Matlab

Download as pdf or txt
Download as pdf or txt
You are on page 1of 152

(March 30, 2010 2:47 p.m.

A MATLAB Tutorial
Ed Overman
Department of Mathematics
The Ohio State University

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 3
1 Scalar Calculations . . . . . . . . . . . . . . . . . . . . . . 6
1.1 Simple Arithmetical Operations . . . . . . . . . . . . . . . . . 6
1.2 Variables . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Round-off Errors . . . . . . . . . . . . . . . . . . . . . . 9
1.4 Formatting Printing . . . . . . . . . . . . . . . . . . . . . 10
1.5 Common Mathematical Functions . . . . . . . . . . . . . . . . . 11
1.6 Complex Numbers . . . . . . . . . . . . . . . . . . . . . . 13
1.7 Script M-files . . . . . . . . . . . . . . . . . . . . . . . 13
1.8 Help! . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.9 Be Able To Do . . . . . . . . . . . . . . . . . . . . . . . 16
2 Arrays: Vector and Matrix Calculations . . . . . . . . . . . . . . . 16
2.1 Generating Matrices . . . . . . . . . . . . . . . . . . . . . 17
2.2 The Colon Operator . . . . . . . . . . . . . . . . . . . . . 21
2.3 Manipulating Matrices . . . . . . . . . . . . . . . . . . . . 21
2.4 Simple Arithmetical Operations . . . . . . . . . . . . . . . . . 25
2.5 Operator Precedence . . . . . . . . . . . . . . . . . . . . . 29
2.6 Be Careful! . . . . . . . . . . . . . . . . . . . . . . . . 29
2.7 Common Mathematical Functions . . . . . . . . . . . . . . . . . 31
2.8 Data Manipulation Commands . . . . . . . . . . . . . . . . . . 32
2.9 Advanced Topic: Multidimensional Arrays . . . . . . . . . . . . . . 34
2.10 Be Able To Do . . . . . . . . . . . . . . . . . . . . . . 35
3 Anonymous Functions, Strings, and Other Data Types . . . . . . . . . . 36
3.1 Anonymous Functions . . . . . . . . . . . . . . . . . . . . 37
3.2 Passing Functions as Arguments to Commands . . . . . . . . . . . . . 38
3.3 Strings . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.4 Advanced Topic: Cell Arrays and Structures . . . . . . . . . . . . . . 40
4 Graphics . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.1 Two-Dimensional Graphics . . . . . . . . . . . . . . . . . . . 43
4.2 Three-Dimensional Graphics . . . . . . . . . . . . . . . . . . 49
4.3 Advanced Graphics Techniques: Commands . . . . . . . . . . . . . . 51
4.4 Advanced Graphics Techniques: Handles and Properties . . . . . . . . . . 55
4.5 Be Able To Do . . . . . . . . . . . . . . . . . . . . . . . 56
5 Solving Linear Systems of Equations . . . . . . . . . . . . . . . . 57
5.1 Square Linear Systems . . . . . . . . . . . . . . . . . . . . 57
5.2 Catastrophic Round-Off Errors . . . . . . . . . . . . . . . . . . 60
5.3 Overdetermined and Underdetermined Linear Systems . . . . . . . . . . . 61
6 File Input-Output . . . . . . . . . . . . . . . . . . . . . . 63
7 Some Useful Linear Algebra Commands . . . . . . . . . . . . . . . 65
8 Programming in MATLAB . . . . . . . . . . . . . . . . . . . 71
8.1 Flow Control and Logical Variables . . . . . . . . . . . . . . . . 71
8.2 Matrix Relational Operators and Logical Operators . . . . . . . . . . . . 75
8.3 Function M-files . . . . . . . . . . . . . . . . . . . . . . 79
8.4 Odds and Ends . . . . . . . . . . . . . . . . . . . . . . 88
8.5 Advanced Topic: Vectorizing Code . . . . . . . . . . . . . . . . . 90
9 Sparse Matrices . . . . . . . . . . . . . . . . . . . . . . . 92
10 Initial-Value Ordinary Differential Equations . . . . . . . . . . . . . 95
10.1 Basic Commands . . . . . . . . . . . . . . . . . . . . . . 95
10.2 Advanced Commands . . . . . . . . . . . . . . . . . . . . 100
11 Boundary-Value Ordinary Differential Equations . . . . . . . . . . . 106
12 Polynomials and Polynomial Functions . . . . . . . . . . . . . . . 110
13 Numerical Operations on Functions . . . . . . . . . . . . . . . . 112
14 Discrete Fourier Transform . . . . . . . . . . . . . . . . . . . 115
15 Mathematical Functions Applied to Matrices . . . . . . . . . . . . . 122
Appendix: Reference Tables . . . . . . . . . . . . . . . . . . . . 125
Solutions To Exercises . . . . . . . . . . . . . . . . . . . . . . 137
ASCII Table . . . . . . . . . . . . . . . . . . . . . . . . . 139
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

2
Introduction

MATLAB is an interactive software package which was developed to perform numerical calculations
on vectors and matrices. Initially, it was simply a MATrix LABoratory. However, today it is much more
powerful:
• It can do quite sophisticated graphics in two and three dimensions.
• It contains a high-level programming language (a “baby C”) which makes it quite easy to code com-
plicated algorithms involving vectors and matrices.
• It can numerically solve nonlinear initial-value ordinary differential equations.
• It can numerically solve nonlinear boundary-value ordinary differential equations.
• It contains a wide variety of toolboxes which allow it to perform a wide range of applications from sci-
ence and engineering. Since users can write their own toolboxes, the breadth of applications is quite
amazing.
Mathematics is the basic building block of science and engineering, and MATLAB makes it easy to handle
many of the computations involved. You should not think of MATLAB as another complication program-
ming language, but as a powerful calculator that gives you fingertip access to exploring interesting prob-
lems in science, engineering, and mathematics. And this access is available by using only a small number
of commands and function† because MATLAB’s basic data element is a matrix (or an array).
This is a crucial feature of MATLAB — it was designed to group large amounts of data in arrays and
to perform mathematical operations on this data as individual arrays rather than as groups of data. This
makes it very easy to apply complicated operations to the data, and it make it very difficult to do it
wrong. In high-level computer languages you would usually have to work on each piece of data separately
and use loops to cycle over all the pieces. In MATLAB this can frequently do complicated “things” in
one, or a few, statements (and no loops). In addition, in a high-level language many mathematical op-
erations require the use of sophisticated software packages, which you have to find and, much worse, to
understand since the interfaces to these packages are frequently quite complicated and the documentation
must be read and mastered. In MATLAB, on the other hand, these operations have simple and consistent
interfaces which are quite easy to master. For an overview of the capabilities of MATLAB, type
>> demo
in the Help Navigator and click on MATLAB.
This tutorial is designed to be a concise introduction to many of the capabilities of MATLAB. It makes
no attempt to cover either the range of topics or the depth of detail that you can find in a reference man-
ual, such as Mastering MATLAB 7 by Duane Hanselman and Bruce Littlefield (which is over 850 pages
long) or MATLAB Guide, 2nd edition by Desmond and Nicholas Higham (which is almost 400 pages long).
This tutorial was initially written to provide students with a free “basic” overview of commands which
are useful in an undergraduate course on linear algebra. Over the years it has grown to include courses in
ordinary differential equations, mathematical modelling, and numerical analysis. It also includes an intro-
duction to two- and three-dimensional graphics because graphics is often the preferred way to present the
results of calculations.
In this tutorial MATLAB is first introduced as a calculator and then as a plotting package. Only after-
wards are more technical topics discussed. We take this approach because most people are quite familiar
with calculators, and it is only a small step to understand how to apply these same techniques to matrices
rather than individual numbers or varibles. In addition, by viewing MATLAB as a simple but powerful
calculater, rather than as a complicated software package or computer language, you will be in the correct
frame of mind to use MATLAB.
You should view MATLAB as a tool that you are “playing with” — trying ideas out and seeing how


There is a technical distinction between a command and a function in MATLAB: input arguments to com-
mands are not enclosed in parentheses (they are separated by spaces) and there are no output arguments (i.e., a
command cannot be on the right-hand side of an equal sign). In reality, this is a very fine distinction since many
commands can be written as functions by putting the arguments between parentheses and separating them with
commas. We will generally use the term functions unless there is a reason to make a distinction.

3
they work. If an idea works, fine; if it doesn’t, investigate further and figure out why. Maybe you mis-
understood some MATLAB command, or maybe your idea needs some refinement. “Play around” in-
teractively and figure it out. There are no hard and fast rules for figuring it out — try things and see
what happens. Don’t be afraid to make mistakes; MATLAB won’t call you an idiot for making a mistake.
When you first learned to ride a bicycle, you fell down a lot — and you looked pretty silly. But you kept
at it until you didn’t fall down. You didn’t study Newton’s laws of motion and try to analyze the motion
of a bicycle; you didn’t take classes in how to ride a bicycle; you didn’t get videos from the library on how
to ride a bicycle. You just kept at it, possibly with the assistance of someone who steadied the bicycle and
gave you a little push to get you started. This is how you should learn MATLAB.
However, this tutorial is not designed for “playing around”. It is very ordered, because it has been de-
signed as a brief introduction to all the basic topics that I consider important and then as a reference
manual. It would be very useful for students to have a document which uses this “play around” approach
so you would learn topics by using them in exploring some exercise. This is how workbooks should be
written: present some exercise for students to investigate, and let them investigate it themselves. And
these exercises should be interesting, having some connection to physical or mathematical models that
the students — or at least a reasonable fraction thereof — have some knowledge of and some interest in.
This tutorial is designed to be a reference manual that could be used alongside such a workbook — if only
someone would write it.

Summary of Contents
We have tried to make this tutorial as linear as possible so that the building blocks necessary for a sec-
tion are contained in preceding sections. This is not the best way to learn MATLAB, but it is a good way
to document it. In addition, we try to separate these building blocks and put them in short subsections so
that they are are easy to find and to understand. Next, we collect all the commands discussed in a sub-
section and put them in a box at the end along with a very brief discussion to make it easy to remember
these commands. Finally, we collect all these commands and put them in the appendix again boxed up by
topic. MATLAB has a HUGE number of commands and functions and this is one way to collect them
for easy reference.
Warning: Usually we do not dis uss the omplete behavior of these ommands, but only their most \use-
ful" behavior. Typing
>> help <command>
or
>> doc <command>
gives you omplete information about the ommand.
Notation: help <command> means to enter whatever command you desire (without the braces).
help command means to type these two words as written.
Section 1 of this tutorial discusses how to use MATLAB as a “scalar” calculator, and section 2 how to
use it as a “matrix” calculator. Following this, you will be able to set up and solve the matrix equation
Ax = b where A is a square nonsingular matrix.
Section 4 discusses how to plot curves in two and three dimensions and how to plot surfaces in three di-
mensions. These three sections provide a “basic” introduction to MATLAB. At the end of each of these
three sections there is a subsection entitled “Be Able To Do” which contains sample exercises to make sure
you understand the basic commands discussed. (Solutions are included.)
You have hopefully noticed that we skipped section 3. It discusses a number of minor topics. Since they
are useful in generating two- and three-dimensional plots, we have included it here.
The following sections delve more deeply into particular topics. Section 5 discusses how to find any and
all solutions of Ax = b where A ∈ Cm×n need not be a square matrix; there might be no solutions, one
solution, or an infinite number to this linear system. When no solution exists, it discusses how to calcu-
late a least-squares solution (i.e., the “best” approximation to a solution). In addition, it discusses how
round-off errors can corrupt the solution, and how to determine if this is likely to occur.
Section 6 is quite brief and discusses advanced commands to input data into MATLAB and output it
to a file. (The basic commands are discussed in subsection 4.1.) This is useful if the data is being shared

4
between various computer programs and/or software packages.
Section 7 discusses a number of commands which are useful in linear algebra and numerical linear alge-
bra. Probably the most useful of these is calculating some or all of the eigenvalues of a square matrix.
Section 8 discusses MATLAB as a programming language — a very “baby C”. Since the basic data ele-
ment of MATLAB is a matrix, this programming language is very simple to learn and to use. Most of this
discussion focuses on writing your own MATLAB commands, called function m-files (which are similar to
functions in C and to functions, more generally subprograms, in Fortran). Using these functions, you can
code a complicated sequence of statements such that all these statements as well as all the the variables
used by these commands are hidden and will not affect the remainder of your MATLAB session. The only
way to pass data into and out of these functions is through the argument list.
Section 9 discusses how to generate sparse matrices (i.e., matrices where most of the elements are zero).
These matrices could have been discussed in section 2, but we felt that it added too much complexity at
too early a point in this tutorial. Unless the matrix is very large it is usually not worthwhile to gener-
ate sparse matrices — however, when it is worthwhile the time and storage saved can be boundless.
Section 10 discusses how to use MATLAB to numerically solve initial-value ordinary differential equa-
tions. This section is divided up into a “basic” part and an “advanced” part. It often requires very little
effort to solve even complicated odes; when it does we discuss in detail what to do and provide a num-
ber of examples. Section 11 discusses how to use MATLAB to numerically solve boundary-value ordinary
differential equations.
Section 12 discusses how to numerically handle standard polynomial calculations such as evaluating
polynomials, differentiating polynomials, and finding their zeroes. Polynomials and piecewise polynomials
can also be used to interpolate data.
Section 13 discusses how to numerically calculate zeroes, extrema, and integrals of functions.
Section 14 discusses the discrete Fourier transform and shows how it arises from the continuous Fourier
transform. We also provide an example which shows how to recover a simple signal which has been
severely corrupted by noise.
Finally, section 15 discusses how to apply mathematical functions to matrices.
There is one appendix which collects all the commands discussed in this tutorial and boxes them up by
topic. If a command has more than one use, it might occur in two or more boxes.
This tutorial closes with an index. It is designed to help in finding things that are “just on the tip of
your tongue”. All the MATLAB commands discussed here are listed at the beginning of the index, as well
as alphabetically throughout the index.
Notation: A variable, such as x, can represent any number of types of data, but usually it represents a
scalar, a vector, or a matrix. We distinguish them by using the lowercase x when it is a scalar
or a vector, and the uppercase X when it is a matrix. Also, in MATLAB vectors can be either
row vectors, e.g., (1, 2, 3) or column vectors (1, 2, 3)T (where “ T ” is the transpose of a vec-
tor or matrix). In a linear algebra setting we always define x to be a column vector. Thus,
for example, matrix-vector multiplication is always written as A ∗ x and the inner product of
the two vectors x and y is x′ *y, i.e., x1 y1 + x2 y2 + . . . + xn yn (where “ ′ ” is the MATLAB
command to take the transpose of a real vector or matrix.

5
1. Scalar Calculations
1.1. Simple Arithmetical Operations
MATLAB can be used as a scientific calculator. To begin a MATLAB session, type matlab or click
on a MATLAB icon and wait for the prompt, i.e., “ >> ”, to appear. (To exit MATLAB, type exit or
quit.) You are now in the MATLAB workspace.
You can calculate 3.17 · 5.7 + 17/3 by entering
>> 3.17*5.7 + 17/3
and 220 by entering
>> 2 20
And
P12 ˆ
j=1 1/j can be entered as
>> 1 + 1/2 + 1/3 + 1/4 + 1/5 + 1/6 + 1/7 + 1/8 + 1/9 + 1/10 + 1/11 + 1/12
You can enter a number in scientific notation using the “ ” operator. For example, you can enter
2×10−20 by
ˆ
>> 2*10 -20
ˆ
MATLAB, however, uses “e” to represent “10 ” so that MATLAB displays
2.0000e-20
ˆ
The “standard” way to input 2×10−20 is as 2e-20 or 2E-20 or 2.e-20 or 2.E-20 (even
2.0000000e-00020 is acceptable).
Warning: 10−20 cannot be input as e-20, but must be input as 1e-20 or 1E-20 or 1.e-20 or 1.E-20
or . . . √
MATLAB can also handle complex numbers, where i or j represents −1 . For example, 5i can be
input as 5i or as 5*i, while 5×1030 i can be input as 5e30i or as 5e30*i or as 5*10 30*i, but not
as 5*10 30i (which MATLAB considers to be 5×1030i ). To calculate (2 + 2i)4 , enter ˆ
ˆ
>> (2 + 2i) 4
ˆ
and MATLAB returns −64.
You can also save all of your input to MATLAB and most of the output (plots are not saved) by using
the diary command. This archive of your work can be invaluable when you are solving homework prob-
lems. You can later use an editor to extract the part you want to turn in, while “burying” all the false
starts and typing mistakes that occur. Conversely, if you are involved in a continuing project, this archive
can be invaluable in keeping a record of your progress.
If you do not specify a file, this archive is saved to the file diary (no extension) in the present direc-
tory. If the file already exists, this is appended to the end of the file (i.e., the file is not overwritten).
Because of this feature you can use the diary command without fear that crucial work will be overwrit-
ten.
If you are entering a line and make a mistake, there are a number of ways you can correct your error:
• you can use the backspace or delete key to erase all the text back to your mistake,
• you can use the left-arrow key, i.e., “ ←”, and the right-arrow key, i.e., “ →”, to move back and forth
in the line, or
• you can use the mouse to move back and forth in the line.
Frequently, you will want to reexecute the previous line, or another previous line. For example, you
might have made a mistake in the previous line and so it did not execute, or did not execute correctly.
Of course, you can just retype the line — but, if it is very long, this can get very time-consuming. Instead,
you can use the up-arrow key, i.e., “ ↑”, to move backward, one statement at a time (or the down-arrow
key, i.e., ↓” to move forward). Then hit the enter (or the return) key to execute the line.

6
1.2. Variables

Arithmetical Operations

a+b Addition. a/b Division.


a-b Subtraction. a\b Left division, (this is exactly the same as
a*b Multiplication. b/a ).
a b Exponentiation (i.e., ab ).
ˆ
diary Saves your input to MATLAB and most of the output to disk. This command toggles
diary on and off. (If no file is given, it is saved to the file diary in the current direc-
tory.)
diary on turns the diary on.
diary off turns the diary off.
diary ′ <file name>′ saves to the named file.

↑ The up-arrow key moves backward in the MATLAB workspace, one line at a time.

1.2. Variables

Notation: We always use lowercase letters to denote scalar variables.


Variables can be used to store numerical values. For example, you can store the value 21/3 in the vari-
able x by entering
>> x = 2 (1/3)
ˆ
This variable can then be used on the right-hand side of an equation such as
>> fx = 3*x 6 - 17*x 3 + 79
ˆ ˆ
There can also be more than one command on a line. For example, if you type
>> x = 2 (1/3); fx = 3*x 6 - 17*x 3 + 79; g = 3/fx;
ˆ ˆ ˆ
then all three commands will be executed. Nothing will be printed out because semicolons follow each
command. If you want everything printed out then type
>> x = 2 (1/3), fx = 3*x 6 - 17*x 3 + 79, g = 3/fx
ˆ ˆ ˆ
Thus, you can separate statements on a line by commas or semicolons. If semicolons are used, the results
of the statement are not displayed, but if commas are used, the results appear on the computer screen.
Warning: A variable can be overwritten at will. For example, at present x = 21/3 . If you now type
>> x = x + 5
then x becomes 21/3 + 5. No warning messages are printed if a variable is overwritten, just as
in a programming language.
Although we do not discuss vectors and matrices until the next section, it is important to understand
that MATLAB considers scalar variables to be vectors of length one or matrices of size 1×1. For exam-
ple, if you type
>> fx
the number 57 is returned. But you can also type
>> fx(1)
or
>> fx(1,1)
and obtain the same result.
Character strings can also be stored in variables. For example, to store the string “And now for some-
thing completely different” in a variable, enter
>> str = ′ And now for something completely different′
(We discuss text variables in more detail in section 3.)
Note: To put a single quote mark into the string, use two single quote marks.
You can change a variable from a scalar to a vector or a matrix whenever you desire — or whenever
you forget that the variable has already been defined. Unlike C, for example, variables do not needed to

7
1.2. Variables

be declared (or typed). A variable springs into existence the first time it is assigned a value, and its type
depends on its context.
At start-up time, MATLAB also contains some predefined variables. Many of these are contained in the
table below. Probably the most useful of these is pi.
Warning: Be careful since you can redefine these predefined variables. For example, if you type
>> pi = 2
then you have redefined π — and no error messages will be printed out!
Another very useful predefined variable is ans, which contains the last calculated value which was not
stored in a variable. For example, it sometimes happens that you forget to put a value into a variable.
Then MATLAB sets the expression equal to the variable ans. For example, if you type
>> (3.2*17.5 - 5/3.1) 2
ˆ
but then realize that you wanted to save this value, simply enter
>> x = ans
and x now contains (3.2 · 17.5 − 5/3.1)2 .
In MATLAB it is trivial to display a variable: simply type it. For example, if x has the value −23.6
then
>> x
returns

x =

-23.6000
It is sometimes useful to display the value of a variable or an expression or a character string without
displaying the name of the variable or ans. This is done by using disp. For example,
>> disp(x)
>> disp(pi 3)
ˆ
>> disp(′ And now for something completely different′)
>> disp(′ ------------------------------------------′)
displays
-23.6000

31.0063

And now for something completely different


------------------------------------------
(The command fprintf, which will be discussed in section 6, allows much finer formatting of variables.)
Note: When disp displays a variable or an array or an expression, it follows with a blank line. However,
when it displays a string or a string variable, it does not.
Variables can also be deleted by using clear. For example, to delete x type
>> clear x
Warning: This is a very dangerous ommand be ause it is so easy to lose a great deal of work.
If you mean to type
>> clear x
but instead you type
>> clear
you will delete all the variables you have reated in the workspa e!

8
1.3. Round-off Errors

Predefined Variables

ans The default variable name when one has not been specified.
pi π.
eps Approximately the smallest positive real number on the computer such that
1 + eps 6= 1 .
Inf ∞ (as in 1/0 ). You can also type inf.
NaN Not-a-Number (as in 0/0 ). You can also type nan.

i √−1 .
j −1 (the same as i because engineers often use these interchangeably).
realmin The smallest “usable” positive real number on the computer. This is “approximately”
the smallest positive real number that can be represented on the computer (on some
computer realmin/2 returns 0 ).
realmax The largest “usable” positive real number on the computer. This is “approximately” the
largest positive real number that can be represented on the computer (on most computer
2*realmax returns Inf ).

About Variables

Variables: are case sensitive (so xa is not the same as Xa ).


can contain up to 31 characters (but this is certainly “overkill”).
must start with a letter, and can then be followed by any number of letters, numbers,
and/or underscores (so z 0 is allowed).
do not need to be declared or typed.
To display a variable, type it alone on a line.
To delete a variable, type clear <variable>.
This is a very dangerous ommand | use it at your own risk.
disp Displays a variable or an expression without printing the variable name or ans.
, Separates multiple statements on the same line. The results appear on the screen.
; When this ends a MATLAB command, the result is not printed on the screen. This can
also separate multiple statements on the same line.

1.3. Round-off Errors


The most important principle for you to understand about computers is the following.
Principle 1.1. Computers annot add, subtra t, multiply, or divide orre tly!
Computers do integer arithmetic correctly (as long as the numbers are not too large to be stored in the
computer). However, computers cannot store most floating-point numbers (i.e., real numbers) correctly.
For example, the fraction 1/3 is equal to the real number 0.3333 . . . Since a computer cannot store this
infinite sequence of threes, the number has to be truncated.
eps is “close to” the difference between the exact number 1/3 and the approximation to 1/3 used in
MATLAB. It is defined to be the smallest positive real number such that 1 + eps > 1 (although it is
not actually calculated quite this accurately). For example, in MATLAB 1 + 0.1 is clearly greater than
1; however, on our computer 1 + 1e-40 is not. To see this, when we enter
>> (1 + .1) - 1
we obtain 0.1000 as expected.
Note: MATLAB guarantees that the expression in parentheses is evaluated first, and then 1 is subtracted
from the result.
However, when we enter

9
1.4. Formatting Printing

>> (1 + 1.e-40) - 1
MATLAB returns 0 rather than 1.e-40. The smallest positive integer n for which
>> (1 + 10 (-n)) - 1
ˆ
returns 0 is computer dependent. (On our computer it is 16.) What is not computer dependent is that
this leads to errors in numerical calculations. For example, when we enter
>> n = 5; ( n (1/3) ) 3 - n
ˆ ˆ
MATLAB returns -1.7764e-15 rather than the correct result of 0. If you obtain 0, try some different
values of n. You should be able to rerun the last statement executed without having to retype it by us-
ing the up-arrow key. Alternatively, on a Mac or a PC use the copy command in the menu; in Unix enter
p.
ˆ
Note: It might not seem important that MATLAB does not do arithmetical operations pre isely. However,
you will see in subsection 5.2 that there are simple examples where this can lead to very incorrect
results.
One command which is occasionally useful when you are just “playing around” is the input command,
which displays a prompt on the screen and waits for you to enter some input from the keyboard. For ex-
3
ample, if you want to try some different values of n in experimenting with the expression (n1/3 ) − n,
enter
>> n = input(′n = ′ ); ( n (1/3) ) 3 - n
ˆ ˆ
The argument to the command input is the string which prompts you for input, and the input is stored
in the variable n; the semicolon keeps the result of this command from being printed out. You can easily
rerun this line for different values of n (as we described above) and explore how round-off errors can affect
simple expressions.
Warning: eps and realmin are very different numbers. realmin is approximately the smallest positive
number that can be represented on the computer, whereas eps is approximately the small-
est positive number on the computer such that 1 + eps 6= 1. ( eps/realmin is larger than the
total number of atoms in the known universe.)

Request Input

input(′<prompt>′) Displays the prompt on the screen and waits for you to enter whatever is
desired.

1.4. Formatting Printing


3
The reason that (n1/3 ) − n can be nonzero numerically is that MATLAB only stores real numbers to a
certain number of digits of accuracy. Type
>> log10(1/eps)
and remember the integer part of this number. This is approximately the maximum number of dig-
its of accuracy of any calculation performed in MATLAB. For example, if you type 1/3 in MATLAB the
result is only accurate to approximately this number of digits. You do not see the decimal representation
of 1/3 to this number of digits because on start-up MATLAB only prints the result to four decimal digits
— or five significant digits if scientific notation is used (e.g., the calculation 1/30000 is displayed in sci-
entific notation). To change how the results are printed out, use the format command in MATLAB. Use
each of these four format functions and then type in 1/3 to see how the result is printed out.

10
1.5. Common Mathematical Functions

Format Options

format short The default setting.


format long Results are printed to approximately the maximum number of digits of accuracy
in MATLAB.
format short e Results are printed in scientific notation using five significant digits.
format long e Results are printed in scientific notation to approximately the maximum number
of digits of accuracy in MATLAB.
format short g Results are printed in the best of either format short or format short e.
format long g Results are printed in the best of either format long or format long e.

1.5. Common Mathematical Functions


MATLAB contains a large number of mathematical functions. Most are entered exactly as you would
write them mathematically. For example,
>> sin(3)
>> exp(2)
>> log(10)
return exactly what you would expect. As is common in programming languages, the trig functions are
evaluated in radians.†
Almost all the functions shown here are built-in functions. That is, they are coded in C so they execute
very quickly. The one exception is the factorial function, i.e., n! = 1 · 2 · 3 . . . n, which is calculated by
>> factorial(n)
Note: This function is actually calculated by generating the vector (1, 2, . . . , n) and them multiplying all
its elements together by prod([1:n]). (We discuss the colon operator in section 2.2.)
There is an important principle to remember about computer arithmetic in MATLAB.
Principle 1.2. If all the numbers you enter into MATLAB to do some al ulation are \rea-
sonably large" and the result of this al ulation is one or more numbers whi h are \ lose to"
eps , it is very likely that the number or numbers should be zero.
As an example, enter
>> deg = pi/180; th = 40; 1 - ( cos(th*deg) 2 + sin(th*deg) 2 )
ˆ ˆ
The result is 1.1102e-16. Clearly, all the numbers entered into this calculation are “reasonable” and the
result is approximately eps. Obviously, the result is supposed to be zero since, from the Pythagorean
theorem
cos2 θ + sin2 θ = 1
for all angles θ. MATLAB tries to calculate the correct result, but it cannot quite. It is up to you to
interpret what MATLAB is trying to tell you.
Note: If you obtained zero for the above calculation, try
>> th = input(′angle = ′ ); 1 - ( cos(th*deg) 2 + sin(th*deg) 2 )
ˆ
for various angles.‡ Some of these calculations should be nonzero.
ˆ
There are a number of occasions in this overview where we reiterate that MATLAB cannot usually cal-
culate results exactly. Sometimes these errors are small and unimportant — other times they are very
important.
Warning: There is one technical detail about functions that will trip you up occasionally: how does MAT-
LAB determine whether a word you enter is a variable or a function? The answer is that MAT-
LAB first checks if the word is a variable and only if it fails does it check if the word is a func-
tion. For example, suppose you enter
>> sin = 20

A simple way to calculate sin 40◦ is to type
>> deg = pi/180; sin(40*deg)

Be sure to define deg = pi/180 beforehand.

11
1.5. Common Mathematical Functions

by mistake (possibly you meant bin = 20 but were thinking about something else). If you now
type
>> sin(3)
MATLAB will reply
??? Index exceeds matrix dimensions.
because it recognizes that sin is a variable. Since MATLAB considers a variable to be a vector
of length one, its complaint is that you are asking for the value of the third element of the vec-
tor sin (which only has one element). Similarly, if you enter
>> sin(.25*pi)
MATLAB will reply
Warning: Subscript indices must be integer values.
because it thinks you are asking for the .25π-th element of the vector sin. The way to undo
your mistake is by typing
>> clear sin

Some Common Real Mathematical Functions

abs(x) The absolute value of x. factorial(n) n! for n a non-negative integer.


acos(x) arccos x. fix(x) If x ≥ 0 this is the largest inte-
acosd(x) arccos x where the result is in ger which is ≤ x.
degrees. If x < 0 this is the smallest
acosh(x) arccosh x. integer which is ≥ x.
acot(x) arccot x. floor(x) This is the largest integer which
acotd(x) arccot x where the result is in is ≤ x.
degrees. log(x) The natural log of x, i.e., loge x.
acsc(x) arccsc x.
acscd(x) arccsc x where the result is in log10(x) The common log of x, i.e.,
degrees. log10 x.
asin(x) arcsin x. mod(x, y) The modulus after division.
asind(x) arcsin x where the result is in That is, x − n ∗ y where
degrees. n = floor(y/x).
asinh(x) arcsinh x. rem(x, y) The remainder of x/y. This is
atan(x) arctan x. almost the same as mod(x, y).
atand(x) arctan x where the result is in Warning: be careful if x < 0.
degrees. round(x) The integer which is closest to
atan2(x, y) arctan y/x where the angle is in x.
(−π, +π]. sec(x) sec x.
atanh(x) arctanh x. secd(x) sec x where x is in degrees.
ceil(x) The smallest integer which is sign(x) If x > 0 this returns +1,
≥ x. if x < 0 this returns −1, and
cos(x) cos x. if x = 0 this returns 0.
cosd(x) cos x where x is in degrees. sin(x) sin x.
cosh(x) cosh x. sind(x) sin x where x is in degrees.
cot(x) cot x. sinh(x) sinh
√ x.
cotd(x) cot x where x is in degrees. sqrt(x) x.
csc(x) csc x. tan(x) tan x.
cscd(x) csc x where x is in degrees. tand(x) tan x where x is in degrees.
exp(x) ex . tanh(x) tanh x.

12
1.7. Script M-files

1.6. Complex Numbers


MATLAB can work with complex numbers as easily as with real numbers. For example, to find the
roots of the quadratic polynomial x2 + 2x + 5 enter
>> a = 1; b = 2; c = 5;
>> x1 = ( -b + sqrt( b 2 - 4*a*c ) ) / (2*a)
ˆ
>> x2 = ( -b - sqrt( b 2 - 4*a*c ) ) / (2*a)
The output is
ˆ
-1.0000 + 2.0000i
and
-1.0000 - 2.0000i
As another example, to calculate eiπ/2 enter
>> exp(i*pi/2)
and obtain
0.0000 + 1.0000i
There are standard commands for obtaining the real part, the imaginary part, and the complex conju-
gate† of a complex number or variable. For example,
>> x = 3 - 5i
>> real(x)
>> imag(x)
>> conj(x)
returns 3, -5, and 3.0000 + 5.0000i respectively.
Note that many of the common mathematical functions can take complex arguments. Above, MATLAB
has calculated eiπ/2 , which is evaluated using the formula

ez = ex+iy = ex (cos y + i sin y) .

Similarly,
eiz + e−iz eiz − e−iz
cos z = and sin z = .
2 2i

Some Common Complex Mathematical Functions

abs(z) The absolute value of z = x + iy. conj(z) z∗ = x − iy.


angle(z) The angle of z. This is calculated imag(z) The imaginary part of z, i.e., y.
by atan2(y, x). real(z) The real part of z, i.e., x.

1.7. Script M-files


So far we have always entered MATLAB statements directly into the text window so that they are ex-
ecuted immediately. However, if we want to repeatedly execute a number of statements we have to put
them all on one line and reexecute the whole line. This line can get very l o o o n n n g! The solution is to
type the sequence of statements in a separate file named <file name>.m. It is easy to edit this file to
remove any errors, and the sequence can be executed whenever desired by typing
>> <file name>
The MATLAB statements themselves are not printed out, but the result of each statement is, unless
a semicolon ends it. This type of file is called a script m-file: when MATLAB executes the command


If a is a complex number, then its complex conjugate, denoted by a∗ is obtained by changing the sign of i
whenever it appears in the expression for a. √For example, if a = 3 + 17i, then a∗ = 3√− 17i; if a = eiπ/4 , then
a∗ = e−iπ/4 ; if a = (2 + 3i) sin(1 + 3i)/(3 − 5 i), then a∗ = (2 − 3i) sin(1 − 3i)/(3 + 5 i).

13
1.8. Help!

<file name> the contents of the file “ <file name>.m” are executed just as if you had typed them into
into the text window. We will not emphasize script m-files further, but you will find many occasions where
they are very helpful.
However, there is one point we need to emphasize. Make sure your file name is not the same as
one of MATLAB's commands or functions. If it is, your file will not execute — MATLAB’s will! To
check this, you can enter
>> type <file name>
which types out the entire file file name.m if it is written in MATLAB and types out
<file name> is a built-in function.
(A built-in function is written in C or Fortran and so cannot be viewed directly.) It your file is typed out,
fine; if not, change the name of your file.
A long expression can be continued to a new line by typing three periods followed by the “enter (or
P20
“return”) key. For example, j=1 1/j can be entered as
>> 1 + 1/2 + 1/3 + 1/4 + 1/5 + 1/6 + 1/7 + 1/8 + 1/9 + 1/10 + 1/11 + 1/12 + ...
1/13 + 1/14 + 1/15 + 1/16 + 1/17 + 1/18 + 1/19 + 1/20
although there are much better ways to obtain this same expression with many fewer keystrokes (as you
will see in subsection 2.8). Lines can also be continued in the MATLAB workspace by using three periods,
but it is much more common to use continuation in an m-file.
If your m-file is very long, it is often valuable to include comments to explain what you are doing. Each
line of comments must begin with the percent character, i.e., “%”. Comments can appear alone on a line
or they can follow a statement that you have entered.
Note: By the way, comment lines are particularly useful if you write a number of script m-files. It is very
easy to forget what this file is supposed to be doing.

Odds and Ends

type Displays the actual MATLAB code for a command or function or m-file
... Continue an expression onto the next line.
% Begin a comment

1.8. Help!
Before discussing how to obtain help in MATLAB, here is a good place to discuss a very frustrating
situation where you desperately need help: how do you abort a MATLAB statement which is presently
executing. The answer is simply to type C (that is, hold down the control key and type “c”).
ˆ
The on-line help facility in MATLAB is quite extensive. If you type
>> help
you will get a list of all the topics that you can peruse further by typing help followed by the name of
the topic. If you want help on a specific command, simply type help followed by the name of the com-
mand, i.e.,
help <command>
For example, if you forget the exact form of the format command, just type
>> help format
and you will see all the various ways that the output can be formatted.
Note: Typing
>> help ?
gives you lots of information about arithmetical and relational and logical operators and special
characters.
There is a more general command that can help you determine which commands might be of use. The
command lookfor searches through the first line of all MATLAB help entries for a particular keyword.
It is case insensitive so capital letters need not be used. For example,

14
1.8. Help!

>> lookfor plot


returns all the MATLAB commands that have something to do with plots. (There are over one hundred.)
This command may be useful — or it may not be. However, it is worth a try if you cannot remember the
name of the command you want to use.
Warning: All of the thousands of MATLAB commands have to be checked, so this command might run
slowly.
Note: The keyword need not be a complete word. For example, the keyword compl is contained in the
words “complement”, “complex”, “complete”, “completion”, and “incomplete” — and in the capi-
tals of all these words.
If you want to find out more about a specific command, enter
>> type <command>
If the command is written in MATLAB’s programming language (as discussed in section 8), the entire
function will be typed out. (The type command does not work on internal MATLAB commands, called
built-in function, which are coded in C.)
MATLAB also has an entire reference manual on-line which can be accessed by entering
>> doc
or
>> helpdesk
This HTML documentation is displayed using your Web browser. It generally gives much more informa-
tion than the help command, and in a more easily understood format.
After working for a while, you may well forget what variables you have defined in the workspace. Sim-
ply type who or whos to get a list of all your variables (but not their values). who simply returns the
names of the variables you have defined, while whos also returns the size and type of each variable. To
see what a variable contains, simply type the name of the variable on a line.
By the way, the demonstrations available by running demo show many of the capabilities of MATLAB
and include the actual code used. This is always a good place to look if you are not sure how do do some-
thing.
Two commands that don’t quite fit in any category are save and load. However, since these com-
mands are occasionally very helpful, this is a good place to discuss them. Occasionally, you might need to
save one or more MATLAB variables: it might have taken you some time to generate these variables and
you might have to quit your MATLAB session without finishing your work — or you just might be afraid
that you will overwrite some of them by mistake. The save command saves the contents of all your vari-
ables to the file “ matlab.mat”. Use help or doc to learn how to save all the variables to a file of your
own choice and how to save just some of the variables. The load command loads all the saved variables
back into your MATLAB session.† (As we discuss in subsection 4.1, the load command can also be used
to input our own data into MATLAB.)


These variables are saved in binary format; when loaded back in using load the variables will be exactly the
same as before. The contents of this file can be viewed by the user with an editor — but the contents will appear
to be gibberish. The contents can only be interpreted by the load command.

15
2. Arrays: Vector and Matrix Calculations

Getting Help

help On-line help.


help lists all the primary help topics.
help <command> displays information about the command.
doc On-line help reference manual in HTML format.
doc accesses the manual.
doc <command> displays information about the command.
helpdesk Accesses the main page of the on-line reference manual.
type <command> Displays the actual MATLAB code for this command.
lookfor <keyword> Searches all MATLAB commands for this keyword.
who Lists all the current variables.
whos Lists all the current variables in more detail than who.
demo Runs demonstrations of many of the capabilities of MATLAB.
save Saves all of your variables.
load Loads back all of the variables which have been saved previously.
C Abort the command which is currently executing (i.e., hold down the control
ˆ key and type “c”).

1.9. Be Able To Do
After reading this section you should be able to do the following exercises. The MATLAB statements
are given on page 137.
1. Consider a triangle with sides a, b, and c and corresponding angles ∠ab, ∠ac, and ∠bc.
(a) Use the law of cosines, i.e.,
c2 = a2 + b2 − 2ab cos ∠ab ,
to calculate c if a = 3.7, b = 5.7, and ∠ab = 79◦ .
(b) Then show c to its full accuracy.
(c) Use the law of sines, i.e.,
sin ∠ab sin ∠ac
= ,
c b
to calculate ∠ac in degrees and show it in scientific notation.
(d) What MATLAB command should you have used first if you wanted to save these results to the file
triangle.ans?

2. Calculate 3 1.2×1020 − 1220 i.
3. Analytically, cos 2θ = 2 cos2 θ − 1. Check whether this is also true numerically when using MATLAB by
using a number of different values of θ. Use MATLAB statements which make it as easy as possible to do
this.
4. How would you find out information about the fix command?

2. Arrays: Vector and Matrix Calculations


In the previous section we discussed operations using single numbers, i.e., scalars. In this section we dis-
cuss operations on sets of numbers called arrays. Until the advanced subsection at the end, we restrict our
attention to one-dimensional arrays, which are called vectors, and two-dimensional arrays, which are called
matrices. In this section we will generally refer to these sets of numbers specifically as vectors or matrices
rather than use the more inclusive term “arrays”. MATLAB was originally developed specifically to work
with vectors and matrices and that is still one of its primary uses. For example, the multiplication of a
matrix and a vector has a very specific meaning whereas the multiplication of two arrays does not.
Notation: We will always write matri es using apital letters and ve tors using lower ase let-
ters.
16
2.1. Generating Matrices

This makes it much easier to understand MATLAB operations. This is also a good practice
for you to use.
In addition, when we write “vector” we mean a column vector and so it is immediately obvious
that A*x is a legitimate operation of a matrix times a vector as long as the number of rows of
the matrix A equals the number of columns of the column vector x. Also, x*A is illegitimate
because the column vector x has only one column while the matrix A is expected to have more
than one row. On the other hand, x′ A is legitimate ( x′ denotes the conjugate transpose of the
vector x ) as long as the row vector x′ has the same number of columns as the number of rows
of the matrix A.
In addition, we have very specific notation for denoting vectors and matrices and the elements of each.
We collect all this notation here.
Notation: Rm denotes all real column vectors with m elements and Cm denotes all complex column vec-
tors with m elements.
Rm×n denotes all real m×n matrices (i.e., having m rows and n columns) and Cm×n de-
notes all complex m×n matrices.
Notation: In this overview the word “vector” means a column vector so that Cm = Cm×1 . Vectors are de-
noted by boldface letters, such as x; we will write a row vector as, for example, xT , where T
denotes the transpose of a matrix or vector (that is, the rows and columns are reversed.)
Notation: A = (aij ) means that the (i, j)-th element of A (i.e., the element in the i-th row and the j-th
column) is aij .
x = (xi ) means that the i-th element of x is xi .
By the way MATLAB works with complex matrices as well as it does real matrices. To remind you of
this fact, we will use C rather than R unless there is a specific reason not to. If there is a distinction be-
tween the real and complex case, we will first describe the real case and then follow with the complex case
in parentheses.

2.1. Generating Matrices


To generate the matrix  
1 2 3
A = 4 5 6
7 8 9
in MATLAB type
>> A = [1 2 3; 4 5 6; 7 8 9]
(where “ ” denotes one or more spaces) or
>> A = [ 1 2 3 ; 4 5 6 ; 7 8 9]
or
>> A = [1,2,3; 4,5,6; 7,8,9]
or
>> A = [ 1 , 2 , 3 ; 4 , 5 , 6 ; 7 , 8 , 9 ]
In other words, either spaces or commas can be used to delineate the elements of each row of a matrix;
semicolons are required to separate rows. (Any number of spaces can be put around commas or semicolons
to improve the readability of the expression.)
Notation: Since we prefer spaces, we will generally use them rather than commas to separate elements in a
row.
Rows can also be separated by beginning each on a separate line. For example, the matrix A can also
be entered by
>> A = [1,2,3
4,5,6
7,8,9]
However, we consider this to be more work than simply using semicolons and will not use it again. The
more complicated matrix  √ 
1 2+ 3 3 sin 1
C =  e2 17/3
√ π+3 
1/3 2 − 3 −7 cos π/7

17
2.1. Generating Matrices

can be entered by typing


>> C = [ 1 2+sqrt(3) 3*sin(1); exp(2) 17/3 pi+3; 1/3 2-sqrt(3) -7*cos(pi/7) ]
or
>> C = [ 1, 2+sqrt(3), 3*sin(1); exp(2), 17/3, pi+3; 1/3, 2-sqrt(3), -7*cos(pi/7) ]
Warning: When an element of a matrix consists of more than one term, it is important to enter all the
terms without spaces — unless everything is enclosed in parentheses. For example,
>> x1 = [1 pi+3]
is the same as
>> x2 = [1 pi+ 3]
and is the same as
>> x3 = [1 (pi +3)]
but is not the same as
>> x4 = [1 pi +3] % not the same as the previous three statements
(Try it!) In other words, MATLAB tries to understand what you mean, but it does not always
succeed.
Definition
The transpose of a matrix A ∈ Cm×n , denoted by AT , is obtained by reversing the rows and columns
of A. That is, if A = (aij ) then AT = (aji ). (For example, the (2, 4) element of AT , i.e., i = 2 and
j = 4, is a42 .)
A square matrix A is symmetric if AT = A.
The conjugate transpose of a matrix A ∈ Cm×n , denoted by AH , is obtained by reversing the rows
and columns of A and then taking the complex conjugates of all the elements. That is, if A = (aij )
then AH = (a∗ji ), where ∗ denotes the complex conjugate of a number.
A square matrix A is Hermitian if AH = A.
Note: In MATLAB AT is calculated by A.′ (i.e., a period followed by a single quote mark).
In MATLAB AH is calculated by A′ (i.e., just a single quote mark.)
A vector can be entered in the same way as a matrix. For example, the vector
 
1
2
 
3
x =   = (1, 2, 3, 4, 5, 6)T
4
 
5
6

can be entered as
>> x = [1; 2; 3; 4; 5; 6]
However, this requires many semicolons; instead, take the transpose of a row vector by entering
>> x = [1 2 3 4 5 6].′
where the MATLAB command for the transpose, i.e., “ T ”, is “ .′ ” (i.e., a period followed by a single
quote mark). There is one further simplification that is usually observed when entering a vector. The
MATLAB command for the conjugate transpose, i.e., “ H ”, of a matrix is “ ′ ” (i.e., just a single quote
mark), which requires one less character than the command for the transpose. Thus, x is usually entered
as
>> x = [1 2 3 4 5 6]′
Aside: In fact, x should be entered as
>> x = [1:6]′
since this requires much less typing. (We will discuss the colon operator shortly.)
Warning: xT → x.′ while xH → x′ so that you an only al ulate xT by x′ if x is real. This has
bitten us occasionally!
Sometimes the elements of a matrix are complicated enough that you will want to simplify the process
p √ √ √ p T
of generating the matrix. For example, the vector r = 2/3 , 2 , 3 , 6 , 2/3 can be entered by
typing

18
2.1. Generating Matrices

>> s2 = sqrt(2); s3 = sqrt(3); r = [ s2/s3 s2 s3 s2*s3 s2/s3 ]′


We have now discussed how to enter matrices into MATLAB by using square parentheses, i.e., [...].
You work with individual elements of a matrix by using round parentheses, i.e., (...). For example, the
element aij of the matrix A is A(i,j) in MATLAB. Suppose you want to create the matrix
 
1 2 3
B = 4 5 6 
7 8 10

without having to enter all nine elements. If A (see the beginning of this section) has already been gener-
ated, the simplest way is to type
>> B = A; B(3,3) = 10
Also, the element xi of the vector x is x(i) in MATLAB. For example, to create the column vector

x = (1, 2, 3, . . . , 47, 48, 49, 51)T ∈ R50

enter
>> x = [1:50]′; x(50) = 51
or
>> x = [1:50]′; x(50) = x(50) + 1
or
>> x = [1:50]′; x(length(x)) = x(length(x)) + 1
where length returns the number of elements in a vector.
MATLAB also has a number of commands that can generate matrices. For example,
>> C = zeros(5)
or
>> C = zeros(5, 5)
generates a 5×5 zero matrix. Also,
>> C = zeros(5, 8)
generates a 5×8 zero matrix. Finally, you can generate a zero matrix C with the same size as an already
existing matrix, such as A, by
>> C = zeros(size(A))
where size(A) is a row vector consisting of the number of rows and columns of A. This command is fre-
quently used to preallocate a matrix of a given size so that MATLAB does not need to “guess” how large
to make it.
Similarly, you can generate a matrix with all ones by ones(n) or ones(m, n) or ones(size(D)).
You can also generate the identity matrix, i.e., the matrix with ones on the main diagonal and zeroes off of
it, by using the command eye with the same arguments as above.
Another useful matrix is a random matrix, that is, a matrix whose elements are all random num-
bers. This is generated by the rand command, which takes the same arguments as above. Specifically,
the elements are uniformly distributed random numbers in the interval (0, 1). To be precise, these are
pseudorandom numbers because they are calculated by a deterministic formula which begins with an ini-
tial “seed” — which is called the state , not the seed. Every time that a new MATLAB session is started,
the default seed is set, and so the same sequence of random numbers will be generated. However, every
time that this command is executed during a session, a different sequence of random numbers is generated.
If desired, a different seed can be set at any time by entering
>> rand(′ state′, <non-negative integer state number>)
Note: You can use the word “seed”, rather than “state”, as the first argument to rand but this causes
MATLAB to use an older random number generator which has a period of 231 − 2. The period of
the current default random number generator is 21492 .
Random matrices are often useful in just “playing around” or “trying out” some idea or checking out
some algorithm. The command randn generates a random matrix where the elements are normally dis-
tributed (i.e., Gaussian distributed) random numbers with mean 0 and standard deviation 1.
There is another “random” function which is useful if you want to reorder a sequence. The command
randperm(n) generates a random permutation of the integers 1, 2, . . . , n.

