20 Solving linear systems of ODE with constant coefficients.

Part I
20.1 General theory
Consider a homogeneous system of ODE with constant coefficients,

ẏ = Ay, A = [aij ]n×n ∈ Mn . (1)

together with the initial condition

y(t0 ) = y 0 ∈ Rn . (2)
My goal in this lecture is to show that the set of solutions to (1) forms an n dimensional vector space.
This fact alone allows to fully describe all possible solutions to system (1) by presenting a basis for this
vector space. First I state, without proof, the existence and uniqueness theorem for the IVP (1)–(2).

Theorem 1. IVP (1)–(2) has a unique solution y(t) defined for −∞ < t < ∞.

Note the global character of the theorem.

Proposition 2. The set S of solutions to system (1) is a vector space.

Proof. First, S is nonempty since 0 ∈ S. Let y 1 and y 2 belong to S, the set of solutions to (1). For
any constants α1 , α2 ∈ R consider y = α1 y 1 + α2 y 2 . We need to show that y ∈ S. This follows from
the fact that

(α1 y 1 + α2 y 2 )· = A(α1 y 1 + α2 y 2 ) ⇐⇒ α1 ẏ 1 + α2 ẏ 2 = α1 Ay 1 + α2 Ay 2 .

Checking the axioms of the vector space is left as an exercise. 

Theorem 3. Vector space S of the solutions to (1) is n-dimensional.

Proof. I will prove this theorem by presenting a basis for S that has exactly n elements.
Consider n initial value problems for (1), where

y(t0 ) = ei , i = 1, . . . , n,

and ei are standard unit vectors in Rn (i.e., vectors with one at the i-th position and zeros everywhere
else). Due to Theorem 1, we have n unique solutions, which I denote as

y i (t), i = 1, . . . , n.

Claim 1: {y 1 (t), . . . , y n (t)} span S. Indeed, assume that x(t) ∈ S is a solution to (1) with the
initial condition x(t0 ) = x0 = (x01 , . . . , x0n )⊤ ∈ Rn . Consider also

y(t) = x01 y 1 (t) + . . . x0n y n (t).

y(t) is a solution to (1) as a linear combination of solutions, and at the point t0 y(t0 ) = x(t0 ), hence,
by Theorem 1,
x(t) ≡ y(t),
MATH266: Intro to ODE by Artem Novozhilov, e-mail: [email protected]. Spring 2024

1
i.e., any solution to (1)–(2) can be represented as a linear combination of {y 1 (t), . . . , y n (t)} .
Claim 2: {y 1 (t), . . . , y n (t)} is a linearly independent set. Let β1 , . . . , βn ∈ R be such that

β1 y 1 (t) + . . . βn y 2 (t) = 0,

for any t (this is the definition of linear independence). Rewrite the last equality as a system in the
matrix form:
Φ(t)β = 0,
where Φ(t) is the matrix having y i (t) as its i-th column, and β ⊤ = (β1 , . . . , βn ).
Since the last equality has to be true for any t then it is true for t0 , but for t = t0 Φ(t0 ) = I, and
the only solution to
Iβ = 0
is a trivial one β ⊤ = (0, . . . , 0). Therefore, {y 1 (t), . . . , y n (t)} is a linearly independent set. Since
{y 1 (t), . . . , y n (t)} span S and is linearly independent, then it is a basis for S. 

Remark 4. A few remarks are in order here.

ˆ Any set of vector functions {y 1 (t), . . . , y n (t)} that form a basis for S is called a fundamental
solution set.

ˆ The matrix  
y11 (t) . . . yn1 (t)


Φ(t) = y 1 (t)| . . . |y n (t) =  ... ..
. ...  ,
y1n (t) . . . ynn (t)
where yij (t) is the j-th component of the i-th vector function y i (t), is called a fundamental
matrix solution if {y 1 (t), . . . , y n (t)} is a basis. This fundamental matrix solution satisfies the
matrix differential equation
Φ̇(t) = AΦ(t).

