An Introduction to Linear Algebra.pdf

An Introduction to Linear Algebra
✁
✁
✁
✁
✁
✁
✁
✁
✕
b
✏✏✏✏✏✏✏✏
✶ x1
✄
✄
✄
✄
✎
x2
❍❍❍
❍
❥ p = Ax̃
e = b − Ax̃
✏✏✏✏✏✏✏✏✏✏✏✏✏✏✏✏✏✏✏✏
✏
✏✏✏✏✏✏✏✏✏✏✏✏✏✏✏✏✏✏✏✏
❍
❍
❍
❍
❍
❍
❍
❍
❍
❍
❍
❍
❍
❍
❍
❍
❍
❍
❍
❍
R(A)
Andreas Jakobsson
Lund University
Version 080515

i
An Introduction to Linear Algebra
These notes were written as a part of a graduate level course on transform the-
ory offered at King’s College London during 2002 and 2003. The material is heavily
indebt to the excellent textbook by Gilbert Strang [1], which the reader is referred
to for a more complete description of the material; for a more in-depth coverage,
the reader is referred to [2–6].
Andreas Jakobsson
andreas.jakobsson@ieee.org

CONTENTS
1 LINEAR ALGEBRA 1
1.1 Solving a set of linear equations 1
1.2 Vectors and matrices 2
1.3 Fundamental subspaces 6
1.4 Orthogonality 9
1.5 Determinants 14
1.6 Eigendecomposition 15
1.7 Singular value decomposition 19
1.8 Total least squares 21
A BANACH AND HILBERT SPACES 24
B THE KARHUNEN-LOÈVE TRANSFORM 27
C USEFUL FORMULAE 29
BIBLIOGRAPHY 30
ii

Chapter 1
LINEAR ALGEBRA
Dimidium facti, qui coepit, habet
Horace
1.1 Solving a set of linear equations
We introduce the topic of linear algebra by initially examining how to go about
solving a set of linear equations; say



2u + v + w = 5
4u − 6v = −2
−2u + 7v + 2w = 9
(1.1.1)
By subtracting 2 times the first equation from the second equation, we obtain



2u + v + w = 5
− 8v − 2w = −12
−2u + 7v + 2w = 9
(1.1.2)
Term this step (i). We then add the first equation to the third equation in (1.1.2),



2u + v + w = 5
− 8v − 2w = −12
8v + 3w = 14
(1.1.3)
We term this step (ii). Finally, in step (iii), we add the second and third equation
in (1.1.3), obtaining



2u + v + w = 5
− 8v − 2w = −12
w = 2
(1.1.4)
1

2 Linear Algebra Chapter 1
This procedure is known as Gaussian1
elimination, and from (1.1.4) we can easily
solve the equations via insertion, obtaining the solution (u, v, w) = (1, 1, 2). An
easy way to ensure that the found solution is correct is to simply insert (u, v, w)
in (1.1.1). However, we will normally prefer to work with matrices instead of the
above set of equations. The matrix approach offers many benefits, as well as opening
up a new way to view the equations, and we will here focus on matrices and the
geometries that can be associated with matrices.
1.2 Vectors and matrices
We proceed to write (1.1.1) on matrix form


2
4
−2

 u +


1
−6
7

 v +


1
0
2

 w
| {z }
2
6
6
4
2 1 1
4 −6 0
−2 7 2
3
7
7
5
2
6
6
4
u
v
w
3
7
7
5
=


5
−2
9

 (1.2.1)
Note that we can view the columns of the matrix as vectors in a three-dimensional
(real-valued) space, often written R3
. This is only to say that the elements of the
vectors can be viewed as the coordinates in a space, e.g., we view the third column
vector, say a3,
a3 =


1
0
2

 (1.2.2)
as the vector starting at origin and ending at x = 1, y = 0 and z = 2. Using a
simplified geometric figure, we can depict the three column vectors as
✑
✑
✑
✑
✑
✑
✑
✑
✸
✲
✁
✁
✁
✁
✁
✁
✕ a2
a1
a3
where the vectors are seen as lying in a three-dimensional space. The solution
to the linear set of equations in (1.1.1) is thus the scalings (u, v, w) such that if we
1Johann Carl Friedrich Gauss (1777-1855) was born in Brunswick, the Duchy of Brunswick
(now Germany). In 1801, he received his doctorate in mathematics, proving the fundamental
theorem of algebra [7]. He made remarkable contributions to all fields of mathematics, and is
generally considered to be the greatest mathematician of all times.

Section 1.2. Vectors and matrices 3
combine the vectors as
ua1 + va2 + wa3 = b, (1.2.3)
we obtain the vector b, in this case b =

5 −2 9
T
, where (·)T
denotes the
transpose operator. This way of viewing the linear equations also tells us when a
solution exists. We say that the matrix formed from the three column vectors,
A =

a1 a2 a3

, (1.2.4)
span the column space, i.e., the vector space containing all points reachable by any
linear combination of the column vectors of A. For example, consider
B =

b1 b2

=


1 0
0 1
0 0

 (1.2.5)
In this case, all the points in the plane b1-b2 belongs to the column space of B;
this as by combining b1 and b2 all points in the plane might be reached. However,
the point b3 =

0 0 1
T
does not lie in the column space since there is no way
to combine b1 and b2 to reach b3. Thus, it holds that
A linear set of equations is solvable if and only if the vector b lies
in the column space of A.
Viewing the linear equations as vectors will be central to our treatment, and we will
shortly return to this way of viewing the problem. Before doing so, we will examine
how we can view the above Gaussian elimination while working on the matrix form.
We can view the elimination in step (i) above as a multiplication by an elemen-
tary matrix, E,


1 0 0
−2 1 0
0 0 1


| {z }
E


2 1 1
4 −6 0
−2 7 2




u
v
w

 = E


5
−2
9

 (1.2.6)
which yields (cf. (1.1.2))


2 1 1
0 −8 −2
−2 7 2




u
v
w

 =


5
−12
9

 (1.2.7)
Similarly, we can view step (ii) and (iii) as multiplications with elementary matrices


1 0 0
0 1 0
0 1 1


| {z }
G


1 0 0
0 1 0
1 0 1


| {z }
F


2 1 1
0 −8 −2
−2 7 2




u
v
w

 = GF


5
−12
9

 (1.2.8)

yielding (cf. 1.1.4))


