Notes on Nonlinear Dynamics, CMI, Spring 2020

Govind S. Krishnaswami, 22 April, 2020

Please let me know at [email protected] of any comments or corrections

Contents

1 Introductory remarks on dynamical systems 2

2 Flows in 1D: Vector field on the real line 3

2.1 Fixed points, phase portrait, linear stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 More examples: RC circuit, Logistic equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 1d Flows as gradient flows and general features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2.4 Existence and uniqueness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

3 Bifurcations of vector fields on the real line 8

3.1 Saddle-node bifurcation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.2 Transcritical bifurcation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.3 Pitchfork bifurcations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.3.1 Supercritical pitchfork bifurcation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.3.2 Subcritical pitchfork bifurcation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

4 Vector fields on a circle 13

5 Linear vector fields on the plane and their fixed points 14

5.1 Examples of fixed points: center, node, saddle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

5.2 Analysis of phase portrait using eigenvalues and eigenvectors of coefficient matrix . . . . . . . . . . . . . . . . . . . . . . . . 17
5.2.1 Damped harmonic oscillator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
5.2.2 Degenerate eigenvalues and deficient coefficient matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
5.2.3 Trace-determinant classification of fixed points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

6 Nonlinear vector fields, especially in two dimensions 23

6.1 Linearization around fixed points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

6.1.1 Robustness of the linear (stability) theory and hyperbolic fixed points . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
6.1.2 Competitive Lotka-Volterra model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
6.2 Conservative systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
6.2.1 Lotka-Volterra prey-predator model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
6.3 Hamiltonian systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

6.3.1 Some general properties of Hamiltonian systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

6.3.2 Poincaré recurrence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
6.4 Index theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
6.4.1 Properties of the index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
6.5 Limit cycles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
6.5.1 Criteria for non-existence of limit cycles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

6.5.2 Poincaré-Bendixon Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

6.6 Bifurcations in two dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
6.6.1 Saddle-node, transcritical and pitchfork bifurcations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
6.6.2 Hopf bifurcations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

7 Lorenz’s 3d model of convection and chaos 48

1
1 Introductory remarks on dynamical systems

We will be concerned with time evolution (or dynamics) of systems with finitely many degrees
of freedom. A typical example is a mechanical systems of finitely many (N ) particles, that is
described by ordinary differential equations (ODEs), say for the positions r1 (t), · · · , rN (t) of the
particles, with only one independent variable (time). This is in contrast to fields (such as velocity
v(x, t) and density ρ(x, t)) that occur, say, in fluid mechanics or gravity or electromagnetism and
are modeled by partial differential equations (PDEs) with more than one independent variable
(x and t). There are two main types of dynamical systems we will consider: continuous time
and discrete time. The former are modeled using (systems of first order) ordinary differential
equations (ODEs), e.g., the three equations

dx
= v(x, t) for t ≥ 0, with x, v ∈ R3 and initial condition x(0) = x0 (1)
dt
This ‘initial value problem’ (IVP) is, in general, a nonlinear system of equations, as the vector
field v need not be a linear function of x. Time is the only independent variable. The space
in which the dependent variables x take values (R3 here) is called the state space or phase
space. The number of dependent variables is the dimension of the phase space. The curve x(t)
parametrized by time is called the phase trajectory or the integral curve of the vector field. The
system is called autonomous if v does not depend explicitly on time. We can think of v as the
prescribed1 velocity field in a fluid flow and of x as the position of a test particle that is carried
around by the flow. The system is autonomous if the fluid flow is steady. We note that a higher
order ODE can be reduced to a system of 1st order ODEs by introducing additional dependent
variables. E.g., Newton’s equation for the pendulum’s angle of deflection from the vertical,
mθ̈ = −(g/l) sin θ becomes the pair of autonomous nonlinear first order ODEs by introducing
p = ml2 θ̇ as a new variable
p/ml2
   
θ̇
= (2)
ṗ −gl sin θ
They become linear in the simple harmonic approximation sin θ ≈ θ . A non-autonomous system
of n ODES ẋi = v i (x, t) can be turned into an autonomous one by introducing an extra
dependent variable xn+1 = t, so that
 i   i
(v (x, xn+1 ))1≤i≤n

(ẋ )1≤i≤n
= . (3)
ẋn+1 1

In other words, a time-dependent vector field on Rn may be viewed as a time-independent vector

field on Rn+1 .
Discrete time dynamical systems are modeled by (systems of) difference equations. The
primary examples are iterated maps of a real or complex variable, such as

zn+1 = f (zn ) for n = 0, 1, 2, · · · , where f (z) = z 2 + i with z0 = 1. (4)

n = 0, 1, 2, . . . plays the role of a discrete time variable and the sequence zn is the analogue
of the phase trajectory. Maps sometimes arise by observing a continuous time system at cer-
tain discrete instants of time, just as the ODEs for some dynamical systems (e.g. the Lorenz
1
v(r, t) is given. It is not dynamical in the sense that we are not given an evolution equation for v .

2
model) arise as simplifications/reductions of PDEs describing evolving fields (e.g. atmospheric
convection). Indeed, Poincaré proposed to obtain information about solutions to a system of
ODEs by examining the points where a trajectory intersects a fixed Poincaré surface in phase
space. A trajectory beginning at t = t0 at a point x0 on the surface [and that is transversal
(goes off it) to the surface] may first return to a point x1 on the surface at a later time t1 . The
map x0 7→ x1 is called the Poincaré first return map and it defines a discrete time dynamical
system. Iterating the return map, we record all locations at which the trajectory intersects the
surface to obtain its Poincaré section. For instance, the Poincaré section of a periodic trajectory
consists of a finite number of points (among which the discrete dynamics periodically circulates).
Discrete time dynamical systems also arise via finite difference approximations that are used in
numerical schemes to solve ODEs. For instance, the ODE ẋ = v(x) may be discretized as

x(t + δt) − x(t) ≈ δt v(x(t)) for small δt (5)

Denoting x(t) = xn and x(t + δt) = xn+1 we define the map

xn+1 = xn + δt v(xn ) ≡ f (xn ) (6)

whose successive iterates (for sufficiently small δt) can provide an approximation to the phase
trajectory. The first-order nature of the underlying ODE ensures that (at least in this ‘forward
Euler’ scheme) xn+1 depends only on xn (and not, say, xn−1 ).
• See S H Strogatz, Nonlinear Dynamics and Chaos for a good introduction to this subject.

2 Flows in 1D: Vector field on the real line

We begin by considering the autonomous first order ODE ẋ = v(x) where the dynamical variable
x is real and subject to the initial condition x(0) = x0 . Here v may be viewed as the velocity
field of a fluid flowing steadily on the line and x(t) the trajectory of a test particle that is
released into the flow at x0 . Such an equation can be reduced to quadrature (evaluation of
integrals)
Z t Z x
′ dx′
dt = t = ′
. (7)
0 x0 v(x )
In favorable cases, the integral can be evaluated in closed form to give t as a function of x. We
would still have to invert the result to express x as a function of time. This is not always easy
to do, as the example v(x) = − sin x reveals
| csc x0 + cot x0 |
−t = log . (8)
| csc x + cot x|
The example ẋ = − sin x, which we will further examine below, arises in an overdamped limit of
a pendulum. A pendulum subject to damping may be modeled by the second order Newtonian
ODE ml2 θ̈ = −γ θ̇ − mgl sin θ . γ > 0 is the damping coefficient. In the overdamped limit, we
ignore the inertia/acceleration term altogether to arrive at θ̇ = −(mgl/γ) sin θ . In this limit,
Newton’s equation changes character, it goes from being second-order to first order2 .
Aside from their intrinsic interest and as a simple starting point for the study of dynamical
systems, flows in 1d can also capture some essential features of flows in higher dimensions. For
2
The inertia term is said to be a singular perturbation of the overdamped equation.

3
instance, due to dissipation, higher dimensional flows can sometimes undergo a ‘dimensional
reduction’ and behave in effect like 1d flows.

2.1 Fixed points, phase portrait, linear stability

Fixed points and phase portrait: Even without an explicit expression for x(t), much qual-
itative insight may be gained by studying the geometric nature of v(x), as emphasized by
Poincaré. If x∗ is a zero of v(x), then it is a ‘stagnation’ point, where the fluid is at rest.
So we have a static/equilibrium/constant solution x(t) ≡ x∗ for all t at each zero (or ‘fixed
point’) of v . On the other hand, if v(x) 6= 0, then the particle moves to the right/left at x
depending on whether v(x) is positive or negative. Moreover, if |v(x)| is increasing, then the
particle speeds up while it decelerates if |v(x)| is decreasing. These observations are enough to
get a qualitative understanding of trajectories. They are usefully displayed on a phase portrait,
which is a diagram of the phase space showing the qualitatively different trajectories.
• For instance, for v(x) = − sin x, the fixed points are at x = 0, ±π, ±2π, . . .. The flow is to
the left for 2nπ < x < (2n + 1)π and to the right on ((2n − 1)π, 2nπ) for any integer n. This
may be displayed on a phase portrait, which in this case, is the real line with fixed points at
integer multiples of π and the direction of trajectories between successive fixed points indicated
by arrows. Draw the phase portrait and also roughly plot some trajectories on the t − x plane.
Non-constant trajectories monotonically approach a fixed point as t → ∞ and emerge from a
neighboring fixed point as t → −∞. We will see why it takes infinitely long. In particular, the
trajectories do not display any oscillations.
• The fixed points may be classified by their stability. In the above example, odd multiples of
π are unstable as the flow is directed away from them in both directions. On the other hand,
even multiples of π are stable fixed points.

Figure 1: Phase portrait and trajectories for ẋ = − sin x .

• A third possibility is provided by the fixed point at the origin for the vector field v(x) = x2 .
In this case, the flow is directed rightward at all x 6= 0. So x = 0 is ‘half-stable’: stable on the
left but unstable on the right. Notice that trajectories either go off to ∞ or approach the fixed
point x = 0 in a monotonic fashion.
Linear stability: The stability/instability of a fixed point may be quantified by a de-
cay/growth rate that controls how fast the phase point approaches/leaves the fixed point from
its immediate vicinity. It is obtained by linearizing the vector field in the neighborhood of the
fixed point x∗ :

ẋ = v(x) = v(x∗ ) + v ′ (x∗ )(x − x∗ ) + . . . ≈ v ′ (x∗ )(x − x∗ ), (9)

4
where we used v(x∗ ) = 0 and dropped the higher order terms in the Taylor series assuming
v ′ (x∗ ) 6= 0. The vector field has now become a linear function of x. The solution with IC
x(0) = x0 ≈ x∗ is
′
x(t) = x∗ + (x0 − x∗ )ev (x∗ )t for small |x − x∗ |. (10)
Thus we see that the trajectory moves away/towards x∗ according as v ′ (x∗ ) is positive or
negative. We say the fixed point x∗ is linearly unstable/stable in these two cases. When
v ′ (x∗ ) < 0 the above solution can be trusted for all t > 0 and in fact becomes increasingly
accurate with x → x∗ as t → ∞. This explains why it took infinitely long for the non-
constant solutions to ẋ = − sin x to approach the stable fixed points. On the other hand, when
v ′ (x∗ ) > 0, the exponentially growing solution cannot be trusted for very long, as |x − x∗ | would
become large. Ignored nonlinear effects would become important. In the above example, this
exponential growth away from the unstable fixed points is eventually superseded by nonlinear
saturation effects as x(t) approaches a stable fixed point value as x → ∞.
• When v ′ (x∗ ) vanishes, this linear stability analysis is inconclusive and the fixed point is said
to be of marginal stability. One could include higher order terms in the Taylor expansion to try
to understand its stability in a nonlinear sense.
• When v ′ (x∗ ) 6= 0, the characteristic time scale 1/v ′ (x∗ ) is the exponential growth or decay
rate. It controls how quickly a trajectory leaves/approaches x∗ .

2.2 More examples: RC circuit, Logistic equation

• Example. RC circuit. The dynamical variable is the charge Q(t) that builds up on a
capacitor of capacitance C , when connected in series at t = 0 to a resistance R and a DC
battery with constant voltage V . The battery voltage V is then the sum of the voltage drops
IR and Q/C across the resistor and capacitor where I = Q̇ is the current. The resulting
evolution equation is
V Q
Q̇ = − = v(Q). (11)
R RC
There is a unique fixed point Q∗ = CV with the flow being towards it both from Q > Q∗
and Q < Q∗ . Thus Q∗ is a stable fixed point, in fact it is globally stable as the charge on
the capacitor always tends asymptotically to Q∗ irrespective of the initial charge. Moreover,
v ′ (Q∗ ) = −1/RC < 0 so that the fixed point is linearly stable as we would expect. Notice that
the charge Q monotonically approaches Q∗ , it does not oscillate.
• Example. Logistic equation: The logistic equation is a simple model for the dynamics
of the population N (t) of a species. It postulates that the per-capita growth rate Ṅ /N is a
constant r > 0 for small N so that small populations grow exponentially. However, as the
population grows, pressure on resources leads to a decrease in the per-capita grow rate which
eventually drops to zero when the population reaches the carrying capacity K . This is modeled
via the ODE
Ṅ = rN (1 − N/K) = v(N ). (12)
It is clear that the vector field has two zeros: at N = 0 and N = K . The flow is to the right
for 0 < N < K where v > 0 and to the left for N > K where v < 0. Draw a phase portrait.
Moreover v ′ (0) = r > 0 while v ′ (K) = −r . Thus, while a zero population is an unstable
fixed point, the carrying capacity is a stable fixed point: any non-zero population always tends
monotonically to the carrying capacity as t → ∞.

5
2.3 1d Flows as gradient flows and general features

1d flows as gradient flows: The lack of oscillations and monotonic approach to equilibrium
in 1d flows suggests that they describe dissipative processes where some ‘potential energy’ is
monotonically decreasing along the flow. Can we find such a monotonically decreasing function?
Yes! In fact, suppose −φ is an anti-derivative for v , i.e., v(x) = −φ′ (x). φ may be viewed as a
potential energy (or velocity potential in fluid mechanics). Then the flow equation becomes

ẋ = −φ′ (x). (13)

In other words, we have expressed the vector field as the gradient of a potential: this is always
possible for vector fields on the real line. Such a flow is called a gradient flow. Now consider
dφ dx
= φ′ (x) = −φ′ (x)2 ≤ 0. (14)
dt dt
Thus we see that the potential φ is monotonically decreasing (non-increasing) along the flow.
It is an example of a ‘Lyapunov’ function.
• Main features of flows on a line: Inspired by the examples we may now summarize some
key features of 1d flows on the line. Fixed points occur at points where the graph of v(x) crosses
the x axis. Between successive fixed points (or between a fixed point and ±∞) v has a definite
sign so that x is either increasing or decreasing. In particular, x cannot oscillate and there
are no non-constant periodic solutions in 1d. Thus, the phase point, if it does not start at a
fixed point, will approach a fixed point or ±∞ monotonically. Furthermore, the graph of x as
a function of t is convex from below if v > 0 and v ′ > 0 or if v < 0 and v ′ < 0. It is concave
from below if v > 0 and v ′ < 0 or if v < 0 and v ′ > 0.
Moreover, such a flow is a gradient flow with the anti-derivative of v monotonically increasing
in time. We may view ẋ = −φ′ (x) as an overdamped limit (inertia m → 0) of Newton’s equation
mẍ = −γ ẋ − φ′ (x) for a particle moving in the potential φ(x) subject to a damping force with
coefficient γ = 1. By ignoring the inertia term, we have reduced the order of the equation by
one and exaggerated the effect of damping, thereby eliminating the possibility of oscillations.

2.4 Existence and uniqueness

• Given the 1d vector field v(x), when can we guarantee that the solution to the IVP ẋ = v(x),
x(0) = x0 exists (for some time t) and is unique? Let us first give some examples that illustrate
how existence and/or uniqueness can fail.
• Failure of existence: Let us consider the vector field
(
−c if x > 0
v(x) = (15)
c if x ≤ 0 where c > 0 is a constant speed,

and the corresponding IVP ẋ = v(x) with x(0) = 0. We notice that x = 0 is not a fixed point,
so x can’t stay there. Since v(0) = c, the phase point wants to move to the right. However, for
arbitrarily small positive x, v(x) = −c, so that the phase point must move to the left. These
two tendencies cannot both be satisfied: there is no solution for this IC. In other words, the
trajectory does not know what to do at x = 0. On the other hand, if we start from a non-zero

6
IC, say x(0) = x0 > 0, then there is a solution x(t) = x0 − ct. However, this solution ceases
to exist at t = x0 /c when the phase point approaches the origin. In this case, we may trace
the lack of existence to the discontinuity of v at x = 0. v points in different directions as one
approaches x = 0 from above or below. This would be okay if v → 0 as x → 0 so that x = 0
is a fixed point, but that is not the case here.
• Interestingly, discontinuous vector fields do not always fail to admit solutions to an IVP. The
problem in the above example is that the flow was ‘attracted’ to x = 0, but once it got there, it
did not know what to do, due to the discontinuity in v . By contrast, the discontinuous vector
field 
c
 if x > 0
v(x) = c sgn x = 0 if x = 0 (16)

−c if x < 0 where c > 0 is a constant speed,


has solutions (x(t) = x0 + ( sgn x0 ) ct) for any IC that exist for all time! x = 0 here is an
unstable fixed point, a repeller so that trajectories are not troubled by the discontinuity at
x = 0. Thus, while continuity may help, it is not necessary to guarantee existence of solutions.
√
• Failure of uniqueness: Consider ẋ = x for x ≥ 0 subject to the IC x(0) = 0. x = 0
is clearly a fixed point. However, the constant solution x(t) ≡ 0 is not the only possibility.
√
Indeed, integrating, and imposing the IC we find 2 x = t, leading to the non-constant solution
x(t) = t2 /4. What went wrong? Notice that the vector field v is not differentiable at x = 0,
√
its derivative v ′ (x) = 1/2 x diverges there. As a consequence, a linear stability analysis would
say that the fixed point x = 0 is infinitely unstable. In fact, there are infinitely many solutions
to the above IVP. Indeed, we verify that the following is a solution for any t0 ≥ 0:
(
0 for 0 ≤ t ≤ t0
x(t) = 2
(17)
(t − t0 ) /4 for t ≥ t0 ,

