Linear Algebra

David Cherney, Tom Denton,

Rohit Thomas and Andrew Waldron
2

Edited by Katrina Glaeser and Travis Scrimshaw

First Edition. Davis California, 2013.

This work is licensed under a

Creative Commons Attribution-NonCommercial-
ShareAlike 3.0 Unported License.

2
 Contents

1 What is Linear Algebra? 9

1.1 Organizing Information . . . . . . . . . . . . . . . . . . . . . . 9
1.2 What are Vectors? . . . . . . . . . . . . . . . . . . . . . . . . 12
1.3 What are Linear Functions? . . . . . . . . . . . . . . . . . . . 15
1.4 So, What is a Matrix? . . . . . . . . . . . . . . . . . . . . . . 20
1.4.1 Matrix Multiplication is Composition of Functions . . . 25
1.4.2 The Matrix Detour . . . . . . . . . . . . . . . . . . . . 26
1.5 Review Problems . . . . . . . . . . . . . . . . . . . . . . . . . 30

2 Systems of Linear Equations 37

2.1 Gaussian Elimination . . . . . . . . . . . . . . . . . . . . . . . 37
2.1.1 Augmented Matrix Notation . . . . . . . . . . . . . . . 37
2.1.2 Equivalence and the Act of Solving . . . . . . . . . . . 40
2.1.3 Reduced Row Echelon Form . . . . . . . . . . . . . . . 40
2.1.4 Solution Sets and RREF . . . . . . . . . . . . . . . . . 45
2.2 Review Problems . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.3 Elementary Row Operations . . . . . . . . . . . . . . . . . . . 52
2.3.1 EROs and Matrices . . . . . . . . . . . . . . . . . . . . 52
2.3.2 Recording EROs in (M |I ) . . . . . . . . . . . . . . . . 54
2.3.3 The Three Elementary Matrices . . . . . . . . . . . . . 56
2.3.4 LU , LDU , and P LDU Factorizations . . . . . . . . . . 58
2.4 Review Problems . . . . . . . . . . . . . . . . . . . . . . . . . 61

3
4

2.5 Solution Sets for Systems of Linear Equations . . . . . . . . . 63

2.5.1 The Geometry of Solution Sets: Hyperplanes . . . . . . 64
2.5.2 Particular Solution + Homogeneous Solutions . . . . . 65
2.5.3 Solutions and Linearity . . . . . . . . . . . . . . . . . . 66
2.6 Review Problems . . . . . . . . . . . . . . . . . . . . . . . . . 68

3 The Simplex Method 71

3.1 Pablo’s Problem . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.2 Graphical Solutions . . . . . . . . . . . . . . . . . . . . . . . . 73
3.3 Dantzig’s Algorithm . . . . . . . . . . . . . . . . . . . . . . . 75
3.4 Pablo Meets Dantzig . . . . . . . . . . . . . . . . . . . . . . . 78
3.5 Review Problems . . . . . . . . . . . . . . . . . . . . . . . . . 80

4 Vectors in Space, n-Vectors 83

4.1 Addition and Scalar Multiplication in R n
. . . . . . . . . . . . 84
4.2 Hyperplanes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.3 Directions and Magnitudes . . . . . . . . . . . . . . . . . . . . 88
4.4 Vectors, Lists and Functions: RS . . . . . . . . . . . . . . . . 94
4.5 Review Problems . . . . . . . . . . . . . . . . . . . . . . . . . 97

5 Vector Spaces 101

5.1 Examples of Vector Spaces . . . . . . . . . . . . . . . . . . . . 102
5.1.1 Non-Examples . . . . . . . . . . . . . . . . . . . . . . . 106
5.2 Other Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.3 Review Problems . . . . . . . . . . . . . . . . . . . . . . . . . 109

6 Linear Transformations 111

6.1 The Consequence of Linearity . . . . . . . . . . . . . . . . . . 112
6.2 Linear Functions on Hyperplanes . . . . . . . . . . . . . . . . 114
6.3 Linear Di↵erential Operators . . . . . . . . . . . . . . . . . . . 115
6.4 Bases (Take 1) . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.5 Review Problems . . . . . . . . . . . . . . . . . . . . . . . . . 118

7 Matrices 121
7.1 Linear Transformations and Matrices . . . . . . . . . . . . . . 121
7.1.1 Basis Notation . . . . . . . . . . . . . . . . . . . . . . 121
7.1.2 From Linear Operators to Matrices . . . . . . . . . . . 127
7.2 Review Problems . . . . . . . . . . . . . . . . . . . . . . . . . 129

4
 5

7.3 Properties of Matrices . . . . . . . . . . . . . . . . . . . . . . 133

7.3.1 Associativity and Non-Commutativity . . . . . . . . . 140
7.3.2 Block Matrices . . . . . . . . . . . . . . . . . . . . . . 142
7.3.3 The Algebra of Square Matrices . . . . . . . . . . . . 143
7.3.4 Trace . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
7.4 Review Problems . . . . . . . . . . . . . . . . . . . . . . . . . 146
7.5 Inverse Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
7.5.1 Three Properties of the Inverse . . . . . . . . . . . . . 150
7.5.2 Finding Inverses (Redux) . . . . . . . . . . . . . . . . . 151
7.5.3 Linear Systems and Inverses . . . . . . . . . . . . . . . 153
7.5.4 Homogeneous Systems . . . . . . . . . . . . . . . . . . 154
7.5.5 Bit Matrices . . . . . . . . . . . . . . . . . . . . . . . . 154
7.6 Review Problems . . . . . . . . . . . . . . . . . . . . . . . . . 155
7.7 LU Redux . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
7.7.1 Using LU Decomposition to Solve Linear Systems . . . 160
7.7.2 Finding an LU Decomposition. . . . . . . . . . . . . . 162
7.7.3 Block LDU Decomposition . . . . . . . . . . . . . . . . 165
7.8 Review Problems . . . . . . . . . . . . . . . . . . . . . . . . . 166

8 Determinants 169
8.1 The Determinant Formula . . . . . . . . . . . . . . . . . . . . 169
8.1.1 Simple Examples . . . . . . . . . . . . . . . . . . . . . 169
8.1.2 Permutations . . . . . . . . . . . . . . . . . . . . . . . 170
8.2 Elementary Matrices and Determinants . . . . . . . . . . . . . 174
8.2.1 Row Swap . . . . . . . . . . . . . . . . . . . . . . . . . 175
8.2.2 Row Multiplication . . . . . . . . . . . . . . . . . . . . 176
8.2.3 Row Addition . . . . . . . . . . . . . . . . . . . . . . . 177
8.2.4 Determinant of Products . . . . . . . . . . . . . . . . . 179
8.3 Review Problems . . . . . . . . . . . . . . . . . . . . . . . . . 182
8.4 Properties of the Determinant . . . . . . . . . . . . . . . . . . 186
8.4.1 Determinant of the Inverse . . . . . . . . . . . . . . . . 190
8.4.2 Adjoint of a Matrix . . . . . . . . . . . . . . . . . . . . 190
8.4.3 Application: Volume of a Parallelepiped . . . . . . . . 192
8.5 Review Problems . . . . . . . . . . . . . . . . . . . . . . . . . 193

9 Subspaces and Spanning Sets 195

9.1 Subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
9.2 Building Subspaces . . . . . . . . . . . . . . . . . . . . . . . . 197

5
6

9.3 Review Problems . . . . . . . . . . . . . . . . . . . . . . . . . 202

10 Linear Independence 203

10.1 Showing Linear Dependence . . . . . . . . . . . . . . . . . . . 204
10.2 Showing Linear Independence . . . . . . . . . . . . . . . . . . 207
10.3 From Dependent Independent . . . . . . . . . . . . . . . . . . 209
10.4 Review Problems . . . . . . . . . . . . . . . . . . . . . . . . . 210

11 Basis and Dimension 213

11.1 Bases in Rn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
11.2 Matrix of a Linear Transformation (Redux) . . . . . . . . . . 218
11.3 Review Problems . . . . . . . . . . . . . . . . . . . . . . . . . 221

12 Eigenvalues and Eigenvectors 225

12.1 Invariant Directions . . . . . . . . . . . . . . . . . . . . . . . . 227
12.2 The Eigenvalue–Eigenvector Equation . . . . . . . . . . . . . . 233
12.3 Eigenspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
12.4 Review Problems . . . . . . . . . . . . . . . . . . . . . . . . . 238

13 Diagonalization 241
13.1 Diagonalizability . . . . . . . . . . . . . . . . . . . . . . . . . 241
13.2 Change of Basis . . . . . . . . . . . . . . . . . . . . . . . . . . 242
13.3 Changing to a Basis of Eigenvectors . . . . . . . . . . . . . . . 246
13.4 Review Problems . . . . . . . . . . . . . . . . . . . . . . . . . 248

14 Orthonormal Bases and Complements 253

14.1 Properties of the Standard Basis . . . . . . . . . . . . . . . . . 253
14.2 Orthogonal and Orthonormal Bases . . . . . . . . . . . . . . . 255
14.2.1 Orthonormal Bases and Dot Products . . . . . . . . . . 256
14.3 Relating Orthonormal Bases . . . . . . . . . . . . . . . . . . . 258
14.4 Gram-Schmidt & Orthogonal Complements . . . . . . . . . . 261
14.4.1 The Gram-Schmidt Procedure . . . . . . . . . . . . . . 264
14.5 QR Decomposition . . . . . . . . . . . . . . . . . . . . . . . . 265
14.6 Orthogonal Complements . . . . . . . . . . . . . . . . . . . . 267
14.7 Review Problems . . . . . . . . . . . . . . . . . . . . . . . . . 272

15 Diagonalizing Symmetric Matrices 277

15.1 Review Problems . . . . . . . . . . . . . . . . . . . . . . . . . 281

6
 7

16 Kernel, Range, Nullity, Rank 285

16.1 Range . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 286
16.2 Image . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
16.2.1 One-to-one and Onto . . . . . . . . . . . . . . . . . . 289
16.2.2 Kernel . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
16.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
16.4 Review Problems . . . . . . . . . . . . . . . . . . . . . . . . . 299

17 Least squares and Singular Values 303

17.1 Projection Matrices . . . . . . . . . . . . . . . . . . . . . . . . 306
17.2 Singular Value Decomposition . . . . . . . . . . . . . . . . . . 308
17.3 Review Problems . . . . . . . . . . . . . . . . . . . . . . . . . 312

A List of Symbols 315

B Fields 317

C Online Resources 319

D Sample First Midterm 321

E Sample Second Midterm 331

F Sample Final Exam 341

G Movie Scripts 367

G.1 What is Linear Algebra? . . . . . . . . . . . . . . . . . . . . . 367
G.2 Systems of Linear Equations . . . . . . . . . . . . . . . . . . . 367
G.3 Vectors in Space n-Vectors . . . . . . . . . . . . . . . . . . . . 377
G.4 Vector Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
G.5 Linear Transformations . . . . . . . . . . . . . . . . . . . . . . 383
G.6 Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
G.7 Determinants . . . . . . . . . . . . . . . . . . . . . . . . . . . 395
G.8 Subspaces and Spanning Sets . . . . . . . . . . . . . . . . . . 403
G.9 Linear Independence . . . . . . . . . . . . . . . . . . . . . . . 404
G.10 Basis and Dimension . . . . . . . . . . . . . . . . . . . . . . . 407
G.11 Eigenvalues and Eigenvectors . . . . . . . . . . . . . . . . . . 409
G.12 Diagonalization . . . . . . . . . . . . . . . . . . . . . . . . . . 415
G.13 Orthonormal Bases and Complements . . . . . . . . . . . . . . 421

7
8

G.14 Diagonalizing Symmetric Matrices . . . . . . . . . . . . . . . . 428

G.15 Kernel, Range, Nullity, Rank . . . . . . . . . . . . . . . . . . . 430
G.16 Least Squares and Singular Values . . . . . . . . . . . . . . . . 432

Index 432

8
 What is Linear Algebra?
1
Many difficult problems can be handled easily once relevant information is
organized in a certain way. This text aims to teach you how to organize in-
formation in cases where certain mathematical structures are present. Linear
algebra is, in general, the study of those structures. Namely

Linear algebra is the study of vectors and linear functions.

In broad terms, vectors are things you can add and linear functions are
functions of vectors that respect vector addition. The goal of this text is to
teach you to organize information about vector spaces in a way that makes
problems involving linear functions of many variables easy. (Or at least
tractable.)
To get a feel for the general idea of organizing information, of vectors,
and of linear functions this chapter has brief sections on each. We start
here in hopes of putting students in the right mindset for the odyssey that
follows; the latter chapters cover the same material at a slower pace. Please
be prepared to change the way you think about some familiar mathematical
objects and keep a pencil and piece of paper handy!

1.1 Organizing Information

Functions of several variables are often presented in one line such as
f (x, y) = 3x + 5y .

9
10 What is Linear Algebra?

But lets think carefully; what is the left hand side of this equation doing?
Functions and equations are di↵erent mathematical objects so why is the
equal sign necessary?

A Sophisticated Review of Functions

If someone says

“Consider the function of two variables 7 13b.”

we do not quite have all the information we need to determine the relationship
between inputs and outputs.

Example 1 (Of organizing and reorganizing information)

You own stock in 3 companies: Google, N etf lix, and Apple. The value V of your
stock portfolio as a function of the number of shares you own sN , sG , sA of these
companies is
24sG + 80sA + 35sN .
0 1
1
Here is an ill posed question: what is V 2A?
@
3
The column of three numbers is ambiguous! Is it is meant to denote

• 1 share of G, 2 shares of N and 3 shares of A?

• 1 share of N , 2 shares of G and 3 shares of A?

Do we multiply the first number of the input by 24 or by 35? No one has specified an
order for the variables, so we do not know how to calculate an output associated with
a particular input.1
A di↵erent notation for V can clear this up; we can denote V itself as an ordered
triple of numbers that reminds us what to do to each number from the input.
1
Of course we would know how to calculate an output if the input is described in
the tedious form such as “1 share of G, 2 shares of N and 3 shares of A”, but that is
unacceptably tedious! We want to use ordered triples of numbers to concisely describe
inputs.

10
1.1 Organizing Information 11

0 1 0 1
1 1
Denote V by 24 80 35 and thus write V @2A = 24 80 35 @2A
3B 3

to remind us to calculate 24(1) + 80(2) + 35(3) = 334

because we chose the order G A N and named that order B

0 1
sG
so that inputs are interpreted as @ sA A .
sN

If we change the order for the variables we should change the notation for V .

0 1 0 1
1 1
@
Denote V by 35 80 24 and thus write V 2 A = 35 80 24 2A
@
3 B0 3

to remind us to calculate 35(1) + 80(2) + 24(3) = 264.

because we chose the order N A G and named that order B 0

0 1
sN
so that inputs are interpreted as @ sA A .
sG

The subscripts B and B 0 on the columns of numbers are just symbols2 reminding us
of how to interpret the column of numbers. But the distinction is critical; as shown
above V assigns completely di↵erent numbers to the same columns of numbers with
di↵erent subscripts.
There are six di↵erent ways to order the three companies. Each way will give
di↵erent notation for the same function V , and a di↵erent way of assigning numbers
to columns of three numbers. Thus, it is critical to make clear which ordering is
used if the reader is to understand what is written. Doing so is a way of organizing
information.
2
We were free to choose any symbol to denote these orders. We chose B and B 0 because
we are hinting at a central idea in the course: choosing a basis.

11
12 What is Linear Algebra?

This example is a hint at a much bigger idea central to the text; our choice of
order is an example of choosing a basis3.

The main lesson of an introductory linear algebra course is this: you

have considerable freedom in how you organize information about certain
functions, and you can use that freedom to

1. uncover aspects of functions that don’t change with the choice (Ch 12)

2. make calculations maximally easy (Ch 13 and Ch 17)

3. approximate functions of several variables (Ch 17).

Unfortunately, because the subject (at least for those learning it) requires
seemingly arcane and tedious computations involving large arrays of numbers
known as matrices, the key concepts and the wide applicability of linear
algebra are easily missed. So we reiterate,

Linear algebra is the study of vectors and linear functions.

In broad terms, vectors are things you can add and linear functions are
functions of vectors that respect vector addition.

1.2 What are Vectors?

Here are some examples of things that can be added:

Example 2 (Vector Addition)

(A) Numbers: Both 3 and 5 are numbers and so is 3 + 5.

0 1 0 1 0 1
1 0 1
(B) 3-vectors: @1A + @1A = @2A.
0 1 1
3
Please note that this is an example of choosing a basis, not a statement of the definition
of the technical term “basis”. You can no more learn the definition of “basis” from this
example than learn the definition of “bird” by seeing a penguin.

12
1.2 What are Vectors? 13

(C) Polynomials: If p(x) = 1 + x 2x2 + 3x3 and q(x) = x + 3x2 3x3 + x4 then
their sum p(x) + q(x) is the new polynomial 1 + 2x + x2 + x4 .

(D) Power series: If f (x) = 1+x+ 2!1 x2 + 3!1 x3 +· · · and g(x) = 1 x+ 2!1 x2 1 3
3! x +· · ·
1 2 1 4
then f (x) + g(x) = 1 + 2! x + 4! x · · · is also a power series.

(E) Functions: If f (x) = ex and g(x) = e x then their sum f (x) + g(x) is the new
function 2 cosh x.