ˆ The proof of the theorem provides a simple test how to find out that n solutions to (1) are linearly
independent. Indeed, consider {y 1 (t), . . . , y n (t)} be n (non necessarily linearly independent)
solutions to (1). Compose matrix Φ(t) as above and consider

W (t) = det Φ(t),

the determinant of this matrix, which depends on t. This determinant is called the Wronskian.
The proof of the theorem actually shows that if there is t0 ∈ R such that W (t0 ) ̸= 0 then
{y 1 (t), . . . , y n (t)} is a linearly independent set.

ˆ For the proof I chose ei as the initial conditions for y i (t). It should be clear that exactly the
same result follows if we just take n linearly independent arbitrary vectors in Rn .

Example 5. Let us show that

   
e−t te−t
y 1 (t) = , y 2 (t) =
−e−t (1 − t)e−t

2
form a basis for the set of solutions to the system

ẏ1 = y2 ,
ẏ2 = −y1 − 2y2 .

The system can be written in the matrix form

   
0 1 y1
ẏ = Ay, A= , y= .
−1 −2 y2

First we need to show that both y 1 and y 2 are solutions. Start with y 1 . Plugging its components
instead of y1 and y2 in the system, we find that y 1 solves the system. The same is true for y 2 (check!).
Now we need to show that they form a linearly independent set. For this, pick t = 0 and consider two
vectors    
1 0
x1 = y 1 (0) = , x2 = y 2 (0) = .
−1 1
These two vectors are linearly independent (because, e.g., the determinant of the matrix composed of
these two vectors, is not zero), hence {y 1 (t), y 2 (t)} is a linearly independent set and forms a basis.
Hence, {y 1 , y 2 } is a fundamental solution set for our problem, the matrix composed from these
vector-functions is a fundamental matrix solution, and the general solution to the system can be
written as  
C1 e−t + C2 te−t
y(t) = C1 y 1 (t) + C2 y 2 (t) = ,
−C1 e−t + C2 (1 − t)e−t
where C1 , C2 are arbitrary real constants.

20.2 Eigenvectors and eigenvalues

Now we know, thanks to the general theory in the previous subsection, that the set of solutions to (1)
is an n-dimensional vector space S. The actual question is how to find a basis for S. Remembering
our success in solving linear ODE of the n-th order using a guess in the form eλt , consider an ansatz
 
v1
λt  ..  λt
y(t) = ve =  .  e .
vn

Plugging this ansatz into (1), we find that

λveλt = Aveλt ⇐⇒ Av = λv,

which should be familiar to you from your linear algebra course.

Definition 6. Nonzero vector v ∈ Cn is called an eigenvector of A ∈ Mn if there exists a scalar

λ ∈ C such that
Av = λv.
λ is called the eigenvalue of A corresponding to v.

3
 Note that according to the guess above, if we have eigenvalue λ with the eigenvector v, then

y(t) = veλt

is a solution to (1).
How to find eigenvalues and eigenvectors? We have

Av − λv = (A − λI)v = 0,

which is a homogeneous system of n linear algebraic equations with n unknowns. We are looking for
any solution other than zero, therefore, we have to require that

det(A − λI) = 0.

The last equality has on the left a polynomial of degree n depending on λ. This polynomial is called
the characteristic polynomial. Hence, the eigenvalues can be found as the roots of the characteristic
polynomial. After we find λ-s, we can look for the eigenvectors one by one.
Example 7. Find the eigenvalues and eigenvectors of
 
4 2
A= .
−1 1

First, find the characteristic polynomial:

4−λ 2
det(A − λI) = = λ2 − 5λ + 6 = 0,
−1 1 − λ

which means that we have two real eigenvalues λ1 = 2 and λ2 = 3.

λ1 = 2. Hence we have
    
2 2 v1 0
(A − 2I)v = 0 ⇐⇒ = .
−1 −1 v2 0

In coordinates we have the system

2v1 + 2v2 = 0,
−v1 − v2 = 0.

