Вы находитесь на странице: 1из 9

Introduction to Nonlinear Physics Summary Notes

Matthew Evans
January 7, 2014
A set of notes based on the lecture course of the same name given by Dr T Galla at the University
of Manchester in Autumn 2013. Please email matthew@charmedxi.co.uk if you spot an error.

One-dimensional first-order systems

Graphical analysis

Systems of the form x = f (x) are called autonomous, in contrast to non-autonomous systems with
x = f (x, t). The solution is some function x(t) which can normally be found by separation of variables.
This gets us so far, but it is often much easier to think about the system graphically. We define points
where x = 0 as fixed points. The flow is towards a fixed point if x is decreasing as it goes through
the FP; we call this a stable FP, an attractor (denoted on plots as a solid circle). If x is increasing as it
crosses the FP, we have an unstable FP, a repeller (hollow circle). If the FP is an extremum, it will be
semi-stable (half-filled in circle).

Linear stability

Consider x = f (x) with f (x ) = 0. Introducing a small perturbation x (t) at x allows us to test the
. As x is small, we can
stability of a fixed point algebraically. We now have x(t) = x + x (t) with x = x

Taylor expand x = x = f (x + x ) to yield


= f 0 (x ) x .
x

(1.1)

This equation can be solved immediately by x (t) = x (0) exp ( f 0 (x )t), given that f 0 (x ) is a constant.
This leaves 2 cases:
if f 0 (x ) > 0, the perturbation grows exponentially so the FP is unstable,
if f 0 (x ) < 0, the perturbation decays exponentially so the FP is stable.

Potentials

Given x = f (x) for smooth


f , we can always find a function such that f (x) = dV /dx. We call V
R
a potential, V (x) = x f (x 0 )dx 0 . The extrema of the potential tell us about the fixed points of the
system and their stability in a trivial way; maxima are unstable, minima are stable, saddle points are
semi-stable.

Impossibility of oscillations in first-order systems

The only things that can happen for a 1D vector field are that the trajectories approach a fixed point, or
diverge to . Overshoot and damping can never occur in a first-order system: there are no periodic
solutions to x = f (x).

II

Bifurcations
Bifurcations in 1D continuous systems

Dynamical systems often depend on external parameters. A bifurcation is a change of the qualitative
behaviour of the system as a result of the variation of a parameter, e.g. change in number/stability of
FP.

Saddle-node bifurcation

Saddle-node bifurcations involve the creation and annihilation of fixed points. As a parameter is varied,
two fixed points move towards each other, collide and annihilate. An example of a system with a
saddle-node bifucation is x = r + x 2 : altering r changes the vertical displacement of the quadratic in
x, eventually the quadratic will not cross the x = 0 line and thus there will be no fixed points. This can
be visualised by plotting x = x 2 then drawing on different values of r for r = x 2 . We can also make r
continuous and draw
p a plot of x against r. This is called a bifurcation diagram. This would consist of
two lines, x = (r), corresponding to the two fixed points.

snbif.png

Figure 1: Examples of saddle-node, transcritical, supercritical pitchfork and subcritical pitchfork (left
to right) bifurcation diagrams .
A saddle-node bifurcation occurs when f (x) is locally bowl-shaped at the bifurcation point, i.e.
parabolic when it crosses the x-axis. This can be seen by Taylor expanding: saddle-node bifurcations
have the form x = f (x) = C + d x 2 + ...

Transcritical bifurcation

Transcritical bifrucations involve the changing of stability of a fixed point. The normal form for a
transcritical bifurcation is x = r x x 2 . A physical example of this is population growth, there is
always a fixed point at zero population, but it exhibits a change in stability when the growth rate (the
parameter) transitions from being positive to negative.
2

Pitchfork bifurcation

Pitchfork bifurcations are common in physical problems that have symmetry. Fixed points can appear
and disappear in symmetric pairs. There are two types of pitchfork bifurcation. The first is called
supercritical pitchfork bifurcation which has form x = r x x 3 (parity invariant). At r < 0, the only
FP is the origin, and it is stable. When r = 0, the origin is still stable, but perturbations no longer decay
exponentially. When r > 0, the origin has become unstable and two more fixed points appear either
p
side of the origin at x = x. Conversely, the subcritical pitchfork bifurcation has form x = r x + x 3
p
which leads to a stable origin and two unstable FP at r, which transform into just one unstable FP
at the origin past the bifurcation point.