There are clearly di↵erent kinds of vectors. Stacks of numbers are not the
only things that are vectors, as examples C, D, and E show. Vectors of
di↵erent kinds can not be added; What possible meaning could the following
have? ✓ ◆
9
+ ex
3
In fact, you should think of all five kinds of vectors above as di↵erent
kinds, and that you should not add vectors that are not of the same kind.
On the other hand, any two things of the same kind “can be added”. This is
the reason you should now start thinking of all the above objects as vectors!
In Chapter 5 we will give the precise rules that vector addition must obey.
In the above examples, however, notice that the vector addition rule stems
from the rules for adding numbers.
When adding the same vector over and over, for example

x + x, x + x + x, x + x + x + x, ... ,

we will write
2x , 3x , 4x , . . . ,
respectively. For example
0 1 0 1 0 1 0 1 0 1 0 1
1 1 1 1 1 4
4 1 = 1 + 1 + 1 + 1 = 4A .
@ A @ A @ A @ A @ A @
0 0 0 0 0 0

Defining 4x = x + x + x + x is fine for integer multiples, but does not help us

make sense of 13 x. For the di↵erent types of vectors above, you can probably

13
14 What is Linear Algebra?

guess how to multiply a vector by a scalar. For example

0 1 011
1
1 B C B 31 C
@1 A = @ 3 A .
3
0 0
A very special vector can be produced from any vector of any kind by
scalar multiplying any vector by the number 0. This is called the zero vector
and is usually denoted simply 0. This gives five very di↵erent kinds of zero
from the 5 di↵erent kinds of vectors in examples A-E above.
(A) 0(3) = 0 (The zero number)
0 1 0 1
1 0
(B) 0 1 = 0A (The zero 3-vector)
@ A @
0 0
(C) 0 (1 + x 2x2 + 3x3 ) = 0 (The zero polynomial)
1 2
(D) 0 1 + x 2!
x + 3!1 x3 + · · · = 0+0x+0x2+0x3+· · · (The zero power series)
(E) 0 (ex ) = 0 (The zero function)
In any given situation that you plan to describe using vectors, you need
to decide on a way to add and scalar multiply vectors. In summary:

Vectors are things you can add and scalar multiply.

Examples of kinds of vectors:

• numbers
• n-vectors
• 2nd order polynomials
• polynomials
• power series
• functions with a certain domain

14
1.3 What are Linear Functions? 15

1.3 What are Linear Functions?

In calculus classes, the main subject of investigation was the rates of change
of functions. In linear algebra, functions will again be the focus of your
attention, but functions of a very special type. In precalculus you were
perhaps encouraged to think of a function as a machine f into which one
may feed a real number. For each input x this machine outputs a single real
number f (x).

In linear algebra, the functions we study will have vectors (of some type)
as both inputs and outputs. We just saw that vectors are objects that can be
added or scalar multiplied—a very general notion—so the functions we are
going to study will look novel at first. So things don’t get too abstract, here
are five questions that can be rephrased in terms of functions of vectors.

Example 3 (Questions involving Functions of Vectors in Disguise)

(A) What number x satisfies 10x = 3?

0 1 0 1
1 0
(B) What 3-vector u satisfies4 @1A ⇥ u = @1A?
0 1
R1 R1
(C) What polynomial p satisfies 1 p(y)dy = 0 and 1 yp(y)dy = 1?

d
(D) What power series f (x) satisfies x dx f (x) 2f (x) = 0?
4
The cross product appears in this equation.

15
16 What is Linear Algebra?

(E) What number x satisfies 4x2 = 1?

All of these are of the form

(?) What vector X satisfies f (X) = B?

with a function5 f known, a vector B known, and a vector X unknown.

The machine needed for part (A) is as in the picture below.

x 10x

This is just like a function f from calculus that takes in a number x and
spits out the number 10x. (You might write f (x) = 10x to indicate this).
For part (B), we need something more sophisticated.

0 1 0 1
x z
@yA @ z A,
z y x

The inputs and outputs are both 3-vectors. The output is the cross product
of the input with... how about you complete this sentence to make sure you
understand.
The machine needed for example (C) looks like it has just one input and
two outputs; we input a polynomial and get a 2-vector as output.

0 R1 1
1
p(y)dy
p @R A.
1
1
yp(y)dy

This example is important because it displays an important feature; the

inputs for this function are functions.
5
In math terminology, each question is asking for the level set of f corresponding to B.

16
1.3 What are Linear Functions? 17

While this sounds complicated, linear algebra is the study of simple func-
tions of vectors; its time to describe the essential characteristics of linear
functions.
Let’s use the letter L to denote an arbitrary linear function and think
again about vector addition and scalar multiplication. Also, suppose that v
and u are vectors and c is a number. Since L is a function from vectors to
vectors, if we input u into L, the output L(u) will also be some sort of vector.
The same goes for L(v). (And remember, our input and output vectors might
be something other than stacks of numbers!) Because vectors are things that
can be added and scalar multiplied, u + v and cu are also vectors, and so
they can be used as inputs. The essential characteristic of linear functions is
what can be said about L(u + v) and L(cu) in terms of L(u) and L(v).
Before we tell you this essential characteristic, ruminate on this picture.

The “blob” on the left represents all the vectors that you are allowed to
input into the function L, the blob on the right denotes the possible outputs,
and the lines tell you which inputs are turned into which outputs.6 A full
pictorial description of the functions would require all inputs and outputs
6
The domain, codomain, and rule of correspondence of the function are represented by
the left blog, right blob, and arrows, respectively.

17
18 What is Linear Algebra?

and lines to be explicitly drawn, but we are being diagrammatic; we only

drew four of each.
Now think about adding L(u) and L(v) to get yet another vector L(u) +
L(v) or of multiplying L(u) by c to obtain the vector cL(u), and placing both
on the right blob of the picture above. But wait! Are you certain that these
are possible outputs!?
Here’s the answer

The key to the whole class, from which everything else follows:

1. Additivity:
L(u + v) = L(u) + L(v) .

2. Homogeneity:
L(cu) = cL(u) .

Most functions of vectors do not obey this requirement.7 At its heart, linear
algebra is the study of functions that do.
Notice that the additivity requirement says that the function L respects
vector addition: it does not matter if you first add u and v and then input
their sum into L, or first input u and v into L separately and then add the
outputs. The same holds for scalar multiplication–try writing out the scalar
multiplication version of the italicized sentence. When a function of vectors
obeys the additivity and homogeneity properties we say that it is linear (this
is the “linear” of linear algebra). Together, additivity and homogeneity are
called linearity. Are there other, equivalent, names for linear functions? yes.

7
E.g.: If f (x) = x2 then f (1 + 1) = 4 6= f (1) + f (1) = 2. Try any other function you
can think of!

18
1.3 What are Linear Functions? 19

Function = Transformation = Operator

And now for a hint at the power of linear algebra. The questions in
examples (A-D) can all be restated as

Lv = w

where v is an unknown, w a known vector, and L is a known linear transfor-

mation. To check that this is true, one needs to know the rules for adding
vectors (both inputs and outputs) and then check linearity of L. Solving the
equation Lv = w often amounts to solving systems of linear equations, the
skill you will learn in Chapter 2.
A great example is the derivative operator.
Example 4 (The derivative operator is linear)
For any two functions f (x), g(x) and any number c, in calculus you probably learnt
that the derivative operator satisfies
d d
1. dx (cf ) = c dx f,
d d d
2. dx (f + g) = dx f + dx g.

If we view functions as vectors with addition given by addition of functions and with
scalar multiplication given by multiplication of functions by constants, then these
familiar properties of derivatives are just the linearity property of linear maps.

Before introducing matrices, notice that for linear maps L we will often
write simply Lu instead of L(u). This is because the linearity property of a

19
20 What is Linear Algebra?

linear transformation L means that L(u) can be thought of as multiplying

the vector u by the linear operator L. For example, the linearity of L implies
that if u, v are vectors and c, d are numbers, then

L(cu + dv) = cLu + dLv ,

which feels a lot like the regular rules of algebra for numbers. Notice though,
that “uL” makes no sense here.

Remark A sum of multiples of vectors cu + dv is called a linear combination of

u and v.

1.4 So, What is a Matrix?

Matrices are linear functions of a certain kind. They appear almost ubiqui-
tously in linear algebra because– and this is the central lesson of introductory
linear algebra courses–

Matrices are the result of organizing information related to linear

functions.

This idea will take some time to develop, but we provided an elementary
example in Section 1.1. A good starting place to learn about matrices is by
studying systems of linear equations.

Example 5 A room contains x bags and y boxes of fruit.

20
1.4 So, What is a Matrix? 21

Each bag contains 2 apples and 4 bananas and each box contains 6 apples and 8
bananas. There are 20 apples and 28 bananas in the room. Find x and y.

The values are the numbers x and y that simultaneously make both of the following
equations true:
2 x + 6 y = 20
4 x + 8 y = 28 .
Here we have an example of a System of Linear Equations.8 It’s a collection
of equations in which variables are multiplied by constants and summed, and
no variables are multiplied together: There are no powers of variables (like x2
or y 5 ), non-integer or negative powers of variables (like y 1/7 or x 3 ), and no
places where variables are multiplied together (like xy).

Reading homework: problem 1

Information about the fruity contents of the room can be stored two ways:
(i) In terms of the number of apples and bananas.
(ii) In terms of the number of bags and boxes.
Intuitively, knowing the information in one form allows you to figure out the
information in the other form. Going from (ii) to (i) is easy: If you knew
there were 3 bags and 2 boxes it would be easy to calculate the number
of apples and bananas, and doing so would have the feel of multiplication
(containers times fruit per container). In the example above we are required
to go the other direction, from (i) to (ii). This feels like the opposite of
multiplication, i.e., division. Matrix notation will make clear what we are
“multiplying” and “dividing” by.
The goal of Chapter 2 is to efficiently solve systems of linear equations.
Partly, this is just a matter of finding a better notation, but one that hints
at a deeper underlying mathematical structure. For that, we need rules for
adding and scalar multiplying 2-vectors;
✓ ◆ ✓ ◆ ✓ ◆ ✓ 0◆ ✓ ◆
x cx x x x + x0
c := and + := .
y cy y y0 y + y0
✓ ◆
8 x
Perhaps you can see that both lines are of the form Lu = v with u = an unknown,
y
v = 20 in the first line, v = 28 in the second line, and L di↵erent functions in each line?
We give the typical less sophisticated description in the text above.

21
22 What is Linear Algebra?

Writing our fruity equations as an equality between 2-vectors and then using
these rules we have:
✓ ◆ ✓ ◆ ✓ ◆ ✓ ◆ ✓ ◆
2 x + 6 y = 20 2x + 6y 20 2 6 20
() = () x +y = .
4 x + 8 y = 28 4x + 8y 28 4 8 28

Now we introduce a function which takes in 2-vectors9 and gives out 2-vectors.
We denote it by an array of numbers called a matrix .
✓ ◆ ✓ ◆✓ ◆ ✓ ◆ ✓ ◆
2 6 2 6 x 2 6
The function is defined by := x +y .
4 8 4 8 y 4 8

A similar definition applies to matrices with di↵erent numbers and sizes.

Example 6 (A bigger matrix)

0 1
0 1 x 0 1 0 1 0 1 0 1
1 0 3 4 B C 1 0 3 4
@ 5 0 3 4A B y C := x @ 5A + y @0A + z @3A + w @4A .
@ z A
1 6 2 5 1 6 2 5
w

Viewed as a machine that inputs and outputs 2-vectors, our 2 ⇥ 2 matrix

does the following:

✓ ◆ ✓ ◆
x 2x + 6y
.
y 4x + 8y

Our fruity problem is now rather concise.

Example 7 (This time in purely mathematical language):

✓ ◆ ✓ ◆✓ ◆ ✓ ◆
x 2 6 x 20
What vector satisfies = ?
y 4 8 y 28
9
✓To ◆be clear, we will use the term 2-vector to refer to stacks of two numbers such
7
as . If we wanted to refer to the vectors x2 + 1 and x3 1 (recall that polynomials
11
are vectors) we would say “consider the two vectors x3 1 and x2 + 1”. We apologize
through giggles for the possibility of the phrase “two 2-vectors.”

22
1.4 So, What is a Matrix? 23

This is of the same Lv = w form as our opening examples. The matrix

encodes fruit per container. The equation is roughly fruit per container
times number of containers equals fruit. To solve for number of containers
we want to somehow “divide” by the matrix.
Another way to think about the above example is to remember the rule
for multiplying a matrix times a vector. If you have forgotten this, you can
actually guess a good rule by making sure the matrix equation is the same
as the system of linear equations. This would require that
✓ ◆✓ ◆ ✓ ◆
2 6 x 2x + 6y
:=
4 8 y 4x + 8y

Indeed this is an example of the general rule that you have probably seen
before ✓ ◆✓ ◆ ✓ ◆ ✓ ◆ ✓ ◆
p q x px + qy p q
:= =x +y .
r s y rx + sy r s

Notice, that the second way of writing the output on the right hand side of
this equation is very useful because it tells us what all possible outputs a
matrix times a vector look like – they are just sums of the columns of the
matrix multiplied by scalars. The set of all possible outputs of a matrix
times a vector is called the column space (it is also the image of the linear
function defined by the matrix).

Reading homework: problem 2

Multiplication by a matrix is an example of a Linear Function, because it

takes one vector and turns it into another in a “linear” way. Of course, we
can have much larger matrices if our system has more variables.

Matrices in Space!

Thus matrices can be viewed as linear functions. The statement of this for
the matrix in our fruity example is as follows.
✓ ◆ ✓ ◆ ✓ ◆✓ ◆
2 6 x 2 6 x
1. = and
4 8 y 4 8 y

23
24 What is Linear Algebra?

✓ ◆ ✓ ◆ ✓ 0 ◆ ✓ ◆✓ ◆ ✓ ◆ ✓ 0◆
2 6 x x 2 6 x 2 6 x
2. + = + .
4 8 y y0 4 8 y 4 8 y0
These equalities can be verified using the rules we introduced so far.
✓ ◆
2 6
Example 8 Verify that is a linear operator.
4 8
The matrix-function is homogeneous if the expressions on the left hand side and right
hand side of the first equation are indeed equal.
✓ ◆ ✓ ◆ ✓ ◆✓ ◆ ✓ ◆ ✓ ◆
2 6 a 2 6 a 2 6
= = a + b
4 8 b 4 8 b 4 8
✓ ◆ ✓ ◆ ✓ ◆
2 a 6bc 2 a+6 b
= + =
4 a 8bc 4 a+8 b

while
✓ ◆✓ ◆  ✓ ◆ ✓ ◆ ✓ ◆ ✓ ◆
2 6 a 2 6 2a 6b
=c a +b = +
4 8 b 4 8 4a 8b
✓ ◆ ✓ ◆
2a + 6b 2 a+6 b
= = .
4a + 8b 4 a+8 b
The underlined expressions are identical, so the matrix is homogeneous.

The matrix-function is additive if the left and right side of the second equation are
indeed equal.

✓ ◆ ✓ ◆ ✓ ◆ ✓ ◆✓ ◆ ✓ ◆ ✓ ◆
2 6 a c 2 6 a+c 2 6
+ = = (a + c) + (b + d)
4 8 b d 4 8 b+d 4 8
✓ ◆ ✓ ◆ ✓ ◆
2(a + c) 6(b + d) 2a + 2c + 6b + 6d
= + =
4(a + c) 8(b + d) 4a + 4c + 8b + 8d
which we need to compare to
✓ ◆✓ ◆ ✓ ◆✓ ◆ ✓ ◆ ✓ ◆ ✓ ◆ ✓ ◆
2 6 a 2 6 c 2 6 2 6
+ =a +b +c +d
4 8 b 4 8 d 4 8 4 8
✓ ◆ ✓ ◆ ✓ ◆ ✓ ◆ ✓ ◆
2a 6b 2c 6d 2a + 2c + 6b + 6d
= + + + = .
4a 8b 4c 8d 4a + 4c + 8b + 8d
Thus multiplication by a matrix is additive and homogeneous, and so it is, by definition,
linear.

24
1.4 So, What is a Matrix? 25

We have come full circle; matrices are just examples of the kinds of linear
operators that appear in algebra problems like those in section 1.3. Any
equation of the form M v = w with M a matrix, and v, w n-vectors is called
a matrix equation. Chapter 2 is about efficiently solving systems of linear
equations, or equivalently matrix equations.

1.4.1 Matrix Multiplication is Composition of Functions

What would happen if we placed two of our expensive machines end to end?

The output of the first machine would be fed into the second.

✓ ◆ ✓ ◆ ✓ ◆
x 2x + 6y 1.(2x + 6y) + 2.(4x + 8y)
y 4x + 8y 0.(2x + 6y) + 1.(4x + 8y)
✓ ◆
10x + 22y
=
4x + 8y

Notice that the same final result could be achieved with a single machine:

✓ ◆ ✓ ◆
x 10x + 22y
.
y 4x + 8y

There is a simple matrix notation for this called matrix multiplication

✓ ◆✓ ◆ ✓ ◆
1 2 2 6 10 22
= .
0 1 4 8 4 8
Try review problem 6 to learn more about matrix multiplication.
In the language10 of functions, if
f :U !V and g : V !W
10
The notation h : A ! B means that h is a function with domain A and codomain B.
See the webwork background set3 if you are unfamiliar with this notation or these terms.

25
26 What is Linear Algebra?

the new function obtained by plugging the outputs if f into g is called g f ,

g f :U !W

where
(g f )(u) = g(f (u)) .
This is called the composition of functions. Matrix multiplication is the tool
required for computing the composition of linear functions.

1.4.2 The Matrix Detour

Linear algebra is about linear functions, not matrices. The following presen-
tation is meant to get you thinking about this idea constantly throughout
the course.

Matrices only get involved in linear algebra when certain

notational choices are made.

To exemplify, lets look at the derivative operator again.

Example 9 of how matrices come into linear algebra.

Consider the equation ✓ ◆
d
+2 f =x+1
dx
where f is unknown (the place where solutions should go) and the linear di↵erential
d
operator dx + 2 is understood to take in quadratic functions (of the form ax2 + bx + c)
and give out other quadratic functions.
Let’s simplify the way we denote the quadratic functions; we will
0 1
a
denote ax2 + bx + c as @ bA .
c B

The subscript B serves to remind us of our particular notational convention; we will

compare to another notational convention later. With the convention B we can say
0 1
✓ ◆ a ✓ ◆
d @ A d
+2 b = + 2 (ax2 + bx + c)
dx dx
c B

26
1.4 So, What is a Matrix? 27

= (2ax + b) + (2ax2 + 2bx + 2c) = 2ax2 + (2a + 2b)x + (b + 2c)

0 1 20 1 0 13
2a 2 0 0 a
@
= 2a + 2b A = 4 @ 2 2 0 A @ b A5 .
b + 2c B 0 1 2 c B

That is, our notational convention for quadratic functions has induced a notation for
d
the di↵erential operator dx + 2 as a matrix. We can use this notation to change the
way that the following two equations say exactly the same thing.

