Cheat Sheet MAT230

Non-linear differential equations – Tommy Odland – Last edit: June 6, 2017

• Picard iteration
Given ẋ = f (x) and x(t0 ) = x0 , we estimate the
GENERAL NOTES solution in the vicinity of x0 by a sequence of ap-
proximations:
BPreliminaries x0 = x0
Z t
• The general framework for ODEs is x1 = x0 + f (x0 )dt
t0
ẋ = f (x), x(0) = x0 . Z t
x2 = x0 + f (x1 )dt
– Time is just a variable xi , and higher-order t0
derivatives can be reduced to first order deriva- .. ..
.=.
tives by introducing new variables. Z t
– A well posed problem has a unique solution with xn = x0 + f (xn−1 )dt
continuous dependence on the initial conditions. t0

– If f is continuous and all its partial derivatives

are continuous, a unique solution exists.
• A function satisfies a Lipschitz condition if there BBifurcations
exists an A such that |f (x2 ) − f (x1 )| ≤ A|x2 − x1 |.
If f 0 (x) is limited, such an A exists. If this condi- • A bifurcation is a qualitative change in behavior.
tion is satisfied, a unique solution exists. • Normal forms of the most common bifurcations:
– Continuous derivative implies Lipschitz, but not
vice-versa. – Saddle node – Pitchfork
• Since explicit solutions are difficult to produce and
ẋ = r + x2 ẋ = rx − x3
hard to interpret, the common procedure of anal-
ysis is instead the following:
– Subcritical pitch-
– Transcritical fork
System → Trajectories → Solution info

• Flows on the line are governed by ẋ = f (x), i.e. ẋ = rx − x2 ẋ = rx + x3

one dimensional maps.
• A fixed point x∗ is a point such that f (x∗ ) = 0. • The purpose of dimensional analysis is to write
an equation in dimensionless form. Two advan-
• The logistics equation is given by Ṅ =
tages are (1) “small” can be defined as  1 and
rN (1 − N/K), where r > 0 is a growth rate and
(2) the number of parameters are reduced by in-
K is a carrying limit. The solution is N (t) =

−1 troducing dimensionless groups.
K 1 + e−kt , and the behavior can be deduced
– How to non-dimensionalize: (1) set a new time
from looking at the fixed points f (N ∗ ) = 0.
scale τ = t/T and a new dependent variable
• The idea of linear stability analysis is to lin- N = x/a, (2) find the derivatives with respect to
earize around a fixed point x∗ . Introducing η(t) = τ , (3) make every term non-dimensional and (4)
x(t) − x∗ implies ẋ = η̇, and we obtain put the unknowns into non-dimensional groups.
η̇ = ẋ = f (x∗ + η) ≈ ηf 0 (x∗ ),
and f 0 (x∗ ) (or F 0 (x∗ )) determines the linearized BLinear systems
behavior around x∗ .
• A linear system is of the form
• Potentials can be used to infer behavior from the
system. The potential V (x) is defined as ẋ = Ax, A ∈ R2×2 .
dV
f (x) = − , • The eigenvalues of A determine the behavior in the
dx
phase plane, since the general solution is
and we have
x(t) = C1 eλ1 t v1 + C2 eλ2 t v2
V (x∗ ) = min ⇒ x∗ is stable,
V (x∗ ) = max ⇒ x∗ is unstable. • In general: nodes, saddles and spirals can occur.
 • A rough sketch is:1 • A conserved quantity is a real valued, non-zero,
– real(λ) > 0 implies growth. continuous function E(x) such that Ė(x) = 0. The
quantity E(x) is conserved along trajectories.
– real(λ) < 0 implies decay .
– imag(λ) 6= 0 implies a spiral/center. – A conservative system cannot have any attract-
ing fixed points.
– If λ1 = λ2 we have a star or degenerate node, de-
– If a point x is a local minimum for E(x), then x
pending on the number of eigenvectors. If λi = 0
is a (perhaps non-linear) center.
we have a non-isolated fixed point.
• A Hamiltonian system has

