Motorola Digital Signal Processing

_.
rJ1
~
=
~
Digital S·
~ AND THE MICR
_.
~
~
'J:J
'J:J
=
~
»
z
0
-i
I
m
$:
-
0
:0
0
0
0 T
Z
-i
:0
0
r
r
m
:0 •
Digital
Signal
Processing
and
The Microcontroller
Mark McQuilken
Microprocessor and Memory Technologies Group
Motorola, Inc.
Austin, Texas
James P. LeBlanc
Electrical Engineering Department
Cornell University
Ithaca, New York
All rights reserved.
Reproduction or translation of any part of this

work beyond that permitted by written permission
of Motorola, Inc. is unlawful.
ACKNOWLEDGMENTS
The Motorola Advanced Microcontroller Division gratefully acknowledges the con-

tributions of the following people in the creation of this book. Many thanks to:
Scott Anderson and Tom Marischen for their commitment to "doing the right thing"
regardless of division boundaries.
Arie Brish for reviewing many of the chapters, even over the holidays, for technical
accuracy. His responsiveness and willingness to "dig-in" and help resolve production
schedule crises made delivery of this work possible. Arie, should we have reserved this
acknowledgement for the next edition?
Francis Christian for sharing his field experience and insights, and for technically
reviewing some of the material.
James LeBlanc, former Motorolan, who left the "rat race" to obtain his PhD. and further
his already extensive knowledge of DSP. Despite a myriad of personal pressures, he
still managed to fulfill his commitment to Motorola and provide us with high quality
material for chapters four through seven (and then some!). Without his contributions,
this book would not have been possible. Now if we could just get him to stop listening
to "The Buzzcocks".
Steve Marsh, who mandated that this book be written, left for "greener pastures", and
thenstood back and watched the fun! Despite this, he and his wife Jackie most graciously
edited, proofread, and critiqued various versions of the manuscript.
Mark McQuilken, who deserves much more than a thank you for the enormous amount
of effort and time spent, much of it at his family'S expense, in authoring chapters one
through three, editing and reviewing all material, preparing figures and drawings, and
coordinating the whole project. His never-ending patience was much appreciated. And,
special thanks go to Mark's wife, Beth, for so graciously putting up with it all, and for
her help in editing some of the material.
Judy Racino, who got saddled, during the last two weeks ofthe production cycle, with
the responsibility to complete this work on-time. Her honest, clear-minded approach to
problem solving avoided many potential obstacles. She willingly spent many long hours
and did "whatever it took" (e.g., writing, editing, proofreading, going for sandwiches,
etc.) to deliver this book to the printer on-time.
Tanya Tussing, who spent untold time and effort in editing and reviewing, including the
removal of thousands of prepositional phrases, to ensure technical accuracy and under-
standability.
John Waite, who provided the glossary of technical terms.
Brian Wilkie for his negotiating skills, financial and moral support of this project.
A "tip of the hat" to Tim Ahrens, Robert Chretien, Mark Glenewinkel, Mark Johnson,
and Craig Haller for helpful edits, suggestions, and moral support.
Special thanks go to Staats Falkenberg & Partners for their assistance with many of the
figures and for designing the cover art: especially to Jack Robberson for his willingness
to always help in tight situations and for being so very organized.
Admundson and Associates for providing the required last minute resources, and
especially to (Iast-but-not-least) Wendy Heath for her editing!layout skills, creativity,
shortened honeymoon, and for sticking with us to the end (well, at least the end of this
edition). Wendy, what time did you finish last night?!
PREFACE
Microcontrollers are signal processors. This may come as a surprise to some, but in the
strictest interpretation it's true: microcontrollers (rviCUs) are almost always used to
evaluate inputs (signals) from sensors and the like, then manipulate (process) the digital
form of those inputs and produce some relevant output. If you're like most of us
however, we don't think of MCUs when we think of digital signal processing (DSP).
That's probably because of two general misconceptiom: I) DSP may only be accom-
plished with ultra-fast processors that have been designed for the purposes of doing DSP,
and 2) you have to be a brilliant person just to understand DSP, let alone use it. This
guidebook will show that both of these are truly misconceptions.
The smaller geometries achieved by the semiconductor industry has made possi ble more
highly integrated microcontrollers; practically a full computer system on one chip.
Because of these advances in technology, digital signal processing (DSP) functions are
being integrated as part of the microcontroller. As a result of high integration levels and
diversity offunctions now available on a single chip, electrical engineers are becoming
exposed to areas that were previously covered by specialists. Digital signal processing
is a prime example.
This book is intended to help bridge the knowledge gap between the two technologies:
digital signal processors and microcontrollers. The target audience is design engineers,
programmers, technicians, and students who deal with embedded control applications,
but need to have a working knowledge of DSP concepts. Most digital signal processing
books are highly theoretical and focus on the mathematics behind DSP. The style of this
book is a little different - it is written in a more casual tone. We believe this type of
presentation will help the reader enjoy and understand the topics discussed.
Chapter Descriptions
Chapter 1 introduces the microcontroller in general, and in particular, discusses the traits,
and extols the space-and-cost-saving virtues of the Motorola MC68HC705C8 (C8). The
C8 is then used in a proto-typical application, a household thermostat, to illustrate an
MCU's usage in the "real-world". As part of this illustration, a noise problem is
hypothetically encountered which is resolved by having the MCU average the readings
coming from the Analog-to-Digital Converter (ADC). This averaging process is then
graphically developed into a "lattice diagram", a graphic device which displays the
process of traditional digital filtering. The rudiments of Analog-to-Digital conversion
and specific ADC types are then briefly presented.
Chapter 2 discusses the strengths and weaknesses of "doing things digitally" with MCV s.
The digital domain strengths are shown two ways: a) through the development of an
MCV-based magnetic tape noise reduction system, and, b) through the implementation
of an MCV-based "digital delay". Although both the noise reduction and delay units
have typically and traditionally been implemented using analog circuits, the digital
MCV versions are shown to be superior because of the inherent strengths of the digital
domain. The details, of each application, even including assembly code segments are
shown. This chapter is closed with a brief overview of the disadvantages of working in
the digital domain.
Chapter 3 shows the benefits and practical details of remapping our "oscilloscope-view"
of signals (classically referred to as the "time-domain") to a frequency response view
(classically referred to as the "frequency domain") for use by an MCV. This remapping
is accomplished by using the Fourier transform (FT). The FT is presented as a set of
paired signals (transform pairs; time domain and frequency domain versions of a given
signal) and a set of rules, including FT properties, rather than as mathematical manipu-
lations of the messiest kind. The transform pairs and rules are then used to review the
thermostat example of Chapter 1 as a digital filter. In addition, the Nyquist criteria for
digitally sampled systems and signal aliasing are "proven" by intuitively applying the
FT. The chapter is closed with a brief list of recommendations for continuing to learn
Digital Signal Processing.
Chapter 4 discusses the fundamental concepts of signal processing by reviewing some

of the basic building blocks required to apply signal processing: the characteristics of
sinusoids and their spectra, the delta function, real and complex modulation, imaginary
and complex numbers, Euler's relation, and phasors. These basic ideas are then applied
to produce a heuristic presentation of the Fourier transform and its basic properties which
are presented, unlike Chapter 3, with some of the mathematics of the FT. The chapter
is closed with a discussion of FT transform pairs - all necessary fare to understand the
contents of chapter 5.
Chapter 5 presents elements of signal filters, both ideal and "real-world" implementa-
tions, and transfer functions. Topics include: using the time domain to describe filters,
signal convolution, and the duality of the FT. The specific characteristics of an ideal
filter are presented with an eye toward defining the critical parts of a filter's response.
Since ideal filters, like perfect circles, are not realized in the real-world, the topic shifts
from ideal to real-world filters and the considerations which must be made, like
"windowing", to understand and develop real filters. The chapter ends with a review of
typical signal classifications band-widths one is likely to encounter when developing
filters.
Chapter 6 begins the discussion of digital signal processing with an emphasis, as with
previous chapters, on filtering. The moving average filter, like that of Chapter 1's
thermostat example, is presented along with the mathematics. The relationship between
this filter type and the digital version of convolution (the convolution sum) is defined.
A connection between the concept of impulse response and frequency response is
presented. By combining the impulse and frequency response concepts with the transfer
function, two classes of filter types are constructed which are expressed through the
lattice diagram and the "z -1" unit time delay. The last item covered in this chapter is a
characterization of the frequency response of a discrete time system.
Chapter 7 discusses the development of digital filters. A comparison between FIR and
IIR filter types is presented. This chapter is the culmination of basic signal processing
building blocks and requires some understanding of the topics in previous chapters.
Notes
Some micro controller prices are mentioned in this book. These are general price ranges
and not actual price quotes. If you are interested in obtaining prices for Motorola
microcontrollers, please contact your nearest Motorola sales office or Motorola author-
ized distributor.
Practical implementations of DSP theory sometimes requires a few extra steps, which
frequently are not included in this book. For more rigorous discussions of 1) all the
required assumptions about signals and systems for DSP theory to be true, and 2)
additional transfer functions that do front-end signal manipulations required for real-
time DSP functions, consult one or more of the books listed in our reference section (at
the end of this book).
Digital Signal Processing And The Microcontroller is a preliminary edition especially
developed as an educational tool and specifically to be part of the "HC16 Tool Kit."
Motorola plans to have future editions available, which will be published by Prentice-
Hall. We appreciate your constructive comments and suggestions. After reviewing this
guidebook, please take time to fill out the business reply card included with the kit and
return it to Motorola.
CONTENTS
Chapter 1
The Microcontroller
What is an MCU? 1-1
A Sample MCU 1-2
What A MCU Is Not 1-5
The Human Factor 1-7
ATypicalMCU Application 1-11
ADC Basics 1-24
Chapter 2
Digital Domain
TheAdvantages 2-1
Digital Domain Strengths 2-3
The MCU As Analog Circuit 2-10
HCIINR and Digital Advantages 2-11
A Perfect Solution 2-19
Chapter 3
Representation of Signals
Pulling It All Together 3-18
Digital Strengths Revisited 3-25
Averaging Filter Summary 3-26
Chapter 4
Fundamentals of DSP
Continuous Time Systems 4-1
Sinusoids and Spectra 4-1
The Frequency Description 4-3
The Sinusoid and Delta Function 4-4
Amplitude and Phase Description 4-7
Sinusoid Basis of Signals 4-14
Real and Complex Modulation 4-15
Frequency Shifting Properties 4-18
Imaginery Numbers 4-19
Chapter 5
Filters and Transfer Functions

Filters and Transfer Functions 5-1
Filter Specification 5-19
Chapter 6
General DSP Concepts

What Exactly is DSP? 6-1
Convolution Sum and the Moving Average 6-7
Frequency Response of the Moving Average Operator 6-9
Infinite Impulse Response Filters 6-19
Chapter 7
Digital Filter Design

How To get the Filter You Desire 7-1
1
C HAP T E R
The Microcontroller
----------_.- ---------------
T oday, the microcontroller (MCU)

is regarded as a single piece of
silicon that contains all of the
WhatlsAn
MCU?
major elements of a computer system. The MCU not only includes the central
processing unit (CPU), comprised of the arithmetic logic unit (ALU) and execu-
tion unit (EU), but the peripheral devices and memory arrays as well. The common
MCU peripherals and memories coresident with the CPU include such diverse
devices as:
• Analog-to-Digital Converters (ADCs)
• Digital-to-Analog Converters (DACs)
• Digital input/output lines
• H-hridge motor drivers
• Voltage comparators
• Math coprocessors
• Light-emitting Diode (LED) drivers
• Random-Access-Memory (RAM)
• Read-Only-Memory (ROM)
• Electrically-Erasable Programmable ROM (EEPROM)
• Electrically-Programmable ROM (EPROM)
• Timer/counters
• Watchdog timers
• Low-voltage detectors
• High-speed Synchronous Peripheral Interface (SPI)
• Universal Asynchronous ReceiverlTransmitters (UART)
This list is certainly not exhaustive and
changes almost weekly. It does, however, represent a good cross section of the
diversity of peripherals available on today's MCUs.
Typically, these peripherals and memo-
ries are not on a single chip; although this is technically feasible, each periph-
eral/memory adds more costs to the MCU. As we will reveal later, "an expensive
MCU" is an oxymoron like "jumbo shrimp."
1-1
Chapter 1 - The Microcontroller
, - - , - - . . _ PAl
Vpp PAS
PA5
PM
PAl
PA2
PAl
'---'--"~ P'"
, - - , - - . . _ PS7
PB6
PBS
PB4
P83
P82
PSI
'---'--"~ PlIO
, - - , - - . . - PC7
PCS
PC,
Pel
PC3
PC2
PCI
'---'--1'- PCO
~~~IP07
PO,
P04
PO,
PO,
POI
OSCI POO
OSC2~~bd
TCIAP
TCAP
Figure 1-1
C8 Block Diagram
A Sample MCU A typical combination of peripherals is

one of the most popular MCVs these days, the Motorola MC68HC70SC8 (see
Figure 1-1). To understand why the MCV is becoming increasingly popular, we'll
tabulate the approximate number and type of peripheral devices displaced by
on-chip peripherals as we look at MCV features. Keep in mind that all of the
features outlined below are on a single chip whose package, a 44-pin plastic leaded
chip carrier (PLCC) in this case, requires no more area than .S square inch.
Features Already Built-in to This MCU:

• On-chip driver for crystal or ceramic resonator to build a stable oscilla-
tor. This feature is very common today, so a separate, external crystal
oscillator, which requires a 14-pin DIP inverter and some discrete
components, is not needed.
1-2
• Memory-mapped digital 110: 24 bidirectional and 7 input-only lines. The

bidirectional lines are software configurable, one bit at a time. This is pretty
nice if the exact mix of digital I/O is unknown at design time. As long as
you don't exceed the maximum count, your software can determine the
110 mix, on-the-fly. This function used to be provided by such devices as
the MC6821 Peripheral Interface Adapter (PIA), a 40-pin DIP.
• 7,744 bytes of EPROM for user program/data storage. This is roughly
equivalent to having a 27C64 type EPROM, a 28-pin DIP, on the same
chip as the CPU.
• Up to 304 bytes of fully static RAM. Since this amount of RAM is small
even when compared to some ofthe smallest available, we'll not count this
in the package savings.
• Software-selectable memory configurations which allow RAMIROM
space tradeoff under program control. This is a neat feature if certain parts
of your program need more RAM at times and more ROM at other times.
• Watchdog timer. This is a safeguard against a runaway program and
potentially catastrophic system performance. Until it was integrated on
devices like this MCU, the watchdog function was handled by an external
circuit, a retriggerable monostable timer. This on-chip retriggerable
monostable is digitally implemented. Its timebase is derived directly from
the crystal oscillator, which avoids drift problems associated with analog-
based retriggerable monostables. I've often seen analog retriggerable
monostables implemented with a 555-type timer, an 8-pin mini-DIP.
• Serial communications interface (SCI). This is equivalent to something
like the old tried and true MC6850, a 24-pin DIP. This self-contained
system uses the CPU crystal timebase to derive various standard baud
rates. With the addition of RS-232 level shifters connected to the receive
and transmit pins, this is a complete system for communicating to termi-
nal-like devices or other MCUs over long wires.
• Serial peripheral interface (SPI). This high-speed (up to 1Mbits/second on
some devices) interface is a four-wire system: 2 data lines, a serial clock
line, and a select line. This is useful for expanding I/O capability by
hanging Serial-In, Parallel-Out (SIPO) or Parallel-In, Serial-Out (PISO)
shift register devices on this SPI "bus." Plus, there are many peripherals
available, such as EEPROM or ADC, that "hang" directly on this interface,
expanding the capabilities of this MCU even further. The SPI system
1-3
peripheral was not really used in the old days, probably because CPUs
had their buses laid bare for the world to use. These buses could be easily
and directly connected to the peripheral, rather than through an inter-
mediate interface system like the SPI. Since a single-chip MCU is
self-contained when used as a single-chip, and the internal CPU bus is
not available to the off-chip world, there is a considerable benefit to
expanding I/O through such an SPI system.
• 16-bit timer subsystem. This system gives us the ability to time external
events relative to CPU activity as well as to generate events. This is
similar in function to a MC6840 Programmable Timer Module ... yes,
another old part and a 28-pin DIP.
Let's count up the packages we can
eliminate by using an MCU along with the type, size, pin count, and area each
package occupies. Table 1-1 contains this information.
Table 1-1
Multi-chip equivaleut of 68HC705 C8's peripherals
Chip Size Type Length x Width Area
8-pin DIP Analog timer 0.4 x 0.3 0.12
14-pin DIP Inverter 0.7 xO.3 0.21
24-pin DIP UART 1.2 x 0.6 0.72
28-pin DIP EPROM 1.4 x 0.6 0.84
28-pin DIP Timer 1.4 x 0.6 0.84
40-pin DIP PIA 2.0 x 0.6 1.20
1-4
Using individual peripherals and memo-

ries, a minimum total area of 3.93 inches 2 would be used. This excludes the
real-estate of the printed circuit board (PCB) which would be consumed just to
electrically connect these components. Compare 3.93 inches 2 to the .5
inches 2 the entire MCU consumes (see Figure 1- 2). Based on this comparison
alone, the MCU is very useful and economical.
3.93 sq. in.
·--1
0.5 sq.ii
J
MCU CPU
Solution Solution
Figure 1-2
A comparison of areas required by like-functioned MCU and CPU systems.
Now that we've mentioned the remark- What An MCU

able features of a specific MCU and what an MCU is in general, we will discuss Is Not
what an MCU is not:
It is not expensive. Although "expen-
siveness" is relative, MCUs would not be considered expensive by most. Some
Motorola MCUs, for example, are available for as little as 90¢ in large quantities
(yes, for the complexity of even this smaller MCU, that is a very reasonable price).
Of course, for a logic gate like the MC74HCOO, 90¢ is a lot of money. But if you
find a MC68020 for that price (and it's still functional), buy it! The MC68020
CPU sells routinely for as much as $170. By comparison, the MCU is an
inexpensive problem solver!
1-5
It is not bulky. Offered in various pack-

ages, a reasonably complex MCV can be obtained in a package as small as a
16-lead SOIC (.4 in. x .3 in., or .12 in2). Even very complex microcontrollers,
like the MC68332 32-bit MCV, similar in performance to some older mainframe
computers like the VAX PDP-l 1, are still available in packages that take up no more
room than 1.2 in 2. This is, literally, mainframe computing power in the palm of
your hand!
It is not a "power pig". Advances in
CMOS technology have continued to increase MCV speeds while reducing power
consumption. One measure of microcontroller/microprocessor performance is
the "instructions per second per watt"(lPS/W). The higher this quantity, the more
efficiently the MCV is using each unit of power. Calculating this quantity for a
popular NMOS CPV of days gone by, the Motorola MC6800, we get 1 million
IPS/W maximum. Compared to the same calculation for the newer
MC68HCllA8, an even more complex product than the relatively primitive
6800, we get 13.3 million IPSIW - over 13 times more efficient in the use of
available power! Of course, at this point we could argue about the technical term
"power pig" and the classification of one or another MCV as such a pig, but we
would miss the point: The MCV of today uses power more efficiently than any
of its predecessors.
Is not difficult to use. Building a micro-
controller application is still a technically oriented process. It requires many levels
of understanding in highly technical disciplines. It requires a leap of faith to
mentally follow the process that generates ASCII data on a word processor and
then assembles, downloads, and emulates the ASCII text file - almost incom-
prehensible to most non-engineers. It's a lot easier and less expensive to develop
MCV applications today than 15 years ago. In these modern times, a complete
in-circuit emulator with software, both assembler and PC-host software for the
emulator, whose performance and features were available ten years ago for
$lO,OOO,.can now be bought for about $500. I've seen some people work with
the new Evaluation Module (EVM), Motorola's answer to low-cost, quick-
learning environments, and program small applications on MCVs the same day!
Pretty amazing considering this was accomplished by a non-electrical engineer.
1-6
It is not a panacea. You should be aware

that an MCU can't be everything to everyone. Perhaps this isn't the time to
question the MCU's potential usefulness, but we will discuss this later in the
chapter. The limiting factor for the performance of an MCU-based system is
usually the MCU's maximum bandwidth.
At this point, you may have a suspicion The Human

that this MCU stuff actually may be useful. From where I sit, I'd say that you're Factor
right. But usefulness, like other things, depends on the observer. Consequently,
as we discuss the potential usefulness ofMCUs, we will compare it to the vantage
points considered important in the electronics industry. Let's start with the most
important points:
An Engineering View
If you've been engineering for a while,
you know engineers never make mistakes. In fact, this may actually be Newton's
long-lost fourth law. In any case, whenever there is a schedule slip:
a) It's usually the technician's fault, and/or,
b) The MegaBug lO00X compiler/assembler keeps
randomly inserting HCF (Halt and Catch Fire)
opcodes in your code, and/or,
c) The purchasing department still hasn't placed the
part order you submitted a year ago last Tuesday.
Whatever the cause, the engineer always
has to modify the design at the last minute. Fortunately, one of the advantages of
MCU-based designs is they are easily modified. This is because, like most
computers, the bulk of the MCU systems' performance lies in its programming.
Consequently, from an engineer's perspective, the MeU is a hedge against
schedule slips from other people's screw-ups.
Trying New Things

During the times engineers are not sav-
ing projects by our ingenious "on-the-fly" redesigns, we are often looking for
new, innovative, and creative ways to accomplish old, boring, and unnecessary
functions. For example, a colleague of mine recently designed a programmable
power supply for generating EPROM and EEPROM programming voltages for
1-7
some Motorola MCUs. He worked very hard to design in an MCU at the core of
the power supply. His boss later pointed how much easier his job would have
been had he used an adjustable linear regulator, a potentiometer instead of an
MCU, and a half dozen other ICs and components. In all seriousness, the design
did have some advantages - the output voltage could be set under computer (PC
type) control by using the MCUs built-in serial interface. That's a feature we pay
a lot for these days in commercial power supply units.
Easy Prototyp,ing
Actually, the reason I mentioned this
power supply is to illustrate the MCU's usefulness to the engineer as a very
friendly and quick-prototyping environment. Even in the case of the programma-
ble power supply, the real value of that approach is to provide an experimental
platform for similar projects whose characteristics, such as rise and fall times,
could be quickly modified. This quick tum-around of ideas in solving problems
can make or break its success. During the early development stages of production
test stations, we would often use the Motorola EVM in conjunction with "experi-
menter's boards" to generate signals which looked similar to the device under test
(DUT). This allowed us to easily validate the performance of the production tester
with known, measurable excitation. It took us no more than an hour with this
approach compared to the time to do the same thing by wiring a bunch of counters
and gates together.
One Part Fits All ... Almost

Imagine having to stock only one IC in
your engineering stockroom/workshop. It's an exaggeration to sayan MCU is so
flexible it would replace all the other ICs in your parts bins. If you've got to choose
only one digital IC for prototyping, however, the MCU should be that device.
Looking in the parts bins in my lab, I have devices like digital-to-analog
converters (e.g., AD7524 CMOS DAC), analog-to-digital converters (like Na-
tional's ADC0800), display drivers (MC14499 for example), analog multipliers
(AD536), op-amps (TL074s, NE5534, etc.), EPROMs (2732s, 2764s), LSITL
and HCMOS logic (' 138s,'00s, etc.), a few generic PALs (16R8s), and, of course,
some MCUs (MC68HC705C8s, MC68705R3s, MC68HC8l1E2s, etc.). When
trying out ideas, I use the MCUs most frequently. Let's say we need a two-digit
LED display for a new product. I can either hook up an MCU, which feeds the
MC14499 display driver via the SPI, or I can hook up the MCU in the configu-
ration shown in Figure 1-3. In this form, the MCU drives the seven-segment
LEDs directly, and the software instructs the MCU to decode and multiplex the
1-8
MCU
Figure 1-3
MCV drives LED directly
display. This eliminates the middleman, that is, the middlecircuit, consisting of
the MC14499 and its supporting components. Since the MCV software decides
the decoding sent to the display, characters appear that aren't supported by the
MC14499.
Despite the flexibility of the MCU in a
prototyping and brainstorming environment, don't throwaway all your non-MCU
components. There are some things other components can do that the MCU can't
- it's just the MCU may be the easiest, most flexible component in your bag of
tricks. Some feature or function implemented with the MCU, to one degree or
another, can replace each of the devices already mentioned.
Product Enhancements
Designing an MCU into a product af-

fords the same flexibility we had for validating our test equipment. A typical
situation might be to use an MCU to interface to an "aural annunciator" (that's a
25-cent word for the lO-cent word, "buzzer") like that shown in Figure 1- 4. The
characteristics of the buzzer will be changed sometime during the life of the
product (see the preceding reasons a, b, and c for schedule slips, page 1-6). If
this circuit had been done with the ubiquitous 555 monolithic timer, one or both
of the timing elements (i.e., resistor and/or capacitor) will need to be physically
changed sometime during its lifetime. Replacing a single resistor and/or capacitor
is probably not a big deal. But imagine having to change some resistors and
1-9
Vpiezo
ID
MCU
Figure 1-4
Very often a product change means software change only
capacitors in lO,OOO final assemblies before shipping your product! An MCU can
avoid many, but not all, uncomfortable situations like this one: an MCU-based
unit could be easily removed and replaced with updated software without risking
any damage the to PCB/assembly. There are many other situations where
easily-altered MCUs can make a lot of cost-effective and face-saving sense.
MCUs Promote Proper Design

Most of us dread the day we make a
mistake ... that is, design-in an undocumented feature. Fortunately, this is more
difficult when using an MCU. Many processes that can be brainlessly put together
with other technologies (discrete, linear, small-scale integration logic, etc.) have
to be carefully considered with an MCU. That is not to sayan MCU is difficult,
it just requires some thought. If you think about it in advance, there is another
advantage to using MCUs: performance is more bounded and predictable than
other choices. This is due to two reasons:
a) One of the advantages of the digital domain is
bounded and predictable performance, and
b) The MCU, like most standard computational ma-
chines, is a sequential machine ... basically, it can
only make one mistake at a time.
1-10
Chapter 1 - The Microcontrolier
We'll revisit this topic when discussing

the advantages of the digital domain in the next chapter.
The View From Production

Production organizations appear to have
two common traits: a) production people intensely dislike rework, and b)
production managers typically don't like to inventory vast amounts of parts stock.
The features of the MCV can turn your production people into "shiny, happy
people."
A product can be designed around the
MCV almost exclusively. The production manager will likely have fewer prob-
lems with a product having few components compared to one with many.
Let's take a look at the MCV in action. ATypical MCU

This way you can experience first hand the MCV's value. Most importantly, Application:
however, this provides a good transition to describe some advantages of working A Household
in the digital domain. Thermostat
Functionally, the household thermostat
is nothing more than a "window comparator," a mechanism that evaluates the
temperature of a room and operates the heating or cooling system (see Figure
1-5). In other words, if a room's temperature is greater or less than a predefined
range of temperatures (called the "comfort zone"), the thermostat switches the
heat or air conditioner on or off until the temperature is again within the comfort
zone.
A To air conditioner
Vsensor~---'
B To heater
Figure 1-5
A Window Comparator
1-11
Comparators may be used as the main

measuring function of the thermostat as indicated in Figure 1-5.
By adding a few more components, this
function diagram could become the "real thing" in Figure 1-. The simplest version
Vee
"Temperature Set"
To AC relay
Temp.
Sensor
To heater relay
Figure 1-6
A typical household thermostat circuit
of a thermostat is a metallic strip usually wound in a coil used as a lever
(contractor) in a single-pole double-throw switch. When the metal in the strip
contracts from the cold, the lever is pulled one direction toward another contact
point, which then completes a circuit and turns on the heat. The heat expands the
metallic strip, and the lever moves in the opposite direction completing another
circuit to tum on the air conditioner. Although this will vary from design to design,
this describes the basic operation of a mechanical thermostat.
When such simple mechanisms are so
reliable, an engineer must have many compelling reasons to use more complex
and costly circuitry. Or the engineer has to be just plain crazy. Jim Sibigtroth, the
original designer of an MCU-based thermostat, had some compelling reasons to
design such an application. The block diagram for this thermostat is shown in
Figure 1-7 .
1-12
Keyboard 14c-)l~ LCD

Display
MCU
Figure 1-7
An MeV-based thermostat
Basic MCU Thermostat Operation
The microcontroller (MCU) reads the
current room temperature through a solid-state temperature sensor whose output
is translated by an Analog-to-Digital Converter (ADC) for the MCU. The ADC' s
output is interpreted by the MCU's software to determine the appropriate action
for any temperature. Through a keypad, someone specifies the "trip points" that
define the "comfort zone" for the heater and air conditioner which is indicated by
liquid crystal display (LCD). The MCU acknowledges this operation through a
piezobuzzer.
Furthermore, the MCU' s serial interface
is made accessible to allow some debugging without affecting the entire assembly.
This is a very typical MCU application. Figure 1- shows the flowchart of the
main control program.
1-13
MAIN
....
~~
Schedule next timer "tick"

(TIC) to occur in 50ms
~
'f
Timed out?
Yes
TIC =TIC+ 1
TIC = 20?
~
Yes
Clear TIC To Zero
.....
1) Update Time And Day
2) Service Keypad
3) Service Beeper
4) Check For User Entry
5) Service AID Temp. Sensors
6) Update Hvac Outputs
7) Service Lcd Display
Figure 1-8
Main Control Program for Honsehold Temperature Control
1-14
A Hypothermal Hypothetical
Let's say doing an MCU-based thermo-
stat appeals to your sense of HTF (High Toy Factor) and has enticed you to try to
build this thermostat project. So you send away to your area's Motorola sales
office for the M68HC05 Applications Guide (M68HC05AG/AD) that has the
details for the project. After painstakingly building the hardware, writing the
software, and installing the thermostat in your wall, you discover the thermostat
is acting strangely. At first, you're not even sure anything is wrong ... but then ... you
notice: it's December, you live in Alaska, and your air conditioner has just kicked
on! But you can to hook up your PC to the thermostat through a serial port and
RS-232 level-shifters. Consequently, you can debug the problem without having
to gouge the thermostat out ofthe wall. Fortunately, you are able to "look at" the
data from the ADC and thus can ascertain, at the very least, if the problem is a
malfunction with the temperature sensor or ADC. So, you set up your PC to take
readings for a 12-hour period and retire to your igloo with a few extra dogs that
night (a three-dog night, thanks to the air conditioner).
After a bone-chilling night of restless
sleep, you get up to compile and graph your data from the previous 12 hours. This
is the graph in Figure 1-. Being an astute observer, you notice two things: a)
the word "comfort" is misspelled due to your typing "i" instead of "0" because of
your hypothermally-induced shaking, and b) the readings from the sensor are
Too Hot--turn on AC
SP.M. 12 A.M. SA.M.

Figure 1-9
A temperature plot of a malfunctioning air condtioner system
1-15
rapidly changing. Based upon a rough thermodynamic calculation (done on the

back of a napkin, of course) and your personal experience, you conclude the actual
temperature can't be varying as wildly nor as frequently as indicated by your graph.
It must be noise. What's an engineer to do? Here are some options:
• Low-pass filter the voltage(s) from the sensors. This makes sense but
isn't indicated since you can't determine whether the noise is coming
from the sensor(s) or there is a malfunction in the ADC. Besides it's
going to get awfully messy to try and remove the unit from your igloo
wall to put the resistors and capacitors needed to low-pass the sensor
output(s). For these reasons, you decide this isn't a reasonable alterna-
tive.
