0% found this document useful (0 votes)
24 views41 pages

Elec9731 LM1

This document provides an overview of linear and robust control systems topics, including matrix methods. It discusses matrix and vector properties such as norms, linear dependence and rank, eigenvalues and eigenvectors, and partitioned matrix inversion. Eigenvalue decomposition and singular value decomposition are introduced. The summary discusses: 1. Key matrix properties including norms, linear dependence, rank, eigenvalues, eigenvectors, and inverses. 2. The four fundamental subspaces and the fundamental theorem of matrix algebra relating row and column rank. 3. Finding eigenvalues by solving the characteristic polynomial and the eigenvalue decomposition representing a matrix as a linear combination of eigenvectors weighted by eigenvalues. 4. The Cayley-Hamilton theorem relating a matrix to

Uploaded by

wyx840927627
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
24 views41 pages

Elec9731 LM1

This document provides an overview of linear and robust control systems topics, including matrix methods. It discusses matrix and vector properties such as norms, linear dependence and rank, eigenvalues and eigenvectors, and partitioned matrix inversion. Eigenvalue decomposition and singular value decomposition are introduced. The summary discusses: 1. Key matrix properties including norms, linear dependence, rank, eigenvalues, eigenvectors, and inverses. 2. The four fundamental subspaces and the fundamental theorem of matrix algebra relating row and column rank. 3. Finding eigenvalues by solving the characteristic polynomial and the eigenvalue decomposition representing a matrix as a linear combination of eigenvectors weighted by eigenvalues. 4. The Cayley-Hamilton theorem relating a matrix to

Uploaded by

wyx840927627
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 41

LINEAR AND ROBUST CONTROL SYSTEMS

ELEC 9731

1 - Matrix Methods Review

c Prof. Victor Solo


School of Electrical Engineering and
Telecommunications
UNSW
Sydney

Slide 1
Topics

1 Matrix and Vector - Basic Properties


2 Linear Dependence and Rank
3 Eigenvalues and Eigenvectors:
Eigenvalue Decomposition (EVD) and
Singular Value Decomposition (SVD)
4 Functions of Matrices
5 Partitioned Matrix Inversion
6 Generalised Inverses
7 Examples

References
G. Strang. Applied Linear Algebra.
T. Kailath. Linear Systems. Appendices.

Slide 2
(A) Matrix and Vector - Basic Properties
Matrix (Rectangular)
A = Am×n = [aij ]; m rows, n columns
Column vector: x = xn×1 = [xi ]
Adjoint: AT = ATn×m = [aji ]
Conformable product
P
C = AB ↔ cij = k
aik bkj
Inner product
T
Pn
a b= 1
ai bi
Outer product (dyadic product)
cm×1 dpT×1 = [ci dj ]m×n
Positive Definite Matrix
xT Ax > 0, x 6= 0
Positive Semi-Definite Matrix
xT Ax ≥ 0
Symmetric Matrix
A = AT
Hermetian transpose = complex conjugate transpose
AH = (A∗ )T

Slide 3
(A) Matrix and Vector - Basic Properties -
Continued
Norms
v
u n
√ uX
kxk = k x k2 = xT x = t x2i
1
n
X p p1
k x kp = ( |xi | )
1

Angle between vectors


xT y
cos θx,y = kxkkyk

Orthogonal vectors
Π
θx,y = 2 or xT y = 0

Slide 4
(B) Linear Dependence and Rank
[B1] Linear dependence and vector spaces
A set of m-vectors c1 , · · · , cn is said to be linearly
independent if there is no linear combination
Pn
1 civi that is zero.
The vector space spanned by c1 , · · · , cn is the set
of vectors obtainable from c1 , · · · , cn by addition
and scalar multiplication.
Collect c1 , · · · , cn into a matrix An×m . Then note
that we can express A in terms of its column
vectors or row vectors
 
T
r1
 
 : 
Am×n = [c1 , · · · , cn ] = 
 

 : 
 
