Rose-Hulman Undergraduate Mathematics Journal

Volume 16 Article 5
Issue 2

Linearization and Stability Analysis of Nonlinear Problems

Robert Morgan
Wayne State University

Follow this and additional works at: https://scholar.rose-hulman.edu/rhumj

Recommended Citation
Morgan, Robert (2015) "Linearization and Stability Analysis of Nonlinear Problems," Rose-Hulman
Undergraduate Mathematics Journal: Vol. 16 : Iss. 2 , Article 5.
Available at: https://scholar.rose-hulman.edu/rhumj/vol16/iss2/5
Rose-
Hulman
Undergraduate
Mathematics
Journal

Linearization and Stability

Analysis of Nonlinear Problems

Robert Morgana

Volume 16, No. 2, Fall 2015

Sponsored by

Rose-Hulman Institute of Technology

Department of Mathematics
Terre Haute, IN 47803
Email: [email protected]
a Wayne State University, Detroit, MI
http://www.rose-hulman.edu/mathjournal
 Rose-Hulman Undergraduate Mathematics Journal
Volume 16, No. 2, Fall 2015

Linearization and Stability Analysis of

Nonlinear Problems

Robert Morgan

Abstract. The focus of this paper is on the use of linearization techniques and lin-
ear differential equation theory to analyze nonlinear differential equations. Often,
mathematical models of real-world phenomena are formulated in terms of systems
of nonlinear differential equations, which can be difficult to solve explicitly. To
overcome this barrier, we take a qualitative approach to the analysis of solutions
to nonlinear systems by making phase portraits and using stability analysis. We
demonstrate these techniques in the analysis of two systems of nonlinear differential
equations. Both of these models are originally motivated by population models in
biology when solutions are required to be non-negative, but the ODEs can be un-
derstood outside of this traditional scope of population models. In fact, allowing
solutions for these equations to be negative provides some very interesting math-
ematical problems, and demonstrates the utility of the analysis techniques to be
described in this article. We provide stability analysis, phase portraits, and numer-
ical solutions for these models that characterize behaviors of solutions based only
on the parameters used in the formulation of the systems. The first part of this pa-
per gives a survey of standard linearization techniques in ODE theory. The second
part of the paper presents applications of these techniques to particular systems
of nonlinear ODEs, which includes some original results by extending the analy-
sis of solutions lying anywhere in the plane R2 , rather than only those in the first
quadrant.

Acknowledgements: The author would like to thank his advisor, Dr. Jarod Hart, for
his patient and thoughtful guidance throughout the writing of this paper. This paper and
the author’s involvement in other research projects would not have been possible without
him. The author would also like to thank the referee for insightful comments that improved
the paper.
Page 68 RHIT Undergrad. Math. J., Vol. 16, No. 2

1 Introduction
The focus of this paper is on the use of linearization techniques and linear differential equation
theory to analyze nonlinear differential equations. Often, mathematical models of real-world
phenomena are formulated in terms of systems of nonlinear differential equations, which can
be difficult to solve explicitly. To overcome this barrier, we take a qualitative approach to
the analysis of solutions to nonlinear systems by making phase portraits and using stability
analysis. We demonstrate these techniques in the analysis of two systems of nonlinear
differential equations. Both of these models are originally motivated by population models
in biology when solutions are required to be non-negative, but the ordinary differential
equations (ODEs) can be understood outside of this traditional scope. In fact, allowing
solutions for these equation to be negative provides some very interesting mathematical
problems, and demonstrates the utility of the analysis techniques to be described in this
article.
Generally, given any system of first-order ODEs (possibly nonlinear), we want to create a
qualitative characterization of the behavior of solutions depending on their initial conditions.
We do this by first finding equilibrium solutions, and then using stability analysis to make
a phase portrait for the system that indicates the general behavior of solutions depending
on their initial conditions. The stability analysis carried out at each equilibrium involves
linearization techniques and solution methods for systems of linear ODEs, which further
involve ideas from linear algebra, including computing eigenvalues and eigenvectors, among
other topics. There are some foundational ideas that we will not discuss in depth so that we
can move more quickly to the focus of our work. The theory behind these topics is crucial
to our work, but our analysis begins where the theory ends.
Although most of the theory discussed in this paper can be extended to arbitrary di-
mensions, for simplicity we will restrict to analysis in two dimensions. For more information
on the extension to arbitrary dimensions, see for example the text by Boyce and DiPrima
[BD]. We further remark that although we limit our analysis to ODEs, linearization tech-
niques also play a role in the solution of partial differential equations (PDEs). Though there
are differences between the application of linearization techniques to ODEs and PDEs, the
ideas of simplifying the problem at hand and analyzing eigenvalues to gain a general under-
standing of the solutions are shared. This can be observed, for instance, in separation of
variable techniques applied to solve the heat equation and the wave equation. See the text
by Haberman [H] for examples, and the references therein for more information.
Our article is organized as follows. In Section 2, we establish some general notation,
formulate ODE problems, and discuss the pertinent existence and uniqueness theory. In
Section 3, we discuss explicitly solving systems of linear ODEs and linearization of nonlinear
systems of ODEs. Section 4 is used to define and describe stability of linear and nonlinear
systems. Finally, in Section 5, we use the linearization and stability analysis described in the
previous sections to analyze two systems of ODEs, which includes phase portraits, stability
analysis of equilibrium solutions, and numerical simulations.
All graphs were generated in Matlab using open source software written by J. C. Polking
RHIT Undergrad. Math. J., Vol. 16, No. 2 Page 69

and J. Castellanos [PC], and available http://math.rice.edu/~dfield/dfpp.html.

2 Systems of Ordinary Differential Equations

An autonomous first order system of differential equations is of the form

x01 (t) = F1 (x1 (t), x2 (t)),

(1)
x02 (t) = F2 (x1 (t), x2 (t)),

or simply x0 = F (x), where F (x) = (F1 (x), F2 (x)) is defined for x = (x1 , x2 ) ∈ R2 . Here the
variables x1 = x1 (t) and x2 = x2 (t) are all parameterized by a single, independent variable
t ∈ R, typically chosen to represent time. A solution to the system of ODEs is of the form
x = φ(t), where
x1 = φ1 (t), x2 = φ2 (t),

each function φj is differentiable and φ0j (t) = Fj (φ1 (t), φ2 (t)) for j = 1, 2. In this work, we
will only work with autonomous systems of ODEs. There are non-autonomous systems of
ODEs as well, which are formulated by allowing the functions F1 , F2 to depend on t and
x1 , x2 , but this is outside the scope of our analysis.
For a given system of ODEs of the form (1), an initial value problem (IVP) associated
with the system is formulated by the equations in (1) combined with an initial condition
x(0) = x0 for some fixed x0 = (x01 , x02 ) ∈ R2 .

2.1 Existence and Uniqueness of Solutions

Before analyzing systems of ordinary differential equations, we should first establish that
the ODE has a solution for us to further analyze. This is the purpose of the existence
and uniqueness theory for systems of differential equations. The appropriate existence and
uniqueness result for our setting is the following theorem, which can be found as stated here
in the text by Boyce and DiPrima [BD].

Theorem 1. Let each of the functions Fj for j = 1, 2 as well as their partial derivatives
∂Fj /∂xk for j = 1, 2 and k = 1, 2 be continuous in a region R of t, x1 , x2 space defined by
α < t < β, α1 < x1 < β1 , α2 < x2 < β2 , and let the point (t0 , x01 , x02 ) be in R. Then there is
an interval |t − t0 | < h in which there exists a unique solution x1 = φ1 (t), x2 = φ2 (t) of the
system of differential equations that also satisfies the initial condition x(t0 ) = (x01 , x02 ).

