Math 312 Lecture Notes

Linear Two-dimensional Systems of Differential Equations

Warren Weckesser
Department of Mathematics
Colgate University

February 2005∗

In these notes, we consider the linear system of two first order differential equations
dx
= ax + by
dt (1)
dy
= cx + dy
dt
or equivalently,    
d~x x a b
= A~x, where ~x = , and A = . (2)
dt y c d

Solving the System

To solve this system, we first recall that the solution to the single equation
dx
= ax (3)
dt
is x(t) = ceat : a constant multiplied by an exponential. We also recall that the last problem of
Homework 2 was a linear system, and the solution to that problem can be written in vector form
as  y0   y0 
2 − 2 + x0 −3t
~x(t) = −t
e + e (4)
y0 0
This solution is a combination of constant vectors multiplied by an exponential. This suggests that
to solve (2), we look for a solution of the form

~x(t) = ~veλt (5)

where ~v is a constant vector, and λ is a number. We note that ~x(t) = ~0 is a solution to (2), which
we call the trivial solution. We would like to find nontrivival solutions, so we assume ~v 6= ~0.
By substituting our “guess” into the differential equation, we find
 
λ~veλt = A ~veλt = A~veλt (6)
∗
This version: 26 February 2005

1
or, after canceling eλt and rearranging,
A~v = λ~v. (7)
This is the eigenvalue problem for the matrix A. To solve (2), we must find the eigenvalues (λ) and
the corresponding eigenvectors (~v) of A.
Recall from linear algebra that an n × n matrix has at most n eigenvalues, and always has
at least one eigenvalue. Thus for our 2 × 2 system, we expect to find one or two eigenvalues. A
“typical” or “generic” 2 × 2 matrix will have two eigenvalues, and that is the case we will study.
The case where there is just one eigenvalue is also important, but we will not pursue it in this
course.
An eigenvalue may be real or complex. We note that if λ is a complex eigenvalue of A, then the
corresponding eigenvector ~v must also be complex. Moreover, by taking the complex conjugate of
both sides of (7), we find

(A~v)∗ = (λ~v)∗
(8)
A~v∗ = λ∗ ~v∗

(where z ∗ is the complex conjugate of z). The second equation follows from the first by using the
multiplicative property of the conjugate (wz) ∗ = w∗ z ∗ , and by noting that A is a real matrix, so
A∗ = A. The second equation says that λ∗ is also an eigenvalue, with a corresponding eigenvector
~v∗ . The short summary is, for a real matrix A, complex eigenvalues always occur in complex
conjugate pairs.
Finally, it is not hard to verify that if ~x1 (t) and ~x2 (t) are solutions to (2), then so is c 1 ~x1 (t) +
c2 ~x2 (t) for any constants c1 and c2 . Equation (2) is a pair of coupled first order equation, so we
expect the general solution to have two arbitrary constants. In fact, the general solution must have
the form
~x(t) = c1 ~x1 (t) + c2 ~x2 (t). (9)
We’ll find the general solution in two cases: real and distinct eigenvalues, and complex eigenvalues.

Real, distinct eigenvalues. Suppose that A has real eigenvalues λ 1 and λ2 , and λ1 6= λ2 . Let
~v1 and ~v2 be corresponding eigenvectors. The general solution is

~x(t) = c1 ~v1 eλ1 t + c2 ~v2 eλ2 t (10)

where c1 and c2 are arbitrary constants.

Complex eigenvalues. Because the matrix A is real, we know that complex eigenvalues must
occur in complex conjugate pairs. Suppose λ 1 = µ + iω, with eigenvector ~v1 = ~a + i~b (where ~a and
~b are real vectors). If we use the formula for real eigenvalues, we obtain

~x = c1~v1 eλ1 t + c2 ~v∗1 eλ1 t

∗
(11)

where z ∗ denotes the complex conjugate of z. This expression is a solution to the differential
equations. However, for arbitrary c 1 and c2 , this expression will generally be complex-valued, and
we want a real-valued solution. We derive such a solution based on the following observation: If the
matrix A has only real elements, and ~x(t) is a complex solution to the linear system of differential