T
rm

Slide 5
(B) Linear Dependence and Rank -
Continued
Forming linear combinations of columns of A is
equivalent to multiplying A into a vector since
 
v1
  n
 :  X
Av = [c1 , · · · , cn ]  = ci vi
 
 : 
  1

vn

Similarly vector pre multiplication generates


linear conbinations of row vectors
 
r1T
  m
 :  X
wT A = [w1 , · · · , wm ]  = wj rjT
 
 : 
  1
T
rm

Slide 6
(B) Linear Dependence and Rank -
Continued II
[B2] The four fundamental sub spaces
(i) Column space
The column (vector) space of A is the vector space
spanned by the columns of A denoted cs(A). It is also
called the range of A and denoted R(A).
(ii) Row space
The row space is the space spanned by the rows.
Since (wT A)T = AT w it is also the column space of
AT and so denoted cs(AT ) or R(AT ).
(iii) Right Null space
or kernel of A is the space spanned by the null vectors
vr such that Avr = 0. It is denoted N (A) or ker(A).
(iv) Left Null space

or left kernel is the space spanned by the null vectors


wk such that wkT A = 0. Since (wkT A)T = AT wk this is
also the right null space of AT and denoted N (AT ).

Slide 7
(B) Linear Dependence and Rank -
Continued III
[B3] Rank & Nullity

r = (column) rank ofA


= dim. column space ofA
= dim.cs(A) = dim.R(A)
ν = (column) nullity ofA
= dim. right null space ofA
= dim.N (A)

Clearly r = # linearly independent columns


ν = # dependent columns
So r + ν = n
Similarly
rT = row rank of A = dim R(AT )
νT = row nullity of A = dim N (AT )
and rT + νT = m

Slide 8
(B) Linear Dependence and Rank -
Continued IV
[B4] Fundamental Theorem of Matrix Algebra
We show every n-vector is in N (A) or R(AT )
If this holds then ν + rT = n but
ν + r = n ⇒ r = rT i.e. column rank=row rank
and this is the fundamental result.
The result is obvious once we write Av = 0 as
 
r1T
 :  v = 0 ↔ riT v = 0
 
 
T
rm
So every null vector is ⊥ to every row vector (and
linear combinations thereof) and any vector ⊥ to
every row vector is automatically a null vector.
Thus the rank of a matrix is its common row and
column rank.
If A is a square matrix we say it has full rank of
rank = order. It is then called non-singular.
Otherwise it is singular.

Slide 9
- Continued V
[B5] Determinants.
A has a determinant if f it is square. If a linear
combination of columns is added to another
column the determinant is unchanged. So if A is
singular we can make one of its columns all 0′ s.
Then it has zero determinant. So det A = |A| = 0
iff A is singular.
[B6] Inverse
A has a left inverse AL if AL A = I. A has a right
inverse AR if AAR = I
If A is square it has an inverse iff it has full rank.
[Suppose A−1 A = I but for some c 6= 0 ,
Ac = 0 ⇒ A−1 0 = c = 0 a contradiction ].
N.B. A−1 =adjoint(A)/detA
(ABC)−1 = C −1 B −1 A−1

Slide 10
(C) Eigenvalues and Eigenvectors
[C1] If An×n is a square matrix, a vector p
satisfying Ap = λp for some scalar λ is called a
right eigenvector of A and λ is called an
eigenvalue (eval).
[C2] Similarly a left eigenvector satisfies
′ ′
q T A = λ q T ↔ AT q = λ q

and λ is another eigenvalue.
[C3] Left and right eigenvectors are orthogonal for
distinct eigenvalues.
Proof: Let q be a left eigenvector with eval λ.