19
2.1. Generating Matrices

Note: randperm changes the state of rand.


MATLAB also makes it convenient to assemble matrices in “pieces”, that is, to put matrices together to
make a larger matrix. That is, the original matrices are submatrices of the final matrix. For specificity, let
us continue with A (see the beginning of this section). Suppose you want a 5×3 matrix whose first three
rows are the rows of A and whose last two rows are all ones. This is easily generated by
>> [ A ; ones(2, 3) ]
(The semicolon indicates that a row has been completed and so the next rows consist of all ones. The fact
that A is a matrix in its own right is immaterial. All that is necessary is that the number of columns of A
be the same as the number of columns of ones(2, 3).) This matrix could also be generated by
>> [ A ; ones(1, 3) ; ones(1, 3) ]
or by
>> [ A ; [1 1 1] ; [1 1 1] ]
or even by
>> [ A ; [1 1 1 ; 1 1 1] ]
Similarly, to generate a 3×4 matrix whose first three columns are the columns of A and whose last
column is (1, 5, 9)T type
>> [A [1 5 9]′ ]
(The space following the A indicates that the next column is to follow. The fact that the next entry is
a column vector is immaterial. All that is necessary is that the number of rows of A be the same as the
number of rows in the new last column.)

Elementary Matrices

zeros(n) Generates an n×n matrix with all elements being 0.


zeros(m, n) Generates an m×n matrix.
zeros(size(A)) Generates a zero matrix with the same size as A.
ones Generates a matrix with all elements being 1.
The arguments are the same as for zeros.
eye Generates the identity matrix, i.e., the diagonal elements are 1 and the off-
diagonal elements are 0.
The arguments are the same as for zeros.
rand Generates a matrix whose elements are uniformly distributed random numbers
in the interval (0, 1). Each time that this function is called during a session it
returns different random numbers.
The arguments are the same as for zeros.
The initial seed is changed by rand(′seed′ , <seed number>).
randn Generates a matrix whose elements are normally (i.e., Gaussian) distributed
random numbers with mean 0 and standard deviation 1. Each time that this
function is called during a session it returns different random numbers.
The arguments are the same as for zeros.
randperm(n) Generates a random permutation of the integers 1, 2, . . . , n.

size(A) The size of a matrix as the row vector (m, n).


Also, size(A,1) returns the number of rows (the first element of A ) and size(A,2)
returns the number of columns (the second element of A ).
length(x) The number of elements in a vector.

A.′ Transpose, i.e., AT . A′ Conjugate transpose, i.e., AH .

20
2.3. Manipulating Matrices

2.2. The Colon Operator


For real numbers a and b the MATLAB command
>> [a:b]
or, more simply,
>> a:b
generates the row vector (a, a + 1, a + 2, . . . , a + k) where the integer k satisfies a + k ≤ b and
a + (k + 1) > b. Thus, the vector x = (1, 2, 3, 4, 5, 6)T should be entered into MATLAB as
>> x = [1:6]′
or even as
>> x = [1:6.9]′
(although we can’t imagine why you would want to do it this way). If c is also a real number the MAT-
LAB command
>> [a:c:b]
or
>> a:c:b
generates a row vector where the difference between successive elements is c. Thus, we can generate num-
bers in any arithmetic progression using the colon operator. For example, typing
>> [18:-3:2]
generates the row vector (18, 15, 12, 9, 6, 3). while typing
>> [ pi : -.2*pi : 0 ]
generates the row vector (π, .8π, .6π, .4π, .2π, 0).
Warning: There is a slight danger if c is not an integer. As an oversimplified example, entering
>> x = [.02 : .001 : .98]′
should generate the column vector (0.02, 0.021, 0.022, . . . , 0.979, 0.98)T . However, because of
round-off errors in storing floating-point numbers, there is a possibility that the last element in
x will be 0.979. The MATLAB package was written specifically to minimize such a possibility,
but it still remains.† We will discuss the command linspace which avoids this difficulty in sec-
tion 4. An easy “fix” to avoid this possibility is to calculate x by
>> x = [20:980]′/1000

2.3. Manipulating Matrices


For specificity in this subsection we will mainly work with the 5×6 matrix
 
1 2 3 4 5 6
 7 8 9 10 11 12 
 
E =  13 14 15 16 17 18  ,
 
19 20 21 22 23 24
25 26 27 28 29 30

which can be generated by


>> E = [ 1:6 ; 7:12 ; 13:18 ; 19:24 ; 25:30 ]
Note: Spaces will frequently be used in MATLAB commands in this subsection for readability.
You can use the colon notation to extract submatrices from E. For example,
>> F = E( [1 3 5] , [2 3 4 5] )
extracts the elements in the first, third, and fifth rows and the second, third, fourth, and fifth columns
of E; thus,  
2 3 4 5
F =  14 15 16 17  .
26 27 28 29

This possiblity is much more real in the programming language C. For example, the statement
for ( i = 0.02; i <= 0.98; i = i + .001 )
generates successive values of i by adding 0.001 to the preceding value. It is possible that when i should have
the value 0.98, due to round-off errors the value will be slightly larger; the condition i <= 0.98 will be false
and the loop will not be evaluated when i should be 0.98.

21
2.3. Manipulating Matrices

You can generate this submatrix more easily by typing


>> F = E( 1:2:5 , 2:5 )
There is an additional shortcut you can use: in a matrix a colon by itself represents an entire row or
column. For example, the second column of F is F(:,2) and the second row is F(2,:). To replace the
second column of F by two times the present second column minus four times the fourth column enter
>> F(:,2) = 2*F(:,2) - 4*F(:,4)
And suppose you now want to double all the elements in the last two columns of F. Simply type
>> F(:,3:4) = 2*F(:,3:4)
There is a last-additional shortcut you can use. Suppose you want the matrix F to consist of the odd
rows of E and the second to the last column, as we did above. You might be changing the size of E and
not want to have to remember how large it is. This can be easily done by
>> F = E( [1:2:end] , [2:end] )
The keyword end designates the last element of the dimension: 5 for the rows of E and 6 for the
columns.
Note: The keyword end has a number of meanings. It also ends a block of code begun with a if, for,
while, or switch (as we will see later). Finally, it can also terminate a primary or secondary func-
tion, and it must terminate a nested function (as we will also see later).
Returning to “ :”, entering E(:,:) prints out exactly the same matrix as entering E (as does
E(1:end,1:end) ). This is not a very useful way of entering E, but it shows how the colon operator
can work. On the other hand, entering
>> G = E( : , 6:-1:1 )
generates a matrix with the same size as E but with the columns reversed, i.e.,
 
6 5 4 3 2 1
 12 11 10 9 8 7 
 
G =  18 17 16 15 14 13  .
 
24 23 22 21 20 19
30 29 28 27 26 25

Finally, there is one more use of a colon. Entering


>> f = E(:)
generates a column vector consisting of the columns of E (i.e., the first five elements of f are the first
column of E, the next five elements of f are the second column of E, etc.).
Note: On the right side of an equation, E(:) is a column vector with the elements being the columns of
E in order. On the left side of an equation, E(:) reshapes a matrix. However, we will not discuss
this reshaping further because the reshape command described below is easier to understand.
The colon operator works on rows and/or columns of a matrix. A different command is needed to work
on the diagonals of a matrix. For example, you extract the main diagonal of E by typing
>> d = diag(E)
(so d is the column vector (1, 8, 15, 22, 29)T ), one above the main diagonal by typing
>> d1 = diag(E, 1)
(so d1 is the column vector (2, 9, 16, 23, 30)T ), and two below the main diagonal by typing
>> d2 = diag(E, -2)
(so d2 is the column vector (13, 20, 27)T ).
The MATLAB command diag transforms a matrix (i.e., a non-vector) into a column vector. The con-
verse also holds: when diag is applied to a vector, it generates a symmetric matrix. The command
>> F = diag(d)
generates a 5×5 matrix whose main diagonal elements are the elements of d, i.e., 1, 8, 15, 22, 29, and
whose off-diagonal elements are zero. Similarly, entering
>> F1 = diag(d1, 1)
generates a 6×6 matrix whose first diagonal elements (i.e., one above the main diagonal) are the elements

22
2.3. Manipulating Matrices

of d1, i.e., 2, 9, 16, 23, 30, and whose other elements are zero, that is,
 
0 2 0 0 0 0
0 0 9 0 0 0 
 
0 0 0 16 0 0 
F1 =  .
0 0 0 0 23 0 
 
0 0 0 0 0 30
0 0 0 0 0 0

Finally, typing
>> F2 = diag(d2, -2)
generates a 5×5 matrix whose −2-nd diagonal elements (i.e., two below the main diagonal) are the ele-
ments of d2, i.e., 13, 20, 27, and whose other elements are zero, i.e.,
 
0 0 0 0 0
 0 0 0 0 0
 
F2 =  13 0 0 0 0 .
 
0 20 0 0 0
0 0 27 0 0

You can also extract the upper triangular or the lower triangular part of a matrix. For example,
>> G1 = triu(E)
constructs a matrix which is the same size as E and which contains the same elements as E on and above
the main diagonal; the other elements of G1 are zero. This command can also be applied to any of the
diagonals of a matrix. For example,
>> G2 = triu(E, 1)
constructs a matrix which is the same size as E and which contains the same elements as E on and above
the first diagonal, i.e.,
 
0 2 3 4 5 6
 0 0 9 10 11 12 
 
G2 =  0 0 0 16 17 18  .
 
0 0 0 0 23 24
0 0 0 0 0 30
The similar command tril extracts the lower triangular part of a matrix.
As an example of the relationship between these three commands, consider the square random matrix F
generated by
>> F = rand(6)
All the following MATLAB commands calculate F anew:
>> triu(F) + tril(F) - diag(diag(F))
>> triu(F, 1) + diag(diag(F)) + tril(F, -1)
>> triu(F) + tril(F, -1)
>> triu(F, 2) + diag(diag(F, 1), 1) + tril(F)
Note: Numerically the first command might not generate exactly the same matrix as the following three
because of round-off errors.
By the way, diag, triu and tril cannot appear on the left-hand side of an equation. Instead, to
zero out all the diagonals above the main diagonal of F enter
>> F = F - triu(F, 1)
and to zero out just the first diagonal above the main diagonal enter
>> F = F - tril(triu(F, 1), 1)
What if you want to insert numbers from the upper right-hand corner of a matrix to the lower left-hand
corner? There is no explicit function which does this but there are a number of indirect functions:
fliplr(A) flips the matrix from left to right, i.e., reverses the columns of the matrix;
flipud(A) flips the matrix up and down, i.e., reverses the rows of the matrix;
rot90(A) rot90 rotates the matrix 90◦ ; and

23
2.3. Manipulating Matrices

rot90(A,k) rotates the matrix k×90◦ .


MATLAB has a command which is useful in changing the shape of a matrix while keeping the same
numerical values. The statement
>> K = reshape(H, m, n)
reshapes the matrix H ∈ Cp×q into K ∈ Cm×n where m and n must satisfy mn = pq (or an error mes-
sage will be generated). A column vector is generated from H, as in H(:), and the elements of K are
taken columnwise from this vector. That is, the first m elements of this column vector go in the first col-
umn of K, the second m elements go in the second column, etc. For example, the matrix E which has
been used throughout this subsection can be easily (and quickly) generated by
>> E = reshape([1:30], 6, 5)′
Occasionally, there is a need to delete elements of a vector or rows or columns of a matrix. This is easily
done by using the null matrix []. For example, entering
>> x = [1 2 3 4]′
>> x(2) = []
results in x = (1, 3, 4)T . As another example, you can delete the even columns of G by
>> G( : , 2:2:6 ) = []
The result is  
6 4 2
 12 10 8 
 
G =  18 16 14  .
 
24 22 20
30 28 26

Also, occasionally, there is a need to replicate or tile a matrix. That is, the command
>> B = repmat(A, m, n)
generates a matrix B which contains m rows and n columns of copies of A. (If n = m then repmat(A,
m) is sufficient.) If A is a p by q matrix, then B ∈ Rmp×nq . This even works if A is a scalar, in which
case this is the same as
>> B = A*ones(m, n)
(but it is much faster if m and n are large since no multiplication is involved).

24
2.4. Simple Arithmetical Operations

Manipulating Matrices

A(i,j) ai,j .
A(:,j) the j-th column of A.
A(i,:) the i-th row of A.
A(:,:) A itself.
A(?1,?2) There are many more choices than we care to describe:
?1 can be i or i1:i2 or i1:i3:i2 or : or [i1 i2 ... ir] and
?2 can be j or j1:j2 or j1:j3:j2 or : or [j1 j2 ... jr].
A(:) On the right-hand side of an equation, this is a column vector containing the
columns of A one after the other.
o
diag(A) A column vector of the k-th diagonal of the matrix (i.e., non-vector) A. If k is
diag(A, k) not given, then k = 0.
o
diag(d) A square matrix with the k-th diagonal being the vector d. If k is not given,
diag(d, k) then k = 0.
o
triu(A) A matrix which is the same size as A and consists of the elements on and
triu(A, k) above the k-th diagonal of A. If k is not given, then k = 0.
o
tril(A) The same as the command triu except it uses the elements on and below the
tril(A, k) k-th diagonalof A. If k is not given, then k = 0.
fliplr(A) Flips a matrix left to right.
flipud(A) Flips a matrix up and down.
Rotates a matrix k×90◦ . If k is not given, then k = 1.
o
rot90(A)
rot90(A, k)
repmat(A, m, n) Generates a matrix with m rows and n columns of copies of A. (If n = m
the third argument is not needed.)
reshape(A, m, n) Generates an m×n matrix whose elements are taken columnwise from A.
Note: The number of elements in A must be mn.
[] The null matrix. This is also useful for deleting elements of a vector and rows
or columns of a matrix.

2.4. Simple Arithmetical Operations


Matrix Addition:
If A, B ∈ Cm×n then the MATLAB operation
>> A + B
means A + B = (aij ) + (bij ) = (aij + bij ). That is, the (i, j)-th element of A + B is aij + bij .
Matrix Subtraction:
If A, B ∈ Cm×n then the MATLAB operation
>> A - B
means A − B = (aij ) − (bij ) = (aij − bij ).
Matrix Multiplication by a scalar:
If A ∈ Cm×n then for any scalar c the MATLAB operation
>> c*A
means cA = c(aij ) = (caij ). For example, the matrix q = (0, .1π, .2π, .3π, .4π, .5π)T can be gener-
ated by
>> q = [ 0 : .1*pi : .5*pi ]′
but more easily by
>> q = [ 0 : .1 : .5 ]′ *pi
or
>> q = [0:5]′ *.1*pi
Matrix Multiplication:
If A ∈ Cm×ℓ and B ∈ Cℓ×n then the MATLAB operation
>> A*B

25
2.4. Simple Arithmetical Operations
P 

means AB = (aij )(bij ) = k=1 aik bkj . That is, the (i, j)-th element of AB is ai1 b1j + ai2 b2j +
. . . + aiℓ bℓj .
Matrix Exponentiation:
If A ∈ Cn×n and p is a positive integer, then the MATLAB operation
>> A p
ˆ
means Ap = AA . . . A.
| {z }
p times
Matrix Exponentiation is also defined when p is not an integer. For example,
>> A = [1 2; 3 4]; B = A (1/2)
ˆ
calculates a complex matrix B whose square is A. (Analytically, B2 = A, but numerically
>> B 2 - A
ˆ
returns a non-zero matrix — however, all of its elements are less than 10 · eps in magnitude.)
Note: For two values of p there are equivalent MATLAB commands:
A1/2 can also be calculated by sqrtm(A) and
A−1 can also be calculated by inv(A).
Matrix Division:
The expression
A
B
makes no sense in linear algebra: if B is a square non-singular matrix it might mean B−1 A or it
might mean AB−1 . Instead, use the operation
>> A\b
to calculate the solution of the linear system Ax = b (where A must be a square non-singular ma-
trix) by Gaussian elimination. This is much faster computationally than calculating the solution of
x = A−1 b by
>> x = inv(A)*b
Similarly,
A\B
solves AX = B by repeatedly solving Ax = b where b is each column of B in turn and x is the corre-
sponding column of X.
It is also possible to solve xT A = bT for xT by
b.′ /A
and to solve XA = B by
B/A
(This is the same as solving AT XT = BT .)
Elementwise Multiplication:
If A, B ∈ Cm×n , then the MATLAB operation
>> A.*B
means (aij bij ). That is, the (i, j)-th element of A.*B is aij bij . Note that this is not a matrix
operation, but it is sometimes a useful operation. For example, suppose y ∈ Rn has been defined pre-
viously and you want to generate the vector z = (1y1 , 2y2 , 3y3 , . . . , nyn )T . You merely type
>> z = [1:n]′ .* y
(where the spaces are for readability). Recall that if y ∈ Cn you will have to enter
>> z = [1:n].′ .* y
because you do not want to take the complex conjugate of the complex elements of y.
Elementwise Division:
If A, B ∈ Cm×n , then the MATLAB operation
>> A./B
means (aij /bij ).
Elementwise Left Division:
If A, B ∈ Cm×n , then the MATLAB operation
>> B.\A

26
2.4. Simple Arithmetical Operations

means the same as A./B


Elementwise Exponentiation:
If A ∈ Cm×n , then
>> A. p
ˆ
means (apij ) and
>> p. A
ˆ
means (paij ). Also, if A, B ∈ Cm×n , then
A. B
ˆ b 
means aijij .
Where needed in these arithmetic operations, MATLAB checks that the matrices have the correct size.
For example,
>> A + B
will return an error message if A and B have different sizes, and
>> A*B
will return an error message if the number of columns of A is not the same as the number of rows of B.
Note: There is one exception to this rule. When a scalar is added to a matrix, as in A + c, the scalar is
promoted to the matrix cJ where J has the same size as A and all its elements are 1. That is,
>> A + c
is evaluated as
>> A + c*ones(size(A))
This is not a legitimate expression in linear algebra, but it is a very useful expression in MATLAB.
For example, you can represent the function

y = 2 sin(3x + 4) − 5 for x ∈ [2, 3]

by 101 data points using


>> x = [2:.01:3]′;
>> y = 2*sin(3*x + 4) - 5
This is much more intelligible than calculating y using
>> y = 2*sin(3*x + 4*ones(101, 1)) - 5*ones(101, 1)
In some courses that use vectors, such as statics courses, the dot product of the real vectors ~a and ~b is
defined by
n
X
~a q~b = ai b i .
i=1

In linear algebra this is called the inner product and is defined for vectors a, b ∈ Rn by aT b. It is calcu-
lated by
>> a′ *b
(If a, b ∈ Cn the inner product is aH b and is calculated by a′ *b.) The outer product of these two vectors
is defined to be abT and is calculated by
>> a*b′
(If a, b are complex the outer product is abH and is calculated by a*b′ .) It is important to keep these
two products separate: the inner product is a scalar, i.e., aT b ∈ R (if complex, aH b ∈ C ), while the outer
product is an n×n matrix, i.e., abT ∈ Rn×n (if complex, abH ∈ Cn×n ).
In linear algebra we often work with “large” matrices and are interested in the amount of “work” re-
quired to perform some operation. Previously, MATLAB kept track of the number of flops, i.e., the num-
ber of fl oating-pointoperations, performed during the MATLAB session. Unfortunately, this disappeared
in version 6. Instead, we can calculate the amount of CPU time† required to execute a command by us-
ing cputime. This command returns the CPU time in seconds that have been used since you began your


The CPU, Central Processing Unit, is the “guts” of the computer, that is, the hardware that executes the
instructions and operates on the data.

27
2.4. Simple Arithmetical Operations

MATLAB session. This time is frequently difficult to calculate, and is seldom more accurate than to
1/
100 -th of a second. Here is a simple example to determine the CPU time required to invert a matrix.
>> n = input(′n = ′ ); time = cputime; inv(rand(n)); cputime - time
Warning: Remember that you have to subtract the CPU time used before the operation from the CPU
time used after the operation.
By the way, you can also calculate the wall clock time required for some sequence of commands by using
tic and toc . For example,
>> tic; <sequence of commands>; toc
returns the time in seconds for this sequence of commands to be performed.
Note: This is very different from using cputime. tic followed by toc is exactly the same as if you had
used a stopwatch to determine the time. Since a timesharing computer can be running many dif-
ferent processes at the same time, the elapsed time might be much greater than the CPU time.
Normally, the time you are interested in is the CPU time.

Arithmetical Matrix Operations

A+B Matrix addition. A.*B Elementwise multiplication.


A-B Matrix subtraction. A. p Elementwise exponentiation.
A*B Matrix multiplication.
ˆ
p. A
A n Matrix exponentiation.
ˆ
A. B
ˆ
A\b The solution to Ax = b by Gaussian
ˆ
A./B Elementwise division.
elimination when A is a square non- B.\A Elementwise left division, i.e., B.\A is
singular matrix. exactly the same as A./B.
A\B The solution to AX = B by Gaussian
elimination.
b/A The solution to xA = b where x and
b are row ve tors.
B/A The solution to XA = B.

cputime Approximately the amount of CPU time (in seconds) used during this session.
tic, toc Returns the elapsed time between these two commands.

28
2.6. Be Careful!

2.5. Operator Precedence


It is important to list the precedence for MATLAB operators. That is, if an expression uses two or more
MATLAB operators, in which order does MATLAB do the calculations? For example, what is 1:n+1? Is
it (1:n)+1 or is it 1:(n+1)? And if we solve ACx = b by A*C\b, does MATLAB do (A*C)\b or A*(C\b)?
The former is C−1 A−1 b while the latter is A−1 C−1 b — and these are different unless A and C commute.
The following table shows the precedence of all MATLAB operators, that is, the order in which it evalu-
ates an expression. The precedence is from highest to lowest. Within each precedence level, operators are
evaluated from left to right in the expression.

Operator Precedence (highest to lowest)


(operators with same precedence are separated by spaces)
1 ( )
′ ′
2 . .
3
ˆ
+ [unary plus]
ˆ
- [unary minus]
4 .* ./ .\ * /
˜\
5 + [addition] - [subtraction]
6 :
7 < <= > >= ==
8 &
˜=
9 |
10 &&
11 ||

The unary plus and minus are the plus and minus signs in x = +1 and x = −1. The plus and mi-
nus signs for addition and subtraction are, for example, x = 5 + 1 and x = 10 − 13. Thus, 1:n+1 is
1:(n+1) because “ +” has higher precedence than “ :”.† Also, A*C\b = (A*C)\b because “ *” and “ \”
have the same precedence and so the operations are evaluated from left to right.

2.6. Be Careful!
Be very areful: o asionally you might misinterpret how MATLAB displays the elements
of a ve tor or matrix. For example, the MATLAB command eig calculates the eigenvalues of a square
matrix. (We discuss eigenvalues in section 7.) To calculate the eigenvalues of the Hilbert matrix of or-
der 5, i.e.,  
1 1/ 1/ 1/ 1/
2 3 4 5
 1/2 1/
3
1/
4
1/
5
1/
6

 
 1/3 1/
4
1/
5
1/
6
1/
7
,
 
 1/ 1/ 1/ 1/ 1/ 
4 5 6 7 8
1/ 1/ 1/ 1/ 1/
5 6 7 8 9

(we discuss this matrix in detail in subsection 5.2) enter


>> format short
>> eig(hilb(5))
MATLAB displays the eigenvalues as the column vector


On the other hand, in the statistical computer languages R and S (which are somewhat similar to MATLAB),
“ :” has higher precedence than “ +” and so 1:n+1 is (1:n)+1 = 2:(n+1).

29
2.6. Be Careful!

ans =

0.0000
0.0003
0.0114
0.2085
1.5671
You might think the the first element of this vector is 0. However, if it was zero MATLAB would display
0 and not 0.0000. Entering
>> format short e
>> ans
displays
ans =

3.2879e-06
3.0590e-04
1.1407e-02
2.0853e-01
1.5671e+00
which makes it clear that the smallest eigenvalue is far from zero.
On the other hand, if you enter
>> format short
>> A = [1 2 3; 4 5 6; 7 8 9]
>> eig(A)
MATLAB displays
ans =

16.1168
-1.1168
-0.0000
It might appear from our previous discussion that the last eigenvalue is not zero, but is simply too small
to appear in this format. However, entering
>> format short e
>> ans
displays
ans =

1.6117e+01
-1.1168e+00
-8.0463e-16
Since the last eigenvalue is close to eps, but all the numbers in the matrix A are of “reasonable size”,
you can safely assume that this eigenvalue is zero analytically. It only appears to be nonzero when calcu-
lated by MATLAB because omputers annot add, subtra t, multiply, or divide orre tly!
As another example of how you might misinterpret the display of a matrix, consider the Hilbert matrix
of order two  
1 1/2
H= 1 .
/2 1/3
We write H100 as  
100 10 1.5437 0.8262
H ≈ 10 ,
0.8262 0.4421
while in MATLAB entering

30
2.7. Common Mathematical Functions

>> format short


>> H = hilb(2)
>> H 100
displays
ˆ
ans =

1.0e+10 *

1.5437 0.8262
0.8262 0.4421
It is very easy to miss the term “ 1.0e+10 *” because it stands apart from the elements of the matrix.
Similarly, entering
>> format short
>> H = hilb(2)
>> ( H (1/2) ) 2 - H
ˆ ˆ 2
should result in the zero matrix, since (H1/2 ) = H. However, MATLAB displays
ans =

1.0e-15 *

0.2220 0
0 0
where, again, it is easy to miss the term “ 1.e-15 *” and not realize that this matrix is very small — in
fact, it should be zero.
Be areful: MATLAB has nite memory. You should have no problem creating a matrix by
>> A = zeros(1000)
but you might well have a problem if you enter
>> A = zeros(10000)
The amount of memory available is dependent on the computer and the operating system and is very hard
to determine. Frequently it is much larger than the amount of physical memory on your computer. But,
even if you have sufficient memory, MATLAB may slow to a crawl and become unusable. The whos com-
mand will tell you how much memory you are using and show you the size of all your variables. If you
have large matrices which are no longer needed, you can reduce their sizes by equating them to the null
matrix, i.e., [], or remove them entirely by using clear.
Warning: Recall that the clear command is very dangerous because clear A deletes the variable A
but clear (without anything following) deletes all variables!

2.7. Common Mathematical Functions


In linear algebra mathematical functions cannot usually be applied to matrices. For example, eA and
sin A have no meaning unless A is a square matrix. (We will discuss their mathematical definitions in
section 15.)
Here we are interested in how MATLAB applies common mathematical functions to matrices and vec-
tors. For example, you might want to take the sine of every element of the matrix A = (aij ) ∈ Cm×n , i.e.,
B = (sin aij ). This is easily done in MATLAB by
>> B = sin(A)
Similarly, if you want C = (eaij ), enter
>> C = exp(A)
p 
Also, if you want D = aij type
>> C = sqrt(A)
or
>> C = A. (1/2)
ˆ

31
2.8. Data Manipulation Commands

All the common mathematical functions in the table entitled “Some Common Real Mathematical Func-
tions” in subsection 1.5 can be used in this way.
As we will see in the section on graphics, this new interpretation of mathematical functions makes it
easy in MATLAB to graph functions without having to use the MATLAB programming language.

2.8. Data Manipulation Commands


MATLAB has a number of “simple” commands which are used quite frequently. Since many of them
are quite useful in analyzing data, we have grouped them around this common “theme”.
To calculate the maximum value of the vector x, type
>> m = max(x)
If you also want to know the element of the vector which contains this maximum value, type
>> [m, i] = max(x)
If the elements of the vector are all real, the result of this command is the element which has the maxi-
mum value. However, if any of the elements of x are complex (i.e., non-real), this command has no math-
ematical meaning. MATLAB defines this command to determine the element of the vector which has the
maximum absolute value of the elements of x.
Warning: Make sure you understand the description of max if you every apply it to non-real vectors. For
example, if x = (−2, 1)T then max(x) returns 1 as expected. However, if x = (−2, i)T then
max(x) returns −2. This is because the element which has the largest absolute value is −2.
Thus, if x is a non-real vector, then max(x) is not the same as max(abs(x)).
Since the columns of the matrix A can be considered to be vectors in their own right, this command
can also be applied to matrices. Thus,
>> max(A)
returns a row vector of the maximum element in each of the columns of A if all the elements of A are
real. If any of the elements of A are non-real, this command returns the element in each column which
has the maximum absolute value of all the elements in that column.
To find the maximum value of an entire real matrix, type
>> max(max(A))
or
>> max(A(:))
and to find the maximum absolute value of an entire real or complex matrix, type
>> max(max(abs(A)))
or
>> max(abs(A(:)))
The command min acts similarly to max except that it finds the minimum value (or element with the
minimum absolute value) of the elements of a vector or the columns of a matrix.
To calculate the sum of the elements of the vector x, type
>> sum(x)
sum behaves similarly to max when applied to a matrix. That is, it returns the row vector of the sums
of each column of the matrix. This command is sometimes useful in adding a deterministic series. For
example,
>> 1 + 1/2 + 1/3 + 1/4 + 1/5 + 1/6 + 1/7 + 1/8 + 1/9 + 1/10 + 1/11 + 1/12 + ...
1/13 + 1/14 + 1/15 + 1/16 + 1/17 + 1/18 + 1/19 + 1/20
is entered much more easily as
>> sum(ones(1, 20)./[1:20])
or even as
>> sum(1./[1:20])
The mean, or average, of these elements is calculated by
>> mean(x)

32
2.8. Data Manipulation Commands

where mean(x) = sum(x)/length(x).


std calculates the standard deviation of the elements of a vector. The standard deviation is a measure
of how much a set of numbers “vary” and is defined as
v
u n
u1 X 2
std(x) = t xi − h x i
n i=1

where h x i is the mean of the elements.


MATLAB can also sort the elements of the vector x in increasing order by
>> sort(x)
If the vector is non-real, the elements are sorted in increasing absolute value. (If two elements have the
same absolute value, the one with the smaller absolute angle in polar coordinates is used.)
The MATLAB command diff calculates the difference between successive elements of a vector. For
example, if x ∈ Rn then the command
>> s = diff(x)
generates the vector s ∈ Rn−1 which is defined by si = xi+1 − xi . There are a number of uses for this
command. For example,
• if s has been sorted, then diff(s) == 0 can be used to test if any elements of s are repeated (or
the number that are repeated.
• similarly, all(diff(x)) > 0 tests if the elements of s are monotonically increasing.
• a numerical approximation to the derivative of y = f (x) can be calculated by diff(y)./diff(x).
The MATLAB function which is almost the inverse of diff is cumsum which calculates the cumu-
lative sum of the elements of a vector or matrix. For example, if s ∈ Rn−1 has been generated by s =
diff(x), then
>> c = cumsum(s)
Pi
generates the vector c ∈ Rn−1 where ci = j=1 sj . We can recover x by
>> xrecovered = zeros(length(x),1)
>> xrecovered(1) = x(1)
>> xrecovered(2:length(x)) = x(1) + c

There are also a number of MATLAB commands which are particularly designed to plot data. The
commands we have just discussed, such as the average and standard deviation, give a coarse measure of
the distribution of the data. To actually “see” what the data looks like, it has to be plotted. Two particu-
larly useful types of plots are histograms (which show the distribution of the data) and plots of data which
include error bars. These are both discussed in subsection 4.1.
pAlthough it does not quite fit here, sometimes you want to know the length of a vector x, which is
x21 + x22 + . . . x2n . (Note that this is not length(x) which returns the number of elements in x, i.e.,
n.) This length, which is often called the Euclidean length, can be calculated by entering
>> sqrt( x′ *x )
but it can be entered more easily by
>> norm(x)
(As we discuss in section 7, the norm of a vector is a more general concept than simply the Euclidean
length.)

33
2.9. Advanced Topic: Multidimensional Arrays

Data Manipulation Commands

max(x) The maximum element of a real vector.


[m, i] = max(x) also returns the element which contains the maximum value in i.

max(A) A row vector containing the maximum element in each column of a matrix.
[m, i] = max(A) also returns the element in each column which contains the maxi-
mum value in i.
o
min(x) The sum of the elements of a vector, or a row vector containing the sum of the ele-
min(A) ments in each column in a matrix.
o
mean(x) The mean, or average, of the elements of a vector, or a row vector containing the
mean(A) mean of the elements in each column in a matrix.
norm(x) The Euclidean length of a vector.
norm(A) The matrix norm of A.
Note: the norm of a matrix is not the Euclidean length of each column in the
matrix.
o
prod(x) The product of the elements of a vector, or a row vector containing the product of
prod(A) the elements in each column in a matrix.
o
sort(x) Sorts the elements in increasing order of a real vector, or in each column of a real
sort(A) matrix.
o
std(x) The standard deviation of the elements of a vector, or a row vector containing the
std(A) standard deviation of the elements in each column in a matrix.
o
sum(x) The sum of the elements of a vector, or a row vector containing the sums of the
sum(A) elements in each column in a matrix.
o
diff(x) The difference between successive elements of a vector, or between successive ele-
diff(A) ments in each column of a matrix.
o
cumsum(x) The cumulative sum between successive elements of a vector, or between successive
cumsum(A) elements in each column of a matrix.

2.9. Advanced Topic: Multidimensional Arrays


We have already discussed 1-D arrays (i.e., vectors) and 2-D arrays (i.e., matrices). Since these are two
of the most fundamental objects in linear algebra, there are many operations and functions which can be
applied to them. In MATLAB you can also use multidimensional arrays (i.e., n-D arrays).
A common use for multidimensional arrays is simply to hold data. For example, suppose a company
produces three products and we know the amount of each product produced each quarter; the data nat-
urally fits in a 2-D array, i.e., (product, amount). Now suppose the company has five sales regions so we
split the amount of each product into these regions; the data naturally fits in a 3-D array, i.e., (product,
region, amount). Finally, suppose that each product comes in four colors; the data naturally fits in a 4-D
array, i.e., (product, color, region, amount).
For another example, a 3-D array might be the time evolution of 2-D data. Suppose we record a grey
scale digital image of an experiment every minute for an hour. Each image is stored as a matrix M with
mi,j denoting the value of the pixel positioned at (xi , yj ). The 3-D array Mall can contain all these im-
ages: Mall(i,j,k) denotes the value of the pixel positioned at (xi , yj ) in the k-th image. The entire
k-th image is Mall(:,:,k) and it is filled with the k-th image M by
>> Mall(:,:,k) = M
If you want to multiply M by another matrix A, you can use M*A or Mall(:,:,k)*A; if you want to
average the first two images you can use .5*(Mall(:,:,1)+Mall(:,:,2)).
Many MATLAB functions can be used in n-D, such as ones, rand, sum, and size. The cat func-
tion is particularly useful in generating higher-dimensional arrays. For example, suppose we have four
matrices A, B, C, and D ∈ R2×7 which we want to put into a three-dimensional array. This is easily done
by
>> ABCD = cat(3, A, B, C, D)

34
2.10. Be Able To Do

which concatenates the four matrices using the third dimension of ABCD. (The “3” denotes the third di-
mension of ABCD.) And it is much easier than entering
>> ABCD(:,:,1) = A;
>> ABCD(:,:,2) = B;
>> ABCD(:,:,3) = C;
>> ABCD(:,:,4) = D;
If instead, we enter
>> ABCD = cat(j, A, B, C, D)
then the four matrices are concatenated along the j-th dimension of ABCD. That is,
cat(1, A, B, C, D) is the same as [A, B, C, D] and cat(2, A, B, C, D) is the same as
[A; B; C; D].
Another useful command is squeeze which squeezes out dimensions which only have one element. For
example, if we enter
>> E = ABCD(:,2,:)
(where the array ABCD was created above), then we might think that E is a matrix whose columns consist
of the second columns of A, B, C, and D. However, size(E) = 2 1 4 so that E is a three-dimensional
array, not a two-dimensional array. We obtain a two-dimensional array by squeeze(E).
The command permute reorders the dimensions of a matrix. For example,
>> ABCD = cat(3, A, B, C, D)
>> BCDA = permute( ABCD, [2 3 4 1] )
is the same as
>> BCDA = cat(3, B, C, D, A)
That is, the second argument of permute shows where the original ordering of the dimensions, i.e., 1, 2,
. . . , n, are to be placed in the new ordering. ipermute is the inverse of permute so, for example,
>> BCDA = cat(3, B, C, D, A)
>> ABCD = ipermute( BCDA, [2 3 4 1] )

Multidimensional Array Functions

cat Concatenates arrays; this is useful for putting arrays into a higher-dimensional array.
ipermute The inverse of permute.
permute Reorders the dimensions of a multidimensional array.
squeeze Removes (i.e., squeezes out) dimensions which only have one element.

2.10. Be Able To Do

After reading this section you should be able to do the following exercises. The answers are given on
page 137.
1. Consider the matrix  
1 2 3 4
 5 6 7 8 
A= .
9 10 11 12
13 14 15 16

(a) Enter it in the following three ways:


(i) type in all 16 elements directly.
(ii) since each row is in arithmetic progression, use the colon operator to enter each row.
(iii) since each column is in arithmetic progression, use the colon operator (and the transpose opera-
tor) to enter each column.
(b) Multiply the second row of A by −9/5 , add it to the third row, and put the result back in the second

35
3. Anonymous Functions, Strings, and Other Data Types

row. Do this all using one MATLAB statement.


2. Generate the tridiagonal matrix
 
4 −1
 −1 4

−1 0 

−1 4 −1
 ∈ Rn×n
 
A= . . . . . .
 . . . 
 
−1 4 −1
0 −1 4

where the value of n has already been entered into MATLAB.


3. Generate the tridiagonal matrix
 
1 −1
 e1

4 −1 0 

e2 9 −1
 ∈ Rn×n
 
A= .. .. ..
 . . . 
en−1 (n − 1)2
 
−1
0 en n2

where the value of n has already been entered into MATLAB.


4. Consider the matrix  
1 1 1 1 0 0 0 0 0 −5
1 1 1 1 0 0 0 0 0 0 
 
1 1 1 1 0 0 0 0 0 0 
A= .
1 1 1 1 0 0 0 0 0 0 
 
1 1 1 1 0 0 0 0 0 0
5 1 1 1 0 0 0 0 0 0

(a) Enter it using as few keystrokes as possible. (In other words, don’t enter the elements individually.)
(b) Zero out all the elements of A below the diagonal.
5. Enter the column vector
x = (0, 1, 4, 9, 16, 25, . . . , 841, 900)T
using as few keystrokes as possible. (In other words, don’t enter the elements individually.)
6. (a) Generate a random 5×5 matrix R.
(b) Determine the largest value in each row of R and the element in which this value occurs.
(c) Determine the average value of all the elements of R.
(d) Generate the matrix S where every element of S is the sine of the corresponding element of R.
(e) Put the diagonal elements of R into the vector r.
7. Generate the matrix  
1 2 3
A = 4 5 6  .
7 8 10

(a) Calculate a matrix B which is the square root of A. That is, B2 = A. Also, calculate a matrix C each
of whose elements is the square root of the corresponding element of A.
(b) Show that the matrices you have obtained in (a) are correct by substituting the results back into the
original formulas.

3. Anonymous Functions, Strings, and Other Data Types


Now that we have discussed scalar and matrix calculations, the next important topic is graphics. How-
ever, there are a number of minor topics which are useful in graphics and so we collect them here. First,

36
3.1. Anonymous Functions

anonymous functions allow us to easily define a function which we can then plot. Second, some graphics
functions require that the name of a function be passed as an argument. Third, character strings are nec-
essary in labelling plots. And, finally, cell arrays are occasionally helpful in labelling plots. Cell arrays are
generally used to manage data and since structures are also used to manage data we also include them
here. Another reason is that there are a number of data types in MATLAB including numeric, text vari-
ables, cell arrays, and structures. (A “minor” data type is logical, which we discuss in section 8.1.) We
might as well get the last three out of the way at once.

3.1. Anonymous Functions


In MATLAB it is common to define a mathematical function in a separate file as we discuss in sub-
section 8.3. (This is similar to writing a function or subroutine or subprogram in a high-level computer
language.) However, if the mathematical function is particularly simple, that is, it can be written as one
simple expression, we can define it in MATLAB using an anonymous function. If our function is

f (< arg1 >, < arg2 >, . . . ) = < expression >

the MATLAB statement is


>> f = @(<arg1>, <arg2>, ...) <expression>
For example, we can define the function

f (t) = t5 e−2t cos(3t)

by
>> f = @(t) t. 5 .* exp(-2*t) .* cos(3*t)
and then evaluate it by
ˆ
>> x = [0:.01:1]′
>> fx = f(x)
>> A = rand(5)
>> fA = f(A)
More generally, we can define
g(x, y, a, b, c) = xa e−bx cos(cy)
by
>> g = @(x, y, a, b, c) x. a .* exp(-b.*x) .* cos(c.*y)
ˆ
in which case any of the input arguments can be in R or in Rn . It is also possible — although probably
not very useful — to let g have one vector argument, say x = (x, y, a, b, c)T by
>> g = @(x) x(1) x(3) * exp(-x(4)*x(1)) * cos(x(5)*x(2))
ˆ
(In this example there is no purpose to use .* or . .)
ˆ
Warning: It is quite easy to forget to put dots (i.e., “ .”) before the mathematical operations of multipli-
cation (i.e., * ), division (i.e., / ), and exponentiation (i.e., ). For example, if f is defined by
>> f = @(t) t 5 * exp(-2*t) * cos(3*t)
ˆ
then
ˆ
>> f(3)
is allowed, but not
>> f([1:10])
Be careful!
The syntax for defining an anonymous function is
>> @(<argument list>) <expression>
(Since there is no left-hand side to this expression, the name of this function is ans.) The symbol @ is the
MATLAB operator that constructs a function handle. This is similar to a pointer in C which gives the ad-
dress of a variable or of a function. The name “handle” is used in MATLAB to denote a variable which
refers to some “object” which has been created. Thus, we can think of an anonymous function as being
created by

37
3.2. Passing Functions as Arguments to Commands

(<argument list>) <expression>


and the handle to the function (in C, the address of the function) being returned by using @. By the way,
we can create a function handle to a MATLAB function by, for example,
>> f = @cos
so that f(3) is the same as cos(3). We give an example where this is very useful in section 3.4
It is important to understand that all user-defined variables which appear in <expression> must ei-
ther appear in the argument list or be defined before the function is defined. If the variable does not
appear in the argument list, then its value is xed when the function is defined. For example, if a very sim-
ple function is defined by
>> r = 10
>> h = @(x) r*x
then the function is h(x) = 10x even if r is modified later. Thus,
>> h(5)
returns 50 and so does
>> r = 0
>> h(5)
Warning: Don’t forget that if a variable does not appear in the argument list, then its value is xed when
the function is defined.
A function can also be defined — but don't do it —by the inline command. For example, the func-
tion f defined above can also be defined by
>> f = inline(′t. 5 .* exp(-2*t) .* cos(3*t)′, ′ t′ )
ˆ
In general, if our function is

f (< arg1 >, < arg2 >, . . . ) = < expression >

the MATLAB statement is


>> f = inline(′<expression>′, ′ <arg1>′, ′ <arg2>′, ...)
Since it is quite easy to forget to put dots (i.e., . ) before the mathematical operations of multiplication
(i.e., * ), division (i.e., / ), and exponentiation (i.e., ), the MATLAB command vectorize does it for
you. To continue the first example,
ˆ
>> f = vectorize( inline(′t 5 * exp(-2*t) * cos(3*t)′, ′ t′ ) )
ˆ
is equivalent to the f defined above but does not require you to remember all the dots.
Warning: The inline command is obsolete and should not be used because it is very, very, very, very,
very inefficient. It is mentioned here only because you might find it in “old” codes.

3.2. Passing Functions as Arguments to Commands

It is sometimes necessary to pass the name of a function into a MATLAB function or a function m-file
created by the user. For example, as we discuss in section 4.1, we can plot the function y = f (x) in the
interval [−5, +5] by
fplot(<function "name">, [-5 +5])
But how do we pass this “name”?
We put the name in quotes because we do not pass the name of the function, but its handle. If f has
been defined by an anonymous function, then we enter
fplot(f, [-5 +5])
because f is a variable which we have already defined. If fnc is a MATLAB function or a user-defined
function m-file, then it is not known in the MATLAB workspace so
fplot(fnc, [-5 +5]) % WRONG
will not work. Instead, we use
fplot(@fnc, [-5 +5]) % CORRECT
Note: There are a number of “older” ways to pass function names. For example,
>> fplot(′fnc′ , [-5 +5])
will also work. We can even pass a “simple” function by, for example,
>> fplot(′(x*sin(x) + sqrt(1 + cos(x)))/(x 2 + 1)′ , [-5 +5])
ˆ
38
3.3. Strings

but don't do it! Instead, use anonymous functions.

