Dynamic Systems - State Space Control
- Lecture 3 ∗
September 11, 2012
Recall some basic facts about finite dimensional vector spaces.
Let v be a vector space whose field of scalars is R or C.
There are two laws of combination,
1 Vector Addition
• For v1 , v2 , v3 ∈ V, the laws of vector addition are associative;
(v1 + v2 ) + v3 = v1 + (v2 + v3 )
• For v1 , v2 ∈ V, the laws of vector addition are commutative;
v1 + v2 = v2 + v1
• There exists a zero vector 0 ∈ V, such that;
v + 0 = v for all v ∈ V
• For every v ∈ V, −v ∈ V as well.
∗ This work is being done by various members of the class of 2012
1
Background on Linear Algebra 2
2 Scalar Multiplication
• For any α ∈ R(C) and v ∈ V,
αv ∈ V
• For α, β ∈ R(C)
α(βv) = (αβ)v (scalar multiplication is associative)
• For α ∈ R(C) and v, w ∈ V,
α(v + w) = αv + αw (Distributive over vector spaces)
• For α, β ∈ R(C), v inV
(α + β)v = αv + βv (Distributive over scalar addition)
A basis for a vector space V is a linearly independent set of vectors: v1 , ..., vn
that spans V . Saying that v1 , ..., vn spans V means that for every w ∈ V there
are scalars α1 , ..., αn such that,
W = α1 v1 , ..., αn vn
1, s, s2 , ..., sn − These monomials are a basis of the vector space of polynomials
p(s) having degree ≤ n A vestor space V is finite dimensional of dimension n if
it has a basis with n elements.
A finite dimensional vector space over Rn (resp Cn ) is isomorphic to
Rn (resp Cn ).
THEOREM: Any square matrix with distinct eigenvalues can be put into
a diagonal form by a change of basis.
PROOF: List the n eigenvalues corresponding to distinct eigenvalues. Prove
this is a basis (DO THIS!). With respect to this basis, the matrix has diagonal
form.
Matrices whose eigenvalues are not distinct
Examples:
1.
1 0
0 1
Background on Linear Algebra 3
2.
λ 1
0 λ
Jordan Canonical Form:
Recall that the eigenvalues of a matrix A are the roots of the characteristic polynomial,
p(λ) = det(λI − A)
Suppose p(λ) can be factored
p(λ) = (λ − λ1 )r1 ...(λ − λp )rp
where λi 6= λj for i 6= j (to be sure that such factorisation exists, we must allow
for complex eigenvalues)
For each λj we are interested in the associated generalized eigenspace,
M k = ker(A − λj I)k = {v ∈ V : (A − λj I)k v = 0}
What are these?
M 0 = ker(A − λj I)0 = ker(I) = {0}
M 1 = ker(A − λj I) = eigenspace of λj
For v ∈ M k ,
(A − λj I)k+1 v = (A − λj I)(A − λj I)k v = (A − λj I)0 = 0
Therefore,M k ∈ M k+1 we have an increasing chain of subspaces M 0 ⊂ M 1 ⊂
Since M k ⊂ C for all k, there is some value t such that M k = M t for all
k ≥ t, let t be the smallest integer such that M k = M t for all k ≥ t (call
M t = Mλj )
Another (this time decreasing) subspace chain which is of interest is,
(A − λj I)k Cn = W k
Background on Linear Algebra 4
Here we have Cn = W 0 ⊂ W 1 ⊂ . . .
Let mλj = dim M(λj ) The dim W t = n − m(λj ) (Since the dimensions of the
range space and the null spce of (A − λj I)t must add up to r) We must have,
W k = W t for k ≥ t
And we denote W t by Wλj
Proposition 1: Cn is the direct sum of M(λj ) and W(λj ) (i.e any v ∈ Cn
maybe writen uniquely as W = vm + wm where vm ∈ M(λj ) , wm ∈ W(λj ) )
Proof: Since (A − λj I)W(λj ) = (A − λj I)t+1 Cn = (A − λj I)t Cn = Wλj
We see that A − λj I is nonsingular on W(λj ) . Let v be any vector in Cn . Then
(A − λj I)t v = 0 ∈ W(λj ) . Because (A − λj I)t is nonsingular on W(λj ) . There is
a unique wm = W(λj ) such that (A − λj I)t Wm = γ. Let vm = V − wm . It is
easily seen (A − λj I)t Vm = 0, so that Vm ∈ M(λj ) .