• Move into another igloo. This sounds more like a management decision
than an engineering one. Consequently, it isn't a reasonable alternative.
• Reduce the variations over time in the ADC data. This makes sense,
assuming you can change the MCU code without changing the hardware.
Since this is my hypothetical example, we'll choose this option. Figure
1- shows the ADC service routine for the thermostat with a call to the
low-pass filter routine inserted.
Some Average{ing) Software

A simple jump-to-subroutine placed
within the ADC read routine won't get rid of our noise problem: it's the subroutine
that's going to do the work. What's in it? Look at the listing in Figure 1-. We' 11
take it one chunk of code at a time. Before starting, however, I want to introduce
a graphic aid to help visualize the movement of data in the averaging routine.
We'll use the following diagram to symbolize the MCU memory:
(most reJ;~Ts~~~~11-7 0
________
L-_ _ _ _
.---~
• 64 Bytes
•
TBUfEnd'I-- - - - - - - - 1
(oldest sample)L-----------'
1-16
*********************************************************
* A2D - Check temp. sensors (via SPI and MC145041) *
* If TIC 0, send addr 0 ignore return data *
* 1fT IC 1, send addr 1 ret urn dat a is ch. 0 va I *
* 1fT I C 2, send addr 2 ret urn dat a is ch. 1 va I *
* If TIC> 2, skip A2D routine *
* To compensate for sensor & op-amp offset, AID result *
* wi I I be modified by subtracting an offset constant *
*********************************************************
A2D EOU * Check temp. sensors
LOA TIC If Tic 0,1,or2writetoSPI
CMP #2
BHI XA2D 1fT i c > 2; Ex i t
ASLA Move TIC # 0-2 to upper nibble
ASLA
ASLA
ASLA 4 bit left sh i ft
TST SPSR Reads SPIF (part of SP1F clear)
BCLR 3,PORTC Drive low true SAID CEO to 0
STA SPDR Initiates a transfer
* Requests conversion of next channel and returns data

* from previous channel Ch.O=lndoor Ch. l=Outdoor
SPIFLP BRCLR 7,SPSR,SPIFLP Wait for SPI Xfer complete

BSET 3,PORTC Drive low true SAID CEo to
LOA TIC If O-Exit, 1 or 2 Read AID data
BEO XA2D o so exit
LOA SPDR Get AID data
JSR LOUPRSS Fi Iter input uith 6,/pt aug.
BRSET 1,TIC,ADCHl 1fT i c=2, dat a is Ch. 1

SUB OFFO A/D Ch.O; subtract offset
STA INTMP Update indoor temperature
BRA XA2D & Ex i t
ADCH 1 SUB OFF 1 A/D Ch. 1; subtract offset
STA OUTMP Update outdoor temperature
XA2D RTS ** RETURN from A2D **
Figure 1-10
ADC routine for thermostat
The left-right width of the block is the

data width, which is 8-bits in this case. The top-bottom height of the block is the
number of bytes. Now...back to the code.
1-17
There are two pieces of the code that

actually do the work in averaging the readings. The ftrst is from the label
"MoveSmpls" up to, but not including, the label "AvgLoop." The second is from
AvgLoop to the label "AvgDone."The first block of code basically prepares the
data to calculate the average by making sure that the 64 most-recentADC samples
are placed within the RAM buffer. The second chunk of code, the one that actually
averages the readings, starts at AvgLoop. Starting at the instruction before
MoveSmpls, we note the x-register is initialized with a value equal to the buffer
RAM location with the highest numerical value minus one: "TBufEnd-l." The
first two instructions at the label MoveSmpls move the second-oldest sample from
TBufEnd-1 to "TBufEnd." The sample at TBufEnd is now the oldest ADC sample
in memory. The third instruction after MoveSmpls decrements the x-register so
it points to TButTop-2. Afterwards, the x-register is checked to see if we've
averaged all of our samples, which is, in this case, 64. Since we haven't done that
many yet, our code branches back to MoveSmpls.
If viewed as data movement within our
rectangular memory block, there are 64 data movements sequentially moving
toward the part of the memory block labelled "TimeBufr." Each data movement
copies the contents of each RAM location to the RAM location just below. The
most current ADC sample is always located at TimeBufr. This is schematically
represented in Figure 1- .
TimeBufr ------I
f-I
EJ·[___3
[=____1
Figure 1-11
ADC Samples memory locations
1-18
*****"'***************.*******"'*"*****************-*****"""'<****
*t!ane:LowPasi9
"'Purpose: To reduce uariatlons in ADC data frolltherllol

'" sensors. A6't-point Ilouingaveroge Is used.
"'Use: poss lalest sample from ADC to this routine uio RCCA.
segPAGEORAM
TellpSllpl rnb 1
RccnSB I"lIIb 1
TellpCotc rnb 1
TimeBufr rmbM
TBufTop eQu'"
segProgran
lOIlJPOss staTellpSnpJ jthis contains the latest

; ADCeD"ple. 50ve It.
IdxlTBufTop-1 ,stort l1ouingdato frOllhere.
MoveSllpls IdoO,x ;get sample
911;1I,x ;nove it dOUlnlhe line.
cpx'TilleBufr-l ;check. for latest eOllple location.

bneMoueSlllpts
inx ;Ieaue x pointing to Ti",eBufr
stoO,X
clrAccMSe ;thlspreps the nSB of the
; Dverogeresult.
RugLoop inx ;polnt to next location.
cpxITBufTop+\ ;check far last lacation
; In buffer.
beQScale ;Ieave If last $ample.
add ,x ;begin avg calc.
bccAvglaop ;If no ripple to MSB thru
; cy flag, then add again.
staTellpCalc
ctra
adcAccMSB ;thispropagale$CY fig
; Into AccnSB
staAccMSB
Ida Te .. pCalc
bra AvgLoop
Scale lsI' AccMSB
rora
lsI' AccMSB
rora
IsrAccnSB
1151' AccMSB
1'01'0
IsrAccMSB
l"orQ
lsI' AccMSB
"01'0
AvgDone I"t8
Figure 1-12
64-point moving average routine
Following the flow of the code, if our

MCU has made it to AvgLoop, then the x -register contains the value TimeBufr-l.
Because we want to confine our averaging routine to the ADC data, which resides
between TimeBufr and TButTop inclusive, we first increment the x-register so it
is pointing to TimeBufr instead of TimeBufr-l. The next instruction compares the
value of the x-register to see if we've made it to the top of the RAM buffer where
the newest ADC sample resides. We branch to the next chunk of code IF the
x-register is pointing to the newest ADC sample located at the "top" of the RAM
buffer (TimeBufr). In our next chunk of code, we begin the process of adding the
ADC samples together. The RCD5 CPU is an 8-bit machine that requires some
extra steps for doing 16-bit addition. These steps are the instructions after the
"add,x" and the instruction just before the label "Scale."
1-19
Averaging consists of two separate proc-

esses, addition of a list of numbers and division of the sum by the count of the
numbers in the list. Now it is time to do the division. That's what the lines
inclusively between the labels "Scale" and "AvgDone" do. The 8-bit architecture
of the HCOS requires many instructions to divide. Following the instructions from
Scale to the end, you'll see these perform six 16-bit shifts on the sum of the ADC
values added in the previous part of the code.
To give you a sneak preview of things to
come, let's mess around with the symbols we use to show data movements. This
will show you that what we are talking about now has a direct application to the
fun stuff later on. So, first separate each block and connect with arrows as shown
in Figure L-.
*
I
~
I
Figure 1-13
Now, turn the entire string of rectangles

90° counter-clockwise and reduce the size of the rectangles to make it easier to
display. This results in Figure 1-.
Figure 1-14
1-20
This view of the memory roughly repre-

sents the movement of the data by the averaging routine where each memory
location is represented by arectangle. At the output of each block, which represents
each byte of data, draw arrows to a circle with a "plus" sign in it. One way of
drawing it is in Figure 1- (a).
Below the summation bubble, draw an-
other circle with an "x" in it. .. the symbol for multiplication (Figure 1- (b».
T~
"
Figure 1-15
Let's pause at this point to review a rule

of arithmetic we learned long ago but probably forgot. A rule in arithmetic says
multiplication is distributive over addition. When mUltiplying a number by the
sum of two other numbers it doesn't matter whether you multiply the individual
numbers first and then add or add them first and then multiply. The result is the
same. In symbols this is:
A(B+ C) =AB +AC
For our diagram, move the multiplica-

tion bubble from its current location to each line entering the addition bubble,
like Figure 1-.
1-21
Figure 1-16
Lattice Diagram
Do you recognize this diagram? If you

don't, don't worry. You'll see plenty of it if you work with DSP stuff. For those
who recognized it, you know this as a lattice diagram. This type of diagram is
often used to show the structure of digital filters. Does that mean that the averaging
we just did is in fact a digital filter? Stay tuned.
Thermostat Performance Results

After walking through the code, let alone
writing it to begin with, we want to know "Does it work?". The answer to that
question indicates writing a book has something in common with writing televi-
sion sitcoms: the writer never introduces a problem or question without an
answer... and usually the answer is the obvious one. So yes ... of course our
averaging routine works! To see how effectively the averaging process works on
the actual thermostat data, look at the before and after graphs of the data in Figure
1- .
Had this thermostat been constructed

with just any other technology, the solutions would have required more extensive
hardware changes. Although we could have opted for a pure hardware fIX in this
case, we didn't have to. That's the whole point of this last exercise: it only took
a few strokes of the computer keyboard to rewrite the source code to the MeV,
1-22
Too Hot--turn on AC
w
a:
::I
~
W
a..
:E
w
I-
6 P.M. 12 A.M. 6 A.M.
WithoutAverating
Too Hot--turn on AC
COMFORT ZONE
Too CoJd--turn on Heat
6 P.M. 12 A.M. 6A.M.
WithAveraging
Figure 1-17
Thermal sensor before and after averaging routine
program a new RC05 with the improved software, and then trade the socketed
MCV for the new version. Compare this to pulling the thermostat out of the wall,
desoldering discrete components, and spending untold hours changing resistor-
capacitor values until the problem disappeared. Granted, you have to know about
software, but the ease of working in the digital domain justifies the effort to
becomes fluent with this software stuff.
1-23
ADC Basics We could capture signals in digital form

so the MCU' s software could work on it through an analog-to-digital converter
(A-to-D, or AOC). Most of us are familiar with these devices and many of us
have used one. Even so, we need to discuss these in more detail to capitalize on
the strengths of the digital domain and minimize construction problems. Digitiz-
ing analog signals creates artifacts in the process not present in the input analog
waveforms.
Now, we will begin our discussion about

analog-to-digital conversion by describing the constituent parts of the process,
followed by a brief discussion about various converters.
The Sample-Hold Function

In most modern ADCs, the first step in
our conversion is usually the sample-hold function. Although not absolutely
necessary, this function is, to some degree, a function of the selected ADC as well
as the signal we want to convert. Is this function required for your application?
What happens if we don't use it? We will answer these questions in a minute. First,
let's look at the process of sampling the analog world with our sample/hold
amplifier.
Figure 1- shows an analog wave form.
The sample/hold amplifier function is represented graphically by the following
steps:
1) Move directly up from the "t" axis with a straight

line until intersecting the input wave form.
2) Draw a rectangle, with a width determined by the

distance between t-axis "ticks" and the height de-
termined by (1), with its left-hand corner touching
the analog wave form.
3) Repeat this process for each successive tick on the

horizontal axis or until you run out of input wave
form.
We have a wave form that looks like

stairsteps representing the general shape of the input waveform. This stairstepped
waveform is the sample-holdvalue that is fed to the input stage of the ADC.
1-24
Ih
I
I
fm1
Figure 1-18
Sampling an Analog Signal with Sample and Hold Amplifier
We assume most of the signals we want

to convert to digital numbers include both DC and AC components. Let's feed a
sine wave into an ADC. Taking the derivative of this function, we obtain an
expression for the the sine wave's rate of voltage change per unit time. The
maximum rate of voltage change occurs when a sine wave crosses the horizontal
time axis. In quantitative terms, the rate of change of an A volt peak-valued sine
wave amounts to:
Vet) = Asin(21tft)
Maximum rate of change == ( : 10= 21tAf
fmax = [_1 ~ = [21t(16 1x 10-6) )[~)""

21tA) dt 28
194 Hertz
If we are using an 8-bit converter with a

conversion time of 16 microseconds and a reference voltage of 5V, the maximum
undistorted sine wave frequency that can be converted is less than 40 Hz, although
our conversion frequency is much greater than the frequency of the sine wave!
The Generalized ADe
The analog-to-digital conversion proc-

ess is nothing more than a translation of continuous-time, real-world signals into
a granular, digital approximation. This digital representation is, in most cases,
only an approximation of the actual voltage. In doing this translation, regardless
of the specific ADC device used, a voltage comparator compares the input's analog
voltage to a converging voltage internally generated by the ADC. When these two
1-25
inputs to the comparator are equal, the analog voltage generator stops and the
digital code associated with the internal analog voltage generator at that moment
represents the analog signal. Let's see how this mechanism is applied with the
various converters.
Successive Approximation Converter

One ADC is the successive approxima-
tion converter (SAC). A device responsible for the character of the SAC is the
successive approximation register (SAR), which determines how the voltage is
generated by the ADC. The elements of the SAC are: a) a voltage comparator, b)
a digital-to-analog converter (DAC) and c) a clocked digital counter. The output
of the comparator tells the digital counter to stop once the D/A output and the
input voltage are roughly equivalent. The SAC's distinctive characteristic is the
manner in which the DAC voltage is created: each "guess" at the input voltage,
created by the DAC, is successively halved until equivalence is achieved. This
technique allows a moderately fast (conversion rates> Illsecond ) and moderate
resolution (12-bits or less) ofthe unknown analog voltage.
IntegratingADC
Another popular converter used for
much slower but typically much higher resolution conversions is the integrating
type converter. This converter applies our generalized converter form in a unique
way: it substitutes a charging capacitor for the DAC of our generalized ADC. The
capacitor is charged by the input signal voltage and discharged by a constant
current source. The time it takes for the voltage ramp, created from discharging
the capacitor, to reach zero voltage is directly proportional to the analog input
voltage that originally charged the capacitor. The linear voltage ramp is timed by
a digitally clocked counter until reaching zero, at which time the digital counter
stops. The resulting count shown by the counter is directly proportional to the
input voltage. This type of AID process is good for 10 Ilsec or higher conversion
times and resolutions typically in the 18- and 20-bit range.
Flash Converter
Another conversion process is the
''flash'' converter, which uses a string of voltage comparators. Each comparator
has one input tied to successively increasing reference voltages. The input voltage
is applied to all of the other comparator's inputs that have all been tied together.
The outputs of the comparators are sent through a digital logic array to produce
1-26
a convenient output code as a function of the input voltage. Let's say we have
three voltage comparators. The comparator's individual inverting inputs are
connected to 1.2SV, 2.5V, and SV, respectively, so the comparators are referenced
to 1.2SV, 2.SV, and SY. Let us then apply our input voltage to the other input. If
the input voltage is less than the reference voltages, none of the comparators will
trip. Once the input voltage increases above 1.2SV, the comparator having the
1.2SV reference will change from logic 0 to logic I. The other comparators will
remain unchanged until the input voltage rises above their respective reference
voltages. Because its performance depends primarily upon the comparator's
response speed, the flash converter is very useful for high-speed concertinas.
1-27
2
CHAPTER
The Digital Domain
M CUS processing digitized ana-

log signals are powerful new
tools for solving old and new
problems alike. In Chapter 1 the microcontroller was introduced as a set of very
useful features, including A-to-D conversion. Using these features we solved a
noise problem in a common household thermostat without changing the hardware.
In this chapter, we'll investigate some traits in the digital domain that make it a
very appealing solution.
Listing the strengths and weaknesses of

the digital domain is a straight-forward task. Some of the important subtleties of
those strengths and weaknesses, however, are unclear. Consequently, we will
briefly discuss and then illustrate each strength with actual applications. After-
ward, we will discuss the weaknesses.
We'll begin by discussing benefits and The Advantages
strengths of the digital domain. Remember the example of the thermostat in
Chapter 1 and how the digital domain worked to our advantage.
Doesn't Vary Over Time
It's hard not to notice the consistent per-

formance from your unit once you've designed in the digital domain. For example,
designing a remotely controllable "digital pot" - a potentiometer substitute that
varies the amplitude of a digitized signal - you notice the performance of the
unit doesn't vary over time or temperature. The process of varying a signal's
amplitude consists of multiplying an incoming signal by a value representing the
amplitude, or "volume" setting, you want. There's not much to drift or change in
that situation. Multiplication is multiplication today, tomorrow, or next year.
That's important to designers who want their designs to work "first time and every
time" for everyone else as well as themselves. The same cannot be said for an
analog implementation, which uses something like a transconductance multiplier
for multiplication. Every example of such devices requires the trimming of offsets
for many applications which can drift with time, temperature, and vibration. In
addition, the strategic placement of a temperature-sensitive resistor is often
required to compensate for the transistors' inherent drift that occurs with changes
2-1
Chapter 2 - The Digital Domain
in temperature. Such analog circuit traits are manageable, but ifwe can reduce
drifting parameters, then let's do it. With any carefully implemented digital
domain design, sources of parametric drift are eliminated completely within
specified error bounds.
Doesn't Vary Between Units

The next step with our digital pot design
is to produce the product. It's in this next phase where the advantages of working
in the digital domain are apparent. Even your company's accountant could
appreciate this point: parametric changes between units are predictable and
bounded. Translated into English, this means the differences between products
can be negligible or, at best, known in the design phase. From an accounting
perspective this means more profit from more sales of a better product, fewer
service returns, and a better reputation for your company.
Less Expensive Than Analog

Many times implementing a design in
the digital domain can be less expensive than the equivalent analog implemen-
tation. V sing our digital pot example, however, the analog implementation is less
costly to build than the digital version.
Unnatural Responses
We have assumed the digital pot example
is just a straight translation of the analog pot function. Another advantage of the
digital domain is the ability to achieve "unnatural" responses including extraor-
dinary analog-like responses. The word "unnatural" in this case is meant to imply
a system response not easily implemented with analog components. So, even
though a single $1 pot could replace our more expensive digital example, would
the analog pot be an equally good solution if there were additional "unnatural"
performance criterion? That depends on what it is. If the analog pot must respond
to the commands from a personal computer (PC) via its RS-232 port, which of
the two makes sense? Right, our digital pot. An analog pot doesn't to respond to
your Pc.
Performance Easily Changed

A final advantage of the digital domain
is illustrated with this example. Let's assume that your customers said they
preferred a logarithmic taper to the digital pot instead of the linear taper product
you currently ship. No problem. Just change the software in the MeV. Perform-
2-2
ance is easily altered with a digital domain implementation. Whether you're

addressing feature enhancements or correcting problems with a product's opera-
tion, you will appreciate the easily altered performance of the digital domain.
Although the performance of some analog circuits can be easily modified,
component values must be changed to alter the circuit's performance. It is possible
with a digital domain implementation to make significant enhancements without
changing a single component. Try that with an analog circuit!
We've been somewhat superficial and Digital Domain
academic in our discussion about the strengths of the digital domain. Let's now Strengths
bring it down to earth with a real-world application: creating a magnetic tape Illustrated
noise reduction system.
"Noise 'R' Us" Recording Studios
Suppose I am an owner of a small sound

recording studio that friends and clients use to tape their musical talents. The
studio has a 16-track magnetic tape machine suffering from the evils of tape hiss,
a high frequency sound produced from the tape's physically passing over the tape
machine's playback head. Digitally-based recording systems with performance
like a compact disc make tape hiss a problem of the past.
A digital recording system, however, is

still too expensive for me to justify. We want our system, however, to sound as
much like commercially available CDs.
Fortunately, in recent years much head-

way has been made in reducing the type of tape noise plaguing our system: Dolby
Labs and a company by the name of dbx have both created noise reduction systems
that effectively reduce tape hiss so it's inaudible. There's only one minor hitch:
money, my lack of money. As nice as the commercially available noise reduction
circuits are, they are expensive. Dolby Labs makes a system that is excellent at
reducing the tape hiss and, at $1200 per tape channel, very good at reducing my
available funds. It was the need for low-noise recordings and the high cost of the
commercial noise reduction systems that is leading me to design my own noise
reduction system. To design a workable system, I had to first understand how noise
reduction works and then choose a specific implementation.
2-3
Simple Noise Reduction

Tape hiss is a function of the current state
of the music's signal-to-noise ratio (SNR). If the music is loud, a listener can't
hear the tape hiss through the music - the tape hiss is "masked." If the volume
of the music is low, compared to the tape hiss, then a listener hears the tape hiss.
Consequently, the masking effect from a high signal-to-noise ratio (SNR) is an
effective way of reducing the audible impact of tape hiss.
The SNR is a way to compare the level

of the music to the level of the noise ... tape hiss, in this case. Several record
producers believe noise reduction is not necessary for rock and roll. Why?
Because rock'n'roll tends to have consistently high signal levels. This is particu-
larly true of the "heavier" rock'n'roll tunes; there's so much music that the music
effectively covers up or "masks" the tape hiss at any given time. You could say
the SNR of rock'n'roll is typically very high (some classical music fans would
refer to the rock'n'roll as noise). Recording only high-energy rock'n'roll is one
way, and perhaps the simplest, to reduce the perceptible noise in recordings. But
this doesn't work in our hypothetical situation .. Some of my clients like "bizarre,"
i.e., non-rock'n'roll, music such as jazz and R&B. Consequently, I had to find
some other way to maximize the SNR of my tape system.
Another method that can be used to ad-

dress the tape hiss issue is shown in Figure 2-1. In this noise reduction (NR)
system, the high frequencies are boosted, for example, + 10 dB before the signal
is recorded to tape. On playback if we introduce a 10 dB loss (-10 dB) at the same
frequencies we boosted, the net result is the signal on the tape will sound the same
as before. This is because the net effect of boosting and then cutting the same
frequencies in this complementary way produces the original signal. This is
analogous to an amplifier stage which boosts overall signal gain by a factor of
two only to have a successive stage introduce a gain factor of.5. The net effect is
the signal emerges apparently unaffected, i.e., at unity gain, at the output of the
last stage. This "boost then cut" form of noise reduction does the same thing except
we are boosting and cutting only parts ofthe signal, only in the same area ofthe
frequency spectrum where the tape hiss resides. In this way, during the playback
the tape hiss levels are cut back when the frequencies are cut. Since we have
effectively restored the original signal and cut the tape hiss, by 10 dB in this
example, we have effectively increased the SNR of our tape recording system by
10 dB. This looks like a practical way to reduce tape hiss when recording clients
don't want to record high-energy rock'n'roll.
2-4
Danger of tape
saturalionatlower
levels'causeol~
boosledhighs.
~+12 o~
O~
2.5 kHz 10 kHz
PRE-EMPHASIS FILTER
I I\ I
2.5 kHz
~~2
10 kHz
DE-EMPHASIS FILTER
Audio Audio
highs highs
boosted boosted,lapehiss
normal
Audio
Audio In--to tape highs
normal,tape hiss
reduced
Figure 2-1
Tape Noise Example
There is a significant flaw with this

boost!cut noise reduction system: the tape's limited dynamic range as a storage
medium_ This means the boosted high frequencies will distort the tape at a lower
absolute level than the lower, unboosted frequencies_ So although the boost/cut
scheme maximizes the SNR, we reduce the effective dynamic range of the system
because of the boost. Fortunately, there is a way around this problem to make the
boost/cut method usable_ An amplitude compression can be used to record the
signal with the boosted high frequencies while reducing the chance of overloading
and distorting the tape. Here's how it works. As the energy of the music increases,
an automatic volume control, the amplitude compressor, turns down the volume
of the music in a predetermined way. As an example, if the music increased its
loudness by a factor of two (6 dB for voltage measurement), we could have this
compressor decrease the volume by 1.414 (3 dB for voltage measurement). The
net result is the signal going to the tape has increased by only a factor of
approximately 1.414 (a 2:1 ratio of input to output change in dB) instead of by 6
dB before we compressed the signal.
What does this all mean? Simply this: by
using a compressor we can reduce the chance of tape overload from boosted high
frequencies. Good, but compressing the signal like this introduces a new problem_
It sounds "unnatural." Fixing this is analogous to boosting and cutting specific
2-5
frequencies of our signal during recording and playback. While compressing the
signal during record to avoid premature distortion, we use the complement of
compression, called expansion, to restore the original signal's dynamic range on
playback.
Encode/Decode NR System
Figure 2-2 shows a block diagram for
implementing an encode/decode type of tape noise reduction (NR) system. Within
the box marked "process" is the compressor and preemphasis filter to encode the
signal going to the tape. This system view of the NR system allows you to see the
actual mechanics of how the NR works without getting bogged down in the details.
Basically, the process block produces a known output, a(t), when presented with
the signal, s(t), that we want to record. The signal that actually gets recorded onto
tape, however, is the sum of these two, that is, a(t) + s(t). During playback, we
change the configuration slightly of the process and summing amplifier. The point
of the decode process is to retrieve the original unencoded signal s(t). By inverting
the phase of the process block and taking its input from the output of the summer,
we get a decode process that is the mirror image of the encode process and that
strips the encoded part of the signal, a(t), from the signal stored on the tape.
Remember, the process function produces a(t) if s(t) is presented at its input. It is
also true if s(t) is at the output of the summer and the phase of the process is
inverted that the system-level equations work out:
ENCODE PROCESS
~ O'-----Proce_ss~~ .,.,"
DECODE PROCESS
s(t)+a(t) --------t s(t)
Figure 2-2
EncodeIDecode Tape Noise Reduction
2-6
Once I determined the overall configura-

tion of the NR system by similar reasoning, I needed to identify the characteristics
of the process block. This was relatively easy because the process employs both
compression and preemphasis to do what I needed it to do. The details of the
encoder is shown in Figure 2-3 .
log(vin )
Linear-ta-Log Converter
f-
iF-
RMS·to-DC
-
I
•
--- ~
0----/1+12
r- -----<
500 12 kHz
PRE·EMPHASISlHIGH·PASS FILTER Voltage Controlled Amplifier
2.16
+
1 to
from
SOURC E T APE
Figure 2-3
Noise Reductiou Encoder Detailed Block Diagram
The function blocks marked VCA,

RMS-to-DC, and LogO make up the compressor. The preemphasized signal is
fed to the input of the compressor, while the output of the compressor is summed
with the unaltered signal. This produces the composite signal, set) + aCt), which
is recorded on tape. Analogously, the decode loop uses the same elements; the
only difference is the input to the compressor is now the output of the summer
rather than the signal to be encoded.
2-7
Performance Requirements
After identifying the key concepts for a
noise reduction system, I established a set of performance standards for the NR
system. They were:
• The NR system shall reduce audible tape playback noise to a level

competitive with commercially available units.
• The artifacts from mistracking between the encode/decode processes

will be kept to an absolute minimum or eliminated completely.
• The NR system will have the lowest possible cost, specifically, less than
$25 per channel.
• The tape shall be "listenable" without the NR decode process. Listenable

means that any equalization funnies can be corrected with standard
studio gear. This is a precaution in case a client takes one of our master
tapes to another studio. I'd sure hate to ship my NR system with the
client's tape!
The next step in the design process was

to identify a specific circuit implementation.
The NR Prototype
The analog-based prototype has two
noteworthy features: an RMS-to-DC converter, with built-in logarithmic output
for driving the VCA, and a thermal tracking resistor. The resistor is required to
keep the RMS-to-DC converter's output voltage from drifting.
Performance Of The Prototype

I breadboarded the circuit to verify the
performance of the unit. All the testing I did was subjective in nature. The dynamic
nature of noise reduction makes it difficult to test with standard lab equipment.
The important performance criterion is how good the NR sounds to the ear. I used
a cassette deck as the tape machine under test. The results were very gratifying
because it performed to my subjective expectations.
2-8
NR Implementation Problems
If the subjective performance of the NR

system is so good, then the design must be O.K., right? No. Unfortunately, the
performance of the circuit could falter sometime in the future because of mis-
tracking between the encode and decode modes of operation. This type of
mistracking can be induced by environmental conditions and results in certain
circuit parameters drifting. Remember our encode/decode equations in Figure 2-2
on page 34 assumed perfect encode/decode arithmetic; there were no error terms
shown in the math.
A tracking error between the encode and

decode process is exhibited as a loss or accentuation of certain frequencies during
tape playback. Nonetheless, I wanted to reduce as much as practical. You have
probably experienced this type of phenomena with your home and car stereos, if
you record tapes at home and listen to them in your car. The mismatch between
the encode and decode filters of the two different machines causes the differences
in the frequency response between your home cassette recorder and your car's
playback unit.
One way to correct this problem is to use

components or processes having low errors. Another way is to trim out the error.
I prefer using circuits that don't require trimming because the components used
to trim can become sources of drift themselves. This left me to find devices or
techniques without the inherent errors. While I was surveying the industry for such
devices, I realized that there was yet another problem to consider even while
looking for solutions to the drift problem: price.
In many ways, the drift issue may be

trivial when compared to the price issue. Even with a perfect solution to my tape
hiss problem, like a digital recording system, the availability of that solution is a
function my VISA credit line. Like it or not, the price issue as legitimate a
consideration in product design as the selection of technologies.
So, here was my dilemma: my prototype

worked well, but included a $23 RMS-to-DC converter inherently prone to
temperature-related drift. Both the cost and inclination for drift were at odds with
my performance specifications. That spec included a total product cost per channel
of $25 or less for all the electronics. The spec also required the design must
2-9
minimize, if not completely eliminate, tracking errors. An additional, "hidden"

spec requirement was I didn't want to spend all my time selecting the components
for this project. This project was beginning to look too much like work! But I
stumbled upon the solution in my "defunct" projects box.
Several months ago, I developed an
MCU-based audio device using one of the HCII variants I obtained for around
$11 apiece in some quantity. This device was to control a Voltage Controlled
Amplifier (VCA) using upon certain information contained in the audio. The
device, an audio compressornimiter, was the functional equivalent to the ampli-
tude compressor needed in the NR system. The question is: if I use it in the NR
system, would it address the performance issues?
TheMCUAs Figure 2-4 gives a visual hint about how
An Analog Cir- an HCII applies to this application. By replacing the RMS-to-DC chip with the
cuit HCll, both the RMS-to-DC conversion is replaced as well as the linear-log
weighting of the signal required to properly drive the VCA stage. Would this part
substitution address the drift and cost issues? Yes. Here's how: the RMS-to-DC
process is implemented by performing simple math on ADC results. The results
of these calculations are causal, bounded, possibly convergent, and possess error
margins that are calculable. This on account of the nature of the arithmetic -
even with quantifying the arguments. If you have any doubt about this, perform
an RMS calculation on a set of numbers tomorrow in your spare time. Repeat this
same calculation several times thereafter under different conditions, but using the
same set of numbers. It doesn't matter when or where you perform the calcula-
tions, you'll get the same answer. It's almost absurd to suggest there would be a
difference from the original calculation. In addition, even the numbers in the
calculation are changed or the precision of the arithmetic, the results may be
predicted to lie within particular outer bounds. If so, then the question is, "Is the
bounded error sufficiently low or can it be made sufficient for our application?"