3.3. Strings
Character strings are a very minor part of MATLAB, which is mainly designed to perform numerical
calculations. However, they perform some very useful tasks which are worth discussing now.
It is often important to combine text and numbers on a plot. Since we discuss graphics in the next sec-
tion, now is a good time to discuss how characters are stored in MATLAB variables. A string variable,
such as
>> str = ′ And now for something completely different′
is simply a row vector with each character (actually its ASCII representation as shown on page 139) being
a single element. MATLAB knows that this is a text variable, not a “regular” row vector, and so converts
the numerical value in each element into the corresponding character when it is printed out. For example,
to see what is actually contained in the vector str enter
>> str + 0
or
>> 1*str
Character variables are handled the same as vectors or matrices. For example, to generate a new text
variable which adds “– by Monty Python” to str, i.e., to concatenate the two strings, enter
>> str2 = [str ′ - by Monty Python′]
or
>> str2 = [str, ′ - by Monty Python′]
(which might be easier to read). To convert a scalar variable, or even a vector or a matrix, to a character
variable use the function num2str. For example, suppose you enter
>> x = linspace(0, 2*pi, 100)′
>> c1 = 2
>> c2 = -3
>> y = c1*sin(x) + c2*cos(x)
and want to put a description of the function into a variable. This can be done by
>> s = [ num2str(c1), ′ *sin(x) + ′ , num2str(c2), ′ *cos(x)′]
without explicitly having to enter the values of c1 and c2.
A text variable can also contain more than one line if it is created as a matrix. For example,
>> Str = [′ And ′
′ ′
now
′ ′
for
′ ′
something

completely′

different ′ ]
is four lines long. Since str is a matrix, each row must have the same number of elements and so we have
to pad all but the longest row. (Using cell arrays, we will shortly show how to avoid this requirement that
each row must have the same number of characters.)
Note: We do not usually enter matrices this way, i.e., one column per line. Instead, we simply use “;” to
separate columns. However here we need to make sure that each row has exa tly the same number
of characters — or else a fatal error message will be generated.
If desired, you can have more control over how data is stored in strings by using the sprintf command
which behaves very similarly to the C commands sprintf, fprintf, and printf. It is also very similar
to the fprintf command in MATLAB which is discussed in detail in Section 6. Note that the data can
be displayed directly on the screen by using disp. That is, sprintf(...) generates a character string
and disp(sprintf(...)) displays it on the screen.
There also is a str2num command to convert a text variable to a number and sscanf to do the same
with more control over how the data is read. (This is also very similar to the C command, as discussed in
section 6.)

39
3.4. Advanced Topic: Cell Arrays and Structures

3.4. Advanced Topic: Cell Arrays and Structures


It is occasionally useful in MATLAB to have a single variable contain all the data which is related to
a specific task — and this data might well consist of scalars, vectors and/or matrices, and text variables.
One simple reason for this is that it is easier to pass all the data into and out of functions. A cell array
generalizes the “standard” arrays which were discussed in the previous section. The elements of a “stan-
dard” array are numbers, either real or complex, whereas the elements of a cell array can be any data
type. The primary difference between a cell array and a structure is that in a structure the elements are
named rather than numbered. We consider this an advanced topic not because it is complicated, but be-
cause it is so seldom necessary.
A simple example of a cell array is
>> C = {2+3i, ′ go cells′; [1 2 3]′ , hilb(5) }
and the output is
C =


[2.0000 + 3.0000i] go cells′
[3x1 double] [5x5 double]
The only difference between this and a “standard” array is that here curly braces, i.e., { . . . }, enclose
the elements of the array rather than brackets, i.e., [ . . . ]. Note that only the scalar and the text vari-
able are shown explicitly. The other elements are only described. A second way to generate the same cell
array is by
>> C(1,1) = {2+3i }
>> C(1,2) = {′ go cells′}
>> C(2,1) = {[1 2 3]′ }
>> C(2,2) = {hilb(5) }
and a third way is by
>> C{1,1} = 2+3i
>> C{1,2} = ′ go cells′
>> C{2,1} = [1 2 3]′
>> C{2,2} = hilb(5)
It is important to understand that there is a difference between C(i,j) and C{i,j}. The former is the
cell containing element in the (i, j)-th location whereas the latter is the element itself. For example,
>> C(1,1) 5 % WRONG
ˆ
returns an error message because a cell cannot be raised to a power whereas
>> C{1,1} 5 % CORRECT
ˆ
returns 1.2200e+02 - 5.9700e+02i. All the contents of a cell can be displayed by using the celldisp
function. In addition, just as a “standard” array can be preallocated by using the zeros function, a cell
array can be preallocated by using the cell function. We will not discuss cells further except to state
that cell array manipulation is very similar to “standard” array manipulation.
Warning: In MATLAB you can change a variable from a number to a string to a matrix by simply
putting it on the left-hand side of equal signs. For example,
>> c = pi
>> c = ′ And now for something completely different′
>> c(5,3) = 17
redefines c twice without any difficulty. However, this cannot be done with cells. If you now
try
>> c{3} = hilb(5)
MATLAB will return with the error message
??? Cell contents assignment to a non-cell array object.
In order to use c as a cell (if has been previously used as a non-cell), you have to either clear it
using clear, empty it using [], or explicitly redefine it by using the cell function.
One particularly useful feature of cell arrays is that a number of text variables can be stored in one cell
array. We previously defined a “standard” array of strings in Str on page 39 where each string had to

40
3.4. Advanced Topic: Cell Arrays and Structures

have the same length. Using a cell array we can simply enter
>> Str cell = {′ And′

now′

for′

something′

completely′

different′}
or
>> Str cell = {′ And′ ; ′ now′ ; ′ for′ ; something′; ′ completely′; ′ different′}
and obtain the i-th row by
>> Str cell{i,:}
Structures can store different types of data similarly to cell arrays, but the data is stored by name,
called fields , rather than by number. Structures are very similar to structures in C and C++ . The cell
array we have been using can be written as a structure by
>> Cs.scalar = 2+3i
>> Cs.text = ′ go cells′
>> Cs.vector = [1 2 3]′
>> Cs.matrix = hilb(5)
Typing
>> Cs
returns
Cs =

scalar: 2.0000 + 3.0000i


text: ′ go cells′
vector: [3x1 double]
matrix: [5x5 double]
The structure can also be created using one command by
>> Cs = struct(′scalar′, 2+3i, ′ text′, ′ go cells′ , ...

vector′, [1 2 3]′ , ′ matrix′, hilb(5))
By the way, structures can themselves be vectors or matrices. For example,
>> Cs(2) = struct(′scalar′, pi, ′ text′ , ′ structures rule′, ...

vector′, ones(10,1), ′ matrix′, hilb(5) 2)
Now
ˆ
>> Cs
returns
Cs =

1x2 struct array with fields:


scalar
text
vector
matrix
Warning: As with cells, you cannot change a nonstructure variable to a structure variable. Instead, you
have to either clear it using clear, empty it using [], or explicitly redefine it by using the
struct function.
The field names of a structure can be handled using the following two commands. The command
fieldnames returns a cell array containing all the field names of a structure as strings. The command
isfield determines if a particular name, which is stored in a string, is a field of a particular structure.
Thus,
>> fieldnames(Cs)

41
4. Graphics

returns
ans =


scalar′

text′

vector′

matrix′
and
>> isfield(Cs, ’vector’)
returns 1.
We can also use function handles in cell elements and structures. For example, suppose you want to
work with all six basic trig functions. They can be stored in a cell array by
>> T = {@sin, @cos, @tan, @cot, @sec, @csc}
so that T{2}(0) = 1. They can also be stored in a structure by
>> Tr.a = @sin; Tr.b = @cos; Tr.c = @tan; Tr.d = @cot; Tr.e = @sec; Tr.f = @csc;
so that Tr.b(0) = 1. By the way, we can even store anonymous functions in cell arrays and structures.
For example,
>> C = {@sin, @(x) exp(sin(x)), @(x) exp(exp(sin(x)))}
is allowed — but probably not very interesting.
Note: We cannot store function handles in standard matrices — we can only store numbers.

Other Data Types

num2str(x) Converts a variable to a string. The argument can also be a vector or a matrix.
str2num(str) Converts a string to a variable. The argument can also be a vector or a matrix
string.
sscanf Behaves very similarly to the C command in reading data from a file using any
desired format. (See fscanf for more details.)
sprintf Behaves very similarly to the C command in writing data to a string using any
desired format. (See fprintf for more details.)
cell Preallocate a cell array of a specific size.
celldisp Display all the contents of a cell array.
struct Create a structure. Some useful commands are fieldnames, isfield, setfield,
getfield, orderfields, and rmfield.

4. Graphics
A very useful feature of MATLAB is its ability to generate high quality two- and three-dimensional
plots using simple and flexible commands. All graphical images are generated in a “graphics window”,
which is completely separate from the “text window” in which MATLAB commands are typed. Thus,
non-graphical and graphical commands can be completely intermixed.
Graphical images can be generated both from data calculated in MATLAB and from data which has
been generated outside of MATLAB. In addition, these images can be output from MATLAB and printed
on a wide variety of output devices, including color ink-jet printers and black-and-white and color laser
printers.
There are a number of demonstrations of the graphical capabilities in MATLAB which are invoked by
>> demo

42
4.1. Two-Dimensional Graphics

Since the MATLAB commands which generate the plots are also shown, this demo makes it quite easy to
generate your own graphics. You also can have very fine control over the appearance of the plots. We be-
gin by considering only the basic commands; more advanced graphics commands are discussed in the next
section.
Note: Most MATLAB commands which take vectors as arguments will accept either row or column vec-
tors.

4.1. Two-Dimensional Graphics

The MATLAB command plot is used to constructing basic two-dimensional plots. For example, sup-
cos x
pose you want to plot  the functions y1 = sin x and y2 = e for x ∈ [0, 2π]; also, you want to plot
2
y3 = sin cos(x − x) for x ∈ [0, 8]. First, generate n data points on the curve by
>> n = 100;
>> x = 2*pi*[0:n-1]′/(n-1);
>> y1 = sin(x);
>> y2 = exp(cos(x));
>> xx = 8*[0:n-1]/(n-1);
>> y3 = sin( cos( xx. 2 - xx ) );
We plot these data points by
ˆ
>> plot(x, y1)
>> plot(x, y2)
>> plot(xx, y3)
Note that the axes are changed for every plot so that the curve just fits inside the axes. We can generate
the x coordinates of the data points more easily by
>> x = linspace(0, 2*pi, n);
>> xx = linspace(0, 8, n);
The linspace command has two advantages over the colon operator:
(1) the endpoints of the axis and the number of points are entered directly as
>> x = linspace(<first point>, <last point>, <number of points>)
so it is much harder to make a mistake; and
(2) round-off errors are minimalized so you are guaranteed that x has exactly n elements, and its first
and last elements are exactly the values entered into the command.†
To put all the curves on one plot, type
>> plot(x, y1, x, y2, xx, y3)
Each curve will be a different color — but this will not be visible on a black-and-white output device.
Instead, you can change the type of lines by
>> plot(x, y1, x, y2, ′ --′ , xx, y3, ′ :′ )
where “ --” means a dashed line and “ :” means a dotted line. (We discuss these symbols in detail in
subsection 4.3.) In addition, you can use small asterisks to show the locations of the data points for the
y3 curve by
>> plot(x, y1, x, y2, ′ --′ , xx, y3, ′ :*′ )
These strings are used to modify the color of the line, to put markers at the nodes, and to modify the type
of line as shown in the table below. (As we discuss later in this section, the colors are defined by giving
the intensities of the red, green, and blue components.)


As we discussed previously, it is very unlikely (but it is possible) that round-off errors might cause the
statement
>> x = [0: 2*pi/(n-1): 2*pi]′ ;
to return n − 1 elements rather than n. This is why we used the statement
>> x = 2*pi*[0:n-1]′ /(n-1);
above, which does not suffer from round-off errors because the colon operator is only applied to integers.

43
4.1. Two-Dimensional Graphics

Customizing Lines and Markers

Symbol Color (R G B) Symbol Line Style Marker Description


r red (1 0 0) - solid line (default) + plus sign
g green (0 1 0) -- dashed line o circle
b blue (0 0 1) : dotted line * asterisk
y yellow (1 1 0) -. dash-dot line . point
m magenta (1 0 1) x cross
(a deep purplish red) s square
c cyan (0 1 1) d diamond
(greenish blue) upward pointing
w white (1 1 1)
ˆ triangle
k black (0 0 0) v downward pointing
triangle
> right pointing triangle
< left pointing triangle
p pentagram
h hexagram
(none) no marker

For example,
>> plot(x, y1, ′ r′ , x, y2, ′ g--o′ , x, y3, ′ mp′ )
plots three curves: the first is a red, solid line; the second is a a green, dashed line with circles at the data
points; the third has magenta pentagrams at the data points but no line connecting the points.
We can also plot the first curve, and then add the second, and then the third by
>> plot(x, y1)
>> hold on
>> plot(x, y2)
>> plot(xx, y3)
>> hold off
Note that the axes can change for every new curve. However, all the curves appear on the same plot.
Instead of putting a number of curves on one plot, you might want to put a number of curves individu-
ally in the graphics window. You can display m plots vertically and n plots horizontally in one graphics
window by
>> subplot(m, n, p)
This divides the graphics window into m×n rectangles and selects the p-th rectangle for the current plot.
All the graphics commands work as before, but now apply only to this particular rectangle in the graphics
window. You can “bounce” between these different rectangles by calling subplot repeatedly for different
values of p.
Warning: If you are comparing a number of plots, it is important to have the endpoints of the axes be the
same in all the plots. Otherwise your brain has to try to do the rescaling “on the fly” — which
is very difficult. Of course, you frequently do not know how large the axes need to be until you
have filled up the entire graphics window. The axis command (discussed below) can then be
used to rescale all the plots.
In addition, you can also change the endpoints of the axes by
>> axis([-1 10 -4 4])
The general form of this command is axis([xmin xmax ymin ymax]). If you only want to set some of
the axes, set the other or others to ±Inf ( −Inf if it is the minimum value and +Inf if it is the maxi-
mum). Also, you can force the two axes to have the same scale by
>> axis equal
or
>> axis image
and to have the same length by

44
4.1. Two-Dimensional Graphics

>> axis square


To learn about all the options for these commands, use the help or doc command.
Note: The command axis is generally only in effect for one plot. Every new plot turns it off, so it must
be called for every plot (unless hold is on ).
The plot command generates linear axes. To generate logarithmic axes use semilogx for a logarith-
mic axis in x and a linear axis in y, semilogy for a linear axis in x and a logarithmic axis in y, and
loglog for logarithmic axes in both x and y.
MATLAB has two different commands to plot a function directly rather than plotting a set of points.
Warning: These commands do not always generate the correct curve (or curves) because they know nothing
of the actual behavior of the function. They can have problems with sharp peaks and asymp-
totes and other “strange behavior”. We will show some examples shortly.
The first command we discuss is fplot, which can be executing by simply entering
>> fplot(<function handle>, <limits>)
where the function is usually generated as an anonymous function or a MATLAB function or a user gener-
ated function m-file (as described in section 8.3). The limits are either
[xmin xmax]
in which case the y-axis just encloses the curve or
[xmin xmax ymin ymax]
in which case you are also specifying the endpoints on the y-axis.
Note: Recall in section 3.2 we discussed how to pass a function as an argument.
This function uses adaptive step control to generate as many data points as it considers necessary to
plot the function accurately. You can also store the data points calculated by
>> [x, y] = fplot(<function handle>, <limits>)
rather than having the function plotted directly. You then have complete control over how to plot the
curve using the plot function.
The other command which can plot a function is ezplot, which is more general than fplot. To plot a
function on the interval [−2π, +2π] enter
>> ezplot(<function handle>)
To include limits (as with fplot ) enter
>> ezplot(<function handle>, <limits>)
In addition, a parametrically defined function can be plotted by
>> ezplot(<fnc 1>, <fnc 2>, <limits>)
Finally, this command can also plot an implicitly defined function, i.e., f (x, y) = 0, by
>> ezplot(<2D fnc>, <limits>)
For example,
>> f = @(x, y) (x 2 + y 2) 2 - (x 2 - y 2);
>> ezplot(f)
ˆ ˆ ˆ ˆ ˆ
plots the lemniscate of Bernoulli (basically an “∞” symbol).
Warning: Be particularly careful when plotting implicit functions because they can be really nasty and
occasionally ezplot may not get it right.
There is an important difference between
>> fplot(f, [-5 5])
and
>> ezplot(f, [-5 5])
In the former f(x) is only evaluated for scalar values of x, while in the latter f(x) is evaluated for vec-
tor values of x. Thus, when using ezplot care must be taken if f is evaluated in a function m-file. If
f(x) cannot be evaluated for vector values, the error message
Warning: Function failed to evaluate on array inputs; vectorizing the function may
speed up its evaluation and avoid the need to loop over array elements.
will be generated
fplot and ezplot do not always generate exactly the same curves. For example, in
>> f = @(x) log(x) + 1;
>> fplot(f, [-2*pi 2*pi])
>> ezplot(f)
fplot generates a spurious plot for x ∈ [−2 ∗ pi, 0) where it plots the real part of log x while ezplot

45
4.1. Two-Dimensional Graphics

only plots the function for x ∈ (0, 2π]. Also, in


>> f = @(x) x ./ (x 2 + 0.01);
ˆ
>> fplot(f, [-2*pi +2*pi])
>> ezplot(f)
the vertical axes are different and ezplot is missing part of the curve. Finally, in
f = @(x) x 3/(x 2 + 3*x - 10);
ˆ ˆ
ezplot(f, [-10 +10])
the function blows up at x = −5 and 2 and part of the curve for x ∈ (−5, 2) is not shown.
Polar plots can also be generated by the polar command. There is also an “easy” command for gener-
ating polar plots, namely ezpolar.
Since you often want to label the axes and put a title on the plot, there are specific commands for each
of these. Entering
>> xlabel(<string>)
>> ylabel(<string>)
>> title(<string>)
put labels on the x-axis, on the y-axis, and on top of the plot, respectively. Note that a title can contain
more than one line as was discussed in section 3.
For example, typing title(t) where
t = [′ The Dead ′

Parrot Sketch′]
or
t = {′ The Dead′

Parrot Sketch′}
or
t = {′ The Dead′ ; ′ Parrot Sketch′}
results in a two-line title. The former uses a “standard” array and so requires all the rows to have the
same number of columns, whereas the latter uses a cell array and so each row can have a different length.
There are also a number of ways to plot data, in addition to the commands discussed above. The two
we discuss here are histograms and error bars. To plot a histogram of the data stored in the vector x,
type
>> hist(x)
which draws ten bins between the minimum and maximum values of the elements in x. For example, to
see how uniform the distribution of random numbers generated by rand is, type
>> x = rand(100000, 1);
>> hist(x)
To draw a histogram with a different number of bins, type
>> hist(x, <number of bins>)
and to draw a histogram with the centers of the bins given by the vector c, type
>> hist(x, c)
As another example, to see how uniform the distribution of Gaussian random numbers generated by
randn is, type
>> x = randn(1000, 1);
>> hist(x)
Clearly you need more random numbers to get a “good” histogram — but, at the moment, we are inter-
ested in a different point. If you rerun this command a number of times, you will find that the endpoints
of the histogram fluctuate. To avoid this “instability”, fix the endpoints of the histogram by
>> xmax = 4;
>> nrbin = 20;
>> nrdata = 1000;
>> c = xmax*[ -1+1/nrbin : 2/nrbin : 1-1/nrbin ];
>> x = randn(nrdata, 1);
>> hist(x, c)

46
4.1. Two-Dimensional Graphics

Note that c contains the midpoints of each bin and not their endpoints. Another way to calculate c,
which might be clearer, is
>> c = linspace(-xmax+xmax/nrbin, xmax-xmax/nrbin, nrbin);
Of course, to get a “good” histogram you should increase nrbin, say to 100, and nrdata, say to
100,000. If you now rerun this code you will see a much smoother histogram.
We have already seen how to plot the vector x vs. the vector y by using the plot command. If, ad-
ditionally, you have an error bar of size ei for each point yi , you can plot the curve connecting the data
points along with the error bars by
>> errorbar(x, y, e)
Sometimes the error bars are not symmetric about the y values. In this case, you need vectors l and u
where at xi the error bars extend from yi − li to yi + ui . This is done by
>> errorbar(x, y, l, u)
Note: All the elements of l and u are non-negative.
Data can also be entered into MATLAB from a separate data file. For example,
>> M = csvread(′<file name>′)
reads in data from a file one row per line of input. The numbers in each line must be separated by com-
mas. The data can then be plotted as desired. The command csvwrite writes the elements of a matrix
into a file using the same format. (If desired, you can have much more control over how data is input and
output by using the fscanf and fprintf commands, which are similar to their C counterparts. These
commands are discussed in detail in section 6.)
The load command can also be used to read a matrix into MATLAB from a separate data file. The
data must be stored in the data file one row per line. The difference between this command and csvread
is that the numbers can be separated by commas or by spaces. The matrix is input by entering
>> load(′ <file name>′)
and it is stored in the matrix named <file name-no extension> (i.e., drop the extension, if any, in the
file name).†
Graphics can also be easily printed from within MATLAB. You can print directly from the graphics
window by going into the “File” menu item. If desired, the plot can be sent to a file rather than to an
output device. You can also store the plot in the text window by using the command print. There are
an innumerable number of printer specific formats that can be used. (See help print or doc print for
details.) If you want to save a file in postscript, you can save it in black-and-white postscript by
>> print -deps <file name b&w>
or in color postscript by
>> print -depsc <file name color>
There is a minor, but important, difference between these two files if they are printed on a black-and-
white laser printer. When the black-and-white file is printed, all the non-white colors in the plot become


The load command is a little tricky because it can read in files generated both by MATLAB (using the save
command) and by the user. For example,
>> save allvariables;
>> clear
or
>> save allvariables.mat;
>> clear
saves all the variables to the file allvariables.mat in binary format and then deletes all the variables. Entering
>> load allvariables
or
>> load allvariables.mat
loads all these variables back into MATLAB using the binary format. On the other hand, if you create a file, say
mymatrix.dat, containing the elements of a matrix and enter it into MATLAB using
>> load(′ mymatrix.dat′ )
you obtain a new matrix, called mymatrix, which contains these elements. Thus, the load command determines
how to read a file depending on the extension.

47
4.1. Two-Dimensional Graphics

black. However, when the color file is printed, the colors are converted to different grayscales. This makes
it possible to differentiate lines and/or regions.
Note: The print command is also a MATLAB function where it is called by
>> print(′-deps′, ′ <file name b&w>′ )
The advantage of using the print function is that the arguments can be variables. An
oversimplified example is
>> device = ′ -deps′
>> file = ′ <file name b&w>′
>> print(device, file)
It is oversimplifed because there is no need to use three lines when one will do. However, if many
plots are to be printed then the print device can be changed once rather than in every print
command. Also, if you are printing many plots then you can easily modify the file names as in
>> i = 1
>> file = [′ fiddledum′, num2str(i), ′ .eps′ ]
>> print(device, file)
>> . . .
>> i = i + 1
>> file = [′ fiddlededum′, num2str(i), ′ .eps′ ]
>> print(device, file)
>> . . .

Input-Output

csvread(′<file name>′ ) Reads data into MATLAB from the named file, one row per line of
input; the numbers in each line must be separated by commas.
load(′<file name>′) Reads data into MATLAB from the named file, one row per line
of input; the numbers in each line can be separated by spaces or
commas. The name of the resulting matrix is <file name>.
csvwrite(′<file name>′, A) Writes out the elements of a matrix to the named file using the same
format as csvread.
print Prints a plot or saves it in a file using various printer specific for-
mats.
For example, print -deps <file name> saves the plot in the file
using encapsulated PostScript (so it can be plotted on a PostScript
laser printer).

48
4.2. Three-Dimensional Graphics

Two-Dimensional Graphics

plot(x, y) Plots the data points in Cartesian coordinates. The gen-


eral form of this command is
plot(x1, y1, s1, x2, y2, s2, ...) where s1, s2,
. . . are optional character strings containing information
about the type of line, mark, and color to be used.
Some additional arguments that can be used:
plot(x) plots x vs. the index number of the elements.
plot(Y) plots each column of Y vs. the index number of
the elements.
plot(x,Y) plots each column of Y vs. x.
If z is complex, plot(z) plots the imaginary part of z
vs. the real part.
semilogx The same as plot but the x axis is logarithmic.
semilogy The same as plot but the y axis is logarithmic.
loglog The same as plot but both axes are logarithmic.
fplot(<function handle>, <limits>) Plots the specified function within the limits given. The
limits can be [xmin xmax] or [xmin xmax ymin ymax].
)
ezplot(<function handle> Generates an “easy” plot (similar to fplot ) given the
ezplot(<fnc 1>, <fnc 2>) function f (x).  It can also plot a parametric function,
ezplot(<2D fnc>) i.e., x(t), y(t) , or an implicit function, i.e., f (x, y) = 0.
Limits can also be specified if desired.
polar(r, theta) Plots the data points in polar coordinates.
ezpolar(<function handle>) Generate an “easy” polar plot of
r = < functionname >(θ).
xlabel(<string>) Puts a label on the x-axis.
ylabel(<string>) Puts a label on the y-axis.
title(<string>) Puts a title on the top of the plot.
axis Controls the scaling and the appearance of the axes.
axis equal and axis([xmin xmax ymin ymax]) are
two common uses of this command.
hold Holds the current plot ( hold on ) or release the current
plot ( hold off ).
linspace(a, b, n) Generates n equally-spaced points between a and b
(inclusive).
hist(x) Plots a histogram of the data in a vector using 10 bins.
hist(x, <number of bins>) changes the number of
bins.
hist(x, c) lets you choose the midpoint of each bin.
o
errorbar(x, y, e) The first plots the data points x vs. y with error bars
errorbar(x, y, l, u) given by e. The second plots error bars which need not
be symmetric about y.
subplot(m, n, p) Divides the graphics window into m×n rectangles and
selects the p-th rectangle for the current plot.

4.2. Three-Dimensional Graphics


The MATLAB command plot3 plots curves in three-dimensions. For example, to generate a helix

49
4.2. Three-Dimensional Graphics

enter
>> t = linspace(0, 20*pi, 1000);
>> c = cos(t);
>> s = sin(t);
>> plot3(c, s, t)
and to generate a conical helix enter
>> t = [0 : pi/100 : 20*pi];
>> c = cos(t);
>> s = sin(t);
>> plot3(t.*c, t.*s, t)
Also, you can put a label on the z-axis by using the  zlabel command. There is also an “easy” plot3
command. It generates the curve x(t), y(t), z(t) for t ∈ (0, 2π) by
>> ezplot3(x, y, z)
if x, y, and z have been defined using anonymous functions. Again, you change the domain of t by
specifying the additional argument [tmin, tmax].
MATLAB also plots surfaces z = f(x, y) in three-dimensions with the hidden surfaces removed. First,
the underlying mesh must be created. The easiest way is to use the command meshgrid. This combines
a discretization of the x axis, i.e., { x1 , x2 , . . . , xm }, and
the y axis, i.e., { y1 , y2 , . . . , yn }, into the rect-
angular mesh (xi , yj ) i = 1, 2, . . . , m , j = 1, 2, . . . , n in the x–y plane. The function f can then be
evaluated at these mesh nodes. For example,
>> x = [-3:0.1:3]′;
>> y = [-2:0.1:2]′;
>> [X, Y] = meshgrid(x, y);
>> F = (X + Y).*exp( -X.*X - 2*Y.*Y );
>> mesh(X, Y, F)
generates a colored, wire-frame surface whereas
>> surf(X, Y, F)
generates a colored, filled-in surface. We discuss how to change the colors, and even how to use the colors
as another variable, in the next section.
You can change the view of a three-dimensional plot by using the view command. This command is
called in either of two ways:
• First, you can give the angles from the origin of the plot to your eye by
view(<azimuth>, <elevation>)
where the azimuth is the angle in degrees in the x–y plane measured from the −y axis (so 0◦ is the
−y axis, 90◦ is the x axis, 180◦ is the y axis, etc.) and the elevation is the angle in degrees up
from the x–y plane toward the +z axis (so 0◦ is in the x–y plane, 90◦ is on the +z axis, etc.).
• Second, you can give the coordinates of a vector pointing from the origin of the plot to your eye by
view([x y z]), where you enter the coordinates of the vector.
If you type
>> contour(X, Y, F)
you will see contour plots of the surface. That is, you will be looking down the z axis at curves which
represent lines of constant elevation (i.e., constant z values). If we type
>> contour3(X, Y, F)
you will see contour plots of the surface in three dimensions. You can again change your view of these
curves by using the view command.
If you do not want to bother with generating the mesh explicitly, you can generate “easy” plots by
ezcontour, ezcontour3, ezmesh, and ezsurf.

50
4.3. Advanced Graphics Techniques: Commands

Three-Dimensional Graphics

plot3(x, y, z) Plots the data points in Cartesian coordinates. The general form of this
command is plot(x1, y1, z1, s1, x2, y2, z2, s2, ...) where s1,
s2, . . . are optional character strings containing information about the
type of line, mark, and color to be used.
ezplot3(<fnc 1>, Generates an “easy” plot in 3-D.
<fnc 2>, <fnc 3>)
mesh(X, Y, Z) Plots a 3-D surface using a wire mesh.
ezmesh(<2D fnc>) Generates an “easy” 3-D surface using a wire mesh.
surf(X, Y, Z) Plots a 3-D filled-in surface.
ezsurf(<2D fnc>) Generates an “easy” 3-D filled-in surface.
view Changes the viewpoint of a 3-D surface plot by
view(<azimuth>, <elevation>) or view([x y z]).
meshgrid(x, y) Generates a 2-D grid given the x-coordinates and the y-coordinates of the
mesh lines.
zlabel(<string>) Puts a label on the z-axis.
axis Controls the scaling and the appearance of the axes.
axis([xmin xmax ymin ymax zmin zmax]) changes the endpoints of the
axes.
contour(X, Y, Z) Plots a contour looking down the z axis.
ezcontour(<2D fnc>) Generates an “easy” contour looking down the z axis.
contour3(X, Y, Z) Plots a contour in 3-D.
ezcontour3(<2D fnc>) Generates an “easy” contour in 3-D.
subplot(m, n, p) Remember than subplot can also be called in 3-D to put a number of
plots in one graphics window.

4.3. Advanced Graphics Techniques: Commands


In the previous subsections we have discussed how to use “simple” graphics commands to generate ba-
sic plots. MATLAB can also do much more “interesting” graphics, and even publication quality graphics.
Here we discuss some of the more useful advanced features. We divide the topic into two subsections: the
first discusses the commands themselves and the second discusses how to change some of the properties of
these commands.
Note: The demonstration program shows many more of the graphics capabilities of MATLAB. Enter
demo
and then in Help Navigator click on Graphics.
It is possible to obtain the current position of the cursor within a plot by using the ginput command.
For example, to collect any number of points enter
>> [x, y] = ginput
Each position is entered by pressing any mouse button or any key on the keyboard except for the carriage
return (or enter) key. To terminate this command press the return key. To enter exactly n positions, use
>> [x, y] = ginput(n)
You can terminate the positions at any time by using the return key. Finally, to determine which mouse
button or key was entered, use
>> [x, y, button] = ginput(n)
The vector button contains integers specifying which mouse button (1 = left, 2 = center, and 3 = right)
or key (its ASCII representation) was pressed.
Labels can also be added to a plot. Text can be placed anywhere inside the plot using
>> text(xpt, ypt, <string>)
The text is placed at the point (xpt,ypt) in units of the current plot. The default is to put the center of
the left-hand edge of the text at this point. You can also use the mouse to place text inside the plot using

51
4.3. Advanced Graphics Techniques: Commands

>> gtext(<string>)
The text is fixed by depressing a mouse button or any key.
If more than one curve appears on a plot, you might want to label each curve. This can be done directly
using the text or gtext command. Alternatively, a legend can be put on the plot by
>> legend(<string1>, <string2>, . . . )
Each string appears on a different line preceded by the type of line (so you should use as many strings as
there are curves). The entire legend is put into a box and it can be moved within the plot by using the
left mouse button.
TEX commands can be used in these strings to modify the appearance of the text. The results are
similar, but not quite identical, to the appearance of the text from the TEX program (so do some exper-
imenting). Most of the “common” TEX commands can be used, including Greek letters; also, “ ” and
ˆ
“R ” are used for superscripts and subscripts. For example, the x-axis can be labelled α2 and the y-axis
α
0 f (x) dx by
>> xlabel(′\alpha 2′ )
ˆ
>> ylabel(′\int 0 \pi\betaf(x) dx′ )
ˆ
To see the complete list of TEX commands, enter
>> doc text
and then click on the highlighted word String.
Note: For you TeXers note the funny control sequence “\betaf(x)” which generates βf (x). If you would
have typed “\beta f(x)” you would have obtained β f (x) because MATLAB preserves spaces. If
typing “\betaf(x)” sets your teeth on edge, try “\beta{}f(x)” instead.
It is often essential for the title to include important information about the plot (which would, other-
wise, have to be written down “somewhere” and connected to this specific plot). For example, suppose
you enter
>> x = linspace(0, 2*pi, 100)
>> c1 = 2
>> c2 = -3
>> p1 = 1
>> p2 = 3
>> y = c1*sin(x). p1 + c2*cos(x). p2
>> plot(x, y)
ˆ ˆ
and you want to “play around” with the two coefficients to obtain the most “pleasing” plot. Then you
probably should have the title include a definition of the function — and you should not have to modify
the title by hand every time you change the coefficients. This can be done by
>> t = [num2str(c1), ′ *sin ′ , num2str(p1), ′ (x) + ′ , num2str(c2), ...
′ ˆ
*cos ′ , num2str(p2), ′ (x)′ ]
>> title(t)
ˆ
where we use the text variable t, rather than putting the string directly into title, simply to make
the example easier to read. There is now a permanent record of the function which generated the curve.
(Alright, this isn’t a great example, but it’s better than nothing.)
You can also put plots in a new graphics window by entering
>> figure
where the figures are numbered consecutively starting with one (and appear at the top of the window). Or
enter
>> figure(n)
and the figure will have the specific number n. This creates a new window and makes it the current tar-
get for graphics commands. You can “bounce” between graphics windows by entering
>> figure(n)
where n is the number of the graphics window you want to make current. New plots will now appear in
this figure. In this way much more information can be generated and viewed on the computer terminal.
Occasionally, it is useful to clear a figure. For example, suppose you divide a window into a 2×2 ar-
ray of plotting regions and use subplot to put a plot into each region; you then save the figure into a
file. Next, you only want to put plots into two of these four regions. The difficulty is that the other two
regions will still contain the previous plots. You can avoid this difficulty by clearing the figure using

52
4.3. Advanced Graphics Techniques: Commands

>> clf
which clears the current figure. You can clear a particular figure by clf(n).
All the above MATLAB commands can be used for 3-D graphics except for gtext. The text com-
mand is the same as described above except that the position of the text requires three coordinates, i.e.,
>> text(x, y, z, <string>)
The new command
>> zlabel(<string>)
labels the z-axis.
As we discussed in the previous subsection, the mesh and surf commands allow us to plot a surface
in three dimensions where the colors on the surface represent its height. We can add a rectangle which
contains the correspondence between the color and the height of the surface by adding
>> colorbar
We can also let the colors represent a separate quantity C, which is also defined at each mesh point, by
changing the command to
>> mesh(X, Y, F, C)
or
>> surf(X, Y, F, C)
Each graphics window has a separate color map associated with it. This color map is simply an n×3
matrix, where each element is a real number between 0 and 1 inclusive. In each row the first column
gives the intensity of the color red, the second column green, and the third column blue; these are called
the RGB components of a color. For example, we show the RGB components of cyan, magenta, yellow,
red, blue, green, white, and black in the table “Customizing Lines and Markers” at the beginning of this
section; for further information, enter doc colorspec. The value input to this color map is the row rep-
resenting the desired color.
For mesh or surf the value of F (or of C if there is a fourth argument) is linearly rescaled so its mini-
mum value is 1 and its maximum value is n. To see the current color map, enter
>> colormap
To change the color map, enter
>> colormap(<color map>)
where <color map> can be an explicit n×3 matrix of the desired RGB components or it can be a string
containing the name of an existing color map. The existing color maps can be found by typing
>> doc graph3d
A useful color map for outputting to laser printers is ′ gray′ . In this colormap all three components of
each row have the same value so that the colors change gradually from black (RGB components [0 0 0])
through gray [.5 .5 .5]) to white [1 1 1]).
MATLAB can also plot a two-dimensional image (i.e., a picture) which is represented by a matrix
X ∈ Rm×n . The (i, j)-th element of X specifies the color to use in the current color map. This color
appear in the (i, j)-th rectilinear patch in the plot. For example, to display the color image of a clown
enter
>> load clown
>> image(X);
>> colormap(map)
The image command inputs the matrix X and the colormap map from clown.mat. Then the image is
displayed using the new color map. Similarly,
>> load earth
>> image(X);
>> colormap(map);
>> axis image
displays an image of the earth. (The axis command forces the earth to be round, rather than elliptical.)
(In the demonstration program, after clicking on “Visualization” double-click on “Image colormaps” to see
the images which you can access in MATLAB and the existing color maps.)

53
4.3. Advanced Graphics Techniques: Commands

MATLAB can also fill-in two-dimensional polygons using fill or three-dimensional polygons using
fill3. For example, to draw a red circle surrounding a yellow square, enter
>> t = linspace(0, 2*pi, 100);
>> s = 0.5;
>> xsquare = [-s s s -s]′ ;
>> ysquare = [-s -s s s]′ ;
>> fill(cos(t), sin(t), ′ r′ , xsquare, ysquare, ′ y′ )
>> axis equal;
To obtain a more interesting pattern replace the above fill command by
>> colormap(′hsv′ );
>> fill(cos(t), sin(t), [1:100], xsquare, ysquare, [100:10:130])
Rather than entering polygons sequentially in the argument list, you can enter
>> fill(X, Y, <color>)
where each column of X and Y contain the endpoints of a different polygon. Of course, in this case the
number of endpoints of each polygon must be the same, by padding if necessary. For example, to draw a
cube with all the faces having a different solid color, input the matrices
     
0 1 1 0 0 0 0 0 1 1 0 0 0 0 1 0 0 1
1 1 0 0 1 1 0 1 1 0 0 0 0 0 1 0 0 1
X=  , Y=  , Z= .
1 1 0 0 1 1 0 1 1 0 1 1 1 1 1 1 0 1
0 1 1 0 0 0 0 0 1 1 1 1 1 1 1 1 0 1

Then enter
>> fill3(X, Y, Z, [1:6])
>> axis equal
Change your orientation using view to see all six faces. Read the documentation on fill and fill3 for
more details.

Advanced Graphics Features: Plots

clf Clear a figure (i.e., delete everything in the figure)


colorbar Adds a color bar showing the correspondence between the value and
the color.
colormap Determines the current color map or choose a new one.
demo Runs demonstrations of many of the capabilities of MATLAB.
figure Creates a new graphics window and makes it the current target.
figure(n) makes the n-th graphics window the current target.
fill(x, y, <color>) Fills one or more polygons with the color or colors specified by the
vector or string <color>.
fill3(x, y, z, <color>) Fills one or more 3D polygons with the color or colors specified by the
vector or string <color>.
image Plots a two-dimensional image.

Advanced Graphics Features: Text and Positioning

ginput Obtains the current cursor position.


o
text(x, y, <string>) Adds the text to the location given in the units of the current plot.
text(x, y, z, <string>)
gtext(<string>) Places the text at the point given by the mouse.
legend(<string 1>, ...) Places a legend on the plot using the strings as labels for each type of
line used. The legend can be moved by using the mouse.

54
4.4. Advanced Graphics Techniques: Handles and Properties

4.4. Advanced Graphics Techniques: Handles and Properties


In this subsection we briefly discuss handle graphics. This is a collection of low-level graphics commands
which do the actual work of generating graphics. In the previous parts of this section we have mainly
discussed “high-level” graphics commands which allow us to create useful and high quality graphical im-
ages very easily. The low-level commands allow us to customize these graphical images, but at the cost of
having to get much more involved in how graphical images are actually created. This subsection will be
quite short because we do not want to get bogged down in this complicated subject. Instead, we will only
discuss a few of — what we consider to be — the more useful customizations.
In handle graphics we consider every component of a graphical image to be an object, such as a subplot,
an axis, a piece of text, a line, a surface, etc. Each object has properties and we customize an object by
changing its properties. Of course, we have to be able to refer to a particular object to change its prop-
erties, and a handle is the unique identifier which refers to a particular object. (Each handle is a unique
floating-point number.)
We will use a small number of examples to explain handle graphics. There are many properties of the
text that can be changed in the text command by
>> text(xpt, ypt, <string>, ′ <Prop 1>′ , <Value 1>, ′ <Prop 2>′ , <Value 2>, . . . )
or
>> h = text(xpt, ypt, <string>);
>> set(h, ′ <Prop 1>′ , <Value 1>, ′ <Prop 2>′ , <Value 2>, . . . )
where <Prop ?> is the name of one of the properties for the text object and <Value ?> is one of the
allowed values. (We show some names and values in the following table.) We have shown two ways to cus-
tomize the properties. In the former all the properties are set in the text command. In the latter the
text command creates an object, using its default properties, with handle h. The set command then
changes some of the properties of the object whose handle is h. For example, entering
>> set(h, ′ Color′, ′ r′ , ′ FontSize′, 16, ′ Rotation′, 90)
results in a large, red text which is rotated 90◦ . You can also change the default properties for gtext,
xlabel, ylabel, zlabel, and title.

Text Properties

Clipping on — (default) Any portion of the text that extends outside the axes rect-
angle is clipped
off — No clipping is done.
Color A three-element vector specifying a color in terms of its red, blue, and
green components, or a string of the predefined colors.
FontName The name of the font to use. (The default is Helvetica.)
FontSize The font point size. (The default is 10 point.)
HorizontalAlignment left — (default) Text is left-justified
center — Text is centered.
right — Text is right justified.
Rotation The text orientation. The property value is the angle in degrees.
VerticalAlignment top — The top of the text rectangle is at the point.
cap — The top of a capital letter is at the point.
center — (default) The text is centered vertically at the point.
baseline — The baseline of the text is placed at the point.
bottom — The bottom of the text rectangle is placed at the point.

55
4.5. Be Able To Do

The more common way of customizing parameters is by using the set command. The two functions
get and set are used to obtain the value of one parameter and to set one or more parameters. For ex-
ample, to get the font which is presenting being used enter
>> s = get(h, ′ FontName′)
and the string s contains the name of the font. The two arguments to get are the handle of the object
desired and the name of the property.
There are two other commands which can obtain a handle:
>> hf = gcf
returns the handle of the current figure and
>> ha = gca
returns the handle of the current axes in the current figure. There is one case where we frequently use
handle graphics. When a figure is printed, the graphical images do not fill the entire page. The default
size is approximately 6.5 inches wide and 5.5 inches high. When we want to use the full size of a sheet of
paper we use
>> figure(′PositionPaper′, [0 0 8.5 11])
or
>> figure(n)
>> set(gcf, ′ PositionPaper′, [0 0 8.5 11])
since the default units for this property are inches. This sets the graphical images to use the full paper
size (the position is given as [left bottom width height] ) with a one inch border. This is frequently
useful if subplot is being used to put a number of plots on a page.
Finally, if subplot is being used it is sometimes useful to put a title on the entire page, not just in
each subplot. This can be done by
>> axes handle = axes(′Position′, [0 0 1 0.95], ′ Visible′, ′ off′ );
>> title handle = get(axes handle, ′ title′);
>> set(title handle, ′ String′, <title>, ′ Visible′, ′ on′ );
The first line specifies a rectangle for the axes in normalized units (so that (
[left bottom width height] = [0 0 1 1] is the full size of the figure). The axes are invisible because
they are only being created so that a title can go on top. The second line gets the handle for the title
object of the new axes. The third line puts <title> into the title object and makes it visible.

Advanced Graphics Features: Properties

get(<handle>, ′ <Prop>′) Return the current value of the property of the object with this
handle.
set(<handle>, ′ <Prop 1>′ , Set the property, or properties, of the object with this handle.
<Value 1>, ...)
gca The current axes handle.
gcf The current figure handle.

4.5. Be Able To Do

After reading this section you should be able to do the following exercises. The answers are given on
page 137.
1. Plot ex and one of its Taylor series approximations.
(a) Begin by plotting ex for x ∈ [−1, +1].
(b) Then plot
x x2 x3
p3 (x) = 1 + + +
1! 2! 3!
on the same graph.

56
5.1. Square Linear Systems

(c) Also plot the difference between ex and this cubic polynomial, i.e., ex − p3 (x) on the same graph.
(d) Next, generate a new graph containing all three curves by using only one plot command, force the
axes to be to the same scale, and let all three curves have different colors. Put labels on the x and y
axes and a silly title on the entire plot.
(e) The above plot is not very instructive because ex − p3 (x) is much smaller than either ex or p3 (x).
Instead, use two plots. The first plot contains ex and p3 (x) and the second plot, which is immediately
below the first, contains ex − p3 (x). These two plots should fill an entire sheet of paper.
2. Consider the function
f (x, y) = (x2 + 4y 2 ) sin(2πx) sin(2πy) .

(a) Plot this function for x, y ∈ [−2, +2].


Note: Make sure you use the “ .∗” operator in front of each sine term. What does the surface look like if
you don’t?
(b) This surface has high peaks which interfere with your view of the surface. Change your viewpoint so
you are looking down at the surface at such an angle that the peaks do not block your view of the central
valley.
Note: There are an infinite number of answers to this part.

5. Solving Linear Systems of Equations


One of the basic uses of MATLAB is to solve the linear system

a11 x1 + a12 x2 + . . . + a1n xn = b1


a21 x1 + a22 x2 + . . . + a2n xn = b2
.. ..
. .
aj1 x1 + aj2 x2 + . . . + ajn xn = bj
... ...
am1 x1 + am2 x2 + . . . + amn xn = bm ,

or the equivalent matrix equation


Ax = b .
Note that there are m equations in n unknowns so that there may be zero solutions to this linear sys-
tem, one solution, or an infinite number of solutions. We will discuss the case where m 6= n in detail in
subsection 5.3. Here we concentrate on m = n.

