EE263 Autumn 2015 S. Boyd and S.

Lall

Dynamic interpretation of eigenvectors

I invariant sets

I complex eigenvectors & invariant planes

I left eigenvectors

I modal form

I discrete-time stability

1
Dynamic interpretation

suppose Av = λv, v 6= 0
if ẋ = Ax and x(0) = v, then x(t) = eλt v
several ways to see this, e.g.,

(tA)2
 
x(t) = etA v = I + tA + + ··· v
2!
(λt)2
= v + λtv + v + ···
2!
= eλt v

(since (tA)k v = (λt)k v)

2
Dynamic interpretation

I for λ ∈ C, solution is complex (we’ll interpret later); for now, assume λ ∈ R

I if initial state is an eigenvector v, resulting motion is very simple — always

on the line spanned by v

I solution x(t) = eλt v is called mode of system ẋ = Ax (associated with

eigenvalue λ)

I for λ ∈ R, λ < 0, mode contracts or shrinks as t ↑

I for λ ∈ R, λ > 0, mode expands or grows as t ↑

3
Invariant sets

a set S ⊆ Rn is invariant under ẋ = Ax if whenever x(t) ∈ S, then x(τ ) ∈ S for

all τ ≥ t
i.e.: once trajectory enters S, it stays in S
trajectory

vector field interpretation: trajectories only cut into S, never out

4
Invariant sets

suppose Av = λv, v 6= 0, λ ∈ R

I line { tv | t ∈ R } is invariant
(in fact, ray { tv | t > 0 } is invariant)

I if λ < 0, line segment { tv | 0 ≤ t ≤ a } is invariant

5
Complex eigenvectors

suppose Av = λv, v 6= 0, λ is complex

for a ∈ C, (complex) trajectory aeλt v satisfies ẋ = Ax
hence so does (real) trajectory
 
x(t) = < aeλt v
  
cos ωt sin ωt α
= eσt vre
 
vim
− sin ωt cos ωt −β

where
v = vre + ivim , λ = σ + iω, a = α + iβ

I trajectory stays in invariant plane span{vre , vim }

I σ gives logarithmic growth/decay factor

I ω gives angular velocity of rotation in plane

6
Dynamic interpretation: left eigenvectors

suppose wT A = λwT , w 6= 0
then
d T
(w x) = wT ẋ = wT Ax = λ(wT x)
dt
i.e., wT x satisfies the DE d(wT x)/dt = λ(wT x)
hence wT x(t) = eλt wT x(0)

I even if trajectory x is complicated, w T x is simple

I if, e.g., λ ∈ R, λ < 0, halfspace { z | w T z ≤ a } is invariant (for a ≥ 0)

I for λ = σ + iω ∈ C, (<w)T x and (=w)T x both have form

eσt (α cos(ωt) + β sin(ωt))

7
Summary

I right eigenvectors are initial conditions from which resulting motion is simple
(i.e., remains on line or in plane)

I left eigenvectors give linear functions of state that are simple, for any initial
condition

8
Example

 
−1 −10 −10
ẋ =  1 0 0 x
0 1 0
block diagram:

x1 x2 x3
1/s 1/s 1/s

−1 −10 −10

X (s) = s3 + s2 + 10s + 10 = (s + 1)(s2 + 10)

√
eigenvalues are −1, ± i 10

9
Example

trajectory with x(0) = (0, −1, 1):

2

x1 1

−1
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
0.5
t
0
x2

−0.5

−1
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
1
t
0.5
x3

−0.5
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
t

10
Example

left eigenvector associated with eigenvalue −1 is

 
0.1
g= 0 
1

let’s check g T x(t) when x(0) = (0, −1, 1) (as above):

1

0.9

0.8

0.7

0.6
gT x

0.5

0.4

0.3

0.2

0.1

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

t
11
Example

√
eigenvector associated with eigenvalue i 10 is
 
−0.554 + i0.771
v =  0.244 + i0.175 
0.055 − i0.077

so an invariant plane is spanned by

   
−0.554 0.771
vre =  0.244  , vim =  0.175 
0.055 −0.077

12
Example

for example, with x(0) = vre we have

1

x1
0

−1
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
0.5
t
x2

−0.5
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
0.1
t
x3

−0.1
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
t

13
Example: Markov chain

probability distribution satisfies p(t + 1) = P p(t)