The first two solutions we found are the limiting cases t0 → ∞ and t0 = 0. For t0 > 0 we
notice that the solution has a discontinuous second derivative at t = t0 .
It is possible to avoid this non-uniqueness if v is sufficiently regular. There are various
sufficient conditions that guarantee existence and uniqueness of solutions. One of these is given
in the theorem below (the proof can be found in most texts on ODEs.)
• Existence-Uniqueness Theorem: Suppose v(x) and v ′ (x) are continuous on an open
interval containing x0 . We say that v is continuously differentiable or C 1 . Then the ODE
ẋ = v(x) with x(0) = x0 has a unique solution for some time interval |t| < τ .
• Picard Iteration is a method to construct a solution to the IVP ẋ = v(x), x(0) = x0 and
to prove the existence-uniqueness theorem when v is continuously differentiable. It proceeds by
defining a sequence of functions xk (t) for k = 0, 1, 2, . . . inductively via
Z t
x0 (t) = x0 and xk+1 (t) = x0 + v(xk (t′ )) dt′ (18)
0

which are then shown to converge to the solution. The idea behind this scheme is to first convert
the ODE into an integral equation by integrating from 0 to t:
Z t Z t
′ ′
x(t) − x(0) = v(x(t )) dt ⇒ x(t) = x0 + v(x(t′ )) dt′ . (19)
0 0

7
This integral equation is a compact form of the IVP: it encodes both the differential equation
and the IC. However, x(t) appears both on the left and the right. The idea is to begin with
the zeroth approximant x0 (t) = x0 and use it toR generate the next approximant by replacing
x(t) by x0 (t) on the RHS. Thus x1 (t) = x0 + v(x0 (t′ ))dt′ . Repeating this we obtain the
Picard iterates (18). Picard iteration is used to solve the time-dependent Schrödinger equation
in quantum mechanics, the resulting series is called the Born series. Truncating it leads to the
Born approximation. A similar procedure is employed in quantum field theory and goes by the
name Dyson series.
• Can a trajectory x(t) reach or emerge from a fixed point x∗ in a finite time?
In the examples v(x) = − sin x, the RC circuit and the logistic equation we saw that it takes
infinitely long for a solution to either reach or emerge from a fixed point. If a solution reaches or
emerges from a fixed point x∗ in finite time, then the IVP through x∗ would have at least two
√
solutions and would not be unique. For instance, in the above example v = x, in addition to
the constant solution x(t) ≡ 0 we have the non-constant solution x(t) = t2 /4 that emerges from
√
the fixed point x∗ = 0 in finite time. Alternatively, the time-reversed flow ẋ = − x with IC
√
x(0) = x0 ≥ 0 has a fixed point at x∗ = 0, which is reached by the solution x(t) = ( x0 − t/2)2
√
in finite time t∗ = 2 x0 .
• Finite-time blow-ups: Though the theorem guarantees the existence of a solution for
some time, the solution need not exist for all time, it could blow up at a singularity. For
example, consider the smooth vector field v(x) = 1 + x2 . It has no zeros and the above theorem
guarantees existence and uniqueness of solutions for any initial state x0 . However, the solution
x(t) = tan(t + arctan x0 ) diverges as t + arctan x0 approaches the nearest odd multiple of π/2
from below. This is an example of a finite time blow-up. Roughly, the reason this happened is
that the vector field v(x) = 1 + x2 grows without bound rather quickly. Notice that this did
not happen for a bounded vector field like v(x) = sin x or one that grows linearly v(x) = x. In
the latter case, the solution x(t) = x0 et grows but diverges only as t → ∞.
• Consider the vector field on the positive real line v(x) = x1+ǫ (with x > 0) for some ǫ > 0.
Find the solution with IC x(0) = x0 > 0. Show that the solution blows up in finite time for any
ǫ > 0. Find the blow-up time.

3 Bifurcations of vector fields on the real line

• We are often interested not in a single vector field but in a family of vector fields depending
continuously on a control parameter, say r . There could be more than one control parameter in
a problem, they may be masses of particles or coupling constants or material constants, growth
rates etc. Even in 1d, the nature of fixed points and the structure of the phase portrait can
undergo qualitative changes (known as ‘bifurcations’) as the control parameter is varied across
a critical value rc , the bifurcation point. Bifurcations are like phase transitions.
• For instance, at a bifurcation, fixed points may be created or destroyed or their stability may
change.
• We will examine three types of bifurcations in 1d: saddle-node, transcritical and pitchfork.
• The buckling/bending of a horizontal rectangular card compressed between two opposite edges
using two fingers provides an example. The compressional stress is the control parameter while a
measure of departure of the card from flatness (e.g. height of center relative to compressed edges)

8
is the dynamical variable. For a range of small compressional forces, the card has only one stable
equilibrium: it is flat. At a critical compression, it bends either upwards or downwards. Thus,
for compression above the threshold value, we have two stable equilibria (cards bent upwards
or downwards) and one unstable equilibrium (flat card). This is an example of a pitchfork
bifurcation. It is associated with spontaneous breaking of the up-down reflection symmetry.
• The transition from laminar to turbulent flow of a fluid or plasma can involve bifurcations.

3.1 Saddle-node bifurcation

• In a saddle-node bifurcation, two fixed points merge and annihilate leaving no fixed point,
as a control parameter is increased across rc . Decreasing r across rc would describe the ‘pair
production’ of two fixed points from a situation where there are no fixed points. The name
saddle-node bifurcation comes from the generalization of this bifurcation to flows on the plane.
• The prototypical family of vector fields displaying such a bifurcation is v(x; r) 2
√ = r + x . The
bifurcation point
√ is rc = 0. Notice that for r < 0, v has two fixed points: x = −r is unstable
while x = − −r is stable. They coalesce at the half-stable fixed point x = 0 as r → 0− . For
r > 0 there are no fixed points. Draw a bifurcation diagram showing the locations of the fixed
points on the r -x plane (with r along the horizontal axis). It should indicate why a saddle-node
bifurcation is also called a fold or turning point bifurcation.
• Similarly, v(x; r) = r − x2 displays a saddle-node bifurcation where a pair of fixed points
emerge as r increases beyond rc = 0.
• The above quadratic vector fields are not just simple examples, they are canonical/normal
forms of vector fields displaying a saddle-node bifurcation (in the neighborhood of the bifurcation
point and in the vicinity of the merging fixed points). Indeed, without loss of generality we may
take the bifurcation point to be rc = 0 and the point of coalescence of fixed points to be x = 0.
Then v(x; r) must be parabolic in x for x and r near zero in order for there to be two merging
fixed points. This may also be seen by Taylor expanding a vector field v(x; r) that displays a
saddle-node bifurcation at (x∗ , rc ). Indeed,
1
v(x; r) = v(x∗ ; rc ) + (x − x∗ )∂x v(x∗ ; rc ) + (r − rc )∂r v(x∗ ; rc ) + (x − x∗ )2 ∂x2 v(x∗ ; rc ) + · · · . (20)
2
Now v(x∗ , rc ) = 0 as x∗ is a fixed point when r = rc . Moreover, ∂x v(x∗ ; rc ) = 0 as the graph
of v must be tangent to the x-axis when r = rc so that v can have a double zero corresponding
to the coalescence of fixed points. Neglecting higher order terms, we get the canonical form
v(x; r) = a(r − rc ) + b(x − x∗ )2 .

3.2 Transcritical bifurcation

This describes a situation where a fixed point x∗ exists for all parameter values, but whose
stability changes as r is varied. For instance, x∗ may go from being stable to unstable as r
crosses rc . This is achieved by making another fixed point ‘pass through’ x∗ and ‘exchange
stabilities’ with it. The following normal form demonstrates the phenomenon. Consider

v(x) = rx − x2 = x(r − x) = −(x − r/2)2 + r2 /4. (21)

9
x∗ = 0 is clearly always a fixed point. A second fixed point located at x∗∗ = r moves to the
right as r increases past rc = 0. Graph v(x) for various values of r to see the following. For
r < 0, the fixed point at the origin is stable while the one at r is unstable. At r = 0, v = −x2
and the fixed points merge at the origin which is half-stable. Finally, for r > 0 the origin
becomes unstable while x∗∗ = r is stable. Thus, the fixed points have exchanged stabilities
via the bifurcation at r = 0. Indicate the locations of the fixed points as a function of r on a
bifurcation diagram.
• The logistic equation, now considered for real N , K and r ,

Ṅ = rN (1 − N/K) = (r/K)N (K − N ), (22)

displays a transcritical bifurcation of the N∗ = 0 (zero population) fixed point as the carrying
capacity K slides past Kc = 0 provided r̃ = r/K > 0 is held fixed. For K < 0, N∗ = 0 is
stable as v ′ (0) = K r̃ < 0 while N∗ = K is unstable. The two fixed points exchange stabilities at
Kc = 0 with the zero population fixed point becoming unstable and N∗ = K becoming stable
for K > 0.
• There are variants of this transcritical bifurcation where the stability of a fixed point changes
as a parameter is varied.

1. For instance, for the linear vector field v(x) = rx, x∗ = 0 changes from being stable for
r < 0 to unstable for r > 0. At r = 0 we have a line of fixed points: the vector field
vanishes. Draw the bifurcation diagram.

3.3 Pitchfork bifurcations

Pitchfork bifurcations are named after an agricultural tool. The latter looks like a trishul or
trident, it has a long handle and three tines or prongs. Pitchfork bifurcations are typically seen
in systems with a discrete parity or left-right (Z/2Z = Z2 ) symmetry. In the bending of a card
mentioned earlier, one stable flat (parity invariant) equilibrium configuration becomes unstable
at the bifurcation point (critical compressional stress) where two stable equilibria with opposite
parities are born. Though the word bifurcation is standard, pitchfork trifurcation describes the
situation more accurately.
• There are two types of pitchfork bifurcations: supercritical and subcritical. The former is
associated with the bending of a card or buckling of a column under a weight as well as to
second order or continuous phase transitions while the latter is associated with first order or
discontinuous phase transitions in thermal physics.

3.3.1 Supercritical pitchfork bifurcation

The normal form in this case is

ẋ = rx − x3 = x(r − x2 ) (23)

This ODE has an x → −x parity symmetry. This means that every trajectory is either parity
neutral or comes paired with another trajectory related to it by a parity transformation. This
applies to fixed points as well (which are static trajectories). For r < 0, the origin x = 0 is the

10
only fixed point, and it is stable. v ′ (0) = r so that trajectories approach the origin exponentially
fast with time constant 1/r . When r = 0, the vector field has a triple zero at x = 0, which is
still a stable fixed point, though the linear decay time constant diverges. So trajectories that
start at x0 at t = 0, i.e., x(t) = sgn (x0 ) (2t + 1/x20 )−1/2 approach the origin like a power-law
rather than exponentially fast. This is similar to critical slowing-down in the vicinity of a 2nd
order phase transition in classical statistical mechanics. Finally, when r > 0, we have three
√
fixed points, an unstable one at x = 0 and two stable ‘non-trivial’ fixed points at x = ± r . A
bifurcation diagram on the r -x plane showing the fixed points looks like a horizontal pitchfork
with prongs pointing to the right. The connection to the bending of a card is evident. This
pitchfork bifurcation is called supercritical to signify that the nontrivial fixed points that are
born at the bifurcation are stable (also, two nontrivial fixed points exist only above the critical
point if r is increased.).
• It is instructive to look at the potential for the vector field v(x) = rx − x3 . It is a quartic
potential φ(x) = − 21 rx2 + 14 x4 with v = −φ′ . Viewed as a potential for a particle moving on the
line, we see that the quartic term is attractive/stabilizing. For r < 0, φ has only one extremum,
a global minimum at x∗ = 0, the minimum becomes very flat φ = x4 /4 at the critical point
r = 0 while it turns into a double-well potential for r > 0.
• The basin of attraction of a fixed point x∗ is the set of ICs x0 for which x(t) → x∗ as t → ∞.
If the basin of attraction is the whole phase space, we say that x∗ is a global attractor. In the
√
above case, x∗ = 0 is a global attractor for r ≤ 0. For r > 0, the two fixed points x∗ = ± r
have basins of attraction consisting of the positive and negative half lines x0 > 0 and x0 < 0,
while x0 = 0 is the only point in the basin of attraction of the unstable fixed point at x∗ = 0.
• A system with a Z2 symmetry displaying a second order phase transition is a 2d Ising spin
system where ‘spins’ which can either point up or down are located on a square lattice. Neigh-
boring spins tend to align due to the exchange interaction but thermal fluctuations can prevent
all spins from being aligned the same way. The system as a whole is invariant under reversal
of the sign of all spins. The system displays a 2nd order paramagnetic to ferromagnetic transi-
tion as the temperature (control parameter) is lowered below the Curie temperature Tc . In the
paramagnetic phase the net magnetization (∝ total spin) is zero (due to thermal fluctuations)
corresponding to the single fixed point x∗ = 0. In the ferromagnetic phase there is a non-zero
net ‘spontaneous’ magnetization with all spins in a magnetic domain pointing either up or down
√
(corresponding to the stable fixed points x∗ = ± r ). In the vicinity of the 2nd order phase
transition at Tc , there are long-range fluctuations, one of whose effects is that one needs to
average over a very large number of configurations to reliably calculate mean values in a Gibbs
ensemble. This is one manifestation of critical slowing down. Moreover, correlation functions
decay as a power law at the critical temperature while away from the critical temperature, they
decay exponentially.

3.3.2 Subcritical pitchfork bifurcation

Reversing the sign of the quartic term in the potential φ or the cubic term in v leads to a
subcritical pitchfork bifurcation. Thus we consider

ẋ = rx + x3 = x(r + x2 ) (24)

11
 Figure 2: Potential and Bifurcation diagram for ẋ = v(x) = rx + x3 − x5 .

1 2 1 4
which corresponds to the quartic potential √ φ(x) = − 2 rx − 4 x . For−r < 0, there are three
fixed points x∗ = 0 (stable) and x∗ = ± −r (unstable). When r → 0 the three fixed points
merge into a single unstable fixed point at x∗ = 0 which then survives as an unstable fixed
point for r > 0. The bifurcation diagram for a subcritical pitchfork bifurcation looks like a
laterally inverted version of the supercritical one, except that the non-trivial fixed points are
unstable in the subcritical case, while they were stable in the supercritical case. It √ is called
a subcritical pitchfork bifurcation to convey that the nontrivial fixed points x∗ = ± −r are
unstable (also, the non-trivial fixed points exist only for r below the critical value rc = 0
for our setup). What is more, for r ≥ 0, the origin is very unstable and all non-constant
trajectories blow-up in finite time. This is easily seen for r = 0 where the solution to the IVP
is x(t) = sgn (x0 )(1/x20 − 2t)−1/2 . We see that x(t) → ±∞ according as sgn (x0 ) = ±1 at the
blowup time tb = 1/2x20 for any non-zero x0 .
• In practice, additional nonlinear effects would become important once the solutions grow
sufficiently. The simplest model incorporating these (that retains the x → −x symmetry) is

ẋ = v(x) = rx + x3 − x5 . (25)

It is clear that x∗ = 0 is always a fixed point, it is linearly stable for r < 0 and unstable for
r > 0. The other fixed points are roots of the quartic equation r + x2 − x4 = 0. They are at
1/2
√

±± 1
x∗ = ± (1 ± 1 + 4r) . (26)
2
For r < rs = −1/4, x∗ = 0 is the only fixed point. For −1/4 < r < 0 there are four non-
±,−
trivial fixed points x±±
∗ . While x∗ are unstable, x±,+
∗ are stable. For r > 0 we have only
±,+
two non-trivial fixed points x∗ , both of which are stable. This vector field displays both
a subcritical pitchfork bifurcation at r = 0 as well as a pair of saddle-node bifurcations at
rs = −1/4 where the stable (x±,+
∗ ) and unstable (x±,−
∗ ) large-amplitude fixed points are born.
Draw the bifurcation diagram.
• It is also instructive to draw the potential φ(x) = −rx2 /2 − x4 /4 + x6 /6 in the three qual-
itatively different regions: (a) r < −1/4, (b) −1/4 < r < 0 and (c) r > 0. For r < −1/4,
the quartic term does not have a chance to play a qualitative role (for small x it is dominated
by the quadratic term with large coefficient, while for large x it is dominated by the x6 term!)

12
Figure 3: Plot of magnetization vs applied magnetic field for various temperatures in a magnetic domain. For
T > Tc (paramagnetic phase) M = χH for small H (linear response regime) while M saturates for large H . For
T ≤ Tc (ferromagnetic phase) there is a residual (spontaneous) magnetization even when H → 0± . The memory
effect (hysteresis loop) is also shown. Thanks to Sonakshi Sachdev for help with the Figure.

and φ has only one extremum, a global minimum at x∗ = 0. For −1/4 < r < 0, there is a
local minimum at x∗ = 0, a pair of local maxima at x±,−
∗ and a pair of global minima at x±,+
∗ .
Finally, for r > 0, φ has a local maximum at x∗ = 0 and a pair of global minima at x±,−
∗ .
• Interestingly, this system can display jumps and hysteresis. Thus, unlike the supercritical
bifurcation that is associated to second order (continuous) phase transitions, the subcritical one
is associated to first order (discontinuous) phase transitions. Indeed, suppose we begin with the
system in its unique stable equilibrium x∗ = 0 for r < −1/4 and imagine slowly increasing r .
As the trivial fixed point is linearly stable up to r = 0− , the system remains at x∗ = 0. At
r = 0, the slightest perturbation causes the system to make a discontinuous transition or ‘jump’
from x∗ = 0 to one of the two stable fixed points x∗ = x±,+ ∗ . Suppose it jumps to the ‘upper’
branch x∗ . With further increase in r the system remains in the stable equilibrium x+,+
+,+
∗ .
Now suppose r is decreased. Interestingly, the system remains in the state x+,+ ∗ even when r
is decreased below r = 0, as it is a stable state. In other words, the system does not retrace its
path even when the control parameter retraces its path! Hysteresis is the name for this lack in
reversibility, it is seen in magnets and are associated to first order phase transitions. A second
jump occurs when r is decreased down to rs = −1/4 at which point the non-trivial fixed points
cease to exit and the system jumps back to the stable equilibrium x∗ = 0. Draw a figure.

4 Vector fields on a circle

Vector fields on a circle will provide us the simplest examples of dynamical systems that can
display oscillations.
• The integral curves of a vector field v(θ) on a circle are solutions of the first order ODE
θ̇ = v(θ). Here, the 2π -periodic coordinate θ parametrizes points on a circle, with θ increasing
counter-clockwise. In particular, v(θ) must be 2π -periodic so that the vector field is single-
valued.
• Uniform circular motion: The constant vector field v(θ) = ω (for ω 6= 0) describes
uniform oscillations. Indeed, the solution to the IVP θ = θ0 + ωt describes a phase point
rotating uniformly (counter-clockwise if ω > 0 and clockwise if ω < 0) at angular velocity ω .