While more general versions of this theorem do exist, for our purposes it is enough to
guarantee that the systems we work with have unique solutions. Thus, we’ve assured the
existence and uniqueness of solutions to our systems of ODEs. So we can move forward with
our analysis of systems of ordinary differential equations.
Page 70 RHIT Undergrad. Math. J., Vol. 16, No. 2

2.2 Linear Systems of ODEs

Linear systems of first-order ordinary differential equations are of the form x0 = F (x), where
each of the functions F1 , F2 describe linear relationships between x1 , x2 . In this scenario, we
write x0 = F (x) as x0 = Ax where A is a matrix such that each entry aij corresponds to the
coefficient of xj in the function Fi ; for this reason A is referred to as the coefficient matrix
of the linear system. Thus, linear systems can be written as

x01
     
x1 a11 x1 + a12 x2
=A = , (2)
x02 x2 a21 x1 + a22 x2

where aij is the ith row and j th column entry of A. If one were to allow the matrix A to
depend on t, the system would be a non-autonomous linear system of ODEs, which again
we will not address in this paper.

2.3 Nonlinear Systems of ODEs

In the last section, we described systems of linear differential equations, but many of the
problems in ODE theory come from nonlinear equations. If the variables in the equations
from (1) do not exhibit linear relationships, then the system is called nonlinear. As we
will see, nonlinear systems can be very complicated, if not impossible, to solve explicitly;
however, when it comes to modeling real-world phenomena, the majority of the systems
that arise are nonlinear. To be able to analyze these systems we will linearize them at their
equilibria, and then construct phase portraits to visualize the trajectories of the solutions
to the system. It is in this way that we can overcome the barrier of not being able to solve
nonlinear systems explicitly. The remainder of this paper will discuss this process in great
detail.

2.4 Equilibria
A relatively simple aspect of systems of ODEs that can be observed without a great deal of
analysis is the locations of the equilibria. Equilibria for a system of the form (1), which are
sometimes also referred to as critical points of the system, are points X ∈ R2 where

F1 (X) = F2 (X) = 0.

It is easy to notice that if X ∈ R2 is an equilibrium solution of the system (1), then the
constant functions x1 (t) = X1 , x2 (t) = X2 define a solution x(t) = (X1 , X2 ) to the system
of equations (hence the name equilibrium solution). Finding these equilibrium solutions
corresponds to solving the nonlinear algebraic equations F (x) = 0. This process starts to
create a picture of the behavior of solutions to a system like (1).
RHIT Undergrad. Math. J., Vol. 16, No. 2 Page 71

2.5 Stability
The majority of our analysis of systems of ODEs will focus on whether or not the systems
have stable equilibria. We characterize an equilibrium as stable or unstable based on the
behavior of solutions whose initial conditions are in the neighborhood of the equilibrium.
If solutions near a critical point of a system stay close to the critical point as t approaches
infinity, we think of the critical point as stable; if this condition is not met then the critical
point is unstable. Furthermore, we call a stable critical point asymptotically stable if, over
time, the solutions actually approach the critical point as opposed to simply staying within
a certain radius. More precisely, we give the following standard definition that can be found
in the text by Boyce and DiPrima [BD]:

Definition 1. Let X ∈ R2 be a critical point of a system of ODEs of the form (1).

• The critical point X is stable if, for any  > 0, there is a δ > 0 such that if a solution
x = φ(t) satisfies ||φ(0) − X|| < δ, then

||φ(t) − X|| <  for all t > 0.

p
Here ||x|| = x21 + x22 denotes the Euclidean norm on R2 .

• The critical point X is unstable if it is not stable as defined above.

• The critical point X is asymptotically stable if there exists a δ > 0 such that if a
solution x = φ(t) satisfies ||φ(0) − X|| < δ, then

lim φ(t) = X.
t→∞

Note that in our definition of stable equilibrium and asymptotically stable equilibrium,
we implicitly assume that the solution φ(t) exists for all time t > 0. The existence of a
solution for all t > 0 given an initial condition x(0) = x0 is not guaranteed to us by the
existence and uniqueness theorem stated earlier in this section; the theorem only assures us
that a solution exists locally in time. For us, the existence of φ(t) for all time t > 0 for
solutions near a stable or asymptotically stable equilibrium is a matter of definition; if an
equilibrium point X is stable (or asymptotically stable), then there must be a δ > 0 for
which any solution φ(t), satisfying ||φ(0) − X|| < δ, must exist for all time in addition to the
stability conditions above. One can delve further into the global existence theory of systems
of ODEs to develop a more complete picture of existence of solutions, but for our purposes
the above definition is sufficient and going deeper into the general theory would detract
from the main purpose of this article. So we leave our discussion of stability, existence, and
uniqueness at that. With our definitions now set, we can now begin to look at systems of
ODEs and classify their equilibria based on the behaviors of solutions.
Page 72 RHIT Undergrad. Math. J., Vol. 16, No. 2

3 Methods of Analysis
Our main goal when analyzing systems of ordinary differential equations is to gain an under-
standing of the behaviors of the solutions to the systems. The natural approach for analyzing
a system is to solve it explicitly, and this method works well if the system is linear. If the
system is not linear, then solving explicitly can be very complicated (maybe impossible), so
we instead linearize the system at its equilibria and gain a qualitative understanding of the
solutions by analyzing the linearized system.

3.1 Solving Explicitly

To demonstrate how to solve a linear system of ODEs and to introduce some general termi-
nology about types of equilibria, we will work through some examples of linear systems in
two dimensions. Let’s begin with the simple system
 
0 1 1
x = x. (3)
16 1
| {z }
A

It can be shown that solutions to (3) are of the form

c1 e5t + c2 e−3t
 
(1) λ1 t (2) λ2 t
x = c1 ξ e + c2 ξ e = , (4)
4c1 e5t − 4c2 e−3t
where λ1 = 5 and λ2 = −3 are the eigenvalues of A, ξ (1) = (1, 4) and ξ (2) = (1, −4) are the
eigenvectors of A, and c1 , c2 ∈ R are arbitrary constants. See the text by Boyce and DiPrima
[BD] for more information about these techniques.
Overall, solutions with initial conditions near the critical point do not stay close to the
critical point, so we call this equilibrium unstable, and furthermore we call this specific type
(one positive eigenvalue and one negative eigenvalue) of point a saddle point. Two other
types of critical points that can arise from real eigenvalues of A are sources and sinks, which
are influenced by the eigenvalues of A in the same way saddle points are. Considering a
linear system of the form x0 = Ax, if both the eigenvalues λ1 and λ2 are positive, then all
solutions to the IVP with a nonzero initial condition tend to infinity in magnitude as t tends
to infinity, so we call the equilibrium at the origin a source. When both λ1 < 0 and λ2 < 0,
we call the equilibrium at the origin a sink, and in this situation all solutions tend to zero
as t tends to infinity because of the exponential decay that is introduced into the expression
of the solutions.
Next we discuss the situation where the eigenvalues of A are complex. This will, in turn,
lead to two new types of critical points: spirals and centers. We will demonstrate the former
with an example here and the latter will be presented in a later section.
Consider the system:
 
0 −1/2 1
x = x. (5)
−1 −1/2
RHIT Undergrad. Math. J., Vol. 16, No. 2 Page 73

The eigenvalues for this system are λ1 = −1/2 + i and λ2 = −1/2 − i. Finding the cor-
responding eigenvectors ξ (1) , ξ (2) yields two possible solutions of the form x(j) = ξ (j) eλj t for
j = 1, 2:
   