2
equations, then the real and imaginary parts of ~x(t) are also solutions to the differential equation.
(We take the “imaginary part” to be the coefficient of i, so the imaginary part is also a real-valued
solution.) This claim is easily verified: Assume ~x(t) = ~p(t) + i~q(t) is a solution to d~x/dt = A~x.
Then
d~p d~q
+i = A(~p + i~q) = A~p + iA~q (12)
dt dt
Two complex expressions are equal if and only if their real and imaginary parts are equal, so this
equation implies
d~p d~q
= A~p and = A~q. (13)
dt dt
Now consider the complex solution

~x1 (t) = ~v1 eλ1 t = (~a + i~b)e(µ+iω)t (14)

To separate this into its real and imaginary parts, we use Euler’s formula: e iθ = cos θ + i sin θ.
Then

~x1 (t) = ~v1 eλ1 t

= (~a + i~b)e(µ+iω)t
= (~a + i~b)eµt eiωt
= eµt (~a + i~b)(cos ωt + i sin ωt) (15)
 
= eµt ~a cos ωt − ~b sin ωt + i(~a sin ωt + ~b cos ωt)
h  i h  i
= eµt ~a cos ωt − ~b sin ωt + i eµt ~a sin ωt + ~b cos ωt
 
The real part of this solution is ~u(t) = eµt ~a cos ωt − ~b sin ωt and the imaginary part (i.e. the
 
coefficient of i) is w~ (t) = eµt ~a sin ωt + ~b cos ωt . Each of these is real-valued, and by the obser-
vation given above, each is a solution to the differential equations. We may then write the general
solution as
   
~ (t) = c1 eµt ~a cos ωt − ~b sin ωt + c2 eµt ~a sin ωt + ~b cos ωt .
~x(t) = c1 ~u(t) + c2 w (16)

Note that we never had to write down the second (complex conjugate) eigenvalue or eigenvector.

3
Classifying the Equilibrium and Its Stability
We now consider several subcases of the real and complex cases. Our goal is to understand the
possible behaviors of solutions in the various cases. An important property of an equilibrium point
is its stability. The following describes three basic cases.

• If all solutions that start close to an equilibrium converge to the equilibrium asymptotically
as t → ∞, we say the equilibrium is asymptotically stable.

• If all solutions in sufficiently small neighborhoods of the equilibrium remain close to the
equilibrium, we say the equilibrium is stable.
Note that asymptotic stability implies stability.

• If every neighborhood of the equilibrium contains solutions arbitrarily close to the equilibrium
that leave the neighborhood, we say the equilibrium is unstable.

Real, distinct eigenvalues. Let λ1 and λ2 be real eigenvalues, λ1 6= λ2 , and let ~v1 and ~v2 be
respective eigenvectors. We consider the following cases.

1. λ1 < λ2 < 00.

When both eigenvalues are negative, it is clear from (10) that lim ~x(t) = ~0. All trajectories
t→∞
approach ~0 asymptotically.
The equilibrium ~0 is called a sink. It is asymptotically stable.

2. 0 < λ1 < λ2 .
In this case, all solutions diverge from the origin. However, for any solution ~x(t), we have
lim ~x(t) = ~0, so all solutions converge to ~0 backwards in time.
t→−∞

The equilibrium ~0 is called a source. It is unstable.

3. λ1 < 0 < λ2 .
Looking at (10), we see that if c2 = 0, the solution will converge to ~0 as t → ∞. If c2 6= 0,
then eventually the term c2 ~v2 eλ2 t will dominate, and the solution will grow exponentially.
Thus, there are solutions that converge to ~0, but “most” (i.e. those with c2 6= 0) will not.
The equilibrium ~0 is called a saddle point, and it is unstable.

4. Zero eigenvalue. It is also possible that one of the eigenvalues is zero. The analysis of this
case is left as an exercise.

