fact extremely useful because of the
behavior of conditional statements in M ATLAB (i.e., >, <, and ==). The command a > 5 will return an array with
the same size as a, but with each element either 1 or 0 depending on whether or not it is greater than 5. Using find
on this array will provide the indices of elements greater than 5. One particularly good use of the find command is

8 The University of Michigan, All rights reserved

 7 Data Visualization

the following contexts. Suppose you wish to set all negative elements in a matrix to zero. You can do this with a single
command like so:

>> m = [-1 5 10; 3 -8 2; -4 -9 -1];

>> m(find(m < 0)) = 0;

Alternately, if you wish to square every element that is greater or equal to 4, you can use the find command twice in a
single line, like this:

>> m(find(m >= 4)) = m(find(m >= 4)).ˆ2;

7 Data Visualization
7.1 Using “plot”
So now we know how to build arbitrarily large arrays, populate them with interesting things, and get individual
elements or sections of them. However, pouring over pages and pages of numbers is generally not much fun. Instead,
we would really like to be able to visualize our data somehow. Of course, MATLAB provides many options for this.
Let’s start by building a vector we can use throughout this section, and then looking at it. Execute the following
commands:

>> x = sin(2*pi*(1:200)/25);
>> plot(x);
>> zoom on;

The first command builds up a sine wave, and the second command plots it. A window should have popped up with
a sine wave in it. Notice the y-axis extents from -1 to 1 as we would expect. Using this form of plot, the x-axis is
labeled with the index number; that is, our vector has 200 elements, and so the data is plotted from 1 to 200. The third
command turned on M ATLAB’s zooming capabilities. If you left-click on the figure, it will zoom in; right-clicking5
will zoom out. You can also left-click and drag to produce a zoom box that lets you control where the figure zooms.
Experiment with this zoom tool until you’re comfortable with it. Depending on the version of M ATLAB that you are
using, there may also be an icon of a magnifying glass with a + in it above the figure; clicking this icon will also enable
and disable zoom mode. Also try zoom xon for zooming only the horizontal axis.

7.2 Interpolation; line and point styles

If you zoomed in closely enough on the plot, you probably noticed that the signal isn’t perfectly smooth. Instead, it
is made up of line segments. This is because our vector, x, is made up of a finite collection of numbers. M ATLAB
defaults to interpolating between these points on the plot. You can tell M ATLAB to show you where the data points
are, or to not interpolate, by changing the line and point styles. Try each of these commands and look at the results
before executing the next one:

>> plot(x,'x-')
>> plot(x,'o')
>> plot(x,'rd:')

help plot lists the various combinations of characters that you can use to change line styles, point styles, and
colors.
5 For Mac users, I believe you double-click to zoom out all the way.

The University of Michigan, All rights reserved 9

An Introduction to M ATLAB

7.3 Axis labels and titles

Often, we want to indicate what each axis of a plot represents or add a figure title. The commands xlabel, ylabel,
and title do this for us. For instance:

>> xlabel('Time (seconds)');

>> ylabel('Amplitude');
>> title('Plot of x[n]');

Note that the single tick marks, ', delimit strings that are passed to these commands.

7.4 Commands related to “plot”

There are a few similar commands for plotting vectors as well. Try these commands, and make sure you zoom in on
each one so you can see the results:

>> stem(x)
>> stairs(x)
>> bar(x)

In this course, you will most often be using the plot and stem commands. Each is useful in a somewhat different
context.

7.5 Plotting with an x-axis

When you checked the help for plot (you did look at the help, didn’t you?), most likely you noticed that there are
some more explicit ways to use the function. There is an optional first parameter that gives the x-position of each data
point. Thus, we use plot for x-y scatter plots and other things. Calling plot without the first parameter is equivalent to
the following command:

>> plot(1:length(x),x,'x-');

Sometimes, we’ll have a time axis that we want to plot against. For instance,

>> t = 0:.01:1.99;
>> plot(t,x);

This scales the time axis to match t. We will find this very useful when working with sampled signals.

7.6 Plotting multiple vectors on the same figure

It possible (and often desirable) to plot multiple vectors simultaneously. One way (which is probably the easiest to
remember) requires a set of parameters for each vector. Execute the following commands:

>> y = .8*sin(2*pi*(1:200)/14 + 0.5);

>> plot(t,x,'go-',t,y,'rx--');

This plots x and y versus t on the same figure with different line types. Note that the line style arguments are optional;
without them, M ATLAB will plot each curve using a different color.
The hold command provides another method of plotting several curves on the same figure. When we type
hold on, an old figure will not be erased before a new one is plotted. To add a curve to the plot we produced above,
use the commands:

10 The University of Michigan, All rights reserved

 7.7 Legends

>> hold on;

>> plot(t,.3*x,'ks:');
>> hold off;

A third way to plot multiple lines simultaneously makes use of the fact that plot will plot the columns of a matrix
as separate lines. Execute the following commands.

>> plot([x; y]');

7.7 Legends
You can add a legend to a plot using the legend command like this:

>> legend('Data set 1', 'Data set 2');

The legend command can take any number of parameters; usually, though, you want one string for each data set on
your plot.

7.8 Putting several axes on one figure

Often we’ll want to plot two vectors next to one another but not on the same set of axes. To do this, we use the
subplot command. subplot takes three parameters: the number of rows, the number of columns, and the figure
number. Thus, the following command the fourth subplot in an array of subplots with three rows and two columns.

>> subplot(3,2,4);
(Notice that it opens the fourth counting across the rows, as you would read a page. This is notably different from
single number indexing of M ATLAB arrays.)
Now, to put several plots in subplots like this, we simply execute several subplot commands like this:

>> subplot(2,1,1);
>> plot(1:10, (1:10).ˆ2);
>> subplot(2,1,2);
>> plot(1:10, (1:10).ˆ3);

7.9 Two-dimensional arrays

You’re probably not surprised by now that M ATLAB also has facilities for visualizing two dimensional arrays. Let’s
look at some of them. First, we need an interesting matrix to look at. Execute the following command:

>> z = membrane(1,50);

We now have a 101x101 matrix of numbers. The most straightforward way to look at this data is using the imagesc
command, which displays the matrix as though it were an image. Execute the following commands:

>> imagesc(z); axis xy; colorbar;

Our surface has been displayed in color. Notice the colorbar along the right side of the image, which tells what values
the various colors map to. This type of display, where different colors are used to represent different values, is known
as a pseudocolor display. If we look at the image we’ve got a “high” spot in the lower right that tapers off to “low”
regions around the outside. The surface also has an overall L-shape. Another way to visualize this uses the contour
command. Try this:

The University of Michigan, All rights reserved 11

An Introduction to M ATLAB

>> contour(z,20); colorbar;

This display, the contour plot, shows us lines of constant height. This is the way that meteorologists usually display
atmospheric pressure on weather maps.
We also have some more interesting options. Try each of the following commands separately:

>> mesh(z); rotate3d on

>> surf(z); rotate3d on

Now we have some “3-D” visualizations of our surface. If you click-and-drag the plot, you should be able to rotate
the surface so that you can see it from various directions. Experiment with this until you’re comfortable with how it
works. Notice what happens if you look at the surface from directly above.
M ATLAB has some very powerful tools for data visualization; here, you’ve seen only a small sampling. There
many more. If you’re interested in exploring this topic further, check help graph2d, help graph3d, and help
specgraph.

8 Programming in M ATLAB
Programming in M ATLAB is really just like using the M ATLAB command line. The only difference is that commands
are placed in a file (called an M-file) so that they can be executed by simply calling the file’s name. We’ll also see
that M ATLAB has many of the same control flow structures, like loops and conditionals, as other, more traditional
programming languages.

8.1 Paths and working directories

Before we jump into programming in M ATLAB we need to make a few comments about files in M ATLAB. M ATLAB
has access to a machine’s file system in roughly the same way a command-line based operating system like DOS
or UNIX. It has a “present working directory” (which you can see with the command pwd); any files in the present
working directory can be seen by M ATLAB. You can change the present working directory in roughly the same way
that you do in DOS or UNIX, using the cd command (for “change directory”). M ATLAB also has a “path,” like the
path in DOS or UNIX, which lists other directories that contain files that M ATLAB can see. The path command will
list the directories in the path. We’ll be making a few files in this tutorial, and you’ll need to store commands in files
when doing the laboratories. You’ll probably want to make a directory somewhere in your personal workspace, cd to
that directory, and store your files there. Unless you’re working on your own system, do not store them in the main
M ATLAB directory; if you do, the system’s administrator will probably become very irritated with you.

8.2 Types of command files in M ATLAB

There are two types of files containing commands that M ATLAB can call, scripts and functions. Both use the “.m” file
extension (and, thus, are called m-files. A script is nothing but a list of commands. When you call the script (by simply
typing in the script’s filename), M ATLAB will execute all of the commands in the file and return to the command line
exactly as if you had typed the commands in by hand. Functions are different in that they have their own workspace
and variables. We pass information to a function by means of input parameters, and receive information from the
function through output parameters.

12 The University of Michigan, All rights reserved

 8.2 Types of command files in M ATLAB

M ATLAB scripts

Start the M ATLAB editor using the command edit6 . Then, place the following lines in the text file and save it as
“hello.m”.

% hello.m -- Introductory 'Hello World' script

% These lines are comments, because they start with '%'

hw = 'Hello World!'; % Comments can appear on the same lines

disp(hw); % as commands, again after a '%'

Now execute it by typing hello at the M ATLAB prompt. (Remember that the file needs to be in your present working
directory or on the path for M ATLAB to see it – cd to the correct directory if necessary). As a result of executing the
script, you should now have a variable ’hw’ in your workspace (remember, who lists variables in your workspace).
Note that scripts make use of (and possibly overwrite) variables in your base workspace. For further information on
scripts, type help script.

M ATLAB functions

The second type of file that we can put commands in is called a function. A function communicates with the current
workspace by passing variables called parameters. It also creates a separate workspace so that it’s variables don’t get
mixed up with whatever variables you have in your current workspace. Note that most M ATLAB commands are also
functions, and the M-file code is available for most of them. You can see the code by using the type command, for
instance as type fliplr.
Using your text editor, make a new file that contains the following lines and save it as “hello2.m.”

% hello2.m -- Introductory 'Hello World' function

% Try typing 'help hello2' when you're done, and see what happens
%
% function output_param = hello2(input_param)

function output_param = hello2(input_param)

% The line above tells MATLAB that this is a function
% with one input and one output parameter

hw2 = ['Hello World! x' num2str(input_param)];

disp(hw2);
output_param = hw2;

To call this function, type hello2(2). Note that once you’ve done this, the variable hw2 does not show up in
your workspace. However, the data that was stored in output_param (the output parameter) has been placed in
ans. This is exactly what happens if you called a M ATLAB built-in function without supplying an output parameter.
Similarly, the ‘2’ is an input parameter which is passed into the function. When a function is executed, it will not have
any variables defined except those defined inside the function itself and the input parameters. Note that a function does
not need to have either input parameters or output parameters. For further help on this, type help function at the
M ATLAB prompt.
6 While you can use any text editor for editing M ATLAB code, the M ATLAB editor has a number of useful features for doing so. UNIX versions

of M ATLAB prior to version 6 did not include a built-in editor.

The University of Michigan, All rights reserved 13

An Introduction to M ATLAB

Scripts versus functions

There are some situations when scripts are more convenient to use, and others where functions are more useful.
Scripts are useful for automating some set of commands that you would otherwise need to type into the command line
repeatedly. Scripts have full access to your variables, which can be both positive and negative. On the one hand, we do
not need to explicitly pass every variable needed to the script. On the other hand, scripts are generally dependent upon
the state of the workspace variables. When running scripts, we also risk overwriting variables that we do not wish to
overwrite.
Functions, on the other hand, are good for when we wish to perform some task repeatedly with different data. If
we want to run a script many times with a different variable setting, we may need to change the variable by hand in
the code. With a function, we simply pass that variable into the function as a parameter. Because of their separate
workspace, we are guaranteed that a function is only dependent upon the parameters we pass into it. This makes
a function more portable from one situation to another (since we don’t need to worry about the state of the calling
workspace), and generally forces the programmer to be clear about variable initialization and the like. One downside
of functions is that it is somewhat harder to see the results of “internal” computations without resorting to debugging
(see Section 9).
The writing of functions versus scripts is very much a matter of personal preference. However, we tend to prefer
using functions in any situation where doing so is not prohibitively difficult. The encapsulation of data allows for the
reuse of functions much more readily than scripts. Perhaps it is telling that nearly all built-in M ATLAB commands are
functions rather than scripts.

8.3 Control Structures

In M ATLAB we also have a number of programming constructs at our disposal. While primarily used in M-files,
these constructs can also be used at the command line. However, anything complicated enough to need a loop or an
if-statement is usually worth putting into an M-file. Let’s look at the most typical types of programming constructs.

Loops
The for loop is used to execute a set of commands a certain number of times, while also providing an index variable.
Consider the simple loop here:

for index = 1:10

disp(index);
end

This loop executes the disp command ten times. The first time it is executed, index is set to 1. Thereafter, it is
incremented by one each time the commands in the loop are executed. Note that the colon form of the for loop is
not mandatory; any row-vector can be used in its place, and the index (which, of course, can be renamed) will be
sequentially set to each of the elements in the vector from left to right.
We can use while loops in a similar manner. Consider this:

ct = 10;
while ct > 0.5
ct = ct/2;
disp(ct);
end

As long as the conditional after while is true, the loop will be executed.

14 The University of Michigan, All rights reserved

 8.4 Strings and string output

Conditional statements
A more traditional method of conditional execution comes from the if-else statement. Consider this:
if pi > 4
disp('Pi is too big!');
elseif pi < 3
disp('Pi is too small!');
else
disp('Pi is just about right.');
end
Here, M ATLAB will first check the conditional, pi > 4. If this is true, the first display command will be executed
and the remainder of the if-else statement will be skipped (that is, none of the other conditionals will be tested).
If the first conditional is false, M ATLAB will begin to check the remaining conditionals. There can be any number
of elseif statements in this construct (including none), and the else statement is entirely optional. If you have a
large number of chained conditionals, you might consider using the switch-case construct (type help switch
or help case).

8.4 Strings and string output

In hello2 above, we constructed a string and displayed it. Though not so useful at the command line, in programming
we often want do work with strings and display them. In MATLAB, strings are delimited by the single tick-mark '.
Thus, 'STRING' is treated as a literal string, rather than being interpreted as a variable. Strings, though, are just
row-vectors of characters. This means that we can build strings using the same vector concatenation operators that we
presented earlier. Thus, the following command:
>> [ 'string' 'test' ]
outputs the string 'stringtest'.
Rather than echoing strings (or numbers, for that matter) by omitting the semicolon, we can also use the disp
command. Notice the difference when we call this command:
>> disp([ 'string' 'test' ]);
Also, for any C programmers in the audience, note that you can perform formatted string output with fprintf and
sprintf.
It is often useful to convert numbers to strings. We can use the num2str command to do this. Consider this:
>> for counter = 1:10
>> disp(['Percent completed: ' num2str(10*counter) '%']);
>> end
In this way, we can produce formatted output without using fprintf or sprintf.
For more information on strings, look at help strings and help strfun.

9 Debugging your M ATLAB code

Inevitably, when you put M ATLAB commands into a file as a script or a function, you will make mistakes and need to
locate them. Because of its interpreted environment, M ATLAB is actually one of the most pleasant languages to debug.
And, as is always the case when debugging code, there are many ways to accomplish this.

The University of Michigan, All rights reserved 15

An Introduction to M ATLAB

If you are executing a script or function and M ATLAB encounters an error, it will immediately print the line number
of the function on which the error occurred. If the error occurs in a file other than the calling file, a call stack will be
printed. This listing shows which file called which other files and on what line number. This allows us to pinpoint the
source of the error quickly.
One of the simplest ways to debug is a method you are probably familiar with from other programming languages.
We can force M ATLAB to print strings or variables using the disp command or by placing the variable name on a line
by itself without a semicolon. This way, we can display for loop counters or other relevant variables to determine
what they contain and exactly when in the program flow the code “breaks.”
The real power of M ATLAB debugging comes from our ability to ”break” at any point in the code and then proceed
to execute any M ATLAB commands. There are a number of ways to do this. For instance, you can tell M ATLAB
to stop and enter “debug mode” whenever an error is encountered. When you’re in debug mode, the command line
changes to K>>. You will then have access to all of the variables that are in scope at the time. Turn on this option with
the command

>> dbstop if error

To turn it off again, use the command

>> dbclear if error

We can also set and clear breakpoints elsewhere in the code using the same commands. To set (and then clear) a
breakpoint in hello2.m at line 11, call

>> dbstop in hello2 at 11

>> dbclear in hello2 at 11

dbstatus will show all breakpoints that are currently active. Note that if you try to set a breakpoint at a non-
command line (such as a comment), the breakpoint will be set at the next valid command.
Another useful command is dbstep, which advances one command in the m-file. If you call dbstep in or
dbstep out, you can step into and out of called functions (that is, you traverse up and down the call stack, which
contains a list of which functions have been called to reach the current point in the code). dbstack lists the current
call stack including your current file and the line number in this file. dbtype types all or parts of an m-file. Eventually,
you’ll want to get out of debug mode, so you can call dbquit to halt execution of the file or dbcont to continue
execution until the end of the file or the next breakpoint. In general, help debug is the starting point in the help
system for learning about the M ATLAB command line debugger.
If you are running M ATLAB on a Windows system (or possibly a Macintosh), the debugger is also available
through the built in editor. The exact implementation depends on your system and the version of M ATLAB, but usually
breakpoints will show up as red circles next to commands. In debug mode, the current command will be pointed to
with an arrow, so you can follow where you are in the code. There are typically shortcut keys and menu items to insert
and remove breakpoints, step through the code, and toggle flags such as stop-if-error.
If you save a file that has breakpoints, you may find that your breakpoints disappear. This can be very annoying, so
there is an alternative method of entering debug mode. Placing the command keyboard into your code is effectively
the same as placing a breakpoint in the code, such that you can execute commands before returning to program
execution (with the command return).
There are a number of error types that you are likely to encounter. One very good rule of thumb says that if an
error occurs inside a M ATLAB function, the error is almost assuredly in the calling function. Usually this means
that the function is being passed improper parameters; check the call stack or dbstep out until you find the line in
your program which is causing problems. Other common errors include indexing errors (indexing with 0 or a number
greater than the length of the indexed dimension of a variables) and assignment size mismatches. MATLAB is usually
pretty descriptive with its error messages once you figure out how to interpret what it is saying. As is usually the

16 The University of Michigan, All rights reserved

 9 Debugging your M ATLAB code

case when debugging, an error message at a particular line may in fact indicate an error that has occurred several lines
before.

M ATLAB reference material

For a useful quick reference for using M ATLAB check the end of this laboratory manual starting on page 19. Included
are various helpful pieces of information for working with M ATLAB.

The University of Michigan, All rights reserved 17

An Introduction to M ATLAB

18 The University of Michigan, All rights reserved

Commonly Used M ATLAB commands
Elementary Math Functions (help elmat)
abs atan exp log rem sqrt
acos ceil fix log10 round tan
angle conj floor mod sign
asin cos imag real sin

Graphing and Plotting Functions (help plot)

axis figure line print stem xlabel
bar grid loglog semilogx subplot ylabel
clf hold plot semilogx text zoom
close legend polar shg title

Relational and Logical Functions (help ops)

all eq (==) ge (>=) isempty isnan not (˜)
and (&) exist gt (>) isfinite le (<=) or (k)
any find ischar isinf lt (<) strcmp

Working With Variables (help general, help elmat)

who length clear exist nan zeros
whos size end isinf inf ones

General Purpose Functions (help general)

exit quit help helpdesk which lookfor

Programming and Control Flow (help lang)

break disp end if pause try
case else error input return warning
catch elseif for otherwise switch while

File and Directory Functions (help general, help iofun)

cd fclose load path save what
dir fopen mkdir rmdir type

Text Input/Output (help iofun, help strfun)

input keyboard sprintf disp return num2str

Debugging Commands (help debug)

dbclear dbquit dbstatus dbstop dbup
dbcont dbstack dbstep dbtype keyboard

Special Symbols (help ops, help punct)

+ - * .* / ./
ˆ .ˆ , ; : ...
’ = nan inf %

The University of Michigan, All rights reserved 19

10 Useful M ATLAB Facts

1. M ATLAB starts indexing it’s arrays from 1 rather than from 0.

2. Use the up-arrow to recall previous commands. If you type in a few characters and then hit the up-arrow,
M ATLAB will try to find a previous command that started with those characters.
3. When indexing matrices, the indices are always given a(row,column).
Similarly, size(a) returns a two-element vector [num_rows, num_columns].
4. Semicolons at the end of a line are not necessary; they simply suppress output.
5. If I multiply (or divide, or exponentiate) two arrays without using the dot-operators, I probably won’t get what
I’m expecting (unless I want to do matrix multiplication).
6. We concatenate arrays (and strings) using square brackets. To do so horizontally, we separate the arrays with
spaces or commas:

>> [ones(3), zeros(3)])

To do so vertically, we separate the arrays with a semicolon:

>> [ones(3); zeros(3)])

7. When a function returns multiple parameters, we use square brackets to retrieve them:

>> [max_value, index] = max([4.3, 2.9, 8.6, 6.3, 1.0])

Otherwise, only one parameter is returned.

8. Most M ATLAB commands (like min, max, sum, prod, and a host of others) work on matrices by operating
down each column individually. Thus, after executing this command:

>> [max_value, index] = max(eye(6))

max_value has a vector of six ones (since the maximum value in each column is 1) and index is a vector
containing the row number of the 1 in each column.
9. The end keyword is exceptionally useful when indexing into arrays of unknown size. Thus, if I want to return
all elements in a vector but the first and last one, I can use the command:

>> x(2:end-1)

which is equivalent to the command:

>> x(2:length(x)-1)

10. M ATLAB automatically resizes arrays for you. Thus, if I want to add an element on to the end of a vector, I can
use the command:

>> x(end+1) = 5;

20 The University of Michigan, All rights reserved

Laboratory 1

Signals, Signal Statistics, and Signal

Detection I

1.1 Introduction
In everyday language, a signal is anything that conveys information. We might talk about traffic signals or smoke
signals, but the underlying purpose is the same. In the study of signals and systems engineering, however, we adopt a
somewhat more specific notion of a signal. In this field, a signal is a numerical quantity that varies with respect to one
or more independent variables. One can think of a signal as a functional relationship, where the independent variable
might be time or position.
As an example, one signal might be the voltage on the wires from a microphone as it varies with time. Another
signal might be the light intensity as it varies with position along a sensor array. The important aspect of these signals,
though, is the mathematical representation, not the underlying medium. That is, the voltage and light signals might be
mathematically the same, despite the fact that the signals come from two very different physical sources. In signals
and systems engineering, we recognize that the most important aspects of signals are mathematical. Thus, we don’t
necessarily need to know anything about the physical behavior of voltage or light to deal with these signals.
What purposes do signals serve? Let us highlight a few of the many important ones. First, a signal can embody a
sensory experience, as in a sound that we would like to hear or a picture that we would like to see. Second, a signal
can convey symbolic information, as in the text of a newspaper. Third, a signal can serve to control some system. For
example, in a typical modern automobile, an electronic control signal determines how much gasoline is emitted by
the fuel injectors. Last, we mention that a signal can embody an important measurement, for example, the speed of a
vehicle or the EKG of some patient.
What is the advantage of having a sound or a picture or text or control information or a measurement embodied
in a signal? For one thing, it enables us to transmit it to a remote location or to record it. In many, but not all, cases,
these are done electronically, either with analog or digital hardware. For another, the signals we encounter frequently
need to be processed, which can also be done electronically with analog or digital hardware. For example, a signal
may contain unwanted noise that needs to be removed; this is an example of what is called noise reduction or signal
recovery. Alternatively, the desired information or sensory experience may need to be extracted from the signal, as in
the case of AM and FM radio signals, which need to be demodulated before we can listen to them, or in the case of
CT scan signals, which need extensive processing before an X-ray like image can be viewed.
Finally, in many situations, the purpose of signals is to permit decisions to be made. This kind of signal processing
is variously called signal classification, signal recognition, or signal detection. As examples, a radar system processes
the signal received from its antenna to determine whether or not it contains a reflected pulse, which would indicate

The University of Michigan, All rights reserved 21

Laboratory 1. Signals, Signal Statistics, and Signal Detection I

the presence of an airplane in the direction to which the antenna is pointed. The bill changer in a vending machine
processes the signal produced by its optical sensor to determine if the inserted piece of paper is a valid dollar bill. A
speech recognition system processes the signal produced by a microphone to determine the words that are spoken.
A speaker recognition system processes the same signal to determine the identity of the person speaking. A heart
monitoring system processes an electrical EKG signal to determine if arythmia is occuring. A digital modem processes
the received signal to determine what bits are being transmitted. These are a just a few of the situations in which signals
must be processed to make decisions.
As one of part of this lab assignment, we will implement and tune a simple signal processing system for detecting
whether or not a recorded signal contains a spoken sound. In later lab assignments, we will develop more sophisticated
decision making systems – for detecting the presence or absence of radar pulses, for decoding a sequence of key presses
from the signal produced by a touchtone telephone, and for deciding which of several vowel sounds has been spoken
into a microphone.
Throughout this course we will develop tools for analyzing, modifying, processing and extracting information from
signals mathematically. One of the most basic (and sometimes most useful) methods involves the calculation of signal
statistics. Calculating signals statistics provides us a substantial amount of useful information about a signal. These
statistics allow us to determine “how much” signal is present (i.e., the signal strength), how long a signal lasts, what
values the signal takes on, and so on. We will use signal statistics to develop measures of signal quality (with respect
to a reference signal) and also to perform signal detection (by determining when a signal contains useful information
rather than just background noise).

1.1.1 “The Questions”

• How can I quantitatively determine a signal’s “quality”?

• How can I detect the presence of “speech” within a segment of a speech signal?

1.2 Background
1.2.1 Continuous-time and discrete-time signals
In its most elementary form, a signal is a time-varying numerical quantity, for example, the time-varying voltage
produced by a microphone. Equivalently, a signal is a numerically valued function of time. That is, it is an assignment
of a numerical value to each instance of time. As such, it is customary to use ordinary mathematical function notation.
For example, if we use s to denote the signal, i.e. the function, then s(t) denotes the value of the signal at time instance
t. In common usage, the notation s(t) also has an additional interpretation — it may also refer to the entire signal.
Usually, the context will make clear which interpretation is intended.
We will deal with many different signals and to keep them separate we will use a variety of symbols to denote
them, such as such as r, x, y, z, x0 . Occasionally, we will use other symbols to denote time, such as t0 , s, u. In some
situations, the independent parameter t represents something other than “time”, such as “distance”. This happens, for
example, when pictures are considered to be signals.
As illustrated in Figure 1.1, there are two basic kinds of signals. When the time variable t ranges over all real
numbers, the signal s(t) is said to be a continuous-time signal. When the time variable t ranges only over the set
of integers {. . . , −2, −1, 0, 1, 2, . . .}, the signal s(t) is said to be a discrete-time signal. To distinguish these, from
now on we will use a somewhat different notation for discrete-time signals. Specifically, we will use one of the
letters i, j, k, l, m, n to denote the time variable, and we will enclose the time-variable in square brackets, rather than
parentheses, as in s[n]. Thus, for example, s[17] denotes the value of the discrete-time signal s[n] at time n = 17. Note
that for discrete-time signals, the time argument has no “units”. For example, s[17] simply indicates the 17th sample
of the signal. When the independent parameter t or n represents something other than time, for example distance, then

22 The University of Michigan, All rights reserved

 1.2.2 Describing Signals

(A)

(B)

Figure 1.1: (A) A continuous-time signal. (B) A discrete-time signal.

the signal can be said to be continuous-space or discrete-space, respectively, or more generally, continuous-parameter
or discrete-parameter.
It is important to reemphasize the inherent ambiguity in the notation s(t) and s[n]. Sometimes s(t) refers to the
value of the signal at the specific time t. At other times, s(t) refers to the entire signal. Usually, the intended meaning
will be clear from context. The same two potential interpretations apply to s[n].

1.2.2 Describing Signals

Some continuous-time signals can be described with formulas, such as s(t) = sin(t) or

0 t<0
s(t) = (1.1)
cos(t) t ≥ 0.
For other signals, there are no such formulas. Rather they might simply be measured and recorded, as with an analog
tape recorder. Similarly, some discrete-time signals can be described with formulas, such as s[n] = sin(n) or

0 n<0
s[n] = , (1.2)
cos[n] n ≥ 0
and some are described simply by recording their values for all values of n.
Often, a discrete-time signal is obtained by sampling a continuous-time signal. That is, if Ts is a small time
increment, then the discrete-time signal s[n] obtained by sampling s(t) with sampling interval or sampling period Ts
is defined by
s[n] = s(nTs ) , −∞ < n < ∞ (1.3)
For example, if s(t) = sin(t) and Ts = 3, then the discrete-time signal obtained by sampling with sampling interval
Ts is s[n] = sin(3n). The reciprocal of Ts is called the sampling rate or sampling frequency and denoted fs = 1/Ts .
Its units are samples per second. The discrete-time signal in Figure 1.1 was obtained by sampling the continuous-time
signal shown above it.
In the above example, we have allowed the time parameter to be negative as well as positive, which begs the
question of how to interpret negative time. Time 0 is generally taken to be some convenient reference time, and
negative times simply refer to times prior to this reference time.

The University of Michigan, All rights reserved 23

Laboratory 1. Signals, Signal Statistics, and Signal Detection I

Nowadays, signals are increasingly processed by digital machines such as computers and DSP chips. These are
inherently discrete-time machines. They can record and work with a signal in just two ways: as a formula or as a
sequence of samples. The former applies to continuous-time and discrete-time signals.√ For example, a computer can
easily compute the value of the continuous-time signal s(t) = sin(t) at time t = 2 or the value of the discrete-time
signal s[n] = cos(n) at time n = 17. However, the latter works only with discrete-time signals. Thus, whenever a
digital machine works with a continuous-time signal, it must either use a formula or it must work with its samples. That
is, it must work with the discrete-time signal that results from sampling the continuous-time signal. This admonition
applies to us, because in this and future lab assignments, many of the signals in which we are interested are continuous-
time, yet we will process them exclusively with digital machines, i.e. ordinary computers.
Except in certain ideal cases, which never apply perfectly in real-world situations, sampling a continuous-time sig-
nal entails a “loss”. That is, the samples only partially “capture” the signal. Alternatively, they constitute an approxi-
mate representation of the original continuous-time signal. However, as the sampling interval decreases (equivalently,
the sampling rate increases), the loss inherent in the sampled signal decreases. Thus in practical situations, when the
sampling interval is chosen suitably small, one can reliably work with a continuous-time signal by working with its
samples, i.e. with the discrete-time signal obtained by sampling at a sufficiently high rate. This will be the approach
we will take in this and future lab assignments, when working with continuous-time signals that cannot be described
with formulas.
When digital machines are used to process signals, in addition to sampling, one must also quantize, or round,
the sampled signal values to the limited precision with which numbers are represented in the machine, e.g. to 32-bit
floating point. This engenders another “loss” in the signal representation. Fortunately, for the computers we will use
in performing our lab experiments, this loss is so small as to be negligible. For example, M ATLAB uses 64-bit double-
precision floating point representation of numbers. (Lab 5 is an exception; in that lab, we will consider systems that
are designed to produce digital signal representations with as few bits as possible.)

1.2.3 Signal support and duration

The support of a signal is the smallest time interval that includes all non-zero values of the signal. For example, the
continuous-time signal s(t) = cos(t), 0 ≤ t ≤ 3, s(t) = 0, else has support interval [0, 3]. The discrete-time signal
s[n] = cos(n), 0 ≤ n ≤ 3, has support interval [0, 3] = {0, 1, 2, 3}. The duration of a signal is simply the length of its
support interval. In the previous examples, the duration of s(t) is 3, and the duration of s[n] is 4. Note that the support
and duration of a signal can be either finite or infinite.

1.2.4 Periodicity
Periodicity is a property of many naturally occurring or man-made signals. A continuous-time signal s(t) is said to be
periodic with period T , where T is some positive real number, if

s(t + T ) = s(t) , for all t (1.4)

If s(t) is periodic with period T , then it is also periodic with period 2T , 3T , . . . . The fundamental period To of s(t)
is the smallest T such that s(t) is periodic with period T .
Similarly, a discrete-time signal s[n] is said to be periodic with period N , where N is some positive integer, if

s[n + N ] = s[n] , for all n (1.5)

If s[n] is periodic with period N , then it is also periodic with period 2N , 3N , . . . . The fundamental period No is the
smallest N such that s[n] is periodic with period N .

24 The University of Michigan, All rights reserved

 1.2.5 Signals in M ATLAB

1.2.5 Signals in M ATLAB

While signals can be represented by formulas or by recorded signal values, when working in M ATLAB, we generally
use the latter. That is, we represent a signal as a vector (i.e., a one-dimensional array) of numbers.

Discrete-time signals
We begin with an example. Suppose we want to represent the following discrete-time signal as an array in M ATLAB:
 2
n 5 ≤ n ≤ 15
s[n] = (1.6)
0 else

In M ATLAB, we do this by creating two vector: a support vector and a value or signal vector. The support vector
represents the support interval of the signal, i.e. the set of integers from the first time at which the signal is nonzero to
the last. For this example, the support vector can be created with the command
>> n = 5:15
This causes n to be the array of 11 numbers 5, 6, . . . , 15. Next, the signal vector can be created with the command
>> s = n.ˆ2
which causes s to be the array of 11 numbers 25, 36, . . . , 225.
Note that as in the above example, we usually only specify the signal within the range of times that it is nonzero.
That is, we usually do not include zero values outside the support interval in the signal vector.
It is often quite instructive to plot signals. To plot the discrete-time signal s[n], use the stem command:
>> stem(n,s)
You can also use the plot command; however, plot draws straight lines between plotted points, which may not be
desirable.
It is important to note that in M ATLAB, when i is an integer then s(i) is not necessarily the signal value at time i.
Rather it is the signal at time n(i). Thus, stem(n,s) and stem(s) result in similar plots with different labelings
of the time axis. Occasionally, it will happen that n(i) = i, in which case s(i) = s(n(i)) and stem(n,s)
and stem(s) result in identical plots with identical time axis labels.

Continuous-time signals
We begin with an example. Suppose we wish to represent the following continuous-time signal as an array in M ATLAB:
 2
t 5 ≤ t ≤ 15
s(t) = (1.7)
0 else

We first choose a sampling interval Ts with a command such as

>> Ts = 1/20
We then create a support vector with the command
>> t = 5:Ts:15
Finally, we create a signal vector with the command
>> s = t.ˆ2

The University of Michigan, All rights reserved 25

Laboratory 1. Signals, Signal Statistics, and Signal Detection I

What have we done? To represent s(t), we have created a support vector t that contains the sample times 5, 5 +
1/20, 5 + 2/20, 5 + 3/20, . . . , 15 , and we have created a signal vector s that contains the samples s(5), s(5 +
1/20), s(5 + 2/20), s(5 + 3/20), . . . , s(15). That is, for n = 1, . . . , 301, s(n) contains the signal value at time
t(n) = 5 + (n-1)/20.
Note that when representing a continuous-time signal as an array, it is usually important to choose the sampling
interval Ts small enough that the signal changes little over any time interval of Ts seconds.
As with discrete-time signals, it is frequently instructive to plot a continuous-time signal. This is done with the
command

>> plot(t,s)

which plots the points (t(1),s(1)), (t(2),s(2)), . . ., and connects them with straight lines. Connecting these
points in this manner produces a plot that approximates the original continuous-time signal s(t), which takes values at
all times (not just integers) and which usually does not change significantly between samples (assuming Ts is chosen
to be small enough). Note that plot(s) produces a similar plot, but the horizontal axis is labeled with sample
“indices” (i.e., the number of the corresponding sample) rather than sample times. When working with continuous-
time signals, it is important that you always use plot(t,s) rather than plot(s). It also important that your plot
indicates what the axes represent, which can be done using the xlabel and ylabel commands.

1.2.6 Signal Statistics

When dealing with a signal, it is often useful to obtain a rough sense of the range of values it takes and of the average
size of its values. We do this by computing one or more signal statistics.
The following lists a number of common signal statistics. It gives the defining formula for each for both continuous-
time and discrete-time signals. Also included is M ATLAB code for calculating the statistic1 for a discrete-time signal.
If we wish to compute a statistic for a continuous-time signal when we only have a sampled representation, we can
use the discrete-time statistic to approximate the continuous statistic. The formulas needed for this approximation are
included here with the label “sampled.” (In most cases, this approximation becomes better as the sampling interval Ts
decreases.) For completeness, signal support and duration are also defined below.

1. Support Interval. A signal’s support interval (also occasionally known as just the signal’s support or its
interval) is the smallest interval that includes all non-zero values of the signal.

Continuous-time: t1 ≤ t ≤ t2 (1.8)
Discrete-time: n 1 ≤ n ≤ n2 (1.9)
M ATLAB: n = n1:n2 for a signal s. (1.10)

2. Duration. The duration of a signal is simply the length of the support interval.

Continuous-time: t2 − t 1 (1.11)
Discrete-time: n 2 − n1 + 1 (1.12)
M ATLAB: Assumed length(s) for a signal s. (1.13)
Sampled: (t2 − t1 ) = (n2 − n1 + 1)Ts (1.14)
1 This code assumes that the signal vector s is defined only over the range of times for which we wish to compute the statistic. More generally,

if n is the support vector and n1 and n2 define a subset of the support vector over which we wish to calculate our statistic, we can compute the
statistic over only this range, n1:n2, by replacing the signal s with the shorter signal s((n1:n2)-n(1)+1).

26 The University of Michigan, All rights reserved

 1.2.6 Signal Statistics

3. Periodicity. Periodicity was described in section 1.2.4. The key formulas are included here.

Continuous-time: s(t) = s(t + T ) (1.15)

Discrete-time: s[n] = s[n + N ] (1.16)
Sampled: T ≈ N Ts (1.17)

4. Maximum and Minimum Value. These values are the largest and smallest values that a signal takes on over
some interval defined by n1 and n2 . In M ATLAB these values are found using the min and max commands.

M ATLAB: M aximum(s) = max(s) (1.18)

M ATLAB: M inimum(s) = min(s) (1.19)

5. Average Value. The average value, M , is the value around which the signal is “centered” over some interval.
Z
1 t2
Continuous-time: M (s(t)) = s(t)dt (1.20)
t2 − t1 t1

1 X
n2
Discrete-time: M (s[n]) = s[n] (1.21)
n2 − n1 + 1 n=n
1

M ATLAB: M (s) = mean(s) (1.22)

Sampled: M (s(t)) ≈ M (s[n]) (1.23)

6. Mean-squared value. The mean-squared value (or MSV) of a signal, M S, is defined as the average squared
valued of the signal over an interval. The MSV is also called the average power, because the squared value of
a signal is considered to be the instantaneous power of the signal.
Z
1 t2
Continuous-time: M S(s(t)) = s2 (t)dt (1.24)
t2 − t1 t1

1 X
n2
Discrete-time: M S(s[n]) = s2 [n] (1.25)
n2 − n1 + 1 n=n
1

M ATLAB: M S(s) = mean(s.ˆ2) (1.26)

Sampled: M S(s(t)) ≈ M S(s[n]) (1.27)

7. Root mean squared value. The root mean squared value (or RMS value) of a signal over some interval is
simply the square root of mean squared value.
s Z
1 t2
Continuous-time: RM S(s(t)) = s2 (t)dt (1.28)
t2 − t1 t1
v
u
u 1 X
n2
Discrete-time: RM S(s[n]) = t s2 [n] (1.29)
n2 − n1 + 1 n=n
1

M ATLAB: RM S(s) = sqrt(mean(s.ˆ2)) (1.30)

Sampled: RM S(s(t)) ≈ RM S(s[n]) (1.31)

The University of Michigan, All rights reserved 27

Laboratory 1. Signals, Signal Statistics, and Signal Detection I

−4 −2 0 2 4 −4 −2 0 2 4
Signal Value Distribution Histogram Approximation

Figure 1.2: Signal value distribution and a discrete histogram approximation

8. Signal Energy. The energy of a signal, E, indicates the strength of a signal is present over some interval. Note
that energy equals the average power times the length of the interval.
Z t2
Continuous-time: E(s(t)) = s2 (t)dt (1.32)
t1
X
n2
Discrete-time: E(s[n]) = s2 [n] (1.33)
n=n1
M ATLAB: E(s) = sum(s.ˆ2) (1.34)
Sampled: E(s(t)) ≈ E(s[n])Ts (1.35)

9. Signal Value Distribution. The signal value distribution is a plot indicating the relative frequency of occur-
rence of values in a signal. There is no closed-form definition of the signal value distribution, but it can be
approximated using a histogram. A histogram counts the number of samples that fall within particular ranges,
or bins. Note that the y-axis is effectively arbitrary, and that the coarseness of the approximation is determined
by the number of histogram bins that are used. Figure 1.2 shows an example of a signal value distribution and
the histogram approximation to that distribution.

M ATLAB: hist(s,num_bins); (1.36)

1.2.7 Measuring signal distortion and error

Suppose that we wish to transmit a signal from one location to another. This is a common task in communication
systems. A common problem is that the signal is often modified or distorted in the communication process. Thus,
the received signal is not the same as the transmitted signal. Typically, we want to reduce the amount of distortion as
much as possible. However, this requires that we have a method of measuring the amount of distortion in a signal. In
order to develop such a measure, we’ll look at a signal plus noise model of signal distortion.
Suppose we are transmitting a signal s[n] over FM radio. Someone tunes in to our radio station and receives a
modified version of our signal, r[n]. We can represent this modification mathematically as the addition of an error

28 The University of Michigan, All rights reserved

 1.2.8 Signal detection

signal, v[n], like this:

r[n] = s[n] + v[n]. (1.37)
Assuming that we have both s[n] and r[n], we can easily calculate v[n] as

v[n] = r[n] − s[n]. (1.38)

Note that if s[n] and r[n] are identical, v[n] will be zero for all n. This suggests that we can simply measure the signal
strength of v[n] by using one of the energy or power statistics.
Mean squared value is a natural choice because it normalizes the error with respect to the length of the signal.
Sometimes, though, the RMS value is more desirable because it produces error values that are directly comparable to
the values in v[n]2 . When we measure the MSV of an error signal, we sometimes call it the mean squared error or
MSE. Similarly, the RMS value of an error signal is often called the root mean squared error or RMSE.
In M ATLAB, we will usually want to calculate the MSE or RMSE over the entire length of the signals that we have.
Supposing that we are given a signal s and a modified version s_mod (with the same size), we can calculate the MSE
and RMSE like this:

>> mse = mean((s - s_mod).ˆ2);

>> rmse = sqrt(mean((s - s_mod).ˆ2));

Notice that we could also subtract s from s_mod; the order doesn’t matter because of the square operation. Also note
that you must include the period before the exponentiation operator in order to correctly square each sample.

1.2.8 Signal detection

Suppose that we are designing a continuous speech recognition and transcription system for a personal computer. The
computer has a microphone attached to it, and it “listens” to the user’s speech and tries to produce the text that was
spoken. However, the user is not speaking continuously; there are periods of silence between utterances. We don’t
want to try to recognize speech where there is silence, so we need some means of determining when there is a speech
signal present.
This is an example of signal detection. There are many different types of signal detection. Sometimes signal
detection involves finding a signal that is obscured by noise, such as radar detection. In other applications, we need to
determine if a particular signal exists in a signal that is the sum of many signals. The “signal present” detector for our
speech recognition system is a simpler form of signal detection, but it still important in many applications. As another
example, some digital transmission systems send bits using what is known as on/off keying. They send an electrical
pulse to represent “1” and send nothing at all to represent “0”. The receiver for such a system uses a “pulse present”
detector.
In this laboratory, we will consider the design of a “signal present” detector to identify spoken segments of a
speech signal. Note, however, that the detector we will design can be used for many other applications as well. Figure
1.3 shows a block diagram of such a detector. The detector consists of two blocks. The first block computes one or
more statistics that we will use to perform the detection. The second block uses the computed statistic(s) to make the
final decision.

Specifying the detector’s operation

The first step is to specify what our system needs to do. From the description above, we know that we will receive a
signal as an input. For simplicity, we’ll assume that we are given an entire discrete-time signal. What must our system
do? We need some sort of indication as to when speech is present in a signal. However, the signal that we are given
2 Mean square values are comparable to the square of the values in v[n]

The University of Michigan, All rights reserved 29

Laboratory 1. Signals, Signal Statistics, and Signal Detection I

Statistic Decision
x[n] d
Calculator Maker
Input Binary
Decisio
Signal

Figure 1.3: An “overview” block diagram for a “signal present” detector.

x1 [n]
RMS( ) ?(> c)
Break x2 [n] RMS( ) ?(> c)
signal Recombine
x[n] into x3 [n] RMS( ) decisions d[m]
...