2 1 1
0 −8 −2
0 0 1




u
v
w

 =


5
−12
2

 (1.2.9)
The resulting matrix, which we will term U, will be an upper-triangular matrix.
We denote the sought solution x =

u v w
T
, and note that we have by the
matrix multiplications rewritten
Ax = b ⇒ Ux = c, (1.2.10)
where U = (GFE) A and c = (GFE) b. As U is upper-triangular, the new matrix
equation is obviously trivial to solve via insertion.
We proceed by applying the inverses of the above elementary matrices (in reverse
order!), and note that
E−1
F−1
G−1
| {z }
L
GFEA
| {z }
U
= A. (1.2.11)
This is the so-called LU-decomposition, stating that
A square matrix that can be reduced to U (without row inter-
changes2
) can be written as A = LU.
As the inverse of an elementary matrix can be obtain by simply changing the sign
of the off-diagonal element3
,
E−1
=


1 0 0
−2 1 0
0 0 1


−1
=


1 0 0
2 1 0
0 0 1

 (1.2.12)
we can immediately obtain the LU-decomposition after reducing A to U,
A = LU =


1 0 0
2 1 0
−1 −1 1




2 1 1
0 −8 −2
0 0 1

 (1.2.13)
Note that L will be a lower-triangular matrix. Often, one divides out the diagonal
elements of U, the so-called pivots4
, denoting the diagonal matrix with the pivots
on the diagonal D, i.e.,
A = LU = LDŨ. (1.2.14)
2Note that in some cases, row exchanges are required to reduce A to U. In such case, A can
not be written as LU. See [1] for further details on row interchanges.
3Please note that this only holds for elementary matrices!
4We note that pivots are always nonzero.

Section 1.2. Vectors and matrices 5
A bit confusing, Ũ is normally just denoted U, forcing the reader to determine
which of the matrices are referred to from the context. For our example,
A = LDŨ =


1 0 0
2 1 0
−1 −1 1




2 0 0
0 −8 0
0 0 1




1 0.5 0.5
0 1 0.25
0 0 1

 (1.2.15)
The LU-decomposition suggests an efficient approach to solve a linear set of equa-
tions by first factoring the matrix as A = LU, and then solve Lc = b and Ux = c.
It can be shown that the LU-decomposition of a n × n matrix requires n3
/3 oper-
ations, whereas the solving of the two triangular systems requires n2
/2 operations
each. Thus, if we wish to solve the same set of equations (same A matrix) with
a new b vector, this can be done in only n2
operations (which is significantly less
than n3
/3+n2
). Often, the inherent structure of the A matrix can further simplify
calculations. For example, we will often consider symmetric matrices, i.e., matrices
satisfying
A = AT
, (1.2.16)
or, more generally, Hermitian5
matrices, i.e., matrices satisfying
A = AH
, (1.2.17)
where (·)H
denotes the Hermitian, or conjugate transpose, operator. Obviously, a
Hermitian matrix is also symmetric, whereas the converse is in general not true. If
A is Hermitian, and if it can be factored as A = LDU (without row exchanges to
destroy the symmetry), then U = LH
, i.e.,
A square Hermitian matrix that (without row interchanges) can be
reduced to U, can be written as A = LDLH
, where D is diagonal,
and L lower-triangular.
This symmetry will simplify the computation of the LU-decomposition, reducing
the complexity to n3
/6 operations. Further, in the case when the diagonal elements
of D are all positive, the LU-decomposition can be written as
A = LDLH
=

LD1/2

LD1/2
H
= CCH
, (1.2.18)
where C is the (matrix) square root of A. However, if C is a square root of A, then
so is CB, i.e.,
A = (CB) (CB)
H
, (1.2.19)
5Charles Hermite (1822-1901) was born in Dieuze, France, with a defect in his right foot, making
it hard for him to move around. After studying one year at the Ecole Polytechnique, Hermite was
refused the right to continue his studies because of this disability. However, this did not prevent
his work, and in 1856 he was elected to the Académie des Sciences. After making important
contributions to number theory and algebra, orthogonal polynomials, and elliptic functions, he
was in 1869 appointed professor of analysis both at the Ecole Polytechnique and at the Sorbonne.

for any matrix B such that
BBH
= BH
B = I. (1.2.20)
A matrix satisfying (1.2.20) is said to be a unitary matrix 6
. Hence, there are an
infinite number of square roots of a given matrix satisfying (1.2.18). Two often used
choices for square roots are
(i) The Hermitian square root: C = CH
. The Hermitian square root is unique.
(ii) The Cholesky7
factor. If C is lower-triangular with non-negative diagonal
elements, then C is called the Cholesky factor of A. This is the famous
Cholesky factorization of A.
We will now return to the discussion on vectors and vector spaces.
1.3 Fundamental subspaces
Recall that Ax = b is solvable if and only if the vector b lies in the column space
of A, i.e., if it can be expressed as a combination of the column vectors of A.
We will now further discuss vector spaces8
, as well as different ways to represent
such a space. For instance, how can we represent this vector space using as few
vectors as possible? Can the matrix A be used to span any other vector spaces?
To answer these questions, we will discuss what it means to span a space, and some
matters that concern vector spaces. We begin with the central definition of linear
dependence. If
c1v1 + · · · + ckvk = 0 only happens when c1 = c2 = · · · = ck = 0 (1.3.1)
then the vectors v1, . . . , vk are linearly independent. Otherwise, at least one of the
vectors is a linear combination of the others. The number of linearly independent
columns of a matrix is termed the column rank of the matrix. If the matrix consists
of only linearly independent column vectors, one say that it has full column rank.
For example, consider
A =


1 3 3 2
2 6 9 5
−1 −3 3 0

 (1.3.2)
6In the case of real-valued data, (1.2.20) can be written as BBT = BT B = I. Such a matrix is
said to be an orthogonal matrix. Here, the term orthonormal matrix would be more appropriate,
but this term is not commonly used.
7Andre-Louis Cholesky (1875-1918) was born in Montguyon, France. He was a French military
officer involved in geodesy and surveying in Crete and North Africa just before World War I. He
developed the method now named after him to compute solutions to the normal equations for
some least squares data fitting problems arising in geodesy. He died in battle during the war, his
work published posthumously.
8See appendix A for a further discussion on vector spaces.