5.1. Square Linear Systems


As we discussed previously, when m = n the MATLAB operation
>> x = A\b
calculates the unique solution x by Gaussian elimination when A is nonsingular. We can also solve it by
>> x = linsolve(A, b)
The advantage of using linsolve is that it can be much faster when A has a particular property. The
third argument to linsolve gives the particular property. For our purposes the most important proper-
ties are lower triangular, upper triangular, symmetric, and positive definite. Enter
>> x = linspace(A, b, prop)
where prop is a logi al structure with the following elements:
LT – the matrix is lower triangular,
UT – the matrix is upper triangular,
SYM – the matrix is symmetric triangular, and
POSDEF – the matrix is positive definite.
Normally, all the elements are false; set the property you want to true by
>> prop.??? = true

57
5.1. Square Linear Systems

where ??? is one of the above properties. To turn this property back off, enter
>> prop.??? = false
(We discuss logical variables in section 8.1.) If A has one (or more) of these properties, it can be solved
much faster than using x = A\b.
When A is singular there are either zero solutions or an infinite number of solutions to this equation
and a different approach is needed. The appropriate MATLAB command system is now rref. It begins
by applying Gaussian elimination to the linear system of equations. However, it doesn’t stop there; it con-
tinues until it has zeroed out all the elements it can, both above the main diagonal as well as below it.
When done, the linear system is in reduced row echelon form:
• The first nonzero coefficient in each linear equation is a 1 (but a linear equation can be simply 0 = 0
in which case it has no nonzero coefficient).
• The first nonzero term in a particular linear equation occurs later than in any previous equation.
That is, if the first nonzero term in the j-th equation is xkj and in the j+1-st equation is xkj+1 ,
then kj+1 > kj .
To use rref, the linear system must be written in augmented matrix form, i.e.,

x1 x2 ... xn = rhs
 
a11 a12 ... a1n b1
a
 21 a22 ... a2n b2 
 ... ... .. ... ...  .

.
am1 am2 ... amn bm

Warning: It is very important to realize that an augmented matrix is not a matrix (because the
operations we apply to augmented matrices are not the operations we apply to matrices). It is
simply a linear system of equations written in shorthand: the first column is the coefficients of
the x1 term, the second column is the coefficients of the x2 term, etc., and the last column is
the coefficients on the right-hand side. The vertical line between the last two columns repre-
sents the equal sign. Normally, an augmented matrix is written without explicitly writing the
header information; however, the vertical line representing the equal sign should be included to
explicitly indicate that this is an augmented matrix.
rref operates on this augmented matrix to make as many of the elements as possible zero by using al-
lowed operations on linear equations — these operations are not allowed on matrices, but only on linear
systems of equations. The result is an augmented matrix which, when written back out as a linear system
of equations, is particularly easy to solve. For example, consider the system of equations

x1 + 2x2 + 3x3 = −1
4x1 + 5x2 + 6x3 = −1
7x1 + 8x2 + 10x3 = 0 ,

which is equivalent to the matrix equation Ax = b where


   
1 2 3 −1
A = 4 5 6  and b =  −1  .
7 8 10 0

The augmented matrix for this linear system is

x1 x2 x3 = rhs
 
1 2 3 −1
 4 5 6 −1  .
7 8 10 0

(We have included the header information for the last time.) Entering

58
5.1. Square Linear Systems

>> rref([A b])


returns the augmented matrix  
1 0 0 2
0 1 0 −3  .
0 0 1 1
Clearly, the solution of the linear system is x1 = 2, x2 = −3, and x3 = 1.
Of course, you could just as easily have found the solution by
>> x = A\b
so let us now consider the slightly different linear system

x1 + 2x2 + 3x3 = −1
4x1 + 5x2 + 6x3 = −1
7x1 + 6x2 + 9x3 = −1 ,

This is equivalent to the matrix equation Ax = b where


   
1 2 3 −1
A = 4 5 6 and b =  −1  .
7 8 9 −1

Since A is a singular matrix, the linear system has either no solutions or an infinite number of solutions.
The augmented matrix for this linear system is
 
1 2 3 −1
4 5 6 −1  .
7 8 9 0

Entering
>> rref([A b])
returns the augmented matrix  
1 0 −1 1
0 1 2 −1  ,
0 0 0 0
so the solution of the linear system is x1 = 1 + x3 and x2 = −1 − 2x3 for any x3 ∈ R (or C if desired).
In vector form, the solution is
           
x1 1 + x3 1 x3 1 1
 x2  =  −1 − 2x3  =  −1  +  −2x3  =  −1  + x3  −2  .
x3 x3 0 x3 0 1

Suppose you modify the matrix equation slightly by letting b = (−1, −1, 0)T . Now entering
>> rref([A b])
results in the augmented matrix  
1 0 −1 1
0 1 2 −1  .
0 0 0 1
Since the third equation is 0 = 1, there is clearly no solution to the linear system.
Warning: The command rref does not always give correct results. For example, if
 
0.95 0.03
C=
0.05 0.97

59
5.2. Catastrophic Round-Off Errors

then the matrix I − C is singular (where I is the identity matrix). However, if you solve
(I − C)x = 0 by
>> C = [0.95 0.03; 0.05 0.97];
>> rref([eye(size(C))-C [0 0]′ ])
MATLAB displays
ans =
1 0 0
0 1 0
which indicates that the only solution is x = 0. On the other hand, if you enter
>> C = [0.95 0.03; 0.05 0.97]; b = 1;
>> rref([eye(size(C))-C [b 0]′ ])
then MATLAB realizes that I − C is singular. Clearly there is some value of b between 0
and 1 where MATLAB switches between believing that I − C is non-singular and singular.†

Solving Linear Systems

linsolve(A, b, <properties>) Solve the linear system of equations Ax = b where A has certain
properties.
rref Calculates the reduced row echelon form of a matrix or an aug-
mented matrix.

5.2. Catastrophic Round-Off Errors

We have mentioned repeatedly that omputers annot add, subtra t, multiply, or divide or-
re tly! Up until now, the errors that have resulted have been very small. Now we present two examples
where the errors are very large.
In this first example, the reason for the large errors is easy to understand. Consider the matrix
 
1 2 3
Aǫ =  4 5 6 ,
7 8 9+ǫ

which is singular when ǫ = 0 and nonsingular otherwise. But how well does MATLAB do when ǫ ≪ 1?
Enter
>> eps = input(′eps = ′ ); A = [1 2 3; 4 5 6; 7 8 9+eps]; inv(A)*A - eye(size(A))
so that the final matrix should be O. Begin by letting ǫ = 0 and observe that the result displayed is
nowhere close to the zero matrix! However, note that MATLAB is warning you that it thinks something is
wrong with the statement


To understand this “switch”, look at the actual coding of rref. It uses the variable tol to determine
whether an element of the augmented matrix
„ «
0.05 −0.03 b1
−0.05 0.03 b2
is “small enough” that it should be set to 0. tol is (essentially) calculated by
tol = max(size(<augmented matrix>)) * eps * norm(<augmented matrix>, inf);
The maximum of the number of rows and columns of the augmented matrix, i.e., max(size(...)), is multiplied
by eps and this is multiplied by the “size” of the augmented matrix. ( norm in section 7.) Since b is the last
column of the augmented matrix, the “size” of this matrix depends on the size of the elements of b. Thus, the
determination whether a number “should” be set to 0 depends on the magnitude of the elements of b.
You can obtain the correct answer to the homogeneous equation by entering
>> rref([eye(size(C))-C [0 0]′ ], eps)
which decreases the tolerance to eps.

60
5.3. Overdetermined and Underdetermined Linear Systems

Warning: Matrix is close to singular or badly scaled.


Results may be inaccurate. RCOND = 1.541976e-18.
( RCOND is its estimate of the inverse of the condition number. See cond in section 7 for more details.)
Now choose some small nonzero values for ǫ and see what happens. How small can ǫ be before MAT-
LAB warns you that the matrix is “close to singular or badly scaled”? In this example, you know that the
matrix is “close to singular” if ǫ is small (but nonzero) even if MATLAB does not. The next example is
more interesting.
For the second example, consider the Hilbert matrix of order n, i.e.,

1 1/2 1/3 . . . 1/n

 1/2 1/3 1/4 . . . 1/(n + 1) 

Hn =  1/3 1/4 1/5 . . . 
1/(n + 2)  ,
 .. .. .. . . .. 
. . . . .
1/n 1/(n + 1) 1/(n + 2) . . . 1/(2n − 1)

which is generated in MATLAB by


>> H = hilb(n)
There does not seem to be anything particularly interesting, or strange, about this matrix; after all,
hij = 1/(i + j − 1) so the elements are all of “reasonable” size. If you type
>> n = 10; H = hilb(n); ( H (1/2) ) 2 - H
ˆ ˆ
the result is not particularly surprising. The resulting matrix should be the zero matrix, but, because of
round-off errors, it is not. However, every element is in magnitude less than 10−15 , so everything looks
fine.
However, suppose you solve the matrix equation

Hx = b

for a given b. How close is the numerical solution to the exact solution? Of course, the problem is: how
can you know what the analytical solution is for a given b? The answer is to begin with x and calcu-
late b by b = Hx. Then solve Hx = b for x and compare the final and initial values of x. Do this in
MATLAB by
>> n = 10; x = rand(n, 1); b = H*x; xnum = H\b
and compare x with xnum by calculating their difference, i.e.,
>> x - xnum
The result is not very satisfactory: the maximum difference in the elements of the two vectors is usually
somewhere between 10−5 and 10−3 . That is, even though all the calculations have been done to approx-
imately 16 significant digits, the result is only accurate to three to ve significant digits! (To see how
much worse the result can be, repeat the above commands for n = 12.)
It is important to realize that most calculations in MATLAB are very accurate. It is not that solving
a matrix equation necessarily introduces lots of round-off errors; instead, Hilbert matrices are very “un-
stable” matrices — working with them can lead to inaccurate results. On the other hand, most matrices
are quite “stable”. For example, if you repeat the above sequence of steps with a random matrix, you find
that the results are quite accurate. For example, enter
>> n = 1000; R = rand(n); x = rand(n, 1); b = R*x; xnum = R\b; max(abs(x - xnum))
The results are much more reassuring, even though n is 100 times as large for this random matrix as for
the Hilbert matrix — and even though there are over 600,000 times as many floating point operations
needed to calculate x by Gaussian elimination for this random matrix!
Note: By entering all the commands on one line, it is easy to repeat this experiment many times for differ-
ent random numbers by simply rerunning this one line.

5.3. Overdetermined and Underdetermined Linear Systems


If A ∈ Cm×n where m > n, Ax = b is called an overdetermined system because there are more equa-
tions than unknowns. In general, there are no solutions to this linear equation. (However, to be sure use
rref.) However, you can find a “best” approximation by finding the solution for which the vector

r = Ax − b

61
5.3. Overdetermined and Underdetermined Linear Systems

which is called the residual, is smallest in Euclidean length; that is,


v
u n
uX
norm(r) ≡ t r2 i
i=1

is minimized. (The norm function is discussed in section 7.) This is called the least-squares solution. This
best approximation is calculated in MATLAB by typing
>> A\b
Analytically, the approximation can be calculated by solving

AT Ax = AT b .

However, numerically this is less accurate than the method used in MATLAB.
Note that this is the same command used to find the solution to a square linear system. This cannot be
the intent here since A is not a square matrix. Instead, MATLAB interprets this command as asking for
the least-squares solution. Again, this command only makes sense if there is a unique solution which mini-
mizes the length of the vector Ax − b. If there are an infinite number of least-squares solutions, MATLAB
warns you of this fact and then returns one of the solutions. For example, if
   
1 2 3 2
 4 5 6  1
A=  and b =  
7 8 9 2
10 11 12 4

then Ax = b has no solutions, but has an infinite number of least-square approximations. If you enter
>> A\b
the response is
Warning: Rank deficient, rank = 2 tol = 1.4594e-14.
It also returns the solution (−1/4 , 0, 29/60 )T (after using the MATLAB command rats which we discuss
below), which is one particular least-squares approximation. To find all the solutions, you can use rref
to solve AT Ax = AT b. (If A is complex, solve AH Ax = AH b.)
Occasionally, if there are an infinite number of least-squares approximations, the solution desired is the
“smallest” one, i.e., the x for which the length of the vector x is minimized. This can be calculated using
the pseudoinverse of A, denoted by A+ . Since A is not square, it cannot have an inverse. However, the
pseudoinverse is the unique n×m matrix which satisfies the Moore-Penrose conditions:
• AA+ A = A
• A+ AA+ = A+
• (AA+ )T = AA+
• (A+ A)T = A+ A
In particular, if A is a square nonsingular matrix, then A+ is precisely A−1 . This pseudoinverse is calcu-
lated in MATLAB by entering
>> pinv(A)
The reason for mentioning the pseudoinverse of A is that the least-squares approximation to Ax = b can
also be calculated by
>> pinv(A)*b
If there are an infinite number of least-squares approximations, this returns the one with the smallest
length.
Next, suppose that A ∈ Cm×n with m < n. Ax = b is called an underdetermined system because there
are less equations than unknowns. In general, there are an infinite number of solutions to this equation.
We can find these solutions by entering
>> rref([A b])
and solving the result for x. We can find one particular solution by entering
>> A\b

62
6. File Input-Output

This solution will have many of its elements being 0. We can also find the solution with the smallest
length by entering
>> pinv(A)*b
Warning: It is possible for an overdetermined system to have one or even an infinite number of solutions
(not least-squares approximations). It is also possible for an underdetermined system to have
no solutions. The way to check the number of solutions is to use the rref command.
One command which is occasionally useful is rats. If all the elements of A and b are rational num-
bers, then the solution and/or approximation obtained is usually a rational number, although stored as
a floating-point number. This command displays a “close” rational approximation to the floating-point
number, which may or may not be the exact answer. For example, entering
>> rats(1/3 - 1/17 + 1/5)
results in the text variable 121/255, which is the correct answer.
Warning: Be careful when using this command. rats(sqrt(2)) makes no sense (as was known in 500
BC).

Solving Linear Systems

A\b When Ax = b is an overdetermined system, i.e., m > n where A ∈ Cm×n , this is the
least-squares approximation; when it is an underdetermined solution, i.e., m < n, this is
a solution which has 0 in many of its elements.
pinv(A) The pseudoinverse of A.

rats(x) Calculates a “close” approximation to the floating-point number x. This is frequently


the exact value.

6. File Input-Output
In section 4.1 we discussed the csvread and csvwrite commands which allow simple input from and
output to a file. The MATLAB commands fscanf and fprintf, which behave very similarly to their C
counterparts, allow much finer control over input and output. Before using them a file has to be opened
by
>> fid = fopen(′<file name>′, <permission string>)
where the file identifier fid is a unique nonnegative integer attached to the file. (Three file identifiers al-
ways exist as in C: 0 is the standard input, 1 is the standard output, and 2 is the standard error.) The
permission string specifies how the file is to be accessed:
′ ′
r read only from the file.
′ ′
w write only to the file (anything previously contained in the file is overwritten). If necessary, the
file is created.
′ ′
a append to the end of the file (everything previously contained in the file is retained).

r+′ read from and write to the file (anything previously contained in the file is overwritten).

w+′ read from and write to the file (anything previously contained in the file is overwritten). If neces-
sary, the file is created.
If the fopen command fails, −1 is returned in the file identifier. Enter
>> fclose(fid)
if a file needs to be closed.
To write formatted data to a file, enter
>> fprintf(fid, <format string>, <variable 1>, <variable 2>, ...)
The elements contained in the variables are written to the file specified in a previous fopen command ac-
cording to the format string. If fid is omitted, the output appears on the screen. The format string is
very similar to that of C, with the exception that the format string is cycled through until the end of the
file is reached or the number of elements specified by size is attained.
To briefly review some of the C format specifications, the conversion characters are:

63
6. File Input-Output

d – The argument is converted to decimal notation.


c – The argument is a single character.
s – The argument is a string.
e – The argument is a floating-point number in “E” format.
f – The argument is a floating-point number in decimal notation.
g – The argument is a floating-point number in either “E” or decimal notation.
Each conversion character is preceded by “%”. The following may appear between the “%” and the con-
version character:
• A minus sign which specifies left adjustment rather than right adjustment.
• An integer which specifies a minimum field width.
• If the maximum field width is larger than the minimum field width, the minimum field width is pre-
ceded by an integer which specifies the maximum field width, and the two integers are separated by a
period.
fprintf can also be used to format data on the screen by omitting the fid at the beginning of the
argument list. Thus, it is possible to display a variable using as little or as much control as desired. For
example, if x contains −23.6 three different ways to display it are
>> x
>> disp([′x = ′ , num2str(x)])
>> fprintf(′%12.6e\n′ , x)
and the results are

x =

-23.6000

x = -23.6000
-2.360000e+01
Note: It is easy to print the matrix A in the MATLAB workspace as we just described. However, it is
a little more difficult to print it to a file. The following works and can be entered on one line, al-
though it is actually a number of statements.
>> Str = num2str(A); for i = [1:size(Str, 1)] fprintf(fid, ′ %s\n′ , Str(i,:));
end
To read formatted data from a file, enter
>> A = fscanf(fid, <format string>, <size>)
The data is read from the file specified in a previous fopen command according to the format string and
put into the matrix A. The size argument, which puts an upper limit on the amount of data to be read, is
optional. If it is a scalar, or is not used at all, A is actually a vector. If it is [m n], then A is a matrix of
this size.

Advanced Input-Output

fopen(′<file name>′, Opens the file with the permission string determining how the
<permission string>) file is to be accessed. The function returns the file identifier,
which is a unique nonnegative integer attached to the file.
fclose(fid) Closes the file with the given file identifier.
fscanf(fid, <format string>) Behaves very similarly to the C command in reading data from
a file using any desired format.
fprintf(fid, <format string>, Behaves very similarly to the C command in writing data to a
<variable 1>,...) file using any desired format.
fprintf(<format string>, Behaves very similarly to the C command in displaying data on
<variable 1>,...) the screen using any desired format.

64
7. Some Useful Linear Algebra Commands

7. Some Useful Linear Algebra Commands


We briefly describe in alphabetical order some of the MATLAB commands that are most useful in lin-
ear algebra. Most of these discussions can be read independently of the others. Where this is not true, we
indicate which should be read first.
chol
Let A ∈ Rn×n be symmetric and positive definite† . Then there exists an upper triangular matrix R
such that RT R = A. R is calculated by
>> R = chol(A)
If A is not positive definite, an error message is printed. (If A ∈ Cn×n then RH R = A.)
cond
Note: Read the discussion on norm below first.
The condition number of A ∈ Cn×n , which is denoted by cond(A), is a positive real number which is
always ≥ 1. It measures how “stable” A is: if cond(A) = ∞ the matrix is singular, while if cond(A) = 1
the matrix is as nice a matrix as you could hope for — in particular, cond(I) = 1. To estimate the num-
ber of digits of accuracy you might lose in solving the linear system Ax = b, enter
log10(cond(A))
In subsection 5.2 we discussed the number of digits of accuracy you might lose in solving Hx = b where
H is the Hilbert matrix of order 10. In doing many calculations it was clear that the solution was only ac-
curate to 3 to 5 significant digits. Since cond(H) is 1.6×1013 , it is clear that you should lose about 13 of
the 16 digits of accuracy in this calculation. Thus, everything fits.
If A is nonsingular, the condition number is defined by

condp (A) = kAkp kA−1 kp for p ∈ [1, ∞]


or
condF (A) = kAkF kA−1 kF .

It is calculated in MATLAB by
>> cond(A, p)
where p is 1, 2, Inf, or ′ fro′ . If p = 2 the command can be shortened to
>> cond(A)
Note that the calculation of the condition number of A requires the calculation of the inverse of A.
The MATLAB command condest approximates the condition number without having to calculate this
inverse. See the discussion of this command below for further information on when it might be preferable.
Note: Sometimes we want to solve, or find the “best” approximation to, Ax = b when A ∈ Cm×n is not
a square matrix. (This is discussed in detail in subsection 5.3.) Since we still want to know the ac-
curacy of any solution, we want to generalize the condition number to nonsquare matrices. This is
done by defining the condition number of a nonsquare matrix in the 2-norm to be the ratio of the
largest to the smallest singular value of A, i.e., σ1 /σmin{m,n} .
condest
Note: Read the discussion on cond above first.
The calculation of the condition number of A ∈ Cn×n requires the calculation of its inverse. There are
two reasons this might be inadvisable.
• The calculation of A−1 requires approximately 2n3 flops, which might take too long if n is very
large.
• If A is a sparse matrix (i.e., most of its elements are zero), we discuss in section 9 how to store only
the nonzero elements of A to conserve storage. (For example, if n = 10,000 and A is tridiagonal‡,
the number of nonzero elements in A is approximately 30,000 but the total number of elements in A

A ∈ Rn×n is positive definite if xT Ax ≥ 0 for all x ∈ Rn and xT Ax = 0 only if x = 0. In practical terms, it
means that all the eigenvalues of A are positive. ( A ∈ Cn×n is positive definite if xH Ax ≥ 0 for all x ∈ Cn and
xT Ax = 0 only if x = 0.)

A matrix is tridiagonal if its only nonzero elements occur on the main diagonal or on the first diagonal above
or below the main diagonal

65
7. Some Useful Linear Algebra Commands

is 100,000,000.) Since the inverse of a sparse matrix is generally much less sparse (in fact it may have
no zero elements at all), MATLAB may not be able to store A−1 .
The command condest calculates a lower bound to the condition number of a matrix in the 1-norm with-
out having to determine its inverse. This approximation is almost always within a factor of ten of the
exact value.
When MATLAB calculates A\b or inv(A), it also calculates condest(A). It checks if its estimate
of the condition number is large enough that A is likely to be singular. If so, it returns an error message
such as
Warning: Matrix is close to singular or badly scaled.
Results may be inaccurate. RCOND = 2.055969e-18.
where RCOND is the inverse of condest(A).
det
Let A ∈ Cn×n . The determinant of A is calculated by
>> det(A)
det(A) = 0 if and only if A is singular. However, due to round-off errors it is very unlikely that you will
obtain 0 numerically unless all the entries to A are integers. For example, consider the matrix
 
0.95 0.03
C= .
0.05 0.97

I − C is singular (where I is the identity matrix) but


>> C = [0.95 0.03; 0.05 0.97]; det( eye(size(C)) - C )
does not return 0. However, the number it returns is much smaller than eps and so it seems “reason-
able” that I − C is singular. On the other hand,
>> det(hilb(10))
returns 2.2×10−53 , but the Hilbert matrix is not singular for any n. (The singular value decomposition,
which is described below, is a much better method for determining if a square matrix is singular.)
eig
Let A ∈ Cn×n . A scalar λ ∈ C is an eigenvalue of A if there exists a nonzero vector v ∈ Cn such that

Av = λv ;

v is called the eigenvector corresponding to λ. There are always n eigenvalues of A, although they need
not all be distinct. MATLAB will very happily calculate all the eigenvalues of A by
>> eig(A)
It will also calculate all the eigenvectors by
>> [V, D] = eig(A)
D ∈ Cn×n is a diagonal matrix containing the n eigenvalues on its diagonal and the corresponding eigen-
vectors are found in the same columns of the matrix V ∈ Cn×n .
A matrix is defective if it has less eigenvectors than eigenvalues. MATLAB normally cannot determine
when this occurs. For example, the matrix
 
1 1
B=
0 1

is defective since it has two eigenvalues, both of which are 1, but it only has one eigenvector, namely
(1, 0)T . If you enter
>> B = [1 1; 0 1]; [V, D] = eig(B)
MATLAB calculates the two eigenvalues correctly, but it finds the two eigenvectors (1, 0)T and
T
(−1, 2.2×10−16 ) . Clearly the latter eigenvector should be (−1, 0)T so that, in fact, there is only one
eigenvector.

66
7. Some Useful Linear Algebra Commands

Note: If A is a sparse matrix (see Section 9), you cannot use eig. You either have to use the function
eigs or do eig(full(A)).
eigs
Note: Read the discussion on eig above first.
Frequently, you do not need all the eigenvalues of a matrix. For example, you might only need the
largest ten in magnitude, or the five with the largest real part, or the one which is smallest in magnitude,
or . . . In addition, you might need some eigenvalues of the generalized eigenvalue problem

Ax = λBx

where B is a symmetric positive definite matrix. (If B is complex, it must be Hermetian.) eigs can do
all of this. Of course, this means that there are numerous possible arguments to this function so read the
documentation carefully.
Why not just use eig anyway? Calculating all the eigenvalues of A ∈ Rn×n requires (very) approx-
imately 10n3 flops, which can take a very long time if n is very large. On the other hand, calculating
only a few eigenvalues requires many, many fewer flops. If A is a full matrix, it requires cn2 flops where
c is of “reasonable” size; if A is a sparse matrix (see Section 9), it requires cn flops.
Note: If A is sparse, you cannot use eig — you will first have to do eig(full(A)).
Also, this command generates lots of diagnostic output. To calculate the largest 3 eigenvalues of A in
magnitude without generating any diagnostics, enter
>> op.disp = 0
>> eigs(A, 3, ′ LM′ , op)
( op.disp is a structure, which was discussed in section 3.4.)
inv
To calculate the inverse of the square matrix A ∈ Cn×n enter
>> inv(A)
The inverse of A, denoted by A−1 , is a matrix such that AA−1 = A−1 A = I, where I ∈ Rn×n is the iden-
tity matrix. If such a matrix exists, it must be unique.
MATLAB cannot always tell whether this matrix does, in fact, exist. For example, the matrix
 
1 2
A=
2 4

does not have an inverse. If you try to take the inverse of this matrix, MATLAB will complain that
Warning: Matrix is singular to working precision.
It will display the inverse matrix, but all the entries will be Inf.
The above matrix was very simple. The matrix
 
1 2 3
A = 4 5 6 (7.1)
7 8 9

also does not have an inverse. If you ask MATLAB to calculate the inverse of A, it will complain that
Warning: Matrix is close to singular or badly scaled.
Results may be inaccurate. RCOND = 2.055969e-18.
( RCOND is the inverse of a numerical approximation to the condition number of A; see condest above.)
That is, MATLAB is not positive that A is singular, because of round-off errors, but it thinks it is likely.
However, MATLAB still does try to calculate the inverse. Of course, if you multiply this matrix by A
the result is nowhere close to I. (Try it!) In other words, be careful — and read (and understand) all

67
7. Some Useful Linear Algebra Commands

warning messages.
lu
Let A ∈ Cn×n . Then there exists an upper triangular matrix U, a unit lower triangular matrix L † , and
a permutation matrix P ‡ such that
LU = PA .
The MATLAB command lu calculates these matrices by entering
>> [L, U, P] = lu(A)
If A is invertible, all the elements of U on the main diagonal are nonzero. If you enter
>> A = [1 2 3; 4 5 6; 7 8 9]; [L, U, P] = lu(A)
where A is the singular matrix defined earlier, u33 should be zero. Entering
>> U(3,3)
displays 1.1102e-16, which clearly should be zero as we discussed in subsection 1.5.
Note: This is the first time we have had a function return more than one argument. We discuss this no-
tation in detail in section 8.3. For now, we simply state that when [V, D] occurs on the right side
of the equal sign it means the matrix whose first columns come from V and whose last columns
come from D. However, on the left side of the equal sign it means that the function returns two
arguments where the first is stored in the variable V and the second in D.
norm
The norm of a vector or matrix is a nonnegative real number which gives some measure of the “size” of
the vector or matrix. The p-th norm of a vector is defined by
 !1/p
 Xn


 |xi |p if p ∈ [1, ∞)
kxkp = i=1



 max |x | if p = ∞ .
i
1≤i≤n

For p = 1, 2, or ∞ it is calculated in MATLAB by entering


>> norm(x, p)
where p is 1, 2, or Inf. If p = 2 the command can be shortened to
>> norm(x)
The p-th norm of a matrix is defined by
kAxkp
kAkp = max for p ∈ [1, ∞]
x6=0 kxkp

and is calculated in MATLAB by entering


>> norm(A, p)
where again p is 1, 2, or Inf. If p = 2 the command can be shortened to
>> norm(A)
There is another matrix norm, the Frobenius norm, which is defined for A ∈ Cm×n by
 1/2
m X
X n
kAkF =  |aij |2 
i=1 j=1

and is calculated in MATLAB by entering


>> norm(A, ′ fro′ )
null
Let A ∈ Cn×n . We can calculate an orthonormal basis for the null space of A by

A unit lower triangular matrix is lower triangular and, in addition, all the elements on the main diagonal
are 1.

P is a permutation matrix if its columns are a rearrangement of the columns of I.

68
7. Some Useful Linear Algebra Commands

>> null(A)
orth
Let A ∈ Cn×n . We can calculate an orthonormal basis for the columns of A by
>> orth(A)
qr
Let A ∈ Rm×n . Then there exists an orthogonal matrix Q ∈ Rm×m † and an upper triangular matrix
R ∈ Rm×n such that
A = QR .
(If A ∈ Cm×n then there exists an unitary matrix Q ∈ Cm×m and an upper triangular matrix R ∈ Cm×n
such that A = QR.) We calculate Q and R in MATLAB by entering
>> [Q, R] = qr(A)
It is frequently preferable to add the requirement that the diagonal elements of R be decreasing in magni-
tude, i.e., |ri+1,i+1 | ≤ |ri,k | for all i. In this case

AE = QR

for some permutation matrix E and


>> [Q, R, E] = qr(A)
One reason for this additional requirment on R is that you can immediately obtain an orthonormal ba-
sis for the range of A and the null space of AT . If rk,k is the last nonzero diagonal element of R, then
the first k columns of Q are an orthonormal basis for the range of A and the final n−k columns are
an orthonormal basis for the null space of AT . The command orth is preferable if all you want is an
orthonormal basis for R(A).
rank
Let A ∈ Cm×n . The rank of A is the number of linearly independent columns of A and is calculated by
>> rank(A)
This number is calculated by using the singular value decomposition, which we discuss below.
svd
Let A ∈ Rm×n . A can be decomposed into

A = UΣVT

where U ∈ Rm×m and V ∈ Rn×n are orthogonal matrices and Σ ∈ Rm×n is a diagonal matrix (although
not necessarily square) with real nonnegative elements in decreasing order. That is,

σ1 ≥ σ2 ≥ . . . ≥ σmin{m,n} ≥ 0 .

(If A ∈ Cm×n then U ∈ Cm×m and V ∈ Cn×n are unitary matrices and Σ ∈ Rm×n is a diagonal matrix
with real nonnegative elements in decreasing order.) These matrices are calculated by
>> [U, S, V] = svd(A)
The diagonal elements of Σ are called the singular values of A. Although A need not be a square ma-
trix, both AT A ∈ Rn×n and AAT ∈ Rm×m are square symmetric matrices. (If A is complex, AH A and
AAH are both square Hermitian matrices.) Thus, their eigenvalues are nonnegative.‡ Their nonzero eigen-
values are the squares of the singular values of A.§ In addition, the eigenvectors of AT A are the columns of


Q ∈ Rm×m is orthogonal if Q−1 = QT . ( Q ∈ Cm×m is unitary if Q−1 = QH .)

The eigenvalues of a real square symmetric matrix are nonnegative. (The eigenvalues of a complex square
Hermitian matrix are real and nonnegative.)
§
For example, if m > n there are n singular values and their squares are the eigenvalues of AT A. The m
eigenvalues of AAT consist of the squares of these n singular values and m−n additional zero eigenvalues.

69
7. Some Useful Linear Algebra Commands

V and those of AAT are the columns of U. (If A is complex, the eigenvectors of AH A are the columns of
V and those of AAH are the columns of U.)
The best numerical method to determine the rank of A is to use its singular values. For example, to see
that  
1 2 3 4
A = 5 6 7 8 
9 10 11 12
has rank 2, use the svd command to find that the singular values of A are 25.4368, 1.7226,
and 8.1079×10−16 . Clearly the third singular value should be 0 and so A has 2 nonzero singular values
and so has a rank of 2. On the other hand, the Hilbert matrix of order 15 has singular values

1.8×100 , 4.3×10−1 , 5.7×10−2 , 5.6×10−3 , 4.3×10−4 , 2.7×10−5 , 1.3×10−6 , 5.5×10−8 ,


1.8×10−9 , 4.7×10−11 , 9.3×10−13 , 1.4×10−14 , 1.4×10−16 , 1.2×10−17 , and 2.4×10−18

according to MATLAB. Following Principle 1.2, you can see there is no separation between the singular
values which are clearly not zero and the ones which are “close to” eps. Thus, you cannot conclude that
any of these singular values should be set to 0. Our “best guess” is that the rank of this matrix is 15.†

Some Useful Functions in Linear Algebra

chol(A) Calculates the Cholesky decomposition of a symmetric, positive definite square


matrix.
cond(A) Calculates the condition number of a square matrix.
cond(A, p) calculates the condition number in the p-norm.
condest(A) Calculates a lower bound to the condition number of A in the 1-norm.
det(A) Calculates the determinant of a square matrix.
eig(A) Calculates the eigenvalues, and eigenvectors if desired, of a square matrix.
eigs Calculates some eigenvalues, and eigenvectors if desired, of a square matrix. There
are numerous possible arguments to this function so read the documentation
carefully.
inv(A) Calculates the inverse of a square invertible matrix.
lu(A) Calculates the LU decomposition of a square invertible matrix.
norm(v) Calculates the norm of a vector.
norm(v, p) calculates the p-norm.
norm(A) Calculates the norm of a matrix.
norm(A, p) calculates the p-norm.
null(A) Calculates an orthonormal basis for the null space of a matrix.
orth(A) Calculates an orthonormal basis for the range of a matrix.
qr(A) Calculates the QR decomposition of a matrix.
rank(A) Estimates the rank of a matrix.
svd(A) Calculates the singular value decomposition of a matrix.


In fact, it can be proven that the Hilbert matrix of order n is nonsingular for all n, and so its rank is
truly n. However, if you enter
>> rank( hilb(15) )
you obtain 12, so that MATLAB is off by three.

70
8.1. Flow Control and Logical Variables

8. Programming in MATLAB
Using the commands we have already discussed, MATLAB can do very complicated matrix operations.
However, sometimes there is a need for finer control over the elements of matrices and the ability to test,
and branch on, logical conditions. Although prior familiarity with a high-level programming language is
useful, MATLAB’s programming language is so simple that it can be learned quite easily and quickly.

8.1. Flow Control and Logical Variables

MATLAB has four flow control and/or branching instructions: for loops, while loops, if-else
branching tests, and switch branching tests.
Notation: All of these instructions end with an end statement, and it is frequently difficult to determine
the extent of these instructions. Thus, it is very important to use indentation to indicate the
structure of a code, as we do in the remainder of this tutorial. This greatly increases the read-
ability of the code for human beings.
The general form of the for loop is
>> for <variable> = <expression>
<statement>
...
<statement>
end
where the variable is often called the index of the loop. The elements of the row vector <expression> are
stored one at a time in the variable and then the statements up to the end statement are executed.† For
example, you can define the vector x ∈ Rn where xi = i sin(i2 π/n) by
>> x = zeros(n, 1);
>> for i = 1:n
x(i) = i * sin( i 2 *pi/n );
end
ˆ
(The first line is not actually needed, but it allows MATLAB to know exactly the size of the final vector
before the for loops begin. This saves computational time and makes the code more understandable; it is
discussed in more detail in section 8.5.) In fact, the entire for loop could have been entered on one line
as
>> for i = 1:n x(i) = i * sin( i 2 *pi/n ); end
ˆ
However, for readability it is best to split it up and to indent the statements inside the loop. Of course,
you can also generate the vector by
>> x = [1:n]′ .* sin( [1:n]′ . 2 *pi/n )
ˆ
which is certainly “cleaner” and executes much faster in MATLAB. √
Warning: In using i as the index of the for loop, i has just been redefined to be n instead of −1 .
Caveat emptor!
A more practical example of the use of a for loop is the generation of the Hilbert matrix of order n,
which we have already discussed a number of times. This is easily done using two for loops by
>> H = zeros(n);
>> for i = 1:n
for j = 1:n
H(i,j) = 1/(i + j - 1);
end
end
Warning: In using i √and j as the indices of the for loops, i and j have just been redefined to be n
instead of −1 . Caveat emptor!


<expression> can be a matrix in which case each column vector is stored one at a time in i.

71
8.1. Flow Control and Logical Variables

for loops often have branches in them. For this we need the if branch, which we now describe. The
simplest form of the if statement is
>> if <logical expression>
<statement>
...
<statement>
end
where the statements are evaluated as long as the <logical expression> is true. The
<logical expression> is generally of the form
<arithmetic expression-left> rop <arithmetic expression-right>
where rop is one of the r elational operators shown below. Some examples of logical expressions are
i == 5
x(i) >= i
imag(A(i,i)) = 0
˜
sin(1) - 1 > x(1) + x(i) 3
ˆ
Warning: String variables cannot be easily compared by == or =.† Instead, if a and b are text vari-
ables, enter ˜
>> strcmp(a, b)
The result is true if the two character strings are identical and false otherwise.

Relational Operators

< Less than. > Greater than.


<= Less than or equal to. >= Greater than or equal to.
== Equal. = Not equal to.
˜ strcmp(a, b) Compares strings.

A second form of the if statement is


>> if <logical expression>
<statement group 1>
else
<statement group 2>
end
where statement group 1 is evaluated if the <logical expression> is true and statement group 2 is
evaluated if it is false. The final form of the if statement is
>> if <logical expression 1>
<statement group 1>
elseif <logical expression 2>
<statement group 2>
elseif <logical expression 3>
<statement group 3>
...
elseif <logical expression r>
<statement group r>
else
<statement group r+1>
end
where statement group 1 is evaluated if the <logical expression 1> is true, statement group 2 is evalu-

Compare the results of
>> ′ Yes′ == ′ yes′
and
>> ′ Yes′ == ′ no′

72
8.1. Flow Control and Logical Variables

ated if the <logical expression 2> is true, etc. The final else statement is not required. If it occurs
and if none of the logical expressions is true, statement group r+1 is evaluated. If it does not occur and if
none of the logical expressions is true, then none of the statement groups are executed.
When a logical expression such as
>> i == 5
is evaluated, the result is either the logical value “TRUE” or “FALSE”. MATLAB calculates this as a nu-
merical value which is returned in the variable ans. The value is 0 if the expression is false and 1 if it is
true.
MATLAB also contains the logical operators “AND” (denoted by “&”), “OR” (denoted by “ |”),
“NOT” (denoted by “ ”), and “EXCLUSIVE OR” (invoked by the function xor ). These act on false
˜
or true statements which are represented by numerical values: zero for false statements and nonzero for
true statements. Thus, if a and b are real numbers then
• the relational equation
>> c = a & b
means that c is true (i.e., 1 ) only if both a and b are true (i.e., nonzero); otherwise c is false
(i.e., 0 ).
• the relational equation
>> c = a | b
means that c is true (i.e., 1 ) if a and/or b is true (i.e., nonzero); otherwise c is false (i.e., 0 ).
• the relational equation
>> c = a
˜
means that c is true (i.e., 1 ) if a is false (i.e., 0 ); otherwise c is false (i.e., 0 ).
• the relational command
>> c = xor(a, b)
means that c is true (i.e., 1 ) if exactly one of a and b is true (i.e., nonzero); otherwise c is false
(i.e., 0 ).
In the above statements c is a logical variable which has the logical value “TRUE” or “FALSE”. Fre-
quently — but not always — the variable can be set by c = 1 or c = 0; but c is now not a logical
variable, but a numerical variable. Frequently — but not always — a numerical variable can be used in-
stead of a logical variable. The preferred ways to set a logical variable are the following. The logical vari-
able can be set by c = logical(1) or c = logical(0) — and now c is a logical variable. A simpler
way to set the logical variable c is c = true or c = false.
There are second logical operators “AND” (denoted by && ) and “OR” ( || ) which are rarely needed.
The statement
>> c = a && b
returns a (scalar) logical true if both inputs evaluate to true (so if they are variables they both must be
scalars). The difference from & is that if a is false then b is not evaluated. Similarly,
>> c = a || b
returns a (scalar) logical true if either input evaluates to true. If a is true then b is not evaluated
(as in C, C++, and Java).

Logical Operators

A&B AND. a && b Short-circuit AND. Returns logical 1


A|B OR. (true) or 0 (false). Only evaluates b if
A NOT. a is true.
˜xor(A, B) EXCLUSIVE OR. a || b Short-circuit OR. Returns logical 1
(true) or 0 (false). Only evaluates b if
a is false.

73
8.1. Flow Control and Logical Variables

The second MATLAB loop structure is the while statement. The general form of the while loop is
>> while <logical expression>
<statement>
...
<statement>
end
where the statements are executed repeatedly as long as the <logical expression> is true. For exam-
ple, eps can be calculated by
>> ep = 1;
>> while 1 + ep > 1
ep = ep/2;
end
>> eps = 2*ep
It is possible to break out of a for loop or a while loop from inside the loop by using the break
command as in C. This terminates the execution of the innermost for loop or while loop.
The continue statement is related to break. It causes the next iteration of the for or while loop
to begin immediately.
The switch command executes particular statements based on the value of a variable or an expression.
Its general form is
>> switch <variable or expression>
case <Value 1>,
<statement group 1>
case {<Value 2a>, <Value 2b>, <Value 2c>, ..., <Value 2m>},
<statement group 2>
...
case <value n>,
<statement group r>
otherwise,
<statement group r+1>
end
where statement group 1 is evaluated if the variable or expression has <Value 1>, where statement group
2 is evaluated if the variable or expression has values <Value 2a> or <Value 2b> or <Value 2c>, etc.
(Note that if a case has more than one value, then all the values must be surrounded by curly brackets.)
The final otherwise is not required. If it occurs and if none of the values match the variable or expres-
sion, then statement group r+1 is evaluated. If it does not occur and if none of the values match, then
none of the statement groups are executed.
Warning: The switch command is different in MATLAB than in C in two ways:
First, in MATLAB the case statement can contain more than one value; in C it can only con-
tain one.
And, second, in MATLAB only the statements between the selected case and the following one
or the following otherwise or end (whichever occurs first) are executed; in C all the state-
ments following the selected case are executed up to the next break or the end of the block.

74
8.2. Matrix Relational Operators and Logical Operators

Flow Control

break Terminates execution of a for or while loop.


case Part of the switch command. The statements following it are executed if its value or
values are a match for the switch expression.
continue Begins the next iteration of a for or while loop immediately.
else Used with the if statement.
elseif Used with the if statement.
end Terminates the for, if, switch, and while statements.
for Repeats statements a specific number of times.
if Executes statements if certain conditions are met.
otherwise Part of the switch command. The statements following it are executed if no case
value is a match for the switch expression.
switch Selects certain statements based on the value of the switch expression.
while Repeats statements as long as an expression is true.

Elementary Logical Matrices

true Generates a logical matrix with all elements having the logical value true.
Use true or true(n) or true(m, n).
false Generates a logical matrix with all elements having the logical value false.
Use false or false(n) or false(m, n).

8.2. Matrix Relational Operators and Logical Operators


Although MATLAB does have a quite powerful programming language, it is needed much less fre-
quently than in typical high-level languages. Many of the operations and functions that can only be
applied to scalar quantities in other languages can be applied to vector and matrices in MATLAB. For
example, MATLAB’s relational and logical operators can also be applied to vectors and matrices. In this
way, algorithms that would normally require flow control for coding in most programming languages can
be coded using simple MATLAB commands.
If A, B ∈ Rm×n then the relational equation
>> C = A rop B
is evaluated as cij = aij rop bij , where rop is one of the relational operators defined previously. C is a
logical array, that is, its data type is “logical” not “numeric”. The elements of C are all 0 or 1: 0 if
aij rop bij is a false statement and 1 if it is a true one. Also, the relational equation
>> C = A rop c
is defined when c is a scalar. It is evaluated as if we had entered
>> C = A rop c*ones(size(A))
Similar behavior holds for logical operators:
>> C = A & B
means cij = aij & bij ,
>> C = A | B
means cij = aij | bij ,
>> C = A
˜
means cij = aij , and
˜
>> C = xor(A, B)
means cij = xor(aij , bij ). Again the elements of C are all 0 or 1.
To show the power of these MATLAB commands, suppose we have entered
>> F = rand(m, n)

75
8.2. Matrix Relational Operators and Logical Operators

and now we want to know how many elements of F are greater than 0.5. We can code this as
>> nr elements = 0;
>> for i = 1:m
for j = 1:n
if F(i,j) > 0.5
nr elements = nr elements + 1;
end
end
end
>> nr elements
However, it can be coded much more simply, quickly, and efficiently since the relational expression
>> C = F > 0.5;
or, to make the meaning clearer,
>> C = (F > 0.5);
generates the matrix C where n
1 if fij > 0.5
cij =
0 otherwise.
Since the number of ones is the result we want, simply enter
>> sum( sum( F > 0.5 ) )
or
>> sum(sum(C))
or
>> sum(C(:))
And suppose we want to replace all the elements of F which are ≤ 0.5 by zero. This is easily done by
>> F = F.*(F > 0.5)
The relational expression F > 0.5 generates a matrix with zeroes in all the locations where we want to
zero the elements of F and ones otherwise. Multiplying this new matrix elementwise with F zeroes out all
the desired elements of F. We can also replace all the elements of F which are ≤ 0.5 by −π using
>> C = (F > 0.5)
>> F = F.*C - pi*( C)
˜
Shortly we will present two easier ways to do this.
There is even a MATLAB function which determines the location of the elements of a vector or a ma-
trix where some property is satisfied. The command
>> find(x)
generates a column vector containing the indices of x which are nonzero. (Recall that nonzero can also
mean “TRUE” so that this command finds the elements where some condition is true.) For example, if
x = (0, 4, 0, 1, −1, 0, π)T then the resulting vector is (2, 4, 5, 7)T . We can add 10 to every nonzero ele-
ment of x by
>> ix = find(x);
>> x(ix) = x(ix) + 10;
Note: If no element of the vector is nonzero, the result is the empty matrix [].
find can also be applied to a matrix. The command
>> find(A)
first transforms A to a column vector (i.e., A(:) ) and then determines the locations of the nonzero ele-
ments. Instead we can work with the matrix directly by entering
>> [iA, jA] = find(A)
The two column vectors iA and jA contain the rows and columns, respectively, of the nonzero elements.
We can also find the locations of the nonzero elements and their values by
>> [iA, jA, valueA] = find(A)
As a simple example of the power of this command we can add 10 to every nonzero element of A by
>> ijA = find(A); A(ijA) = A(ijA) + 10
Note: ijA contains the locations of the nonzero elements of A when considered to be a column vector.
Since A(k) has no meaning in linear algebra if k is a scalar (since an element of A requires both

76
8.2. Matrix Relational Operators and Logical Operators

a row and a column number), MATLAB assumes that this is the element number of A as a column
vector.
We can also find the elements of a vector or a matrix which satisfy a more general property than being
nonzero. For example, to find the locations of all the elements of x which are greater than 5 enter
>> find(x > 5)
and to find the locations of all the elements of x which are greater than 5 and less than 8 enter
>> find( (x > 5) & (x < 8) )
We can find the number of elements which satisfy this last property by entering
>> length( find( (x > 5) & (x < 8) ) )
Previously, we showed how to replace all the elements of F which are ≤ 0.5 by −π. A method which
does not require any multiplication is
>> ijF = find(F <= 0.5);
>> F(ijF) = -pi
or even
>> F( find(F <= 0.5) ) = -pi
The “beauty” of MATLAB commands such as these is they are so easy to use and to understand (once
you get the hang of it) and they require so few keystrokes.
Another, slightly different method uses the matrix
>> D = (F <= 0.5)
rather than the vector ijF. Recall that ijF is a vector which contains the actual locations of the el-
ements we want to zero out, whereas D is a matrix of ones and zeroes which explicitly shows which el-
ements should be zeroed. We can use D to determine which elements of F should be replaced by zero
by
>> F(D) = -pi
(We can even use
>> F(F <= 0.5) = -pi
to combine everything into a single statement.) This requires some explanation. The matrix D is being
used here as a “mask” to determine which elements of F should be replaced by −π: for every element of
D which is nonzero, the corresponding element of F is replaced by −π; for every element of D which is
zero, nothing is done.
How does MATLAB know that D should be used to “mask” the elements of F? The answer is that D
is a logical matrix because it was defined using a logical operator, and only logical matrices and vectors
can be used as “masks”. To see that D is a logical variable and F is not, enter
>> islogical(D)
>> islogical(F)
And to see what happens when you try to use a non-logical variable as a “mask”, enter
>> F(2*D)
We can also convert a non-logical variable to a logical one by using the MATLAB command logical.
To explain logical arrays more clearly, we take a specific and very simple example. Enter
>> v = [0:.25:1];
>> c = (v >= .5);
so that v = [0 .25 .5 .75 1.0] and c = [0 0 1 1 1] where “0” denotes false and “1” denotes true.
The result of
>> v(c)
is [.5 .75 1.0]. That is, c is a logical vector and v(c) deletes the elements of v which are “false”.
On the other hand
>> iv = find(v < .5);
returns iv = [1 2] and
>> v(iv) = [];
returns v = [.5 .75 1.0]. The difference between c and iv is that c is a logical vector and iv is a
scalar vector. If you enter
>> v([0 0 1 1 1]) % WRONG
instead of
>> v(c)

77
8.2. Matrix Relational Operators and Logical Operators

you obtain the error message


??? Subscript indices must either be real positive integers or logicals.
because [0 0 1 1 1] is a numeric vector and so must contain the numbers of the elements of v which
are desired — but there is no element “0”.
MATLAB also has two functions that test vectors and matrices for logical conditions. The command
>> any(x)
returns 1 if any element of the vector x is nonzero (i.e., “TRUE”); otherwise 0 is returned. When
applied to a matrix, it operates on each column and returns a row vector. For example, we can check
whether or not a matrix is tridiagonal by
>> any( any( triu(A, 2) + tril(A, -2) ) )
Here we check all the elements of A except those on the main diagonal and on the two adjacent ones. A
result of 1 means that at least one other element is nonzero. If we want a result of 1 to mean that A is
tridiagonal we can use
>> any( any( triu(A, 2) + tril(A, -2) ) )
˜
instead. The command
>> any(A)
operates columnwise and returns a row vector containing the result of any as applied to each column.
The complementary function all behaves the same as any except it returns 1 if all the entries are
nonzero (i.e., “TRUE”). For example, you can determine if a matrix is symmetric by
>> all( all(A == A.′ ) )
A result of 1 means that A is identical to AT .
For completeness we mention that MATLAB has a number of other functions which can check the
status of variables, the status of the elements of vectors and matrices, and even of their existence. For ex-
ample, you might want to zero out all the elements of a matrix A which are Inf or NaN. This is easily
done by
>> A( find( isfinite(A) ) ) = 0
˜
where isfinite(A) generates a matrix with 1 in each element for which the corresponding element of A
is finite. To determine if the matrix A even exists, enter
exist(′A′ )
See the table below for more details and more functions.