We can say a little more about cases 1 and 2. Consider the sink, where λ 1 < λ2 < 0. We can
write the general solution as
h i
~x(t) = eλ2 t c1~v1 e(λ1 −λ2 )t + c2~v2 (17)

Suppose c2 6= 0. Since λ1 − λ2 < 0, the first term in the square brackets goes to zero as t increases,
and for large positive t we have
~x(t) ≈ c2~v2 eλ2 t . (18)

4
This says that as the solution converges to zero, it does so along the direction of ~v2 . That is, ~x(t)
will approach ~0 along a curve that is tangent to the eigenvector ~v2 .
In the case of a source, where 0 < λ1 < λ2 , we can write the general solution as
h i
~x(t) = eλ1 t c1~v1 + c2 ~v2 e(λ2 −λ1 )t (19)

Suppose c1 6= 0. For large negative t, the second term in the square brackets goes to zero while the
first is constant, so we have
~x(t) ≈ c1~v1 eλ1 t . (20)
This says that as t → −∞, the solution converges to ~0 along the direction of ~v1 . That is, ~x(t) will
approach ~0 (as t → −∞) along a curve that is tangent to the eigenvector ~v1 .
We can summarize these observations by saying that in the case of a source or sink, curved
trajectories in the phase plane approach ~0 along curves that are tangent to the eigenvector of the
eigenvalue closest to zero. By stating it this way, we don’t have to worry about which eigenvalue
we called λ1 or λ2 . This observation will be very useful when we sketch phase portraits.

Complex eigenvalues. We assume one eigenvalue is λ 1 = µ+iω, and a corresponding eigenvector

is ~v1 = ~a + i~b. We can rewrite the general solution (16) as
h    i
~x(t) = eµt c1 ~a cos ωt − ~b sin ωt + c2 ~a sin ωt + ~b cos ωt (21)

Note that the only t dependence in the expression in the square brackets is through sin ωt and
cos ωt. Thus the expression in the square brackets is periodic, with period T = 2π/ω. All solutions
therefore have an oscillatory behavior, but whether the oscillation grows or decays depends on the
sign of µ. We have the following subcases.

1. µ < 00.
In this case, the amplitude of the oscillation decays exponentially. The trajectories spiral into
the origin as t increases, and lim ~x(t) = ~0.
t→∞

The equilibrium ~0 is called a spiral sink. It is asymptotically stable.

2. µ = 00.
Since e0 = 1, the solution is periodic. In the phase plane, the trajectories are ellipses.
The equilibrium ~0 is called a center, and it is stable.

3. µ > 00.
In this case, the amplitude of the oscillation grows exponentially. The trajectories spiral out
of the origin as t increases. Considering backwards time, we have lim ~x(t) = ~0.
t→−∞

The equilibrium ~0 is called a spiral source. It is unstable.

5
Sketching Phase Portraits for Planar Linear Systems
We summarize the procedure for sketching the phase portraits for the linear system
   
d~x x a b
= A~x, where ~x = , and A = .
dt y c d

Procedure

1. Find the eigenvalues of the matrix, and classify the equilibrium as a saddle, sink, source,
spiral source, spiral sink, or center. (There are a few other special cases that we did not
cover.)

2. If the eigenvalues are real, find the associated eigenvectors, and draw the straight-line solu-
tions.

3. Sketch the nullclines, and indicate the directions of the vector field on the nullclines.
The x nullcline is the line where dxdt = 0, so ax + by = 0. Along this line, the vector field is
vertical, so trajectories must have a vertical tangent when they cross this line.
The y nullcline is the line where dydt = 0, so cx + dy = 0. Along this line,the vector field is
horizontial, so trajectories must have a horizontal tangent when they cross this line.

4. Indicate the direction of the vector field on the x and y axes. Note that at the points (1, 0)
and (0, 1), the vector field is given by the first and second columns of A, respectively.

5. If the eigenvalues are real, distinct, and of the same sign, all trajectories except the straight-
line solutons will approach the origin tangent to the eigenvector of the eigenvalue closest to
zero.

6. Use the above information to sketch several representative trajectories.

6
Examples
In the following examples, we will find the general solution and sketch the phase portrait for the
given system of differential equations.

Example 1. 1 
d~x 2 −1
= A~x, where A =
dt 1 −1
or equivalently,
dx 1
= x−y
dt 2
dy
=x−y
dt
√
7
1. The characteristic polynomial is λ 2 + 21 λ + 12 , and the eigenvalues are λ = − 41 ± i 4 . The
eigenvalues are complex, and the real part is negative, so the origin is a spiral sink.