20 1 0 13 0 1
✓ ◆ 2 0 0 a 0
d 4 @ A @ A5
+2 f =x+1, 2 2 0 b = 1A .
@
dx
0 1 2 c B
1 B

Our notational convention has served as an organizing principle to yield the system of
equations

2a =0
2a + 2b = 1
b + 2c = 1

0 1
0
1
with solution 2 A , where the subscript B is used to remind us that this stack of
@
1
4 B
numbers encodes the vector 12 x+ 14 , which is indeed the solution to our equation since,
d
substituting for f yields the true statement dx + 2 ( 12 x + 14 ) = x + 1.

It would be nice to have a systematic way to rewrite any linear equation

as an equivalent matrix equation. It will be a little while before we can learn
to organize information in a way generalizable to all linear equations, but
keep this example in mind throughout the course.
The general idea is presented in the picture below; sometimes a linear
equation is too hard to solve as is, but by organizing information and refor-
mulating the equation as a matrix equation the process of finding solutions
becomes tractable.

27
28 What is Linear Algebra?

d
A simple example with the knowns (L and V are dx and 3, respectively) is
shown below, although the detour is unnecessary in this case since you know
how to anti-di↵erentiate.

To drive home the point that we are not studying matrices but rather lin-
ear functions, and that those linear functions can be represented as matrices
under certain notational conventions, consider how changeable the notational
conventions are.

28
1.4 So, What is a Matrix? 29

Example 10 of how a di↵erent matrix comes into the same linear algebra problem.

Another possible notational convention is to

0 1
a
denote a + bx + cx2 as @ bA .
c B0

With this alternative notation

0 1
✓ ◆
a ✓ ◆
d @ A d
+2 b = + 2 (a + bx + cx2 )
dx dx
c B0

= (b + 2cx) + (2a + 2bx + 2cx2 ) = (2a + b) + (2b + 2c)x + 2cx2

0 1 20 1 0 13
2a + b 2 1 0 a
@
= 2b + 2c A = 4 @0 2 2 A @ bA5 .
2c B0
0 0 2 c B0

Notice that we have obtained a di↵erent matrix for the same linear function. The
equation we started with
20 1 0 13 0 1
✓ ◆ 2 1 0 a 1
d
+ 2 f = x + 1 , 4@0 2 2A @ bA5 = @1A
dx
0 0 2 c B0
0 B0

2a + b = 1
, 2b + 2c = 1
2c = 0
011
4
B C
has the solution @ 12 A. Notice that we have obtained a di↵erent 3-vector for the
0
same vector, since in the notational convention B 0 this 3-vector represents 14 + 12 x.

One linear function can be represented (denoted) by a huge variety of

matrices. The representation only depends on how vectors are denoted as
n-vectors.

29
30 What is Linear Algebra?

1.5 Review Problems

You probably have already noticed that understanding sets, functions and
basic logical operations is a must to do well in linear algebra. Brush up on
these skills by trying these background webwork problems:
Logic 1
Sets 2
Functions 3
Equivalence Relations 4
Proofs 5
Each chapter also has reading and skills WeBWorK problems:
Webwork: Reading problems 1 ,2

Probably you will spend most of your time on the following review questions:
1. Problems A, B, and C of example 3 can all be written as Lv = w where
L:V !W,
(read this as L maps the set of vectors V to the set of vectors W ). For
each case write down the sets V and W where the vectors v and w
come from.
2. Torque is a measure of “rotational force”. It is a vector whose direction
is the (preferred) axis of rotation. Upon applying a force F on an object
at point r the torque ⌧ is the cross product r ⇥ F = ⌧ :

30
1.5 Review Problems 31

Remember that the cross product of two 3-vectors is given by

0 1 0 01 0 0 1
x x yz zy 0
@ y A ⇥ @ y 0 A := @zx0 xz 0 A .
z z0 xy 0 yx0
Indeed, 3-vectors are special, usually vectors an only be added, not
multiplied.
Lets find the force
0 1 F (a vector) one must apply
0 to1 a wrench lying along
1 0
the vector r = @1A ft, to produce a torque @0Aft lb:
0 1
0 1
a
(a) Find a solution by writing out this equation with F = @ b A.
c
(Hint: Guess and check that a solution with a = 0 exists).
0 1
1
(b) Add 1A to your solution and check that the result is a solution.
@
0
(c) Give a physics explanation of why there can be two solutions, and
argue that there are, in fact, infinitely many solutions.
(d) Set up a system of three linear equations with the three compo-
nents of F as the variables which describes this situation. What
happens if you try to solve these equations by substitution?

3. The function P (t) gives gas prices (in units of dollars per gallon) as a
function of t the year (in A.D. or C.E.), and g(t) is the gas consumption
rate measured in gallons per year by a driver as a function of their age.
The function g is certainly di↵erent for di↵erent people. Assuming a
lifetime is 100 years, what function gives the total amount spent on gas
during the lifetime of an individual born in an arbitrary year t? Is the
operator that maps g to this function linear?

4. The di↵erential equation (DE)

d
f = 2f
dt
31
32 What is Linear Algebra?

says that the rate of change of f is proportional to f . It describes

exponential growth because the exponential function

f (t) = f (0)e2t

satisfies the DE for any number f (0). The number 2 in the DE is called
the constant of proportionality. A similar DE
d 2
f= f
dt t
has a time-dependent “constant of proportionality”.

(a) Do you think that the second DE describes exponential growth?

(b) Write both DEs in the form Df = 0 with D a linear operator.

5. Pablo is a nutritionist who knows that oranges always have twice as

much sugar as apples. When considering the sugar intake of schoolchil-
dren eating a barrel of fruit, he represents the barrel like so:
fruit

(s, f )

sugar

Find a linear operator relating Pablo’s representation to the “everyday”

representation in terms of the number of apples and number of oranges.
Write your answer as a matrix.
Hint: Let represent the amount of sugar in each apple.

Hint

32
1.5 Review Problems 33

6. Matrix Multiplication: Let M and N be matrices

✓ ◆ ✓ ◆
a b e f
M= and N = ,
c d g h
and v the vector ✓ ◆
x
v= .
y
If we first apply N and then M to v we obtain the vector M N v.

(a) Show that the composition of matrices M N is also a linear oper-

ator.

(b) Write out the components of the matrix product M N in terms of

the components of M and the components of N . Hint: use the
general rule for multiplying a 2-vector by a 2⇥2 matrix.
(c) Try to answer the following common question, “Is there any sense
in which these rules for matrix multiplication are unavoidable, or
are they just a notation that could be replaced by some other
notation?”
(d) Generalize your multiplication rule to 3 ⇥ 3 matrices.
7. Diagonal matrices: A matrix M can be thought of as an array of num-
bers mij , known as matrix entries, or matrix components, where i and j
index row and column numbers, respectively. Let
✓ ◆
1 2
M= = mij .
3 4
Compute m11 , m12 , m21 and m22 .
The matrix entries mii whose row and column numbers are the same
are called the diagonal of M . Matrix entries mij with i 6= j are called
o↵-diagonal. How many diagonal entries does an n ⇥ n matrix have?
How many o↵-diagonal entries does an n ⇥ n matrix have?
If all the o↵-diagonal entries of a matrix vanish, we say that the matrix
is diagonal. Let
✓ ◆ ✓ 0 ◆
0 0 0
D= and D = .
0 µ 0 µ0

33
34 What is Linear Algebra?

Are these matrices diagonal and why? Use the rule you found in prob-
lem 6 to compute the matrix products DD0 and D0 D. What do you
observe? Do you think the same property holds for arbitrary matrices?
What about products where only one of the matrices is diagonal?

(p.s. Diagonal matrices play a special role in in the study of matrices

in linear algebra. Keep an eye out for this special role.)
8. Find the linear operator that takes in vectors from n-space and gives
out vectors from n-space in such a way that
(a) whatever you put in, you get exactly the same thing out as what
you put in. Show that it is unique. Can you write this operator
as a matrix?
(b) whatever you put in, you get exactly the same thing out as when
you put something else in. Show that it is unique. Can you write
this operator as a matrix?
Hint: To show something is unique, it is usually best to begin by pre-
tending that it isn’t, and then showing that this leads to a nonsensical
conclusion. In mathspeak–proof by contradiction.
9. Consider the set S = {⇤, ?, #}. It contains just 3 elements, and has
no ordering; {⇤, ?, #} = {#, ?, ⇤} etc. (In fact the same is true for
{1, 2, 3} = {2, 3, 1} etc, although we could make this an ordered set
using 3 > 2 > 1.)
(i) Invent a function with domain {⇤, ?, #} and codomain R. (Re-
member that the domain of a function is the set of all its allowed
inputs and the codomain (or target space) is the set where the
outputs can live. A function is specified by assigning exactly one
codomain element to each element of the domain.)

(ii) Choose an ordering on {⇤, ?, #}, and then use it to write your
function from part (i) as a triple of numbers.

(iii) Choose a new ordering on {⇤, ?, #} and then write your function
from part (i) as a triple of numbers.

34
1.5 Review Problems 35

(iv) Your answers for parts (ii) and (iii) are di↵erent yet represent the
same function – explain!

35
36 What is Linear Algebra?

36
 Systems of Linear Equations
2
2.1 Gaussian Elimination
Systems of linear equations can be written as matrix equations. Now you
will learn an efficient algorithm for (maximally) simplifying a system of linear
equations (or a matrix equation) – Gaussian elimination.

2.1.1 Augmented Matrix Notation

Efficiency demands a new notation, called an augmented matrix , which we
introduce via examples:
The linear system ⇢
x + y = 27
2x y = 0,
is denoted by the augmented matrix
✓ ◆
1 1 27
.
2 1 0
This notation is simpler than the matrix one,
✓ ◆✓ ◆ ✓ ◆
1 1 x 27
= ,
2 1 y 0

although all three of the above denote the same thing.

37
38 Systems of Linear Equations

Augmented Matrix Notation

Another interesting rewriting is
✓ ◆ ✓ ◆ ✓ ◆
1 1 27
x +y = .
2 1 0
✓ ◆
1
This tells us that we are trying to find the combination of the vectors and
✓ ◆ ✓ ◆ ✓ ◆ ✓ ◆ 2
1 27 1 1
adds up to ; the answer is “clearly” 9 + 18 .
1 0 2 1

Here is a larger example. The system

1x + 3y + 2z + 0w = 9
6x + 2y + 0z 2w = 0
1x + 0y + 1z + 1w = 3 ,
is denoted by the augmented matrix
0 1
1 3 2 0 9
@ 6 2 0 2 0A ,
1 0 1 1 3
which is equivalent to the matrix equation
0 1
0 1 x 0 1
1 3 2 0 B C 9
@ 6 2 0 2 A B y C = @0 A .
@ zA
1 0 1 1 3
w
Again, we are trying to find which combination of the columns of the matrix
adds up to the vector on the right hand side.
For the the general case of r linear equations in k unknowns, the number
of equations is the number of rows r in the augmented matrix, and the
number of columns k in the matrix left of the vertical line is the number of
unknowns, giving an augmented matrix of the form
0 1 1 1
a1 a2 · · · a1k b1
B 2 2 C
Ba1 a2 · · · a2k b2 C
B C
B .. .. .. ..C .
@ . . . .A
a1 a2 · · · ak b r
r r r

38
2.1 Gaussian Elimination 39

Entries left of the divide carry two indices; subscripts denote column number
and superscripts row number. We emphasize, the superscripts here do not
denote exponents. Make sure you can write out the system of equations and
the associated matrix equation for any augmented matrix.

Reading homework: problem 1

We now have three ways of writing the same question. Let’s put them
side by side as we solve the system by strategically adding and subtracting
equations. We will not tell you the motivation for this particular series of
steps yet, but let you develop some intuition first.
Example 11 (How matrix equations and augmented matrices change in elimination)
✓ ◆✓ ◆ ✓ ◆ ✓ ◆
x + y = 27 1 1 x 27 1 1 27
, = , .
2x y = 0 2 1 y 0 2 1 0
With the first equation replaced by the sum of the two equations this becomes
✓ ◆✓ ◆ ✓ ◆ ✓ ◆
3x + 0 = 27 3 0 x 27 3 0 27
, = , .
2x y = 0 2 1 y 0 2 1 0
Let the new first equation be the old first equation divided by 3:
✓ ◆✓ ◆ ✓ ◆ ✓ ◆
x + 0 = 9 1 0 x 9 1 0 9
, = , .
2x y = 0 2 1 y 0 2 1 0
Replace the second equation by the second equation minus two times the first equation:
✓ ◆✓ ◆ ✓ ◆ ✓ ◆
x + 0 = 9 1 0 x 9 1 0 9
, = , .
0 y = 18 0 1 y 18 0 1 18
Let the new second equation be the old second equation divided by -1:
✓ ◆✓ ◆ ✓ ◆ ✓ ◆
x + 0 = 9 1 0 x 9 1 0 9
, = , .
0 + y = 18 0 1 y 18 0 1 18

Did you see what the strategy was? To eliminate y from the first equation
and then eliminate x from the second. The result was the solution to the
system.
Here is the big idea: Everywhere in the instructions above we can replace
the word “equation” with the word “row” and interpret them as telling us
what to do with the augmented matrix instead of the system of equations.
Performed systemically, the result is the Gaussian elimination algorithm.

39
40 Systems of Linear Equations

2.1.2 Equivalence and the Act of Solving

We now introduce the symbol ⇠ which is called “tilde” but should be read as
“is (row) equivalent to” because at each step the augmented matrix changes
by an operation on its rows but its solutions do not. For example, we found
above that
✓ ◆ ✓ ◆ ✓ ◆
1 1 27 1 0 9 1 0 9
⇠ ⇠ .
2 1 0 2 1 0 0 1 18

The last of these augmented matrices is our favorite!

Equivalence Example

Setting up a string of equivalences like this is a means of solving a system

of linear equations. This is the main idea of section 2.1.3. This next example
hints at the main trick:

Example 12 (Using Gaussian elimination to solve a system of linear equations)

✓ ◆ ✓ ◆ ✓ ◆ ⇢
x+ y = 5 1 1 5 1 1 5 1 0 2 x+0 = 2
, ⇠ ⇠ ,
x + 2y = 8 1 2 8 0 1 3 0 1 3 0+y = 3

Note that in going from the first to second augmented matrix, we used the top left 1
to make the bottom left entry zero. For this reason we call the top left entry a pivot.
Similarly, to get from the second to third augmented matrix, the bottom right entry
(before the divide) was used to make the top right one vanish; so the bottom right
entry is also called a pivot.

This name pivot is used to indicate the matrix entry used to “zero out”
the other entries in its column; the pivot is the number used to eliminate
another number in its column.

2.1.3 Reduced Row Echelon Form

For a system of two linear equations, the goal of Gaussian elimination is to
convert the part of the augmented matrix left of the dividing line into the
matrix ✓ ◆
1 0
I= ,
0 1

40
2.1 Gaussian Elimination 41

called the Identity Matrix , since this would give the simple statement of a
solution x = a, y = b. The same goes for larger systems of equations for
which the identity matrix I has 1’s along its diagonal and all o↵-diagonal
entries vanish:
0 1
1 0 ··· 0
B 0 1 0 C
B C
I=B .. .. .. C
@ . . . A
0 0 ··· 1

Reading homework: problem 2

For many systems, it is not possible to reach the identity in the augmented
matrix via Gaussian elimination. In any case, a certain version of the matrix
that has the maximum number of components eliminated is said to be the
Row Reduced Echelon Form (RREF).

Example 13 (Redundant equations)

) ! ! (
x + y = 2 1 1 2 1 1 2 x + y = 2
, ⇠ ,
2x + 2y = 4 2 2 4 0 0 0 0 + 0 = 0

This example demonstrates if one equation is a multiple of the other the identity
matrix can not be a reached. This is because the first step in elimination will make
the second row a row of zeros. Notice that solutions still exists (1, 1) is a solution.
The last augmented matrix here is in RREF; no more than two components can be
eliminated.

Example 14 (Inconsistent equations)

) ! ! (
x + y = 2 1 1 2 1 1 2 x + y = 2
, ⇠ ,
2x + 2y = 5 2 2 5 0 0 1 0 + 0 = 1

This system of equation has a solution if there exists two numbers x, and y such that
0 + 0 = 1. That is a tricky way of saying there are no solutions. The last form of the
augmented matrix here is the RREF.

41
42 Systems of Linear Equations

Example 15 (Silly order of equations)

A robot might make this mistake:
) !
0x + y = 2 0 1 2
, ⇠ ··· ,
x + y = 7 1 1 7

and then give up because the the upper left slot can not function as a pivot since the 0
that lives there can not be used to eliminate the zero below it. Of course, the right
thing to do is to change the order of the equations before starting
) ! ! (
x + y = 7 1 1 7 1 0 9 x + 0 = 9
, ⇠ ,
0x + y = 2 0 1 2 0 1 2 0 + y = 2.

The third augmented matrix above is the RREF of the first and second. That is to
say, you can swap rows on your way to RREF.

For larger systems of equations redundancy and inconsistency are the ob-
structions to obtaining the identity matrix, and hence to a simple statement
of a solution in the form x = a, y = b, . . . . What can we do to maximally
simplify a system of equations in general? We need to perform operations
that simplify our system without changing its solutions. Because, exchanging
the order of equations, multiplying one equation by a non-zero constant or
adding equations does not change the system’s solutions, we are lead to three
operations:

• (Row Swap) Exchange any two rows.

• (Scalar Multiplication) Multiply any row by a non-zero constant.

• (Row Addition) Add one row to another row.

These are called Elementary Row Operations, or EROs for short, and are
studied in detail in section 2.3. Suppose now we have a general augmented
matrix for which the first entry in the first row does not vanish. Then, using
just the three EROs, we could1 then perform the following.

1
This is a “brute force” algorithm; there will often be more efficient ways to get to
RREF.

42
2.1 Gaussian Elimination 43

Algorithm For Obtaining RREF:

• Make the leftmost nonzero entry in the top row 1 by multiplication.
• Then use that 1 as a pivot to eliminate everything below it.
• Then go to the next row and make the leftmost nonzero entry 1.
• Use that 1 as a pivot to eliminate everything below and above it!
• Go to the next row and make the leftmost nonzero entry 1... etc
In the case that the first entry of the first row is zero, we may first interchange
the first row with another row whose first entry is non-vanishing and then
perform the above algorithm. If the entire first column vanishes, we may still
apply the algorithm on the remaining columns.
Here is a video (with special e↵ects!) of a hand performing the algorithm
by hand. When it is done, you should try doing what it does.