13
Notice that this vector field has no fixed points! The phase point returns to its initial location
after a time period T = 2π/|ω|. ω represents a constant torque that drives the phase point
round and round.
• Overdamped pendulum: The example ẋ = − sin x considered earlier is in fact better
viewed as a dynamical system on the circle. If we regard x = θ as an angle then θ = 0, π
(mod 2π ) are the only fixed points, they lie at the bottom and top of the circle. The flow is
clockwise on the right semi-circle and counter-clockwise on the left. Thus θ = 0 is stable while
θ = ±π is unstable. In this case, all trajectories other than the constant one at θ = π eventually
end up at the attractive fixed point θ = 0. There is no oscillatory behavior and no non-trivial
periodic solutions exist. The equation describes an overdamped pendulum. θ = 0 is the stable
equilibrium with bob hanging down and θ = π is the unstable equilibrium with bob pointing
upwards.
• Overdamped pendulum subject to constant torque: Let us now ‘combine’ the above
two examples and consider the vector field

θ̇ = ω − a sin θ (27)

with ω ≥ 0 and a ≥ 0. This equation arises in the overdamped limit of a pendulum driven by
a constant torque τ :
ml2 θ̈ = −γ θ̇ + τ − mgl sin θ, (28)
as well as in other contexts (e.g. Josephson junctions and charge density waves, see Strogatz).
We may think of ω as a measure of the counter-clockwise torque and of a as a measure of the
downward acceleration due to gravity, which tries to make the bob come down to θ = 0. On
the other hand, ω drives the bob uniformly counterclockwise. For definiteness, we will hold
ω > 0 fixed and imagine increasing the control parameter a. When the gravitational force is
absent (a = 0), we have uniform circular motion at frequency ω and time period 2π/ω . When
0 < a < ω , the motion around the circle becomes non-uniform. Draw a phase portrait and plot
v(θ). Since v(θ) > 0, the motion is still counterclockwise and periodic: there are no fixed points.
The constant torque opposes gravity maximally at the bottleneck around θ = π/2 where the
bob slows down. The two forces reinforce each other most at θ = −π/2 where the bob speeds
up. For a ≪ ω the non-uniformity is small. When a . ω , the pendulum moves very slowly near
θ = π/2 but covers the rest of the circle quickly. At a = ac = ω , a half-stable fixed point is born
at θ = π/2 in a saddle-node bifurcation. Draw the corresponding phase portraits. For a > ω ,
we have two fixed points θ∗± on either side of θ = π/2 at sin θ∗ = ω/a. Show (using linear
stability analysis) that θ∗+ (the one to the right of π/2) is unstable while the one to its left (θ∗− )
is stable. In this phase, the motion is no longer oscillatory and all non-constant trajectories are
attracted to the stable fixed point θ∗− whose basin of attraction is the circle with θ∗+ removed.

5 Linear vector fields on the plane and their fixed points

Non-constant integral curves of vector fields on the plane display richer behavior (e.g. periodicity,
oscillations, limit cycles, homo and heteroclinic orbits etc) than on a line where they had to
monotonically approach a stable fixed point or ±∞ without oscillation or the possibility of
periodic motion. We will begin by examining linear systems in 2d as they are simpler and also
help to classify fixed points of more general flows on a plane.

14
• A (homogeneous) linear vector field on the plane v = Ar, where r is a 2-component vector
and A a 2 × 2 real (not necessarily symmetric) matrix leads to the 2d homogeneous linear
system     
ẋ a b x
= or ṙ = v = Ar. (29)
ẏ c d y
The IVP now requires two initial conditions x(0) and y(0). It is clear that the origin x∗ = (0, 0)
is always a fixed point. There could be non-trivial fixed points r∗ 6= 0 with Ar∗ = 0, if A is
singular (i.e. if det A = 0 or equivalently if it has a non-trivial kernel or null-space). More
precisely, the origin is the only fixed point if det A 6= 0, we have a line of fixed points along
the null space of A if the null space is one-dimensional and a plane of fixed points if A has a
two-dimensional null-space (in which case A = 0). The trajectories are directed parametrized
curves (x(t), y(t)) on the x-y phase plane whose velocity vector is v . Poincaré classified the
possible phase portraits of such systems. Let us begin with some examples that illustrate the
types of fixed points that can occur.
• It is noteworthy that the time-dependent Schrödinger equation of quantum mechanics i~ψ̇ =
Hψ is a linear equation of the above sort, with the Hamiltonian operator H defining a vector
field on the state space. However, the state space in QM is in general a complex vector space
whose dimension could be finite or infinite and H is restricted to be hermitian.

5.1 Examples of fixed points: center, node, saddle

• Center (simple harmonic oscillator): Linear systems in 2d naturally arise when studying
the Newtonian dynamics of a particle moving on a line subject to a linear force. In the language
of mechanics, such systems have one degree of freedom and a two-dimensional phase space. A
familiar example is the simple harmonic oscillator mẍ = −kx that arises in describing small
oscillations of amplitude x, of a mass m attached to a spring with force constant k . Introducing
the linear momentum p = mẋ we write this as the pair of coupled linear equations for the
coordinates (x, p) on the phase plane
      
ẋ p/m 0 1/m x
= = . (30)
ṗ −kx −k 0 p

In this case, the coefficient matrix is non-singular (det A = k/m = ω 2 > 0) so that the origin
is the only fixed point. It corresponds to the static solution (x = p = 0) where the mass is
always at rest. More generally, show that the phase trajectories are clockwise-directed closed
curves centered at the origin and display them on a phase portrait. These periodic trajectories
describe an oscillating mass. In fact, the general solution to the above ODEs is

x(t) = A cos(ωt + ϕ) and p(t) = −mωA sin(ωt + ϕ) (31)

where ϕ and A are constants determined by initial conditions. Unlike with flows on a line,
which had to be gradient flows with a monotonically decreasing ‘potential energy’, this is a
conservative system. It is easily verified that the energy or Hamiltonian H = p2 /2m + kx2 /2 is
constant along trajectories. In fact, the trajectories coincide with the level curves of H , which
are ellipses.
What is the nature of the fixed point r∗ at the origin? It is called a center or an elliptic fixed
point or an O-point. It is neither an attractive fixed point nor a repulsive one: trajectories that

15
begin nearby do not approach it nor go away from it asymptotically (as t → ∞). However, a
trajectory that begins nearby remains nearby: this is called Lyapunov stability, after the Russian
mathematical physicist. Furthermore, we say that the fixed point at the origin is neutrally stable
since it is Lyapunov stable but not attracting. Coincidentally, the matrix A is not diagonalizable
over the reals: it has purely imaginary eigenvalues (±iω ). We will see later how this is related
to the stability of r∗ = 0.
• We have introduced two notions of stability here. The first concerns asymptotic behavior: a
fixed point r∗ is said to be attractive if trajectories that begin sufficiently close to r∗ approach r∗
as t → ∞. On the other hand, Lyapunov stability concerns behavior at all times: r∗ is Lyapunov
stable if trajectories that begin sufficiently close to r∗ remain close to r∗ at all subsequent times3 .
A fixed point that is both attracting and Lyapunov stable is called asymptotically stable. Though
the two often occur together, they need not. The origin in the harmonic oscillator is Lyapunov
stable but not attracting. On the contrary, a fixed point may be attracting but not Lyapunov
stable, if trajectories that begin nearby go far away before eventually reaching r∗ . An example
is given by the half-stable fixed point θ∗ = π/2 of the vector field v(θ) = ω(1 − sin θ) on the
circle. It is attracting, since all trajectories eventually approach θ = π/2. However, it is not
Lyapunov stable as the trajectories have to go ‘round the circle’ before eventually falling into
the fixed point.
• Node, star, saddle and line of fixed points: Consider the family of linear systems on
the plane parametrized by a real number α:
    
ẋ α 0 x
= with x(0) = x0 and y(0) = y0 . (32)
ẏ 0 −1 y

As long as α 6= 0, the origin is the only fixed point. When α = 0, every point on the x axis is
a fixed point. In both cases, the equations decouple and the solutions are easily written down:

x(t) = x0 eαt and y(t) = y0 e−t . (33)

It is clear that y monotonically decays to zero while the behavior of x depends on α. Draw
the phase portrait for five values of α: α < −1, α = −1, −1 < α < 0, α = 0 and α > 0.
(i) When α < −1 the origin is the only fixed point. All non-constant trajectories approach it
while becoming asymptotically tangent to the y -axis, as x decays faster than y . The fixed point
at the origin is called a stable node. It is both Lyapunov stable and attracting and therefore
asymptotically stable. (ii) When α = −1, A = −I and x and y approach zero at the same rate.
All non-constant trajectories are straight lines that approach the origin as t → ∞. The origin
is still an asymptotically stable node but in this case it is called a symmetrical node or star.
(iii) For −1 < α < 0 the origin remains an asymptotically stable node, but with trajectories
approaching it along the x-axis as y decays faster. (iv) If α = 0, x is independent of time
and we have a line of fixed points along the x axis. Non-constant trajectories are vertical lines
approaching one of these fixed points either from the upper or lower half plane. Each of these
fixed points r∗ is neutrally stable (Lyapunov stable but not attracting), as trajectories that
3
More precisely, r∗ is Lyapunov stable if given any tolerance ǫ > 0 there exists a δ > 0 so that we can
guarantee that |r(t) − r∗ | < ǫ for all t > 0 provided we start close enough ( |r(0) − r∗ | < δ ). If we do not have
notions of distance but only the topological notion of open sets, this condition can be formulated in terms of open
neighborhoods. r∗ is Lyapunov stable if given any non-empty neighborhood N of r∗ , there exists a non-empty
neighborhood V of r∗ such that if the initial condition lies within V then r(t) is guaranteed to lie within N for
all t ≥ 0 .

16
begin nearby remain nearby but may not approach r∗ . (v) Finally, when α > 0, the fixed
point at the origin is a saddle point, also called an X-point. It is an example of what is called
a hyperbolic fixed point4 . Trajectories with x0 = 0 approach it along the y axis. But all other
non-constant trajectories asymptotically approach (x = ±∞, y = 0). Thus, the saddle point is
neither Lyapunov stable nor attractive. For α = 1, the phase portrait consists of a family of
hyperbolae with xy constant (in general xy α is constant on trajectories). For a saddle point
(and more generally for hyperbolic fixed points), one defines the notions of stable and unstable
manifolds. The stable manifold Ws of a hyperbolic fixed point r∗ is defined as the set of ICs
(x0 , y0 ) (points on phase space) for which the trajectory tends to r∗ as t → ∞. The unstable
manifold Wu is similarly the set of ICs for which the trajectory tends to r∗ as t → −∞. In the
case at hand, Ws is the y -axis while Wu is the x-axis. We observe that a typical trajectory (in
this case, every trajectory) approaches the unstable/stable manifold as t → ±∞.

5.2 Analysis of phase portrait using eigenvalues and eigenvectors of coefficient matrix

• One reason we could easily analyze the second example above is that the equations for x
and y decoupled as A was a diagonal matrix. The horizontal and vertical directions were
eigendirections of A. In general, an IC in an eigendirection leads to simple evolution. Suppose
w is a (real) eigenvector of A with (real) eigenvalue λ, Aw = λw . Then if the initial state
r(0) = r0 w is in this eigendirection, then the trajectory remains along w at all times. Indeed,
we see that the solution is given by r(t) = eλt r0 w since ṙ = λr(t) is the same as Ar(t).
• A with two linearly independent eigenvectors: Somewhat more generally, suppose the
real matrix A = (ab|cd) has two distinct eigenvalues
1h p i
λ± = (a + d) ± (a + d)2 − 4(ad − bc) . (34)
2
Then we are guaranteed5 that the corresponding eigenvectors w± are linearly independent.
Thus, if we expand the initial state as r0 = r− w− + r+ w+ , then we can immediately write down
the solution to the IVP:
r(t) = r− eλ− t w− + r+ eλ+ t w+ . (35)
Notably, this formula works even if the eigenvalues and eigenvectors are not real! However r0
must be real; a convenient way of ensuring this is to exploit the fact that λ± must be complex
conjugates to take w± to also be complex conjugates and r+ = r− ∗ . An immediate consequence

is that if the real parts of the eigenvalues λ± are negative, then the origin is an attractor while
it is unstable if ℜλ± > 0. However, when the eigenvalues are not real, the eigenvectors are also
not real and the corresponding eigenspaces are not lines through the origin on the plane; they
cannot be interpreted as invariant subspaces for the dynamics on the plane.
On the other hand, if the eigenvalues are real, then the eigenvectors can be taken real
and define invariant subspaces for the dynamics on the phase plane. They are unstable/stable
depending on whether the corresponding eigenvalue is positive/negative. (a) If both are of the
4
Somewhat confusingly, the name hyperbolic fixed point is used for a wider class of fixed points that includes
saddles, nodes and spirals, as we will clarify in the next section.
5
Suppose Av = λv and Aw = µw are two (non-zero) eigenvectors corresponding to distinct eigenvalues
λ 6= µ . Suppose v = αw are linearly dependent. Then Av = λv and also Av = Aαw = αµw = µv so that
(λ − µ)v = 0 or λ = µ , contradicting the hypothesis.

17
same sign, the origin is either a stable or unstable node. (b) It is a saddle point if they have
opposite signs. (c) If λ+ = λ− 6= 0, and they nevertheless have linearly independent eigenvectors,
then A = λ+ I and we have a stable/unstable star or symmetrical node depending on whether
the eigenvalues are positive or negative. (d) If one eigenvalue, say λ− is zero and the other is
non-zero, then we have a line of fixed points along the 0-eigenspace. The fixed points are
stable if λ+ < 0 and unstable if λ+ > 0. (e) Perhaps not surprisingly, if both eigenvalues vanish
but A has two linearly independent eigenvectors, then A must be the zero matrix and every
point on the plane is a neutrally stable fixed point.
(f) On the other hand, if the eigenvalues are purely imaginary λ± = ±iω , then the trajectory

r(t) = r− e−iωt w− + r+ eiωt w+ (36)

is periodic with period 2π/ω , as in the case of the harmonic oscillator and the origin is a
neutrally stable center. In fact, for the SHO A = (0, 1/m| − k, 0) has eigenvalues ±iω where
ω = k/m. The corresponding eigenvectors can be taken as w± = (1, ±imω)t . We notice
p

that the eigenvectors cannot be taken real, so they do not define invariant subspaces on the x-p
phase plane. However, reality of A implies that w± can be taken to be complex conjugates so
that reality of r then implies that r± must also be complex conjugates. The solution r = (x, p)
in the case of the SHO can then be written as

x(t) = r+ (cos ωt + i sin ωt) + r− (cos ωt − i sin ωt) = 2(ℜr+ ) cos ωt − 2(ℑr+ ) sin ωt and
p(t) = 2mω [−ℜr+ sin ωt − ℑr+ cos ωt] . (37)

Summing the squares we find that (x/2)2 + (p/2mω)2 is a constant (depending on the complex
number r+ , which is determined by the ICs). Thus, the trajectories of the SHO are ellipses
centered at the origin.
(g) Spiral sink or source: A center is not structurally stable. What this means is that a
center is generally nor preserved under a perturbation of system parameters a, b, c and d. Typi-
cal perturbations (that preserve the reality of A) would cause the pure imaginary eigenvalues to
move off the imaginary axis and become truly complex and conjugate to each other. In fact, if
the eigenvalues are not real or imaginary, they must be of the form λ± = γ ± iω (with γ, ω 6= 0).
The solution is then given by

r(t) = eγt r− e−iωt w− + r+ eiωt w+

(38)

In this case, we have a stable or unstable spiral (spiral sink or source, sometimes called a
focus) according as γ is negative or positive. γ is called the growth or decay rate. Physically,
the motion is like that of an underdamped oscillator.

5.2.1 Damped harmonic oscillator

This gives us an opportunity to take a quick look at the damped harmonic oscillator defined by
Newton’s equation for a particle whose displacement from equilibrium is denoted x:

mẍ = −kx − γ ẋ. (39)

p
The linear restoring force is −kx and we define the frequency ω = k/m as before. Here γ is
the damping coefficient, it must be positive to ensure that the damping force acts in a direction

18
opposite to the instantaneous velocity, thereby slowing the particle down. Introducing p = mẋ,
(39) can be written as a first order system on the x-p phase plane
    
ẋ 0 1/m x
= , (40)
ṗ −k −γ/m p

with initial conditions x(0) = x0 and p(0) = p0 . The eigenvalues of the coefficient matrix are
1  p 
λ± = −γ ± γ 2 − 4m2 ω 2 . (41)
2m
It is useful to note that det A = λ+ λ− = ω 2 where A is the above coefficient matrix. Since
ℜλ± = −γ/2m < 0, the origin is an attractor: the particle eventually comes to rest at the
equilibrium point x = 0. There are three qualitatively different parameter regimes.

1. If γ < 2mω (weak damping), then, λ± are complex and the origin is a spiral sink. The
motion is said to p
be underdamped: x oscillates infinitely often (with ‘time period’
2π/|ℑλ± | = 4πm/ 4m2 ω 2 − γ 2 ) with decreasing amplitude as it settles down to x = 0.
The phase trajectory winds clockwise around the origin.

2. When γ = 2mω , the eigenvalues coincide λ± = −γ/2m = −ω . However, in this case A

is not proportional to the identity, and there is only one linearly independent eigenvector,
which can be taken as (1, −mω)t . The motion is said to be critically damped and the
origin is a degenerate node. We will look at this case in more detail shortly.

3. If γ > 2mω (strong damping), then λ± are real, distinct and negative (λ− < λ+ < 0).
The origin is an asymptotically stable node. The motion is said to be overdamped: the
particle performs less than one full oscillation before coming to rest at x = 0.
In fact, the eigenvectors corresponding to λ± can be taken as
 
λ∓
w± = . (42)
mω 2

Since λ− < λ+ < 0, w+ points roughly towards west-north-west (WNW) while w−

points in a NNW direction. They enclose a ‘small’ wedge in the 2nd quadrant with acute
opening angle while −w− and −w+ do so in the 4th quadrant. Similarly, w− and −w+
(or −w− and w+ ) enclose a ‘large’ wedge with obtuse opening angle. The trajectory
r(t) = r+ eλ+ t w+ + r− eλ− t w− approaches the origin along w+ as t → ∞. As t → −∞
both components of r become unbounded, but the w− component grows faster, so r points
along ±w− as t → −∞. Sketch the phase portrait. We see that x either monotonically
approaches 0 or changes sign once (i.e. passes through the equilibrium point once) before
eventually reaching 0 as t → ∞. The first case occurs if the IC is in one of the two small
wedges with acute opening angles and the latter possibility occurs if the IC lies in one of
the two large wedges with obtuse opening angles.