By subjective test, under a variety of conditions and after guesstimating a
quantitative impact, the answer is a resounding "Yes!"
2-10
log(vin ) f-- ~----LLl(J)~mCl).E~CJt3::Sfj

I-~ -c[8§~»CLo...o..a..
8 XTAL Qa::: LL
9 pca
Linear-to-Log Converter 10 PCl
PC2
--
12 PC3
13 PC4
14 PCS MC68HC11A1
15 PC6
16 PC?
17 Resel\
18 XIRQI
191RQI
JJ,:')~
20 PDO
f---
RMS-to-OC
Figure 2-4
68 HCll Can Replace Both the RMS-TO-DC
Chip and the Linear Log Weighting
The linear-to-Iog converter has the same

features the RMS-to-DC process does: the process is bounded and has a predict-
able error contribution. This, too, turned out to be acceptable for this application
even when combined with the errors ofthe RMS-to-DC conversion process.
Before our discussion about the NR sys- HC11 NRAnd
tem, we talked about the strengths of the digital domain. These strengths include: Digital Domain
Advantages
a) Minimizing parametric changes within a unit over

time and temperature,
b) Predictable and bounded performance between

units,
c) Lower costs than an equivalent analog version,

usually having a subset of analog performance,
d) "Unnatural" responses, including extraordinary

analog-like responses, and
e) Flexible performance.
2-11
The digitally-controlled NR system cer-

tainly takes advantage of these features. If you agree our RMS-to-DC converter
and linear-to-Iog converter functions will not drift or change original performance
when implemented digitally, then we can agree that parametric changes relative
to time and temperature are minimized within the same device because of the
digital implementation. It's not a large leap in logic to assert the same charac-
teristics that give us (a) also give us (b). Parametric changes between units are
predictable and bounded. The latter point is really a restatement of the former.
The cost of the HC II-based NR system
per channel, around $15, is much less than the analog counterpart even though it
includes the cost of the DAC required to interface to the VCA. Compare this $15
price tag with the $23 price tag for just the RMS-to-DC converter. This exemplifies
another advantage of the digital domain: a digital implementation is less expen-
sive than analog but usually with a subset of analog performance. In this case,
the digitally implemented RMS-to-DC converter does not have the 10 MHz
bandwidth of the monolithic RMS-to-DC converter. The bandwidth of the
digitally implemented system is lower than the analog version ...but so is the price.
The HCll-based NR system is also a
good example of how the digital domain can be used to obtain "unnatural"
responses. In the NR system, the unnatural response is a nonlinear, logarithmic
response. The same function may be accomplished using analog components like
the RMS-to-DC chip and its log output, but the accuracy ofthe output is affected
by temperature. This is only one example. Many functions in the analog domain
could be accomplished with great difficulty or many parts that are much more
easily accomplished in the digital domain. This is because of the nonlinear or
complex nature of the required function. Functions like waveform generation fall
in the "it's-easier-to-do-in-digital-than-in-analog" category. Good examples are
waveforms that are not particularly periodic or that have other characteristics not
be easily processed or generated by continuous-time devices.
Finally, digital system performance is
more easily changed than that of an analog system. While the NR system we've
described is useful to illustrate several advantages of using the digital domain,
another process is useful to illustrate how a digital system is more easily modified.
And that's an analog building-block, the all-pass filter.
2-12
Easily Modified Performance

The operational amplifier schematic in
Figure 2-5 shows an "all-pass" filter. We won't discuss the practical uses ofthe
all-pass filter. Instead, we will contrast the circuit's ability to manipulate certain
parameters of its input signal. Then we will indicate some possible uses for the
circuit.
R R
Signal >~-_-'\N'v -_-----'VV'v--<~_C>
...... Signal
In Out
I
ITOw)1 =
11 • 'wRCI
J =
-J (1)' + (·jwRC) 2
= 1
11+jwRCI -J(1)'+OWRC) 2
Figure 2-5
All Pass Filter
The equations in the figure reveal two

performance aspects of the all-pass network. One, quantity, the magnitude ofthe
network function, is equated to a value of one. This means this circuit produces
a signal whose amplitude is identical to the input. Normally, an amplifier with this
characteristic would be called a voltage follower. But this is not a voltage follower.
The second mathematical expression the figure shows is the time delay of a signal
passing through this circuit will be affected as a function of the input frequency
of that signal. The expression given is called the group delay of the network.
Let's ignore any reason why we would
change the time delay of a signal- we'll get back to that later. How would we
duplicate this same type of function in the digital domain? One answer is a simple,
single-stage delay illustrated in Figure 2-6 . Let's walk through the description of
this diagram. The input signal is digitized with an ADC, processed by the CPU,
and output to a DAC. To realize a delay, the processor must perform three steps:
(a) Move data from a RAM location to the DAC.
(b) Move data from the ADC to the RAM location.
2-13
(c) Go back to step (a).
7 0 7 0
"A"IIIIIIIII.a ~.IIIIIIIII
a) Move data from
"A" DAC.
location to
• b) Move data from
ADC
to location "A".
c) GO BACK TO STEP "a"
Figure 2-6
Single Stage Delay
That approach is straightforward. How

long is the time delay? It's however long it takes the CPU to perform steps (a),
(b), and (c). How long is the delay of the analog circuit compared to this? Well,
I'm not sure but I sure don't want to figure it out! But the expression for the
analog circuit's time delay is much messier than the digital version. What if we
wanted to change the time delay value? Basically we only have to shuffle the ADC
samples in RAM around a bit more than we are currently doing.
As shown in Figure 2-7 , to achieve a
longer time delay, lengthen the RAM buffer from what was previously one
location to "n" locations and employ the following procedure:
(a) Transfer the last ("nth" sample) RAM location to
theDAC,
(b) Move each previous sample into the next higher
sample position of the RAM array, and
(c) Fill the top position with the latest ADC sample.
2-14
7 o
a) Take last sample in buffer

and send to DAC
\• • • •
7 o o 7 o o
w ~
~
samples ~ .,
b) Move each lower sample

into the next higher sample
starting at the "bottom".
c) Get latest sample from

the ADC and place at the
"top' of the buffer.
~ •
7 • o} "n"
samples
d) Return to step "a".
Figure 2-7
Longer RAM Buffer is Used to Achieve Longer Delay
To change the delay time of this digital

delay, we only have to change the number of bytes in the RAM array for the ADC
values. The source code for implementing this digital delay line on the Motorola
MC68HC05 MCU is shown in Figure 2-8 .
2-15
**************************************************************
* Name: Ti meOe I ay
*
* Purpose: To illustrate a 61 sample digital delay,
*
**************************************************************
w PRGEORRM
7 o
TempSmp I rmb I
TimeBufr rmb 64
TBufEnd equ *
se8 Program
•
Ti meDe I ay sta TempSmpl
Ida TBufEnd
;this contains the latest
; ROC !amp Ie, Save it,
; Get very last samp Ie,
•••
MoveSmp I s
sta
I dx
Ida
sta
dex
ORC
ITBufEnd-1
O,x
I, x
;Send it to ORC,
;start moving data from here,
; get samp I e
; move it down t he line,
TBulEnd_
cpx 'TimeBufr-1 ;check for latest sample location,
bne MoveSmp I s
inx ; leave x point ing to TimeBufr
Ida RDCResu I t
sta O,x ;put latest result into buffer,
rts
Figure 2-8
Implementing Digital Delay Line with 68HCOS
How do we change the time delay char-

acteristics in the analog all-pass filter in Figure 2-5 (page 2-13)? A circuit with
an approximate time delay of lOOllsec, for most frequencies, is shown in Figure
2-9. Although the expression for the first analog all-pass filter's time delay (Figure
2-5) wasn't too complex, it was nothing compared to this all-pass network's
2-16
expression. The transfer function for this mUlti-opampcircuit is so messy we won't

present it here, let alone attempt to modify the circuit elements to obtain a different
time delay! Can it be done? Yes. Would I want to do it? No. The digital implemen-
tation is much more convenient to modify.
1430Q
Q 01~F
2740Q
100 kQ 10.0kQ
.01 ~F
1---.--1+
10.0 kQ
2150Q
Figure 2-9
100 Ilsec delay all-pass filter
Uses Of Time Delay

We've skirted answering the question,
"Why would anyone want to delay a signal?" Here's One answer that's to the
point is: to delay signals. Another, is many analog filters mess up a signal's group
delay that can be corrected by an all-pass network. In the past when passive audio
signal equalizers were the only game in town, some people believed the signal
would sound better if the weird time delay characteristics of some equalizers were
corrected.
A more contemporary use of the time

delay function is to hide the fact that most rock'n'roll singers can't sing. Yet, even
this is not the primary use for such a device.
Time delay is an integral part of the

process of averaging data. We saw this first hand with the thermostat described
earlier in this chapter. But at that point we didn't call the data movement "time
delay," and that's exactly what we're doing with the data. Look at Figure 2-10 .
The left side of this figure is the HCOS source code from Figure 1-11 in Chapter
2-17
1, and the right side is the averaging routine from our discussion on filtering
thermostat data. The heavy-lined rectangle highlights the common code segments
between the time delay routine and the averaging routine. The time delay code is
exactly duplicated in the lowpass filtering routine. Time delay is an important
and integral part of the filtering process.
**************************************************************
* Hame:TimeDelay
*
* Purpose: To illustrate a 64 sample digital delay.
*
**************************************************************
seg PAGEORAM
TempSmpl rmb
TimeBufr rmb 64
TBufEnd equ *
seg Program
TimeDelay sta TempSmpl j th is conta i ns the latest

j ADC sample. Save it.
Ida TBufEnd jGet very last sample.
sta DAC jSend it to DAC.
Idx IITBufEnd-l jstart moving data from here.

MoveSmpls Ida O,x jget sample
sta 1, x j move it down t he line.
dex
cpx IITimeBufr-l jcheck for latest sample location.
bne MoveSmpls
inx j leave x pointing to TimeBufr
Ida ADCResuit
st a 0, x j put Iat est resu Itint 0 bu ff er.
rts
Figure 2-10
Time Delay is an Integral Part of Averaging Data
2-18
We hope you are convinced, if not enthu- A Perfect

siastic, about "doing things digitally." After all, digital techniques appear to have Solution?
a lot of advantages over many analog methods. So, it appears there is an ideal
solution to most problems. Well ... not quite. There are some disadvantages to using
digital processes. Some are practical considerations like cost, and others are
problems inherent in the process of digital conversion and manipulation. Let's
take them one at a time. These are not listed in order of importance since their
significance depends on your system.
Limited resolution of data. Even the

output of high resolution A-to-D converters, like 18- and 20-bit converters, are
still only approximations of the real-world. The process of sampling with limited
resolution and finite sampling intervals produces artifacts not present before
converting the analog signal. Usually, the artifacts are unwanted and must be
minimized to avoid corrupting the system's specified performance.
Limited bandwidth. The maximum sig-

nal frequency that may be processed by a digitally sampled system is considerably
lower than that of an analog system. Many analog system frequency responses
run into the gigahertz region, while high-resolution (8-bits or higher) systems are
pushing the current technology to run in the area of tens of megahertz. Because
of this technology/performance ceiling, we must deliberately minimize the band-
width of the incoming signal to avoid sampling artifacts. This may severely limit
a purely digital implementation, depending on the application. The bandwidth
issue is commonly linked to the issue of limited resolution. As the speed of data
conversion goes up, the resolution tends to go down. Whereas you may obtain
30-megasamples per second from an 8-bit converter, a l6-bit converter may not
produce the same conversion rate. The two issues are inversely related.
Limited dynamic range. Dynamic range

is the measure of a system's ability to handle very small and very large amplitude
signals. There are really two reasons the dynamic range is limited with digital
systems: the inherent resolution limitations of the ADC process and the arithmetic
characteristics of the given CPU. If an 8-bit converter is used, for example, then
the maximum ratio of the smallest to largest signals that can be digitized is 255: 1.
A 16-bit system, on the other hand, can handle a maximum ratio of 65535: I! The
dynamic range is significant. If an input signal is larger or smaller than the
dynamic range, the signal will be either completely ignored or severely distorted
since the signal isn't large enough for the digital system to "see" it.
2-19
Dynamic range is also detennined by the

arithmetic capabilities of the CPU. Given identical throughput and functional
performance requirements, an 8-bit CPU will handle a wider input signal dynamic
range than a 4-bit CPU and produce less erosion of the dynamic range. For
example, let's add the number "I" twenty times to itself. After the sixteenth
addition, the 4-bit CPU would either overflow or invoke a special overflow
handling routine. Doing the same thing on an 8-bit CPU requires nothing special
until the 255th addition! Although there are other, more subtle arithmetic processes
like convergent rounding that affect the dynamic range and precision of a digital
system, quantizing the arguments degrades dynamic range the most.
High cost for high resolution and
throughput systems. Basically, the more closely we're able to digitally represent
a signal and/or at a higher bandwidth, the more money it is going to cost. When
conversion time is constant, then the higher the resolution conversion, the greater
the cost. With system throughput remaining constant, the higher the bandwidth
you need, the greater the cost. These are detennined by the practicalities of making
high-speed, high resolution ADCs and CPUs.
2-20
3
C HAP T E R
The Representation of Signals
N ow that we know how various

ADCs perform their magic, it's
time to talk about some different
Repre-
sentationof
Signals
ways of looking at wave forms. This is a topic that many of us take for granted,
and yet it is as fundamentally important to understanding signal processing as
Ohm's law is to understanding AC and DC circuits. If you're a practicing engineer
or even just have some experience poking around a circuit with an oscilloscope,
you've seen a signal's familiar voltage versus time representation on the 'scope.
Well, whether you're experienced or not, we will now stop taking this repre-
sentation for granted. There is information we've overlooked to help us understand
digital manipulation of signals a bit better. At the worst (or at best, depending on
your view), we'll avoid some problems because we won't make certain assump-
tions.
As just mentioned, voltage versus time
is probably the most common view. Our lab instruments continually enforce this
view by showing us how signals change over time. Consequently, we develop
skills and, particularly, thought processes centered around this view. How many
times have you watched the rise time of a square wave degraded by an "integrator"
similar to Figure 3-1 ? Or observed the characteristic tilt of the high-pass differ-
entiator on a square wave or a tone burst as shown in Figure 3-2? In these cases,
the type and extent of filtering is evaluated according to the time domain view.
There is another way. You might ask, "Why do we need another view? After all,
our trusty 'scope usually helps us solve the problem and what's wrong with my
thought processes anyway?!" As we will see, the idea is to make things easier for
us and to make us more effective. The concepts we are about to discuss will be
very helpful as we start to move from the analog realm into the digital world.
Figure 3-1 Figure 3-2

An Integrated Square Wave A Square Wave Through A High
Pass Filter
3-1
Chapter 3 - The Representation of Signals
There are really many different ways of

looking at the same signals. In fact, in many cases different views can provide
much more information about the signals in question than the 'scope view or, as
it is also called, the time domain description. There is at least one new and exciting
way of looking at signals in addition to the standard time domain view. For
example, let's recall the log function. Aside from giving us grief during our first
calculus courses, this function proved to be very useful in the early days of
computers. Then a computer could only add and subtract. (According to all you
purists out there, it is really only addition, since subtraction is addition with signed
numbers.) Any multiplication and division had to be performed by sequential
additions or subtractions. You may even recall the restoring and nonrestoring
division algorithms, which used iterative subtractions and were popular for such
machines.
In any case, a much faster and non-itera-
tive way to do the same type of functions was to map the arguments into the log
domain, that is, to use the identity: Log(xy)=Log(x) + Log(y). By taking the
exponential of both sides, we're left with xy=l O(Log(x)+Log(y». Roughly translated
into English, this means that we can determine the product of two numbers (x and
y in this case) by adding the log of the numbers and exponentiating the result.
The advantage to this approach, in this example, is we can multiply two numbers
by using the addition capability of our computers. "So what?!", you say. What
does this have to do with the representation of signals? Simply this: by changing
our view of the numbers, i.e., by mapping into the log domain, we are able to
quickly perform a function that our computer otherwise isn't capable of perform-
ing quickly. In other words, we can do something relatively easily with the
numbers which we couldn't do easily before, just by remapping the problem
from one domain to the other.
So, what is this new and exciting way of
viewing signals? Sad to say, it's neither new nor very exciting, but it is extremely
useful. It's the frequency response view. Yes, the frequency response view can be
a very compact and useful way of characterizing a system's performance. If you're
a stereo enthusiast, you know your equipment's frequency response is a key
measurement of your system's ability to accurately reproduce your favorite music.
One thing you may not have realized, however, is the frequency response of your
stereo is a shorthand method of describing the stereo's response to a whole bunch
3-2
of time domain signals. For example, my stereo, if excited with a 50 hertz sine
wave of .775 Vrms, will put out a 50 hertz sine wave of such-and-such size.
Imagine doing this for a collection of sine waves of various frequencies. It might
look something like Table 3-1.
Table 3-1
25Hz 50Hz 100Hz 200Hz 1kHz 4kHz 10kHz 20kHz
.700 .780 .905 .825 .790 .803 .653 .420
Table 3-1 certainly illustrates the ability

of my stereo to handle each frequency, but it also took a while to fill in this table.
The table is also hard to read and understand. A better way to show the same
information might have been to graph the magnitude of the sine waves, at the
output of my stereo, versus its frequency value, as shown in Figure 3-3. Easier
to read, isn't it? These two ways of showing stereo performance are also equivalent
in meaning. Notice that the frequency response plot is merely another way of
describing how the stereo responds to sine waves.
1.0
0.9
0.8
0.7
en
::2; 0.6
es 0.5
gJ
0 0.4
>
0.3
0.2
0.1
0.0
10 100 1000 10000 100000
Frequency (in Hertz)
Figure 3-3
Frequency Plot of a Stereo Amplifier
3-3
If there were nothing to be gained by the

frequency response view, it would not be worth discussing. Since we would never
put anything wasteful in this book, this view must offer a lot of potential gain.
Earlier in this chapter, we discussed the use of averaging to effectively filter noisy
data. Would you believe there is an easier way to filter the noise which uses our
new frequency response representation? The answers to the following two
questions give a clue to this easier way:
• What answer do you get when you multiply a number by 1?
• What answer do you get when you multiply a number by O?
The answer to the first question is, of
course, the number itself, and the second answer is zero. Now, imagine taking a
frequency response, such as shown in Figure 3-4(a), and mUltiplying it by the
rectangular shape, as shown in Figure 3-4(b). In other words, for any frequency
that has nonzero amplitude in both figures, we plot the first frequency response
in the resultant graph. Otherwise, we plot a 0 in the resultant graph. After we
perform this short procedure, we end up with the graph shown in Figure 3-5.
fx= 5kHz

(a) An Input Signal Spectrum and Result of Multiplying Fig. 3-4
(b) A Low Pass Filter Spectrum (b) by Fig.3-4 (b)
The plot of Figure 3-4(a) contains only

those frequency components at or below 5 kHz. After mUltiplying the original
frequency response, containing frequencies up to 15 kHz, by a rectangular pulse
with a width of 5 kHz, we end up with a frequency response containing frequency
components no greater than 5 kHz. In other words, we have low-pass filtered the
original frequency response plot. At this point, you would probably guess that it
is possible to filter a signal by simply multiplying it's frequency representation
by a rectangular shape. Your guess would be perfectly correct! Couple this idea
with your knowledge of how effectively and easily we can multiply numbers with
microprocessors these days.
3-4
Before we start multiplying, there are

actually two challenges at this point:
• Transform the real-world signals from their normal, analog state into a
form that a microprocessor can manipulate.
• Transform the digital time domain data into a frequency domain repre-
sentation.
We already covered converting analog
signals into digital ones a little earlier. Let's discuss time and frequency domain
converSIOns.
We have, at least implicitly, agreed that
the scope and frequency response representations are equivalent. If they really
are (and they are!), then the process of going from time-like data into frequency-
like data must also be reversible. So maybe instead of transforming the time-based
data into the frequency domain, we could translate our rectangular frequency
response into time-based data. Either approach is valid because the time and
frequency domain representations of the signal are equivalent. From a practical
view, the point is to identify which representation is easier and/or cost effective
to manipulate. In the earlier thermostat example, we actually implemented the
filtering function in the time domain. We transformed (unknown to most of us at
the time) the desired frequency response shape into a set of time-based procedures
consisting of additions and multiplications. If we had graphically presented the
actual frequency response shape (which we hadn't done at that point), you would
have seen something like Figure 3-6 - much less stark and extreme than the
rectangular frequency shape. This "rounder" frequency response shape ultimately
translates to a more gradual slope in the transition band - a less effective filter.
Figure 3-6
Low-pass Frequency Response of Thermostat
Averaging Filter of
3-5
Time To Frequency And Back Again

What is it that allows us to move from
the time-data to frequency-data and vice versa? As we saw with the logarithm
example, there are many potential remapping techniques. However, the most
widely accepted mapping technique between the frequency and time domains is
the Fourier Transform (Fr). A function has two representations: time domain
and frequency domain. The combination of these two representations is called a
transform pair.
Believe it or not, you don't really have to
know the details of the Fourier Transform for it to be useful to you. Though it
would probably benefit you to work through the math of an Fr at least once, we
won't go through it here. You may review the FT in Chapter 4 which shows the
actual mechanics of the Fourier Transform. To become somewhat conversant in
manipulating signals in the digital domain, we need to discuss some of the results
of applying the Fr.
By now, we've presented several con-
cepts possibly new to you. Without jumping ahead in the book, you may have to
take the concepts and their manipulation on faith. However, if you are squeamish
about "taking things on faith," then please read ahead to Chapter 4. There you
will find the detailed logical argument to thoroughly understand and accept the
concepts. Otherwise, continuing here will at least expose you to some very
important concepts and "grease the skids" for later. Just seeing the application
of the FT, without the nitty-gritty mathematical details, can be immediately useful
to you. We will now present some of the more useful transform pairs along with
some important signal constructs.
3-6
The Impulse
An impulse is a concept represented by
a vertical arrow:
f
The impulse is a periodic event the con-
verter uses to initiate taking a sample of the incoming analog wave form. It also
can represent a sample clock, like that used by the ADC process. Actually, a single
impulse is not used to model the ADC, but a "train" of impulses is used. The train
of impulses is usually regarded as stretching infinitely in time. Two useful
characteristics of these impulses are:
1) The magnitude can be 1, and when multiplied by
the input signal gives us a rudimentary under-
standing of the process of digitally sampling ana-
log signals.
2) Also, the time-based infinite train of impulses
gives us a corresponding infinite train of impulses
in the frequency domain, and the distance between
impulses is given by the inverse relationship
f=I!f, where "f' denotes the frequency and "T"
the period of the time signal.
Looking at this graphically as in Figure
3-7 , we see as the distance in the time-based train of impulses is decreased, the
distance of the frequency-based train of impulses increases. Conversely, as the
distance between the time domain impulses increases, the distance between the
frequency domain impulses decreases. The importance of this characteristic will
become evident shortly. Keep it in mind.
3-7
~
I-- 31.---J
...
":1 t
!
t... ..
i
j f4- ~T
4 t t t t t t t t f: t
• i
1--.57T --I I-- 1.751 0
.:ltt.... t.t ..t.:: t ... t .. J..... J::

Frequent Domain I
Figure 3-7
Relationship Between Time & Frequency Impulses
Multiplication
Earlier we discussed the frequency re-
sponse of a stereo system. We went through a brief exercise of multiplying a
frequency response by a rectangular shape. We just discussed how mUltiplying a
train of impulses by an analog signal models the process of sampling that signal.
Based on just these two scenarios, the process of multiplication in either the time
or frequency domain is useful. But are they really the same thing? That is, is
multiplication in the frequency domain the same type of multiplication in the time
domain? Let me say definitively: yes and no. An explanation is in order.
The mechanics of multiplication are the
same regardless of the domain you are working in. The same old multiplication
you and I have been doing for many years now is implemented the same in either
domain. That is not to say if we multiply something in one domain we are also
simultaneously mUltiplying in the other domain. On the contrary. Remember we
are using the FT to go between the time and frequency domains. Because of this
mapping process, multiplication in one domain translates into a different opera-
3-8
tion in the other domain, just as adding in the log domain results in multiplication
in the linear domain. The results are always the same. It's just that the steps, or
mechanics of the processes are different depending upon which domain you are
working in.
Convolution
If you derive an output in the time do-
main by multiplying a signal by a function, what is the equivalent operation used
on that signal and function in the frequency domain? It's called convolution. If
you applied convolution to an input and a function in the time domain, what is
the corresponding operation used in the frequency domain? It's multiplication.
In DSP lingo:
y(t) = x(t) * h(t) => Y(f) = X(f) * H(f)
and
Y(f) = X(f) * H(f) => y(t) = x(t) * h(t)

Armed with this technotrivia, you can
liven up your convoluted technical discussions with all kinds of bad puns like this
one. Once you know the mechanics of a convolution, you can put it to work
livening up your MCU-based product designs. Let's look at the mechanics of
convolution by looking at a special case involving a train of impulses.
Simply stated, convolving a wave shape
with a train of impulses duplicates that wave shape for every frequency there is
an impulse. For example, let's convolve, in the time domain, the following two
signals: a rectangular pulse and a train of impulses as shown in Figure 3-8. Now
apply the convolution "rule" we just stated on the rectangular pulse and the train
of impulses. Take the rectangular pulse of Figure 3.8(a)and "paste" its shape at
every frequency having an impulse (Figure 3.8(b). Notice that just as the input
rectangular pulse was centered around the vertical axis, a rectangle is centered at
each impulse after convolution. If done properly, we end up with the graph in
Figure 3-9. Not very informative or useful you say? True ... but that's just an
example. Let's take it one step further with an example with a more useful
outcome.
3-9
.... l . . . . t . . . . J"""
,
Figure 3-8
Rectangular Pulse (a) and a Train ofImpulses (b)
Figure 3-9
Convolution of Fig. 3-8 (a) & 3-8 (b)
3-10
Take the sine wave in Figure 3-1 O( a) and

multiply it by the train of impulses Figure 3-1O(b). The result is like that in Figure
3-11. Obviously we have a train of impulses with the original sine wave's shape.
Now, convolve this sinusoidally-weighted train of impulses with the rectangular
pulse shown in Figure 3-12. What's the result? The time-domain plot shown in
Figure 3-13. You astute observers will recognize this as the sample-and-hold
function applied to a sine wave. The rest of us, particularly the management types,
will describe the resultant wave form as "a bumpy sine wave." Who am I to argue?
.. ...tlttttltt~tttt!ttt ..·~
",
~
Figure 3-10
A sine wave (a) and a train of impulses (b) to sample it
ft
Figure 3-11
The Sine Wave of Fig. 3-10 Multiplied by a Train of Impulses
n n
II nil nil
Ii' ...J I I I ..J III
I III IIIIII IIIIII
I I I I II II I I I I
IIII IIII
U II UII
w

A Short Rectangular Pulse Sine Wave of Fig. 3-10 (a) after Sam-
3-11
Nevertheless, this convolution's result

closely represents a real world sample-hold function sampling a sine wave. As we
discussed earlier, the sample-hold function is usually an analog to digital con-
verter's first step in the process of going from the analog to the digital world. By
using the impulse's characteristics and convolution mechanics, we can model
what happens to a signal in the analog to digital process.
A Quick Convolution Example

If you understand the concepts presented
thus far, you actually have enough information to start processing real-world
signals. Convolving functions in the frequency domain is the real-world process-
ing for which we're prepared. Summarized below are the FT's features and
relevant facts discussed so far about the frequency domain:
Important point #1: Convolving a rec-

tangular shape with a train of impulses results in duplicating that rectangular shape
everywhere there is an impUlse.
Important point #2: A rectangular

shape centered around the vertical frequency axis in the frequency domain is a
low-pass filter.
Important point #3: MUltiplying a sig-

nal by a train of impulses in the time domain represents the digital sampling
process for an analog wave form.
Given these tidbits of information, try to

determine the answer to the following problem:
Convolve a 1OOO-hertz wide rectangular shape cen-

tered at zero frequency, with an impulse train in the
frequency domain. What minimum distance be-
tween frequency impulses is required to guarantee
no overlap between the resultant rectangular
shapes?
The answer follows this reasoning:
First, the lOOO-hertz wide rectangular
shape is centered around zero frequency, so it stretches SOO hertz on either side
of the vertical axis. Once the impulse train is convolved with the rectangular shape,
it will stretch SOD-hertz on either side of each impulse.
3-12
We know that convolving a train of im-

pulses with another function basically duplicates that function everywhere there
is an impulse. Therefore, the result of this impulse train convolved with this
rectangular shape is a series of 1000-Hz rectangular shapes, centered at each
impulse of the impulse train. The minimum distance between the duplicated
rectangular pulses will be 500 hertz, for a given impulse, plus 500 Hz for each
adjacent impulse. As you probably know, 500 Hz plus 500 Hz equals 1000 Hz.
Thus, in this case, the minimum distance between impulses, to avoid overlap of
the rectangular shapes, must be at least 1000 Hertz.
What have wejustreasoned through? It's

called the Nyquist Criteria for digitally sampled systems. You've probably heard
of it. It's an important operating consideration when working with digitally
sampled systems. A typical textbook definition of the Nyquist Criteria states
that our system sampling frequency must be greater than twice the highest
frequency component of the input signal. Why? When this "rule" is not followed
(as in the example above), the frequency spectrum associated with one impulse
starts to "collide" with the frequency spectrum associated with the impulse next
to it. Overlapping spectra is not usually something that we want, because over-
lapping spectra means the spectra are interfering with each other. In fact, the
overlap creates frequencies not present in the original spectrum. These products
of overlapping spectra are called aliased components.
By applying the rudiments of the Fourier

transform, we've seen all digitally sampled systems are bound by two important
concepts: The Nyquist Criteria and aliasing. Despite their importance, we have
been able to deduce their existence, and we've even guesstimated a quantitative
system impact.
The Rectangular Pulse
Most of this material may appear to be

"black magic," but I assure you the rules and transform pairs we have discussed
are a direct consequence of the FT. We've only looked at signals graphically, but
a mathematical approach supports the FT as well. Without actually performing
the mathematics of the FT, it's pretty tough to predict the time-frequency pairs.
Time-frequency pairs are the two functions that represent a signal or operation in
both the time and frequency domain. For example, given a rectangular pulse in
the time domain, the FT transformation results in a rather strange-looking, obtuse
function called the sinc function (pronounced "sink"). It looks like the graph of
3-13
The Sine And Cosine Functions

This transform pairis one you've already
seen but probably didn't realize it. In fact, we opened our discussion on transform
pairs with it: the Ff of the sine or cosine function results in an impulse in the other
domain. For a time-based sinusoid, the frequency plot will be an impulse located
at the frequency of the sinusoid like that shown in Figure 3-15. As you may
remember, the difference between the sine and cosine functions is slight. They
are really the same wave form but 90° out-of-phase with respect to one another.
To be truthful, both sine and cosine may simply be represented with a single
impulse arrow as shown in Figure 3-15. That's only part of the story, however.
There is an equivalent frequency domain representation to designate the phase
differences between the two, but we'll leave that for further discussion until
ChapterXX. It is sufficient at this time to represent either time-based function
by a single impulse in the frequency domain.
f
Figure 3-15
The Fourier Transfonn of a Sine or Cosine Wave
Frequency Representation of Direct Current (DC)

A DC wave form is really no wave form
at all. Electrical engineers understand DC to mean the current or voltage amplitude
does not vary with time-there is zero frequency. This produces a horizontal line
on an amplitude versus time plot as shown in Figure 3-16. The equivalent
frequency domain is a single impulse located at the zero frequency point.