Beginner Elimination
This algorithm and its variations is known as Gaussian elimination. The
endpoint of the algorithm is an augmented matrix of the form
0 1
1 ⇤ 0 ⇤ 0 · · · 0 ⇤ b1
B 0 0 1 ⇤ 0 · · · 0 ⇤ b2 C
B C
B 0 0 0 0 1 · · · 0 ⇤ b3 C
B C
B . . . . . . C
. .
B . . . . .
. .
. .
. C
B C
B C.
B 0 0 0 0 0 ··· 1 ⇤ b C k
B C
B 0 0 0 0 0 · · · 0 0 bk+1 C
B C
B . . . . . . . . C
@ .. .. .. .. .. .. .. .. A
0 0 0 0 0 · · · 0 0 br
This is called Reduced Row Echelon Form (RREF). The asterisks denote
the possibility of arbitrary numbers (e.g., the second 1 in the top line of
example 13).
Learning to perform this algorithm by hand is the first step to learning
linear algebra; it will be the primary means of computation for this course.
You need to learn it well. So start practicing as soon as you can, and practice
often.

43
44 Systems of Linear Equations

The following properties define RREF:

1. In every row the left most non-zero entry is 1 (and is called a pivot).
2. The pivot of any given row is always to the right of the pivot of the
row above it.
3. The pivot is the only non-zero entry in its column.
Example 16 (Augmented matrix in RREF)
0 1
1 0 7 0
B0 1 3 0C
B C
@0 0 0 1A
0 0 0 0

Example 17 (Augmented matrix NOT in RREF)

0 1
1 0 3 0
B0 0 2 0C
B C
@0 1 0 1A
0 0 0 1
Actually, this NON-example breaks all three of the rules!

The reason we need the asterisks in the general form of RREF is that
not every column need have a pivot, as demonstrated in examples 13 and 16.
Here is an example where multiple columns have no pivot:
Example 18 (Consecutive columns with no pivot in RREF)
✓ ◆ ✓ ◆
x + y + z + 0w = 2 1 1 1 0 2 1 1 1 0 2
, ⇠
2x + 2y + 2z + 2w = 4 2 2 2 1 4 0 0 0 1 0
⇢
x + y + z = 2
,
w = 0.
Note that there was no hope of reaching the identity matrix, because of the shape of
the augmented matrix we started with.

With some practice, elimination can go quickly. Here is an expert showing

you some tricks. If you can’t follow him now then come back when you have
some more experience and watch again. You are going to need to get really
good at this!

44
2.1 Gaussian Elimination 45

Advanced Elimination
It is important that you are able to convert RREF back into a system
of equations. The first thing you might notice is that if any of the numbers
bk+1 , . . . , br in 2.1.3 are non-zero then the system of equations is inconsistent
and has no solutions. Our next task is to extract all possible solutions from
an RREF augmented matrix.

2.1.4 Solution Sets and RREF

RREF is a maximally simplified version of the original system of equations
in the following sense:
• As many coefficients of the variables as possible are 0.
• As many coefficients of the variables as possible are 1.
It is easier to read o↵ solutions from the maximally simplified equations than
from the original equations, even when there are infinitely many solutions.
Example 19 (Standard approach from a system of equations to the solution set)
9 0 1 0 1
x + y + 5w = 1 = 1 1 0 5 1 1 0 0 3 5
y + 2w = 6 , @0 1 0 2 6A ⇠ @0 1 0 2 6A
;
z + 4w = 8 0 0 1 4 8 0 0 1 4 8
8 9
8 9 >
> x = 5 3w >
>
< x + 3w = 5 = >
< y >
=
= 6 2w
, y + 2w = 6 ,
: ; >
> z = 8 4w >
>
z + 4w = 8 >
: >
;
w = w
0 1 0 1 0 1
x 5 3
B yC B 6C B 2C
B C B C B C
,B C=B C+wB C.
@ zA @ 8A @ 4A
w 0 1
There is one solution for each value of w, so the solution set is
80 1 0 1 9
>
> 5 3 >
>
>
<B 6C B 2C >
=
B C B C
B C + ↵ B C : ↵ 2 R .
>
> @ 8A @ 4A >
>
>
: >
;
0 1

45
46 Systems of Linear Equations

Here is a verbal description of the preceeding example of the standard ap-

proach. We say that x, y, and z are pivot variables because they appeared
with a pivot coefficient in RREF. Since w never appears with a pivot co-
efficient, it is not a pivot variable. In the second line we put all the pivot
variables on one side and all the non-pivot variables on the other side and
added the trivial equation w = w to obtain a system that allowed us to easily
read o↵ solutions.

The Standard Approach To Solution Sets

1. Write the augmented matrix.

2. Perform EROs to reach RREF.

3. Express the pivot variables in terms of the non-pivot variables.

There are always exactly enough non-pivot variables to index your solutions.
In any approach, the variables which are not expressed in terms of the other
variables are called free variables. The standard approach is to use the non-
pivot variables as free variables.
Non-standard approach: solve for w in terms of z and substitute into the
other equations. You now have an expression for each component in terms
of z. But why pick z instead of y or x? (or x + y?) The standard approach
not only feels natural, but is canonical, meaning that everyone will get the
same RREF and hence choose the same variables to be free. However, it is
important to remember that so long as their set of solutions is the same, any
two choices of free variables is fine. (You might think of this as the di↵erence
between using Google MapsTM or MapquestTM ; although their maps may
look di↵erent, the place hhome sici they are describing is the same!)
When you see an RREF augmented matrix with two columns that have
no pivot, you know there will be two free variables.

Example 20 (Standard approach, multiple free variables)

46
2.1 Gaussian Elimination 47

0 1
1 00 47 ⇢
B0 14 1C
3
B C, x + 7z =4
@0 00 00A y + 3z+4w = 1
0 00 00
8 9
0 1 0 1 0 1 0 1
>
> x = 4 7z >
> x 4 7 0
< =
B y C B1C B 3C B 4C
y = 1 3z 4w
, ,B C B C B C B C
@ z A = @0A + z @ 1A + w @ 0A
>
> z = z >
>
: ;
w = w w 0 0 1

so the solution set is

80 1 0 1 0 1 9
>
> 4 7 0 >
>
<B C B =
B1C + z B 3C B
C+wB 4C
C : z, w 2 R .
> @ 0A @ 1 A @ 0A >
>
: >
;
0 0 1

From RREF to a Solution Set

You can imagine having three, four, or fifty-six non-pivot columns and
the same number of free variables indexing your solutions set. In general a
solution set to a system of equations with n free variables will be of the form

1 + µ2 x2 + · · · + µn xn : µ1 , . . . , µn 2 R}.
{xP + µ1 xH H H

The parts of these solutions play special roles in the associated matrix
equation. This will come up again and again long after we complete this
discussion of basic calculation methods, so we will use the general language
of linear algebra to give names to these parts now.

Definition: A homogeneous solution to a linear equation Lx = v, with

L and v known is a vector xH such that LxH = 0 where 0 is the zero vector.

If you have a particular solution xP to a linear equation and add

a sum of multiples of homogeneous solutions to it you obtain another
particular solution.

47
48 Systems of Linear Equations

Particular and Homogeneous Solutions

Check now that the parts of the solutions with free variables as coefficients
from the previous examples are homogeneous solutions, and that by adding
a homogeneous solution to a particular solution one obtains a solution to the
matrix equation. This will come up over and over again. As an example
d2
without matrices, consider the di↵erential equation dx 2 f = 3. A particular

solution is 32 x2 while x and 1 are homogeneous solutions. The solution set is

{ 32 x2 + ax + c1 : a, b 2 R}. You can imagine similar di↵erential equations
with more homogeneous solutions.

You need to become very adept at reading o↵ solutions sets of linear

systems from the RREF of their augmented matrix; it is a basic skill for
linear algebra, and we will continue using it up to the last page of the book!

Worked examples of Gaussian elimination

2.2 Review Problems

Reading problems 1 ,2
Augmented matrix 6
Webwork: 2 ⇥ 2 systems 7, 8, 9, 10, 11, 12
3 ⇥ 2 systems 13, 14
3 ⇥ 3 systems 15, 16, 17
1. State whether the following augmented matrices are in RREF and com-
pute their solution sets.
0 1
1 0 0 0 3 1
B0 1 0 0 1 2 C
B C
@0 0 1 0 1 3 A ,
0 0 0 1 2 0
0 1
1 1 0 1 0 1 0
B0 0 1 2 0 2 0C
B C,
@0 0 0 0 1 3 0A
0 0 0 0 0 0 0

48
2.2 Review Problems 49

0 1
1 1 0 1 0 1 0 1
B0 0 1 2 0 2 0 1C
B C
B0 0 0 0 1 3 0 1C
B C.
@0 0 0 0 0 2 0 2A
0 0 0 0 0 0 1 1
2. Solve the following linear system:
2x1 + 5x2 8x3 + 2x4 + 2x5 = 0
6x1 + 2x2 10x3 + 6x4 + 8x5 = 6
3x1 + 6x2 + 2x3 + 3x4 + 5x5 = 6
3x1 + 1x2 5x3 + 3x4 + 4x5 = 3
6x1 + 7x2 3x3 + 6x4 + 9x5 = 9

Be sure to set your work out carefully with equivalence signs ⇠ between
each step, labeled by the row operations you performed.
3. Check that the following two matrices are row-equivalent:
✓ ◆ ✓ ◆
1 4 7 10 0 1 8 20
and .
2 9 6 0 4 18 12 0
Now remove the third column from each matrix, and show that the
resulting two matrices (shown below) are row-equivalent:
✓ ◆ ✓ ◆
1 4 10 0 1 20
and .
2 9 0 4 18 0
Now remove the fourth column from each of the original two matri-
ces, and show that the resulting two matrices, viewed as augmented
matrices (shown below) are row-equivalent:
✓ ◆ ✓ ◆
1 4 7 0 1 8
and .
2 9 6 4 18 12
Explain why row-equivalence is never a↵ected by removing columns.
4. Check that the system of equations corresponding to the augmented
matrix 0 1
1 4 10
@3 13 9A
4 17 20

49
50 Systems of Linear Equations

has no solutions. If you remove one of the rows of this matrix, does
the new matrix have any solutions? In general, can row equivalence be
a↵ected by removing rows? Explain why or why not.

5. Explain why the linear system has no solutions:

0 1
1 0 3 1
@0 1 2 4 A
0 0 0 6

For which values of k does the system below have a solution?

x 3y = 6
x + 3z = 3
2x + ky + (3 k)z = 1

Hint

6. Show that the RREF of a matrix is unique. (Hint: Consider what

happens if the same augmented matrix had two di↵erent RREFs. Try
to see what happens if you removed columns from these two RREF
augmented matrices.)

7. Another method for solving linear systems is to use row operations to

bring the augmented matrix to Row Echelon Form (REF as opposed to
RREF). In REF, the pivots are not necessarily set to one, and we only
require that all entries left of the pivots are zero, not necessarily entries
above a pivot. Provide a counterexample to show that row echelon form
is not unique.
Once a system is in row echelon form, it can be solved by “back substi-
tution.” Write the following row echelon matrix as a system of equa-
tions, then solve the system using back-substitution.
0 1
2 3 1 6
@0 1 1 2 A
0 0 3 3

50
2.2 Review Problems 51

8. Show that this pair of augmented matrices are row equivalent, assuming
ad bc 6= 0: ! !
a b e 1 0 de bf
ad bc
⇠
c d f 0 1 af ce
ad bc

9. Consider the augmented matrix:

✓ ◆
2 1 3
.
6 3 1

Give a geometric reason why the associated system of equations has

no solution. (Hint, plot the three vectors given by the columns of this
augmented matrix in the plane.) Given a general augmented matrix
✓ ◆
a b e
,
c d f
can you find a condition on the numbers a, b, c and d that corresponds
to the geometric condition you found?
10. A relation ⇠ on a set of objects U is an equivalence relation if the
following three properties are satisfied:
• Reflexive: For any x 2 U , we have x ⇠ x.
• Symmetric: For any x, y 2 U , if x ⇠ y then y ⇠ x.
• Transitive: For any x, y and z 2 U , if x ⇠ y and y ⇠ z then x ⇠ z.
Show that row equivalence of matrices is an example of an equivalence
relation.
(For a discussion of equivalence relations, see Homework 0, Problem 4)

Hint

11. Equivalence of augmented matrices does not come from equality of their
solution sets. Rather, we define two matrices to be equivalent if one
can be obtained from the other by elementary row operations.
Find a pair of augmented matrices that are not row equivalent but do
have the same solution set.

51
52 Systems of Linear Equations

2.3 Elementary Row Operations

Elementary row operations are systems of linear equations relating the old
and new rows in Gaussian elimination:

Example 21 (Keeping track of EROs with equations between rows)

We refer to the new kth row as Rk0 and the old kth row as Rk .

R10 =0R1 + R2 +0R3

0 1 R20 = R1 +0R2 +0R3 0 1 0 01 0 10 1
0 1 1 7 R30 =0R1 +0R2 + R3 2 0 0 4 R1 0 1 0 R1
@2 0 0 4A ⇠ @0 1 1 7A @R20 A = @1 0 0A @R2 A
0 0 1 4 0 0 1 4 R30 0 0 1 R3
R10 = 12 R1 +0R2 +0R3
R20 =0R1 + R2 +0R3 0 1 0 0 1 01 10 1
R30 =0R1 +0R2 + R3 1 0 0 2 R1 2 0 0 R1
⇠ @0 1 1 7A @R20 A = @ 0 1 0A @R2 A
0 0 1 4 R30 0 0 1 R3
R10 = R1 +0R2 +0R3
R20 =0R1 + R2 R3 0 1 0 01 0 10 1
R30 =0R1 +0R2 + R3 1 0 0 2 R1 1 0 0 R1
⇠ @0 1 0 3A @R20 A = @0 1 1 A @ R2 A
0 0 1 4 R30 0 0 1 R3

On the right, we have listed the relations between old and new rows in matrix notation.

Reading homework: problem 3

2.3.1 EROs and Matrices

Interestingly, the matrix that describes the relationship between old and new
rows performs the corresponding ERO on the augmented matrix.

52
2.3 Elementary Row Operations 53

Example 22 (Performing EROs with Matrices)

0 10 1 0 1
0 1 0 0 1 1 7 2 0 0 4
@1 0 0A @2 0 0 4A = @0 1 1 7A
0 0 1 0 0 1 4 0 0 1 4

⇠
01 10 1 0 1
0 0
2 2 0 0 4 1 0 0 2
@ 0 1 0A @0 1 1 7A = @0 1 1 7A
0 0 1 0 0 1 4 0 0 1 4

⇠
0 10 1 0 1
1 0 0 1 0 0 2 1 0 0 2
@0 1 1A @0 1 1 7A = @0 1 0 3A
0 0 1 0 0 1 4 0 0 1 4

Here we have multiplied the augmented matrix with the matrices that acted on rows
listed on the right of example 21.

Realizing EROs as matrices allows us to give a concrete notion of “di-

viding by a matrix”; we can now perform manipulations on both sides of an
equation in a familiar way:

Example 23 (Undoing A in Ax = b slowly, for A = 6 = 3 · 2)

6x = 12
, 3 1 6x = 3 1 12

, 2x = 4
, 2 1 2x = 2 1 4
, 1x = 2

The matrices corresponding to EROs undo a matrix step by step.

53
54 Systems of Linear Equations

Example 24 (Undoing A in Ax = b slowly, for A = M = ...)

0 10 1 0 1
0 1 1 x 7
@2 0 0 A @ yA = @4A
0 0 1 z 4
0 10 10 1 0 1 0 1
0 1 0 0 1 1 x 0 1 0 7
, @1 0 0A @2 0 0A @ y A = @1 0 0A @4A
0 0 1 0 0 1 z 0 0 1 4
0 10 1 0 1
2 0 0 x 4
, @0 1 1A @ y A = @7A
0 0 1 z 4
01 10 10 1 01 1 0 1
2 0 0 2 0 0 x 2 0 0 4
, @ 0 1 0 A @ 0 1 1 A @ y A = @ 0 1 0 A @ 7A
0 0 1 0 0 1 z 0 0 1 4
0 10 1 0 1
1 0 0 x 2
, @ 0 1 1 A @ yA = @7A
0 0 1 z 4
0 10 10 1 0 10 1
1 0 0 1 0 0 x 1 0 0 2
, @0 1 1A @0 1 1A @ y A = @0 1 1A@7A
0 0 1 0 0 1 z 0 0 1 4
0 10 1 0 1
1 0 0 x 2
, @0 1 0A @ y A = @3A .
0 0 1 z 4

This is another way of thinking about Gaussian elimination which feels more
like elementary algebra in the sense that you “do something to both sides of
an equation” until you have a solution.

2.3.2 Recording EROs in (M |I )

Just as we put together 3 1 2 1 = 6 1 to get a single thing to apply to both
sides of 6x = 12 to undo 6, we should put together multiple EROs to get
a single thing that undoes our matrix. To do this, augment by the identity
matrix (not just a single column) and then perform Gaussian elimination.
There is no need to write the EROs as systems of equations or as matrices
while doing this.

54
2.3 Elementary Row Operations 55

Example 25 (Collecting EROs that undo a matrix)

0 1 0 1
0 1 1 1 0 0 2 0 0 0 1 0
@ 2 0 0 0 1 0 A ⇠ @ 0 1 1 1 0 0 A
0 0 1 0 0 1 0 0 1 0 0 1
0 1 0 1
1 0 0 0 12 0 1 0 0 0 12 0
⇠ @ 0 1 1 1 0 0 A ⇠ @ 0 1 0 1 0 1 A.
0 0 1 0 0 1 0 0 1 0 0 1

As we changed the left side from the matrix M to the identity matrix, the
right side changed from the identity matrix to the matrix which undoes M .
Example 26 (Checking that one matrix undoes another)
0 10 1 0 1
0 12 0 0 1 1 1 0 0
@ 1 0 1 A@ 2 0 0 A=@ 0 1 0 A.
0 0 1 0 0 1 0 0 1

If the matrices are composed in the opposite order, the result is the same.
0 10 1 0 1
0 1 1 0 12 0 1 0 0
@ 2 0 0 A@ 1 0 1 A = @ 0 1 0 A.
0 0 1 0 0 1 0 0 1

Whenever the product of two matrices M N = I, we say that N is the

inverse of M or N = M 1 . Conversely M is the inverse of N ; M = N 1 .