Input blocks ... ?(> c)

Binary
signal decisio
RMS( ) ?(> c) signal
xk [n]

Calculate
Compare to
RMSV threshold

Figure 1.4: A detailed block diagram for the “signal present” detector.

will contain both periods of silence and periods with speech. For some integer N , let us break the signal into blocks
of N samples, such that the first N samples make up the first block, the next N samples make up the second block,
and so on. Then, we will make a “speech present” or “speech not present” decision separately for each block. The
output of our system will consist of a signal with a 0 or 1 for each block, where a 1 denotes “speech present” and a
0 denotes “speech not present”. To describe this signal in M ATLAB, our system will produce a signal support vector
containing containing the index of the first sample of each block and a signal vector containing the 0 or 1 for each
block. Choosing the support vector in this way will allow us to plot the decisions on the same figure as the signal
itself.
How will we make the decision for each block? Since we can assume a signal that is relatively free of noise, we can
simply calculate the energy for each block and compare the result to a threshold. If the statistic exceeds the threshold,
we decide that speech is present. Otherwise, we decide that speech is not present. Using signal energy, though, is not
ideal; the necessary threshold will depend on the block size. We may want to change the block size, and we should be
able to keep the threshold constant. Using average power is a better option, but we would like our threshold to have
a value comparable to the values of the signal. Thus, the RMS value seems to be an ideal choice, and this is what we
will use. In summary, for each block the detector computes the RMS value R and compares it to a threshold c. The
decision is

signal present, if R ≥ c
signal not present, if R < c

Note that our detector system will has two design parameters. One is the block size N , and the other is the threshold
c. To tune the system, we will need to find reasonable values for these parameters when we make the detector itself.
A more detailed block diagram of the detector can be found in Figure 1.4.

30 The University of Michigan, All rights reserved

 1.2.8 Signal detection

Detector algorithm
Now that we’ve specified the behavior of the detector, let’s come up with an algorithm for performing the detection.
Of course, this is not the only way to implement this detector.
It is assumed that the input signal is contained in a signal vector x whose support vector is simply [1, 2, ...,
length(x)].
• Define a variable called block_size, representing N .
• Define a variable called threshold, representing c.
• Calculate the number_of_blocks.
• Determine the support vector support_output for the output signal.
• For each block of the signal:
– Calculate the RMS value of the current block
– Compare the RMS value to the threshold
– Store the result in the output array
• Return the output array
What follows are some details about the algorithm:
1. First, note that we want number_of_blocks to be an integer. For this calculation, recall that the function
length returns the number of samples in a vector. Also, note that the floor command rounds down to the
nearest integer
2. Suppose that the block size is 512. Then the vector support_output should contain the numbers [1,
513, 1025, ...] and so on. You can generate support_output with a single line of code by using the
: operator3.
3. There are actually two separate ways to implement the “for each block” part of this algorithm in M ATLAB. One
involves using a for loop, while the other makes use of the reshape command and M ATLAB’s vector and
matrix arithmetic capabilities. Both take roughly the same amount of code, but the second way is somewhat
faster. You can implement whichever version of the algorithm that you choose in the lab assignment.
(a) If you implement the algorithm using a for loop, you should first initialize the output array to “empty”
using the command “output =[];”. Then, loop over the values in support_output. Within the
loop, you need to determine what values of n1 and n2 to use in the RMS value calculation for a given
value of the loop counter. Then, compare the RMS value that you calculate to the threshold and append
the result to the end of output4 .
(b) An alternative to the for loop is to use the reshape command to make a matrix out of our signal with
one block of the signal per column. If you choose to use reshape, you first need to discard all samples
beyond the first block_size × number_of_blocks samples of the input signal. reshape this
shorter signal into a matrix with block_size rows and number_of_blocks columns5. Then, use
.ˆ to square each element in the matrix, use mean to take the mean value of each column, and take
the sqrt of the resulting vector to produce a vector of RMS values. Finally, compare this vector to
threshold to yield your output vector.
3 Type help colon if you need assistance with this operator.
4 Use either output(end+1) = result; or output = [output, result];
5 Remember to assign the output of reshape to something! No M ATLAB function ever modifies its input parameters unless you explicitly

reassign the output to the input.

The University of Michigan, All rights reserved 31

Laboratory 1. Signals, Signal Statistics, and Signal Detection I

1.3 Some M ATLAB commands for this lab

• Zooming on Figures: In M ATLAB, you can interactively zoom in and out on figure windows. To do so, you can
either find a “+ magnifying glass” icon on the figure window, or you can type zoom on at the command lines.
Then, you can click and drag a zoom box on the figure window to get a closer look at that portion of the figure.
Also very useful is the zoom xon command, which enables zooming only in the x-direction; this is usually
how we will want to zoom in on our signals.
• Using line styles and legend: Whenever you plot two or more signals on the same set of axes, you must
make sure that the signals are distinguishable and labeled. Generally, we do this using line styles and the
legend command. The plot command gives you a wide range of options for changing line styles and
colors. For instance, the commands

>> hold on
>> plot(1:10,1:10,'-')
>> plot(1:10,2:11,':')
>> plot(1:10,3:12,'--')

plot lines using solid, dotted, and dashed lines. Type help plot for more details about using different line
styles and colors. The legend command adds a figure legend for labeling the different signals. For instance,
the command

>> legend('Solid (lower)', 'Dotted (middle)', 'Dashed (higher)')

adds a legend with labels for each of the three signals on the figure. Note that signal labels are given in the order
that the signals were added to the figure.
• Labeling Figures: Any time that you create a figure for this laboratory, you need to include axis labels, a figure
number, and a caption:

>> xlabel('This is the x-axis label');

>> ylabel('This is the y-axis label');
>> title('Figure 1: This is a caption describing the figure');

Note that it is recommended that you use your word processor to produce figure numbers and captions, rather
than using the title command. You also need to include the code that you used to produce the figures,
including label commands. Note that each subplot of a figure must include its own axis labels.
• Function Headers: At the top of the file containing a function declaration, you must have a line like this:

function [out1, out2, out3] = function_name(in1, in2, in3)

where in1, in2, and in3 are input parameters and out1, out2, and out3. Note that you can name the
parameters anything you like, and there can be any number of them. The word function is a M ATLAB
keyword. Also, you do not need to explicitly return the output parameters. Instead, M ATLAB will take their
values at the end of the function’s execution and return them to the calling function.
• sum, mean, min, and max: Given a vector (i.e., a one-dimensional array), these M ATLAB functions calculate
the sum, mean, minimum, and maximum (respectively) of the numbers in the array and returns a single number.
If these functions are given a matrix (i.e., a two-dimensional array), they calculate the appropriate statistic on
each column of the matrix and return a row-vector containing one result for each column.

32 The University of Michigan, All rights reserved

 1.4 Demonstrations in the Lab Session

• for loops: Most for loops in M ATLAB have the following form

>> for index = 1:20

>> % This loop is executed twenty times
>> end

In this case, the loop counter, index, is set to 1 on the first execution loop, 2 on the second, and so on. The
loop will execute a total of 20 times. We can, of course, use any variable for the loop counter.
A more general form of the for loop is given by

>> for index = row_vector

>> % Code in the loop goes here
>> end

where index is the loop counter and row_vector is a row vector that contains all the values that will be
assigned to index on successive iterations of the loop. Thus, the loop will execute size(row_vector,2)
times6 .
• Reshaping arrays: The reshape command is used to change the shape of an array:

>> new_array = reshape(array,[new_rows, new_columns]);

array must have new_rows*new_columns elements7 . new_array will have dimensions of new_rows
× new_columns.
• Logical operators: The logical operators (>, <, >=, <=, ==, ˜=) perform a test for equality or various forms
of inequality. They all operate in the same manner by evaluating to 1 (“true”) or 0 (“false”) depending upon the
truth value of the operator. After executing the following statement, for instance,

>> result = (x > 0);

result will contain a one or a zero if x is a single number. If x is an array, result will be an array of ones
and zeros with a size equal to x, where each element indicates whether the corresponding element in x is in fact
greater than 0.

1.4 Demonstrations in the Lab Session

• Laboratory policies
• Signals in M ATLAB – sampling
• Signal statistics
• Approximating continuous-time signal statistics with sampled signals
• Model of discrete-time signal as signal plus noise
• The “signal present” detector
6 If row vector is actually a matrix (or a column vector), each column of row vector will be assigned to index in turn.
7 The number of elements in array can be checked using the command prod(size(array)).

The University of Michigan, All rights reserved 33

Laboratory 1. Signals, Signal Statistics, and Signal Detection I

1.5 Laboratory Assignment

Note that in this and all following laboratory assignments, the bullets (i.e, •) indicate items that you must include in
your laboratory.

1. (A simple signal and its statistics) Use the following M ATLAB commands to create a signal:

>> n = 1:50;
>> s = sin(2*pi*n/50);

(a) (Plotting a signal with labels) Use stem to plot the signal. Make sure that you include8:
• [2] The figure itself9 .
• [1] An x-axis label and a y-axis label.
• [1] A figure number and a caption that describes the figure.
• [1] The code you used to produce the signal and the figure. This should be included in an appendix at
the end of your report10 . Make sure you clearly indicate which problem the code belongs to.
(b) (Calculating signal statistics) Calculate the following statistics over the length of the signal (i.e., let n1 = 1
and n2 = length(s)), and include your results in your report11 .
• [2] Maximum value
• [2] Minimum value
• [2] Mean value
• [2] Mean squared value
• [2] RMS value
• [2] Energy
(c) (Approximating continuous-time statistics in discrete-time) Suppose that s is the result of sampling a
continuous-time signal with a sampling interval Ts = 1/100. Use the discrete-time statistics to estimate
the following statistics for the continuous-time signal s(t) = sin(4πt):
• [2] Signal duration
• [2] Energy
• [2] Average power
• [2] RMS Value

2. (Statistics of real-world signals) Download the file lab1_data.mat from the course web page. Place it in the
present working directory or in a directory on the path, and type

>> load lab1_data

8 Notethat every figure that you produce in a laboratory for this class must include these things!
9 OnWindows systems, you can select “Copy Figure” from Edit menu on the figure window to copy the figure to the clipboard and then paste
it into your report. Also, to make your report compact, you should make all figures as small as possible, while being just large enough that the
important features are clearly discernable.There are two ways to shrink plots, you can shrink them in your lab report document, or you can shrink
the M ATLAB window before copying and pasting. Shrinking the M ATLAB window is generally preferable because it does not shrink the axis labels.
Note, you may need to specify the appropriate copy option, so that what is in fact copied is the shrunk rather than original version of the plot
10 You should include all M ATLAB code that you use in the appendix. However, you do not need to include code that is built into M ATLAB or

code that we provide to you.

11 Remember to include the code you used to calculate these in your M ATLAB appendix.

34 The University of Michigan, All rights reserved

 1.5 Laboratory Assignment

This file contains two signals which will be loaded into your workspace. You will use the signal clarinet12
in this problem and in Problem 3. The other signal, mboc, will be used in Problem 4.

(a) (Plot the real-world signal) Define the support vector for clarinet as 1:length(clarinet). Then,
use plot to plot the signal clarinet.
• [4] Include the figure (with axis labels, figure number, caption, and M ATLAB code) in your report.
(b) (Zoom in on the signal) Zoom in on the signal so that you can see four or five “periods13 .”
• [4] Include the zoomed-in figure (with axis labels, figure number, caption, and M ATLAB code) in your
report.
(c) (Find the signal’s period) Estimate the “fundamental period” of clarinet. Include:
• [3] Your estimate for the discrete-time signal (in samples).
• [3] Your estimate for the original continuous-time signal (in seconds).
(d) (Approximate the SVD) Use the hist command to estimate the signal value distribution of clarinet.
Use 50 bins.
• [4] Include the figure (with axis labels, code, etc.) in your report.
• [1] From the histogram, make an educated guess of the MSV and RMSV. Explain how you arrived at
these guesses.
(e) (Calculate statistics) Calculate the following (discrete-time) statistics over the length of the signal:
• [2] Mean value
• [2] Energy
• [2] Mean squared value
• [2] RMS value

3. (Looking at and measuring signal distortion) In this problem, we’ll measure the amount of distortion introduced
to a signal by two “systems.” Download the two files lab1_sys1.m and lab1_sys2.m. Apply each system
to the variable clarinet using the following commands:

>> sys1_out = lab1_sys1(clarinet);

>> sys2_out = lab1_sys2(clarinet);

(a) (Examine the effects of the systems) Use plot14 and M ATLAB’s zoom capabilities to display roughly one
“period” of:
• [3] The input and output of lab1_sys1 on the same figure.
• [3] The input and output of lab1_sys2 on the same figure.
(b) (Describe the effects of the systems) What happens to the signal when it is passed through these two
systems? Look at your plots from the previous section and describe the effect of:
• [3] lab1_sys1.m on clarinet.
• [3] lab1_sys2.m on clarinet.
(c) (Measure the distortion) Calculate the RMS error introduced by each system.
12 This is a one-second recording of a clarinet, recorded at a sampling frequency of 22,050 Hz. To listen to the sound, use the command
soundsc(clarinet,22050).
13 This signal, like all real-world signals, is not exactly periodic; however, it is approximately periodic.
14 Make sure the two signals are easily distinguishable by using different line styles. Also, any time that you plot multiple signals on a single set

of axes, you must use legend or some other means to label the signals!

The University of Michigan, All rights reserved 35

Laboratory 1. Signals, Signal Statistics, and Signal Detection I

• [3] RMS error introduced by lab1_sys1.

• [3] RMS error introduced by lab1_sys2.
• [1] Which system introduces the least error? Is this what you would have expected from your plots?

4. (Developing an energy detector) In this problem, we will develop a detector that identifies segments of a speech
signal in which speech is actually present. Download the files sig_nosig.m and lab1_data.mat (if you
haven’t already) from the course web page. The first is a “skeleton” m-file for the signal/no signal detector
function. The second contains a speech signal, mboc15 , that we will use to test the detector.

(a) (Write the detector function) Following the detector description given in Section 1.2.8, complete the func-
tion in sig_nosig.m. Use a threshold of 0.2 and a block size of 512. Verify the operation of your
completed function on the mboc signal by comparing its output to that of sig_nosig_demo.dll16.
• [15] Include the code for your completed version of sig_nosig.m in the appendix of your lab
report.
(b) (Plot the results of your function) Call sig_nosig17 like this:
>> [detection,n] = sig_nosig(mboc);
Then, plot the output of sig_nosig (using “stairs(n,detection,'k:');”) and the signal mboc
(with plot) on the same figure.
• [4] Include this plot in your report.
(c) (Adjusting the threshold) The threshold given above isn’t very good; it causes the detector to miss signif-
icant portions of the signal. Change the threshold until all significantly visible portions of the signal are
properly marked as “speech” by the detector, but the regions between these portions are marked as “no
speech.” (Note that you cannot use the compiled demo function for this part of the assignment.)
• [4] What threshold did you find?
• [6] Include the figure (like the one you generated in Problem 4b) that displays the output of your
detector with this new threshold.

5. On the front page of your report, please provide an estimate of the average amount of time spent outside of lab
by each member of the group.

15 This signal has also been recorded with a sampling frequency of 22,050 Hz.
16 sig nosig demo.dll is a completed version of the function in sig nosig.m. This demo function has been compiled for use on Matlab
version 6 on Windows-based systems ONLY.
17 If you did not successfully complete this function, you may use the compiled demo function for this part of the assignment.

36 The University of Michigan, All rights reserved

Laboratory 2

Signal Correlation and Detection II

2.1 Introduction
In Lab 1, we designed an energy-based signal/no-signal detector for determining when a desired signal is present. This
type of detector has a wide variety of applications, from speech analysis to communication, but it has two weaknesses.
First, an energy-based detector is very susceptible to noise, especially when the signal’s energy is small compared to
the energy of the noise. Second, such a detector cannot distinguish between different types of signals that are mixed
together.
In this laboratory, we will examine an alternative detection method that addresses these concerns. It uses a com-
putation called correlation to detect the presence of a signal with a known form. In general, correlation measures the
similarity between two signals. Using correlation for detection has significant applications. For instance, it allows
several signals to be sent over a single communications channel simultaneously. It also allows the use of radar and
sonar in noisy environments. Later in this course, we will see that correlation forms the basis for one of the most
important tools in signals and systems engineering, the spectrum.

2.1.1 “The Questions”

• How can we transmit and receive bits from several different users on the same communication channel?

• How can we develop a radar detection scheme that is robust to noise, and how do we characterize its perfor-
mance?

2.2 Background
2.2.1 Correlation
Suppose that we have two discrete-time signals, x[n] and y[n]. We compute the correlation1 between these two signals,
C(x, y), using the formula
Xn2
C(x, y) = x[n]y[n] (2.1)
n=n1
1 We will occasionally refer to this operation as “in-place” correlation to distinguish it from “running” correlation. Sometimes this is also called
an “inner product” or “dot product.”

The University of Michigan, All rights reserved 37

Laboratory 2. Signal Correlation and Detection II

Positively Correlated Uncorrelated Negatively Correlated

4 4 4
2 2 2

x[n]
0 0 0
−2 −2 −2
−4 −4 −4
0 5 10 0 5 10 0 5 10
4 4 4
2 2 2
y[n]

0 0 0
−2 −2 −2
−4 −4 −4
0 5 10 0 5 10 0 5 10
10 4 10
2
x[n]*y[n]

0 0 0
−2
−10 −4 −10
0 5 10 0 5 10 0 5 10
C(x,y) = 38 C(x,y) = 0 C(x,y) = −38

Figure 2.1: Examples of positively correlated, uncorrelated, and anticorrelated signals.

where n1 and n2 define the interval over which we are calculating the correlation. In words, we compute a correlation
by multiplying two signals together and then summing the product. The result is a single number that indicates the
similarity between the signals x[n] and y[n].
What values can C(x, y) take on, and what does this tell us about the signals x[n] and y[n]? Let us consider the
examples in Figure 2.1. For the signals shown in the first column, C(x, y) > 0, in which case, the signals are said to
be positively correlated. Basically, this means that the signals are more similar than they are dissimilar. In the second
column, we can see an example where C(x, y) is zero. In this case, the two signals are uncorrelated. One might
say that uncorrelated signals are “equally” similar and dissimilar. Notice, for instance, that the signal x[n] × y[n] is
positive as often as it is negative. Knowledge of the value of signal x[n] at time n indicates little about the value of
y[n] at time n. Finally, in the third column we see an example where C(x, y) < 0, which means that x[n] and y[n] are
negatively correlated. This means the signals are mostly dissimilar.
Note that the positively correlated signals given in Figure 2.1 are actually identical. This is a special case; from
equation (2.1), we can see that in this case the correlation is simply the energy of x[n], i.e.

C(x, x) = E(x) . (2.2)

Sometimes, it is more useful to work with normalized correlation, as defined by

C(x, y) 1 X
n2
CN (x, y) = p =p x[n]y[n]. (2.3)
E(x)E(y) E(x)E(y) n=n1

Normalized correlation is somewhat easier to interpret. The well known Cauchy-Schwartz inequality shows that the
normalized correlation varies between -1 and +1. That is, for any two signals

−1 ≤ CN (x, y) ≤ 1. (2.4)

Thus, signals that are as positively correlated as possible have normalized correlation 1 and signals that are as nega-
tively correlated as possible have normalized correlation -1. Moreover, it is known that two signals have normalized

38 The University of Michigan, All rights reserved

 2.2.2 Running correlation

correlation equal to 1 when and only when one of the signals is simply the other multiplied by a positive number. In
this case, the signals are said to be perfectly correlated. Similarly, two signals have normalized correlation equal to
-1 when and only when one is simply the other multiplied by a negative number, in which case they are said to be
perfectly anticorrelated.

2.2.2 Running correlation

In many situations, it is quite useful to correlate a signal y[n] with a sequence of delayed versions of another signal
x[n]. That is, we wish to correlate y[n] with x[n], with x[n − 1], with x[n − 2], etc. In such cases, we perform running
correlation of y[n] with x[n], which produces the correlation signal

r[k] = C(x[n − k], y[n]) , k = 0, 1, 2, . . . (2.5)

X
= x[n − k]y[n] , k = 0, 1, 2, . . . (2.6)
n

Note that since n was used as the time variable for x and y, we have introduced a new time variable, k, for r.
Suppose, for example, that we want to know the distance to a certain object, like an airplane. We transmit a radar
pulse, x[n], and receive a signal, y[n], that contains the reflection of our pulse off of the object. For simplicity, let’s
assume that we know y[n] is simply a delayed version of x[n], that is2 ,

y[n] = x[n − n0 ], (2.7)

However, we do not know the delay factor, n0 . Since n0 is proportional to the distance to our object, this is the quantity
that we wish to estimate. We can use correlation to help us determine this delay, but we need to use running correlation
rather than simply in-place correlation.
Suppose that we first guess that n0 is equal to zero. We correlate y[n] with x[n] and record the resulting correlation
value as one sample of a new signal, r[0]. Then, we guess that n0 is equal to one, shift x[n] over by one sample,
and correlate y[n] with x[n − 1]. We record this correlation value as r[1]. We can continue this shift-and-correlate
procedure, building up the new signal r[k] according to the formula
∞
X
r[k] = C(x[n − k], y[n]) = x[n − k]y[n]. (2.8)
n=−∞

Once we find a value of r[k] that equals E(x), we have found the value of n0 . This procedure of building up the signal
r[k] is known as running correlation or sliding correlation. We will refer to the resulting signal (r[k] above) as the
correlation signal.
As an example, Figure 2.2 shows a radar pulse, a received signal containing two delayed versions of the radar
pulse (one without noise and one with noise), and the running correlation produced by correlating the pulse with the
received signal.
Let us note a couple important features of the correlation signal. First, the limits of summation in equation (2.8)
are infinite. Usually, though, the support of x[n] and y[n] will be finite, so we do not actually need to perform an
infinite summation. Instead, the duration of the correlation signal will be equal to the sum of the durations of x[n]
and y[n] minus one3 . There will also be transient effects (or edge effects) at the beginning and end of the correlation
signal. These transient effects result from cases where x[n − k] only partially overlaps y[n]. Finally, notice that the
value of the correlation signal at time k = 0 is just the in-place correlation C(x[n], y[n]).
2 Recall that a signal x[n − n0 ] is equal to the signal x[n] shifted n0 samples to the right.
3 Suppose that support interval of x[n] is nx1 ≤ n ≤ nx2 , while the support interval of y[n] is ny1 ≤ n ≤ ny2 . For this general case, we can
see that the first nonzero sample of r[k] will occur at k = ny1 − nx2 . Similarly, the last nonzero sample will fall at k = ny2 − nx1 . Thus, the
duration of r[k] is (ny2 − ny1 ) + (nx2 − nx1 ) + 1 = (ny2 − ny1 + 1) + (nx2 − nx1 + 1) − 1 = duration(x) + duration(y) − 1.

The University of Michigan, All rights reserved 39

Laboratory 2. Signal Correlation and Detection II

1
A
0.5

0
−40 −20 0 20 40 60 80 100

1
B
0.5
0
−0.5

100 200 300 400 500 600 700 800 900 1000
1

0.5 C

0
100 200 300 400 500 600 700 800 900 1000 1100

Figure 2.2: (A) A radar pulse. (B) A received sequence from the radar system, containing two pulses and noise. (C)
The running correlation produced by correlating the radar pulse with the received signal.

2.2.3 Using correlation for signal detection

Whenever we wish to use correlation for signal detection, we use a two-part system. The first part of the system
performs the correlation and produces the correlation value or correlation signal, depending upon whether we are
doing in-place or running correlation. The second part of the system examines the correlation or correlation signal and
makes a decision or sequence of decisions. See the block diagram given in Figure 2.3.

Correlation
Signal
Input Correlation Decision Decision
Signal Calculator Maker

Figure 2.3: A generalized block diagram for a correlation-based detection system.

In the radar example used to motivate running correlation in Section 2.2.2, we simply checked to see if the corre-
lation signal at a given point equals the energy of the transmitted signal. While this will work for the idealized system
presented, real systems are usually much less ideal. We may have multiple reflections, distorted reflections, reductions
in reflection amplitude, and various kinds of environmental noise. In order to address such problems in a wide variety
of systems, we commonly use a simple threshold comparison as our decision maker. For instance, if we compute a
running correlation signal r[n], we might choose a constant c called a threshold and make a decision for each sample
based on the following formula:
1
r ? c (2.9)
0

That is, when the correlation value r is greater than the threshold, c, we decide 1, or “signal present.” If the value
is less than the threshold, we decide 0, or “signal absent.” In our radar example, for instance, we might select the
threshold to be c = E(x)/2.

40 The University of Michigan, All rights reserved

 2.2.4 Using correlation for detection of signals transmitted simultaneously with other signals

1
0
−1
10 20 30 40 50 60 70 80 90 100
1
0
−1
10 20 30 40 50 60 70 80 90 100
1
0
−1
10 20 30 40 50 60 70 80 90 100

Figure 2.4: Example code signals for simultaneous communications

2.2.4 Using correlation for detection of signals transmitted simultaneously with other sig-
nals
Suppose that we wish to design a multi-user wireless communication system that permits several users to simultane-
ously transmit a sequence of message bits. That is, each user will transmit a signal across a common communication
channel (for example, a wire or a small portion of the electromagnetic spectrum) that conveys his/her message bits,
and despite the fact that these signals are received on top of one another, it should be possible to decode the message
bits produced by any one of the users. The users of this system are completely uncoordinated; so no user has any
idea who else might be using the system at any given time. How can we design a system so that each user can use
the system without experiencing interference from the other users? This is the problem faced by the designers of cell
phones and cordless phones, for example.
It turns out that we can use a correlation-based detector to address this problem. To begin, suppose that each user
is trying to send a one bit message to a friend, and suppose each user has a distinct code signal, like those shown in
Figure 2.4. Each code signal is made up of some number of binary chips, which are regions of constant signal value;
the signals shown here each consist of ten chips.4 To send a “one” message bit, the user transmits his or her code
signal. To send a “zero” message bit, the user instead transmits a negated version of his or her code signal (which is is
perfectly anticorrelated with the code signal).
Now, to send a sequence of message bits, the user concatenates these positive and negative versions of the code
signal into a sequence of code signals, which is called the transmitted signal and which is input to the communication
channel. Other users transmit their own message bits in the same fashion, except that, of course, they use different
code signals. For example, Figure 2.5 shows a transmitted signal conveying eight message bits using the top code
signal from Figure 2.4. It also shows this signal with the transmitted signals from three other users added to it. Notice
that the signal in the upper panel is obscured in the lower panel.
When someone, say the user’s friend, receives the signal from the communication channel5 and wishes to decode
the user’s message bits, the friend correlates the received signal with the user’s code signal. Specifically, in-place
correlation of the received signal with the code signal produces a value with which a decision about the first message
bit can be decided. Then in-place correlation of the received signal with a delayed version of the code signal produces
a value from which the second message bit can be decided, and so on. Since each of these correlation values would
equal plus or minus the energy of the code signal if there were no other signals or noise present, it is natural to have
4 Though the code signals clearly have a binary nature, we use the term “chip” to distinguish binary segments of the code signal from binary

message elements, which we call “bits”.

5 For simplicity, we assume the communication channel does not attenuate or otherwise distort the transmitted signal.

The University of Michigan, All rights reserved 41

Laboratory 2. Signal Correlation and Detection II

2
A
0

−2
100 200 300 400 500 600
5
B
0

−5
100 200 300 400 500 600

Figure 2.5: (A) Example of a transmitted signal. (B) The sum of the transmitted signals from four users.

the decision maker decide that a message bit is “one” if the correlation value is positive and “zero” if the correlation
value is negative. That is, a threshold of zero is chosen.
A communication system of this form is said to be a code-division, multiple-access (CDMA) system or a direct-
sequence, spread-spectrum (DSSS) system. They are used, for example, in 900 Mhz cordless telephones. Such
systems work best when the code signals are as different as possible, i.e. when the normalized correlation between the
code signals of distinct users are as near to zero as possible, which is what system designers typically attempt to do.
Consider the examples in Figure 2.4. The first two code signals are completely uncorrelated, as are the second two.
The first and third signals are slightly anticorrelated. The normalized correlation between these signals is only -0.2,
which is small enough that these two code signals will not interfere much with one another.
Above, we’ve indicated that our detection system uses in-place correlation. This means that this system is syn-
chronous; that is, the receiver knows when bits are sent. However, we can actually save ourselves some work by
using running correlation, and then sampling the resulting correlation signal at the appropriate times. This is how
we will implement this communication system in the laboratory assignment. Using the running correlation algorithm
presented in this lab, the “appropriate times” occur in the correlation signal at the end of each code signal. That is, if
our code signals are N samples long, we want to pick off the (k × N )th sample out of the running correlator to decide
the k th message bit.
The threshold used to decode bits in this detection system, which we have chosen to be zero, is actually a design
parameter of the system. If it should happen, for instance, that the system’s noise is biased in a way that we tend to get
slightly positive correlations when no signal is sent, then we would be able to improve performance of the system by
using a positive threshold, rather than a threshold of zero. Alternatively, we might want to decide that no bit has been
sent if the magnitude of the correlation is below some threshold. In this case, we actually have two thresholds. One
separates “no signal” from a binary “one;” the other separates “no signal” from a binary “zero.”

2.2.5 Noise, detector errors, and setting the threshold

Detectors, such as the radar and DSSS detectors we have discussed, must typically operate in the presence of noise.
Here, we begin by discussing the radar example of Section 2.2.2, and conclude with a brief discussion of the DSSS
detector.
When the radar pulse is x[n], the typical received radar signal has the form

y[n] = x[n − n0 ] + w[n] (2.10)

where x[n−n0 ] is the reflected radar pulse and w[n] is noise, i.e. an unpredictable, usually wildly fluctuating signal that

42 The University of Michigan, All rights reserved

 2.2.5 Noise, detector errors, and setting the threshold

normally is little correlated with x[n] or any delayed version of x[n]. To estimate n0 , we perform running correlation
of y[n] with x[n], and the resulting correlation signal is

∞
X
r[k] = x[n − k]y[n]
n=−∞
X∞
= x[n − k](x[n − n0 ] + w[n])
n=−∞
X∞ ∞
X
= x[n − k]x[n − n0 ] + x[n − k]w[n]
n=−∞ n=−∞
= r0 [k] + rw [k] , (2.11)

where r0 [k] is the running correlation of x[n − n0 ] with x[n]. Note that r0 [k] is what r[k] would be if there were no
noise, as given in equation (2.8). rw [k] is the running correlation of the noise w[n] with x[n], which is added to r0 [k].
This shows that the effect of noise is to add rw [k] to r0 [k]. Though in a well designed system rw [k] is usually close to
zero, it will occasionally be large enough to influence the decision made by the decision maker.
In Section 2.2.3, we argued that a threshold-based decision maker was useful for such systems. Then, for example,
when r[k] > c, the decision is that a radar pulse is present at time k, whereas when r[k] < c, the decision is that
no radar pulse is present at time k. Since in the absence of noise r[k] = E(x) when there is a radar pulse at time k,
and since r[k] = 0 when there is no pulse at time k, it is natural, as mentioned in Section 2.2.3 to choose threshold
c = E(x)/2.
Though it makes good sense to use a threshold detector, such a detector will nevertheless occasionally make an
error, i.e. the wrong decision. Indeed, there are two types of errors that a detector can make. First, it could detect a
reflection of the transmitted signal where no actual reflection exists. This is called a false alarm. It occurs at time k
when r0 [k] = 0 and rw [k] > c, i.e. when the part of the correlation due to noise is larger than the threshold. The other
type of error occurs when the detector fails to detect an actual reflection because the noise causes the correlation to drop
below the threshold even though a signal is present. This type of error is called a miss. It occurs when r0 [k] = E(x)
and r[k] = E(x) + rw [k] < c, which in turn happens when rw [k] < c − E(x). In summary, a false alarm occurs when
there is no radar pulse present, yet the noise causes rw [k] > c, and a miss occurs when there is a radar pulse present,
yet the noise causes rw [k] < c − E(x).
Depending on the detection system being developed, these two types of error could be equally undesirable or one
could be more undesirable than another. For instance, in a defensive radar system, false alarms are probably preferable
to misses, since the former are decidedly less dangerous. We can trade off the likelihood of these two types of error
by adjusting the threshold. Raising the threshold decreases the likelihood of a false alarm, while lowering it decreases
the likelihood of a miss.
It is often useful to know the frequency of each type of error. There is a simple way to empirically estimate these
frequencies. First, we perform an experiment where we do not send any radar pulses, but simply record the received
signal y[n], which contains just environmental noise w[k]. We then compute its running correlation r[k] with the radar
pulse x[n], which is just rw [k]. We count the number of times that rw [k] exceeds the threshold c and divide by the
total number of samples. This gives us an estimate the false alarm rate, which is the frequency with which the detector
will decide a radar pulse is present when actually there is none. We can also use this technique to estimate the miss
rate. When a radar pulse is present, an error occurs when rw [k] < c − E(x). Thus, we can estimate the miss rate by
counting the number of times the already computed correlation signal rw [k] is less than c − E(x), and dividing by the
total number of samples.
The signal value distribution is also useful here. If we plot the histogram of values in rw [k], we can use this plot
to determine the error rate estimates. The estimate of false alarm rate is the area of the histogram above values that

The University of Michigan, All rights reserved 43

Laboratory 2. Signal Correlation and Detection II

exceeds c, divided by the total area of the histogram6. Similarly, the estimate of the miss rate is the area of the hisogram
above values that are less than c − E(x)
Assuming that the distribution of rw [k] is symmetric about 0, we can minimize the total error rate (which is simply
the sum of the false alarm and miss rates) by setting a threshold that yields the same number of false alarms as misses.
Since the distribution of rw [k] is assumed to be symmetric, we get an equal number of false alarms and misses when
c = E(x)/2, which is the threshold value suggested earlier.
Next, it is important to note how the error rates depend on the energy of the radar pulse. Consider first the
false alarm rate, which corresponds to the frequency with which the noise induced correlation signal rw [k] exceeds
c = E(x)/2. Suppose for example that the radar pulse is amplified by a factor of two. Then its energy increases by a
factor of four, andR consequently, the threshold c increases by a factor of four. On the other hand, one can see from the
formula rw [k] = y[n]x[n − k] dx that the noise term rw [k] will be doubled. Since the threshold is quadrupled but the
noise term is only doubled, the frequency with which the noise term exceeds the threshold will be greatly decreased,
i.e. the false alarm rate is greatly decreased. A similar argument shows that the miss rate is also greatly decreased.
Thus, we see that what matters is the energy of the radar pulse, in relation to the strength of the noise. If the energy
of the signal increases, but the typical values of the noise w[n] remain about the same, the system will make fewer
errors. By making the energy sufficiently large, we can make the error rate as small as we like. In the lab assignments
to follow, we will observe situations where the noise w[n] is so strong that it completely obscures the radar pulse
x[n − n0 ], yet the radar pulse is long enough that it has enough energy that a correlation detector will make few errors.
Finally, we comment on the effects of noise on the DSSS detector. In this case, instead of deciding whether a
pulse is present or not, the detector decides whether a positive or negative code signal is present. As with the radar
example, this must ordinarily be accomplished in the presence of noise. However, in this case there are two kinds of
noise: environmental noise, similar to that which affects radar, and multiple user noise, which is due to other users
transmitting their own code signals. In the absence of any noise, the in-place correlation r(x, y) computed by the
detector will be +E(x) when the message bit is “one” and −E(x) when the message bit is “zero” , where x[n] is the
user’s code signal. For this reason, using a decision threshold c = 0 is natural. When the message bit is zero, an error
occurs when r(x, y) > 0, which happens when the correlation term rw due to noise exceeds E(x). Similarly, when the
message bit is one, an error occurs when r(x, y) < 0, which happens when rw < −E(x). As with the radar example,
errors occur less frequently when the signal energy becomes larger. This will be evident in the lab assignment, when
code signals of different lengths, and hence different energies, are used.

2.2.6 An algorithm for running correlation

Here, we provide an algorithm for running correlation. One of its primary benefits is that it is easy to understand.
In this algorithm, we imagine the filter as a box into which we drop one new sample of the “incoming” signal and a
corresponding new sample of the correlation signal comes out. This allows the algorithm to be used in real-time: as
samples of our signal arrive (from a radar detector, for instance), we can process the resulting signal with almost no
delay.
In this algorithm we refer to the signal we are looking for (i.e., the transmitted radar signal) as x[n], following
(2.8). The algorithm goes like this:

1. Initialize an input buffer, which is simply an array with length equal to the duration of x[n], to all zeros.

2. For each sample that comes in:

(a) Update the buffer by doing the following:

i. Discard the sample at the beginning of the buffer.
ii. Shift the rest of the samples one place towards the beginning of the buffer.
6 That is, we sum the values in this region of the histogram and divide by the sum of all values in the histogram

44 The University of Michigan, All rights reserved

 2.3 Some M ATLAB commands for this lab

iii. Insert the incoming sample at the end of the buffer.

(b) Initialize a running sum variable to zero.
(c) For each position, n, in the buffer:
i. Multiply the nth position in the input buffer by the nth sample of x[n].
ii. Add the resulting product to the running sum.
(d) Output the running sum as the next sample of the correlation signal.
In the laboratory assignment, you will be asked to complete an implementation of this algorithm. Note that
significant portions of this algorithm can be implemented very simply in M ATLAB. For instance, all of (a) can be
accomplished using a single line of code. Similarly, parts (b) through (d) can all be accomplished in a single line using
one of M ATLAB’s built-in functions and its vector arithmetic capabilities.

2.3 Some M ATLAB commands for this lab

• Calculating in-place correlation: If you have two signals, x and y, that you wish to correlate, simply use the
command

>> c_xy = sum(x.*y);

Note that x and y must be the same size; otherwise M ATLAB will return an error.
• The subplot command: To put several plots on the same figure in M ATLAB, we use the subplot command.
subplot creates a rectangular array of axes in a figure. Figure 2.6 has an example figure with such an array.
Each time you call subplot, you activate one of the axes. subplot takes three input parameters. The first
and second indicate the number of axes per row and the number of axes per column, respectively. The third
parameter indicates which of the axes to activate by counting along the rows7 . Thus the command:

>> subplot(2,3,5)

activates the plot with the circle in Figure 2.6.

• The axis command: The command axis([x_min, x_max, y_min, y_max]) allows us to set the
axis display range for a particular plot. If you wish to change the display range of the currently active plot (or
subplot) so that the x-axis ranges from 5 and 10 and the y-axis ranges from -100 to 100, simply execute the
command

>> axis([ 5, 10, -100, 100]);

Other useful forms of the axis command include axis tight, which fits the axis range closely around the
data in a plot, and axis equal, which assures that the x- and y-axes have the same scale.
• Buffer operations in M ATLAB: It is often useful to use M ATLAB’s vectors as buffers, with which we can shift
values in the buffer towards the beginning or end of the buffer by one position. Such an operation has two parts.
First, we discard the number at the beginning or end of the buffer. If our buffer is a vector b, we can do this
using either b = b(2:end) or b = b(1:end-1). Then, we append a new number to the opposite end of
the buffer using a standard array concatenation operation. Note that we can easily combine these two steps into
a single command. For instance, if b is a row vector and we wish to shift towards the end of the buffer, we use
the command
7 Note in particular that this is the opposite of M ATLAB’s usual convention!

The University of Michigan, All rights reserved 45

Laboratory 2. Signal Correlation and Detection II

1 1 1

0 0 0

−1 −1 −1
0 100 200 0 100 200 0 100 200
1 1 1

0 0 0.5

−1 −1 0
0 100 200 −1 0 1 50 100 150 200

Figure 2.6: Example of a M ATLAB figure with subplots.

>> b = [ new_sample, b(1:end-1) ];

• Counting elements that meet some condition: Occasionally we may want to determine how many elements in
a vector meet some condition. This is simple in M ATLAB because of how the conditional operators are handled.
Recall that for a vector, v, (v == 3) will return a vector with the same size as v, the elements of which are
either 1 or 0 depending upon the truth of the conditional statement. Thus, to count the number of elements in v
that equal 3, we can simply use the command

>> count = sum(v == 3);

2.4 Demonstrations in the Lab Section

• Detector error types

• Why use correlation? Or, when energy detectors break down.

• “In-place” correlation as a similarity measure

• Running correlation

• Multi-user communication

2.5 Laboratory Assignment

In this lab assignment, all signals are discrete-time and their support is assumed to be of the form {1, 2, . . . , N }.

1. (Computing and interpreting in-place correlations) Download the file code_signal.m and use it to create the
following signals:

>> code1 = code_signal(75,10);

>> code2 = code_signal(50,10);
>> code3 = code_signal(204,10);

46 The University of Michigan, All rights reserved

 2.5 Laboratory Assignment

(a) (Plotting code signals) Use subplot and stairs to plot the three code signals on three separate axes
in the same figure. After plotting each signal, make sure that the signal is visible by calling axis([1,
100, -1.5, 1.5]).
• [4] Include your figure, with axis labels on each subplot, a figure number and caption, and the gener-
ating code in your report.
(b) (Calculate statistics) For each of the three signals generated above, calculate:
• [3] Their mean values.
• [3] Their energies.
(c) (Calculate correlations) Calculate the “in-place” correlation and normalized correlation for the following
pairs of signals.
• [2] code1 and code2
• [2] code1 and code3
• [2] code2 and code3
(d) (Classify correlations) For each of the signal pairs given in problem 1c:
• [3] Identify each pair as positively correlated, uncorrelated, or negatively correlated.

2. (Implementing and interpreting running correlation) Download the file run_corr.m, which is a “skeleton”
file for an implementation of the “real-time” running correlation algorithm described in Section 2.2.2. It accepts
two input signals, performs running correlation on them, and produces the correlation signal with a length equal
to the sum of the lengths of the input signals minus one.

(a) (Write the code) Complete the function, following the algorithm given in Section 2.2.2. You can use the
completed demo version of the function, run_corr_demo.dll to check your function’s output8 .
• [10] Include your code in the M ATLAB appendix of your report.
(b) (Compute running correlations) Use run_corr.m to compute the running correlation between the fol-
lowing pairs of signals, and plot the resulting correlation signals on the same figure using subplot.
• [2] code1 and code2.
• [2] code3 and itself.
(c) (Interpret a running correlation) When performing running correlation with a signal and itself, the resulting
correlation signal has some special properties. Look at the correlation signal that you computed between
code3 and itself.
• [1] Is the correlation signal symmetric? (It can be shown that it should be.)
• [2] What is the maximum value of the correlation signal? How does the maximum value relate relate
to the energy of code3?

3. (Using correlation to decode DSSS signals transmitted simultaneously with other signals.) Download the file
lab2_data.mat and load it into your workspace. The file contains the variable which represents a received
signal that is the sum of several message carrying signals, one from each of four users. The message carrying
signal from each user conveys a sequence of bits using the direct-sequence spread-spectrum technique, described
in Section 2.2.4, in which each bit is conveyed by sending a code signal or its negative. Each user has a different
code signal. One of the code signals is the ten chip signal corresponding to the integer 170, while another is the
six chip signal corresponding to the integer 25. The other two code signals are unknown to us. In this problem,
we will try to extract the bit sequences conveyed by the known code signals from dsss. Start by generating the
following code signals:
8 If you cannot get your function working properly, you may use run corr demo.dll to complete the rest of the assignment.

The University of Michigan, All rights reserved 47

Laboratory 2. Signal Correlation and Detection II

>> cs1 = code_signal(170,10);

>> cs2 = code_signal(25,6);

(a) (Plot the signals) First, let’s look at the signals we’re given.
• [3] Use subplot and stairs to plot dsss, cs1, and cs2 on three separate axes of the same
figure.
(b) (Decoding the bits of the user with the longer code signal) Start by using run_corr to correlate the
received signal dsss with the longer code signal cs1. Call the resulting signal cor1. Now, to decode
the sequence of message bits from this user, we need to extract the appropriate samples from cor1. That
is, we need to extract just those samples of the running correlation that correspond to the appropriate
in-place correlations. We can do this in M ATLAB using the following command:
>> sub_cor1 = cor1(length(cs1):length(cs1):length(cor1));
Each sample of sub_cor1 is used to make the decision about one of the user’s bits. When it is greater
than zero, i.e. the correlation of the received signal with the code signal is positive, the decoder decides
the bit is 1. When it is less than zero, the decoder decides the user’s bit is 0.
• [3] On two subplots of the same figure, use plot to plot cor1, and stem to plot sub_cor1.
• [4] Decode the sequence of bits. (You can do this visually or with M ATLAB ˙) (Hint: The sequence is
10 bits long, and the first 3 bits are “011”.)
(c) (Decoding the bits of the user with the shorter code signal) Repeat the procedure in a and b above, this
time using the code signal cs2. Call your correlation signal cor2, and the vector of extracted values
sub_cor2.
• [3] On two subplots of the same figure, use plot to plot cor2, and stem to plot the signal
sub_cor2.
• [4] Decode the sequence of bits. (Hint: there are 17 bits in this sequence.)
• [2] Since the code signal cs2 has less energy (because it is shorter), there is a greater chance of error.
Are there any decoded bits that you suspect might be incorrect? Which ones? Why?

4. (Using running correlation to detect reflected radar pulses) lab2_data.mat also contains three other signals:
radar_pulse, radar_received, and radar_noise. The received signal contains several reflections
of the transmitted radar pulse and noise. The signal radar_noise contains noise with similar characteristics
to the noise in the received signal without the reflected pulses.

(a) (Examining the radar signals) First, let’s take a look at the first two signals.
• [2] Calculate the energy of radar_pulse, E(x).
• [3] Use subplot to plot radar_pulse and radar_received in separate axes of the same
figure.
• [1] Can you identify the reflected pulses in the received signal by visual inspection alone?
(b) (Perform running correlation) Use run_corr to correlate radar_received with radar_pulse.
• [3] Plot the resulting correlation signal.
• [2] Where are the received pulses? Visually identify sample locations of each pulse in the correlation
signal.
• [2] Compare the heights of the peaks to the energy of the radar pulse and explain why the peaks are
larger/smaller than the energy.