Let p be a right evec with eval λ

⇒ q T Ap = λq T p

= λ qT p

T
⇒q p = 0 ifλ 6= λ

Slide 11
(C) Eigenvalues and Eigenvectors -
Continued
[C4] Finding eigenvalues
If A is square we must solve (A − λI)p = 0 to find
λ. This means A − λI is singular and this implies
|A − λI| = 0
Now this equation is an nth degree polynomial in
λ and so has n roots which are the eigenvalues.
Thus there may be repeated eigenvalues. This
polynomial is the characteristic polynomial of A.
[C5] (i) Neither p nor q need be unit vectors when
pT q = 0

(ii) If A is symmetric the evals are real since


Ap = λp ⇒ Ap∗ = λ∗ p∗
⇒ pH Ap = λpH p
also pH Ap = λ∗ pH p
⇒λ = λ∗

Slide 12
(C) Eigenvalues and Eigenvectors -
Continued II
[C6] Eigenvalue Decomposition (EVD)
(also called spectral decomposition).
Suppose the evals are distinct : λ1 , · · · , λn
Let P = (p1 , · · · , pn ), Q = (q1 , · · · , qn ) be the
corresponding right and left evecs. These are
mutually orthogonal so
QT P = [qiT pj ] = [δij ] = I
⇒ QT = P −1
(The pi are linearly independent so p has full rank
and so has an inverse). Continuing
⇒ P QT = I
 
q1T
 
⇒ (p1 , · · · , pn )  : 

 = I
qnT
n
X
⇒ pi qiT = I
1

Slide 13
(C) Eigenvalues and Eigenvectors -
Continued III
From this we can get the EVD.

Api = λi pi , i = 1, · · · , n
⇒ Api qiT = λi pi qiT , i = 1, · · · , n
n
X Xn
⇒A pi qiT = λi pi qiT
1 1
n
X
⇒ A = P ΛQT = λi pi qiT
1

Slide 14
(C) Eigenvalues and Eigenvectors -
Continued IV
[C7] Cayley Hamilton Theorem
From the EVD
A2 = P ΛQT P ΛQT = P Λ2 QT
induction ⇒ Am = P Λm QT
Pn
Now suppose |A − λI| = 1 au λu
Pn
Cayley Hamilton Theorem: Then 1 au Au = 0

Slide 15
(C) Eigenvalues and Eigenvectors -
Continued V
Proof:
n
X n
X
au Au = au P Λu QT
1 1
n
X
= P( au Λu )QT
1
n
X n
X
= P diag( au λu1 , · · · , au λun )QT
1 1

= P diag(0, · · · , 0)QT
= 0

NB
(i) If A is symmetric then left evecs=right evecs
so Q = P .
(ii) Again note very carefully that in the EVD
even if pi have unit length qi will not (and vice
versa). We have only pTi qi = 1.

Slide 16
(C) Eigenvalues and Eigenvectors - SVD
[C8] Singular Value Decomposition (SVD)
Am×n is rectangular m > n
So rank (A) ≤ n
Then AATm×m , AT An×n are square and AAT has
rank ≤ n , so it has at least m − n null vectors
⇒ m − n zero evals. Next
λI − AAT = |λI − AT A|
so the remaining n evals of AAT are the same as
those of AT A.
Suppose these n evals are distinct. Since AT A is
positive semidefinite these are non negative so we
denote them ρ21 , · · · , ρ2n .

Slide 17
(C) Eigenvalues and Eigenvectors - SVD
Continued II
Let u1 , · · · , un be the corresponding evecs. They
are orthogonal (we make them unit length),

AT Aui = ρ2i ui , i = 1, · · · , n (0.1)


let vi = Aui /ρi , (soviT vi = 1) (0.2)
⇒ AT vi = AT Aui /ρi = ρi ui (0.3)

Then consider that further

AAT vi = Aui ρi = ρ2i vi (0.4)

So vi are evecs of AAT with evals ρ2i and they are


mutually orthogonal.
We can now establish the
singular value decomposition of A
n
X
SVD: A = ρi vi uTi = V ΛU T (0.5)
1

(not to be confused with the EV D : they agree


when A is symmetric.)

Slide 18
(C) Eigenvalues and Eigenvectors - SVD
Continued III
Consider the n × n matrix
n
X n
X
(A − ρi vi uTi )T (A − ρi vi uTi )
1 1
n
X n
X
= (AT − ρi ui viT )(A − ρi vi uTi )
1 1
Xn
= AT A − ρ2i ui uTi
1
n
X n
X
− ρ2i ui uTi + ρ2i ui uTi
1 1
n
X
= AT A − ρ2i ui uTi
1
= 0 by EVD

This establishes the result.

