Ordinary Differential Equations – Linear differential equations…

Ordinary Differential Equations

5. Examples of linear differential equations and their

applications

We consider some examples of systems of linear differential equations with constant

coefficients

y1′ = a11 y1 + ... + a1n yn

y2′ = a21 y1 + ... + a2 n yn
(5.1)
#
yn′ = an1 y1 + ... + ann yn

and the linear n-th-order differential equation (n > 1) with constant coefficients

x ( n ) + an −1 x ( n −1) + ... + a1 x′ + a0 x = 0 (5.2)

We discuss suitable methods to compute solutions and the fundamental matrix. In

particular, we compute the fundamental matrix for all planar systems (n = 2) and classify
these systems according to their properties in phase space. We also introduce the notion
of stability for linear differential equations.

5.1. Planar systems

1. Equation of small oscillations. We study the linear second-order differential equation

x + a1 x + a0 x = 0
 (5.3)

for real coefficients a0, a1. This equation arises from problems in mechanics and
electricity.

Oscillations of a spring. A mass m coupled to a spring is submitted to a linear force.

Newton’s equation of motion reads mx = − Dx where D is a positive constant. It is of the
form (5.3) with a0 = D/m and a1= 0. If friction is taken into account then a1 is a positive
constant.

Small oscillations of the mathematical pendulum. We consider a weightless rod of

length l attached at one end and bearing a point mass m at the other end. If x denotes the
x = −mgl sin ( x ) . For small
angle by which the rod deviates from the vertical then ml 2 
deviations sin(x) ≈ x and we get an equation of the form (5.3) with a0 = g/l and a1= 0.

1
 Ordinary Differential Equations – Linear differential equations…

Again, if we take into account a force due to friction that is proportional to the velocity
then a0 = a1 is a positive constant.

Inverted pendulum. If we turn the pendulum upside down, then the sign of the force
due to gravitation changes sign and therefore a0 = - g/l.

Electrical circuits. The charge y of a condenser of capacity C in a closed electrical

circuit with inductivity L and resistance R satisfies the second order differential equation
Ly + Ry + C −1 y = 0 which is of the form (5.3) with a0 = 1/LC and a1= R/L.

1.1. Transformation to a first-order system. The second-order differential equation

(5.3) is equivalent to a two-component linear first-order system with matrix A given by

 0 1 
A=  (5.4)
 −a0 −a1 

It is convenient to rewrite A using the trace and the determinant or its eigenvalues. Then

 0 1   0 1 
A= = (5.5)
 − det ( A ) tr ( A )   −λ1λ2 λ1 + λ2 

where

tr ( A ) − tr ( A ) − 4 det ( A )
2

λ1 =
2 (5.6)
tr ( A ) + tr ( A ) − 4 det ( A )
2

λ2 =
2

are the (possibly complex) eigenvalues of A. If there is no friction (no resistance), then
tr(A) = 0 which means that the system is area preserving in phase space.

1.2. Equation in phase space. The equation (5.3) describes a motion in the two-
dimensional phase space with coordinates x, y = x . The matrix A defines the direction
field, which is a composition of a rotation and a dilation of the plane. The orbit of a
solution is the set of points {( x(t ), y (t ) ) , t ∈ \} or if we consider y as a function of x the
set of points {x, y(x)} where y is (at least formally) a solution of the non-linear first-order
differential equation

dy ( x) x x
= − a0 − a1 = − det ( A ) + tr ( A )
dx y ( x) y ( x)

2
 Ordinary Differential Equations – Linear differential equations…

which is of homogeneous type and can be solved (at least in principle) by elementary
integration methods (see ch.2.3). Another important quantity defined on the phase space
is the energy of the system. The energy functional associated to (5.3) is defined by

E ( x, x ) = 12 x 2 + 12 a0 x 2 (5.7)

If x(t) is a solution of (5.3) then

d
E ( x, x ) = x ( 
x + a0 x ) = − a1 x 2 = tr ( A ) x 2 (5.8)
dt

If there is no friction ( a1 = -tr(A) = 0) the energy of a solution remains constant and the
solution follows the lines of constant energy in phase space. In this case the system is
called conservative. Since the equation is linear the energy defines a quadratic form and
the orbits are either ellipses (det(A) = a0 > 0) , hyperbolas (det(A) = a0 < 0), or straight
lines (det(A) = a0 = 0). In the presence of friction in a mechanical system or electrical
resistance in an electrical circuit ( a1 = -tr(A) > 0) the energy is decreasing.