48 The University of Michigan, All rights reserved

 2.5 Laboratory Assignment

• [4] Given that the speed of light is 3 × 108 m/s and the sampling frequency of the detector is 107
samples per second, what is the approximate distance to each object9 ?
(c) (Estimating error rates, as in Section 2.2.5) In a real radar detector, the correlation signal would be com-
pared to a threshold to perform the detection. To estimate the error rates for such a detector, let’s consider
a threshold that is equal to one-half the energy of the transmitted pulse, i.e. c = E(x)/2. Perform running
correlation between radar_pulse and radar_noise call the resulting correlation signal noise_c.
• [2] Plot noise_c.
• [3] For how many samples is noise_c greater than this threshold? Use this value to estimate the
false alarm rate.
• [3] For how many samples is noise_c less than this threshold minus the energy of the transmitted
pulse? Use this value to estimate the miss rate.
• [2] What is the total error rate for this threshold?
(d) (Determining error rates from a histogram) As discussed in Section 2.2.5, we can use a histogram to judge
the number of errors as well.
• [3] Plot the histogram of noise_c using 100 bins.
• [3] Describe how you could derive the error numbers in problem 4c from the histogram.
(e) (Setting the threshold to achieve a particular error rate) Suppose that detector false alarms are considered
to be more serious than detector misses. Thus, we have determined that we want to raise the threshold so
that we achieve a false alarm rate of approximately 0.004. Find a threshold that satisfies this requirement.
• [4] What is your threshold?
• [3] What is the false alarm rate on this noise signal with your threshold?
• [3] What is the miss rate on this noise signal with your threshold?
• [2] What is the total error rate for the new threshold? Compare this to the total error rate of the
threshold used in problem 4c.

5. On the front page of your report, please provide an estimate of the average amount of time spent outside of lab
by each member of the group.

9 Remember that the radar pulse must travel to the object and then back again.

The University of Michigan, All rights reserved 49

Laboratory 2. Signal Correlation and Detection II

50 The University of Michigan, All rights reserved

Laboratory 3

Sinusoids and Sinusoidal Correlation

3.1 Introduction
Sinusoids are important signals. Part of their importance comes from their prevalence in the everyday world, where
many signals can be easily described as a sinusoid or a sum of sinusoids. Another part of their importance comes from
their properties when passed through linear time-invariant systems. Any linear time-invariant system whose input is a
sinusoid will have an output that is a sinusoid of the same frequency, but possibly with different amplitude and phase.
Since a great many natural systems are linear and time-invariant, this means that sinusoids form a powerful tool for
analyzing systems.
Being able to identify the parameters of a sinusoid is a very important skill. From a plot of the sinusoid, any
student of signals and systems should be able to easily identify the amplitude, phase, and frequency of that sinusoid.
However, there are many practical situations where it is necessary to build a system that identifies the ampli-
tude, phase, and/or frequency of a sinusoid — not from a plot, but from the actual signal itself. For example, many
communication systems convey information by modulating, i.e. perturbing, a sinusoidal signal called a carrier. To
demodulate the signal received at the antenna, i.e. to recover the information conveyed in the transmitted signal, the
receiver often needs to know the amplitude, phase, and frequency of the carrier. While the frequency of the sinusoidal
carrier is often specified in advance, the phase is usually not specified (it is just whatever phase happens to occur when
the transmitter is turned on), and the amplitude is not known because it depends on the attenuation that takes place
during transmission, which is usually not known in advance. Moreover, though the carrier frequency is specified in
advance, no transmitter can produce this frequency exactly. Thus, in practice the receiver must be able to “lock onto”
the actual frequency that it receives.
Doppler radar provides another example. With such a system, a transmitter transmits a sinusoidal waveform at
some frequency fo . When this sinusoid reflects off a moving object, the frequency of the returned sinusoid is shifted
in proportion to the velocity of the object. A system that determines the frequency of the reflected sinusoid will also
be able to determine the speed of the moving object.
How can a system be designed that automatically determines the amplitude, frequency and phase of a sinusoid?
One could imagine any number of heuristic methods for doing so, each based on how you would visually extract
these parameters. It turns out, though, that there are more convenient methods for doing so – methods which involve
correlation. In this lab, we will examine how to automatically extract parameters from a sinusoid using correlation.
Along the way, we will discover how complex numbers can help us with this task. In particular, we will make use of
the complex exponential signal and see the mathematical benefits of using an “imaginary” signal that does not really
exist.

The University of Michigan, All rights reserved 51

Laboratory 3. Sinusoids and Sinusoidal Correlation

3.1.1 “The Question”

• How can we design a system that automatically determines the amplitude and phase of a sinusoid with a known
frequency?

• How can we design a system that automatically determines the frequency of a sinusoid?

3.2 Background
3.2.1 Complex numbers
Before we begin, let us quickly review the basics of complex numbers.
√ Recall the a complex number z = x + jy is
defined by its real part, x, and its imaginary part, y, where j = −1. Also recall that we can rewrite any complex
number into polar form1 or exponential form, z = rejθ , where r = |z| is the magnitude of the complex number and
θ = angle(z) is the angle. We can convert between the two forms using the formulas

x = r cos(θ) (3.1)
y = r sin(θ) (3.2)

and
p
r = x2 + y 2 (3.3)


tan−1 xy
, x≥0
θ = (3.4)
tan−1 xy + π, x < 0

A common operation on complex numbers is the complex conjugate. The complex conjugate of a complex number,
z ∗ , is given by

z∗ = x − jy (3.5)
= re−jθ (3.6)

Conjugation is particularly useful because zz ∗ = |z|2 .

Euler’s2 formula is a very important (and useful) relationship for complex numbers. This formula allows us to
relate the polar and rectangular forms of a complex number. Euler’s formula is

ejθ = cos(θ) + j sin(θ) (3.7)

Equally important are Euler’s inverse formulas:

ejθ + e−jθ
cos(θ) = (3.8)
2
ejθ − e−jθ
sin(θ) = (3.9)
2j

It is strongly recommended that you commit these three equations to memory; you will be using them regularly
throughout this course.
1 Sometimes the polar form is written as z = r \θ, which is a mathematically less useful form. This form, however, is useful for suggesting the

interpretation of r as a radius and θ as an angle.

2 Pronounced “oiler’s”.

52 The University of Michigan, All rights reserved

 3.2.2 Sinusoidal and complex exponential signals in continuous time

3.2.2 Sinusoidal and complex exponential signals in continuous time

Recall that a continuous-time sinusoid in standard form, s(t), is given by the formula

s(t) = A cos(ω0 t + φ), (3.10)

where A > 0 is the sinusoid’s amplitude, ω0 is the sinusoid’s frequency given in radian frequency (radians per second),
and φ is the sinusoid’s phase. It is also common to represent such a sinusoid in the following form

s(t) = A cos(2πf0 t + φ), (3.11)

where f0 is the sinusoid’s frequency given in Hertz (Hz, or cycles per second). Note that ω0 = 2πf0 . The frequency
of a sinusoid is generally restricted to be positive.
The notation for sinusoids also extends to a special signal known as the complex exponential signal3 . Complex
exponential signals are very similar to sinusoids, and have the same three parameters. We define a continuous-time
complex exponential signal, c(t), in standard form as

c(t) = Aej(ω0 t+φ) (3.12)

It is generally useful to consider that sinusoids are composed of a sum of complex exponential signals by using Euler’s
inverse formulas. Thus, a sinusoid in standard form can be rewritten in several different ways:

s(t) = A cos(ω0 t + φ) (3.13)

A h j(ω0 t+φ) i
= e + e−j(ω0 t+φ) (3.14)
2
A
= (c(t) + c∗ (t)) (3.15)
2 n o
= Re Aej(ω0 t+φ) (3.16)

Using Euler’s formula, we can also interpret a complex exponential signal c(t) as the sum of a real cosine wave and
an imaginary sine wave:
c(t) = A cos(ω0 t + φ) + jA sin(ω0 t + φ) (3.17)
Sometimes it is useful to visualize a complex exponential signal as a “corkscrew” in three dimensions, as in Figure
3.1. Note that it is common to permit complex exponential signals to have either positive or negative frequency. The
sign of the frequency determines the “handedness” of the corkscrew.

3.2.3 Finding the amplitude and phase of a sinusoid with known frequency
We’ve suggested that we can use correlation to help us determine the amplitude and phase of a sinusoid with known
frequency. Suppose that we have a continuous-time sinusoid (the target sinusoid)

s(t) = A cos(ω0 t + φ) (3.18)

with known frequency ω0 , but unknown amplitude A and phase φ, which we would like to find. We can perform in-
place correlation4 between this sinusoid and a reference sinusoid, u(t), with the same frequency and known amplitude
3 These are sometimes referred to simply as complex exponentials.
R
4 In-place correlation between two real, continuous-time signals, x(t) and y(t) is defined as C(x, y) = ab x(t)y(t)dt. The length (b − a) is
the correlation length.

The University of Michigan, All rights reserved 53

Laboratory 3. Sinusoids and Sinusoidal Correlation

0.5
Imaginary
0

−0.5
1
0.5
−1 0
0
5 −0.5
10
15 −1 Real
20
Time

Figure 3.1: Three-dimensional plot of a complex exponential signal.

and phase. Without loss of generality, let u(t) have A = 1 and φ = 0. Then5 ,
Z t2
C(s, u) = A cos(ω0 t + φ) cos(ω0 t)dt (3.19)
t1
Z t2
A
= cos(φ) + cos(2ω0 t + φ)dt (3.20)
2 t1
 t2
A 1
= cos(φ)t + sin(2ω0 t + φ) (3.21)
2 2ω0 t1

Since we know the frequency, ω0 , we can easily set the limits of integration to include an integer number of funda-
mental periods of our sinusoids. In this case, the second term evaluates to zero and the correlation reduces to

A
C(s, u) = cos(φ)(t2 − t1 ) (3.22)
2

This formula is a useful first step. If we happen to know the phase φ, then we can readily calculate the amplitude A of
s(t) from C(s, u). Similarly, if we know the amplitude A, we can narrow the phase φ down to one of two values. If
both amplitude and phase are unknown, though, we cannot uniquely determine them.
If the interval over which we correlate is not a multiple of the fundamental period of u(t), then the second term in
equation (3.21) may be nonzero. However, if as commonly happens ω0 is much greater than one, then the second term
will be so small that it can be ignored, and equation (3.22) holds with approximate equality.
To resolve the ambiguity when both amplitude and phase are unknown, one common approach to correlate with a
second reference sinusoid that is π2 out of phase with the first. Here, though, we will explore a different method which
is somewhat more enlightening. Notice what happens if we use a complex exponential,

c(t) = ejω0 t (3.23)

1 1
5 Recall that cos(A) cos(B) = 2
cos(A − B) + 2
cos(A + B).

54 The University of Michigan, All rights reserved

 3.2.3 Finding the amplitude and phase of a sinusoid with known frequency

as our reference signal6 :

Z t2
C(s, c) = s(t)c∗ (t) dt (3.24)
t1
Z t2
= A cos(ω0 t + φ)e−jω0 t dt (3.25)
t1
Z t2
A h j(ω0 t+φ) i
= e + e−j(ω0 t+φ) e−jω0 t dt (3.26)
t1 2
Z t2
A
= ejφ + e−j(2ω0 t+φ) dt (3.27)
2 t1
 t
A jφ −1 −j(2ω0 t+φ) 2
= e t+ e (3.28)
2 2jω0 t1

If we again assume that we are correlating over an integer number of periods of our target sinusoid, then the second
term goes to zero and we are left with
A jφ
C(s, c) = e (t2 − t1 ). (3.29)
2
Our correlation has resulted in a simple complex number whose magnitude is directly proportional to the amplitude of
the original sinusoid and whose angle is identically equal to its phase! We can easily turn the above formula inside-out
to obtain
2
A = |C(s, c)| (3.30)
t2 − t1
φ = angle(C(s, c)) (3.31)

We can also see from equation (3.29) that in correlating with a complex exponential signal, we have effectively
calculated the phasor7 representation of our sinusoid.
As with the case of correlating with a sinusoid, we note that when the interval over which we correlate is not
a multiple of the fundamental period of c(t), then the second term in equation (3.28) is not zero. However, if as
commonly happens ω0 is much greater than 1, then the second term will again be small enough that it can be ignored,
and equations (3.29), (3.30), and (3.31) hold with approximate equality.

The Amplitude and Phase Calculator

In this lab we will implement a system that estimates the amplitude and phase of a sinusoid with a known frequency.
Since we will do this using a computer and M ATLAB, we must necessarily work with sampled version of the signals
s(t) and c(t). Specifically, if Ts denotes the sampling period, then we work with the discrete-time signals

s[n] = s(nTs ) = A cos(ω0 Ts n + φ) (3.32)

c[n] = c(nTs ) = ejω0 Ts n (3.33)

As shown below, when Ts is small, the correlation between s(t) and c(t) can be approximately computed from the
correlation between s[n] and c[n]. Let {n1 , . . . , n2 } denote the discrete-time interval corresponding to the continuous-
time interval [t1 , t2 ], and let N = n2 − n1 + 1 denote the number of samples taken in the interval [t1 , t2 ], so that
R
6 Notice that we conjugate our complex exponential here. This is because correlation between two complex signals is defined as x(t)y ∗ (t)dt.
7 When jφ
we represent a sinusoid with amplitude A and phase φ as the complex number Ae to simplify the calculation of a sum of two or more
sinusoids, this complex number is known as a phasor.

The University of Michigan, All rights reserved 55

Laboratory 3. Sinusoids and Sinusoidal Correlation

Signal vector
Amplitude and Amplitude
Support vector Phase Calculator
(APC) Phase
Frequency

Figure 3.2: System diagram for the “amplitude and phase calculator.”

t2 − t1 ≈ N Ts . Then,
Z t2
C(s, c) = s(t) c∗ (t) dt (3.34)
t1
X
n2 Z (n+1)Ts
= s(t) c∗ (t) dt (3.35)
n=n1 nTs

X
n2 Z (n+1)Ts
≈ s(nTs ) c∗ (nTs ) dt (3.36)
n=n1 nTs

X
n2
= s(nTs ) c∗ (nTs ) Ts (3.37)
n=n1
X
n2
= s[n] c∗ [n] Ts (3.38)
n=n1
= Cd (s, c) Ts (3.39)

where the approximation leading to the third relation is valid because Ts is small, and consequently the signals s(t) and
c(t) change little over each Ts second sampling interval, and where we use Cd (s, c) to denote the correlation between
the discrete-time signals s[n] and c[n], to distinguish it from the correlation between continuous time signals s(t) and
c(t). We see from this derivation that the continuous-time correlation is approximately the discrete-time correlation
multiplied by the sampling interval, i.e.

C(s, c) ≈ Cd (s, c)Ts (3.40)

We will use this value of correlation in equations (3.30) and (3.31) to estimate the amplitude and phase of a continuous-
time sinusoid.
In the laboratory assignment, we will be implementing an “amplitude and phase calculator” (APC) as a M ATLAB
function. A diagram of this system is shown in Figure 3.2. The system takes three input parameters. The first is the
signal vector which contains the sinusoid itself. The second is the support vector for the sinusoid. The third input
parameter is the frequency of the reference sinusoid in radians per second. Note that for the system’s output to be
exact, the input sinusoid must be defined over exactly an integer number of fundamental periods.
The system outputs the sinusoid’s amplitude and its phase in radians. The system calculates these outputs by
first computing the in-place correlation given by equations (3.37) or (3.38). Then, this correlation value is used with
equations (3.30) and (3.31) to compute the amplitude and phase. Note that in equation (3.30), we need to replace
t2 − t2 with N = n2 − n1 + 1 when implementing in discrete time.

56 The University of Michigan, All rights reserved

 3.2.4 Determining the frequency of a target sinusoid

3.2.4 Determining the frequency of a target sinusoid

Suppose now that we are given the task of making a system that automatically determines the frequency of a target
sinusoid. It turns out that correlation can help us with this problem as well. Consider the following case. Let the target
sinusoid be defined by s(t) = A cos(ωs t + φ), where ωs , A, and φ are all unknown. We correlate s(t) with a complex
exponential signal, c(t) = ejωc t , with frequency ωc , where ωs 6= ωc :
Z t2
C(s, c) = s(t)c∗ (t) dt (3.41)
t1
Z t2
= A cos(ωs t + φ)e−j(ωc t) dt (3.42)
t1
Z t2
A h j(ωs t+φ) i
= e + e−j(ωs t+φ) e−j(ωc t) dt (3.43)
t1 2
Z t2
A
= ej[(ωs −ωc )t+φ] + e−j[(ωs +ωc )t+φ] dt (3.44)
2 t1
 t2
A 1 1
= ej[(ωs −ωc )t+φ] + e−j[(ωs +ωc )t+φ] (3.45)
2 ωs − ωc ωs + ωc t1

Here, let us make a simplifying assumption and assume that (ωs + ωc ) is sufficiently large that we can neglect the
second term. Then, we have
A h i
C(s, c) ≈ ej[(ωs −ωc )t2 +φ] − e−j[(ωs −ωc )t1 +φ] (3.46)
2(ωs − ωc )

The resulting equation depends primarily on the frequency difference (ωs − ωc ) between the target sinusoid and our
reference signal. Though it is not immediately apparent, the value of this correlation converges to the value of equation
(3.29) as the (ωs − ωc ) approaches zero.
e c), defined as
Consider now the length-normalized correlation, C(s,

e c) = C(s, c) .
C(s, (3.47)
t2 − t1
One can see from equation (3.29) that when the reference and target signals have the same frequency, the length-
normalized correlation does not depend on the length of the signal. However, when the signals have different frequen-
cies, one can see from equations (3.46) and (3.47) that the magnitude of the length-normalized correlation becomes
smaller as we correlate over a longer period of time. (This happens more slowly as the frequency difference becomes
smaller.) In the limit as the correlation length goes to infinity, the length-normalized correlation goes to zero unless
the frequencies match exactly. This is a very important theoretical result in signals and systems.
Another special case occurs when we correlate over a common period of the target and reference signals. This
occurs when our correlation interval includes an integer number of periods of both the target signal and reference
signal. In this case, the correlation in equation (3.46), for signals of different frequencies, is identically zero8 . Of
course, the correlation is not zero when the frequencies match. Note that this is the same condition required for
equation (3.29) to be exact.
How does all of this help us to determine the frequency of the target sinusoid? The answer is perhaps less elegant
than one might hope; basically, we “guess and check”. If we have no prior knowledge about possible frequencies
for the sinusoid, we need to check the correlation with complex exponentials having a variety of frequencies. Then,
whichever complex exponential yields the highest correlation, we take the frequency of that complex exponential as
our estimate of the frequency of the target signal. In the next section, we will formalize this algorithm.

The University of Michigan, All rights reserved 57

Laboratory 3. Sinusoids and Sinusoidal Correlation

A frequency estimation algorithm

Suppose that we have a continuous-time target sinusoid s(t) with support [0, T ] with unknown amplitude, frequency,
and phase. To estimate these parameters, we’ll calculate the length-normalized correlation between this signal and
reference complex exponentials with various frequencies over the signal’s T second length. Then, we look for the
frequency that produces the largest correlation.
We choose the frequencies of these complex exponentials to be multiples 1/T so that the correlation is over an
integer number of periods of each complex exponential. That is, the frequencies will be 1/T, 2/T, . . .. As in the
previous subsection, we’ll need to approximately compute the correlation from samples of s(t) and each reference
exponential, taken with some small sampling interval Ts . For convenience we’ll take N samples and choose Ts =
T /N for some large even integer N . With samples taken at intervals of Ts seconds, we cannot hope to do a good job
of correlating with complex exponentials with very high frequency. The rule of thumb that you will learn in Chapter 4
is that, at the very least, two samples are needed from each period of the signal being sampled. Therefore, the highest
frequency with which we will correlate is, approximately, 1/(2Ts ). Specifically, we will correlate s(t) with complex
exponentials at frequencies
1 2 N 1
, , ..., = (3.48)
T T 2T 2Ts
Then, for k = 1, 2, . . . , N/2, the length normalized correlation of s(t) with the complex exponential at frequency k
T
is (using equations (3.37), (3.38) and (3.47))
N −1
1 X k
X[k] ≈ s(nTs ) e−j2π T nTs Ts (3.49)
T n=0
N −1
1 X k
= s[n] e−j2π N n (3.50)
N n=0

where we have used the fact that Ts /T = 1/N and where we have denoted the result X[k] because this is the notation
used in future labs for the last formula above. Thus, the output output of these correlations is the set of N/2 numbers
X[1], . . . , X[N/2]. Remember that X[k] will generally be complex. To estimate the frequency of the target sinusoid,
we simply identify the value of k for which |X[k]| is largest. With kmax denoting this value, our estimated frequency,
ωest , is
kmax kmax
ωest = 2π = 2π (3.51)
T N Ts
Now that we have estimated the frequency, we should also be able to estimate the amplitude and phase as well. In
fact, we have almost calculated these estimates already. From equations (3.30) and (3.31), they are:

Aest = 2|X[kmax]| (3.52)

φest = angle(X[kmax ]) (3.53)

There is one potential problem here, however. Previously, we assumed the frequency was known exactly when deter-
mining the amplitude and phase; now, we only know the frequency approximately. In the laboratory assignment, we
will see the effect of this approximation.
In the laboratory assignment, we will be developing a system that can automatically estimate the amplitude, phase,
and frequency of a sinusoid. A block diagram of the “frequency, amplitude, and phase estimator” (FAPE) system
is given in Figure 3.3. Unlike the APC, this system takes only two input parameters: a signal vector and the corre-
sponding support vector. The system has four output parameters. The first three are the estimates of the frequency,
amplitude, and phase of the input sinusoid. The fourth is the vector of correlations X[1], . . . , X[N/2] produced by the
correlations. It is often useful to examine this vector to get a sense of what the system is doing.

58 The University of Michigan, All rights reserved

 3.3 Some M ATLAB commands for this lab

Signal vector Frequency, Frequency

Amplitude, and Amplitude
Support vector Phase Estimator Phase
(FAPE) Vector of
Correlations

Figure 3.3: System diagram for the “frequency, amplitude, and phase estimator.”

Estimating doppler shift

When a sinusoidal transmitted signal reflects off an object moving towards the transmitter at speed vo , the signal
returned to the transmitter is again a sinusoid, but with the higher frequency
v
ωr = ωt ∗ (3.54)
v − vo
where ωt is the original frequency of the transmitted signal and v is the speed of propagation in the given medium.
This is called the Doppler shift phenomenon. If one measures ωr , e.g. with FAPE, then one can use equation (3.54)
to compute the speed of the object vs , assuming of course, that ωt and v are known. Indeed, this is how radar systems
are able to measure the speed of automobiles, airplanes, baseballs, wind, etc. In the lab assignment you will be asked
to estimate the velocity of an underwater object from a reflection of a sonar signal.

3.3 Some M ATLAB commands for this lab

• Constructing complex numbers: M ATLAB represents all complex numbers in rectangular form. To √ enter
a complex number, simply type 5 + 6*j (for instance). Note that both i and j are used to represent −1
(unless you have used one or the other as some other variable). To enter a complex number in polar form, type
2*exp(j*pi/3) (for instance).
You may be wondering how M ATLAB actually works with complex numbers, given that complex numbers are,
in general, the sum of a real number and an imaginary one. The fact is that the imaginary component of a
complex number is in fact a real number, which M ATLAB stores in the usual way. It thinks of a complex
number as a pair of floating point numbers, one to be interpreted as the real part and the other to be interpreted
as the imaginary part. And it knows the rules of arithmetic to apply to such pairs of numbers in order to do what
complex arithmetic is supposed to do.
• Extracting parts of complex numbers: If z contains a complex number (or an array of complex numbers),
you can find the real and imaginary parts using the commands real(z) and imag(z), respectively. You can
obtain the magnitude and angle of a complex number (or an array of complex numbers) using the commands
abs(z) and angle(z), respectively.
• Complex conjugation: To compute the complex conjugate of a value (or array) z, simply use the M ATLAB
command conj(z).
• Finding the index of the maximum value in a vector: Sometimes we don’t just want to find the maximum
value in a vector; instead, we need to know where that maximum value is located. The max command will do
this for us. If v is a vector and you use the command

>> [max_value, index] = max(v);

The University of Michigan, All rights reserved 59

Laboratory 3. Sinusoids and Sinusoidal Correlation

the variable max_value will contain largest value in the vector, and index contains position of max_value
in v.

• M ATLAB commands to help you visually determine the amplitude, frequency, and phase of a sinusoid:
Sometimes you may need to determine the frequency, phase, and amplitude of a sinusoid from a MATLAB
plot. In these cases, there three commands that are quite useful. First, the command grid on provides in-
cludes a reference grid on the plot; this makes it easier to see where the sinusoid crosses zero (for instance).
The zoom command is also useful, since you can drag a zoom box to zoom in on any part of the sinusoid.
Finally, you can use axis in conjunction with zoom to find the period of the signal. To do so, simply
zoom in on exactly one period of the signal and type axis. M ATLAB will return the current axis limits as
[x_min, x_max, y_min, y_max].

• Calling apc: The function apc, which you will be writing in this laboratory, estimates amplitude and phase
of a continous-time target sinusoid from its samples. The input parameters are a (sampled) target sinusoid s,
the sinusoid’s support vector t, and the continuous-time frequency w0 in radians per second. We call apc like
this:

>> [A,phi] = apc(s,t,w0);

Note that a compiled version of this function, called apc_demo.dll, is also available.

• Calling fape: The function fape, which you will be writing in this laboratory, implements the frequency,
amplitude, and phase estimator system. This function accepts the samples of a target continuous-time sinusoid
s and it’s support vector t, like this:

>> [frq,A,phi,X] = apc(s,t);

where frq is the estimated frequency in radians per second, A is the estimated amplitude, phi is the estimated
phase, and X is the vector of correlations, X[1], . . . , X[N/2] between s and each reference complex exponential.
Note that a compiled version of this function called, fape_demo.dll, is also available.

3.4 Demonstrations in the Lab Section

• Complex Numbers in M ATLAB

• Sinusoids and complex exponentials in M ATLAB

• Sinusoidal correlation: matching frequencies

• Sinusoidal correlation: different frequencies

• FAPE: the Frequency, Amplitude, and Phase Estimator

3.5 Laboratory Assignment

1. (Understanding sinusoids) Execute the following commands:

>> t = linspace(-0.5, 2, 1000);

>> plot(t,cos(linspace(-7.5,27,1000)),'k:');

60 The University of Michigan, All rights reserved

 3.5 Laboratory Assignment

(a) (Extracting sinusoid parameters) Visually identify the amplitude, continuous-time frequency, and phase of
the continuous-time (sampled) sinusoid that you’ve just plotted.
• [4] Include your estimated values in your report. Reduce your answers to decimal form.
• [3] What is the phasor that corresponds to this sinusoid? Write it in both rectangular and polar form.
(Again, keep your answers in decimal form. You should use MATLAB to perform these calculations.)
(b) (Checking your parameters) Verify your answers in the previous problem by generating a sinusoid using
those parameters and plotting them on the above graph using hold on. Use t as your time axis/support
vector. The new plot should be close to the original, but it does not need to be exactly correct.
• [3] Include the resulting graph in your report. Remember to include a legend.

2. (The Amplitude and Phase Calculator) In this problem we will complete and test a function which implements
the “Amplitude and Phase Calculator”, as described in Section 3.2.3. Download the file apc.m. This is a
“skeleton” M-file for the “amplitude and phase calculator”. Also, generate the following sinusoid (s_test)
with its support vector (t_test):

>> t_test = 0:0.01:.99;

>> s_test = 1.3*cos(t_test*10*pi + 2.8);

(a) (Identify sinusoid parameters by hand) What are the amplitude, frequency in radians per second, and phase
of s_test?
• [2] Include your answers in your lab report.
(b) (Write the APC) Complete the function apc. You should use the signal s_test to test the operation of
your function. You may also wish to use the compiled function apc_demo.dll to test your results on
other sinusoids.
• [10] Include the code for apc in your M ATLAB appendix.
(c) (Test APC on a sinusoid with unknown parameters)
Download the file lab3_data.mat. This .mat file contains the support vector (t_samp) and signal
vector (s_samp) for a sampled continuous-time sinusoid with a continuous-time frequency of ω0 = 200π
radians.
• [4] From t_samp, determine the sampling period, Ts , of this signal.
• [3] Use apc9 to determine the amplitude and phase of the sinusoid exactly.
(d) (APC in a non-ideal case) What happens if we use apc to correlate over a non-integral number of periods
of our target sinusoid? We will investigate this question in this problem and the next. First, let’s examine
a single non-integral number of periods. Generate the following sinusoid:
>> apc_support = 0:0.1:8;
>> apc_test = cos(apc_support*2*pi/3);
2π
This is a sinusoid with a frequency of ω0 = 3 radians per second, unit amplitude, and zero phase shift.
• [2] Plot apc_test and include the plot in your report.
• [2] What is the fundamental period of apc_test?
• [2] Approximately how many periods are included in apc_test?
• [2] Use apc to estimate the amplitude and phase of this sinusoid. What are the amplitude and phase
errors for this signal?
9 If you failed to correctly complete apc.m, you may use apc demo.dll for the following problems. If you use the demo function, please

indicate this in your lab report.

The University of Michigan, All rights reserved 61

Laboratory 3. Sinusoids and Sinusoidal Correlation

(e) (APC in many non-ideal cases) Now we wish to examine a large number of different lengths of this
sinusoid. You will do this by writing a for loop that repeats the previous part for many different values of
the length of the incoming sinusoid. Specifically, write a for loop with loop counter support_length
ranging over values of 1:0.1:50 seconds. In each iteration of the loop, you should
i. Set apc_support equal to 0:0.1:(support_length-0.1),
ii. Recalculate apc_test using the new apc_support,
iii. Use apc to estimate the amplitude and phase of apc_test, and
iv. Store these estimates in two separate vectors.
Put your code in an M-file script so that you can run it easily.
• [8] Include your code in the M ATLAB appendix.
• [4] Use subplot to plot the amplitude and phase estimates as a function of support length in two
subplots of the same figure. You should be able to see both local oscillation of the estimates and a
global decrease in error with increased support length.
• [3] At what support lengths are the amplitude estimates correct (i.e., equal to 1)?
• [3] What minimum support length do we need to be sure that the phase error is less than 0.01 radians?

3. (The Frequency, Amplitude, and Phase Estimator) In this problem, we’ll explore the frequency, amplitude and
phase estimator, as described in Section 3.2.4. Download the file fape.m. This is a “skeleton” M-file for the
“frequency, amplitude, and phase estimator” system.

(a) (Write the FAPE) Complete the fape function. You can use t_test and s_test from Problem 2 along
with the compiled fape_demo.dll to check your function’s results.
• [16] Include the completed code in your report’s M ATLAB appendix.
• [3] What are the frequency (in radians per second), amplitude, and phase estimates returned by fape
for t_test and s_test? Are these estimates correct?
• [3] Use stem and abs to plot the magnitude of the vector of correlations returned by fape versus
the associated frequencies.
• [3] What do you notice about this plot? What can you deduce from this fact? (Hint: Consider what
this plot tells you about the returned estimates.)
(b) (Running FAPE on in a non-ideal case) In this problem, we’ll see what happens to FAPE when the tar-
get sinusoid does not include an integral number of periods. lab3_data.mat contains the variables
fape_test_t (a support vector) and fape_test_s (its associated sinusoidal signal). Run fape on
this signal.
• [3] What are the frequency in radians per second, amplitude, and phase estimates that are returned?
• [3] Use stem and abs to plot the magnitude of the returned vector of correlations.
• [4] Plot fape_test_s and a new sinusoid that you generate from the parameter estimates returned
by FAPE on the same figure (using hold on). Use fape_test_t as the support vector for the
new sinusoid. Make sure you use different line types and include a legend.
• [3] What can you say about the accuracy of estimates returned by FAPE?
• [3] Compare the plot of the correlations generated in this problem and in Problem 3a. What do these
different plots tell you?
Food for thought: Investigate the error characteristics of fape as you did with apc in problem 2e. Do the
frequency, amplitude, and phase estimates improve as we use longer support lengths? Which parameter is
exhibits the most error? What does the vector of correlations, X[k], tell you about these estimates?

62 The University of Michigan, All rights reserved

 3.5 Laboratory Assignment

(c) Measuring speed via Doppler shift. A sonar transmitter in the ocean emits a sinusoidal signal with fre-
quency 1000 Hz, and the signal reflects off an object moving toward the transmitter. The received signal
can be found in the M ATLAB workspace lab3_data.mat. The signal vector is called s_sonar and
the support vector is t_sonar. The speed of sound in salt water is approximately 1450 meters/second.
(Note: because the signal is rather long, it may take a little while for FAPE to run.)
• [4] Estimate the speed of the object.
Food for thought: Use randn to add some random noise to s sonar and observe how your estimate
changes. How much noise do you need to add to produce an error? Does the system degrade gracefully?
(That is, is the amount of error proportional to the amount of noise?)

4. On the front page of your report, please provide an estimate of the average amount of time spent outside of lab
by each member of the group.

The University of Michigan, All rights reserved 63

Laboratory 3. Sinusoids and Sinusoidal Correlation

64 The University of Michigan, All rights reserved

Laboratory 4

Fourier Series and the DFT

4.1 Introduction
As emphasized in the previous lab, sinusoids are an important part of signal analysis. We noted that many signals
that occur in the real world are composed of sinusoids. For example, many musical signals can be approximately
described as sums of sinusoids, as can some speech sounds (vowels in particular). It turns out that any periodic signal
can be written exactly as a sum of amplitude-scaled and phase-shifted sinusoids. Equivalently, we can use Euler’s
inverse formulas to write periodic signals as sums of complex exponentials. This is a mathematically more convenient
description, and the one that we will adopt in this laboratory and, indeed, in the rest of this course. The description of
a signal as a sum of sinusoids or complex exponentials is known as the spectrum of the signal.
Why do we need another representation for a signal? Isn’t the usual time-domain representation enough? It turns
out that spectral (or frequency-domain) representations of signals have many important properties. First, a frequency-
domain representation may be simpler than a time-domain representation, especially in cases where we cannot write an
analytic expression for the signal. Second, a frequency-domain representation of a signal can often tell us things about
the signal that we would not know from just the time-domain signal. Third, a signal’s spectrum provides a simple way
to describe the effect of certain systems (like filters) on that signal. There are many more uses for frequency-domain
representations of a signal, and we will examine many of them throughout this course. Spectral representations are
one of the most central ideas in signals and systems theory, and can also be one of the trickiest to understand.
Consider the following problem. Suppose that we have a signal that is actually the sum of two different signals.
Further, suppose that we would like to separate one signal from the other, but the signals overlap in time. If the signals
have frequency-domain representations that do not overlap, it is still possible to separate the two signals. In this way,
we can see that frequency-domain representations provide another “dimension” to our understanding of signals.
In this laboratory, we will examine two tools that allow us to use spectral representations. The Fourier Series is
a tool that we use to work with spectral representations of periodic continuous-time signals. The Discrete Fourier
Transform (DFT) is an analogous tool for periodic discrete-time signals. Each of these tools allow both analysis (the
determination of the spectrum of the time-domain signal) and synthesis (the reconstruction of the time-domain signal
from its spectrum). Though you may not be aware of it, you have already performed DFT analysis; the “frequency,
amplitude, and phase estimator” system that you implemented in Laboratory 3 actually performs DFT analysis.

4.1.1 “The Questions”

• How can we determine the spectral content of signals?

• How can we separate signals that overlap in time?

The University of Michigan, All rights reserved 65

Laboratory 4. Fourier Series and the DFT

(B)
(A)

−5ω0 −4ω0 −3ω0 −2ω0 −ω0 ω0 2ω0 3ω0 4ω0 5ω0

Figure 4.1: (A) A time-domain representation of a signal. (B) A frequency-domain representation of the same signal
produced with the Fourier Series.

4.2 Background
4.2.1 Frequency-domain representations
This section provides an overview to the Fourier series approach of the frequency-domain representation of continuous-
time signals.
So far, we have typically thought of signals as time-varying quantities, like s(t). When we plot these signals, we
generally place time along the horizontal axis and signal value along the vertical axis. The idea behind the frequency-
domain representation of a signal is similar. Rather than plotting signal value versus time, we plot a spectral value
versus frequency. Doing this involves a transformation of the signal. Figure 4.1 shows an example of a time-domain
and frequency-domain representation of a signal. Note that we can think of the result of the transform as a signal as
well, a signal whose independent variable is frequency rather than time.
The frequency domain representation of a signal (i.e., its spectrum) is easy to construct when the signal is composed
of a sum of simple complex exponential signals. In this case, the spectrum consists of a few isolated spectral lines
(“spikes”) on the frequency axis at the frequencies of those complex exponentials. These spectral lines are complex-
valued, and their magnitudes and angles equal the amplitudes and phases of the corresponding complex exponentials.
Alternatively, we may draw two separate spectral line plots — one showing the magnitude and the other showing their
angles.
If we add more complex exponentials to our signal, then we simply add more spectral lines to its frequency-domain
representation. Eventually, if we add enough complex exponentials (possibly an infinite number), we can create any
signal that we might want. This includes signals that do not look very sinusoidal, like square waves and sawtooth
waves. We will use this result for periodic signals in this laboratory assignment.

4.2.2 Periodic Continuous-Time Signals — The Fourier Series

Suppose that we have a periodic continuous-time signal s(t) with period T seconds. We have claimed that any such
signal can be represented as a sum of complex exponential signals. We now assert that these complex exponentials
have harmonically related frequencies. Specifically, their frequencies (in radians per second) form a harmonic series

. . . , −3ω0 , −2ω0 , −ω0 , 0, ω0 , 2ω0 , 3ω0 , . . . , (4.1)

where
2π
ω0 = (4.2)
T
is the fundamental frequency. The frequency kω0 , k ≥ 2, is called the k-th harmonic of the fundamental frequency,
or the k-th harmonic frequency for short.

66 The University of Michigan, All rights reserved

 4.2.2 Periodic Continuous-Time Signals — The Fourier Series

Next we assert that the representation of s(t) in terms of complex exponentials with these frequencies is given by
the Fourier Series synthesis formula1 :
2π(−2) 2π(−1) 2π0 2π1 2π2
s(t) = . . . α−2 ej T t + α−1 ej T t
+ α0 ej T t + α1 ej T t + α2 ej T t + ...
X ∞
2πk
= αk ej T t , (4.3)
k=−∞

where the αk ’s, which are called Fourier coefficients. The Fourier coefficients are determined by the Fourier series
analysis formula Z
1 2πk
αk = s(t)e−j T t dt , (4.4)
T hT0 i
R
where hT i indicates an integral over any T second interval2 . In other words, the Fourier synthesis formula shows that
2πk
the complex exponential component of s(t) at frequency T is
2πk
αk ej T t . (4.5)

Similarly, the Fourier analysis formula shows how the complex exponential components can be determined from s(t),
even when no exponential components are evident.
In general, the Fourier coefficients, i.e. the αk ’s, are complex. Thus, they have a magnitude |αk | and a phase
or angle \αk . The magnitude |αk | can be viewed as the strength of the exponential component at frequency kω0 =
2πk/T , while the angle \αk gives the phase of that component. The coefficient α0 is the DC term; it measures the
average value of the signal over one period.
Once we know the αk ’s, the spectrum of s(t) is simply a plot consisting of spectral lines at frequencies

. . . , −2ω0 , −ω0 , 0, ω0 , 2ω0 , . . . .

The spectral line at frequency kω0 is drawn with height indicating the magnitude |αk | and is labeled with the complex
value of αk . Alternatively, two separate spectral line plots can be drawn — one showing the |αk |’s and the other
showing the \αk ’s.
Notice that the Fourier synthesis formula is very similar to the formula given in Lab 3 for the correlation between
a sinusoid and a complex exponential. Indeed it has the same interpretation: in computing αk we are computing the
correlation3 between the signal s(t) and a complex exponential with frequency 2πk/T . Thought of another way, this
correlation gives us an indication of how much of a particular complex exponential is contained in the signal s(t).

Partial Series
Notice the infinite limits of summation in the synthesis formula (4.3). This tells us that, for the general case, we need
an infinite number of complex exponentials to represent our signal. However, in practical situations, such as in this
lab assignment, when we use the synthesis formula to determine signal values, we can generally only include a finite
number of terms in the sum. For example, if we use only the first N positive and negative frequencies plus the DC
term (at k = 0), our approximate synthesis equation becomes

X
N
2πk
s(t) ≈ αk ej T t . (4.6)
k=−N
1 This is the exponential form of the Fourier series synthesis formula. There is also a sinusoidal form, which is presented later in this section.
2πk
2 Because s(t)e−j T t is periodic with period T , this integral evaluates to the same value for any interval of length T .
3 Actually, here we are computing what we called the length-normalized correlation.

The University of Michigan, All rights reserved 67

Laboratory 4. Fourier Series and the DFT

Fortunately, Fourier series theory shows that this approximation becomes better and better4 as N −→ ∞. Alterna-
tively, it is known that the mean-squared value of the difference between s(t) and the approximation tends to zero as
N −→ ∞. Specifically, it can be shown that
!
XN
2πk
XN
M S s(t) − αk ej T t = M S(s(t)) − |αk |2
k=−N k=−N
−→ 0 as N −→ ∞ . (4.7)
How large must N be for the approximation to be good? There is no simple answer. However, you will gain some
idea by the experiments you perform in this lab assignment.

T -Second Fourier Series

If a signal s(t) is periodic with period T , then it is also periodic with period 2T , and period 3T , and so on. Thus when
applying Fourier series, we have a choice as to the value of T . Often, we will choose T to be the smallest period,
i.e. the fundamental period of s(t). However, there are also situations where we will not. For example, suppose we
wish to perform spectral analysis/synthesis of two or more periodic signals that have different fundamental periods.
We could of course form a separate Fourier series for each signal. In this case, each Fourier series will be based on
a different harmonic series of frequencies. Wouldn’t it be nicer if we could base each series on a common harmonic
series of frequencies? We can do this by choosing T to be a multiple of the fundamental periods of both signals.
When we want to explicitly specify the value of T that is used in a Fourier series, we will say T -second Fourier
series. What then is the relationship between Fourier series corresponding to different values of T ? To see what is
happening, let us compare a T -second Fourier series to a 2T -second Fourier series. The T -second Fourier series has
components at the frequencies
. . . , −2ω0 , −ω0 , 0, ω0 , 2ω0 , . . . , (4.8)
where
2π
ω0 = (4.9)
T
and the 2T -second Fourier series has components at the frequencies.
ω0 ω0
. . . , −2ω00 , −ω00 , 0, ω00 , 2ω00 , . . . = . . . , −ω0 , − , 0, , ω0 , . . . , (4.10)
2 2
where
2π ω0
ω00 =
= . (4.11)
2T 2
From this we see that the 2T -second Fourier series decomposes s(t) into frequency components with half the sepa-
ration of that of the T -second Fourier series. However, since s(t) is periodic with period T , its spectrum is actually
concentrated at frequencies that are multiples of ω0 (or a subset thereof). Hence, the “additional” coefficients in the
2T -Fourier series must be zero, and it turns out that the nonzero coefficients are the same as for the T -second Fourier
series. Specifically, it can be shown that with αk and α0k denoting the T -second and 2T -second Fourier coefficients,
respectively, then 
αk/2 , k even
α0k = (4.12)
0, k odd
In summary, Fourier series analysis/synthesis can be performed over one fundamental period or over any number
of fundamental periods. Usually, when Fourier series is mentioned, the desired number of periods interval will be
clear from context. In any case, the spectrum is not affected by the choice of T .
4 It is known that under rather benign assumptions about the signal s(t), the approximation converges to s(t) as N −→ ∞ at all times t where

s(t) is continuous, and at times t where s(t) has a jump discontinuity, the approximation converges to the average of the values immediately to the
left and right of the discontinuity.

68 The University of Michigan, All rights reserved

 4.2.2 Periodic Continuous-Time Signals — The Fourier Series

Aperiodic Continuous-Time Signals

Next, we briefly discuss how Fourier series can also be applied when the signal s(t) is not periodic. In this case, we
can nevertheless determine the spectrum of a finite segment of the signal, say from time t1 to time t2 , by performing
Fourier series analysis/synthesis on just this segment. That is, if we find Fourier coefficients
Z
1 t2
2πk
αk = s(t)e−j T t dt , (4.13)
T t1

where T = t2 − t1 , then we have

∞
X 2πk
s(t) = αk ej T t , for t1 ≤ t ≤ t2 . (4.14)
k=−∞

This will give us an idea of the frequency content of the signal during the given time interval. It is important to
emphasize, however, that the synthesis equation (4.14) is valid only when t is between t1 and t2 . Outside of this time
interval, the synthesis formula will not necessarily equal s(t). Instead, it describes a signal that is periodic with period
T , called the periodic extension of the segment between t1 and t2 .

Properties of the Fourier Coefficients

We conclude our discussion of the Fourier series with a list of useful properties, some of which have already been
mentioned. A few of these will be useful in this lab assignment. The rest are included for completeness. These prop-
erties are stated without derivations. However, each can be derived straightforwardly from the analysis and synthesis
formulas. Though not required in this laboratory, you may want to confirm some of these properties using the Fourier
analysis and synthesis programs described in Section 4.3.

1. (Fourier series analysis) The T -second Fourier series analysis of a periodic signal s(t) with period T produces
a set of Fourier coefficients αk , k = . . . , −2, −1, 0, 0, 1, 2, . . ., which are, in general, complex valued.

2. (Frequency components) If αk are the coefficients of the T -second Fourier series of the periodic signal s(t) with
j 2πk
period T , then the frequency or spectral component of s(t) at frequency 2πk
T is αk e
T t.

3. (DC component) The coefficient α0 equals the average or DC value of s(t).

4. (One-to-one relationship) There is a one-to-one relationship between periodic signals and Fourier coefficients.
Specifically, if s(t) and s0 (t) are distinct5 periodic signals, each periodic with period T , then their T -second
Fourier coefficients are not entirely identical, i.e. αk 6= α0k for at least one k. It follows that one can recognize
a periodic signal from its Fourier coefficients (and its period).