Logical Functions

all True if all the elements of a vector are true; operates on the columns of a matrix.
any True if any of the elements of a vector are true; operates on the columns of a
matrix.
exist(′<name>′) False if this name is not the name of a variable or a file. If it is, this function re-
turns:
1 if this is the name of a variable,
2 if this is the name of an m-file,
5 if this is the name of a built-in MATLAB function.
find The indices of a vector or matrix which are nonzero.
logical Converts a numeric variable to a logical one.
ischar True for a character variable or array.
isempty True if the matrix is empty, i.e., [].
isfinite Generates a matrix with 1 in all the elements which are finite (i.e., not Inf or
NaN ) and 0 otherwise.
isinf Generates a matrix with 1 in all the elements which are Inf and 0 otherwise.
islogical True for a logical variable or array.
isnan Generates a matrix with 1 in all the elements which are NaN and 0 otherwise.

78
8.3. Function M-files

8.3. Function M-files


We have already discussed script m-files, which are simply an easy way to collect a number of state-
ments and execute them all at once. Function m-files, on the other hand, are similar to functions or pro-
cedures or subroutines or subprograms in other programming languages. Ordinarily, variables which are
created in a function file exist only inside the file and disappear when the execution of the file is completed
— these are called local variables. Thus you do not need to understand the internal workings of a function
file; you only need to understand what the input and output arguments represent.
Note: The generic term for script files and function files is m-files, because the extension is “m”.
Unlike script files, function files must be constructed in a specific way. The first line of the file
<file name>.m must begin with the keyword function. Without this word, the file is a script file. The
complete first line, called the function definition line, is
function <out> = <function name>(<in 1>, ..., <in n>)
or
function [<out 1>, ..., <out m>] = <file name>(<in 1>, ..., <in n>)
where the name of the function must be the same as the name of the file (but without the extension). The
input arguments are <in 1>, <in 2>, . . . The output arguments must appear to the left of the equal
sign: if there is only one output argument, i.e., <out>, it appears by itself; if there is more than one, i.e.,
<out 1>, etc., they must be separated by commas and must be enclosed in square brackets.
Variables in MATLAB are stored in a part of memory called a workspace. The base workspace contains
all the variables created during the interactive MATLAB session, which includes all variables created in
script m-files which have been executed. Each function m-file contains its own function workspace which
is independent of the base workspace and every other function workspace. The only way to “connect”
these workspaces is through the arguments of a function or by using the global command (which we will
discuss shortly).
There is great flexibility in the number and type of input and output arguments; we discuss this topic in
great detail later. The only detail we want to mention now is that the input arguments are all passed “by
value” as in C. (That is, the values of the input arguments are stored in temporary variables which are
local to the function.) Thus, the input arguments can be modified in the function without affecting any
input variables in the calling statement.†
Warning: The name of the le and the name of the fun tion must agree. This is also the name of the om-
mand that exe utes the fun tion.
Comment lines should immediately follow. A comment line begins with the percent character, i.e., “%”.
All comment lines which immediately follow the function definition line constitute the documentation for
this function; these lines are called the online help entry for the function. When you type
>> help <function name>
all these lines of documentation are typed out. If you type
type <function name>
the entire file is printed out. In addition, the first line of documentation, i.e., the second line of the file,
can be searched for keywords by entering
>> lookfor <keyword>
Make sure this first comment line contains the name of the command and important keywords which de-
scribe its purpose.
Note: Comments can be placed anywhere in an m-file, including on a line following a MATLAB state-
ment. The initial comment lines in a script file and the comment lines in a function file which


If you are worried because passing arguments by value might drastically increase the execution time of the
function, we want to reassure you that this does not happen. To be precise, MATLAB does not actually pass all
the input arguments by value. Instead, an input variable is only passed by value if it is modified by the func-
tion. If an input variable is not modified, it is passed “by reference”. (That is, the input argument is the actual
variable used in the calling statement and not a local copy.) In this way you get the benefit of “call by value”
without any unnecessary overhead. And how does MATLAB know if an input argument is modified? It can only
be modified if it appears on the left-hand side of an equal sign inside the function!

79
8.3. Function M-files

immediately follow the first line are special: they appear on the screen when you type
>> help <function name>
Before discussing functions at great length, there is one technical detail it is important to consider be-
fore it trips you up: how does MATLAB find the m-files you have created? Since MATLAB contains
thousands of functions, this is not an easy task. Once MATLAB has determined that the word is not a
variable, it searches for the function in a particular order. We show the order here and then discuss the
items in detail throughout this subsection.
(1) It checks if <function name> is a built-in function (i.e., coded in C).
(2) It checks if <function name> is a function, i.e., the primary function, a subfunction, or a nested
function in the current scope, in the current file. (We discuss all these terms shortly.)
(3) It checks if the file <function name>.m exists in the current directory.
(4) It checks if the current directory has a subdirectory called “private”; if it does, MATLAB checks if
the file <function name>.m exists in this subdirectory.
(5) It searches the directories in the search path for the file <function name>.m.
Note from (3) that MATLAB searches in the current directory for the function by searching for the m-
file with the same name. If the m-file is not in the current directory, the simplest way to enable MATLAB
to find it is have the subdirectory in your search path. If you type
>> path
you will see all the directories that are searched. If you have created a subdirectory called “matlab” in
your main dire tory, this is usually the first directory searched (unless the search path has been modified).
Thus, you can put your m-files in this subdirectory and be sure that MATLAB will find them. You can
also add directories to the search path by
>> path(′ new directory′, path)
or
>> path(path, ′ new directory′)
(The former puts “new directory” at the beginning of the search path while the latter puts it at the end.)
Warning: When you begin a MATLAB session, it always checks if the subdirectory “matlab” exists in
your main directory. If you create this subdirectory after you start a MATLAB session, it will
not be in the search path.
Now we return to our discussion of creating functions. We begin with a simple example of a function file
which constructs the Hilbert matrix (which we have already used a number of times).
function H = hilb local(n)
% hilb local: Hilbert matrix of order n (not from MATLAB)
% hilb local(n) constructs the n by n matrix with elements 1/(i+j-1).
% This is one of the most famous examples of a matrix which is
% nonsingular, but which is very badly conditioned.
H = zeros(n);
for i = 1:n
for j = 1:n
H(i,j) = 1/(i+j-1);
end
end
The input argument is n and the output argument is H. The first line of the documentation includes the
name of the function as well as a brief description that lookfor uses. The following lines of documenta-
tion also appear on the screen if we enter
>> help hilb local
Note: The above code is not presently used in MATLAB (although it was in early versions.) The actual
MATLAB code for this function is shown in subsection 8.5.
We follow by defining H to be an n×n matrix. Although not essential, this statement can greatly in-
crease the speed of the function because space can be preallocated for the matrix. For example, consider

80
8.3. Function M-files

the following code.


function a = prealloc(n, which)
% prealloc: testing how well preallocating a vector works
% n = the size of the vector
% which = 1 - preallocate the vector
% = 2 - do not
if which == 1
a = zeros(n,1);
end
a(1) = 1;
for i = 2:n
a(i) = a(i-1) + 1;
end
If which = 0 the vector a is not preallocated, while if which = 1 it is. We find that
>> prealloc(100000, 1)
runs over 4000 (that’s right, four thousand) times as fast as
>> prealloc(100000, 0)

Note that i and j are redefined from −1 since they appear as for loop indices. However, since i
and j are local to this function, this does not have any effect when this command is executed. Also, the
variable H is local to the function. If we type
>> Z = hilb local(12)
then the matrix Z contains the Hilbert matrix and H is undefined.
Normally functions are completed when the end of the file is reached (as above). If the flow control in a
function file is complicated enough, this might be difficult to accomplish. Instead, you can use the return
command, which can appear anywhere in the function and force an immediate end to the function. In
addition, you can force the function to abort by entering
error(<string>)
If the string is not empty, the string is displayed on the terminal and the function is aborted; if the string
is empty, the statement is ignored.
One feature of function files which is occasionally very useful is that they can have a variable number of
input and output variables. For example, the norm of a vector x can be calculated by entering
>> norm(x, p)
if p = 1, 2, or inf or, more simply, by
>> norm(x)
if p = 2. Similarly, if only the eigenvalues of a matrix A ∈ Cn×n are desired, enter
>> eigval = eig(A)
However, if both the eigenvalues and eigenvectors are desired, enter
>> [V, D] = eig(A)
where D ∈ Cn×n is a diagonal matrix containing the n eigenvalues on its diagonal and the corresponding
eigenvectors are found in the same columns of the matrix V ∈ Cn×n .
Note: On the right side of an equation, [V D] or [V, D] is the matrix whose initial columns come from
V and whose final columns come from D. This requires that V and D be matrices which have the
same number of rows. On the left side, [V, D] denotes the two output arguments which are re-
turned by a function. V and D can be completely different variables. For example, one can be a
character variable and the other a matrix.
MATLAB can also determine the number of input and output arguments: nargin returns the number
of input arguments and nargout returns the number of output arguments. For example, suppose we want
to create a function file which calculates
2
f (x, ξ, a) = e−a(x−ξ) sin x .

We can “spruce” this function up to have default values for ξ and a and also to calculate its derivative

81
8.3. Function M-files

with the following function file.


function [out1, out2] = spruce(x, xi, a)
% spruce: a silly function to make a point, f(x,b,a) = sin(x)*exp(-a*(x-b) 2)
% if only x is input, xi = 0 and a = 1
ˆ
% if only x and xi are input, a = 1
% if only one output argument, f(x,xi,a) is calculated
% if two output arguments, f(x,xi,a) and f’(x,xi,a) are calculated
if nargin == 1
xi = 0;
a = 1;
elseif nargin == 2
a = 1;
end
out1 = exp(-a.*(x-xi). 2).*sin(x);
if nargout == 2
ˆ
out2 = exp(-a.*(x-xi). 2).*(cos(x) - 2.*a.*(x-xi).*sin(x));
end
ˆ
If there is only one input argument then ξ is set to 0 and a is set to 1 (which are useful default val-
ues) while if there are only two input arguments then a is set to 1. If there is only one output argument
then only f (x) is calculated, while if there are two output arguments then both f (x) and f ′ (x) are
calculated.
Also, note that x can be a scalar (i.e., a single value) or it can be a vector. Similarly, ξ and a can
each be a scalar or a vector. If x is a vector, i.e., (x1 , x2 , . . . , xn )T , while ξ and a are scalars, then the
function is 2
f (xi , ξ, a) = sin(xi )e−a(xi −ξ) for i = 1, 2, . . . , n ,
and all the values can be calculated in one call to spruce. If, on the other hand, x, ξ, and a are all vec-
tors, then the function is
2
f (xi , ξi , ai ) = sin(xi )e−ai (xi −ξi ) for i = 1, 2, . . . , n ,

and, again, all the values can be calculated in one call to spruce.
We have now presented all the essential features of the MATLAB programming language, and it cer-
tainly is a “minimal” language. MATLAB can get away with this because most matrix operations can
be performed directly — unlike in most other programming languages. You only need to write your own
function if MATLAB cannot already do what you want. If you want to become proficient in this language,
simply use the type command to look at the coding of some functions.
Now that we have discussed the essentials of programming in MATLAB, it is necessary to discuss how
to program correctly. When you are entering one statement at a time in the text window, you immediately
see the result of your calculation and you can determine whether or not it is correct. However, in an m-file
you have a sequence of statements which normally end with semicolons so that you do not see the inter-
mediate calculations. What do you do if the result is incorrect? In other words, how do you debug your
m-file?
There are a number of simple techniques you can use and we discuss them in turn. In a script m-file
intermediate calculations are normally not printed out, but they are still available to look at. This can fre-
quently lead to an understanding of where the calculation first went wrong. However, this is not true of
function m-files since all the local variables in the function disappear when the function ends. Of course,
with any m-file you can selectively remove semicolons so that intermediate results are printed out. This is
probably the most common method of debugging programs — no matter what programming language is
being used.
When loops are involved (either using for or while statements), the printed output can seem to be
endless (and it is endless if you are in an infinite loop). And it is usually impossible to read the out-
put since it is zipping by at (what appears to be) nearly the speed of light! The pause command can
slow down or even stop this output. By itself pause stops the m-file until some key is pressed while

82
8.3. Function M-files

pause(<floating-point number>) pauses execution for this many seconds — or fractions of a second.
(This is computer dependent, but pause(0.01) should be supported on most platforms.) You can even
turn these pauses on and off by using pause on and pause off.
The echo command is also useful for debugging script and function files, especially when if state-
ments are involved. Typing
>> echo on
turns on the echoing of statements in all script files (but not printing the results if the statements end
with semicolons), and echo off turns echoing off again. However, this does not affect function files. To
turn echoing on for a particular function, type
>> echo <function name> on
and to turn echoing on for all functions, type
>> echo on all
The keyboard command is also very useful for debugging m-files. It stops execution of the m-file,
similar to the pause command. However, it returns complete control to the user to enter any and all
MATLAB commands. In particular, you can examine any variables in the function’s workspace. If de-
sired, you can also change the value of any of these variables. The only way you will recognize this is not a
“standard” MATLAB session is that the prompt is
K>>
for Keyboard. To terminate the “keyboard” session and return control to the m-file, enter
K>> return
To terminate both the “keyboard” session and the execution of the m-file, enter
K>> dbquit
When using the debugger, you are not running your program. Instead, you are running the debugger
which is running your program. Thus, many of the commands you enter are commands for the debugger;
to distinguish these commands from “normal” MATLAB commands they begin with db. There are two
ways to run the debugger: you can type the debugger commands into the workspace, or in a MATLAB
window you can use the mouse and click on the commands.
In addition, you are still in the workspace so many MATLAB commands can still be executed. For ex-
ample, to see the values of variables just type the variable name into the workspace. (Alternatively, you
can move the mouse over the variable name in the window and its description and value(s) will be shown.)
Also, you can run most MATLAB commands as long as you do not try to create new variables (but you
can modify existing variables).
We will not discuss the commands in this debugger in detail, but only provide a brief description of
each one, because these are similar to commands in any debugger. If you have experience with using a
debugger, help or doc will give you complete details.

Debugging Commands

keyboard Turns debugging on.


dbstep Execute the next executable line.
dbstep n Execute the next n lines.
dbstep in The same as dbstep except that it will step into another m-file (rather than over it).
dbstep out Executes the remainder of the current function and stops afterwards.
dbcont Continue execution.
dbstop Set a breakpoint.
dbclear Remove a breakpoint.
dbup Change the workspace to the calling function or the base workspace.
dbdown Change the workspace down to the called function.
dbstack Display all the calling functions.
dbstatus List all the breakpoints.
dbtype List the current function, including the line numbers.
dbquit Quit debugging mode and terminate the function.
return Quit debugging mode and continue execution of the function.

83
8.3. Function M-files

Now we want to discuss the arguments in a MATLAB function, since they are are used somewhat differ-
ently than in other programming languages. For example, in
function out = funct1(a, t)
a and t are the input arguments and out is the output argument. Any and all input variables are lo-
cal to the function and so can be modified without affecting the arguments when the function funct1 is
called. (This is true no matter what type of variables they are.) In
function [out1, out2, out3] = funct2(z)
z is the only input argument and there are three output arguments, each of which can be any type of vari-
able. There is no requirement that all three of these output arguments actually be used. For example, the
calling statement might be any of the following:
>> art = funct2(1.5)
>> [physics, chemistry] = funct2([1 2 3])
>> [math, philosophy, horticulture] = funct2(reshape([1:30], 6, 5))
(just to be somewhat silly).
In a programming language such as C, Fortran, or Pascal funct1 would be written in a form such as
function funct1(a, t, out)
or even as
function funct1(a, out, t)
Similarly, funct2 would be written as
function funct2(z, out1, out2, out3)
or even as
function funct2(out3, out2, z, out1)
It would be up to the user to control which were the input arguments, which were the output arguments,
and which were both (i.e., one value on input and another on output). In MATLAB input arguments oc-
cur on the right side of the equal sign and output arguments occur on the left. Arguments which are to be
modified by the function must occur on both sides of the equal sign in the calling statement. For example,
in funct2 if z is modified and returned in out1 then the calling sequence should be
>> [a, b, c] = funct2(a)
where a appears on both sides of the equal sign. (There is an alternative to this awkward use of param-
eters which are modified by the function: you can make a variable global, as we discuss at the end of this
section. However, this is not usually a good idea.)
There is another difference between MATLAB and most other programming languages where the type
of each variables has to be declared, either explicitly or implicitly. For example, a variable might be an
integer, a single-precision floating-point number, a double-precision floating-point number, a character
string, etc. In MATLAB, on the other hand, there is no such requirement. For example, the following
statements can follow one another in order and define x to be a string variable, then a vector, then a
scalar, and finally a matrix.
>> x = ′ WOW?′
>> x = x + 0
>> x = sum(x)
>> x = x*[1 2; 3 4]
It is particularly important to understand this “typelessness” when considering output arguments. For ex-
ample, there are three output arguments to funct2 and any of them can contain any type of variable. In
fact, you can let the type of these arguments depend on the value or type of the input argument. This is
probably not something you should want to do frequently, but it is sometimes very useful.
Occasionally, there is a need to pass values from the workspace to a function or to pass values between
different functions without using the input arguments. (As we discussed earlier, this may be desirable if a
variable is modified by a function.) In C this is done by using global variables. MATLAB also has global
variables which are defined by declaring the variables to be global using
>> global <variable 1> <variable 2> <variable 3> ...
By the way, a variable is a global variable if it appears in a global statement and a local variable if it does
not. (Note that a variable can be a local variable in one function and a global variable in another.)
Warning: Spaces, not commas, must separate the variables in a global statement
This statement must appear in every function which is to share the variables. If the workspace is also to
share these variables, you must type this statement (or be put into a script file which you execute) before

84
8.3. Function M-files

these variables are used.


Instead of using a global variable, it is frequently preferable to save the value of a local variable between
calls to the function. Normally, local variables come into existence when the function is called and dis-
appear when the function ends. Sometimes it is very convenient to be able to “save” the value of a local
variable so that it will still be in existence when the function is next called. In C, this is done by declaring
the variable static. In MATLAB it is done by declaring the variable persistent using
>> persistent <variable 1> <variable 2> <variable 3> ...
Warning: Spaces, not commas, must separate the variables.
Note: The first time you enter the function, a persistent variable will be empty, i.e., [], and you can test
for this by using isempty.
We now present a simple example where persistent variables are very helpful. Suppose we want to write
a function m-file to evaluate  
y2
z(y) =
y1 (1 − βy12 ) − αy2 + Γ cos ωt
where α, β, Γ, and ω are parameters which will be set initially and then left unchanged during a run. We
might be studying a mathematical model where this function will be evaluated many, many times for dif-
ferent values of y. For each experiment these parameters will be fixed, but they will be different for each
experiment. We do not want to “hardcode” the values in the function because we would have to repeat-
edly change the function — which is very undesirable. Certainly we can write the function as
function z = fncz1(y, alpha, beta, Gamma, omega)
z = [ y(2) ; -y(1)*(1-beta*y(1) 2)-alpha*y(2)+Gamma*cos(omega*t) ];
ˆ
but then we have to include these four parameters in each call to the function. We can always simplify
this function by combining the four parameters into one vector by
function z = fncz2(y, par)
% alpha = par(1), beta = par(2), Gamma = par(3), omega = par(4)
z = [ y(2) ; -y(1)*(1-par(2)*y(1) 2)-par(1)*y(2)+par(3)*cos(par(4)*t) ];
ˆ
but then it is harder to read the equation. (If this function was more complicated it would be much
harder to read.) To make this last function easier to read we could write it as
function z = fncz3(y, par)
alpha = par(1)
beta = par(2)
Gamma = par(3)
omega = par(4)
z = [ y(2) ; -y(1)*(1-beta*y(1) 2)-alpha*y(2)+Gamma*cos(omega*t) ];
but we want to propose another alternative.
ˆ
There are many reasons that evaluating z(y) using fncz1, fncz2, or fncz3 might not be desirable
— or even practical. As one example, this function might be called repeatedly inside a general purpose
function m-file, say general, which we have written. In general we only want to call the function as z
= fnctn(y) and not have to worry about how parameters are passed to the function. If we instead write
the function m-file for z(y) as
function z = fncz4(y, alpha, beta, Gamma, omega)
persistent alpha p beta p Gamma p omega p
if nargin > 1
alpha p = alpha;
beta p = beta;
Gamma p = Gamma;
omega p = omega;
end
z = [ y(2) ; -y(1)*(1-beta p*y(1) 2)-alpha p*y(2)+Gamma p*cos(omega p*t) ];
ˆ
we will initially call the function as fncz4(y, alpha, beta, Gamma, omega) and then afterwards call
it as fncz2(y). In the initial call all the parameters are saved in persistent variables. Later calls do
not need to input these parameters because they have been saved in the function. That is, we would call
general by
fncz4(y, alpha, beta, Gamma, omega);
general(′fncz4′, ...)

85
8.3. Function M-files

where general would be written as


function general(fnctn, ...)
...
z = feval(fnctn, y)
...
The function feval(fnctn,y), which is discussed in the next section, means fnctn(y). It is used when
the name of a function has been passed as an argument into another function. Our discussion of saving
parameters in function has been somewhat lengthy, but it has many uses.
The final technical detail about function m-files concerns an important element of programming style
in any computer language. It frequently happens that a block of code occurs two or more times in a func-
tion. Sometimes these blocks can be combined by using a loop, but, even if possible, this often makes
the code unwieldly. Instead, this block of code can be put into a new function and called from the orig-
inal function. Another reason for splitting a block of code off into a new function is when the function
has grown large enough to be hard to comprehend. The remedy is to split the code up into a number of
functions, each of which can be easily understood and debugged. In MATLAB functions normally have
to be separated into different files so that each function and its file name agree; otherwise, MATLAB can-
not find the function. This can be annoying if a number of files have to be created: for example, it can be
difficult to remember the purpose of all these functions, and it can be difficult to debug the primary func-
tion. MATLAB has a feature to handle this proliferation of files; function m-files can contain more than
one function. The first function in the file is called the primary function and its name must agree with the
name of the file. There are two further types of functions: Any remaining functions are called subfunc-
tions and nested functions. (At the end of subsection 10.2 we code the function gravity using a number
of nested functions.)
Note: The primary function or a subfunction begins with the function definition line (i.e., the line which
begins with the keyword function ). It is possible to end the primary function and each subfunc-
tion with the command end, but this is not necessary because MATLAB recognizes that a function
has ended when it encounters the next function statement. However, if a nested function is used
then it — and all other functions — must end with the end statement.
First, we discuss subfunctions, which are quite simple. They are placed following the primary function
and between or following other subfunctions. They are only visible to the primary function and to other
subfunctions in the same file. Thus, different m-files can contain subfunctions with the same name. Also,
the help, lookfor, and type commands can only access the primary file. It is crucial to understand
that variables in the primary function or in a subfunction are local to that function and unknown out-
side it (unless they are declared to be global ). The only way to pass variables between these functions is
through the argument list.
Usually, subfunctions are sufficient — and they are much easier to describe. When they are not suffi-
cient, we have nested functions. Its main advantage (as far as we are concerned) is that variables can be
passed into and out of a nested function without being in the argument list. Nested functions are more
complicated than subfunctions and we will only provide a brief discussion.†
To make this discussion specific, consider the following function m-file.


They are similar to internal functions in Fortran 95, and they are somewhat related to internal classes in Java
— but not in C++.

86
8.3. Function M-files

function [p1, p2, p3] = nested ex(x, y, z) % 1


p1 = x; % 2
p2 = y; % 3
nest 1; % 4
function nest 1 % 5
n1a = p1 + z; % 6
n1b = p2 + z; % 7
p1 = n1a + n1b; % 8
p2 = n1a*n1b; % 9
p3 = sub 1(p1, p2); % 10
nest 2; % 11
end % 12
function nest 2 % 13
n2a = p1; % 14
p1 = n2a 2; % 15
p2 = p2 2;
ˆ % 16
ˆ
% p3 = n1a; % WRONG % 17
end % 18
end % 19
function s3 = sub 1(a, b) % 20
s1 = 10; % 21
s2 = a + b; % 22
s3 = s2 + nest 3(s1); % 23
function n3b = nest 3(n3a) % 24
n3b = s1 2 + n3a; % 25
end
ˆ % 26
end % 27

A nested function is within another function. For example, the nested functions nest 1 and nest 2 are
nested within the primary function nested ex, and the nested function nest 3 is nested within the sub-
function sub 1. (Nested functions can have other nested functions within them, but enough is enough.)
The important concept to understand when using nested functions is the scope of variables in the func-
tion m-file. The scope of a variable is the context within which it is defined, i.e., where it can be set,
modified, and used. Now let us consider a function workspace. The workspace of the primary function
is also independent of the workspace of each subfunction. However, since a nested function is within one
or more other functions, it is within the workspace of this function or these functions. In the function
nested ex the nested functions nest 1 (lines 5–12) and nest 2 (lines 13–19) have access to the vari-
ables p1, p2, and p3 of the primary function (lines 1–20). They also have access to the subfunction
sub 1 (lines 20–27) as shown in line 10. Note that nest 2 also has access to nest 1 as shown in line 11.
It is very important to understand that a nested function also has its own workspace. For example, the
variables n1a and n1b in nest 1, n2a in nest 2, and n3a and n3b in nest 3 are local to their re-
spective functions. In particular, n1a cannot be accessed in nest 2 as shown in line 17. (If n1a really
needs to be passed to nest 2, then it must be in the workspace of nested ex. This could be done by
adding n1a = 0 after line 3.) Finally, note that nested functions can also have arguments passed in the
argument list as in nest 3 on line 24.
Now let us return to the topic of how MATLAB finds a function. As we stated previously (but did not
discuss), when a function is called from within an m-file, MATLAB first checks if the function named is
the primary function or a subfunction in the current file. If it is not, MATLAB searches for the m-file in
the current directory. Then MATLAB searches for a private function by the same name (described below).
Only if all this fails does MATLAB use your search path to find the function. Because of the way that
MATLAB searches for functions, you can replace a MATLAB function by a subfunction in the current

87
8.4. Odds and Ends

m-file — but make sure you have a good reason for doing so!†
In the previous paragraph we described how to create a subfunction to replace one function by another
of the same name. There is another, more general, way to handle this replacement: you can create a sub-
directory in your current directory with the special name “private”. Any m-files in this subdirectory are
visible only to functions in the current directory. The functions in this subdirectory are called private
functions. For example, suppose we are working in the directory “personal” and have created a number of
files which use rref to solve linear systems. And suppose we have written our own version of this com-
mand, because we think we can calculate the reduced row echelon of a matrix more accurately. The usual
way to test our new function would be to give it a new name, say myrref, and to change the call to rref
in every file in this directory to myrref. This would be quite time-consuming, and we might well miss
some. Instead, we can code and debug our new function in the subdirectory “private”, letting the name of
our new function be rref and the name of the m-file be rref.m. All calls in the directory to rref will
use the new function we are testing in the subdirectory “private”, rather than MATLAB’s function. Even
more important, any function in any other directory which calls rref will use the MATLAB function and
not our “new, improved version”.

Function Commands

function Begins a MATLAB function.


end Ends a function. This statement is only required if the function m-file con-
tains a nested function, in which case it must be used by all the functions in
the file.
error(′<message>′) Displays the error message on the screen and terminates the m-file
immediately.
echo Turns echoing of statements in m-files on and off.
global Defines a global variable (i.e., it can be shared between different functions
and/or the workspace).
persistent Defines a local variable whose value is to be saved between calls to the
function.
keyboard Stops execution in an m-file and returns control to the user for debugging
purposes. The command return continues execution and dbquit aborts
execution.
return Terminates the function immediately.
nargin Number of input arguments supplied by the user.
nargout Number of output arguments supplied by the user.
pause Halts execution until you press some key.

8.4. Odds and Ends


In MATLAB it is possible for a program to create or modify statemennts “on the fly”, i.e., as the pro-
gram is running. Entering
>> eval(<string>)
executes whatever statement or statements are contained in the string. For example, entering
>> s = ′ x = linspace(0, 10, n); y = x.*sin(x).*exp(x/5); plot(x, y)′
>> eval(s)


Since MATLAB contains thousands of functions, this means you do not have to worry about one of your sub-
functions being “hijacked” by an already existing function. When you think up a name for a primary function
(and, thus, for the name of the m-file) it is important to check that the name is not already in use. However,
when breaking a function up into a primary function plus subfunctions, it would be very annoying if the name of
every subfunction had to be checked — especially since these subfunctions are not visible outside the m-file.

88
8.4. Odds and Ends

executes all three statements contained in the string s. In addition, if an executed statement generates
output, this is the output of eval. For example, if we type
>> A = zeros(5,6);
>> [m, n] = eval(′size(A)′);
then m is 5 and n is 6.
There is a very practical applications for this command since it can can combine a number of state-
ments into one. For example, suppose we want to work with the columns of the Hilbert matrix of size n
and we want to create variables to hold each column, rather than using H(:,i). We can do this by hand
by typing
>> c1=H(:,1);
>> c2=H(:,2);
...
which gets tiring very quickly. Instead, we can do this by typing
>> for i = 1:n
eval( [′ c′ num2str(i) ′ =H(:,i)′] )
end
This requires some explanation. It might be a little clearer if we separate the statement inside the for
loop into two statements by
s = [′ c′ , num2str(i), ′ =H(:,i)′]
eval(s)
(where we include commas in the first statement for readability). s is a text variable which contains
c1=H(:,1) the first time the loop is executed, then c2=H(:,2) the second time, etc. (To understand
how s is created, recall that s is really just a row vector with each element containing the ASCII repre-
sentation of the corresponding character.)
Finally, there is a very esoteric application for this command that allows it to catch errors. This is sim-
ilar to the “catch” and “throw” commands in C++ and Java. To use this feature of eval, call it using
two arguments as
>> eval(<try string>, <catch string>)
The function executes the contents of <try string> and ignores the second argument if this execution
succeeds. However, if it fails then the contents of <catch string> are executed. (This might be a call
to a function which can handle the error.) If there is an error, the command lasterr returns a string
containing the error message generated by MATLAB.
A MATLAB command which is occasionally useful in a function is feval. It executes a function, usu-
ally defined by an m-file, whose name is contained in a string by
>> feval(<string>, x1, x2, ..., xn)
(See below for other ways to pass the function in the argument list.) Here x1, x2, ..., xn are the
arguments to the function. For example, the following two statements are equivalent
>> A = zeros(5,6)
>> A = feval(′zeros′, 5, 6)
Suppose that in the body of one function, say sample, we want to execute another function whose
name we do not know. Instead, the name of the function is to be passed as an argument to sample.
Then feval can be used to execute this text variable. For example, suppose in function sample we want
to generate either linear or logarithmic plots. We can input the type of plot to use by
function sample(<type of plot>)
...
feval(<type of plot>, x, y1, x, y2, ′ --′ , xx, y3, ′ :′ )
...

There are two common ways to pass the function <type of plot> in the argument list:
(1) use a character string, e.g., feval( ′ loglog ′ , ...), or
(2) use a function handle, e.g., feval(@logval, ...), or
Note: eval and feval serve similar purposes since they both evaluate something. In fact, feval can al-
ways be replaced by eval since, for example, feval( ′ zeros ′ , 5, 6) can always be replaced by
eval( ′ zeros(5,6) ′ ). However, there is a fundamental difference between them: eval requires

89
8.5. Advanced Topic: Vectorizing Code

the MATLAB interpreter to completely evaluate the string, whereas feval only requires MATLAB
to evaluate an already existing function. feval is much more efficient, especially if the string must
be evaluated many times inside a loop.

Odds and Ends

eval Executes MATLAB statements contained in a text variable. Can also


“catch” an error in a statement and try to fix it.
feval Executes a function specified by a string. (Can be used to pass a function
name by argument.)
lasterr If eval “catches” an error, it is contained here.

8.5. Advanced Topic: Vectorizing Code


As long as your MATLAB code executes “quickly”, there is no need to try to make it faster. However, if
your code is executing “slowly”, you might be willing to spend some time trying to speed it up.† There are
three standard methods to speed up a code:
(0) Preallo ate matri es as shown in the function prealloc on page 80. This is very simple and very
effective if the matrices are “large”.
(1) Use MATLAB functions, whenever possible, rather than writing your own. If a MATLAB function is
built-in, then it has been written in C and is faster than anything you can do. Even if it is not, much
time has been spent optimizing the functions that come with MATLAB; you are unlikely to do better.
(2) Replace flow control instructions with vector operations. We have already discussed this topic at
length in subsection 8.2. Here we will focus on some advanced techniques.
Of course, we should first determine how much CPU time is really being expended in our program
— and that is the profile command. It will happily show you the percentage of CPU time spent in
each line of each function in your program! And it will return the results in a very readable fashion! The
commands
>> profile on
>> ..........
>> profile viewer
begin and end profiling and display the results in the Profiler window.
As a simple example of method (0), consider the function hilb on page 80. hilb local(2000) runs
over 300 times slower if the line H = zeros(n) is omitted.
Continuing with this example, currently the MATLAB function hilb is written as
J = 1:n; % J is a row vector
J = J(ones(n, 1),:); % J is now an n by n matrix with each row being 1:n
I = J′ ; % I is an n by n matrix with each column being 1:n
E = ones(n, n);
H = E./(I+J-1);
as you can see by entering
>> type hilb
In the past this code ran nearly 20 times as fast as hilb local. However, now hilb local runs 50%
faster. The reason is that MATLAB has greatly improved its handling of for and while statements.
Thus, it is frequently not necessary to convert simple loops into complicated vector code.
As a realistic example of method (2), suppose you have a large vector y which is the discretization of
a smooth function and you want to know some information about it. In particular, consider the inter-
vals in y where yi > R. What is the average length of these intervals and what is their standard devia-
tion? Also, only include intervals which lie completely within y (i.e., ignore any intervals which begin or

We have put “quickly” and “slowly” in quotes because this is quite subjective. Remember that your time is
valuable: if it takes you longer to optimize your code than you will save in running it more quickly, stifle the urge
to muck around with it. Also remember that the amount of time it actually takes to optimize a code is usually a
factor of two or three or . . . longer than the time you think it will take before you get started.

90
8.5. Advanced Topic: Vectorizing Code

end y ). It is not difficult to write such a code using flow control statements:
function ylen intvl = get intervals slow(y, R)
n = length(y);
if y(1) > R % check if the first point is in an interval
in intvl = 1; % yes
intvl nr = 1;
yin(intvl nr) = 1;
else
in intvl = 0; % no
intvl nr = 0;
end
for i = [2: n] % check the rest of the points
if in intvl == 1 % we are currently in an interval
if y(i) <= R % check if this point is also in the interval
yout(intvl nr) = i; % no, so end the interval
in intvl = 0;
end
else % we are currently not in an interval
if y(i) > R % check if this point is in the next interval
intvl nr = intvl nr + 1; % yes, so begin a new interval
yin(intvl nr) = i;
in intvl = 1;
end
end
end
if y(1) > R % check if we have begun in an interval
yin(1) = []; % yes, so delete it
yout(1) = [];
end
if length(yin) > length(yout) % check if we have ended in an interval
yin( length(yin) ) = []; % yes, so delete it
end
ylen intvl = yout - yin;
When completed, yin and yout contain the element numbers where an interval begins and where it
ends, respectively. This is straightforward — but very slow if y has millions of elements.

91
9. Sparse Matrices

To write a vectorized code, we have to think about the problem differently:


(1) We do not care about the actual values in y, only whether they are greater than R or not. So we
construct a logical matrix corresponding to y by yr = (y > R).
(2) We do not actually care about the 0’s and 1’s — only about where the value changes because these
mark the boundaries of the intervals. So we take the difference between adjacent elements of yr by
yd = diff(yr).
(3) We actually only need to know the elements which contain nonzero values so we find the element
numbers by ye = find(yd), i.e., ye = find(yd =0).
˜
(4) We do not care about the actual locations of the beginning and end of each interval, only the lengths
of these intervals. So we take the difference again by ylen = diff(ye).
(5) Finally, ylen contains the lengths of both the intervals and the distances between successive inter-
vals. So we take every other element of ylen. We also have to be a little careful and check whether
y begins and/or ends in an interval.
Here is the code:
function ylen intvl = get intervals fast(y, R)
yr = (y > R); % (1)
yd = diff(yr); % (2)
ye = find(yd); % (3)
ylen = diff(ye); % (4)
if y(1) > R % (5), check if we begin in an interval
ylen(1) = []; % yes
end
ylen intvl = ylen( 1:2:length(ylen) ); % get every other length

Finally, the question remains: is the time savings significant? For “large” y the CPU time is reduced by
approximately 50.

Improving Efficiency of Function

profile Profile the execution time of a MATLAB code. This is very useful for improving the
performance of a code by determining where most of the CPU time is spent.

9. Sparse Matrices
Many matrices that arise in applications only have a small proportion of nonzero elements. For exam-
ple, if T ∈ Cn×n is a tridiagonal matrix, then the maximum number of nonzero elements is 3n−2. This
is certainly a small proportion of the total number of elements, i.e., n2 , if n is “large” (which commonly
means in the hundreds or thousands or . . . )
For full matrices (i.e., most of the elements are nonzero) MATLAB stores all the elements, while for
sparse matrices (i.e., most of the elements are zero) MATLAB only stores the nonzero elements: their
locations (i.e., their row numbers and column numbers) and their values. Thus, sparse matrices require
much less storage space in the computer. In addition, the computation time for matrix operations is sig-
nificantly reduced because zero elements can be ignored.
Once sparse matrices are generated, MATLAB is completely responsible for handling all the details
of their use: there are no special commands needed to work with sparse matrices. However, there are a
number of commands which are inappropriate for sparse matrices, and MATLAB generally generates a
warning message and refers you to more appropriate commands. For example, cond(S) has to calculate
S−1 , which is generally a full matrix; instead, you can use condest which estimates the condition number
by using Gaussian elimination. You have two alternatives: first, use full to generate a full matrix and
use the desired command; or, second, use the recommended alternative command.
There are two common commands in MATLAB for creating sparse matrices. You can enter all the
nonzero elements of S ∈ Cm×n individually by

92
9. Sparse Matrices

>> S = sparse(i, j, s, m, n)
where i and j are vectors which contain the row and column indices of nonzero elements and s is the
vector which contains the corresponding values. For example, the square bidiagonal matrix
 
n −2


n−1 −4 0 

 n−2 −6 
S= .. .. 
 . . 
 
2 −2n + 2
0 1
has the following nonzero elements

i j si,j i j si,j
1 1 n 1 2 −2
2 2 n−1 2 3 −4
3 3 n−2 3 4 −6
... ... ... ... ... ...
n−1 n−1 2 n−2 n−1 −2n + 4
n n 1 n−1 n −2n + 2
A simple way to generate this matrix is by entering
>> S = sparse([1:n], [1:n], [n:-1:1], n, n) + ...
sparse([1:n-1], [2:n], [-2:-2:-2*n+2], n, n)
We could, of course, generate S using one sparse command, but it would be more complicated. The
above command is easier to understand, even if it does require adding two sparse matrices. Since the out-
put from this command is basically just the above table, it is difficult to be sure that S is precisely what
is desired. We can convert a sparse matrix to full by
>> full(S)
and check explicitly that S is exactly what is shown in the above matrix.
In addition, a full (or even an already sparse) matrix A can be converted to sparse form with all zero
elements removed by
>> S = sparse(A)
Finally, a zero m×n matrix can be generated by
>> SZ = sparse(m, n)
which is short for
>> SZ = sparse([], [], [], m, n)
The second common command for generating sparse matrices is
>> S = spdiags(B, d, m, n)
which works with entire diagonals of S. B is an min{ m, n }×p matrix and its olumns become the di-
agonals of S specified by d ∈ Cp . (For example, if d = (0, 1)T then the first column of B contains the
elements on the main diagonal and the second column contains the elements on the diagonal which is one
above the main diagonal.) Thus, we can also generate the matrix S given above by
>> B = [ [n:-1:1]′ [0:-2:-2*n+2]′ ]
>> S = spdiags(B, [0 1]′ , n, n)
Warning: Be Careful! The command spdiags is somewhat similar to diag but must be handled more
carefully. Note that the element b1,2 is 0, which does not appear in S. The difficulty is that
the number of rows of B is generally larger than the lengths of the diagonals into which the
columns of B are to be placed and so some padding is required in B. The padding is done so
that all the elements in the k-th row of B come from the k-th olumn of S.
For example, the matrix
 
0 0 6 0 0
1 0 0 7 0
 
S1 =  0 2 0 0 8 
 
0 0 3 0 0
0 0 0 4 0

93
9. Sparse Matrices

can be generated as a sparse matrix by


>> A = diag([1:4], -1) + diag([6:8], 2)
>> S1 = sparse(A)
or by
>> B = [ [1:4] 0; 0 0 [6:8] ]′
>> S1 = spdiags(B, [-1 2], 5, 5)
In the latter case note that the columns of B have to be padded with zeroes so that each col-
umn has five elements, whereas in the former case the vector which becomes the particular
diagonal precisely fits into the diagonal. The element s1,3 of S1 contains the value 6. It ap-
pears in the 3-rd row of B because it occurs in the 3-rd olumn of S1. Note that the element
bn,2 is not used since it would go into the element sn,n+1 .
A slight variation of the above command is
>> T = spdiags(B, d, S)
where T is equated to S and then the columns of B are placed in the diagonals of T specified by d.
Thus, a third way to generate the matrix S given above is
>> S = spdiags([n:-1:1]′, [0], n, n)
>> S = spdiags([0:-2:-2*n+2]′, [1], S)
Just as with the diag command, we can also extract the diagonals of a sparse matrix by using
spdiags. For example, to extract the main diagonal of S, enter
>> B = spdiags(S, [0])
The number of nonzero elements in the sparse matrix S are calculated by
>> nnz(S)
(Note that this is not necessarily the number of elements stored in S because all these elements are
checked to see if they are nonzero.) The locations and values of the nonzero elements can be obtained
by
>> [iA, jA, valueA] = find(A)
The locations of the nonzero elements is shown in the graphics window by entering
>> spy(S)
These locations are returned as dots in a rectangular box representing the matrix which shows any struc-
ture in their positions.
All of MATLAB’s intrinsic arithmetic and logical operations can be applied to sparse matrices as well
as full ones. In addition, sparse and full matrices can be mixed together. The type of the resulting matrix
depends on the particular operation which is performed, although usually the result is a full matrix. In
addition, intrinsic MATLAB functions often preserve sparseness.
You can generate sparse random patrices by sprand and sparse, normally distributed random matri-
ces by sprandn. There are a number of different arguments for these functions. For example, you can
generate a random matrix with the same sparsity structure as S by
>> sprand(S)
or you can generate an m×n matrix with the number of nonzero random elements being approximately
ρmn by
>> sprand(m, n, rho)
Finally, you can generate sparse random symmetric matrices by sprandsym; if desired, the ma-
trix will also be positive definite. (There is no equivalent command for non-sparse matrices so use
full(sprandsym(...))
Additionally, sparse matrices can be input from a data file with the spconvert command. Use
csvread or load to input the sparsity pattern from a data file into the matrix <sparsity matrix>.
This data file should contain three columns: the first two columns contain the row and column indices of
the nonzero elements, and the third column contains the corresponding values. Then type
>> S = spconvert(<sparsity matrix>)
to generate the sparse matrix S. Note that the size of S is determined from the maximum row and the
maximum column given in <sparsity matrix>. If this is not the size desired, one row in the data file
should be “ m n 0” where the desired size of S is m×n. (This element will not be used, since its value is
zero, but the size of the matrix will be adjusted.)