This is a homogeneous system of linear algebraic equations which always has solution (0, 0)⊤ . However,
since we are looking for eigenvectors, we are not interested in this solution. Note that the eigenvalue
was chosen such that this system has infinitely many solutions. In other words, the vectors composing
the matrix of the system are linearly dependent. To solve a system having infinitely many solutions
we need to choose some variables that we call free. Other variables should be expressed through the
free variables. In our case we can consider

v1 = −v2 ,

from the first equation (exactly the same follows from the second equation, as expected). In this case
I choose v2 be the free variable. Hence any vector of the form v 1 = (−v2 , v2 )⊤ solves my system. I

4
need only one such vector, so I am free to pick any value for v2 , e.g., v2 = 1. Therefore, my first
eigenvector, which is defined up to a multiplicative constant, is
 
−1
v2 = .
1

λ2 = 3. In this case the system is

v1 + 2v2 = 0,
−v1 − 2v2 = 0,

and the second eigenvector can be chosen as

 
−2
v2 = .
1

Now consider a system of ODE

ẏ = Ay,
where A as above. We know two eigenvalues and two eigenvectors, therefore, we have two solutions
to this system:    
λ1 t −1 2t λ2 t −2 3t
y 1 (t) = v 1 e = e , y 2 (t) = v 2 e = e .
1 1
We can check directly in this example that vectors v 1 and v 2 are linearly independent, hence y 1 (t)
and y 2 (t) are linearly independent and form a basis. Therefore the general solution to our problem is
given by
     
2t 3t −1 2t −2 3t −C1 e2t − 2C2 e3t
y(t) = C1 v 1 e + C2 v 2 e = C1 e + C2 e = .
1 1 C1 e2t + C2 e3t

In the last example I had to check that {v 1 , v 2 } is a linearly independent set to prove that {y 1 , y 2 }
is a basis for S. There is a general fact, which is useful in this respect:

Proposition 8. If λ1 , . . . , λm are distinct eigenvalues of A ∈ Mn with the corresponding eigenvectors

v 1 , . . . , v m , then {v 1 , . . . , v m } is a linearly independent set.

You can skip the proof, but for a mathematically inclined student I will present it here.

Proof. Let me prove this proposition by induction. Check this for m = 1. v 1 is linearly independent
because it is by definition non-zero. Now assume it is true for m = j, i.e., we assume that any set of j
eigenvectors corresponding to distinct eigenvalues is linearly independent. Now I would like to prove
that it is also true for j + 1 eigenvectors. Consider the linear combination

α1 v 1 + . . . + αj v j + αj+1 v j+1 = 0.

Multiply both sides of it by A from the left and use the fact that v i are eigenvectors:

α1 λ1 v 1 + . . . + αj λj v j + αj+1 λj+1 v j+1 = 0.

5
Now multiply the first equality by λ1 and subtract from the second one:

α2 (λ2 − λ1 )v 2 + . . . + αj+1 (λj+1 − λ1 )v j+1 = 0,

which a linear combination of j eigenvectors, which form linearly independent set by assumptions.
This means that
αi (λi − λ1 ) = 0, i = 2, . . . j + 1,
and since all λ are distinct
αi = 0, i = 2, . . . , j + 1,
which leaves us with
α1 v 1 = 0 =⇒ α1 = 0,
therefore {v 1 , . . . , v j+1 } is independent. 

Example 9. Find the general solution for

 
1 −1 4
ẏ = 3 2 −1 y.
2 1 −1

First, find eigenvalues:

 
1 − λ −1 4
det(A − λI) = det  3 2−λ −1  = λ3 − 2λ2 − 5λ + 6.
2 1 −1 − λ

Recall that the integer roots of a polynomial can be found among the integers that divide 6. These
are ±1, ±2, ±3. You can find that already λ1 = 1 works. You can proceed checking the roots one by
one and find that λ2 = 3 and λ3 = −2, or you can divide the characteristic polynomial by (λ − 1) and
find that the result is the quadratic polynomial λ2 − λ − 6 that has roots 3 and −2.
Now to the eigenvectors.
Case λ1 = 1. We have  
0 −1 4
A − λ1 I = 3 1 −1 ,
2 1 −2
hence we need to solve

−v2 + 4v3 = 0,
3v1 + v2 − v3 = 0,
2v1 + v2 − 2v3 = 0.