1.3. Explicit solutions of the second order equation. According to the results of the
previous chapter (theorem 4.7) the general complex valued solution is of the form

x ( t ) = c1 exp ( λ1t ) + c2 exp ( λ2t ) , λ1 ≠ λ2

(5.9)
x ( t ) = c1 exp ( λ t ) + c2t exp ( λt ) , λ1 = λ2 ≡ λ

The corresponding real solutions are listed in the following table.

λ1, λ2 real , tr(A)2 ≥ 4det(A)

λ1 ≠ λ2 tr(A)2 > 4det(A) λ1 = λ2 = λ, tr(A)2 = 4det(A)

x ( t ) = c1 exp ( λ1t ) + c2 exp ( λ2t ) x ( t ) = c1 exp ( λ t ) + c2t exp ( λ t )

λ1, λ2 complex conjugate, tr(A)2 < 4det(A)

λ1 = µ – iω, λ2 = µ + iω λ1 = – iω, λ2 = iω, tr(A) = 0

x ( t ) = exp ( µ t ) ( c1 sin (ω t ) + c2 cos (ω t ) ) or x ( t ) = c1 sin (ω t ) + c2 cos (ω t ) or

x ( t ) = C exp ( µ t ) sin (ω t + δ ) x ( t ) = C sin (ω t + δ )

The inverted pendulum belongs to the first class since det(A) = a0 < 0. In all other models
det(A) = a0 > 0 and damping -a1 = tr(A) ≤ 0 For weak damping i.e. small tr(A) ≤ 0 there
are two complex conjugate, critical damping corresponds to a double real eigenvalue If
both eigenvalues are real and different the system is called overdamped.

3
 Ordinary Differential Equations – Linear differential equations…

1.4. Computing exp(tA). A complex valued fundamental matrix of the associated first
order system is easily computed if the eigenvalues are different. The matrix A defined in
(5.4) resp. (5.5) has eigenvectors

1 1
v1 =   , v2 =  
 λ1   λ2 

Therefore a fundamental matrix Y(t) is given by

 exp ( λ1t ) exp ( λ2t ) 

Y ( t ) = ( v1 exp ( λ1t ) v2 exp ( λ2t ) ) =   (5.10)
 λ1 exp ( λ1t ) λ2 exp ( λ2t ) 
Since

 1  0
λ2 v1 − λ1v2 = ( λ2 − λ1 )   , v2 − v1 = ( λ2 − λ1 )  
0  1 
we get

λ1t λ2t
1  λ2 e − λ1 e eλ2t − eλ1t 
exp ( tA ) =   (5.11)
 (
λ2 − λ1  λ2 λ1 eλ1t − eλ2t ) λ2 eλ2t − λ1 eλ1t 


If λ1 = µ – iω, λ2 = µ + iω (5.11) reads

e µ t  ω cos (ω t ) − µ sin (ω t ) sin (ω t ) 

exp ( tA ) =   (5.12)
(
ω  sin (ω t ) µ 2 + ω 2 ) ω cos (ω t ) + µ sin (ω t ) 

To compute exp(tA) in case λ1 = λ2 = λ we let λ1 → λ2 = λ . We obtain

1 − λt t 
exp ( tA ) = eλt   (5.13)
 −t λ 1 + λ t 
2

The direct computation goes as follows: By theorem 4.7. the functions exp(λt)
and t⋅exp(λt) are solutions of the second order equation. The corresponding solution
vectors are

1  t 
v1 = eλt   , v2 = eλt  
λ  1 + λt 

from which we easily construct a fundamental matrix.

4
 Ordinary Differential Equations – Linear differential equations…

2. General first order systems and their classification. Let A be an arbitrary real 2×2
matrix. Then A is either similar to the matrix given in (5.5) or A is a multiple of the unit
matrix E. In the first case all results are similar to the results for the second order
equation. To classify the systems it is sufficient to consider their normal forms, which are
one of the following real matrices:

 λ1 0   λ 1   µ ω 
 , ,  (5.14)
 0 λ2   0 λ   −ω µ 

where we used the notations introduced above.