III

Numerical methods for first-order ODEs


Euler method

We can solve first-order ODEs numerically by discretising time in steps of h to get numerical values of
x(t 1 ), x(t 2 ) up to x(t f ) that we want to be as close to the true solution as possible. Writing x n = x(t n ),
the simplest approximation is to use the slope at the point x n to find x n+1 . The Euler-forward method
consists of iterative application of the following formula:
x n+1 = x n + h f (x n , t n ),

(3.1)

i.e. a straight line from x n to x n+1 . Expanding x(t n + h) yields x(t n ) + h f (x n , t n ) + h2 x (t n )/2 + ..., so
the Euler method introduces an error of order h2 at each step. This is a first-order method as the total
error is of order h (the number of steps required is proportional to h1 with error h2 per step).

Midpoint method

The midpoint method improves on the Euler method by approximating the slope midway between two
points x n and x n+1 to find x n+1 , it follows the formula


h
h
x n+1 = x n + h f x n + f (x n , t n ), t n +
.
(3.2)
2
2
This is a second-order method, the error at each step is of order h3 so the total error is O (h2 ).

Fourth-order Runge-Kutta

We could write the midpoint method as k1 = h f (x n , t n ), k2 = h f (x n +k1 /2, t n +h/2) and x n+1 = x n +k2 .
The Runge-Kutta method is a generalisation of this, with 4 terms:
k1 = h f (x n , t n ),

k1
k2 = h f x n + , t n +
2

k2
k3 = h f x n + , t n +
2

(3.3)
h

2
h

(3.4)

(3.5)

k4 = h f (x n + k3 , t n + h),
3

(3.6)

and finally,
x n+1 = x n +

k1
6

k2
3

k3
3

k4
6

(3.7)

This is a fourth-order method, the error per step is of order h5 .

IV

Higher-dimensional linear systems


Two-dimensional linear systems

Two-dimensional linear systems have the form x = a x + b y, y = c x + d y, for parameters a, b, c, d. In


vector notation,
!
a b
x=
x.
(4.1)
c d
If we have solutions x1 and x2 , then as the system is linear, any linear combination of these is also
a solution. We can take second-order ODEs in one dimension and reduce them to first order ODEs
in two dimensions by defining x = y, e.g. harmonic oscillator mh = kh, let x = h, y = h, so
x = p
y, y = k x/m. We end up with the solutions x(t) = h0 cos (t) and y(t) = h0 sin (t), for
= k/m.

Phase space

If we plot the two solutions (x(t), y(t)), we end up with the parametric plot in the x y-plane, called
the trajectory in the phase plane. At each point on the trajectory, the velocity of the phase space point
is v = (
x , y ). Plotting all these points gives us the phase portrait of the system. In 2D, fixed points are
now the points in phase space where v = (0, 0).
For some general system X1 = f (X 1 , X 2 ), X2 = g(X 1 , X 2 ), the curves defining X1 = 0 and X2 = 0 are
called the nullclines of the system.
Trajectories in phase space never intersect. If they did, there would exist points that have two
distinct velocity vectors. Broadly speaking, there are only four types of trajectory in 2D:
closed curves (cycles),
heteroclinic orbits, curves of finite lengths connecting two different fixed points,
homoclinic orbits, a curve leaving a fixed point then very slowly tending back towards it (perhaps
never reaching it)
open curves, a curve where at least one end diverges to .

Linear stability analysis in higher dimensions

Suppose we have x = f (x, y), y = g(x, y), where (x , y ) denotes a fixed point. Again, we consider a
small perturbation on each coordinate and Taylor expand:





f
f
1 2f 2 1 2f
2f

2




x y +... (4.2)
y +
y +
f (x + x , y + y ) = f (x , y )+
x +
x +
x FP
y FP
2 x 2 FP
2 y 2 FP
x y FP

As the perturbations are small, we ignore terms of order 2 in the perturbation; noticing that each
function vanishes at the FP allows us to write,
!
!
!
d x (t)
x f y f
x (t)
=