BPhase plane (R2 ) ∂H ∂H

ẋ = ẏ = − ,
∂y ∂x
• A good website for plotting phase planes is http: where H(x, y) is the Hamiltonian. H(x, y) is con-
//comp.uark.edu/~aeb019/pplane.html. stant along solution paths, since Ḣ(x, y) = 0.
• Quantitative behavior with exact formulas is often • A system is reversible if it’s invariant under a
unattainable, so we settle for qualitative behavior. change of variables x 7→ R(x), t 7→ −t.
• Nullclines are curves where either ẋ = 0 or ẏ = 0. – One common mapping is y 7→ −y, t 7→ t. (Re-
• Linearizing around f (x) = 0 gives correct informa- flection over x-axis.)
tion unless Re(λi ) = 0 for some i in the Jacobian • The index of a curve C is how many coun-
matrix of the linearization. If Re(λi ) = 0 the point terclockwise rotations the vector f (x) makes as
is “fragile”—linearizing does not always give the x moves in a counterclockwise closed trajectory
correct answer. around C once.
• The basin of attraction for a point x∗ is the sub- 1
set of the phase plane which sends all trajectories IC = [φ]
2π C
to x∗ .
Colloquially, IC measures the “winding” of f (x)
• Stability along the curve C.
– Poincare stability (stability of path) – As long as a continuous deformation C 7→ C 0
Let H∗ be the half path for the solution x∗ (t). does not cross through a fixed point, the index
Then H∗ is Poincare stable iff for every  > 0 is the same, i.e. IC = IC 0 .
there exists a δ() > 0 such that
– If C does not enclose fixed points, IC = 0.
||x(t0 ) − x∗ (t0 )|| < δ ⇒ max dist(H∗ , H) <  – For a saddle, Ix = −1. For other types of fixed
x
points Ix = +1.
where x(t) is a neighboring solution with path
– If IC encloses fixed points x∗1 , x∗2 , . . . , x∗n then
H.
n
– Liapunov stability (stability of solution) X
IC = Ii
The solution x∗ (t) is Liapunov stable for t ≥ t0 i=1
iff for every  > 0 there exists a δ(t, ) > 0 such
that
BAsymptotics
||x(t0 ) − x∗ (t0 )|| < δ ⇒ ||x(t) − x∗ (t)|| < 
• Big O and little o
for all t ≥ t0 , where x(t) is any neighboring so-
– f (x) = O(g(x)) if limx→x0 f (x)/g(x) < ∞, in
lution.
terms of order f (x) ≤ g(x) as x → x0 .
– Uniform stability (stability for all t)
– f (x) = o(g(x)) if limx→x0 f (x)/g(x) = 0, in
If a solution x∗ (t) is stable and δ if the definition
terms of order f (x) < g(x) as x → x0 .
of Liapunov stability is independent of t0 then
x∗ (t) is uniformly stable. I.e. the path is always • An asymptotic sequence of functions {φn } has
close to neighboring paths for every t0 . φn+1 (x) = o (φn (x)) as x → x0 .
– Asymptotic stability (stability in the limit) • An asymptotic series is of the form
If x∗ (t) is a stable solution, and in addition there n
X
exists a δ(t0 ) > 0 such that f (x; ) ∼ ak (x)φk (),  → 0 .
k=0
||x(t0 ) − x∗ (t0 )|| < δ ⇒ lim ||x(t) − x∗ (t)|| = 0
t→∞ The coefficients are calculated using
then x∗ (t) is asymptotically stable. x(t) is a
Pn−1
f (x; ) − k=0 ak (x)φk ()
neighboring solution. an (x) = lim .
→0 φn (x)
1
A dynamic website on linear systems: http://mathlets.org/mathlets/linear-phase-portraits-matrix-entry/
 • An asymptotic expansion does not have to con- • Regular perturbation theory consists of writ-
verge to be useful: as n → ∞ for a fixed  it does ing the solution as
not have to converge for all x, but for a fixed N it
should converge as  → 0 for all x. x(t, ) = x0 (t) + x1 (t) + 2 x2 (t) + O(3 ).
• A sequence becomes non-uniform when φn does
not dominate φn+1 , i.e. when two adjacent terms BBoundary layers
are of the same order. The region of non-
uniformity for 1 + x + 2 x2 + . . . is x = O(1/). • Boundary layers arise when  is in front of the high-
• Non-uniformity arises from: est order derivative, this causes a singularity.
– Infinite domains, due to secular terms.
–  in front of highest derivative. BDiscrete maps and fractals
• Optimal truncation: truncate in front of small-
• A fixed point x∗ = f (x∗ ) has the following stability
est term of the asymptotic expansion.
properties:
– |f 0 (x∗ )| < 1 ⇒ Stable
BLimit cycles(closed orbits) – |f 0 (x∗ )| > 1 ⇒ Unstable
– |f 0 (x∗ )| = 1 ⇒ Linear analysis inconclusive
• Ruling out closed orbits
• Cobwebs can be used to picture iterated maps.
– If ẋ = −∇V (gradient system) there are no
closed orbits. • The logistics map xn+1 = rxn (1 − xn ) exhibits
chaos when r ≈ 3.5699.
– If there exists a Liapunov function there is no
closed orbit. A Liapunov function has: • The similarity dimension of an object/fractal is
i. V (x) > 0 for all x 6= x∗ (pos.def). ln(m)
ii. V̇ (x) < 0 for all x (downhill flow). D= ,
ln(r)
And V (x) must be continuously differentiable,
i.e. continuous and continuous first derivatives. where r is the shrinking factor and m is the number
– Dulac’s criterion states that if ∇·(g ẋ) has one of boxes needed to cover the object.
sign in Ω there are no closed orbits in Ω, where • The box dimension is
g is a scalar function. This is from the div. thm:
ln (N ())
ZZ I D = lim ,
→0 ln (1/)
∇ · F dA = F · n̂ dr
where N () is the number of boxes of side length
Some choices for g are 1, 1/(xa y b ), eax and eby .  needed to cover the object.
– Bendixson’s negative criterion is Dulac’s cri- – The Cantor set has dimension D = ln(2)/ ln(3).
terion with g = 1.
• Finding closed orbits BChaos, attractors and Liapunov
– The Poincaré-Bendixson theorem states
that if one can construct a trapping region then • Chaos is aperiodic behavior in a deterministic sys-
Ω must have a limit cycle if Ω has no fixed points. tem with sensitive dependence on the initial con-
ditions.
• An attractor is a closed set A with the following
properties:
Ω (a) All trajectories starting in A stay in A.
(b) A attracts and open set of initial conditions.
(c) A is minimal with respect to these properties.
• Assume that we have two solution trajectories x(t)
and x(t) + δ(t), then ||δ(x)|| ∼ ||δ0 ||eλt , where λ is
• Relaxation oscillations operate on two time
the Liapunov exponent. If λ > 1 there is sensi-
scales: a slow buildup and a fast release.
tive dependence of the initial conditions.
• The general weakly non-linear equation is