94
10.1. Basic Commands

Sparse Matrix Functions

speye Generates a sparse identity matrix. The arguments are the same as
for eye.
sprand Sparse uniformly distributed random matrix.
sprand Sparse uniformly distributed random symmetric matrix; the matrix
can also be positive definite.
sprandn Sparse normally distributed random matrix.
sparse Generates a sparse matrix elementwise.
spdiags Generates a sparse matrix by diagonals or extracts some diagonals
of a sparse matrix.
full Converts a sparse matrix to a full matrix.
find Finds the indices of the nonzero elements of a matrix.
nnz Returns the number of nonzero elements in a matrix.
spfun(′<function name>′, A) Applies the function to the nonzero elements of A.
spy Plots the locations of the nonzero elements of a matrix.
spconvert Generates a sparse matrix given the nonzero elements and their
indices.
sprandsym Generates a sparse uniformly distributed symmetric random ma-
trix; the matrix can also be positive definite.

10. Initial-Value Ordinary Differential Equations


Most initial-value ordinary differential equations cannot be solved analytically. Instead, using MATLAB
we can obtain a numerical approximation to the ode system

d
y = f(t, y) for t ≥ t0
dt
where y ∈ Rn with initial condition y(t0 ) = y0 . The basic MATLAB commands are easily learned. How-
ever, the commands become more involved if we want to explore the trajectories in more detail. Thus, we
divide this section into the really basic commands which are needed to generate a simple trajectory and
into a more advanced section that goes into many technical details. We also provide a large number of ex-
amples, many more than in other sections of this overview, to provide a template of how to actually use
the advanced features.

10.1. Basic Commands


In this subsection we focus on the particular example

y ′′ + αy ′ − y(1 − βy 2 ) = Γ cos ωt ,

which is called Duffing’s equation. This ode has many different types of behavior depending on the values
of the parameters α, β, Γ, and ω.
As written, this is not in the form of a first-order system. To transform it we define y1 = y and
y2 = y1′ = y ′ so that

y1′ = y2
y2′ = y1′′ = y ′′ = y1 (1 − βy12 ) − αy2 + Γ cos ωt
or
 ′  
y1 y2
= .
y2 y1 (1 − βy12 ) − αy2 + Γ cos ωt

95
10.1. Basic Commands

Note: This same “trick” can be applied to an n-th order by defining y1 = y, y2 = y1′ , y3 = y2′ , . . . ,

yn = yn−1 .
Before discussing how to solve Duffing’s equation specifically, we discuss the commands which solve
time-evolution odes. To obtain a numerical solution to a time-evolution first-order ode system, enter
>> <ode solver>(<function handle>, tspan, y0)
or
>> [t, Y] = <ode solver>(<function handle>, tspan, y0)
or
>> sol = <ode solver>(<function handle>, tspan, y0)
First, we have to choose which of the ode solvers to use, which is discussed in detail shortly. It would be
possible for MATLAB itself to decide which numerical method to employ. However, there are good rea-
sons why the decision should be left in the hand of the user.
Warning: Make sure you understand how to enter the name of the function handle. This is discussed at
length in section 3.2, and we also briefly discuss it below.
All of the solvers use the same input and output arguments, which we now discuss. The input parame-
ters are:
function The name of the function handle that calculates f(t, y).
tspan The vector that specifies the time interval over which the solution is to be calculated. If this
vector contains two elements, these are the initial time and the final time; in this case the
ode solver determines the times at which the solution is output. If this vector contains more
than two elements, these are the only times at which the solution is output.
Note: the final time can be less than the initial time, in which case the trajectory is moving
backwards in time.
y0 The vector of the initial conditions for the ode.
If there are no output parameters, the individual elements of the solution, i.e., y1 (t), y2 (t), . . . , yn (t),
are plotted vs. t on a single plot. The circles on the trajectories show the actual times at which the solu-
tion is calculated.
If there are two output parameters, these are:
t The column vector of the times at which the solution is calculated.†
Y The matrix which contains the numerical solution at the times corresponding to t. The first
column of Y contains y1 , the second column y2 , etc.‡
If there is one output parameter, then it is a structure. The output is now
sol.x The column vector of the times at which the solution is calculated.
sol.y The matrix which contains the numerical solution at the times corresponding to t.
There are seven distinct ode solvers which can be used, as shown in the table below. All these ode
solvers use an adaptive step size to control the error in the numerical solution. Each time step is chosen
to try to keep the local error within the prescribed bounds as determined by the relative error and the
absolute error tolerances (although it does not always succeed). That is, ei , which is the error in yi , is
supposed to satisfy
ei ≤ max{ RelTol · |yi |, AbsTol(i) }
where the default value of RelTol is 10−3 and of the vector AbsTol is 10−6 for each element. (How-
ever, there is no guarantee that the error in the numerical calculation actually satisfies this bound.)


The t in [t, Y] is unrelated to the t argument in the function duffing.

We have capitalized the Y in [t, Y] to indicate that the output is a matrix whereas the argument y is a
vector in the function.

96
10.1. Basic Commands

ODE Solvers

ode45 Non-stiff ode solver; fourth-order, one-step method.


ode23 Non-stiff ode solver; second-order, one-step method.
ode113 Non-stiff ode solver; variable-order, multi-step method.
ode15s Stiff ode solver; variable-order, multi-step method.
ode23s Stiff ode solver; second-order, one-step method.
ode23t Stiff ode solver; trapezoidal method.
ode23tb Stiff ode solver; second-order, one-step method.

It is up to you to decide which ode solver to use. As a general rule, unless you believe that the ode is
stiff (which we discuss in the next paragraph), try ode45 or ode113. For a given level of accuracy, these
methods should run “reasonably fast”. (Which one runs faster is very dependent on the ode.) If you know
(or believe) that the ode is stiff, or if these two non-stiff solvers fail, then try ode15s.
And what is a stiff ode? There is no precise definition. Instead, we say it is stiff if the time step re-
quired to obtain a stable and accurate solution is “unreasonably” small. The best way to explain this
rather vague impression is through some simple examples.
Consider the second-order time-evolution ode

y ′′ + 999y ′ + 1000y = 0 for t ≥ 0

with the initial conditions y(0) = η1 and y ′ (0) = η2 . The solution to this ode is

y(t) = c1 et + c2 e−1000t

where
1 1
c1 = (η1 − η2 ) and c2 = (1000η1 + η2 ) .
1001 1001
There are two time scales in this solution: there is a rapid decay due to the e−1000t term and there is a
slow growth due to the et term. Initially, the time step will be “very small” so that the rapid decay is
calculated accurately (i.e., ∆t ≪ 1/1000 ). However, soon it will be negligible and the time step should in-
crease so that it calculates the slow growth accurately (i.e., ∆t ≪ 1 ). However, if a non-stiff solver, such
as ode45 or ode23, is used, the time step must always be “very small”. That is, it must accurately track
the rapidly decaying term — even after this term has disappeared in the numerical solution. The reason
is that a numerical instability will cause the trajectory to blow up if the time step increases. However, if
a stiff solver is used, the time step can increase by many orders of magnitude when the rapidly decaying
term has disappeared.
The same is true for the ode
y ′′ + 1001y ′ + 1000y = 0
whose solution is
y(t) = c1 e−t + c2 e−1000t .
Initially, the time step will be “very small” so that the rapid decay is calculated accurately (i.e.,
∆t ≪ 1/1000 ). However, soon it will be negligible and the time step should increase so that it calculates
the slowly decaying mode accurately (i.e., ∆t ≪ 1 ).
On the other hand, consider the ode

y ′′ − 1001y ′ + 1000y = 0

whose solution is
y(t) = c1 et + c2 e1000t .
the time step must always be “very small” so that the rapidly growing mode e1000t is calculated accu-
rately (i.e., ∆t ≪ 1/1000 ). Thus, this is not a stiff ode.
The above examples are very simple. They are only designed to show that an ode is stiff if there is a
rapidly decaying mode and any growth in the solution occurs on a much slower time scale. (This fre-
quently happens in chemical reaction models, where some reactions occur on a very fast time scale and

97
10.1. Basic Commands

and other occur on a much slower time scale.) In the next subsection we discuss van der Pol’s equation, a
second-order ode which is either non-stiff or stiff depending on the value of one parameter. You can plot
the solution and observe the separation of the fast scale and the slow scale as this parameter increases.
One difficulty with a stiff ode solver is that you may well have to supply the Jacobian of the ode your-
self. The Jacobian of f(t, y) is the n×n matrix
 
∂fi
J(t, y) = (t, y) ,
∂yj

i.e., the element in the i-th row and j-th column of J is

∂fi
.
∂yj

Any of the stiff methods can approximate this matrix numerically. However, if the ode is “bad” enough,
this may not be enough. You may have to calculate all these partial derivatives yourself and include them
in your function file. (We show an example of this later.)
The reason for this large choice of ode solvers is that some odes are very, very, very nasty. It is possible
that most of the ode solvers will fail and only one, or maybe two, will succeed. SAY MORE???
To conclude this subsection, we return to Duffing’s equation. Suppose we want to solve the ode for
t ∈ [0, 100] with initial conditions y = (2, 1)T and plot the results. Since this is a very well-behaved ode
for the parameters given, we can use ode45. The simplest approach is to use an anonymous function to
input the right-hand side.
>> alpha = 0.05;
>> beta = 1.0;
>> Gamma = 0.5;
>> omega = 1.0;
>> duffing a = @(t, y)[ y(2) ; y(1)*(1-beta*y(1) 2)-alpha*y(2)+Gamma*cos(omega*t) ];
>> ode45(duffing a, [0 100], [2 1]);
ˆ
(The “a” denotes the fact that duffing a is an anonymous function handle.) The solution will now be
plotted as y1 and y2 vs. t. (This plot is rather “cluttered” because, not only is the trajectory plotted,
but in addition markers are put at each of the points of the numerical solution.)
Warning: There are a number of parameters which are needed by the function and these must be de-
fined before the function is created. Also, the function handle duffing a will always use these
parameters, even if they are later changed.
Note: Since duffing a is already a function handle, we merely need to use its name as the first argument
to ode45.
To obtain complete control over what is plotted, you should let ode45 output the trajectory and do the
plots yourself. This is easily accomplished by changing the last line of the previous code to
>> [t, Y] = ode45(duffing a, [0 100], [2 1]);
>> figure(1)
>> subplot(2, 1, 1)
>> plot(t, Y(:,1))
>> subplot(2, 1, 2)
>> plot(t, Y(:,2))
>> figure(2)
>> plot(Y(:,1), Y(:,2))
This results in a plot of y vs. t and a separate plot of y ′ vs. t, so that both plots are visible even if
they have vastly different scales. There is also a separate plot of y ′ vs. y, which is called a phase plane.
The next simplest approach is to use a nested function, and so there must also be a primary function.

98
10.1. Basic Commands

function duffing ode(alpha, beta, Gamma, omega, y0, final time)


ode45(@duffing n, [0 final time], y0);
function deriv = duffing n(t, y)
% duffing n: Duffing’s equation, nested function
deriv = [ y(2) ; y(1)*(1-beta*y(1) 2)-alpha*y(2)+Gamma*cos(omega*t) ];
end
ˆ
end
(The “n” denotes the fact that duffing n is a nested function.) Note that the parameters are input to the
primary function and so are immediately accessible to the nested function. Clearly, this second approach
(of using a nested function) requires more coding that the first approach (of using an anonymous func-
tion). However, the first approach only works if the right-hand side can be defined using one MATLAB
statement. If the right-hand side is more complicated, then a nested function is the simplest choice.
Note: Since duffing n is a function, and not a function handle, we have to include ”@” before the name of
the function.
The third, and oldest, approach is to create a separate function m-file (i.e., a primary function) which
calculates the right hand side of this ode system.
function deriv = duffing p(t, y)
% duffing p: Duffing’s equation, primary function
alpha = 0.05;
beta = 1.0;
Gamma = 0.5;
omega = 1.0;
deriv = [ y(2) ; y(1)*(1-beta*y(1) 2)-alpha*y(2)+Gamma*cos(omega*t) ];
ˆ
(The “p” denotes the fact that duffing p is a primary function.) Note that all the parameters are defined
in the m-file so that it will have to be modified whenever we want to modify the parameters. This is a
very bad approach because this file will have to be repeatedly modified.
Note: Since duffing p is a function, and not a function handle, we have to include ”@” before the name of
the function.
Finally, it is very inconvenient that the parameters in Duffing’s equation are determined in the function
itself. We should be able to “explore” the rich behavior of Duffing’s equation without having to constantly
modify the function — in fact, once we have the function exactly as we want it, we should never touch
it again. (This is not only true for esthetic reasons; the more we fool around with the function, the more
likely we are to screw it up!)
This is easily done by adding parameters to the function file.
function deriv = duffing p2(t, y, alpha, beta, Gamma, omega)
% duffing p2: Duffing’s equation, primary function
% with coefficients passed through the argument list
deriv = [ y(2) ; y(1)*(1-beta*y(1) 2)-alpha*y(2)+Gamma*cos(omega*t) ];
ˆ
(The “p2” denotes the fact that duffing p2 is another primary function.) However, this function cannot
be called directly by the ode solver. Instead it is called indirectly by
>> alpha = 0.05;
>> beta = 1.0;
>> Gamma = 0.5;
>> omega = 1.0;
>> duffing c = @(t, y) duffing p2(t, y, alpha, beta, Gamma, omega)
>> ode45(duffing c, [0 100], [2 1]);
Notice that the function duffing c takes only two arguments: t and y. But the function it invokes
is duffing p2 which takes six arguments. Thus, ode45 thinks it is only passing two arguments to
duffing c, but it is actually passing six arguments to duffing p2. In computer science this is a very
simple example of closure† . (The “c” denotes the fact that duffing c is an anonymous function handle


A closure is a complicated term to explain. In this context it means that the parameters used when the func-
tion duffing c was defined are saved and can be referenced inside ode45.

99
10.2. Advanced Commands

which is also a closure.)


To see a sampling of the different type of behavior in Duffing’s equation, let α = 0.15, β = 1, Γ = 0.3
and ω = 1, and let the initial condition be y(0) = (0, 1)T . After a short initial transient, the solution set-
tles down and appears to be “regular” by t = 100: in fact, it appears to be exactly periodic with a period
of 2π due to the 0.3 cos t term. (In fact, to the accuracy of the computer it is exactly periodic.) How-
ever, if we merely change the initial condition to y = (1, 0)T the behavior appears to be chaotic, even at
t = 1000. Here is an example of a ode which has periodic motion for one initial condition and is chaotic
for another! If we now change α from 0.15 to 0.22 we find periodic motion with a period of 6π. This is
just a sampling of the behavior of Duffing’s equation in different parameter regions.
By the way, to separate the initial transient behavior from the long-time behavior, you can use the
script m-file
initial time = ???
final time = ???
y0 = ???
alpha = ???;
beta = ???;
Gamma = ???;
omega = ???;
duffing a = @(t, y)[ y(2) ; y(1)*(1-beta*y(1) 2)-alpha*y(2)+Gamma*cos(omega*t) ]; [t,
Y] = ode45(duffing a, [0 initial time], y0);
ˆ
figure(1)
plot(Y(:,1), Y(:,2))
l t = length(t);
[t, Y] = ode45(duffing c, [t(l t) final time], Y(l t,:));
figure(2)
plot(Y(:,1), Y(:,2))

10.2. Advanced Commands


There are a number of parameters that we can use to “tune” the particular ode solver we choose. The
MATLAB function odeset is used to change these parameters from their default values by
>> params = odeset(′<Prop 1>′ , <Value 1>, ′ <Prop 2>′ , <Value 2>, ...)
where each parameter has a particular name and it is followed by the desired value. The result of this
command is that the parameters are contained in the variable params. You include these parameters in
the ode solver by adding this variable to the argument list of the ode solver function as
>> [t, Y] = <ode solver>(<function handle>, tspan, y0, params)
Some of the more common parameters are shown in the table below; they will be discussed further later.
To determine all the parameters, their possible values and the default value, enter
>> odeset

100
10.2. Advanced Commands

Initial-Value ODE Solver Parameters

odeset(′<Prop 1>′ , Assigns values to properties; these are passed to the ode solver when it is
<Value 1>, ...) executed.

AbsTol The absolute error tolerance. This can be a scalar in which case it applies to all the
elements of y or it can be a vector where each element applies to the corresponding
element of y. (Default value: 10−6 .)
Events A handle to a function which determines when an event occurs.
Jacobian A handle to a function which returns the Jacobian.
JPattern A sparse matrix whose nonzero elements (which should be 1 ) correspond to the possi-
ble nonzero elements of the Jacobian. This is only used when the Jacobian is calculated
numerically, i.e., when the Jacobian property is not used.
OutputFcn A handle to a function which is called after each successful time step. For example, a
plot of the trajectory can be generated automatically as it is being calculated.
Useful MATLAB functions are:

odeplot′ which generates a plot of time versus all the components of the trajectory,
i.e., t vs. y1 , y2 , . . . , yn ;

odephas2′ which generates a plot of y1 vs. y2 , i.e., Y(:,1) vs. Y(:,2);

odephas3′ which generates a plot of y1 vs. y2 vs. y3 , i.e., Y(:,1) vs. Y(:,2)
vs. Y(:,3).
It is possible to plot different components of y using OutputSel.
OutputSel A vector containing the components of Y which are to be passed to the function speci-
fied by the OutputFcn parameter.
Refine Refines the times which are output in t. This integer value increases the number of
times by this factor. (Default value: 1 for all ode solvers except ode45, 4 for ode45.)
RelTol The relative error tolerance. (Default value: 10−3 ).
Stats Whether statistics about the run are output on the terminal (value: ′ on′ ) after the
trajectory is calculated or they are not (value: ′ off′ ). (Default value: ′ off′ .)

For example, if you want to use ode45 with the relative error tolerance set to 10−6 for Duffing’s equa-
tion, enter
>> params = odeset(′RelTol′, 1.e-6);
>> [t, Y] = ode45(duffing a, tspan, y0, params);
The trajectory will be more accurate — but the command will run slower. If you also want the statistics
on the performance of the particular ode solver used, enter
>> params = odeset(′RelTol′, 1.e-6, ′ Stats′ , ′ on′ );
>> [t, Y] = ode45(@duffing a, tspan, y0, params);
and the number of successful steps, the number of failed steps, and the number of times f(t, y) was eval-
uated will be printed on the terminal. This might be useful in “optimizing” the performance of the ode
solver if the command seems to be running excessively slowly. For implicit methods where the Jacobian
needs to be calculated, the number of times the Jacobian was evaluated, the number of LU decompositions,
and the number of times the linear system was solved will also be returned.
The ode solver can also record the time and the location when the trajectory satisfies a particular con-
dition: this is called an event. For example, if we are calculating the motion of the earth around the sun,
we can determine the position of the earth when it is closest to the sun and/or farthest away; or, if we are
following the motion of a ball, we can end the calculation when the ball hits the ground — or we can let it
continue bouncing. Enter
>> ballode
to see a simple example.
For example, suppose we want to record where and when a trajectory of Duffing’s equation passes
through y1 = ±0.5. That is, we define an “event” to be whenever the first component of y passes

101
10.2. Advanced Commands

through −0.5 or +0.5. This can be done by modifying the primary function duffing ode and replac-
ing the ode45 statement by
params = odeset(′RelTol′, 1.e-6, ′ Events′, @duffing event);
[t, Y, tevent, Yevent, indexevent] = ode45(@duffing n, tspan, y0, params);
where we create a new nested function in the primary function duffing ode.
function [value, isterminal, direction] = duffing event(t, y)
value = [y(1)+0.5; y(1)-0.5]; % check whether y(1) passes through ±0.5
isterminal = [0; 0]; % do not halt when this occurs
direction = [0; 0]; % an event occurs when y(1) passes through
% zero in either direction
end
Note that we can define the right-hand side of Duffing’s equation by using duffing a, duffing n,
duffing p, or duffing p2 and duffing c. We have chosen duffing n since we have created the nested
function duffing event. (We could let duffing event be a primary function, but there is no reason to
do so.)
There are a number of steps we have to carry out to turn “events” on. First, we have to use the odeset
command. However, this only tells the ode solver that it has to watch for one or more events; it does not
state what event or events to watch for. Instead, we describe what an event is in this new function. Three
vector arguments are output:
value – A vector of values which are checked to determine if they pass through zero during a time
step. No matter how we describe the event, as far as the ode solver is concerned an event
only occurs when an element of this vector passes through zero. In some cases, such as this
example it is easy to put an event into this form. In other cases, such as determining the
apogee and perigee of the earth’s orbit, the calculation is more complicated.
isterminal – A vector determining whether the ode solver should terminate when this particular
event occurs: 1 means yes and 0 means no.
direction – A vector determining how the values in value should pass through zero for an event to
occur:
1 means the value must be increasing through zero for an event to occur,
−1 means the value must be decreasing through zero for an event to occur, and
0 means that either direction triggers an event.
The final step is that the left-hand side of the calling statement must be modified to
[t, Y, tevent, Yevent, index event] = ode45(...);
Any and all events that occur are output by the ode solver through these three additional variables:
tevent is a vector containing the time of each event,
Yevent is a matrix containing the location of each event, and
index event is a vector containing which value in the vector value passed through zero.
If the result is stored in the structure sol, the new output is
sol.xe is a vector containing the time of each event,
sol.ye is a matrix containing the location of each event, and
sol.ie is a vector containing which value in the vector value passed through zero.
Since the function duffing event might appear confusing, we now discuss how an event is actually
calculated. At the initial time, t and y are known and duffing event is called so that the vector
 (0) 
(0) y1 + 0.5
e = (0) ,
y1 − 0.5

i.e., value, can be calculated. In addition, isterminal and direction are returned. Next, duffing is
called and the solution y(1) is calculated at time t(1) . duffing event is called again and e(1) is calcu-
lated and compared elementwise to e(0) . If the values have different signs in some row, then direction
is checked to determine if the values are passing through zero in the correct direction or if either direction
is allowed. If so, the time at which the element is zero is estimated and the ode is solved again to obtain

102
10.2. Advanced Commands

a more accurate estimate. This procedure continues until the zero is found to the desired accuracy. Then
isterminal is checked to see if the run should be continued or should be stopped.
Another interesting ode is van der Pol’s equation

y ′′ − µ(1 − y 2 )y ′ + y = 0

where µ > 0 is the only parameter. As a first order system it is


 ′  
y1 y2
=
y2 µ(1 − y12 )y2 − y1

and its Jacobian is  


0 1
J= .
−2µy1 y2 − 1 µ(1 − y12 )
The right-hand side can be coded as a nested function inside a primary function by
function vdp ode(mu, y0, final time)
ode45(@vdp n, [0 final time], y0);
function deriv = vdp n(t, y)
% vdp n: van der Pol’s equation
deriv = [y(2); mu*(1-y(1) 2)*y(2)-y(1)];
end
ˆ
end
This is not stiff unless µ is “large”. For example, let µ = 1 and solve the ode with initial conditions
y(0) = 1 and y ′ (0) = 0 for t ∈ [0, 100] using ode45. Then, plot the result and note the number of ele-
ments in t. Repeat this procedure using µ = 10 and increase the final time, if necessary, so that you still
see a few complete oscillations. Then let µ = 100, etc., until the time required to plot a few oscillations
becomes “very large”. Then use ode15s and note the huge difference in the time required. ADD ALL
VDPS???
There is no need to use the ode solver parameters JPattern or Jacobian in this example because this
ode is so “nice”. However, since they might be needed for a nastier ode, we include them by using
Vdp pattern = sparse([1 2 2], [2 1 2], [1 1 1], 2, 2);
params = odeset(′Jacobian′, @vdpj n, ′ JPattern′, Vdp pattern):
[t, Y] = <ode solver>(@vdp n, tspan, y0, opt);
where the Jacobian is calculated numerically using the nested function
function J = vdpj n(t, y)
% vdpj n: Jacobian for van der Pol’s equation
J = [ 0 1; -2*mu*y(1)*y(2)-1 mu*(1-y(1) 2) ];
end
ˆ
for the elements determined by Vdp pattern. By the way, if we use the property JPattern but not
Jacobian then the Jacobian is calculated numerically just for the elements determined by the sparse ma-
trix.
Note: Plotting the trajectory by
plot(t, Y)
is not very instructive. Instead, use
subplot(2,1,1)
plot(t, Y(:,1))
subplot(2,1,2)
plot(t, Y(:,2))
Our final example is slightly more complicated. Suppose we kick a ball into the air with initial speed
s and at an angle of α, and we want to follow its motion until it hits the ground. Let the x axis be the
horizontal axis along the direction of flight and z be the vertical axis. Using Newton’s laws we obtain the
ode system
x′′ = 0 and z ′′ = −g

103
10.2. Advanced Commands

where g = 9.8 meters/second is the acceleration on the ball due to the earth’s gravity. The initial condi-
tions are
x(0) = 0 , x′ (0) = s cos α , z(0) = 0, and z ′ (0) = s sin α
where we assume, without loss of generality, that the center of our coordinate system is the initial location
of the ball. We also want to determine four “events” in the ball’s flight: the highest point of the trajectory
of the ball and the time it occurs, the distance it travels and the time it hits the ground, and the x val-
ues and times when the ball reaches the height h > 0. But beware because the ball may never attain this
height!
Although these odes can be solved analytically (consult any calculus book), our aim is to give an ex-
ample of how to use many of the advanced features of MATLAB’s ode solvers. (If we would include the
effects of air resistance on the ball, then these odes would become nonlinear and would not be solvable
analytically.) We convert Newton’s laws to the first-order system
 ′  
y1 y2
 y2   0 
 = 
y3 y4
y4 −g

by letting y1 = x, y2 = x′ , y3 = z, and y4 = z ′ The initial conditions are

y1 (0) = 0 , y2 (0) = s cos α , y3 (0) = 0 , and y4 (0) = s sin α .

One complication with solving this system numerically is that we do not know when the ball will hit
the ground, so we cannot give the final time. Instead, we use a time, 10s/g which is much greater than
needed and we let the program stop itself when the ball hits the ground. In addition, we want the relative
error to be 10−6 . Finally, we want the trajectory (i.e., z vs. x ) to be plotted automatically.
The following is a completely self-contained example using nested functions.

104
10.2. Advanced Commands

function [times, values] = gravity ode(speed, angle, height)


% gravity ode: The trajectory of a ball thrown from (0,0) with initial
% speed and angle (in degrees) given.
% times: (1) = time ball at peak, (2) = time ball hits ground
% (3,4) = time ball attains height
% values: (1) = z value at peak, (2) = x value when ball hits ground
% (3,4) = x values when ball attains height
% Note: (3,4) will not be used if height > z value at peak
g = 9.8;
gravity init()
[t, Y, tevent, Yevent, index event] = ode45(@gravity, tspan, y0, params);
if length(t) == 2
times = tevent;
values = [Yevent(3,1) Yevent(1,2)];
else
times = tevent([2 4 1 3]);
values = [Yevent(3,2) Yevent(1,4) Yevent(1,1) Yevent(1,3)];
end
%%%%% nested functions follow
function gravity init
% gravity init: Initialize everything
tspan = [0 10*speed/g];
y0 = [ 0; speed*cos(angle*pi/180); 0; speed*sin(angle*pi/180) ];
params = odeset(′RelTol′, 1.e-6, ...

Events′, @gravity event, ...

Refine′, 20, ...

OutputFcn′, ′ odephas2′, ...

OutputSel′, [1 3]);
end
function deriv = gravity(t, y)
% gravity: Calculates the right-hand side of the ode
deriv = [y(2); 0; y(4); -g];
end
function [value, isterminal, direction] = gravity event(t, y)
% gravity event: determines the events
value = [y(3) y(3)-height y(4)]; % z = 0, z-height = 0, z’ = 0
isterminal = [1 0 0]; % halt only when z = 0
direction = [-1 0 -1]; % an event occurs when z or z’ decrease through 0
% or z-height passes through 0 in either direction
end
end
Note that the parameters g, speed, angle, and height do not need to be passed into the nested func-
tions. Similarly, tspan, y0, and params do not need to be passed out.
MATLAB also has the function ode15i which solves fully implicit odes. It is very similar to the func-
tions we have already discussed, but there is one important difference. Although it is a very powerful
function, we only provide a very simple example which uses it.
We consider a linear second-order ode in a neighborhood of a regular singular point. Consider the ode

P (t)y ′′ (t) + Q(t)y ′ (t) + R(t)y(t) = 0

where P (t), Q(t), and R(t) are polynomials with no common factors. The singular points of this
ode are the values of t for which P (t) = 0. If t0 is a singular point, it is a regular singular point if
limt→t0 (t − t0 )Q(t)/P (t) and limt→t0 (t − t0 )2 R(t)/P (t). A “common” ode of this type is Bessel’s equa-
tion
t2 y ′′ (t) + ty ′ (t) + (t2 − n2 )y(t) = 0 for t ≥ 0 (10.1)

105
11. Boundary-Value Ordinary Differential Equations

where n is a nonnegative integer and the initial condition is given at t = 0. The solution is denoted by
Jn (t) and, for specificity, we will concentrate on n = 1. At t = 0 the ode reduces to −y(0) = 0 and so
we require y(0) = 0. The free initial condition is y ′ (0) and for this example we choose y ′ (0) = 1.
If we write Bessel’s equation as
 
′′ 1 ′ n2
y (t) + y (t) + 1 − 2 y(t) = 0 (10.2)
t t

we clearly have a problem at t = 0 and for t ≈ 0. The ode solvers we discussed previously can han-
dle (10.2) for t ≥ 1 with the initial conditions that y(1) and y ′ (1) are given. However, a completely
different method of solution is required for t ≥ 0 and the form (10.1) is preferred to (10.2).
When we convert Bessel’s equation to the first order system we again let y1 (t) = y(t) and y2 (t) = y ′ (t)
and leave the t2 in the numerator to obtain
 ′   
y1 y2
=
t2 y2′ −ty2 − (t2 − 1)y1

Previously, we have always written the first-order system as y′ = f(t, y), but this form has a problem
when t = 0. Instead, we write it as g(t, y, y′ ) = 0 so that
 
y1′ − y2
g(t, y, y′ ) = .
t y2 + ty2 + (t2 − 1)y1
2 ′

Finally, we not only have to input the initial condition y(0) = (0, 1)T , but we also have to input
T
y′ (0) = y1′ (0), y2′ (0) . It is easy to calculate y1′ (0) = y2 (0), but y2′ (0) ≡ y ′′ (0) is more complicated.
Differentiate (10.1) with respect to t to obtain

t2 y ′′′ (t) + 3ty ′′ (t) + t2 y ′ (t) + 2ty(t) = 0

and differentiate it again to obtain

t2 y ′′′′ (t) + 5ty ′′′ (t) + (t2 + 3)y ′′ (t) + 4ty ′ (t) + 2y(t) = 0 .

Now set t = 0 to obtain y ′′ (0) = 0. We can solve Bessel’s equation for t ∈ [0, 10] by
>> g = @(t, y, yp) [yp(1)-y(2);t 2 *yp(2)+t*y(2)+(t 2 -1)*y(1)];
>> tspan = [0 10]
ˆ ˆ
>> y0 = [0;1]
>> yp0 = [1;0]
>> [t,Y] = ode15i(g, tspan, y0, yp0)
>> plot(t, Y(:,1))

Implicit ODE Solver

ode15i Stiff ode solver for the fully implicit ode f(t, y, y ′ ) = 0.

11. Boundary-Value Ordinary Differential Equations


In addition to initial-value ordinary differential equations there is a second type of odes that MATLAB
can solve numerically. Boundary-value odes are also odes of the form

d
y = f(x, y) for x ∈ [a, b]
dx
106
11. Boundary-Value Ordinary Differential Equations

where y ∈ Rn but conditions are given at both ends of the interval. If the boundary conditions are sep-
arated, then k conditions are given at x = a and n − k other conditions are given at x = b. If the
boundary conditions are non-separated, then the conditions at x = a and at x = b are related. To allow
any of these boundary conditions we write the boundary conditions as φ y(a), y(b) = 0 where φ ∈ Rn .
For simplicity, we will only consider two closely related second-order odes, i.e., n = 2. This example
should enable you to study any boundary-value ode. Consider the two nonlinear boundary-value ordinary
differential equations

d2 y dy
(x) + 2 (x) + ǫey(x) = 0 (11.1a)
dx2 dx
and
d2 y dy
ǫ (x) + 2 (x) + ey(x) = 0 (11.1b)
dx2 dx

for x ∈ [0, 1] where ǫ > 0. Our boundary conditions are


 
 y(0)
φ y(0), y(1) = =0, (11.2)
y(1)

which are called Dirichlet boundary conditions. These two odes are quite simple, but also quite interesting
and challenging to solve for certain intervals in ǫ.
We could use the Neumann boundary conditions y ′ (0) = 4 and y ′ (1) = −7 by
 
 y ′ (0) − 4
φ y(0), y(1) = =0. (11.3)
y ′ (1) + 7

Or we could use the mixed boundary conditions y(0) − y ′ (0) = 1 and y(1) + 2y ′ (1) = 3 by
 
 y(0) − y ′ (0) − 1
φ y(0), y(1) = =0. (11.4)
y(1) + 2y ′ (1) − 3

Finally, we could use periodic boundary conditions, which are non-separated, by


 
 y(1) − y(0)
φ y(0), y(1) = =0. (11.5)
y ′ (1) − y ′ (0)

The primary MATLAB function is bvp4c. However, the functions bvpinit and deval are also
needed. We solve the boundary value problem by
>> sol = bvp4c(<right-hand side>, <boundary conditions>, <initial guess>)
There are two functions we need to write: odefun is f(x, y) and bcfun is the boundary conditions. For
our example the ode given by
function yp = nlode(x, y)
global which ode eps
if which ode == 1
yp = [y(2); -eps*exp(y(1))-2*y(2)];
else
yp = [y(2); -(exp(y(1))+2*y(2))/eps];
end
where we use global to input which ode to use and ǫ. The boundary condition is given by
function bc = nlode bc(ya, yb)
bc = [ya(1); yb(1)];
Since these boundary conditions are particularly simple, we also include the function
function bc = nlode bc2(ya, yb)
bc = [ya(1)-ya(2)-1; yb(1)+2*yb(2)-3];

107
11. Boundary-Value Ordinary Differential Equations

for mixed boundary conditions (11.4). In addition, we have to choose an initial guess for y(x) using
bvpinit by either
>> bvpinit(x, y init)
or
>> bvpinit(x, <initial guess function>)
For example, if we want the initial iterate to be a parabola which is zero at x = 0 and 1 and has maxi-
mum value A then y(x) = y1 (x) = 4Ax(1 − x) and y ′ (x) = y2 (x) = 4A(1 − 2x) then we can write
>> x = linspace(0, 1, 21);
>> solinit = bvpinit(x, @nlode y ic);
where nrode y ic is written as
function y ic = nlode y ic(x)
global A
y ic = [4*A*x.*(1 - x); 4*A*(1-2*x)];
The only alternative is to write
>> x = linspace(0, 1, 21);
>> y1 val = ???;
>> y2 val = ???;
>> solinit = bvpinit(x, [y1 val; y2 val]);
where y1 val and y2 val are scalar values. Thus the initial guess is y1 = y1 val*ones(size(x)) and y2
= y2 val*ones(size(x)). This is often unacceptable because constant initial guesses may be so far from
the solution that convergence cannot be obtained. What we would like to do is
>> x = linspace(0, 1, 21);
>> y1 = 4*A*x.*(1 - x);
>> y2 = 4*A*(1 - 2*x);
>> solinit = bvpinit(x, [y1; y2]); % WRONG
This fails because y1 and y2 must be scalar variables and not vectors. If you really, really need y1 and
y2 to be vectors, then do not use bvpinit. Instead, specify the structure solinit directly by entering
>> x = linspace(0, 1, 21);
>> y1 = 4*A*x.*(1 - x);
>> y2 = 4*A*(1 - 2*x);
>> solinit.x = x;
>> solinit.y = [y1;y2];
Warning: This is dangerous because future versions of Matlab might change the fieldnames of the struc-
ture solinit. However, it works for now.
We are finally ready to solve this ode by
>> global which ode eps
>> global A
>> which ode = 1;
>> A = 1;
>> eps = 3;
>> x = linspace(0, 1, 21);
>> solinit = bvpinit(x, @nlode y ic);
>> sol = bvp4c(@nlode, @nlode bc, solinit);
The solution is contained in sol and is extracted by deval. For example, if xi = (i − 1)∆x where
x1 = 0 and xn = 1 then we determine, and plot, the numerical solution y by
>> xpt = linspace(0, 1, 101);
>> Ypt = deval(sol, xpt);
>> plot(xpt, Ypt(1,:), xpt, Ypt(2,:), ′ r′ )
Having done all this work, we now combine everything into the function m-file nlode all to show how

108
11. Boundary-Value Ordinary Differential Equations

much easier it is to use nested functions and to combine everything into one primary function.
function sol = nlode all(which ode, eps, A, nr points)
% nlode all: boundary-value solver using bvp4c
x = linspace(0, 1, nr points);
solinit = bvpinit(x, @nlode y ic);
sol = bvp4c(@nlode, @nlode bc, solinit);
xpt = linspace(0, 1, 101);
Ypt = deval(sol, xpt);
plot(xpt, Ypt(1,:), xpt, Ypt(2,:), ′ r′ )
%%%%% nested functions follow
function y ic = nlode y ic(x)
y ic = [4*A*x.*(1 - x); 4*A*(1-2*x)];
end
function yp = nlode(x, y)
if which ode == 1
yp = [y(2); -eps*exp(y(1))-2*y(2)];
else
yp = [y(2); -(exp(y(1))+2*y(2))/eps];
end
end
function bc = nlode bc(ya, yb)
bc = [ya(1); yb(1)];
end
end
This m-file is easy to read and easy to debug and easy to modify. Also, the solution is returned so it can
be used in the MATLAB workspace.
The reason we chose these particular odes is to “check out” bvp4c. For the ode (11.1a) there are two
solutions for 0 ≤ ǫ / 3.82 and no solutions for ǫ ' 3.82. This is a good test of any boundary-value
solver. For the ode (11.1a) there is one solution for any ǫ > 0. The “interesting” feature of this ode is
that for ǫ ≪ 1 the solution rises rapidly from y(0) = 0 to y(x) ≈ log 2 for x = O(ǫ) and then decays
gradually so that y(1) = 0. It is very challenging for a boundary-value solver to be able to capture this
rapid rise.
One final point needs to be emphasized. Sometimes, any “halfway decent” initial choice of y will con-
verge to a solution. In fact, this is true for our example — but it is not true for many examples. Some-
times it takes a “good” initial choice to obtain convergence; a “bad” choice will never converge to the de-
sired solution. The standard method to use to obtain a “good” initial iterate is the continuation method.
Frequently there are values of the parameter(s) for which “good” initial iterates are known. For example,
for the ode (11.1a) if ǫ ≪ 1 we can approximate ǫey by the Taylor series expansion ǫ(1 + y) and solve
the resulting linear ode. If ǫ = 0.1 the resulting analytical solution is a very good approximation to the
numerical solution. You can use this solution as the initial guess for ǫ = 0.2. The numerical solution can
then be used as an initial guess for a larger value of ǫ, etc.
The only difficulty with this method is that there might be more solutions. When ǫ = 0.1 there is a
second solution whose maximum is over 8. For this solution y ′ (0) ≈ 35 which indicates how rapidly the
solution is growing at the left endpoint. This solution can only be found by trying “large” initial guesses
(e.g., choosing A to be large in nlode y ic ).
For the ode (11.1b) it is very difficult to determine “good” initial guesses when ǫ ≪ 1 since the solu-
tion grows so rapidly. Again, the continuation method is very helpful. Start with a “large” value of ǫ,
say ǫ = 1, and choose a “reasonable” initial guess. (Since the two odes are identical when ǫ = 1 you can
use the solution you found to ode (11.1a).) Then slowly decrease ǫ. For example, when ǫ = 0.01 we have
y ′ (0) ≈ 130 and when ǫ = 0.001 we have y ′ (0) ≈ 1300. In conclusion, we want to remind you that for
the odes we have discussed here almost any “halfway reasonable” initial choice for the ode (11.1a) will
converge to one of the two solutions and for the ode (11.1b) will converge to the single solution. However,
you might well find an ode for which this is not true.

109
12. Polynomials and Polynomial Functions

Boundary-Value Solver

bvp4c(<right-hand side>, Numerically solves y′ (x) = f(x, y) for x ∈ [a, b]


<boundary conditions>, with given boundary conditions and an initial
<initial guess>) guess for y. The user supplied functions are
f(x, y) = right hand side(x,y) and
boundary conditions(ya,yb) where ya = y(a) and
yb = y(b).
o
bvpinit(x,y) Calculates the initial guess either by giv-
bvpinit(x, <initial guess function>) ing y directly or by using a function y =
initial guess function(x).
deval(x,y soln) Interpolate to determine the solution at x.

12. Polynomials and Polynomial Functions


In MATLAB the polynomial

p(x) = c1 xn−1 + c2 xn−2 + . . . + cn−1 x + cn .

is represented by the vector q = (c1 , c2 , . . . , cn )T . You can easily calculate the roots of a polynomial by
>> r = roots(q)
Conversely, given the roots of a polynomial you can recover the coefficients of the polynomial by
>> q = poly(r)
Warning: Note the order of the coefficients in the polynomial. c1 is the coefficient of the highest power of
x and cn is the coefficient of the lowest power, i.e., 0.
The polynomial can be evaluated at x by
>> y = polyval(q, x)
where x can be a scalar, a vector, or a matrix. If A is a square matrix, then

p(A) = c1 An−1 + c2 An−2 + . . . + cn−1 A + cn

is calculated by
>> polyvalm(q, A)
(See section 15 for more details on this type of operation.)
A practical example which uses polynomials is to find the “best” fit to data by a polynomial of a partic-
ular degree. Suppose the data points are

(−3, −2), (−1.2, −1), (0, −0.5), (1, 1), (1.8, 2)

and
 we want to find the “best” fit by a straight line. Defining the data points more abstractly as
(xi , yi ) i = 1, 2, . . . , n and the desired straight line by y = c1 x + c2 , the matrix equation for the
straight line is
   
x1 1   y1
 x2 1  c1  y2 
 .. ..  =  ..  .
. . c2 .
xn 1 yn
In general, there is no solution to this overdetermined linear system. Instead, we find the least-squares
solution c = (c1 , c2 )T by
>> c = [x ones(n, 1)] \ y

110
12. Polynomials and Polynomial Functions

We can plot the data points along with this straight line by
>> xx = linspace(min(x), max(x), 100);
>> yy = polyval(c, xx);
>> plot(xx, yy, x, y, ′ o′ )
We can find the “best” fit by a polynomial of degree m < n, i.e., y = c1 xm + c2 xm−1 + . . . + cm+1 , by
calculating the least-squares solution to
Vc = y
where
xm xm−1 ...
   
1 1 x1 1 c1
 xm xm−1 ... x2 1 c 
V =  ..2 2.
.. ... ...  and c =  ..2  .
. .
xmn
m−1
xn ... xn 1 cn
The matrix V is called a Vandermonde matrix . The statement
>> V = vander(x);
generates the square Vandermonde matrix with m = n − 1. To generate the n×(m − 1) Vandermonde
matrix we want, enter
>> V = vander(x)
>> V(: , 1:m-1) = [];
This entire procedure can be carried out much more easily by entering
>> q = polyfit(x, y, m-1)
where the third argument is the order of the polynomial (i.e., the number of coefficients in the polyno-
mial).
Warning: The Vandermonde matrix is approximately as badly conditioned as the Hilbert matrix
which was discussed in section 5.2. For example, cond(vander([1 : 10])) = 2×1012 whereas
cond(hilb(10)) = 2×1013 .
You can also find a local maximum or minimum of the polynomial p(x) by finding the zeroes of p′ (x).
The coefficients of p′ (x) are calculated by
>> polyder(q)
where q is the vector of the coefficients
 of p(x).
Given a set of data points (xi , yi ) there is sometimes a need to estimate values that lie within these
data points (this is called interpolation) or outside them (this is called extrapolation). This estimation is
generally done by fitting data which is “near” the desired value to a polynomial and then evaluating this
polynomial at the value.
There are a number of commands to interpolate data  points in any number
of dimensions. The simplest
command in one dimension to interpolate the points (xi , yi ) 1 ≤ i ≤ n is
>> yvalues = interp1(x, y, xvalues, <method>)
where xvalues is a vector of the values to be interpolated, yvalues is the vector of the interpolated
values, and <method> is an optional argument specifying the method to be used. One additional require-
ment for this command is that the elements of x are monotonic, i.e., either all in increasing order or in
decreasing order, to make it easy for the function to determine which data points are “near” the desired
value. Five of the interpolation methods which can be used are the following:

nearest′: The interpolated value is the value of the nearest data point.

linear′: Linear splines are used to connect the given data points. That is, straight lines connect
each pair of adjacent data points. (This is the default.)

spline′: Cubic splines are used to connect the given data points. That is, cubic polynomials con-
nect each pair of adjacent data points. The additional constraints needed to obtain unique
polynomials are that the the two polynomials which overlap at each interior data point
have the same first and second derivatives at this point.

pchip′ : Piecewise cubic Hermite polynomials connect each pair of adjacent data points. This is simi-
lar to spline but the second derivatives need not be continuous at the interior data points.
Instead, this interpolation is better at preserving the shape of the data. In particular, on in-
tervals where the data is monotonic so is the piecewise polynomial, and on intervals where
the data is concave up or down so is the piecewise polynomial.

111
13. Numerical Operations on Functions


cubic′ : The same as pchip.
An alternate way to interpolate these points is by using the two commands
>> pp = spline(x, y)
>> yvalues = ppval(pp, xvalues)
to generate and interpolate the cubic spline or
>> pp = pchip(x, y)
>> yvalues = ppval(pp, xvalues)
to generate and interpolate the piecewise cubic Hermite polynomials. The first command generates the
structure pp which contains all the information required to obtain a unique piecewise polynomial. The
second command interpolates the piecewise polynomial at the x values given by the vector xvalues.
Interpolation really means interpolation. If a value lies outside the interval [x1 , xn ] then, by default,
NaN is returned. This can be changed by adding a fifth argument:
• If the fifth argument is a number, this value is returned whenever the value lies outside the interval.
• If the fifth argument is ′ extrap′ , extrapolation (using the same method) is used.
The command spline can be used instead of using interp1 with the method spline. With it you
can specify precisely the boundary conditions to use. Similarly, the command pchip can be used instead
of using interp1 with the method pchip or cubic.

