Dynamical systems 2019 [email protected].

se

4 Phase plane (Strogatz 6)

The previous lecture dealt with linear two-dimensional flows. This
lecture considers non-linear two-dimensional flows living in a phase
space of dimensionality two: the phase plane.

4.1 Geometrical approach: Phase portraits

Consider a general dynamical systems of dimensionality two:
ẋ = f
with x = (x1, x2) and f (x) = (f1(x), f2(x)). To have a cleaner
notation without indices, we often use

ẋ = f (x, y)
x = x1 , y = x2 , f = f1 , g = g2 , ⇒
ẏ = g(x, y)
The trajectory x(t) depends on the initial condition x(0):

In non-linear systems it is usually not possible to find x(t) analytically.

Phase portraits are typically much more complicated compared to the
linear flows considered so far. One example: (Strogatz Fig. 6.1.2)

1
Dynamical systems 2019 [email protected]

• Fixed points (A,B,C)

f (x∗) = 0

• Closed orbits (D) [periodic solution x(t) = x(t + T )].

• Arrangement of trajectories near different fixed points and dif-
ferent closed orbits may differ:
– A, C saddle
– B spiral
• Stability
– A, B, C unstable
– D stable
As for the one-dimensional case: if the flow is smooth (both f and
its partial derivatives ∂ifj are continuous) in some open connected
domain D ∈ Rn, then for x0 ∈ D the initial-value problem has a
unique solution in some time interval around t = 0.
As a consequence different trajectories cannot intersect. If they did,
there would be two solutions starting from the point of intersection,
i.e. breaks the uniqueness condition.

4.1.1 Numerical computation of phase portraits

Using a low-level language such as C++ without suitable external li-
braries, one may use a Runge-Kutta integration scheme (Strogatz 6.1)
to compute trajectories from a set of initial conditions. Using Matlab
or Mathematica, it is more convenient to use the built-in functions,
e.g. StreamPlot[] in Mathematica to plot the flow, or NDSolve[] to
find the trajectories.

2
Dynamical systems 2019 [email protected]

4.1.2 Sketching the phase portrait by hand

To draw a phase portrait by pen and paper, it is often instructive to
first determine the nullclines. These are the curves defined by

ẋ = 0 or ẏ = 0 .

Along the nullclines the flow is either vertical (ẋ = 0) or horizontal

(ẏ = 0). They divide the phase plane into regions where direction of
flow is known or approximately known:
ẋ < 0 ẋ = 0 ẋ > 0
ẏ < 0 . ↓ &
ẏ = 0 ← · →
ẏ > 0 - ↑ %
Intersection points between a nullcline with ẋ = 0 and one with ẏ = 0
give the fixed points of the flow.
Since trajectories are not allowed to cross, the information given by
the nullclines often allows to make a qualitative plot of the dynamics.

Linear example

ẋ = 5x + y
ẏ = −x − y

Nullclines:

ẋ = 0 : y = −5x
ẏ = 0 : y = −x

3
Dynamical systems 2019 [email protected]

From the plotted trajectories we see that the fixed point at the
intersection of the nullclines is a saddle point.
Consistency check:
 
5 1
A= ⇒ ∆ = det A = −4 ⇒ Saddle point
−1 −1

4.2 Analytical approach: Linear stability analysis

A dynamical system of dimensionality two

ẋ = f (x, y)
ẏ = g(x, y)

has fixed points (x∗, y ∗) where f (x∗, y ∗) = g(x∗, y ∗) = 0. Linearize

around the fixed point (c.f. Lecture 2):

η = x − x∗ , µ = y − y∗
∂f ∂f
!
d η
   
η
= J(x∗, y ∗) + ... , with J(x∗, y ∗) = ∂x
∂g
∂y
∂g (1)
dt µ µ ∂x ∂y

where J is the stability matrix (Jacobian matrix in Strogatz) and the

derivatives are evaluated at the fixed point (x∗, y ∗).
In linear stability analysis, we neglect the higher-order terms and
the deviation (η, µ) satisfies a linear system that can be analyzed and
classified as in Lecture 3.