.
(4.3)
dt y (t)
y (t)
x g y g FP
The matrix of partial derivatives is called the Jacobian. Each partial derivative is calculate at the FP, so
they will just yield numbers. Let us make the following definition:
!
!
J11 J12
x f y f
=
= J .
(4.4)
J
J
g g
21

22

FP

Thus, we can rewrite the previous equation more compactly as


= J x

.
x

(4.5)

Performing the matrix multiplications yield the following components,


= J11 x + J12 y ,
x

(4.6)

y = J21 x + J22 y ,

(4.7)

a 2x2 linear system with two linearly independent solutions. To solve this, we make an ansatz of
x = et and y = et . Substituting this ansatz into the above yields the following eigenvalue problem:
!
!


.
(4.8)

=J

This means that the vector (, ) must be an eigenvector of the Jacobian with eigenvalue . If the
Jacobian has two distinct eigenvalues, the general solution to the original problem will then be a linear
combination of the two eigenvectors, for example
= Ae1 t
x

1
1

!
+ Be2 t

2
2

!
.

(4.9)

The eigenvectors of this system define two eigendirections which have stability that depends on
their respective eigenvalue. We call the eigendirection with the smallest absolute eigenvalue the slow
eigendirection as this is the direction that trajectories approach as t .
This can be easily extended to higher dimensions by considering the N xN Jacobian of a system of
N equations.

Classification of fixed points

It can be seen from Equation (??) that we have the following cases:
(i) If both eigenvalues are real and negative, we have a stable fixed point.
(ii) If both eigenvalues are real and positive, we have an unstable fixed point.
5

(iii) If both eigenvalues are real with one positive and one negative, we have a saddle point, considered unstable.
However, we can also have complex eigenvalues, these themselves lead to three possible cases: a centre
(phase portrait of concentric circles), an unstable spiral (spiralling away from FP) or a stable spiral
(spiralling into FP). If the eigenvalues are complex, then they form a conjugate pair. We can write
1,2 = i, where the new parameters and determine the amplitude and oscillation of the
spirals respectively. If is negative, we have a stable spiral as the amplitude decays. If it is positive, we
have an unstable spiral, and if it is zero we have a centre.
A general 2x2 Jacobian has the following characteristic equation for its eigenvalues:
2 TrJ + det J = 0,

(4.10)

if we denote the trace as and the determinant as , we can straight away right the formula for the
eigenvalues:
r

2
=
.
(4.11)
2
4
Everything is then determined by and , which give us 1 2 and 1 + 2 respectively. We now have
the previous cases again, reparameterised:
(i) Saddle points: < 0, 2 4 > 0.
(ii) Nodes: > 0, 2 4 > 0. The node is unstable if sgn = 1 and stable if sgn = 1.
(iii) Spirals and centres: if > 0 and 2 4 < 0, we have a complex conjugate pair. If < 0, we
have a stable spiral, if > 0, we have an unstable spiral and if = 0, we have a centre. There
are some other cases, borderlines, stars and degenerate nodes on the boundaries of these cases.

Effects of nonlinear terms

Pattern formation

Mass-action kinetics

Many systems in nature can be modelled as individual-based systems. The law of mass action states
that the rate at which a reaction occurs is proportional to the product of the concentrations of the
reactants. The general case: a reaction of 3 chemicals A, B and C with concentrations a, b and c is
given by the equation:
A + mB k nB + pC,
(5.1)
will have rates of change in concentration of
a = ka b m ,

b = (n m)ka b m ,

c = pka b m .

(5.2)

Or in the most general case,


n1 A1 + n2 A2 + ... k n01 A1 + n02 A2 + ...

(5.3)

yields
ai = (n0i ni )k
6

i
Y

ai i .

(5.4)

VI

Instability in PDEs and pattern formation


Homogeneous solutions

Consider the PDE,


= Du00 + F (u),
u

(6.1)

where F (u) is some reaction term, e.g. Ginzburg-Landau equation has F (u) = ru u . We look
= u00 = 0. This means that
for homogeneous solutions in space and time, u(x, t) = uh, yielding u
p
F (uh) = 0, so, for Ginzburg-Landau,uh = 0 or r, assuming r > 0. Is this stable? Consider a small
perturbation v(x, t), such that u(x, t) = uh + v(x, t). Substituting this in,

dF
00

v = Dv (x, t) +
v(x, t).
(6.2)
du u=uh
3

We make an ansatz: v(x, t) = et e ik x for real .



dF

v = v = D(ik) v +
du
2

v,

(6.3)

= (k).