Slide 19
(C) Eigenvalues and Eigenvectors - SVD
Continued IV
To summarise
(i) (0.2) ⇒ Aui = ρi vi ↔ AT Aui = ρ2i ui
(ii) (0.4) ⇒ AT vi = ρi ui ↔ AAT vi = ρ2i vi
compare to EVD
T
Pn
Api = λi pi , A qi = λi qi , A = 1 λi pi qiT
when A is symmetric vi = ui and ρi = λi

Slide 20
(C) Eigenvalues and Eigenvectors - SVD
Continued V
(iii) SVD tells about near singularity. Suppose
one singular value ρ0 is small. Then consider the
relation:

Au0 = ρ0 v0
 
u01
 
↔ (c1 , · · · , cn ) 
 : 
 = ρ0 v0
uon
n
X
⇒ ci uoi = ρ0 v0
1

Now v0 is a unit vector so this equation gives a


linear combination of columns of A that is nearly
0.
Similarly the equation AT v0 = ρ0 u0 gives a linear
combination of rows of A that is nearly zero.

Slide 21
D. Matrix Functions
We already defined f (A) = P f (Λ)QT
[D1] Matrix exponential can then be defined in
several ways. With distinct eigenvalues

eAt = P eΛt QT = P diag(eλi t )QT



X (λi t)r T
= P diag( )Q
0
r!
∞ r
X t
= P diag(λri )QT
0
r!
∞ r
X t r
= A
0
r!
 
0 1
e.g. A =   ⇒ A2 = 0 ⇒ Am = 0, m ≥ 2
0 0
 
At
1 t
⇒e = 
0 1
If A has repeated ergenvalues Am = P J m QT etc..
J is the Jordan form.

Slide 22
[D2] Matrix Calculus
(a) Matrices
If A(t) = [aij (t)] can then easily show
d
dt A(t)B(t) = A(t) dB dA
dt + B(t) dt
d −1
dt A = −A−1 dA
dt A−1

Similarly
Z ∞
L(A(t)) = A(s) = A(t)e−st dt
0
Z ∞
= [ aij (t)e−st dt]
0
= [aij (s)]

e.g. L(eAt ) = (sI − A)−1

Slide 23
[D2] Matrix Calculus - Continued
(b) Vectors
If φ = f (x) = f (x1 , · · · , xn )
 T
dφ ∂f ∂f
Then dx = ∂x1 , · · · , ∂xn
If φ=f (x) , then
 
∂φ1 ∂φ2 ∂φm
∂x1 ∂x1 ··· ∂x1
 

 : 

dφ T  
dx =
 : 

 

 : 

∂φ1 ∂φn
∂xn ··· ··· ∂xn

Then can show


d T
c x = c
dx
d
(Ax)T = AT
dx
d T
x Ax = Ax + AT x = 2Ax(Asymmetric)
dx

Slide 24
[D2] Matrix Calculus - Continued II
[D3] Matrix Integration
RT
0
A(s)ds = [sT0 aij (s)ds]
[D4] Signal Norms
Z ∞
1
k s k2 = ( s2 (t)dt) 2
−∞
Z ∞
p
k s kp = ( |s(t)| dt)1/p ↑k s k∞ = sup |s(t)|
−∞ t

Slide 25
[E] MATRIX INVERSION
[E1] Partitioned Matrix Inversion
If A,B are square matrices and the indicated
inverses exist, then
 −1
A D
  =
B C
 
A−1 + A−1 D∆c−1 BA−1 −A−1 D∆c−1
 
−1 −1 −1
−∆c BA ∆c

where ∆c = C − BA−1 D
The result is easily checked by direct
multiplication. In a similar way show
 −1
A D
  =
B C
 
−1
∆A −∆−1A DC −1
 
−C −1 B∆−1 A C −1
+ C −1
B∆ −1 −1
A C

where ∆A = A − DC −1 B

Slide 26
[E] MATRIX INVERSION - Continued
[E2] Matrix Inversion Lemma
Equating the two inverses in section E1 gives
(A − DC −1 B)−1 = A−1 + A−1 D∆c−1 BA−1
or in a more usual notation (change −C −1 with C
and interchange B and D)
(A + BCD)−1 = A−1 − A−1 B∆−1 DA−1
∆ = C −1 + DA−1 B