3-15
Figure 3-16
DC and Its Spectrum
The consequence of this is twofold: a) A

train of impulses that has an impulse located at the zero frequency point means
there is a DC component in the time domain signal, b) A horizontal line in the
frequency domain is plotted in the time domain as an impulse at t=O. This is an
important point. The implication is a horizontal line in the frequency domain
suggests that all frequencies are at the same amplitude. We interpret the single
time domain impulse as a very short-lived signal. A short time domain impulse
produces a frequency plot containing all frequencies.
Time Or Frequency: It's All The Same

Up to this point we've freely inter-
changed a signal's representation in the time and frequency domains and vice
versa. That is, if a rectangular pulse in the time domain results in a sinc function
in the frequency domain, we've also assumed the opposite was true: a sinc function
in the time domain produces a rectangular wave shape in the frequency domain.
We haven't attempted to prove this, we've just casually presented this as accept-
able. Fortunately, it is not only acceptable, it's correct. This pairing of wave
shapes, regardless of the domains in which they reside, is called duality in signal
processing lingo. This is actually one of the most important consequences of the
FT. Keep it in mind as you would the transform pairs we've discussed - it's very
important.
3-16
Other Important FT Stuff
We've only talked about one of the prop-

erties of the FT so far: duality. There are many more. We can get along well, for
most day-to-day engineering work, without these additional details. This is
primarily because some of the properties, like the linearity property, are so
common in engineering that we assume it applies here too. Most of us haven't
been concerned with other properties, like the time and frequency shifting
properties, until now. What follows is a list of the properties with a short
explanation.
Linearity property. The FT of the sum

of two functions is equivalent to the sum the FT of each function.
Because the FT is a linear function, many

of the other arithmetic properties also apply and often come in handy when
solving problems. Two of those properties I've found particularly useful are the
commutative and associative properties.
Symmetry property. This is another

term for duality.
Time scaling property. This is one of

the properties establishing the inverse relationship between the time and fre-
quency domains. Basically, this property states that if our time variable "t" is
multiplied by a constant "k" then the transformed "f' variable is divided by the
constant "k." As it turns out, the amplitude of the transformed function is also
affected by this property.
Frequency scaling. A complementary

statement of the time scaling property starting at the frequency domain. You could
"prove" this by applying the time scaling and the symmetry properties. In
symbols, that is, frequency scaling symbols.
Time shifting property. This property

is responsible for the fact that time delay in the time domain is equivalent to phase
shifting in, you guessed it, the frequency domain.
Frequency shifting property. The con-

cept of modulation is a direct consequence of this property, which states a
frequency may be shifted by multiplying the corresponding time domain function
by a certain exponential function.
3-17
Other properties. Even functions, odd

functions, real functions, and imaginary functions are properties that, at this point,
require little more than this quick acknowledgment. Although ultimately very
important to those of us who will become deeply involved in signal processing,
these are not absolutely necessary to apply signal processing to many engineering
problems.
Putting It All At the end of this book is a list of some

Together really informative books (not including this one of course), where the curious
among you may delve more deeply into these things.
We will now put our new knowledge to
work by taking the concepts we have learned so far to develop a digital filter. We
will develop the details of the digital filter we constructed for the thermostat that
is the averaging filter. Although this is a very specific filter, the mechanics
presented are useful for constructing any type of digital filter that does not utilize
feedback in the filtering process.
Earlier in the chapter we said filtering an
input signal could be accomplished by multiplying the input signal spectrum by
a rectangular wave shape. But, while we didn't indicate how to obtain the input
signal spectrum, we did infer that maybe we didn't have to. Instead, we assumed
we could work in the time domain. By working in the time domain we avoid the
transformation of an input signal into a frequency spectrum. Then we found the
rectangular low-pass filter shape in the frequency domain was identical to the sinc
function in the time domain. The rectangular shape and the sinc function are
transform pairs. These concepts give us a good starting point to build a digital
filter.
Let's recount these principles and state
them the way you might see them in a signal processing textbook.
Multiplication oftwo frequency spectra
in the frequency domain is equivalent to the convolution in the time domain of
Uhe inverse FT of each spectrum. This is the kind of statement you see in a
signal processing textbook. It's short, concise, and confusing. We'll translate it.
Multiplication oftwo frequency spectra
in the frequency domain. In the case of our low-pass filter, this means we will
mUltiply the rectangular wave shape by the input signal spectrum. But we don't
know what the input signal's spectrum is or how to get it. So, we're much better
off working in the time domain. How? ...
3-18
Is convolution in the time domain ofthe

inverse FT of each spectrum. Earlier we discussed specific Fourier Transform
pairs. We said the FT of a rectangular time-based pulse is a sinc function in the
frequency domain. Because of the symmetry property, we also are able to say a
rectangular shape in the frequency domain is a sinc function in the time domain.
The process of translating a time domain function to its frequency domain version
is usually referred to as "taking the Fourier transform" of the time-based func-
tions. The process of translating the frequency domain function into its time
domain version is referred to as "taking the inverse FT" of the function. We have
been freely going both directions because we have been applying the symmetry
property. Thus, all we have to do is con vol ve the sinc pulse with a time-based input
signal to filter it.
Simply stated, if our lowpass filter is the The lhermosIat

mUltiplication of a rectangular wave shape with the input signal's frequency Example Re-
spectra, then we can filter in the time domain by convolving the time-based input visited
signal by time domain representation of the rectangle wave. In this case, it is the
sinc function. But how do we convolve an input signal with a sinc function? We
already have. We have already convolved an input signal with another function.
Remember the thermostat averaging routine and the lattice diagram? Both the
routine and the diagram demonstrate convolution in the time domain! Let's walk
through the details of this thermostat filter implementation.
First, we digitally sample the output of
the analog temperature sensor. We now know we may approximate a signal's
digitization by multiplying the input waveform by a train of time domain im-
pulses. Because we don't really know what the time or frequency domain output
of the sensor looks like, we'll make something up. We'll assume temperature
slowly varies, and for all intents and purposes, it's a DC value. To demonstrate
the characteristics of the filter, we'll superimpose a 60 Hz sine wave on top of the
DC output of the sensor. The output ofthe sensor would then look like this:
From the view of the sensor output, the
functional interpretation of Figure 3-17 is the temperature is relatively stable.
Unfortunately, "riding on top" of the sensor output voltage is some 60-cyc1e noise,
possibly induced by the house wiring. The corresponding spectrum for this time
domain wave shape is:
3-19
Figure 3-17
An Imagined Output of a Thermal Sensor Plus Noise (60Hz)
For the sake of brevity, we will designate

this composite function TempNnoyz(t). The plot in Figure 3-18 shows an FT
property and two of the transform pairs we looked at earlier. The figure reveals
the FT's linearity property. Briefly, the linearity property states the FT of the
sum of two signals is equivalent to the sum of each function's FT.
10 20 30 40 50 60 70 80 90 100
Figure 3-18
Spectrum of Thermal Sensor Output Plus Noise of Fig. 3-17
The first transform pair is shown in Fig-

ure 3-16. It is a constant time function, and its corresponding impulse is located
at the frequency plot's origin, like so:
The second transform pair is a sine wave
and its corresponding FT, an impulse located at the frequency of the sine wave:
3-20
By applying the linearity property, we

end up with the composite spectrum TempNnoyz(t) in Figure 3-18.
Now that the output voltage of the sensor
.... F
has been defined, the signal must be digitized. In this case, we will use a sampling
rate 10 times greater than the frequency of the 60 Hz sine wave. Remember,
I I I I I t I
10 203040 50 60 70 80 90 100 F
I I I ~
Figure 3-19
A 60 Hz Sine Wave and Its Spectrum
digitally sampling a signal can, for our purposes, be adequately modeled by

multiplying the input signal by a train of impulses in the time domain. The time
domain plot of the sampled temperature profile looks like Figure 3-20 on the
following page:
3-21
_~]]lllllllllllllllilli
Figure 3-20
Digitally Sampled Thermal Sensor Output
We'll call this TempNnoyz(nT). Notice
the use of "nT" to distinguish the signal's sampled timebase from the continuous
time designation of "t." What does the spectrum of this digitized function look
like? You are correct if you guessed the composite function of Figure 3-18
duplicated wherever there is an impulse in the frequency domain. This function
is Figure 3-21 and occurs because time domain multiplication translates to
frequency domain convolution. And convolving a wave shape by a train of
impulses duplicates the wave shape everywhere there is an impulse. To sample
TempNnoyz(t), we multiplied it by a train of time domain impulses which resulted
in TempNnoyz(nT). The frequency domain translation of this time domain activity
is to convolve the FT TempNnoyz(t), which is given by TempNnoyz(f), by a
frequency domain impulse train. This process of sampling translates the continu-
ous time function into a stream of numbers. We then need to feed these into the
process represented by the lattice diagram. Let's look at the lattice diagram we
developed earlier:
t+ L t+..,
Spectrum of Sampled A Lattice Diagram
Sensor Output
3-22
Data moves from left to right. The sam-

ples coming from the ADC enter on the left, and they leave our averaging process
on the right. As shown in the lattice diagram, each piece of data held in RAM is
first multiplied by a number or coefficient. If you'll remember from our discussion
placement of this coefficient is a consequence of the distributive property of
multiplication over addition. Suffice it to say in our thermostat averaging exam-
ple:
a) There are 64 "blocks" in our main data
flow in the lattice diagram.
b) All 64 of the coefficients happen to be
the same. They all had the value of 1164.
The data moves through the lattice dia-
gram so that each data sample is shifted left to right, once each sample clock. We
could view the output of each block, called taps, as being one sample clock apart.
Since each sample coefficient has an amplitude of 1164 and a sample clock delay,
we could picture combining these two basic elements for every tap. The plot would
look like this:
Figure 3-23
Representation of Lattice Diagram Coefficients as a Sampled Rectangular
Pulse
Does this look familiar? It looks like a

rectangular pulse multiplied by a train of impulses. Since we are convolving this
wave shape by the digitized input signal, TempNnoyz(nT), this convolution also
looks like the input signal's FT being multiplied by the FT of the sampled
rectangular pulse. The spectra for a sampled rectangular pulse looks like this:
3-23
Figure 3-24
Spectrum of Sampled Rectangular Pulse of Fig. 3-23
So, how does this filter work? If we look

morec1osely at Figure 3-24, we can see the details. Figure 3-25 shows figure3-14
(the sinc function) magnified a little. The graph's left hand side, the "negative"
frequencies, is cropped from view. In addition, it only shows its characteristic sinc
shape up to where it first passes through the horizontal axis. We did this to simplify
the final step of multiplying the input spectra by this wave shape. Compare Figure
3-25 sinc-type spectrum with the shape of the ideallowpass filter:
-'.
Figure 3-25
Sinc Function Low-pass Filter
Figure 3-26
Ideal Low-pass Filter Frequency Response
3-24
The two spectra in this figure are similar.

The amplitudes in the lower frequency are affected less that those in the higher
frequency. But the sinc function's gradual slope implies it is going to be less
effective as a filter. The final step in understanding this averaging type filter is
to multiply the input signal spectrum by the filter's spectrum:
1020304050607080 90100f 10 2030 40 50 60 70 80 90 100 f
Figure 3-27
Multiplying Thennal Sensor Output Spectrum (Plus Noise) by Sinc Func-
tion Results in Reduced Noise Component
We can see from the graph that applying

this averaging process has significantly reduced the magnitude of unwanted noise
in our system.
Digital
It may not be obvious that there is a lot
of underlying mathematics involved in building the averaging filter. Yet it is Strengths Re-
precisely the mathematics that gives this averaging filter its performance and visited
appeal. Solving the problem of the noisy thermostat illustrates its performance,
but how about its appeal? The following indicates the appeal of using the
averaging filter as well as other digital filters:
• The performance of the filter doesn't vary over time or temperature.
• Parametric changes between units are predictable and bounded.
• Performance of a digital domain implementation is easily altered.
This averaging filter exemplifies these

"strengths" perfectly. Here's why: the point where the sincfunction passes through
the horizontal axis is set by the relationship f=1/(2T), where T is the width of the
rectangular pulse either side of the vertical axis in the time domain. By increasing
the ''T'' value, i.e., making the width of the pulse wider, the corresponding width
of the sine pulse becomes narrower. If you needed more roll-off of the noise than
we obtained, you would widen the rectangular pulse. Since the width of the
3-25
rectangular pulse is determined by the number of averaging points, we would have

to increase the number of averaging points to narrow the sine pulse. In any case,
the characteristics of the filter are a function of the mathematics, the number of
points, and the arithmetic used to combine the data and are unaffected by the
ravages of time and temperature.
Averaging Fil- We have just constructed an averaging
ter Summary filter by doing two major things:
Learning new concepts. We learned how
to look at signals from a different perspective; the Ff has given us a new way of
looking at signals. The Ff has also given us a new methods, like convolution, to
give us a very effective means of manipulating signals.
Reinterpreting what we already knew.
The same old process of averaging which many of us have done time and again
was reviewed and translated to include the new concepts brought to us by the
Fourier Transform.
Where Do We Before we started to talk about the details
Go From of the FT, I said we could do useful things without getting bogged down in the
Here? mathematics of the FT. We built a digital filter and have "proven" the Nyquist
criteria and aliasing. Both of these are substantial accomplishments. They indicate
a true understanding of some of the more difficult signal processing concepts like
convolution.
From here, you can apply the type of
reasoning we learned from the averaging filter. This style ofthought and problem
solving will always have its place no matter how detailed your involvement in the
mathematics of signal processing. I must give you a word of caution: although
our accomplishments to this point are very real and very useful, you must progress
to the next level of understanding to be truly creative, and wholly accurate, in
applying signal processing techniques. In order to increase your understanding of
signal processing, I recommend that you read Chapters 4-7 of this book. These
chapters review the concepts we've discussed and presents many others like
Infinite Impulse Response CUR) filters and the discrete Fourier Transformjust to
name two in more detail.
3-26
Then read every signal processing text

you can get your hands on. At first, the mathematics can be intimidating. If you
try to relate the math to the graphical and intuitive treatment of the subject in this
book, you won't get so bogged down or lost. While many of us may be involved
in signal processing for recreation, I believe signal processing is a problem-solv-
ing tool which could enrich all our lives.
Talk to anyone and everyone about what
you are learning. By hearing yourself talk about what you know and what you're
learning, you will solidify your understanding.
3-27
C HAP T E R
Fundamentals of DSP
4
T he purpose of this chapter is to
. present some basic concepts
needed for the use and apply Digi-
tal Signal Processing (DSP). The presentation of this chapter may be different
than the reader is accustom to. I have attempted to present topics through the use
of arguments and concepts. Mathematics is omitted where verbal and graphical
presentations may better convey an idea. For more in-depth applications, you are
directed to one of several texts on the subject.
Before leaping into the digital domain of Continuous

DSP, take a minute to look at the world around us, a continuous real-time world, Time
the analog domain. By continuous, we mean things move from one state to another Systems
by passing through all the states in between. At all times there can be a state
associated with this process. The way the sun moves from sunrise to sunset is a
good example of a continuous time process. The sun does not stay in the same
position all morning and then jump instantaneously to its sunset position. Nor
does it jump 1112 of its journey at the top of each hour. It appears to move
continuously across the sky. This is analog. What we see, hear and touch are
analog signals. This being the case, many readers may wonder why anyone would
be interested in digital (non-analog) signal processing. Digital signal processing's
advantages are made clearer later, but ftrst let's investigate continuous time analog
signals and the concepts of the analog domain.
Much literature is dedicated to the study of Sinusoids

sinusoids. Sinusoids go by many names: sine waves, pure tones, simple harmonic And Spectra
signals, etc. Taking our lead from this, perhaps we should look at the sinusoids
and their spectra. Many of the everyday events involve sinusoids, vibrating guitar
strings, rotating motors, even the apparent motion ofthe sun. The most convenient
example might be the ordinary dial tone of the telephone. This signal is a sine
wave which oscillates back and forth. The word "tone" is used because a person
perceives the signal as a single pitch at 440 cycles each second (440 Hz).
Notice that the dial tone was described two
different ways. The ftrst, as an oscillating sine wave, describes what is happening
with respect to time. This is known as the time domain description of the signal.
The second description indicated something is going back and forth in a cyclic
4-1
Chapter 4 - Fundamentals of DSP
manner and as a tone, tells us how fast or how often it is moving. This is an example
of the frequency domain description of the signal. Both the time and frequency
domain descriptions accurately described the phenomena. To understand digital
signal processing, we will need to accept, understand, and use the two different
yet intimately related descriptions of signals in the time andfrequency domains.
100 200 300 400 500

frequency in Hertz
Figure 4-1 Domain Figure 4-2 Frequency

Description of Cosine Domain Description of
Function Cosine Signal
The Cosine Function

The cosine function is a sine wave with a
phase of 90 degrees. The term "sinusoid" is often used to refer to a cosine function
as well as a sine function. This reference is generally accepted. Thus, for the
remainder of the book, the term sinusoid may refer to either a sine or cosine. At
times one may be preferable to the other; some mathematical relationships are
more eJegantly stated using one instead of the other.
4-2
Figure 4-3
Cosine Function is equal to sine wave

with a phase shift of 90 degrees
Few engineers would argue that we The

describe a dial tone by a time domain description or a frequency domain Frequency
description, that is, a tone with a frequency of 440 Hertz and no other frequencies. Description
Some readers may not be so sure we can give a frequency domain description for
any arbitrary signal. For example, take the signal coming from the needle of your
stereo when playing a favorite record. It has a time domain description that can
be viewed on a scope. We may guess that it also has a frequency domain
description since that's what the bass and treble adjusts on the stereo. What is
really meant when we talk about the frequency domain description of this
complicated signal? Let's use a hand waving argument for a moment. Let x(time)
describe what happens to x over time, that is, how the recording varies with time.
Now let X(frequency) describe how often these "things" happen to the recording
x(time). Frequency is synonymous with "how often something happens." The
question remains, "What should we use as a basis to measure how often these
things occur?"
4-3
We are able to accept that a sinusoid con-

sists of a single frequency. The simplicity of this description makes the sinusoid
the ideal concept to use as a basisforfrequency domain descriptions. This means
that we may think of the complete frequency domain description as a collection
of sinusoids, just as we view the time domain description as a collection of points
in time. It is also possible to view this the other way around: a collection of
sinusoids (a frequency domain description) may be used to construct a collection
of time points. The time or frequency domain descriptions are equally valid and
there is nothing lost going from one to the other. Based upon this reasoning, the
highly complicated waveform of our favorite record does indeed have a valid and
unique frequency domain description.
<=> X(Q ~IIIIIIIIIIr'-,

f·
Figure 4-4 Time Domain Figure 4-5 Frequency
Description of Signal x(t) Domain Description of
Signal X(t)
The Sinusoid Continuing with the frequency domain de-
And The Delta scription of the sinusoid, it can be said that thefrequency domain description (or
Function spectra) ofa sinusoid is zero everywhere except at the frequency of that sinusoid.
At this point we have skirted the issue of what the value of the spectra is at the
frequency of the sinusoid.
Let's not worry about the height of the
spectrum at the sinusoid's frequency. Our concern is with the area under the curve
at the sinusoid's frequency. Let's define the spectra of a sinusoid as having an
amplitude ofzero everywhere except at the sinusoid'sfrequency, where the area
under the curve of the spectra has an area of one. See Figures 4-6 and 4-7 for
an illustration.
A sinusoid with an amplitude of two can
be considered as having zero spectra everywhere except at the sinusoid's fre-
quency where the area under the curve is equal to two. The mathematical notation
for this type of function is known as the Delta function.
4-4
Area under
1 ------
spike = 1
<=>
~ ' - - - v - ... f
zero zero
Figure 4-6 Sine Figure 4-7 Spectra
Wave with
Area under
spike = 2
<=>
'---v-' ' - - - - - v - - ....... f
zero zero
Figure 4-8 Figure 4-9 Spectra

of Sine Wave
Sine Wave with from Figure 4-8
Amplitude 2 and
Frequency 1/2 of
Fig. 4-6
The Delta Function

The Delta junction equals zero every-
where except when its argument is zero. At the zero point, the value ofthe function
is not specified, but rather, the area underneath the curve is specified. This area
is specified with a value ofone. One way to understand this is by looking at various
rectangles. The area of a rectangle is equal to the base times the height. Consider
a sequence of rectangles all with an area of one. As shown in Figure 4-10, start
4-5
with a square (base = 1, height = 1, thus area = 1 x 1). Notice that for each
successive rectangle we have shortened the base and are forced to increase the
height in order to maintain a total area of one. It is in our interest to continue this
process until our rectangle is so narrow that its base is zero. How tall would the
last rectangle have to be to maintain its area of one? In a rough sense, it would
have to have infinite height to make up for its zero width.
Area = 1
I
~}
AREA
~
base = 1
he;ght _ 1
W
A
E
A
-
base
he;ght _
1
= 2'"
2
base
1
= 4""
f
base
he;ght-oo
= 0
Figure 4-10
Delta Function has zero width, infinite height, and area one
If we place the last rectangle on an axis, we
can consider this as a function that was zero everywhere except at a certain point
where the area under the curve is one. This is precisely the Delta function.
Figure 4-11 Example of Figure 4-12 Example of the

the Delta Function Delta Function
It is also worth noting that typically the

Delta function is plotted with an arrow at the top of the spike. This arrow helps
remind us the height is infinite.
4-6
Earlier we discussed that the sinusoid with

amplitude one has an area of one under the spectral spike, and the sinusoid with
amplitude two has an area of two. Since we know the arrow reminds us the height
is infinite, we are free to draw the length of the line to be proportional to the area
under the spike. This is exactly the convention used.
You may now, for the time being, forget

about the Delta function and breathe a sigh of relief. It was included so that some
loose ends can be avoided when the Fourier transform is presented.
The process of specifying a signal is a bit Amplitude

more complicated than saying there is a tone at a particular frequency. To be And Phase
thorough, we need to discuss two other parameters: amplitude and phase. Descriptions
Amplitude is the maximum level a sinusoid reaches anywhere in its cycle. Think
of the sinusoid's height in Figure 4-13. For physical phenomena, the amplitude
can be related to qualities such as loudness, brightness, and voltage. The higher
the amplitude, as seen in the time domain, the higher the amplitude is in the
frequency domain represented as a "spike".
Let's look at phase, a simple idea so often
made difficult. Take a look at the signals shown in Figures 4-13. These are all
sinusoids of the same frequency and amplitude, yet they are obviously different.
Not only is the time domain description we were using up to now inadequate to
resolve these differences, Figure 4-14 shows the frequency domain description
is also lacking. The last descriptor we need to help us out of this quandary is the
descriptor of phase.
Looking once more at Figure 4-13, we see
the only difference between the sinusoids is each starts at a different point in its
cycle. Wave one starts at zero degrees (at t = 0), wave two starts at 90 degrees,
and wave three starts at 180 degrees. Calling this starting point "the phase" of
the sinusoid completes the description nicely. We will also need to provide for
phase in the frequency description as shown in Figures 4-15 through 4-24.
4-7
Figure 4-14
Figure 4-13
A: ~~ x(i)
~~ <->
--~--- -
Phase
I.
Figure 4-15
Figure 4-16
B:
~_u<=>'(fj
Phase
c: x(i)
~~
---~I
---<=> Phase
Figure 4-19 Time Domain Figure 4-20 Frequency Domain
4-8
Importance Of Phase
If the previous description of phase is any
indication, the concept of phase doesn't seem too difficult, then why is it so
misunderstood? The difficulty doesn't begin with the phase, but with the question
of when does t = O. Viewing Figure 4-13, it is easy to know when t = O. Simply
look at the plot, point to where t =0, and you have it. Look at the dial tone example
in Figures 4-21 through 4-23. Let's define t =0 as the exact moment you pick up
the receiver. When the receiver is picked up you can "look" at the waveform,
determine its phase as shown in the figures, and complete the amplitude, phase
and frequency descriptions. Sometimes it can be picked up exactly when the
waveform equals zero and increasing (0 degrees phase, Figure 4-21), and other
times when it equals one and decreasing (90 degrees phase). Simple enough?
Suppose someone left the receiver off the hook before you arrived. You are asked·
to give the waveform description as before. How do you determine the phase?
Well, you can't. Sometimes phase doesn't matter, ~only the amplitude and
frequency are important.
Xa
Figure 4-21
... / ~~""-- ...

Figure 4-22
Xc
Figure 4-23
4-9
Relative Phase
As we just discussed, sometimes it is not

possible to determine the absolute phase of a signal due to the ambiguity of when
t = o. Therefore, we must accept the fact that we cannot determine phase of a signal
and go about our business. However, there's more to consider than just the
absolute phase. We must also consider relative phase. The idea comes into play
when there is more than one signal. Let's use three for illustrative purposes. We
may not know the absolute phase ofany ofthe three, but we can easily see a phase
difference between them. This phase difference is known as relative phase.
The three signals of Figures 4-24 through

4-26 have the same frequency but different phases. We know this because we can
see their peaks and valleys do not coincide. We do not know the absolute phase
of A, B, or C, but we can determine theirrelative phase. Let's arbitrarily pick signal
A as our reference signal. We can find a point on A where the wave equals zero
and is headed positive. Use this point as a time reference and determine the other
two signals' phases in relation to this point. Remember that since we chose the
reference point, it is the relative phase we are determining, not the absolute phase.
Now a quick look at the importance of

relative phase. In Figures 4-27 and 4-28 we added signals A and B (Figures 4-24
through 4-28 to get signal X (Figure 4-28). Notice how signal X has the same
frequency, but the amplitude has become larger. We also added signals Band C
to get signal Y (Figure 4-28). This equals zero all the time. We have two very
different results from adding signals of like frequency but of different relative
phases. Therefore, relative phase is important.
... J K7
Figure 4-24
... ~ ...
Figure 4·25
...~
Figure 4-26
....
4-10
Figure 4-27
... -----...;....------
Figure 4-28
Time Delay And Phase Shift
By now, you've had it with sinusoids, phase, etc. What has this all have to do with
DSP? It is time for a technical teaser, a simple concept that is at the very root of
filtering (including digital filtering). It also starts us thinking about the interplay
between time and frequency. Consider Figures 4-29. The sinusoid shows a phase
equal to zero (solid line). Also found is a delayed version of the same (dotted line).
The time delay between the two is "t" seconds. We can visually estimate the
relative phase between the two to be approximately 3/4 1t (about 120 degrees).
Let's add the delayed version to its original and look at the output. Notice the
amplitude has been effected. The original and delayed amplitude both had an
amplitude of one, but their sum has a smaller amplitude.
Continuing the teaser, observe the same

experiment in figures 4-31 and 4-32. The only thing different is the frequency of
the sinusoid. The time delay is the same but the relative phase is different! We
1t
will estimate the relative phase to be 2" here. Again, summing the two, we find the
amplitude of the result is different from the inputs (Figure 4-32). But more
importantly, we see the resulting amplitude is different from that offigures 4-30.
The frequency dependency of the relation-

ship between the inputs and outputs of our simple setup shows the characteristics
of a filter. For our constant time delay of ''t,'' there are input frequencies for which
constructive interference occurs which results in larger amplitudes, and there are
other input frequencies for which destructive interference occurs which results in
smaller amplitudes. This futering effect is made possible by the fact that for a
constant time delay, the relative phase is dependent upon the inputfrequency.
The frequency dependency of the relative phases can be used to affect different
frequencies in different manners.
What about DSP? The experiment shows
the makings of afi/ter. The first operation was a delay, the second the summation.
Although DSP hasn't actually been discussed yet, those of you with an inkling
may see the two operations (delay and addition) are the types a computer (or DSP)
would be very good at. To delay, one could store something in memory to be used
again later. To add, well, it should be obvious that computers can be somewhat
useful with addition and storage.
Fig u re 4-30
Figure 4-29
1t •••••••••••••••••
LlO=2
Input
L .. d::>-Output
I Delay ~
4-12
Frequency Descriptions
We have been adding signals of the same
frequency to each other without a problem. Let's investigate two signals of
different frequencies. Figures 4-33,4-35, and 4-37 show signal A, alowfrequency
sinusoid, added to B, a higher frequency sinusoid, to produce C which exhibits
sinusoids of both frequencies. The time domain plot of C visibly shows the two
frequencies. The large amplitude-slow fluctuations are due to A, while the small
and fast fluctuations are due to signal B. The frequency domain representations
may also describe the phenomena here Figures 4-34 through 4-36).
r\A
Xa(t)
A:
Figure 4-33
<=>
11------Figure 4-34
---i~
Xh(t)
B:
r---
Figure 4-35
<=>
J----.--Figure 4-36
f--
Xc =Xa +Xb
c: r'tFtFigure 4-37
<=> LL-L- Figure 4-38
.1'~
D:
. .ft;r~J\:. <=> Jlt,tcu ___._
4-13
Notice how the frequency domain plots are

very simple and quickly show signal spectral components. To further the point,
consider signal D (Figure 4-33). Can we determine what's really happening here?
Not easily. However, the frequency domain representation ofD clearly shows its
components (Figure 4-40)
Sinusoidal Until now, we have been discussing the
Basis of sinusoid. The greatest advantage of this signal is it is so easy to determine the
Signals frequency domain description: a single spectral line with its height proportional
to its amplitude. But, hasn't this been a waste of time since "real world" signals
are so much more complicated? Not really.
A classic example of this is the square
wave. Figure 4-41 shows how a square wave can be built by adding the sinusoids
of the proper frequency, phase, and amplitude. Moreover, from this construction
we are given a sense of what spectral content of the square wave is. Notice how
we get closer to the desired square wave with each additional step. Using all of
the odd harmonics of the original sinusoid, we eventually get a square wave. This
means that a perfect square wave needs infinite terms, thus, has infinite spectral
content. But as you can see, the amplitude needed for each addition gets smaller.
So, for many purposes, a square wave
which is "good enough" can be achieved with a limited number of spectral
components. The need for an infinite number of sinusoids where only a limited
number may be physically tolerated is called the Gibb's Phenomena. This is why
many square waves we see on oscilloscopes actually look more like C, D or F than
G. Many non-sinusoidal signals can be treated as the summation ofsinusoids.
This is how we are able to think of frequency components of non-sinusoidal
signals.
Figure 4-1
4-14
Il, tit I tit I.
We added various sinusoids and saw what Real And

can happen in both the time and frequency descriptions. What happens when we Complex
mUltiply various sinusoids? Before we answer this, let's walk through a thought Modulation
experiment. If we took a constant and multiplied it by a time varying signal, our
result will be another time varying signal. We can say that by multiplying the
constant by this time varying signal we affected the frequency of the original
constant. Obviously, if this is true, our result is no longer constant. Now, what
happens if the original signal wasn't constant, but a time varying signal as well?
It seems completely likely that we will still affect its frequency by multiplying
it with another time varying signal. This idea is known as modulation.
Our thought experiment is graphically
shown in Figure 4-44 with a teaser of things to come. The left columns are time
plots, the right columns are the frequency domain plots. We can think of constant's
original spectra (located at zero) as jumping up to a new location determined by
the multiplying signal. The real question is what happens in the second case when
there are two time-varying signals?