(6.4)

u=uh


dF

= k D +
du
2

uh

We can split our solution into real and imaginary parts, then consider the boundary conditions: x u(0, t) =
x u(L, t) = 0 = x v(0, t) = x v(L, t) = 0. The imaginary part, v = et sin (k x) will never work, so
the solution is v = et cos (k x) with k constrained to k = n/L with (k) = F 0 (uh) Dk2 . We now have
two cases:
(i) F 0 (uh) > 0: the modes near k = 0 are unstable, their growth rate (k) is positive. The most
unstable node is k = 0. The resulting instability will be dominated by this mode; there is no
dependence on x, so no patterns are observed.
(ii) F 0 (uh) < 0: (k) < 0 k. All modes are stable; perturbations of any wavenumber will decay.
When we make the et e ik x ansatz, we are going into Fourier space, where 2 k2 . This can be
called diagonalising the Laplacian by Fourier transforming. In principle, k could be complex, but these
solutions are unphysical. From now on, the ansatz will be used with the understanding that the real
part is the only physically relevant part.

The Turing mechanism

The Turing mechanism was an attempt to answer questions posed by nature such as how is symmetry
broken in the development of an embryo? We now consider instabilities in multi-dimensional PDEs:
!
!
!
!
1
u
D1 0
u001
f1 (u1 , u2 )
=
u
=
+
.
(6.5)
2
u
0 D2
u002
f2 (u1 , u2
If uh = (u1,h, u2,h), consider the effect of a perturbation v(x, t) = (v1 (x, t), v2 (x, t)). These result in the
following linearised equations:
!
!
D1 0
v100
=
v
+ Jhv,
(6.6)
D2 0
v200
7

where Jh is the Jacobian of the system evaluated at uh, which we shall assume is equal to
!
a b
Jh =
.
c d

(6.7)

Let us first consider the non-spatial model where 1,2 = f1,2 (1 , 2 ) and 1,2 = 1,2 (t) for systems
which have a stable FP, such that Tr(Jh) = a + d < 0 and det (Jh) = ad bc > 0. Making the ansatz
v = A1,2 e t cos (k x) yields the following expression,
v=
and so

A1

k2 D1

k2 D2

A2

!
v + Jhv,

k2 D1 + a

k2 D2 + d

(6.8)

!
,

(6.9)

so (k) are eigenvalues of M(k), the matrix above:


(k) =

i
p
1h
Tr(M) (Tr(M)2 4 det (M) .
2

(6.10)

As we assumed the trace of the Jacobian was negative, the trace of M is also negative. The mode with
wavenumber k is unstable iff det (M(k)) < 0:
= det (M(k)) = D1 D2 k4 k2 (D1 d + D2 a) + (ad bc),

(6.11)

so, if ad bc > 0 as assumed, and D1 d + D2 a > 0 then (k) has regions where nodes are stable and regions where nodes are unstable ((k) < 0), the perturbations grow. They cannot grow forever though,
in our linear stability analysis we assumed that v was small, which breaks down as the amplitude grows;
nonlinearities eventually kick in and limit the growth.
The unstable region is defined by kmin < k < kmax . Remember, for systems with a finite domain,
only modes with k = n/L are allowed. For small L, this could mean that there are no allowed modes
inside the region and thus perturbations always decay. For larger L, there may be multipled allowed
modes for which the perturbations grow. This is important to consider as the location of the unstable
modes is independent of L. Stripes form for L > L C for some critical value L C . The Turing mechanism
has been shown to apply to several natural systems, for example activator-inhibitors in biology, certain
chemical reactions and perhaps also animal coat patterns.

VII

Discrete systems in 1D

Maps

Fixed points and stability

Cobweb construction

Stability of 2D maps

Periodic cycles in 1D maps and their stability

Feigenbaum universality

VIII

Chaos and fractals

Definition and Liapunov exponents

Fractals

Self-similarity and power laws

Fractal dimensions

Strange attractors

Stretch and fold

Lorenz attractor

Volumes in phase space

Вам также может понравиться