5.2.2 Degenerate eigenvalues and deficient coefficient matrix

The example of the critically damped oscillator encourages us to consider the special case where
the two eigenvalues of A are equal but there is only one linearly independent eigenvector. This

19
degenerate case is somewhat exceptional (non-generic) as a small perturbation of matrix elements
would make the eigenvalues distinct. Nevertheless, it is an interesting case as the eigenvectors
do not furnish a basis in which to expand a general initial state and the matrix A may be said
to be deficient. For instance, consider the ‘strictly upper-triangular’ system
 
0 1
ẋ = y, ẏ = 0 corresponding to A = . (43)
0 0

A has only one eigenvalue 0, with algebraic multiplicity two. The corresponding null-space is
spanned by the one linearly independent eigenvector, say x̂ = (1, 0)t . Thus, the horizontal axis
is a line of fixed points. The solution to the IVP is y = y0 and x = x0 + y0 t. The flow is
horizontal and to the right in the upper-half plane and to the left in the lower-half plane. Draw
the phase portrait. None of the fixed points is stable, as a small perturbation either upwards or
downwards would go off to x = ±∞.
• More generally, consider the ‘upper-triangular’ system
 
λ 1
ẋ = λx + y, ẏ = λy corresponding to A = with λ < 0. (44)
0 λ

As A is invertible, the origin is the only fixed point. λ is a doubly degenerate eigenvalue with
only one linearly independent eigenvector, say (1, 0)t , so that the x axis is an invariant subspace
with points driven towards the origin: (x(0), 0) 7→ (x(0)eλt , 0). More, generally, since A is upper
triangular, y evolves independently of x and is given by y(t) = y(0)eλt . We are left with an
inhomogeneous equation for x:

ẋ = λx + y(0)eλt with solution x(t) = x(0)eλt + y(0)teλt . (45)

Draw the corresponding phase portrait noting that the slope of trajectories is given by y(t)/x(t) =
y(0)/(x(0) + y(0)t). The slope approaches 0− in the far past and 0+ in the far future, so the
flow is generally ‘clockwise’. The origin is an asymptotically stable degenerate node. In that
case, the motion is like that of a critically damped oscillator. The same formulae apply also
when λ > 0, in which case the origin is an unstable degenerate node.
• Let us look at the critically damped oscillator with γ = 2mω in more detail
    
ẋ 0 1/m x
= . (46)
ṗ −k −2ω p

As noted, the coefficient matrix A = (0, 1/m| − k, −2ω) has the doubly degenerate eigenvalue
λ = −ω with only one linearly independent eigenvector which we take as w = (1, −mω)t . To
facilitate solving the equations we will make a similarity transformation to a basis where A
is upper/lower triangular (essentially in Jordan form). This basis consists of eigenvectors and
generalized eigenvectors of A. Eigenvectors w are non-trivial solutions to (A − λI)w = 0,
they lie in the kernel of (A − λI). Generalized eigenvectors are vectors that lie in the kernel of
(A − λI)p for some p > 1 but not in the kernel of (A − λI)q for q < p. In our case (A − λI)2 is
identically zero so we can take our generalized eigenvector to be any vector linearly independent
of w , say w′ = (0, 1)t .
The resulting similarity matrix is S = (w, w′ ) = (1, 0| − mω, 1) with inverse S −1 =
(1, 0|mω, 1). Applying the similarity transformation we get J = S −1 AS = (−ω, 1/m|0, −ω)

20
which is upper triangular. The ODE
 
−1 −1 −1 −1 x
ṙ = Ar = SJS r now becomes S ṙ = J(S r) where S r= . (47)
mωx + p

It is natural to define y = mωx + p, which evolves independently of x due to the triangular

structure:
    
ẋ −ω 1/m x y
= or ẏ = −ωy and ẋ = −ωx + . (48)
ẏ 0 −ω y m

Thus  
−ωt −ωt y0 t
y(t) = y0 e and x(t) = e x0 + where y0 = mωx0 + p0 . (49)
m
The solution to the IVP may now be expressed as
 
−ωt p0 t
x(t) = e x0 (1 + ωt) + and p(t) = mẋ = e−ωt [−kx0 t + p0 (1 − ωt)] . (50)
m

The origin (x = 0, p = 0) is an asymptotically stable degenerate node. w is the only invariant

subspace. Trajectories that don’t start along w make roughly half a turn clockwise around the
origin before reaching it. The phase portrait in the critically damped case may be viewed as
resulting from the overdamped portrait when the eigenvectors w± become collinear and the
smaller wedges shrink letting the larger wedges become half-planes. Every trajectory that does
not start along w crosses the equilibrium point x = 0 once (at t = −mx0 /(mωx0 + p0 ), which
may be negative) before eventually reaching the fixed point as t → ∞. Similarly, every (possibly
extended) trajectory comes to rest momentarily once at t = p0 /(kx0 + ωp0 ) before eventually
coming to rest at the fixed point.

5.2.3 Trace-determinant classification of fixed points

• While the formulation in terms of eigenvalues allows us, for the most part, to classify fixed
points and determine their stability, the (possibly complex) eigenvalues themselves depend only
on two real numbers: the trace τ = a + d and determinant ∆ = ad − bc of A = (ab|cd). Thus,
by and large, one obtains a more economical classification in terms of τ and ∆. In other words,
though the space of linear systems defined by A = (ab|cd) is four dimensional, the character of
the system is largely determined6 by just two real control parameters τ and ∆. In fact, the
characteristic equation for A (λ2 − τ λ + ∆ = 0) implies that the eigenvalues are
1 p
λ± = (τ ± τ 2 − 4∆) with τ = λ + + λ− and ∆ = λ+ λ− . (51)
2
Thus, we may try to determine the nature of the fixed points and phase portrait if we are given
a (τ, ∆) pair. To begin with, the sign of the discriminant D = τ 2 − 4∆ tells us that λ± are (a)
distinct and complex conjugates if D < 0, (b) real and distinct if D > 0 and (c) real and equal
if D = 0. Evidently, the location of (τ, ∆) relative to the parabola ∆ = τ 2 /4 plays a crucial
role.
6
There are exceptional cases where the trace and determinant are not adequate to determine the character of
the phase portrait. This happens if A has repeated eigenvalues, as we will see.

21
 At points above the parabola ∆ > τ 2 /4, D < 0 and the distinct complex conjugate eigen-
values with real part τ /2 imply that there are two main possibilities: (i) (Stable) spiral sink if
τ < 0 and (ii) (Unstable) spiral source if τ > 0. We will deal with the boundary τ = 0 below.
At points below the parabola (∆ < τ 2 /4) we have D > 0. The eigenvalues are real and
distinct and lead to the following main possibilities: (1) saddle if ∆ < 0, since in this case the
real eigenvalues have opposite signs, (2) unstable node if ∆ > 0 and τ > 0 since the eigenvalues
are both positive and (3) stable node if ∆ > 0 and τ < 0 since the eigenvalues are both negative.
Mark these on the τ − ∆ plane.
We have thus identified five open regions in the τ − ∆ plane, corresponding to spiral sources
and sinks, stable and unstable nodes and finally saddles. A matrix A corresponding to a point
in any one such region is generic in the sense that a small change in system parameters a, b, c, d
(i.e., while maintaining linearity and reality) does not change the character of the phase portrait.
Special circumstances prevail on the boundaries between the above five regions, with the
behavior not always completely determined by the values of τ and ∆. These are the possible
borderline cases:

1. Boundary between spiral sources and sinks: On the ray τ = p 0, ∆ > 0 (symmetry
axis of the parabola), the eigenvalues are both imaginary λ± = ±i |D| and the origin is
a center. It is neutrally stable, as befits a system that is borderline between spiral sources
and spiral sinks.

2. When τ = ∆ = 0, the eigenvalues vanish and the origin x = y = 0 is no longer an isolated

fixed point: we may either have a plane of fixed points (if A has two linearly independent
eigenvectors so that A = 0) or a line of fixed points along the null space of A if A has
only one eigenvector (e.g. A = (01|00)). The fixed points are neutrally stable in the
eigendirection and unstable in any other direction.

3. The parabola ∆ = τ 2 /4 lies on the boundary between spirals and nodes. On the
parabola, the eigenvalues are real and equal λ± = τ /2. Let us suppose τ, ∆ 6= 0. If A has
two linearly independent eigenvectors, then A is a (non-zero) multiple of the identity and
the origin is an unstable star node if τ > 0 and a stable star node if τ < 0. On the other
hand, if A is deficient and has only one eigenvector, then the origin is a stable/unstable
degenerate node according as τ is negative or positive. This is illustrated by the example
A = (λ1|0λ) where λ = τ /2 6= 0. Visually, the phase portrait of a degenerate node arises
as the phase portrait of a node is morphed to lose one eigenvector. When both eigenvectors
cease to be real, the node becomes a spiral.

4. On the other hand, the pair of rays ∆ = 0, τ 6= 0 lie on the boundary between saddles
and nodes. In this case we have two distinct real eigenvalues λ− = 0 and λ+ = τ . We
again have a line of fixed points along the null space of A. They are neutrally stable
along the null space and unstable/stable in any other direction depending on whether τ
is positive or negative.

These borderline cases are of two sorts. Centers lie on the dividing line between spiral sinks
and spiral sources. A small change in (τ, ∆) can change the stability of a center. Non-isolated
fixed points are similarly on the dividing line between stable/unstable nodes and saddles so that

22
 Figure 4: Nature of fixed points on the trace-determinant plane for a 2d linear system.

a small change in (τ, ∆) can alter the stability. By contrast, the stability of symmetrical and
degenerate nodes is robust, although they lie on the borderline between nodes and spirals.
In summary, aside from these exceptional cases, we may immediately obtain much informa-
tion about the phase portrait from the trace and determinant of A. For instance, a negative
determinant ∆ implies a saddle at the origin, while for positive ∆, we have either a source or
a sink according as τ is positive or negative.
We may regard the above regions of the trace-determinant plane as encoding bifurcation
diagrams for 2d linear systems. A 1-parameter family of such linear systems would define a
curve in the τ -∆ plane. When the curve crosses any of the above boundaries (∆ = 0, ∆ = τ 2 /4
and τ = 0 for ∆ > 0), the character of the phase portrait changes. For instance, upon crossing
the parabola, the origin may go from being a node to a spiral while upon crossing the τ axis,
the origin may go from a saddle to a node. In a damped harmonic oscillator, if we increase
the damping coefficient γ holding the mass m and frequency ω fixed, the system goes from
being underdamped to critically damped at γ = 2mω and finally overdamped as the number of
linearly independent real eigenvectors of A goes from zero to one to two. This corresponds to a
qualitative change in the phase portrait, with the fixed point at the origin transitioning from a
spiral sink to a stable node via a stable degenerate node at the bifurcation point γ = 2mω .

6 Nonlinear vector fields, especially in two dimensions

Given a vector field v = (f (r), g(r)) (with r = (x, y)) on the x-y plane, we have the associated
IVP
ṙ = v(r) or ẋ = f (x, y) and ẏ = g(x, y) with r(0) = r0 = (x0 , y0 ). (52)

The corresponding trajectory is now a plane curve. As in one dimension, sufficient conditions
for existence and uniqueness of solutions are that the components of v be differentiable with
continuous first partial derivatives. If the components f and g are C 1 in some open connected
region R ∈ R2 containing r0 , then a unique solution to the above IVP exists for some time

23
|t| < τ . Uniqueness means trajectories cannot cross. This does not prevent two trajectories
from approaching each other asymptotically, as they can do at a fixed point.
A special topological feature of dynamics on a plane is that a periodic trajectory (a closed
curve) separates the phase plane into an inside and an outside. Trajectories that begin inside
cannot wander outside while those that begin outside cannot come in. We will see that the
Poincaré Bendixon theorem restricts what they can do. One new possibility that we did not
encounter in 2d linear systems is that of a limit cycle: an isolated periodic trajectory that
is approached asymptotically (as t → ∞ or −∞) by trajectories from the inside and outside.
Another new feature is the possibility of multiple isolated fixed points and the related
possibility of homoclinic and heteroclinic trajectories. A heteroclinic trajectory approaches
a pair of distinct fixed points as t → ±∞ while a homoclinic trajectory approaches the same
fixed point as t → ±∞. Draw examples of phase portraits displaying these.
Though 2d flows arising from autonomous systems are quite interesting and possible to
analyze, they do not display chaos, as the trajectories do not have enough room to maneuver.
One needs a 3d phase space to find examples of chaos in autonomous systems.
• Example 1: As a first example of a nonlinear vector field on the plane, consider Strogatz’s
system ẋ = x + e−y and ẏ = −y . There is only one fixed point: (x∗ = −1, y∗ = 0). Since
y(t) = y0 e−t decays to zero, trajectories must asymptotically approach the x axis or lie along
the x axis. Thus for large times, ẋ ≈ x + 1 so that x ≈ constant × et → ±∞ in the far future,
provided we do not begin at the fixed point. This suggests that the fixed point is unstable even
though the flow is towards the x axis. This is confirmed by the fact that along the x axis,
the solution x(t) = (x0 + 1)et + 1 tends to ±∞ according as x0 is greater or lesser than −1.
In sketching phase portraits, it sometimes helps to draw the nullclines. Nullclines are curves
along which the flow is either horizontal or vertical. For instance, the flow is horizontal on the
nullcline ẏ = −y = 0, (the x-axis) as already noted. On the other hand, the flow is vertical on
the nullcline ẋ = x + e−y = 0 which corresponds to the graph of the function y = − log(−x),
which lies in the left half-plane. Along this nullcline, the flow is downward for y > 0 and upward
for y < 0. In general, nullclines intersect at the fixed points, in this case there is just one fixed
point. What is more, the nullclines divide the phase plane into regions where ẋ and ẏ have
a fixed sign. Here there are four such regions roughly corresponding to the four quadrants.
Proceeding from the first quadrant counter-clockwise, the flow is roughly to the south-east,
south-west, north-west and north-east. It is important to recognize that nullclines are in general
not trajectories7 . Trajectories can cross nullclines as long as they are not trajectories (in our
example, the nullcline y = 0 happens to be the union of two trajectories and a fixed point). We
may use these facts to sketch the phase portrait, showing that (−1, 0) is a nonlinear version of
a saddle. Check the portrait by plotting it numerically using a computer. Having done this,
one wonders whether we could have inferred that the phase portrait in the neighborhood of
(−1, 0) looks like a saddle with lesser effort. This is indeed possible, by studying the linear
approximation to the vector field near the fixed point.
7
A ẏ = 0 nullcline is a union of trajectories provided it is horizontal and an ẋ = 0 nullcline is a union of
trajectories if it is vertical

24
 1.0 2 2

0.5 1 1

0.0 0 0

-0.5 -1 -1

-1.0 -2 -2
-2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 -2 -1 0 1 2 -2 -1 0 1 2

Figure 5: (a) Nonlinear saddle (b) Dipole field mistaken for a plane of fixed points (c) Spiral mistaken for a
center.

6.1 Linearization around fixed points

As in 1d, to understand the behavior of a 2d vector field v(r) = (f, g) near a fixed point
r∗ = (x∗ , y∗ ), it helps to linearize it. Writing x = x∗ + δx and y = y∗ + δy for small (δx, δy),
the linearization of the ODEs ṙ = v reads
!  
  ∂f ∂f
d δx δx
= ∂x ∂g
∂y
∂g . (53)
dt δy ∂x ∂y
δy
r∗

The coefficient matrix of this linear system is the Jacobian matrix of first partials, evaluated at
the fixed point (indicated by the r∗ subscript).
A natural question is whether we may trust the linearization of v to correctly capture the
qualitative nature of the immediate vicinity of the fixed point. A useful result (more on this in
the next section) states that the linearization may be trusted as long as it predicts a fixed point
corresponding to a point away from the parabola ∆ = τ 2 /4 line ∆ = 0 or ray τ = 0, ∆ ≥ 0 on
the τ − ∆ plane. In other words, a linear spiral, node or saddle is a spiral, node or saddle of the
corresponding nonlinear system (in the sense that the phase portrait of the nonlinear system in
a neighborhood of the fixed point can be continuously deformed into that of its linearization).
On the other hand, a linear center or non-isolated fixed point may or may not survive when the
effects of nonlinearities are included8 . Let us consider examples that illustrate some of these
features.
• E.g. 1. A dipole mistaken for a plane of fixed points: A dipole field is, for instance,
the magnetic field around a point magnetic dipole moment on a plane that contains the axis of
the dipole. The equations for the field lines (of a dipole at the origin and pointing along x) are
given by the system
ẋ = x2 − y 2 and ẏ = 2xy. (54)
These equations are the real and imaginary parts of ż = z 2 where z = x + iy . The origin (0, 0)
is the only fixed point. However, the linearization around it leads to a Jacobian matrix that is
identically zero. Thus, linear theory would wrongly suggest a plane of fixed points. This happens
8
Degenerate nodes and star nodes are continuously deformable into ordinary nodes, so the distinction between
node, degenerate node and symmetric node is not useful beyond the linear theory.

25
because the origin is a ‘double zero’ of the vector field z 2 (similar things happen for vector fields
with higher order zeros). The nullclines are the coordinate axes and the lines y = ±x: they
intersect that the origin. The flow is upwards on y = x and downwards on y = −x. It is to
the right on the x-axis and to the left on the y -axis. We can now sketch the phase portrait,
which should be familiar from the field lines of an electric or magnetic dipole. In the upper
half-plane, all trajectories begin and end at the origin and proceed counter-clockwise: they are
homoclinic orbits. The lower half-plane is a mirror image, it consists of clockwise homoclinic
orbits. This phase portrait does not resemble that of any linear 2d system: an attempt at a
linear approximation does a poor job by producing a plane of fixed points. Homoclinic orbits
are an essentially nonlinear phenomenon.
• E.g. 2. Spiral mistaken for a center: As noticed in §5.2.3, centers lie on the borderline
between spiral sinks and spiral sources. Perhaps not surprisingly, linearization can mistake a
spiral for a center. For instance, consider the system θ̇ = 1 and ṙ = −arn in plane polar
coordinates r = (x2 + y 2 )1/2 ≥ 0, θ = arctan(y/x). For a = 0 it is clear that trajectories are
concentric circles and we have a (nonlinear) center at the origin. More generally, suppose a > 0
and let n ≥ 1 be an integer. It is clear that r decreases monotonically to zero while θ grows
linearly. Hence, the phase point spirals counter-clockwise towards the origin, which is a spiral
sink (or source for a < 0).
˙ = −a δr has a 1 × 1 Jacobian with negative eigenvalue
For n = 1, the linearized equation δr
and correctly suggests that the origin is a spiral sink for a > 0. However, for n ≥ 2, the
˙ = 0 incorrectly suggests that the origin is a center. Strictly speaking,
linearized equation δr
polar coordinates breakdown at the origin (where θ is not uniquely defined), so to be on the
safe side, we should check these conclusions by linearizing the vector field in, say, Cartesian
coordinates.