4-15
Xa(t)
t time
t frequency
t
time frequency
Xc = Xa(t) * Xb(t)
---Ih /\ /---
V~~
time
1
frequency
Figure 4-44
Was here: ; •. ~ 1 Is now here.
Figure 4-45
Xg(t)
---~~~~L;: <=> L - -_ _ _ _ _ _ _ --lt_
t========= ========.. <=>1 L-_ _ _ _..L.t- --
Xi=Xg 'Xh
r"=7c=:==-,'C:7-==----'C7c='\" <=>
Figure 4-46
4-16
Real Modulation
The mathematical basis for exactly how
the frequency is affected is given by the trigonometric identity found in at the
bottom of the page. However, even without using the identity, we can visualize
some of its qualities. We start with two sinusoids of frequency f and 3f Figures
4-47 through 4-52. These two are multiplied together. Notice the resulting wave
looks a bit different, it doesn't really match either of the original two waves.
Take a closer look. Its not too hard to
imagine that within the resulting wave we see something happening twice as often
as f and maybe something else happening at four times as often as f. Just for grins,
let's plot waves with frequencies of 2f and 4f. As a check, we added the waves of
f = 2 and f = 4 and verified their sum to be the same as when we multiplied waves
withf= 1 andf=3.
~1\1[\/\J[\I\A
4-17
Earlier we graphically demonstrated how Frequency

modulation can be interpreted in the time domain. In most cases, the goal of Shifting
modulation is to make use of its frequency shifting properties. To illustrate this, Properties
the corresponding frequency domain plots are shown in figure 4-53 - 4-59. To
keep things simple, only the cases where alpha - beta and alpha + beta results in
positive numbers as shown. We will discuss negative frequency cases later.
Frequency shifting is most relevant to
modulation. It is this modulation effect that allows audio signals (voice, music,
etc.) with frequencies between 20 to 20,000 Hertz to be shifted up to the hundreds
of thousands of Hertz or even megahertz where these signals may be transmitted
in the form of radio waves.
tz-Jl
I ~f I, t I >
Figure 4·54
Figure 4·53
tz+Jl
tz
v···t·····"'1 ~ II t ~
Figure 4·55 Figure 4·56
Signal with fl Signal with f2 - fl and h + fl

---+)@----+
~ Signal with f2
Figure 4·59
4-18
Imaginary The philosophy adopted for this book has

Numbers been to provide conceptual examples and heuristic arguments first to create an
understanding. Mathematics should follow where they cement these concepts
into a firm foundation. However, this is not so easy with the introduction of
imaginary or complex numbers. Historically, there has been some debate of the
philosophical aspects of the use of imaginary numbers. The reader is asked to
take a slight leap offaith concerning the use of imaginary numbers. Please believe
that their use can be valid and they are not uncomfortable to work with. More
importantly, they allow us to perform analysis in an easier manner.
The basis of imaginary numbers is the unit

i (or sometimes j - used by electrical engineers who wish to avoid confusion with
the symbol (i) which is used to represent current). "i" represents the square root
of -1.
If we define vC1 to be equal to a number

we will call "i," then "i" exists by definition. First we will view some of the basic
properties of "i." This begans to make more sense after it's been explained with
some of the examples on the next few pages.
If i = -Y-f , then f = -1 and ~I = -i
Some imaginary numbers: 5i -4i 2.2li sqrt(2)i
Q: What is 2i + 3i ?
A: 5i.
Q: What is 5i . i ?
A: 5' i2 = 5' (-1) = -5.
Let's look at the next step - complex

numbers. Complex numbers are numbers which may have real and imaginary
parts. Generally, we write complex numbers as the sum of these two parts.
Some complex numbers: 2 + i 4 + 3i -2 - 2i
Q: What is (2+i) + (4+3i) ?
4-19
A: Add the real parts together and add the imaginary parts together to get 6+4i.
Q: What is (2+i) . (4+3i) ?
A: 2"4 + 2'3i + i"4 + i'3i
=8 + 6i + 4i + 3i 2
=8 + 10i + (-3)
= 5 + 10i
The letter z is often used to denote a
complex variable. Further convention uses x to represent the real component and
y to represent the imaginary component (foreshadowing?):
z =X + iy
Real(z) =X
Imag(z) = y
The Z-Plane (Cartesian Form)

Okay, the foreshadowing relates to how
conveniently the components of complex numbers can be represented on a plane,
This Cartesian plane is known as the Z-plane (or the Complex plane), Here a
complex number's real component is the x-coordinate, and the imaginary
component is the y-coordinate. Therefore each complex number can be
represented as a point on this plane.
-----7
Figure 4-60
4-20
The Z-Plane (Polar Form)
Every point on the Z-plane can be ex-

pressed in Cartesian (XY) coordinates. Alternatively polar coordinates could be
used. Here, the point is defined by its distance from the origin (magnitude or "r")
and the angle from the positive real axis (angle or <1».
Figure 4-61
Cartesian And Polar Relationships
Points on the Z-plane can be represented

in either coordinate system. Both are used. The relationship between the can be
expressed by Figure 4-.0 The superscribed asterisk denotes complex conjugate.
Figure 4-62
4-21
Euler's Formula
We're going to need to look at one more

mathematical tidbit before we can complete this section. It's a version of euler's
formula, named after Leonhard Euler 18th century mathematician and physicist.
This tidbit is quite useful. But, more than that it is really elegant. So elegant that
we'll provide a summary of its derivation. Consider the exponential function and
its series expansion,
Also remember:
I·0 =1, ·1
I
.
= I, ·2
I =-1, I·3 =-I,. ·4
I =1, I·5 =I. ...
Letting "a" be replaced with "ia" we have:
Grouping real and imaginary components together:
The real grouping is the series expansion

for the cosine function and the imaginary grouping is the expansion for the sine
function! So:
eia = cosa + i sina (EulersFormula)
4-22
Now does it seem like it might be useful to

us? Sure. Previous page tells us we can describe the complex number z by
cartesian (XY) or polar (magnitude angle) systems:
Z= x+ jy
but,
X= ICOS(<1»
Y= fSin(<1»
so,
Z = r [cos(<1» + i sin(<1»]
\/ eH
Since,
ei ($) = cos<1> + isin <1>
Then,
Remember Euler now?
With Euler's formula we can see that for

any z we have a concise representation of z:
Not only do we have this concise repre-

sentation, but since we've shown thatthis is mathematically valid as well, we know
that we can perform mathematical operations on this representation and expect
them to be true.
4-23
Phasors And Unit Circle

Let's plot a group of z's on the complex
plane. For the plot we choose z's with a constant unit magnitude (r=1) and allow
for different (and increasing <\>'s.)
Figure 4-63
With <\> increasing with time, we have the

concept of a phasor. This phasor can be thought of as rotating around and around
at some frequency. The frequency of rotation depends on how quickly cp increases.
For our chosen z' s with r= 1, this phasor will trace out a set of points known as the
unit circle. The points 1, -1, i, -i all lie on the unit circle.
Figure 4-64
4-24
Some points to consider about phasor's.

They are described by quantities such as magnitude (r) and frequency (rate of
rotation). This sounds a bit familiar to the qualities of sinusoids, true? Yeah,
maybe.
Another thought, if a phasor has a fre-
quency of "f' turning counterclockwise, then a phasor with "-f' would tum the
other way wouldn't it? Yes.
Figure 4-65
Figure 4-66
What About Phasors?
Let's draw a picture using two counter-ro-

tating (one with +f, one with -f) phasors A and B.
Figure 4-66
4-25
Figure 4-67
Also, review the relationships
Now, let's add the phasors A and B.

Graphically we can see that this results in cancellation of the imaginary part (one
is positive, the other negative), and the real part is doubled. From above we also
know the real part (x) is r cos phi.
Figure 4-68
In a similar manner, B could be subtracted

from A. Now the real parts cancel, leaving two time the imaginary components.
But, the imaginary component (y) is equal to sin(<I».
Figure 4-69
4-26
Negative Frequencies In Sinusoids?
Rewriting and restating we have:
cos(2rrJt) = [ 2
l
ei2.1tf + e-i2.1tf
ei2.1tf - e-i2.1tfj
sin(2nft) = j ( 2
Figure 4-70
Remembering our sinusoid we see that this

could now be expressed as consisting of two phasors. One phasor has positive
frequency, one has negative frequency. Well, in our earlier plots for the frequency
domain we haven't shown that. Let's correct for that omission now and redraw
the frequency domain plot for a sinusoid with frequency of "f."
What about the phase for this negative

frequency? Let's take a look at the two phasors at t=O. (Which seems fair enough
since we look at the points around t=O to determine the phase of a sinusoid in a
time plot.)
ei2.1tf + e-i2.1tfj
cos(2nft) = ( 2
4-27
but t=O,
cos(O) = (
eiO +
2
e-iOj
So one phasor is at eiO (but e iO=l) and
the other is at e -iO (but e- iO=l), so plotting these phasors we have the following:
Figure 4-71
From our definition of angles we can agree

the angles are °
for both.
Thus, representing both the phase and
magnitude of the particular sinusoid, we have the given phase for the positive
frequency and the opposite phase for the negative frequency.
4-28
This page intentionally left blank.
4-29
Let's put the complex representation of the Complex

sinusoid to use in a quick review of the modulation theorem and verify that the Reprsentation
new representation works.
When a sine wave is multiplied with an-
other sine wave, we still get waveforms of the sum and differences of the two
original frequencies.
1
cos(x) cos(y) = 2 [coS(X- y) + cos(x+ y)]
This product can be written as:
[~~e--) ["'Y~e-"Y)
mUltiplying term by term,
rearranging,
1
2[COS(X + y) + cos(x - y)]
t t
Figure 4-72 Spectrum of cosx
4-30
r_ -1" _
Figure 4-73 Spectrum of cosx cosy
So why bother to use this complex repre-

sentation if you get the same thing? We have used complex number notations to
represent strictly real sinusoids. Remember that for the complex phaser of the
sinusoid, the imaginary parts exactly cancel and what is left is the real part. Since
only the representation changed, one would not expect anything different to
happen. However, if we want to, we can do something different.
With the use of complex representation, we
now have the freedom to multiply our real sinusoid by a complex waveform; one
with both real and imaginary components where the imaginary components don't
have to cancel. Notice that both of the resulting phasers had their frequencies
reduced by a factor of y. We no longer get the sum and the difference, only the
sum; for example, (x + -y) and (-x + -y). The result shows that we have one phaser
rotating in the positive direction at rate (x+y), and one phaser rotating in the
negative direction (x-y). What have we produced? The multiplication by the single
phaser gave us a means to perform a frequency shift in only one direction. We can
choose this shift to be up or down by the sign of the exponential.
Looking at the spectral plot in figure 4-74,
we see that it is not symmetric about f = O. This means that for each spike on the
left there is not a corresponding spike on the right. The lack of pairing means that
the phaser does not have a counter-rotating pair which cancels the imaginary part.
The resulting waveform is complex.
What happens if we take the real part of the
resulting waveform? Let's toss out the 112 for clarity.
Figure 4-74
4-31
Real [cos(x - y) + j sin(x - y) + above_ cos(x + Y) - j cos(x + y) ]}
We have the sum and difference again, just

as we would if we had done a strictly real modulation.
Integral Of Sine Wave

The last bit of information you need to
truly appreciate and comprehend the Fourier Transform is below. This transform
is the basis of spectral analysis. Bear in mind that the figure on this page is a
stepping stone to greater things and is not introducing a concept that stands
alone. By the way, for those who get queasy with the mention of integral calculus,
breath easy ... this material is straight forward.
We think of the integral of a waveform as,
on a XY plot, the integral of the curve Y as it is plotted along the X axis is the
area under the curve. An addendum is that the area above the x axis is + and the
area below is-.
Let's use a sinusoid for the curve. What's
the integral, the area under the curve, for one complete cycle of a sinusoid? In
{r=:J
j
Figure 4-75
!
Figure 4-76
4-32
figure 4-77 we can see that there is the same amount of positive area above the x
axis as there is negative area below the axis. These two are equal; however,
opposite sign areas cancel each other. The integral of the sinusoid's cycle is zero.
We can extend this reasoning for a sinusoid that continues to infinity. Still there
are equal positive and negative areas that cancel each other; thus, the integral of
the infinite sine wave is zero. Notice, that this result is true regardless of the
frequency, amplitude or phase of the sine wave. All infinite duration sinusoids
have an integral of zero.
Figure 4-77
All Sinusoids Have Zero Valued Integers?
One clarification needs to be made for a

special case. What if the sinusoid's frequency is zero? A zero frequency means
that it is unchanging or a constant. Electrical Engineers recognize this as
another way of saying DC. Figure 4-78 shows us that a constant integrated over
infinity is not zero, it is infinite. So a sinusoid with zero frequency, a constant or
DC if you prefer, is different.
Fourier We have all the arguments in place to

Transforms understand a most magical concept called the Fourier Transform. We do hope,
however, that it will not seem too mysterious after the explanation.
t
Figure 4-78
Area
4-33
The Fourier Transfonn is what enables us

to look at signals in either the time or frequency domain. It allows us to switch
back and forth or perfonn, as it is usually said, a transfonnation between the two
domains. Before actually defining the Fourier Transfonn and proceeding with an
example, let's introduce some of the language and symbols associated with the
transfonn.
If we describe a signal by its time domain
description, we use a small letter followed by (t); for example, x(t). Alternatively,
when a signal is described by its frequency domain representation, a capital letter
is used followed by (f); for example, X(f). The same letter is used in both the time
and frequency descriptions if both descriptions refer to the same signal. For
instance, x(t);: time domain description of signal x and X(f) =frequency domain
description of the same signal x.
The variables x(t) and X(f) are intimately
related since they both describe the same signal. The relationship between the two
is given a name, Fourier Transform pairs, or just transform pairs. In fact, the
relationship is also given a symbol. The symbol means that the variables on each
side are indeed transfonn pairs. x(t) <=> X(f) means that x(t) is a transfonn pair
with X(f). This is written as:
X(f) =F [x(t)] where F means Fourier Transfonn
Taking the Fourier Transfonn of some-

thing really means that you give me the time domain description of x, and I'll give
you the Frequency Domain description. More explicitly, we can say X(f) is
described by taking the Fourier Transfonn ofx(t). See figure 4-79 for an illustra-
tion.
F[x(t)]
8~0
Figure 4-79
4-34
Alternatively, it can be described as you

give me the frequency domain description and I'll give you the time domain
description. In other words, going the other direction. This is called the inverse
Fourier1'ransform. This is written as:
x(t) = F -\ [X(f) I where F -\ means Inverse Fourier Transform

JT! [X(f)]
~y ~~
~ ~
Figure 4-80
Why the Fourier Transform?
Good question. Here are a few reasons

why.
Reason 1: In signal processing, a common

problem involves the transmission of signals or their reception. This transmission
and reception takes place through a channel. For radio, the channel is the sky;
for a musical concert, the channel is the room in which the concert takes place;
for underwater sonar, it is the water. The point is, each one of these channels has
its own characteristics. Some may really corrupt the signal of interest. The channel
characteristics may have echoes (and other frequency-dependent effects). Time
delay, which we have seen, inherently effects frequencies. With the Fourier
Transform we have a better understanding of the frequency content of the signals
we are sending and receiving. It provides a quantitative measure of the spectral
corruption of the channel. With this measure, we are better equipped to use our
signal processing to correct for the corruption, analyze the effects of such
corruption, or assess performance.
4-35
Reason 2: In many applications we look

for the presence of a signal or the lack thereof. This sounds simple enough, but
in a noisy environment or channel it may be difficult. The Fourier Transform can
help. Assume that our signal of interest is a sinusoid with a frequency off With
time plot 4-81, noise has been added. However, looking at the Fourier Transform
of plot 4-82, it is apparent that the sinusoid is there. Voila! The Fourier Transform
in action.
Figure 4-81
Figure 4-82
4-36
The Mathematics Of The Fourier Transform
Now it's time to mathematically define

this transform. Don't get caught up with the form of the equation itself, the
concepts needed to understand it have already been covered. The Fourier
Transform is defined as:
X(f) = F[x(f)]
X(~ = f ctoo
x(f) e-j2rcft dt
It's not as bad as it looks. Let's dissect it a

little. First, ignore the integral and look at the inside. We have a signal x(t)
multiplied by a complex exponential.
x( f) e-j2rctt dt
Do you recognize this expression? Isn't

that complex modulation? Yes, it is. Remember that complex modulation is a
one-way shift of the spectra. In fact, the value of "I' chosen is what determines
how much the spectra shifted and in which direction.lnfact, after the modulation
is complete, the spectra that was originally atfnow resides at the frequency of
zero.
Shifted by Ie
Figure 4-83 Spectrum of x(t) Figure 4-84
Let's look at the integral in figure 4-85 and

4-86. We are taking the integral over infinite time which is the area under the curve
4-37
Let's look at the integral in figure 4-85 and

4-86. We are taking the integral over infinite time which is the area under the curve
of a signal composed of sinusoids. Didn't we see that the integral for sinusoids
over infinite time equals zero? Yes, unless the frequency of the sinusoid is zero.
First the complex modulation allows us to choose f which performs a shift of the
spectra. The chosenfis now at zero frequency. The integration makes everything
zero except for what was shifted down to zero frequency. Basically, that's all the
Fourier Transform is doing. We pick a frequency of interest and modulate the
signal by that frequency so that after integration we have a measure of how much
of the signal was at the chosen frequency. kes everything zero except for what has
been shifted down to zero frequency.
,,
,,
,,
,,
,,
,,
Figure 4-85
Figure 4-86
Basically, that's all the Fourier Transform Intuitive

is doing, we select a frequency of interest, modulate the signal by that frequency, Example Of
so that after we integrate (cancel all the non-zero frequency content), we have a The Fourier
measure of how much of the signal was at our chosen frequency. Transform
Some may wish to accuse us of circular
reasoning so let's look at a different example. For this example, we will not do
the actual calculus to perform the integration. Instead, keep in mind an assumption
and two rules previously learned to follow an intuitive approach.
Assumption: Any signal describable in
the time domain has a description in the frequency domain. We may not know
exactly what its frequency domain description is right now, but we can assume
that it does exist.
4-38
Rule #1) Multiplication by a complex si-

nusoid represents a one-way shift of a spectrum. The process is called complex
modulation.
Rule #2) Integration of non-zero fre-
quency sinusoids over all time equals O. By choosing to integrate something over
all time, we deliberately force ourselves to look at the DC component which is
the only component that may have a non-zero component. We effectively blind
ourselves to any time-varying quantities.
We can now follow a simple procedure to construct the spectra of any given real
waveform x(t).
Let's integrate the given x(t) over all time.
Using rule two, we know the zero frequency component and can fill in the f = 0
position on the spectral plot.
Figure 4-87
I I
>1 I I
We can choose f, say f = 3, and modulate,

using rule one, the given signal x(t) and call the result x'(t). By using rule two,
we can integrate x'(t) and determine the DC component of x'(t). Furthermore, we
4-39
know that by rule number one that the modulation shifted the spectra of x(t) by
f=3. So that the DC component ofx '(t) is the f=3 component of the original x(t).
Thus, we can now fill in the point on the spectral plot for f=3.
X(O)
X(O)
-ccc:--+-,-+, +~-+-,_t!---+~--1TL-/~'- __
,.
Figure 4-88
lDIII'I,,.
Figure 4-89
-,----------
Figure 4-90 Entire spectrum of x(t)
4-40
This procedure can be continued for other

values off(f = 1, f = 2, f = 2.002, f = 2.0024, f = 3.14159 ... , etc.) Once all the
values off are viewed, and the entire spectrum of x(t) is determined, we find X(f)
from x(t).
Very nice, but there are a couple of prob-

lems with this approach; for example, we have to integrate a signal over all time.
Who can wait that long? Furthermore, does this have to be done for each frequency
component? Keep in mind that this presentation has tried to heuristically show
what the mathematical operations that comprise the Fourier Transform are actu-
ally doing, and how these operations may give us the spectral composition of a
signal from its time domain description.
Determining the Fourier Transform of a
simple mathematical formula may be done analytically. Finding the Fourier
Transform of a real-life signal may be different. We will address these differences
later under the heading Fourier Transfonn ofReal-Life Signals. If you would like
more math, continue with the Fourier Transform Pairs section.
A More Mathematical Approach To The Fourier Transform
This page presents a slightly more mathe-

matical demonstration of the Fourier transform. If you've no interest please turn
the page, you will not need this for following material.
The term "slightly more mathematical"

was used because the development will utilize some straight forward elementary
calculus. We're not going to provide a development of the delta function.
Let's look at the Fourier Transform of a

sinusoid with phase <I> and non-zero frequency fc:
x(t) = cos(21tfct+ <1»
XCf) = J+OOcos(21tfct+ <1» e-j2nft dt
4-41
Notice that if "f' is not equal to +/- fe,

we've got the infinite time integral of sinusoids which we know are zero. What
about if f = +/- fc?
For f =+fc we have,
1 ·f+co 1+=
= 2" eJ<l> -co 1 dt+ 2" f_ cocos (21tft + <1» dt
1+00
f
+ 2" _=sin(2nft + <1» dt
Similarly for f =-fc we have,
4-42
Notice the sinusoidals terms are zero since

they are the infinite integral of sinusoids with non-zero frequency.
Now we're left with the just the first term
for f=+fc and the last term when f=-fc
1 +00 1
X(t) = - ej$f 1 dt= - e j$ ;1(+00) - t(-oo) = 00 for f= fc
2 -DO 2
1 . +00 1·
+ - e-1$f 1 dt= - efCP ;1(+00) - t(-oo) = 00 for f= -fc
2 -00 2
So we have infinity? What the does that

mean?
Well, first let's look at the whole thing.

True we've infinity when f is the frequency of the sinusoid (fc or -fc) but we also
figured that the spectrum if zero for all fnot equal to +/-fc. Putting these two ideas
together, what does this sound like? It's the Delta function we saw earlier. Using
this substition we can see:
F[cos(2nfct+ <l»] = e H 8(f- fc) + e-H 8(f- fc)
To illustrate, let <l> = O. Then we have the

famaliar transform of the cosine function below:
cos(2nfct) <=> 8(f - fc) + 8(f + fc)
1
-!c
1
Ie
Figure 4-3
4-43
What about using the sine function instead

of the cosine function?
sin(2nft) = ??
We could start from scratch again and fol-

low the same path as we did for cosine or we take a short cut and can use the
trigonometric identity sinx = cos(x - ~),
Figure 4-91
F[sin(2nfct)] = F[cos(2nfct- ~)]
r ,n
-}-
,n
}-
l
_I !!_~§(f- fc) + e 2 8(f+ fc) 1
-L 2 J
= [ -j 8(f- fc) 2+ j 8(f+ fc)]
4-44
Notice the Fourier transform of the sine

function is purely imaginary. Don't let this bother you, the imaginary parts and
real parts are just used to describe the phase. Notice that the only difference
between the cosine and sine function is phase. In fact, we just made use of this
fact!
Phase And Magnitudes Of The Fourier Transforms
jy
cos(2rif<.t) = .! ~21tf,t + .! e-j21tf,l

2 2
Figure 4-92
The preceding pages plots the Fourier

transforms for the sine function and the cosine function. The cosine was strictly
real and the sine was stricly imaginary. Typically, however, we tend to view things
in terms of amplitude and phase. Especially when we aren't lucky enough to have
just strictly real or just strictly imaginary. Viewing the cosine and sine functions
Fourier transforms with amplitude and phase we get...
jy
sin(21t" .t) = 1. ef11if,t _ 1. e-J11tf,!

~'2j 2j
Figure 4-93
4-45
We have come full circle at this point. This

is exactly the description for sinusoids that was at the very begining of the chapter
when time and frequency domain were first mentioned!
What about those phasors, do they still
have relevence? Yes, in fact they tie everything together quite nicely. Below are
the phasor diagrams for both the cosine and sine function.
In coming full circle we have now ce-
mented the relationship between the two domains with a mathematical relation-
ship, the Fourier transform. The main idea here is that the heuristic presentation
that worked in the first few pages continues to work after a more rigouress
development. Inspite of the integral calculus and complex mathematics of this
transform, it is indeed possible to make sense and utilize Fourier transforms.
4-46
mag
.11
t t
Figure 4-94 sin(2n:fct) <=> ~ j [(f-Jcl- (f+Jcll
Pair: Sine Function
sin(2n:fct) <=> ~ j [if +fc)l
4-47
Duality
A quick note on a concept that can save us
tim. This concept is called duality.
Two points to consider:
Point I: Did you notice that the Fourier

transform of a Delta function ( (t» in time resulted in a constant values in the
frequency domain? In other words, a spike at t=O has frequency components of
equal magnitudes for every frequency f.
Point 2: Did you notice the the Inverse
fourier transform of a spike in the frequency domain at f=O results in a constant
in time domain? In other words a DC component remains constant over all time.
·Notice how if one ignores the axis labels
they are the same? Notice how it doesn't matter if we are talking about time or
frequency. The fact remains that the "shape" on the left side of the plots transforms
to the same "shape" on the right side of the plots regardless of the name we give
to the domain? Well, this is an example of what is meant by duality.
We can be more specific andmore mathe-
matical:
+00
F[x(t)] = f x(t) e-j2rcft dt= X(fJ
(We already knew this, just a review)
F- 1[X(fJ] = f
+00 .
X('> e J2rcft df= x(t)
(We knew this too)
Now, not only is the above true, but now we also have:
x(-t) = f+OO X(fJ e-j2rcft df
and
4-48
The concept of a transform pair is valid

regardless which domain is actually being considered. The transform of a spike
is a constant, regardless of whether the spike is in the frequency or the time
domain. A rectangular pulse transforms into a sinc pulse, regardless of whether
the rectangular pulse refers to a time signal or a spectral content. Etc, etc, etc ...
This tends to make a lot of sense when one

realizes that the only difference between the Fourier transform and the Inverse
Fourier transform is the sign of the exponential (direction of (de)modulation).
The amplitude, phase and frequency are

useful measures of the sinusoid. However, not all signals are simple sinusoids,
so there are other metrics that are useful in describing these other signals.
4-49
Filters and Transfer Functions
C HAP T E R
5
W e briefly mentioned (signal)
filters earlier. The signal filter
in concept is no different than
Filters and
Transfer
Functions
a coffee filter or an oil filter. A coffee filter lets the coffee through but filters out
the grinds. A signal filter lets some components of the signal through but filters
out the other components of the signal (hopefully the noise.) But let's make it a
little more formal.
A signal filter (from now on called a filter)
is typically drawn as a box (Figure 5-1). By conventionx(t) represents the filter's
input signal, y(t) the output signal. The filter is described by a function that is
usually given the letter h(t). Filter impulse response is the name given to h(t) .
...Y(t)
x(t) h(t)
input Filter output
Figure 5-1 A signal Filter in Time Domain
More on this and how exactly the three

(x(t), y(t) and h(t» are related later. How are x(t),y(t) and h(t) related? Typically
the frequency domain makes more sense. Often it is the frequency components
of a signal that we desire to affect. (remember bass/treble controls on the stereo?).
Let's use the frequency domain, now we
have the input spectra labeled by X(f), the output spectra by Y(f). The filter is
now described by a function H(f). This is exactly the transform of h(t). H(f) is
known as the filter frequency response.
x(f)
input
H(f)
Filter output
Figure 5-2 A signal Filter in the Frequency Domain
5-1
Chapter 5 - Filters and Transfer Functions
The advantage of using the frequency do-

main is that the relationship between X(f), Y(f), and H(f) is simple. Use multipli-
cation. As shown below, Y(f) is equal to X(f) times H(f).
XCD
input
11(1)
Filter
YCD = HCD XCD
output
-
Figure 5-3
You'll see a good reason as to why H(f) is

called the transfer function if the equation is re-written as ...
Y(f)
H(f)=-
X(f)
It's understood that H(f) describes how the

input (X(f)) is related to the output (Y(f)). In other words "how the input is
'transferred' to the output by the function H(f)."
Remember, since H(f) is a Fourier trans-

form ofh(t), we know that H(f) could well be complex even ifh(t) is strictly real.
Thus, we'll be interested in the magnitude of H(f) as well as the phase of H(f).
Discussions often focus more on the magnitude. This is common, but don't get
lulled into thinking the phase doesn't matter, as we know, sometimes it does.
What Exactly Is h(t)?
Let's try a thought experiment, if someone

handed us an unknown filter (H(f) is unknown), how could we figure out what it
was?
5-2
A simple way would be to put in a wave-

form that consisted of equal amounts of all frequencies. By doing that, we could
look at the resulting output from the unknown filter and see how each of these
input frequencies were affected. We could identify the filter's frequency response,
since it's the ratio of the input and the output. In fact, for X(t)= 1, then H(t)=Y (t)!
Thus, we've directly measured the filter's frequency response H(t).
Let's use X(f) YCf) HCf) XCf) but XCf) t,
- - - - -.~1
l. H(Q ?
YCf)
..
HCf) • 1 = HCf)
t
Determination of H(f)
Figure 5-4
Simple enough, but how can we put in such

an input spectrum (X(t)=l for all t)? Looking at the Fourier transform pairs, we
find that such a spectrum exists if the input is a Delta function. So, using an
impulse as input, the time domain description of the above block diagram is ...
[~(f)-;---:'=-; iW·
___ l
------~-~IL___H_Cf_)_~_- _ _____ ~--------
:a
---::::.=--=:.. •. i-::..:::·
Determination of h(t)
h(t) is output from impulse
Figure 5-5
5-3
Two Views Of The Same Thing
When using frequency domain we used

X(f)=l as an input and got the frequency description of the filter as the output is
H(f). H(j) has been called the filter's "transfer function."
Alternatively, the time domain descrip-

tions could be viewed. Here we used an input Delta function to give us the time
domain description of the filter. Because of this, the time descriptor h(t) is known
as the filter's impulse response (the response due to an impulse being used as
input).
Either description, h(t) or H(f), may be

used to completely describe a filter. Essentially, they describe the same thing
since they are uniquely related to each other by the Fourier transform. Both views
have there merits and uses.
What About Using The Time Domain To Describe Filters?
The question remains, "If Y(j) = H(j) X(f),

what can we say about how the time domain descriptions x(t), h(t), and yet) are
related."
We can figure this out. Remember the

Inverse Fourier transform? Let's use that and what we learned previously. It's
not necessarily evident what's going on here for those with a dislike for calculus,
but it's not too terribly difficult if you are interested.
Mathematical Derivation Of The x(t), y(t), and h(t) Relationship
Let's start with what we have accepted so far:
x(t) X(f), y(t) Y(f), h(t) H(f) and,
Y(f) = H(f) X(f) let's take the 1FT (Inverse Fourier transform)
y(t) = F -1 [Y(f)] = F -1 [H(f) X(f)]
5-4
by definition, we get:
y(t) = r y(~ d2 r H(~ X(~ d2

1tft df= 1tftdt
but we know X(f)=F[x(t')] is used for dummy variable
so rewriting we have,
rewriting again,
y(t)=r r H(~ x(f) e-~1tf( d2 1tft dt' df
now let's relate t' and t by 't = t - t', substitute and rewrite
y(t) = r r H(~ x(t-'t) e-~1tft d't df
separate and reorder integral,
y(t) = r x(t-'t) [r H(~ d2 1tft df] d't
this is the IFf of H(f), so it equals h('t).