Section 1.3. Fundamental subspaces 7
The matrix A clearly does not have full column rank, as the column vectors are
linearly dependent: note that
a2 = 3a1
a4 = a1 +
1
3
a3
Only the first and the third column vectors are linearly independent, and the column
rank of A is thus two. We are now ready to clarify what we mean by the spanning
of a space: if a vector space X consists of all linear combinations of the particular
vectors x1, x2, . . . , xn, then these vectors span the space. In other words, this means
that every vector v ∈ X can be expressed as some combination of these vectors:
v = c1x1 + c2x2 + . . . + cnxn, (1.3.3)
for some coefficients {ck}. Note that a set of vectors that span a space does not
need to be linearly independent. We can for instance use the column vectors of A
in (1.3.2) to span the column space of A, normally denoted R(A). Obviously, there
is an infinite number of vectors that we could use to span R(A), but how many are
really needed? This question leads to the idea of a basis.
A basis for a vector space is a set of vectors satisfying:
(i) It is linearly independent.
(ii) It spans the space.
The number of vectors spanning a basis is called the dimension of the space. For
the matrix in (1.3.2), the first and the third column vectors are sufficient to span
the space, and these two constitute a basis for R(A). The dimension of R(A),
which is (always) the same as the column rank of the matrix A, is thus two. As a
result, the column space will be a subspace of R3
, this as the column vectors lie in
a three-dimensional (real-valued) space. Note that R(A), being two-dimensional,
will not fill the entire R3
, immediately raising the question what fills the remaining
part of this space. We will return to this point shortly.
There is clearly an infinite number of bases spanning any given space, for instance,
any scaled version of a1 and a3 would do. We will eventually examine how to pro-
ceed to pick an (in some sense) appropriate basis, but will first spend a moment on
the row vectors. Considering (1.3.2), we can obviously view it as constructed from
a set of row vectors,
A =


1 3 3 2
2 6 9 5
−1 −3 3 0

 △
=


b1
b2
b3

 (1.3.4)

where
b1 =

1 3 3 2

(1.3.5)
b2 =

2 6 9 5

(1.3.6)
b3 =

−1 −3 3 0

(1.3.7)
Note that the row vectors lie not in R3
, but in R4
. This as they contain four,
not three, (real-valued) indices. These vectors must for this reason reside in a
different vector space than the column vectors; we will get back to this point shortly.
Similarly to our definition of column rank, we define the row rank as the number of
linearly independent row vectors, and speak of full row rank if all the row vectors
are linearly independent. What is the row rank of the matrix in (1.3.4)? We can
easily answer this using the following remarkable fact:
The number of independent column vectors is always equal to the
number of independent row vectors!
This means that one (and only one) of the row vectors in (1.3.4) must be linearly
dependent. Can you see this? This also means that the row rank and column ranks
are always equal. For this reason, we normally only refer to the row and the column
rank as the rank of the matrix9
,
rank(A) = rank(AT
). (1.3.8)
Consequently, we say that the matrix is full rank when all the column vectors and
the row vectors are linearly independent. For obvious reasons, only square matrices
can be full rank. Let us study some examples:


1 1 −1
−1 1 1
1 1 −1




1 1 −1
1 1 1
−1 1 −1




−1 1 1
1 −1 1
1 −1 −1


What are the dimensions of the column spaces spanned by these matrices? What
are the ranks of the matrices? Which of the matrices are bases?
We are now ready to examine the question if A will span other spaces than R(A).
From the previous discussion, it might seem as spanning one space is plenty: in fact
A will span four different spaces, the so-called fundamental subspaces. Consider a
matrix A containing m row vectors and n column vectors, normally referred to as
a m × n matrix, with rank r. Which are then these fundamental subspaces? The
two first are easily grasped:
(i) The column space, R(A), spanned by the column vectors of A.
9We note that the rank of a matrix is equal to the number of pivots.

Section 1.4. Orthogonality 9
(ii) The row space, R(AT
), spanned by the row vectors of A.
Both these spaces will have dimension r, as the number of linearly independent
column vectors equal the number of linearly independent row vectors. The two
remaining subspaces consist of the so-called nullspaces. Consider a vector x such
that
Ax = 0. (1.3.9)
Obviously the null vector, x = 0, will satisfy this, but other vectors may do so as
well. Any vector, other than the null vector, satisfying (1.3.9) is said to lie in the
nullspace of A. Similarly, we denote the space spanned by the vectors y, such that
yT
A = 0, (1.3.10)
the left nullspace of A, this to denote that the vector y multiply A from the left10
.
The two final fundamental subspaces are thus:
(iii) The nullspace, N(A), is spanned by the vectors x satisfying (1.3.9).
(iv) The left nullspace, N(AT
), is spanned by the vectors y satisfying (1.3.10).
For example, for the matrix
B =

1 0 0
0 0 0

(1.3.11)
the columns space, R(B), and row space, R(BT
), are the lines (both having dimen-
sion r = 1) through

1 0
T
respectively

1 0 0
T
Further, the nullspace, N(B), is a plane containing

0 1 0
T
and

0 0 1
T
and the left nullspace, N(BT
), is the line through

0 1
T
. We note that the
nullspaces have different dimensions; how are these dimensions related to the matrix
B? To answer this, we will need examine the notion of orthogonality.
1.4 Orthogonality
We begin by introducing the inner product between two vectors x and y as
hx, yi
△
= xH
y. (1.4.1)
The inner product is a measure of similarity between the vectors, and is propor-
tional to the degree of closeness between x and y; the more similar the vectors are,
10Thus, the ordinary nullspace can be thought of as the “right” nullspace.

the bigger the inner product. If the vectors are parallel (either direct or reverse),
the magnitude of the inner product reaches its maximum. If the vectors are perpen-
dicular, the inner product is zero (more on this in the next section). For example,
the inner product between a1 and a2 in (1.3.2) is
ha1, a2i = aT
1 a2 = 3 aT
1 a1 = 18.
From geometry, it is well known that one can compute the angle between two vectors
using the inner product between the vectors and their lengths as
cos θ =
|hx, yi|
||x||2 ||y||2
. (1.4.2)
where the length of a vector is (in Euclidean space) the 2-norm11
of the vector, i.e.,
for a vector in l2(C),
||x||2
△
=
p
hx, xi =
√
xHx. (1.4.3)
From (1.4.3), we conclude that no vector (excluding the null vector) can have zero
length. Considering (1.4.2), we see that two vectors are orthogonal if and only if
their inner product hx, yi = 0. For example, the vectors
x =

2 2 −1
T
y =

−1 2 2
T
are orthogonal, as their inner product is zero. Also note that the vectors x and y
are linearly independent! In general, it holds that:
Orthogonal vectors are linearly independent.
Further, we say that two subspaces are orthogonal to each other if every vector in
the first subspace V is orthogonal to every vector in the second subspace W. The
subspace W is then said to lie in the orthogonal complement of V, denoted V⊥
(pronounced “V perp”). Similarly, V will lie in W⊥
. Recalling the definitions of
the fundamental subspaces, we see that
(i) The nullspace is the orthogonal complement of the row space.
(ii) The left nullspace is the orthogonal complement of the column space.
This explains the dimensions of the subspaces spanned by B in (1.3.11): every
vector that does not lie in the row space will belong to the nullspace, every vector
not in the column space will belong to the left nullspace! Thus, the dimensions of
the four fundamental subspaces for a m × n matrix A are
11See appendix A for a further discussion on norms and spaces. To make the discussion more
general, we hereafter consider only complex-valued vectors unless otherwise specified, with real-
valued vectors being a special case. Note that according to the Cauchy-Schwartz inequality, as
stated in (C.1.2), (1.4.2) is limited to 0 ≤ θ ≤ π
2
.