Polynomial Functions

interp1(x, y, xvalues, <method>) Interpolates any number of values using the given data
points and the given method.
interp2 Interpolates in two dimensions.
interp3 Interpolates in three dimensions.
interpn Interpolates in n dimensions.
pchip Cubic Hermite interpolation.
poly(<roots>) Calculates the coefficients of a polynomials given its roots.
polyder(q) Calculates the derivative of a polynomial given the vector of
the coefficients of the polynomial.
polyfit(x, y, <order>) Calculates the coefficients of the least-squares
 polynomial
of
a given order which is fitted to the data (xi , yi ) .
polyval(q, x) Evaluates the polynomial p(x).
polyvalm(q, A) Evaluates the polynomial p(A) where A is a square matrix.
ppval evaluates the piecewise polynomial calculated by pchip or
spline.
roots(q) Numerically calculates all the zeroes of a polynomial given
the vector of the coefficients of the polynomial.
spline Cubic spline interpolation.
vander Generates the Vandermonde matrix.

13. Numerical Operations on Functions


MATLAB can also find a zero of a function by
>> fzero(<function handle>, x0)

>> fzero(<function handle>, x0)


x0 is a guess as to the location of the zero. Alternately,
>> fzero(<function handle>, [xmin xmax])

112
13. Numerical Operations on Functions

finds a zero in the interval x ∈ (xmin, xmax) where the signs of the function must differ at the endpoints
of the interval.
Note: The function must cross the x-axis so that, for example, fzero cannot find the zero of the function
f (x) = x2 .
The full argument list is
>> fzero(<function handle>, xstart, <options>)
where xstart is either x0 or [xmin xmax], as we discussed previously. We can “tune” the zero find-
ing algorithm by using the function optimset to create a structure which changes some of the default
parameters for fzero. That is,
>> opt = optimset(′<Prop 1>′ , <Value 1>, ′ <Prop 2>′ , <Value 2>, ...)
changes the options included in the argument list and
>> fzero(<function handle>, xstart, opt, <arg 1>, <arg 2>, . . . )
executes fzero with the new options. Enter
>> help optimset
for a discussion of how optimset works and
>> optimset(@fzero)
to see the default parameters.
Frequently, the function will have parameters that need to be set. For example, we can find a zero of
the function f (x) = cos ax + bx by using an anonymous function
>> a = ???;
>> b = ???;
>> fcos a = @(x) cos(a*x) + b*x;
>> yzero = fzero(fcos a, xstart);
or by using a nested function
function fzero example(a, b, xstart)
yzero = fzero(@fcos n, xstart);
function y = fcos n(x)
% fcos: f(x) = cos(a*x) + b*x
y = a*cos(x) + b*x;
end
end
It sometimes happens that the function has already been coded in a separate file, i.e., it is a primary
function m-file, such as
function y = fcos p(x, a, b)
% fcos: f(x) = cos(a*x) + b*x
y = a*cos(x) + b*x;

Then we can use closure, as already discussed in section [Macro:[ode: basic]chap], so that the parame-
ters can be set outside of fzero. This is easily done by entering
>> a = ???;
>> b = ???;
>> fcos c = @(x) fcos p(x, a, b);
>> yzero = fzero(fcos p, xstart);
The parameters a and b are determined when the function fcos c is generated and so are passed indi-
rectly into fzero
MATLAB can also find a local minimum of a function of a single variable in an interval by
>> fminbnd(<function handle>, xmin, xmax)
As with fzero, the full argument list is
>> fminbnd(<function handle>, xmin, xmax, options)
MATLAB can also find a local minimum of a function of several variables by
>> fminsearch(<function handle>, iterate0)
where iterate0 is a vector specifying where to begin searching for a local minimum. For example, if we
enter

113
13. Numerical Operations on Functions

>> fnctn = @(x) (x(1) - 1) 2 + (x(2) + 2) 4;


>> fminsearch(fnctn, [0 0]′ )
ˆ ˆ
we obtain (1.0000 − 2.0003)T (actually (1.00000004979773, −2.00029751371046)T ). The answer might
not seem to be very accurate. However, the value of the function at this point is 1.03×10−14 , which is
quite small. If our initial condition is (1, 1)T , the result is (0.99999998869692, −2.00010410231166)T.
Since the value of funct at this point is 2.45×10−16 , the answer is about as accurate as can be expected.
In other words, the location of a zero and/or a local minimum of a function might not be as accurate as
you might expect. Be areful. To determine the accuracy MATLAB is using to determine the minimum
value type
>> optimset(@fminsearch)
The value of TolX, the termination tolerance on x, is 10−4 and the value of TolFun, the termination
tolerance on the function value, is the same.
There is no direct way to find zeroes of functions of more than one dimension. However, it can be done
by using fminsearch. For example, suppose we want to find a zero of the function
 
x1 + x2 + sin(x1 − x2 )
f(x) = .
x1 − x2 + 2 cos(x1 + x2 )

Instead, we can find a minimum of g(x) = f12 (x) + f22 (x). If the minimum value is 0, we have found
a zero of f — if it is not zero, we have not found a zero of f. For example, if f is defined an an anony-
mous function the result of
>> xmin = fminsearch(f, [0 0])
is xmin = (−.1324 . . . , 1.0627 . . . ). We are not done since we still have to calculate g(xmin ). This is
≈ 2.4×10−9 which is small — but is it small enough? We can decrease the termination tolerance by
>> opt = optimset(′TolX′ , 1.e-8, ′ TolFun′, 1.e-8)
>> xmin = fminsearch(f, [0 0], opt)
Since g(xmin ) = 2.3×10−17 we can assume that we have found a zero of f.
MATLAB can also calculate definite integrals using three commands. The first is quad which uses
Rb
adaptive Simpson’s method. To evaluate a f (x) dx by Simpson’s method enter
>> quad(<function handle>, a, b)
The full argument list is
>> quad(<function handle>, a, b, tol, trace)
where tol sets the relative tolerance for the convergence test and information about each iterate is
printed if trace is non-zero.
The second is quadl which uses adaptive Gauss-Lobatto quadrature, which is a variant of Gauss
quadrature.
quadl uses the more accurate formula and so should require many fewer function evaluations. For ex-
ample, quad calculates the exact integral (up to round-off errors) for polynomials of degree five whereas
quadl calculates the exact integral (up to round-off errors) for polynomials of degree nine.
The third is quadgk which uses adaptive Gauss-Kronrod quadrature. This command is more general
than the previous two because it is much more general:
• The interval can be half-infinite (i.e., a = −∞ or b = +∞ ) or fully infinute ( a = −∞ and
b = +∞ ).
• In addition the integrand can have an integrable singularity.
For example, the error in
>> f = @(x) 1./(1 + x. 2);
>> quadgk(f, 0, inf)
ˆ
is 8.8818×10−16 (the actual value is 2π ) and the error in
>> g = @(x) exp(sqrt(x))/sqrt(x);
>> quadgk(g, 0, 1)
is −2.3670×10−13 (the actual value is 2(e − 1) ).
MATLAB can also calculate the double integral
Z xmax Z ymax
f (x, y) dxdy
xmin ymin

114
14. Discrete Fourier Transform

by
>> dblquad(<function handle>, xmin, xmax, ymin, ymax)
It can also calculate the triple integral
Z xmax Z ymax Z zmax
f (x, y, z) dxdy dz
xmin ymin zmin

by
>> triplequad(<function handle>, xmin, xmax, ymin, ymax, zmin, zmax)

Numerical Operations on Functions

dblquad(<function handle>, a, b, c, d) Numerically evaluates a double integral.


fminbnd(<function handle>, xmin, xmax) Numerically calculates a local minimum of a one-
dimensional function given the endpoints of the
interval in which to search
fminsearch(<function handle>, iterate0) Numerically calculates a local minimum of a multi-
dimensional function given the the initial iterate
vector.
fzero(<function handle>, x0) Numerically calculates a zero of a function given
the initial iterate. x0 can be replaced by a 2-
vector of the endpoints of the interval in which a
zero lies.
optimset Allows you to modify the parameters used by
fzero, fminbnd, and fminsearch.
quad(<function handle>, a, b) Numerically evaluates an integral using Simpson’s
method.
quadgk(<function handle>, a, b) Numerically evaluates an integral using the adap-
tive Gauss-Kronrod method. The interval can be
infinite and/or the function can have an integrable
singularity.
quadl(<function handle>, a, b) Numerically evaluates an integral using the adap-
tive Gauss-Lobatto method.

14. Discrete Fourier Transform


There are a number of ways to define the discrete Fourier transform; we choose to define it as the dis-
cretization of the continuous Fourier series. In this section we show exactly how to discretize the contin-
uous Fourier series and how to transform the results of MATLAB’s discrete Fourier transform back to
the continuous case. We are presenting the material in such detail because there are a few slightly differ-
ent definitions of the discrete Fourier transform; we present the definition which follows directly from the
real Fourier series. xdi A “reasonable” continuous function f which is periodic with period T can be
represented by the real trigonometric series
∞  
X 2πkt 2πkt
f (t) = a0 + ak cos + bk sin for all t ∈ [0, T ] (14.1)
T T
k=1

115
14. Discrete Fourier Transform

where Z T
1
a0 = f (t) dt
T 0
Z T

2 
ak = f (t) cos kt dt 


T 0
Z T
for k ∈ N[1, ∞) .
2 

bk = f (t) sin kt dt 

T 0

The coefficients a0 , a1 , a2 , . . . and b1 , b2 , . . . are called the real Fourier coefficients of f , and ak and bk
are the coefficients of the k-th mode. The power of the function f (t) is†
Z T
1
f (t) 2 dt
P =
T 0

so that ∞
1X 
P = |a0 |2 + |ak |2 + |bk |2 .
2
k=1

The power in each mode, i.e., the power spectrum, is



|a0 |2 if k = 0
Pk = 1 2 2

2 |ak | + |bk | if k > 0

and the frequency of the k-th mode is k/T cycles per unit time.
Since
eiαt + e−iαt eiαt − e−iαt
cos αt = and sin αt = ,
2 2i
we can rewrite the real Fourier series as the complex Fourier series

X  
f (t) = a0 + 1
2 (ak − ibk )e2πikt/T + 21 (ak + ibk )e−2πikt/T
k=1

so that

X
f (t) = ck e2πikt/T for all t ∈ [0, T ] (14.2)
k=−∞

where


The term “power” is a misnomer because the function f need not be related to a physical quantity for which
the power makes any sense. However, we will stick to the common usage.
To understand the physical significance of power, we begin with the definition of work. Consider a particle
which is under the influence of the constant force F ~ . If the particle moves from the point P0 to P1 then the
work done to the particle is F ~ q~r , where ~r is the vector from P0 to P1 . The power of the particle is the work
~ q~v where ~v = ~r/t.
done per unit time, i.e., F
Next, consider a charge q which is moving between two terminals having a potential difference of V . The
work done on the charge is W = qV = ItV , where I is the current and t is the time it takes for the charge to
move between the two terminals. If R is the resistance in the circuit, V = IR and the power is
W V2
= IV = I 2 R =
P = .
t R
Thus, if we consider f (t) to be the voltage or the current of some signal, the instantaneous power in the signal is
proportional to f 2 (t) and the average power is proportional to
1 T ˛˛
Z
˛2
f (t)˛ dt .
T 0

116
14. Discrete Fourier Transform

c0 = a 0
)
ck = 21 (ak − ibk ) (14.3)
for k > 0 .
c−k = 12 (ak + ibk )

The coefficients . . . , c−2 , c−1 , c0 , c1 , c2 , . . . are called the complex Fourier coefficients of f , and ck and
c−k are the coefficients of the k-th mode. (Note that these Fourier coefficients are generally complex.) We
can also calculate ck directly from f by
Z T
1
ck = f (t)e−2πikt/T dt for k = . . . , −2, −1, 0, 1, 2, . . .
T 0

Note that if f is real, then c−k = c∗k (by replacing k by −k in the above equation). The power of f (t)
is

X 
P = |c0 |2 + |ck |2 + |c−k |2
k=1

and the power in each mode is 


|c0 |2 if k = 0
Pk = 
|ck |2 + |c−k |2 if k > 0 .

We can only calculate a finite number of Fourier coefficients numerically and so we truncate the infinite
series at the M -th mode. We should choose M large enough that

M
X
f (t) ≈ ck e2πikt/T for all t ∈ [0, T ] .
k=−M

There are now


N = 2M + 1
unknowns (which is an odd number because of the k = 0 mode). We require N equations to solve for
these N unknown coefficients. We obtain these equations by requiring that the two sides of this ap-
proximation be equal at the N equally spaced abscissas tj = jT /N for j = 0, 1, 2, . . . , N − 1 (so that
0 = t0 < t1 < . . . < tN −1 < tN = T ).† That is,

M
X
f (tj ) = γk e2πiktj /T for j = 0, 1, 2, . . . , N − 1
k=−M
or, written as a first-order system,
M
X
fj = γk e2πijk/N for j = 0, 1, 2, . . . , N − 1 (14.4)
k=−M

where fj ≡ f (tj ). This linear system can be solved to obtain

N −1
1 X
γk = fj e−2πijk/N for k = −M, −M + 1, . . . , M . (14.5)
N j=0

The reason we have replaced the coefficients c−M , c−M+1 , . . . , cM−1 , cM by γ−M , γ−M+1 , . . . , γM−1 , γM
is that the c’s are the coefficients in the continuous complex Fourier series, eq. (14.2), and are calculated
by (14.3). The γ’s are the coefficients in the discrete complex Fourier series, eq. (14.4), and are calcu-
lated by (14.5).


Note that tN is not used because f (tN ) has the same value as f (t0 ) and so does not provide us with an
independent equation.

117
14. Discrete Fourier Transform

Note: To repeat: the discrete Fourier coefficient γk is a function of M , i.e., γk (M ), and is generally not
equal to the continuous Fourier coefficient ck . However, as M → ∞ we have γk (M ) → ck . For a
fixed M we generally only have γk (M ) ≈ ck as long as |k| is “much less than” M . Of course, it
takes practice and experimentation to determine what “much less than” means.
We define the discrete Fourier series by

M
X
fFS (t) = γk e2πikt/T for all t ∈ [0, T ] .
k=−M

It is our responsibility (using our experience) to choose M large enough that f (t) ≈ fFS (t). Given
f = (f0 , f1 , f2 , . . . , fN −1 )T , the Fourier coefficients are calculated in MATLAB by
>> fc = fft(f)/N
where the coefficients of the discrete Fourier transform are contained in fc in the order
T
γ0 , γ1 , . . . , γM−1 , γM , γ−M , γ−M+1 , . . . , γ−2 , γ−1 .

The command fftshift changes the order to


T
γ−M , γ−M+1 , . . . , γ−2 , γ−1 , γ0 , γ1 , . . . , γM−1 , γM .

The original function, represented by the vector f, is recovered by


>> f = N*ifft(fc)
and the order is changed by ifftshift
It is important to check the Fourier transform to make sure it returns the results we expect. We began
with the real trigonometric series (14.1) and derived the complex trigonometric series (14.2) from it. The
nonzero Fourier coefficients of f (x) = cos x = (eix + e−ix )/2 are a1 = 1 and c−1 = c1 = 1/2, whereas the
nonzero Fourier coefficients of f (x) = sin x = (eix − e−ix )/(2i) are a1 = 1 and c−1 = i/2 but c1 = −i/2.
The code
>> n = 9;
>> x = linspace(0, 2*pi, n+1);
>> x(n+1) = [];
>> c c = fft(cos(x));
>> d c = fftshift(c c);
>> c s = fft(sin(x));
>> d s = fftshift(c s);
>> ci c = ifft(cos(x));
>> di c = ifftshift(c c);
>> ci s = ifft(sin(x));
>> di s = ifftshift(c s);
returns the vectors
cc = (0, 4.5, 0, 0, 0, 0, 0, 0, 4.5),
dc = (0, 0, 0, 4.5, 0, 4.5, 0, 0, 0),
cs = (0, −4.5i, 0, 0, 0, 0, 0, 0, 4.5i),
ds = (0, 0, 0, 4.5i, 0, −4.5i, 0, 0, 0),
(i)
cc = (0, 0.5, 0, 0, 0, 0, 0, 0, 0.5),
(i)
dc = (0, 0, 0, 0, 0.5, 0, 0.5, 0, 0),
(i)
cs = (0, 0.5i, 0, 0, 0, 0, 0, 0, −0.5i), and
(i)
ds = (0, 0, 0, 0, −0.5i, 0, 0.5i, 0, 0).
Notice that fft and ifft both return the correct coefficients for cos x (up to the scaling of n ), but only
the fft returns the correct coefficients for sin x. Thus, the command fft is correct, but it multiplies the
coefficients by N .

118
14. Discrete Fourier Transform

Also, notice that fftshift correctly shifts the coefficients, whereas ifftshift does not — but
ifftshift correctly shifts the coefficients back. That is, ifftshift is the inverse of fftshift so
ifftshift(fftshift(c s)) = c s.
Warning: One of the most common mistakes in using fft is forgetting that the input is in the order

f0 , f1 , f2 , . . . , fN −1

while the output is in the order

γ0 , γ1 , . . . , γM−1 , γM , γ−M , γ−M+1 , . . . , γ−2 , γ−1 ,

not
γ−M , γ−M+1 , . . . , γ−2 , γ−1 γ0 , γ1 , . . . , γM−1 , γM .

There is only one difficulty with our presentation. As we have already stated, the vector f has
N = 2M + 1 elements, which is an odd number. The Fast Fourier Transform (FFT, for short), which
is the method used to calculate the discrete Fourier coefficients by fft and also to recover the original
function by ifft, generally works faster if the number of elements of f is even, and is particularly fast if
it a power of 2.
The figure below shows the cputime needed to calculate fft(f) as a function of N . Since the verti-
cal axis is logarithmic, it is clear that there is a huge difference in the time required as we vary N .The
dashed lines show the minimum and maximum asymptotic times as cn log2 n.

cputime to calculate fft(f,n)


−3
10

−4
10
cpu time (seconds)

−5
10

−6
10
dashed lines are c*n*log2n; c = 2.7e−09, 2.0e−08

−7
10
0 200 400 600 800 1000 1200 1400 1600 1800 2000
n

119
14. Discrete Fourier Transform

For N to be even, we have to drop one coefficient, and the one we drop is γM . Now

N = 2M

is even. The discrete complex Fourier series is

M−1
X
fFS (t) = γk e2πikt/T for all t ∈ [0, T ]
k=−M

and the discrete Fourier coefficients are calculated by

N −1
1 X
γk = fj e−2πijk/N for k = −M, −M + 1, . . . , M − 2, M − 1 .
N j=0

As before, given f = (f0 , f1 , f2 , . . . , fN −1 )T , the Fourier coefficients are calculated by


>> fc = fft(f)/N
The coefficients of the discrete Fourier transform are now contained in fc as
T
fc = γ0 , γ1 , . . . , γM−2 , γM−1 , γ−M , γ−M+1 , . . . , γ−2 , γ−1 .

The original function, represented by the vector f, is again recovered by


>> f = N*ifft(fc)
Note: Since there are now an even number of Fourier coefficients, we can reorder them by using fftshift,
which switches the first half and the last half of the elements. The result is
T
fftshift(fc) = γ−M , γ−M+1 , . . . , γ−2 , γ−1 , γ0 , γ1 , . . . , γM−2 , γM−1 , .

Also, ifftshift is the same as fftshift if N is even.


Warning: Remember that if you reorder the elements of fc by
>> fc shift = fftshift(fc)
you will have to “unorder” the elements by applying
>> fc = fftshift(fc shift)
again before you use ifft.
Note: When N is even we cannot recover γM and so we only know one of the two coefficients of the
M -th mode. Thus, we cannot determine the M -th mode correctly. Although we cannot give a sim-
ple example, it occasionally happens that this causes difficulties. The solution is to set γ−M = 0 so
that the M -th mode is dropped completely.
Here is a simple example of the use of Fourier coefficients from The Student Edition of MATLAB: User’s
Guide. We begin with a signal at 50 and 120 hertz (cycles per unit time), y0, and then we perturb it
by adding Gaussian noise, ypert. We plot the periodic unperturbed signal, and then the perturbed sig-
nal, vs. time. (If you enter all these commands into an m-file, put a pause command between each of the
plot commands.)
>> time = .6;
>> N = 600;
>> t = linspace(0, time, N);
>> y0 = sin(2*pi*50*t) + sin(2*pi*120*t); % unperturbed signal
>> ypert = y0 + 2*randn(size(t)); % perturbed signal
>> figure(1)
>> plot(t, y0, ′ r′ ), axis([0 time -8 8])
>> hold on
>> plot(t, ypert, ′ g′ )
Clearly, once the random noise has been added, the original signal has been completely lost — or has it.

120
14. Discrete Fourier Transform

We now look at the Fourier spectrum of y0 by plotting the power at each frequency. First, we plot the
unperturbed power, power0, and then the perturbed power, powerpert, vs. the frequency at each mode,
freq. The two spikes in the plot of the unperturbed power are precisely at 50 and 120 hertz, the signa-
ture of the two sine functions in y0. (For simplicity in the discussion, we have deleted the power in the
M -th mode by fc(N/2 +1) = [] so that power0(k) is the power in the k−1-st mode.)
>> fc0 = fft(y0)/N; % Fourier spectrum of unperturbed signal
>> figure(2)
>> fc0(N/2 +1) = []; % delete k = N/2 +1 mode
>> power0(1) = abs(fc0(1)). 2;
ˆ
>> power0(2:N/2) = abs(fc0(2:N/2)). 2 + abs(fc0(N-1:-1:N/2 +1)). 2;
>> freq = [1:N]′ /time;
ˆ
% the frequency of each mode
ˆ
>> plot(freq(1:N/2), power0, ′ r′ ), axis([0 freq(N/2) 0 .5])
>> fcpert = fft(ypert)/N; % Fourier spectrum of perturbed signal
>> hold on
>> powerpert(1) = abs(fcpert(1)). 2;
ˆ
>> powerpert(2:N/2) = abs(fcpert(2:N/2)). 2 + abs(fcpert(N-1:-1:N/2 +1)). 2;
>> plot(freq(1:N/2), powerpert, ′ g′ )
ˆ ˆ
Clearly, the original spikes are still dominant, but the random noise has excited every mode.
To see how much power is in the unperturbed signal and then the perturbed signal, enter
>> sum(power0)
>> sum(powerpert)
The perturbed signal has about five times as much power as the original signal, which makes clear how
large the perturbation is.
Let us see if we can reconstruct the original signal by removing any mode whose magnitude is “small”.
By looking at the power plots, we see that the power in all the modes, except for those corresponding
to the spikes, have an amplitude / 0.1. Thus, we delete any mode of the perturbed Fourier spectrum,
i.e., fcpert, whose power is less than this value; we call this new Fourier spectrum fcchop. We then
construct a new signal ychop from this “chopped” Fourier spectrum and compare it with the original
unperturbed signal.
>> fcchop = fcpert; % initialize the chopped Fourier spectrum
>> ip = zeros(size(fcpert)); % construct a vector with 0’s
>> ip(1:N/2) = ( powerpert > 0.1 ); % where fcchop should be
>> ip(N:-1:N/2 +2) = ip(2:N/2); % zeroed out
>> fcchop( find( ip) ) = 0; % zero out "small" modes
˜
>> ychop = real( N*ifft(fcchop) ); % signal of "chopped" Fourier spectrum
>> figure(1)
>> plot(t, ychop, ′ b′ )

( ychop is the real part of N*ifft(fcchop) because, due to round-off errors, the inverse Fourier trans-
form returns a “slightly” complex result.) The result is remarkably good considering the size of the per-
turbation. If you have trouble comparing y0 with ychop, reenter
>> plot(t, y0, ′ r′ )

Discrete Fourier Transform

fft(f) The discrete Fourier transform of f.


ifft(fc) The inverse discrete Fourier transform of the Fourier coefficients fc.
fftshift(fc) Switches the first half and the second half of the elements of fc.
ifftshift(cf) Unswitches the first half and the second half of the elements of fc. (fftshift and
ifftshift are the same if the number of elements is even.

121
15. Mathematical Functions Applied to Matrices

15. Mathematical Functions Applied to Matrices


As we briefly mentioned in subsection 2.7, mathematical functions can generally only be applied to
square matrices. For example, if A ∈ Cn×n then eA is defined from the Taylor series expansion of ea .
That is, since
a a2 a3
ea = 1 + + + + ...
1! 2! 3!
we define eA to be
A A2 A3
eA = 1 + + + + ... .
1! 2! 3!
(Thus, if A ∈ Cm×n where m 6= n then eA does not exist because Ak does not exist if A is not a square
matrix.)
If A is a square diagonal matrix eA is particularly simple to calculate since
p
ap11
  
a11
 a22
. . 0   ap22
. . 0 
Ap = 
   
.  = . .
apn−1,n−1
   
an−1,n−1
0 ann 0 apnn

Thus,
ea11
 
 e a22
. . 0 
eA = 
 
. .
ean−1,n−1
 
0 eann
The MATLAB command
>> expm(A)
calculates eA if A is a square matrix. (Otherwise, it generates an error message.)
A simple example where eA occurs is in the solution of first-order ode systems with constant coeffi-
cients. Recall that the solution of
dy
(t) = ay(t) for t ≥ 0 with y(0) = yic
dt
is
y(t) = yic eat .
Similarly, the solution of

y1 (t)
 
a11 a12 ... a1n

y1 (t)

d  y2 (t)   a21 a22 ... a2n   y2 (t) 
 ..  =  .. ... .. ...   ...  for t ≥ 0 with y(0) = yic
dt . . .
yn (t) an1 an2 ... ann yn (t)

i.e., y′ (t) = Ay(t), is


y(t) = eAt yic .
To calculate y(t) for any time t, you only need enter
>> expm(A*t) * yic
Note: The above statement gives the exact solution to the ode system at t = 10 by
>> expm(A*10) * yic
You could also use numerical methods, as discussed in section 10, to solve it. However, you would
have to solve the ode for all t ∈ [0, 10] in order to obtain a numerical approximation at the final
time. This
√ would be much more costly than simply using the analytical solution.
Similarly, B is calculated in MATLAB by entering

122
15. Mathematical Functions Applied to Matrices

>> sqrtm(A)
Finally, log B is calculated in MATLAB by entering
>> logm(A)
These are the only explicit MATLAB commands for applying mathematical functions to matrices. How-
ever, there is a general matrix function for the other mathematical functions. The command
>> funm(A, <function handle>)
evaluates <function name>(A) for the MATLAB functions exp, sin, cos, sinh, and cosh as well as
user-defined functions.

Matrix Functions

expm(A) Calculates eA where A must be a square matrix.



sqrtm(A) Calculates A where A must be a square matrix.
logm(A) Calculates log A where A must be a square matrix.
funm(A, <function handle>) Calculates <function name>(A) where A must be a square
matrix.

123
15. Mathematical Functions Applied to Matrices

124
Appendix: Reference Tables

These tables summarize the functions and operations described in this tutorial. The number (or num-
bers) shown give the page number of the table where this entry is discussed.

Arithmetical Operators

+ Addition. (p. 7, 28)


- Subtraction. (p. 7, 28)
* Scalar or matrix multiplication. (p. 7, 28)
.* Elementwise multiplication of matrices. (p. 28)
/ Scalar division. (p. 7, 28)
./ Elementwise division of matrices. (p. 28)
\ Scalar left division, i.e., b\a = a/b. (p. 7)
\ The solution to Ax = b for A ∈ Cm×n : when m = n and A is nonsingular this
is the solution Gaussian elimination; when m > n this is the least-squares ap-
proximation of the overdetermined system; when m < n this is a solution of the
underdetermined system. (p. 28, 63)
.\ Elementwise left division of matrices i.e., B.\A = A./B. (p. 28)
Scalar or matrix exponentiation. (p. 7, 28)
ˆ. Elementwise exponentiation of matrices. (p. 28)
ˆ

Special Characters

: Creates a vector by a:b or a:c:b; subscripts matrices. (p. 25)


; Ends a statement without printing out the result; also, ends each row when enter-
ing a matrix. (p. 9)
, Ends a statement when more than one appear on a line and the result is to be
printed out; also, separates the arguments in a function; also, can separate the
elements of each row when entering a matrix. (p. 9)
... Continues a MATLAB command on the next line. (p. 14)
% Begins a comment. (p. 14)
↑ The up-arrow key moves backward in the MATLAB workspace, one line at a time.
(p. 7)

125
Appendix: Reference Tables

Getting Help

demo Runs demonstrations of many of the capabilities of MATLAB. (p. 16, 54)
doc On-line reference manual. (p. 16)
help On-line help. (p. 16)
helpdesk Loads the main page of the on-line reference manual. (p. 16)
load Loads back all of the variables which have been saved previously. (p. 16)
lookfor Searches all MATLAB commands for a keyword. (p. 16)
profile Profile the execution time of a MATLAB code. This is very useful for improving
the performance of a code by determining where most of the CPU time is spent.
(p. 92)
save Saves all of your variables. (p. 16)
type Displays the actual MATLAB code. (p. 14, 16)
who Lists all the current variables. (p. 16)
whos Lists all the current variables in more detail than who. (p. 16)
C Abort the command which is currently executing (i.e., hold down the control key
ˆ and type “c”). (p. 16)

Predefined Variables

ans The default variable name when one has not been specified. (p. 9)
pi π. (p. 9)
eps Approximately the smallest positive real number on the computer such that
1 + eps 6= 1. (p. 9)
Inf ∞ (as in 1/0 ). (p. 9)
NaN Not-a-Number
√ (as in 0/0 ). (p. 9)
i √ −1 . (p. 9)
j −1 . (p. 9)
realmin The smallest “usable” positive real number on the computer. (p. 9)
realmax The largest “usable” positive real number on the computer. (p. 9)

Format Options

format short The default setting. (p. 11)


format long Results are printed to approximately the maximum number of digits of accuracy in
MATLAB. (p. 11)
format short e Results are printed in scientific notation. (p. 11)
format long e Results are printed in scientific notation to approximately the maximum number of
digits of accuracy in MATLAB. (p. 11)
format short g Results are printed in the best of either format short or format short e.
(p. 11)
format long g Results are printed in the best of either format long or format long e. (p. 11)

126
Appendix: Reference Tables

Input-Output Functions

csvread Reads data into MATLAB from the named file, one row per line of input. (p. 48)
csvwrite Writes out the elements of a matrix to the named file using the same format as
csvread. (p. 48)
diary Saves your input to MATLAB and most of the output from MATLAB to disk.
(p. 7)
fopen Opens the file with the permission string determining how the file is to be accessed.
(p. 64)
fclose Closes the file. (p. 64)
fscanf Behaves very similarly to the C command in reading data from a file using any
desired format. (p. 64)
fprintf Behaves very similarly to the C command in writing data to a file using any de-
sired format.
It can also be used to display data on the screen. (p. 64, 64)
input Displays the prompt on the screen and waits for you to enter whatever is desired.
(p. 10)
load Reads data into MATLAB from the named file, one row per line of input. (p. 48)
print Prints a plot or saves it in a file using various printer specific formats. (p. 48)

127
Appendix: Reference Tables

Some Common Mathematical Functions

abs Absolute value (p. 12, 13) cotd Cotangent (argument in degrees).
acos Inverse cosine. (p. 12) (p. 12)
acosd Inverse cosine (result in degrees). csc Cosecant. (p. 12)
(p. 12) cscd Cosecant (argument in degrees).
acosh Inverse hyperbolic cosine. (p. 12) (p. 12)
acot Inverse cotangent. (p. 12) exp Exponential function. (p. 12)
acotd Inverse cotangent (result in de- factorial Factorial function. (p. 12)
grees). (p. 12) fix Round toward zero to the nearest
acoth Inverse hyperbolic cosine. (p. ) integer. (p. 12)
acsc Inverse cosecant. (p. 12) floor Round downward to the nearest
acscd Inverse cosecant (result in degrees). integer. (p. 12)
(p. 12) imag The imaginary part of a complex
angle Phase angle of a complex number. number. (p. 13)
(p. 13) log The natural logarithm, i.e., to the
asec Inverse secant. (p. ) base e. (p. 12)
asec Inverse secant (result in degrees). log10 The common logarithm, i.e., to the
(p. ) base 10. (p. 12)
asin Inverse sine. (p. 12) mod The modulus after division. (p. 12)
asin Inverse sine (result in degrees). real The real part of a complex number.
(p. 12) (p. 13)
asinh Inverse hyperbolic sine. (p. 12) rem The remainder after division.
atan Inverse tangent. (p. 12) (p. 12)
atand Inverse tangent (result in degrees). round Round to the closest integer. (p. 12)
(p. 12) sec Secant. (p. 12)
atan2 Inverse tangent using two argu- secd Secant (argument in degrees).
ments. (p. 12) (p. 12)
atanh Inverse hyperbolic tangent. (p. 12) sign The sign of the real number. (p. 12)
ceil Round upward to the nearest inte- sin Sine. (p. 12)
ger. (p. 12) sind Sine (argument in degrees). (p. 12)
conj Complex conjugation. (p. 13) sinh Hyperbolic sine. (p. 12)
cos Cosine. (p. 12) sqrt Square root. (p. 12)
cosd Cosine (argument in degrees). tan Tangent. (p. 12)
(p. 12) tand Tangent (argument in degrees).
cosh Hyperbolic cosine. (p. 12) (p. 12)
cot Cotangent. (p. 12) tanh Hyperbolic tangent. (p. 12)

Arithmetical Matrix Operations

A+B Matrix addition. (p. 7, 28) A.*B Elementwise multiplication. (p. 28)
A-B Matrix subtraction. (p. 7, 28) A. p Elementwise exponentiation. (p. 28)
A*B Matrix multiplication. (p. 7, 28) p. A
ˆ
A n Matrix exponentiation. (p. 7, 28) A. B
ˆ
ˆ
A\b The solution to Ax = b by Gaussian A./B
ˆ Elementwise division. (p. 28)
elimination when A is a square non- B.\A Elementwise left division, i.e., B.\A is
singular matrix. (p. 28, 63) exactly the same as A./B. (p. 28)
A\B The solution to AX = B by Gaussian
elimination. (p. 28)
b/A The solution to xA = b where x and
b are row ve tors. (p. 7, 28)
B/A The solution to XA = B. (p. 28)

128
Appendix: Reference Tables

Elementary Matrices

eye Generates the identity matrix. (p. 20)


false Generates a logical matrix with all elements having the value false. (p. 75)
ones Generates a matrix with all elements being 1. (p. 20)
rand Generates a matrix whose elements are uniformly distributed random numbers in
the interval (0, 1). (p. 20)
randn Generates a matrix whose elements are normally (i.e., Gaussian) distributed ran-
dom numbers with mean 0 and standard deviation 1. (p. 20)
randperm(n) Generates a random permutation of the integers 1, 2, . . . , n. (p. )
speye Generates a Sparse identity matrix. (p. 95)
sprand Sparse uniformly distributed random matrix. (p. 95, 95)
sprandsym Sparse uniformly distributed symmetric random matrix; the matrix can also be
positive definite. (p. 95)
sprandn Sparse normally distributed random matrix. (p. 95)
true Generates a logical matrix with all elements having the value true. (p. 75)
zeros Generates a zero matrix. (p. 20)

Specialized Matrices

hilb Generates the hilbert matrix. (Defined on p. 61.)


vander Generates the Vandermonde matrix. (Defined on p. 111.)

Elementary Matrix Operations

size The size of a matrix. (p. 20)


length The number of elements in a vector. (p. 20)
.′ The transpose of a matrix. (p. 20)

The conjugate transpose of a matrix. (p. 20)

Manipulating Matrices

cat Concatenates arrays; this is useful for putting arrays into a higher-dimensional
array. (p. 35)
clear Deletes a variable or all the variables. This is a very dangerous ommand.
(p. 9)
diag Extracts or creates diagonals of a matrix. (p. 25)
fliplr Flips a matrix left to right. (p. 25)
flipud Flips a matrix up and down. (p. 25)
ipermute The inverse of permute. (p. 35)
permute Reorders the dimensions of a multidimensional array. (p. 35)
spdiags Generates a sparse matrix by diagonals. (p. 95)
repmat Tiles a matrix with copies of another matrix. (p. 25)
reshape Reshapes the elements of a matrix. (p. 25)
rot90 Rotates a matrix a multiple of 90◦ . (p. 25)
squeeze Removes (i.e., squeezes out) dimensions which only have one element. (p. 35)
triu Extracts the upper triangular part of a matrix. (p. 25)
tril Extracts the lower triangular part of a matrix. (p. 25)
[] The null matrix. This is also useful for deleting elements of a vector and rows or
columns of a matrix. (p. 25)

129
Appendix: Reference Tables

Odds and Ends

path Viewing and changing the search path. (p. )


cputime Approximately the CPU time (in seconds) used during this session. (p. 28)
tic, toc Returns the elapsed time between these two commands. (p. 28)
pause Halts execution until you press some key. (p. 88)
rats Converts a floating-point number to a “close” rational number, which is frequently
the exact value. (p. 63)

Two-Dimensional Graphics

plot Plots the data points in Cartesian coordinates. (p. 49)


fill Fills one or more polygons. (p. 54)
semilogx The same as plot but the x axis is logarithmic. (p. 49)
semilogy The same as plot but the y axis is logarithmic. (p. 49)
loglog The same as plot but both axes are logarithmic. (p. 49)
ezplot Generates an “easy”plot (similar to fplot ). It can also plot a parametric func-
tion, i.e., x(t), y(t) , or an implicit function, i.e., f (x, y) = 0. (p. 49)
polar Plots the data points in polar coordinates. (p. 49)
ezpolar Generates an “easy” polar plot. (p. 49)
linspace Generates equally-spaced points, similar to the colon operator. (p. 49)
xlabel Puts a label on the x-axis. (p. 49)
ylabel Puts a label on the y-axis. (p. 49)
title Puts a title on the top of the plot. (p. 49)
axis Controls the scaling and the appearance of the axes. (p. 49)
hold Holds the current plot or release it. (p. 49)
hist Plots a histogram. (p. 49)
errorbar Plots a curve through data points and also the error bar at each data point.
(p. 49)
subplot Divides the graphics window into rectangles and moves between them. (p. 49, 51)

130
Appendix: Reference Tables

Three-Dimensional Graphics

plot3 Plots the data points in Cartesian coordinates. (p. 51)


ezplot3 Generates an “easy” plot in 3-D. (p. 51)
fill3 Fills one or more 3D polygons. (p. 54)
mesh Plots a 3-D surface using a wire mesh. (p. 51)
ezmesh Generates an “easy” 3-D surface using a wire mesh. (p. 51)
surf Plots a 3-D filled-in surface. (p. 51)
ezsurf Generates an “easy” 3-D filled-in surface. (p. 51)
view Changes the viewpoint of a 3-D surface plot. (p. 51)
meshgrid Generates a 2-D grid. (p. 51)
zlabel Puts a label on the z-axis. (p. 51)
axis Controls the scaling and the appearance of the axes. (p. 51)
contour Plots a contour looking down the z axis. (p. 51)
ezcontour Generates an “easy” contour looking down the z axis. (p. 51)
contour3 Plots a contour in 3-D. (p. 51)
ezcontour3 Generates an “easy” contour in 3-D. (p. 51)
subplot Divides the graphics window into rectangles and moves between them. (p. 49, 51)
colorbar Adds a color bar showing the correspondence between the value and the color.
(p. 54)
colormap Determines the current color map or choose a new one. (p. 54)

Advanced Graphics Features

clf Clear a figure (i.e., delete everything in the figure (p. 54)
demo Runs demonstrations of many of the capabilities of MATLAB. (p. 16, 54)
figure Creates a new graphics window and makes it the current target. (p. 54)
fplot Plots the specified function within the limits given. (p. 49)
gtext Places the text at the point given by the mouse. (p. 54)
image Plots a two-dimensional image. (p. 54)
legend Places a legend on the plot. (p. 54)
text Adds the text at a particular location. (p. 54)
ginput Obtains the current cursor position. (p. 54)
get Returns the current value of the property of an object. (p. 56)
set Sets the value of the property, or properties of an object. (p. 56)
gca The current axes handle. (p. 56)
gcf The current figure handle. (p. 56)

String Functions, Cell Arrays, and Structures

num2str Converts a variable to a string. (p. 42)


sprintf Behaves very similarly to the C command in writing data to a text variable using
any desired format. (p. 42)
sscanf Behaves very similarly to the C command in reading data from a text variable
using any desired format. (p. 42)
str2num Converts a string to a variable. (p. 42)
strcmp Compares strings. (p. 72)
cell Preallocate a cell array of a specific size. (p. 42)
celldisp Display all the contents of a cell array. (p. 42)
struct Create a structure. (p. 42)

131
Appendix: Reference Tables

Data Manipulation Commands

errorbar Plots a curve through data points and also the error bar at each data point.
(p. 49)
hist Plots a histogram of the elements of a vector. (p. 49)
max The maximum element of a vector. (p. 34)
min The minimum element of a vector. (p. 34)
mean The mean, or average, of the elements of a vector. (p. 34)
norm The norm of a vector or a matrix. (p. 34)
prod The product of the elements of a vector. (p. 34)
sort Sorts the elements of a vector in increasing order. (p. 34)
std The standard deviation of the elements of a vector. (p. 34)
sum The sum of the elements of a vector. (p. 34)

Some Useful Functions in Linear Algebra

chol Calculates the Cholesky decomposition of a symmetric, positive definite matrix.


(p. 70)
cond Calculates the condition number of a matrix. (p. 70)
condest Calculates a lower bound to the condition number of a square matrix. (p. 70)
det Calculates the determinant of a square matrix. (p. 70)
eig Calculates the eigenvalues, and eigenvectors if desired, of a square matrix. (p. 70)
eigs Calculates some eigenvalues and eigenvectors of a square matrix. (p. 70)
inv Calculates the inverse of a square invertible matrix. (p. 70)
linsolve Solve a square matrix equation where the matrix can have certain properties to
increase the CPU time. (p. 60)
lu Calculates the LU decomposition of a square invertible matrix. (p. 70)
norm Calculates the norm of a vector or matrix. (p. 70)
null Calculates an orthonormal basis for the null space of a matrix. (p. 70)
orth Calculates an orthonormal basis for the range of a matrix. (p. 70)
pinv Calculates the pseudoinverse of a matrix. (p. 63)
qr Calculates the QR decomposition of a matrix. (p. 70)
rank Estimates the rank of a matrix. (p. 70)
rref Calculates the reduced row echelon form of a matrix. (p. 60)
svd Calculates the singular value decomposition of a matrix. (p. 70)

Logical and Relational Operators

& Logical AND. (p. 73) < Less than. (p. 72)
| Logical OR. (p. 73) <= Less than or equal to. (p. 72)
Logical NOT. (p. 73) == Equal. (p. 72)
˜xor Logical EXCLUSIVE OR. (p. 73) > Greater than. (p. 72)
&& A short-circuiting logical AND. (p. 73) >= Greater than or equal to. (p. 72)
|| A short-circuiting logical OR. (p. 73) = Not equal to. (p. 72)
˜
strcmp Comparing strings. (p. 72)

132
Appendix: Reference Tables

Flow Control

break Terminates execution of a for or while loop. (p. 75)


case Part of the switch command. (p. 75)
continue Begins the next iteration of a for or while loop immediately. (p. 75)
else Used with the if statement. (p. 75)
elseif Used with the if statement. (p. 75)
end Terminates the scope of the for, if, switch, and while statements. (p. 75, 88)
error Displays the error message and terminates all flow control statements. (p. 88)
for Repeat statements a specific number of times. (p. 75)
if Executes statements if certain conditions are met. (p. 75)
otherwise Part of the switch command. (p. 75)
switch Selects certain statements based on the value of the switch expression. (p. 75)
while Repeats statements as long as an expression is true. (p. 75)

Logical Functions

all True if all the elements of a vector are true; operates on the columns of a matrix.
(p. 78)
any True if any of the elements of a vector are true; operates on the columns of a ma-
trix. (p. 78)
exist False if this name is not the name of a variable or a file. (p. 78)
find The indices of a vector or matrix which are nonzero. (p. 78)
ischar True if a vector or array contains character elements. (p. 78)
isempty True if the matrix is empty, i.e., []. (p. 78)
isfinite Generates a matrix with 1 in all the elements which are finite (i.e., not Inf or
NaN ) and 0 otherwise. (p. 78)
isinf Generates a matrix with 1 in all the elements which are Inf and 0 otherwise.
(p. 78)
islogical True for a logical variable or array. (p. 78)
isnan Generates a matrix with 1 in all the elements which are NaN and 0 otherwise.
(p. 78)
logical Converts a numeric variable to a logical one. (p. 78)

133
Appendix: Reference Tables

Programming Language Functions

echo Turns echoing of statements in m-files on and off. (p. 88)


end Ends a function. Only required if the function m-file contains a nested function.
(p. 75, 88)
error Displays the error message and terminates the function. (p. 88)
eval Executes MATLAB statements contained in a text variable. (p. 90)
feval Executes a function specified by a string. (p. 90)
function Begins a MATLAB function. (p. 88)
global Defines a global variable (i.e., it can be shared between different functions and/or
the workspace). (p. 88)
lasterr If eval “catches” an error, it is contained here. (p. 90)
persistent Defines a local variable whose value is to be saved between calls to the function.
(p. 88)
keyboard Stops execution in an m-file and returns control to the user for debugging purposes.
(p. 83, 88)
nargin Number of input arguments supplied by the user. (p. 88)
nargout Number of output arguments supplied by the user. (p. 88)
return Terminates the function immediately. (p. 83, 88)

Debugging Commands

keyboard Turns debugging on. (p. 83, 88)


dbstep Execute one or more lines. (p. 83)
dbcont Continue execution. (p. 83)
dbstop Set a breakpoint. (p. 83)
dbclear Remove a breakpoint. (p. 83)
dbup Change the workspace to the calling function or the base workspace. (p. 83)
dbdown Change the workspace down to the called function. (p. 83)
dbstack Display all the calling functions. (p. 83)
dbstatus List all the breakpoints. (p. 83)
dbtype List the current function, including the line numbers. (p. 83)
dbquit Quit debugging mode and terminate the function. (p. 83)
return Quit debugging mode and continue execution of the function. (p. 83, 88)

Discrete Fourier Transform

fft The discrete Fourier transform. (p. 121)


fftshift Switches the first half and the second half of the elements of a vector. (p. 121)
ifft The inverse discrete Fourier transform. (p. 121)
fftshift Unswitches the first half and the second half of the elements of a vector. (p. 121)

134
Appendix: Reference Tables

Sparse Matrix Functions

speye Generates a Sparse identity matrix. (p. 95)


sprand Sparse uniformly distributed random matrix. (p. 95, 95)
sprandn Sparse normally distributed random matrix. (p. 95)
sparse Generates a sparse matrix elementwise. (p. 95)
spdiags Generates a sparse matrix by diagonals. (p. 95)
full Converts a sparse matrix to a full matrix. (p. 95)
find Finds the indices of the nonzero elements of a matrix. (p. 95)
nnz Returns the number of nonzero elements in a matrix. (p. 95)
spfun Applies the function to a sparse matrix. (p. 95)
spy Plots the locations of the nonzero elements of a sparse matrix. (p. 95)
spconvert Generates a sparse matrix given the nonzero elements and their indices. (p. 95)

Time Evolution ODE Solvers

ode45 Non-stiff ode solver; fourth-order, one-step method for the ode y′ = f(t, y). (p. 97)
ode23 Non-stiff ode solver; second-order, one-step method. (p. 97)
ode113 Non-stiff ode solver; variable-order, multi-step method. (p. 97)
ode15s Stiff ode solver; variable-order, multi-step method. (p. 97)
ode23s Stiff ode solver; second-order, one-step method. (p. 97)
ode23t Stiff ode solver; trapezoidal method. (p. 97)
ode23tb Stiff ode solver; second-order, one-step method. (p. 97)
ode15i Stiff ode solver; variable-order, multi-step method for the fully implicit ode
f(t, y, y′ ) = 0. (p. 106)
odeset Assigns values to properties of the ode solver. (p. 101)

Boundary-Value Solver

bvp4c Numerically solves y′ (x) = f(x, y) for x ∈ [a, b] with given boundary conditions
and an initial guess for y. (p. 110)
bvpinit Calculates the initial guess either by giving y directly or by using a function
y = initial guess function(x). (p. 110)
deval Interpolate to determine the solution desired points. (p. 110)

Numerical Operations on Functions

dblquad Numerically evaluates a double integral. (p. 115)


fminbnd Numerically calculates a local minimum of a one-dimensional function. (p. )
fminsearch Numerically calculates a local minimum of a multi-dimensional function. (p. 115)
optimset Allows you to modify the parameters used by fzero, fminbnd, and fminsearch.
(p. 115)
fzero Numerically calculates a zero of a function. (p. 115)
quad Numerically evaluates an integral using Simpson’s method. (p. 115, 115)
quadgk Numerically evaluates an integral using the adaptive Gauss-Kronrod method. The
interval can be infinite and/or the function can have an integrable singularity.
(p. 115)
quadl Numerically evaluates an integral using the adaptive Gauss-Lobatto method. (p. )

135
Appendix: Reference Tables

Numerical Operations on Polynomials

interp1 Does one-dimensional interpolation. (p. 112)


interp2 Does two-dimensional interpolation. (p. 112)
interp3 Does three-dimensional interpolation. (p. 112)
interp4 Does n-dimensional interpolation. (p. )
pchip Cubic Hermite interpolation. (p. 112)
poly Calculates the coefficients of a polynomial given its roots. (p. 112)
polyder Calculates the derivative of a polynomial. (p. 112)
polyfit Calculates the least-squares polynomial of a given order which fits the given data.
(p. 112)
polyval Evaluates a polynomial at a point. (p. 112)
polyvalm Evaluates a polynomial with a matrix argument. (p. 112)
ppval interpolates a piecewise polynomial calculated by pchip or spline. (p. 112)
roots Numerically calculates all the zeroes of a polynomial. (p. 112)
spline Cubic spline interpolation. (p. 112)

Matrix Functions

expm Matrix exponentiation. (p. 123)


funm Evaluate general matrix function. (p. 123)
logm Matrix logarithm. (p. 123)
sqrtm Matrix square root. (p. 123)

136
Solutions To Exercises

These are the solutions to the exercises given in subsections 1.9, 2.10, and 4.5.