(1) 1 (−1/2+i)t (2) 1
x = e and x = e(−1/2−i)t .
i −i

However, in this case, the real and imaginary parts of x(2) are linear combinations of the real
and imaginary parts of x(1) , so we can express the whole solution with just x(1) in the form
Re x(1) + iIm x(1) , or more simply as x(1) = u(t) + iv(t):
 −t/2   −t/2 
(1) e cos t e sin t
x = +i .
−e−t/2 sin t e−t/2 cos t
| {z } | {z }
u(t) v(t)

As a result of this discussion, any solution to (5) can be written as a linear combination of
u and v.
While our phase portrait of the system will reveal its spiral nature, it is clear that the
solutions differ from the previous straight-line solutions because of the trigonometric factors
that are involved. The direction of rotation can also be determined based on aij . In fact, with
some elementary computations we can see that the solutions spiral in a clockwise direction
when a12 > a21 , and in a counterclockwise direction when a12 < a21 . Thus, for this system,
since a12 = 1 > −1 = a21 , the solutions will rotate clockwise about the critical point at the
origin.
In Figure 1b below, we have included a phase portrait for the system of equations in (5),
along with a few plots of numerical solutions. This phase portrait allows us to see the actual
behaviors of solutions to system (5). We should note here that the eigenvalues of A alone

Figure 1

(a) A phase portrait of system (3) (b) A phase portrait of system (5)
Page 74 RHIT Undergrad. Math. J., Vol. 16, No. 2

determine the limiting behavior of solutions to the linear system as t tends to infinity. This
behavior can be determined independent of the eigenvectors, and we will rely on analyzing
the eigenvalues of a linear system when solving explicitly becomes too time consuming or
impossible. This occurrence tends to manifest when we are faced with a nonlinear system,
and therefore, linearization is another useful tool in the analysis of systems of first-order
ODEs that allows us to apply an understanding of eigenvalues to nonlinear systems of ODEs.

3.2 Linearization
In the context of this paper, our goal is to begin with a nonlinear system, linearize it at its
critical points, then use linear ODE techniques to understand the approximate behavior of
solutions to the linearized system, and finally apply our understanding of the behavior to
the nonlinear system. As detailed in the previous sections, solving the system x0 = Ax is a
straightforward process, but the situation is severely complicated if the functions F1 , F2 in
(1) are not linear. In this section we discuss a technique known as linearization for systems
of ODEs, which is a way to approximate nonlinear systems much in the same way that a
tangent line can be used to approximate a smooth function.
Consider again the system of ODEs in (1), where the functions F1 , F2 can be nonlinear in
the variables x1 , x2 . We also assume that F1 , F2 are continuously differentiable so that our
linearization techniques are applicable. Given an equilibrium solution X, we linearize the
system at the equilibrium by replacing x by u = x − X and replacing Fj (x) by its tangent
plane at X for each j. Note that Fj (X) = 0 since X is an equilibrium solution to the system,
and hence the linearized system is given by the following system of ODEs:
 0    
u1 ∂x1 F1 (X) ∂x2 F1 (X) u1
= , (6)
u02 ∂x1 F2 (X) ∂x2 F2 (X) u2
| {z }
A

where uj (t) = xj (t) − Xj for j = 1, 2. Here, we use the partial derivatives of F1 , F2 to

linearize the system; that is, the tangent plane of Fj at X is

y = Fj (X) + ∇Fj (X) · (x − X) = ∂x1 Fj (X)u1 + ∂x2 Fj (X)u2 .

Note that Fj (X) = 0 here since X is an equilibrium solution. The formula can also be
written as the total derivative of the nonlinear system:

u0 = DF (X) u,
| {z }
A

where DF (x) = ∂(F 1 ,F2 )

∂(x1 ,x2 )
(x) and X ∈ R2 is a critical point of the system.
We note here that the method of linearization is not applicable to all nonlinear, first-order
systems of ODEs. The matrix of partial derivatives, A, must be defined at the equilibria. To
guarantee the existence of A, F1 and F2 must be differentiable, and they must be continuously
differentiable on an open neighborhood of the critical point for much of the existence and
RHIT Undergrad. Math. J., Vol. 16, No. 2 Page 75

uniqueness theory for ODEs to apply. There are also issues that arise with the lack of
invertibility, or equivalently vanishing eigenvalues, of A when the critical point is not isolated.
This situation presents itself below when we encounter an infinite number of critical points
that are not isolated.

4 Stability of Linear and Nonlinear Systems

At this point, we have the tools to analyze the equilibria of a system of ODEs based on
the coefficient matrix A of either a linear or linearized system. We have seen how different
systems can exhibit different types of behaviors at certain critical points, and we have seen
that the eigenvalues of A are the determining factors in which type of behavior a critical
point will exhibit. Now, we can concisely classify the types of equilibria a system will have
as well as their stability based solely on the eigenvalues of the matrix A.

4.1 Classifying Equilibria

In general, we can predict the behavior of the solutions to a linear or linearized system
based on the eigenvalues of the matrix A. As discussed earlier, negative eigenvalues cause
a solution to tend to zero, positive eigenvalues cause the solution to approach infinity, and
imaginary eigenvalues lead to a spiraling behavior of solutions.
For a concise description of the different possibilities for eigenvalues and their correspond-
ing effects on the end behaviors of solutions, see the table below. We omit the case where one
eigenvalue is zero and the other less than or equal to zero, however, it is possible to continue
the same type of analysis for this situation. We also omit discussion of explicit solutions to
x0 = Ax when A has repeated eigenvalues. These solutions can be found, but involve ideas
of Jordan forms for the matrix A. See, for example, the text by Boyce and DiPrima [BD]
for more information on these solutions.

Eigenvalues Type of Critical Point Possible end behavior for x0 6= 0

λ1 ≥ λ2 > 0 Source ||x(t)|| → ∞
λ1 ≤ λ2 < 0 Sink ||x(t)|| → 0
λ1 > 0 > λ2 Saddle Point ||x(t)|| → 0 or ||x(t)|| → ∞
λ1 , λ2 = r ± iµ and r > 0 Outward Spiral ||x(t)|| → ∞
λ1 , λ2 = r ± iµ and r < 0 Inward Spiral ||x(t)|| → 0
λ1 , λ2 = ±iµ Center ||x(t)|| = c for some c ≥ 0

For the last line of this table, corresponding to a center, one can conclude that ||x(t)|| = c
by direct computation since the solution is given by
   
cos(µt) sin(µt)
x(t) = c1 + c2
− sin(µt) cos(µt)
Page 76 RHIT Undergrad. Math. J., Vol. 16, No. 2

and

||x(t)||2 = (c1 cos(µt) + c2 sin(µt))2 + (−c1 sin(µt) + c2 cos(µt))2 = c21 + c22 = c2 .

Because linearization is a good approximation in regions very close to the equilibria, our
analysis can also be applied to the nonlinear system in the neighborhood of each critical
point. However, when it comes to stability, it is slightly more difficult to directly link the
outcomes of the linear analysis to the behavior of the nonlinear system. We are able to
make some conclusions about stability of nonlinear equations based on the stability of the
associated linearization, but in some situations the linearization is not sufficient to make
such conclusions.