4
 Dynamical systems 2019 [email protected]

4.2.1 Example on phase-plane analysis

Analyze the dynamical system:

ẋ = x(3 − 2x − y)
ẏ = y(2 − x − y) .

The nullclines are

ẋ = 0 : x = 0 or x = (3 − y)/2
ẏ = 0 : y = 0 or y = 2 − x .

On the nullclines the flow is one-dimensional and therefore straight-

forward to analyze:
Nullcline x=0 x = (3 − y)/2 y=0 y =2−x
Flow ẏ = y(2 − y) ẏ = y(1 − y)/2 ẋ = x(3 − 2x) ẋ = x(1 − x)

The system has 4 fixed points at the intersections of the two types of
nullclines:

(x∗1, y1∗) = (0, 0) , (x∗2, y2∗) = (0, 2) , (x∗3, y3∗) = (3/2, 0) , (x∗4, y4∗) = (1, 1) .

The nullclines give a rough picture of the flow, but it is complicated

to figure out what happens close to the fixed points using nullclines

5
 Dynamical systems 2019 [email protected]

only. Therefore, use linear stability analysis for the fixed points:
 
(3 − 2x − y) + x(−2) x(−1)
J=
y(−1) (2 − x − y) + y(−1)
 
3 − 4x − y −x
=
−y 2 − x − 2y
Fixed point (x∗, y ∗) (0, 0) (0, 2) (3/2, 0) (1, 1)
∗ ∗
τ ≡ trJ(x , y ) 5 -1 -5/2 -3
∆ ≡ det√J(x∗, y ∗) 6 -2 -3/2 1√
λ1,2 = (τ ± τ 2 − 4∆)/2 (2, 3) (−2, 1) (−3, 1/2) (−3 ± 5)/2
Type Unstable node Saddle Saddle Stable√node
v1 (0, 1) (0, 1) (1, 0) (1 − √5, 2)
v2 (1, 0) (−3, 2) (−3/7, 1) (1 + 5, 2)

Stable|unstable directions The real part of the eigenvalues deter-

mine the stability of a fixed point. Small deviations from a fixed point
along an eigenvector v u corresponding to a real positive eigenvalue,
λu > 0, remain in the direction of v u and grow exponentially fast.
This follows from Eq. (1) using (η, µ) = (t)v u with   1:
d
v u = J(x∗, y ∗)v u = λuv u ⇒ (t) = (0)eλut .
dt
The (normed) eigenvector v u is an unstable direction of the fixed point.
Similarly, a normed eigenvector v s corresponding to λs < 0 is a stable
direction of the fixed point: small deviations in this direction shrink
exponentially fast. For complex eigenvalues, all directions are either
unstable (Re λ > 0) or stable (Re λ < 0) directions of the fixed point
(for dimensionality 2).

Stable|unstable manifolds The stable manifold Ms of a fixed point

is either a point, curve, or surface in the phase-plane. It is defined as
the set of points (including the fixed point) that approach the fixed
point in the limit t → ∞. Similarly, the unstable manifold Mu consists
of the set of points that approach the fixed point in the limit t → −∞,
i.e. if the flow is reversed, then Ms and Mu switch stability.

6
Dynamical systems 2019 [email protected]

• In a linear system the stable|unstable manifold is given by

the subspace spanned by the set of stable|unstable directions.
For example, a saddle point has one negative and one positive
eigenvalue, it attracts along the stable direction, but repels along
the unstable direction. Its stable and unstable manifolds are
lines in these directions. Attractors|repellers are stable|unstable
in all directions and the stable|unstable manifold is a surface
(the entire phase plane).
• For a non-linear system, Ms and Mu approach the manifolds
of the linearized fixed point close to it, but may deviate further
away due to non-linear effects. Example for a saddle point:

The stable|unstable manifold approaches the stable|unstable di-

rection vs|vu of the fixed point close to the fixed point.
The two-dimensional stable|unstable manifold of an attractor|repeller
may become bounded.
To numerically evaluate the stable|unstable manifold: start close to
the fixed point in the stable|unstable direction and integrate the sys-
tem backwards|forward in time.
Coming back to our example, from the table, the stable|unstable
directions close to the fixed points are:

7
Dynamical systems 2019 [email protected]

As shown by the nullclines, the flow aligns with the coordinate axes.
Therefore, since trajectories cannot cross, the four quadrants are iso-
lated from each other.
Consider first the upper-right quadrant. Since the flow is negative
for large values of x or y, trajectories do not escape to infinity, and
must therefore be attracted by the stable node at (1, 1) (its stable
manifold or basin of attraction is the upper right quadrant). In par-
ticular, the unstable manifolds of the saddle points must connect with
the stable node. As a consequence, trajectories become trapped on
either side of these manifolds:

This is a generic behaviour, the manifolds of saddle points often di-

vide the phase space into regions of qualitatively different long-term
dynamics (C.f. Strogatz 6.4).
One example is the stable manifolds of the saddle points along the
coordinate axes: these separate dynamics that are attracted to the
stable node, from the rest of the phase plane where trajectories run

8
Dynamical systems 2019 [email protected]

off to infinity. The stable manifolds of the saddle points are examples
of separatrices (singular: separatrix): they divide the phase space into
regions of different long-term behaviour.
Outside the upper-right quadrant, the unstable manifolds of the sad-
dle points must run away to infinity (no attractor can attract them).
Note: The stable|unstable manifolds and the nullclines can some-
times coincide (the coordinate axes in the example above), but in
general they are different curves, also close to the fixed point.

4.2.2 Effect of small non-linear terms

When is it safe to neglect quadratic terms in the stability analysis?
Linear stability analysis gives a qualitatively correct picture if the
fixed point is a node, spiral, or saddle (as in the Example in Sec-
tion 4.2.1).
For the border-line cases (degenerate node, star, center, or not iso-
lated) non-linear terms may (or may not) change the dynamics qual-
itatively from the border-line case to one of the neighbouring cases
in the ∆-τ diagram. To find what type the fixed point is, one must
analyze the non-linear dynamics close to the fixed point, either ana-
lytically, or using a geometric approach.

Example Consider the system

ẋ = −y + ax(x2 + y 2) ≡ f (x, y)
ẏ = x + ay(x2 + y 2) ≡ g(x, y)

We have one fixed point (x∗, y ∗) = (0, 0) and stability matrix

3ax2
 
−1 + 2axy
J(x, y) = .
1 + 2axy 3ay 2

At the fixed point

 
0 −1
J(0, 0) = .
1 0

9
Dynamical systems 2019 [email protected]

with eighenvalues λ1,2 = ±i. The fixed point is a center according to

linear stability analysis.
Analyse the non-linear behaviour of the system. In polar coordinates
p
r= x2 + y 2 , θ = atan(y/x)
1
⇒ ṙ = √ 2 (2xẋ + 2y ẏ)
2 x + y2
1
= (x[−y + ax(x2 + y 2)] + y[x + ay(x2 + y 2)])
r
1
= (ax2r2 + ay 2r2) = ar3 .
r
d 1 1 dy

θ̇ = atan(x) = =
dx 1 + x2 1 + (y/x)2 dt x
1 ẏ y 1
 
= − ẋ = (xẏ − y ẋ)
1 + (y/x)2 x x2 x2 + y 2
1 1
= 2 x[x + ay(x2 + y 2)] − y[−y + ax(x2 + y 2)] = 2 x2 + y 2 = 1



r r
The dynamics ṙ = ar3 and θ̇ = 1 is simpler to analyze. If a = 0
the system is linear and we have a center. Otherwise, we have a spiral
(unstable if a > 0 or stable if a < 0).
In conclusion, the non-linear terms change the stability in this case.
Similarly, stars and degenerate nodes can be changed into spirals or
nodes by small non-linearities, but their stability does not change.
More generally:
• If both Re λ1 and Re λ2 are non-zero (attractors, repellers, sad-
dle points), the qualitative dynamics is robust to small per-
turbations. The fixed point is hyperbolic and the dynamics is
structurally stable.
• If Re λ1 and/or Re λ2 are zero (centers or non-isolated fixed
points), the dynamics is marginal and not structurally stable:
a small perturbation may change closed orbits and the number
of fixed points.

10