The correct way to deal with such systems is to use the process of gaussian elimination to put the
system in the upper echelon form. Here, however, you can see from the first equation that v2 = 4v3 ,
which implies that v1 = −v3 from both second and the third equation. Hence the general solution
here is (−v3 , 4v3 , v3 )⊤ , where v3 is chosen as the free variable. An eigenvector can be taken as
v 1 = (−1, 4, 1)⊤ .

6
Case λ2 = 3. An eigenvector can be taken as

v⊤
2 = (1, 2, 1).

Case λ3 = −2. An eigenvector can be taken as

v⊤
3 = (−1, 1, 1).

We know that since v 1 , v 2 , v 3 are eigenvectors corresponding to distinct eigenvalues, they are
linearly independent, and hence the general solution to the problem can be written as
     
−1 1 −1
y(t) = C1 4 e + C2 2 e + C3 1  e−2t .
  t   3t 
1 1 1

20.3 Summary
Let us put together the information we now have:
ˆ The vector space S of solutions to (1) is n-dimensional. Hence, to provide the general solution,
we need a basis, which consists of a linearly independent set of n solutions to (1).

ˆ If λ is an eigenvalue of A and v is a corresponding eigenvector, then y(t) = veλt is a solution

to (1).

ˆ Eigenvectors corresponding to distinct eigenvalues form a linearly independent set, hence if we

are lucky to have n distinct eigenvalues λi , i = 1, . . . , n, with the corresponding eigenvectors
v i , i = 1, . . . , n, then {y i (t) = v i eλi t }ni=1 is a basis, and the general solution has the form:

y(t) = C1 v 1 eλ1 t + . . . + Cn v n eλn t .

Is the problem solved? Not really. Recall that the eigenvalues are the roots of the characteristic
polynomial det(A − λI), which has degree exactly n. As we remember, it means that this polynomial
has exactly n roots, counting multiple. Hence the first issue: What to do with the eigenvalues, which
are the roots of the characteristic polynomial with multiplicity more then one. We cannot have two
the same eigenvectors, because they are clearly linearly dependent. Second issue is that we might
have complex eigenvalues. Therefore, the corresponding eigenvector will be also complex (do you see
why?). On the other hand the matrix of our system is generally real-valued, and we would prefer to
deal with real-valued solutions. We will deal with these two questions in the next lectures.

∗
20.4 For a mathematically inclined student
In the introductory lecture on the systems of ODE I said that the usual way to solve systems is to look
for a coordinate change, such that in the new coordinates the system becomes decoupled and we can
solve the equations one by one. Opposite to what I said, I never mentioned any change of variables
in the theory and computations outlined above. Let me fix this issue here and show how exactly the
same solution can be obtained from a different point of view.
Consider system
ẏ = Ay

7
and change of variables y = V x, where matrix V has to be determined. We obtain, tacitly assuming
that V is invertible, that
V ẋ = AV x ⇐⇒ ẋ = V −1 AV x.
Now we need to answer the question how to pick V such that V −1 AV is a “simple” matrix. By
“simple” I mean, e.g., a diagonal matrix, but this is not always possible. Here is a very important
result.

Proposition 10. Let A ∈ Mn have n distinct eigenvalues λi with the corresponding eigenvectors
v i , i = 1, . . . , n. Compose the matrix V from the eigenvectors such that the i-th vector takes the i-th
column. Then
A = V ΛV −1 , V −1 AV = Λ,
where Λ is the diagonal matrix with (λi )ni=1 on the main diagonal.

Proof.
AV = Av 1 + . . . + Av n = λ1 v 1 + . . . + λn v n = V Λ,
and since {v 1 , . . . , v n } forms a linearly independent set, V is invertible. 

Therefore, for new variables x, we obtain the decoupled system

ẋ = Λx,

which has the solution

xi (t) = Ci eλi t , i = 1, . . . , n.
To return to the original variable, we have

y(t) = V x = C1 v 1 eλ1 t + . . . + Cn v n eλn t ,

exactly as we already obtained before.