Diagonal case
λ1 < λ2 < 0 λ1 < 0 < λ2 0 < λ1 < λ2

Stable node Saddle point Unstable node

λ1 = λ2 < 0 λ1 < 0 = λ2 0 < λ1 = λ2

Stable star Marginally stable Unstable star

5
 Ordinary Differential Equations – Linear differential equations…

Jordan case
λ1 = λ2 = λ < 0 λ1 = λ2 = λ = 0 λ1 = λ2 = λ = 0

Stable node Marginally unstable Unstable node

Complex case
µ<0,ω>0 µ=0,ω>0 µ>0,ω>0

Stable focus Center Unstable focus

2.1. Stability. We precise the meaning of stability and instability of linear equations. We
note that y(t) = 0 is always a stationary solution of a linear homogeneous system. We
define stability and instability as a property of solutions with initial conditions close to
the zero solution. We say that y(t) = 0 is stable if all solutions remain bounded,
asymptotically stable if all solutions tend to zero as t tends to infinity, and unstable if
there is an unbounded solution of the linear equation. We have seen that y(t) = 0 is stable
if there is no eigenvalue with positive real part ( with exception of the marginally
unstable Jordan block), asymptotically stable if all eigenvalues have negative real part,
and unstable if there is an eigenvalue with positive real part or a marginally unstable
Jordan block. A more appropriate formulation of stability of a stationary solution, which
also extends to nonlinear systems is the concept of Lyapunov stability.

Definition: The stationary solution y(t) = 0 is called stable or Lyapunov stable if for
every ε > 0 there is a δ > 0 such that for every y0 with | y0| < δ the solution y(t) with

6
 Ordinary Differential Equations – Linear differential equations…

initial condition y(t) = y0 exists for all t > 0 and satisfies | y(t) | < ε for all t > 0. The
stationary solution is called asymptotically stable if in addition all solutions satisfy
lim y (t ) = 0 .
t →∞

As an application of the linear result we present the following result for nonlinear planar
autonomous systems.

Theorem 1: Let A ∈ M ( 2 × 2, \ ) and g : \ 2 → \ 2 , g ( y ) = O y ( ) a differentiable

2

function. Consider the differential equation

y′ = Ay + g ( y ) (5.15)

If all eigenvalues of A have negative real parts then the stationary solution y = 0 of (5.15)
is asymptotically stable.

Proof: The proof of the stability result will be a beautiful application of Gronwall-s
inequality. If y(t) is a solution of (5.15) then it satisfies the integral equation

t
y ( t ) = exp ( At ) y0 + ∫ exp ( A(t − s ) ) g ( y ( s ) ) ds (5.16)
0

Let |.| be a norm on R2. Since all eigenvalues of A have negative real parts there exist
positive constants a, C such that

exp ( At ) x ≤ Ce − at x

for any x ∈ R2.

By the assumption on g there is a positive constant K such that g ( y ) ≤ K y if y ≤ r .
2

Therefore, if y ≤ r , then

g ( y ) ≤ Kr y

Later we shall choose r sufficiently small such that the constant Kr will be small. Taking
the norm on both sides of (5.16) and using these estimates we obtain the following
integral inequality which holds for all t such that y ≤ r in [0,t]:

t
y ( t ) ≤ Ce − at y0 + CKr ∫ e
− a( t − s )
y ( s ) ds
0

7
 Ordinary Differential Equations – Linear differential equations…

We set z (t ) = y (t ) e at . Then z(t) satisfies the integral inequality

t
z (t ) ≤ C y0 + CKr ∫ z ( s )ds
0

By Gronwall’s inequality z (t ) ≤ C y0 e CKrt

.
We choose r sufficiently small such that CKr = a/2. We get

y (t ) ≤ C y0 e − at 2

If we choose y0 such that y0 ≤ r C , then the above inequality shows in particular that
y (t ) ≤ r for all t > 0 concluding the proof.