4.2 Classifying Stability

In the context of this paper, we have seen how the eigenvalues of A influence the behavior
of solutions to the linear system, but now we can apply our analysis of the eigenvalues to
the stability of the critical points of linear and nonlinear systems. In general, based on our
conditions of stability, the types of critical points which attract solutions to them (sinks and
inward spirals) are asymptotically stable, the types of critical points which have any sort
of repelling nature (sources, saddles, and outward spirals) are unstable, and the remaining
type of critical point (centers) only meets the criterion of a stable critical point. Using our
criterion of stability, we can adapt the table from the previous section to include stability
information about each type of critical point.
Our next task, then, is to broaden our understanding of stability to a nonlinear system.
We introduce a little notation first. Given a system of ODEs x0 = F (x) with critical point
X ∈ R2 , we say that the system is linearly stable, linearly unstable, or linearly asymptotically
stable if the associated linearized system at X is stable, unstable, or asymptotically stable
respectively. We will need the next theorem to make stability conclusions about nonlinear
systems based on stability analysis of associated linear systems. This theorem can be found
for example in the text by Boyce and DiPrima [BD].
Theorem 2. Assume F (x) is a continuously differentiable function on R2 , and that X is
an equilibrium solution of x0 = F (x). If the linear system x0 = DF (X)x is asymptotically
stable, then the system is asymptotically stable at x = X. If the linear system x0 = DF (X)x
is unstable, then the system is unstable at X.
Therefore, looking at the eigenvalues of the linear system is a good way to approximate
the end behavior of solutions to the nonlinear system in most scenarios. However, because
we cannot conclude asymptotic stability, stability, or instability for a critical point whose
linearization yields a center using Theorem 2, we cannot conclude anything about its sta-
bility in a nonlinear system without going much more in depth. Hence, we can rewrite our
eigenvalue table, and here we use the term “linear stability” to describe the stability of crit-
ical points in linear or linearized system whereas the term “nonlinear stability” refers to the
stabilities of critical points in a nonlinear system.
RHIT Undergrad. Math. J., Vol. 16, No. 2 Page 77

Eigenvalues of DF (X) Type of Critical Point Linear Stability Nonlinear Stability

λ1 ≥ λ2 > 0 Source Unstable Unstable
λ1 ≤ λ2 < 0 Sink Asymptotically Stable Asymptotically Stable
λ1 > 0 > λ2 Saddle Point Unstable Unstable
λ1 , λ2 = r ± iµ, r > 0 Outward Spiral Unstable Unstable
λ1 , λ2 = r ± iµ, r < 0 Inward Spiral Asymptotically Stable Asymptotically Stable
λ1 , λ2 = ±iµ Center Stable Indeterminate

The classification of equilibria as different types of critical points and as stable or unstable,
which is summarized in the table above, is one of the goals of this paper. Now, using linear
differential equation theory, we are able to understand a great deal about a system based
solely upon the eigenvalues of its coefficient matrix. We will use this sort of analysis in the
following phase-portrait characterizations.

5 Phase-Portrait Characterizations
Understanding linear differential equation theory makes it possible to, given a complicated
nonlinear system, linearize it at its equilibria, classify those equilibria based on the eigen-
values of the system’s coefficient matrix A, and determine the stability of those equilibria as
well as the behaviors of solutions to the system near those equilibria. Therefore, using these
techniques, we will analyze two nonlinear models to predict the behaviors of the systems
based upon certain parameters. The models we will present can be initially understood in
terms of population models in biology. In those models, it is typically only used when solu-
tions that lie in the first quadrant, which correspond to nonnegative values for populations,
but we will use our analysis techniques to create phase portraits for these systems on the
entire plane. Extending our analysis to all solutions in the plane for these equations is a
new contribution to ODE theory, which involves a substantial amount of work resulting in
interesting applications of the linearization techniques and new phase plane characterizations
for the systems of ODEs introduced later in the this section.

5.1 A Simple Nonlinear System of ODEs

In this section we use the techniques detailed in the previous sections to study the following
system of ODEs,

dx/dt = x(a − αy),

(7)
dy/dt = y(−c + γx).

Here we assume that the parameters a, c, α, and γ are positive. This system is commonly
used as a predator-prey model in biology when the variables x and y are restricted to be
nonnegative. In this context, the system (7) is known as the Lotka-Volterra equations. Given
an initial condition (x0 , y0 ) with nonnegative entries, this system can be interpreted as a
Page 78 RHIT Undergrad. Math. J., Vol. 16, No. 2

population model where x represents the population of the prey, y represents the population
of the predator, and the parameters α and γ are determined by the birth/death rates and
interaction of the two species. More information on this can be found in [BD] and the
references therein. For the purposes of our analysis, we simply assume a, c, α, γ > 0 and
consider solutions of the equation that allow x and y to be arbitrary real-valued solutions,
with initial conditions (x0 , y0 ) lying anywhere in the plane R2 . All of the results discussed in
this section pertaining to nonnegative solutions to (7), as well as a more detailed explanation
of the physical meanings of the parameters can be found in the text by Boyce and DiPrima
[BD].
To analyze system in (7) it helps to approach it in more general terms first, and then
to look at specific situations. Looking at system (7), one can see that there are critical
points at (0, 0) and at (c/γ, a/α), and applying the linearization in (6) at those points one
can observe the approximate behavior of the solutions over time when an initial condition is
close to (0, 0) or (c/γ, a/α). We start by computing the derivative matrix
 
a − αy −αx
DF (x, y) = . (8)
γy −c + γx
The linearization at the equilibrium solution (0, 0) is given by DF (0, 0),
 0    
u a 0 u
= .
v0 0 −c v
The eigenvalues for this system, then, are λ1 = a and λ2 = −c. Because a and c are both
positive, it becomes clear that (0, 0) is a saddle point (since one eigenvalue is positive and
the other is negative). Hence (0, 0) is an unstable equilibrium of the predator-prey system.
At the point (c/γ, a/α), the system can be analyzed in a similar fashion. The linearized
system is
0 −cα
 0    
u γ u
= ,
v0 aγ
α
0 v
√ √
for which the eigenvalues are λ1 = ca i and λ2 = − ca i. Looking at our stability table, we
see that this critical point is a center; hence the equilibrium (c/γ, a/α) is linearly stable, but
we are unable to conclude the stability for the general nonlinear system at this equilibrium.
Our next step, then, is to look at the shape of the center in the phase plane to get a
better picture of how the levels of the two populations relate to one another. The shape of
the solution near (c/γ, a/α) can be determined through observing the relationship dv/du.
Starting with the relation
dv aγ 2 u
=− 2 , (9)
du cα v
we can simplify and integrate to obtain the equation
u2 v2
+ = 1,
K/(aγ 2 ) K/(cα2 )
RHIT Undergrad. Math. J., Vol. 16, No. 2 Page 79

where K is a constant of integration.

One can easily recognize the form of this equation as an ellipse, and it makes sense that
this form would appear as it corresponds to the general shape of the center. Looking at the
equation for the ellipse, now, we can assess the placement of the major and minor axes based
on the values for aγ 2 and cα2 . That is, if aγ 2 > cα2 , then we will observe a vertical stretch
in the elliptical shape of solutions around the point (c/γ, a/α), and if aγ 2 < cα2 , then we
will observe a horizontal stretch. If aγ 2 = cα2 , then the solutions will take the shapes of
circles about the point (c/γ, a/α).
Seeing that (c/γ, a/α) is a center, we can find which way the solutions of the two-
dimensional system will move about the critical point. Based on the analysis done for
system (5), for the point (c/γ, a/α) if cαγ
− aγ
α
> 0, then the solutions will move clockwise
cα aγ
about the critical point, and if γ − α < 0, then the solutions will move counterclockwise
about the critical point.
One final aspect of the behaviors of solutions that we can identify qualitatively is the
eccentricity of the solutions. That is, the ratio of the lengths of the major axes to those of
the minor axes. If we express the length of the horizontal axis as z and the length of the
vertical axis as w, then we can define the eccentricity of the ellipse as
s
z cα2
= .
w aγ 2