2. In the complex case, we don’t need the eigenvectors to sketch the phase portrait, but we do
need an eigenvector
√ ~v1 associated with √ λ1 = µ + iω to find the general solution. In this case,
λ1 = − 14 + i 47 , so µ = − 14 and ω = 47 , and
  √   " √
1 1 7
#
2 − − 4 + i 4 −1 3
− i 7
−1
A − λ1 I =   √  = 4 4 √ (22)
1 −1 − − 41 + i 47 1 − 34 − i 47

From the first row of A − λ1 I we see that we can choose

" #   " #
1√ 1 0√
~v1 = 3 7 = 3 +i (23)
4 −i 4 4 − 47

So " #
0√
 
1
~a = 3 and ~b = (24)
4 − 47
The general solution is then
(  " # )
1 √  0 √ 
~x(t) = c1 e−(1/4)t 3 cos 7t/4 − √
7 sin 7t/4
4 − 4
(  " # )
1 √  0 √ 
+ c2 e−(1/4)t 3 sin 7t/4 + √
7 cos 7t/4 (25)
4 − 4

3. The x nullcline is
1 1
x − y = 0, or y= x.
2 2
At the point (1, 1/2), we have dy
dt = 1/2 > 0, so the vectors on the x nullcline in the first
quadrant point in the positive y direction. (Since this problem is linear, they must point in
the opposite direction in the opposite quadrant.)

7
 The y nullcline is
x − y = 0, or y = x.
At the point (1, 1), we have dx
dt = −1/2, so the vectors on the y nullcline in the first quadrant
point in the negative x direction.
   
1/2 −1
4. The vector field at (1, 0) is , and at (0, 1) it is . Vectors in these directions are
1 −1
drawn on the positive x and y axes, respectively (and vector in the opposite directions are
drawn on the negative x and y axes).

5. (The eigenvalues are complex, so this guideline is not applicable.)

6. Here is the phase portrait. The dashed lines are the nullclines. We know the equilibrium is
a spiral sink, so the trajectories will spiral into the origin. We use the vector field on the
nullclines and the axes to make a “pretty good” sketch of two of the spiral trajectories.

−1 1

−1

8
Example 2. We consider the linear system of differential equations with
 
4 −2
A=
3 −3

1. The characteristic polynomial is λ 2 − λ − 6, and the eigenvalues are λ1 = −2 and λ2 = 3. We

have real eigenvalues with opposite signs, so ~0 is a saddle point.
   
1 2
2. The corresponding eigenvectors are ~v1 = and ~v2 = . These will correspond to
3 1
straight-line solutions with slope 3 and 1/2, respectively, in the phase plane.
The general solution is    
1 −2t 2 3t
~x(t) = c1 e + c2 e . (26)
3 1

3. The x nullcline is 4x − 2y = 0, or y = 2x, and the y nullcline is 3x − 3y = 0, or y = x. These

are shown as dashed lines in the phase plane.
   
4 −2
4. The vector field at (1, 0) is , and the vector field at (0, 1) is . Vectors in the same
3 −3
direction as these are shown on the positive x and y axes, respectively. (They are in the
opposite directions on the negative axes.

5. (This guideline is not applicable, since the eigenvalues have opposite signs.)

6. Here is the phase portrait:

−1 1

−1

9
Example 3. We’ll look at one more example, this time with
 
4 2
A=
1 3
1. The characteristic polynomial is λ 2 − 7λ + 10, and the eigenvalues are λ1 = 2 and λ2 = 5.
We have positive real eigenvalues, so ~0 is a source.
   
1 2
2. The corresponding eigenvectors are ~v1 = and ~v2 = . These will correspond to
−1 1
straight-line solutions with slope −1 and 1/2, respectively, in the phase plane.
The general solution is    
1 2t 2 5t
~x(t) = c1 e + c2 e . (27)
−1 1
3. The x nullcline is 4x + 2y = 0, or y = −2x, and the y nullcline is x + 3y = 0, or y = −x/3.
These are the dashed lines in the phase plane.
   
4 1
4. The vector field at (1, 0) is , and the vector field at (0, 1) is . Vectors in the same
1 3
direction as these are shown on the positive x and y axes, respectively.
5. We have two distinct eigenvalues with the same sign, so we know that the equilibrium is a
source, and all trajectories (except the straight-line solution associated with the bigger eigen-
value) will come out of the origin tangent to the eigenvector associated with the eigenvalue
 
1
closest to zero. In this case, the trajectories will be tangent to the direction of ~v1 = .
−1
6. Here is the phase portrait:

−1 1

−1

10