5. (Conjugate symmetry) If s(t) is a real-valued signal, i.e. its imaginary part is zero, then for any integer k

α−k = α∗k (4.15)

|α−k | = |αk | (4.16)
\α−k = −\αk . (4.17)
5 By “distinct”, we mean that s(t) and s0 (t) are sufficiently different that s(t) 6= s0 (t) for all times t in some interval with (t , t ), with nonzero
1 2
length. They are not “distinct” if they differ only at a set of isolated points. To see why we need this clarification, observe that if s(t) and s0 (t)
differ only at time t1 , then they have the same Fourier coefficients, because integrals, such as those defining Fourier coefficients, are not affected by
changes to their integrands at isolated points. Likewise, s(t) and s0 (t) will have the same Fourier coefficients if they differ only at isolated times
t1 , t2 , . . .. However, if s(t) 6= s0 (t) for all t in an entire interval, no matter how small, then αk 6= α0k for at least one k.

The University of Michigan, All rights reserved 69

Laboratory 4. Fourier Series and the DFT

6. (Conjugate pairs) If αk ’s are the T -second Fourier coefficients for a real-valued signal s(t), then for any k the
sum of the complex exponential components of s(t) corresponding to αk and α−k is a sinusoid at frequency
2πk/T . Specifically, using the inverse Euler relation,
2πk
t + \αk ) .
2πk 2πk
αk ej T t + α−k e−j T t = 2|αk | cos( (4.18)
T
7. (Sinusoidal form of the Fourier synthesis formula) The previous property leads to the sinusoidal form of the
Fourier synthesis formula:
∞
X 2πk
s(t) = α0 + 2|αk | cos( t + \αk ) . (4.19)
T
k=−∞

8. (Linear combinations) If s(t) and s0 (t) have T -second Fourier coefficients αk and α0k , respectively, then as(t) +
bs0 (t) has T -second Fourier coefficients aαk + bα0k .
9. (Fourier series of elementary signals) The following lists the T -second Fourier coefficients of some elementary
signals.
2πm
(a) Complex exponential signal: s(t) = ej T t =⇒

1, k = m
αk = . (4.20)
0, k =
6 m

(b) Cosine: s(t) = cos( 2πm

T t) =⇒  1
2, k = ±m
αk = . (4.21)
0, k 6= ±m
(c) Sine: s(t) = sin( 2πm
T t) =⇒  j
 −2, k = m
αk = j
, k = −m . (4.22)
 2
0, k 6= ±m
(d) General sinusoid: s(t) = cos( 2πm
T t + φ) =⇒
 1 jφ
 2e , k=m
1 −jφ
αk = e , k = −m . (4.23)
 2
0, k 6= ±m
10. (T -second Fourier series) If a periodic signal s(t) has period T and T -second Fourier coefficients αk , then the
nT -second Fourier coefficients are

αk/n , k = multiple of n
α0k = (4.24)
0, else

11. (Parseval’s relation) If αk ’s are the T -second Fourier coefficients for signal s(t), then the mean-squared value
of s(t), equivalently the power, equals the sum of the squared magnitudes of the Fourier coefficients. That is,
Z X∞
1
M S(s) = |s(t)|2 dt = |αk |2 (4.25)
T hT i
k=−∞

12. (Uncorrelatedness/orthogonality of complex exponentials) The T -second correlation between complex expo-
2πm 2πn
nential signals ej T t and ej T t , m 6= n, is zero. This property is used in the derivation of the previous and
other properties.

70 The University of Michigan, All rights reserved

 4.2.3 Periodic Discrete-Time Signals — The Discrete Fourier Transform

4.2.3 Periodic Discrete-Time Signals — The Discrete Fourier Transform

This section overview the discrete Fourier transform approach to the frequency-domain representation of discrete-time
signals.
Consider a periodic discrete-time signal s[n] with period N . As with continuous-time signals, we wish to find
its frequency-domain representation, i.e. its spectrum. That is, we wish to represent s[n] as a sum of discrete-time
complex exponential signals. Again, by analogy to the continuous-time case we will use frequencies that are multiples
of
2π
b0 =
ω . (4.26)
N
However, unlike the continuous-time case, we now use only a finite number of such frequencies. Specifically, we use
the N harmonically related frequencies:

b0 , 2b
0, ω ω0 , . . . , (N − 1)b
ω0 . (4.27)

The reason is that any complex exponential signal with the frequency kb ω0 is in fact identical to a complex exponential
signal with one of the N frequencies listed above6. Notice that this set of frequencies ranges from 0 to 2π(N N
−1)
, which
is just a little less than 2π.
We now assert that the representation of s[n] in terms of complex exponentials with the above frequencies is given
by the discrete-time Fourier series synthesis formula or as we will usually call it, the the Discrete Fourier Transform
(DFT) synthesis formula
2π0 2π1 2π2 2π(N −1)
s[n] = S[0]ej N n + S[1]ej N n + S[2]ej N n + . . . + S[N − 1]ej N n

X
N −1
2πk
= S[k]ej N n , (4.28)
k=0

where the S[k]’s, which are called DFT coefficients, are determined by the DFT analysis formula
1 X 2πk
S[k] = s[n]e−j N n , k = 0, 1, 2, 3, . . . , N − 1 (4.29)
N
hN i

where hN i indicates a sum over any N consecutive integers7 , e.g. the sum over 0, . . . , N .
As with the continuous-time Fourier series, the DFT coefficients are, in general, complex. Thus, they have a
magnitude |S[k]| and a phase or angle \S[k]. The magnitude |S[k]| can be viewed as the strength of the exponential
component at frequency kb ω0 = 2πk/N , while \S[k] is the phase of that component. The coefficient S[0] is the DC
term; it measures the average value of the signal over one period.
Once we know the S[k]’s, the spectrum of s[n] is simply a plot consisting of spectral lines at frequencies

b 0 , 2b
0, ω ω0 , . . . , (N − 1)b
ω0 .

The spectral line at frequency kbω0 is drawn with height indicating the magnitude |S[k]| and is labeled with the complex
value of S[k]. Alternatively, two separate spectral line plots can be drawn — one showing the |S[k]|’s and the other
showing the \S[k]’s.
Since the sums in the synthesis and analysis formulas are finite, there are no convergence-of-partial-sum issues,
such as those that arise for the continuous-time Fourier series.
6 If ω0 is not in this range, then k = mN + l where m 6= 0 and 0 ≤ l < N . It then follows that the complex exponential with this frequency
kb
2πk n 2π(mN +l) 2πl n 2πl n
is ej N = ej N
n
= ej2πmn ej N = ej N , which is an exponential with one of the N frequencies in the list above.
7 Because −j 2πk n
s[n]e N is periodic with period N , the sum is the same for any choice of N consecutive integers.

The University of Michigan, All rights reserved 71

Laboratory 4. Fourier Series and the DFT

Often the DFT coefficients S[0], . . . , S[N ] are said to be the “DFT of the signal s[n]” and the process of com-
puting them via the analysis equation (4.29) is called “taking the DFT” of s[n]. Conversely, applying the synthesis
equation (4.28) is often called “taking the inverse DFT” of S[0], . . . , S[N ].
Notice that the DFT analysis formula (4.29) is identical to equation (3.45) in Lab 3. That is, in computing the set
of correlations between a signal s[n] and the various complex exponentials in Lab 3, we were actually taking the DFT
of s[n]. Indeed, it continues to be helpful to view the DFT analysis as the process of correlating s[n] with various
complex exponentials. Those correlations that lead to larger magnitude coefficients indicate frequencies where the
signal has larger components.
In some treatments, the DFT analysis and synthesis formulas differ slightly from those√given above in that the 1/N
factor is moved from the analysis formula to the synthesis formula8, or replaced by a 1/ N factor multiplying each
formula. All of these approaches are equally valid. The choice between them is largely a matter of taste. For example,
our approach is the only one for which S[0] equals the average signal value. For the other approaches, the average is
S[0] multiplied by a known constant. The only cautionary note is that one should never use the analysis formula from
one version with the synthesis formula from another. In this course, we will always use the analysis and synthesis
formulas shown above.
Although we will always take 0, ω b0 , 2bω0 , . . . , (N − 1)b
ω0 as the analysis frequencies produced by the DFT, it is
important to point out that every frequency ω b in the upper half of this range, i.e. between π and 2π, is equivalent
to a frequency ωb − 2π, which lies between −π and 0. By “equivalent,” we mean that a complex exponential with
frequency ωb with π < ω b < 2π equals the complex exponential with frequency ω b − 2π. Thus, it is often useful to think
of frequencies in the upper half of our designated range as representing frequencies in the range −π to 0.
For example, let us look at the DFT of a sinusoidal signal, s[n] = cos( 2πm N n), with 0 < m < 2 . The DFT
N

coefficients, S[k], are given by

(S[0], . . . , S[N − 1]) = (0, . . . , 0, 1/2, 0, . . . , 0, 1/2, 0, . . . , 0), (4.30)

where S[m] = S[N − m] = 1/2 and S[k] = 0 for other k’s. In the synthesis formula, the coefficient S[m] multiplies
2πm 2π(N −m)
the complex exponential ej N n , and the coefficient S[N − m] multiplies the complex exponential ej N n =
2πm
e−j N n . Thus, these two coefficients can be viewed as multiplying exponentials at frequencies ± 2πm
N , which by the
inverse Euler formula sum to yield s[n] = cos( 2πm
N n).

N -point DFT
As with continuous-time signals, if a discrete-time signal s[n] is periodic with period N , then it also periodic with
period 2N , and period 3N , and so on. Thus, when applying the DFT, we have a choice as to the value of N . Sometimes
we choose it to be the the smallest period, i.e. the fundamental period, but sometimes we do not. When we want to
explicitly specify the value of N used in a DFT, we will say N -point DFT.
The relationship between the N -point and 2N -point DFT is just like the relationship between the T -second and
2T -second Fourier series. That is, whereas the N -point DFT has components at frequencies

b0 , 2b
0, ω ω0 , . . . , (N − 1)b
ω0 , (4.31)

the 2N -point DFT has components at the frequencies

b0
ω b0
ω
ω00 , . . . , (2N − 1)b
b 00 , 2b
0, ω ω00 = 0, b0 , . . . , (2N − 1)
,ω .. (4.32)
2 2
where
2π ω0
b0 =
ω = (4.33)
2N 2
8 The DSP First textbook does this in Chapter 9.

72 The University of Michigan, All rights reserved

 4.2.3 Periodic Discrete-Time Signals — The Discrete Fourier Transform

From this we see that the separation between frequency components has been halved. Moreover, it can be shown that
the relationship between the original and new coefficients is

0 S[k/2], k even
S [k] = (4.34)
0, k odd

In summary, DFT analysis/synthesis can be performed over one fundamental period or over any number of fun-
damental periods. Usually, when the DFT is mentioned, the desired number of periods interval will be clear from
context. In any case, the spectrum is not affected by the choice of N .

Aperiodic Discrete-Time Signals

Next, we briefly discuss how the DFT can also be applied when the signal s[n] is not periodic. In this case, we can
nevertheless determine the spectrum of a finite segment of the signal, say from time n1 to time n2 , by performing DFT
analysis/synthesis on just this segment. That is, if we find DFT coefficients
1 X 2πk
S 0 [k] = s[n]e−j N n , k = 0, 1, 2, 3, . . . , N − 1 (4.35)
N
hN i

where N = n2 − n1 , then we have

X
N −1
2πk
s[n] = S 0 [k]ej N n , k = 0, 1, 2, 3, . . . , N − 1 . (4.36)
k=0

This will give us an idea of the frequency content of the signal during the given time interval. It is important to
emphasize, however, that the synthesis equation (4.36) is valid only at times n from n1 to n2 . Outside of this time
interval, the synthesis formula will not necessarily equal s[n]. Instead, it describes a signal that is periodic with period
N , called the periodic extension of the segment from n1 to n2 .

Approximating Fourier series coefficients with the DFT

Frequently, we are interested in finding the spectrum of some continuous-time signal s(t), but for practical reasons,
we sample the signal and work with the resulting discrete-time signal s[n]]. Can we find, at least approximately,
the spectrum of s(t) by working with the discrete-time signal s[n]? As discussed below there is a close relationship
between the Fourier series coefficients of s(t) and the DFT of s[n].
Suppose s(t) is periodic with period T , and suppose we sample s(t) with sampling interval Ts = T /N , where N
is an integer, resulting in the discrete-time signal s[n] = s(nTs ), which is easily seen to be periodic with period N .
Let αk denote the T -second Fourier coefficients of s(t), and let S[k] denote the N -point DFT of s[n]. Then it can be
shown that if N is very large, then
αk ≈ S[k] , when k << N (4.37)
Moreover, it can be shown that if it should happen that s(t) has no spectral components at frequencies greater than
1/(2Ts ), then 
 S[k], 0 ≤ k ≤ N/2
αk = S[N − k + 1], −N/2 ≤ k < 0 (4.38)

0, |k| > N/2
The above two equation shows how the DFT can be used to compute, at least approximately, the Fourier series
coefficients. In fact, the Fourier series analysis program described in the in the MATLAB section of this assignment
uses the DFT to compute the Fourier coefficients.

The University of Michigan, All rights reserved 73

Laboratory 4. Fourier Series and the DFT

Properties of the DFT coefficients

The following are a number of useful properties of the DFT with which you should be familiar. A few of these
will be useful in this lab assignment. Others will be used in future assignments. These properties are stated without
derivations. However, each can be derived straightforwardly from the analysis and synthesis formulas. Though not
required in this laboratory, you may want to confirm some of these properties using the DFT analysis and synthesis
programs described in Section 4.3.

1. (DFT analysis) The N -point DFT of a periodic signal s[n] with period N produces a vector of N DFT co-
efficients S[0], . . . , S[N − 1], which are, in general, complex valued. Equivalently, the coefficients may be
considered to be determined by a set of N signal samples.

2. (Frequency components) If S[k] is N -point DFT of the periodic signal s[n] with period N , then the frequency
j 2πk
or spectral component of s[n] at frequency 2πk
N is S[k]e
N n . The component of the signal at frequency
−2πk
N
−j 2πk
is S[N − k]e N . N

3. (DC component) The coefficient S[0] equals the average value or DC value of s[n].

4. (One-to-one relationship) There is a one-to-one relationship between discrete-time signals with period N (equiv-
alently, sequences of N signal samples) and sequences of N DFT coefficients. Specifically, if s[n] and s0 [n]
are distinct periodic signals with period N , i.e. s[n] 6= s0 [n] for some value of n, then their N -point DFT
coefficients are not entirely identical, i.e. S[k] 6= S 0 [k] for at least one k. It follows that one can recognize a
discrete-time periodic signal from its DFT coefficients (and N ).

5. (Conjugate symmetry) If s[n] is a real-valued signal, i.e. its imaginary part is zero, then for any integer k

S[N − k] = S ∗ [k] (4.39)

|S[N − k]| = |S[k]| (4.40)
\S[N − k] = −\S[k] . (4.41)

These facts indicate that we are usually only interested in the first half of the DFT coefficients. In particular,
note that when we plot the DFT, the location of the origin and the appearance of the symmetry is different than
when we plot the Fourier Series. See Figure 4.2 for an example of the relation between the two.

6. (Conjugate pairs) If S[k] is the N -point DFT of a real-valued signal s[n], then for any k the sum of the com-
plex exponential components of s[n] corresponding to S[k] and S[N − k] is a sinusoid at frequency 2πk/N .
Specifically, using the inverse Euler relation,

2πk
n + \S[k]) .
2πk 2πk
S[k]ej N n + S[N − k]e−j N n = 2|S[k]| cos( (4.42)
N

7. (Linear combinations) If s[n] and s0 [n] have N -point DFT S[k] and S 0 [k], respectively, then as[n] + bs0 [n] has
N -point DFT aS[k] + bS 0 [k].

8. (Sampled continuous-time signals) If the discrete-time signal s[n] comes from sampling a continuous-time
signal s(t) with sampling interval Ts , i.e. if s[n] = s(nTs ), then the continuous-time frequency represented by
DFT coefficient S[k] is 2πk
N fs , where fs = 1/Ts samples per second is the sampling rate.

9. (DFT of elementary signals) The following lists the N -point DFT of some elementary signals.

74 The University of Michigan, All rights reserved

 4.2.3 Periodic Discrete-Time Signals — The Discrete Fourier Transform

(A) |Ck| X[k] (B)

−5ω −4ω −3ω −2ω −ω ω 2ω 3ω 4ω 5ω 1 2 3 4 5 N−5 N−4 N−3 N−2 N−1

0 0 0 0 0 0 0 0 0 0

ω −−> k −−>

Figure 4.2: (A) The magnitude of the Fourier Series coefficients αk for a periodic continuous-time signal. (B) The
DFT of a periodic discrete-time version of the same signal. Note that the origin for the Fourier Series coefficients is in
the middle of the plot, but the origin for the DFT is to the left.

2πm
(a) Complex exponential signal: s[n] = ej N n =⇒
(S[0], . . . , S[N − 1]) = (0, . . . , 0, 1, 0, . . . , 0) , (4.43)
where the nonzero coefficient is S[m].


(b) Cosine: s[n] = cos 2πm
N n =⇒
1 1
(S[0], . . . , S[N − 1]) = (0, . . . , 0, , 0, . . . , 0, , 0, . . . , 0) , (4.44)
2 2
where the nonzero coefficients are S[m] and S[N − m].


(c) Sine: s[n] = sin 2πm
N n =⇒
j j
(S[0], . . . , S[N − 1]) = (0, . . . , 0, − , 0, . . . , 0, , 0, . . . , 0) , (4.45)
2 2
where the nonzero coefficients are S[m] and S[N − m].


(d) General sinusoid: s[n] = cos 2πmN n+φ =⇒
1 1
(S[0], . . . , S[N − 1]) = (0, . . . , 0, ejφ , 0, . . . , 0, e−jφ , 0, . . . , 0) , (4.46)
2 2
where the nonzero coefficients are S[m] and S[N − m].
 
(e) Not quite periodic sinusoid: s[n] = cos 2π(m+)
N n where (m + ) is non-integer =⇒ The resulting
9
S[k]’s will all be nonzero , typically with small magnitudes except those corresponding to frequencies
closest to 2π(m+)
N .
(f) Period contains unit impulse period: s[n] = (1, 0, . . . , 0) =⇒
 
1 1
(S[0], . . . , S[N − 1]) = ,..., . (4.47)
N N
10. (N -point DFT) If S[k] is the N -point DFT of the periodic signal s[n] with period N , then the mN -point DFT
coefficients are 
S[k/m], k = multiple of m
S[k] = (4.48)
0, else
9 This is the same effect that you saw in lab 3 when you ran fape over a non-integer number of periods of the sinusoid.

The University of Michigan, All rights reserved 75

Laboratory 4. Fourier Series and the DFT

11. (Parseval’s relation) If S[k] is the N -point DFT of s[n], then

−1
1 X X
N
2
M S(x) = |s[n]| = |S[k]|2 . (4.49)
N
hN i k=0

This shows that the power in the signal s[n] equals the energy of the DFT coefficients.

12. (Uncorrelatedness/orthogonality of complex exponentials) The N -point correlation between complex exponen-
2πm 2πl
tial signals ej N n and ej N n , m 6= l, is zero. This property is used in the derivation of the previous one.

4.2.4 Separating Signals Based on Differing Harmonic Series

We’ve already suggested that there are many nearly-periodic signals that occur in the real world, with two notable
examples being many musical signals and vowels in speech signals. These sort of signals can be analyzed using the
Fourier Series or the DFT (applied to samples). We will use the DFT, principally because if we wanted to use the
Fourier series, we would anyway approximately compute the Fourier coefficients with the DFT. In particular, let us
consider a note played on a musical instrument like a flute or clarinet. Such a signal is nearly periodic with some
fundamental period. If the note is played at “concert pitch,” for instance, it has a fundamental frequency of 440 Hz and
a fundamental period of 1/440 seconds. Few musical signals, though, are purely sinusoidal. From our development
of the Fourier series, we know that a periodic signal can be described as a sum of complex exponentials (or sinusoids)
with harmonically-related frequencies. That is, the spectrum of our musical note is composed of a harmonic series. In
particular, if the fundamental frequency is 440 Hz, higher harmonics will be at 880 Hz, 1320 Hz, 1760 Hz, and so on.
Figure 4.3 shows a stemp plot of the DFT of an example harmonic series.
Suppose that we have two instruments playing different notes (i.e., the two signals have different fundamental
periods) at the same time. The signal coming from each instrument is a single harmonic series, but a listener “hears”
a signal which is the sum of these two signals. By the linear combination properties of the Fourier Series and DFT,
we know that the spectrum of the combined signal is simply the sum of the spectra of the separate signals. We can use
this property to separate the two signals in the frequency-domain, even though they overlap in the time-domain.
Suppose that we wish to simply remove one of the notes from the combined signal. We’ll assume that we have
recorded and sampled the signal, so we’re working in discrete-time. We’ll also assume that the combined signal is
also periodic10 with some (fairly long) fundamental period N0 . If we take the N0 -point DFT of a segment of the
combined signal, we can identify the coefficients that make up each harmonic series. Then, we simply zero-out all of
10 In the “real-world,” this is a somewhat questionable assumption. However, we can approximate this behavior quite well by simply using a long

DFT. In this case, each harmonic may be “spread” over several DFT coefficients, so to remove a harmonic we need to zero-out all of coefficients
associated with it. This spreading behavior is the same as what you saw in Lab 3 when running fape over non-periodic signals.

0.8

0.6
|X[k]|

0.4

0.2

0
0 10 20 30 40 50 60 70
Coefficient Number, k

Figure 4.3: The DFT of a harmonic series. Note that only the first half of the DFT coefficients are shown in this figure.

76 The University of Michigan, All rights reserved

 4.3 Some M ATLAB commands for this lab

the coefficients corresponding to the harmonics of the note we wish to remove. When we resynthesize the signal with
the inverse DFT, the resulting signal will contain only one of the two notes.
We can extend this procedure to more complicated signals, like melodies with many notes. In this case, we simply
analyze and resynthesize each note individually. Of course, with more simultaneously-sounding notes and more
complicated music, this procedure becomes rather difficult. In this lab, we will implement this procedure to remove a
“corrupting” note held throughout a simple, easily analyzed melody. Though somewhat idealized, the problem should
help to motivate the use of the DFT and the frequency domain.

4.3 Some M ATLAB commands for this lab

• Fourier Series Synthesis in M ATLAB: The function fourier_synthesis is a function that we provide to
compute the approximate T -second Fourier series synthesis formula, equation (4.6). Its inputs are the period T
and a set of 2N + 1 Fourier coefficients. Its output is the synthesized signal. The calling command is

>> [ss,tt] = fourier_synthesis(CC, T, periods, Ns);

where CC is a vector containing the Fourier coefficients, T is the interval (in seconds) over which the Fourier
series is applied. periods is the (integer) number of periods to include in the resynthesis; periods defaults
to a value of 1 if not provided. The optional parameter Ns specifies how many samples per period to include in
the output signal.
It is assumed that CC contains the coefficients α−N . . . αN . (N is implicitly determined from the length of CC.)
Thus, CC has length 2N + 1, the CC(n) element contains the Fourier series coefficient αn−N −1 . Further, note
that the α0 coefficient falls at CC(N+1).
The two returned parameters are the signal vector ss and the corresponding signal support vector tt.

• Fourier Series Analysis in M ATLAB: The function fourier_analysis is the complement to the function
fourier_synthesis. It performs T -second Fourier series analysis on an input signal. The calling command
is

>> [CC,ww] = fourier_analysis(ss,T,N);

where ss is a vector containing the signal samples, T is the interval T in seconds over which the Fourier series is
to be computed, and N is the number of positive harmonics to include in the analysis. (2N+1 is the total number
of harmonics.) It is assumed that ss contains samples of the signal to be analyzed over the interval [0, T ].
The outputs are the vectors CC, which contains the 2N + 1 Fourier coefficients11 , and ww, which contains the
frequencies (in Hertz) associated with each Fourier coefficient.

• DFT Analysis in M ATLAB: In order to calculate an N -point DFT using M ATLAB, we use the fft command12.
The specific calling command is

>> XX = fft(xx)/length(xx);
11 Because fourier analysis is given only samples of the desired continuous-time signal, it cannot compute the Fourier coefficients exactly.

Rather it computes an approximation by using the DFT.

12 FFT stands for the Fast Fourier Transform, which is a fast implementation of the DFT. Calculating the DFT from its definition requires

O(N 2 ) computations, but the FFT only requires O(N log N ). Additionally, the FFT is faster when N is equal to a power of two (i.e., N =
256, 512, 1024, 2048, etc.).

The University of Michigan, All rights reserved 77

Laboratory 4. Fourier Series and the DFT

This computes the N -point DFT of the signal vector xx , where N is the length of xx, and where the signal is
assumed to have support 0, 1, . . . , N − 1. Since the M ATLAB command fft does not include the factor 1/N in
the analysis formula, as in equation (4.29), we must divide by length(xx) to obtain the N DFT coefficients
XX.
• DFT Synthesis in M ATLAB: The synthesis equation for the DFT is computed with the command ifft. If we
have computed the DFT using the above command, we must also remember to multiply the result by N :

>> xx = ifft(XX)*length(XX);

Note that the ifft command will generally return complex values even when the synthesis should exactly
be real. However, the imaginary part should be negligible (i.e., less than 1 × 10−14 ). You can eliminate this
imaginary part using the real command.
• Indexing the DFT: Since M ATLAB begins its indexing from 1 rather than 0, remember to use the following
rules for indexing the DFT:

X[0] ⇒ X(1)
X[1] ⇒ X(2)
X[k] ⇒ X(k+1)
X[N − k] ⇒ X(N-k+1)
X[N − 1] ⇒ X(N)

4.4 Demonstrations in the Lab Section

• Approximating signals as sums of sinusoids, as in Problem 1.
• “Mapping out” this week’s background section
• Relating the Fourier Series to the DFT
• T -second Fourier Series and the N -point DFT
• The DFT in M ATLAB

4.5 Laboratory Assignment

1. (Building signals from sinusoids) In this problem, you will “hand tune” the amplitudes and phases of three
sinusoids so that their sum matches a “target” periodic signal as well as possible. The signals are considered to
be continuous-time. One could do this task analytically or numerically using the Fourier series analysis formula,
but we want you to gain the insight that results from doing it manually. A graphical MATLAB program has been
written to facilitate this procedure.
Download the files sinsum.m and sinsum.fig and execute sinsum13 . M ATLAB will bring up a GUI
window with three sinusoids (colored, dotted lines), the sum of these three sinusoids (the black, dashed line),
and a target periodic signal (the black, solid line). The frequencies of the sinusoids are ω0 , 2ω0 , and 3ω0 , where
ω0 is the fundamental frequency of the target signal.
13 Note that this function will only work under M ATLAB 6 and higher. It is highly recommend that you use a Windows-based PC for this problem,

since you need to copy the figure window into your report. Using the Windows clipboard simplifies this task significantly.

78 The University of Michigan, All rights reserved

 4.5 Laboratory Assignment

As stated earlier, the goal of this problem is to adjust the amplitudes and phases of the three sinusoids to
approximate the target signal as closely as possible. You can enter the amplitude and phase for each sinusoid
in the spaces provide in the GUI window, or using the mouse, you can click-and-drag each sinusoid to change
its amplitude and phase. In addition to displaying the three sinusoids, their sum, and the target signal, the GUI
window also shows the mean-squared error between the sum and the target signals.
Use sinsum.m to hand tune the amplitudes and phases of the three sinusoids to make the mean-squared error
as small as you can.
(Hint: You should be able achieve an MSE less than 0.24. You will receive +2 bonus points if you can achieve
an MSE less than 0.231.)
(Hint: In attempting to minimize the MSE you might try to adjust one sinusoid to minimize the MSE, then
another, then another. After doing all three, go back and see if readjusting them in a “second round” has any
benefits.)

• [16(+2)] Include the resulting figure window in your report. (On Windows systems, use the “Copy to
Clipboard” button to copy the figure, then you can simply paste it into a Word or similar document. There
is also a “Print Figure” button for other systems if you can’t get access to a PC.)

Food for thought14: Did you try the procedure suggested in the hint above, in which you tune each sinusoid one
at a time and then return to each for a “second round” of tuning? If so, can you explain why the second round
did or did not lead to any improvements? (Hint: Consider Fourier series property 12.)
Food for thought: By executing sinsum(1), sinsum(2), and sinsum(3), you can match different signals
with sinusoids. Find MSE’s that are as small as possible for each of these other signals.

2. (Applying Fourier series synthesis)

In this problem you will simply apply fourier_synthesis to a given set of Fourier coefficients and find
the resulting continuous-time signal. Download the file fourier_synthesis.m. Use it to generate an
approximation to the signal with the following Fourier coefficients:


2 2
− πk k = ±1, ±3, ±5, . . .
αk = (4.50)
0 k = 0, ±2, ±4, . . .

Let T = 0.1 seconds, and generate 5 periods of the signal. Use N = 20, giving you 41 Fourier series coef-
ficients. (Hint: First, define a frequency support vector, kk=-20:20. Then, generate CC from kk and set all
even harmonics to zero.)

• [4] Use stem to plot the magnitude of the Fourier coefficients. Use your kk vector as the x-axis.
• [3] Use plot to plot samples of the continuous-time signal that fourier_synthesis returns versus
time in seconds.
• [2] What kind of signal is this?

3. (Applying Fourier series analysis) In this problem you will use the Fourier series analysis and synthesis formula
to see how the accuracy of the approximate synthesis formula (4.6) depends on N .
Download the files lab4_data.mat and fourier_analysis.m. lab4_data.mat contains the vari-
ables step_signal and step_time, which are the signal and support vectors for the samples of a periodic
continuous-time signal with fundamental period T0 = 1 second. Note that there are Ns = 16384 samples in
14 “Food for thought” items are not required to be read or acted upon. There is no extra credit for involved. However, if you include something in

your report, your GSI will read and comment on it. Alternatively, you can discuss “food for thought topics” in office hours.

The University of Michigan, All rights reserved 79

Laboratory 4. Fourier Series and the DFT

one fundamental period. (step_signal and step_time include several fundamental periods, but you’ll
be dealing with only one period in several parts of this problem. As such, you might find it useful to create a
one-period version of step_signal.)

(a) (Look at the signal to be analyzed) First, let us examine step_signal.

• [3] Use plot to plot step_signal versus its support vector.
• [3] Compute the mean-squared value of step_signal.
(b) (Perform FS analysis) Use fourier_analysis to perform a T0 second Fourier series analysis over a
single period of step_signal with with N = 50.
• [4] Use subplot and stem to plot the magnitude and phase of the resulting Fourier series coeffi-
cients. Make sure that your x-axis is given in frequency.
(c) (Resynthesize FS approximations) Use fourier_analysis and fourier_synthesis to generate
an approximations of step_signal with N = 25, 50, 100, and 200. (Perform T0 -second Fourier
analysis and synthesis over a single period of the signal for each N . Be sure to resynthesize a single period
with Ns = 16384 samples.)
• [4] Use plot and subplot to plot your resynthesized signals for each N in separate panels of a
subplot array.
• [3] Calculate the mean-squared error of the resynthesis for each value of N .
• [3] Compute the sum of the squared magnitudes of CC for each value of N .
• [3] Find and document a relationship between the mean-squared errors and the sum of squared mag-
nitudes of CC you have computed. (Hint: Consider the mean-squared value that you computed for
step_signal. You might also want to look in the Properties of Fourier Coefficients subsection.)
(d) (Meet an MSE target) Find the smallest value of N for which the mean-squared error of the resynthesis is
less than 0.5% of the mean-squared value of step_signal.
• [4] Include this value in your report.

Food for thought: Try repeating Part (b) with the Fourier analysis performed over two fundamental periods of
the signal, and compare to the previous answer to Part (b). Do the new Fourier coefficients turn out as expected?

4. (Using the DFT to describe a signal as a sum of discrete-time sinusoids) In this problem, you will simply
apply the DFT to a particular discrete-time signal, which is also contained in lab4_data.mat, namely,
signal_id. signal_id is considered to be a periodic discrete-time signal with fundamental period N0 =
128 = length(signal_id). Take the N0 -point DFT of signal_id.

• [3] Use stem to plot the magnitude of the DFT versus the DFT coefficient index, k.
• [12] Use the DFT to describe signal_id as a sum of discrete-time sinusoids. That is, for each sinusoid,
give the amplitude, frequency (in radians per sample), and phase.

5. (Use the DFT to remove undesired components from a signal) In this problem you will use the technique
described in Section 4.2.4 to eliminate a noise signal from a desired signal. This signal, melody, is also
contained in lab4_data.mat. This variable contains samples of a continuous-time signal sampled at rate
fs = 8192 samples/second. It contains a simple melody with one note every 1/2 second. Unfortunately, this
melody is corrupted by another “instrument” playing a constant note throughout. We would like to remove this
second instrument from the signal, and we will use the DFT to do so.
It is a good idea to begin by listening to melody using the soundsc command.

80 The University of Michigan, All rights reserved

 4.5 Laboratory Assignment

(a) (Examine DFT of first note) In order to remove the corrupting instrument, we need to determine where it
lies in the frequency domain. Let’s begin by looking at just the first note (i.e the first 0.5 seconds or 4096
samples). This “note” consists of the sum of two notes — one is the first note of the melody, the other is
the constant note from the corrupting instrument. Each of these notes has components forming a harmonic
series. The fundamental frequencies of these harmonic series are different, which is the key to our being
able to remove the corrupting note. Take the DFT of the first 0.5 seconds (4096 samples) of the signal.
• [3] Use stem to plot the magnitude of the DFT for the first note.
• [3] Identify the frequencies contained in each of the two harmonic series present in signal. What are
the fundamental frequencies?
(b) (Examine DFT of second note) By comparing the spectra of the first two notes, we can identify the cor-
rupting instrument. Take the DFT of the second 0.5 seconds (samples 4097 through 8192).
• [3] Use stem to plot the magnitude of the DFT for the second 0.5 seconds.
• [2] What are the fundamental frequencies (in Hz) of the two harmonic series in this note?
• [2] We know that the melody changes from the first note to the second, but the corrupting instrument
does not. Thus, by comparing the harmonic series found in this and the previous part, identify which
fundamental frequency belongs to the melody and which to the corrupting instrument.
(c) (Identify the DFT coefficients of the corrupting signal) In order to remove the “corrupting” instrument, we
simply need to zero-out the coefficients corresponding to the harmonics of the note from the corrupting
instrument. This is done directly on the DFT coefficients of each 0.5 seconds of the signal. Then, we
resynthesize the signal from the modified DFT coefficients.
• [4] Based on this, and your results from the previous parts of this problem, which DFT coefficients
need to be set to zero in order to remove the corrupting instrument from this signal? (Hint: Remember
the conjugate pairs.)
(d) (Complete the function that removes the corrupting instrument) Finally, we’d like to remove the corrupting
instrument from our melody. Download the file fix_melody.m. This function contains the code that
you’ll use to remove the corrupting instrument from the melody signal. For each note of the melody, the
function takes the DFT, zeros out the appropriate coefficients (which you must provide), and resynthesizes
the signal.
• [4] Complete the function by setting the variable zc equal to a vector containing the DFT coefficients
that must be zeroed-out.
• [1] Execute the function using the command
>> result = fix_melody(melody);
Listen to the resulting signal. Have you successfully removed the corrupting instrument?
(e) (Check your result with the spectrogram) Finally, we’d like to be able to visually check our result. Down-
load the function melody_check.m. melody_check produces an image called a spectrogram that
you can use to check your work. Basically, the spectrogram works by taking the DFT of many short seg-
ments of a signal and arranging them as the columns of an image. Note that the x-axis is time and the
y-axis is frequency. The color of each point on the image represents the strength of the spectral compo-
nent (in decibels) at that time and frequency. The dark horizontal bands show the presence of sinusoidal
components in the signal at the associated times.
• [4] Execute melody_check by passing it melody. Include the resulting figure in your report.
• [1] Can you identify the components of the corrupting instrument on this spectrogram?
• [4] Now, execute melody_check by passing it result. Include the resulting figure in your report.

The University of Michigan, All rights reserved 81

Laboratory 4. Fourier Series and the DFT

• [2] Compare the spectrogram of melody to the spectrogram of result. What differences do you
see? Is this what you expect to see?

6. On the front page of your report, please provide an estimate of the average amount of time spent outside of lab
by each member of the group.

82 The University of Michigan, All rights reserved

Laboratory 5

Images, Compression, and Coding

5.1 Introduction

A common application of signals and systems is in the production, manipulation, storage and distribution of images.
For example, image transmission is an important aspect of communication (especially on the internet), and we would
like to be able to distribute high quality images quickly over low-bandwidth connections. To do so, images must be
encoded into a sequence or file of bits, which can be digitally transmitted or stored. When display of the image is
required, the sequence/file of bits must be decoded into a reproduction of the image. A block diagram of a general
data compression system, with an encoder and decoder, is shown in Figure 5.1.
Systems or algorithms that do the encoding and decoding are called source coders, coders, data compressors, or
compressors. They are called source coders because they encode the data from a source, e.g. a camera or scanner.
They are also called data compressors, because their encoders usually produce fewer bits than were produced by the
original data collector. For example, JPEG is a commonly used, standardized image compressor. You’ve probably
downloaded many JPEG encoded images over the internet — any image with filename extension .jpg. FAX machines
use a different image compression algorithm.
In this lab, we will experiment with some basic data compression techniques as applied to images. Typically, there
is a tradeoff between the number of bits an encoder produces and the quality of the decoded reproduction. With more
bits we can usually obtain better quality at the expense of greater storage or bandwidth requirements. When we assess
how well these techniques work, we will count the number of bits their encoders produce(fewer is better), and as a
measure of quality, and we will compute the mean-squared or RMS error as a measure of the quality of the decoded
reproduction (low error means high quality, or equivalently, low distortion).

bits
Signal Encoder/ Decoder/ Reconstructed
Compressor Decompressor Signal

Data Compression System

Figure 5.1: A block diagram of a general data compression system.

The University of Michigan, All rights reserved 83

Laboratory 5. Images, Compression, and Coding

5.1.1 “The Question”

• How can we compress images so that they take up less storage space and/or less bandwidth?

5.2 Background
5.2.1 Images
So far, we have dealt entirely with one-dimensional signals. That is, these signals are indexed by only one independent
variable (usually time). In this lab, we will start to consider two-dimensional signals. An image is an example of a
two-dimensional signal. In an image, we usually index the signal based on horizontal and vertical position — two
dimensions that are needed to find the “signal value” at any given point.
In this lab, we will generally restrict our attention to gray-scale images1 . We mathematically represent such an
image (in continuous-space) as a signal x(t, s), where 0 ≤ t ≤ H, 0 ≤ s ≤ W . x(t, s) denotes the intensity,
brightness, or value of the image at the position with vertical coordinate t and horizontal coordinate s, and H and W
are the height and width of the image, respectively. The values of x(t, s) are generally nonnegative. Thus, a small
value of x(t, s) (close to zero) corresponds to black while larger values correspond to progressively lighter shades of
gray.
In digital image processing, the image is assumed to be sampled at regularly spaced intervals creating a discrete-
space image x[m, n]:
x[m, n] = x(mTs , nTs ), (5.1)
where Ts is the sampling interval, given in units of distance. Thus, in discrete-space, an image is simply an M × N
array or matrix of numbers x[m, n], where m and n are integers in the range [1, M ] and [1, N ], respectively. Each
x[m, n] is called a pixel. We adopt the usual convention that x[1, 1] is the upper left pixel, x[1, N ] is the upper right,
x[M, 1] is the lower left, and x[M, N ] is the lower right.
We shall also adopt the common, but not universal, convention of digital image processing that pixel values, often
called levels, are integers ranging from 0 to 255. The reason the pixel values are integers is that computers cannot store
real-valued quantities. Instead the raw pixel values must be quantized to values from a finite set. The usual practice
is to scale the raw image pixel values by some constant so the maximum value is close to 255 and then to round each
pixel value to the nearest integer, thereby obtaining an image whose values are integers between 0 and 255. Why 0
to 255? There are two reasons. One is that these values can be conveniently represented with one byte, i.e. 8 bits.2
Another reason is that the effects of rounding to 256 possible levels are not ordinarily observable, whereas rounding
to a significantly smaller number, say 128, is sometimes noticeable.

5.2.2 Signal statistics for images

Two-dimensional signals in general, and images in particular, have the same sorts of statistics that one-dimensional
images have. Generalizing from the one-dimensional case is often quite straightforward. We will also introduce two
new statistics for both one- and two-dimensional signals.

1. Average Value. The mean or average value, M , of a discrete-space image x[m, n] is

1 XX
M N
M (x) = x[m, n] (5.2)
N M m=1 n=1
1 Color
images are often represented as three separate signals (or channels), one each for red, green, and blue.
an integer in a computer, it must be represented with a binary sequence. If binary sequences of length b are used, then 2b levels can be
2 To store

represented, because there are 2b distinct binary sequences of length b. Thus, it takes 8 bits to store the 256 levels from 0 to 255.

84 The University of Michigan, All rights reserved

 5.2.2 Signal statistics for images

2. Mean-squared value. The mean-squared value (or MSV), M S, of a discrete-space image x[m, n] is

1 XX 2
M N
M S(x) = x [m, n] (5.3)
N M m=1 n=1

3. Root mean-squared value. The root mean-squared value (or RMS value) of a discrete-space image x[m, n] is
v
u
u 1 X M X N
RM S(x) = t x2 [m, n] (5.4)
N M m=1 n=1
p
= M S(x) (5.5)

4. Variance. The variance of a discrete-space image x[m, n] is

1 XX
M N
V ar(x) = (x[m, n] − M (x))2 (5.6)
N M m=1 n=1
= M S(x − M (x)) (5.7)
= M S(x) − (M (x))2 (5.8)

5. Standard deviation. The standard deviation of a discrete-space image x[m, n] is

v
u
u 1 X M XN
Std(x) = t (x[m, n] − M (x))2 (5.9)
N M m=1 n=1
p
= V ar(x) (5.10)
= RM S(x − M (x)) (5.11)
p
= M S(x) − (M (x))2 (5.12)

Notice the relationship between the variance and standard deviation, equation (5.10), and the relationship between
these statistics and the MS and RMS values3 , equations (5.8) and (5.12). The variance and standard deviation measure
how widely varying are the values of a signal. If they are small, it means that the signal values (and thus the signal
value distribution) is tightly clustered around the mean value, while if they are large, the signal values range widely.
Recall from Laboratory 1 that we often use the MS and RMS values to measure distortion of a signal. We will
be doing this for images in this laboratory. If y[m, n] is a distorted version of x[m, n], then we can measure the
mean-squared error (MSE) and root mean-squared error (RMSE), using

1 XX
N M
M SE = (x[m, n] − y[m, n])2 (5.13)
N M n=1 m=1
√
RM SE = MSE. (5.14)
3 Equation (5.8) is not something that is immediately obvious, but it is something that can be straightforwardly derived. (Doing so is an interesting

exercise.)

The University of Michigan, All rights reserved 85

Laboratory 5. Images, Compression, and Coding

x[m,n] Transform Quantizer Binary bits

Encoder

Image Encoder

Figure 5.2: A block diagram of a general image encoder/compressor.

5.2.3 Data compression

There are two primary types of data compressors: lossless and lossy. A lossless compressor will encode and decode
in such a way that the decoded reproduction is exactly the same as the original (8 bits per pixel) image. A lossy
compressor will encode and decode in such a way that the decoder produces only an approximation to the original
image. On the one hand, lossless is better because it is, well, lossless. This is essential when compressing computer
files. UNIX compress, gzip, and the PKZip compression formats are all lossless.
On the other hand, if a small amount of distortion is permitted, lossy compressors can attain much larger amounts
of compression, i.e. their encoders can produce many fewer bits. For multimedia, lossy compression is often accept-
able. Examples of lossy compression schemes that you may have used include MP3 (for audio), JPEG (for photos),
and MPEG (for movies). These three schemes are all examples of transform coding methods; we will examine a
simple transform coding scheme in this laboratory. MP3 encoding is also an example of so-called perceptual cod-
ing. Perceptual coding methods often introduce significant amounts of distortion, but do so in a way that is nearly
imperceptible.
Consider the generalized image encoder/compressor shown in Figure 5.2. It consists of three main components:
the transform, the quantizer and the binary encoder. These are described briefly now, and in more detail in the next
three subsections. The input to the encoder is a sampled image, x[m, n]. We generally consider that the pixels of
x[m, n] take on a continuum of real values.
The first component, the transform, is optional. When it is included, it is usually a spatial-domain to frequency-
domain transformation like the Discrete Fourier Transform (DFT). Applying the transform to short segments or blocks
of the signal tends to concentrate the energy of the signal into just a few coefficients, which permits the quantizer and
binary encoder to be more effective.
The next component is quantization. Quantization takes the input sample/pixel and “rounds” it to one of a finite set
of levels. It is a lossy or noninvertible operation in the sense that one cannot recover the original sample/pixel from the
quantized sample/pixel. As an example, we noted previously that digital images often have pixel values ranging from
0 to 255. This is because each raw pixel value, as produced by some camera, has already been rounded to the nearest
of a set of 256 quantization levels. Lossless image compression schemes work by operating on image that are already
quantized; additional quantization is not permitted. However, in lossy compression schemes, additional quantization
is performed, in order to obtain greater compression. For example, each image pixel may be quantized to the nearest
of a set of only 64 levels.
The final component is binary coding, which assigns a sequence of bits called a codeword, to each level produced
by the quantizer. In some systems, called fixed-length coders, the number of bits used to represent each pixel is known
in advance (e.g. 2 bits per pixel). This is the simplest type of coder. Other systems, called variable-length coders,
assign binary codewords of different lengths to each pixel value, usually based on the frequency of occurrence. More
frequently occuring levels are assigned shorter codewords. This allows the compression system to achieve additional
compression. Many advanced compression systems, including JPEG, use variable length coders. MP3 coding has
provisions for both fixed-length and variable-length coding.
As illustrated in Figure 5.3, the decoder/decompressor corresponding to the encoder/compressor just described
has two components. The input to the decoder is the bits produced by the encoder. The first component, the binary

86 The University of Michigan, All rights reserved

 5.2.4 Transformation

bits Binary Inverse ^x[m,n]

Encoder Transform

Image Decoder

Figure 5.3: A block diagram of a general image decoder/decompressor.

decoder, inverts the operation of the binary encoder, and produces the levels originally produced by the quantizer. The
quantization operation produced by the encoder is not invertible, so there is not corresponding decoding step. Instead
the last step is the inverse transform, which as the name suggests performs the inverse of the encoding transform. The
output of the decoder, called an encoded or decoded image or reproduction can be displayed on a monitor or printed
on paper, as desired.