In abstract generality, let M be some matrix and, as always, let I stand

for the identity matrix. Imagine the process of performing elementary row
operations to bring M to the identity matrix:

(M |I) ⇠ (E1 M |E1 ) ⇠ (E2 E1 M |E2 E1 ) ⇠ · · · ⇠ (I| · · · E2 E1 ) .

The ellipses “· · · ” stand for additional EROs. The result is a product of

matrices that form a matrix which undoes M

· · · E2 E1 M = I .

This is only true if the RREF of M is the identity matrix.

Definition: A matrix M is invertible if its RREF is an identity matrix.

55
56 Systems of Linear Equations

1
How to find M
1
(M |I) ⇠ (I|M )

Much use is made of the fact that invertible matrices can be undone with
EROs. To begin with, since each elementary row operation has an inverse,

M = E1 1 E2 1 · · · ,

while the inverse of M is

1
M = · · · E2 E1 .

This is symbolically verified by

1
M M = · · · E2 E1 E1 1 E2 1 · · · = · · · E2 E2 1 · · · = · · · = I .

Thus, if M is invertible, then M can be expressed as the product of EROs.

(The same is true for its inverse.) This has the feel of the fundamental
theorem of arithmetic (integers can be expressed as the product of primes)
or the fundamental theorem of algebra (polynomials can be expressed as the
product of [complex] first order polynomials); EROs are building blocks of
invertible matrices.

2.3.3 The Three Elementary Matrices

We now work toward concrete examples and applications. It is surprisingly
easy to translate between EROs and matrices that perform EROs. The
matrices corresponding to these kinds are close in form to the identity matrix:

• Row Swap: Identity matrix with two rows swapped.

• Scalar Multiplication: Identity matrix with one diagonal entry not 1.

• Row Sum: The identity matrix with one o↵-diagonal entry not 0.

Example 27 (Correspondences between EROs and their matrices)

56
2.3 Elementary Row Operations 57

• The row swap matrix that swaps the 2nd and 4th row is the identity matrix with
the 2nd and 4th row swapped:
0 1
1 0 0 0 0
B0 0 0 1 0C
B C
B0 0 1 0 0C .
B C
@0 1 0 0 0A
0 0 0 0 1

• The scalar multiplication matrix that replaces the 3rd row with 7 times the 3rd
row is the identity matrix with 7 in the 3rd row instead of 1:
0 1
1 0 0 0
B0 1 0 0C
B C
@0 0 7 0A .
0 0 0 1

• The row sum matrix that replaces the 4th row with the 4th row plus 9 times
the 2nd row is the identity matrix with a 9 in the 4th row, 2nd column:
0 1
1 0 0 0 0 0 0
B0 1 0 0 0 0 0C
B C
B0 0 1 0 0 0 0C
B C
B0 9 0 1 0 0 0C .
B C
B0 0 0 0 1 0 0C
B C
@0 0 0 0 0 1 0A
0 0 0 0 0 0 1

We can write an explicit factorization of a matrix into EROs by keeping

track of the EROs used in getting to RREF.
Example 28 (Express M from Example 25 as a product of EROs)
Note that in the previous example one of each of the kinds of EROs is used, in the
order just given. Elimination looked like
0 1 0 1 0 1 0 1
0 1 1 2 0 0 1 0 0 1 0 0
E E E
M = @ 2 0 0 A ⇠1 @ 0 1 1 A ⇠2 @ 0 1 1 A ⇠3 @ 0 1 0 A=I,
0 0 1 0 0 1 0 0 1 0 0 1
where the EROs matrices are
0 1 0 1 1 0 1
0 1 0 2 0 0 1 0 0
E1 = @ 1 A
0 0 , E2 = @ A
0 1 0 , E3 = @ 0 1 1 A.
0 0 1 0 0 1 0 0 1

57
58 Systems of Linear Equations

The inverse of the ERO matrices (corresponding to the description of the reverse row
maniplulations)
0 1 0 1 0 1
0 1 0 2 0 0 1 0 0
E1 1 = @ 1 0 0 A , E2 1 = @ 0 1 0 A , E3 1 = @ 0 1 1 A .
0 0 1 0 0 1 0 0 1

Multiplying these gives

0 10 10 1
0 1 0 2 0 0 1 0 0
E1 1 E2 1 E3 1 = @ 1 0 0 A @ 0 1 0 A@ 0 1 1 A
0 0 1 0 0 1 0 0 1
0 10 1 0 1
0 1 0 2 0 0 0 1 1
= @ 1 0 0 A@ 0 1 1 A=@ 2 0 0 A=M.
0 0 1 0 0 1 0 0 1

2.3.4 LU , LDU , and P LDU Factorizations

The process of elimination can be stopped halfway to obtain decompositions
frequently used in large computations in sciences and engineering. The first
half of the elimination process is to eliminate entries below the diagonal
leaving a matrix which is called upper triangular. The elementary matrices
which perform this part of the elimination are lower triangular, as are their
inverses. But putting together the upper triangular and lower triangular
parts one obtains the so-called LU factorization.

Example 29 (LU factorization)

0 1 0 1
2 0 3 1 2 0 3 1
B 0 1 2 2C B 2C
M =B C E⇠1 B0 1 2 C
@ 4 0 9 2A @0 0 3 4A
0 1 1 1 0 1 1 1
0 1 0 1
2 0 3 1 2 0 3 1
E2 B0 1 2 2C E3 B0 1 2 2C
⇠ B C ⇠ B C := U ,
@0 0 3 4A @0 0 3 4A
0 0 3 1 0 0 0 3

58
2.3 Elementary Row Operations 59

where the EROs and their inverses are

0 1 0 1 0 1
1 0 0 0 1 0 0 0 1 0 0 0
B0 1 0 0C B0 1 0 C
0C B0 1 0 0C
E1 = B@2 0 1 0A ,
C E2 = B@0 , E3 = B C
0 1 0A @0 0 1 0A
0 0 0 1 0 1 0 1 0 0 1 1
0 1 0 1 0 1
1 0 0 0 1 0 0 0 1 0 0 0
B 0 1 0 0C C B C B 0C
E1 1 = B 1 B0 1 0 0C
, E3 1 = B
0 1 0 C.
@ 2 0 1 0A , E2 = @0 0 1 0A @0 0 1 0A
0 0 0 1 0 1 0 1 0 0 1 1

Applying inverse elementary matrices to both sides of the equality U = E3 E2 E1 M

gives M = E1 1 E2 1 E3 1 U or
0 1 0 10 10 10 1
2 0 3 1 1 0 0 0 1 0 0 0 1 0 0 0 2 0 3 1
B 0 1 2 2C B 0 1 0 0CB 0CB 0CB 2C
B C=B CB0 1 0 CB0 1 0 CB0 1 2 C
@ 4 0 9 2A @ 2 0 1 0 @0 0
A 1 0 @0 0
A 1 0 @0
A 0 3 4A
0 1 1 1 0 0 0 1 0 1 0 1 0 0 1 1 0 0 0 3
0 10 10 1
1 0 0 0 1 0 0 0 2 0 3 1
B 0 1 0 0C B0 1 0 0C B0 1 2 2C
=B
@ 2
CB CB C
0 1 0A @0 0 1 0A @0 0 3 4A
0 0 0 1 0 1 1 1 0 0 0 3
0 10 1
1 0 0 0 2 0 3 1
B 0 1 0 0C B0 1 2 2C
=B
@ 2
CB C.
0 1 0A @0 0 3 4A
0 1 1 1 0 0 0 3

This is a lower triangular matrix times an upper triangular matrix.

59
60 Systems of Linear Equations

What if we stop at a di↵erent point in elimination? We could multiply

rows so that the entries in the diagonal are 1 next. Note that the EROs that
do this are diagonal. This gives a slightly di↵erent factorization.

Example 30 (LDU factorization building from previous example)

0 1 0 1 0 3 11
2 0 3 1 2 0 3 1 1 0 2 2
B 0 1 2 2C E3 E2 E1 B0 1 2 2C B0 1 2 2C
M =B C ⇠ B C E⇠4
B C
@ 4 0 9 2A @0 0 3 4A @0 0 3 4A
0 1 1 1 0 0 0 3 0 0 0 3
0 3 11 0 3 11
1 0 2 2 1 0 2 2
B0 1 2 2C B 2C
E5
⇠ B C 6 B0
E
4A ⇠ @
1 2 C
4 A =: U
@0 0 1 0 0 1
3 3
0 0 0 3 0 0 0 1

The corresponding elementary matrices are

01 1 0 1 0 1
20 0 0 1 0 0 0 1 0 0 0
B0 1 0 0C B0 1 0 0C B0 1 0 0C
E4 = B
@0 0
C, E5 = B C, E6 = B C,
1 0A @0 0 1
3 0
A @0 0 1 0A
1
0 0 0 1 0 0 0 1 0 0 0 3
0 1 0 1 0 1
2 0 0 0 1 0 0 0 1 0 0 0
B0 1 0 C
0C B0 1 0 C
0C B0 1 0 0C
E4 1 =B
@0 , E5 1 = B , E6 1 = B C.
0 1 0A @0 0 3 0A @0 0 1 0A
0 0 0 1 0 0 0 1 0 0 0 3

The equation U = E6 E5 E4 E3 E2 E1 M can be rearranged as

M = (E1 1 E2 1 E3 1 )(E4 1 E5 1 E6 1 )U.

We calculated the product of the first three factors in the previous example; it was
named L there, and we will reuse that name here. The product of the next three
factors is diagonal and we wil name it D. The last factor we named U (the name means
something di↵erent in this example than the last example.) The LDU factorization
of our matrix is
0 1 0 10 10 1
2 0 3 1 1 0 0 0 2 0 0 0 1 0 32 12
B 0 1 2 2C B 1 0 0C B 0C B C
B C=B 0 C B0 1 0 C B0 1 2 24 C .
@ 4 0 9 2 A @ 2 0 1 0 A @ 0 0 3 0 A @0 0 1 3A
0 1 1 1 0 1 1 1 0 0 0 3 0 0 0 1

60
2.4 Review Problems 61

The LDU factorization of a matrix is a factorization into blocks of EROs

of a various types: L is the product of the inverses of EROs which eliminate
below the diagonal by row addition, D the product of inverses of EROs which
set the diagonal elements to 1 by row multiplication, and U is the product
of inverses of EROs which eliminate above the diagonal by row addition.
You may notice that one of the three kinds of row operation is missing
from this story. Row exchange may be necessary to obtain RREF. Indeed, so
far in this chapter we have been working under the tacit assumption that M
can be brought to the identity by just row multiplication and row addition.
If row exchange is necessary, the resulting factorization is LDP U where P is
the product of inverses of EROs that perform row exchange.

Example 31 (LDP U factorization, building from previous examples)

0 1 0 1
0 1 2 2 2 0 3 1
B 2 0 3 1C P B 0 1 2 2C
M =B C⇠ B C E6 E5 E⇠
4 E3 E2 E1
L
@ 4 0 9 2 A @ 4 0 9 2 A
0 1 1 1 0 1 1 1

0 1
0 1 0 0
B1 0 0 0C
P =B
@0
C=P 1
0 1 0A
0 0 0 1

M = P (E1 1 E2 1 E3 1 )(E4 1 E5 1 E6 1 )(E7 1 )U = P LDU

0 1 0 10 10 10 3 11
0 1 2 2 1 0 0 0 2 0 0 0 0 1 0 0 1 0 2 2
B 2 0 3 1C B 0 1 0 0CB 0CB 0CB 2C
B C=B CB0 1 0 CB1 0 0 CB0 1 2 C
@ 4 0 9 2 A @ 2 0 1 0 @0
A 0 3 0 @0
A 0 1 0 @0
A 0 1 4A
3
0 1 1 1 0 1 1 1 0 0 1 3 0 0 0 1 0 0 0 1

2.4 Review Problems

Reading problems 3
Webwork: Matrix notation 18
LU 19

61
62 Systems of Linear Equations

1. While performing Gaussian elimination on these augmented matrices

write the full system of equations describing the new rows in terms of
the old rows above each equivalence symbol as in Example 21.
0 1
✓ ◆ 1 1 0 5
2 2 10
, @ 1 1 1 11A
1 2 8
1 1 1 5

2. Solve the vector equation by applying ERO matrices to each side of

the equation to perform elimination. Show each matrix explicitly as in
Example 24.

0 10 1 0 1
3 6 2 x 3
@5 9 4 A @ y A = @ 1 A
2 4 2 z 0

3. Solve this vector equation by finding the inverse of the matrix through
(M |I) ⇠ (I|M 1 ) and then applying M 1 to both sides of the equation.
0 10 1 0 1
2 1 1 x 9
@1 1 1A @ y A = @6A
1 1 2 z 7

4. Follow the method of Examples 29 and 30 to find the LU and LDU

factorization of
0 1
3 3 6
@3 5 2A .
6 2 5

5. Multiple matrix equations with the same matrix can be solved simul-
taneously.

(a) Solve both systems by performing elimination on just one aug-

mented matrix.
0 10 1 0 1 0 10 1 0 1
2 1 1 x 0 2 1 1 a 2
@ 1 1 1 A @ y A = @1 A , @ 1 1 1 A @ b A = @1 A
1 1 0 z 0 1 1 0 c 1

62
2.5 Solution Sets for Systems of Linear Equations 63

(b) Give an interpretation of the columns of M 1 in (M |I) ⇠ (I|M 1

)
in terms of solutions to certain systems of linear equations.

6. How can you convince your fellow students to never make this mistake?
R10 =R1 +R2
0 1 R20 =R1 R2
0 1
1 0 2 3 1 1 4 6
R30 =R1 +2R2
@0 1 2 3 A ⇠ @1 1 0 0A
2 0 1 4 1 2 6 9

7. Is LU factorization of a matrix unique? Justify your answer.

1. If you randomly create a matrix by picking numbers out of the blue,

it will probably be difficult to perform elimination or factorization;
fractions and large numbers will probably be involved. To invent simple
problems it is better to start with a simple answer:

(a) Start with any augmented matrix in RREF. Perform EROs to

make most of the components non-zero. Write the result on a
separate piece of paper and give it to your friend. Ask that friend
to find RREF of the augmented matrix you gave them. Make sure
they get the same augmented matrix you started with.
(b) Create an upper triangular matrix U and a lower triangular ma-
trix L with only 1s on the diagonal. Give the result to a friend to
factor into LU form.
(c) Do the same with an LDU factorization.

2.5 Solution Sets for Systems of Linear Equa-

tions
Algebraic equations problems can have multiple solutions. For example x(x
1) = 0 has two solutions: 0 and 1. By contrast, equations of the form Ax = b
with A a linear operator (with scalars the real numbers) have the following
property:
If A is a linear operator and b is known, then Ax = b has either

1. One solution

63
64 Systems of Linear Equations

2. No solutions

3. Infinitely many solutions

2.5.1 The Geometry of Solution Sets: Hyperplanes

Consider the following algebra problems and their solutions.

1. 6x = 12 has one solution: 2.

2a. 0x = 12 has no solution.

2b. 0x = 0 has infinitely many solutions; its solution set is R.

In each case the linear operator is a 1 ⇥ 1 matrix. In the first case, the linear
operator is invertible. In the other two cases it is not. In the first case, the
solution set is a point on the number line, in case 2b the solution set is the
whole number line.
Lets examine similar situations with larger matrices: 2 ⇥ 2 matrices.
✓ ◆✓ ◆ ✓ ◆ ✓ ◆
6 0 x 12 2
1. = has one solution: .
0 2 y 6 3
✓ ◆✓ ◆ ✓ ◆
1 3 x 4
2a. = has no solutions.
0 0 y 1
✓ ◆✓ ◆ ✓ ◆ ⇢✓ ◆ ✓ ◆
1 3 x 4 4 3
2bi. = has solution set +y :y2R .
0 0 y 0 0 1
✓ ◆✓ ◆ ✓ ◆ ⇢✓ ◆
0 0 x 0 x
2bii. = has solution set : x, y 2 R .
0 0 y 0 y
Again, in the first case the linear operator is invertible while in the other
cases it is not. When a 2 ⇥ 2 matrix from a matrix equation is not invertible
the solution set can be empty, a line in the plane, or the plane itself.
For a system of equations with r equations and k veriables, one can have a
number of di↵erent outcomes. For example, consider the case of r equations
in three variables. Each of these equations is the equation of a plane in three-
dimensional space. To find solutions to the system of equations, we look for
the common intersection of the planes (if an intersection exists). Here we
have five di↵erent possibilities:

64
2.5 Solution Sets for Systems of Linear Equations 65

1. Unique Solution. The planes have a unique point of intersection.

2a. No solutions. Some of the equations are contradictory, so no solutions

exist.

2bi. Line. The planes intersect in a common line; any point on that line
then gives a solution to the system of equations.

2bii. Plane. Perhaps you only had one equation to begin with, or else all
of the equations coincide geometrically. In this case, you have a plane
of solutions, with two free parameters.

Planes

2biii. All of R3 . If you start with no information, then any point in R3 is a

solution. There are three free parameters.

In general, for systems of equations with k unknowns, there are k + 2

possible outcomes, corresponding to the possible numbers (i.e., 0, 1, 2, . . . , k)
of free parameters in the solutions set, plus the possibility of no solutions.
These types of solution sets are hyperplanes, generalizations of planes that
behave like planes in R3 in many ways.

Reading homework: problem 4

Pictures and Explanation

2.5.2 Particular Solution + Homogeneous Solutions

Lets look at solution sets again, this time trying to get to their geometric
shape. In the standard approach, variables corresponding to columns that
do not contain a pivot (after going to reduced row echelon form) are free. It
is the number of free variables that determines the geometry of the solution
set.