Remark: It is clear that theorem 1 also holds for higher dimensional autonomous
systems.

For the second order differential equation (5.3) the trivial solution x = 0 is asymptotically
stable if det(A) = a0 > 0 and -a1 = tr(A) < 0. In this case the energy functional defined in
(5.7) is positive definite and strictly decreasing for nontrivial solutions. Functionals on
phase space having this property will be called Lyapunov functionals. The existence of an
appropriate Lyapunov functional will imply the stability result.

3. Applications
3. 1. An example from pharmacokinetics. The reaction equation


k21
A1 ← → A2 
k0
→0
k12

describes a basic model for studying the dynamics of a drug in the body where A1 and A2
represent the drug concentration in tissue and blood, respectively. The rate k0 is due to
removal from the body by the kidneys. The corresponding system of differential
equations is then

y1 = k12 y2 − k21 y1

(5.17)
y 2 = − k12 y2 + k21 y1 − k0 y2

Hence

 −k k12 
A =  21  (5.18)
 k21 − k12 − k0 

8
 Ordinary Differential Equations – Linear differential equations…

and tr(A) < 0, det(A) = k21k0 > 0. The trivial solution is asymptotically stable: the drug
will completely disappear as one expects.

3. 2. The mathematical pendulum with friction. The equation for the mathematical
pendulum with friction is given by

x + γ x = − gl −1 sin ( x )
 (5.19)

By theorem 1 the stationary solution x = 0 is asymptotically stable.

5.2. How to solve linear equations with constant coefficients

We consider the differential equation

y′ = Ay

for a real n×n matrix A. While it is straightforward to find all solutions if all eigenvalues
of A have algebraic multiplicity one (see theorem 4.5) the computation of solutions in the
case of multiple eigenvalues is a little bit more involved.

1. A solution algorithm. We describe a routine to compute a fundamental system of any

linear differential equations with constant coefficients.

(step 1) Compute the characteristic polynomial PA(σ) and compute its zeros which are the
eigenvalues of A.

(step 2) For simple eigenvalues λ find the corresponding eigenvectors w. construct the
solution as in theorem 4.5, that is set y (t ) = eλt w and take real and imaginary part in case
of complex conjugate eigenvalues λ , λ .

(step 3)Let λ be an eigenvalue with multiplicity k > 1. First of all, solve for corresponding
eigenvectors, i.e. solve

( A − λE) w = 0
The number of linear independent solutions n1 is equal to the dimension of the nullspace
(or kernel) of A – λE .
The solutions of the differential equation are constructed as in (step 2) yielding n1 linearly
independent solutions for the eigenvalue λ.

(step 4) If n1 = k, then you have found all solutions for this eigenvalue.
If n1 < k, compute the nullspace of (A – λE)2, i.e. solve

9
 Ordinary Differential Equations – Linear differential equations…

( A − λE)
2
w=0

The number of linear independent solutions n2 is equal to the dimension of the nullspace
(or kernel) of (A – λE)2. Take the n2 – n1 linear independent solutions, which do not
belong to the nullspace of A – λE , i.e. ( A − λ E ) w ≠ 0 .
The corresponding solution y(t) of the differential equation is given by y(t) = exp(At)w. It
can be explicitly computed as follows:

exp ( tA ) w = exp ( t ( λ E + A − λ E ) ) w
= exp ( t λ E ) exp ( t ( A − λ E ) ) w

( )
= eλt E + t ( A − λ E ) + 12 t 2 ( A − λ E ) + ... w
2

= eλt w + teλt ( A − λ E ) w

The last equality is a consequence of ( A − λ E ) w = 0 for all n > 1. Again take real and
n

imaginary part in case of complex conjugate eigenvalues λ , λ .

(step 5) If n2 = k, then you have found all solutions for this eigenvalue.
If n2 < k, then compute the nullspace of (A – λE)3 and proceed as before.

(step 6) Repeat the procedure for the following powers of A – λE until the dimension of
the nullspace and hence the number of independent solution of the differential equation
equals k.

2. Examples. We present a few explicit computations of solution in the case of multiple

eigenvalues.