5.2.4 Transformation
Efficient lossy data compressors typically perform some sort of preprocessing on the data to be compressed. One very
common preprocessing step is a transform, and such compressors are called transform coders. For example, JPEG
is a transform coder based on the discrete cosine transform (DCT), which is a spectral transformation similar to the
DFT. The transform is typically applied to small groups of pixels called blocks. In this lab, we will experiment with a
simple DFT-based transform coder that uses short 1 × 8 pixel blocks. That is, we use an N -point DFT with N = 8.
Recall that the synthesis and analysis formulas for an 8-point DFT are given by
7
X 2πk
x[n] = X[k]ej 8 n (5.15)
k=0
7
1X 2πk
X[k] = x[n]e−j 8 n (5.16)
8 n=0
2πk
Here, X[k]ej 8 n is the “spatial” frequency component at frequency ω̂ = 2πk
8 . The DFT synthesis formula shows
that an image block x[n] can be viewed as the sum of such components. More specifically,
(x[0], x[1], x[2], . . . , x[7]) = X[0] (1, 1, 1, . . . , 1)
2π 2π 2π 2π
+ X[1] ( ej N 0 , ej N 1 , ej N 2 , . . . , ej N 7 )
2π2 2π2 2π2 2π2
+ X[2] ( ej N 0 , ej N 1 , ej N 2 , . . . , ej N 7 )
+ ...
2π2 2π2 2π2 2π2
+ X[7] ( ej N 0 , ej N 1 , ej N 2 , . . . , ej N 7 ) (5.17)
Note that we are NOT presuming that these blocks are in any way periodic.
Why do we perform a spectral transformation before compressing a signal? As suggested in Section 5.2.3, the
transformation serves to shift around the energy in a given block so that it is easier to compress. Consider, for instance,
a single block of an image given by
x[n] = ( 165, 168, 167, 166, 167, 165, 168, 166 )
This block is roughly constant, so we expect its 8-point DFT to have a large X[0] (i.e., DC) component. Since there
is little other variation, though, the rest of the DFT coefficients will be relatively small. By only storing the X[0]

The University of Michigan, All rights reserved 87

Laboratory 5. Images, Compression, and Coding

th
coefficient, for instance, and throwing away the rest, we only need 1/8 as much storage as if we had stored all of the
coefficients for this block; further, we have introduced only a small amount of distortion (as measured by the mean-
squared error). While we won’t go so far as to throw away the rest, this example suggests the basic idea for how to use
the transform to compress a signal. If some transformed coefficients tend to be smaller than others, we can store them
more efficiently.
An 8-point DFT produces 8 complex numbers, X[0], X[1], . . . , X[7]. This actually translates into 16 real numbers
(the 8 real and 8 imaginary parts) that we need to consider storing. Thus taking the transfrom would at first seem to be
a bad idea, because we now need twice as much storage to represent a block! However, there are symmetry properties
that we can exploit so that we only need to store 8 of these numbers. Since the input signal is real, recall that the
8-point DFT, X[k] has the conjugate symmetry property:

X[8 − k] = X ∗ [k] (5.18)

This means that knowing the real and imaginary parts of X[1], for instance, completely determines the real and imag-
inary parts of X[N − 1]. Thus, though we need to store the real and imaginary parts of X[0], X[1], X[2], X[3], X[4],
we do not need to store X[5], X[6], X[7], because these values can be recovered from the previous four. Further, the
coefficients X[0] and X[4] are purely real (that is, X[0] = X ∗ [0] and X[4] = X[8 − 4] = X ∗ [4]). Thus, X[0] and
X[4] each require the storage of a single real number. From this argument, we can see that the 8-point DFT produces
a total of only eight real numbers that must be stored.
In our implementation of the transform encoder, the eight numbers c[0], . . . , c[7] that we choose to represent the
8-point DFT X[0], . . . , X[7], of an image block are

c[0] = X[0]
√
c[1] = 2 Re {X[1]}
√
c[2] = 2 Re {X[2]}
√
c[3] = 2 Re {X[3]}
c[4] = X[4]
√
c[5] = 2 Im {X[1]}
√
c[6] = 2 Im {X[2]}
√
c[7] = 2 Im {X[3]}
√
The 2 factors have been included where one coefficient is, in effect, standing in for two. It can be shown that with
these factors
X7 X7
x2 [1, n] = 8 c2 [k], (5.19)
n=0 k=0

which is an often useful fact. This is derived using Parseval’s relation, as given in Lab 44 .

5.2.5 Quantization
Quantization is the most elementary form of lossy data compression, while also forming a fundamental part of more
advanced lossy compression schemes such as transform coding. We may quantize an image directly, or we may
quantize the results of a transformation as described in Section 5.2.4. When a number x is quantized to L levels, we
4 For this derivation, see the document titled “Notes: The Distortion of Transform Coding” by D.L. Neuhoff.

88 The University of Michigan, All rights reserved

 5.2.6 Binary coding

mean that its value is replaced by (or quantized to) the nearest member of a set of L quantization levels. Here, we
consider uniform quantization5. For the uniform quantization used here:

• We define a quantizer range defined by values xmin and xmax

xmax −xmin
• We divide this range into L equally sized segments, each with size ∆ = L .

• We place the quantization level for a given segment in the middle of that segment.

The quantizer is illustrated with the figure shown below, which shows L = 8 segments of width ∆ = (xmax −xmin )/8
as thick lines and the corresponding levels within each segment as circles.

level 1level 2level 3level 4level 5level 6level 7level 8

xmin xmax

Given a pixel x[m, n], the quantizer operates by outputing the nearest level. Equivalently, if x[m, n] lies in the ith
segment, then quantizer outputs the ith level. If x is larger than xmax , then x is quantized to the largest level, namely,
xmax − ∆/2. Similarly, if x is smaller than xmin , then x is quantized to the smallest level, namely, xmin + ∆/2.
One can see that if x is within the quantizer range, then its quantized value will differ from x by at most ∆/2,
so that the quantizer introduces only a small error. On the other hand, when x is outside the range, the quantizer
can introduce a large error. Thus, when designing a quantizer it is important to choose the quantizer range so that
it includes most values of x. Making the range large will do this. However, we don’t want to make the range too
large. Larger ranges mean that ∆ = (xmax − xmin )/L is larger, which in turn increases the maximum possible error
introduced when x lies within the range of the quantizer.

5.2.6 Binary coding

The output of a data compression encoder must always be bits, not quantized samples or pixels. Thus, the quantizer
is always followed by a binary encoder, as illustrated in Figure 5.4. A compressor that consists simply of a quantizer
followed by a binary encoder will be called a direct quantizer, in contrast to a transform coder or some other coder
that involves a preprocessing step.
A binary encoder operates by assigning a distinct sequence of bits, called a codeword to each level of the quantizer.
For example, an assignment of codewords to levels is shown below

000 001 010 011 100 101 110 111

level 1level 2level 3level 4level 5level 6level 7level 8

xmin xmax

5 There are sometimes advantages to using quantizers with unequal level spacings, but we will not deal with such quantizers in this lab. Uniform

quantizers are sometimes called uniform scalar quantizers to distinguish them from more sophisticated quantizers that do not operate independently
on successive data samples.

The University of Michigan, All rights reserved 89

Laboratory 5. Images, Compression, and Coding

^
x[n,m]
Binary bits Binary
x[n,m] Quantizer ^
x[n,m]
Encoder Decoder

Encoder Decoder

Figure 5.4: Block diagram of a direct quantizer

Such binary codewords are the output of the encoder when quantizing the data. The decoder will, eventually,
receive a binary codeword and output the corresponding quantization level as the reproduction of the original piece of
data. For instance, if the image pixel x[m, n] lies in the third segment of the quantizer, the binary encoder will produce
010, which when received, the decoder will produce level 3 as the reproduction of x[m, n].
If, as often happens, the number of levels is a power of two, i.e. M = 2b where b is an integer, then the simplest
approach is to make each codeword have b bits. It does not matter which b-bit sequence is assigned to which level, but
the usual scheme, as illustrated above, is to assign the binary sequence representing 0 to the smallest level, the binary
sequence representing 1 to the next largest level, and so on. With this type of binary coding, the encoder is fixed-length
(or fixed-rate) in the sense described earlier. Often, a better scheme is to use shorter codewords for the quantization
levels that occur more frequently, and longer ones for those that are used less frequently. Such variable-length codes
are used in JPEG and other high efficiency schemes.

5.2.7 Performance
There are two ways that we measure the performance of a compression system. First, we want to know how many
bits are required to store an image. The total number of bits produced by the encoder is equal to the number of blocks
multiplied by the number of bits required to encode one block. More commonly, we report the number of bits required
to store a single pixel. This is called the coding rate, R. The coding rate is equal to the number of bits required to code
a single block divided by the number of pixels in a block. Naturally, we prefer a lower coding rate.
The second performance measure is the amount of distortion introduced by the coder. Generally, we measure this
distortion by computing the mean-squared (MSE) or RMS error (RMSE). We also prefer to have low distortion, and
equivalently low error.
Unfortunately, we generally have to trade off between these two performance measures. That is, we can produce
a highly compressed (with a low coding rate) image, but this generally introduces a large RMS error. Alternatively,
we can have a very high-quality representation of an image (with low distortion), but such a representation requires
many bits to encode. Figure 5.5 shows an illustration of the tradeoff between the two performance measures. In the
laboratory assignment, you will produce a plot similar to this for compression using uniform quantization.

Performance of direct quantizers

Let us now analyze the performance of a direct quantizer, where the quantizer is uniform with L = 2b levels and range
[xmin , xmax ].
Since the binary encoder for such a system assigns b bits to each level, the coding rate is

R = b bits/pixel (bpp) (5.20)

Elementary theory predicts that when the quantizer range includes most values of the image x[m, n] and when ∆
is much smaller than the standard deviation of the image, then the MSE induced by quantizing with level spacing ∆

90 The University of Michigan, All rights reserved

 5.2.7 Performance

can be approximated as follows6 /

1 2
M SE ≈ ∆ (5.21)
12
 2
1 xmax − xmin
= (5.22)
12 L
1
= (xmax − xmin )2 2−2R , (5.23)
12

This shows that if we were to shrink ∆ by a factor of 2, as would happen if L were doubled and the range were held
constant, then the MSE would decrease by a factor of four. Equivalently, the last equation shows that this factor of
four reduction comes by increasing the coding rate by one bit per pixel.
When a quantizer is applied to data whose signal value distribution is fairly constant over a given range, then it is
usually good practice to choose the quantizer range to match the data range. This is generally the case when directly
quantizing images, so we will generally choose xmin = 0 and xmax = 255.
On the other hand, when quantizing data whose signal value distribution is quite uneven, then it may be best to
choose the quantizer range to be a subset of the data range. For example, in transform coding, it often happens most
of the data to be quantized is near zero but there are a few very, very large values. In such cases, experience has shown
that to design a quantizer with small MSE, one should normally choose the width of the p range to be proportional to
the standard deviation of the data being quantized, i.e. (xmax − xmin ) = c × Std(x) = c V ar(x). The constant of
proportionality c is usually between 2 and 6. Smaller values of c work well for smaller values of L, and larger values
of c work well for large values of L. Using this relation in (5.21), we find
 2
1 2 1 xmax − xmin
MSE ≈ ∆ = (5.24)
12 12 L
2
c V ar(x)
≈ . (5.25)
12 L2
c2
≈ V ar(x)2−2R . (5.26)
12

This shows that quantizer MSE is proportional to the variance of the data and inversely proportional to L2 .

6 See the document “Note: The ∆2 /12 Formula” by D.L. Neuhoff.

RMS Error

Coding Rate (bits/pixel)

Figure 5.5: There is an inherent tradeoff between coding rate and distortion.

The University of Michigan, All rights reserved 91

Laboratory 5. Images, Compression, and Coding

c[0] ^
c[0] bits ^c[0]
Quantizer 1 Binary Encoder Binary Decoder
c[1] ^
c[1] ^c[1]
DFT Quantizer 2 Binary Encoder Binary Decoder IDFT
x[m,n] Block Block ^x[m,n]

...

...
c[7] ^
c[7] ^c[7]
Quantizer 8 Binary Encoder Binary Decoder

Encoder Decoder

Figure 5.6: A block diagram of a transform coder. The encoder divides the incoming image, x[m, n] into 1 × 8 blocks
and transforms each block into a sequence of 8 coefficients c[0], . . . , c[7]. These coefficients are then quantized to yield
c[0], . . . , b
b c[7], and encoded into a binary representation. The decoder creates a reconstruction of the image, x b[m, n] by
decoding the binary codewords and inverting the transformation.

5.2.8 Designing a transform coder

In the previous sections, we have described the three main components of transform type data compression system.
In particular, in Section 5.2.4, we described a transform that uses an 8-point DFT of 1 × 8 blocks of image pixels. A
block diagram of the system based on this transform can be found in Figure 5.6.
To make this transform coder work well, though, the quantizers must be individually designed for each of the
eight types of (independent) coefficients. Indeed, if we quantize all eight types of coefficients with the same number
of levels, then the transform coder will not work substantially better than direct quantization (quantization without
preprocessing). Thus, for each of the eight types of coefficients, we must carefully choose the number of quantization
levels, L, and the quantizer range limits, xmin and xmax .
Let Lk be the number of levels for each coefficient c[k]. Further, let each Lk be a power of two such that Lk = 2bk ,
where bk is the number of bits that we allocate to the transformed coefficient c[k].
It should be clear that choosing large Lk ’s will permit the transform coder to encode with less distortion. However,
2
the total number of bits produced by the encoder is the number of blocks, N8 , times the number of bits to encode
P7
one block, k=0 bk . Thus, higher Lk ’s require more bits to store the signal, and thus a higher coding rate. For this
transform coder, we can calculate the coding rate, R, as
7
1X
R= bk bpp . (5.27)
8
k=0

In many situations, we are given a desired coding rate R, e.g. R = 2 bpp. In this case, the question becomes how
we should divide these bits among the eight types of coefficients, i.e. how to choose the bk ’s, so they average to the
desired coding rate R, yet cause the distortion in the reproduction produced by the transform code to be as small as
possible.
Using (5.19), it can be shown that7 .
X7
MSE = MSE[k] , (5.28)
k=0

where MSE[k] is the MSE of the quantizer for c[k]. In other words, the MSE of the transform coder is approximately
the sum of the MSE’s of the quantizers for the different coefficients.
7 See the document titled “Notes: The Distortion of Transform Coding” by D.L. Neuhoff.

92 The University of Michigan, All rights reserved

 5.3 Some M ATLAB commands for this lab

Let us first consider a transform coder where each type of coefficient is quantized with the same number of
bits/pixel, i.e. b0 = b1 = . . . = b7 . We assert without proof that such a transform coder has roughly the same
MSE as that of direct quantization with the same number of bits/pixel. Now, we will now argue that changing the
b[k]’s so that some are larger than others will make the transform coder work better than direct quantization.
From (5.26) we have that
1 2
MSE[k] ≈ c V ar(c[k])2−2bk , (5.29)
12
where V ar(c[k]) denotes the variance of the c[k] values. One can see from the above that the coefficients with larger
variance will be quantized with larger mean-squared error. In particular the DC coefficients C[0] usually have the
largest variance; so they will have the largest MSE. On the other hand, the c[3]’s and c[7]’s usually have the smallest
variance and distortion.
Now suppose we increase b0 by one and decrease b7 by one. From (5.27) we see that this will have no net effect
on the number of bits produced by the coder. However, from (5.29) we see that this decreases the (large) MSE of
the DC coefficients c[0] by a factor of 4, and increases the (small) MSE of the c[7] coefficients by a factor of 4. Is
it beneficial to decrease one MSE by 4, when another one increases by 4? We can see from (5.28) that indeed it is
beneficial. Decreasing a larger MSE by the factor 4 decreases the average in (5.28) more than increasing a small MSE
by the factor of 4 increases the average.8 Thus, what we want to do is shift bits towards the coefficients with larger
variances. This will make MSE smaller than if all coefficients were quantized with the same number of bits and,
therefore, smaller than the distortion of direct quantization.
More generally, in a well designed transform code, all of the MSE[k]’s will be approximately the same. If they
were quite different, we could move a bit from a coefficient with small MSE to one with large MSE and achieve a
net decrease in overall MSE. In this light, we can see that the role of the transform is to make the variances of the
coefficients as different as possible. Some should be large, and others should be small.

5.3 Some M ATLAB commands for this lab

• Making a matrix into a vector. Especially when working with images, it is often useful to be able to convert a
matrix into a vector containing the same elements. In M ATLAB, we can do this for a matrix x in the following
manner:

>> y = x(:);

After this operation, y contains a “vectorized” version of x. Specifically, if x is an M × N matrix, y is a vector

whose first M elements are the first column of x, whose second M elements are the second column of x, and
so on. This is especially useful for calculating many of the signal statistics presented in this laboratory.

• Calculating signal statistics on images. To compute these statistics on images, we first need “vectorize” the
image.

1. Average value, M (x):

>> M = mean(x(:))
2. Mean-squared value, M S(x):
>> MS = mean(x(:).ˆ 2)
3. Root-mean squared value, RM S(x):
8 For example, 24 + 4 is larger than 24/4 + 4 × 4.

The University of Michigan, All rights reserved 93

Laboratory 5. Images, Compression, and Coding

>> RMS = sqrt(mean(x(:).ˆ 2))

4. Variance, V ar(x).
>> variance = var(x(:));
5. Standard deviation, Std(x).
>> std_dev = std(x(:));
6. Signal value distribution. To compute a histogram with 256 bins centered at integers from 0 to 255, use
the command
>> hist(x(:),0:255);
• Loading images. Images are generally stored in some sort of standard file format, like TIFF or JPEG. To load
such an image file into M ATLAB, we use the command imread. Unfortunately, imread generally returns
images as arrays of integers. This is unfortunate because M ATLAB puts some heavy restrictions on the use of
integers. In particular, to prevent integer overflow you cannot perform arithmetic on integers. Thus, we need to
convert our loaded images into double precision arrays using the command double. To load and convert an
image in the file my_img.tif, for instance, use the command

>> x = double(imread('my_img.tif'));

Note that the imread command will load many standard image file formats, including JPEG, PNG, BMP, TIFF,
PCX, and a host of others.
• Displaying images. To display an image in M ATLAB, there are actually a number of commands that must be
used simultaneously. To display the image itself, we use imagesc command. To tell M ATLAB to display the
image as a gray-scale image, we use command colormap(gray). To set the axes so that the aspect ratio is
correct, use the command axis image. Finally, to add a “color bar” that relates image values to colors, use
the colorbar command. Every image that you produce for this course must have a color bar; you will lose
points for every image you display without a color bar. To do all of these things at once to display an image x,
use the following code:

>> imagesc(your_img); colormap(gray); axis image; colorbar

You will be using this sequence of commands often, so you might wish to write a short function that executes
all of these commands simultaneously.
• Quantizing an image: The function quantize_fcn.m, which we provide to you for this lab, implements a
uniform quantizer for images and transform coefficients. It takes a signal, the desired number of quantization
levels (L), and the two numbers that define the quantization range, xmin and xmax . For instance, to quantize an
image, img, to 64 levels, use the command

>> [q_img, delta] = quantize_fcn(img,64,0,255);

q_img contains the quantized image, while delta contains the ∆ value used for quantization. Here, note that
xmin = 0 and xmax = 255. This separately quantizes each pixel of img to one of 64 levels, in accordance with
the procedure described in the background section.
• Using the DFT Coder: The DFT-based transform coder that we have described in this laboratory is provided
as three separate functions.

94 The University of Michigan, All rights reserved

 5.4 Demonstrations in the Lab Section

– dft_block.m breaks the image into 1 × 8 blocks and computes the DFT of each block. If the image
is N × N , this function produces a series of eight band images. For k = 1, . . . , 8, the k th band image
contains the c[k − 1] coefficients for each block. For example, the k = 1 band image, contains the c[0], or
DC, coefficients from each block. Each band image has size M × N/8.
The eight band images are returned as a three-dimensional array. To produce the band images for an image,
img and then access the third band image, for instance, we would use the commands
>> A = dft_block(img);
>> A(:,:,3);
Note that except for the first one, each band image contains both positive and negative values. However,
we can still display them using imagesc.
– inverse_dft_block.m reconstructs the image from the matrix of band images returned by the func-
tion dft_block.m.
– dft_coder.m puts both of these blocks together by calling dft_block, quantizing the coefficient
matrix, and reconstructing the image with inverse_dft_block.
dft_coder takes several input parameters, all of which are optional except the first one. The first
parameter is the image to encode. The second is a vector of bit allocations, bk . For instance, if we call
dft_coder like this
>> coded = dft_coder(img,[8 6 6 6 6 4 4 4]);
we quantize our c[0] (DC) coefficients using 8 bits, the next four (real) coefficients with 6 bits each, and
the last three (imaginary) coefficients using 4 bits each9 .
Note that the number of bits required to encode a single pixel is equal to the average value of all of the
bk ’s. Thus, the example above uses 5.5 bits per pixel.
When run, dft_coder returns the decoded image and also displays a table of useful statistics corre-
sponding to each coefficient c[k]. To see this table, make sure that you put a semicolon at the end of your
call to dft_coder.

5.4 Demonstrations in the Lab Section

1. Images are signals too.
2. Signal compression
3. The “Almost JPEG” DFT Coder
4. Designing the coder

5.5 Laboratory assignment

1. (Images in M ATLAB) In this problem, you’ll familiarize yourself with the image capabilities of M ATLAB along
with one particular image, the “cameraman.”

(a) (Display an image) Load the image “cameraman.tif”. (If your computer does not have the Image Process-
ing Toolbox, you’ll need to download the file from the web page).
9 Though more advanced coders may allow the allocation of fractions of bits, for this coder you must allocate a whole number of bits to each

coefficient. You can, however, assign no bits to a coefficient. In this case, that coefficient is simply set to a constant value.

The University of Michigan, All rights reserved 95

Laboratory 5. Images, Compression, and Coding

• [3] Display the image and include the resulting figure in your report.
• [1] Calculate the size of the image (the number of rows and columns) and the total number of pixels
in the image.
• [2] Find the minimum and maximum pixel values, xmin and xmax in the image.
(b) (Produce and interpret a histogram) Estimate the signal value distribution of this image by generating a
histogram with 256 bins centered at integers from 0 to 255.
• [3] Include the resulting plot in your report.
• [2] From this histogram, what signal values occur the most often in this image?
• [2] In words, describe which part(s) of the image corresponds to these signal values.
(c) (Examining signal values) It is useful to be able to think of images in terms of the signal values that make
them up. Download the M-file display_square.m, which will help this process. Use this function to
display the pixel values in several rectangular segments of the “cameraman” image. Find, approximately,
the smallest rectangle of pixels that includes the black tip of the camera lens.
• [3] Include in your report a plot from display_square.m showing the pixel values of the rectangle
you found.
• [2] From this display, what are the row and column indices of this rectangle?
• [2] From this display, what are minimum and maximum values within this rectangle?
(d) (Signal representations) We know that this image takes on only integer values over a finite range, but there
are still a few different ways we can represent the image. In the original file, for instance, each pixel is
represented using 8 bits. In M ATLAB, though, we convert the image into 64-bit double precision values.
• [1] How many bits are required to describe the entire image at 8 bits per pixel?
• [1] How many bits are required to describe the entire image at 64 bits per pixel?
• [1] How many possible pixel values can a 64-bit number represent?

2. (Direct quantization) In this problem, we will experiment with direct quantization as an image compression
mechanism. Download the function quantize_fcn.m.
Assume xmin = 0 and xmax = 255 throughout this problem.

(a) (Quantize an image) Use quantize_fcn to quantize the “cameraman” image using 64 levels, 16 levels,
and 4 levels.
• [4] Display and include in your report the three resulting quantized images along with the original
using subplot. Again, make sure that you indicate which image is which.
• [2] Describe the effects of the quantization in these plots.
(b) (Plot quantization functions)
Use M ATLAB to make a plot of the function being implemented by quantize_fcn.m. For example, for
the 64 level quantizer, run quantize_fcn(u,64,0,255) for an input vector u with elements ranging
from 0 to 255, and plot the resulting quantized values versus u.
• [2] Plot the quantization function for the 16 level quantizer.
• [2] Also, plot the histogram of image quantized with 16 levels, using 256 bins centered at integers
from 0 and 255.
(c) (Quantization as compression) For the 4, 16, and 64 level quantizers,
• [2] How many bits are needed to represent each pixel in one of these images?
• [2] How many bits are needed to represent each of these quantized images?

96 The University of Michigan, All rights reserved

 5.5 Laboratory assignment

(d) (Measuring quantization error) Find the “error image” (the difference between the original image and the
quantized image) corresponding to each of these quantized images.
• [4] Using subplot, display and include in your report the three error images in the same plot.
• [2] Can you see aspects of the original images in these plots?
• [3] Calculate the RMS error for each quantization of the image.
(e) (Evaluating RMS error predictions) Now, we want to compare the actual RMS error for “cameraman”
versus the predicted RMS error (based on the derivation in Section 5.2.7) for quantizers with 2, 4, 8, 16,
32, 64, and 128 levels.
• [4] Calculate the actual RMS error for each of these quantizers.
• [4] Calculate the predicted RMS errors for these quantizers.
• [4] Plot both the actual and predicted RMS error values versus the required number of bits per pixel.
• [2] For what number of bits per pixel is this prediction most accurate?

3. (Compression using a transform coder) In this problem, you will experiment with the DFT-based transform
coder that is described in the background section.

(a) (Create and examine band images) Download the M-file dft_block.m. Use it to generate the matrix of
band images for the “cameraman” image.
• [4] Use subplot to display all eight band images simultaneously. Use axis square rather than
axis image when you display these band images.
• [2] Discuss the appearances of the various band images. For example, can you see any features of the
original cameraman image in any or all of them?
(b) (Reconstruct a coded image) Download the M-file inverse_dft_block.m. Use this function to re-
construct the original image from the set of band images produced by dft_block.
• [2] Compute the RMS error between the original and the transformed/inverse transformed image. (It
should be negligibly small.)
(c) (Designing coders for image compression) Download the M-file dft_coder.m. Our goal in using
dft_coder is to find appropriate parameters for the eight quantizers when compressing the “camerman”
image. Through intelligent design, we hope to achieve lower RMS error than with direct quantization of
the image using the same number of bits. We do this by allocating bits to each of our eight quantizers
independently.
i. (Design a 4 bpp coder) Find a 4 bits per pixel design with as small an RMS error as you can. You
should be able to get an RMS error less than 4. (Hint: As a general rule of thumb from Section 5.2.8,
bigger coefficients should get more bits.)
• [4] What bit allocation did you use, and what was the resulting RMS error?
• [3] Display the reconstructed image and the error image on the same figure using subplot.
• [2] Compare your RMS error to the RMS error of 4 bits per pixel uniform quantization that you
performed in problem 2e.
• [1] Compare the qualitative appearance of the reconstruction produced by the transform coder to
that produced by the direct quantizer.
ii. (Design a 3 bpp coder) Find a 3 bits per pixel design with as small an RMS error as you can. You
should be able to get an RMS error less than 6.4.
• [4] What bit allocation did you use, and what was the resulting RMS error?

The University of Michigan, All rights reserved 97

Laboratory 5. Images, Compression, and Coding

• [3] Display the reconstructed image and the error image on the same figure using subplot.
• [2] Compare your RMS error to the RMS error of the 3 bits per pixel uniform quantization that
you performed in problem 2e.
• [1] Compare the qualitative appearance of the reconstruction produced by the transform coder to
that produced by the direct quantizer. (Note: You have not yet displayed the 3 bpp image, so you
will need to generate it for comparison.)
iii. (Design a 2 bpp coder) Find a 2 bits per pixel design with as small an RMS error as you can. You
should be able to get an RMS error less than 10.8.
• [4] What bit allocation did you use, and what was the resulting RMS error?
• [3] Display the reconstructed image and the error image on the same figure using subplot.
• [2] Compare your RMS error to the RMS error of a 2 bits per pixel uniform quantization that you
performed in problem 2e.
• [1] Compare the qualitative appearance of the reconstruction produced by the transform coder to
that produced by the direct quantizer.
iv. (Comment on coder design) Given your experimentation with this transform coder,
• [2] Comment on the relative performances of direct quantization and transform coding as the
number of bits/pixel changes.

Food for Thought: In this lab, we’ve used a 1-dimensional transform for our coder. We can achieve significantly
better compression if we use a 2-dimensional transform. M ATLAB implements a two-dimensional DFT with the
command fft2. As a challenging project, consider modifying the transform coder provided here to work on 4
× 4 or 8 × 8 blocks of an image. How much compression can you achieve with this modified coder?
4. On the front page of your report, please provide an estimate of the average amount of time spent outside of lab
by each member of the group.

Postscript: JPEG Compression

In this lab, we’ve presented a transform coder here which uses some of the same basic ideas as JPEG compression.
However, JPEG achieves much higher compression rates that what we have seen, and with much less distortion. How
is this achieved? There are several modifications used in JPEG coding.
1. JPEG uses a two-dimensional transform. This allows much greater compaction of the data into a few transform
coefficients.
2. JPEG uses a transform called the discrete cosine transform, which purely real, rather than the DFT. This removes
some of the redundancies in our coding method.
3. JPEG uses a technique called run-length encoding. This allows a coder to store a “run” of similar values by
indicating the value and the number of repetitions.
4. JPEG uses a variable-length coding scheme (often Huffman coding, which you may study in an intermediate
programming course on data structures and algorithms) to produce a bit stream for the final coded representation.
All of these improvements allow images to be significantly compressed with relatively small distortion. For more
information about JPEG coding, you might wish to look at the JPEG Tutorial:
http://www.ece.purdue.edu/˜ace/jpeg-tut/jpegtut1.html

98 The University of Michigan, All rights reserved

Laboratory 6

FIR Filtering and Image Processing

6.1 Introduction
Digital filters are one of the most important tools that signal processors have to modify and improve signals. Part of
their importance comes from their simplicity. In the days when analog signal processing was the norm, almost all
filtering was accomplished with RLC circuits. Now, a great deal of filtering is accomplished digitally with simple (and
extremely fast) routines that can run on special digital signal processing hardware or on general purpose processors.
So why do we filter signals? There are many reasons. One of the biggest is noise reduction (which we have
called signal recovery). If our signal has undesirable frequency components, e.g. it contains noise in a frequency
range where there is little or no desired signal, then we can use filters to reduce the relative amplitude of the signal
at such frequencies. Such filters are often called frequency blocking filters, because they block signal components
at certain frequencies. For example, lowpass filters block high frequency signal components, highpass filters block
low frequency signal components, and bandpass filters block all frequencies except those in some particular range (or
band) of frequencies.
There are a wide range of uses for filtering in image processing. For example, they can be used to improve the
appearance of an image. For instance, if the image has granular noise, we might want to smooth or blur the image to
remove such. Typically such noise has components at all frequencies, whereas the desired image has components at
low and middle frequencies. The smoothing acts as a lowpass filter to reduce the high frequency components, which
come, predominantly, from the noise. Alternatively, we might want to sharpen the image to make its edges stand out
more. This requires a kind of highpass filter.
In this lab, we will experiment with a class of filters called FIR (finite impulse response) filters. FIR filters are
simple to implement and work with. In fact, an FIR filtering operation is almost identical to the operation of running
correlation which you have worked with in Laboratory 2. In particular, we will examine the use of FIR filters for image
processing, including both smoothing and sharpening. We will also examine their use on simple one-dimensional
signals.

6.1.1 “The Question”

• How do we implement FIR filters in M ATLAB?

• How can we improve the appearance of an image? Specifically, how can we remove noise or “sharpen” an
image?

The University of Michigan, All rights reserved 99

Laboratory 6. FIR Filtering and Image Processing

6.2 Background
6.2.1 Implementing FIR Filters
FIR filters are systems that we apply to signals. An FIR filter takes an input signal x[n], modifies it by the application
of a mathematical rule, and produces an output signal y[n]. This rule is generally called a difference equation, and it
tells us how to compute each sample of the output signal y[n] as a weighted sum of samples of the input signal x[n].
A common form of the difference equation is given as
X
M
y[n] = bk x[n − k] (6.1)
k=0
= b0 x[n] + b1 x[n − 1] + b2 x[n − 2] + . . . + bM x[n − M ] (6.2)
The bk ’s are called the FIR filter coefficients, and M is the order of the FIR filter. The set of FIR filter coefficients
completely specifies an FIR filter. Different choices of the order and the coefficients leads to different kinds of filters,
e.g. to lowpass, highpass and bandpass filters.
Equation (6.1) defines the class of causal FIR filters. A more general form is given by
X
M2
y[n] = bk x[n − k] (6.3)
k=−M1
= b−M1 x[n + M1 ] + . . . + b−1 x[n + 1]
+ b0 x[n] + b1 x[n − 1] + . . . + bM2 x[n − M2 ] , (6.4)
where M1 and M2 are nonnegative integers. Here, the order of the filter is M1 + M2 . When M1 > 0, the FIR filter is
non-causal. To calculate the “present” value of y[n0 ], a causal FIR filter only requires “present” (n = n0 ) and “past”
(n < n0 ) values of x[n]. Non-causal filters, on the other hand, require “future” (n > n0 ) values of x[n]. Thus, a
filter with difference equation given by y[n] = x[n] + x[n − 1] is causal, but a filter with difference equation given
by y[n] = x[n] + x[n + 1] is non-causal. The distinction between causal and non-causal filters is necessary if we
wish to implement one of these filters in real-time. Causal filters can be implemented in real-time, but to implement
non-causal filters we generally need all of the data for a signal before we can filter it.
Compare equation (6.3) with the equation for performing running correlation between a signal b[n] and x[n]:
∞
X
y[n] = C(b[k], x[k − n]) = b[k]x[k − n]. (6.5)
k=−∞

Recall that we thought of running correlation as a procedure where we “slid” one signal across the other, calculating
the in-place correlation at each step. If we consider that the bk ’s of an FIR filter form a signal, then the application
of an FIR filter uses the same procedure with two minor differences. First, when we apply an FIR filter, we are only
“correlating” over a finite range; however, we typically assume bk = 0 for k outside the range [M1 , M2 ]. Thus, we can
change the limits of summation to range over (−∞, ∞) without changing the result. Second, when applying a filter,
the signal x[n] is time-reversed with respect to the bk coefficients1 . This is not the case for running correlation.
From the definition alone, it is not easy to see how a filter “works.” With the connection to correlation, though, we
can suggest an intuitive graphical understanding of this process which is shown in Figure 6.1. To calculate a single
sample of y[n], we time-reverse the signal formed by the bk coefficients (by flipping it across the n = 0 axis). Then,
we shift this time-reversed signal by n samples and perform in-place correlation. The result is the nth sample of y[n].
To build up the entire signal y[n], we do this repeatedly, “sliding” one signal across the other and calculating in-place
correlations at each point.
1 That is, x[n − k] is a time-reversed version of x[k − n], just as s[−n] is a time-reversed version of s[n]. Note that we can “time-reverse” the

bk coefficients rather than x[n] and achieve the same result.

100 The University of Michigan, All rights reserved

 6.2.2 Edge effects and delay

bk x[n]

x[n] and
6-nb

y[n] S
=x[k]b
k-n

Figure 6.1: A graphical illustration of filtering. The filter coefficients, bk and signal to be filtered, x[n], are shown on
the top axis. The middle axis shows x[n] and a time-reversed and shifted version of bk . We multiply these two signals
and sum the result to yield a single sample of the output, y[n], which is shown on the bottom axis. For example, to
compute the y[6] sample, we multiply the samples of x[n] by b6−n and sum the result.

You may find it useful to go back to Lab 2 and review the algorithm for in-place correlation. In that description of
the algorithm, we used x[n] where here we wish to use the signal formed by the bk ’s. We can use this algorithm when
implementing FIR filters, as well. Note, however, that we want to time-reverse the bk coefficients when we multiply
them by the incoming signal samples. That is, we always want to multiply the b−M1 coefficient by the newest sample
in the buffer.

6.2.2 Edge effects and delay

Suppose that we consider filtering a signal, x[n], with a causal filter whose difference equation is given by

1 1 1 1 1
y[n] = x[n] + x[n − 1] + x[n − 2] + x[n − 3] + x[n − 4]. (6.6)
5 5 5 5 5
That is, bk = ( 15 , 15 , 15 , 15 , 15 ), M1 = 0, and M2 = 4. We can think of the operation of this filter as replacing each
sample of x[n] with the average of that sample and the past four samples. As such, we often call filters like this
moving-average filters. The result of this filtering for a particular signal is shown in Figure 6.2.
In this particular case, x[n] has a support interval of [0, 28] and is zero outside of this range. Consider what happens
to the output signal, y[n], near the edges of this range. First, the output sample at y[0] will be dominated by zeros from
outside of the support interval, because

1 1 1 1 1
y[0] = x[n] + 0 + 0 + 0 + 0 . (6.7)
5 5 5 5 5
Similarly, y[1], y[2], y[3], and y[4] will also be affected by these zeros, but to a lesser extent. This effect can be seen
in Figure 6.2 as y[n] “ramps up” to the nominal values of x[n]. This effect is known as a start-up transient. A similar
effect occurs beyond y[28], where the signal takes a few samples to “die off”. This effect is known as an ending
transient. Both of these transients are known as edge effects, and need to be considered when filtering.

The University of Michigan, All rights reserved 101

Laboratory 6. FIR Filtering and Image Processing

Input x[n]
6

0
0 5 10 15 20 25 30
Sample Number

6
Output y[n]

0
0 5 10 15 20 25 30
Sample Number

Figure 6.2: Input and output of a 5-point moving average filter.

What do the edge effects do to the support length of the output signal? Well, from Figure 6.2 we can see that y[n]
has a support length four samples longer than that of x[n]. In general, the length of the output signal which is non-zero
will be equal to the length of the input signal plus the order of the FIR filter.
There is one additional point that should be examined. Look at the location of the “dip” in Figure 6.2. In x[n], the
“dip” occurs at sample 18, but in y[n] it occurs at sample 20. In fact, the entire support interval of y[n] has not only
gotten larger, it has also been shifted over to the right (or delayed) by two samples. Why is this? The delay introduced
by this filter results from the fact that each output sample is an average of samples to its left. If we instead define
this filter so that it considers two samples to both the right and left, we can eliminate this delay. That is, we define a
different difference equation:

1 1 1 1 1
y[n] = x[n + 2] + x[n + 1] + x[n] + x[n − 1] + x[n − 2]. (6.8)
5 5 5 5 5
This modification, though, has taken a causal filter and made it non-causal.
Delay is a common problem for causal filters. In fact, the only causal filter that does not introduce delay is a
zero-order amplifier system with a difference equation y[n] = b0 x[n]. This system changes the amplitude of a signal,
but does nothing else. Compare this to the system with difference equation y[n] = x[n − N ]. This system’s only effect
is to delay the signal by N samples. In some circumstances, the delay introduced by a causal filter does not affect
the operation of the system. For our purposes in this laboratory, we will need to be careful to account for the delay
introduced by FIR filters when comparing two signals with a mean-squared or RMS distortion measure.

6.2.3 Noise and distortion

One of the most common reasons to apply a filter is to attempt to remove noise. There is no single definition of noise,
but the most general usage describes noise as any unwanted component of a signal. For instance, a common type of
electrical noise has a sinusoidal characteristic with a frequency of 60 Hz. This noise arises from the frequency of the
alternating current used to distribute electricity. This 60 Hz signal can “leak” into other systems and corrupt sensor
measurements. Another common type of noise is random noise. This sort of noise typically has a jagged-looking
characteristic. It typically manifests itself as static in audio signals and “snow” in images.
Filtering gives us a means of reducing the noise in a signal through frequency blocking. In general, filters operate
by attenuating (i.e., blocking) certain frequencies in a signal while passing others with relatively little attenuation.
Note that removing noise in this way requires the noise to have a different frequency-domain description than the
signal of interest.

102 The University of Michigan, All rights reserved

 6.2.3 Noise and distortion

s[n] Recovery ^ ^ ^
x[n] + s[n] = x[n] + v[n]
Filter

v[n]

Figure 6.3: A block diagram of additive noise and a recovery filter that attempts to remove the noise.

For instance, consider the example of 60 Hz sinusoidal noise. If our signal of interest is composed of frequencies
above 60 Hz, we can treat this component as low frequency noise and attempt to remove it with a filter that blocks low
frequencies. This sort of filter is generally called a highpass filter. If our signal has components above and below 60
Hz, we might try to remove the corrupting signal by only eliminating frequencies near 60 Hz. This would require a
bandpass filter
Random noise typically has frequency components all over the spectrum. However, a good portion of these com-
ponents will usually have higher frequency than the frequencies in our signal. Thus, we might consider the application
of a lowpass filter that blocks high frequencies to reduce the amplitude of noise components.
Consider the block diagram in Figure 6.3. This is a model where a signal of interest, x[n], is corrupted by the
addition of a noise signal, v[n]. We apply a recovery filter to try to remove the noise component from s[n] = x[n] +
v[n]. The resulting signal is ŝ[n] = x̂[n] + v̂[n], where x̂[n] is the filtered signal of interest (which we hope will be as
similar to x[n] as possible) and v̂[n] is the filtered noise signal (which we hope will be as small as possible). Often we
can tune the noise-removal filter to increase its “strength” (by, for instance, increasing the length of a moving average
filter). The “stronger” the filter, the more noise we can eliminate. Unfortunately, the filter also distorts the signal of
interest; a stronger filter will distort the signal of interest more. Thus, the use of filters to remove noise can be thought
of as finding a tradeoff between two types of distortion. The goal, then, is to find the point where the total distortion (as
measured by the mean-squared error or RMS error between x[n] and ŝ[n]) is minimized as a function of filter strength.

Nonlinear filtering
While standard FIR filters can be useful for noise reduction, in some cases we may find that they distort the desired
signal too much. An alternative is to use nonlinear filters. Nonlinear filters have the potential to remove more noise
while introducing less distortion to the desired signal; however, the effects of these filters are much more difficult to
analyze.
Consider the case of an image, for instance. One of the most important features of images of natural scenes
are edges. Edges in images are usually just sharp transitions where one object ends and another begins. If we are
attempting to remove high-frequency noise from an image, we will often apply a lowpass filter. Edges, though, have
considerable high-frequency content, so the edges in resulting image will be smoothed out. To get around this problem,
we can consider the application of a common nonlinear filter called a median filter. Median filters replace each sample
of a signal with the median (i.e., the most central value) of a block of samples around the original sample. That is, we
can describe the operation of the median filter as

y[n] = M edian(x[n + M1 ], . . . , x[n], . . . , x[n − M2 ]) (6.9)

where 
x((N +1)/2) N odd
M edian(x1 , . . . , xN ) = 1 (6.10)
2 (x(N/2) + x(N/2+1) ) N even
and where x(n) is the nth smallest of the values x1 through xN . The order of the median filter is given by M1 + M2 ,
and it determines how many samples will be included in the median calculation. Note that the filter is noncausal

The University of Michigan, All rights reserved 103

Laboratory 6. FIR Filtering and Image Processing

because its output depends on future, as well as past and present, inputs. Unlike lowpass filters, median filters tend
to preserve edges in signals very well. These filters are also very powerful for removing certain types of noise while
introducing relatively little distortion. In this laboratory, we will examine the effect of applying nonlinear filters to
two-dimensional signals.

6.2.4 Filtering two-dimensional signals

The above discussions of filtering are for one-dimensional signals. Suppose we would like to filter two-dimensional
signals like images instead of just one-dimensional signals. There are three ways to approach this.
The first approach simply applies one-dimensional filters to each of the rows (or each of the columns) of an image.
This approach tends to produce an “uneven” filtered signal that is, for instance, smoothed in one dimension but not the
other. This unevenness is generally not desirable and motivates a second approach.
The second approach is somewhat “stronger” than the first. This approach applies one-dimensional filters to both
the rows and the columns. In this lab we will adopt the convention that we first filter the columns, and then filter the
rows of the resulting image. Most types of filters that we use in one dimension can be extended to two dimensions in
this fashion. For example, if we apply the moving average filter with difference equation give in equation (6.6), for
instance, this will have the effect of smoothing the image. Note that the edge effects and delay issues discussed earlier
also apply to two-dimensional filtering done in this fashion.
If we apply that moving average filter with order 4 to the columns and then the rows of the image, what is the
mathematical effect of the operation? It is not too difficult to see that each sample of the image has been replaced by
the average of a 5 × 5 block of pixels. This suggests that we could describe this filtering operation in terms of two-
dimensional set of filtering coefficients. For instance, the difference equation for this two-dimensional filter would
be
X 4 X 4
1
y[m, n] = x[m − k, n − l]. (6.11)
25
k=0 l=0
1
This operation is equivalent to filtering with a two-dimensional set of coefficients bk,l , where bk,l = 25 for k = 0, . . . , 4
and l = 0, . . . , 4.
This result suggests the third, most general, approach to FIR filtering of two-dimensional signals. The general
difference equation for this approach is

X
M2 X
N2
y[m, n] = bk,l x[m − k, n − l]. (6.12)
k=−M1 l=−N1

[−M1 , M2 ] and [−N1 , N2 ] define the range of nonzero coefficients. Note that a filter, such as the one defined by
equation 6.11, is causal if M1 and N1 are non-negative. However, we should also note that in image processing,
causality is rarely important. Thus, two-dimensional FIR filters typically have coefficients centered around b0,0 . A
schematic of such a set of filter coefficients is shown in Figure 6.4.