Then let's rewrite now and obtain,
y(t) = r x(t-'t) h('t) d't
5-5
Certainly we are getting close. What else

can we do? Unfortunately (or maybe fortunately), this is the end of the road, we're
stuck with this. However, this integral is so useful its given a name and a symbol
of its own. It's known as the convolution integral. The asterisk (*) is used to
denote convolution, please don't confuse this with multiplication, it's not the
same.
yCt) = x(t) * h(t) = r x(t-'t) h('t) de
Let's redraw the filter block diagram with

this new information to tie it together:
x(~) hCt) I yet) = xCt) * hCt)

----; YCf) =noRCf) XCf)
XCf) L-_R--'C'-.-'.f)_ _-------'
Figure 5-6
Convolution Integral, Intuitive Perspec-

tive
The Fourier transform was an integral

equation, yet it made sense on a basic level. The same can be said for the
convolution integral. This time we'll work backwards. Iinstead of dissecting the
integral equation to understand it, let's synthesize it from a string of arguments.
Argument #1) Input an impulse into a
filter and the output is the impulse response, h(t). We established that a filter is
completely defined by its impulse response.
Argument #2) Input signal as a collection
of impulses:
5-6
Imagine the waveform x(t) as shown -
Figure 5-8
Let's draw some impulses, say every sec-

ond, equal to the value of x(t) at that time (in other words, weighted impulses ).
Figure 5-7
Try drawing more of these impulses this

time, sayan every half second. This still doesn't look like much.
T I r r1
Figure 5-9
5-7
Now, try every tenth of a second. This is

starting to look pretty good. It's not too bad of a representation of the original
signal x(t).
Figure 5-10
What if we drew the impulses every infi-

nitesimal time interval? We would get exactly the original x(t).
Figure 5-11
We can indeed view the signal x(t) as a

collection of individual impulses.
Argument #3) Without explicitly stating it, we assumed our discussion is based
on linear time invariant systems. (Non-linear systems and their analysis can get
quite complicated and are beyond the scope of this book.) What linear means to
us is a linear system that has the property that "the sum of the input components
produce an output equal to the sum of the outputs if the individual components."
This is known as superposition. Pictorially it is simple:
5-8
xa(t)
·1 h(t)
Ya(t)
~
Figure 5-12
Xb(t)
·1 h(t)
Yb (t)
~
Figure 5-13
xa(t) + Xb(t) - h(t)

-
Ya(t) + Yb(t)
Figure 5-14
Time invariant means the systems im-

pulse response does not change with time. Impulses coming in at different times
would still have the same response (assuming equivalent states).
Combine #1 and #2 and #3 to form the
logical string. Input an impulse into the filter and the output is the impulse
response (argument #1).
Every impulse train that the input signal x

is constructed from, argument #2, will have a corresponding impulse response
generated by it, arguments #1 and #2.
The actual output of the filter due to signal

x(t) is the additive action of all these impulse responses (argument # 3).
5-9
Assume the original signal x(t) and imag-

ine an impulse response that looks like the example below.
-1 h(t) ~
x(t) ~
Let's recreate the "a signal is a collection

of impulse" plots with the filter's output y(t) represented by the addition of these
impulse responses.
Draw some impulse responses, say every
second, equal to the value ofx(t) at that time. As expected, there is not too much
to say about this.
Figure 5-17
Try drawing more of these impulse re-

sponses this time, say on every half second. Again, it still doesn't look like much.
Figure 5-18
5-10
Now try every tenth of a second. This is

starting to look pretty good.
Figure 5-19
What was meat when we drew the im-

pulses every infinitesimal time interval?Actuaily we would get exactly the true
output signal yet).
lZ5 Figure 5-20
Take a minute to review how these pre-

sented arguments and accompanying plots lead us to the convolution integral.
After that, we'll step up the above argument with a little more mathematics.
Each point of our input signal x(t) can be

represented by an weighted impulse (by weighted we mean an impulse whose
area under the curve (i.e. integral) is equal to x(t) at that time. Say, for any time
= tau, we get:
y(1:) = r 8 (t-1:) h(f) dt
One impulse at time t=1: ( (t-'t)) into the

filter will generate one impulse response starting at time t=1: (h(t-1:».
By superposition, the output of a collec-

tion of inputs equals the collection of the outputs of the individual inputs. This
5-11
This means the output signal yet) is the collection of the outputs of each of the
impulse inputs. (Since again we've got an infinite number, we'll get an integral.)
Voila, the convolution integral. In spite of
the looks of this integral equation, it really does make sense.
Convolution In The Time Domain

Multiplication In The Frequency Domain
In describing a filter we just naturally

assumed that a filter would have the characteristics that Y(t) =H(t) X(f). This is
because we tend to think as filters as acting on frequencies, by attenuating or
amplifying certain frequency components. This characteristic come about from
our definition and concept of a filter.
In order to have the frequency domain
characteristic of Y(f) = H(f)· X(t), we found that the time domain characteristics
were given by the convolution integral. This was done by use of the definition of
the relation between the time and frequency domains (the Fourier transform).
As seen earlier, convolution is so prevalent

that it' given a shorthand notation, the asterisk:
x(t) * h(t) = r h(t-'t)x(j)df
Putting this more succinctly we have:
If yet) = x(t) * h(t), then Y(f) = H(f)X(f)
"Convolution in the time domain equals

multiplication in the Frequency domain!"
Duality Revisited
Could duality come into play with regards

to the above quote? If duality means that we swap "f' and "t," does it also mean
that "convolution'" and "multiplication" could get swapped as well? Does
"multiplication in the time domain equal convolution int the frequency domain?"
5-12
Yes. The proof of this is left as an exercise

to the reader.
t t t t
\
t t t t
Figure 5-21
A familiar example of this would be the

multiplication of two sinusoids x(t) and y(t).
The reason that convolution came up was
because we were talking about filters. Let's get back to that.
Ideal Filter Types

It is impossible to draw a perfect circle,
nonetheless, the concept of a perfect circle is commonly used.
It is much the same with the ideal filters
presented. They are impossible to construct, but they are useful idealizations.
The basic ideal filter types are; low-pass, high-pass, bandpass, and bandstop.
Ideal Low-Pass Filter: A ideal lowpass
filter will allow all frequency components below the cutoff frequency to pass
unharmed, but prevent the passage of frequency components above the cutoff
frequency. The set of frequencies allowed to pass through is called the Passband.
The set of frequencies stopped is the Stopband.
5-13
Figure 5-22
Ideal High-Pass Filter: Merely the oppo-

site of a lowpass filter. Highs go through but the low frequencies get stopped.
Again the boundary between these two is determined by the cutoff frequency.
Figure 5-23
Ideal Bandpass Filter: Here only the

frequencies that lie between the upper bandedge and the lower bandedge get
through. Thus, there are two stopbands and one passband.
Figure 5-24
5-14
Ideal Bandstop Filter: As you probably

guessed, it is the opposite of the bandpass. Frequencies between the lower
bandedge and upper bandedge get stopped. Now there's two passbands and one
stopband.
Figure 5-25
Why Can't You Make An Ideal Filter? (Causality)

It's a good question. Also, its gives us one
of the first opportunities to use what we learned about Fourier transforms.
Let's look closer at the ideallowpass filter.
Figure 5-26
We know that we can determine the

impulse response h(t) from its frequency description H(f). We can do so by
calculating the inverse Fourier transform, but let's save ourselves the bother and
refer back at the Transform Pair section where we'd find the pair:
5-15
Thus, from transform pairs and duality we

now know that the ideal filter has an impulse response defined by the sine pulse
(shown below)
Figure 5-27
Great, so now we can go ahead and build

the ideal filter, right? No. Take a look at the sine pulse plot, notice the dots at the
left and right. They signify that the impulse response continues for ever and ever.
Well, this is one reason the ideal filter is impossible. Eternal time systems are a
bit difficult to build.
Remember, when we plot impulse re-
sponses, by convention, we say the input impulse "hits" the filter at time t=O.
Look again at the impulse response and notice how it has appreciable content to
the left of t=O. This corresponds to the time before the input impulse even get to
the filter! More specifically, the ideal filter would have to start reacting to
something that hasn't even occurred yet! (A difficult task at best). A term to
describe this task is "non-causal." To build a non-causal filter is beyond our
means.
We'd really like the filters to be causal,
which is to say h(t) = 0 for t = o.
The observant reader may say, "We could
just build a delay in the filter." This could be thought of as shifting the impulse
response plot to the right. If we could delay this enough, maybe we could move
the whole impulse response to the right of t=O, then it'd be causal and we'd be
able to build it. Great! But, don't forget the impulse response needed for the
ideal LPF is infinite in duration. Thus, we'd have to add a lot of delay, infinite
in fact. So much for that idea.
However, the idea of putting delay into the
impulse response is not an entirely bad idea. It does, indeed, get us an impulse
response closer to what we needed.
5-16
Almost Realistic Filter Types

Here we take a step towards a little more
realism in our description of filters. One thing that made the ideal filters so
impossible was the discontinuity, or jump in the filter's frequency response from
o in the stopband to 1 in the passband. Let's get over this hurdle by allowing
values in between 0 and 1. This added concept is the transition band, bordered
by the stopband edge and the passband edge.
Almost Realistic Low-Pass:
\
Figure 5-28
Almost Realistic High-Pass:
Figure 5-29
5-17
Realistic Filters
I Figure 5-30
To be able to build a filter we must allow

some additional concessions.
Figure 5-31
First, we allow for variation in the pass

band. The pass-band response is allowed to vary somewhat around the desired
response ofH(f)=l. It could be greater than 1 or a less (of course we'd generally
like it to be really close to 1). This is sometimes called passband ripple.
Secondly, we relax the constraint of the

stop-band. Now, it doesn't have to exactly equal zero (of course we'd like
generally like it to be really close to 0). The amount of signal allowed through
the stopband is called stopband attentuation.
Filter Now, we have all the basic parameters we

Specification need to specify a basic filter that we wish to build. These are:
These are the parameters generally re-

garded as filter design specifications. (Some may argue that I've left out such
5-18
things as phase response, overshoot, etc. Fine, but I will still call the above
parameters the basic filter design parameters.)
Type (Lowpass, Highpass, etc.)
Passband Edges
/{
Passband Ripple
Stopband Edges
Stopband Attenuatio"'n_ _+-______-+~
Figure 5-32
Implementing The Filter From These Parameters:

From a set of filter design parameters,
there are a variety of techniques for implementing an actual filter. Some of the
classical filter implementation types for continuous time filters are Chebychev,
Butterworth, Elliptic. Traditionally, continuous time implementations of these
filters used capacitors, resistor, inductors in passive filters or active filters with
the inclusion of op-amps.
Our goal is Digital Signal Processing. I
know we still haven't formally started the DSP discussion yet - hang in there, the
continuous time basics are important. Let's delve into filter implementation
issues as digital filter implementation issues and save this for when we get to the
strict DSP sections.
5-19
Return To Fourier Transforms
We defined the Fourier transform with an

equation. We investigated the motives for such an equation and it seemed to
make sense. Next, we took a survey of Fourier transform pairs of some simple
mathematical functions of interest.
Let's apply the concept to common sig-

nals. Let's use our knowledge to see what happens when we want use the Fourier
transform to determine the spectra of signals that we'll find in real-life. By
real-life, we simply mean something that we might measure over finite time.
Finite Measuring Time

We actually give a technical term and dis-
cussion about the fact that we'll use a finite measurement time on the signal we
wish to apply Fourier analysis. The Fourier integral has plus infinity and and
minus infinity as the limits of integration. When we limit our time "window" we
see that we aren't doing a strict Fourier analysis. We're kind of "fudging" a bit.
Thus, the results we get from our limited Fourier analysis will be somewhat
different. Let's demonstrate. Below, you'll find the sinusoid in its infinitely
long time representation along with its spikey Fourier representation.
5-20
X(j) = F(sin(2nft»
Figure 5-34
Figure 5-33
Let's consider what happens if we take

only a small portion of that sinusoid. In other words, we'll use a finite window
instead of minus infinite to plus infinite. Let's say we'll examine the "sinusoid"
from 8:21 AM to 8:23 AM. We might draw this Be (before coffee) time domain
waveform as shown below:
Xhc(t) = sin(2nft) [when t is from

8:21 a.m. to 8:23 a.m.]
Figure 5-35
The question remains, what's its frequency

domain description?
More On Windowing
Let's define xbc(t) a little more properly.
5-21
a.m.
8:21 a.m.
Figure 5-36
Now with this proper definition we could

calculate the Fourier transform of xbc(t) by direct application of the Fourier
Transform Equation and get an answer. Who wants to go through the steps of
calculus? There is no need to do that, there's a more illustrative solution. We'll
consider Xbc(t) to be made up by the product of two other "simple" signals, w(t)
and s(t).
Xbc(t)= w(t)"s(t)
where: w(t) is a rectangular pulse (we've
examined these before)
s(t) is an infinite sinusoid
The three signals xbc(t), w(t), and s(t) are
plotted below to verify that, indeed, the product of x(t) and s(t) is a valid
representation.
~I
Figure 5-37
5-22
wet),
\.10 1 i
Figure 5-38
Xbc(t)= w(t)·s(t)
Figure 5-39
By viewing Xbc(t) composed in this way,

we can take an shortcut around the calculus computations using the dual of
"convolution in the time domain equals multiplication in the frequency domain."
The dual of this states:
"Multiplication in the time domain is con-
volution in the frequency domain."
Knowing this, we can directly construct
the frequency domain representation of Xbc(t) is (Xbc(f)).
To convolve the spectra of set) with wet),
written as S(f)*W(f). We'll need to know S(f) and W(f), luckily we've already
seen them in the Fourier transform pairs sections. Here they are,
S(f)
I
5-23
Think of convolution as it was presented

in the filtering section paralleled below. Signal S is convoloved with W by
considering W to be the "impulse response" of a filter. (If the thought of sticking
frequency domain signals into filters bothers you, just forget about what domain
they are in and consider them by their shapes, that's why we've stripped off the
"f' in S(f). Remember we're only interested in the convolution process.
·1 '--------------'
Figure 5-42
Finally, we can see what the spectrum of

this windowed sinusoid look like.
Observation #1: Notice that Xbc(f) is not

a simple spike like the infinite time sinusoid was. It's been spread out to infinity.
5-24
Observation #2: Here we chose to view the

windowing effect as convolving the frequency responses of the window and the
input sinusoid to get the resulting plot. Since we chose the input as a sinusoid
we could have looked at it in a different way. We know multiplication of a signal
by a sinusoid is modulation. Verify that the spectrum of the window could also
be viewed as having been modulated up and down by the sinusoid. It's amazing
how you can view the same thing from different angles and it still makes sense.
Observation #3: There's an interesting
interplay between time and frequency stemming from the fact that time and
frequency are reciprocals of each other. Remember the duality issue between
convolution and multiplication? Well, here's a close relative - "A finite time signal
has infinite frequency content." We uSed S to stand for the the signal and W for
the window.
Windowing Interplay
Use the same setup as before but change
the length of the time window that we'll use to examine the sinusoid. Instead of
a couple of minutes before coffee, let's examine it all morning from 9am to 11:30
am. We'll call this Xam(t), what's its spectrum look like now?
Figure 5·45
(JO(\Of) (JO(\Of)
O~VO O~VO
5-25
With the a different window function w

we'11 have to re-work our determination of its frequency description W. We see
that the longer w(t) is "on" (has an amplitude of 1), the narrower its frequency
domain description is. Since we end up convolving our spikey (S) with a narrower
(W), see below. Notice now we get a representation closer to what the actual
Fourier determination gives us, this makes sense.
Figure 5-49
Figure 5-50
Figure 5-51
Now let's go the other way. Let's choose

to examine the signal over a one second intervals call this signal Xlsec(t). Again
we guess what the new window's frequency description will look like and do the
convolution. With such a short duration window, w(t) we get a very broad spectral
representation, W(f). Thus, we might get something like this:
Now the frequency determination is not as
good. We can still determine the sinusoids exact frequency by picking the biggest
the "peak" of the very broad bump.
But, if there was the slightest amount of
noise in our measurements we'd have something resembling the plot below. Try
to tell the exact frequency of the signal.
5-26
Figure 5-52
This bit of noise wouldn't have bothered

us as much in the signal xam(t).
Windows Wrap-up
Hopefully the preceding pages have driven
home the point that the chosen window can have a big impact on your spectral
analysis. It's a simple point, but one that is sometimes forgotten. For instance,
forgetful engineers have been known to try to do spectral analysis in the 20-50
Hz region using a window only a few millisecond long. The figure below may
give an intuitive feel to how accurate their results might have been.
Figure 5-53
Alternative Windows
We may have lead the reader erroneously
to believe that the only window function to use is the rectangular window that
we've been discussing. This is not true. It just happens to be the simplest to use
and easiest for illustrative purposes. In the rectangular window we presented,
you're either looking at the signal or you're not. There's a negative point about
5-27
the rectangular window and that is its frequency domain description. W(t) is
bumpy with many local rninimas and local maximas (sidelobes).
A better approach may be to avoid using
the rectangular window and its "I'm looking, I'm not looking" approach. A
different windowing function will allow for a more gradual transition between
looking and not looking.
See the example of the Hamming Window.
Instead of multiplying the signal set) by the rectangular window wet) of duration
of two minutes, we'll use the hamming window Whamm(t) of the same duration.
Start viewing at the signal a little bit at first, then more and more till we taper the
view back down to zero.
\J~ \J \J~ C7
Figure 5-54
Figure 5-55
f\hf\
Figures 5-56
Notice that the frequency description of

Whamm(t) (Whamm(t)) is a smoother function.
5-28
... - ... , ,
,,
Figure 5-57
Thus, when we convolve S(f) and

Whamm(f) to get the spectral estimate, we get a better plot. By better, we mean
lower sidelobe levels.
\ , ,,
'. \ J l ..
Figures 5-58
Spectrum of x using a Hamming Window
Windows Caveat
Compare the plots of the spectral estimates
for the two minute rectangular window and the two minute hamming window on
the next page.
5-29
Figures 5-59
Indeed, the Hamming Window signal's

spectra have lower sidelobes, but notice how it has broadened as well.
Conversely the rectangular windowed sig-
nal's spectra is narrower but the sidelobes are higher.
That's the window tradeoff: sidelobe sup-
pression vs. spectral widening. There are many possible window functions to
choose from. Each falls in its own niche within the window tradeoff. Some of
the more common window functions are, rectangular, hamming, hanning, gaus-
sian, blackman harris.
The Bandwidth Concept

Armed with the Fourier transform and
choosing our windows properly, we may determine the spectra of various signals.
Voice Signals
Let's look at an arbitrary voice spectra.
We'll find most of its content is located in the maybe 200Hz to 3200 Hz range.
This total span of frequencies where there are appreciable spectral components,
3200 Hz, is called the Signal Band Width. Thus, any system which is used to
carry voice signals should have the ability to carryall these frequencies. (In fact,
the frequencies listed here are the frequencies generally used for telephone
applications). If this is so, we say that this system has a System Band Width of
3200 Hz. In general, it's a good idea for the system bandwidth at least as large
as the signal bandwidth for good results.
Musical Signals
The range of human hearing is generally
20 to 20,000 Hz. So, a good musical reproduction system should have a
bandwidth of 20,000 Hz.
5-30
Why does an (mono) FM radio signal have

only a bandwidth of 15000 Hz? It would be better if it was 20,000 Hz. However,
to allow for this much bandwidth, the broadcast station would have to use up more
of the radio spectra (or radio band). The bandwidth of 15000 Hz was seen as a
good compromise between getting a "good" quality signal to broadcast and
conserving more of the band. So by it's bandlimiting approach (system band-
width smaller than the input signal bandwidth) we don't get a "perfect" musical
reproduction.
The Baseband Concept:

The term baseband is used to describe an
original signal before any modulation takes place. In the above example of
musical signals in a radio broadcast we can regard 0 to 15,000 Hz as the baseband
of the signal. After this music is modulated up to radio frequencies (in the
hundreds of megahertz range) it is no longer considered in the baseband. In order
for the listeners to enjoy the music, their radios must de-modulate the radio signal
to bring the music back to baseband.
Modulation does not effect the signal
bandwidth. This bandwidth remains constant throughout the modulation and
demodulation process and it is not effected by where it is modulated.
A general rule of thumb: For a signal at
baseband, its bandwidth can generally be regarded as the frequency for which the
signal has any appreciable content.
5-31
6
C HAP T E R
General DSP Concepts
I n dissecting the term digital signal

processing, we know that digital is a
given. Let's examine what is meant by
What Exactly
Is DSP?
signal processing.
Signal processing is a general term used to
describe the various functions are performed on signals usually to improve or
enhance the original signal. The specific improvement or enhancement differs
greatly from application to application and the type of signals involved. Perhaps
a quick sketch of various applications of signal processing, specific signals and
types of processing will help. Let's start with a few classic examples.
General Signal Processing
Communications Example: A communica-

tions engineer may be foremost concerned with the proper transfer and reception
of information from point A to point B over a channel or medium. Any signal will
is affected in some way by the channel. To extract the maximum information, it
is necessary to counteract the channel effects. That is, filtering may be used to
filter out the unwanted added noise or interference. Automatic gain control may
possibly be used to compensate for fading.
Control Example: In control engineering
(e.g., manufacturing quality/process control, chemical process control, vehicle
stabilization) the actual control performances done by taking measurements of
the process, then initiating a corrective action based upon the measurement.
Signal processing becomes involved when this measurement may be affected by
noise and require filtering for accurate, stable results. Further, signal generation
may be required to initiate the desired corrective action.
Storage or Retrieval Example: storage or
Retrieval is a fancy way of saying recording or signal storage. Here, the goal is
to archive the signal so that extracting information is maximized on storage and
on playback. Again, filtering is helpful.
The word "classic" is used to describe the
signal processing used in the examples because the problems with their theory
6-1
Chapter 6 - General DSP Concepts
and solutions have been known for a long time, that is since the beginning of
radio. Simple components such as tubes, transistors, inductors, capacitors,
resistors and vinyl or magnetic tape, are used to compose the elements of the
signal processing gear. The functional elements of signal processing are filters,
variable gain amplifiers, mixers, and oscillators.
nsp composes the most elemental and

most important aspects of signal processing today and it offers solutions to these
classic problems as welL However, with the advantages that DSP offers, signal
processing has moved beyond the "same old thing." Signal processing ideas
dabbled with in research labs and labeled "interesting but unrealistic" have found
new life in a variety of consumer as well as military and industrial products.
DSP's Newer Uses
We will look at those necessary and effec-

tive classic approaches to see how DSP is used to implement those functions in
later sections.
Communications Example: In addition to

the noise and interference introduced by the channel, an echo is often introduced
as well. Have you ever seen the "ghosting" of images on your TV set or heard an
echo on a long distance telephone call? DSP made possible the implementation
of "echo cancellers" of quality previously beyond the reach of the analog
electronics.
Storage/Retrieval Example: Think of the

compact audio disc. This is perhaps one of the most visible examples of a digital
storage and playback system. Earlier forms of storage and playback such as the
LP or cassette tapes demanded, accurate and expensive motors to minimize the
warbling sound resulting from motor speed variations. Each time a LP or cassette
is played back, mechanical wear degrades the quality of the achieving medium.
Additionally, the medium itself generates additional components of noise; surface
noise for LPs and tape hiss. Storing digitally eliminates the need for expensive
motors and no archival degradation takes place on playback. The medium itself
introduces no additional noise to the system. The market acceptance of the
compact disc demonstrates that these advantages can be achieved within market
cost constraints.
6-2
A major factor in the introduction of DSP

into a variety of areas has been the decreasing cost of digital micro-electronics
and digital memories. The presence of new DSP implementations will likely
continue as micro-electronic components continue to increase in performance
while decreasing in cost.
We may be accused of skipping over the
tempting "wow factor" and "wonder toys" associated with DSP; those machines
that recognize and respond to the spoken word, machines that speak for them-
selves; real time signature analysis; and machine vision. While such DSP func-
tioning does exist today it is to a large degree, so application-specific that it's
understanding would not be very portable from one to the next. This is a
publication concerning DSP and not the complexities of voice recognition algo-
rithms. Realize that DSP applications are based, to a great extent, on the classical
techniques already introduced such as with filtering.
Defining Digital Filtering

It should be made clear exactly what is
meant by the term digital filtering. The digital filtering is filtering which takes
place by the mathematical manipulation of the numbers which representing the
signal. In fact, for the balance of the presentation, we assume that the sampling
and reconstruction conditions are properly met by the data conversion sub-sys-
tems. Additionally, the appropriate choice of sample word-length is assumed based
upon the sampled signals SNR and the dynamic range. Focusing strictly on the
digital portion of the system, we will deal with the effects of numerical manipu-
lations on the signal samples.
Introduction To Digital Filtering

The moving average operator is the first
hint of a digital filter. The filter needs an input signal to operate on, an operation
to perform, and an output signal. We chose an arbitrary sample sine wave of O.lfs
as the input. See figure 6-1.
6-3
10 Samples per cycle
?rrrrllr. pllrrrIL
Figure 6-1 An Arbitrary Sampled Input Sigual
For the moving average operator, the op-

eration performed is an averaging of the samples. We sum up a few samples, then
divide the sum by the number of samples that we added together, and voila, an
average! Moving means that we find an average for one sample, then move to
another, find an average, and so on. The output is formed by looking at the
sequential values obtained by each step of the averaging process.
Deciding how many samples are averaged,

determines the order of the filter. Imagine this with the input and output signals
below.
x(n) = input signal
y(n) = output signal = moving average of x(n)
10 Samples per cycle

Avera.ge these 10 .. _
'"~~ 11111T[O' 11 [U !J ", [q TIF' """

the"e 10 (etc.)
Figure 6-2
The input and output signals differ. Is this

really a filter in terms of what we have previously discussed? If it is, why haven't
we discussed its frequency characteristics? Where is the convolution in the time
domain? To answer these questions, we step up our notation used to represent and
manipulate sampled signals.
6-4
Representation Of Sampled Signals

Sampled signals, or more accurately, dis-
crete signals, have acquired a representation of their own. Below is an overview.
x(n); This is a signal called x. With a value
it takes at integer value of n. If n is explicitly stated, (i.e., x(S» means the fifth
sample of signal x. If n, is left as an unspecified integer value (i.e., we have x(n»,
then we are generally speaking of the entire signal x.
x(n-k); This is a delayed version of the
signal x(n). It is delayed by k samples. To see this, view the sketch in figure 6-3
and let's use k = 4. Now, compare the signals x(n) and x(n-4). Notice that x(n-4)
eventually takes on the same values of x(n) just four samples later. For instance,
for n = 10, k = 4 we have if, x(lO) = 100, then x(lO) = x(14-4) = 100.
o = x(n)
_ = x(n-4)
I I
I I
.
I ..
I I
I ..
Figure 6-3 x(n) Delayed by Four Units
N yx(i); Summation of samples of signal

x. Here, the samples x(l), x(2), x(3), ... , N are summed together.
Moving Average Operator
We can now continue the investigation of

the moving average operation adding a more mathematical description. For the
input, we chose a sine wave sampled at O.1fs. This can be written as;
x(n) = sin( 21tnTs ), where Ts is the sampling period
= sin( 21tnf/fs ), but f=O.1fs so this is also
= sin( 21tnO.1 )
6-5
Defining the output waveform y(n) be the

moving average of ten samples of x(n.) Rewriting this mathematically gives
insight.
y(n) = 1/10 [x(n) + x(n-1) + x(n-2) + x(n-3) + ... + x(n-8) + x(n-9)]
It should be obvious that this is an averag-

ing process. x(n-l) is the sample that occurred just before x(n). Similarly, x(n-2)
is the sample value that occurred two sample periods before x(n). So the formula
shows that the output is simply formed by adding the present input sample, x(n),
to its nine preceding sample's value, then dividing by ten. Let's use the summation
version of this equation.
1 9
y(n) = 10 L x(n - k)
k=O
Let's take a not-so-obvious mathematical

tum. Instead of adding the samples x(n-k)s and then dividing by ten, let's bring
the 1110 inside the summation sign. So we multiply all of the samples by 1110
first. This doesn't change a thing, we still have the same equation.
1 1 1 1 1
yCn) = 10 x(n) + 10 x(rr-1) + 10 x(rr-2) + 10 x(rr-3) + ... + 10 x(rr-9)
In fact, let's give the 1110's a variable

name. How about calling it "h?" Better yet, let's have ten h's all with the value of
1110. Again, using the summation method of writing the exact same thing:
9 1
yCn) = L h(k) * x(rr-k) where h(k) = -10 for k=O, 1, 2, ... 9
k=O
Nothing has changed by rewriting. How-

ever, now we have the discrete version of something with which we are very
familiar. Remember the convolution integral from the continuous time and
filtering discussion? We will rewrite it with a generalized version of our new
formula.
yCt)=f h('t)*x(t-'t)d't
6-6
y(n) = I k=Q
h(k) * x(n-k)
(For a particular pattern we have h(k) = /0 for k=O, 1, ... 9 and zero every-
where else.)
Notice how strong the parallel is between

the convolution integral and the summation formula above. In both cases, we have
the input signal x and its past values, multiplied by a function h. We then take the
sum. Remember that the integral is nothing more than an infinite sum. Indeed,
the summation formula is precisely the discrete version of the convolution
integral. It is known as the convolution sum.
How far can we push the parallel between Convolution

the convolution integral, continuous time filters, conOlolution sum, and discrete SumAndThe
time filters? Let's play some games with this moving average. Moving
Average
For the summation formula representation

of the moving average, we chose h(k) = 1110 for k = 0, 1,2, ... 9. If this is truly
like the convolution formula we saw earlier, we would expect that h(k) represents
the impulse response of the filter. That's what is represented for the continuous
time filters. Let's check that out. The impulse input means:
x(n) = 1 for n=O

o else
Notice that this implies,
x(n-k) = 1 for n=k (since if n=k, then n-k=O)

o else
So our output yen) is,
y(n) = average of ten x(n)'s

= [1/10 x(n) + 1/10 x(n-1) + 1/10 x(n-2) + 1/10 x(n-3)
+ ... + 1/10 x(n-9) 1
So we can determine yen) for each n from our input x(n).
6-7
Ifn<O, then all then-I, n-2, n-3, .,. indices

are strictly less than zero. This means allx(n-k)'s are zero.
For n=O, only x(n) is not zero,

y(O) = [1/10 * 1 + 1/10 * 0 + 1/10 * 0 + 1/10 * 0 + ... + 1/10 * 0]
For n = 1, only x(n-1) is not zero,

y(1) = [ 1/10 * 0 + 1/10 * 1 + 1/10 * 0 + 1/10 * 0 + ... + 1/10 * 0]
For n = 2, only x(n-2) is not zero,

y(2) = [ 1/10 * 0 + 1/10 * 0 + 1/10 * 1 + 1/10 * 0 + ... + 1/10 * 0]
and so on through n =9.
Ifn =9, then all the n-I, n-2, n-3, ... , n-9,
are now strictly greater than zero, thus all other x(n-k)'s are again zero.