Through this qualitative analysis we have now determined that (0, 0) is an unstable
equilibrium of the linearized system and that (c/γ, a/α) is a stable, but not asymptotically
stable, equilibrium of the linearized system that exhibits a center-like behavior. Now given
particular values for the constants of the system, we can immediately predict the long-term
behavior of the two populations. To observe the results of our analysis, consider the system:

dx/dt = x(1.5 − 0.5y)

(10)
dy/dt = y(−0.5 + x)

We can immediately observe that a = 1.5, α = 0.5, c = 0.5, and γ = 1. The critical points,
then, are (0, 0) and (c/γ, a/α) = (0.5, 3). The former critical point will be a saddle. The
latter critical point will be a center, and if we do some minor calculations we can see that
aγ 2 = 1.5 and cα2 = 0.125, so the solutions will tend to be vertically stretched. If we want
to know how vertically stretched the solutions are, we can look at the ratio z/w, which in
this case can be computed as follows:
r
z 0.125 1
= =√ .
w 1.5 12
Therefore, the vertical axes of the
√ elliptical solutions will be longer than the corresponding
horizontal axes by a factor of 12 ≈ 3.46. Thus, the major axes can be expressed by w
and the minor axes can be expressed by z, and we should note here that as we look at
Page 80 RHIT Undergrad. Math. J., Vol. 16, No. 2

solutions that are further from the critical point, the ratio z/w will become a less accurate
representation of a relationship between the two lengths because we are working with a
nonlinear system that we have approximated with our linearization techniques. We can also
observe that cαγ
− aγ
α
= 0.75 − 1 = −0.25 < 0, so the vertically-stretched elliptical solutions
will move counterclockwise about the critical point.
A phase portrait can be used to verify our analysis, and by looking at Figures 2a and 2b,
we can see that the critical points, the direction and magnitude of stretching at the center
point, as well as the direction of the solutions all match our predictions. Here, the figure on
the left is a phase plane of the system that reveals the effects of the equilibria on solutions
in all quadrants of the x-y plane. The figure on the right shows only the first quadrant since
historically the purpose of this model is to describe positive values for the two populations.
In Figure 2b, we also present the elliptical solutions of the linearized system which (not so
surprisingly) become more accurate representations of solutions to the nonlinear system in
regions closer to the equilibrium (c/γ, a/α).

Figure 2

(b) First quadrant including linearized

(a) A phase portrait of system (10) solutions

After some analysis, it is possible to determine the behaviors of solutions to the system
near the equilibria with very few calculations. It is difficult, though, to make claims about
the long-term behaviors of solutions since our analysis relies on approximations that are
only valid in the neighborhoods of the equilibria. While we cannot necessarily extrapolate
the long-term behavior from our analysis, we can still know about the locations and types
of equilibria, as well as the general shapes and directions of solutions near those equilibria.
This predator-prey model is one example of the power of being able to properly analyze
linear systems because the simple techniques we apply to the linear system allow us to know
a great deal about the behaviors of solutions to the nonlinear system.
As a closing remark, it it important to note that in the case of a center in the linear system,
the stability of the nonlinear system cannot be directly determined from the techniques
RHIT Undergrad. Math. J., Vol. 16, No. 2 Page 81

discussed in this paper. We can easily predict the behavior of solutions in the neighborhood
of critical points, but not much can be said about the stability of centers in nonlinear systems
without going more in-depth.

5.2 A More Complicated Nonlinear System of ODEs

The second system of ODEs we consider is
dx/dt = x(α1 − β1 x − γ1 y),
(11)
dy/dt = y(α2 − β2 x − γ2 y),
where α1 , α2 , β1 , β2 , γ1 , γ2 > 0. This model again is founded in biological population model-
ing. When one restricts to nonnegative solutions x, y ≥ 0, this system can be interpreted as a
competing-species model where x and y represent the populations of two species competing
for a common food source. In this setting, the parameters α1 , α2 , β1 , β2 , γ1 , γ2 > 0 would
be determined by environmental factors and characteristics of the particular species being
modeled. More information on this can be found in the text by Boyce and DiPrima [BD]
and the references therein. We will consider solutions to the equation in the whole plane R2 ,
rather than restricting to the first quadrant, which yields some interesting new results in the
behavior of solutions to the system. The results below can again be found in the text by
Boyce and DiPrima [BD] when solutions are restricted to being nonnegative. Although, here
we formulate a full analysis of (11) for arbitrary solutions (not just nonnegative solutions),
which as we will see, poses a more formidable challenge.
We start our analysis by making a change of variables with u = β1 x and v = γ2 y. Thus,
we arrive at the system:
u0 = u(α1 − u − (γ1 /γ2 )v),
(12)
v 0 = v(α2 − (β2 /β1 )u − v),

If we let γ = γ1 /γ2 and β = β2 /β1 , the system simplifies to a point where we can more easily
apply the linearization techniques we have been discussing. We note that the solutions to
equation (11) and (12) are identical up to a rescaling of the variables, in particular using
the substitution u = β1 x and v = γ2 y. So from this point on, we only consider equation
(12), with the understanding that solutions to (11) are completely described by those of the
simplified ones in (12). The simplified system in (12) written in a more concise notation
becomes
 0   
u u(α1 − u − γv)
= F (u, v) = ,
v0 v(α2 − βu − v)

and we can find the equilibria by looking at where u0 = v 0 = 0.

Observe that u0 = 0 where u = 0 and where u = α1 −γv, and that v 0 = 0 where v = 0 and
where v = α2 − βu. These equations are lines in the u-v plane, and wherever a line on which
u0 = 0 crosses a line on which v 0 = 0, the system will have an equilibrium. For simplicity,
we will be referring to u = 0 as L1 , u = α1 − γv as L2 , v = 0 as L3 , and v = α2 − βu as L4 .
Page 82 RHIT Undergrad. Math. J., Vol. 16, No. 2

The system will have a certain number of equilibria that is dependent upon how the
given values for the parameters α1 , α2 , β, and γ cause L2 and L4 to be oriented in the plane.
Because we are looking at two lines in a plane, there are only three possibilities: L2 and L4
could be parallel, they could be one and the same line, or they could intersect at one point.
If L2 and L4 are parallel, then the system only has three equilibria: where L2 intersects
L3 at (α1 , 0), where L1 intersects L4 at (0, α2 ), and where L1 intersects L3 at the origin. For
L2 and L4 to be parallel (but not coincide), it must be the case that βγ = 1 and β 6= α1 /α2 .
Under these conditions, the u-v plane is pictured in Figure 3a below.
By picking initial conditions in each region of the plane we can gain a qualitative un-
derstanding how solutions to the system will behave, and the behaviors of the system are
illustrated by the arrows. With this type of analysis is it possible to see that near the origin
all solutions will be repelled, near (0, α2 ) solutions will be attracted in one direction and
repelled in the other, and near (α1 , 0) all solutions will be attracted. Notice how there is no
equilibrium located where both of the populations have positive numbers of members, so we
know that two competing species could not coexist if βγ = 1 and β 6= α1 /α2 , according to
this model. However, in a broader scope we can appreciate that (0, 0) will be a source, that
(0, α2 ) will be a saddle, and that (α1 , 0) will be a sink.
The second possibility for L2 and L4 in the u-v plane is that they are one and the same
line. For this to happen, βγ = 1 and β = α1 /α2 . When L2 and L4 coincide, there would
exist one equilibrium at the origin and infinitely many equilibria along L2 (or L4 ), which is
to say that the equilibria exist at (0, 0) and at (t, α2 − βt) for all t ∈ R. The phase portrait
for the u-v plane in this case is given by Figure 3b.
The numerical solution in Figure 3b indicates that solutions to the system are drawn
toward L2 (or L4 ), however, this conclusion cannot be reached using the linearization tech-
niques of this paper. It is not hard to see that since the equilibria along L2 are not isolated,
they cannot be asymptotically stable. Hence it is clear that Theorem 2 is not well-suited to
analyze the stability in this situation. In fact, the numerical simulations indicate that these
equilibria are stable (though we cannot make this conclusion rigorously). So linearization
techniques do not appear to be appropriate for analysis of the ODE when L2 = L4 . In our
context, though, we can at least make the observation that (0, 0) will be a source and that
to gain a better understanding the behaviors of solutions near L2 (or L4 ) we will have to
turn to other methods.
The final possibility would be when L2 and L4 intersect at only one point, which occurs
when βγ 6= 1. In this scenario, there are a total of four equilibria at the points P1 (0, 0),
1 −γα2 α2 −βα1
P2 (α1 , 0), P3 (0, α2 ), and P4 ( α1−βγ , 1−βγ ). The last equilibrium is of particular interest
because it is the only equilibrium to have a chance of existing at a location where both
of the populations would have a nonzero number of members. In saying that, the fourth
equilibrium could exist in the first, second, or fourth quadrants of the u-v plane, because
both L2 and L4 have negative slopes and are positive-valued when intersecting the v-axis.
Though, no matter the quadrant in which P4 is located, we can determine the type of critical
point by examining the parameters of the system as we did with the predator-prey model.
Thus, we will linearize our model and apply the techniques of this paper. To linearize the
RHIT Undergrad. Math. J., Vol. 16, No. 2 Page 83