(i) dim{R(A)} = rank{A} = r
(ii) dim

R(AT
) = r
(iii) dim{N(A)} = n − r.
(iv) dim

N(AT
) = m − r.
The notion of orthogonality can be of further use; recall that Ax = b is only solvable
if b ∈ R(A). What happens if b /
∈ R(A)? The obvious answer is that there exists
no vector x such that Ax = b. However, we can still find a vector x̃ such that Ax̃
will (in some sense) lie as close as possible to b. Often, one choose the vector x̃ such
that the Euclidean distance kAx̃ − bk2
2 is minimized, in which case x̃ is referred to
as the least squares solution12
, i.e., the vector minimizing the squared error. Put
differently,
x̃ = arg min
x
kAx − bk2
2 (1.4.4)
Geometrically, this is the same as finding the vector p = Ax̃, lying in R(A), such
that the error, e, between b and Ax̃, is as short as possible. Thus, if b lies in R(A),
the error is zero and the exact solution x̃ = x will exist. Otherwise, the error will
be minimized if chosen such that it is perpendicular to the column space13
. This
is a very important point; spend a moment examining the figure below and make
sure that you understand it!
✁
✁
✁
✁
✁
✁
✁
✁
✕
b
✏✏✏✏✏✏✏✏
✶ x1
✄
✄
✄
✄
✎
x2
❍❍❍
❍
❥ p = Ax̃
e = b − Ax̃
✏✏✏✏✏✏✏✏✏✏✏✏✏✏✏✏✏✏✏✏
✏
✏✏✏✏✏✏✏✏✏✏✏✏✏✏✏✏✏✏✏✏
❍
❍
❍
❍
❍
❍
❍
❍
❍
❍
❍
❍
❍
❍
❍
❍
❍
❍
❍
❍
R(A)
As the error, e = b − Ax̃, will be orthogonal to R(A), the vector p = Ax̃ can
be viewed as the projection of b onto R(A). To see this, note that as e ⊥ R(A),
AH
e = AH
(b − Ax̃) = 0 ⇒ AH
Ax̃ = AH
b. (1.4.5)
12The least squares solution was introduced by C. F. Gauss, who used it to determine the orbit
of the asteroid Ceres in 1821 by formulating the estimation problem as an optimization problem.
13This fact is often exploited in estimation theory, where it is termed the orthogonality principle.

Using the fact that
If the A has linearly independent columns, then AH
A is square,
Hermitian and invertible.
we obtain
x̃ = AH
A
−1
AH
b
△
= A†
b, (1.4.6)
where A†
denotes the Moore-Penrose pseudoinverse of A. The projection of b onto
R(A) is thus
p = Ax̃ = A AH
A
−1
AH
b = AA†
b
△
= ΠAb, (1.4.7)
where ΠA is the projection matrix onto R(A). What would happen if you would
project p onto R(A)? In other words, what is Π2
Ab? How can you explain this
geometrically? Another point worth noting is that
e = b − Ax̃ = b − ΠAb = (I − ΠA) b
△
= Π⊥
Ab, (1.4.8)
where Π⊥
A is the projection matrix onto the space orthogonal to R(A), i.e., Π⊥
A is
the projection matrix onto the left nullspace. Minimizing the error vector is thus
the same thing as minimizing the projection of b onto the left nullspace, i.e., we
write b = ΠAb + Π⊥
Ab and find the vector x̃ such that
x̃ = arg min
x
kΠ⊥
Abk2
2, (1.4.9)
which is equivalent to the expression in (1.4.4). As an example, lets examine an
overdetermined (m n) system


1 1
1 −1
−1 1


| {z }
A
x =


1
1
1


| {z }
b
(1.4.10)
As b does not lie in R(A) (show this!), no solution will exist. The closest solution,
in the least squares sense, can be found as
x̃ = A†
b =

0.5
0.5

(1.4.11)
which corresponds to
p = Ax̃ = ΠAb =

1 0 0
T
(1.4.12)
Verify that the error vector is orthogonal to the column space by computing the
angle between b − Ax̃ and the columns of A!

We now proceed to discuss how to find an orthogonal basis for a given space,
say R(A); as commented earlier, a basis is formed by a set of linearly independent
vector spanning the space. By necessity, this set must contain r = rank {A} vectors
(why?), but obviously there exist an infinite number of such vector sets. We are
here interested in finding an orthogonal set of vectors using A. This can be done in
a series of projection termed the Gram-Schmidt orthogonalization which will now
be described briefly; consider the matrix
A =

a1 . . . an

(1.4.13)
and choose the first basis vector, q1, as
q1 =
a1
ka1k2
(1.4.14)
Note that due to the normalization, q1 has unit length, we say it is a unit vector.
The second basis vector should be orthogonal to q1, and for this reason we form
q′
2 = a2 − hq1, a2iq1 (1.4.15)
and choose the second basis vector as
q2 =
q′
2
kq′
2k2
(1.4.16)
The second basis vector is thus formed from a2, but with the component in the
direction of q1 subtracted out. Note that (1.4.15) can be written as
q′
2 = a2 − qH
1 a2q1 = a2 − q1 qH
1 q1
−1
qH
1
| {z }
Πq1
a2 = Π⊥
q1
a2 (1.4.17)
where we used the fact that kq1k = 1. Thus, the second vector is nothing but
the projection of a2 onto the space orthogonal to R(q1); as a result it must be
orthogonal to q1. Due to the normalization in (1.4.14) and (1.4.16), the two vectors
are orthonormal, i.e., they are orthogonal unit vectors. We proceed to construct
Q =

q1 q2

(1.4.18)
and form q3 to be orthonormal to the vectors in Q, i.e.,
q3 =
Π⊥
Qa3
Π⊥
Qa3
2
(1.4.19)
By adding q3 to Q and repeating the projections, we can in this manner construct
the orthonormal basis. Written in matrix form, this yields
A = Q R =

q1 . . . qn






hq1, a1i hq1, a2i . . . hq1, ani
hq2, a2i . . . hq2, ani
...
.
.
.
0 hqn, ani





(1.4.20)