Basically, we input an impulse and got ten
~ rrrrrrrrrr~~::-.--o-"
-2 -1 a 1 2 3 4 5 6 7 8 9 10 11 12 • • •
Figure 6-4 Impulse Response of y(n)
values. The actual impulse response is ten values, each equal to 1110. These output
values are exactly described by the terms of h(i). Indeed, h(i) is the impulse
response. So what about frequency responses?
6-8
We saw from continuous time filters that Frequency

if we have the impulse response of a filter (h(t), then the frequency response Response Of
(H(f) of the filter is determined by the Fourier transform of h(t). We are not ready The Moving
to make bold statements about our moving average operator. It is appropriate to Average
start looking at the moving average operation from a frequency domain perspec-
Operator
tive.
Averaging is performed to get a better view
of longer term trends, in other words, to remove the day-to-day or moment-to-
moment fluctuations by filtering out the higher frequency noise.
The moving averaging process is a digital
low pass filter. We can use our Fourier transform process to see the frequency
response.
Since averaging is easy to compute, let's
take a look at how various frequency inputs are affected by this averaging. For
the following figures, the input amplitude remained constant with a value of one;
only the input frequencies were varied. All that has been done here is the
computation of the moving average (averaging ten consecutive samples) of the
input sinusoid for the various frequencies.
We can take a shortcut and remember that
the average of a constant value is merely the constant value. So it's no surprise to
see that we get the same amplitude on the output as we had for the input.
fre =0
0.5
-0.5
-1
0 50 100
Figure 6-5
6-9
fre = 0.04
0.5
-0.5
50 100
Figure 6-6
fre = 0.08
0.5 -
-0.5
-10~----------~5~0----------~100
Figure 6-7
fre = 0.1
0.5
o --------
-0.5
-I ------"------------'
o 50 100
Figure 6-8
fre = 0.12
0.5
_IL ___________ ~ __________ ~
o 50 100
Figure 6-9
6-10
fre =0.14
0.5
f ::: 0; 14 output C()IltinueS to

-0.5 increase in amplitude.
-I _ . _ . -..._ ....
o 50
Figure 6-10
0.5
o f 0.2 Zero output again.

-0.5
_IL-_ _ _ _ _ _ _ _ _ _ _ ~ ~
o 50 100
Figure 6-11
We could continue, but let's assume we

have actually run a set of sampled sine waves of various frequencies through our
moving average operator. We chose to let sampled sinusoids have frequencies
from zero to half the sampling frequency our limit given by the Nyquist criterion.
We can now present a plot of the resulting amplitudes for the various frequency
inputs. We now have a rough plot of the frequency response of the moving average
operator.
Out ut Am vs. In ut Fre
.g
~ 0.8
~ 0.6
0.4
0.2
0
0 0.5
flfs Frequency
Figure 6-12
Frequency Response of a
Moving Average Operator
6-11
Since we can discuss the frequency

response of this moving average operator, we will throw out the word operator
and call it a moving average filter.
The Moving Average Filter
Let's reflect for a moment on this moving

average filter. In fact, since we are interested in exactly how the output waveforms
are produced, let's zoom in on ten samples and overlay a plot of the h( k) values
on the plot of input samples x(n).
3 4 6 7 8 10
Figure 6-13
This type of diagram comes in handy for

understanding many types of filters and gives a great intuitive feel for exactly
how the filtering process works. Please become familiar with it! Notice that it
shows the actual convolution process. The figure below shows values ofx(n) for
an input frequency of O. We can use this graphic for other input frequencies to
verify the frequency dependence of the averaging process.
Since an interesting aspect of the moving

average filter's frequency response is the value of zero it takes on at O.1fs, let's
re-draw the diagram using an input frequency of 0.1 fs.
Figure 6-14
6-12
It is now possible to see how the averaging

process produces a zero output. There are waveform samples that are positive, and
five samples that are negative. In fact, the pairs of fives are identical except for
being opposite in sign. Thus, when all ten samples are added together, a value of
zero is produced. In a like fashion, the other frequencies that lead to zero amplitude
exhibit a similar kind of symmetry. The sum of the positive samples in the
averaging window is equal and opposite in sign to the sum ofthe negative samples
in the averaging window. Thus, the positive samples cancel the negative samples
and the sum is zero. For this perfect cancellation to take place, the input waveform
must conform to the moving average window.
Perhaps the easiest case to see why
multiples of O.lfs are zero, is the case of f=0.5fs. Here, we recognize that the
sample values alternate between + 1 and -1. Therefore, the moving average is
always adding five samples of + 1 with five values of -1. The sum equals O. Thus,
the moving average always equals zero.
Figure 6-15
Construction VS. Destruction

The averaging process is a low pass
process and part of the averaging requires the summation of neighboring samples.
In a low frequency input, neighboring samples have the same sign. Each of these
samples is multiplied by an hen) value. Since these are all the same sign, the
summation of these values is always constructive in nature.
>crII]JIr~ >< - LoVV' frequency

input da.ta.
Figure 6-16
6-13
Conversely, a high frequency signal has

samples of both positive and negative signs over the same size neighborhood.
MUltiplying by the constant h(n) values will have no effect on the sign of these
samples. Therefore, we are left with both positive valued and negative valued
samples. As such, their summation is destructive in nature.
x - High frequency input data
Few neighboring samples have

opposite sign.
Figure 6-17
It is e~actly the frequency dependent na-

ture of the constructive or destructive properties of the summation calculation
that is exploited to yield a filtering process.
Conjecture Relating Impulse Response To Frequency Response
In the previous section, the h(n)' s were all

of the same sign. This acted to produce a low pass filter. What happens if we were
to change the h(n) values to something that alternated signs? Can we conjecture
the following?
Low Frequency Signal Impulse Response ====> Constructive Addition

Same Signed Neighbors Of Alternating Sign
High Frequency Signal Impulse Response ====> Destructive Addition

Alternating Signed Of Alternating Sign
Neighbors
Conjecture Investigated
This conjecture certainly has an attractive,

symmetric quality to it. To investigate, use two standard waveforms, a DC signal
6-14
(constant) and one with a frequency of fs/2. These are chosen simply because
they are so easy to work with and still provide useful information.
What happens if the impulse response al-

ternates in sign? Let's take the previous impulse response and change the sign of
hen) when n is an odd number. Thus, we have h(n) when n is an odd number.
0.1 o - Filter impulse response
-0.1
Figure 6-18
Impulse Response that Alternates in Sign
We can easily see that a DC waveform

undergoes a destructive summation effect in this case, due to the alternating sign
of the hen). What about the case of fs12? Let's look again using the signal
superimposed on the impulse response diagram.
_ - Filter impulse response
x - Data samples
Figure 6-19
h(n) and x(n-k) lined up so each pair is either
positive or negative
Here we have the h(n)'s and x(n-k)'s "lin-

ing up" so that each pair is either both positive or both negative. Thus, the result
of the multiplication step is always positive, and the sum is constructive. The
output sample has a positive value. Its appreciable magnitude conveys a high pass
filter. That may be correct for figure 6-19, but look at what happens for the next
output sample.
6-15
For the next sample, we can view the input

as delayed, or moved to the right, within the filter window. The last sample from
the previous diagram has exited to the right, while the new sample has appeared
to the left.
-Hl1[11--
o - Filter impulse response
x - Data samples (1 sample later than

previous plot)
Figure 6-20
What occurs when hen) and x(n-k) are lined up with
opposite signs.
In cases like this, we always have positive

samples lined up with negative h(n)'s or negative samples lined up with positive
hen)' s. So now the result of each of the multiplications is negative. The sum of a
collection of negative numbers is negative. We still have constructive addition
and an appreciable magnitude on the output sample. It just has a negative sign,
that's all.
There is absolutely no problem that the

output signal alternates signs every other sample. This is exactly what the input
signal did.
We have seen how this alternating sign,

h(n), can indeed attenuate low frequency inputs while passing high frequencies.
Our experiment, based upon rudimentary examples, demonstrates the underlying
principles involved with digital filtering. However, few filters are simple moving
average or alternating sign filters.
The moving average filter discussion was

used to introduce the topic of digital filtering. Our everyday idea of averaging
was expanded to a moving average. This lead to the notion of a convolution
summation. This convolution sum was shown to possess a frequency-dependent
character which certainly appears suitable for a filtering operation. Moreover, it
has become apparent that different choices for the impulse response (h(n)'s) can
greatly effect the frequency characteristics of the convolution sum. To design a
6-16
digital filter of a desired frequency response, you must determine the correspond-
ing impulse response.
Determining The Right Impulse Response
The moving average filter had a given

impulse response and determined frequency response as shown in figure 6-12. If
we were to make a continuous time analogy to the h(n), by connecting the dots
to get h(t), we would have a continuous time version of the impulse response.
Since this is now a continuous time impulse response, the dots have become
connected. We can apply the Fourier transform to determine the frequency
response.
In our familiarity with the continuous time

Fourier transform, we realize that the filter's needed impulse response can be
found simply by taking the inverse Fourier transform of the desired frequency
response.
In light of the similarity between discrete

and continuous convolutions we can determine the digital filter's impulse re':
sponse. In other words, take a desired continuous frequency response, call it H(f)
and compute its impulse response, h(t) by the inverse Fourier transform. We
merely take a sample value of h(t). These samples are taken every sample period
corresponding to the rate the digital system is intended to use. We call these
samples h(n), and use h(n) to implement the digital filter. This is known as the
impulse invariant filter design.
Implementing Response
There are a variety of ways to determine a

proper filter given the desired frequency response. We will look at other methods
of impulse response determination later.
We presently need to address another im-

portant aspect of digital filtering; implementation. It is correct to do so at the
present time since the eventual implementation will invariably impact the filter
determination as well.
6-17
Finite Impulse Response Filters
The finite impulse response fIlter (FIR)

is the first filter type we will view. Actually, the FIR filter is nothing more than a
generalization of the moving average filter we discussed. The difference is that
we no longer require the h(n)'s to have the same value.
The limits of the summation go from zero

to N-I. This helps to denote the that the impulse response is finite. Since there are
N terms in this summation, N is known as the order of the filter. This convolution
sum becomes more penetrable in its block diagram form. The block diagram is a
direct implementation of the actual convolution sum above. For this reason, the
block diagram, shown in figure 6-21, is known as the direct form or canonical
form of the FIR.
Z -1 has
the same
meaning as
Delay and
is discussed
on page 6-20.
Figure 6-21
The Direct Form or Canonical Form of FIR Filter
We can see that each delayed sample is put

through a gain (multiplied by hen)). The output consists of the summation of N
of these delayed and gained samples. The direct form implementation is closely
related to analog tap delay lines used in communications engineering and has
borrowed much of its nomenclature. The string of digital delays along the top is
known as a delay line. The points available within this delay line are called taps.
The values the samples are multiplied by are the tap gains, tap coefficients, and
tap weights.
The FIR filter structure is a very standard

and common filter structure. The most notable being:
1) They are inherently stable (as opposed to the next filter type to be pre-
sented).
6-18
2) The dynamic range of the states of the filter is easily computed. This is
important to prevent clipping of the signal.
3) It can easily be constructed to have the desirable property of linear phase.
4) FIR tap gains for a given desired frequency response and can be easily
computed by Fourier transform techniques.
Let's look at another standard digital filter

implementation.
We do not need an infinite number of taps Infinite

to construct an infinite impulse response (I1R). The reason for this is that the IIR Impulse
does not use the same structure as the FIR. The concept of feedback is used for Response
the I1R. Here, the idea is to re-circulate some of the input signal. It should be Filters
obvious what is meant by looking at the I1R block diagram shown in figure 6-23.
To simplify things, we drew an elementary, IIR filter.
Z -1 has
the same
meaning as
Delay and
is discussed
on page 6-20.
Figure 6-23 IIR Direct Form
To realize the infinite nature of the filter's

impulse response, look at figure 6-23 using a feedback gain of 0.9 for the filter.
For the first output of the impulse response we have the same value as the impulse.
The next output value is the sum of the input plus 0.9 times the previous output.
However, all following input values are zero, so we keep multiplying the present
output by 0.9 to get the next output value. The frequency response of this IIR filter
can be computed in the same manner as was done for the moving average filter
in figure 6-2.
6-19
The Delay
In both the FIR and IIR filter descriptions,

we made casual use of an element labeled delay or sample delay. Whatever
sample value you enter eventually comes out. Often we talk of the unit sample
delay, which means the time between input and output is exactly one sample
period of the digital system.
This sample delay appeared in both of the

block diagrams and in the convolution summation. The only reason we bring the
delay up again is to introduce an alternate description. This new description gives
us a more abstract and powerful representation. It provides us with a direct link
between the digital impulse response and the digital frequency response. In
other words, we won't have to use continuous Fourier Transforms and sampling
as a bridge to determine the digital filter.
The z-1 Description Of Delay
Although the actual reasoning behind the

use of the z-l to represent a delay is not completely intractable, it is one of those
discussions that is beyond our scope. Let's agree to use it as an alternative and
identical description of the sample delay. Using this notation, instead of writing
the sample delay ofx( n) as x( n-l), we will now write the delayed version as z-l x( n).
Although we have glossed over the how's

and whys of this notation, we can still use it as an effective device (albeit a blind
rule) to gain insight into DSP. This shouldn't bother us too much since the idea
of following blind recipes is exactly what quantum physicists have been doing
for years. To acquaint ourselves with the use of this new notation, let's rewrite a
familiar equation, the convolution sum of the moving average filter, using this
z-l form.
The Frequency Response Of Discrete Time Signals And Systems
The previous pages have begun to discuss

digital filtering. As filters are, for the most part, thought of as effecting the
frequency components of signals, we will need to present some further ideas on
what frequency really means in terms of digital (or discrete) signals.
6-20
Up to this point, we have allowed to our-

selves to think of a discrete sequence having certain frequency components by
imagining that we are merely talking about the sampled version of some continu-
ous waveform with those frequency components. This has an intuitive appeal. A
true mathematical expression for the frequency description of the actual discrete
time signals was deliberately delayed until now. We now seek to rid ourselves of
this excess baggage by re-examining the Fourier transform. Consider the continu-
ous time signal x(t) and its frequency description X(f) given by the Fourier
transform of x(t):
xu) = AAt)] = r x(t) e-/2rcft dt
Now, consider the sampled version of x(t)

obtained by sampling x(t) every Ts seconds, call this xs(n). Here n=kTs where k
takes integer values. This is just the same sampling process we've used all along.
xs(n) = xs(kTs) = x(t) 8 (t-kTs)
We are still at liberty to take the Fourier

transform of this, so let's do just that:
xu) = FIxs(n)] = F[x(t) * (5 (t-kTs)] = r x(t) * 8 (t-kTs) e-/2rcft dt
Since 8(t-kTs) is zero everywhere except

for when t-kTs is an integer. For simplicity let's call this integer n. We need to
consider only the values in the integral. This reduces the integral merely to the
summation:
xu) = L x(kTs) e-/2rcfkTs = L oo

x(n) e-j2rcfn
~-oo ~
This expression gives us the frequency

components of the discrete, or digital, signal directly from the samples of the
signal x(n). We no longer need to imagine a link to the continuous signal; the
frequency components are strictly determined by the samples. This is the Fourier
transform of the sequence x(n) now given by an infinite sum.
6-21
Often in DSP, engineers talk not of the

Fourier transform, but of the "Z-transform." Indeed, the Z-transform is prevalent
and will be presented and defined next. Although the reasons behind the use of
the Z-transform and Z notation are not completely intractable, it is just one of
those discussions that is beyond the scope of this book.
6-22
C HAP T E R
Digital Filter Design

7
To construct this digital filter there are a How To Get
variety of well-defined options available once the desired frequency response is The Filter You
decided upon. To a large extent, the filter design methods used will depend on Desire
the filter parameters that have been specified for the desired response.
For instance, we've seen from our discus-
sion on continuous time filters that the general parameters needed to specify an
analog filter (and digital) consist of:
• Pass band Edge(s)
• Stopband Edge(s)
• Stopband Attenuation
A bit simplified. For the analog filter, we

also have to decide the order of the filter, that is how many inductors and
capacitors will be required. This order determination is required for our digital
filter as well. For the digital implementation, the filter order may be thought of
as the number of "delay" operators (or the highest power of z in the H(z)
polynomial--- it's the same thing.) So, we'll add that to our list:
• Filter Order
Notice that the first three quantities are

generally readily apparent to us based on our filtering needs. Whereas, the value
of our filter order is dependent on our first three constraints, as well as our chosen
filter design technique and filter implementation. Generally, from a cost consid-
eration, we seek to keep the filter order as small as possible while still doing the
job.
The problem is the value of our filter order
depends on the design method we use. It also depends on the type of filter we
choose to build, IIR or FIR.
In general, the first decision to make on the
filter implementation is to choose either an IIR or FIR. Each has its own set of
advantages and disadvantages and the two can be quite different. Faced with this,
7-1
Chapter 7 - Digital Filter Design
we are required to present some basic facts and generalities on the issue of llR
vs. FIR.
FIR vs. IIR "FACT" SHEET
FIR Filter IIR Filter
#1) - may require high + impressive performance

order for relatively achievable with modest order
modest filtering
requirements
#2) + always stable -(?) stability issues must be

considered
#3) + simple limits to - internal states may

state dynamic rang acquire very high dynamic range
#4) - sensitivity to coefficient quantization
#5) - susceptibility limit cycles
#6) +(?) linear phase -(?) non-linear phase response

response
Here are a few points.
#1) The first point is short and sweet.

When it comes to filtering performance, as measured in great stopband attenu-
ation, sharp transition regions, the IIR generally can out perform the FIR, for any
given filter order. The performance difference (or equivalently the order differ-
ence) can be dramatic. However, the advantage of IIR' s in this area may be offset
by the other factors. Obviously, the continued popularity of the FIR filter
demonstrates that filtering efficiency is not the only metric used to decide between
IIR or FIR.
7-2
#2) Unstable filters (or systems) are to be

avoided. Instability may arise only from a feedback mechanism. This feedback
is inherent in IIR designs. The stability liability of the IIR may be a bit misleading
at first glance. To be sure, we don't want an unstable filter. With proper design,
all filter poles can be made to fall inside the unit circle. Thus, proper stability will
be guaranteed. Conversely, FIR filters have no feedback, and are thus inherently
stable. The FIR transfer function has no poles, consisting strictly of zeros.
#3) The state of a filter is just another way
of saying a filter memory location (or register) used to hold the delayed samples.
These states have a dynamic range, just as we say a signal possesses a dynamic
range. We've already seen that the dynamic range of a signal is limited by the
number of bits used in this signal's representation. Let's examine the state's of
both the FIR and IlR filter, keeping the dynamic range in mind. We'll use a
canonical form of the FIR and IIR block diagrams to make our point here.
Note: Quotation marks around the word
'facts' is deliberate; this table is formed by generalities. The +'s and -'s above
are meant to give some indication of the desirability, or , of the listed filter
qualities.
FIR:
Figure 7-1 The Direct Form Realization of FIR Filter
This case is straight forward. Here, the

dynamic range of the states is exactly the same as the dynamic range of the signal.
(This is seen by realizing that the signal itself passes through the delay line.) The
same number of bits may be used in the filter state storage as is used in the signal
representation. In fact, the only point of concern here may be at the summer.
7-3
Since we could be adding a potentially

large number of delayed samples (obviously this number depends on the filter
order). The concern for possible overflow at summer stage can be mitigated in
many situations by appropriate scaling of the tap gain values.
IIR:
Figure 7-2 The IIR Direct Form
In contrast to the FIR filter, the IIR filter

does not have only the original input signal flowing through its filter states.
Because of the feedback mechanism of the IIR filter, many delayed and super-
imposed versions of the input signal circulate through the delay line. If this
superimposition is constructive in nature the amplitudes of the filter states
(dynamic range) may greatly exceed that of the input signal. This may allow for
the filter states' dynamic range to be much larger than the original input signal.
For such cases, the same number of bits may not be used in the filter state storage
as is used in the signal representation. Poor results, such as clipping, may result.
Care must be taken not only for the summer nodes but also the filter state.
#4) Sensitivity to Coefficient Quantiza-

tion. In the following sections we'll discuss how to determine the proper
coefficients for a desired frequency response. For now let's imagine we have
determined the proper coefficients by one of these methods. Frequently, these
coefficient values will be determined through the use of a computer program

(with rather high numerical precision - perhaps 32-bit floating point). Since our
filter will require real-time operation (fast) at reasonable cost (cheap), the filter
will probably be implemented on a machine (programmable DSPmicroprocessor,
7-4
or fixed hardware state machine) with lesser numerical accuracy. In this case, we
cannot implement the "true 32 bit floating point" desired coefficients, but are
forced to accept something reasonably close. The term used for, not surprisingly,
known as coefficient quantization. Typical quantization values for fixed point
coefficients are 12, 14, 16, or 24 bits. In making this transition from high accuracy
design to limited accuracy implementation we must take care. Some implemen-
tation forms are more sensitive to coefficient quantization than others. To be a
bit simplistic, we'll say here that FIR structure implementations are generally less
susceptible to this quantization than I1R filters. Furthermore, among IIR imple-
mentations there are some forms more sensitive than others.
#5) Limit Cycles: Limit Cycles are a phe-

nomena strictly limited to the IIR filters since they are not possible in the FIR
implementations. This again is due to the feedback nature of the I1R. As we've
seen, we fully expect a stable filter to eventually reach a zero output condition for
the zero input case. [Think of the example where the impulse response was (1
-0.9 (_0.9)2 (_0.9)3 ... ), here the output asymptomatically approaches zero.] Note
that this has been discussed in reference to infinite precision processes, where we
make no quantizations on either the coefficient (0.9) nor the state of the filter.
The point here is that re-examining this with respect to quantized coefficients and
quantized filter states we get an interesting result. This may be best shown by an
example.
0=5 Bit Quantized System

* = Unquantized System
-1 "-----'------'-----'
Figure 7 - 3 The Limit Cycle
7-5
In figure 7-3, we have resurrected a

first order (at z=-O.9) IIR filter. However, now the state and coefficient quanti-
zation have been included. For illustrative purposes, we have chosen a 5-bit
quantization (see other figure). The impulse response has been plotted for the
first 25 output values. Notice how the quantized IIR's output "gets stuck" in a
fixed value oscillation, while the unquantized version continues on its journey
towards zero. This occurs because in the quantized system the result of the
multiplication of the (quantized) state by the (quantized) alpha yields the same
magnitude (after quantization) as the previous state. This is a simple example of
a limit cycle. The point to note is that this limit cycle has produced an oscillation.
Its behavior has been greatly magnified for our example here by using a severely
limited quantization scheme of only five bits. However, even for more realistic
quantizations limit cycles may still persist. These limit cycles may become more
complex as the IIR structure increases in order.
#6) In the preceding chart, the "Linear

Phase vs. Non-linear Phase Response" of FIR versus IIR filters, a question mark
was placed next to the entry. It's presence was meant to make the reader reflect
on whetherlinear phase might be a good, bad, or even a "don't care" characteristic.
The answer to this depends on the application. However, it might be said that
linear phase response is given more credit than it deserves as a desirable filter
quality. To be sure, some applications require linear phase. However, if you're
not explicitly aware of phase properties of your signals, chances are you don't
need a linear phase filter. First, some background:
• We haven't paid too much attention to the concept of phase since it

Was first discussed in the continuous-time section. Reviewing the
basic points will remind us that is phase is always relative measure
between two sinusoids. In a filter, this relative measurement is
taken between the sinusoidal input and the sinusoidal output of a
filter. The "Phase Response" of the filter is result of the collection
of all such relative measurements for all possible input frequen-
cies. As such, a filter's phase response defines a relative phase (or
phase delay) as a function of frequency.
• In our continuous-time filter discussion, we concentrated on the

magnitude response of ideal and non-ideal filters. The reason for
this concentration (beyond the fact that magnitude response is a
slightly simpler concept that phase response) is that most often it
is exactly the magnitude response that is desired to be affected.
Below is an example.
7-6
Example: You want the bass guitar on your stereo to be

louder, so you turn up the bass knob (which increases
the magnitude response of low frequencies of your ste-
reo.) Nobody even asks about the phase response con-
siderations involved in turning this knob.
• What does linear phase mean? Linear phase means that the phase
delay between filter input and output is a linear function with re-
spect to frequency. For example, the phase delay at 200 Hz will be
some constant times the phase delay at 100 Hz. The actual mean-
ing of linear phase becomes easier to see with a small excursion,
looking at a few equations with our linear phase filter. Consider a
simple, sinusoidal input:
input: output:
cos(w ·t) -----> A· cos(w . t + <p)
Here, the corresponding output is the ex-

pected sinusoid of the same frequency with a different amplitude and phase. We're
interested only in the phase for this discussion. But, we actually know more than
this. From the fact that this is a linear phase filter we know that <p is a linear
function of frequency (w), writing this out:
<p = - a·w
now, rewriting the above equations:
cos("t) -----> Aco.s(3w·t-a·3·w)
-----> Acos( W· (t-a))
Remember that for any function f(t), the

function f(t-td) is the same function delayed by some time td. The result is really
no surprise here: We just get a delayed and amplitude scaled version of our original
7-7
sinusoid as output. What happens if the input was of a different frequency? Let's
see, using 3. 00 instead of 00:
cos(3oot) -----> Acos(3oot+<p)
but <p==a3oo (by linear phase filter),
cos(3oo· t) ----> Acos(3oot-a3oo)
-----> Acos(3oo(t-a))
We get the same time delay as the case

before. For two different frequencies (00 and 3(0), we get the same time delay (a).
In fact, for any input frequency, we'll get the same time delay between input and
output, as long as the linear phase constraint of <p = a 00 is met! It may make more
sense to call a linear phase filter a constant time delay filter since all spectral
components are delayed by the same amount of time.
Okay, so with the above points in mind, we

may consider the linear phase argument with respect to FIR and IIR filtering. It
turns out that the linear phase, or constant time delay property can be easily met
with an FIR filter implementation. This linear phase condition can be guaranteed
by constraining the filter's impulse to be symmetric about the center tap. For
instance, assuming an N tap filter where N is odd for our picture, we have:
• (N/2)-1 =(N/2)+ 1
(N/2)-2 = (N/2)+2 picture of symmetric hen)
(NI2)-3 = (NI2)+3 etc.
The FIR can easily be made to have linear

phase response. The same does not hold for IIR filters. One quick way to see this
is to consider the symmetry argument of impulse response that guarantees linear
phase. For the IIR we'd seek to answer the question, "How does one build a
symmetric version of something with an infinite tail?" (Don't forget the first "I"
in IIR stands for "infinite") Obviously, we'll require an infinite duration 'head'
7-8
before this infinite tail. But, this would give us problems with respect to causality.
So, the goal of finding a causal, linear phase IIR filter seems elusive.
Keep in mind that continuous-time (ana-

log) filters have never been linear phase. And yet, in spite of their lack of linear
phase, they have been useful tools for a variety of functions over the years. So,
be careful not to relegate non-linear phase filters to the junk pile of obsolescence,
it just isn't so.
Should You Choose An FIR Or An IIR Filter?
Wouldn't it be great if we could simply

"always use an FIR filter? The truth is that there is no one, global answer. Each
filter design must be considered within the framework of its implementation. The
list presented some of the general tradeoffs involved in this selection.
Typically, the engineer may perform a
rough filter design using estimates of the desired filter response for both an FIR
implementation and an IIR implementation. The results ofthis initial design gives
an estimate of the necessary filter orders. (And correspondingly, the number of
numerical operations per second required for such a filter.) Sometimes the limited
horsepower of DSP machinery used in the filter's implementation will dictate the
use of the IIR filter at this point. However, in general, the selection may have to
be based on other issues mentioned. This design process typically is an iterative
process where the engineer unravels the desired filters implementation through a
series of small investigations. Luckily, as we'll soon see, much of the drudgery
of this series of small investigations is mitigated by the use of computer programs.
A great variety ofDSP filter design computer packages makes this job easier. This
by no means that an exhaustive computer search should be used to find the "best"
filter implementation!
"FIR vs. IIR", some readers may believe

there is a prejudice against FIR filters. This is definitely not so, although, we have
deliberately played down the linear phase advantage of the FIR over the IIR. This
is because the "linear phase advantage" is truly an advantage only in those
applications where linear phase is important. For a large number of applications
this may not be the case. Thus, if one were to be limited to the choice of an FIR
because of its linear phase properties (when it is of no consequence for the
application at hand) the engineer has placed an unnecessary constraint on the
design problem (and thus not obtained the optimal solution). The result of such
7-9
a false constraint could have a tremendous impact on the system solution, as IIR
and FIR filters of similar performance can be of considerably different order.
7-10
GLOSSARY
AC (Alternating Current): Refers to the cyclical nature of electric current when it

has a frequency associated with it.
ACCOUNTANT-ESE: Terms which may be used by an accountant.
ADC: Analog to digital converter.
ALL-PASS FILTER: A filter that allows all frequencies to pass through the circuit.
AMPLITUDE: Amplitude refers to the intensity of an electrical signal in terms of its

voltage (or potential), current or power level at any given point in time or
frequency.
AMPLITUDE COMPRESSION: An action to shrink a signal' s amplitude. Amplitude

compression is performed on music signals before they are sent across the
airwaves, for example.
AMPLITUDE 1 (OR, AMPLITUDE OF ONE): When this phrase is used, the author
is referring to the maximum intensity of unit 1 that a signal reaches. It may
refer, for example to voltage, current, or, power signals. In the case of a signal
which repeats itself periodically, such as a sine wave, this phrase refers to the
intensity that a signal may reach on its upward or downward swing.
ANALOG: This concept refers to an unbroken stream of any natural or man-made

phenomena, such as a river, or the electrical current flowing through a high
power line. In signal processing, the continuous action always refers to the
continuous action of an electrical signal or signals in which we have taken
interest. The idea of brokenness, or discontinuity, as it is associated to signal
processing, is discussed under both "Discrete" and "Digital."
ANALOG DOMAIN: This phrase refers to the category under which all continuous
actions fall; or to everything in the universe that can be said to be continuous
and analog. For example, we can say that the motion of the sun is part of the
analog universe. However, in the unlikely event that something breaks the
motion of the sun from its normal path, we would cease to think of its motion
as continuous and analog in that moment (see "Discrete" and "Digital").
G-1
Glossary
ANALOG SIGNALS: An analog signal refers to something in the analog domain

that contains information (see "Signal"). For example, the position of
the sun is an analog signal if we view it as something which contains
information about the time of day.
ANALOG-TO-DIGITAL-CONVERTER: This device does just as it says; it

converts analog waveforms into digital numbers suitable for storage in
computer memory.
ARITHMETIC LOGIC UNIT (ALU): The ALU is the portion of the Central
Processing Unit (CPU) which interprets and performs arithmetic and
logic commands which we issue to the CPU.
ASCII: This acronym stands for American Standard Code for Information
Interchange. This is the standard which defines certain digital bit patterns
to stand for common symbols such as our alphabet and our number
system. The bit pattern, 0011000, for example, stands for the number
zero.
ATTENUATION: This is what happens when a signal tapers off. We are

attenuating the music on our radio when we tum down the volume.
BANDPASS FILTER: This is a filter which attenuates all frequencies higher

than the upper bandedge, and lower than the lower bandedge; while
allowing all those frequencies in between to "pass." It is the opposite of
a bandstop filter. The ideal filter has an absolute transition to zero, rather
than the non-ideal attenuation.