Figure 3

(a) Simulation of the u-v plane when (b) Simulation of the u-v plane for L2 =
L2 k L4 L4

system we take the total derivative of F (u, v):

 
α1 − 2u − γv −γu
DF (u, v) = . (13)
−βv α2 − βu − 2v
Thus, at P1 the linearized system is:
   
0 α1 0 w1
w = DF (0, 0)w = w, where w = .
0 α2 w2
The eigenvalues for DF (0, 0) are λ1 = α1 and λ2 = α2 , which must both be positive.
Therefore, the critical point will be unstable because it will repel all nearby solutions.
We should note here that for our future analysis we will let c1 = α1 − γα2 and c2 =
α2 − βα1 . At P2 the linearized system is:
   
0 −α1 −γα1 −α1 −γα1
w = DF (α1 , 0)w = w= w.
0 α2 − βα1 0 c2
The eigenvalues for DF (α1 , 0) are λ1 = −α1 and λ2 = c2 . In this case, if c2 < 0, then
λ1 , λ2 < 0 and the critical point is asymptotically stable. However, if c2 > 0, then λ1 < 0 <
λ2 , so the critical point will be a saddle point and hence be unstable.
At P3 the system is:
   
0 α1 − γα2 0 c1 0
w = DF (0, α2 )w = w= w.
−βα2 −α2 −βα2 −α2
The eigenvalues for DF (0, α2 ) are λ1 = c1 and λ2 = −α2 . Therefore, if c1 < 0, then the
critical point will be linearly stable because λ1 , λ2 < 0. If c1 > 0, however, then the critical
point will be unstable because λ2 < 0 < λ1 .
Page 84 RHIT Undergrad. Math. J., Vol. 16, No. 2

From our analysis of the first three equilibria and how they are affected by the signs of
c1 and c2 , we can learn about the stability of the equilibria by finding to which quadrant of
the c1 -c2 plane the parameters of the system correspond. The results of this analysis can be
organized into the following table.

Location of (c1 , c2 ) P1 P2 P3
c1 , c2 > 0 Source Saddle Saddle
c1 < 0 < c2 Source Saddle Sink
c1 , c2 < 0 Source Sink Sink
c1 > 0 > c2 Source Sink Saddle

Table 1: Stability of P1 , P2 , and P3 based on location in the c1 -c2 plane.

We can look at P4 in the same fashion as P1 , P2 , and P3 , however, because the system
will have more complicated eigenvalues, the stability analysis will be a little more extensive.
For P4 the linearized system is:
   
0 c1 c2 1 γα2 − α1 γ(γα2 − α1 )
w = DF , w= w
1 − βγ 1 − βγ 1 − βγ β(βα1 − α2 ) βα1 − α2
 
1 c1 γc1
= w,
βγ − 1 βc2 c2
 
c1 c2
where again c1 = α1 − γα2 and c2 = α2 − βα1 . The eigenvalues of DF 1−βγ , 1−βγ can be
found by using the quadratic formula to solve for the roots of the characteristic polynomial
c1 c2
of the coefficient matrix as follows: The characteristic polynomial of DF ( 1−βγ , 1−βγ ) is
  
c1 c2 βγc1 c2 c1 + c2 c1 c2
λ− λ− − 2
= λ2 − λ− ,
βγ − 1 βγ − 1 (βγ − 1) βγ − 1 βγ − 1
and the roots of the characteristic polynomial are
p
(c1 + c2 ) ± (c1 + c2 )2 + 4(βγ − 1)c1 c2
λ1 , λ2 = .
2(βγ − 1)
To determine which values of each constant will cause the critical point to be a source, sink,
saddle, etc., we need to look at the eigenvalues λ1 , λ2 more closely. The first distinction we
make is where the eigenvalues are real and where they are complex. To accomplish this task,
we let

g(c1 , c2 ) = (c1 + c2 )2 + 4(βγ − 1)c1 c2 = c21 + c22 + 2(2βγ − 1)c1 c2 ,

which is the discriminant of the quadratic equation above, and assess where g(c1 , c2 ) ≥ 0 and
where g(c1 , c2 ) < 0. If βγ < 1, it follows that g(c1 , c2 ) ≥ 0 for all c1 , c2 ∈ R. If βγ > 1, then
it is possible that g(c1 , c2 ) is negative. If βγ = 1, then we fall into one of the two situations
RHIT Undergrad. Math. J., Vol. 16, No. 2 Page 85

above, where either L2 and L4 are parallel or equal (see Figures 3a and 3b). So we assume
that βγ 6= 1, and split into two situations, where βγ < 1 and βγ > 1.
For future reference, we also define
p
c1 + c2 ± g(c1 , c2 )
f± (c1 , c2 ) = ,
2(βγ − 1)

which gives the eigenvalues λ1 , λ2 = f± (c1 , c2 ).

5.2.1 The Case βγ < 1

We first note that by our assumptions, it follows that α1 , α2 , β, γ > 0. So in the case that
βγ < 1, it cannot be the case that both c1 < 0 and c2 < 0. This is because when βγ < 1, it
follows that c1 = α1 − γα2 < 0 implies that α1 < γα2 and hence

c2 = α2 − βα1 > α2 − βγα2 = (1 − βγ)α2 > 0.

Similarly c2 < 0 implies c1 > 0. Therefore if βγ < 1, the only possible values for (c1 , c2 ) lie
in the first, second, and fourth quadrants. This can be seen geometrically as the fact that
c1 c2
the order pair ( 1−βγ , 1−βγ ) is the intersection of lines L2 and L4 from the previous section.
Since L2 and L4 cannot intersect in the third quadrant, it is impossible for both c1 and c2
to be negative when βγ < 1.
When βγ < 1, it follows that g(c1 , c2 ) ≥ 0 for all c1 , c2 ∈ R. Therefore λ1 and λ2
are always real. It is also not hard to check that f± (c1 , c2 ) = 0 exactly when c1 = 0 or
c2 = 0. Therefore we consider the sign of f+ (c1 , c2 ) and f− (c1 , c2 ) in each of the three viable
quadrants of the c1 -c2 plane (the first, second, and fourth p quadrants). We already know
that g(c1 , c2 ) is nonnegative in this situation. Note that g(c1 , c2 ) < |c1 + c2 | when c1 and
c2 have the same sign (that is in the first quadrant). Therefore λ1 = f+ (c1 , c2 ) < 0 and
λ2 = f− (c1 , c2 ) < 0 when (c1 , c2 )pis in the first quadrant. Also if (c1 , c2 ) is in either the
second or fourth quadrants, then g(c1 , c2 ) > |c1 + c2 |. Then for (c1 , c2 ) in either the second
or fourth quadrants, we have λ2 = f− (c1 , c2 ) < 0 < f+ (c1 , c2 ) = λ1 . Thus, we have the
stability for P4 in Table 2 for the case βγ < 1. That is, Table 2 characterizes the stability of
equilibrium P4 when βγ < 1 based on the location of (c1 , c2 ) in the c1 -c2 plane.