where the m × n matrix Q forms the orthonormal basis, and R is upper triangular
and invertible. This is the famous Q-R factorization. Note that A does not need
to be a square matrix to be written as in (1.4.20); however, if A is a square matrix,
then Q is a unitary matrix. In general:
If A ∈ Cm×n
, with m ≥ n, there is a matrix Q ∈ Cm×n
with
orthonormal columns and an upper triangular matrix R ∈ Cn×n
such that A = QR. If m = n, Q is unitary; if in addition A is
nonsingular, then R may be chosen so that all its diagonal entries
are positive. In this case, the factors Q and R are unique.
The Q-R factorization concludes our discussion on orthogonality, and we proceed
to dwell on the notion of determinants.
1.5 Determinants
As we have seen above, we often wish to know when a square matrix is invertible14
.
The simple answer is:
A matrix A is invertible, when the determinant of the matrix is
non-zero, i.e., det{A} 6= 0.
That is a good and short answer, but one that immediately raises another question.
What is then the determinant? This is less clear cut to answer, being somewhat
elusive. You might say that the determinant is a number that capture the soul of
a matrix, but for most this is a bit vague. Let us instead look at the definition of
the determinant; for a 2 by 2 matrix it is simple:
det

a11 a12
a21 a22

=
a11 a12
a21 a22
= a11a22 − a12a21 (1.5.1)
For a larger matrix, it is more of a mess. Then,
det A = ai1Ai1 + · · · + ainAin, (1.5.2)
where the cofactor Aij is
Aij = (−1)i+j
det Mij (1.5.3)
and the matrix Mij is formed by deleting row i and column j of A (see [1] for further
examples of the use of (1.5.2) and (1.5.3)). Often, one can use the properties of the
determinant to significantly simplify the calculations. The main properties are
(i) The determinant depends linearly on the first row.
a + a′
b + b′
c d
=
a b
c d
+
a′
b′
c d
(1.5.4)
14We note in passing that the inverse does not exist for non-square matrices; however, such
matrices might still have a left or a right inverse.

Section 1.6. Eigendecomposition 15
(ii) The determinant change sign when two rows are exchanged.
a b
c d
= −
c d
a b
(1.5.5)
(iii) det I = 1.
Given these three basic properties, one can derive numerous determinant rules. We
will here only mention a few convenient rules, leaving the proofs to the interested
reader (see, e.g., [1,6]).
(iv) If two rows are equal, then det A = 0.
(v) If A has a zero row, then det A = 0.
(vi) Subtracting a multiple of one row from another leaves the det A unchanged.
(vii) If A is triangular, then det A is equal to the product of the main diagonal.
(viii) If A is singular, det A = 0.
(ix) det AB = (det A)(det B)
(x) det A = det AT
.
(xi) |cA| = cn
|A|, for c ∈ C.
(xii) det A−1

= (det A)
−1
.
(xiii) |I − AB| = |I − BA|.
where A, B ∈ Cn×n
. Noteworthy is that the determinant of a singular matrix is
zero. Thus, as long as the determinant of A is non-zero, the matrix A will be
non-singular and invertible. We will make immediate use of this fact.
1.6 Eigendecomposition
We now move to study eigenvalues and eigenvectors. Consider a square matrix A.
By definition, x is an eigenvector of A if it satisfies
Ax = λx. (1.6.1)
That is, if multiplied with A, it will only yield a scaled version of itself. Note that
(1.6.1) implies that
(A − λI) x = 0, (1.6.2)
requiring x to be in N(A−λI). We require that the eigenvector x must be non-zero,
and as a result, A − λI must be singular15
, i.e., the scaling λ is an eigenvalue of A
if and only if
det (A − λI) = 0. (1.6.3)
15Otherwise, the nullspace would not contain any vectors, leaving only x = 0 as a solution.

The above equation is termed the characteristic equation, and also points to how
one should go about computing the eigenvalues of a matrix. For example, let
A =

4 −5
2 −3

(1.6.4)
and thus,
det (A − λI) =
4 − λ −5
2 −3 − λ
= (4 − λ)(−3 − λ) + 10 = 0, (1.6.5)
or λ = −1 or 2. By inserting the eigenvalues into (1.6.1), we obtain
x1 =

1
1

and x2 =

5
2

(1.6.6)
The eigendecomposition is a very powerful tool in data processing, and is frequently
used in a variety of algorithms. One reason for this is that:
If A has n distinct eigenvalues, then A has n linearly independent
eigenvectors, and is diagonalizable.
If A has n linearly independent eigenvectors, then if placed as columns of a matrix
S, we can diagonalize A as
S−1
AS = Λ
△
=



λ1
...
λn


 ⇔ A = SΛS−1
. (1.6.7)
That is, A is similar to a diagonal matrix. In general, we say a matrix B is similar
to a matrix A is there exists a nonsingular matrix S such that
B = SAS−1
. (1.6.8)
The transformation A → SAS−1
is often called a similarity transformation by the
similarity matrix S. We note that:
If A and B are similar, then they have the same eigenvalues.
To illustrate the importance of (1.6.7), we note as an example that it immediately
implies that
Ak
= SΛk
S−1
. (1.6.9)
Check this for A2
using (1.6.4). Similarly,
A−1
= SΛ−1
S−1
, (1.6.10)

Section 1.6. Eigendecomposition 17
with Λ−1
having

λ−1
k as diagonal elements. Check! We proceed with some obser-
vations; the eigenvalues of AT
are the same as those of A, whereas the eigenvalues
of AH
are the complex conjugate of those of A. Further,
n
Y
k=1
λk = det A (1.6.11)
n
X
k=1
λk = tr A (1.6.12)
where tr A denotes the trace16
of A, i.e, the sum of the diagonal elements of A.
Check this for the example in (1.6.4). Note that eigenvalues can be complex-valued,
and can be zero. As mentioned, we are often encountering Hermitian matrices. In
this case:
A Hermitian matrix has real-valued eigenvalues and eigenvectors
forming an orthonormal set.
Thus, for a Hermitian matrix, (1.6.7) can be written as
A = UΛUH
=
n
X
k=1
λkxkxH
k (1.6.13)
where U is a unitary matrix whose columns are the eigenvectors of A. Equation
(1.6.13) is often referred to as the spectral theorem. Note that xkxH
k is nothing but
the projection onto the kth eigenvector, i.e.,
Πxk
= xkxH
k . (1.6.14)
Using (1.6.14), one can therefore view the spectral theorem as
A =
n
X
k=1
λkΠxk
(1.6.15)
The eigenvectors form a basis spanning R(A). The spectral theorem is closely re-
lated to the Karhunen-Loève transform, briefly described in Appendix B.
We proceed by briefly discussing quadratic minimization; this problem is so fre-
quently reoccurring (recall for instance the least squares problem in (1.4.4)) that it
deserves a moment of our attention. As an example, consider the minimization of
the function
f = 2x2
1 − 2x1x2 + 2x2
2 − 2x2x3 + 2x2
3 (1.6.16)
16An often used result is the fact that tr {AB} = tr {BA}.