1.9.1a) >> a = 3.7; b = 5.7; deg = pi/180; ab = 79*deg;


>> c = sqrt(a 2 + b 2 - 2*a*b*cos(ab))
answer: 7.3640
ˆ ˆ
b) >> format long
>> c
answer: 7.36398828251259
c) >> format short e
>> asin( (b/c)*sin(ab) ) / deg
answer: 4.9448e+01
d) >> diary ′ triangle.ans′
1.9.2) >> (1.2e20 - i*12 20) (1/3)
answer:
ˆ ˆ
1.3637e+07 - 7.6850e+06i
1.9.3) >> th = input(′th =′ ); cos(2*th) - (2*cos(th) 2 - 1)
1.9.4) help fix or doc fix.
ˆ
2.10.1a) >> A = [1 2 3 4; 5 6 7 8; 9 10 11 12; 13 14 15 16]
>> A = [1:4; 5:8; 9:12; 13:16]
>> A = [ [1:4:13]′ [2:4:14]′ [3:4:15]′ [4:4:16]′ ]
b) >> A(2,:) = (-9/5)*A(2,:) + A(3,:)
2.10.2) >> A = 4*eye(n) - diag( ones(n-1,1), 1 ) - diag( ones(n-1,1), -1 )
2.10.3) >> A = diag([1:n]. 2) - diag( ones(n-1,1), 1 ) - diag( exp([2:n]), -1 )
2.10.4a) >>
ˆ
A = [ ones(6,4) zeros(6) ]; A(6,1) = 5; A(1,10) = -5
b) >> A = A - tril(A,-1)
2.10.5) >> x = [0:30]′. 2 % or x = [0:30]. 2′
2.10.6a) >> R = rand(5)
ˆ ˆ
b) >> [m, im] = max(R′)
c) >> mean(mean(R)) % or mean(R(:))
d) >> S = sin(R)
e) >> r = diag(R)
2.10.7a) >> A = [1 2 3; 4 5 6; 7 8 10]
>> B = A .5 % or B = sqrtm(A)
>> C =
ˆ
A. .5 % or C = sqrt(A)
ˆ
b) >> A - B 2
>> A -
ˆ
C. 2
ˆ
4.5.1a) >> x = linspace(-1, +1, 100);
>> y = exp(x);
>> plot(x, y)
b) >> z = 1 + x + x. 2 /2 + x. 6 /6
>> hold on
ˆ ˆ
>> plot(x, z)

137
Solutions To Exercises

c) >> plot(x, y-z)


d) >> hold off
>> plot(x, y, ′ r′ , x, z, ′ g′ , x, y-z, ′ m′ )
>> axis equal
>> xlabel(′x′ )
>> ylabel(′y′ )
>> title(′e i\pi = -1 is profound′)
ˆ
e) >> subplot(2, 1, 1)
>> hold off
>> plot(x, y, ′ r′ , x, z, ′ g′ )
>> axis equal
>> xlabel(′x′ )
>> ylabel(′y′ )
>> title(′e i\pi = -1 is profound′)
>>
ˆ
subplot(2, 1, 2)
>> plot(x, y-z)
4.5.2a) >> x = linspace(-3, 3, 91);
>> y = x;
>> [X, Y] = meshgrid(x, y); % or just do [X, Y] = meshgrid(x, x);
>> Z = (X. 2 + 4* Y. 2) .* sin(2*pi*X) .* sin(2*pi*Y);
>>
ˆ
surf(X, Y, Z);
ˆ
b) One particular choice is
>> view([1 2 5]) % or view([63 66])

138
ASCII Table

ASCII Table

Control Control
Octal Decimal Sequence Description Octal Decimal Sequence Description
000 0 @ Null character 100 64 @ AT symbol
001 1
ˆ A Start of header 101 65 A
002 2
ˆ B Start of text 102 66 B
003 3
ˆ C End of text 103 67 C
004 4
ˆ D End of transmission 104 68 D
005 5
ˆ E Enquiry 105 69 E
006 6
ˆ F Acknowledgment 106 70 F
007 7
ˆ G Bell 107 71 G
010 8
ˆ H Backspace 110 72 H
011 9
ˆ I Horizontal tab 111 73 I
012 10
ˆ J Line feed 112 74 J
013 11
ˆ K Vertical tab 113 75 K
014 12
ˆ L Form feed 114 76 L
015 13
ˆ M Carriage return 115 77 M
ˆ
016
017
14
15
ˆ
ˆ
N
O
Shift out
Shift in
116
117
78
79 ˜
N
O
020 16 P Data link escape 120 80 P
021 17
ˆ Q Device control 1 (often XON) 121 81 Q
022 18
ˆ R Device control 2 122 82 R
023 19
ˆ S Device control 3 (often XOFF) 123 83 S
024 20 ˆ T Device control 4 124 84 T
025 21
ˆ U Negative acknowledgement 125 85 U
026 22
ˆ V Synchronous idle 126 86 V
027 23
ˆ W End of transmissions block 127 87 W
030 24
ˆ X Cancel 130 88 X
031 25
ˆ Y End of medium 131 89 Y
032 26
ˆ Z Substitute 132 90 Z
033 27
ˆ [ Escape 133 91 [ Left bracket
034 28
ˆ \ File separator 134 92 \ Back slash
035 29
ˆ ] Group separator 135 93 ] Right bracket
036 30
ˆ Record separator 136 94 Caret
037 31
ˆˆ Unit separator 137 95
ˆ Underscore
040 32
ˆ Space 140 96 ‘ Opening single quote
041 33 ! 141 97 a
042 34 ” Double quote 142 98 b
043 35 # Numer sign 143 99 c
044 36 $ Dollar sign 144 100 d
045 37 % Percent 145 101 e
046 38 & Ampersand 146 102 f
047 39 ’ Closing single quote (apostrophe) 147 103 g
050 40 ( Left parenthesis 150 104 h
051 41 ) Right parenthesis 151 105 i
052 42 * Asterisk 152 106 j
053 43 + Plus sign 153 107 k
054 44 , Comma 154 108 l
055 45 - Minus sign or dash 155 109 m
056 46 . Dot 156 110 n
057 47 / Forward slash 157 111 o
060 48 0 160 112 p
061 49 1 161 113 q
062 50 2 162 114 r
063 51 3 163 115 s
064 52 4 164 116 t
065 53 5 165 117 u
066 54 6 166 118 v
067 55 7 167 119 w
070 56 8 170 120 x
071 57 9 171 121 y
072 58 : Colon 172 122 z
073 59 ; Semicolon 173 123 { Left brace
074 60 < Less than sign 174 124 | Vertical bar
075 61 = Equal sign 175 125 } Right brace
076 62 > Greather than sign 176 126 Tilde
077 63 ? Question mark 177 127
ˆ? Delete

American Standard Code for Information Interchange (ASCII) specifies a correspondence between bit pat-
terns and character symbols. The octal and decimal representations of the bit patterns are shown along

139
ASCII Table

with a description of the character symbol. The first 32 codes (numbers 0–31 decimal) as well as the last
(number 127 decimal) are non-printing characters which were initially intended to control devices or pro-
vide meta-information about data streams. For example, decimal 10 ended a line on a line printer and
decimal 8 backspaced one character so that the preceding character would be overstruck. The control
sequence column shows the traditional key sequences for inputting these non-printing characters where
the caret ( ) represents the “Control” or “Ctrl” key which must be held down while the following key is
depressed.
ˆ

140
Index

Note: In this index MATLAB commands come first, followed by symbols, and only then does the index begin
with “A”.
Note: All words shown in typewriter font are MATLAB commands or predefined variables unless it is specifically
stated that they are defined locally (i.e., in this document).
Note: If an item is a primary topic of a section, an appendix, or a subsection, this is indicated as well as the page
number (in parentheses).

MATLAB functions cotd, 12, 128


cputime, 27, 28, 128, 130
abs, 12, 13, 128 csc, 12, 128
acos, 12, 128 cscd, 12, 128
acosd, 12, 128 csvread, 47, 48, 63, 127
acot, 12, 128 csvwrite, 47, 48, 63, 127
acotd, 12, 128 cumsum, 33, 34
acosh, 12, 128 dblquad, 115, 135
acsc, 12, 128 dbclear, 83, 134
acscd, 12, 128 dbcont, 83, 134
all, 78, 133 dbdown, 83, 134
angle, 13, 128 dbquit, 83, 134
any, 78, 133 dbstack, 83, 134
asec, 128 dbstatus, 83, 134
asecd, 128 dbstep, 83, 134
asin, 12, 128 dbstop, 83, 134
asind, 12, 128 dbtype, 83, 134
asinh, 12, 128 dbup, 83, 134
atan, 12, 128 demo, 3, 15, 16, 42, 51, 54, 126, 131
atand, 12, 128 det, 66, 70, 132
atanh, 12, 128 deval, 108, 110, 135
atan2, 12, 128 diag, 22, 25, 129
axis, 44, 49, 51, 53, 130, 131 diary, 6, 7, 127
ballode, 101 diff, 33, 34
break, 74, 75, 133 disp, 8, 9, 39, 64
bvp4c, 107, 108, 110, 135 doc, 4, 15, 16, 126
bvpinit, 108, 110, 135 echo, 83, 88, 134
case, 74, 75, 133 eig, 29, 66, 70, 81, 132
different than in C, 74 eigs, 67, 70, 132
cat, 34, 35, 129 else, 72, 75, 133
ceil, 12, 128 elseif, 72, 75, 133
cell, 40, 42, 131 end, 22, 71, 72, 74, 75, 86, 88, 133, 134
celldisp, 40, 42, 131 error, 81, 88, 133, 134
chol, 65, 70, 132 errorbar, 47, 49, 130, 132
clear, 8, 9, 12, 31, 129 eval, 88, 89, 90, 134
danger in using, 8 exist, 78, 133
clf, 53, 54, 131 exp, 12, 13, 128
colorbar, 53, 54, 131 expm, 122, 123, 136
colormap, 53, 54, 131 eye, 19, 20, 129
cond, 61, 65, 70, 132 ezcontour, 50, 51, 131
condest, 65, 70, 132 ezcontour3, 50, 51, 131
conj, 13, 128 ezmesh, 50, 51, 131
continue, 74, 75, 133 ezplot, 45, 49, 130
contour, 50, 51, 131 ezplot3, 50, 51, 131
contour3, 50, 51, 131 ezpolar, 46, 49, 130
cos, 12, 128 ezsurf, 50, 51, 131
cosd, 12, 128 factorial, 11, 12, 128
cosh, 12, 128 false, 73, 75, 129
cot, 12, 128 fclose, 63, 64, 127

141
feval, 86, 89, 90, 92, 134 linspace, 43, 47, 49, 130
fft, 118, 121, 134 load, 15, 16, 47, 48, 53, 126, 127
fftshift, 118, 120, 121, 134 be careful, 47
fieldnames, 42 log, 12, 128
figure, 52, 54, 131 logical, 77, 78, 133
fill, 54, 130 loglog, 45, 49, 130
fill3, 54, 131 logm, 123, 136
find, 76, 77, 78, 94, 95, 133, 135 log10, 12, 128
fix, 12, 128 lookfor, 14, 16, 79, 86, 126
fliplr, 23, 25, 129 lu, 68, 70, 132
flipud, 23, 25, 129 max, 32, 34, 132
floor, 12, 128 mean, 32, 34, 132
fminbnd, 113, 115, 135 mesh, 50, 51, 53, 131
fminsearch, 113, 115, 135 meshgrid, 50, 51, 131
fopen, 63, 64, 127 min, 34, 132
for, 22, 71, 75, 133 mod, 12, 128
format, 10, 11, Subsect. 2.6 (29), 126 nargin, 81, 88, 134
fplot, 38, 45, 54, 131 nargout, 81, 88, 134
fprintf, 8, 39, 47, 63, 64, 127 nnz, 94, 95, 135
fscanf, 47, 63, 64, 127 norm, 33, 34, 68, 70, 81, 132
full, 93, 95, 135 null, 68, 70, 132
function, 79, 86, 88, 134 num2str, 39, 42, 48, 64, 131
funm, 123, 136 odeset, 100, 101, 105, 135
fzero, 112, 113, 115, 135 ode113, 97, 135
gca, 56, 131 ode15i, 105, 106, 135
gcf, 56, 131 ode15s, 97, 135
get, 56, 131 ode23, 97, 135
getfield, 42 ode23s, 97, 135
ginput, 51, 54, 131 ode23t, 97, 135
global, 84, 88, 134 ode23tb, 97, 135
gtext, 52, 53, 54, 55, 131 ode45, 97, 135
help, 4, 14, 16, 79, 86, 126 ones, 19, 20, 129
helpdesk, 15, 126 optimset, 113, 115, 135
hilb, 29, 31, 66, 90, 129 orth, 69, 70, 132
hist, 46, 49, 130, 132 orderfields, 42
hold, 44, 49, 130 otherwise, 74, 75, 133
if, 22, 72, 75, 133 path, 80, 130
ifft, 118, 121, 134 pause, 82, 88, 130
ifftshift, 118, 120, 121, 134 pchip, 112, 136
imag, 13, 128 permute, 35, 129
image, 53, 54, 131 persistent, 85, 88, 134
inline, 38 pinv, 62, 63, 132
input, 10, 127 plot, 43, 44, 45, 49, 111, 130
interp1, 112, 136 plot3, 50, 51, 131
interp2, 112, 136 polar, 46, 49, 130
interp3, 112, 136 poly, 110, 112, 136
interpn, 112, 136 polyder, 111, 112, 136
inv, 26, 67, 70, 132 polyfit, 111, 112
ipermute, 35, 129 polyval, 110, 111, 112, 136
ischar, 78, 133 polyvalm, 110, 112, 136
isempty, 78, 85, 133 ppval, 112, 136
isfield, 42 print, 47, 48, 127
isfinite, 78, 133 prod, 34, 132
isinf, 78, 133 profile, 90, 92, 126
islogical, 77, 78, 133 qr, 69, 70, 132
isnan, 78, 133 quad, 114, 115, 135
keyboard, 83, 88, 134 quadgk, 114, 115, 135
lasterr, 89, 90, 92, 134 quadl, 114, 115, 135
legend, 52, 54, 131 rand, 19, 20, 46, 61, 129
length (number of elements in), 19, 20, 33, 77, 129 randn, 19, 20, 46, 129
linsolve, 57, 60, 132 randperm, 19, 20, 129

142
rank, 69, 70, 132 true, 73, 75, 129
rats, 63, 130 type, 14, 15, 16, 79, 86, 126
real, 13, 128 vander, 111, 112, 129
rem, 12, 128 vectorize, 38, 42
repmat, 24, 25, 129 view, 50, 51, 131
reshape, 22, 24, 25, 129 while, 22, 74, 75, 133
return, 81, 83, 88, 134 who, 15, 16, 126
rmfield, 42 whos, 15, 16, 126
roots, 110, 112, 136 xlabel, 46, 49, 52, 55, 130
rot90, 25, 129 xor, 73, 75, 132
round, 12, 128 ylabel, 46, 49, 52, 55, 130
rref, Sect. 5 (57), 60, 88, 132 zeros, 19, 20, 129
save, 15, 16, 126 zlabel, 51, 53, 55, 131
sec, 12, 128
secd, 12, 128 Symbols
semilogx, 45, 49, 130
semilogy, 45, 49, 130 +, 7, 25, 28, 125, 128
set, 55, 56, 131 exception to, 27
setfield, 42 -, 7, 25, 28, 125, 128
sign, 12, 128 *, 7, 25, 28, 125, 128
sin, 12, 128 .*, 26, 28, 125, 128
sind, 12, 128 /, 7, 26, 28, 125, 128
sinh, 12, 128 warning about matrix division, 26
size, 19, 20, 129 ./, 26, 28, 125, 128
sort, 33, 34, 132 \, 7, 28, 57, 59, 62, 63, 125, 128
sparse, 93, 94, 95, 135 .\, 26, 28, 128
spconvert, 94, 95, 135
ˆ , 6, 7, 26, 28, 125, 128
spdiags, 93, 95, 135
differences from diag, 93
ˆ
. , 27, 28, 128

, 7, 18, 20, 129
speye, 95, 129, 135 . ′ , 18, 20, 129
spfun, 95, 135 ..., 14, 125
spline, 112, 136 %, 14, 125
sprand, 94, 95, 129, 135 ,, 7, 9, 17, 25, 125
sprandn, 94, 95, 129, 135 ;, 7, 9, 17, 25, 125
sprandsym, 94, 95, 129 :, 18, Subsect. 2.2 (21), Subsect. 2.3 (21), 25, 125
sprintf, 39, 42, 131 <, 72, 132
spy, 95, 135 <=, 72, 132
sqrt, 12, 31, 128 >, 72, 132
sqrtm, 26, 123, 136 >=, 72, 132
squeeze, 35, 129 ==, 72, 132
sscanf, 39, 42, 131 =, 72, 132
std, 33, 34, 132 ˜
&, 73, 75, 132
strcmp, 72, 131, 132 &&, 73, 132
str2num, 39, 42, 131 |, 73, 75, 132
struct, 41, 42, 131 ||, 73, 132
subplot, 44, 49, 51, 130, 131 , 73, 75, 132
sum, 34, 76, 132 ˜
!, See factorial
surf, 50, 51, 53, 131 [], 24, 25, 31, 129
svd, 69, 70, 132 @, 37
switch, 22, 74, 75, 133 ↑ up-arrow key, 6, 7, 125
different than in C, 74
tan, 12, 128 A
tand, 12
tanh, 12, 128 AH , See Conjugate transpose
text, 51, 53, 54, 55, 131 AT , See Transpose
tic, 28, 128, 130 A+ , See Pseudoinverse
title, 46, 49, 55, 130 Abort statement, 14
toc, 28, 128, 130 abs, 12, 13, 128
tril, 23, 25, 129 Accuracy, 10
triplequad, 115 principle, 11
triu, 23, 25, 129 acos, 12, 128

143
C
acosd, 12, 128
acosh, 12, 128 ˆCC,(programming
14, 16, 126
language), 7, 21, 37, 41, 42, 47, 54, 63,
acot, 12, 128 64, 74, 84, 85, 127, 131
acotd, 12, 128 C++ (programming language), 41, 89
acsc, 12, 128 Calculator, Subsect. 1.1 (6)
acscd, 12, 128 case, 74, 75, 133
all, 78, 133 different than in C, 74
AND (logical operator), 73, 75, 132 Case sensitive, 9
angle, 13, 128 cat, 34, 35, 129
Anonymous functions, See Function Catching errors, 89
ans, 8, 9, 126 ceil, 12, 128
any, 78, 133 cell, 40, 42, 131
Arithmetic progression, 21 Cell array, 37, Subsect. 3.4 (40), 46
Arithmetical operations, Subsect. 1.1 (6), Subsect. 2.4 celldisp, 40, 42, 131
(25), 125 Character string, 7, Subsect. 3.3 (39), 131
+, 7, 25, 28, 125, 128 appending to, 39
exception to, 27 concatenating, 39
-, 7, 25, 28, 125, 128 converting to, 39
/, 7, 26, 28, 125, 128 comparing strings, 72
warning about matrix division, 26 executing, 88
./, 26, 28, 125, 128 multiline, 39
*, 7, 25, 28, 125, 128 TEX commands in, 52
.*, 26, 28, 125, 128 chol, 65, 70, 132
\, 7, 28, 57, 59, 62, 63, 125, 128 Cholesky decomposition, 65
.\, 26, 28, 128 clear, 8, 9, 12, 31, 129
ˆ , 7, 26, 28, 125, 128 danger in using, 8
ˆ
. , 27, 28, 128
elementwise, 26, 27
Clear (a figure), 52
See also clf
Array, Sect. 2 (16) clf, 53, 54, 131
See also Matrix, Multidimensional array, Vector, or Closure, 99, 113
Cell array Clown, 53
ASCII character representation, 39, 51, 139 Colon operator, 18, Subsect. 2.2 (21), Subsect. 2.3 (21),
asec, 128 25, 125
asecd, 128 possible floating-point errors in, 21, 43
asin, 12, 128 See also linspace
asind, 12, 128 colorbar, 53, 54, 131
asinh, 12, 128 Color map, 53
atan, 12, 128 colormap, 53, 54, 131
atand, 12, 128 Colors, See RGB components
atanh, 12, 128 Command, 3
atan2, 12, 128 See also Function
Augmented matrix form, 57–60 Comment character, 14, 125
See also Matrix Complex conjugate, 13
Average value, 32 Complex numbers, 6, Subsect. 1.6 (13)
axis, 44, 49, 51, 53, 130, 131 Conchoid of Nicodemes
cond, 61, 65, 70, 132
B condest, 65, 70, 132
Condition number, See Matrix
Ball, 103–105 conj, 13, 128
ballode, 101 Conjugate transpose, 18
Bessel’s equation, See Initial-value ordinary differential See also Transpose
equations Continuation (of a line), 14, 125
Binary format, 15, 47 Continuation method, 109
Boundary-value ordinary differential equations, Sect. 11 continue, 74, 75, 133
(106), 110, 135 contour, 50, 51, 131
continuation method, 109 Contour plot, 50
break, 74, 75, 133 contour3, 50, 51, 131
bvp4c, 107, 108, 110, 135 Control flow, See Programming language
bvpinit, 108, 110, 135 cos, 12, 128

144
cosd, 12, 128 duffing a (locally defined), 98, 101
cos z, 13 duffing c (locally defined), 99, 100
cosh, 12, 128 duffing event (locally defined), 102
cot, 12, 128 duffing n (locally defined), 99, 102
cotd, 12, 128 duffing ode (locally defined), 99
CPU, 27 duffing p (locally defined), 99
cputime, 27, 28, 128, 130 duffing p2 (locally defined), 99
csc, 12, 128
cscd, 12, 128 E
csvread, 47, 48, 63, 127
csvwrite, 47, 48, 63, 127 ez , 13
Cubic splines, See Interpolation Earth, 53
cumsum, 33, 34 echo, 83, 88, 134
Cursor eig, 29, 66, 70, 81, 132
entering current position, 51 Eigenvalues, 29, 65, 66, 67, 69, 81
definition of, 66
D Eigenvectors, 66, 67, 69, 81
eigs, 67, 70, 132
Data else, 72, 75, 133
best polynomial fit to, 110 elseif, 72, 75, 133
closing files, 63 end, 22, 71, 72, 74, 75, 86, 88, 133, 134
manipulation, Subsect. 2.8 (32), 132 eps, 9, 10, 74, 126
opening files, 63 See also Machine epsilon
reading into MATLAB, 47, 48, 63, 94, 127 Erase (a figure), 52
writing from MATLAB, 47, 48, 63, 127 See also clf
Data types, 37 error, 81, 88, 133, 134
dblquad, 115, 135 Error bars, 46, 47
dbclear, 83, 134 errorbar, 47, 49, 130, 132
dbcont, 83, 134 Euclidean length, See Length of a vector
dbdown, 83, 134 eval, 88, 89, 90, 134
dbquit, 83, 134 EXCLUSIVE OR (logical operator), 73, 75, 132
dbstack, 83, 134 exist, 78, 133
dbstatus, 83, 134 exp, 12, 13, 128
dbstep, 83, 134 expm, 122, 123, 136
dbstop, 83, 134 Exponentiation, 6, 7, 26
dbtype, 83, 134 Extrapolation, 111
dbup, 83, 134 See also Interpolation
Debugging m-files, See Function files and Script files eye, 19, 20, 129
demo, 3, 15, 16, 42, 51, 54, 126, 131 ezcontour, 50, 51, 131
Demonstration program, 3, 15, 51, 53 ezcontour3, 50, 51, 131
det, 66, 70, 132 ezmesh, 50, 51, 131
Determinant, 66 ezplot, 45, 49, 130
deval, 108, 110, 135 ezplot3, 50, 51, 131
diag, 22, 25, 129 ezpolar, 46, 49, 130
Diagonals, See Matrix ezsurf, 50, 51, 131
diary, 6, 7, 127
diff, 33, 34 F
Digits of accuracy, 10
disp, 8, 9, 39, 64 factorial, 11, 12, 128
Display Factorial function, 11
formatting the, Subsect. 1.4 (10) false, 73, 75, 129
misinterpreting, Subsect. 2.6 (29) FALSE (result of logical expression), 73
suppressing, 7, 9, 17, 25, 125 Fast Fourier transform, See Fourier transform
variable, 8, 9, 64 fclose, 63, 64, 127
See also disp, fprintf feval, 86, 89, 90, 92, 134
doc, 4, 15, 16, 126 fft, 118, 121, 134
Documentation (MATLAB), 15 fftshift, 118, 120, 121, 134
Dot product, 27 Field, See Structure
Duffing’s equation, See Initial-value ordinary fieldnames, 42
differential equations figure, 52, 54, 131
duffing (locally defined) fill, 54, 130

145
fill3, 54, 131 Fun tion ( ont.)
find, 76, 77, 78, 94, 95, 133, 135 passing function name in argument list, Subsect. 3.2
Finite differences, 33 (38), 89
fix, 12, 128 passing arguments indirectly, See Closure
Floating-point numbers, 9, 21 primary, 86
Floating-point operations, See Flops private, 88
fliplr, 23, 25, 129 return, 81, 83, 88, 134
flipud, 23, 25, 129 saving parameters in, 85–86
floor, 12, 128 subfunctions in, 80, 86
Flops (fl oating-point operations), 27 Function handle, 37, 42
Flow control, See Programming language Functions (mathematical)
fminbnd, 113, 115, 135 See also Polynomials
fminsearch, 113, 115, 135 common mathematical, Subsect. 1.5 (11)
fopen, 63, 64, 127 definite integrals of, 114
for, 22, 71, 75, 133 “hijacked”, 88
format, 10, 11, Subsect. 2.6 (29), 126 local minimum of, 113
Format options (in format command), 10, 11, 126 numerical operations on, 112, 115, 135
Format specifications (in fprintf, fscanf, sprintf, and zeroes of, 112, 114
sscanf), 63 funm, 123, 136
Fourier series, Sect. 14 (115) fzero, 112, 113, 115, 135
complex, 116
real, 115 G
Fourier transform, Sect. 14 (115)
discrete, Sect. 14 (115), 121, 134 Gauss-Kronrod quadrature (for numerical integration),
fast (FFT), 119 114
fplot, 38, 45, 54, 131 Gauss-Lobatto quadrature (for numerical integration),
fprintf, 8, 39, 47, 63, 64, 127 114
printing a matrix, 64 Gaussian elimination, 57, 61
specifications (format), 63 gca, 56, 131
Frequency, See Power gcf, 56, 131
fscanf, 47, 63, 64, 127 Generalized eigenvalue problem, 67
specifications (format), 63 get, 56, 131
full, 93, 95, 135 get intervals fast (locally defined), 92
function, 79, 86, 88, 134 get intervals slowly (locally defined), 91
Function, 22, Subsect. 8.3 (79) getfield, 42
anonymous, Subsect. 3.1 (37), 42 ginput, 51, 54, 131
warning, 38 global, 84, 88, 134
built-in, 11, 14, 15 Graphics, Sect. 4 (42)
commands in, 83, 88, 134 advanced techniques, Subsect. 4.3 (51), 131
comments in, 79 changing endpoints, 44
conflict between function and variable name, 11 customizing lines and markers, 44
debugging, 82, 83 demonstration, 42
definition line, 79 handle, Subsect. 4.4 (55)
differences from command, 3 holding the current plot, 44
end statement, 86 labelling, 51–54
ending, 86 text properties, 55
error, 81, 88, 133, 134 using TEX commands, 52
example using multiple input and output arguments, multiple plots, 44
81 multiple windows, 52
function (required word), 79, 86, 88, 134 object, 55
inline, 38 handle for an, 55
warning, 38 printing, 47, 48, 127
input and output arguments, 79, 83–84 properties, Subsect. 4.4 (55)
pass by reference, 79 saving to a file, 47, 48, 127
pass by value, 79 two-dimensional, Subsect. 4.1 (43), 130
variable number of, 81 three-dimensional, Subsect. 4.2 (49)
name of, 13 window, 42
warning about user-defined m-files, 14 Gravity, 104
nested, 80, 86 gravity (locally defined), 105
order in which MATLAB searches for functions, 80, gravity event (locally defined), 105
87 gravity init (locally defined), 105

146
gravity ode (locally defined), 105
gtext, 52, 53, 54, 55, 131 Inline functions, See Function
Inner product, 27
H input, 10, 127
Integration, numerical, 114
H
, See Conjugate transpose Interpolation, 111, 112
Handle, See Function handle cubic, 111
Handle graphics, See Graphics cubic splines, 111, 112
Helix, 50 Hermite cubic interpolation, 111
help, 4, 14, 16, 79, 86, 126 how to do extrapolation, 112
Help facility, Subsect. 1.8 (14) linear splines, 111
keyword, 14 interp1, 112, 136
getting help, 16, 126 interp2, 112, 136
helpdesk, 15, 126 interp3, 112, 136
Hermite polynomials, See Interpolation interpn, 112, 136
hilb, 29, 31, 66, 90, 129 inv, 26, 67, 70, 132
hilb local (locally defined), 80 ipermute, 35, 129
Hilbert matrix, 29, 30, 61, 66, 70, 80, 111 ischar, 78, 133
function file for, 80, 90 isempty, 78, 85, 133
hist, 46, 49, 130, 132 isfield, 42
Histogram, 46 isfinite, 78, 133
hold, 44, 49, 130 isinf, 78, 133
islogical, 77, 78, 133
I isnan, 78, 133

I, See Identity matrix J


i, 6, 9, 126
Identity matrix, 19 j, 6, 9, 126
See also eye Java (programming language), 89
if, 22, 72, 75, 133
ifft, 118, 121, 134 K
ifftshift, 118, 120, 121, 134
imag, 13, 128 keyboard, 83, 88, 134
image, 53, 54, 131 Keyword, 14
Imaginary numbers, 6, 9, 126
Inf, 9, 44, 126 L
Initial-value ordinary differential equations, Sect. 10
(95) lasterr, 89, 90, 92, 134
Bessel’s equation, 105 Left division, See \
Duffing’s equation, 95–103 legend, 52, 54, 131
first-order system, 95 Lemniscate of Bernoulli, 45
with constant coefficients, 122 length (number of elements in), 19, 20, 33, 77, 129
solvers, 96, 97, 135 Length of a vector (i.e., Euclidean length), 33
ode113, 97, 135 See also norm
ode15i, 105, 106, 135 Linear splines, See Interpolation
ode15s, 97, 135 Linear system of equations, Sect. 5 (57), Subsect. 5.3
ode23, 97, 135 (61)
ode23s, 97, 135 least-squares solution, 62, 110
ode23t, 97, 135 overdetermined, Subsect. 5.3 (61), 110
ode23tb, 97, 135 solving by \, 26, 57, 62
ode45, 97, 135 solving by linsolve, 57
absolute error, 96 solving by rref, Sect. 5 (57)
adaptive step size, 96 underdetermined, Subsect. 5.3 (61)
events, 102 linsolve, 57, 60, 132
passing parameters to, 99 linspace, 43, 47, 49, 130
properties of, 101 load, 15, 16, 47, 48, 53, 126, 127
relative error, 96 be careful, 47
statistics for, 101 log, 12, 128
stiff, 97, 103 logical, 75, 77, 78, 133
Van der Pol’s equation, 103 LOGICAL AND (short circuiting logical operator), 73,
Inline, 38 132

147
LOGICAL OR (short circuiting logical operator), 73, Matrix ( ont.)
132 QR decomposition, 69
Logical (data type), 37, 75 positive definite, 94
Logical expression, 72 preallocation of, 19, 40, 80
result of, 73 pseudoinverse of, 62
Logical functions, 78, 133 replicating, 24
Logical operators, 73, 132 reshaping, 22, 24
AND (&), 73, 75, 132 singular, 59, 60, 65, 66
AND (short-circuit) (&&), 73, 132 warning of, 67
applied to matrices, Subsect. 8.2 (75) singular value decomposition, 69
result of, 75 sparse, Sect. 9 (92), 135
EXCLUSIVE OR (xor), 73, 75, 132 specialized, 129
NOT ( ), 73, 75, 132 sum of elements, 32
˜
OR (|), 73, 75, 132 SVD, See Singular value decomposition (above)
OR (short-circuit) (||), 73, 132 symmetric, 18, 94
loglog, 45, 49, 130 tridiagonal, 65, 92
logm, 123, 136 unitary, 69
log10, 12, 128 upper triangular part of, 23
lookfor, 14, 16, 79, 86, 126 Vandermonde, See Vandermonde matrix
lu, 68, 70, 132 max, 32, 34, 132
LU decomposition, 68 Maximum value, 32
mean, 32, 34, 132
M Mean value, 32
Memory (of variables), 31
Machine epsilon (eps), 9, 126 mesh, 50, 51, 53, 131
calculation of, 74 meshgrid, 50, 51, 131
Mathematical functions, Subsect. 1.5 (11), 13, Subsect. M-file, 79
2.7 (31), 128 min, 34, 132
Matrix Minimum value, 32
augmented, 57–60 mod, 12, 128
is not a matrix, 58 Monotonicity, test for, 33
Cholesky decomposition, 65 Monty Python, 39
condition number, 65 Moore-Penrose conditions, 62
approximation to, 65 Mouse location, See ginput
defective, 66 Multidimensional arrays, Subsect. 2.9 (34)
deleting rows or columns, 24 permute order, 35
determinant of, See Determinant
diagonals of, 22, 93, 94 N
elementary, 20, 129
elementary operations, 129 NaN, 9, 126
empty, See See null below nargin, 81, 88, 134
extracting submatrices, 21 nargout, 81, 88, 134
full, 92 nested ex (locally defined), 87
generating, Subsect. 2.1 (17), Subsect. 2.3 (21) Newton’s laws, 103
individual elements, 19 nlode (locally defined), 107
by submatrices, 20 nlode all (locally defined), 109
Hermitian, 18 nlode bc (locally defined), 107
Hilbert, See Hilbert matrix nlode bc2 (locally defined), 107
identity, 19 nlode yic (locally defined), 108
inverse of, 67 nnz, 94, 95, 135
Jacobian, 98, 103 norm, 33, 34, 68, 70, 81, 132
lower triangular part of, 23, 68 Norm
unit, 68 matrix, 68
LU decomposition, 68 Frobenius, 68
manipulation, Subsect. 2.3 (21), 129 p -norm, 68
“masking” elements of, 77 vector, 68
maximum value, 32 NOT (logical operator), 73, 75, 132
minimum value, 32 Notation (for vectors and matrices), 5
multidimensional, Subsect. 2.9 (34) null, 68, 70, 132
null, 24, 25, 31, 129 Null matrix, 24, 25, 129
orthogonal, 69 Null space, 68

148
num2str, 39, 42, 48, 64, 131 Polynomials ( ont.)
order of, 111
O representing by vector, 110
roots of, 110
Ode, See Initial-value ordinary differential equations polyval, 110, 111, 112, 136
odeset, 100, 101, 105, 135 polyvalm, 110, 112, 136
ode113, 97, 135 Positive definite matrix, See Matrix
ode15i, 105, 106, 135 Power, 116, 117
ode15s, 97, 135 average, 116
ode23, 97, 135 definition of, 116
ode23s, 97, 135 frequency of, 116
ode23t, 97, 135 in each mode, 116, 117
ode23tb, 97, 135 instantaneous, 116
ode45, 97, 135 spectrum, 116
ones, 19, 20, 129 ppval, 112, 136
Operator precedence, Subsect. 2.5 (29) prealloc (locally defined), 81
optimset, 113, 115, 135 Precedence, See Operator precedence
OR (logical operator), 73, 75, 132 Predefined variables, See Variables
Ordinary differential equations, See Initial-value Principles about computer arithmetic, 9, 11
ordinary differential equations and Boundary-value print, 47, 48, 127
ordinary differential equations Printing, See Display
orderfields, 42 prod, 34, 132
orth, 69, 70, 132 Product
Orthonormal basis, 69 dot, See Dot product
otherwise, 74, 75, 133 inner, See Inner product
Outer product, 27 outer, See Outer product
Overdetermined system, See Linear system of equations profile, 90, 92, 126
Programming language (MATLAB), Sect. 8 (71)
P flow control, Subsect. 8.1 (71), 75, 133
break out of, 74
Parentheses, 9 continue loop, 74
path, 80, 130 for loops, 71
Path, See Search path if statement, 71
pause, 82, 88, 130 switch statement, 22, 74, 75, 133
pchip, 112, 136 different than in C, 74
permute, 35, 129 while loops, 73
persistent, 85, 88, 134 needed less frequently, 75
uses of, 85 Pseudoinverse, See Matrix
Phase plane, See Plotting Pseudorandom numbers, See Random numbers
pi, 8, 9, 126 Pythagorean theorem, 11
Piecewise polynomials, See Interpolation
pinv, 62, 63, 132 Q
plot, 43, 44, 45, 49, 111, 130
Plot, generating a, See Graphics QR decomposition, 69
Plotting qr, 69, 70, 132
a curve, 43, 49 quad, 114, 115, 135
a function, 45 Quadratic polynomial, roots of, 13
a parametric function, 45 quadgk, 114, 115, 135
an implicit function, 45 quadl, 114, 115, 135
in polar coordinates, 46 Quote mark, 7
phase plane, 98
plot3, 50, 51, 131 R
polar, 46, 49, 130
Polar coordinates, 46 rand, 19, 20, 46, 61, 129
poly, 110, 112, 136 randn, 19, 20, 46, 129
polyder, 111, 112, 136 Random matrix, 19, 23, 61, 95, 135
polyfit, 111, 112 Random numbers, 19
Polynomials, Sect. 12 (110), 136 Gaussian distribution, 19, 46
differentiating, 111 normal distribution, 19
evaluating, 110 pseudorandom numbers, 19
finding minimum and maximum of, 111 seed, 19

149
Random numbers ( ont.) Singular value decomposition, 69
state, 19 sinh, 12, 128
uniform distribution, 19, 46 size, 19, 20, 129
randperm, 19, 20, 129 sort, 33, 34, 132
rank, 69, 70, 132 Sort numbers, 33
Rank of matrix, 69 sparse, 93, 94, 95, 135
rats, 63, 130 spconvert, 94, 95, 135
Rational approximation to floating-point number, 63, spdiags, 93, 95, 135
130 differences from diag, 93
RCOND, 61, 66, 67 speye, 95, 129, 135
real, 13, 128 spfun, 95, 135
realmax, 9, 126 spline, 112, 136
realmin, 9, 10, 126 Splines, See Interpolation
Reduced row echelon form, 58 sprand, 94, 95, 129, 135
round-off errors in, 59 sprandn, 94, 95, 129, 135
Relational operators, 72, 132 sprandsym, 94, 95, 129
<, 72, 132 sprintf, 39, 42, 131
<=, 72, 132 specifications (format), 63
>, 72, 132 spruce (locally defined function), 82
>=, 72, 132 spy, 95, 135
==, 72, 132 sqrt, 12, 31, 128
=, 72, 132 sqrtm, 26, 123, 136
˜matrix, Subsect. 8.2 (75) squeeze, 35, 129
result of, 75 sscanf, 39, 42, 131
rem, 12, 128 specifications (format), 63
Remainder, 12, 128 Standard deviation, 33
Request input, 10 Statements
repmat, 24, 25, 129 executing in text variables, 88
reshape, 22, 24, 25, 129 rerunning previous, 10
return, 81, 83, 88, 134 separating on a line, 7, 9, 17, 25, 125
RGB components (of a color), 53 std, 33, 34, 132
rmfield, 42 Stiff ode, 97, 103
roots, 110, 112, 136 strcmp, 72, 131, 132
rot90, 25, 129 String, See Character string
round, 12, 128 str2num, 39, 42, 131
Round-off errors, Subsect. 1.3 (9), 11, 21, 23, 26, 30, 31, struct, 41, 42, 131
43, 59, Subsect. 5.2 (60), 61, 67 Structure, 37, 41, 113
rref, Sect. 5 (57), 60, 88, 132 field, 41
Subfunctions, See Function files
S subplot, 44, 49, 51, 130, 131
warning, 44
save, 15, 16, 126 sum, 34, 76, 132
Save terminal commands, 6 surf, 50, 51, 53, 131
Save work, 6 Surface plot, 50
Scientific notation, 6 changing view, 50
Scope, See Variables filled-in, 50
Script files, 13, 79, 83 wire-frame, 50
debugging, 82, 83 svd, 69, 70, 132
names of, 13 SVD, See Singular value decomposition
Search path, 80, 87 switch, 22, 74, 75, 133
sec, 12, 128
secd, 12, 128 T
semilogx, 45, 49, 130
T
semilogy, 45, 49, 130 , See Transpose
set, 55, 56, 131 tan, 12, 128
setfield, 42 tand, 12
sign, 12, 128 tanh, 12, 128
Simpson’s method (of numerical integration), 114 Taylor series expansion, 122
sin, 12, 128 TEX, See Character string
sind, 12, 128 text, 51, 53, 54, 55, 131
sin z, 13 Text properties, 55

150
Variables, string ( ont.)
Text window, 42
See also Character string
tic, 28, 128, 130
typeless, 8, 84
Time, See cputime, tic, toc
vdp n (locally defined), 103
title, 46, 49, 55, 130
vdp ode (locally defined), 103
multiline, See Character string, multiline
vdpj n (locally defined), 103
Title
Vector
for entire figure, 56
average value of elements, 32
toc, 28, 128, 130
column vs. row, 17
Transpose, 18, 20, 129
deleting elements, 24
conjugate, 18, 20, 129
generating, Subsect. 2.1 (17)
Trigonometric functions, Subsect. 1.5 (11), Subsect. 2.7
individual elements, 19
(31)
“masking” elements of, 77
tril, 23, 25, 129
maximum value, 32
triplequad, 115
mean value of elements, 32
triu, 23, 25, 129
minimum value, 32
true, 73, 75, 129
preallocation of, 19, 40, 80
TRUE (result of logical expression), 73
repeated elements, testing for, 33
type, 14, 15, 16, 79, 86, 126
sort elements, 33
standard deviation of elements, 33
U sum of elements, 32
vectorize, 38, 42
Underdetermined system, See Linear system of
Vectorizing code, Subsect. 8.5 (90)
equations profile execution time, 90
view, 50, 51, 131
V
W
Van der Pol’s equation, See Initial-value ordinary
differential equations
while, 22, 74, 75, 133
vander, 111, 112, 129 who, 15, 16, 126
Vandermonde matrix, 111 whos, 15, 16, 126
Variables, Subsect. 1.2 (7) Workspace, 6, 83
about, 9
case sensitive, 9 X
conflict between variable and function name, 11
deleting, 9 xlabel, 46, 49, 52, 55, 130
global, 84 xor, 73, 75, 132
inputting, 10
list of, 15 Y
loading, 15
local, 79, 84 ylabel, 46, 49, 52, 55, 130
logical, 77
modifying, 84 Z
overwriting, 7
persistent, 85 zeros, 19, 20, 129
predefined, 8, 9, 126 zlabel, 51, 53, 55, 131
ans, 8, 9, 126
eps, 9, 10, 74, 126
i, 6, 9, 126
Inf, 9, 44, 126
j, 6, 9, 126
NaN, 9, 126
overwriting, 8, 71
pi, 8, 9, 126
realmax, 9, 126
realmin, 9, 10, 126
saving, 15
saving local variables in functions, 85
scope of, 87
special cases of vectors or matrices, 7
static, 85
string, 7, Subsect. 3.3 (39)

151
152

You might also like