Slide 27
[E] MATRIX INVERSION - Continued II
[E3] Determinant Identities
Expand in rows to see that
   
A D A 0
det = det A det C = det
0 C B C

Now use the identity


    
−1
A D A 0 I A D
=
B C B I 0 ∆c

to find that (with ∆c as given in section E1)


 
A D
det = det A det ∆c
B C

Similarly show that this also is


 
A D
det = det C det ∆A
B C

Now set A = I m , C = I n to see from the last two identities


that

det I n − BD = det(I m − DB)

In particular, if b, d are vectors,

det(A + bdT ) = det[(A)(I + A−1 bdT )]


= det(A) det(I + A−1 bdT )
= (1 + dT A−1 b) det A

Slide 28
F. Generalised Inverses
[Fi)] We say G−m×p is a generalised inverse
(g-inverse) of Gp×m if GG− G = G
T − baT
Example G = ab , then G = aT abT b

Proof:
− T baT T
GG G = ab aT abT b ab = abT = G
[F(ii)] solving Gx = b when G is singular.
Suppose G has rank r and let φ1 , · · · , φk be
linearly independent null vectors i.e.
Gφt = 0, t = 1, · · · , k
If G− is any g-inverse of G, then a general
solution to Gx = b is (assuming there is one)

Pk
x = G b + 1 αr φr
where αr are free parameters.

Slide 29
F. Generalised Inverses - Continued
Proof. Let x0 be a solution so that Gx0 = b , then
Gx = GG− b + 0 = GG− Gx0 = Gx0 = b
Let x1 be another solution so that Gx1 = b, then

G(x1 − x0 ) = b − b = 0
k
X ′
⇒ x1 − x0 = θr φr for some θr s
1
k
X
⇒ x1 = x0 + θr φ r
1
k
X
= G− b + αr φr
1

recovering the general form.

Slide 30
F. Generalised Inverses - Continued II
[F(iii)] Moore - Penrose g-inverse
Let G have SVD
G U Λ VT
=
p×m p×n n×n n×m

n =rankG ≤ min(p, m). Note that


U T U = In = V T V
Then a g-inverse is
G+ V Λ−1 UT
=
mx×p m×n n×n n×p

It is called the Moore-Penrose g-inverse


Proof

GG+ G = U ΛV T (V Λ−1 U T )U ΛV T
= U ΛΛ−1 ΛV T
= U ΛV T = G

Slide 31
F. Generalised Inverses - Continued III
[F(iv)] Moore-Penrose Properties

G+ G = V Λ−1 U T U ΛV T
= V Λ−1 ΛV T
T
= V Vm×m (6= Im )
GG+ = U ΛV T V Λ−1 U T
= U ΛΛ−1 U T
T
= U Up×p (6= Ip )

[F(v)] Projection g-inverse


Suppose p > m; if G has rank m then a g-inverse
is

G− = (GT G)−1 GT

Proof: GG− G = G(GT G)−1 GT G = G

Slide 32
F. Generalised Inverses - Continued IV
[F(vi)] Projection Matrices
Suppose Xn×r has rank r(< n). Find
X⊥ (n − r) × n of rank n − r , orthogonal to X
T
i.e. X⊥ X=0
i.e. all columns of X⊥ are orthogonal to all
columns of X and W = [X, X⊥ ] has full rank.
Projection Lemma:
In×n = X(X T X)−1 X T + X⊥ (X⊥
T
X⊥ )−1 X⊥
T

Proof: Let
 
(X T X)−1 XT
M =  
T
(X⊥ X⊥ )−1 T
X⊥
 
(X T X)−1 X T
⇒ MW =   (X, X⊥ )
T
(X⊥ X⊥ )−1 X⊥
T

 
Ir×r 0
=  
0 I(n−r)×(n−r)