65
66 Systems of Linear Equations

Example 32 (Non-pivot variables determine the gemometry of the solution set)

0 1
0 1 x1 0 1 8
1 0 1 1 B C 1 < 1x1 + 0x2 + 1x3 1x4 = 1
x2 C @ A
@0 1 1 1A B
@x3 A = 1 , 0x1 + 1x2 1x3 + 1x4 = 1
:
0 0 0 0 0 0x1 + 0x2 + 0x3 + 0x4 = 0
x4

Following the standard approach, express the pivot variables in terms of the non-pivot
variables and add “empty equations”. Here x3 and x4 are non-pivot variables.
9 0 1 0 1 0 1 0 1
x1 = 1 x3 + x4 > > x1 1 1 1
= B C B C B C B
x2 = 1 + x3 x4 x2 C B1C 1C 1C
,B @ A = @ A + x3 B @ A + x4 B
@
C
x3 = x3 >
> x3 0 1 0A
;
x4 = x4 x4 0 0 1

The preferred way to write a solution set S is with set notation;

80 1 0 1 0 1 0 1 9
>
> x1 1 1 1 >
>
<B C B C B =
x2 C B1C 1C B 1C
S= B @ A = @ A + µ1 B @
C + µ2 B
A @
C : µ 1 , µ2 2 R .
>
> x 0 1 0A >
>
: 3 ;
x4 0 0 1

Notice that the first two components of the second two terms come from the non-pivot
columns. Another way to write the solution set is

S = xP + µ1 xH H
1 + µ 2 x 2 : µ 1 , µ2 2 R ,

where 0 1 0 1 01
1 1 1
B 1C B 1C B 1C
xP = B C
@0A , xH
1 =B C
@ 1A , xH
2 =B C
@ 0A .
0 0 1
Here xP is a particular solution while xH H
1 and x2 are called homogeneous solutions.
The solution set forms a plane.

2.5.3 Solutions and Linearity

Motivated by example 32, we say that the matrix equation M x = v has
1 + µ2 x2 | µ1 , µ2 2 R}. Recall that matrices are linear
solution set {xP + µ1 xH H

operators. Thus

M (xP + µ1 xH H P H H
1 + µ2 x 2 ) = M x + µ1 M x 1 + µ2 M x 2 = v ,

66
2.5 Solution Sets for Systems of Linear Equations 67

for any µ1 , µ2 2 R. Choosing µ1 = µ2 = 0, we obtain

M xP = v .

This is why xP is an example of a particular solution.

Setting µ1 = 1, µ2 = 0, and subtracting M xP = v we obtain

M xH
1 = 0.

Likewise, setting µ1 = 0, µ2 = 1, we obtain

M xH
2 = 0.

Here xH H
1 and x2 are examples of what are called homogeneous solutions to
the system. They do not solve the original equation M x = v, but instead its
associated homogeneous equation M y = 0.
We have just learnt a fundamental lesson of linear algebra: the solution
set to Ax = b, where A is a linear operator, consists of a particular solution
plus homogeneous solutions.

{Solutions} = {Particular solution + Homogeneous solutions}

Example 33 Consider the matrix equation of example 32. It has solution set
80 1 0 1 0 1 9
>
> 1 1 1 >
>
<B C B 1C B 1C =
B 1 C B C B C
S = @ A + µ 1 @ A + µ 2 @ A : µ 1 , µ2 2 R .
>
> 0 1 0 >
>
: ;
0 0 1
0 1
1
B 1C
Then M xP = v says that B C
@0A is a solution to the original matrix equation, which is
0
certainly true, but this is not the only solution.
0 1
1
B C
M xH B 1C
1 = 0 says that @ 1A is a solution to the homogeneous equation.

67
68 Systems of Linear Equations

0 1
1
B 1C
M xH
2 = 0 says that B C
@ 0A is a solution to the homogeneous equation.
1
Notice how adding any multiple of a homogeneous solution to the particular solution
yields another particular solution.

Reading homework: problem 4

2.6 Review Problems

Reading problems 4 ,5
Webwork: Solution sets 20, 21, 22
Geometry of solutions 23, 24, 25, 26

1. Write down examples of augmented matrices corresponding to each

of the five types of solution sets for systems of equations with three
unknowns.

2. Invent a simple linear system that has multiple solutions. Use the stan-
dard approach for solving linear systems and a non-standard approach
to obtain di↵erent descriptions of the solution set. Is the solution set
di↵erent with di↵erent approaches?

3. Let 0 1 0 1 1
a11 a12 · · · a1k x
B 2 2 2C B 2 C
B a1 a2 · · · ak C Bx C
M =B
B .. .. .. C
C and x = B
B ..
C.
C
@. . .A @ . A
r r r
a1 a2 · · · ak xk

Note: x2 does not denote the square of the column vector x. Instead
x1 , x2 , x3 , etc..., denote di↵erent variables (the components of x);
the superscript is an index. Although confusing at first, this nota-
tion was invented by AlbertPEinstein who noticed that quantities like
k
a21 x1 + a22 x2 · · · + a2k xk =: 2 j
j=1 aj x , can be written unambiguously
as a2j xj . This is called Einstein summation notation. The most im-
portant thing to remember is that the index j is a dummy variable,

68
2.6 Review Problems 69

so that a2j xj ⌘ a2i xi ; this is called “relabeling dummy indices”. When

dealing with products of sums, you must P remember to introduce a
i i i i
new dummyP for each term;
P i.e., a i x b i y = i ai x bi y does not equal
ai x i b j y j = i ai x
i j
j bj y .
Use Einstein summation notation to propose a rule for M x so that
M x = 0 is equivalent to the linear system

a11 x1 +a12 x2 · · · +a1k xk = 0

a21 x1 +a22 x2 · · · +a2k xk = 0
.. .. .. ..
. . . .
r 1 r 2 r k
a1 x +a2 x · · · +ak x = 0

Show that your rule for multiplying a matrix by a vector obeys the
linearity property.

4. The standard basis vector ei is a column vector with a one in the ith
row, and zeroes everywhere else. Using the rule for multiplying a matrix
times a vector in problem 3, find a simple rule for multiplying M ei ,
where M is the general matrix defined there.

5. If A is a non-linear operator, can the solutions to Ax = b still be written

as “general solution=particular solution + homogeneous solutions”?
Provide examples.

6. Find a system of equations whose solution set is the walls of a 1 ⇥ 1 ⇥ 1

cube. (Hint: You may need to restrict the ranges of the variables; could
your equations be linear?)

69
70 Systems of Linear Equations

70
 The Simplex Method
3
In Chapter 2, you learned how to handle systems of linear equations. However
there are many situations in which inequalities appear instead of equalities.
In such cases we are often interested in an optimal solution extremizing a
particular quantity of interest. Questions like this are a focus of fields such as
mathematical optimization and operations research. For the case where the
functions involved are linear, these problems go under the title linear pro-
gramming. Originally these ideas were driven by military applications, but
by now are ubiquitous in science and industry. Gigantic computers are dedi-
cated to implementing linear programming methods such as George Dantzig’s
simplex algorithm–the topic of this chapter.

3.1 Pablo’s Problem

Let us begin with an example. Consider again Pablo the nutritionist of
problem 5, chapter 1. The Conundrum City school board has employed
Pablo to design their school lunch program. Unfortunately for Pablo, their
requirements are rather tricky:

Example 34 (Pablo’s problem)

The Conundrum City school board is heavily influenced by the local fruit grower’s
association. They have stipulated that children eat at least 7 oranges and 5 apples
per week. Parents and teachers have agreed that eating at least 15 pieces of fruit per
week is a good thing, but school janitors argue that too much fruit makes a terrible
mess, so that children should eat no more than 25 pieces of fruit per week.

71
72 The Simplex Method

Finally Pablo knows that oranges have twice as much sugar as apples and that apples
have 5 grams of sugar each. Too much sugar is unhealthy, so Pablo wants to keep the
children’s sugar intake as low as possible. How many oranges and apples should Pablo
suggest that the school board put on the menu?

This is a rather gnarly word problem. Our first step is to restate it as

mathematics, stripping away all the extraneous information:

Example 35 (Pablo’s problem restated)

Let x be the number of apples and y be the number of oranges. These must obey

x 5 and y 7,

to fulfill the school board’s politically motivated wishes. The teacher’s and parent’s
fruit requirement means that
x + y 15 ,
but to keep the canteen tidy
x + y  25 .
Now let
s = 5x + 10y .
This linear function of (x, y) represents the grams of sugar in x apples and y oranges.
The problem is asking us to minimize s subject to the four linear inequalities listed
above.

72
3.2 Graphical Solutions 73

3.2 Graphical Solutions

Before giving a more general algorithm for handling this problem and prob-
lems like it, we note that when the number of variables is small (preferably 2),
a graphical technique can be used.
Inequalities, such as the four given in Pablo’s problem, are often called
constraints, and values of the variables that satisfy these constraints comprise
the so-called feasible region. Since there are only two variables, this is easy
to plot:

Example 36 (Constraints and feasible region) Pablo’s constraints are

x 5
y 7

15  x + y  25 .

Plotted in the (x, y) plane, this gives:

You might be able to see the solution to Pablo’s problem already. Oranges
are very sugary, so they should be kept low, thus y = 7. Also, the less fruit
the better, so the answer had better lie on the line x + y = 15. Hence,
the answer must be at the vertex (8, 7). Actually this is a general feature

73
74 The Simplex Method

of linear programming problems, the optimal answer must lie at a vertex of

the feasible region. Rather than prove this, lets look at a plot of the linear
function s(x, y) = 5x + 10y.

Example 37 (The sugar function)

Plotting the sugar function requires three dimensions:

The plot of a linear function of two variables is a plane through the origin.
Restricting the variables to the feasible region gives some lamina in 3-space.
Since the function we want to optimize is linear (and assumedly non-zero), if
we pick a point in the middle of this lamina, we can always increase/decrease
the function by moving out to an edge and, in turn, along that edge to a
corner. Applying this to the above picture, we see that Pablo’s best option
is 110 grams of sugar a week, in the form of 8 apples and 7 oranges.
It is worthwhile to contrast the optimization problem for a linear function
with the non-linear case you may have seen in calculus courses:

74
3.3 Dantzig’s Algorithm 75

Here we have plotted the curve f (x) = d in the case where the function f is
linear and non-linear. To optimize f in the interval [a, b], for the linear case
we just need to compute and compare the values f (a) and f (b). In contrast,
for non-linear functions it is necessary to also compute the derivative df /dx
to study whether there are extrema inside the interval.

3.3 Dantzig’s Algorithm

In simple situations a graphical method might suffice, but in many applica-
tions there may be thousands or even millions of variables and constraints.
Clearly an algorithm that can be implemented on a computer is needed. The
simplex algorithm (usually attributed to George Dantzig) provides exactly
that. It begins with a standard problem:

Problem 38 Maximize f (x1 , . . . , xn ) where f is linear, xi 0 (i = 1, . . . , n) sub-

ject to
0 1
x1
B C
Mx = v , x := @ ... A ,
xn
where the m ⇥ n matrix M and m ⇥ 1 column vector v are given.

This is solved by arranging the information in an augmented matrix and

then applying EROs. To see how this works lets try an example.

75
76 The Simplex Method

Example 39 Maximize f (x, y, z, w) = 3x 3y z + 4w subject to constraints

c1 := x+y+z+w = 5
c2 := x + 2y + 3z + 2w = 6 ,

where x 0, y 0, z 0 and w 0.

The key observation is this: Suppose we are trying to maximize f (x1 , . . . , xn )

subject to a constraint c(x1 , . . . , xn ) = k for some constant k (c and k would
be the entries of M x and v, respectively, in the above). Then we can also
try to maximize
f (x1 , . . . , xn ) + ↵c(x1 , . . . , xn )
because this is only a constant shift f ! f + ↵k. Choosing ↵ carefully can
lead to a simple form for the function we are extremizing.

Example 40 (Setting up an augmented matrix):

Since we are interested in the optimum value of f , we treat it as an additional
variable and add one further equation

3x + 3y + z 4w + f = 0 .

We arrange this equation and the two constraints in an augmented matrix

0 1 8
1 1 1 1 0 5 < c1 = 5
>
B C
@ 1 2 3 2 0 6 A , c2 = 6 .
>
:
3 3 1 4 1 0 f = 3x 3y z + 4w

Keep in mind that the first four columns correspond to the positive variables (x, y, z, w)
and that the last row has the information of the function f . The general case is depicted
in figure 3.1.

Now the system is written as an augmented matrix where the last row
encodes the objective function and the other rows the constraints. Clearly we
can perform row operations on the constraint rows since this will not change
the solutions to the constraints. Moreover, we can add any amount of the
constraint rows to the last row, since this just amounts to adding a constant
to the function we want to extremize.

76
3.3 Dantzig’s Algorithm 77

variables (incl. slack and artificial) objective

z }| { z}|{
0 1
B C constraint equations
B C
@ A
objective equation
"
objective value

Figure 3.1: Arranging the information of an optimization problem in an

augmented matrix.

Example 41 (Performing EROs)

We scan the last row, and notice the (most negative) coefficient 4. Naı̈vely you
might think that this is good because this multiplies the positive variable w and only
helps the objective function f = 4w + · · · . However, what this actually means is
that the variable w will be positive and thus determined by the constraints. Therefore
we want to remove it from the objective function. We can zero out this entry by
performing a row operation. For that, either of the first two rows could be used.
To decide which, we remember that we still have to solve solve the constraints for
variables that are positive. Hence we should try to keep the first two entries in the
last column positive. Hence we choose the row which will add the smallest constant
to f when we zero out the 4: Look at the last column (where the values of the
constraints are stored). We see that adding four times the first row to the last row
would zero out the 4 entry but add 20 to f , while adding two times the second row
to the last row would also zero out the 4 but only add 12 to f . (You can follow this
by watching what happens to the last entry in the last row.) So we perform the latter
row operation and obtain the following:
0 1
1 1 1 1 0 5 c1 = 5
B C
@ 1 2 3 2 0 6 A c2 = 6
1 7 7 0 1 12 f = 12 + x 7y 7z .
We do not want to undo any of our good work when we perform further row operations,
so now we use the second row to zero out all other entries in the fourth column. This
is achieved by subtracting half the second row from the first:
0 1
1 1
2 0 2 0 0 2 c1 12 c2 = 2
B C
@ 1 2 3 2 0 6 A c2 = 6
1 7 7 0 1 12 f = 12 + x 7y 7z .

77
78 The Simplex Method

Precisely because we chose the second row to perform our row operations, all entries
in the last column remain positive. This allows us to continue the algorithm.
We now repeat the above procedure: There is a 1 in the first column of the last
row. We want to zero it out while adding as little to f as possible. This is achieved
by adding twice the first row to the last row:
0 1
1 1
2 0 2 0 0 2 c1 12 c2 = 2
B C
@ 1 2 3 2 0 6 A c2 = 6
0 7 6 0 1 16 f = 16 7y 6z .

The Dantzig algorithm terminates if all the coefficients in the last row (save perhaps
for the last entry which encodes the value of the objective) are positive. To see why
we are done, lets write out what our row operations have done in terms of the function
f and the constraints (c1 , c2 ). First we have

f = 16 7y 6z

with both y and z positive. Hence to maximize f we should choose y = 0 = z. In

which case we obtain our optimum value

f = 16 .

Finally, we check that the constraints can be solved with y = 0 = z and positive
(x, w). Indeed, they can by taking x = 4, w = 1.

3.4 Pablo Meets Dantzig

Oftentimes, it takes a few tricks to bring a given problem into the standard
form of example 39. In Pablo’s case, this goes as follows.

Example 42 Pablo’s variables x and y do not obey xi 0. Therefore define new

variables
x 1 = x 5 , x2 = y 7 .
The conditions on the fruit 15  x + y  25 are inequalities,

x1 + x2 3, x1 + x2  13 ,

so are not of the form M x = v. To achieve this we introduce two new positive
variables x3 0, x4 4 and write

c1 := x1 + x2 x3 = 3 , c2 := x1 + x2 + x4 = 13 .

78
3.4 Pablo Meets Dantzig 79

These are called slack variables because they take up the “slack” required to convert
inequality to equality. This pair of equations can now be written as M x = v,
0 1
✓ ◆ x1 ✓ ◆
1 1 1 0 B C
Bx2 C = 3 .
1 1 0 1 @x3 A 13
x4

Finally, Pablo wants to minimize sugar s = 5x + 10y, but the standard problem
maximizes f . Thus the so-called objective function f = s + 95 = 5x1 10x2 .
(Notice that it makes no di↵erence whether we maximize s or s + 95, we choose
the latter since it is a linear function of (x1 , x2 ).) Now we can build an augmented
matrix whose last row reflects the objective function equation 5x1 + 10x2 + f = 0:
0 1
1 1 1 0 0 3
@ 1 1 0 1 0 13 A .
5 10 0 0 1 0

Here it seems that the simplex algorithm already terminates because the last row only
has positive coefficients, so that setting x1 = 0 = x2 would be optimal. However, this
does not solve the constraints (for positive values of the slack variables x3 and x4 ).
Thus one more (very dirty) trick is needed. We add two more, positive, (so-called)
artificial variables x5 and x6 to the problem which we use to shift each constraint

c1 ! c1 x5 , c2 ! c 2 x6 .

The idea being that for large positive ↵, the modified objective function

f ↵x5 ↵x6

is only maximal when the artificial variables vanish so the underlying problem is un-
changed. Lets take ↵ = 10 (our solution will not depend on this choice) so that our
augmented matrix reads
0 1
1 1 1 0 1 0 0 3
@ 1 1 0 1 0 1 0 13 A
5 10 0 0 10 10 1 0
0 1
1 1 1 0 1 0 0 3
R30 =R3 10R1 10R2
⇠ @ 1 1 0 1 0 1 0 13 A .
15 10 10 10 0 0 1 160

Here we performed one row operation to zero out the coefficients of the artificial
variables. Now we are ready to run the simplex algorithm exactly as in section 3.3.