BANDSTOP FILTER: This filter passes frequencies higher than the upper
bandedge, and lower than the lower bandedge, while attenuating all those
in between the bandedges; this is commonly referred to as a "notch
filter." This filter is the opposite of a bandpass Filter. The ideal filter has
an absolute transition to zero, rather than the non-ideal attenuation:
BANDWIDTH: A bandwidth is a range of frequencies. For example, the lowest

audible frequency is 20 Hertz, while the highest audible frequency is
20000 Hertz; together, these frequencies define the bandwidth of audible
frequencies.
G-2
Glossary
BASEBAND: This is an often-used term to refer to the original form of a signal before
it was modified by a DSP filtering technique.
BLACKMAN HARRIS WINDOW: (See "Windows") -C- CERAMIC RESONA-
TOR: This is a frequency-producing device whose ceramic material is so
strong and responsive that it can be made to resonate, or vibrate, at a very
high frequency when an electrical signal is applied (see "Natural Fre-
quency").
BREADBOARDED: To build up the circuit with the intent of trying it to see if it
works.
CARRIER FREQUENCY: This can best be understood in terms of the example
given in the textbook. We know that the audible range of musical signals is
from 20 Hertz to 20000 Hertz. However, it turns out that when FM radio
stations want to broadcast these signals, they do so at a considerably higher
frequency; somewhere between 88000 Hertz and 108000 Hertz. The musical
signals are being 'carried' by another signal; hence the term "Carrier Fre-
quency."
CAUSAL: This concept refers to the idea that any signal which we create, and which
we begin to see at the output of our system, was created from previous inputs.
That is, the causal system never anticipates what the input might be; but it
depends on a new value at the input to create another output value.
CENTRAL PROCESSING UNIT (CPU): This is the "brain of the computer. The
CPU analyzes and interprets the commands that we give it, and then it
executes those commands in an orderly fashion.
COEFFICIENT: This is also known as a Coefficient Multiplier. This is a number by
which each point in a signal is multiplied in order to alter it in predetermined
way. More than one coefficient is generally used to accomplish the signal
processing objective.
CONVERSION TIME: Length of time to convert a signal from analog to digital.
CONVOLUTION: This is the process whereby 2 or more signals are mixed together
to form a new signal.
G-3
Glossary
CONVOL UTION INTEGRAL: This is the mathematical formula, or algorithm,

which describes how 2 or more signals are mixed to synthesize a new
one. In a more detailed sense, this integral describes how signals,
themselves composed of 1 or more sinusoids, may be systematically
meshed together for the purpose of creating a new set offinely integrated
sinusoids, which together compose one new signal. This is also known
as the Superposition Integral.
COSINE FUNCTION: Below we see a graphically description of a cosine

function: Figure 1. A cosine function
D.C. (OR, DIRECT CURRENT): Whenever the term "dc," or "level," is used,
it refers simply to the constant voltage level of a signal which exists
without a frequency component.
DAC: Digital to analog converter.
DELAY: Delay is one of the main strengths of DSP. In an analog system, it is

difficult to control how long one event must precede another, faithfully,
time after time. A digital system, however, incorporates memory devices
which can store a value for just about as long as we want to before using
it again. Computers, and their ability to store and synchronize events,
are the reason for Digital Signal Processing; and they give us the
flexibility to use delay effectively.
G-4
Glossary
DELTA FUNCTION: The delta function is a tool used in many areas of Electrical
Engineering to describe a single, infinitely small occurrence of a signal
whether it be in the Frequency Domain, or in the Time Domain. It can be
viewed as the smallest, simplest element of any signal, out of which all signals
are built. The delta function is to the atom, as signals are to matter.
DIGITAL: Digital simply means having a value of "1" or '0", as opposed to analog,
which could have any value in between. It often implies computers or
computing devices. The word digital is best thought of in connection with the
words analog and discrete. A signal being received by a computer is at first
analog in nature since it is generally a continuous wavefonn. The analog
signal then passes through a device which converts the analog signal into a
set of discrete quantifiable sub-signals. These sub-signals are then put into
memory within the computer where they can officially be termed "digital."
DIGITAL SIGNAL PROCESSING, OR, DSP: This is what this book is all about
and aims to explain. Briefly, DSP refers to the way in which computers are
used to alter signals which have already passed out of the analog domain,
into the discrete domain, and finally into the digital domain. Once in the
digital domain, the DSP engineer uses the processing power of the computer
to alter the properties of the signal before it is passed back out of the computer
into the analog world again.
DIGITAL-TO-ANALOG-CONVERTER: A device which converts digital numbers

into analog output.
DISCRETE: Discrete implies a broken stream of of a nonnally-thought -to-be

continuous action or thing. For example, old timepieces had second hands
that swept across the face of the watch in one continuous, analog motion. The
newer watches, we noticed, had second hands that skipped across the face of
the watch at regular intervals to count out the seconds. This action would be
considered discrete (broken ).
DOMAIN: In Digital Signal Processing, the engineer or enthusiast frequently speaks

in tenns of domains - for example, the frequency domain, or the digital
domain. The word domain came from a branch of mathematics called Set
Theory; but generally it is used define the boundaries of the set to which we
are limiting ourselves.
DRIFT: Change in perfonnance.
G-5
Glossary
DUALITY: Without meaning to sound cosmic about it, this is a concept which
refers to the "twoness" in the universe. But more important to DSP, it
refers to the reciprocal relationship between the time and the frequency
domains which makes possible the the very powerful tool of the Fourier
Transform Pairs.
DUMPING: To get rid of; stop using.
DYNAMIC RANGE: This refers to the frequency range of a signal. Since most
signals are composed of many sub- signals, it is helpful to speak of its
dynamic range, or the range of frequencies of which it is composed.
EEPROM: Electrically, erasable programmable memory.
EPROM: Erasable programmable memory.
EXECUTION UNIT: This part of the Central Processing Unit (CPU) coordi-
nates and synchronizes all actions which the Arithmetic Logic Unit
(ALU) performs.
FEEDBACK: In signal processing. quite often the output of a system is re-used

in order to help make it more stable, or simply to produce a desired effect;
this is known as feedback. FILTER (OR, SIGNAL FILTER): In the
broadest sense, a filter is anything that modifies whatever it is that we
are trying to modify. For example, a singer modifies, or filters, sounds
in order to make the tones pleasing to the ear. We can see that filtering
is actually a point of view. However, in the DSP sense, filters are ways
that we can change an incoming signal, such as the signal being received
from a telephone line, or the music we hear from our CD player.
FLASH CONVERTER: This is a high speed Analog-to-Digital Converter (See

"Successive Approximation Converter").
FOURIER TRANSFORM: The Fourier Transform is a process based on the

mathematics of calculus which is used to move back and forth between
the frequency and the time domains in examining the signal.
G-6
Glossary
FOURIER TRANSFORM PAIRS: Table XX given on page YY lays them out for
perusal. Such frequent use of these particular bridges, so to speak, between
the frequency and the time domains (and back again)is made, that it finally
became prudent simply to put one of these tables in just about every DSP
reference manual or textbook in existence.
FREQUENCY: Intuitively, we know that this word means how often something
happens. In the DSP sense, we mean how quickly a signal finishes one of its
cycles (see "Sine Function").
FREQUENCY DOMAIN: This domain gives information about how often a signal
completes one of its cycles, and how many major sub-signals there are
contained in the signal. A signal can be composed of many individual signals
where each one of those signals has its own cycle, or, frequency, associated
with it ( see "Sine Function," or, "Cosine Function" ). So, one can imagine
that, upon viewing a complicated signal in the frequency domain, one would
see a full range of main lobes, indicating that the signal has many concen-
trations of single frequencies within it.
FREQUENCY DOMAIN DESCRIPTION: (see "Frequency Domain").
FREQUENCY RESPONSE: Completely defines the way a system responds to

frequency input.
FUNDAMENTAL FREQUENCY: This refers to the dominant frequency among a

group of related frequencies (See "Harmonics").
GAUSSIAN WINDOW: (See "Windows").
GmB'S PHENOMENON: These are oscillations in the time or the frequency domain
which occur at any sharp transition point in a signal. The famous mathema-
tician, Josiah Gibbs, proved that when we are attempting to create a square
wave (which is a wholly owned subsidiary of the human mind -they do not
occur in nature) or, whenever we are trying to create a signal in the frequency
or the time domain which exhibits sharp transitions, small oscillations start
to occur. These oscillations are also known as ringing (See "Ringing").
GROUP DELAY: In the DSP sense, this takes each increment of phase change, and
ratios it with the amount of frequency change during the same increment.
When the signal is demonstrating Gibb's Phenomenon in the frequency
domain, for example, the ratio is not constant; but for as long as the amplitude
was flat, the ratio was a constant.
G-7
Glossary
H- H-BRIDGE MOTOR DRIVER: This is a microcontroller peripheral which

is a configuration of transistors used to control motors. Yes, the circuit
does look like an "H."
HAMMING WINDOW: (See "Windows").
HANG: Connect; attach to.
HANNING WINDOW: (See "Windows").
HARMONICS: When we pluck a guitar string we produce a sound which is

composed of many different sub- signals each of which has its own
individual frequency. Each one of these frequencies is associated with
an integer multiple of the main, or fundamental, frequency that the
plucked string made. Harmonics, then, are the related sub-frequencies
that a signal may produce. In a frequency domain plot, we would see the
fundamental frequency as a main lobe, and each of its harmonics would
be the side lobes, which taper away from it on either side.
HERTZ: The unit used to measure electrical signal frequency.
HEURISTIC: Refers to the exploratory method of engineering sciences. But it

is really just a fancy name for "rules of thumb." Heuristic notions are
also common sense notions, or notions about how to do things based on
the experiences of life; for example, we learn not to touch a live wire by
either being zapped, or being warned before it can happen. This is known
as negative feedback, and it is the most important part of the heuristic
method.
HIGHPASS FILTER: This is a software or hardware device which will block

the low frequency signals of our choice, and allow us to pass those high
frequency signals which interest us. The ideal filter has an absolute, sharp
transition to zero, rather than the non-ideal attenuation.
IDEAL FILTER: The ideal filter blocks every frequency component of an

incoming signal above, and/or, below, the specified bandedges of the
filter. If it were a bandpass filter, it would look like a this:
G-8
Glossary
Needless to say, the ideal does not occur naturally; but it is a convenient form to tum
to since it is easier to draw when we are trying to explain something.
IDEAL BANDPASS FILTER: (See "Ideal Filter," and "Bandpass Filter").
IDEAL BANDSTOP FILTER: (See "Ideal Filter," and "Bandstop Filter").
IDEAL LOWPASS FILTER: (See "Ideal Filter," and "Low Pass Filter").
IIR FILTER: Infinite Impulse Response Filter
IMPULSE: This is a single discrete sample of a signal (See "Delta Function").
IN-CIRCUIT-EMULATOR: This device is generally built into a computer for the

purpose of allowing the systems designer to use external equipment to
examine, alter, and control the MCU system.
INFINITE IMPULSE RESPONSE FILTER: This kind filter is produced using

feedback. Its major drawback is that it has stability problems.
INTEGRATING CONVERTER: A Digital-to-Analog Converter whose specialty is

making very accurate representations ofthe analog signalfor use in the digital
domain. The drawback is that this kind of converter is slow.
INVERSE FOURIER TRANSFORM: The inverse of the Fourier Transform simply

refers to the process whereby a signal which we are viewing in the frequency
domain, is transformed back to the time domain. (See "Fourier Transform").
LINEARITY: In the DSP sense, and in almost any other sense, this idea refers to the
constant ratio associated with incremental changes in cause and effect. For
example, if time changes and voltage changes with it, and the ratio between
them remains a constant, then we have a linear relationship.
LINEAR-LOG WEIGHTING: (See "Weighting"). This is a conversion used to

emphasize aspects of a signal that come out best when viewed on a logarith-
mic scale as opposed to a linear scale.
LOCAL MINIMAS: When viewing a graph of a waveform in the frequency, or in

the time domain, we may choose to study one partiCUlar area of the waveform,
ignoring all the rest. Local to that one particular area, the waveform may have
a lowest, and a highest point: the local minima is, clearly enough, the lowest
point (See "Local Maximas").
G-9
Glossary
LOCAL MAXIMAS: When viewing a graph of a waveform in the frequency, or

in the time domain, we may choose to study one particular area of the
waveform, ignoring all the rest. Local to that one particular area, the
waveform may have a lowest, and a highest point: the local maxima is,
clearly enough, the highest point (See "Local Minimas ").
LOW-PASS FILTER: This is a device which blocks frequencies higher than a

predetermined point in the frequency spectrum; thereby allowing only
those frequencies lower than that point to pass through, since all the
points higher than that point would be multiplied by zero. The ideal filter
has an absolute transition to zero, rather than the non-ideal attenuation.
Please see the graphical description below:
LOW VOLTAGE DETECTORS: This is circuitry which alerts the Central

Processing Unit, or Microcontroller Peripheral that the amplitude of the
signal in question has dropped below a certain point.
MAGNITUDE: This is a measure of the intensity of a signal.
MAIN LOBE: The main lobe is located by searching, in the frequency domain,
for the largest peak among a group of smaller peaks which taper up to
it. The main lobe represents the area within which most of the signal's
frequency is found. If the signal were composed of exactly one sinusoid
operating at a frequency of exactly 440 Hertz, we would find, ideally
speaking, a spike at 440 Hertz on the frequency plot.
MAPPING: Using an established mathematical relationship to transfer from one

domain to another.
MC6800: A 4-bit Motorola MicroController.
MC68332: A Motorola 32-bitMCU.
MC68HCllA8: An 8-bit Motorola MicroController.
MICROCONTROLLER: A bunch of mostly digital and somewhat analog

circuitry that helps us control mainly mechanical equipment; and, as we
know, also helps us perform low-level DSP operations.
G-10
Glossary
MICROCONTROLLER PERIPHERALS (MCU): These are functions through

which the MCU communicate to the outside world. (Examples: AlD, Timers,
Serial I/O ports, etc.)
MODULATION: This is the process whereby we modify the frequency of an existing
signal. This word describes the act of multiplying two signals together with
the purpose in mind of modifying the frequency properties which each
originally had to arrive at one signal with an entirely new frequency domain
signature - phew! that was a mouthful- (See "Carrier Signal" for a well
known application of this method).
NETWORK FUNCTION: When we are talking about the design of filters in DSP,
this idea refers to the ratio of input to output terms.
NATURAL FREQUENCY: Every material on earth has a frequency at which it
vibrates best as a result of its own physical properties. This is known as its
resonant, or natural frequency.
NON-CAUSAL: If a system is non-causal then it is a system whose output is
independent of the inputs (see "Causal").
NON-CAUSAL FILTER: A device whose signal-altering capabilities do not depend
upon previous inputs to the system.
NYQUIST CRITERIA: The rule which states that the frequency at which we sample
the incoming signal must be twice that of the highest frequency inherent in
that signal.
ODD HARMONICS: (See "Harmonics "). Odd harmonics are individual frequencies
which are odd integer multiples of the fundamental or main frequency.
ONLY GAME IN TOWN: The only way of doing something.
OPAMP: Operational amplifier.
OPERATIONAL AMPLIFIER: A bunch of analog circuitry which can filter, boost
or otherwise process analog signals.
OSCILLOSCOPE: This is device which is used to view and analyze the time domain
activity of any signal.
OUT-OF-PHASE: Shifted 90 degrees.
PARALLEL TRANSFER: The transfer of bit patterns, pattern by pattern. In other
words, the bits are sent in parallel.
PARALLEL INTERFACE ADAPTOR(PIA): This is a device which can link an
MCU to a MCU peripheral device.
G-11
Glossary
PASSBAND EDGE: This is a boundary of a filter within which frequencies are

passed. PASSBANDS: These are ranges of frequencies which are not
filtered out by a digital, or analog, filter.
PASSBAND RIPPLE: (See "Gibbs Phenomenon', or, 'Ringing")
PHASE: Phase is a measure which helps to show us the difference between two
cyclical signals. Below we see a graph of a sine wave and a cosine wave.
If each cycle takes one second, we can see that the time delay between
them is 114 of a second; however, the phase delay would be expressed
not in terms of time, but in terms of cycle differences. In this case we
see that we have a 1/4 cycle difference between the two signals. In other
words, if we moved the sine wave over by 114 of a cycle, we would get
the cosine wave.
PHASE, ABSOLUTE: Under this condition we know exactly when the signal
started up(See "Phase").
PHASE, RELATIVE: This phrase refers to a common condition of not knowing

when a signal has started up; but we do know the relationship between
that signal relative to another signal (See "Phase").
PIEZOBUZZER: A piece of piezo-electric material which acts as a buzzer.
PISO: Parallel In Serial Out.
PLCC: Plastic leaded chip carrier.
PREEMPHASIS: A frequency enhancer; that is, a device that will boost the
frequencies that we choose to boost (See "Weighting").
QUANTIZATION: When we tum a discrete signal into a digital one, this is

known as quantizing.
RAM BUFFER: Random Access Memory (RAM) where we store the quantized
signal and the coefficients we will be using to modify it.
REAL-ESTATE: Surface area, in this case referring to printed circuit board area.
RESONANCE: (See "Natural Frequency").
G-12
Glossary
RIDING ON TOP: Superimposed on

RINGING: This is a phenomenon which results from the energy required to change
from one intensity level, or, amplitude, to another higher or lower one.
Intuitively we can see that as the intensity level shoots up from a lower to a
higher level, it may, and quite often will, overshoot the mark -like letting
go of one end of a loose spring which then will first overshoot the mark, and
then undershoot it as it returns. Finally, it will settle down to the intended
level. This is known as ringing( see "Gibbs Phenomenon").
RMS (ROOT MEAN SQUARE): The fundamental measurement of the magnitude
of the A.C. signal. Mathematically, we compute this by squaring the signal,
taking its average, and then taking the square root of the resulting number.
RMS-TO-DC FUNCTION: The method by which we arrive at the DC equivalent of
an RMS value. The DC value that produces the same amount of heat as its
RMS counterpart is the equivalent value.
ROM: Read only memory.
SAMPLE: One instantaneous value of an analog waveform.
SAMPLE-AND-HOLDAMPLIFIER: This is the analog circuit that takes an instan-
taneous value of an incoming waveform and holds it until it is time to take
another sample (See "Sample-and-Hold Function").
SAMPLE-AND-HOLD FUNCTION: This is the bridge between the discrete form
of a signal and the analog representation of it. This function describes what
a Sample-And-Hold circuit would do; which is to take an instantaneous value
of a waveform and hold it until such time as another sample is taken. During
the time that it is held, the instantaneous value is converted (quantized) into
a digital value and stored in a RAM Buffer (See "Sample-and-Hold Ampli-
fier").
SERIAL: Sequential transfer of bit patterns, one bit at a time; and furthermore, one
right after the other.
SERIAL COMMUNICATIONS INTERFACE: This device allows a bit-by-bit
transfer of information outside the MeU.
SIDELOBES: Whenever we look at a signal's frequency domain plot, we see many
small peaks each getting successively larger and larger until it reaches the
largest one in that group; then, on the other side of the largest one, each peak
gets successively smaller. Sidelobes are all those smaller ones. (See "Main
Lobe").
G-13
Glossary
SIGNAL: A signal is anything that has meaning or significance to the person or

device receiving it. In other words, a signal is anything that contains
information of value to the receiver, like an electrical signal, for instance.
SIGNAL BANDWIDTH: The bandwidth of a signal is defined by the number

of different sinusoids that compose it. If a signal contains three sinusoids
each operating at a different speed, then its bandwidth is defined by the
highest to the lowest of those three contributing frequencies.
SIGNAL FILTER: (See "Filter").
SIGNAL-TO-NOISE RATIO: In designing signal filters, the DSP engineer

must account for noise in the system. This ratio is a characteristic of the
system which the engineer is always trying to maximize by minimizing
the noise with appropriate filtering techniques.
SIGNAL PROCESSING: This the general title for the branch of Electrical
Engineering which is devoted finding new ways to alter electrical signals
for useful ends whether it be by analog, or digital means.
SINC FUNCTION: (1) This function is fundamental to the field of signal

processing. It is often used to show the duality concept when referring
to the relationship between the frequency and the time domains. It is
used this way since the time domain representation of the frequency
domain sinc function is a rectangular pulse; and vice versa, the time
domain sinc function is a frequency domain "rectangular pulse," such
as a bandpass filter(See "Sinc Pulse"). (2) The function used to 'trace
out' a set of discrete impulses at the output of a Digital-to- Analog
Converter, thereby producing an analog waveform.
SINC PULSE: This refers to the dual of the sinc function in the opposite domain
which is always a rectangular pulse(See "Sinc Function").
G-14
Glossary
SINE WAVE: Below we see a graphical description of a sine wave:
For an explanation of the mathematical detail behind these descriptions, please refer
to a Signal Processing text.
SINGLE FREQUENCY: When the author of this chapterrefers to "single frequency"

he means a sinusoid which has only one frequency.
SINUSOIDS: One idea in DSP which one should understand is that of the Sinusoid.
Sinusoids are the signals which when combined in various ways compose all
other signals that occur naturally or synthetically in the universe. These are
the signals which the DSP enthusiast will filter orreshapefor whateverreason.
When the mathematician, Joseph Fourier, examined these signals, he made
the remarkable discovery that all the signal he studied are part of the set of
signals which are composed entirely of sine waves and cosine waves.
SIPO: Serial In Parallel Out.
SNR: Signal to noise ratio.
SOIC: Small Outline Integrated Circuit.
SPEC: Specification.
SPECTRA: This word is the plural form of the word, spectrum (see "Spectrum").
SPECTRAIANALYSIS: In one sense, we refer here to the thought process whereby

we study and dissect the frequency domain signature of a signal in an effort
to gain some understanding of its characteristics. In another sense, it is the
act whereby a frequency domain representation of a signal is broken down
into its component parts in order to return it to its time domain representation;
which, by another name, is the inverse fourier transformation.
G-15
Glossary
SPECTRAL SPIKE: Please refer to the graphic under the glossary heading
"Cosine Wave," or, "Sine Wave." A spectral spike in each of those cases
is that which is described graphically in the second figure of each entry.
It is the single frequency associated with a simple sinusoid.
SPECTRUM: This word refers to the way a signal looks when we view its
frequency characteristics. For example, when we look at a rainbow, we
are actually looking at how often particular light waves finish one cycle
of their life. If one of them finishes one cycle in 15 seconds, then when
we view its frequency domain description, we see that its spectrum, or
range of frequencies fall into the category of blue colors. A spectrum is
simply a range of frequencies.
SPI: Serial Port Interface.
SPIKE: In the frequency domain, this is an area, denoted by a main lobe, where
we would find a major concentration of a particular frequency. In the
time domain, this can be a representation of an instantaneous energy
surge which appears suddenly in time, and disappears just as quickly;
which we would view as a spike on an oscilloscope.
SQUARE WAVE: In this section of his chapter, the author is attempting to show
how all signals are composed of one or more simple sinusoids. The
square wave is a signal which looks like one might expect:
STOPBANDATTENUATION: Everything before the passband edge is passed

frequencies, and everything after it until the stopband edge, is stopband
attenuation. This is the part of the frequency domain representation of a
filter which "rolls off' until it reaches zero amplitude. The major issue
which filter designers contend with is how many frequency bands must
the stopband attenuation pass through before it reaches zero? The fewer,
the better. However, the "faster" stopband attenuation implies more
computation, and therefore, more computing time and power to achieve
it.
STOPBAND EDGE: Somewhere, just before the stopband attenuation reaches

zero, we reach the stopband edge. As in the case of the passband edge,
calculus describes how we reach one or the other edge just before we
reach either the bottom or the top of our amplitude extremes.
G-16
Glossary
SUCCESSIVE APPROXIMATION CONVERTER: ThisisanAnalog-to-Digi-

tal Converter which is used in applications where the speed required is
no more than about one million hertz (which by today's standards is not
slow, but not blazingly fast either).
SUPERPOSITION: This is a mathematical principle which simply put, states

that we can add two signals to get another. This allows us to describe and
understand signals which are composed of many sub- signals. Convolu-
tion also uses the superposition principle.
SUPERPOSITION INTEGRAL: (See "Convolution Integral").
SYSTEM BANDWIDTH: We are referring here to the range of frequencies that

are relevant to a particular system. For example, the system bandwidth
of our telephone system is 4000 Hertz. This means that any signals from
zero to 4000 Hertz in frequency will be heard by whoever is at the other
end of the line.
TIME DELAY: The time delay is the difference in time between two signals.
TIME DOMAIN: When we speak of the time domain, we are signifying that we
are interested in any signal behavior, which we observe over time. We
could, for example, observe this behavior on an oscilloscope.
TIME DOMAIN DESCRIPTION: This kind of description can be a verbal, a

graphical, or a mathematical model, but it must contain elements which
always impinge upon the time-domain signal processing goals for which
we have set out.
TIME-FREQUENCY PAIRS: (See "Fourier Transform Pairs").
TRAIN OF IMPULSES: This refers to a set of equally spaced delta functions

each corresponding to a different point on a signal. This terminology is
generally used to describe the effect of sample-and-hold circuitry upon
the incoming waveform.
G-17
Glossary
TRANSFER FUNCTION: We are referring, here, to a relationship between the input

and the output of our Digital Signal Processing system, whatever that may
be. It is generally expressed as a ratio of the mathematical expression of the
output sequence, to the expression for the input sequence.
TRIP POINTS: A threshold at which something occurs.
UNIVERSAL ASYNCHRONOUS RECEIVER/TRANSMITTER (UART): This

device converts bits that are being transferred one at a time, into a set of bits
that are to be transferred as one packet. It can also convert a packet of bits
into bits that are to be transferred in serial (See "Serial" and "Parallel
Transfer").
VANTAGE POINTS: Views or Perspectives.
VAX PDP-ll: Alargemainframecomputer.
VCA: Voltage controller amplifier.
VOLTAGE FOLLOWER: Circuitry whose voltage output tracks the voltage input.
To be more precise, it makes a fine high current buffer because it has high
input impedance, low output impedance and unity gain.
WATCHDOG TIMER: These timers are used to protect against faulty computer
programs, or situations in which an instruction sequence begins to repeat
itself indefinitely and in a way that is outside the control of the user. This
countdown timer detects these situations by allowing only a certain number
of repetitions to occur before timing out and giving control back to the user.
WEIGHTING: Basically, this is a constant assigned to, say, one frequency item, in
order to indicate its importance. In general, it is any constant assigned to any
part of a signal to emphasize or de- emphasize it.
WINDOWS: A window, in the Digital Signal Processing sense, is a type of aperture

used to filter a given signal (see "Windowing"). The most famous windows
are: Rectangular, Hamming, Hanning, Gaussian, and Blackman-Harris.
When we get to the point where we want to select the right window for our
filtering purposes, we would begin by exercising our ability to juggle with
the tradeoffs inherent in each window's characteristics. In general, the
characteristics which we would attempt to keep aloft in our heads are: (1) the
width of the central peak, or main lobe; (2) the passband edge location; (3)
G-18
Glossary
is relative to the main lobe, and how 'quickly' the other sidelobes taper
off to nothing; and, (4) the amount of area beneath the window function's
frequency domain description. (Adapted from DSP in VLSI, by Richard
1. Higgins).
WINDOWING: Windowing is like taking a photograph. If we are very concerned

about getting the most complete photograph that we can, we concern
ourselves with getting the appropriate film for the occasion (color, or
black and white; high-speed shutter sensitivity; and so on). Then, before
actually pushing the shutter release button, we would try to get the best
possible picture of the event given the size of our frame. It is the same
way with windowing; that is, before we actually start snapping pictures,
we choose the right film (the window) which will have all the right
qualities given the kind of picture we are trying to take of the signal; then
we try to get everything into the picture that we need (choose the width
of the window properly); and then we start taking pictures.
G-19
References
The Fast Fourier Transform. E. Oran Brigham, Prentice-Hall, 1974.
Digital Signal Processing In VLSI, R. J. Higgins, Prentice-Hall, 1988.
TheFourierTransform And Its Awlications. 2nd edition, R. N. Bracewell, McGraw-Hill,

1978.
Signals And Systems. Alan V. Oppenheim and Alan S. Willsky with Ian T. Young,
Prentice-Hall, 1983.
Digital Signal Processing, A. V. Oppenheim and R. W. Schafer, Prentice-Hall, 1975.

NOTES

Motorola Digital Signal Processing

Uploaded by

Copyright:

Available Formats

Motorola Digital Signal Processing

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Motorola Digital Signal Processing

Uploaded by

Copyright:

Available Formats

_.

Reproduction or translation of any part of this

The Motorola Advanced Microcontroller Division gratefully acknowledges the con-

John Waite, who provided the glossary of technical terms.

Chapter 4 discusses the fundamental concepts of signal processing by reviewing some

Filters and Transfer Functions

General DSP Concepts

Digital Filter Design

T oday, the microcontroller (MCU)

A Sample MCU A typical combination of peripherals is

Features Already Built-in to This MCU:

• Memory-mapped digital 110: 24 bidirectional and 7 input-only lines. The

Chip Size Type Length x Width Area

8-pin DIP Analog timer 0.4 x 0.3 0.12

14-pin DIP Inverter 0.7 xO.3 0.21

24-pin DIP UART 1.2 x 0.6 0.72

28-pin DIP EPROM 1.4 x 0.6 0.84

28-pin DIP Timer 1.4 x 0.6 0.84

40-pin DIP PIA 2.0 x 0.6 1.20

Using individual peripherals and memo-

3.93 sq. in.

Now that we've mentioned the remark- What An MCU

It is not bulky. Offered in various pack-

It is not a panacea. You should be aware

At this point, you may have a suspicion The Human

Trying New Things

One Part Fits All ... Almost

Designing an MCU into a product af-

MCUs Promote Proper Design

We'll revisit this topic when discussing

The View From Production

Let's take a look at the MCV in action. ATypical MCU

Comparators may be used as the main

Keyboard 14c-)l~ LCD

Schedule next timer "tick"

Clear TIC To Zero

1) Update Time And Day

4) Check For User Entry

5) Service AID Temp. Sensors

6) Update Hvac Outputs

7) Service Lcd Display

SP.M. 12 A.M. SA.M.

rapidly changing. Based upon a rough thermodynamic calculation (done on the

Some Average{ing) Software

* Requests conversion of next channel and returns data

SPIFLP BRCLR 7,SPSR,SPIFLP Wait for SPI Xfer complete

JSR LOUPRSS Fi Iter input uith 6,/pt aug.

BRSET 1,TIC,ADCHl 1fT i c=2, dat a is Ch. 1

The left-right width of the block is the

There are two pieces of the code that

"'Purpose: To reduce uariatlons in ADC data frolltherllol

lOIlJPOss staTellpSnpJ jthis contains the latest

cpx'TilleBufr-l ;check. for latest eOllple location.

Following the flow of the code, if our

Averaging consists of two separate proc-

Now, turn the entire string of rectangles

This view of the memory roughly repre-

Let's pause at this point to review a rule

A(B+ C) =AB +AC