Finding the minimum is easily done by setting the partial derivatives to zero, yield-
ing x1 = x2 = x3 = 0. But which solution different from the zero solution minimizes
(1.6.16)? Rewriting (1.6.16) as
f =

x1 x2 x3



2 −1 0
−1 2 −1
0 −1 2




x1
x2
x3

 = xT
Ax (1.6.17)
we conclude that it must be A that determines the minimum of f. Let
λ1 ≥ λ2 ≥ . . . ≥ λn
and define the Rayleigh’s quotient
R(x) =
xH
Ax
xHx
(1.6.18)
Then, according to the Rayleigh-Ritz theorem, it holds that
λ1 ≥
xH
Ax
xHx
≥ λn (1.6.19)
and that R(x) is minimized by the eigenvector corresponding to the minimum eigen-
value, λn. Similarly, R(x) is maximized by the dominant eigenvector, corresponding
to the largest eigenvalue, λ1. For our example, the eigenvalues of A are λ1 = 3.41,
λ2 = 2 and λ3 = 0.58. The minimum of f, excluding the zero solution, is thus 0.58,
obtained for
x1 =

0.5 0.7071 0.5
T
Concluding this section, we introduce the notion of positive definiteness.
The Hermitian matrix A is said to be positive definite if and only
if xH
Ax can be written as a sum of n independent squares, i.e.,
xH
Ax 0, ∀x (except x = 0)
If we also allow for equality, A is said to positive semi-definite. Note that
Axi = λixi ⇒ xH
i Axi = λixH
i xi = λ (1.6.20)
as the eigenvectors are orthonormal. We conclude that a positive definite matrix
has positive eigenvalues. The reverse is also true; if a matrix has only positive
eigenvalues, it is positive definite (the proof is left as an exercise for the reader).
The Hermitian matrix A is said to be positive definite if and only
if A has positive (real-valued) eigenvalues.
In passing, we note that if A is positive semi-definite, it implies that there exists a
matrix C (of rank n) such that A = CCH
(compare with the discussion at the end
of Section 1.2). For this case, we note that R(A) = R(C).

Section 1.7. Singular value decomposition 19
1.7 Singular value decomposition
As mentioned in previous section, the eigendecomposition in (1.6.7) is a very pow-
erful tool, finding numerous applications in all fields of signal processing. However,
the decomposition is limited by the fact that it only exists for square matrices. We
will now proceed to study the more general case.
Every matrix A ∈ Cm×n
can be factored as
A = UΣVH
where Σ ∈ Rm×n
is diagonal, and U ∈ Cm×m
, V ∈ Cn×n
are
unitary matrices.
This is the famous singular value decomposition17
(SVD), being one of the most
powerful tools in signal processing. The diagonal elements of Σ,
Σ =










σ1
...
σp
0
...
0










(1.7.1)
are denoted the singular values of A. Here, p = rank(A) ≤ min(m, n), and σk ≥ 0;
in general
σ1 ≥ σ2 ≥ . . . ≥ σp ≥ 0
We note that
AH
A = VΣH
ΣVH
⇒ AH
AV = VΣH
Σ (1.7.2)
which implies that σ2
i are the eigenvalues of AH
A (similarly, it is easy to show that
σ2
i are the eigenvalues of AAH
). In the special case when A is a square matrix, the
17The singular value decomposition has a rich history, dating back to the independent work by
the Italian differential geometer E. Beltrami in 1873 and the French algebraist C. Jordan in 1874.
Later, J. J. Sylvester in 1889, being unaware of the previous work by Beltrami and Jordan, gave
a third proof for the factorization. Neither of these used the term singular values, and their work
was repeatedly rediscovered up to 1939 when C. Eckart and G. Young gave a clear and complete
statement of the decomposition for a rectangular matrix. In parallel to the algebraists work, re-
searchers working with the theory of integral equations rediscovered the decomposition. In 1907,
E. Schmidt published a general theory for symmetric and nonsymmetric kernels, making use of
the decomposition. In 1919, É. Picard’s generalized this theory, coining Schmidt’s “eigenvalues”
as singular values (“valeurs singulières”). It took many years of rediscovering the decomposition,
under different names, until A. Horn in 1954 writing a paper on matrix theory used the term
“singular values” in the context of matrices, a designation that have since become standard ter-
minology (it should be noted that in the Russian literature, one also sees singular values referred
to as s-numbers).

singular values coincide with the modulus of the eigenvalues18
. Further, we note
that as only the first p = rank(A) singular values are non-zero, we can rewrite the
SVD as
A =

U1 U2

Σ1
Σ2

VH
1
VH
2

(1.7.3)
= U1Σ1VH
1 (1.7.4)
=
p
X
k=1
σkukvH
k (1.7.5)
where Σ1 ∈ Rp×p
, U1 ∈ Cm×p
, U2 ∈ Cm×(n−p)
, V1 ∈ Cn×p
, V2 ∈ Cn×(n−p)
, and
uk and vk denote the kth column of U1 and V1, respectively. As the rank of a
matrix is equal to the number of nonzero singular values, Σ2 = 0. Using (1.7.4),
we can conclude that
R(A) =

b ∈ Cm
: b = Ax
=

b ∈ Cm
: b = U1Σ1VH
1 x
=

b ∈ Cm
: b = U1y
= span (U1)
The dominant singular vectors forms a basis spanning R(A)! Using a similar tech-
nique, we find that
N(A) = span (V2)
R(AH
) = span (V1)
N(AH
) = span (U2)
which also yields that
ΠA = U1UH
1 (1.7.6)
Π⊥
A = I − U1UH
1 = U2UH
2 (1.7.7)
For a moment returning to the least squares minimization in (1.4.4),
min
x
kAx − bk2
2 = min
x
kUΣVH
x − bk2
2
= min
x
kΣVH
x − UH
bk2
2
= min
x
kΣy − b̂k2
2 (1.7.8)
where we defined y = VH
x and b̂ = UH
b. Minimizing (1.7.8) yields
y = Σ†
b̂ ⇒ x = VΣ†
UH
b (1.7.9)
18Due to the numerical robustness of the SVD algorithm, it is in practice often preferable to use
the SVD instead of the eigenvalue decomposition to compute the eigenvalues.