ẍ + ω02 x = F (x, ẋ),

and a specific example is the Duffing equation

ẍ + x + x3 = 0.
 a(t) cos(ω0 t + θ(t)), where ω0 is found in the un-
perturbed problem.
RECIPES (TECHNIQUES) 2. Assume that u̇(t) = −ω0 a sin(ω0 t + θ). Compare
with u̇(t)true to obtain one constraint. Compute
d
dt (u̇(t)ass ) and plug into the problem to obtain
BRecipe: Strained coordinates second constraint.
When to use: When straight forward Poincare ex- 3. Obtain differential equations for a(t) and θ(t) from
pansion fails due to non-uniformities. Does not work if the above constraints.
amplitude needs to be adjusted. 4. Averaging the RHS of the differential equations
2
1. Introduce τ = t(1 + w1 +  w2 + . . . ) give the averaging method.
2. Compute derivatives.
3. Insert into equation, collect terms. BRecipe: Boundary layers
4. Use freedom in w1 to cancel terms giving rise to
secular terms, i.e. terms that grow boundlessly as When to use: Boundary value problems with  in
t → ∞. front of highest derivative.
1. Determine location of the boundary layer, deter-
mined by a(x) in
BRecipe: Multiscale
f 00 + a(x)f 0 + b(x)f = c(x), x1 < x < x2 .
When to use: When straight forward Poincare ex-
pansion fails due to non-uniformities. Has the power
to change amplitude and frequency. (a) If a(x) > 0, then the BL is at x = x1 .
1. Construct a fast time T0 = t and a slow time (b) If a(x) < 0, then the BL is at x = x2 .
T1 = t. Functions of the slow time T1 will be (c) If a(x) changes sign, then the BL is at a(x) = 0.
regarded as constant on the fast time scale T0 .
2. Choose s = ±(x − xi )/p near the BL. Choose p
We have
so as to keep as many terms as possible.
x(t, ) = x0 (T0 , T1 ) + x1 (T0 , T1 ) + O(2 ) 3. Use the Prandtl matching condition (here
with BL at x1 )
and the time derivative becomes
dx ∂x ∂x lim f inner = lim f0outer
ẋ = = + . s→∞ 0 x→x0
dt ∂T0 ∂T1
to determine coefficients, the limit is also f0match .
2. Plug the series expansion into the equation, col-
lect powers of . A typical solution for O(1) is 4. The solution becomes
A(T1 )eiT0 + A∗ (T1 )e−iT0 .
f ∼ f0inner + f0outer − f0match .
3. Plug into O(), get differential equation for A(T1 )
by setting resonant terms to zero.
4. Use A(T1 ) := R(T1 )eiθ(T1 ) to get differential equa- BReferences
tions for R(T1 ) and θ(T1 ).
5. Use initial conditions and solve. • Lecture notes.
• Steven H. Strogatz (Steven Henry). Nonlinear Dy-
namics and Chaos: With Applications to Physics,
BRecipe: Averaging Biology, Chemistry, and Engineering. Second edi-
tion. Westview Press, a member of the Perseus Books
When to use: in weakly non-linear oscillators, exem- Group, 2015.
plified by ü + ω02 u = F (u, u̇). • Alan W. Bush. Perturbation Methods for Engineers
1. The unperturbed problem has solution u(t) = and Scientists. CRC Press Library of Engineering
a cos(ω0 t + θ). Assume the solution is u(t) = Mathematics. Boca Raton, Fla: CRC Press, 1992.