6.1.1 Robustness of the linear (stability) theory and hyperbolic fixed points

Linear centers and non-isolated fixed points of the linearization may not survive when small non-
linear effects are accounted for. However, unlike the neutrally stable center which can be a linear
approximation to both a stable and an unstable spiral, the stability of a node (asymmetrical,
symmetrical or degenerate) does not change due to nonlinear effects.
If we are only interested in the stability of a fixed point (and are willing to gloss over
the difference between, say, a stable degenerate node and a stable node which, in any case, are
continuously deformable into each other) then fixed points may be classified into those that have
robust (linear) stability and those that are marginal. The robust ones are sinks (attractors),
sources (‘repellers’ or attractors as t → −∞) and saddles, corresponding to eigenvalues of the
linearization that have positive real parts, negative real parts or opposite signs respectively. The
stability of such a fixed point is unchanged by nonlinear effects. The marginal ones are centers
(with both eigenvalues imaginary) and non-isolated fixed points (where at least one eigenvalue
vanishes): their stability can be altered by nonlinear effects. In other words, the marginal cases
are those where at least one eigenvalue has zero real part.
• It is useful to introduce a common term for the robust cases: a hyperbolic fixed point is one
at which all the eigenvalues of the linearization of the vector field have non-zero real parts (this
applies to real autonomous systems of dimension one, two or higher). In the above examples,
the origin is a hyperbolic fixed point as long as it is an isolated fixed point and not a center.

26
So saddles, nodes (symmetric or otherwise), spirals and degenerate nodes are all examples of
hyperbolic fixed points. The name is a bit misleading: trajectories need not look like hyperbolae
in the vicinity of a hyperbolic fixed point (except if it is a saddle)!
The linear approximation faithfully captures the behavior in the neighborhood of a hyperbolic
fixed point. The Hartman-Grobman theorem says that the phase portrait in the vicinity of a
hyperbolic fixed point is topologically equivalent (or ‘conjugate’) to that of the linearized system.
Topological equivalence means there is a homeomorphism (continuous map with continuous
inverse) that maps the local phase portrait (an open neighborhood of the fixed point) to the
whole phase plane of the linearized system, taking trajectories to trajectories while preserving
the sense of time. Topologically equivalent phase portraits are continuously deformed versions
of each other. For instance, a homeomorphism would take fixed points to fixed points, closed
orbits to closed orbits and preserve the nature of homoclinic or heteroclinic trajectories.
• Hyperbolic fixed points are structurally stable. This means the topology of the local phase
portrait around a hyperbolic fixed point is unchanged by an arbitrarily small perturbation of
the vector field. As noted, the phase portrait of a center is not structurally stable: a small
perturbation can change the topology of the closed orbits, turning them into spirals. Similarly,
a line of fixed points is not structurally stable: a small perturbation can destroy them.
• A feature of hyperbolic fixed points is that for each eigenvalue λ, eλt approaches 0 either
as t → ∞ or −∞. By contrast, if the eigenvalue is purely imaginary, then eλt goes round and
round the unit circle as t → ±∞ and remains at 1 if λ = 0.
• We can define the stable and unstable manifolds Ws and Wu of a fixed point r∗ : the set
of points other than r∗ that approach it as t → ±∞. For a sink or attractor (e.g. spiral or
node), Wu is empty while Ws is the basin of attraction (it includes all points in a sufficiently
small open neighborhood of r∗ ). For a source/repeller Ws is empty while Wu is the ‘basin of
repulsion’ and includes an open neighborhood of r∗ .
• What can you say about Ws and Wu for a center?
• Suppose (x∗ , y∗ ) is part of a line/curve (1-parameter family) of fixed points. Give an example
of a vector field (with formula and phase portrait) for which the stable and unstable manifolds
Ws , Wu of a suitable fixed point (x∗ , y∗ ) are both non-empty. What can you say about Ws and
Wu of (x∗ , y∗ ) if the vector field is linear?
• Saddles have non-empty stable as well as unstable manifolds. For a dipole field, Ws and Wu
coincide except along the axis of the dipole. Draw phase portraits of suitable vector fields to
illustrate these observations.
• Side remark: Poincaré developed a discrete time dynamics on the surfaces of section that
bear his name. The concepts of stable and unstable manifolds can be extended to the dynamics
on a Poincaré surface. He discovered that remarkable things happen when the stable and
unstable manifolds of a hyperbolic fixed point intersect transversally (on a Poincaré section
through the fixed point). Such a point of intersection is called a homoclinic point (it is not
another fixed point). He showed that if there is one homoclinic point, then there is an infinite
sequence of them accumulating at the original hyperbolic fixed point. This leads to a very
complicated flow pattern which we now associate with chaotic systems. However, this does not
happen for autonomous flows in two dimensions. One needs a three or higher dimensional phase
space (in which the Poincaré surface is embedded) for homoclinic phenomena and chaos to be
possible.

27
 2.0

1.5

1.0

0.5

0.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0

Figure 6: Phase portrait of competitive Lotka-Volterra model. Horizontal and vertical axes are rabbit and sheep
populations.

6.1.2 Competitive Lotka-Volterra model

• The competitive Lotka-Volterra equations may be regarded as a generalization of the logistic

equation Ṅ = rN (1 − N/K) for single species population dynamics. It describes competition
between two species that share or compete for resources, though neither preys on the other9 . For
example, suppose r and s are the populations of rabbits and sheep (say in 1000s). In suitable
units, one postulates that

ṙ = r(3 − r − 2s) = 3r(1 − r/3) − 2rs and ṡ = s(2 − s − r) = 2s(1 − s/2) − sr. (55)

Given that rabbits generally multiply quicker, their linear growth rate (3) is more than that for
sheep (2). Left to themselves, the rabbit and sheep populations would approach the respective
carrying capacities Kr = 3 and Ks = 2 reflecting the fact that rabbits require lesser resources.
Encounters between rabbits and sheep occur at a rate proportional to the product of the popula-
tions and decrease the respective growth rates with the smaller rabbits suffering twice as much.
Note that the quadratic terms r2 and s2 arise from encounters between members of the same
species. Find the fixed points and sketch a phase portrait and discuss its qualitative properties.
Solution: There are four fixed points (r∗ , s∗ ) . Three are to be expected from the logistic equation:
(0, 0) corresponds to no animals, (3, 0) and (0, 2) correspond to exclusive rabbit or sheep populations
equal to their carrying capacities with the other species dying out. Finally, there is a nontrivial fixed
point at (1, 1) where rabbits and sheep coexist. To investigate their stabilities, we linearize the system
around the fixed points. The Jacobian
 
3 − 2r − 2s −2r
A= (56)
−s 2 − r − 2s

evaluated at the fixed points is

       
3 0 −3 −6 −1 0 −1 −2
A0,0 = , A3,0 = , A0,2 = and A1,1 = . (57)
0 2 0 −1 −2 −2 −1 −1

Here, the trace and determinant (τ, ∆) are (5,6), (-4,3), (-3,2) and (-2,-1) and are sufficient to characterize
the fixed points. (0,0) is an unstable node, (3,0) and (0,2) are stable nodes while (1,1) is a saddle. Since
9
The model originally introduced by Lotka and Volterra is related but distinct: it deals with predator-prey
systems.

28
all of these are hyperbolic fixed points, the linear theory correctly captures the topological behavior in
their vicinity. To get some additional information,
√ we may find the eigenvectors of the Jacobian √ at the
saddle point. The eigenvalues are λ± = −1 ± 2 with the corresponding eigenvectors w± = a(∓ 2, 1)
pointing roughly in the West-North-Westerly and ENE directions. w± are tangent to the unstable and
stable manifolds at the saddle (1, 1) . Use these facts to qualitatively sketch the phase portrait. There are
two heteroclinic trajectories from the saddle: one to each of the single-species nodes. They constitute the
unstable manifold of the saddle. Similarly, there is a unique heteroclinic trajectory from the no-animal
node (0,0) to the (1,1) saddle point. It forms one part of the stable manifold of the saddle, the other
part is a trajectory that comes in from the ‘north-eastern’ infinity to the saddle point. These two parts
of the stable manifold are called separatrices. In fact, they partition the phase portrait ( r, s ≥ 0 ) into
two regions which are the basins of attraction of the pure sheep and pure rabbit nodes. In conclusion,
starting from a generic IC, one of the two species eventually dies out leaving a population of the other
species equal to the corresponding carrying capacity. Only the 1-parameter family of ICs that lie on the
stable manifold of the saddle point lead to rabbits and sheep coexisting in the far future.

6.2 Conservative systems

Conservative systems are a rather special but important class of autonomous dynamical systems
ṙ = v(r). In essence, they are systems that admit a conserved quantity: a non-trivial dynamical
variable that is conserved along the flow (constant on trajectories). Though we will focus on 2d
systems, many of the features we discuss generalize to higher dimensions. Newtonian mechanical
systems provide interesting examples. Consider Newton’s equation for a particle moving in 1d
subject to the potential V (x): mẍ + V ′ (x) = 0. Multiplying by the integrating factor ẋ we find
that the energy E = 12 mẋ2 + V (x) is conserved.
More generally, we say that a system ṙ = v(r) is conservative if it admits a continuous
real-valued function E(r) that is not constant on any open subset of phase space and which
satisfies Ė = 0. A function that is constant on phase space is always trivially conserved, so
we must exclude it to ensure that our definition does not classify every dynamical system as
conservative!
A simple consequence of the definition is that a conservative system cannot have an attracting
or repelling fixed point (such as a node or a spiral). A fixed point r∗ is attracting if all trajectories
r(t) starting from points in a sufficiently small neighborhood approach r∗ as t → ∞. Replacing
the condition with t → −∞ defines a repeller. Now, if a conservative system had an attracting
fixed point r∗ , then the conserved quantity would be constant throughout its basin of attraction.
Since the basin of attraction includes an open set containing r∗ , the conserved quantity would
be inadmissible. A similar argument applies to repelling fixed points. In the same vein, one
may argue that conservative systems cannot admit limit cycles: a limit cycle would imply that
the conserved quantity is constant on a narrow ribbon-like open neighborhood of the isolated
periodic trajectory. Though spirals and nodes are forbidden, conservative systems can display
centers and saddles as well as lines of fixed points.
• Double-well potential: The motion of a particle of mass m in a double-well potential
V (x) = (g/4)(x2 − a2 )2 with g, a > 0 is governed by Newton’s equation mẍ = −V ′ (x) =
ga2 x − gx3 . Putting p = mẋ, we get the first order system
ẋ = p/m and ṗ = ga2 x − gx3 . (58)
Find the conserved energy, fixed points, linearization around them, their nature and sketch the
phase portrait and indicate any homoclinic orbits.

29
 1.0

0.5

0.0

-0.5

-1.0

-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5

Figure 7: Phase portrait of particle in a double well potential for m = a = g = 1 .

Solution: The double well potential furnishes an example of a conservative system with saddles and
centers. A graph of V shows that there is a potential barrier of height ga4 /4 separating two potential
wells. There are three fixed points (x∗ , p∗ ) = (0, 0), (a, 0) and (−a, 0) . Linearization around these fixed
points leads to the Jacobian matrices
   
0 1/m 0 1/m
J(0,0) = and J(±a,0) = . (59)
ga2 0 −2ga2 0

While the trace τ = 0 in all cases, the determinant ∆(0,0) = −ga2 /m < 0 while ∆(±a,0) = 2ga2 /m > 0 .
Thus (0, 0) is a linear saddle while (±a, 0) are linear centers. While the Hartman-Grobman theorem
guarantees that the phase portrait in the neighborhood of the origin is homeomorphic to a saddle,
it is non-committal on whether the linear centers (±a, 0) survive in the nonlinear theory. It turns
out that (±a, 0) are nonlinear centers: this is a consequence of energy conservation. Since the energy
E = p2 /2m + (g/4)(x2 − a2 )2 is conserved, trajectories must lie along level curves of E . Draw the
level curves of E and thereby obtain the phase portrait for the double well potential. Sketch a graph of
the energy function over the x - p phase plane showing that it has a bowl-like shape near (±a, 0) . The
energy level contours are the curves along which horizontal planes intersect the graph. One finds that
for E < ga4 /4 , the level contours of energy are a pair of closed curves around (±a, 0) , so the latter are
indeed centers. Physically, they correspond to oscillatory periodic motion around the minima x = ±a of
the potential. As E → 0 , the level contours approach a pair of ellipses as is seen by Taylor expanding
around (x = ±a, p = 0) :
p2
E= + ga2 (x ∓ a)2 + O(x ∓ a)3 . (60)
2m
The E = ga4 /4 level contour is a figure-8 shaped separatrix that is the union of two homoclinic orbits
and the saddle point (0, 0) . The homoclinic orbits correspond to trajectories where a particle approaches
x = 0 as t → ±∞ from either of the two wells. For E > ga4 /4 , the level contours of E are ‘large’ closed
curves corresponding to periodic motion that explores both wells. The period of oscillation diverges as
E → ga4 /4 either from above or below. The figure-8 shaped separatrix partitions the phase plane into
regions where motion is confined/bound to one or other potential well and the region where the particle
is unbound/deconfined.
• Nonlinear centers in conservative systems: The double-well potential illustrates a gen-
eral feature of 2d conservative systems ṙ = v(r). Suppose r∗ is an isolated fixed point of v that
lies at a minimum or maximum of the conserved quantity E(r). Then r∗ is a (nonlinear) center.
In other words, all trajectories that start sufficiently close to r∗ are closed and the local phase
portrait is homeomorphic to that of a linear center. The reason is that the graph of E(r) over a
sufficiently small neighborhood of r∗ must look like a bowl with a unique minimum/maximum

30
 1.5

1.0

0.5

0.0

0.0 0.5 1.0 1.5

Figure 8: Phase portrait of Lotka-Volterra prey-predator model for b = p = d = r = 1 . Horizontal and vertical
axes are prey and predator populations x and y .

at r∗ , whence the energy level contours in the neighborhood must be closed contours. Each
such closed level contour must be a trajectory (rather than a union of more than one trajectory)
since there cannot be any fixed point at which the trajectory can stop (on account of r∗ being
isolated).

6.2.1 Lotka-Volterra prey-predator model

In the model introduced by Lotka (1925) and Volterra (1926), x(t) ≥ 0 and y(t) ≥ 0 are the
populations of the prey and predator. The per capita growth rates are assumed linear
ẋ = x(b − py) and ẏ = y(−d + rx) for b, p, d, r > 0. (61)
b and d are the per capita birth and death rates of the prey and predator in the absence of
interactions. py and rx are the per capita predation and growth rates of prey and predator due
to interspecies interactions. Similar equations are also used to model epidemics, combustion etc.
The system has two fixed points corresponding to extinction (x∗ = 0, y∗ = 0) and coexistence
(x∗ = d/r, y∗ = b/p). We verify that the Lotka-Volterra system admits a conserved quantity
C = d log x + b log y − rx − py. (62)
The gradient
(Cx , Cy ) = (d/x − r, b/y − p) (63)
vanishes precisely at the above two fixed points. The Hessian of C is the diagonal matrix
diag(−d/x2 , −b/y 2 ) and evaluates to diag(−r2 /d, −p2 /b) at the coexistence fixed point. The
latter is therefore a local maximum of C and is therefore a center. Trajectories lie along level
contours of C in the 1st quadrant of the x-y (prey-predator) plane (see Fig. 8). These level
contours are closed curves encircling the center at the coexistence fixed point. Thus, generically,
the populations are periodic and oscillate in ‘cycles’ around the coexistence values.

6.3 Hamiltonian systems

• Hamiltonian systems are a physically important class of conservative systems. The simplest
possibility is a Hamiltonian system on the phase plane with canonical coordinates x and p,

31
usually representing position and momentum of a particle. The system is defined by a distin-
guished real-valued (and sufficiently smooth) function H(x, p) called the Hamiltonian, that is
not constant on any open subset of the phase space. The Hamiltonian generates the flow via
Hamilton’s equations
∂H ∂H
ẋ = and ṗ = − . (64)
∂p ∂x
The difference in sign between the two equations and the switch between x and p are essential
and distinguish Hamiltonian flows from gradient flows (ẋ, ẏ) = −(∂x f, ∂y f ). Unlike gradient
flows where f is non-increasing along trajectories, H is conserved in Hamiltonian systems.
Indeed,
∂H ∂H ∂H ∂H ∂H ∂H
Ḣ = ẋ + ṗ = − = 0. (65)
∂x ∂p ∂x ∂p ∂p ∂x
The Hamiltonian vector field VH = (∂p H, −∂x H) is said to be the skew gradient of H . The
transformation of the phase space generated by a Hamiltonian flow is an example of a canonical
transformation (other canonical transformations that are continuously connected to the identity
transformation arise by picking other Hamiltonian functions).
• Somewhat more generally, we may define a Hamiltonian system on a 2d phase space (surface
M that is not necessarily a plane) if we are given a smooth real-valued Hamiltonian function
H and a non-degenerate (invertible) second rank antisymmetric tensor field rab on it. If ξ 1 , ξ 2
are local coordinates on M , then Hamilton’s equations are
∂H ∂H
ξ˙a = rab b = (ω −1 )ab b . (66)
∂ξ ∂ξ
Verify that these equations reduce to Hamilton’s canonical equations on the phase plane if
(ξ 1 , ξ 2 ) = (x, p) and rab = (0, 1| − 1, 0). The skew-symmetric tensor rab is called the Poisson
tensor and its inverse ω = r−1 (ωab rbc = δac ) is called the symplectic form. The Hamiltonian
vector field VHa = rab ∂b H is the symplectic gradient of the Hamiltonian.
• Harmonic and anharmonic oscillators, particle motion in a double well potential, motion of a
simple pendulum or Euler top are some examples of Hamiltonian systems.
• Remark: It is possible to generalize the concept of Hamiltonian dynamical systems to phase
spaces (symplectic manifolds (M, ω)) of any even dimension ≥ 2. The above formula for the
Hamiltonian vector field is unchanged, but one additionally requires that the symplectic 2-form
be closed dω = 0. Even more generally, one may define Hamiltonian flows ( ξ˙a = rab ∂b H ) on
Poisson manifolds (M, r) of any dimension ≥ 2 by requiring that the Poisson tensor rab satisfy
a quadratic relation called the Jacobi identity.