6.2.5 Image processing with FIR filters

If you’ve ever used photo editing software like Adobe Photoshop, you may have seen operations called “smoothing”
and “sharpening”. These and many similar operations are typically implemented using simple two-dimensional FIR
filters. We will consider three such operations in this laboratory: smoothing (or blur), edge finding, and sharpening (or
edge enhancement)
We’ve already suggested that a moving average filter performs a smoothing operation. However, there are more
advanced ways of smoothing. Consider, for instance, a filter that weights samples nearby more strongly than those

104 The University of Michigan, All rights reserved

 6.2.5 Image processing with FIR filters

b2,2 b2,1 b2,0 b2,-1b2,-2

k
b1,2 b1,1 b1,0 b1,-1b1,-2
b0,2 b0,1 b0,0 b0,-1b0,-2
b-1,2 b-1,1 b-1,0 b-1,-1b-1,-2
b-2,2 b-2,1 b-2,0 b-2,-1b-2,-2

Figure 6.4: The coefficients of a two-dimensional moving average filter. In this figure, pixels exist at the intersection
of the horizontal and vertical lines.

that are far away. This performs a “weaker” smoothing, but it also introduces less distortion. Because of this, these
sorts of filters are often more useful for random noise reduction than standard moving average filters.
The “edge finding” filter highlights edges in an image by producing large positive or negative values while setting
constant regions of the image to zero. The most basic edge finding filter is a simple one-dimensional first difference
filter. A first difference filter has the difference equation

y[n] = x[n] − x[n − 1]. (6.13)

This filter will tend to respond positively to increases in the signal and negatively to decreases in the signal. Adjacent
input samples that are identical (or nearly so), though, will tend to cancel one another, causing the output to be zero (or
close to zero). There are various two-dimensional “equivalents” of the first-difference filter, many of which respond
to edges of a particular orientation. One general edge-finding filter has the following difference equation:
1
y[m, n] = 4 x[m+ 1, n + 1] − x[m + 1, n] + 14 x[m + 1, n − 1]
− x[m, n + 1] + 3x[m, n] − x[m, n − 1] (6.14)
+ 14 x[m − 1, n + 1] − x[m − 1, n] + 14 x[m − 1, n − 1]

This filter “finds” edges of almost any orientation by outputting a value with large magnitude wherever an edge occurs.
Both the first difference filter and this general edge-finding filter are examples of highpass filters. Note the “oscillatory”
pattern of bk values such that adjacent coefficients are negatives of one another. This pattern is characteristic of
highpass filters. Note that both of these filters will typically produce both positive and negative values, even if our
input signal is strictly non-negative. Also note that for both of these filters, the average of the bk coefficients is zero;
this means that these filters tend to “reject” constant regions of an input signal by setting them to zero.
The third operation, sharpening, makes use of an edge finding filter as well. Basically, the sharpening filter pro-
duces a weighted sum of the output of an edge-finding filter and the original image. Suppose that x[m, n] is the
original image, and y[m, n] is the result of filtering x[m, n] with the filter defined in equation (6.14). Then, the result
of sharpening, z[m, n] is given by
z[m, n] = x[m, n] + by[m, n], (6.15)

where b controls the amount of sharpening; higher values of b produce a “sharper” image. Note that z[m, n] can also
be viewed as the output of a single filter. For display purposes, we will threshold the resulting signal so that the output
image has the same range of data values as the input image. That is, assuming that our input image has values between

The University of Michigan, All rights reserved 105

Laboratory 6. FIR Filtering and Image Processing

0 and 255, the final output of the sharpening operation, ẑ[m, n] will be

 0 z[m, n] < 0
ẑ[m, n] = z[m, n] 0 ≤ z[m, n] ≤ 255 (6.16)

255 255 < z[m, n]

Note that thresholding is a nonlinear operation, but it is not crucial to the sharpening process. This final result can also
be considered to be the output of a single nonlinear filter.
Sharpening is a useful operation when an image has undergone an undesired smoothing operation. This happens
frequently in optical systems when they are not entirely in focus. Unlike smoothing filters, though, sharpening filters
tend to enhance random noise; often they may make “noise-like” components of a signal visible where they were not
visible before.

6.3 Some M ATLAB commands for this lab

• 1-D Filtering in M ATLAB: The usual method for causal filtering in M ATLAB is to use the filter command,
like this:

>> yy = filter(bb,1,xx);

(We’ll use the second parameter later in the course when we study IIR filters.) xx is a vector containing the
discrete-time input signal to be filtered, bb is a vector of the bk filter coefficients, and yy is the output signal.
The first element of this vector, bb(1), is assumed to be b0 .
By default, filter returns a portion of the filtered signal equal in length to xx. Specifically, the resulting
signal includes the start-up transient but not the ending transient. This means that the output will be delayed by
an amount determined by the coefficients of the filter.
A method for filtering which does not introduce delay is often desirable, i.e. a noncausal filtering method, espe-
cially when calculating RMS error between filtered and original versions of a signal. The command filter2
is meant as a two-dimensional filtering routine, but it can be used for 1-D filtering as well. Further, it can be
instructed to return a “delay-free” version of the output signal. When using filter2, it is important that xx
and bb are either both row vectors or both column vectors. Then, we use the command

>> yy = filter2(bb,xx,'same');

where xx is the input signal vector, yy is the output signal vector, and bb is the vector of filter coefficients. If
the length of the vector bb is odd, the b0 coefficient is taken to be the coefficient at center of the vector bb. If
the length of bb is even, b0 is taken to be just left of the center. The output of filter2 has support equal to
that of the input signal xx.
Though we will not use these additional options, we can also have filter2 return the full length of the filtered
signal (the length of the input signal plus the order of the filter) like this:

>> yy = filter2(bb,xx,'full');

or just the portion not affected by edge effects (the length of the input signal minus twice the order of the filter),
like this:

>> yy = filter2(bb,xx,'valid');

106 The University of Michigan, All rights reserved

 6.3 Some M ATLAB commands for this lab

• 2-D Filtering in M ATLAB: Three approaches to filtering a two-dimensional signal were mentioned in Sec-
tion 6.2.4. The first approach, which simply applies a one-dimensional filter to each row of the image (alterna-
tively, to each column) can be implementing with the MATLAB commands described in the previous bullet.
The second approach applies a one-dimensional filter first to the columns and then to the rows of the image
produced by the first stage of filtering. If the one-dimensional filter is causal with coefficients bk contained in
the M ATLAB vector bb and the image is contained in the 2-dimensional matrix xx, then this approach can be
implemented with the command

>> yy = filter(bb,1,filter(bb,1,xx)')';

Note that we do not need to vectorize the image xx, because when presented with an matrix, filter applies
one-dimensional filtering to each column. However, to perform the second stage of filtering (on rows of the
image produced by the first stage), we need to transpose the image produced by the the first stage of filtering
and then transpose the final result again to restore the original orientation. This approach will introduce edge
effects at the top and on one side of the image; however the resulting image will be the same size as x.
The third approach uses a two-dimensional set of coefficients bk,l . If these coefficients are contained in the
matrix bb and the image is contained in the matrix xx, then the filter can be implemented with the command

>> yy = filter2(bb,xx,'same');

Note that the 'same' parameter indicates that the filter is non-causal and thus the b0,0 coefficient is located as
near to the center of the matrix bb as possible. The same alternate third parameters for filter2 that are listed
in the 1-D filtering section apply here as well.
• Generating filter coefficients: We will be examining the effects of many types of filters in this laboratory.
Some have filter coefficients that can be generated easily in M ATLAB. Others require a function (which we will
provide to you) to generate. Note that the the vectors representing the bk ’s will be column vectors.

1. An L-point moving average filter has filter coefficients given by bb = ones(L,1)/L.

2. A first-difference filter has filter coefficients given by bb = [1; -1];
3. The function g_smooth produces coefficients for a particular type of smoothing (lowpass) filters with
easily tunable “strength”. g_smooth takes a single real-valued parameter, which is the “width” of the
filter, and returns a set of tapered filter coefficients of the corresponding filter, bb. For example,
>> bb = g_smooth(1.2);
returns the coefficients for a filter with “width” 1.2. A g_smooth filter with width 0 will pass the input
signal without modification, and higher widths will smooth more strongly. Good nominal values for the
width range from 0.5 to 2. The bk coefficients for g_smooth filters with several widths are plotted in
Figure 6.5.
4. The function g_smooth2 is the two-dimensional equivalent of g_smooth. It again takes a single input
parameter (the width of the filter) and returns the two-dimensional set of filter coefficients of the corre-
sponding filter. For example,
>> bb = g_smooth(0.8);
returns the coefficients of a filter with width 0.8.
5. In Section 6.2.5, we presented a general-purpose two-dimensional edge-finding filter in equation (6.14).
The coefficients for this filter are given by

The University of Michigan, All rights reserved 107

Laboratory 6. FIR Filtering and Image Processing

0.14
Width = 3
Width = 4
0.12 Width = 5
Width = 6

0.1

Coefficient amplitude
0.08

0.06

0.04

0.02

0
0 5 10 15 20 25 30 35 40 45
Coefficient number

Figure 6.5: The coefficients for g smooth filters with varying widths.

>> bb = [.25, -1, .25; -1, 3, -1; .25, -1, .25];

6. In Section 6.2.5, we also discussed a method for implementing a sharpening filter. Since we include a
threshold operation, this operation is nonlinear and cannot be accomplished using only an FIR filter. Thus,
we provide the sharpen command, which takes an image and a sharpening “strength” and returns a
sharpened image:
>> yy = sharpen(xx,0.7);
The second parameter is the strength factor, b, as discussed in Section 6.2.5. A sharpening strength of 0
passes the signal without modification.
7. As described in Section 6.2.3, median filters are a special type of nonlinear filter, and they cannot be
described using linear difference equations. To use a median filter on a one-dimensional signal, we use the
command2 medfilt1 like this:
>> yy = medfilt1(xx,N);
N is the order of the median filter, which simply describes how many samples we consider when taking
the median. In two dimensions3, we use medfilt1 twice:
>> yy = medfilt1(medfilt1(xx,N)',N)';
Again, N is the order of the median filter. Here, we are using a one-dimensional filter on both the rows and
columns of the image. Note that since medfilt1 operates down the columns, we need to transpose the
image between the filtering operations and again at the end.

6.4 Demonstrations in the Lab Section

• Filtering in M ATLAB.

• FIR filters for noise reduction

2 medfilt1 is a part of the signal processing toolbox.
3 Wecan also use medfilt2, but this function is a part of the Image Processing Toolbox which we do not require for this course. medfilt2
works by outputting the median of an N × N block of the image.

108 The University of Michigan, All rights reserved

 6.5 Laboratory Assignment

• Image processing with FIR filters

• Median filtering

6.5 Laboratory Assignment

1. (Noise reduction in 1-D) In this problem, you investigate noise-reduction on one-dimensional signals. Download
the file lab6_data.mat, which contains various signals for this lab. In this problem, we will consider the
signal simple, which is a noise-free one-dimensional signal, and simple_noise, which is the same signal
with corrupting random noise.

(a) (Effects of delay) First, we’ll examine the delay introduced by the two filtering implementations, filter
and filter2, that we will be using. Filter simple with a 7-point running average filter. Do this twice,
first using filter and then using filter2 with the 'same' parameter4.
• [3] Use subplot and plot to plot the original signal and two filtered signals in three subplots of
the same figure.
• [2] One of the filtering commands has introduced some delay. Which one? How many samples of
delay have been added?
• [3] Compute the mean-squared error between the original signal and the two filtered signals. Which
is lower? Why?
(b) (Measuring distortion in 1-D) Now, use filter2 to apply the same 7-point running average filter to
the signal simple_noise. Referring to Figure 6.3, we consider simple to be the signal of interest
x[n], simple_noise to be the noise corrupted signal s[n], and their difference to be the noise, v[n] =
s[n] − x[n]. Note that the lower of the two mean-squared errors that you computed in Problem 1a is
M S(x̂[n] − x[n]), which is a measure of the distortion of the signal of interest introduced by the filter.
• [2] Compute the mean-squared error between simple and simple_noise. Referring back to
Figure 6.3, this is M S(v[n]), the mean-squared value of the noise.
• [2] Compute the mean-square error between your filtered signal and simple. This value is M S(ŝ[n]−
x[n]), which is a measure of how a good a job the filter has done at recovering the signal of interest.
• [1] Determine the distortion due to noise at the output of your reconstruction filter (i.e., M S(v̂[n]))
by subtracting M S(x̂[n] − x[n]) from M S(ŝ[n] − x[n]).
• [3] Compare M S(v̂[n]) and M S(ŝ[n]−x[n]) to M S(v[n]). What is the dominant source of distortion
in this filtered signal?
(c) (Running average filters in 1-D) Use filter2 to apply a 3-point, a 5-point, and an 9-point moving
average filter to simple_noise.
• [3] Use subplot, subplot, and stem to plot the original signal, the three filtered signals, and the
three sets of filter coefficients, in seven panels of the the same figure. (Use plot for the signals and
stem for the coefficients.)
• [3] Compute the mean-squared error between each filtered signal and simple.
• [2] Which of the four moving average filters that you have applied has the lowest mean-squared error?
Compare this value to M S(v[n]).
(d) (Tapered smoothing filter in 1-D) Download the file g_smooth.m, and use it to generate filter coefficients
with “widths” of 0.5, 0.75, and 1.0. (Note the lengths of the returned coefficient vectors. You should plot
the filter coefficients to get a sense of how the “width” factor affects the them.) Use filter2 to apply
these filters to simple_noise.
4 Henceforth, every time you use filter2 in this laboratory, you should use the ’same’ parameter.

The University of Michigan, All rights reserved 109

Laboratory 6. FIR Filtering and Image Processing

• [3] Use plot and subplot to plot the three filtered signals and the three sets of coefficients in six
panels of the same figure.
• [3] Compute the mean-squared error between each filtered signal and simple.
• [3] Which of these filtered signal has the lowest mean-squared error? Compare this value to the lowest
mean-squared error that you found for the moving average filters and to M S(v[n]).

2. (Noise reduction on images) In this problem, you look at the effects of applying smoothing filters to an image
for noise reduction.
Download the files peppers.tif5 and peppers_noise1.tif. The first is a “noise-free” image, while
the second is an image corrupted by random noise. Load these two images into MATLAB.

(a) (Examining 2-D filter coefficients) We’ll be using the function g_smooth2 to produce filter coefficients
for this problem. To get a sense of what these coefficients look like, generate the coefficients for a
g_smooth2 filter with width 5. In two side-by-side subplots of the same figure:
• [2] Display the coefficients as an image using imagesc.
• [1] Generate a surface plot of the coefficients using the command surf(bb) (assuming your coeffi-
cients matrix is called bb).
(b) (Examine the effects of noise) First, we’ll consider the noisy signal peppers_noise1.
• [3] Use subplot to display peppers and peppers_noise1 side-by-side in a single figure.
Remember to set the color map, set the axis shape, and include a colorbar as you did in lab 4.
• [3] Compute the mean-squared error between these two images.
(c) (Minimizing the MSE) Our goal is to find a g_smooth2 reconstruction filter that minimizes the mean-
squared error between the filtered image and the original, noise-free image. Use filter2 when filtering
signals in this problem.
• [6] Find a filter width that minimizes the mean-squared error. What is this filter width and the corre-
sponding mean-squared error? (Hint: you might want to plot the mean-squared error as a function of
filter width.)
• [2] Display the filtered image with the smallest mean-squared error.
• [2] Look at some filtered images with different widths. Can you find one that looks better than the
minimum mean-squared error image6 ? What filter width produced that image?

3. (Salt and pepper noise in images) Next, we’ll look at methods of removing a different type of random noise from
this image. Download the file peppers_noise2.tif and load it into M ATLAB. This signal is corrupted
with salt and pepper noise, which may result from a communication system that loses pixels.
(a) (Examining the noise) First, let’s see what we’re up against. Salt and pepper noise randomly replaces
pixels with a value of either 0 or 255. In this image, one-fifth of the pixels have been lost in this manner.
• [2] Display peppers_noise2.
• [2] Compute the mean-squared error between this image and peppers.
(b) (Using lowpass filters) Now, let’s try using some g_smooth2 filters to eliminate this noise. Start by using
filter2 to filter peppers_noise2 with a g_smooth2 filter of width 1.3. Note that this is very close
to the optimal width value.
5 Like “cameraman”, “peppers” is a standard image used for testing image processing routines. Our version, however, is smaller than the

traditionally used image.

6 Though mean-squared error is widely used as a measure of signal distortion, it is well known that its judgments of quality do not always

correspond closely to the eye’s judgments of quality.

110 The University of Michigan, All rights reserved

 6.5 Laboratory Assignment

• [2] Display the resulting image.

• [3] Compute the mean-squared error.
(c) (Using median filters) Finally, let’s use a median filter to try to remove this noise. Apply median filters of
order 3 and 5 to peppers_noise2.
• [3] Use subplot to display the two filtered images side-by-side in the same figure.
• [3] Compute the mean-squared errors between the median-filtered signals and peppers.
• [2] Look at the filtered images and describe the distortion that the median filters introduce into the
signal.
• [3] Compare the median filter to the g_smooth2 filters. Discuss both measured distortion and the
appearance of the filtered signals.
4. (Edge-finding and enhancing) In this last problem, we’ll look at edge-finding and sharpening filters.
(a) (Applying a first difference filter) In order to see how edge-finding filters work, let’s start in one dimension.
Use filter to apply a one-dimensional first difference filter to the signal simple (which can be found
in lab6_data.mat).
• [2] Plot the resulting signal.
• [3] There are five non-zero “features” of this signal. (These features should be clear from the plot.)
Describe them and what they correspond to in simple.
(b) (“Finding” edges) Now we’d like to look at the effects of the general edge-finding filter presented in Section
6.2.5. Use filter2 to apply this filter to peppers.
• [2] Display the resulting image.
• [2] Describe the resulting image.
• [2] Zoom in on the filtered image and examine some of the more prominent edges. What do you
notice about these edges? (Hint: Are they just a “ridge” of a single color?)
(c) (Sharpening an image) Download sharpen.m and use the function to display several sharpened versions
of the peppers image.
• [3] Use subplot. to display the sharpened image with a “strength” of 1 next to the original
peppers image.
• [2] Zoom in on this sharpened image. What makes it look “sharper”? (Hint: Again, look at the
prominent edges of the images. What do you notice?)
• [2] The sharpened images (especially for strengths greater than 1) generally appear more “noisy” than
the original image. Speculate as to why this might be the case.
(d) (Using sharpening to remove smoothing) Finally, we want to try using the “sharpen” function to undo a
blurring operation. Download the file peppers_blur.tif and load it into M ATLAB.
• [2] Compute the RMS error between peppers and peppers_blur.
• [6] Use sharpen to “de-blur” the blurred image. Find the sharpening strength that minimizes the RMS
error of the “de-blurred” image. Include this strength and its corresponding RMS error in your report.
• [2] Display the “de-blurred” image with the minimum RMS error and alongside peppers_blur
using subplot. Include the resulting figure in your report.
Note that sharpening is very much a perceptual operation. The minimum distortion sharpened image
may not look terribly much improved. Look at what happens as you increase the sharpening factor even
more. With additional “sharpening,” the (measured) distortion may increase, but the result looks better
perceptually.

The University of Michigan, All rights reserved 111

Laboratory 6. FIR Filtering and Image Processing

5. On the front page of your report, please provide an estimate of the average amount of time spent outside of lab
by each member of the group.

112 The University of Michigan, All rights reserved

Laboratory 7

Decoding DTMF: Filters in the Frequency

Domain

7.1 Introduction

In Lab 6, you examined the behavior of several different filters. Some of the filters were “smoothing filters” that
averaged the signal over many samples. Others were “sharpening” filters that accentuated transitions and edges. While
it is very useful to understand the effects of these filters in the time-domain or (for images) the spatial-domain, it is
often not easy to quantify these effects, especially when we are dealing with more complicated filters. Thus, just as we
did with signals, we would like to obtain a better understanding of the behavior of our filters in the frequency-domain.
Assuming that our filter is linear and time-invariant, we can talk about the filter having a frequency response. We
derive the frequency response in the following way. We know that if we put a complex exponential signal into such
a filter, the output will be a scaled and shifted complex exponential signal with the same frequency. The amount of
scaling and phase shift, though, is dependent on the frequency of the input signal. If we send a complex exponential
signals with some frequency through the filter, we can measure the scaling and phase shifting of that signal. The
collection of complex numbers which corresponds to this scaling and shifting for all possible frequencies is known as
the filter’s frequency response. The magnitude of the frequency response at a given frequency is the filter’s gain at that
frequency.
In this lab, we will be using the frequency response of filters to examine the problem solved by telephone touch-
tone dialing. The problem is this: given a noisy audio channel (like a telephone connection), how can we reliably
transmit and detect phone numbers? The solution, which was developed at AT&T, involves the transmission of a sum
of sinusoids with particular frequencies. In order for this solution to be feasible, we must be able to easily decode the
resulting signal to determine which numbers were dialed. We will see that we can do this easily by considering filters
in the frequency domain.

7.1.1 “The Question”

• How can we decode telephone touch-tone (DTMF) signals?

The University of Michigan, All rights reserved 113

Laboratory 7. Decoding DTMF: Filters in the Frequency Domain

7.2 Background
7.2.1 DTMF signals and Touch ToneTM Dialing
Whenever you hit a number on a telephone touch pad, a unique tone is generated. Each tone is actually a sum of
two sinusoids, and the resulting signal is called a dual-tone multifrequency (or DTMF) signal. Table 7.1 shows the
frequencies generated for each button. For instance, if the “6” button is pressed, the telephone will generate a signal
which is the sum of a 1336 Hz and a 770 Hz sinusoid.

Frequencies 1209 Hz 1336 Hz 1477 Hz

697 Hz 1 2 3
770 Hz 4 5 6
852 Hz 7 8 9
941 Hz * 0 #

Table 7.1: DTMF encoding table for touch tone dialing. When any key
is pressed, the tones of the corresponding row and column are generated.

We will call the set of all seven frequencies listed in this table the DTMF frequencies. These frequencies were
chosen to minimize the effects of signal distortions. Notice that none of the DTMF frequencies is a multiple of
another. We will see what happens when the signal is distorted and why this property is important.
Looking at a DTMF signal in the time domain does not tell us very much, but there is a common signal processing
tool that we can use to view a more useful picture of the DTMF signal. The spectrogram is a tool that allows us to
see the frequency properties of a signal as they change over time. The spectrogram works by taking multiple DFTs
over small, overlapping segments1 of a signal. The magnitudes of the resulting DFTs are then combined into a matrix
and displayed as an image. Figure 7.1 shows the spectrogram of a DTMF signal. Time is shown along the x-axis and
frequency along the y-axis. Note the bars, each of which represents a sinusoid of a particular frequency existing over
some time period. At each time, there are two bars which indicate the presence of the two sinusoids that make up the
DTMF tone. From this display, we can actually identify the number that has been dialed; you will be asked to do this
in the lab assignment.

7.2.2 Decoding DTMF Signals

There a number of steps to perform when decoding DTMF signals. The first two steps allow us to determine the
strength of the signal at each of the DTMF frequencies. We first employ a bank of bandpass filters with center
frequencies at each of the DTMF frequencies. Then, we process the output of each bandpass filter to give us an
indication of the strength of each filter’s output. The third step is to “detect and decode.” From the filter output
strengths, we detect whether or not a DTMF signal is present. If it is not, we refrain from decoding the signal until
a tone is detected. Otherwise, we select the two filters with the largest output strengths and use this information to
determine which key was pressed. A block diagram of the DTMF decoder can be seen in Figure 7.2.

Step 1: Bandpass Filters

From Lab 2, you may recall that correlating two signals provides us with a measure of how similar those two signals
are. Since convolution is just “correlation with a time reversal,” we can use this same idea to design a filter that passes
a given frequency. If our filter’s impulse response “looks like” the signal we want to pass, we should get a large
amplitude signal out; similarly, signals that are different will produce smaller output signals.
1 Note that each segment is some very small fraction of a second, and the segments usually overlap by 25-75%.

114 The University of Michigan, All rights reserved

 7.2.2 Decoding DTMF Signals

4000
60
3500
50
3000
Frequency (Hz)

2500 40

2000 30

1500
20
1000

10
500

0 0
0 0.5 1 1.5 2 2.5 3 3.5 4
Time (sec)

Figure 7.1: A spectrogram of a DTMF signal. Each horizontal bar indicates a sinusoid that exists over some time
period.

When performing DTMF decoding, we want filters that pass only one of the DTMF frequencies and reject all
of the rest. We can make use of the correlation idea above to develop such a bandpass filter. We want our impulse
response to be similar to a signal with the frequency that we wish to pass; this is the filter’s center frequency. This
means that for a bandpass filter with center frequency f , we want our impulse response, h, to be equal to

sin(2πfc k/fs ) 0 ≤ k ≤ M
h[k] = (7.1)
0 else

From this equation, we have an FIR filter with order M . (Note that the support length of the impulse response is
M + 1.) What should M be? M is a design parameter. You may remember from Lab 3 that correlating over a long
time produces better estimates of similarity. Thus, we should get better differentiation between passed frequencies
and rejected frequencies if M is large. There is a tradeoff, though. The longer M is, the more computation that is
required to perform the convolution. Thus for efficiency reasons we would like M to be as small as possible. More
computation also equates to more expensive devices, so we prefer smaller M for reasons of device economy as well.
Since we have seven DTMF frequencies, we will also have seven bandpass filters in our system; in our decoder system,
we will choose a different value of M for each bandpass filter.
Because of the relatively small set of frequencies of concern in DTMF decoding, we will see that larger M do
not necessarily produce better frequency differentiation. In order to judge how good a bandpass filter is at rejecting
unwanted DTMF frequencies, we will define the gain-ratio, R. Given a filter with center frequency fc and frequency
response H, the gain-ratio is
|H(fc )|
R= (7.2)
max |H(fˆ)|
where fˆ is in the set of DTMF frequencies and fˆ 6= fc . In words, we define R to be the ratio of the filter’s gain at its
center frequency to the next-highest gain at one of the DTMF frequencies. Having a high gain-ratio is desirable, since
it indicates that the filter is rejecting the other possible frequencies.
Note that since we will be comparing the outputs of a variety of bandpass filters, we also need to normalize each
filter by the center frequency gain. Thus, we will need to record not only the M that we select but also the center

The University of Michigan, All rights reserved 115

Laboratory 7. Decoding DTMF: Filters in the Frequency Domain

697 Hz Lowpass
Bandpass Filter Rectify Filter

770 Hz Lowpass Detect

Bandpass Filter Rectify Filter
s(t) and Decoded
Number
Decode
DTMF
Signal 1477 Hz Lowpass
Bandpass Filter Rectify Filter

Step 1 Step 2 Step 3

Figure 7.2: A block diagram of the DTMF decoder system. The input is a DTMF signal, and the output is a string of
numbers corresponding to the original signal.

frequency gain. You will be directed to record and include these gains in the lab assignment.

Step 2: Determining filter output strengths

In order to measure the strength of the filter’s output, we actually want to measure (or follow) the envelope of the
filter outputs. To follow just the positive envelope of the signal, we first need to eliminate the negative portions of the
signal. If we simply truncate all parts of the signal below zero, we have applied a half-wave rectifier. Alternately, we
can simply take the absolute value of the signal, in which case we have applied a full-wave rectifier. It is possible to
build rectifiers using diodes, and it turns out that half-wave rectifiers are easier to design. However, full-wave rectifiers
are preferable, and they are no more difficult to implement in M ATLAB. Thus, we will use full-wave rectifiers2 . See
Figure 7.3 to see the effects of these two types of rectifiers.
If we now pass the rectified signal through a smoothing filter, the output will be a nearly constant signal whose
value is a measure of the strength of the filter’s input at the center frequency of the filter. To accomplish this smoothing
we will use a simple moving average filter with impulse response
 1
MLP +1 0 ≤ k ≤ MLP
hLP = (7.3)
0 else

The order of this filter is MLP . The value MLP (and thus the corresponding strength of the smoothing filter) is a design
parameter of the decoder system. When choosing MLP , there is a tradeoff between amount of smoothing and transient
effects. If our filter’s impulse response is not long enough, the output signal will still have significant variations. If
it is too long, transient effects will dominate the output of the filter. If it is too short, the system may “smooth over”
short DTMF tones or periods of silence. Note that in our decoder system, we will apply the same smoothing filter to
the output of each filter. Figure 7.3 shows the results of smoothing for half-wave and full-wave rectified signals.

Step 3: “Detect and Decode”

Once we have processed the outputs of the bandpass filters, we can now detect whether or not a DTMF tone is present
and, if it is, determine which key was pressed to produce it. Ultimately, we want to convert our signal into a sequence
of keys pressed to produce this sequence. The detect-and-decode step itself involves three steps.
2 Note that the names “full-wave rectifier” and “half-wave rectifier” come from the circuit implementation of these systems

116 The University of Michigan, All rights reserved

 7.2.2 Decoding DTMF Signals

Half-Wave
Rectifier

Input signal
Full-Wave Lowpassed
Rectifier Result

Figure 7.3: A comparison of half-wave and full-wave rectification. Notice that full-wave rectification allows us to
achieve a higher output signal level after lowpass filtering.

The first step is to detect whether a DTMF tone is actually present at a particular time. If it is not, we risk making
an error in our decoding of the input signal. We detect the presence of a DTMF tone by comparing the rectified and
smoothed bandpass filter outputs to a threshold, c. If any of the signals are greater than the threshold, then we decide
that a DTMF tone is present. Figure 7.4a shows the rectified-and-smoothed output from one of the bandpass filters
and the threshold to which it is compared. The threshold is a design parameter of the decoder. We generally want to
the threshold to be high enough that noise will not “trigger” the detector during a period of silence, but low enough
that noise won’t pull the signal from a DTMF tone below the threshold. Figure 7.4b shows a noisy DTMF tone with
the threshold.
When the input signal is noisy, there is an additional problem during the transient portions at the beginning and
end of a DTMF tone. Near the threshold crossing, the noise could cause the signal to cross the threshold several times,
as shown in Figure 7.4c; this might cause a single DTMF tone to be decoded as multiple key presses. To avoid this
problem, we do not make a detection decision for every sample of the input signal. Instead, we only make a decision
every 100 samples. This makes it more likely that there will only be one decision made in the vicinity of the threshold
crossing. It also reduces computation time somewhat. Note that the number 100 is somewhat arbitrary. We can choose
a smaller number, but then we increase the risk the multiple-crossing problem. Alternatively, we can make it larger;
however, it we make it too large, our detector may miss short tones or silences.
The second step is to decode of the DTMF tones that we have detected in the previous step. By “decode,” we simply
mean that we must decide which key was pressed to generate a particular DTMF tone. To do this, we determine which
two bandpass filters have the largest output at each time when a DTMF tone was detected. Then, we effectively
perform a table look-up to see which key was pressed at these times. The result is a sequence of decoded numbers
corresponding to key presses. However, each DTMF tone will generally produce a sequence of identical numbers
since it is “decoded” at many times during the duration of the DTMF tone. To translate this sequence of numbers into
a sequence of key presses, we need a third step.
The third step simply combines adjacent, identical numbers in the decoded sequence. That is, a “run” of identical
numbers is replaced by a single number. Through this process, each DTMF tone is finally represented by a single
number. Note that for this process to work correctly, our sequence of numbers must also contain an indication of when
no tone was present. Otherwise, any repeated key press would be decoded as only a single key press.

The University of Michigan, All rights reserved 117

Laboratory 7. Decoding DTMF: Filters in the Frequency Domain

(a) (b) t (c)

Figure 7.4: An illustration of the detector subsystem. (a) A clean DTMF signal is compared to a threshold, c. (b) The
threshold should be set so that noise will not produce false tone detections or miss true tone detections in the presence
of noise. (c) Near the threshold crossing, noise can cause multiple detections.

7.2.3 Decoder Robustness

Whenever designing a communication system, like the DTMF coder/decoder described here, it is important to consider
how the system behaves in the presence of undesirable effects. For instance, the telephone system could corrupt our
DTMF signal with some amount of static. Under such conditions, how well would the decoder work? How much
noise can the system tolerate? These are all questions about the robustness of the decoder system to noise. No system
can work perfectly under less than ideal conditions, so it is important to understand when and how a system will fail.
In the lab assignment, we will examine the robustness of this system under noise.

7.2.4 Sidenote: Searching Parameter Spaces

Quite frequently, you will find yourself in the position of searching for a “good value” for a particular parameter about
which you have no other information. In these cases, there are some techniques that we can employ to speed the search.
The basic idea is that we want to get “in the ballpark” before we worry about finding locally optimum solutions. To
do this, we think about varying parameters over factors of 2 or factors of 10. Thus, you might try parameter values of
0.01, 0.1, 1, 10, and 100 to get a general notion of how the system responds to a parameter. Once we have done this,
we can then isolate a smaller range over which to optimize. This prevents us from spending too much time searching
aimlessly.

7.3 Some M ATLAB commands for this lab

• Computing the frequency response of an FIR filter: The M ATLAB command freqz returns the frequency
response of a filter at a specified number of discrete-time frequencies. The general usage of freqz for causal
FIR filters is

>> [H,w] = freqz(bb,1,n);

Here, bb is the set of filter coefficients (i.e., the impulse response) of the FIR filter, n is the number of points in
the range [0, π) at which to evaluate the frequency response, H is the frequency response, and w is the set of n
corresponding discrete-time frequencies, which are spaced uniformly from 0 to π. The frequency response, H,
is a vector of complex numbers which define the gain (abs(H)) and phase-shift (angle(H)) of the filter at
the given frequencies.

118 The University of Michigan, All rights reserved

 7.3 Some M ATLAB commands for this lab

Alternatively, we can evaluate the frequency response only at a specified set of frequencies by replacing n with
a vector of discrete-time frequencies. Thus, the command

>> H = freqz(bb,1,[pi/3, pi/2, 2*pi/3]);

2π
returns the frequency response at the discrete-time frequencies π3 , π2 , and 3 .
When we apply a filter to a sampled signal with sampling frequency fs (in samples per second), we can evaluate
the frequency response at the discrete-time frequencies corresponding to a specified set of continuous time-
frequencies in Hertz in the following manner:

>> H = freqz(bb,1,[100 200 400 500]/fs*2*pi);

This converts the specified continuous-time frequencies into discrete-time frequencies and evaluates the fre-
quency response at those points.

• Sorting a vector: The M ATLAB command sort sorts a vector in ascending order. Thus, given a vector x, the
command

>> y = sort(x);

produces a vector y such that y(1) is the smallest value in x and y(end) is the largest value in x.

• Creating matrices of ones and zeros: In order to create arrays of arbitrary size containing only ones or only
zeros, we use the M ATLAB ones and zeros commands. Both commands take the same set of input parameters.
If only one input parameter is used, a square matrix with the specified number of rows and columns is generated.
For instance, the command

>> x = ones(5);

produces a 5 × 5 matrix of ones. Two parameters specify the desired number of rows and columns in the matrix.
For instance, the command

>> x = zeros(4, 8);

produces a 4 × 8 matrix (i.e., four rows and eight columns) containing only zeros. To generate column vectors
or row vectors, we set the first or second parameter to 1, respectively.

• The DTMF Dialer: dtmf_dial.m is a DTMF “dialer” function. It takes a vector of key presses (i.e., a
phone number) and produces the corresponding audio DTMF signal. Note that this function as provided is
incomplete; you will be directed to complete it in the laboratory assignment. (The lines of code that you need
to complete are marked with a ?.) To produce the DTMF signal that lets you dial the number 555-2198, use the
command:

>> signal = dtmf_dial([5 5 5 2 1 9 8]);

An optional second parameter will cause the function to display a spectrogram of the resulting DTMF signal:

>> signal = dtmf_dial([5 5 5 2 1 9 8],1);

The University of Michigan, All rights reserved 119

Laboratory 7. Decoding DTMF: Filters in the Frequency Domain

This function assumes a sampling frequency of 8192 samples per second. Each DTMF tone has a length of 1/2
second, and the tones are separated by 1/10 second of silence. Note that the number 10 corresponds to a '#',
11 corresponds to a '0', and 12 corresponds to a '*'.

• The DTMF Decoder: dtmf_decode.m is an (incomplete) DTMF decoder function. (Once again, the lines
of code that you need to complete are marked with a ?.) It takes a DTMF signal (as generated by dtmf_dial)
and returns the sequence of key-presses used to create the signal. Thus, if our DTMF signal is stored in signal,
we decode the signal using the command:

>> decoded = dtmf_decode(signal);

An optional second parameter will cause the function to display a plot of the smoothed and rectified outputs of
each bandpass filter:

>> decoded = dtmf_decode(signal,1);

• Bandpass Filter Characterization: dtfm_filt_char.m is a function that we will use to help us calculate
gain-ratios for the bandpass filters used in the DTMF decoder. We use the function to focus on one of the
bandpass filters at a time. The function takes two parameters: the order, M, of one of the bandpass filter’s impulse
responses and the center frequency in Hertz, frq, of that filter. The function returns a vector containing the gain
(i.e., the magnitude of the frequency response) at each of the DTMF frequencies, from lowest to highest. It also
produces a plot of the frequency response with locations of the DTMF frequencies indicated. Use the following
command to execute the function:

>> gains = dtmf_filt_char(M,frq);

A second optional parameter lets you suppress the plot:

>> gains = dtmf_filt_char(M,frq,0);

• Testing the robustness of the DTMF decoder: dtmf_attack.m is a function that tests the DTMF decoder
in the presence of random noise. This function generates a standard seven digit DTMF signal, adds a specified
amount of noise to the signal, and then passes it through your completed dtmf_decode function. The decoded
string of key presses is compared to those that generated the signal. Since the noise is random, this procedure
is repeated ten times. The function then outputs the fraction of trials decoded successfully. The function also
displays the plot from the last execution of dtmf_decode. (Note: since each call to dtmf_decode takes a
little time, this function is rather slow. Be patient with it.)
For instance, to test the system with a noise power of 2.5, we use the following command:

>> success_rate = dtmf_attack(2.5);

The result is a number that provides the fraction of the 10 trials that were successful.
Although dtmf_attack is a complete function, it calls dtmf_dial and dtmf_decode, each of which you
must complete.

120 The University of Michigan, All rights reserved

 7.4 Demonstrations in the Lab Section

7.4 Demonstrations in the Lab Section

• Examining the frequency response of FIR filters
• Dual tone multi-frequency signals
• Generating “synthetic” DTMF signals.
• Bandpass filters
• The DTMF decoder
• Noise and the DTMF decoder

7.5 Laboratory Assignment

1. (The DTMF dialer.) Before we can decode a DTMF signal, we need to be able to produce DTMF signals. In
this problem, we’ll write a function that takes a phone number and produces the corresponding DTMF signal,
just like the telephone would produce if you dial the number.
Download the function dtmf_dial.m, which is a nearly complete dialer function. You simply need to replace
the question marks by code that completes the function. The first missing line of code generates a DTMF tone
for each number in the input and appends it to the output signal. The second line of code appends a short silence
to the signal to separate adjacent DTMF tones.

• [8] Complete the function and include the code in your lab report.
• [3] Using your newly completed dialer function, execute the following command to create a DTMF signal
and display it’s spectrogram:
>> signal = dtmf_dial([1 2 3 4 5 6 7 8 9 10 11 12],1);
Include the resulting figure in your report. Note how each key press produces a different pattern on the
spectrogram.
• [4] What is the phone number that has been dialed in Figure 7.1?

2. (The bandpass filters of the DTMF Decoder.) As we have noted, a key part of the DTMF decoder is the bank
of bandpass filters that is used to detect the presence of sinusoids at the DTMF frequencies. We have specified
a general form for the bandpass filters, but we still need to choose the filter orders and create their impulse
responses. In this problem you will be identifying good values for M .

(a) (The impulse response of one bandpass filter.) First, we need to be able to create the impulse response for a
bandpass filter. Using equation (7.1) with a sampling frequency fs = 8192 Hz and M = 50, use M ATLAB
to create a vector containing the impulse response, h, of a 770 Hz bandpass filter3 .
• [4] What is the command that you used to create this impulse response?
• [2] Use stem to plot your impulse response.
(b) (The frequency response of one bandpass filter.) When we talk about the response of a filter to a particular
frequency, we can think about filtering a unit amplitude sinusoid with that frequency and measuring the
amplitude and phase shift of the resulting signal. We can certainly do this in M ATLAB, but it’s far simpler
to use the freqz command. Here, you’ll use freqz to examine the frequency response and gain-ratio of
a bandpass filter like the ones we’ll use in the DTMF decoder.
3 Remember that if a filter has order M , the support length of the impulse response should be M + 1.

The University of Michigan, All rights reserved 121

Laboratory 7. Decoding DTMF: Filters in the Frequency Domain

• [4] Use freqz to calculate the frequency response of your 770 Hz bandpass filter at all seven of the
DTMF frequencies4. Calculate the gain at each frequency, and include these numbers in your report.
• [3] From the frequency response of your filter at these frequencies, calculate the gain-ratio, R.
• [2] Do you think that this is a good gain-ratio for our bandpass filters? (Hint: You might want to come
back to this problem after you’ve worked the remainder of this problem.)
(c) (Choosing M for this bandpass filter.) Now, we’d like to see what happens when we change M for your
770 Hz bandpass filter. We’ve provided you with a function that will facilitate this. Download the file
dtmf_filt_char.m. This function will help you to visualize the frequency response of these filters
and to determine their gain at the DTMF frequencies.
• [1] Use this function to verify that the gains you calculated in Problem 2b were correct.
• [2] Include the frequency-response plot that dtmf_filt_char produces in your report.
• [4] The frequency response of this filter is characterized by several “humps” which are typically called
lobes. Describe the frequency response in terms of such lobes. Vary M and examine the plots that
result (you do not need to include these plots). Describe the differences in the frequency response as
M (which represents the length of the filter’s impulse response) is changed.
• [1] What happens to the relative heights of adjacent lobes as M is changed?
• [2] What features of the filter’s frequency response contribute to the gain ratio R?
• [2] For what values of M do we achieve gain ratios greater than 10?
(d) (A function for computing gain ratios.) You’ll need to compute the gain-ratio repeatedly while finding
good design parameters for the bandpass filters, so in this problem you’ll automate this task. Write a
function that accepts a vector of gains (such as that returned by dtmf_filt_char) and computes the
gain ratio, R. (Hint: This is a simple function if you use the sort command. You can assume that the
center frequency gain is the largest value in the vector of gains.)
• [6] Include the code for this function in your report.
(e) (Specifying the bandpass filters.) For each bandpass filter that corresponds to one of the seven DTMF
frequencies, we want to find a choice of M that yields a good gain ratio but also minimizes the computation
required for filtering.
To do this, for each bandpass filter frequency, use dtmf_filt_char and your function from Problem
2d to calculate R for all M between 1 and 200. Then, plot R as a function of M . You can save some
computation time by setting the third parameter of dtmf_filt_char to zero to suppress plotting. You
should be able to identify at least one local maximum5 of R on the plot. The “optimal” value of M that
we are looking for is the smallest one that produces a local maximum of R that is greater than 10.
• [4] Create this plot of R as a function of M for the bandpass filter with a center frequency of 770 Hz.
Include the resulting plot in your report.
• [2] Identify the “optimal” value of M for this filter, the associated center frequency gain, and the
resulting value of R.
• [12] Repeat the above two steps for the remaining six bandpass filters. (You do not need to include the
additional plots in your report.) Create a table in which you record the center frequency, the optimal
M value, the associated center frequency gain, and the resulting value of R.

3. (Completing the DTMF decoder.) Now we have designed the bank of bandpass filters that we need for the
DTMF decoder. In this problem, we’ll use the parameters that we found to help us complete the decoder design.
4 Remember that our system uses a sampling frequency of 8192 Hz
5A local maximum is basically just a point on the plot that is larger than all other values in its vicinity. It may or may not be the highest possible
peak, which is called the global maximum.

122 The University of Michigan, All rights reserved

 7.5 Laboratory Assignment

Download the file dtmf_decode.m. This function is a nearly complete implementation of the DTMF decoder
system described earlier in this lab. There are several things that you need to add to the function.