2.1. Solve the Cauchy problem

 1 2 −1  1
   
y′ =  0 1 1  y, y (0) =  2  (5.20)
0 0 1   3
   

If A denotes the matrix in (5.20) then PA(λ) = (λ – 1)3 and therefore λ = 1 is a triple
eigenvalue of A. Let {e1, e2, e3} denote the standard orthonormal basis of R3. We
compute

10
 Ordinary Differential Equations – Linear differential equations…

 0 2 −1 
 
A− E = 0 0 1 
0 0 0 
 

We see that the nullspace of A – E is spanned by e1. Hence y1(t) = exp(t) e1 solves (5.20).
We compute (A – E)2.

 0 0 2
 
( A− E)
2
=  0 0 0
 0 0 0
 

The nullspace of (A – E)2 is spanned by e1 and e2. Only e2 yields a new solution of the
differential equation. According to the general procedure described above it is given by

y2 (t ) = et e2 + tet ( A − E ) e2 = et e2 + 2tet e1

We note that since (A – E)3 = 0 its nullspace is R3. Take e3 as the third vector spanning
R3. The third linear independent solution is

 0  −1   1
  t   2 t 
y3 (t ) = e e3 + te ( A − E ) e3 + 2 t e ( A − E ) e3 = e  0  + te  1  + t e  0 
t t 1 2 t 2 t

 1 0  0
     

We have computed the fundamental matrix

 1 2t t 2 − t 
 
Y ( t ) = ( y1 (t ), y2 (t ), y3 (t ) ) = et  0 1 t 
0 0 1 


Since Y(0) = E we have Y(t) = exp(tA). The solution of the Cauchy problem is therefore

 1 2t t 2 − t   1  1 + t + 3t 2 
    
y (t ) = et  0 1 t   2  = et  2 + 3t  (5.21)
0 0
 1   3  
 3 


2.2. Solve the Cauchy problem

11
 Ordinary Differential Equations – Linear differential equations…

1 1 1 1
   
y′ =  1 1 1  y, y (0) =  2  (5.22)
   
 −2 −2 −2   − 1

If A denotes the matrix in (5.20) then PA(λ) = λ 3 and therefore λ = 0 is a triple eigenvalue
of A. The nullspace of B is spanned by the vectors

 −1  −1 
   
w1 =  1  , w2 =  0 
0 1
   

Hence y1(t) = w1 and y2(t) = w2 solve (5.22). We note that since A2 = 0 its nullspace is R3.
Take w3 = w2 = (0,1,0) as the third vector spanning R3. The third linear independent
solution is

 0  1   t 
     
y3 (t ) = e2 + tAe2 =  1  + t  1  =  1 + t 
 0   −2   −2t 
     

We have computed the fundamental matrix

 −1 − 1 t 
 
Y ( t ) = ( y1 (t ), y2 (t ), y3 (t ) ) =  1 0 1 + t 
 0 1 −2t 
 

We easily check that

1 + t t t 
 
exp ( tA ) = ( y3 (t ) − y1 (t ), y3 (t ), y2 (t ) − y1 (t ) + y3 (t ) ) =  t 1 + t t 
 −2t −2t 1 − 2t 
 

The solution of the Cauchy problem is therefore

1 + t t t   1   1 + 2t 
    
y (t ) =  t 1 + t t   2  =  2 + 2t  (5.23)
 −2t −2t 1 − 2t   −1  −1 − 4t 
    

12
 Ordinary Differential Equations – Linear differential equations…

2.3. Consider the equations of motion for two coupled pendulums

x = − ax − k ( x − y )

(5.24)
y = −ay − k ( y − x)


where a and k are positive constants. The matrix of the corresponding first-order system
for the vector ( x, x , y, y ) is given by
T

 0 1 0 0
 
−a − k 0 k 0
A=
 0 0 0 1
 
 k 0 −a − k 0 

it has degenerate purely imaginary eigenvalues ±i a , ±i a + 2k and can be solved by the

procedure described above. In this case, however, it is more convenient to “decouple” the
equations of motion by considering the equations for the sum and the difference of x and
y, which are given by

y = −a ( x + y )
x + 

(5.25)
y = − ( a + 2k ) ( x − y )
x − 


which are easily solved.

13