79
80 The Simplex Method

The first row operation uses the 1 in the top of the first column to zero out the most
negative entry in the last row:
0 1
1 1 1 0 1 0 0 3
@ 1 1 0 1 0 1 0 13 A
0 5 5 10 15 0 1 115
0 1
0
1 1 1 0 1 0 0 3
R2 =R2 R1
⇠ @ 0 0 1 1 1 1 0 10 A
0 5 5 10 15 0 1 115
0 1
1 1 1 0 1 0 0 3
R30 =R3 +10R2
⇠ @ 0 0 1 1 1 1 0 10 A .
0 5 5 0 5 10 1 15

Now the variables (x2 , x3 , x5 , x6 ) have zero coefficients so must be set to zero to
maximize f . The optimum value is f = 15 so s = f + 95 = 110 exactly as before.
Finally, to solve the constraints x1 = 3 and x4 = 10 so that x = 8 and y = 7 which
also agrees with our previous result.

Clearly, performed by hand, the simplex algorithm was slow and complex
for Pablo’s problem. However, the key point is that it is an algorithm that
can be fed to a computer. For problems with many variables, this method is
much faster than simply checking all vertices as we did in section 3.2.

3.5 Review Problems

1. Maximize f (x, y) = 2x + 3y subject to the constraints

x 0, y 0, x + 2y  2 , 2x + y  2 ,

by

(a) sketching the region in the xy-plane defined by the constraints

and then checking the values of f at its corners; and,
(b) the simplex algorithm (hint: introduce slack variables).

2. Conoil operates two wells (well A and well B) in southern Grease (a

small Mediterranean country). You have been employed to figure out
how many barrels of oil they should pump from each well to maximize

80
3.5 Review Problems 81

their profit (all of which goes to shareholders, not operating costs). The
quality of oil from well A is better than from well B, so is worth 50%
more per barrel. The Greasy government cares about the environment
and will not allow Conoil to pump in total more than 6 million barrels
per year. Well A costs twice as much as well B to operate. Conoil’s
yearly operating budget is only sufficient to pump at most 10 million
barrels from well B per year. Using both a graphical method and then
(as a double check) Dantzig’s algorithm, determine how many barrels
Conoil should pump from each well to maximize their profits.

81
82 The Simplex Method

82
 Vectors in Space, n-Vectors
4
To continue our linear algebra journey, we must discuss n-vectors with an
arbitrarily large number of components. The simplest way to think about
these is as ordered lists of numbers,
0 1
a1
B C
a = @ ... A .
an

Do not be confused by our use of a superscript to label components of a vector.

Here a2 denotes the second component of the vector a, rather than the number
a squared!
We emphasize that order matters:

Example 43 (Order of Components Matters)

0 1 0 1
7 7
B4C B2C
B C 6= B C .
@2A @4A
5 5

The set of all n-vectors is denoted Rn . As an equation

80 1 9
1
>
< a >
=
B .. C 1
R := @ . A a , . . . , a 2 R .
n n
>
: an >
;

83
84 Vectors in Space, n-Vectors

4.1 Addition and Scalar Multiplication in Rn

A simple but important property of n-vectors is that we can add two n-vectors
together and multiply one n-vector by a scalar:
Definition Given two n-vectors a and b whose components are given by
0 1 0 1
a1 b1
B C B C
a = @ ... A and b = @ ... A
an bn
their sum is 0 1
a1 + b 1
B .. C
a + b := @ . A.
n n
a +b
Given a scalar , the scalar multiple
0 1
a1
B C
a := @ ... A .
an
Example 44 Let 0 1 0 1
1 4
B2C B3C
a=B C B C
@3A and b = @2A .
4 1
Then, for example,
0 1 0 1
5 5
B5C B 0C
a+b=B C
@5A and 3a 2b = B C
@ 5A .
5 10
A special vector is the zero vector . All of its components are zero:
0 1
0
@ .. A
0= . =: 0n .
0
In Euclidean geometry—the study of Rn with lengths and angles defined
as in section 4.3 —n-vectors are used to label points P and the zero vector
labels the origin O. In this sense, the zero vector is the only one with zero
magnitude, and the only one which points in no particular direction.

84
4.2 Hyperplanes 85

4.2 Hyperplanes
Vectors in Rn are impossible to visualize unless n is 1,2, or 3. However,
familiar objects like lines and planes still make sense for any value of n: The
line L along the direction defined by a vector v and through a point P labeled
by a vector u can be written as

L = {u + tv | t 2 R} .

Sometimes, since we know that a point P corresponds to a vector, we will

be lazy and just write L = {P + tv | t 2 R}.

80 1 0 1? 9
>
> 1 1 ?? >
>
<B C B0C? =
B2 C B C ?
Example 45 @ A + t @ A?t 2 R describes a line in R4 parallel to the x1 -axis.
>
> 3 0 ? >
>
: ;
4 0 ?

Given two non-zero vectors u, v, they will usually determine a plane,

unless both vectors are in the same line, in which case, one of the vectors
is a scalar multiple of the other. The sum of u and v corresponds to laying
the two vectors head-to-tail and drawing the connecting vector. If u and v
determine a plane, then their sum lies in the plane determined by u and v.

85
86 Vectors in Space, n-Vectors

The plane determined by two vectors u and v can be written as

{P + su + tv | s, t 2 R} .

Example 46 (A plane in a higher dimensional space)

80 1 0 1 0 1? 9
>
> 3 1 0 ?? >
>
>
> B1C B0C B1C? >
>
>
> B C B C B C ? >
>
<B C B C B C ? =
4 0 0
B C + s B C + t B C?s, t 2 R
> B1C B0C B0C? >
>
> B C B C B C? >
>
>
> @ 5 A @ 0A @ 0A ? >
>
>
: ? >
;
9 0 0 ?

describes a plane in 6-dimensional space parallel to the xy-plane.

Parametric Notation

We can generalize the notion of a plane with the following recursive def-
inition. (That is, infinitely many things are defined in the following line.)

Definition A set of k + 1 vectors P, v1 , . . . , vk in Rn with k  n determines

a k-dimensional hyperplane,
( k
)
X
P+ i vi | i 2 R
i=1

86
4.2 Hyperplanes 87

unless any of the vectors vj lives in the (k 1)-dimensional hyperplane de-

termined by the other k 1 vectors
( k
)
X
0+ i vi | i 2 R .
i6=j

Example 47 (3+1 vectors that do not specify a 3-dimensional hyperplane)

80 1 0 1 0 1 0 1? 9
>
> 3 1 0 1 ?
? >
>
>
> B1C B0C B1C B1C? >
>
>
> B C B C B C B C? >
>
<B C B0C B0C B0C? =
B4 C B C B C B C?
S := B C + s B C + t B C + u B C?s, t, u 2 R
>B1C
>
> B0C B0C B0C? >
>
>
>
> @5A @0A @0A @0A? >
>
>
: ? >
;
9 0 0 0 ?

is not a 3-dimensional hyperplane because

0 1 0 1 0 1 8 0 1 0 1? 9
1 1 0 >
> 1 0 ?? >
>
B1C B0C B1C > >
> B0C B1C? >
>
>
B C B C B C > < B C B C? >
=
B0C B0C B0C B0C B0C?
B C = 1 B C + 1 B C 2 s B C + t B C?s, t 2 R .
B0C B0C B0C > B0C B0C? >
B C B C B C > > B C B C? >
>
@0A @0A @0A > > @0 A @0A? >
>
>
: ? >
;
0 0 0 0 0 ?

In fact, the set could be rewritten as

80 1 0 1 0 1? 9
>
> 3 1 0 ?? >
>
>B1C
> B0C B1C? >
>
>
> B C B C B C? >
>
<B C B0C B0C? =
B 4 C B C B C?
S = B C + (s + u) B C + (t + u) B C?s, t, u 2 R
>
>
> B1C B0C B0C? >
>
>
>
> @5A @0A @0A? >
>
>
: ? >
;
9 0 0 ?

80 1 0 1 0 1? 9
> 3 1 0 ? >
>
> ? >
>
>
> B1C B0C B1C? >
>
> B
<B C C B C B C ? >
=
4 B0 C B 0C ?
B C B C B C ?
= B C + a B C + b B C?a, b 2 R
>
>
> B1C B0C B0C? >
>
>
>
> @ 5A @0 A @0A? >
>
>
: ? >
;
9 0 0 ?

and so is actually the same 2-dimensional hyperplane in R6 as in example 46.

87
88 Vectors in Space, n-Vectors

You might sometimes encounter the word “hyperplane” without the qual-
ifier “k-dimensional. When the dimension k is not specified, one usually as-
sumes that k = n 1 for a hyperplane inside Rn . This is the kind of object
that is specified by one algebraic equation in n variables.

Example 48 (Specifying a plane with one linear algebraic equation.)

The solution set to
0
1 0 1
x1 1 x2 x3 x4 x5
Bx 2 C B x2 C
B C B C
x1 + x2 + x3 + x4 + x5 = 1 , B C B
Bx 3 C = B x3 C
C
@x 4 A @ x4 A
x5 x5

is
80 1 0 1 0 1 0 1 0 1? 9
>
> 1 1 1 1 1 ?? >
>
>
> >
<BB0C
C B
B 1C
C
B
B 0CC
B
B 0CC
B
B 1C
C?
? >
=
B0C + s2 B 0C B 1C B 0C B 0C ? s 2 , s3 , s4 , s5 2 R ,
> B C B C + s3 B C + s4 B C + s5 B C? >
>
> @0A @ 0A @ 0 A @ 1 A @ 0A? >
>
>
: ? >
;
0 0 0 0 1 ?

a 4-dimensional hyperplane in R5 .

4.3 Directions and Magnitudes

Consider the Euclidean length of an n-vector:
v
u n
p uX
kvk := (v 1 )2 + (v 2 )2 + · · · + (v n )2 = t (v i )2 .
i=1

Using the Law of Cosines, we can then figure out the angle between two
vectors. Given two vectors v and u that span a plane in Rn , we can then
connect the ends of v and u with the vector v u.

88
4.3 Directions and Magnitudes 89

Then the Law of Cosines states that:

kv uk2 = kuk2 + kvk2 2kuk kvk cos ✓

Then isolate cos ✓:

kv uk2 kuk2 kvk2 = (v 1 u1 )2 + · · · + (v n un )2

(u1 )2 + · · · + (un )2
(v 1 )2 + · · · + (v n )2
= 2u1 v 1 · · · 2un v n

Thus,
kuk kvk cos ✓ = u1 v 1 + · · · + un v n .

Note that in the above discussion, we have assumed (correctly) that Eu-
clidean lengths in Rn give the usual notion of lengths of vectors for any plane
in Rn . This now motivates the definition of the dot product.

1 00 1
u1 v1
B .. C B .. C
Definition The dot product of u = @ . A and v = @ . A is
un vn

u v := u1 v 1 + · · · + un v n .

89
90 Vectors in Space, n-Vectors

Example 49 of the dot product of two vectors from R100 .

0 1 0 1
1 1
B 2 C B1C
B C B C
B 3 C B1C 1
B C B C
B 4 C · B1C = 1 + 2 + 3 + · · · + 100 = .100.101 = 5050.
B C B C 2
B .. C B ..C
@ . A @ .A
100 1

The sum above is the one Gauß, according to legend, could do in kindergarten.

Definition The length (or norm or magnitude) of an n-vector v is

p
kvk := v v.

Example 50 of the norm of a vector from R101 .

0 1
1
B 2 C v
B C u 101
B 3 C uX p
B C
B 4 C =t i2 = 37, 961.
B C
B .. C i=1
@ . A
101

Definition The angle ✓ between two vectors is determined by the formula

u v = kukkvk cos ✓ .

Example 51 of an 0angle between

1 0two1vectors form R101 .
1 1
B 2 C B0C
B C B C
B 3 C B1C ⇣ ⌘
B C B C 10,201
The angle between B 4 C and B0C is arccos p37,916p51 .
B C B C
B .. C B ..C
@ . A @ .A
101 1

Definition Two vectors are orthogonal (or perpendicular) if their dot

product is zero.

90
4.3 Directions and Magnitudes 91

Example 52 of vectors from R101 that are orthogonal to each other.

0 1 0 1
1 1
B1C B 1C
B C B C
B1C B 1C
B C·B C = 0.
B ..C B ..C
@ .A @ .A
1 1

Notice that the zero vector 0n from Rn is orthogonal to every vector in Rn ;

0n · v = 0 for all v 2 Rn .
The dot product has some important properties; it is

1. symmetric:
u v = v u,

2. Distributive:
u (v + w) = u v + u w ,

3. Bilinear (which is to say, linear in both u and v):

u (cv + dw) = c u v + d u w ,

and
(cu + dw) v = c u v + d w v .

4. Positive Definite:
u u 0,
and u u = 0 only when u itself is the 0-vector.

There are, in fact, many di↵erent useful ways to define lengths of vectors.
Notice in the definition above that we first defined the dot product, and then
defined everything else in terms of the dot product. So if we change our idea
of the dot product, we change our notion of length and angle as well. The
dot product determines the Euclidean length and angle between two vectors.
Other definitions of length and angle arise from inner products, which
have all of the properties listed above (except that in some contexts the
positive definite requirement is relaxed). Instead of writing for other inner
products, we usually write hu, vi to avoid confusion.

91
92 Vectors in Space, n-Vectors

Reading homework: problem 1

Example 53 Consider a four-dimensional space, with a special direction which we will

call “time”. The Lorentzian inner product on R4 is given by hu, vi = u1 v 1 + u2 v 2 +
u3 v 3 u4 v 4 . This is of central importance in Einstein’s theory of special relativity.
Note, in particular, that it is not positive definite. As a result, the “squared-length”
of a vector with coordinates x, y, z and t is kvk2 = x2 + y 2 + z 2 t2 . Notice that
it is possible for kvk2  0 even with non-vanishing v! The physical interpretation of
this inner product depends on the sign of the inner product; two space time points
X1 := (x1 , y1 , z1 , t1 ), X2 := (x2 , y2 , z2 , t2 ) are
p
• separated by a distance hX1 , X2 i if hX1 , X2 i 0.

p
• separated by a time hX1 , X2 i if hX1 , X2 i  0.
In particular, the di↵erence in time coordinates t2 t1 is not the time between the
two points! (Compare this to using polar coordinates for which the distance between
two points (r, ✓1 ) and (r, ✓2 ) is not ✓2 ✓1 ; coordinate di↵erences are not necessarily
distances.)

Theorem 4.3.1 (Cauchy-Schwarz Inequality). For any non-zero vectors u

and v with an inner-product h , i
|hu, vi|
 1.
kuk kvk
The easiest proof would use the definition of the angle between two vectors
and the fact that cos ✓  1. However, strictly speaking speaking we did
not check our assumption that we could apply the Law of Cosines to the
Euclidean length in Rn . There is, however a simple algebraic proof.
Proof. Let ↵ be any real number and consider the following positive, quadratic
polynomial in ↵
0  hu + ↵v, u + ↵vi = hu, ui + 2↵hu, vi + ↵2 hv, vi .
2
Since any quadratic a↵2 +2b↵+c takes its minimal value c ba when ↵ = 2a b
,
and the inequality should hold for even this minimum value of the polynomial
hu, vi2 |hu, vi|
0  hu, ui ,  1.
hv, vi kuk kvk

92
4.3 Directions and Magnitudes 93

Theorem 4.3.2 (Triangle Inequality). For any u, v 2 Rn

ku + vk  kuk + kvk.

Proof.

ku + vk2 = (u + v) (u + v)
= u u + 2u v + v v
= kuk2 + kvk2 + 2 kuk kvk cos ✓
= (kuk + kvk)2 + 2 kuk kvk(cos ✓ 1)
 (kuk + kvk)2 .

That is, the square of the left-hand side of the triangle inequality is  the
square of the right-hand side. Since both the things being squared are posi-
tive, the inequality holds without the square;

ku + vk  kuk + kvk

The triangle inequality is also “self-evident” when examining a sketch of

u, v and u + v.

Example 54 Let 0 1 0 1
1 4
B2C B3C
a=B C B C
@3A and b = @2A ,
4 1
so that
a a = b b = 1 + 22 + 32 + 42 = 30

93
94 Vectors in Space, n-Vectors

p 2 p
) kak = 30 = kbk and kak + kbk = (2 30)2 = 120 .
Since 0 1
5
B5C
a+b=B C
@5A ,
5
we have
2
ka + bk2 = 52 + 52 + 52 + 52 = 100 < 120 = kak + kbk
as predicted by the triangle inequality. p p
Notice also that a b = 1.4 + 2.3 + 3.2 + 4.1 = 20 < 30. 30 = 30 = kak kbk in
accordance with the Cauchy–Schwarz inequality.

Reading homework: problem 2

4.4 Vectors, Lists and Functions: RS

If you were going shopping you might make something like the following list.

We could represent this information mathematically as a set,

S = {apple, orange, onion, milk, carrot} .

94
4.4 Vectors, Lists and Functions: RS 95

There is no information of ordering here and no information about how many

carrots you will buy. This set by itself is not a vector; how would we add
such sets to one another?
If you were a more careful shopper your list might look like the following.

What you have really done here is assign a number to each element of the
set S. In other words, the second list is a function
f : S ! R.
Given two lists like the second one above, we could easily add them – if you
plan to buy 5 apples and I am buying 3 apples, together we will buy 8 apples!
In fact, the second list is really a 5-vector in disguise.
In general it is helpful to think of an n-vector as a function whose domain
is the set {1, . . . , n}. This is equivalent to thinking of an n-vector as an
ordered list of n numbers. These two ideas give us two equivalent notions for
the set of all n-vectors:
80 1 9
1
>
< a >
=
B .. C 1
R := @ . A a , . . . , a 2 R = {a : {1, . . . , n} ! R} =: R{1,··· ,n}
n n
>
: an >
;

The notation R{1,··· ,n} is used to denote the set of all functions from {1, . . . , n}
to R.
Similarly, for any set S the notation RS denotes the set of functions from
S to R:
RS := {f : S ! R} .
When S is an ordered set like {1, . . . , n}, it is natural to write the components
in order. When the elements of S do not have a natural ordering, doing so
might cause confusion.

95
96 Vectors in Space, n-Vectors

Example 55 Consider the set S = {⇤, ?, #} from chapter 1 review problem 9. A

particular element of RS is the function a explicitly defined by

a? = 3, a# = 5, a⇤ = 2.

It is not natural to write

01 0 1
3 2
a = @ 5A or a = @ 3A
2 5

because the elements of S do not have an ordering, since as sets {⇤, ?, #} = {?, #, ⇤}.