(a) (Setting the M ’s and the gains of the bandpass filters.) First, you need to record your “optimized” values
of M and the center frequency gains in the function. Replace the question marks on line 29 by a vector of
your optimized values of M . They should be in order from smallest frequency to largest frequency. Do
the same on line 32 for the variable G, which contains the center frequency gains.
• [2] Make these modifications to the code. (At the end of this problem, make sure that you include
your completed function in your report.)
(b) (Setting the impulse responses of the bandpass filters.) Also, you need to define the impulse response for
each bandpass filter on line 49. Use equation (7.1) for this, where the filter’s order is given by M(i).
• [3] Make this modifications to the code.
(c) (Selecting the order of the post-rectifier smoothing filter.) Next, you need to specify the post-rectifier
smoothing filter, h_smooth. Temporarily set both h_smooth (line 36) and threshold (line 40) equal
to 1 and run dtmf_decode on the DTMF signal you generated in Problem 1. This function displays a
figure containing the rectified and smoothed outputs for each bandpass filter. With h_smooth equal to 1,
no smoothing is done and we only see the results of the rectifier in this figure. We will use moving average
filters of order MLP , as defined by the M ATLAB command
>> h_smooth = ones(M_LP+1,1)/(M_LP+1);
We want the smoothed output to be effectively constant during most of the duration of the DTMF tones,
but we don’t want to smooth so much that we might miss short DTMF tones or pauses between tones.
• [4] Examine the behavior of the smoothed signal when you replace line 36 with moving average
filters with order MLP equal to 20, 200, and 2000. Which filter order, MLP gives us the best tradeoff
between transient effects and smoothing?
• [1] Set h_smooth to be the filter you have just selected.
(d) (Detection threshold.) Finally, you need to identify a good value for threshold. threshold deter-
mines when our system detects the presence of a DTMF signal. dtmf_decode plots the threshold on its
figure as a black dotted line. We want the threshold to be smaller than the large amplitude signals during
the steady-state portions of a DTMF signal, but larger than the signals during the start-up transients for
each DTMF tone. (Hint: When choosing a threshold, consider what might happen if we add noise to the
input signal.)
• [4] By looking at the figure produced by dtmf_decode, what would be a reasonable threshold
value? Why did you choose this value?
• [2] Set threshold to the value you have just selected.
• [2] Now, execute dtmf_decode and include the resulting plot in your report. (Note: You can
include this plot in black and white, if you like.)
• [2] dtmf_decode should output the same vector of “key presses” that was used to produce your
signal. What “key presses” does the function produce? Do these match the ones used to generate the
DTMF signal? If not, you’ve probably made a poor choice of threshold.
(e) Remember to include the code for your completed dtmf_decode function in your report.

4. (Robustness of the DTMF decoder to noise.) In the introduction to this lab, we indicated that we would be
transmitting our DTMF signals over a noisy audio channel. So far, though, we have assumed that the decoder
sees a perfect DTMF signal. In this problem, we will examine the effects of additive noise on the DTMF decoder.

The University of Michigan, All rights reserved 123

Laboratory 7. Decoding DTMF: Filters in the Frequency Domain

(a) Download the file dtmf_attack.m. Execute dtmf_attack with various noise powers. Find a value
of noise power for which some but not all of the trials fail.
• [3] What value of noise power did you find? (Hint: use the parameter searching method discussed in
the background section to speed your search).
• [6] Make a plot of the fraction of successes versus noise power. Include at least 10 values on your plot.
Make sure that your minimum noise power has a success rate at (or at least near) 1 and your maximum
noise power has a success rate at (or near) 0. Try to get a good plot of the transition between high
success rates and low success rates. While making this plot, pay attention to the types of errors that
the decoder is making.
(b) By examining the plots for failure trials and the types of errors that the decoder is making, you should be
able to speculate about the source of the errors.
• [3] What types of errors is the system making when it decodes the noisy signals?
• [2] Speculate about what could you do to the decoder in order to increase the system’s tolerance to
additive noise.

5. On the front page of your report, please provide an estimate of the average amount of time spent outside of lab
by each member of the group.

124 The University of Michigan, All rights reserved

Laboratory 8

Classification and Vowel Recognition

8.1 Introduction
The ability to recognize and categorize things is fundamental to human cognition. A large part of our ability to
understand and deal with the world around us is a result of our ability to classify things. This task, which is generally
known as classification, is important enough that we often want to design systems that are capable of recognition and
categorization. For instance, we want vending machines to be able to recognize the bills inserted into the bill changer.
We want internet search engines to classify web pages based on their relevance to our query. We want computers that
can recognize and classify speech properly so that we can interact with them naturally. We want medical systems that
can classify unusual regions of an x-ray as cancerous or benign. We want high speed digital communication modems
that can determine which sequence of, say, 64-ary signals that was transmitted.
There is a vast array of applications for classification. We have actually already seen some of these applications.
Detection, which we studied in Labs 1 and 2, is a form of classification where we chose from only two possibilities. In
this lab, we consider one popular application of multiple-alternative classification: speech recognition. In particular,
we will focus on a simplified version of speech recognition, namely, vowel classification. That is, we will experiment
with systems that classify a short signal segment of an audio signal which corresponds to a spoken vowel, such as an
“ah”, an “ee”, an “oh”, and so on. (We won’t deal with how one determines that a given segment corresponds to a
vowel.) In the process, we will develop some of the basic ideas behind automatic classification.
One of these basic ideas is that an item to be classified is called an instance. For example, if each of 50 short
segments of speech must be individually classified, then each segment is considered to be one instance. A second
basic idea is that there is a finite set of prespecified classes to which instances may belong. The goal of a classifier
system (or simply a classifier) is to determine the class to which a presented instance belongs. A third basic idea is that
to simplify the process, the classification of a given instance is based on a set of feature values. This set is a relatively
small list of numbers that, to an appropriate degree, describes the given instance. For example, a short segment of
speech might contain thousands of samples, but we will see that vowel classification can be based on feature sets
with as few as two components. A fourth basic idea is that classification is often performed by comparing the feature
values for an instance to be classified with sets of feature values that are representative of each class. The output of
the classifier will be the class whose representative feature values are most similar, in some appropriate sense, to the
feature values of the instance to be classified.

8.1.1 “The Question”

• What is the general framework for performing automatic classification?

The University of Michigan, All rights reserved 125

Laboratory 8. Classification and Vowel Recognition

Feature
Instance Feature Vector Feature Class
Calculator Classifier Label

Class Representative
Feature Vectors

Figure 8.1: Block diagram of a general classifier system.

• How can we recognize and classify vowels in a speech recognition system?

8.2 Background
8.2.1 An Introduction to Classification
You may recall Lab 7, in which we developed a system for decoding DTMF signals into the sequence of key-presses
that produced the original signal. Our DTMF decoder was actually performing classification on each segment of the
DTMF signal. Classification is a process in which we examine instances of some thing (like an object, a number, or a
signal) and try to determine which of a number of groups, or classes, each instance belongs to. We can think of this as
a labeling process. In our DTMF decoder, for example, we looked at a given segment of the signal and labeled it with
a number corresponding to an appropriate key press.
Generally, classification is a two-stage process. Figure 8.1 shows a block diagram of a classifier system. First, we
need some information about the instance that we are considering. This information is traditionally referred to as a set
of features. If we are classifying people, for instance, we might use height, weight, or hair color as features. If we are
classifying signals, we might use power, the output of some filter, or the energy in a certain spectral band as features.
So that we can deal with our features easily, we generally like to have a set of measurable features to which we can
assign numerical feature values. When we are using more than one feature to describe an instance, we typically place
all of the feature values into a feature vector, f = (f1 , f2 , . . . , fN ). N is the number of elements in the feature vector
and is called the dimension of the feature vector. A feature vector is calculated for each instance we wish to classify by
measuring the appropriate aspects of that instance. As shown in Figure 8.1, the first block is the “feature calculator,”
which takes an instance (of a signal, for instance) and produces the set of numerical feature values. For our DTMF
decoder, our features were the spectral strength of a given segment of the signal at each DTMF frequency. That is, the
feature calculator produced a seven-element feature vector, one for each DTMF frequency.
The second stage of classification, the “feature classifier” (which we have previously called a “decision maker”).
uses the feature vectors to decide which class a feature vector belongs to. Generally, we make this decision by
comparing the feature vector for an instance to each member of a set of representative feature vectors, one for each
class under consideration. The idea is that the feature classifier labels the instance as the class that has the most
similar representative feature vector. We will discuss the specifics of the feature classifier after we have presented a
classification example.1
Before we continue, we should note the relationship between what we previously called “detection” and what we
now call “classification”. Detection generally refers to binary “signal present” or “signal not present” decisions. For
instance in Lab 1, we used energy to decide whether a signal was present or not, and Lab 2 we used correlation to
make such decisions. As such, detection, is generally considered to be a special case of the more general notion of
1 The classifier for the DTMF decoder can be viewed as implicitly operating in this fashion. It is an interesting exercise to find the representative

feature vectors implicitly used.

126 The University of Michigan, All rights reserved

 8.2.2 A classification example

classification, which refers to decisions among two or more classes. However, this usage is not universal. For example,
“detection” is sometimes used to describe a system that decides which of 64 potential signals was transmitted to a
modem, each representing a distinct pattern of 6 bits. This lab assignment also generalizes the idea, used in Labs
1 and 2, that decisions are made on a single number or feature. However, as noted before, Lab 7 also used such a
generalization.

8.2.2 A classification example

The easiest way to get a feel for classification problems is to consider an example. Suppose that we have a large
number of flowering plants in our garden, each of which belongs to one of two different types, A and B. We know
which plant belongs to each type, but they all look very similar. Now, we may know which of our plants belong to
which type, but we would also like to be able to classify new plants as either Type A or Type B as we expand our
garden. To do this, we will design a classifier for these plants. The plants in our garden with known type will form our
design set (or training set) and will be used for designing the classifier.
If we happen to know that Type A plants tend to be taller than Type B plants, this suggests that we might be able
to use the plant’s height as a feature for classification. Suppose we measure the heights of all of the plants in our
garden (our design set) and then plot a histogram of this data. We might see something like Figure 8.2. This is an
unusually good case in which we have two readily-distinguished clusters of feature values. The type A plants form a
cluster with heights centered around a mean (i.e. average) of 50 centimeters, and the type B plants form a cluster with
heights centered around a mean of 40 centimeters. Most importantly, the two clusters do not overlap. This suggests
that classification can indeed be based on plant height.
How do we use this information to classify a new plant (i.e., a new instance)? Intuitively, if the new plant’s height
is closer to the Type A mean of 50 cm than to the Type B mean of 40 cm, we should classify the plant as type A rather
than type B. In this case, we can use a simple threshold test. If a new plant’s height is greater than 45 cm (which is
halfway between two mean feature values), we classify the new plant as Type A. Conversely, if it’s height is less than
45 cm we classify it as Type B. In other words, for each class, we use the mean feature value as the class representative,
and we compare the feature value of a new instance to be classified (it’s height) to the two means and decide the class
whose representative feature value (its mean) is closest to the feature value of the given instance.
Figure 8.3 (left) shows a histogram of plant heights in a more troublesome scenario. In this case, Type A plants
still tend to be taller than Type B plants, but there are a significant number of plants that we will confuse (that
is, misclassify) if we decide exclusively using this one feature. Though we will typically need to deal with some

14
Type A
Type B
12

10
Number of flowers

0
36 38 40 42 44 46 48 50 52 54 56
Flower height (cm)

Figure 8.2: A simple example where one feature (plant height) is sufficient to perform classification. This histogram
shows how many plants have a given height.

The University of Michigan, All rights reserved 127

Laboratory 8. Classification and Vowel Recognition

10
Type A
9 Type B 50

7 45
Number of flowers

Leaf length (mm)

6
Type A
5 Type B
40
4

3
35
2

1
30
0
35 40 45 50 55 60 35 40 45 50 55 60
Flower height (cm) Flower height (cm)

Figure 8.3: An example where a histogram of one feature is not sufficient to perform perfect classification (left), but a
scatter plot of two features shows a clear separation between the two classes (right).

classification error, we can often reduce it by adding more features. Suppose we measure not only the height of the
plant but also the average length of its leaves. Now, instead of a histogram, we can look at the training set of features
using a scatter plot, in which we plot a point for each feature vector in our training set. We put one of the two features
along each of the plot’s axes. For example, a scatter plot for the two features just mentioned for each plant is shown
in Figure 8.3 (right). Here we again see two distinct clusters, which suggests that we can classify with little error by
using these two features together.2
How do we design a classifier for this case? We cannot simply use a threshold on one of the features. Instead, we
will use a more general decision rule, which is based on mean feature vectors and distances between an instance and
the mean feature vectors. First, let us consider the two features as a two-dimensional vector f = (f1 , f2 ). Thus, if a
plant is 52 cm tall and has leaves with average length of 44 mm, our feature vector is f = (52, 44). Now, given the
feature vectors from each plant in our design set of one type, we want to calculate a mean feature vector for plants
of that type. Since the mean feature vector indicates the central tendency of each feature in a class, we use it as a
representative of the entire class. To calculate a mean feature vector in this case, we first take the mean, m1 , of all of
the plant heights for plants of one type. Then we take the mean, m2 , of all of the leaf lengths for plants of the same
type. The mean feature vector is then f̄ = (m1 , m2 ). Note that this is the general procedure for calculating the mean
of a set of vectors, regardless of the vector’s dimension. On the scatter plot in Figure 8.3, we’ve plotted the locations
of mean feature vectors with large symbols.
As with the one-feature case, we will classify new instances based on how close they are to each of the mean feature
vectors. To do this, we still need to know how to calculate distances between two feature vectors. For simplicity, we
will calculate distances using the Euclidean distance measure3 . The Euclidean distance between two vectors is simply
the straight-line distance between their corresponding points on a scatter plot like that in Figure 8.3. To calculate the
distance, d, between two feature vectors (f1 , f2 ) and (m1 , m2 ), we simply use the formula
p
d = (f1 − m1 )2 + (f2 − m2 )2 (8.1)

Euclidean distance generalizes to any number of dimensions; the general formula can be found later in equation (8.2).
Note that the Euclidean distance is essentially the RMS difference (i.e., RMS “error”) between two vectors4 , which
2 In this case, classification using either feature by itself will result in many classification errors. That is, by itself, neither feature is sufficient to

separate the two clusters. One can see this by projecting the scatter plot on to either one of the axes. When we do so, we see that the two feature
values of the two classes are intermingled, rather than remaining distinct.
3 There are a wide variety of possible distance measures; Euclidean distance is certainly not the only choice.
4 The two calculations actually differ by a scaling factor, since RMS involves a mean while Euclidean distance involves a sum.

128 The University of Michigan, All rights reserved

 8.2.3 A few more classification examples

Type A
52 Type B

50

48

Leaf length (inches)

46

44

42

40

38

36
40 45 50 55
Flower height (inches)

Figure 8.4: An example where two features are not as clearly separated.

we have used repeatedly throughout this course. Here, though, we refer to the computation as “Euclidean distance”,
rather than RMS difference, to motivate a geometric interpretation of classification.
Now that we have designed a classifier for this case, we can finally consider the classification of a new instance.
To classify a new instance, we first calculate the distances between that instance’s feature vector and the mean feature
vectors of each class. Then, we simply classify the instance as a member of the class for which the distance is smallest.
Consider what this means in terms of the scatter plot. Given a new instance, we can plot it’s feature vector on the scatter
plot. Then, we classify based on the nearest mean feature vector. For a two-class case such as that shown in Figure
8.3, there exists some set of points that are equally far from both mean feature vectors. These points form a decision
line that separates the plane into two halves. We can then classify based on the half of the plane on which a feature
vector falls. For example, in Figure 8.3, any plant with a feature vector that falls above the line will be classified as
type B. Similarly, any plant with a feature vector that falls below the line will be classified as type A.
With this classification rule, we can correctly classify almost all of our training instances. However, note that we’re
not classifying perfectly. There is one rogue type B close to the rest of the type A’s. In general, though, we will need
to accept more error than this.
Of course, two features may not be enough either. If our scatter plot looked like the one in Figure 8.4, then we can
still see the two clusters, but we can’t perfectly distinguish them based only on these two features. The line we draw
for our distance rule will properly classify most of the instances, but many are still classified incorrectly. Once again,
we can either accept the errors that will be made or we can try to find another feature to help us better distinguish
between the two classes. Unfortunately, visualizing feature spaces with more than two dimensions is rather difficult.
However, the intuition we’ve built for two-dimensional feature spaces extends to higher dimensions. We can calculate
mean feature vectors and distances in the roughly the same way regardless of the number of dimensions.

8.2.3 A few more classification examples

We’ve looked at a simple classification task with only two classes, but there are some more examples that are instruc-
tive. Consider Figure 8.5(A). In this example, the two clusters fall right on top of one another, so we will have very
poor classification performance. This is an example where neither of the features assist classification performance very
much. In this case, we need to find better features before we can have much luck with classification. Figure 8.5(B)
shows a similar example. Here, feature 2 will help us to improve our classification performance but feature 1 will not.
(Can you see why?) Note that it may be worse to have a second feature which is bad than to only have one (good)
feature. Unfortunately, determining which features are good and which are bad is nontrivial when we have more than
two (or three) features and can no longer visualize the data.

The University of Michigan, All rights reserved 129

Laboratory 8. Classification and Vowel Recognition

(A) (B)
11.5

10.5
11
Feature 2

Feature 2
10.5
10
10
9.5
9.5

9
9
7 7.5 8 8.5 9 7 8 9
Feature 1 Feature 1

(C) (D)
24 30

22
25
20
Feature 2

Feature 2

18 20

16
15
14

12 10
10 15 20 25 10 15 20 25 30
Feature 1 Feature 1

Figure 8.5: (A) Classes overlap, so the features do not allow much discrimination; these are bad features. (B) Feature 2
aids discrimination, but Feature 1 does not. (C) An example with four distinct classes; decision lines are approximate.
(D) An example with three indistinct classes.

It is also important to realize that we may have more than just two classes in a classification problem. Figure 8.5(C)
shows an example in which we have four classes that have distinct clusters in our feature space. The mean feature
vectors are indicated on these plots with large markers. Again, we can use the same distance-based decision rule to
classify instances. That is, we classify a given instance according to the class whose mean feature vector is closest to
its feature vector. We have included (approximate) decision lines on this plot which partition the feature space (i.e.,
the plane) into four pieces. These indicate which class a given feature vector will be classified as. Of course, multiple
classes can be indistinct, too. Figure 8.5(D) shows an example for three indistinct classes. Here, two of the classes (∗
and ◦) are reasonably distinguishable, but we cannot easily separate the third class (+) from either of the other two.

8.2.4 Formalizing the feature classifier

In the previous sections, we presented some examples of classifier design and operation. Here, we’ll formalize these
ideas with respect to the general classifier block diagram shown in Figure 8.1. In particular, we will expand upon the
feature classifier shown in that block diagram. Figure 8.6 shows an expanded block diagram of the feature classifier5 .
Suppose we need our classifier to decide among C classes and that the classifier will be based on a set of N
5 The main goal of any feature classifier is to determine which of a set of representative feature vectors a new instance is most similar to. In

this lab, we use Euclidean distance to measure similarity, and so we use a distance-based feature classifier. Other types of feature classifier are also
possible, such as a correlation-based feature classifier.

130 The University of Michigan, All rights reserved

 8.2.4 Formalizing the feature classifier

Distances to class
representatives
Feature Distance Decision Class
Vector Calculator Rule Label

Class Representative
Feature Vector

Figure 8.6: Block diagram of a distance-based feature classifier, which makes the decision in a general classifier
system.

features, forming a feature vector f = (f1 , . . . , fN ). Our feature classifier will rely on a set of representative feature
vectors, one for each class. We will denote the representative feature vector for the cth class as f̄c = (f¯c,1 , . . . , f¯c,N ),
where c = 1, . . . , C.
Given a set of representative feature vectors (the choice of such will be discussed later), we can classify new
instances using the feature classifier. The feature classifier (seen in Figure 8.6) has two steps. The first step computes
the distances between the input feature vector and each of the class representatives. As we have done in the previous
sections, we will use Euclidean distance in our system. Equation (8.1) gives the formula for Euclidean distance in two
dimensions. For a general, N -dimensional feature space, we use the following equation. Let u = (u1 , . . . , uN ) and
v = (v1 , . . . , vN ) be two N -dimensional vectors (i.e., arrays with length N ). We calculate the Euclidean distance
between them as
v
uN
uX p
d(u, v) = t (vi − ui )2 = (v1 − u1 )2 + (v2 − u2 )2 + · · · + (vN − uN )2 . (8.2)
i=1

Again, we note that, to within a scaling factor, Euclidean distance is equivalent to the root mean squared error between
two vectors.
The second step of the feature classifier applies a decision rule to select the best class for the input instance. The
decision rule that we will use is the nearest class representative rule. This simply means that the classifier decides the
class whose whose representative feature vector is closest (in Euclidean distance) to the feature vector of the instance
being classified. That is, if f is the feature vector for an instance to be classified, then the decision rule decides class c
if d(f , f̄c ) is less than d(f , f̄c0 ) for all other classes6 c0 . Other decision rules, which may weight the distances from the
various class representatives, are also possible, but they will not be considered here7 .
Let us now discuss how to choose the class representative feature vectors f̄1 , . . . , f̄C . Finding these vectors is the
main aspect in feature classifier design. We have previously suggested that we can find a representative feature vector
for a class by taking the mean across some set of instances that belong to that class. We describe this calculation
formally as follows. Suppose that we have a set of N -dimensional feature vectors from M instances of a given class
c (this the design set of instances for this class). Let f̃i = (f˜i,1 , f˜i,2 , . . . , f˜i,N ) denote the ith such feature vector. We
calculate the mean feature vector, f̄c = (f¯c,1 , . . . , f¯c,N ), for this class as

1 X
M
1
f̄c = f̃i = (f̃1 + f̃2 + · · · + f̃M ). (8.3)
M i=1 M
6 Ifit should happen that f is equally closest to two or more class representatives, then an arbitrary choice is made among them.
7 Note that our DTMF signal classifier from Lab 7 used a simpler “feature classifier” that was based on neither distance nor correlation. However,
with a little extra work it could have been formulated as either of these types of classifier, most likely without a degradation of performance.

The University of Michigan, All rights reserved 131

Laboratory 8. Classification and Vowel Recognition

Alternatively, we can say that the j th element of the mean feature vector, f̄c , is

1 X˜
M
1 ˜
f¯c,j = fi,j = (f1,j + f˜2,j + · · · + f˜M,j ). (8.4)
M i=1 M

8.2.5 Measuring the performance of a classifier

The performance of a classifier is based on the how many errors it makes. One good way to characterize the perfor-
mance of a classifier is with a confusion matrix, K, which simply measures how often members of one class were
confused with members of another class. Specifically, when the classifier recognizes N classes, then the confusion
matrix K = [Ki,j ] is an N × N matrix, whose element Ki,j in the ith row and j th column is the fraction of those times
that class j occurs but the classifier produces class i. That is,
# of class j instances classified as i
Ki,j = (8.5)
# of class j instances
For example, the following is confusion matrix for a hypothetical four-class classifier:
 
.9 .03 .01 .02
 .03 .95 .01 .03 
K =   .05 .01 .96 .1 
 (8.6)
.02 .01 .02 .85

The diagonal elements, Kn,n , show what fraction of instances from the the nth class were correctly classified. The .9
in the upper left corner, for instance, indicated that 90% of instances from the first class were classified as belonging
to the first class. Thus, higher diagonal elements are desirable.
The off-diagonal element Kn,m indicates what fraction of instances from the mth class were misclassified as
belonging to class n. In the example above, for instance, the .02 in the upper right corner indicates that 2% of
instances in the fifth class were incorrectly classified as belonging to the first class. Thus, we hope that off-diagonal
elements are as small as possible. The confusion matrix for a perfect classifier will be an identity matrix (i.e., ones on
the diagonals, zeros elsewhere).

Data usage when designing classifiers

When designing classifiers and testing their performance, it is important to note that classifiers generally perform
better on the training data used in their design than on new data of the same general type. Thus, to objectively assess
the performance of a classifier, one must test it on a different data set, usually called a test set, than the one on which
it was designed. To see why, consider the extreme case in which the training data contains just one feature vector
for each class, which becomes the mean feature vector for its class. In this case, the resulting classifier will perfectly
classify every feature vector in the training set. However, it may not do very well at all when classifying other data. In
more realistic cases where the training data has quite a few instances of each class, the performance of the classifier
on the training data will usually be somewhat (but possibly not significantly) better than on test data. Nevertheless, it
is widely accepted that testing a classifier on independent data is good practice. Thus, when a certain amount of data
is available for design, it is usually divided into two sets — one for training, the other for testing.
To keep things simple, in this lab we will not separate our set of instances into separate design sets and testing
sets. Thus, it is important to know that we may not be accurately characterizing our system’s performance in the “real
world.” You will be given an opportunity to test our vowel classifier and see how well it actually performs on your
voice. Specifically, all of the vowel instances provided in this lab were taken from a single speaker. How does this
affect the performance of the system for other speakers? Can you come up with a better set of representative feature
vectors (possibly by collecting vowel samples from a variety of speakers)?

132 The University of Michigan, All rights reserved

 8.2.6 Vowel Classification

"ee" vowel "ah" vowel

40 40
30
30
20
|S(ω)| (db)

|S(ω)| (db)
20
10
0 10

−10 0
−20
−10
−30
0 1000 2000 3000 4000 0 1000 2000 3000 4000
Frequency (Hz) Frequency (Hz)

"ee" vowel "ah" vowel

40
40
30
30

20 20
|S(ω)| (db)

|S(ω)| (db)

10 10

0 0

−10 −10

−20 −20
0 1000 2000 3000 4000 0 1000 2000 3000 4000
Frequency (Hz) Frequency (Hz)

Figure 8.7: The magnitude spectrum (in decibels) of four vowel signals. The plots on the left correspond to two
instances of an “ee” vowel, as in the word tree. The plots on the right correspond to two instances of an “ah” vowel,
as in the word father. The solid line is a smoothed version of the spectrum, which shows the general trends of the
spectrum.

8.2.6 Vowel Classification

So far, we have discussed a general-purpose framework for performing classification. In this section, we will specifi-
cally discuss how to apply these techniques to the classification of vowels in speech signals.
Vowels in speech are nearly periodic segments of the speech signal. From our studies of the Fourier Series, we
know that these segments of the signal are thus approximately equal to the sum of sinusoids with harmonically related
frequencies. As with the DTMF signals in Lab 7, the time-domain provides relatively little information about the
signal. So, as with the DTMF signals, this suggests that we need to examine vowels in the frequency domain. Figure
8.7 shows examples of the magnitude spectrum (in decibels8 ) of two different vowels. The plots on the left correspond
to an “ee” vowel (as in the word tree), while the plots on the right correspond to an “ah” vowel (as in the word father).
Also shown is a smoothed version of each spectrum, which shows its general trend.
There are a number of interesting things to note about these plots. First, we can see the peaks that correspond
to the harmonics that make up the periodic signal. Notice that the peaks are spaced more closely in some plots than
others, corresponding to a lower fundamental frequency and thus a longer fundamental period. As illustrated by this
figure, though, the fundamental frequency of the signal is independent of the vowel being produced. Notice that the
overall shape of the frequency spectrum is different between the two vowels, but remains relatively constant between
8 To convert a number, x, into decibels, we use the formula xdB = 20 log10 (x).

The University of Michigan, All rights reserved 133

Laboratory 8. Classification and Vowel Recognition

the two instances of each vowel, as can be seen from the smooth versions. This shape determines the timbre9 of the
sound, and, correspondingly, the “sound” of the vowel. Notice that there are peaks in the smoothed spectrum at various
places. These peaks are called formants; it is generally known that the position of these formant is the primary feature
that distinguishes one vowel from another, i.e. that makes one vowel sound different from another.
Unfortunately, there is no solid definition of a “formant,” and they are remarkably difficult to identify automatically.
In fact, there is some disagreement as to what constitutes a formant in some cases. In this lab, we’ll work with two
sets of features that hopefully capture the information contained in the formant positions. In Lab 9, we’ll investigate
the use of another, somewhat more sophisticated feature for vowel recognition. This feature actually models speech
production, and thus should more readily capture the relevant aspects of the vowel signal.
The first feature set that we use in this lab will be the formant features. The formant features attempt to locate the
formants themselves using a simple algorithm. This algorithm first uses the DFT to compute the spectrum of a short
segment of a vowel. Then, the spectrum is smoothed using a weighted averaging filter. Finally, the algorithm returns
frequencies of the largest peaks on the smoothed signal that occur above and below 1500 Hz. Thus, there are two
formant features, so the resulting feature vector is two-dimensional.
The second feature set, the filter bank features, are quite similar to the features used in the DTMF decoder. The
filter bank features compute the energy (in decibels) of the speech signal after it has been passed through a bank of six
bandpass filters. We will use bandpass filters with center frequencies of 600 Hz, 1200 Hz, 1800 Hz, 2400 Hz, 3000
Hz, and 3600 Hz. Thus, the resulting feature vectors are six-dimensional.
Note that there are a large number of vowels that we could possibly consider. However, for simplicity we will
restrict attention to just five vowels: “ee” (as in tree), “ah” (as in father), “ae” (as in fate), “oh” (as in boat), and “oo”
(as in moon). Each of these five vowels will be its own class.

8.3 Some M ATLAB commands for this lab

• Converting a value into decibels: Expressing a numerical value in decibels compresses the range of values
using a logarithmic transformation. Thus allows us to see features that might otherwise not be visible. The
decibel transformation is particularly useful when looking at the magnitude spectrum of audio signals, since
hearing is based on a logarithmic amplitude scale. Given a value x, we convert it to decibels using the command

>> x_dB = 20*log10(x);

This command can be also used to simultaneously convert a vector of values to decibels.

• Calculating features for a vowel signal: As indicated, the features we would like to consider in order to
classify a vowel signal are based on the signal’s spectrum. We provide functions to calculate the two feature
sets described in this laboratory. Each function takes an audio waveform, x, and (optionally) the sampling
frequency in samples per second, fs. (If no sampling frequency is specified, a sampling frequency of 8192
samples per second is assumed.) Both functions return a row vector, y, that contains the features calculated
from the waveform.
To compute the “formant features,” use calc_formants.m:

>> y = calc_formants(x,fs);

Similarly, to compute the “filter bank features,” calc_fbank.m:

>> y = calc_fbank(x,fs);
9 Pronounced “tambor.”

134 The University of Michigan, All rights reserved

 8.3 Some M ATLAB commands for this lab

• Working with features vectors in M ATLAB: In this lab, we will adopt the convention that a feature vector is
a row vector, and that a set of feature vectors, such as a set of class representatives or a set of testing data, is
stored in a matrix such that there is one feature vector per row and one feature per column. This allows us to
easily compute mean feature vectors from such a matrix.
When computing Euclidean distances, note that the computation is almost the same as that which we used for
computing RMS error. The only difference is that we replace the mean operation by a summation.

• Advanced plotting: You may recall from Lab 1 that we can use M ATLAB’s plot command to change the color
and style of plotted lines. A line-style string consists of as many as three parts. One part specifies a color (for
instance, ‘k’ for black or ‘r’ for red). Another part specifies the type of markers at each data point (for instance,
‘*’ uses asterisks while ‘o’ specifies circles). The third part specifies the type of line used to connect the points
(‘:’ specifies a dotted line, while ‘-’ specifies a solid line). Note that these three parts can occur in any order,
and all are optional. If no color is specified, one is chosen automatically. If no marker is specified, a marker will
be not be plotted. If a marker is specified but a line type is not, then lines will not be drawn between data points.
Thus, the command:

>> plot(x1,y1,'rx',x2,y2,'k:');

will plot x1 versus y1 using red verb-x-’s with no connecting line, and also x2 with y2 with a dotted connecting
line but no marker. See help plot for more details.
Additionally, we can change the width of lines and the size of markers using additional parameter-pairs. For
instance, to increase the line width to 2 and the marker size to 18, use the command

>> plot(x1,y1,'rx--','Linewidth',2,'Markersize',18);

• Executing the feature classifier: The function feature_classifer is an incomplete function that you
will use to automatically classify a feature vector (or a set of feature vectors) based on the distances to a set of
representative feature vectors. The function takes two inputs. The first input, M, is a matrix of the representative
feature vectors, with one feature vector per row. Note that the vector on the first row corresponds to the first
class, the second row to the second class, and so on. The second input parameter, fmatrix, can either be a
single feature vector to be classifed (stored as a row vector) or a matrix of feature vectors to be classified, with
one feature vector per row. To call feature_classifier, use the command:

>> labels = feature_classifer(M,fmatrix);

The function outputs a column vector of class labels, labels, with one label for each row of fmatrix. The
labels are numbers that indicate which representative feature vector the corresponding instance is closest to.
Thus, if the first element of labels is a 3, it means that the feature vector in the first row of fmatrix is
closest to the representative feature vector in the third row of M.

• Calculating confusion matrices:

We provide you with a function, confusion_matrix, that computes a confusion matrix for you. Note
that confusion_matrix calls your feature_classifier function, so it will not work until you have
completed that function. To compute a confusion matrix, we need the matrix of representative feature vectors
used by that classifier and a set of testing data for each class. Thus, if our matrix of representative feature vectors
is M and class1, class2, and class3 are matrices that contain feature vectors from our testing set with
instances of each of three classes (with one feature vector per row), we compute the 3 × 3 confusion matrix
using the command:

The University of Michigan, All rights reserved 135

Laboratory 8. Classification and Vowel Recognition

>> K = confusion_matrix(M,class1,class2,class3);

Note that the function confusion_matrix works for any number of classes. The size of the confusion
matrix is determined by the number of input parameters.

8.4 Demonstrations in the Lab Section

• Introduction to Classification
• Evaluating performance of a classifier
• Vowel spectra
• Vowel features

8.5 Laboratory Assignment

1. (Examining one vowel instance.) Download the file lab8_data.mat. This file contains a variable called
vowel1, which is a one half-second recording of a vowel sound with a sampling frequency of 8192 samples
per second.
(a) Use soundsc to listen to this vowel.
• [2] To which of the five vowel classes does this vowel belong?
(b) Take the DFT of vowel1. In two subplots of the same figure, plot the magnitude of the DFT and the
magnitude of the DFT in decibels. Only plot the first half of the DFT coefficients in each plot. Also,
make sure that you label the x-axis with the frequency in Hertz, not the DFT coefficient number. (Hint:
The maximum frequency showing on your plot should be 4096 Hz, which is one half of the sampling
frequency.)
• [4] Include this figure in your report.
• [3] Compare the two plots. Are there aspects of the spectrum that are easier to see in one of the plots
than in the other?
• [2] From this figure, estimate the fundamental frequency of the vowel sound.
• [2] Estimate the frequencies (in Hz) of the three most prominent formants.
(c) Download the files calc_formants.m and calc_fbank.m. You will use them to calculate features
for this vowel.
• [2] Calculate and include the formant feature vector for this vowel.
• [2] Compare the calculated features to your estimate of the formant locations.
• [2] Calculate and include the filter bank feature vector for this vowel.
• [2] What are the center frequencies of the filters that have the greatest output amplitude? Compare
this to your estimated formant locations.
2. (Mean feature vectors and hand classification.) lab8_data.mat also has several other variables, including
matrices containing features for 50 instances of each vowel. The variables ah_form, ee_form, ae_form,
oh_form, and oo_form contain formant feature vectors for each vowel. Each matrix has 50 rows (one for
each instance) and two columns (one for each feature value). Similarly, the variables ah_fbank, ee_fbank,
ae_fbank, oh_fbank, and oo_fbank contain the filter bank feature vectors with one row per instance.

136 The University of Michigan, All rights reserved

 8.5 Laboratory Assignment

(a) (Mean feature vectors) Calculate the mean feature vectors for each vowel class and for both feature classes.
• [4] Include the five mean formant feature vectors in your report. Make sure you label them.
• [4] Include the five mean filter bank feature vectors in your report. Again, make sure you label them.
(b) (Generate a scatter plot) We would like to see how separable the classes are from the formant feature
vectors. To do this, you’ll create a scatter plot that plots the first formant location versus the second
formant location. Plot each of the feature vectors, using a different color and marker symbol for each
vowel. Make sure you include a legend. Also, plot the mean vector for each class on your scatter plot.
(Hint: To make your mean vectors stand out, you should increase the line width and marker size for just
those points.)
• [10] Include the scatter plot in your report.
• [4] Interpret this scatter plot. Are all of the classes distinct and easily separated? Do you expect
expect any vowels to be frequently confused? Do you expect any vowels to frequently be classified
correctly?
Food for thought: What would happen if we only used one of these two features for classification? Which
would give us better classification results? Do you think that a third formant feature might improve class
separation?
(c) (Hand classification) Now, you’ll “classify” the signal vowel1 by hand. To do this, you’ll need to compute
the distance between the instance and the five mean feature vectors.
• [5] Compute the distances between the formant feature vector that you generated for vowel1 and the
five mean formant feature vectors.
• [5] Compute the distances between the filter bank feature vector that you generated for vowel1 and
the five mean filter bank feature vectors.
• [3] Using the nearest-representative decision rule, use the above results to classify vowel1. Do the
results for both feature sets agree? If not, which feature set produces the correct answer?

3. (Complete and use the feature classifier code.) In this problem, you’ll complete and then use the function, called
feature_classifier.m, that does this classification for us automatically. As described in the background
section, the function takes as input a matrix of representative feature vectors and a matrix of instances to classify.
We output a label for each of the instances in the input.
(a) Complete the function. (Hint: You should use two for loops. One loops over the rows of the matrix of
test instances. Then, for each instance, loop over the rows of M and compute the distances. To make the
classification for each instance, find the position of the smallest distance and store it in labels.)
• [12] Include your code in your report.
(b) (Test feature_classifier on the formant features.) Place your mean formant feature vectors into
a matrix, M_form with one feature vector per row. For consistency, put “oo” in the first row, “oh”
in the second row, “ah” in the third row, “ae” in the fourth row, and “ee” in the fifth row. Call your
feature_classifier function using this matrix and ee_form. (Hint: To make sure your function
works correctly, you should compare its output to the completed and compiled function feature_classifier_demo.d
If you did not successfully complete feature_classifier, you can use this demo function through-
out the remainder of the lab.)
• [2] What fraction these instances are properly classified?
• [3] Calculate the fraction of the instances that are misclassified as each of the incorrect classes. That
is, determine the fraction that are misclassified as “ah,” the fraction misclassified as “ae,” and so on.
• [1] From this data, what vowel is “ee” most often misclassified as?

The University of Michigan, All rights reserved 137

Laboratory 8. Classification and Vowel Recognition

(c) Repeat the above with the filter bank features, this time using the matrix ee_fbank and generating the
matrix M_fbank. Use the same order for your classes. (Again, you should compare your function’s
output to the output of feature_classifier_demo.dll).
• [2] What fraction these instances are properly classified?
• [3] Calculate the fraction of the instances that are misclassified as each of the incorrect classes.
• [1] From this data, what vowel is “ee” most often misclassified as?
4. (Compute and interpret confusion matrices.) Download the file confusion_matrix.m. In this problem,
you will compute and interpret confusion matrices for the two feature classes.
(a) Use confusion_matrix to compute the confusion matrix for the formant features. Use M_form as
your set of class representatives. Use the vowel following vowel order for your remaining input parameters:
“oo,” “oh,” “ah,” “ae,” and “ee.” (This should be the same as the order as the classes in M_form).
• [3] Include this confusion matrix in your report. Label each row and column with the corresponding
class.
• [2] In Problem 3b, you computed a portion of the confusion matrix. Identify that portion and verify
that your results were correct.
• [2] From this confusion matrix, determine how many instances of “ee” vowels were misclassified as
“oo” vowels.
• [1] Which vowel is most commonly misclassified using this feature set?
(b) Use confusion_matrix to compute the confusion matrix for the filter bank features. Use M_fbank as
your set of class representatives. Use the vowel following vowel order for your remaining input parameters:
“oo,” “oh,” “ah,” “ae,” and “ee.” (This should be the same as the order as the classes in M_form).
• [3] Include this confusion matrix in your report. Label each row and column with the corresponding
class.
• [2] In problem 3c, you computed a portion of the confusion matrix. Identify that portion and verify
that your results were correct.
• [2] From this confusion matrix, determine how many instances of “ae” vowels were misclassified as
“ah” vowels.
• [1] Which vowel is most commonly misclassified using this feature set?
(c) Finally, compare the two confusion matrices.
• [2] Which feature set has the best performance overall?
• [2] Based on the performance of these classifiers, comment on the spectral similarities between the
various vowels. That is, are any of the vowel classes particularly like any of the other vowel classes?
5. On the front page of your report, please provide an estimate of the average amount of time spent outside of lab
by each member of the group.

Food for thought:

• As was mentioned in the background section, all of the vowel instances used here were taken from a single
speaker. As such, this classifier will probably perform better on that speaker’s vowels than on the rest of the
population. The compiled functions formant_classifier.dll and fbank_classifier.dll let you
test the classifier designed here on your own voice. When you execute either of these functions (which require

138 The University of Michigan, All rights reserved

 8.5 Laboratory Assignment

no input parameters), M ATLAB will immediately record one quarter-second from the microphone, compute the
features, and classify the vowel. Use this function to test the classifier on your own voice for various vowels.
Record how well it does on each vowel. Is performance better or worse than what is indicated by the confusion
matrices you calculated?
• The above function also returns the set of features that it computed from the recorded vowel. If you collect these
features into series of matrices of testing data, you can use confusion_matrix.m to formally compute the
performance of this classifier on your voice.
• How well does the classifier work on other people’s voices? Are there certain people for whom it works very
well? Very poorly? Does it work if we vary the pitch?
• The compiled functions listed above take an optional input parameter, which is a matrix of representative feature
vectors. Since the current classifier is designed using only one speaker’s vowels, maybe you can improve the
performance by coming up with a better set of representative feature vectors. To do this, consider gathering a
set of vowels from a number of different speakers and combining them into a set of mean feature vectors. Can
you improve the performance of the system?

The University of Michigan, All rights reserved 139

Laboratory 8. Classification and Vowel Recognition

140 The University of Michigan, All rights reserved

Laboratory 9

Filter Design, Modeling, and the z-Plane

9.1 Introduction

So far, we’ve been considering filters as systems that we design and then apply to signals to achieve a desired affect.
However, filtering is also something that occurs everywhere, without the intervention of a human filter designer. At
sunset, the light of the sun is filtered by the atmosphere, often yielding a spectacular array of colors. A concert
hall filters the sound of an orchestra before it reaches your ear, coloring the sound and adding pleasing effects like
reverberation. Even our own head, shoulders, and ears form a pair of filters that allows us to localize sounds in space.
Quite often, we may wish to recreate these filtering effects so that we can study them or apply them in different
situations. One way to do this is to model these “natural” filters using simple discrete-time filters. That is, if we can
measure the response of a particular system, we would often like to design a filter that has the same (or a similar)
response.
One of the goals for this laboratory is to introduce the use of discrete-time filters as models of real-world filters.
In particular, we will examine how to apply a modeling approach to understanding vowel signals. This in turn will
suggest a way that we might improve the performance of the vowel classifier we developed in Lab 8 using an automatic
modeling method.
Another goal of this lab is to present a method of filter design called pole-zero placement design. Working with
this method of filter design is extremely useful for building an intuition of how the z-plane “works” with respect to the
frequency domain that you are already familiar with. The design interface that we use for this task should help you to
develop a graphical understanding of how poles and zeros affect the frequency response of a system. We will use this
design methodology both to design a traditional “goal-oriented” lowpass filter, and to do some filter modeling.

9.1.1 “The Question”

• How can we design filters for certain purposes?

• How can we model vowel production using discrete-time filters?

The University of Michigan, All rights reserved 141

Laboratory 9. Filter Design, Modeling, and the z-Plane

9.2 Background
9.2.1 Filters and the z-transform
Previously, we have presented the general time-domain input-output relationship for a causal filter1 given by the
convolution sum: X X
y[n] = x[n] ∗ h[n] = h[k]x[n − k] = x[k]h[n − k] , (9.1)
k k

where x[n] is the input signal, y[n] is the output signal, and h[n] is the filter impulse response. Using the z-transform
techniques described in Chapter 7 of DSP First, we can also describe the input/output relationship in the z-domain as

Y (z) = H(z)X(z) , (9.2)

where X(z) is the z-transform of x[n], which is the complex-valued function, defined on the complex plane2 by
X
X(z) = x[n]z −n , (9.3)
n

where Y (z) is the z-transform of y[n], defined in a similar fashion, and where H(z) is the system function of the filter,
which is a complex-valued function defined on the complex plane by one of the following equivalent definitions:

1. The system function is the z-transform of the filter impulse response h[n], i.e
X
H(z) = h[n]z −n . (9.4)
n

2. For X(z) and Y (z) as defined above, the system function is given by

Y (z)
H(z) = . (9.5)
X(z)

The system function has a very important relationship to the frequency response of a system, H(ω̂). The system
function evaluated at ej ω̂ is equal to the frequency response evaluated at frequency ω̂. That is,

H(ω̂) = H(ej ω̂ ) . (9.6)

P
We can derive this result from equation 9.4. If we let z = ej ω̂ , then we know that H(z) = H(ej ω̂ ) = n h[n]e−j ω̂n .
This is simply the definition of a system’s frequency given its impulse response h[n].

9.2.2 FIR Filters and the z-transform

For a causal FIR filter, one can easily determine the system function using either of the equivalent definitions given
above. However, let us highlight the use of the second definition, which will be useful in the next subsection where
the first definition is difficult to apply. In particular, for a causal FIR filter with coefficients {b0 , . . . , bM }, the general
time-domain input-output relationship for a causal FIR filter is given by the difference equation

y[n] = b0 x[n] + b1 x[n − 1] + b2 x[n − 2] + · · · + bM x[n − M ] . (9.7)

1 Note that in this lab, we will only be concerned with causal filters.
2 The complex plane is simply the set of all complex numbers. The real part of the complex number is indicated by the x-axis, while the imaginary
part is indicated by the y-axis.