Section 1.8. Total least squares 21
Comparing with (1.4.6), yields
A†
= VΣ†
UH
(1.7.10)
and the conclusion that the least squares solution can immediately be obtained
using the SVD. We also note that the singular values can be used to compute the
2-norm of a matrix,
kAk2 = σ1. (1.7.11)
As is clear from the above brief discussion, the SVD is a very powerful tool finding
use in numerous applications. We refer the interested reader to [3] and [4] for further
details on the theoretical and representational aspects of the SVD.
1.8 Total least squares
As an example of the use of the SVD, we will briefly touch on total least squares
(TLS). As mentioned in our earlier discussion on the least squares minimization, it
can be viewed as finding the vector x such that
min
x
kAx − bk2
2 (1.8.1)
Here, we are implicitly assuming that the errors are confined to the “observation”
vector b, such that
Ax = b + ∆b (1.8.2)
It is often relevant to also assume the existence of errors in the “data” matrix, i.e.,
(A + ∆A)x = (b + ∆b) (1.8.3)
Thus, we seek the vector x, such that (b + ∆b) ∈ R(A+∆A), which minimize the
norm of the perturbations. Geometrically, this can be viewed as finding the vector
x minimizing the (squared) total distance, as compared to finding the minimum
(squared) vertical distance; this is illustrated in the figure below for a line.
✲
✻
⋄
⋄
⋄
✲
✻
⋄
⋄
⋄
(a) LS: Minimize vertical (b) TLS: Minimize total
distance to line. distance to line.

Let [A | B] denote the matrix obtained by concatenating the matrix A with the
matrix B, stacking the columns of B to the right of the columns of A. Note that,
0 =

A + ∆A | b + ∆b

x
−1

=

A | b

x
−1

+

∆A | ∆b

x
−1

(1.8.4)
and let
C =

A | b

D =

∆A | ∆b

(1.8.5)
Thus, in order for (1.8.4) to have a solution, the augmented vector [xT
, −1]T
must
lie in the nullspace of C + D, and in order for the solution to be non-trivial, the
perturbation D must be such that C + D is rank deficient19
. The TLS solution
finds the D with smallest norm that makes C + D rank deficient; put differently,
the TLS solution is the vector minimizing
min
∆A,∆b,x
k [∆A ∆b] k2
F such that (A + ∆A)x = b + ∆b (1.8.6)
where
kAk2
F = tr(AH
A) =
min (m,n)
X
k=1
σ2
k (1.8.7)
denotes the Frobenius norm, with σk being the kth singular value of A. Let C ∈
Cm×(n+1)
. Recalling (1.7.5), we write
C =
n+1
X
k=1
σkukvH
k (1.8.8)
The matrix of rank n closest to C can be shown to be the matrix formed using only
the n dominant singular values of C, i.e.,
C̃ =
n
X
k=1
σkukvH
k (1.8.9)
Thus, to ensure that C + D is rank deficient, we let
D = −σn+1un+1vH
n+1 (1.8.10)
19Recall that if C + D is full rank, the nullspace only contains the null vector.

Section 1.8. Total least squares 23
For this choice of D, C + D does not contain the vector20
vn+1, and the solution
to (1.8.4) must thus be a multiple α of vn+1, i.e.,

x
−1

= αvn+1 (1.8.11)
yielding
xT LS = −
vn+1(1 : n)
vn+1(n + 1)
(1.8.12)
We note in passing that one can similarly write the solution to the TLS matrix
equation (A + ∆A)X = (B + ∆B), where B ∈ Cm×k
, as
XT LS = −V12V−1
22 (1.8.13)
if V−1
22 exists, where
V =

V11 V12
V21 V22

(1.8.14)
with V11 ∈ Cn×n
, V22 ∈ Ck×k
and V12, VT
21 ∈ Cn×k
. The TLS estimate yields a
strongly consistent estimate of the true solution when the errors in the concatenated
matrix [A b] are rowwise independently and identically distributed with zero mean
and equal variance. If, in addition, the errors are normally distributed, the TLS
solution has maximum likelihood properties. We refer the interested reader to [4]
and [5] for further details on theoretical and implementational aspects of the TLS.
20The vector vn+1 now lies in N (C + D).

Appendix A
BANACH AND HILBERT
SPACES
Mister Data, there is a subspace communication for you
Star Trek, the Next Generation
In this appendix, we will provide some further details on the notion of vector spaces,
and in particular on the definition of Banach and Hilbert spaces, the latter being
the vector space of primary interest in signal processing.
A vector space is by definition a set of vectors X for which any vectors in the set,
say x1, x2 and x3, will satisfy that x1 + x2 ∈ X and αx1 ∈ X for some (possibly
complex-valued) constant α, as well as the vector space axioms:
(1) x1 + x2 = x2 + x1 (commutability of addition),
(2) x1 + (x2 + x3) = (x1 + x2) + x3 (associativity of addition),
(3) ∃0 ∈ X : 0 + x1 = x1 (existence of null element),
(4) ∀x1 ∈ X ⇒ ∃ − x1 ∈ X : x1 + (−x1) = 0 (existence of the inverse element),
(5) 1 · x1 = x1 (unitarism),
(6) α(βx1) = (αβ)x1 (associativity with respect to number multiplication),
(7) α(x1 + x2) = αx1 + αx2 (distributivity with respect to vector additism),
(8) (α + β)x1 = αx1 + βx1 (distributivity with respect to number additism).
Any element in a vector space is termed a vector. Often, we are primarily interested
in vector spaces that allow some form of distance measure, normally termed the
24

25
norm. Such a space is termed a Banach1
space. A Banach space is a complete2
vector space that admits a norm, kxk. A norm always satisfies
kxk ≥ 0 (A.1.1)
kλxk = |λ| kxk (A.1.2)
kx + yk ≤ kxk + kyk (A.1.3)
for any x, y in the Banach space and λ ∈ C. The so-called p-norm of a (complex-
valued) discrete sequence, x(k), is defined as
kxkp
△
=
∞
X
k=−∞
|x(k)|p
!1
p
, (A.1.4)
for any integer p 0. Similarly, for a (complex-valued) continuous function, f(t),
the p-norm is defined as
kfkp
△
=
Z ∞
−∞
|f(t)|p
dt
1
p
∞. (A.1.5)
Here, we are primarily concerned with the 2-norm, also called the Euclidean norm
as kx − yk2 yields the Euclidean distance between any two vectors, x, y ∈ CN
. For
instance, for
x =