6.3.1 Some general properties of Hamiltonian systems

Lack of nodes, spirals and limit cycles: Since Hamiltonian systems are conservative, it
follows that they do not admit attractive or repulsive fixed points (nodes or spirals) or limit
cycles.
Liouville property: In fact, Hamiltonian systems have a much stronger property, the Liouville
property of preservation of phase space area (or volume in higher dimensions) under the flow.
In the simplest case of canonical coordinates, the area element on the phase plane is defined as
dx ∧ dp. Liouville’s theorem says that this area element is preserved by the flow, so that the

32
area of any region is unchanged as it moves under the flow. To see that dx ∧ dp is preserved,
we consider the infinitesimal canonical transformation generated by a Hamiltonian flow over a
small time ǫ treated to linear order:

x′ = x + ǫ∂p H + O(ǫ2 ), and p′ = p − ǫ∂x H + O(ǫ2 ). (67)

From multivariable calculus, the transformed area element is given by

∂ x x′ ∂ p x′
   
′ ′ 1 + ǫHpx ǫHpp
dx ∧ dp = (det J)dx ∧ dp where J = ≈ (68)
∂ x p′ ∂ p p′ −ǫHxx 1 − ǫHxp

and Hxp = ∂p ∂x H etc. On account of the equality of mixed partials, det J ≈ 1 + ǫ2 (Hxx Hpp −
2 ). Thus, to order ǫ, there is no change in the area element under Hamiltonian evolution.
Hxp
• Remark: On a symplectic surface, the area element is given by the symplectic two-form
ω = 12 ωab dξ a ∧ dξ b . One can show that ω is Lie-dragged by the Hamiltonian vector field
LVH ω = 0. More generally, on a 2n-dimensional symplectic manifold, one defines the volume
element as the nth exterior power ω n of the symplectic form. One shows that LVH ω = 0 which
implies (by the Leibniz rule) that LVH ω n = 0, so that phase space volume is preserved by a
Hamiltonian flow.
• The Liouville property rules out nodes, spirals and limit cycles in Hamiltonian systems: the
flow around an attractive or repulsive fixed point or limit cycle cannot preserve phase space
volume as the entire basin of attraction/repulsion (with non-zero area) is mapped by the flow
(as t → ∞ or −∞) to a single point or a closed curve (having zero area or volume).
• While spirals and nodes are ruled out, planar Hamiltonian systems can have saddles and
centers, as the dynamics in a double well potential shows. What is the nature of the linearization
of the Hamiltonian vector field around a fixed point?

6.3.2 Poincaré recurrence

Recurrence is a generalization of periodic behavior where one gives up exact return to the initial
state for approximate return to the initial state to any prescribed accuracy. We ask for conditions
that will guarantee that a trajectory will eventually return to a prescribed neighborhood of an
initial phase point and ask how many times it will return to that neighborhood. Let us begin
with examples that illustrate some of the possibilities.
• E.g. 1: Linear harmonic oscillator: Here H = (1/2)(p2 + ω 2 x2 ) and we have a center
at the origin. Every non-constant trajectory is periodic with period 2π/ω . Thus the system is
recurrent. We notice that this happens even though the phase space has infinite volume.
• E.g. 2: The double well potential: corresponds to H = p2 + g(x2 − a2 )2 and has a
saddle at the origin along with a pair of centers at (±a, 0). All trajectories other than the
two homoclinic ones are periodic. However, the period varies with energy and diverges as we
approach the separatrix at E = ga4 from above or below. Thus all trajectories other than
those that begin on the homoclinic orbits return to the initial state, and the system is recurrent,
though with a recurrence time that varies with location on phase space.
• E.g. 3: Irrational windings on a torus: Not all Hamiltonian systems need have fixed
points or periodic trajectories. Consider flow on a 2-torus with 2π -periodic coordinates θ1 and

33
θ2 . We can view the phase space as the square 0 ≤ θ1,2 ≤ 2π with the opposite edges identified.
The evolution is defined by the the equations

θ̇1 = ω1 and θ̇2 = ω2 where (ω1 , ω2 ) 6= (0, 0). (69)

Since the frequencies are not both zero, the system does not possess any fixed points. The
trajectories are straight lines on the square:

θ1 = ω1 t + θ1 (0) and θ2 = ω2 t + θ2 (0). (70)

When a trajectory reaches an edge of the square (say at (θ1 = 2π, θ2 )), it continues from the
opposite edge (say (θ1 = 0, θ2 )) with the same slope and direction. So a trajectory on the phase
square typically looks like a collection of oblique parallel lines with slope ω2 /ω1 . It is clear
that the evolution preserves the area on the phase space measured with respect to the element
dθ1 ∧ dθ2 . In fact, if we postulate that θ1 and θ2 are canonically conjugate {θ1 , θ2 } = 1 we may
derive the equations of motion (locally10 ) from the linear Hamiltonian H = −ω2 θ1 + ω1 θ2

θ̇1 = {θ1 , H} = ω1 and θ̇2 = {θ2 , H} = ω2 . (71)

Now both θ1 and θ2 evolve periodically in time, with periods T1 = 2π/ω1 and T2 = 2π/ω2 .
However, the superposition of these two periodic motions need not be periodic. If T1 = 2T2 then
the phase trajectory is periodic with period T1 . More generally, if T1 and T2 are commensurate
(i.e., if there exist integers m, n 6= 0 such that mT1 + nT2 = 0), then the periods have a ‘least
common multiple’ (LCM) and the trajectory is periodic (it is a closed curve on the torus). On
the other hand, if they are incommensurate (T2 /T1 or T1 /T2 is irrational), then the trajectory
is not closed: it is as if the LCM has gone to infinity, the trajectory is aperiodic and keeps
winding around the torus without ever returning to the initial state. In fact, one can show that
for incommensurate frequencies, every trajectory is aperiodic and fills up the torus, in the sense
that it comes arbitrarily close to any given point on the torus. Thus, irrational windings on
a torus provide an example of a Hamiltonian system that displays recurrent motions without
admitting any periodic trajectories.
• E.g. 3: Free particle and Kepler problem. Not all Hamiltonian systems display recur-
rence. In fact, free particle motion on a line is not recurrent in general. If it has non-zero initial
momentum, the particle goes off to infinity without ever returning to a given neighborhood of
the initial location: x(t) = x(0) + p(0)t/m → ±∞ depending on the sign of p(0). Similarly,
dynamics in the Kepler problem H = (1/2)p2 − GM m/|r| is not recurrent. Trajectories with
H < 0 are periodic, corresponding to elliptical orbits in configuration (position) space while
trajectories with H > 0 fly off to infinity along hyperbolic orbits. These two examples indicate
that having a phase space of infinite volume without any force to contain the motion to a region
of finite volume could come in the way of recurrence.
The Poincaré recurrence theorem applies to a Hamiltonian system with a phase space of
any dimension but having a finite phase volume (measured with respect to the Liouville volume
element)11 . The theorem says that given any open set N0 of initial conditions of positive volume
10
Note that this Hamiltonian is not globally well-defined on the torus: it is multi-valued.
11
These are simple sufficient conditions, they can sometimes be relaxed without losing recurrence, as the
examples of the harmonic oscillator and double-well potential indicate. The recurrence theorem was conjectured
by H. Poincaré in 1890 and proved by C. Carathéodory in 1919.

34
 6
5
4
3
2
1

1 2 3 4 5 6
√
Figure 9: Left: Irrational windings on a torus: θ = ω1 t and φ = ω2 t for 0 ≤ t ≤ 100 with ω1 = 1, ω2 = 2 .
The embedding used is x = R + r cos θ cos φ , y = R + r cos θ sin φ and z = r sin θ with r = 1 and R = 3 . Right:
(θ(t), φ(t)) plotted modulo 2π on a square with periodic boundary conditions.

in phase space, almost every trajectory that begins in it (with the possible exception of those
beginning from a subset of zero volume) returns to N0 , and does so infinitely often.
Sketch of the proof: The proof exploits the Liouville property of preservation of phase volume
as well as the assumption of finite phase volume. Suppose we consider ICs lying in an open set
N0 (say an open ball) with necessarily non-zero volume. After a time t, Hamiltonian evolution
takes N0 to a region Nt with the same volume. Now, the family of phase trajectories starting
in N0 trace out a tubular region in phase space whose volume must be an increasing (non-
decreasing) function of time. Step 1: One possibility is that Nt always intersects N0 in a set
of non-zero volume ∩t≥0 Nt . In this case, all trajectories in this intersection remain within it
and the assertion of the theorem is true for this set, and we may proceed to Step 2. On the
other hand, suppose there is a first time at which the volume of Nt ∩ N0 drops to zero. In this
case, we will argue that there must be a later time t1 after which Nt and N0 intersect in a
region of non-zero volume. This is because the volume of the phase tube is increasing but cannot
exceed the total volume of the phase space. Thus, there must come a time when the phase tube
‘intersects itself’. The first such intersection must be of Nt1 with N0 rather than with any
Nt′ >0 since all points of Nt′ evolved back in time start from N0 . Thus, after a time t1 , Nt1
and N0 must intersect in a set of non-zero volume. Thus, trajectories starting from a non-zero
fraction of the volume of N0 return to N0 after t1 . Step 2: We now apply the same argument
to the remaining portion of N0 (i.e., interior of N0 \ Nt1 ) to conclude that there is a time t2
after which trajectories starting from a non-zero fraction of the volume of N0 \ Nt1 return to it.
Proceeding in this manner, one argues that except possibly for a subset of zero volume, every
trajectory starting in N0 returns to N0 . Step 3: Now we repeat the whole argument to show
that almost every trajectory that begins in N0 returns to N0 twice and so on, thus establishing
the theorem.
• Remarks: (a) Note that the time at which the trajectory returns to N0 for the second time
need not be twice the time of first return. Moreover, the recurrence times tend to grow as the
volume of N0 is decreased (this is easily seen for irrational windings on a torus). (b) There can
be initial conditions in N0 that never return to N0 : an example is provided by ICs that lie on

35
homoclinic orbits in the double well potential problem. (c) There can be exceptional ICs and
sets N0 for which a trajectory returns to N0 a finite number of times, and then stops doing so.

6.4 Index theory

• The index of a vector field around a closed curve C measures how much the vector field
rotates as one goes around C . In contrast with the linearization, which typically provides local
information around a fixed point, the index can provide global information on the phase portrait.
It is an integer that contains information on the fixed points that may lie inside C . We may
view C as a probe used to extract information about the vector field, just as a Gaussian surface
or Amèprian loop are probes used to determine the charge or current enclosed in electrostatics
and magnetostatics.
• Suppose C is a (counter-clockwise directed) simple closed curve (no self-intersections and no
retracing) that does not pass through any fixed points of the 2d vector field v . Note that C need
not be a trajectory. Then, at each point of C , the angle φ (measured counter-clockwise) that
v makes with the horizontal x-axis is defined. Now φ changes continuously as we traverse the
curve. The index IC is defined as the net change in φ (in units of 2π ) as we go counter-clockwise
round C once.
• More precisely, suppose v = (f, g) and let C be the curve r(s) = (x(s), y(s)) param-
eterized by 0 ≤ s ≤ 2π . Then the angle v(r(s)) makes with the horizontal is φ(s) =
arctan(g(r(s))/f (r(s))). The infinitesimal change in φ as s is incremented by ds is

dφ
df
f dg − g ds
dφ = ds = ds2 ds. (72)
ds f + g2
Here f and g depend on s through x and y , so with subscripts denoting partial derivatives,
dg df
= gx x′ (s) + gy y ′ (s) and = fx x′ (s) + fy y ′ (s). (73)
ds ds
Note that x′ (s) and y ′ (s) are not to be confused with the velocities ẋ(t) and ẏ(t) which are
defined only along a trajectory. Thus, the change in φ upon going around the contour C is
Z 2π
1 1 f gs − gfs
I
IC = dφ = ds. (74)
2π 2π 0 f 2 + g2
Since v returns to its original direction after one trip around C , the change in φ must be
a multiple of 2π , whence Ic must be an integer. Ic counts the number of counter-clockwise
revolutions that the vector field makes during one circuit around C .
• The index only depends on the vector field on C and not inside or outside C . We may think
of the above formula as an analogue of the LHS in Gauss’ law in electrostatics
Qencl
Z
E · n̂ dS = . (75)
closed surface ǫ0
We wish to find the analogue of the RHS. What plays the role of charge enclosed? Let us now
compute the index in some examples to gain some intuition.
• In simple cases, one can find the index pictorially by following how the vector field rotates
as one traverses C once counterclockwise. One simply draws arrows on C at a few reasonably

36
spaced points showing the direction of v and counts how many full turns it makes. Try this out
for the examples that follow.
• E.g. 1: Constant vector field: A constant vector field, say, v = (f = 1, g = 0) doesn’t have
any fixed points. fs and gs both vanish, so the index of any closed curve vanishes. We could
have obtained this result pictorially by drawing the vector field along the curve and observing
that v does not change direction as we go around C . It is plausible that the index continues
to vanish even if the vector field is not constant but fluctuates a little bit in direction without
‘turning around’.
• E.g. 2: C encloses a node: The linear system ẋ = ax and ẏ = ay has a symmetrical
node at the origin for any a 6= 0. We take C to be the unit circle centered at the origin and
traversed counter-clockwise (x = cos s, y = sin s) for 0 ≤ s ≤ 2π . Thus

gs = gy ys = a cos s = ax and fs = fx xs = −a sin s = −ay. (76)

Thus
2π
dφ (f gs − gfs ) xx − y(−y) 1
Z
= 2 2
= =1 and IC = ds = 1. (77)
ds f +g x2 + y 2 2π 0
Notice that IC = 1 both for positive and negative a (source and sink inside C ).
• Argue pictorially that IC = 1 for a vector field that points roughly radially outwards or
inwards along a closed curve that is roughly a circle enclosing the origin.
• E.g. 3: C encloses a saddle point: Suppose ẋ = f = x and ẏ = g = −y and C is the
counter-clockwise unit circle (x, y) = (cos s, sin s) enclosing the saddle at the origin. Then it is
easily checked that
dφ −xys + yxs
= = −1 ⇒ IC = −1. (78)
ds x2 + y 2

• E.g. 4: C is a circular trajectory around a center: Consider the linear oscillator

with m = k = ω = 1: ẋ = f = p and ṗ = g = −x, whose trajectories are clockwise
directed concentric circles around the center at the origin. Choose C to be the unit circle
(x, y) = (cos s, sin s) (we could even take s = −t). Proceeding as before, we find IC = 1.
• E.g. 5: C is a circular contour around a spiral: Consider the under-damped oscillator

ẋ = p/m and ṗ = −kx − (γ/m)p with γ < 2mω, (79)

which has a spiral sink at x = p = 0. Show that IC = 1 for C a unit circle. Check that the
index does not change if we make the spiral sink into a source. This can be achieved by rotating
the vector field at each point by π , however, this will not change the net change in angle φ.

6.4.1 Properties of the index

• The above examples have prepared us for the following properties of the index.
• If the closed contour C ′ may be obtained from C by continuous deformation without
passing through any fixed points, then IC = IC ′ . To see this, we first note that the index IC
must vary continuously as C is deformed. Since IC is a continuous integer-valued function of
C , it must be a constant.

37
• The index of a closed curve that does not enclose any fixed points is zero. This is
because we may continuously shrink C without changing IC and without encountering any
fixed points till C is just a point and IC = 0. Alternatively, once C is sufficiently small, the
vector field is approximately a constant field over it, so IC = 0 as in Example 1.
• The index IC is additive under sub-division of the contour. Suppose C is a circle and
we draw a diameter (that does not pass through a fixed point), thereby defining two D-shaped
simple closed contours C1 and C2 whose concatenation is C . C1 and C2 traverse the common
diameter in opposite directions so that it does not contribute to the index. Thus we see that
IC = IC1 + IC2 . This holds more generally when C is the concatenation of C1 and C2
• The index IC is unchanged if we reverse the direction of the vector field everywhere, i.e.,
v → −v or (f, g) → (−f, −g) or t → −t. It is clear that IC does not change since dφ =
(f gg − gfs )/(f 2 + g 2 ) is unchanged. This explains why the index of the curve in E.g. 2 was
independent of the sign of a (i.e., whether C enclosed a stable or unstable node).
• More generally, if we rotate the vector field at every point of the phase plane by the same
angle φ0 , then all the angles φ are augmented by φ0 so that the index of any given curve is
unchanged.
• If C is a closed orbit, then IC = 1. We found that this is the case for a trajectory of the
linear harmonic oscillator. A sketch makes this plausible for any closed trajectory: the vector
field rotates once counter-clockwise as we follow it once counter-clockwise around a trajectory.
Note that by convention, we take C to run counter-clockwise irrespective of the sense of the
trajectory.
• Index of a point: We may use the invariance of IC under continuous deformation of C to
define the index of an isolated fixed point or regular point. Suppose r∗ is an isolated fixed point,
then we define its index Ir∗ to be IC for any counter-clockwise closed curve C that encloses r∗
and no other fixed point. On the other hand, if r is a regular point, i.e., not a fixed point and
with no fixed point in any sufficiently small neighborhood, then we define its index Ir to be IC
for any counterclockwise curve C that encloses r and no fixed point. It is clear that the index
of a regular point vanishes.
• Since the local phase portrait of a nonlinear node/spiral or saddle is homeomorphic to that of
a linear node/spiral or saddle, our examples show that Inode = Ispiral = +1 while Isaddle = −1.
Unfortunately, the index cannot be used to distinguish between a node and a spiral. On the
trace-determinant (τ − ∆) plane classifying linear systems, we observe that the index is +1 for
∆ > 0 and −1 for ∆ < 0. Evidently, the index is not directly related to stability but changes
discontinuously as we cross the τ -axis. The notion of the index of a fixed point allows us to
arrive at a simple expression for the index of a closed curve.
• Index of C as the sum of indices of enclosed fixed points. Suppose C is a simple
closed curve that encloses n isolated fixed points r1∗ , r2∗ , . . . , rn∗ with indices Ir1∗ , . . . , Irn∗ .
Then the index of C is the sum of the indices of the enclosed fixed points. Exercise: Sketch
the proof.
Answer: The idea of the proof is to continuously deform C (with out changing IC ) so that it consists
of small nearly closed curves Cj surrounding each of the fixed points rj and narrow ‘two-way’ paths
connecting adjacent Cj s. Draw a picture. In the limit of infinitesimally narrow bridges, their contribution
to IC cancels out when accounting for the upstream and downstream portions, while ICj → Irj∗ . Thus