Hence Cn = M(λj ) + W(λj ) and remains only to show that the expression
W = vm + wm is unique. This follows since dim M(λj ) + dim W(λj ) = n, and
any pair of respective basis.
For M(λj ) and W(λj ) {v1 , . . . , vn }, {w1 , . . . , wn } together define a basis for Cn .
Cn = M(λj ) ⊕ W(λj )
Where, ⊕ is the notation for direct sum.
Background on Linear Algebra 5
Proposition 2: For any i, j, A − λI maps M(λj ) , W(λj ) onto themselves.
Proof: In the case i = j we’ve shown that (A − λj I)M k+1 ⊂ M k ⊂ M k+1 and
(A − λj I)W k = W k+1 ⊂ W k . This implies W(λj ) and M(λj ) are invarient
under (A − λi I)
But from this it follows that these subspaces are invariant under any
polynomial in A − λi I. In particular, under A − λi I = A − λi I + (λi − λj )I.
Proposition 3: For λi 6= λj , M(λi ) ⊂ W(λj )
Proof: Let v ∈ M(λj ) . I claim that v cannot be in kernal of A − λj I. For j 6= i
suppose to the contrary, V ∈ ker(A − λj I). Then,
(λj − λi )ti v = {A − λi I − (A − λj I)}ti v
ti
X ti
ti
= (A − λi I) v + (−1)k
(A − λiI )ti −k (A − λjI )k v
k
k=1
= 0 (the first term = 0 because v ∈ M(λj ) ), the remaining terms would need to
be zero under the assumption v ∈ ker(A − λj I)
Since λi − λj 6= 0, this implies v = 0. This in turn implies (A − λj I)|Mλi (The
linear transformation restricted to subspace M(λi ) ) is non-singular. This
means that A − λj I maps M(λi ) onto it-
self and hence, M(λi ) is contained in the range of (A−λj I)tj which is just W(λj ) .
Proposition 4: Cn = M(λ1 ) ⊕ M(λ2 ) ⊕ · · · ⊕ M(λp )
Remark: The direct sum notation means that any vector v ∈ Cn can be
uniquely expressed as a sum, V = v1 + · · · + vp , vj ∈ M(λj )
Proof: Cn = M(λ1 ) ⊕ W(λ1 ) and since W(λ1 ) ⊃ M(λ2 ) ,
We can write,
Cn = M(λ1 ) ⊕ (W(λ1 ) ∩ [M(λ2 ) ⊕ M(λ1 ) ])
= M(λ1 ) ⊕ ([W(λ1 ) ∩ M(λ2 ) ] ⊕ (W(λ1 ) ∩ W(λ2 ) ))
= M(λ1 ) ⊕ M(λ2 ) ⊕ (W(λ1 ) ∩ W(λ2 ) )
Similiarly we obtain,
Cn = M(λ1 ) ⊕ M(λ2 ) ⊕ M(λ3 ) ⊕ (W(λ1 ) ∩ W(λ2 ) ∩ W(λ3 ) )
Eventually we obtain,
Cn = M(λ1 ) ⊕ M(λ2 ) · · · ⊕ M(λp ) ⊕ (W(λ1 ) ∩ · · · ∩ W(λp ) )
Background on Linear Algebra 6
To complete the proof, we show that
W(λ1 ) ∩ · · · ∩ W(λp ) = {0}
Since each W(λi ) is invarient under all A − λj I, W(λ1 ) ∩ · · · ∩ W(λp ) is invarient
as well.
A − λj I is non-singular on W(λ1 ) ∩ · · · ∩ W(λp ) .Otherwise it would be singular
on all W(λi ) s and W(λj ) in particular. Hence,
(A − λ1 I)r1 (A − λ2 I)r2 . . . (A − λp I)rp
is nonsingular on W(λ1 ) ∩ · · · ∩ W(λp ) . But since this operator maps all vectors
in Cn to zero, the only way it could be nonsingular on W(λ1 ) ∩ · · · ∩ W(λp ) is if
this intersection is the zero subspace.