142 The University of Michigan, All rights reserved

 9.2.2 FIR Filters and the z-transform

Taking the z-transform of both sides of this difference equation yields

Y (z) = b0 X(z) + b1 X(z)z −1 + b2 X(z)z −2 + · · · + bM X(z)z −M

= X(z)(b0 + b1 z −1 + b2 z −2 + b3 z −3 + · · · + bM z −M ) , (9.8)

where we have used the fact that the z-transform of x[n − no ] is X(z)z −no . Dividing both sides of the above by X(z)
gives the system function:

Y (z)
H(z) = = b0 + b1 z −1 + b2 z −2 + b3 z −3 + · · · + bM z −M . (9.9)
X(z)

Notice that H(z) is a polynomial of order M . We can factor the above complex-valued polynomial as3

H(z) = K(1 − r1 z −1 )(1 − r2 z −1 )(1 − r3 z −1 ) · · · (1 − rM z −1 ) , (9.10)

where K is a real number called the gain, and {r1 , . . . , rM } are the M roots or zeros of the polynomial, i.e. the values
r such that H(r) = 0. We typically assume that the filter coefficients bk are real. In this case, the zeros may be real or
complex, and if one is complex, then its complex conjugate is also a zero. That is, complex roots come in conjugate
pairs.
The very important point to observe now from equation (9.10) is that the system function H(z) of a causal FIR
filter is completely determined by its gain and its zeros. Therefore, we can think of {K, r1 , . . . , rM } as one more way
to describe a filter4 . We will see that when it comes to designing an FIR filter to have a certain desired frequency
response, the description of the filter in terms of its gain and its zeros is by far the most useful. In other words, the best
way to design a filter to have a desired frequency response (e.g., a low pass filter) is to appropriately choose its gain
and zeros. One may then find the system function by multiplying out the terms of equation (9.10), and then picking
off the filter coefficients from the system function. For example, the number multiplying z −3 in the system function is
the filter coefficient b3 . The specific procedure will be described shortly.
The fact that we may design the frequency response of a causal FIR filter by choosing its zeros5 stems from the
following principle:
If a filter has a zero r located on the unit circle, i.e. |r| = 1, then H(\r) = 0, i.e. the frequency response
has a null at frequency \r. Similarly, if a filter has a zero r located close to the unit circle, i.e. |r| ≈ 1,
then H(\r) ≈ 0, i.e. the frequency response has a dip at frequency \r. In either case, H(ω̂) ≈ 0, when
ω̂ ≈ \r.
The above fact follows from the property that if ω̂ = \r and |r| = 1, then ej ω̂ = r, and so

H(ω̂) = H(ej ω̂ ) = H(r) = 0 . (9.11)

A similar statement shows H(ω̂) ≈ 0 when |r| ≈ 1 and/or ω̂ ≈ \r.

From this fact, we see that we see that we can make a filter block a particular frequency, i.e. create a null or a dip in
the frequency response, simply by placing a zero on or near the unit circle at an angle equal to the desired frequency6.
On the other hand, the frequency response at frequencies corresponding to angles that are not close to these zeros will
have large magnitude. The filter will “pass” these frequencies. The specific procedure to design such a filter is the
following.
3 The Fundamental Theorem of Algebra guarantees that H(z) factors in this way.
4 Previous ways of describing a filter have included the filter coefficients, the impulse response sequence, the frequency response function, and
the system function.
5 The gain does not affect the shape of the frequency response.
6 Since non-real zeros must occur in conjugate pairs, we must also place a conjugate zero on the unit circle, i.e. a zero whose angle is the negative

of the first.

The University of Michigan, All rights reserved 143

Laboratory 9. Filter Design, Modeling, and the z-Plane

1. Choose frequencies ω̂1 , ..., ω̂L at which the frequency response should contain a null or a dip.
2. Choose zeros ri = ρi ej ω̂i , i = 1, . . . , L, with ρi = 1 or ρi ≈ 1, depending upon whether a null or a dip is
desired at frequency ω̂. For each ω̂i 6= 0 choose also a zero rj that is the complex conjugate of ri . Let M be the
total number of zeros chosen.
3. Form the system function H(z) = K(1 − r1 z −1 ) × · · · × (1 − rM z −1 ), where K is a gain that we also choose.
4. Cross multiply the factors of H(z) found in the previous step so as to express H(z) as a polynomial whose
terms are powers of z −1 .
5. Identify the FIR filter coefficients {b0 , . . . , bM }, which are simply the coefficients of the polynomial found in
the previous step, as shown in equation (9.9).

9.2.3 IIR filters and rational system functions

We now consider IIR filters. The general time-domain input-output relationship for a causal IIR filter is given by the
difference equation
y[n] = b0 x[n] +b1 x[n − 1] + b2 x[n − 2] + · · · + bM x[n − M ]
+a1 y[n − 1] + a2 y[n − 2] + · · · + aN y[n − N ] . (9.12)
Here, we have the usual FIR filter coefficients, bk , but we also have another set of coefficients ak , which multiply past
values of the filter’s output. We will call the bk ’s the feedforward coefficients and the ak ’s the feedback coefficients7.
If the ak ’s are zero, then this filter reduces to a causal FIR filter.
As an example, consider the simple IIR filter with difference equation:
1
y[n] = x[n] + y[n − 1] (9.13)
2
What is the impulse response of this filter? If we assume8 that y[n] = 0 for n < 0, one can straightforwardly show
that the impulse response is  n
1
h[n] = , n ≥ 0, (9.14)
2
which is never zero for any positive n. (Note that the impulse response is generally not so simple to compute; this is
an unusual case where the impulse response can be obtained by inspection.) Thus, by introducing feedback terms into
our difference equation, we have produced a filter with an infinite impulse response, i.e., an IIR filter.
In general, computing the system function by taking the z-transform of the resulting infinite impulse may not be
trivial because of the required infinite sum, and also because it may be difficult to find the impulse response. However,
we can use the fact that H(z) = Y (z)/X(z) to determine the system function. To do this, we first collect the y[n]
terms on the left side of the equation and take the z-transform of the result.
y[n] − a1 y[n − 1] − · · · − aN y[n − N ] = b0 x[n] + b1 x[n − 1] + · · · + bM x[n − M ]
(9.15)
−1 −N −1 −M
Y (z) − a1 Y (z)z − · · · − aN Y (z)z = b0 X(z) + b1 X(z)z + · · · + bM X(z)z
(9.16)
Y (z)(1 − a1 z −1 − · · · − aN z −N ) = X(z)(b0 + b1 z −1 + · · · + bM z −M ) (9.17)
Y (z) b0 + b1 z −1 + · · · + bM z −M
H(z) = = (9.18)
X(z) 1 − a1 z −1 − · · · − aN z −N
7 In some texts (and in M ATLAB ), the feedback coefficients are defined as the negatives of the a coefficients given here. Because of this, you
k
should always be sure to check which convention is used.
8 This assumption is one of the “initial rest conditions” discussed in chapter 8 of DSP First.

144 The University of Michigan, All rights reserved

 9.2.3 IIR filters and rational system functions

Equation (9.18) shows the general form of the system function of an IIR filters. Since it is the ratio of two polynomials,
it is called a rational function9.
Just as we could factor the polynomial in equation (9.9), we can do the same with equation (9.18) to yield

(1 − r1 z −1 )(1 − r2 z −1 )(1 − r3 z −1 ) · · · (1 − rM z −1 )
H(z) = K . (9.19)
(1 − p1 z −1 )(1 − p2 z −2 )(1 − p3 z −1 ) · · · (1 − pN z −1 )

The roots of the polynomial in the numerator, {r1 , . . . , rM }, are again called the zeros of the system function. The
roots of the polynomial in the denominator, {p1 , . . . , pN } are called the poles of the system function. K is again a gain
factor that determines the overall amplitude of the system’s output. As before, the zeros are complex values where
H(z) goes to zero. The poles, on the other hand, are complex values where the denominator goes to zero and thus the
system function goes to infinity10 . Again, we typically assume that the filter coefficients bk and ak are real, so both
the poles and zeros of the system function must be either purely real or must appear in complex conjugate pairs.
Just as we could completely characterize an FIR filter by its gain and its zeros, we can completely characterize
an IIR filter by its gain, its zeros, and its poles. As in the FIR case, this is typically the most useful characterization
when designing IIR filters. As before, if the system function has zeros near the unit circle, then the filter magnitude
frequency response will be small at frequencies near the angles of these zeros. On the other hand, if there are poles
near the unit circle, then the magnitude frequency response will be be large at frequencies near the angles of these
poles. With FIR filters we could directly design filters to have nulls or dips at desired frequencies. Now, with IIR
filters, we can design peaks in the frequency response, as well as nulls. The specific procedure is the following.

1. Choose frequencies ω̂1 , ..., ω̂L at which the frequency response should contain a null, a dip, or a peak.

2. Choose zeros ri = ρi ej ω̂i at those frequencies at which a null or a dip should occur, with ρi = 1 or ρi ≈ 1, as
desired. For each such ω̂i 6= 0, choose also a zero rj that is the complex conjugate of ri . Let M be the total
number of zeros chosen.

3. Choose poles pi = ρi ej ω̂i at those frequencies at which a peak should occur, with ρi = 1 or ρi ≈ 1 as desired.
For each such ω̂i 6= 0 choose also a pole pj that is the complex conjugate of pi . Let N be the total number of
poles chosen.
−1 −1
4. Form the system function H(z) = K (1−r 1z )×···×(1−rM z )
(1−p1 z −1 )×···×(1−pN z −1 ) , where K is a gain that we also choose.

5. Cross multiply the factors of H(z) found in the previous step and express H(z) as the ratio of two polynomials
whose terms are powers of z −1 .

6. Identify the IIR filter coefficients {a0 , . . . , aN , b0 , . . . , bM }, which are simply the coefficients of the polynomials
found in the previous step, as shown in equation (9.18).

Poles and zeros at the origin and at infinity

Here, we have defined our system functions in terms of negative powers of z. This is because our general forms for
FIR and IIR filters are defined in terms of time delays, and multiplication of the z-transform of some signal X(z) by
z −1 is equivalent to a time delay of one sample. However, there are may be “hidden” poles and zeros when we express
a system function in this matter.
Consider first the system function for our FIR filter given by (9.9). If we try to evaluate this system function at
z = 0, we will immediately find that we are dividing by zero. Thus, there is actually a pole at the origin of this system
9 This is a generalization of the terminology that the ratio of two integers is called a rational number.
10 Technically, because of a division by zero, H(z) is undefined at the location of a pole. However, the magnitude of the system function becomes
very large in the neighborhood of a pole.

The University of Michigan, All rights reserved 145

Laboratory 9. Filter Design, Modeling, and the z-Plane

0.8

0.6

0.4

Imaginary Part
0.2
2
0

−0.2

−0.4

−0.6

−0.8

−1
−1 −0.5 0 0.5 1
Real Part

Figure 9.1: A pole-zero plot of an IIR filter.

function. To reveal such “hidden” poles and zeros, we express the system function in terms of positive powers of z.
M
To do so, we multiply by zzM , which yields

b0 z M + b1 z M−1 + b2 z M−2 + b3 z M−3 + · · · + bM

H(z) = . (9.20)
zM
By the Fundamental Theorem of Algebra, we know that the numerator polynomial has M roots, and thus the system
has M zeros. However, the denominator, z M , has M roots as well, all at z = 0. This means that our causal FIR
system function has M poles at the origin.
In some cases, like the previous example, we find extra poles at the origin. In other cases, we find extra zeros at
the origin. For example, the filter y[n] = y[n − 1] + x[n], has H(z) = 1−z1 −1 = z−1 z
, from which we see there is
one zero at the origin. In still other cases we find zeros at infinity . For example, the filter y[n] = x[n − 1], has
11

H(z) = z −1 = z1 , from which we see that there is a zero at infinity. In still other cases, we find combinations of
the previous cases. We will call poles and zeros located at the origin or at infinitytrivial poles and zeros because they
do not affect the system’s magnitude frequency response12 . In this laboratory, we will primarily be concerned with
nontrivial poles and zeros (those not at the origin or at infinity).
Note that there will always be the same number of poles and zeros in a linear time-invariant system, including both
trivial and nontrivial poles and zeros. The total number equals the M or N , whichever is larger. Such facts are useful
for checking to make sure that you have accounted for all poles and zeros in a system.
Note also that if one chooses filter coefficients such that the numerator and denominator contain an identical factor,
i.e. if ri = pj for some i and j, then these factors “cancel” each other, i.e. the filter is equivalent to a filter whose
system function has neither factor.

Pole-zero plots
It is often very useful to graphically display the locations of a system’s poles and zeros. The standard method for this
is the pole-zero plot. Figure 9.1 shows an example of a pole-zero plot. This is a two-dimensional plot of the z-plane
that shows the unit circle, the real and imaginary axes, and the position of the system’s poles and zeros. Zeros are
11 A zero at infinity means that |H(z)| → 0 as |z| → ∞. It can be shown that a causal filter can never have a pole at infinity.
12 Trivial poles and zeros do affect the phase (and thus the delay or time shift) of a system.

146 The University of Michigan, All rights reserved

 9.2.4 Graphical interpretation of the system function

typically marked with an ‘o’, while poles are indicated with an ‘x’. Sometimes, a location has multiple poles and
zeros. In this case, a number is marked next to that location to indicate how many poles or zeros exist there. Figure
9.1, for instance, shows four zeros (two conjugate pairs), two “trivial” poles at the origin, and one other conjugate pair
of poles. Recall that zeros and poles near the unit circle can be expected to have a strong influence on the magnitude
frequency response of the filter.

9.2.4 Graphical interpretation of the system function

If we take the magnitude of H(z), we can think of |H(z)| as defining a (strictly positive) surface over the z-plane for
which the height of the surface is given as a function of the complex number z. Figure 9.2 shows an example of just
such a surface. This system function has two zeros (which form a complex conjugate pair) and two poles at the origin.
Notice that the unit circle is outlined on the surface |H(z)|. The height of the surface at z = ej ω̂ (i.e., on the unit
circle) defines the magnitude of the frequency response, |H(ω̂)|, which is shown to the right of the surface.
On Figure 9.2, we can see two points where the surface |H(z)| goes to zero; these are the zeros of the system
function. Notice how the surface is “pulled down” in the vicinity of these zeros, as though it has been “tacked to
the ground” at the location of the zeros. Near the system’s zeros, the magnitude frequency response has a low point
because of the influence of the nearby zero. Also notice how the surface is “pushed up” at points far from the zeros;
this is another common characteristic of system function zeros. (Since the two poles in this figure are at the origin, they
have no effect on the system’s magnitude frequency response.) Thus, the magnitude frequency response has higher
gain at points far away from the zeros.
Figure 9.3 shows the surface |H(z)| as defined by a different system function. This system function has two poles
(which form a complex conjugate pair) and two zeros at the origin. Notice how the poles “push up” the surface near
them, like poles under a tent. The surface then typically “drapes” down away from the poles, getting lower at points
further from them. The magnitude frequency response here has a point of high gain in the vicinity of the poles. (Again,
the zeros in this system function are located at the origin, and thus do not affect the magnitude frequency response.)
Figure 9.4 shows the surface for a system function which has poles and zeros interacting on the surface. This
system function has four poles and four zeros. Notice the tendency of the poles and zeros to cancel the effects of one
another. If a pole and a zero coincide exactly, they will completely cancel. If, however, a pole and a zero are very
near one another but do not have exactly the same position, the z-plane surface must decrease in height from infinity
to zero quite rapidly. This behavior allows the design of filters with rapid transitions between high gain and low gain.

9.2.5 Poles and stability

System poles cause the system function to go to infinity at certain values of z because we are dividing by zero. On
the one hand, this can have the desirable effect of raising the magnitude frequency response at certain frequencies. On
the the other hand, this can have some undesirable side effects. One somewhat significant problem is introduced if we
have a pole outside the unit circle. Consider the following filter, for instance:

y[n] = x[n] + 2y[n − 1] (9.21)

This filter has a single pole at z = 2. What is this system’s impulse response? If the input is x[n] = δ[n] and y[n] = 0
for n < 0, then at n = 0, y[n] = 1. Then, every y[n] after that is equal to twice the value of y[n − 1]. The value
of this impulse response grows as time goes on! This system is unstable13. Unstable filters cause severe problems,
and so we wish to avoid them at all costs. As a general rule of thumb, you can keep your filters from being unstable
by keeping their poles strictly inside the unit circle. Note that the system’s zeros do not need to be inside the unit to
maintain stability.
13 Technically, it is bounded input, bounded output (BIBO) unstable because the input has a limited magnitude but the output does not.

The University of Michigan, All rights reserved 147

Laboratory 9. Filter Design, Modeling, and the z-Plane

Imaginary Part
(A) (B)
2
0
5
4 −1
−1 0 1
|H(z)|

3 Real Part
2 3
(C)
1

|H(ω)|
2
1
−1 0 1
0
−1 0
1 Imag(z) 0 1 2 3
Real(z) Discrete radian frequency, ω

Figure 9.2: (A) The z-plane surface defined by the system function H(z) = (1 − ejπ/4 z −1 )(1 − e−jπ/4 z −1 ). (B)
The corresponding pole-zero plot. (C) The corresponding magnitude frequency response.

Imaginary Part
(A) (B)
2
0

4 −1
3 −1 0 1
|H(z)|

Real Part
2
1 2 (C)
|H(ω)|

1
−1 1
0
0
−1 0
1 Imag(z) 0 1 2 3
Real(z) Discrete radian frequency, ω

1
Figure 9.3: (A) The z-plane surface defined by the system function H(z) = (1−0.8ejπ/2 z−1 )(1−0.8e−jπ/2 z −1 ) . (B) The

corresponding pole-zero plot. (C) The corresponding magnitude frequency response.

1
Imaginary Part

(A) (B)
0

4
−1
3 −1 0 1
|H(z)|

Real Part
2
1 2 (C)
|H(ω)|

1
−1 1
0
0
−1 0
1 Imag(z) 0 1 2 3
Real(z) Discrete radian frequency, ω

Figure 9.4: (A) The z-plane surface for a complicated system function with four poles and four zeros. (B) The
corresponding pole-zero plot. (C) The corresponding magnitude frequency response.

148 The University of Michigan, All rights reserved

 9.2.6 Filter design using manual pole-zero placement

9.2.6 Filter design using manual pole-zero placement

In this laboratory, we will explore a method of filter design in which we place poles and zeros on the z-plane in order to
match some target frequency response. You will be using a MATLAB graphical user interface to do this. The interface
allows you to place, delete, and move poles and zeros around the z-plane. The frequency response will be displayed
in another figure and will change dynamically as you move poles and zeros. To keep the filter’s coefficients real, you
will design by placing a pair of poles and zeros on the z-plane simultaneously.
The approach for this method depends somewhat on the type of filter that we wish to design. If we want an FIR
filter (i.e., a filter that has no poles), we need to use zeros to “pin down” the frequency response where it is low, and
allow the frequency response to be pushed upwards in regions where there are no zeros. Note that if we put a zero
right on the unit circle, we introduce null in the frequency response at that point. Conversely, the closer to the origin
that we place a zero, the less effect it will have on the frequency response (since it will begin to affect all points on
the unit circle roughly equally). You might use the example of the running average filter and bandpass filters (given in
Chapter 7 of DSP First) as a prototype of how to use zeros to design FIR filters using zero placement.
If we wish to design an IIR filter (with both poles and zeros), it usually makes sense to start with the poles since
they typically affect the frequency response to a greater extent. If the frequency response that we are trying to match
has peaks on it, this suggests that we should place a pole somewhere near that peak (inside the unit circle). Then,
use zeros to try to pull down the frequency response where it is too high. As with zeros, poles near the origin have
relatively little effect on the system’s filter response.
Regardless of which type of filter we are designing, there are a couple of methodological points that should be
mentioned. First, moving a pole or zero affects the frequency response of the entire system. This means that we
cannot simply optimize the position of each pole-pair and zero-pair individually and expect to have a system which is
optimized overall. Instead, after adjusting the position of any pole-pair or zero-pair, we generally need to move many
of the remaining pairs to compensate for the changes. This means that filter design using manual pole-zero placement
is fundamentally an iterative design process.
Additionally, it is important that you consider the filter’s gain. Often we cannot adjust the overall magnitude of
the frequency response using just poles and zeros. Thus, to match the frequency response properly, you may need to
adjust the filter’s gain up or down. The pole-zero design interface that you will use in this Lab includes an edit box
where you can change the gain parameter. Alternately, by dragging the frequency response curve, you can change the
gain graphically. A related idea is that of spectral slope. By having a pair of poles or zeros inside the unit circle and
near the real axis, we can adjust the overall “tilt” of the frequency response. As we move the pair to the right and left
on the z-plane, we can adjust the slope of the system’s frequency response up and down.
Note that there are automatic filter design methods which do not require manual placement of poles and zeros. In
Section 9.2.8 we discuss one such method.

9.2.7 Design of Standard Filter Types

Many of the filters that we wish to design and use belong to one of four standard types: lowpass, highpass, bandpass,
or bandstop. These filters are characterized by a passband (a band of frequencies which are relatively unaltered by
the filter) and a stopband (a band of frequencies which are significantly attenuated, or decreased in amplitude, by the
filter). The locations of the passband and stopband are what characterize the different filter types. For instance, a
lowpass filter has a passband which contains low frequencies and a stopband which contains high frequencies, while
a bandpass filter has a single passband that is surrounded by two stopband regions. Between the passband and the
stopband is a transition band in which the filter’s frequency response changes from high to low. The location of the
transition band in frequency determines the filter’s cutoff frequency. In this lab, we will specify the cutoff frequency
by using the two frequencies that bound the transition band.
When designing these types of filter, there are a number of different design goals that we may attempt to achieve.
For instance, we may wish to have a very flat frequency response in the passband of the filter. Unfortunately, it can

The University of Michigan, All rights reserved 149

Laboratory 9. Filter Design, Modeling, and the z-Plane

Transition
Band
Passband Stopband

Passband Stopband

w )|
Ripple Attenuation
|H(

Frequency Cutoff Frequen

Figure 9.5: An illustration of the various bands of a lowpass filter.

be difficult to achieve a flat frequency response over some frequency region. Instead, the frequency response usually
varies somewhat over that region; this variation is called ripple. Thus, one of design criteria may be to minimize
passband ripple. In this lab, we define the passband ripple as the (positive) decibel value of the ratio between the
maximum and minimum filter gains in the passband.
Another common goal is to try to minimize the gain in the stopband of the filter relative to the gain in the passband.
That is, we wish to maximize the stopband attenuation. In this lab, we define the stopband attenuation as the decibel
ratio between the maximum filter gain in the passband and the maximum filter gain in the stopband.
Figure 9.5 shows the passband, transition band, and stopband for a lowpass filter. The figure also illustrates the
the passband ripple and stopband attenuation. In Problem 2 of this assignment, you will design a lowpass filter to
maximize stopband attenuation.

9.2.8 Modeling Vowel Production

In Lab 8, we discussed some of the properties of vowel production in speech, but we did not examine the mechanisms
behind vowel production. In order to gain a better understanding of vowel production and how we can model it using
discrete-time filters, we need to introduce some theory.
Speech production is primarily governed by the larynx (or voice box) and the vocal tract. Figure 9.6 shows a
diagram of the larynx and vocal tract. When we speak a vowel, the lungs push a stream of air through the larynx and
the vocal folds (commonly referred to as the vocal chords). Given the appropriate muscular tension, this stream of air
causes the vocal folds to vibrate14 . This in turn creates a nearly periodic fluctuation in air pressure passing through
the larynx. The fundamental frequency of vocal fold vibration is typically around 100 Hz for males and 200 Hz for
females.
This fluctuating air stream then passes through the vocal tract, which is the airway leading from the larynx and
through the mouth to the lips. The positions of the tongue, lips, and jaw serve to shape the vocal tract, with different
positions creating different vowel sounds. The different sounds are produced as the vocal tract shapes the spectrum
of the pressure signal coming from the larynx. Depending upon the vocal tract configuration, different frequencies of
the spectrum are emphasized; from Lab 8, we know these frequencies as formants. When whispering a vowel, the
lungs push air through the larynx, but the vocal folds do not vibrate. In this case, the air pressure fluctuation is quite
14 In fact, during normal production the vocal folds open and close completely on each cycle of the vibration.

150 The University of Michigan, All rights reserved

 9.2.8 Modeling Vowel Production

Figure 9.6: A diagram showing the larynx and vocal tract.

x[n]
Glottal Vocal Tract y[n]
Source Filter

Figure 9.7: A block diagram of the source-filter model of speech production.

noise-like and generally is not periodic. Nevertheless, the tongue, lips and jaw shape the vocal tract just as before to
make the various vowel sounds.
Note that the above description is only accurate for vowels and so-called voiced consonants like “m” and “n.”
Most consonant sounds are produced using the tongue, lips, and teeth rather than the vocal cords. We will not consider
consonants in this lab.
It is traditional to model speech production using a source-filter model. Figure 9.7 shows a block diagram of the
source-filter model. The first block is the glottal source, which takes as input a fundamental frequency and produces
a periodic signal (the glottal source signal) with the given fundamental frequency. The signal produced is typically
modeled as a periodic pulse train. To a first approximation, we can assume that spectrum of this pulse train is composed
of equal amplitude harmonics. The glottal source signal is meant to be analogous to the signal formed by the air
pressure fluctuations produced by the vibrating vocal cords. Note that to model whispering, the glottal source signal
can be modeled using random noise rather than a pulse train.
The second block of the source-filter model is the vocal tract filter. This is a discrete-time filter that mimics the
spectrum-shaping properties of the vocal tract. Since we are assuming a source signal with equal-amplitude harmonics,
the vocal tract filter provides the spectral envelope for our output signal. That is, when we filter the source signal with
fundamental frequency ω̂0 radians per sample, the k th harmonic of the output signal will have an amplitude equal to
the filter’s magnitude frequency response evaluated at k ω̂0 . This is illustrated in Figure 9.8 which shows a particular
example. The magnitude spectrum of the glottal source signal is shown on top, the magnitude frequency response of
the vocal tract filter is shown in the center, and the magnitude spectrum of the output signal, which is the signal that
models the specific vowel signal. One may clearly see that, as desired, the envelope of the spectrum of the vowel

The University of Michigan, All rights reserved 151

Laboratory 9. Filter Design, Modeling, and the z-Plane

X(ω)
0.5

0
0 0.5 1 1.5 2 2.5 3

1
H(ω)

0.5

0
0 0.5 1 1.5 2 2.5 3

1
Y(ω)

0.5

0
0 0.5 1 1.5 2 2.5 3
Frequency (radians per sample)

Figure 9.8: A plot of the magnitude spectrum of a glottal source signal, the frequency response of a vocal tract filter,
and the magnitude spectrum of the output signal.

signal model matches the spectrum of the vocal tract filter.

In Problem 3 of this assignment, we will make such source-filter models for particular vowel signals, by measuring
the spectrum of the vowel signal and designing an IIR vocal tract filter whose frequency response approximates this
spectrum.
Typically, our vocal tract filter can have relatively few filter coefficients (i.e., approximately 10-20 coefficients).
Further, the acoustics of the vocal tract suggest that this filter should be IIR. Often, the vocal tract is modeled using
an all-pole filter which has no nontrivial zeros. This is because an acoustic passageway like the vocal tract primarily
affects a sound through resonances. A resonance is a part of a system that tends to vibrate at a certain resonant
frequency, thus amplifying that frequency in signals passed through them. The feedback form of an IIR filter is a
direct implementation of resonance; this is how IIR filters are able to produce high gain at certain frequencies. Using
this simple model of speech production, it is possible to synthesize artificial vowels.

All-pole analysis and vowel classification

In the last lab, we explored some features for vowel classification that were based on two measures of spectral energy
in a vowel signal. The development of the source-filter model, however, suggests an acoustically motivated feature
for vowel classification. If we assume that the vocal tract can be modeled with a low-order discrete-time filter, then
the vocal tract filter captures all of the relevant information about which vowel has been produced. Variations such as
fundamental frequency and type of vowel production (i.e., voiced or whispered) are restricted to the glottal source and
can be neglected. Using samples of the frequency response of the vocal tract filter as features for vowel classification
has been shown to produce good classification results.
Fortunately, there are nice mathematical tools for deriving all-pole filter models automatically from a time-domain
waveform. These tools, fit the spectrum of a time-domain signal with poles in a least-squares sense. Note that these
tools work directly with the time-domain waveform rather than its spectrum; typically, they return the resulting ak
feedback coefficients for a filter with those poles, rather than the locations of the poles themselves. We will explore

152 The University of Michigan, All rights reserved

 9.3 Some M ATLAB commands for this lab

these tools for all-pole analysis in the laboratory assignment, and we will compare classification performance using
features based on these models to the performance we achieved with our other feature sets from Lab 8.

9.3 Some M ATLAB commands for this lab

• Calculating the frequency response of IIR Filters: Previously we have used freqz to compute the frequency
response of FIR filters. We can use the same command to compute the frequency response of an IIR filter. If
our filter is defined by feedforward coefficients bk stored in a vector B and feedback coefficients ak stored in a
vector A, we compute the frequency response at 256 points using the command:

>> [H,w] = freqz(B,A,256);

As with FIR filters, M ATLAB’s convention for the bk coefficients is B(1)= b0 , B(2)= b1 , . . ., B(M+1)= bM .
M ATLAB’s convention for the ak coefficients is A(1)= 1, A(2)= −a1 , . . ., A(N+1)= −aN . Both bk and
ak are given as defined in equation 9.18. H contains the frequency response and w contains the correspond-
ing discrete-time frequencies. Alternatively, we can compute the frequency response only at a desired set of
frequencies. For example, the command

>> [H,w] = freqz(B,A,[pi/4, pi/2, 3*pi/4]);

returns the frequency response of the filter at the frequencies π/4, π/2, and 3π/4.
• Pole-Zero Place 3-D: In this laboratory, we will primarily be exploring filter design using manual pole-zero
placement. To help us do this, we will be using a M ATLAB graphical user interface (GUI) called Pole-Zero
Place 3-D. Pole-Zero Place 3-D allows you to place, move, and delete poles and zeros on the z-plane, and
provides immediate feedback by displaying the filter’s frequency response and the |H(z)| surface. Additionally,
it calculates some useful statistics for assessing the quality of a particular filter design.
To run this program you need to download two different files: pole_zero_place3d.m
and pole_zero_place3d.fig. To begin Pole-Zero Place 3-D, simply execute15 the command

>> pole_zero_place3d;

Once the program starts, the GUI window shown in Figure 9.9 will appear. The axis in the upper left of the
window shows a portion of the z-plane with the unit circle. In the lower left is an axis that displays the frequency
response of the system. In the lower right is a 3-D axis which displays a 3-D graph of the |H(z)| surface16 .
The interface allows you to do a wide variety of things.

1. To add a poles or zeros to the z-plane, click the Add Zeros or Add Poles button and then click on the z-plane
plot in the upper left of the GUI. The state of the Place pair checkbox determines whether a single (real)
pole or zero is added, or whether a conjugate pair is added.
Note that the program also adds the hidden poles and zeros that accompany nontrivial poles and zeros.
Specifically, for each zero that is added a pole is added at the origin, or if there is already at least one
zero at the origin, instead of adding a pole at the origin, one zero at the origin is removed, i.e. cancelled.
Moreover, for each pole that is added, a zero is added at the origin, or if there are already poles at the
origin, one pole is removed, i.e. cancelled. The system does not allow one to place zeros at infinity, and it
can be shown that zeros at infinity will not be induced by any other choices of poles or zeros.
15 This program was designed to run using Windows systems running M ATLAB 6 or higher; it will not work with previous versions of M ATLAB .

It should work with Unix operating systems running M ATLAB 6, but this has not been tested.
16 This surface plot requires significant computation, and thus it can be toggled on and off using the View 3-D checkbox in the upper right.

The University of Michigan, All rights reserved 153

Laboratory 9. Filter Design, Modeling, and the z-Plane

1
0.8
0.6
0.4
0.2
Imag(z) 2
0
−0.2
−0.4
−0.6
−0.8
−1

−1 −0.5 0 0.5 1
Real(z)

2
3

|H(z)|
2
1.5
|H(ω)|

1 1
0 0.5
−1 0
−0.5
0 −0.5
0.5 0.5
1 −1
0 0.5 1 1.5 2 2.5 3 Imag(z)
ω (Discrete radian frequency) Real(z)

Figure 9.9: The GUI window for Pole-Zero Place 3-D.

2. To move a real pole or zero (or a conjugate pair of complex poles or zeros), you must first select the
pole/zero by clicking on one member of the pair. Then, you can drag it around the z-plane, use the arrow
keys to move it, or move it to a particular location by inputting the magnitude and angle (in radians) in the
Magnitude and Angle edit boxes.
3. To delete a pole or zero (or pair), select it and hit the Delete Poles/Zeros button. Again, the system will
maintain an equal number of poles and zeros by also removing poles or zeros from the origin as necessary.
This may also have the effect of no longer cancelling other poles and zeros, and thus the total number of
poles and zeros that appear at the origin will change.
4. To change the filter’s gain, you can either use the Filter Gain edit box or you can click-and-drag the blue
frequency response curve in the lower left.
5. To toggle between linear amplitude and decibel displays in the lower two plots, select the desired radio
button above the Filter Gain edit box.
6. To rotate the 3-D |H(z)| plot, simply click-and-drag the axes in the lower right of the GUI. To enable or
disable the 3-D plot, toggle the View 3-D checkbox in the upper right of the GUI
7. To begin with an initial filter configuration defined by the feedforward coefficients, B, and the feedback
coefficients, A, start the program with the command
>> pole_zero_place3d(B,A);
This is useful if you wish to start continue working on a design that you had previously saved. You may
set either of these parameters to empty ([]) if you do not wish to specify the filter coefficients.
8. To print the GUI window, you can either use the Copy to Clipboard button to copy an image of the figure
into the clipboard17, or you can print the figure using the Print GUI button.
9. To save your current design, use the Export Filter Coefs button. The feedforward and feedback coefficients
will be stored in the variables B_pz and A_pz, respectively.
17 Windows operating systems only

154 The University of Michigan, All rights reserved

 9.3 Some M ATLAB commands for this lab

10. To hear your filter’s response to periodic signal with equal-amplitude harmonics, press the Play Sound
button. This is particularly useful when using the GUI to design vocal tract filters for vowel synthesis.

• Pole-Zero Place 3-D – Filter Matching Mode: In “filter matching mode,” you specify samples of a desired
transfer function at harmonically related frequencies and try to match that transfer function. The GUI plots a
red curve or stem plot along with the frequency response function; this is the response we wish to match. Two
edit boxes labeled Linear Matching Error and Decibel Matching Error indicate how closely your filter matches
the desired frequency response. The matching error values are computed as the RMS error between the desired
frequency response and your filter design in both linear amplitude and in decibels.
To start the GUI in this mode, use the following command:

>> pole_zero_place3d(B,A,filter_gains,fund_frq);

The filter_gains are the values of the desired filter frequency response at harmonically related frequencies,
and fund_frq is the fundamental frequency (in radians per sample) of the harmonic series of frequencies at
which the filter_gains are defined.
• Pole-Zero Place 3-D – Lowpass Design Mode: In “lowpass design mode,” you specify the maximum frequency
of the passband and the minimum frequency of the stopband (both in radians per sample). To start the GUI in
this mode, use the following command:

>> pole_zero_place3d(B,A,[pass_max,stop_min]);

In this mode, the GUI computes the passband ripple and the stopband attenuation of your lowpass filter design.
You can use these measures to evaluate your filter design. The figure in the lower right also displays the passband
and stopband of the filter, with appropriate minima and maxima.
• Converting between filter coefficients and zeros-poles: Given a set of filter coefficients, we often need to
determine the set of poles and zeros defined by those coefficients. Similarly, we often need to take a set of poles
and zeros and compute the corresponding filter coefficients. There are two M ATLAB commands that help us do
this. First, if we have our filter coefficients stored in the vectors B and A, we compute the poles and zeros using
the commands

>> zeros = roots(B);

>> poles = roots(A);

This is because the system zeros are simply the roots of the numerator polynomial whose coefficients are the
numbers in B, while the system zeros are simply the roots of the denominator polynomial whose coefficients are
the numbers in A. We can see this in equation 9.18. To convert back, use the commands

>> B = poly(zeros);
>> A = poly(poles);

Note that we loose the filter’s gain coefficient, K, with both of these conversions.
• Generating pole-zero plots: Frequently, we’d like to use M ATLAB to make a pole-zero plot for a filter. If our
filter is defined by feedforward coefficients B and feedback coefficients A (both row vectors), we can generate a
pole-zero plot using the command:

>> zplane(B,A);

The University of Michigan, All rights reserved 155

Laboratory 9. Filter Design, Modeling, and the z-Plane

Alternately, if we have a list of poles, p, and a list of zeros, z, (both column vectors) we can use the following
command:

>> zplane(z,p);

An example of a pole-zero plot resulting from this command is shown in Figure 9.1.

• Automatic all-pole modeling: Using the M ATLAB command aryule, we can compute an all-pole filter model
for a discrete-time signal. That is aryule automatically finds an all-pole filter whose magnitude frequency
response that in some sense matches the magnitude frequency response of the signal. If signal is given by
signal, the command

>> A = aryule(signal,N);

returns the filter feedback coefficients ak as a vector A. The parameter N indicates how many poles we wish
to use in our filter model. Once we have A, we can compute the filter’s frequency response at 256 points using
freqz as

>> [H,w] = freqz(1,A,256);

9.4 Demonstrations in the Lab Section

• The z-transform, system functions, and IIR filters

• The system function “surface,” |H(z)|

• Using Pole-Zero Place 3-D for filter design

• Vocal tract modeling

9.5 Laboratory Assignment

1. (Fit an FIR filter’s frequency response.) Download the files pole_zero_place3d.m,
pole_zero_place3d.fig, and lab9_data.mat. In this problem, you will get familiar with the Pole-
Zero Place 3-D program for filter design using pole-zero placement.
In lab9_data, the variable FIR_fr contains samples of the frequency response for a simple FIR filter with
six zeros. Execute Pole-Zero Place 3-D using the command

>> pole_zero_place3d([],[],FIR_fr,2*pi/8192);

Use the GUI to find an FIR filter with six nontrivial zeros that matches the frequency response of the original
filter. You should be able to get the linear matching error to be less than 0.1. (Hint: The original filter had all six
of its zeros inside the unit circle, so yours should as well.)

• [6] Include the GUI window with your matching filter in your report. In this and the following problems,
make sure that it is possible to read the filter evaluation scores on your printout. (Note: The easiest way to
include the GUI window is to use the Copy to Clipboard button on a Windows machine. After hitting the
button, wait for the GUI to flash white and then paste the result into your report.)

156 The University of Michigan, All rights reserved

 9.5 Laboratory Assignment

• [2] What are the filter coefficients bk and ak for your filter?
• [2] Where are the zeros on the z-plane? Give your answers in rectangular form.

2. (Design a lowpass filter.) In this problem, we will use the “lowpass design mode” of Pole-Zero Place 3-D
to design some lowpass filters, as described in Section 9.2.7. For the various parts of this problem, use the
command

>> pole_zero_place3d([],[],2*pi*[1500 2000]/8192);

This sets the filter transition band to 1500 Hz to 2000 Hz if we assume a sampling rate of 8192 samples per
second.

(a) (Design an FIR lowpass filter to maximize stopband attenuation.) First, let’s see what we can do with just
zeros (that is, with FIR filters). Using only six nontrivial zeros (i.e., three zero pairs), design a lowpass
filter with a stopband attenuation of at least 30 dB. (Remember, we want our stopband attenuation to be as
large as possible). For now, take note of your filter’s passband ripple, but don’t worry about minimizing it.
• [8] Include the GUI window with your matching filter in your report.
• [2] What are the filter coefficients bk and ak for your filter?
• [2] Where are the zeros on the z-plane? Give your answers in rectangular form.
Food for thought: Using just zeros, try to find a way to minimize the passband ripple. What does this do to
your stopband attenuation? Try this with more zeros, but don’t use any poles.
(b) (Design an IIR lowpass filter to maximize stopband attenuation.) There are two primary benefits to the
use of IIR filters. First, it is very easy to get very high gain at certain frequencies. This lets us design a
lowpass filter with very high stopband attenuation. Using a single pair of nontrivial poles, design a lowpass
filter that has a stopband attenuation greater than 60 dB. Use the same transition band as in the previous
problem. Again, you should take note of the passband ripple, but don’t worry about minimizing it.
• [4] Include the GUI window with your matching filter in your report.
• [2] What are the filter coefficients bk and ak for your filter?
• [2] Where are the poles on the z-plane? Give your answers in rectangular form.
(c) (Design an IIR lowpass filter for both high stobpand attenuation and low ripple.) The second benefit of
IIR filters is the ability to achieve fast transitions between high gain and low gain. Among other things,
this allows us to transition between the passband and stopband more quickly, which in turn allows us to
achieve relatively high stopband attenuation with low passband ripple.
Once again using the same transition band, design a lowpass filter with a passband ripple of less than 2
dB and a stopband attenuation of at least 20dB. You may use as many poles and zeros as you wish, but it
is possible to meet these criteria with only two poles and four zeros. (Hint: use decibel mode to help you
increase the stopband attenuation, and linear mode to help you decrease the passband ripple.)
• [12] Include the GUI window with your matching filter in your report.
• [2] What are the filter coefficients bk and ak for your filter?
• [2] Where are the poles and zeros on the z-plane? Give your answers in rectangular form.
Food for thought: For a more interesting challenge, design a lowpass filter with a passband ripple of less
than 1 dB and a stopband attenuation of 60 dB. This can be done with six poles and six zeros, but you
might want to use more than this.

The University of Michigan, All rights reserved 157

Laboratory 9. Filter Design, Modeling, and the z-Plane

3. (Matching a vowel signal’s spectrum using pole-zero placement.) lab9_data.mat contains the variable
vowel2, which is a short vowel signal sampled at 8192 Hz. In this problem we will find a filter model of the
vocal tract used to produce this vowel.

(a) First, let’s examine the signal itself.

• [2] Use soundsc to listen to vowel2. What vowel does this signal represent?
• [4] As a measure of the spectrum, plot the magnitude DFT coefficients for this signal in decibels
versus frequency in Hertz. Use only the first half of the DFT coefficients, where the maximum value
on the x-axis is one half of the sampling rate.
• [2] From this plot, estimate the fundamental frequency of the vowel signal in Hertz.
(b) (Design a vocal tract filter with both poles and zeros.)
lab9_data.mat contains the variables vowel2_amps, which contains the amplitudes of the harmon-
ics that make up vowel2. We’ll use the amplitudes to find a vocal tract filter for this vowel. Start Pole-Zero
Place 3-D using the command
>> pole_zero_place3d([],[],vowel2_amps,2*pi*frq/8192);
where frq is the fundamental frequency (in Hz) that you estimated. (Hint: The red stems should go
all the way across the frequency response plot. Also, there should be one stem for each element of
vowel2_amps. If not, you have probably estimated your fundamental frequency incorrectly.)
Set the GUI to “decibel plot” and find a filter with six nontrivial poles and six nontrivial zeros that makes
the decibel matching error as small as possible. You should be able to get the decibel error below 4.2. (It
is possible to achieve a decibel error of 3.65.)
• [8] Include the GUI window with your matching filter in your report.
• [2] What are the filter coefficients bk and ak for your filter?
• [2] Where are the poles and zeros on the z-plane? Give your answers in rectangular form.
(c) (Design a vocal tract filter with only poles and zeros.) Now, repeat the above for a filter with 10 poles
nontrivial and no nontrivial zeros (except for those at the origin). You should be able to achieve a decibel
matching error below 2.6. (It is possible to achieve a decibel error of 2.1.)
• [8] Include the GUI window with your matching filter in your report.
• [2] What are the filter coefficients bk and ak for your filter?
• [2] Where are the poles on the z-plane? Give your answers in rectangular form.
If you are working on a computer with audio capability, you should use the Play Sound button to listen to a
synthesis of the vowel signal. Note that the all-pole model produces less error with fewer total coefficients.
This suggests that all-pole filters are more appropriate for vocal tract modeling.

4. (All-pole modeling and classification) In this problem, we’ll look at an automatically generated all-pole model
of vowel2, and we’ll see how we can use this model to help us improve our vowel classifier performance from
Lab 8.

(a) (Automatically generate an all-pole filter model.) Use aryule to compute an all-pole model for vowel2
with 10 poles.
• [2] What are the resulting feedback coefficients?
• [2] Make a pole-zero plot of the resulting filter.

158 The University of Michigan, All rights reserved

 9.5 Laboratory Assignment

• [4] Use freqz to plot the frequency response of this filter and the frequency response of the filter
you found in Problem 3c in two subplots of the same figure. Display the two frequency responses
in decibels. (Note: The two frequency responses in decibels may be offset by some constant, which
corresponds to a scaling of the original spectrum.)
(b) (Construct and evaluate the vowel classifier.) Here, we’ll consider 16 samples of the frequency response
of an all-pole filter to be a potential feature vector for vowel classification. lab9_data.mat contains
five matrices which contain all-pole feature vectors for the same vowel instances we examined in Lab 8.
They are oo_ap, oh_ap, ah_ap, ae_ap, and ee_ap.
• [5] As you did in Lab 8, compute the mean all-pole feature vectors for each of the five vowel classes.
Make sure you label which mean vector belongs with which class.
• [5] Combine the above vectors into a matrix of mean feature vectors. Then, use the function (from
Lab 8)
confusion_matrix.m to calculate the confusion matrix for the all-pole features. As you did in
Lab 8, use the vowel order “oo,” “oh,” “ah,” “ae,” and “ee”. (If you did not complete your function
confusion_matrix in Lab 8, use the compiled function confusion_matrix_demo.dll for
this problem.)
• [4] Compare this confusion matrix with the two confusion matrices that you computed in Lab 8. Is
the performance of the classifier better with this feature class? Note any similarities between the three
confusion matrices.

5. On the front page of your report, please provide an estimate of the average amount of time spent outside of lab
by each member of the group.

The University of Michigan, All rights reserved 159