TA session# 2
Jun Sakamoto
April 26,2016
Contents
1
Matrix differentiation 1
2
Rank 3
3
Inverse matrix 3
4
Eigenvalue and Eigenvector 3
1 Matrix differentiation
Proposition2.1
Let a and β are k×1 vector. Then,
∂a′β
∂β = a Proof
We define a=
a1
... ak
and β=
β1
... βk
. Then, a
′β= a1β1+ a2β2+ · · · + akβk.Thus,
∂a′β
∂β =
∂(a1β1+a2β2+···+akβk)
∂β1
...
∂(a1β1+a2β2+···+akβk)
∂βk
=
a1
... ak
= a
Q.E.D Proposition2.2
Let a and β are k×1 vector. Then,
∂β′a
∂β = a Proof
We define a=
a1
... ak
and β=
β1
... βk
. Then, β
′a= β
1a1+ β2a2+ · · · + βkak.Thus,
∂β′a
∂β =
∂(β1a1+β2a2+···+βkak)
∂β1
...
∂(β1a1+β2a2+···+βkak)
∂βk
=
a1
... ak
= a
Q.E.D Proposition2.3
Let A and β are a k×k matrix and a k×1 vector.Then,
∂β′Aβ
∂β = (A + A
′)β
Proof
We only consider the case β is 2×1 vector and A is 2×2 matrix. Then A =a c b d
and β=β1 β2
.Then, β′Aβ = aβ12+ bβ1β2+ cβ1β2+ dβ12.Thus,
∂β′Aβ
∂β =
∂(aβ12+bβ1β2+cβ1β2+dβ12)
∂β1
∂(aβ12+bβ1β2+cβ1β2+dβ12)
∂β2
!
=2aβ1+ bβ2+ cβ2 bβ1+ cβ1+ 2dβ2
=
2a b+ c b+ c 2d
β1 β2
=(a c b d
+a b c d
)β1
β2
=(A + A′)β
Q.E.D Especially, when A is symmetic,(i.e, A = A′)
∂β′Aβ
∂β = (2A)β
2 Rank
We consider a n×k matrix X. Matrix rank is defined as maximum nunber of linear independent column vectors or row vectors in the matrix.
*Example* A=1 2 3
2 4 5
Aconsists of column vectors–(1, 2)′ , (2, 4)′ and (3, 5)′–and consists of row vectors–(1, 2, 3) and (2, 4, 5).This matrix rank is 2 because column vector (1, 2)′ and (2, 4)′ are liner dependent. Then a number of linear independent column vectors is 2. Moreover row vectors (1, 2, 3) and (2, 4, 5) linear independent, so a number of linear independent column vectors is 2. Then rank(A)=2.
Definition2.1
(1)If rank(X)=min(n,k), then X is a full rank matrix.
(2)If X is a square(i.e, n=k) and a full rank, then X is regular matrix.
3 Inverse matrix
Definition2.2
Let matrix X and Y be a n×n matrix. If XY = Y X=In.T henmatrixXis a inverse matrix of Y .Moreover if matrix X has inverse matrix, then it is called to be regular.
Definition2.3
Matrix X has inverse matrix if and only if the determinant of X is not 0. Proposition 2.4
(1)If X is a regular matrix, then (X−1)−1=X.
(2)If X and Y be a regular matrix, XY , (XY )−1 are a regular matrix and Y−1X−1 is equal to a (XY )−1 Proof
Let X and Y be a regular matrix.
(1)Obvious by definition of inverse matrix.
(2)(XY )(Y−1X−1)=X(Y Y−1)X−1=XIX−1= XX−1=I.
Then XY and (Y−1X−1) are a regular matrix and XY (XY )−1=I. So Y−1X−1 is equal to a (XY )−1. Q.E.D
4 Eigenvalue and Eigenvector
Deffinition2.4
Let a matrix A be square. If a scalar λ and a vector x satisfy the following equation, then there are called eigenvalue and eigenvector respectively.
Ax=λx
Above equation can be written as
Ax− λx=0 (A − λI)x=0.
A condition with a solution which has non-trival solutin is det |A − λI|=0.
Moreover a solution of a above equation are eigenvalues and vice versa. Also eigenvalues(λ1, λ2, ...λn) correspond to eigenvectors(x1, x2, ..., xn).