Slide 33
F. Generalised Inverses - Continued V

= In×n
⇒W = M −1
⇒ WM = I
 
(X T X)−1 X T
= (X, X⊥ )  
T
(X⊥ X⊥ )−1 X⊥
T

= X(X T X)−1 X T + X⊥ (X⊥


T
X ⊥ )− X ⊥
T

as required
N.B. PX = X(X T X)−1 X T is a projection matrix
e.g. if

ξ = Xa + X⊥ b
⇒ PX ξ = X(X T X)−1 X T a + 0
= Xa
∈ column space of X

Slide 34
F. Generalised Inverses - Continued VI
[F(vii)] Miscellaneous
(a) Column interchange is represented by post
multiplication by a permutation matrix
    
a1 b1 0 1 b1 a1
e.g.     =  
a2 b2 1 0 b2 a2
(ii) If the cols of Y are linearly dependent on the
cols of X then Y = XA for some matrix A
Proof. If v is linearly dependent on ξ1 , · · · , ξr ,
Pr
the columns of X then v = 1 αr ξr , for some
coefficients αr
 
α1
 
⇒ v = (ξ1 , · · · , ξr ) 
 :  = Xα

αr
So there are vectors a1 , · · · , am with
Y = (y1 , · · · , ym )
= (Xa1 , · · · , Xam )
= X(a1 , · · · , .am ) = XA

Slide 35
G. Examples
[G(i)] Linear Combination of Columns
    
3 1 2 1 0
    
A =  2 2 1  1  =  0 
    

1 3 2 −2 0
     
3 1 2
     
=1× 2 +1× 2 −2×
   
 1


1 3 2

[G(ii)] 2x2 Matrix Inverse


 −1  
a b 1  d −b
  = 
c d ∆ −c a
∆ = det = ad − bc

[G(iii)] Evals and Evecs


 
2 2
A= 
3 1

Slide 36
G. EXAMPLES - Continued
For evals solve

λ−2 −2
|λI − A| =
−3 λ−1
= (λ − 2)(λ − 1) − 6
= λ2 − 3λ + 2 − 6
= λ2 − 3λ − 4
= (λ − 4)(λ + 1)
⇒ Roots are λ1 , λ2 = 4, −1
For right evecs solve: Api = λi pi i = 1, 2
For λ1
    
(1) (1)
2 2 p1 p1
   = 4 
(1) (1)
3 1 p2 p2
(1) (1) (1)
⇒ 2p1 + 2p2 = 4p1
(1) (1) (1)
3p1 + p2 = 4p2
(1) (1) (1) (1)
⇒ 2p2 = 2p1 ⇒ p2 = p1
Second equation agrees.

Slide 37
G. EXAMPLES - Continued II
We may take
   
(1)
p1 1
 = 
(1)
p2 1

For λ2
    
(2) (2)
2 2 p1 p1
   = −1 ×  
(2) (2)
3 1 p2 p2
(2) (2) (2)
⇒ 2p1 + 2p2 = −p1
(2) (2) (2)
⇒ 3p1 + p2 = −p2
(2) (2)
⇒ 2p2 = −3p1
(2) 3 (2)
⇒ p2 = − p1
2
and second equation agrees.
   
(2)
p1 2
We may take 
(2)
 =  
p2 −3

Slide 38
G. EXAMPLES - Continued III
So
 
1 2
P =  
1 −3
⇒ QT = P −1
 
1  −3 −2 
=
−5 −1 1
 
3 2
5 5
=  
1
5 − 51
 
q1T
=  
q2T

which gives left evecs

Slide 39
G. EXAMPLES IV
 
2 2
Use the same matrix as above A =  
3 1
and find its SVD. We have
  
T 2 3 2 2
A A =    
2 1 3 1
 
8 8
=  
8 10

Repeating the type of calaulations above gives


roots
√ √
ρ1 = 9 + 2 5; ρ2 = 9 − 5 etc...

Slide 40
Symbolic Computation in Matlab
   
a b 1
e.g. given A =   ,x =  .
0 c −d
Find Ax
Matlab code is

>> syms a
>> syms b
>> syms c
>> syms d

A*b
ans =
[ a-b*d, -c*d]

Slide 41

You might also like