38
we obtain
IC = Ir1∗ + Ir2∗ + · · · + Irn∗ . (80)

• Corollary: An immediate consequence is that a closed trajectory must enclose at least one
fixed point, since the index of a closed trajectory is 1. For instance, a closed trajectory can
enclose a single spiral or node but cannot enclose a single saddle or two nodes and no other fixed
points. In favorable cases, we may use this result along with a knowledge of the locations of fixed
points and their indices to rule out closed trajectories. Show that the competitive Lotka-Volterra
model cannot have any periodic solutions.
• There are examples of fixed points with any integer index. For example, the fixed
point at the origin for the dipole field ẋ = x2 − y 2 , ẏ = 2xy has index 2. Show this by sketching
the vector field and choosing C to be a unit circle centered at the origin.
• A fixed point can have index zero. An example is furnished by the fixed point at the
origin for the vector field ẋ = x2 and ẏ = y 2 . The vector field fluctuates between pointing to
the north and to the east without ever rotating. One checks that the index of any closed curve
enclosing the origin vanishes. Thus, regular points are not the only ones with index zero.
• Index as degree: The index can be regarded as the degree of a map between circles.
Consider a planar vector filed. Let C(s) : [0, 2π] → R2 be a simple closed curve as before, it
is homeomorphic to a circle. At each point of C consider the normalized vector field n̂(s) =
v(s)/|v(s)|, which is a unit vector, so it lies on a circle. As we go round C once we follow the
movement of n̂(s) around the unit circle: it must return to its starting point as n̂(0) = n̂(2π).
Thus we get a map from S 1 → S 1 taking s 7→ n̂(s). This map could be many-to-one and the
number of times the curve C covers the target circle is called its degree. If the vector field is
a constant, then n̂(s) ≡ n̂(0) does not move and the degree is zero. If n̂ goes once round the
circle counterclockwise (or clockwise) we say that the map has degree +1 (or -1). Note that
n̂(s) can go back and forth, it may retrace its path sometimes. More generally, the map may
be a double cover, triple cover or n-fold cover. The index, defined as the degree of this map,
allows us to generalize the index to vector fields on curved surfaces. For a surface M embedded
in 3d Euclidean space, n̂ must lie on a 2-sphere and trace out a curve which is an n-fold cover
of a circle.
• Index of vector fields on Riemann surfaces: Consider a smooth vector field on a compact
(closed, bounded) surface without boundaries. Examples include the sphere, torus and higher
handlebodies. These ‘Riemann’ surfaces are classified by their genus, a non-negative integer
which is the number of handles: g = 0 for the sphere, g = 1 for the torus, g = 2 for the
double torus etc. The Poincaré-Hopf index theorem states that for a smooth vector field on a
surface of genus g , the sum of indices of the fixed points of a vector field is equal to the Euler
characteristic χ = 2 − 2g . In particular, every such vector field on a sphere (χ = 2) has to have
at least one fixed point: there is no nonvanishing vector field on the sphere – this is sometimes
called the hairy ball theorem and says that one cannot consistently comb hair (of non-zero
length) tangentially everywhere on a sphere. On the other hand, one can imagine a vector field
on the sphere whose integral curves point longitudinally, emerging from a node at the north pole
and ending up at a node at the south pole (the poles each have index one). Another example is
a vector field with integral curves that are latitudes: it has two fixed points, centers at the poles.
On a torus, χ = 0, so the torus admits vector fields that are nowhere vanishing; for instance
one that points azimuthally everywhere.

39
 4

-2

-4
-4 -2 0 2 4

Figure 10: Limit cycle of the van der Pol oscillator for m = k = 1 and µ = 1/5 . Notice the unstable spiral at
the origin.

6.5 Limit cycles

As noted earlier, a limit cycle is an isolated periodic orbit. Though a limit cycle can exist in a
phase space of any dimension d ≥ 2, we will confine our discussion to two dimensions. In the
case of a plane, the limit cycle partitions the phase space into an inside and an outside. Since a
limit cycle is isolated, trajectories that begin in a sufficiently small tubular neighborhood cannot
be periodic and must spiral either toward or away from the limit cycle. This leads to stable,
unstable and half-stable limit cycles.
Linear systems do not admit limit cycles, since a slightly rescaled periodic orbit (λr(t) for
λ ≈ 1) must also be a periodic orbit of the linear system, so that r(t) cannot be isolated.
Stable limit cycles, when they exist, represent dynamically generated (self-sustained, i.e.,
without external periodic driving) oscillations of a nonlinear system with a dynamically deter-
mined asymptotic (t → ∞) period and waveform, including amplitude. By contrast, periodic
trajectories of a linear system can have any amplitude by a suitable choice of ICs, due to the
freedom to rescale a periodic trajectory. There are several examples of periodic motions that
may be modeled as limit cycles: airplane wing flutter, vibrations in bridges, oscillations in pop-
ulations in predator-prey systems, the heart beat, daily rhythms in body temperature and sleep,
hormone secretion etc.
E.g. 1: Limit circle in plane polar coordinates: It is easy to cook-up a limit cycle in a
nonlinear system on the plane using polar coordinates. In fact, consider the system

ṙ = r(1 − r) and θ̇ = 1. (81)

The angular motion is uniform counterclockwise rotation and decouples from the radial motion
which has a repulsive fixed point at r = 0 and an attractive one at r = 1. Thus, the circle r = 1
is a stable limit cycle with trajectories spiraling towards it from either side. The limit cycle is
the circular orbit (x, y) = (cos t, sin t).
E.g. 2: van der Pol oscillator: A more non-trivial example of a limit cycle arises in the van
der Pol equation:
mẍ = −kx − µ(x2 − 1)ẋ for m, k, µ > 0. (82)
The vdP equation arose (when van der Pol was working with vacuum tubes at Philips in the

40
 3.5

3.0

2.5

2.0

1.5

1.0
0.0 0.5 1.0 1.5 2.0

Figure 11: Limit cycle of the Brusselator for a = 1 and b = 2.2 > 1 + a2 where the fixed point at (a, b/a) is a
spiral source.

Netherlands) in modeling LCR electrical circuits where the ‘resistor’ had a nonlinear V -I charac-
teristic making it behave like a resistor (sink) for currents of large magnitude and an anti-resistor
(source) for small currents. The system may be viewed as an oscillator with a nonlinear damping
coefficient γ(x) = µ(x2 − 1). By design, for |x| > 1, γ > 0 and the motion is damped while for
|x| < 1, γ < 0 and we have ‘anti-damping’. As a first order system, we have
µ 2
ẋ = p/m, and ṗ = −kx − (x − 1)p. (83)
m
The only fixed point is at the origin (x, p) = (0, 0). Linearization around it gives the Jacobian
J = (0, 1/m| − k, µ/m) with ∆ = k/m > 0 and τ = µ/m > 0. So the origin is a spiral
source if µ2 < 4km and an unstable node if µ2 > 4km. Either way, trajectories move outward
from the origin. On the other hand, for large x and p, the nonlinearity dominates and we
have ṗ ≈ −(µ/m)x2 p and ẋ = p/m, indicating decay of p with x remaining bounded. Thus,
due to the balancing effects of damping and anti-damping, we may expect the vdP equation to
display a stable limit cycle. With somewhat more effort, it can be shown that the van der Pol
equation has a unique stable limit cycle (see Fig. 10). One can find its shape numerically, it is
not circular/elliptical (except when µ → 0) and there is no closed-form expression available for
it.
Eg. 3: Brusselator. The Brusselator is a model for an autocatalytic chemical reaction
introduced by Ilya Prigogine and collaborators in Brussels. The name is made by combining
the words ‘Brussels’ and ‘oscillator’. The Brusselator is described by the pair of differential
equations
ẋ = a − (1 + b)x + x2 y and ẏ = bx − x2 y. (84)
Check that for a 6= 0, it has a unique fixed point at (x∗ = a, y∗ = b/a). The Jacobian for the
linearization around this fixed point is

b − 1 a2
 
J= with τ = b − 1 − a2 and ∆ = a2 . (85)
−b −a2

Thus, the fixed point is linearly unstable if b > 1 + a2 . A limit cycle forms around this unstable
fixed point (see Fig. 11) leading to characteristic oscillations with amplitude independent of
initial conditions.

41
6.5.1 Criteria for non-existence of limit cycles

If we suspect that a system does not admit limit cycles, then it may be possible to rule them
out.
Fixed point criterion: From index theory, we have learned that a closed trajectory has index
+1 and must enclose at least one fixed point. This clearly places a constraint on the occurrence
and location of limit cycles. For instance, a limit cycle cannot encircle just two nodes.
Gradient flows do not admit limit cycles: Consider a gradient flow ṙ = v = −∇W in a
planar region with W a single-valued scalar ‘potential’ function. Suppose r(t) for t ∈ [0, T ] is
a closed trajectory (that is not a ring of fixed points). The change in W as we go once round
the closed trajectory
T T T
dW
Z Z Z
∆W = dt = ∇W · ṙ dt = − |∇W |2 dt (86)
0 dt 0 0

is strictly negative. On the other hand, ∆W must vanish, since W is single-valued. Thus we
arrive at a contradiction: we cannot have periodic trajectories in a gradient flow.
Lyapunov Functions: Even if a vector field is not a gradient vector field, we can still rule out
closed orbits if it admits a ‘Lyapunov’ function on phase space that decreases along the flow.
More precisely, the C 1 real-valued function L(r) is said to be a Lyapunov function for the flow
ṙ = v(r) if (a) L is positive on the phase space (b) L vanishes only at fixed points and (c)
L̇(r) < 0 along the flow as long as r is not a fixed point. The same argument that we used for
gradient flows shows that the existence of a Lyapunov function forbids closed trajectories.
Bendixon’s and Dulac’s criteria: For a closed trajectory to exist, heuristically, the vector
field must go ‘round’. On the other hand, if the vector field is ‘spreading out’ (diverging,
∇ · v > 0) or converging, as in the vicinity of a node, then we cannot have a closed orbit there.
This is called Bendixon’s criterion. A slight generalization is called Dulac’s criterion, which we
now describe. Suppose Ω is a simply-connected region in the plane and there exists a smooth
function f (r) such that ∇ · (f (r)v) is either strictly positive or strictly negative, then v cannot
have any closed orbits in Ω. The proof uses Green’s theorem, which is a planar analogue of
Gauss’/Stokes’ theorem. Indeed, suppose there is a closed trajectory γ lying inside Ω, then
Green’s theorem applied to the vector field f v says that
I ZZ
f v · n̂ dl = ∇ · (f v) dxdy (87)
γ=∂G G

where n̂ is the outward pointing unit normal on γ , dl is a line element on it and G is the region
whose boundary is γ . Since γ is a trajectory, v must be tangent to it, so v · n̂ = 0 and the
LHS vanishes. However, by assumption, the RHS is either strictly positive or negative leading
to a contradiction. Bendixon’s criterion is the special case where f = 1. It says that a vector
field with ∇ · v 6= 0 in a simply-connected region cannot have any closed orbits contained in
that region.
• E.g. We may use Dulac’s criterion to rule out closed orbits in the competitive Lotka-Volterra
model
ṙ = r(3 − r − 2s) and ṡ = s(2 − s − r) for r, s > 1. (88)

42
It is easily checked that this vector field has a divergence that can have either sign in the first
quadrant r, s > 0. However, suppose we pick f = 1/rs, then
     
3 − r − 2s 2−s−r 1 1
∇ · (f v) = ∂r + ∂s =− + < 0. (89)
s r r s

Thus, by Dulac’s criterion, the model does not admit any periodic solutions with a positive
number of sheep and rabbits.

6.5.2 Poincaré-Bendixon Theorem

The celebrated Poincaré-Bendixon theorem gives sufficient conditions for the existence of a closed
orbit in 2d and also places restrictions on what a trajectory can do. In particular, it precludes
chaos in 2d autonomous systems. The statement is as follows.
Poincaré-Bendixon Theorem: Suppose the C 1 vector field ṙ = v(r) is defined in an open
subset D of the phase plane and let K ⊂ D be a compact (closed and bounded) subset that
contains no fixed points. Moreover, suppose γ is a ‘trapped’ trajectory that begins in K (at
some time) and remains in K at all subsequent times. Then γ is either a closed orbit or
approaches (spirals towards) a closed orbit as t → ∞. Since K contains all its limit points,
the limiting closed orbit in the latter case must lie within K . Either way, K contains a closed
trajectory.
• From index theory, we know that a closed orbit must enclose a fixed point. Since the theorem
requires that K be fixed-point free, K must be chosen to exclude the fixed points that lie within
the closed orbit that the theorem guarantees. Thus, in applications of the Poincaré-Bendixon
theorem, K is typically chosen to be shaped like an annulus. In practice it may be tricky to
identify a trajectory that remains in K . A way to ensure this is to pick K to be a trapping
region: a connected compact subset of D on whose boundary the vector field points inward. All
trajectories that begin in a trapping region are then confined to it. If a trapping region does not
contain any fixed points, then the Poincaré-Bendixon theorem guarantees that it has a closed
trajectory.
• By a careful choice of a trapping region, one can use the Poincaré-Bendixon theorem to show
that the van der Pol oscillator has a closed trajectory.
• Example: Apply the Poincaré-Bendixon theorem to show that the vector field

ṙ = r(1 − r) + αr sin θ and θ̇ = 1 (90)

admits a closed trajectory for any sufficiently small α > 0.

ANS: We may choose the trapping region to be the annulus rmin = 1 − α ≤ r ≤ rmax = 1 + α
for any α < 1. We need to ensure that ṙ = r(1 − r) + αr sin θ > 0 for r = rmin and ṙ < 0 for
r = rmax . Use the fact that −1 ≤ sin θ ≤ 1 to check that this is satisfied.
• Absence of chaos on the phase plane. The Poincaré-Bendixon theorem restricts the
possible behavior of trajectories on the phase plane. If a trajectory is confined to a compact
region without fixed points, the trajectory must either be periodic or approach a periodic orbit.
In other words, the trajectory cannot wander around endlessly in a bounded region without

43
approaching a fixed point or a limit cycle. In both these cases, there is no sensitivity to initial
conditions, which is a hallmark of chaos12 .
In three and higher dimensions, phase trajectories confined to a bounded region can be
chaotic: they can wander around endlessly without approaching a fixed point or a periodic
orbit and display extreme sensitivity to initial conditions. In addition to fixed points, periodic
orbits and limit cycles, one encounters ‘strange attractors’ that are fractals: this happens for
instance in Edward Lorenz’s 1963 toy-model for atmospheric convection, which has a 3d phase
space. For certain classes of initial conditions, the motion is chaotic/irregular: sensitive to initial
conditions, remaining bounded and never settling down to periodic motion.

6.6 Bifurcations in two dimensions

We briefly examine some bifurcations of two-dimensional vector fields. In addition to the change
in character of fixed points (as in 1d), here we also have the possibility of limit cycles changing
stability or being created/destroyed.

6.6.1 Saddle-node, transcritical and pitchfork bifurcations

The saddle-node, transcritical and pitchfork bifurcations generalize in a straightforward manner

to 2 and higher dimensions. The flow in the all directions other than one is passive (attractive or
repulsive) and the change in character of the vector field is essentially confined to one direction
(say x). Thus, normal forms can be chosen to involve decoupled pairs of differential equations
for x and y .
• A saddle-node bifurcation is exemplified by the normal form

ẋ = a − x2 and ẏ = −y. (91)

as the control parameter a is varied, this vector field displays the basic mechanism by which
fixed points can be produced or annihilated in pairs. While y(t) = y(0)e−t approaches zero
exponentially fast, the dynamics in the x direction is as in §3.1. For a > 0, we have a saddle
√ √
at (− a, 0) and a stable node at ( a, 0). As a decreases to zero, the saddle and node coalesce
at a partly stable-partly unstable fixed point at the origin (whose basin of attraction is the
closed right half plane) leaving behind no fixed points when a < 0. This explains the name
saddle-node bifurcation. However, for a . 0, there is a remnant/ghost of the fixed point around
(0, 0). Though the flow is generally downward and leftwards (towards an unstable direction of
the erstwhile saddle), trajectories are slowed down as the pass through a bottleneck around the
origin.
• Transcritical bifurcation. Similarly, the transcritical bifurcation (§3.2) generalizes to 2d.
The canonical form is
ẋ = x(a − x) and ẏ = −y. (92)
This vector field has fixed points at (0, 0) and (a, 0). For a < 0, the origin is a stable node and
(a, 0) is a saddle. As a increases past 0, the fixed point at (a, 0) passes through the origin and
12
In unbounded regions, nearby trajectories can diverge, but then they may fly off to infinity in a regular
fashion, as in the Kepler problem.

44
 Figure 12: Saddle-Node bifurcation in 2d

exchanges stabilities with it. As a result, for a > 0, the origin becomes a saddle while (a, 0)
becomes a stable node.
Remark: In all these examples of bifurcations, a stable fixed point has undergone a change
in stability as a control parameter is varied. In the linear approximation, the Jacobian of the
vector field around a stable fixed point generically has a pair of eigenvalues with negative real
parts. These eigenvalues could both be negative (as in the case of a stable node) or be complex
conjugates (as in the case of a spiral sink); either way they lie in the left half of the complex
eigenvalue plane. The bifurcation is associated with a change in stability, this typically happens
when one or both eigenvalues reach the imaginary axis or cross over to the right half-plane. In
the above examples, we were dealing with real eigenvalues since nodes and saddles were involved
and one of the eigenvalues reached λ = 0 at the bifurcation point. Thus, they are called zero-
eigenvalue bifurcations. The other possibility where a pair of complex conjugate eigenvalues in
the left half plane together cross the imaginary axis is qualitatively different and leads to Hopf
bifurcations which we turn to next.