Pn
pi (t) = Prob( z(t) = i ) so i=1 pi (t) = 1

Pij = Prob( z(t + 1) = i | z(t) = j ), so n

P
i=1 Pij = 1
(such matrices are called stochastic)
rewrite as:
[1 1 · · · 1]P = [1 1 · · · 1]
i.e., [1 1 · · · 1] is a left eigenvector of P with e.v. 1
hence det(I − P ) = 0, so there is a right eigenvector v 6= 0 with P v = v
it canPbe shown that v can be chosen so that vi ≥ 0, hence we can normalize v so
that n i=1 vi = 1

interpretation: v is an equilibrium distribution; i.e., if p(0) = v then p(t) = v for

all t ≥ 0
(if v is unique it is called the steady-state distribution of the Markov chain)

14
Modal form

suppose A is diagonalizable by T
define new coordinates by x = T x̃, so

T x̃˙ = AT x̃ ⇔ x̃˙ = T −1 AT x̃ ⇔ x̃˙ = Λx̃

15
Modal form

in new coordinate system, system is diagonal (decoupled):

x̃1
1/s

λ1

x̃n
1/s

λn

trajectories consist of n independent modes, i.e.,

x̃i (t) = eλi t x̃i (0)

hence the name modal form

16
Real modal form

when eigenvalues (hence T ) are complex, system can be put in real modal form:

S −1 AS = diag (Λr , Mr+1 , Mr+3 , . . . , Mn−1 )

where Λr = diag(λ1 , . . . , λr ) are the real eigenvalues, and

 
σj ωj
Mj = , λj = σj + iωj , j = r + 1, r + 3, . . . , n
−ωj σj

where λj are the complex eigenvalues (one from each conjugate pair)

17
Real modal form

block diagram of ‘complex mode’:

1/s

−ω

1/s

18
Diagonalization

diagonalization simplifies many matrix expressions

e.g., resolvent:

−1
(sI − A)−1 = sT T −1 − T ΛT −1

−1
= T (sI − Λ)T −1
= T (sI − Λ)−1 T −1
 
1 1
= T diag ,..., T −1
s − λ1 s − λn

powers (i.e., discrete-time solution):

k
Ak = T ΛT −1
= T ΛT −1 · · · T ΛT −1

= T Λk T −1
= T diag(λk1 , . . . , λkn )T −1

(for k < 0 only if A invertible, i.e., all λi 6= 0)

19
Diagonalization

exponential (i.e., continuous-time solution):

eA = I + A + A2 /2! + · · ·

2
= I + T ΛT −1 + T ΛT −1 /2! + · · ·
2 −1
= T (I + Λ + Λ /2! + · · · )T
= T eΛ T −1
= T diag(eλ1 , . . . , eλn )T −1

20
Analytic function of a matrix

for any analytic function f : R → R, i.e., given by power series

f (a) = β0 + β1 a + β2 a2 + β3 a3 + · · ·

we can define f (A) for A ∈ Rn×n (i.e., overload f ) as

f (A) = β0 I + β1 A + β2 A2 + β3 A3 + · · ·

substituting A = T ΛT −1 , we have

f (A) = β0 I + β1 A + β2 A2 + β3 A3 + · · ·
= β0 T T −1 + β1 T ΛT −1 + β2 (T ΛT −1 )2 + · · ·
= T β0 I + β1 Λ + β2 Λ2 + · · · T −1

= T diag(f (λ1 ), . . . , f (λn ))T −1

21
Solution via diagonalization

assume A is diagonalizable
consider LDS ẋ = Ax, with T −1 AT = Λ
then

x(t) = etA x(0)

= T eΛt T −1 x(0)
n
X
= eλi t (wiT x(0))vi
i=1

thus: any trajectory can be expressed as linear combination of modes

22
Interpretation

I (left eigenvectors) decompose initial state x(0) into modal components wiT x(0)

I eλi t term propagates ith mode forward t seconds

I reconstruct state as linear combination of (right) eigenvectors

23
Application

for what x(0) do we have x(t) → 0 as t → ∞?

divide eigenvalues into those with negative real parts

<λ1 < 0, . . . , <λs < 0,

and the others,

<λs+1 ≥ 0, . . . , <λn ≥ 0

from
n
X
x(t) = eλi t (wiT x(0))vi
i=1

condition for x(t) → 0 is:

x(0) ∈ span{v1 , . . . , vs },

or equivalently,
wiT x(0) = 0, i = s + 1, . . . , n

(can you prove this?)

24
Stability of discrete-time systems

suppose A diagonalizable
consider discrete-time LDS x(t + 1) = Ax(t)
if A = T ΛT −1 , then Ak = T Λk T −1
then
n
X
x(t) = At x(0) = λti (wiT x(0))vi → 0 as t → ∞
i=1

for all x(0) if and only if

|λi | < 1, i = 1, . . . , n.

we will see later that this is true even when A is not diagonalizable, so we have
fact: x(t + 1) = Ax(t) is stable if and only if all eigenvalues of A have magnitude
less than one

25