18.03 LA.

10: The Matrix Exponential

[1] Exponentials
[2] Exponential matrix
[3] Fundamental matrices
[4] Diagonalization
[5] Exponential law

[1] Exponentials
What is ex ?
Very bad definition: ex is the xth power of the number e ∼ 2.718281828459045 . . .
Two problems with this: (1)√What is e? (2) What does it mean to raise
a number to the power of, say, 2, or π?
Much better definition: y(x) = ex is the solution to the differential equa-
dy
tion = y with initial condition y(0) = 1.
dx
Now there’s no need to know about e in advance; √ e is defined to be e1 .
And ex is just a function, which can be evaluated at 2 or at π just as easily
as at an integer.
Note the sublety: you can’t use this definition to describe ex for any single
x (except x = 0); you need to define the entire function at once, and then
evaluate that function at the value of x you may want.
As you know, this gives us solutions to other equations: I claim that
dy
y = ert satisfies = ry. This comes from the chain rule, with x = rt:
dt
dy dx dy
= = ry
dt dt dx

A further advantage of this definition is that it can be extended to other

contexts in a “brain-free” way.
A first example is Euler’s definition

eiθ = cos θ + i sin θ

1
We defined x(t) = e(a+bi)t to be the solution to ẋ = (a + bi)x, and then
calculated that
e(a+bi)t = eat (cos(bt) + i sin(bt))

In all these cases, you get the solution for any initial condition: ert x(0)
is the solution to ẋ = rx with initial condition x(0).
[2] Matrix exponential
We’re ready for the next step: We have been studying the equation
dx
= Ax
dt
where A is a square (constant) matrix.
Definition. eAt is the matrix of functions such that the solution to ẋ = Ax,
in terms of its initial condition, is eAt x(0).
How convenient is that!
If we take x(0) to be the vector with 1 at the top and 0 below, the product
e x(0) is the first column of eAt . Similarly for the other columns. So:
At

Each column of eAt is a solution of ẋ = Ax. We could write this:

d At
e = AeAt
dt
eAt is a matrix-valued solution! It satisfies a simple initial condition:
eA0 = I

Not everything about 1 × 1 matrices extends to the general n × n matrix.

But everything about 1×1 matrices does generalize to diagonal n×n matrices.
 
λ1 0
If A = Λ = , the given coordinates are already decoupled: the
0 λ2
equation
  ẋ = Ax is just ẋ1 = λ1x1 and ẋ2 = λ2 x2 . Plug in initialcondition

1 Λt eλ1 t 0
: the first column of e is . Plug in initial condition : the
0   0 1
0
second column is . So
e λ2 t
 λt 
Λt e 1 0
e =
0 eλ2 t

2
Same works for n × n, of course.
[3] Fundamental matrices
Here’s how to compute eAt . Suppose we’ve found the right number (n)
independent solutions of ẋ = Ax: say u1 (t), . . . , un (t). Line them up in a
row: this is a “fundamental matrix” for A:
 
Φ(t) = u1 u2 · · · un
The general solution is  
c1
x(t) = Φ(t) ... 
 
cn
 
c1
Φ(t) may not be quite eAt , but it’s close. Note that x(0) = Φ(0) ... ,
 
cn
 
c1
 .. 
or  .  = Φ(0)−1 x(0). Thus
cn
x(t) = Φ(t)Φ(0)−1 x(0)
So
eAt = Φ(t)Φ(0)−1
for any fundamental matrix Φ(t).
 
0 −1
Example: A = . Characteristic polynomial pA (λ) = λ2 +1, so the
1 0
eigenvalues are ±i. Thephase portrait
 is a “center.”
  Eigenvectors for λ = i
−i 1 1
are killed by A − iI = ; for example . So the exponential
1 −i i
solutions are given by
   
it 1 1
e = (cos t + i sin t)
i i
and its complex conjugate. To find real solutions, take just the right linear
combinations of these to get the real and imaginary parts:
   
cos t sin t
u1 (t) = , u2 (t) =
− sin t cos t

3
These both parametrize the unit circle, just starting at different places. The
corresponding fundamental matrix is
 
cos t sin t
Φ(t) =
− sin t cos t

We luck out, here: Φ(0) = I, so

 
At cos t sin t
e =
− sin t cos t

[4] Diagonalization
Suppose that A is diagonalizable: A = SΛS −1 .
 
1 2
Example: A = . You can find the eigenvalues as roots of the
0 3
characteristic polynomial, but you might as well remember that the eigenval-
ues of an upper (or lower) triangular matrix are the
 diagonal entries: here 1
1
and 3. Also an eigenvalue for 1 is easy: v1 = . For the other, subtract
  0  
−2 2 1
3 from the diagonal entries: kills v2 = .
0 0 1
So    
1 1 1 0
S= , Λ=
0 1 0 3

Suppose A = SΛS −1 . Then we have exponential solutions corresponding

to the eigenvalues:
u1 (t)eλ1 t v1 , . . .
These give a fine fundamental matrix:
 
Φ(t) = eλ1 t v1 . . . eλn t vn
 
λ1
= Se Λt
 
, S = u1 . . . un , Λ = 
 .. 
. 
λn
Then Φ(0) = S, so
eAt = SeΛt S −1

4
 In our example,
   
At 1 1 et 0 1 −1
e =
0 1 0 e3t 0 1

You could multiply this out, but, actually, the exponential matrix is often a
pain in the neck to compute, and is often more useful as a symbolic device.
Just like ex , in fact!

[5] The exponential law

I claim that
eA(t+s) = eAt eAs
This is a consequence of “time invariance.” We have to see that both sides
are equal after multiplying by an arbitrary vector v. Let x(t) be the solution
of ẋ = Ax with initial condtion x(0) = v: so x(t) = eAt v. Now fix s and let

y(t) = x(t + s) = eA(t+s) v

Calculate using the chain rule:

d d
y(t) = x(t + s) = ẋ(t + s) = Ax(t + s) = Ay(t)
dt dt
So y is the solution to ẏ = Ay with y(0) = x(s) = eAs v. That means that
y(t) = eAt eAs v. QED
This is the proof of the exponential law even in the 1 × 1 case; and you
will recall that as such it contains the trigonometric addition laws. Powerful
stuff!

5
 M.I.T. 18.03 Ordinary Differential
Equations
18.03 Extra Notes and Exercises
c Haynes Miller, David Jerison, Jennifer French and M.I.T., 2013