In this important way, RS seems di↵erent from R3 . What is more evident

are the similarities; since we can add two functions, we can add two elements
of RS :

Example 56 Addition in R{⇤,?,#}

If a, b 2 R{⇤,?,#} such that

a? = 3, a# = 5, a⇤ = 2

and
b? = 2, b# = 4, b⇤ = 13
then a + b 2 RS is the function such that

(a + b)? = 3 2 = 1, (a + b)# = 5 + 4 = 9, (a + b)⇤ = 2 + 13 = 11 .

Also, since we can multiply functions by numbers, there is a notion of

scalar multiplication on RS .

Example 57 Scalar Multiplication in RS

If a 2 R{⇤,?,#} such that
a? = 3, a# = 5, a⇤ = 2
then 3a 2 R{⇤,?,#} is the function such that

(3a)? = 3 · 3 = 9, (3a)# = 3 · 5 = 15, (3a)⇤ = 3( 2) = 6.

We visualize R2 and R3 in terms of axes. We have a more abstract picture

of R4 , R5 and Rn for larger n while RS seems even more abstract. However,
when thought of as a simple “shopping list”, you can see that vectors in RS
in fact, can describe everyday objects. In chapter 5 we introduce the general
definition of a vector space that unifies all these di↵erent notions of a vector.

96
4.5 Review Problems 97

4.5 Review Problems

Reading problems 1 ,2
Vector operations 3
Vectors and lines 4
Webwork: Vectors and planes 5
Lines, planes and vectors 6,7
Equation of a plane 8,9
Angle between a line and plane 10

1. When he was young, Captain Conundrum mowed lawns on weekends to

help pay his college tuition bills. He charged his customers according to
the size of their lawns at a rate of 5¢ per square foot and meticulously
kept a record of the areas of their lawns in an ordered list:

A = (200, 300, 50, 50, 100, 100, 200, 500, 1000, 100) .

He also listed the number of times he mowed each lawn in a given year,
for the year 1988 that ordered list was

f = (20, 1, 2, 4, 1, 5, 2, 1, 10, 6) .

(a) Pretend that A and f are vectors and compute A f .

(b) What quantity does the dot product A f measure?
(c) How much did Captain Conundrum earn from mowing lawns in
1988? Write an expression for this amount in terms of the vectors
A and f .
(d) Suppose Captain Conundrum charged di↵erent customers di↵er-
ent rates. How could you modify the expression in part 1c to
compute the Captain’s earnings?

2. (2) Find the angle between the diagonal of the unit square in R2 and
one of the coordinate axes.
(3) Find the angle between the diagonal of the unit cube in R3 and
one of the coordinate axes.
(n) Find the angle between the diagonal of the unit (hyper)-cube in
Rn and one of the coordinate axes.

97
98 Vectors in Space, n-Vectors

(1) What is the limit as n ! 1 of the angle between the diagonal of

the unit (hyper)-cube in Rn and one of the coordinate axes?
✓ ◆ ✓ ◆
cos ✓ sin ✓ x
3. Consider the matrix M = and the vector X = .
sin ✓ cos ✓ y

(a) Sketch X and M X in R2 for several values of X and ✓.

||M X||
(b) Compute ||X||
for arbitrary values of X and ✓.
(c) Explain your result for (b) and describe the action of M geomet-
rically.

4. (Lorentzian Strangeness). For this problem, consider Rn with the

Lorentzian inner product defined in example 53 above.

(a) Find a non-zero vector in two-dimensional Lorentzian space-time

with zero length.
(b) Find and sketch the collection of all vectors in two-dimensional
Lorentzian space-time with zero length.
(c) Find and sketch the collection of all vectors in three-dimensional
Lorentzian space-time with zero length.
(d) Replace the word “zero” with the word “one” in the previous two
parts.

The Story of Your Life

5. Create a system of equations whose solution set is a 99 dimensional

hyperplane in R101 .

6. Recall that a plane in R3 can be described by the equation

0 1
x
n · @yA = n · p
z

98
4.5 Review Problems 99

where the vector p labels a given point on the plane and n is a vector
normal to the plane. Let N and P be vectors in R101 and
0 1
x1
B x2 C
B C
X = B .. C .
@ . A
x101

What kind of geometric object does N · X = N · P describe?

7. Let 0 1 0 1
1 1
B1C B 2 C
B C B C
B1C B 3 C
u = B C and v = B C
B .. C B .. C
@.A @ . A
1 101
Find the projection of v onto u and the projection of u onto v. (Hint:
Remember that two vectors u and v define a plane, so first work out
how to project one vector onto another in a plane. The picture from
Section 14.4 could help.)

8. If the solution set to the equation A(x) = b is the set of vectors whose
tips lie on the paraboloid z = x2 + y 2 , then what can you say about
the function A?

9. Find a system of equations whose solution set is

80 1 0 1 0 1 9
>
> 1 1 0 >
>
<B C B =
B1 C + c 1 B 1C B 0C
C + c 2 B C c 1 , c2 2 R .
> @ 2 A @ 0A @ 1A >
>
: >
;
0 1 3

Give a general procedure for going from a parametric description of a

hyperplane to a system of equations with that hyperplane as a solution
set.

10. If A is a linear operator and both v and cv (for any real number c) are
solutions to Ax = b, then what can you say about b?

99
100 Vectors in Space, n-Vectors

100
 Vector Spaces
5
As suggested at the end of chapter 4, the vector spaces Rn are not the only
vector spaces. We now give a general definition that includes Rn for all
values of n, and RS for all sets S, and more. This mathematical structure is
applicable to a wide range of real-world problems and allows for tremendous
economy of thought; the idea of a basis for a vector space will drive home
the main idea of vector spaces; they are sets with very simple structure.
The two key properties of vectors are that they can be added together
and multiplied by scalars. Thus, before giving a rigorous definition of vector
spaces, we restate the main idea.

A vector space is a set that is closed under addition and

scalar multiplication.

Definition A vector space (V, +, . , R) is a set V with two operations +

and · satisfying the following properties for all u, v 2 V and c, d 2 R:

(+i) (Additive Closure) u + v 2 V . Adding two vectors gives a vector.

(+ii) (Additive Commutativity) u + v = v + u. Order of addition does not

matter.

(+iii) (Additive Associativity) (u + v) + w = u + (v + w). Order of adding

many vectors does not matter.

101
102 Vector Spaces

(+iv) (Zero) There is a special vector 0V 2 V such that u + 0V = u for all u

in V .

(+v) (Additive Inverse) For every u 2 V there exists w 2 V such that

u + w = 0V .

(· i) (Multiplicative Closure) c · v 2 V . Scalar times a vector is a vector.

(· ii) (Distributivity) (c + d) · v = c · v + d · v. Scalar multiplication distributes

over addition of scalars.

(· iii) (Distributivity) c · (u + v) = c · u + c · v. Scalar multiplication distributes

over addition of vectors.

(· iv) (Associativity) (cd) · v = c · (d · v).

(· v) (Unity) 1 · v = v for all v 2 V .

Examples of each rule

Remark Rather than writing (V, +, . , R), we will often say “let V be a vector space
over R”. If it is obvious that the numbers used are real numbers, then “let V be a
vector space” suffices. Also, don’t confuse the scalar product · with the dot product .
The scalar product is a function that takes as its two inputs one number and one
vector and returns a vector as its output. This can be written

·: R ⇥ V ! V .

Similarly
+:V ⇥V !V .
On the other hand, the dot product takes two vectors and returns a number. Suc-
cinctly: : V ⇥ V ! R. Once the properties of a vector space have been verified,
we’ll just write scalar multiplication with juxtaposition cv = c · v, though, to keep our
notation efficient.

5.1 Examples of Vector Spaces

One can find many interesting vector spaces, such as the following:

102
5.1 Examples of Vector Spaces 103

Example of a vector space

Example 58
RN = {f | f : N ! R}
Here the vector space is the set of functions that take in a natural number n and return
a real number. The addition is just addition of functions: (f1 +f2 )(n) = f1 (n)+f2 (n).
Scalar multiplication is just as simple: c · f (n) = cf (n).
We can think of these functions as infinitely large ordered lists of numbers: f (1) =
13 = 1 is the first component, f (2) = 23 = 8 is the second, and so on. Then for
example the function f (n) = n3 would look like this:
0 1
1
B 8 C
B C
B 27 C
B C
f =B .
B . C.
.
C
B C
B n3 C
@ A
..
.

Thinking this way, RN is the space of all infinite sequences. Because we can not write
a list infinitely long (without infinite time and ink), one can not define an element of
this space explicitly; definitions that are implicit, as above, or algebraic as in f (n) = n3
(for all n 2 N) suffice.
Let’s check some axioms.

(+i) (Additive Closure) (f1 + f2 )(n) = f1 (n) + f2 (n) is indeed a function N ! R,

since the sum of two real numbers is a real number.

(+iv) (Zero) We need to propose a zero vector. The constant zero function g(n) = 0
works because then f (n) + g(n) = f (n) + 0 = f (n).

The other axioms should also be checked. This can be done using properties of the
real numbers.

Reading homework: problem 1

Example 59 The space of functions of one real variable.

RR = {f | f : R ! R}

103
104 Vector Spaces

The addition is point-wise

(f + g)(x) = f (x) + g(x) ,
as is scalar multiplication
c · f (x) = cf (x) .
To check that RR is a vector space use the properties of addition of functions and
scalar multiplication of functions as in the previous example.
We can not write out an explicit definition for one of these functions either, there
are not only infinitely many components, but even infinitely many components between
2
any two components! You are familiar with algebraic definitions like f (x) = ex x+5 .
However, most vectors in this vector space can not be defined algebraically. For
example, the nowhere continuous function
(
1, x 2 Q
f (x) = .
0, x 2 /Q

Example 60 R{⇤,?,#} = {f : {⇤, ?, #} ! R}. Again, the properties of addition and

scalar multiplication of functions show that this is a vector space.

You can probably figure out how to show that RS is vector space for any
set S. This might lead you to guess that all vector spaces are of the form RS
for some set S. The following is a counterexample.
Example 61 Another very important example of a vector space is the space of all
di↵erentiable functions: ⇢
d
f: R!R f exists .
dx
From calculus, we know that the sum of any two di↵erentiable functions is dif-
ferentiable, since the derivative distributes over addition. A scalar multiple of a func-
tion is also di↵erentiable, since the derivative commutes with scalar multiplication
d d
( dx (cf ) = c dx f ). The zero function is just the function such that 0(x) = 0 for ev-
ery x. The rest of the vector space properties are inherited from addition and scalar
multiplication in R.

Similarly, the set of functions with at least k derivatives is always a vector

space, as is the space of functions with infinitely many derivatives. None of
these examples can be written as RS for some set S. Despite our emphasis on
such examples, it is also not true that all vector spaces consist of functions.
Examples are somewhat esoteric, so we omit them.
Another important class of examples is vector spaces that live inside Rn
but are not themselves Rn .

104
5.1 Examples of Vector Spaces 105

Example 62 (Solution set to a homogeneous linear equation.)

Let 0 1
1 1 1
M = @2 2 2A .
3 3 3
The solution set to the homogeneous equation M x = 0 is
8 0 1 0 1 9
< 1 1 =
c1 @ A
1 + c2 @ A
0 c 1 , c2 2 R .
: ;
0 1
0 1
1
This set is not equal to R since it does not contain, for example, 0A. The sum of
3 @
0
any two solutions is a solution, for example
2 0 1 0 13 2 0 1 0 13 0 1 0 1
1 1 1 1 1 1
42 @ 1A + 3 @ 0A5 + 47 @ 1A + 5 @ 0A5 = 9 @ 1A + 8 @ 0A
0 1 0 1 0 1

and any scalar multiple of a solution is a solution

2 0 1 0 13 0 1 0 1
1 1 1 1
4 45 @ 1A 3 @ 0A5 = 20 @ 1A 12 @ 0A .
0 1 0 1

This example is called a subspace because it gives a vector space inside another vector
space. See chapter 9 for details. Indeed, because it is determined by the linear map
given by the matrix M , it is called ker M , or in words, the kernel of M , for this see
chapter 16.

Similarly, the solution set to any homogeneous linear equation is a vector

space: Additive and multiplicative closure follow from the following state-
ment, made using linearity of matrix multiplication:

If M x1 = 0 and M x2 = 0 then M (c1 x1 +c2 x2 ) = c1 M x1 +c2 M x2 = 0+0 = 0.

A powerful result, called the subspace theorem (see chapter 9) guarantees,

based on the closure properties alone, that homogeneous solution sets are
vector spaces.
More generally, if V is any vector space, then any hyperplane through
the origin of V is a vector space.

105
106 Vector Spaces

Example 63 Consider the functions f (x) = ex and g(x) = e2x in RR . By taking

combinations of these two vectors we can form the plane {c1 f + c2 g|c1 , c2 2 R} inside
of RR . This is a vector space; some examples of vectors in it are 4ex 31e2x , ⇡e2x 4ex
and 12 e2x .

A hyperplane which does not contain the origin cannot be a vector space
because it fails condition (+iv).
It is also possible to build new vector spaces from old ones using the
product of sets. Remember that if V and W are sets, then their product is
the new set
V ⇥ W = {(v, w)|v 2 V, w 2 W } ,
or in words, all ordered pairs of elements from V and W . In fact V ⇥ W is a
vector space if V and W are. We have actually been using this fact already:

Example 64 The real numbers R form a vector space (over R). The new vector space

R ⇥ R = {(x, y)|x 2 R, y 2 R}

has addition and scalar multiplication defined by

(x, y) + (x0 , y 0 ) = (x + x0 , y + y 0 ) and c.(x, y) = (cx, cy) .

Of course, this is just the vector space R2 = R{1,2} .

5.1.1 Non-Examples
The solution set to a linear non-homogeneous equation is not a vector space
because it does not contain the zero vector and therefore fails (iv).

Example 65 The solution set to

✓ ◆✓ ◆ ✓ ◆
1 1 x 1
=
0 0 y 0
⇢✓ ◆ ✓ ◆ ✓ ◆
1 1 0
is +c c 2 R . The vector is not in this set.
0 1 0

Do notice that if just one of the vector space rules is broken, the example is
not a vector space.
Most sets of n-vectors are not vector spaces.

106
5.2 Other Fields 107

⇢✓ ◆
a
Example 66 P := a, b 0 is not a vector space because the set fails (·i)
b
✓ ◆ ✓ ◆ ✓ ◆
1 1 2
since 2 P but 2 = 2
/ P.
1 1 2

Sets of functions other than those of the form RS should be carefully

checked for compliance with the definition of a vector space.

Example 67 The set of all functions which are nowhere zero

{f : R ! R | f (x) 6= 0 for any x 2 R} ,

does not form a vector space because it does not satisfy (+i). The functions f (x) =
x2 +1 and g(x) = 5 are in the set, but their sum (f +g)(x) = x2 4 = (x+2)(x 2)
is not since (f + g)(2) = 0.

5.2 Other Fields

Above, we defined vector spaces over the real numbers. One can actually
define vector spaces over any field. This is referred to as choosing a di↵erent
base field. A field is a collection of “numbers” satisfying properties which are
listed in appendix B. An example of a field is the complex numbers,

C = x + iy | i2 = 1, x, y 2 R .

Example 68 In quantum physics, vector spaces over C describe all possible states a
physical system can have. For example,
⇢✓ ◆
V = | ,µ 2 C
µ
✓ ◆ ✓ ◆
1 0
is the set of possible states for an electron’s spin. The vectors and describe,
0 1
respectively, ✓an electron
◆ with spin “up” and “down” along a given direction. Other
i
vectors, like are permissible, since the base field is the complex numbers. Such
i
states represent a mixture of spin up and spin down for the given direction (a rather
counterintuitive yet experimentally verifiable concept), but a given spin in some other
direction.

107
108 Vector Spaces

Complex numbers are very useful because of a special property that they
enjoy: every polynomial over the complex numbers factors into a product of
linear polynomials. For example, the polynomial

x2 + 1

doesn’t factor over real numbers, but over complex numbers it factors into

(x + i)(x i) .

In other words, there are two solutions to

x2 = 1,

x = i and x = i. This property has far-reaching consequences: often in

mathematics problems that are very difficult using only real numbers become
relatively simple when working over the complex numbers. This phenomenon
occurs when diagonalizing matrices, see chapter 13.
The rational numbers Q are also a field. This field is important in com-
puter algebra: a real number given by an infinite string of numbers after the
decimal point can’t be stored by a computer. So instead rational approxi-
mations are used. Since the rationals are a field, the mathematics of vector
spaces still apply to this special case.
Another very useful field is bits

B2 = Z2 = {0, 1} ,

with the addition and multiplication rules

+ 0 1 ⇥ 0 1
0 0 1 0 0 0
1 1 0 1 0 1

These rules can be summarized by the relation 2 = 0. For bits, it follows

that 1 = 1!
The theory of fields is typically covered in a class on abstract algebra or
Galois theory.

108
5.3 Review Problems 109

5.3 Review Problems

Reading problems 1
Webwork:
Addition and inverse 2
⇢✓ ◆
x
1. Check that x, y 2 R = R2 (with the usual addition and scalar
y
multiplication) satisfies all of the parts in the definition of a vector
space.

2. (a) Check that the complex numbers C = {x + iy | i2 = 1, x, y 2 R},

satisfy all of the parts in the definition of a vector space over C.
Make sure you state carefully what your rules for vector addition
and scalar multiplication are.

(b) What would happen if you used R as the base field (try comparing
to problem 1).

3. (a) Consider the set of convergent sequences, with the same addi-
tion and scalar multiplication that we defined for the space of
sequences:
n o
V = f | f : N ! R, lim f (n) 2 R ⇢ RN .
n!1

Is this still a vector space? Explain why or why not.

(b) Now consider the set of divergent sequences, with the same addi-
tion and scalar multiplication as before:
n o
V = f | f : N ! R, lim f (n) does not exist or is ± 1 ⇢ RN .
n!1

Is this a vector space? Explain why or why not.

4. Consider the set of 2 ⇥ 4 matrices:

⇢✓ ◆
a b c d
V = a, b, c, d, e, f, g, h 2 C
e f g h

109