*Example* Let a matrix A be
8 1 4 5
.
Then,
det|A − λI|=
8 − λ 1 4 5 − λ
= (8 − λ)(5 − λ) − 4=36 − 13λ + λ2= (λ − 9)(λ − 4)=0. Thus,λ1= 9, λ2= 4.
An eigenvector of correspond to λ1 is
−1 1
4 −4
x1
x2
=0
thus,x1= x2.Then, x1=t 1 1
. (t is a arbitrary constant) Also eigenvector of correspond to λ2is
4 1 4 1
x1
x2
=0
thus,4x1= x2.Then, x1=t1 4
Proposition2.5
(1)Eigenvalues and Eigenvectors of symmetric matrix are real value.
(2)An eigenvector xi corespond to an eigenvalue λiand an eigenvector xj corespond to an eigenvalue λj are orthgonal for all j.
Proof
(1)Let A be a real symmetric matrix.Let n denote the order of A. We take a solution of det|A − λI|=0.
We have an eigenvalue λc with eigenvector xc, perhaps both with complex entries. By definition of 2.4 Axc = λcxc.
Multiplying conjugate complex vector ¯x′c from the left, we have
¯ x′
cAxc = ¯x′cλcxc.
¯ x′
cAxc = λcx¯′cxc. Taking the complex conjecture transpose on the both sides,
¯
x′cAxc = ¯λcx¯′cxc. λcx¯′cxc = ¯λcx¯′cxc.
Therefore,
(λc− ¯λc)(¯x′cxc) = 0 Since,(¯x′cxc) 6= 0,
λc = ¯λc. Thus, λc is a real number.
Q.E.D (2)Let λ1and λ2 are eigenvalues and x1 and x2eigenvectors. Multiplying x1′ and x′2 to Ax2= λ2x2 and Ax1= λ1x1from the left.
x′1Ax2= λ2x′1x2 x′2Ax1= λ1x′2x1 Therefore,
λ2x′1x2= x′1Ax2= (x2′Ax1)′ = λ1(x′2x1)′ Hence,
(λ2− λ1)x′1x2= 0. Thus,if λ26= λ1, then x′1x2=0.
Q.E.D Definition2.5
Let a matrix A be square. If it satisfy the following equation, then it is called as idempotent matrix.
A2=A Proposition2.6
(1)Eigenvalue of idempotent matrix is 0 or 1.
(2)Rank of idempotent matrix is equivalent to trace of matrix. Proof
(1)Let A be a idempotent matrix. By definition of eigen value,
λx= Ax
= AAx
= λAx = λ2x. Therefore,
λx= λ2x (λ − λ2)x = 0 λ(1 − λ)x = 0. Since, x6= 0,λ is 1 or 0.
Q.E.D (2)We consider the case where A is nonnull. An order of idempotent matrix A is n×n, then there exist a n×r matrix B and r×n matrix L such that A = BL. Also rank(B)=rank(L)=r. We have that
BLBL= A(2) = A = BL = BIrL, since
BL= Ir. Thus, we find that
trace(A) = trace(BL) = trace(LB) = trace(Ir) = r
Q.E.D Definition 2.6
Let a matrix A be a n×n. If a matrix A can be diagonalized, that is, P−1AP = D,
where P and D are n×n regular matrix and n×n diagonal matrix, respectively. Proposition2.7
Let A has linear independent eigenvectors–x1, x2,...,xn. Also P consists of (x1, x2,...,xn). Then, P−1AP = D is diagonal matrix and nonzero elemnts are eigenvalue of A.
Proof
P−1AP = P−1A(x1, x2, ..., xn)
= (P−1Ax1, P−1Ax2, ..., P−1Axn)
= (λ1P−1x1, λ2P−1x2, ..., λnP−1xn) Also, P−1xi is ith column vector of P−1P. So P−1xi are unit vector ei
Then, (λ1P−1x1, λ2P−1x2, ..., λnP−1xn) = λ1e1, λ2e2, ..., λnen. Therefore,
P−1AP =
λ1 0
. ..
0 λn
Q.E.D
*Example*
A= 1 3
−2 −4
So,
λ1= −1, x1= t 3
−2
andλ2= −2, x2= t 1
−1
. Then a matrix P is
P = 3 1
−2 −1
, P−1= 1 1
−2 −3
Since
P−1AP =−1 0 0 −2