Location of (c1 , c2 ) Sign of eigenvalues Classification of P4 Stability when βγ < 1

c1 , c2 > 0 λ1 , λ2 < 0 Sink Stable
c1 < 0 < c2 λ2 < 0 < λ1 Saddle Unstable
c1 > 0 > c2 λ2 < 0 < λ1 Saddle Unstable

Table 2: Stability of P4 when βγ < 1. Note that the situation c1 , c2 < 0 is not possible.
Page 86 RHIT Undergrad. Math. J., Vol. 16, No. 2

5.2.2 The Case βγ > 1

Similar to the last case where βγ < 1, we can immediately conclude some additional infor-
mation of about c1 and c2 based only on the information α1 , α2 , β, γ > 0 and βγ > 1. In
this situation, it is impossible for both c1 and c2 be positive. Again, this can be interpreted
c1 c2
geometrically by the following argument. Since ( 1−βγ , 1−βγ ) is the intersection of line L2 and
L4 , it cannot lie in the third quadrant. Since in this case 1 − βγ < 0, it is not possible for
both c1 and c2 to be positive. Hence we immediately rule out this option for consideration
in the stability analysis of equilibrium P4 .
Our process here is to determine which regions in the c1 -c2 plane correspond to each type
of critical point. One can easily check that g(c1 , c2 ) can be factored in the following way
when βγ > 1:
  p    p  
g(c1 , c2 ) = c2 + 2βγ − 1 + 2 βγ(βγ − 1) c1 c2 + 2βγ − 1 − 2 βγ(βγ − 1) c1 .

Therefore, if we solve the equation g(c1 , c2 ) = 0, we find that the c1 -c2 plane becomes divided
by the lines:
p
c2 = mc1 and c2 = m−1 c1 , where m = 2βγ − 1 + 2 βγ(βγ − 1). (14)

Since g(c1 , c2 ) = 0 for any c2 = mc1 , it follows by symmetry that g(c1 , c2 )p= 0 for any
c1 = mc2 . In fact, it it is not hard to show directly that m−1 = 2βγ − 1 − 2 βγ(βγ − 1);
that is, g can be factored g(c1 , c2 ) = (c2 − mc1 )(c2 − m−1 c1 ). This factorization makes it
possible to quickly determine the sign of g(c1 , c2 ) for any values of c1 , c2 . In fact, it splits
the c1 -c2 plane into 4 regions, where c1 − mc1 and√mc2 − c1 are positive or negative.
(c1 +c2 )± g(c1 ,c2 )
By looking more closely at f± (c1 , c2 ) = 2(βγ−1)
, we can identify where the bound-
ary conditions for each type of critical point are. We find that when g(c1 , c2 ) < 0, if c1 +c2 < 0
then we have an inward spiral, if c1 +c2 > 0 then we have an outward spiral, and if c1 +c2 = 0
then we have a center. Using this knowledge we can divide the c1 -c2 plane even further. We
note here that region I is not plausible since c1 and c2 cannot both be positive when βγ > 1.

Figure 4: The c1 -c2 plane divided into eleven (ten possible) regions for βγ > 1.
RHIT Undergrad. Math. J., Vol. 16, No. 2 Page 87

If we solve the equation f± (c1 , c2 ) = 0, we find that f is zero along the c1 -and-c2 axes,
meaning that f± could potentially change sign when crossing those boundaries. These new
boundaries in addition to the boundaries in the imaginary regions will divide the c1 -c2 plane
into eleven regions for us to analyze. By looking at the sign of the real part of f± (which
coincide with looking at the real part of the eigenvalues λ1 , λ2 ) in each of these regions, we
can tell what type of critical point the system will have when its parameters characterize it
as belonging to a certain region of the c1 -c2 plane. In total, we arrive at the decomposition
of the c1 -c2 plane in Figure 4, which we use to determine the sign of Re(f± ) and hence the
stability of P4 when βγ > 1.
At this point we can organize our results for βγ > 1 into two tables. To utilize the
conditions of these tables, recall
p that we set c1 = α1 − γα2 , c2 =pα2 − βα1 , β = β2 /β1 ,
−1
γ = γ1 /γ2 , m = 2βγ − 1 + 2 βγ(βγ − 1), and m = 2βγ − 1 − 2 βγ(βγ − 1). The first
step should be to calculate the quadrant of the c1 -c2 plane to which the parameters of the
system refer. Next, compute the values of m, m−1 , and |c2 |/|c1 |, and compare them using
Table 3 to determine the correct region out of the possible ten. Once the region has been
determined, use Table 4 to determine the behavior of the solutions to the system near the
fourth equilibrium. Hence using Tables 3 and 4 we can determine the stability of equilibrium
P4 based on a few simple computations involving the parameters in (12).

c1 < 0 < c2 c2 < 0 < c1 c1 , c2 < 0

|c2 |
|c1 |
>m II VIII VII
|c2 |
m> |c1 |
>1 III IX VII
|c2 |
|c1 |
=1 IV IV VII
|c2 |
1 > |c1 | > m−1 V X VII
|c2 |
m−1 > |c 1|
VI XI VII

Table 3: Description of ten plausible regions of the c1 -c2 plane when βγ > 1. Note that it is
not possible for c1 , c2 > 0.

5.2.3 Numerical Simulations for Equation 11 when βγ < 1

We can now explore phase portraits for the system in equation (11) based on the parameters
α1 , α2 , β1 , β2 , γ1 , and γ2 . Recall that we’ve defined u = β1 x, v = γ2 y, γ = γ1 /γ2 , β = β2 /β1 ,
c1 = α1 − γα2 , and c2 = α2 − βα1 .
We consider the numerical example of equation (11) given by

dx/dt = x(1.5 − x − 0.5y),

(15)
dy/dt = y(2 − 0.75x − y).

Right away we can see that α1 = 1.5, α2 = 2, β1 = 1, β2 = 0.75, γ1 = 0.5, and γ2 = 1. Thus,
c1 = 0.5, c2 = 0.875, and βγ = 0.375 < 1. We should also note here that, based on these
Page 88 RHIT Undergrad. Math. J., Vol. 16, No. 2

Region Critical Point Nonlinear Stability

II Source Unstable
III Outward Spiral Unstable
IV Center Indeterminate
V Inward Spiral Asymptotically Stable
VI Sink Asymptotically Stable
V II Saddle Unstable
V III Sink Asymptotically Stable
IX Inward Spiral Asymptotically Stable
X Outward Spiral Unstable
XI Source Unstable

Table 4: Stability of P4 when βγ > 1 depending on location in the c1 -c2 plane.

given values of parameters, when we shift to u-v coordinates with the substitutions u = β1 x,
v = γ2 y, γ = γ1 /γ2 , and β = β2 /β1 , the system would coincidentally take the form

du/dt = u(1.5 − u − 0.5v),

dv/dt = v(2 − 0.75u − v).