6.6.2 Hopf bifurcations

Unlike the saddle-node, transcritical and pitchfork bifurcations which involve collisions between
fixed points, Hopf bifurcations do not. A Hopf bifurcation is one way in which a limit cycle
can be created or destroyed. Though Hopf bifurcations are not direct generalizations of 1d
bifurcations, they bear some resemblance to pitchfork bifurcations and can occur for systems
with a state space of dimension two or more.
Supercritical Hopf bifurcation: A supercritical Hopf bifurcation is somewhat analogous to
a supercritical pitchfork bifurcation. Thus, we have a stable fixed point r∗ when the control

45
 Figure 13: Transcritical bifurcation in 2d

2 2

1 1

0 0

-1 -1

-2 -2
-2 -1 0 1 2 -2 -1 0 1 2
Figure 14: Supercritical Hopf bifurcation of (93). Stable spiral for a = −1 < 0 transitions to a stable limit
cycle for a = 1 > 0 . Here b = ω = 1 .

parameter is on one side of a critical value, say a < ac . However, rather than being a stable
node, it is a spiral sink. Thus, for a < ac the system displays damped oscillations before settling
down to the fixed point. On the other hand, for a > ac , the fixed point becomes a spiral source.
However, stability is not entirely lost, since the spiral approaches a stable limit cycle of roughly
elliptical shape that encircles r∗ . What this means in practice is that the exponential decay
to an equilibrium state becomes slower as a → a− c . Beyond ac , the equilibrium state becomes
unstable and the system flutters/oscillates about it with relatively small amplitude. One says
that the supercritical Hopf bifurcation is safe or soft. A prototypical vector field displaying a
supercritical Hopf bifurcation is given by the following system in plane polar coordinates

ṙ = ar − r3 and θ̇ = ω + br2 (93)

For a ≤ 0, r∗ = 0 is a stable spiral. The approach to the origin goes from exponentially fast to
a power law as a → 0− . When a > 0, the origin is a spiral source. However, the unstable spiral

46
growth saturates (due to the stabilizing effect of the −r3 term) at a stable limit cycle given by
√
the circle r = a on which ṙ = 0 and θ̇ = ω + ab. The behavior of eigenvalues of the Jacobian
is most easily examined in Cartesian coordinates x = r cos θ and y = r sin θ where

ẋ = cos θ ṙ − r sin θ θ̇ = x(a − r2 ) − y(ω + br2 ) = ax − ωy + O(x3 , y 3 , · · · ) and

ẏ = sin θ ṙ + r cos θ θ̇ = y(a − r2 ) + x(ω + br2 ) = ωx + ay + O(x3 , y 3 , · · · ). (94)

Thus the linearization around the fixed point at the origin is given by
 
a −ω
J= (95)
ω a

with the eigenvalues λ = a ± iω . Evidently, the eigenvalues cross the imaginary axis from the
left to right half plane as a increases past zero.
Subcritical Hopf bifurcations: are dangerous since after the bifurcation, stable oscillations
are rendered unstable and trajectories jump to some ‘far-away’ attractor (a fixed point, a limit
cycle, infinity or even a chaotic attractor, as in the case of the Lorenz model). This is illustrated
by the vector field
ṙ = ar + r3 − r5 and θ̇ = ω + br2 . (96)
The linearization around the fixed point at the origin,
x
ẋ = ṙ − y θ̇ = x(a + r2 − r4 ) − y(ω + br2 ) ≈ ax − ωy and
r
y
ẏ = ṙ + xθ̇ = y(a + r2 − r4 ) + x(ω + br2 ) ≈ ωx + ay, (97)
r
 
a −ω
leads to the same Jacobian J = as in the above supercritical example. As before, as
ω a
a increases past 0, the eigenvalues a ± iω cross over to the right half plane making the origin
an unstable fixed point. What makes this example different is that the r3 term in (96) is now
destabilizing. As in the subcritical pitchfork example of §3.3.2, this is balanced by a stabilizing
r5 term. The latter will be seen to lead to a large amplitude limit cycle to which trajectories
can flow after the bifurcation.
Indeed, for −1/4 <√a < 0, there is a spiral sink at the origin surrounded by an unstable
limit cycle 2
2
√at r = (1 − 1 + 4a)/2 which is in turn surrounded by a larger stable limit cycle at
r = (1+ 1 + 4a)/2 (see Fig. 15). So for −1/4 < a < 0 the system displays decaying oscillations
for small amplitude initial conditions. As a → 0− , the unstable limit cycle shrinks to the origin
turning it into an unstable spiral. The decaying oscillations morph into growing oscillations at
the subcritical Hopf bifurcation at a = 0. In this √
example, these growing trajectories are then
2
attracted to the far-away limit cycle at r = (1 + 1 + 4a)/2.
As in the one-dimensional example of §3.3.2, this system displays hysteresis. When a is
decreased from positive values, the large amplitude oscillations persist all the way down to
a = −1/4. At this point, the stable and unstable limit cycles collide and annihilate in a saddle-
node bifurcation of cycles and we are left with a spiral sink at the origin for a < −1/4.
In addition to Hopf bifurcations, there are other ways in which limit cycles can be created
or destroyed in 2d. These are classified as global bifurcations as they involve large regions of
the phase plane. (a) We already encountered the saddle-node or fold bifurcation of cycles
at a = −1/4 in the example of (96). Here, a stable and unstable limit cycle that exist for

47
 1.5 1.5

1.0 1.0

0.5 0.5

0.0 0.0

-0.5 -0.5

-1.0 -1.0

-1.5 -1.5

-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5

Figure 15: Subcritical Hopf bifurcation of (96) at a = 0 . Left: stable spiral surrounded by an unstable circular
limit cycle and a larger stable limit cycle for a = −1/5 . Right: Unstable spiral with trajectories attracted to a
large amplitude limit cycle for a = 1/2 . Here b = ω = 1 .

a > −1/4 annihilate and cease to exist for a < −1/4. At a = −1/4 they coalesce to form a
half-stable cycle. (b) Saddle-node infinite period bifurcation: Here, a limit cycle develops
a bottleneck and a half-stable fixed point is born on the cycle at the bifurcation point, leading
to a divergence of the period. Subsequently, the fixed point splits into a pair of fixed points
on the erstwhile cycle in a saddle-node bifurcation. Show that a simple example demonstrating
this phenomenon is ṙ = r(1 − r2 ) and θ̇ = a − sin θ at a = 1. (c) Homoclinic or saddle-loop
bifurcation: In this case, a limit cycle moves close to a saddle point and touches it when the
control parameter reaches its bifurcation point value. As should be plausible from our earlier
examples (see §4), this results in a divergence of the period of the limit cycle and the cycle turns
into a homoclinic orbit.

7 Lorenz’s 3d model of convection and chaos

The Lorenz equations ṙ = v(r) are a system of three ODEs

ẋ = σ(y − x), ẏ = ρx − y − xz and ż = xy − βz. (98)

They were introduced by Edward Norton Lorenz (J. Atmos. Sci. 20, 130 (1963)) as a simplified
model for atmospheric Rayleigh-Bénard convection cells in a shallow 2d fluid layer warmed from
below and cooled from above. He viewed it as a toy system to study the weather and obtained
it by reducing a larger system of 12 ODEs used in studying convection. The modes x and y
are measures of the velocity and temperature convection rates in the convective roll patterns
while z is proportional to the vertical convective heat transfer. The three positive parameters
ρ, σ and β are constants proportional to the Rayleigh number, Prandtl number (fluid dynamic
parameters) and a geometric aspect ratio. Of the three, the Rayleigh number ρ is the main
control parameter while σ and β may be taken fixed under typical conditions. In Lorenz’s

48
original work he took ρ = 28, σ = 10 and β = 8/3. In addition to convection, the Lorenz
equations can also be used to model (geo)dynamos, lasers and water wheels.
• In these variables, the Lorenz equations are quadratically nonlinear and possess the discrete
reflection (clockwise-counterclockwise) symmetry x → −x, y → −y and z → z .
• If we were to ignore all interaction terms, then the linear system ẋ = −σx, ẏ = −y and
ż = −bz would imply that x, y and z each decays monotonically and exponentially to zero.
The full system is much richer.
• Fixed points: For ρ ≤ 1 the trivial fixed point x∗ = y∗ = z∗ = 0 is the only fixed point
while for ρ > 1 two additional nontrivial fixed points emerge. In fact, the conditions for the
Lorenz vector field to vanish are

ẋ = 0 ⇒ x = y, ẏ = 0 ⇒ x(ρ − 1 − z) = 0 and ż = 0 ⇒ x2 − βz = 0. (99)

The second equation implies either x = 0 or z = ρ − 1. If x = 0, then x∗ = y∗ = z∗ = 0 is

the only possibility. If z = ρ − 1 then x2 = β(ρ − 1), which has nontrivial real solutions only if
ρ > 1. Thus, for ρ ≤ 1, r∗ = 0 is the only fixed point. What is more, for ρ < 1, we will use a
Lyapunov function to show that r∗ = 0 is a global attractor so that the dynamics is relatively
benign for ρ < 1.
• On the other hand, for ρ > 1 we have, in addition to the trivial fixed point, two nontrivial
fixed points p
C± = (x∗ = ± β(ρ − 1), y∗ = x∗ , z∗ = ρ − 1). (100)
Notice that C± merge with the trivial fixed point as ρ → 1+ . The trivial fixed point at the origin
corresponds to the absence of convection and no horizontal or vertical temperature variation.
The nontrivial fixed points C± correspond to steady clockwise/counter-clockwise convection,
which explains the symbol C . As we might expect, the system displays a supercritical pitchfork
bifurcation at ρ = 1, which is associated with the spontaneous breaking of the above discrete
symmetry. It is supercritical since C± are stable for ρ & 1.
• Stability of trivial fixed point: To examine linear stability of fixed points, we consider the
Jacobian matrix:  
 i −σ σ 0
∂v
J= = ρ − z∗ −1 −x∗  . (101)
∂xj r∗
y∗ x∗ −β
At the trivial fixed point,  
−σ σ 0
J(r∗ = 0) =  ρ −1 0  (102)
0 0 −β
and the linearized flow in the z direction decouples and is stable. The restriction of J to the
x-y plane is a 2 × 2 matrix with trace τ = −(1 + σ) < 0 (implying stability) and determinant
∆ = σ(1 − ρ). Moreover, the discriminant

D = τ 2 − 4∆ = σ 2 + 2σ + 1 − 4σ + 4ρσ = (σ − 1)2 + 4ρσ > 0 (103)

is positive. Thus, if 0 < ρ < 1, ∆ > 0 and the origin is a stable node while if ρ > 1, it
becomes a saddle point since ∆ < 0. This may also be seen from the characteristic equation

49
(λ + β)(λ2 + (σ + 1)λ + σ(1 − ρ)) = 0 whose roots are
1 √ 
λ± = τ± D and λ3 = −β. (104)
2

• Contraction of volumes: The divergence of the vector field v controls how infinitesimal
volumes in the x-y -z phase space change with time. If (δV )(r) is an infinitesimal volume
located at r,
d
(δV )(r) = (∇ · v)(r) = tr J = −(σ + 1 + β). (105)
dt
Since ∇ · v is strictly negative and independent of location, little volumes located anywhere in
phase space will shrink exponentially in time for any positive σ and β . The system is dissipative
in this sense. Thus, any 3d space of initial conditions will eventually shrink (be attracted) to a
set of zero volume. This attractor could be one or more fixed points or a limit cycle or invariant
surface or some less familiar invariant set of points with zero volume. We will see that the
Lorenz system does not admit any invariant surfaces that enclose a finite volume, but that the
other possibilities can and do occur for various values of ρ.
• Origin is a global attractor when ρ < 1: Consider the non-negative function

1 x2
 
φ= + y2 + z2 . (106)
2 σ

We will show that for ρ < 1, φ is a Lyapunov function for the Lorenz system. It is clear that
φ, vanishes only at the trivial fixed point r∗ = 0 and has concentric ellipses as its level surfaces.
It remains to show that φ is monotonically decreasing along the flow, so that every trajectory
must approach r = 0 as t → ∞. To do so, we compute:
xẋ
φ̇ = + y ẏ + z ż = (xy − x2 ) + (ρxy − y 2 − ✘
xyz)
✘ + (✘ ✘ − βz 2 )
xyz
σ
2
= −(x
" − (ρ+ 1)xy+  y 2 + βz 2 )
2  ! #
ρ+1 2

ρ+1 2 2
= − x− y + 1− y + βz , (107)
2 2

where we have completed a square. Now for ρ < 1, ρ + 1 < 2 so that ((ρ + 1)/2)2 < 1, Thus,
for ρ < 1, φ̇ is the negative of the sum of three squares with positive coefficients. It follows
that φ̇ < 0 for ρ < 1. What is more, we observe that for φ̇ to vanish, each of the three terms in
(107) must vanish. This happens only if x = y = z = 0, so that φ is strictly decreasing along
the flow away from the origin. Thus, for ρ < 1, any trajectory that begins away from r = 0
must approach the trivial fixed point as t → ∞. This implies that for ρ < 1, there cannot be
any limit cycles, nontrivial invariant manifolds or chaos.
• What is the nature of the dynamics for ρ > 1?
• Stability of nontrivial fixed points C± : The nontrivial fixed points C± (which exist only
for ρ > 1) can be stable or unstable depending on the values of ρ, σ and β . For the values
chosen by Lorenz (ρ = 28, σ = 10 and β = 8/3), they are unstable: two of the eigenvalues of J
have positive real parts: λ3 ≈ −13.9, λ± ≈ 0.09 ± 10.2i. On the other hand holding σ, β fixed,
for a smaller ρ (e.g. ρ = 18) they are stable λ3 ≈ −13.21, λ± = −.23 ± 8.3i. So we might
suspect a change in stability of C± at a critical value ρH (named since it turns out to be a Hopf

50
bifurcation) at which λ± cross the imaginary axis. To find where this occurs, we note that the
Jacobians at C± ,

−σ σ 0

p
J(C± ) =  p 1 p −1 ∓ β(ρ − 1) . (108)
± β(ρ − 1) ± β(ρ − 1) −β
have the characteristic polynomial
λ3 + (1 + β + σ)λ2 + β(ρ + σ)λ + 2βσ(ρ − 1). (109)
On the other hand, at the transition, the eigenvalues must be of the form λ = ±iω and λ3 for
ω, λ3 ∈ R leading to the characteristic polynomial
(λ − iω)(λ + iω)(λ − λ3 ) = λ3 − λ3 λ2 + ω 2 λ − ω 2 λ3 . (110)
Comparing, we read off λ3 = −(1 + β + σ) [which corresponds to a stable direction], ω 2 =
β(ρH + σ) and ω 2 λ3 = −2βσ(ρH − 1). The latter two equations admit only one solution
 
σ+β+3
ρH = σ . (111)
σ−β−1
Here we assume σ > β +1 so that ρH > 0 [when ρ > 1 and σ < β +1, the nontrivial fixed points
C± are unstable]. For the Lorenz values, the inequality is satisfied and ρH = 470/19 ≈ 24.7368,
which lies between ρ = 18 and ρ = 28 we considered above. Thus, C± are linearly stable
provided 1 < ρ < ρH .
• Subcritical Hopf bifurcation: As ρ → ρH from below, one can show with more effort, that
the stable fixed points C± undergo a subcritical Hopf bifurcation. This means that unstable
limit cycles shrink to C± as ρ → ρ−H and render the nontrivial fixed points unstable for ρ > ρH .
So for ρ > ρH , generic trajectories do not approach C± nor the trivial fixed point and do not
have the above limit cycles to approach either. What can trajectories do? Can they approach
some other limit cycle? Lorenz argued (using a discrete time dynamics ‘Lorenz map’ concocted
from the Lorenz model) that any limit cycles would have to be unstable. Can they approach
some invariant surface, such as a torus (and become quasi-periodic)?
• No invariant surfaces enclosing finite (non-zero) volume in Lorenz system: We may
argue that the Lorenz system cannot have any invariant surfaces that enclose a finite volume
(e.g. a closed and bounded surface such as a sphere or a torus or a Riemann surface of higher
genus). Indeed, suppose there is such an invariant surface Σ, then the 3d region Ω inside Σ
must also be invariant under the flow. In particular, the volume of Ω must be conserved in
time. However, this contradicts the previously established (105) exponential decrease in volume
of any region of the phase space.
• Trajectories cannot escape to infinity. Can trajectories fly off to infinity? It turns out
this too is not possible. This can be shown by finding a trapping surface. For example, consider
the ellipsoids EK defined by the level surfaces F = K of the function
F (x, y, z) = ρx2 + σy 2 + σ(z − 2ρ)2 . (112)
We will show that if K is chosen sufficiently large, then every trajectory that begins on EK re-
mains inside it. Given any initial condition, there will then be such a trapping surface containing
it, establishing that trajectories cannot escape to infinity.

51
• To see this, we will show that the vector field points inwards on EK for K big enough. Now,
the outward normal on EK and the Lorentz vector field are given by

∇F = (2ρx, 2σy, 2σ(z − 2ρ)) and v = (σ(y − x), ρx − y − xz, xy − βz). (113)

For v to point inwards on EK we need v · ∇F < 0. Now,

✘ − ρx2 + ✘✘ − y2 − ✘ ✘✘ − βz 2 + 2ρβz
 
v · ∇F = 2σ ✘ ρxy ρxy xyz
✘+✘ xyz
✘−✘2ρxy
= −2σ ρx2 + y 2 + β(z − ρ)2 − βρ2 .
 
(114)

For this to be negative we need

ρx2 + y 2 + β(z − ρ)2 > βρ2 . (115)

Here β and ρ are fixed. This condition says that (x, y, z) must lie outside a certain ellipsoid.
Though they have different centers and semi-axes, by choosing K sufficiently big, we may
ensure EK lies outside this ellipsoid thereby ensuring the condition is satisfied. Thus, we have
our trapping ellipsoids guaranteeing that trajectories cannot escape to infinity.
• Thus, for ρ > ρH , trajectories are confined to a large ellipsoid in phase space but are gener-
ically not attracted to any fixed point, limit cycle or invariant surface enclosing a nonzero
volume. This non-periodic and non-quasi-periodic bounded motion may be shown to display ex-
treme sensitivity to initial conditions (with the associated difficulty in making reliable long-term
predictions). The dynamics is said to be chaotic. Due to the shrinking of phase space volume,
trajectories must approach an attractor of zero volume, it is called a strange attractor and is a
fractal set. A typical trajectory for Lorenz’s values of parameters is displayed in Fig. 16. For
more on the Lorenz attractor and chaos in the Lorenz system, see the book by Strogatz.

Figure 16: A ‘chaotic’ trajectory in the Lorenz model (viewed in the x - z plane) for ρ = 28, σ = 10 and
β = 8/3 . The trajectory ‘goes round’ the ‘left lobe’ a few times before switching to the right lobe. The lobes are
located around the unstable fixed points C± . This winding and switching continues indefinitely with no apparent
regularity. Figure from Wikipedia.

52