1 3 2
T
(A.1.6)
y =

0 0 2
T
(A.1.7)
the 2-norms are kxk2 =
√
12 + 32 + 22 =
√
14 and kyk2 = 2. Similarly, the Eu-
clidean distance is kx − yk2 =
√
10. If not specified, we will here assume that
the used norm is the 2-norm. Note that the definitions allows for both infinite-
dimensional vectors and vectors defined as continuous functions.
As is obvious from the above discussion, a Banach space is quite general, and it
is often not enough to just require that a vector lies in a Banach space. For this
reason, we will further restrict our attention to work only with Banach spaces that
has an inner product (this to define angles and orthogonality). A Hilbert space is
1Stefan Banach (1892-1945) was born in Kraków, Austria-Hungary (now Poland). He received
his doctorate in mathematics at Lvov Technical University in 1920, and in 1924, he was promoted
to full professor. He used to spend his days in the cafés of Lvov, discussing mathematics with his
colleagues. He suffered hardship during the Nazi occupation and died of lung cancer shortly after
the liberation. His Ph.D. thesis is often said to mark the birth of modern functional analysis.
2A complete vector space is a vector space where every Cauchy sequence converges to an element
in the space. A sequence is a Cauchy sequence if for any ǫ 0, kx(n) − x(p)k ǫ for large enough
n and p.

26 Banach and Hilbert spaces Appendix A
a Banach space with an inner product, hx, yi. This inner product also defines the
length of a vector in the space,
kxk2 =
p
hx, xi. (A.1.8)
An example of a Hilbert space is l2(C), the space of “square summable sequences”,
i.e., the space of (complex-valued) sequences satisfying kxk2 ∞, where the inner
product is
hx, yi = xH
y =
∞
X
k=−∞
x∗
(k)y(k), (A.1.9)
and where (·)∗
denotes conjugate. Note that although every Hilbert space is also a
Banach space, the converse need not hold. Another example is L2(C), the space of
square integrable functions, i.e., (complex-valued) functions satisfying kfk2 ∞,
where the inner product is
hf, gi =
Z ∞
−∞
f∗
(x)g(x) dx. (A.1.10)
In words, this means simply that the inner-product is finite, i.e., any vector in a
Hilbert space (even if the vector is infinite-dimensional or is a continuous function)
has finite length. Obviously, this is for most practical purposes not a very strong
restriction.

Appendix B
THE KARHUNEN-LOÈVE
TRANSFORM
E ne l’idolo suo si trasmutava
Dante
In this appendix, we will briefly describe the Karhunen-Loève transform (KLT).
Among all linear transforms, the KLT is the one which best approximates a stochas-
tic process in the least square sense. Furthermore, the coefficients of the KLT are
uncorrelated. These properties make the KLT very interesting for many signal pro-
cessing applications, such as coding and pattern recognition.
Let
xN = [x1, . . . , xN ]
T
(B.1.1)
be a zero-mean, stationary and ergodic random process with covariance matrix (see,
e.g., [8]).
Rxx = E{xN xH
N } (B.1.2)
where E{·} denotes the expectation operator. The eigenvalue problem
Rxxuj = λjuj, for j = 0, . . . , N − 1, (B.1.3)
yields the j:th eigenvalue, λj, as well as the j:th eigenvector, uj. Given that Rxx is
a covariance matrix1
, all eigenvalues will be real and non-negative (see also Section
1.6). Furthermore, N orthogonal eigenvectors can always be found, i.e.,
U = [u1, . . . , uN ] (B.1.4)
will be a unitary matrix. The KLT of xN is found as
α = UH
xN (B.1.5)
1A covariance matrix is always positive (semi) definite.
27

28 The Karhunen-Loève transform Appendix B
where the KLT coefficients,
α = [α1, . . . , αN ]
T
, (B.1.6)
will be uncorrelated, i.e.,
E{αjαk} = λjδjk for j, k = 1, . . . , N. (B.1.7)
We note that this yields that
UH
RxxU = Λ ⇔ Rxx = UΛUH
(B.1.8)
where Λ = diag(λ1, . . . , λN ).

Appendix C
USEFUL FORMULAE
This one’s tricky. You have to use imaginary numbers, like eleventeen
Calvin Hobbes
Euler’s formula
eiωt
= cos(ωt) + i sin(ωt) (C.1.1)
Cauchy-Schwartz inequality
For all x, y ∈ Cn
, it hold that
hx, yi ≤
p
hx, xi
p
hy, yi (C.1.2)
with equality if and only if x and y are dependent. Similarly, for two continuous
functions, fx and gx, it holds that
Z ∞
−∞
|fx|2
dx ·
Z ∞
−∞
|gx|2
dx ≥
Z ∞
−∞
f∗
x gx dx
2
(C.1.3)
Some standard integrals
1
2π
Z π
−π
eiωt
dω = δt (C.1.4)
Z ∞
−∞
e−ax2
dx =
r
π
a
(C.1.5)
Z ∞
−∞
x2
e−ax2
dx =
1
2a
r
π
a
(C.1.6)
29

BIBLIOGRAPHY
[1] G. Strang, Linear algebra and its applications, 3rd Ed. Thomson Learning, 1988.
[2] R. A. Horn and C. A. Johnson, Matrix Analysis. Cambridge, England: Cambridge
University Press, 1985.
[3] R. A. Horn and C. A. Johnson, Topics in Matrix Analysis. Cambridge, England:
Cambridge University Press, 1991.
[4] G. H. Golub and C. F. V. Loan, Matrix Computations (3rd edition). The John
Hopkins University Press, 1996.
[5] S. V. Huffel and J. Vandevalle, The Total Least Squares Problem: Computational
Aspects and Analysis. SIAM Publications, Philadelphia, 1991.
[6] P. Stoica and R. Moses, Introduction to Spectral Analysis. Upper Saddle River,
N.J.: Prentice Hall, 1997.
[7] J. C. F. Gauss, Demonstratio Nova Theorematis Omnem Functionem Alge-
braicam Rationalem Integram Unius Variabilis in Factores Reales Primi Vel Secundi
Gradus Resolve Posse. PhD thesis, University of Helmstedt, Germany, 1799.
[8] M. H. Hayes, Statistical Digital Signal Processing and Modeling. New York: John
Wiley and Sons, Inc., 1996.
30

An Introduction to Linear Algebra.pdf

More Related Content

What's hot (20)

Similar to An Introduction to Linear Algebra.pdf (20)

More from Ashley Carter (20)

Recently uploaded (20)

An Introduction to Linear Algebra.pdf