Thus, we can directly apply out analysis to system (15).

Plugging the former parameters into our equations for the critical points we get the points
P1 (0, 0), P2 (1.5, 0), P3 (0, 2), and P4 (0.5, 1.5). Now looking specifically at each critical point,
and the fact that in this case βγ < 1, we can tell from Table 1 that P1 is a source and P2
and P3 are both saddles. We know that (c1 , c2 ) lies in the first quadrant of the c1 -c2 plane
for this example, that is c1 , c2 > 0, and since βγ < 1 as well, it follows that P4 also lies in
the first quadrant. Furthermore, looking at Table 2, it is now easy to see that equilibrium
P4 is a sink. Once we have gone through this sort of extensive analysis, it becomes very
straightforward to characterize a system based on its parameters, and the techniques used
in this section can be applied to other equally-complicated models as well. We can verify
our analysis with a phase portrait of the nonlinear system. In Figure 5a, we can see that the
equilibria not only exist where we predicted them to be, but also that they are the types of
equilibria that we expected.
Using the above analysis of the fourth equilibrium, P4 , we can look more closely at the
behavior of the system specifically when (c1 , c2 ) lies in the second quadrant of the c1 -c2 plane,
c1 c2
that is, c1 < 0 < c2 . Since c1 < 0 < c2 and βγ < 1, we know that 1−βγ < 0 and 1−βγ > 0.
Therefore P4 lies in the second quadrant of the u-v plane. It can also easily be read from
Tables 1 and 2, using c1 < 0 < c2 and βγ < 1, that P1 is a source, P2 is a saddle, P3 is a sink,
and P4 is a saddle. This situation is represented in Figure 5b. This analysis can be done for
(c1 , c2 ) in any viable quadrant for (c1 , c2 )–that is in the first, second, or fourth quadrants–to
analyze the solutions of (15), and more generally solutions of (11) and (12) when βγ < 1.
RHIT Undergrad. Math. J., Vol. 16, No. 2 Page 89

Figure 5

(a) Phase portrait and numerical solu- (b) Phase portrait and numerical solu-
tions for equation (11) when c1 , c2 > 0 tions for equation (11) when c1 < 0 < c2
and βγ < 1. and βγ < 1.

5.2.4 Numerical Simulations for Equation 11 when βγ > 1

In this section, we assume that we are in the situation where βγ > 1. So we can use Tables
1, 3, and 4 to determine the stability behavior of the equilibrium solutions P1 , P2 , P3 , and
P4 , and hence understand the phase portraits associated to the system (11).
To demonstrate the case βγ > 1, consider the following system in x-y coordinates that
we have shifted into u-v coordinates for simplicity:
x0 = x(0.1 − 2x − y), u0 = u(0.1 − u − 0.5v),
or rather,
y 0 = y(2.1 − 8x − 2y), v 0 = v(2.1 − 4u − v).
Looking at the system in u-v coordinates, we can see that α1 = 0.1, α2 = 2.1, β = 4, and
γ = 0.5. Thus, c1 < 0 < c2 . We can also compute m, m−1 , and |c2 |/|c1 | to find that we
have 1 < |c2 |/|c1 | = 34/19 < 5.83 ≈ m. Thus, from Table 3 we can see that P4 is described
by region III of the c1 -c2 plane, and from Table 4 we can see that P4 will be an outward
spiral. We can also use the relationship c1 < 0 < c2 and Table 1 to observe that P1 will be
a source, P2 will be a saddle, and P3 will be a sink. To verify our analysis, a phase portrait
of the system in u-v coordinates is shown in Figure 6a.
We give another example: consider a collection of parameters for equation (11) such that
βγ > 1 and (c1 , c2 ) lies in region V of the c1 -c2 plane, which is pictured in Figure 4. From
Table 1, using that c1 < 0 < c2 when (c1 , c2 ) is in region V, we can conclude that P1 is a
source, P2 is a saddle, and P3 is a sink. Also, using Table 4, we can conclude that P4 is
asymptotically stable whose linearized solution is an inward spiral. This situation is pictured
in Figure 6b for a particular choice of parameters such that βγ > 1 and (c1 , c2 ) lies in V.
Similar analysis can be done to construct the phase planes for solutions of equation (11)
when (c1 , c2 ) lies in any of the viable regions of the c1 -c2 , that is whenever (c1 , c2 ) lies in one
Page 90 RHIT Undergrad. Math. J., Vol. 16, No. 2

of the regions II through XI (not in the first quadrant, as that is not possible when βγ > 1).
Thus, from our analysis, one can approach this model and, using only the parameters, fully
understand the behaviors of solutions near equilibria in most scenarios (again, excluding
centers).

Figure 6

(a) Numerical simulation of the u-v (b) Numerical simulation of the u-v
plane when (c1 , c2 ) is in region III and plane when (c1 , c2 ) is in region V and
βγ > 1. βγ > 1.

6 Conclusion
This last section on phase-portrait characterizations demonstrates the power of linearization
techniques. The systems of nonlinear ODEs discussed in the last section are difficult to solve,
however, through the use of the techniques in this paper, we can simplify the problems
and understand the general behaviors of the solutions. This process involves finding the
equilibria of the nonlinear system, then linearizing the nonlinear system by taking the total
derivative at the critical points. The next step was to use the linearized system to classify
each equilibrium as stable or unstable based upon the eigenvalues of the coefficient matrix
A. In the end, we arrived at a relatively simple characterization of the systems based on the
values of their parameters.
Without a doubt, it would be preferable to find explicit solutions to nonlinear systems of
ODEs, but finding such solutions can be very difficult (if not impossible) in general. Using
linearization and stability analysis to construct qualitative solutions in phase planes is a
viable alternative to searching for explicit solutions, as demonstrated here. Understanding
the types of critical points that exist for a given system allowed us to draw phase portraits
and visually express the behaviors of the solutions. We then applied these techniques to real-
world models and characterized how the parameters of each system influenced the behaviors
RHIT Undergrad. Math. J., Vol. 16, No. 2 Page 91

of the solutions. In this way, we developed a qualitative characterization of solutions to the

equations we considered based only on the equation parameters and given initial conditions.
As a closing remark, this type of linear analysis of nonlinear problems is an accurate
method of understanding the behaviors of solutions in the nonlinear system–with a few
exceptions that were mentioned earlier. Methods of linearization become very useful when
dealing with complicated systems that could be difficult or even impossible to solve, so it is
beneficial to have a firm understanding of how to linearize and analyze real-world modeling
problems as we did in this paper. Linearization is not appropriate in all situations, and it
seems to run into trouble when the total derivative does not behave nicely, that is, if DF (x)
is not continuous or if it is not invertible at the critical point of interest. This situation
manifested itself in our analysis when our equation had non-isolated equilibria.

References
[BD] W. E. Boyce and R. C. DiPrima, Elementary differential equations and boundary value
problems, John Wiley and Sons, Inc., New York-London-Sydney, 1965.

[H] R. Haberman, Applied partial differential equations; with Fourier series and boundary
value problems, Pearson Education, Inc., Upper Saddle River, New Jersey, 2004.

[PC] J. C. Polking and J. Castellanos, dfield and pplane, open source Matlab software,
http://math.rice.edu/~dfield/dfpp.html.

[S] J. E. Sasser, History of ordinary differential equations; the first hundred years, Pro-
ceedings of The Midwest Mathematics History Conferences, Miami University, Oxford,
Ohio, October 2-3, 1992, Douglas E. Cameron and James D. Wine, Editors, ISBN:
1-884905-05-6, Modern Logic Publishing, 1997.