Matrix Analysis: Lecture 5

Feng Wei
daoshuo@bit.edu.cn

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .

(Beijing Institute of Technology) Matrix Analysis: Lecture 5 1 / 21

Finitely Generated Subspaces

Let V be a linear space over a field F and { α1 , · · · , αm } be a collection of

vectors in V . Then

W = { k1 α1 + k2 α2 + · · · + km αm | ki ∈ F, i = 1, 2, · · · , m }

is a subspace of V . W is called a finitely generated subspace of V .

{ α1 , · · · , αm } is called the generators of W . From the point of geometrical
view, we often say that the subspace W is spanned by { α1 , α2 , · · · , αm }. In
this case, W is denoted by

W = Span { α1 , α2 , · · · , αm }
= { k1 α1 + k2 α2 + · · · + km αm | ki ∈ F, i = 1, 2, · · · , m }

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .

(Beijing Institute of Technology) Matrix Analysis: Lecture 5 2 / 21

Finitely Generated Subspaces

All 2-dimensional planes and 3-dimensional spaces in Cartesian coordinate

system are finitely generated spaces. More generally, any n-dimensional
linear space can be considered as a finitely generated space.
A natural question towards to the spanned subspace W is to determine a
basis of W and its dimension. This example enable us to distinguish
collections of vectors from subspaces explicitly. It simultaneously provide us
with the essential connection between collections of vectors and subspaces.
For a collection {α1 , α2 , · · · , αm } of vectors, we do not pay more attention
on its linear operations. We are not sure of whether αi + α j or kαi lie in
this collection. But for the spanned subspace Span { α1 , α2 , · · · , αm }, we
require that it is closed under the addition and scalar multiplication
operations.

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .

(Beijing Institute of Technology) Matrix Analysis: Lecture 5 3 / 21

Finitely Generated Subspaces

The essential connection between collections of vectors and subspaces can

be stated as follows

Rank { α1 , α2 , · · · , αm } = Dim Span { α1 , α2 , · · · , αm }.

An arbitrary maximal linearly independent subset of {α1 , α2 , · · · , αm } can

be chosen to be a basis for Span { α1 , α2 , · · · , αm }.

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .

(Beijing Institute of Technology) Matrix Analysis: Lecture 5 4 / 21

Eigen-Subspaces

Let A ∈ Cn×n and λ be an eigenvalue of A. Then

Vλ = { α | Aα = λ α } ∪ { 0 }

is a subspace of Cn . We call Vλ the eigen-subspace associated with λ . We

define the dimension dimVλ to be the geometric multiplicity of λ . It was
well-known that the dimension dimVλ of Vλ is closely related to the
question of whether A is diagonalizable.

Let A ∈ Cn×n and λ be an arbitrary eigenvalue of A. A is diagonalizable if

and only if the algebraic multiplicity of λ is exactly its geometric
multiplicity.

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .

(Beijing Institute of Technology) Matrix Analysis: Lecture 5 5 / 21

Sums and Intersections of Subspaces

Let V be a linear space over a field F and V1 ,V2 be two subspaces of V .

We define the sum and intersection of V1 and V2 to be

V1 +V2 = { α + β | α ∈ V1 , β ∈ V2 }

and
V1 ∩V2 = { α | α ∈ V1 and α ∈ V2 },
respectively. It is straightforward to check that both of them are subspaces
of V .
Remark
The union V1 ∪V2 of V1 and V2 is not necessarily a subspace of V .

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .

(Beijing Institute of Technology) Matrix Analysis: Lecture 5 6 / 21

 z C1
D1

B1 C
D
A1

B
y
A

x
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .

(Beijing Institute of Technology) Matrix Analysis: Lecture 5 7 / 21

Dimension Formula

Theorem 1
Let V be a linear space over a field F and U,W be two finite-dimensional
subspaces of V . Then

dim (U +W ) + dim (U ∩W ) = dimU + dimW.

This is the so-called dimension formula for finite-dimensional spaces.

Proof: Since U and W are both finite-dimensional subspaces of V ,

dimU ∩W is finite-dimensional. Suppose that dimU = s, dimW = t and
dimU ∩W = r. Let us take a basis of U ∩W

α1 , α2 , · · · , αr

which can be extended to a basis for U and W . That is,

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .

(Beijing Institute of Technology) Matrix Analysis: Lecture 5 8 / 21

Dimension Formula

α1 , α2 , · · · , αr , β1 , · · · , βs−r
is a basis for U. Similarly,

α1 , α2 , · · · , αr , γ1 , · · · , γt−r

is a basis for W . Then we have

U = Span {α1 , · · · , αr , β1 , · · · , βs−r }
W = Span {α1 , · · · , αr , γ1 , · · · , γt−r }

This gives

U +W = Span {α1 , · · · , αr , β1 , · · · , βs−r , γ1 , · · · , γt−r }

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .

(Beijing Institute of Technology) Matrix Analysis: Lecture 5 9 / 21

Dimension Formula

It is enough for us to show that α1 , · · · , αr , β1 , · · · , βs−r , γ1 , · · · , γt−r are

linearly independent.
Suppose that

k1 α1 + · · · + kr αr + p1 β1 + · · · + ps−r βs−r + q1 γ1 + · · · + qt−r γt−r = 0.

This relation can be rewritten as

η = k1 α1 + · · · + kr αr + p1 β1 , + · · · + ps−r βs−r
= −q1 γ1 − · · · − qt−r γt−r

We therefore conclude that η ∈ U ∩W , which give rises to

η = l1 α 1 + · · · + r α r .
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .

(Beijing Institute of Technology) Matrix Analysis: Lecture 5 10 / 21

Dimension Formula

Thus we get
l1 α1 + · · · + r αr = −q1 γ1 − · · · − qt−r γt−r ,
which is equivalent to

l1 α1 + · · · + r αr + q1 γ1 + · · · + qt−r γt−r = 0.

This implies that

l1 = · · · = lr = q1 = · · · = qt−r = 0

and that
η = 0.

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .

(Beijing Institute of Technology) Matrix Analysis: Lecture 5 11 / 21

Dimension Formula

And hence

0 = η = k1 α1 + · · · + kr αr + p1 β1 + · · · + ps−r βs−r .

Since α1 , · · · , αr , β1 , · · · , βs−r are linearly independent, we obtain

k1 = · · · = kr = p1 = · · · = ps−r = 0.

Now we can say that α1 , · · · , αr , β1 , · · · , βs−r , γ1 , · · · , γt−r are linearly

independent.

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .

(Beijing Institute of Technology) Matrix Analysis: Lecture 5 12 / 21

Dimension Formula

Example 1
Let U be the subspace of upper triangular matrices in R2×2 and W be the
subspace of lower triangular matrices in R2×2 . That is,
{[ ] }
a b
U= a, b, c ∈ R .
0 c

and {[ ] }
d 0
W= d, e, f ∈ R .
e f

One can check the validity of dimension formula in an easy way.

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .

(Beijing Institute of Technology) Matrix Analysis: Lecture 5 13 / 21

Dimension Formula
In a 3-dimensional Cartesian coordinate system, please check the validity
of dimension formula.
Example 2
z

x
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .

(Beijing Institute of Technology) Matrix Analysis: Lecture 5 14 / 21

Direct Sums of Subspaces

One special case for the sums of subspaces is the so-called direct sum of
subspaces.
Let V be a linear space over the field F and S, T be two subspaces of V .
We say that U = S ⊕ T is the direct sum of S and T if
1 S ∩ T = { 0 }, and
2 S+T =U (in general, Si ∩ (∑i̸= j S j ) = { 0 } and ∑i Si = U).

The subspaces S and T are said to be complements of each other in U.

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .

(Beijing Institute of Technology) Matrix Analysis: Lecture 5 15 / 21

Direct Sums of Subspaces

Proposition 2
The following statements are equivalent for the direct sum U = S ⊕ T :
1 U = S ⊕ T is direct sum;
2 For any u ∈ U, u can be uniquely written in the form u = s + t with
s ∈ S and t ∈ T ;
3 dim (S + T ) = dim S + dim T ;
4 Suppose that {α1 , · · · , αk } is a basis of S and that {β1 , · · · , βl } is a
basis of T . Then
{α1 , · · · , αk , β1 , · · · , βl }
is a basis of T .

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .

(Beijing Institute of Technology) Matrix Analysis: Lecture 5 16 / 21

Direct Sums of Subspaces

Example 3
Let U be the subspace of symmetric matrices in Rn×n and W be the
subspace of skew-symmetric matrices in Rn×n . That is,

U = { AT = A, A ∈ Rn×n }.

and
W = { AT = −A, A ∈ Rn×n }.

Then Rn×n = U ⊕W .
Question 1
What are the dimensions of U and W in Examples 3 ? Please find their
bases.
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .

(Beijing Institute of Technology) Matrix Analysis: Lecture 5 17 / 21

Direct Sums of Subspaces

Example 4
Let U be the subspace of even functions in RR and W be the subspace of
odd functions in RR . That is,

U = { f (x) = f (−x), f (x) ∈ RR }.

and
W = { f (x) = − f (x), f (x) ∈ RR }.

Then RR = U ⊕W .

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .

(Beijing Institute of Technology) Matrix Analysis: Lecture 5 18 / 21

Direct Sums of Subspaces

Example 5
Let K n×n be the set of all n × n matrices over a number field K. Suppose
that K0n×n is the subset of all matrices with zero trace and that Span {I} is
the subset of all scalar matrices. It is not difficult to prove that K0n×n and
Span {I} are subspaces of K n×n . Moreover,

K n×n = K0n×n ⊕ Span {I}

For an arbitrary n × n matrix A,

 
a11 a12 · · · a1n
 a21 a22 · · · a2n  n
 
A= . .. . . . , Tr(A) = ∑ aii
 .. . . ..  i=1
an1 an2 · · · ann
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .

(Beijing Institute of Technology) Matrix Analysis: Lecture 5 19 / 21

Direct Sums of Subspaces
we have the following decomposition
 
a11 − Tr(A) a12 ··· a1n
 n

 a21 a22 − Tr(A) ··· a2n  Tr(A)
A=  n + I
.. .. .. ..  n
 . . . . 
an1 an2 ··· ann − Tr(A)
n
= B +C,

where
 
a11 − Tr(A) a12 ··· a1n
 n

 a21 a22 − Tr(A) ··· a2n 
B= n  , C = Tr(A) I
 .. .. .. ..  n
 . . . . 
an1 an2 ··· ann − Tr(A)
n
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .

(Beijing Institute of Technology) Matrix Analysis: Lecture 5 20 / 21

Direct Sums of Subspaces
Clearly, B ∈ K0n×n ,C ∈ Span {I}. This implies that

K n×n = K0n×n + Span {I}.

Furthermore,
K0n×n ∩ Span {I} = {0}.
We assert that
K n×n = K0n×n ⊕ Span {I}.

Question 2
What are the dimensions of K0n×n and Span {I} in Examples 5 ? Please
find a basis for K0n×n .

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .

(Beijing Institute of Technology) Matrix Analysis: Lecture 5 21 / 21