0% found this document useful (0 votes)
31 views23 pages

Yu.A. Kuznetsov (Utrecht University, NL)

This document provides an overview of numerical continuation methods for solving algebraic and differential equations. It discusses: 1) Equilibria of autonomous ODEs and how to determine stability of equilibria. 2) Algebraic continuation problems (ALCPs) and definitions related to regular, limit, and branching points on solution curves. 3) The Moore-Penrose numerical continuation method which uses tangent predictions and Newton-Moore-Penrose corrections to iteratively compute points on the solution curve.
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
31 views23 pages

Yu.A. Kuznetsov (Utrecht University, NL)

This document provides an overview of numerical continuation methods for solving algebraic and differential equations. It discusses: 1) Equilibria of autonomous ODEs and how to determine stability of equilibria. 2) Algebraic continuation problems (ALCPs) and definitions related to regular, limit, and branching points on solution curves. 3) The Moore-Penrose numerical continuation method which uses tangent predictions and Newton-Moore-Penrose corrections to iteratively compute points on the solution curve.
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 23

Lecture 1

Continuation problems.
Numerical continuation of
equilibria and limit cycles of
ODEs

Yu.A. Kuznetsov (Utrecht University, NL)

May 5, 2009

1
Contents

1. Equilibria of autonomous ODEs.

2. Algebraic continuation problems.

3. Moore-Penrose numerical continuation.

4. Limit cycles of autonomous ODEs.

5. Boundary-value continuation problems.

6. Discretization via orthogonal collocation.

2
1. Equilibria of autonomous ODEs
• Consider a system of autonomous ODEs
depending on one parameter:
u̇ = f (u, α), u ∈ Rn , α ∈ R,
where f : Rn × R → Rn is smooth.
• Equilibrium manifold:

f (u, α) = 0

• Let u0 ∈ Rn be an equilibrium at parameter


value α0 and A0 = fu(u0, α0).

If ℜ(λ) < 0 for each eigenvalue λ of A0, u0


is stable.
If ℜ(λ) > 0 for at least one eigenvalue λ of
A0, u0 is unstable.

3
2. Algebraic continuation problems

• Def. 1 ALCP: Find a curve M ⊂ RN +1,


implicitly defined by a smooth function F

F (x) = 0, F : RN +1 → RN ,
starting from a point x0 ∈ M .

Finding an equilibrium manifold is an exam-


ple of ALCP with N = n,
!
u
x = (u, α) ≡ ∈ Rn+1, F (x) = f (u, α).
α
• Def. 2 A point p ∈ M is called regular for
ALCP if rank Fx(p) = N .

p
x0

• Near any regular point p, the ALCP defines a


solution curve M that passes through p and
is locally unique and smooth.

4
• If p ∈ M is a regular point, then the linear
equation

Jv = 0, J = Fx(p),
has a unique (modulus scaling) solution v ∈
RN +1.

• Lemma 1 A tangent vector v to M at p


satisfies
Jv = 0.

Indeed, let x = x(s) be a smooth parametriza-


tion of M , such that x(0) = p and ẋ(0) = v.
The differentiation of F (x(s)) = 0 yields at
s = 0:
d
F (x(s)) = Fx(x(0))ẋ(0) = Jv = 0
ds s=0

• Def. 3 A regular point p ∈ M is a limit


point for ALCP with respect to a coordi-
nate xj if vj = 0.

5
x1
 
w
v=
0

xN +1

If p is a limit point w.r.t. xN +1, then the N × N


matrix
!
∂Fi(p)
A=
∂xj i,j=1,...,N
has eigenvalue λ = 0. Indeed, let x = x(s) be a
smooth parametrization of M , such that x(0) =
p and ẋ(0) = v with
!
w
v= 6= 0, w ∈ RN .
0
Then
∂F (p)
Jv = Aw + vN +1 = Aw = 0.
∂xN

6
• Def. 4 A point p ∈ M is called a branching
point for ALCP if rank Fx(p) < N .

Let p = 0 be a branching point. Write


1
F (x) = Jx + B(x, x) + O(kxk3),
2
where J = Fx(p). Introduce the null-spaces

N (J) = {v ∈ RN +1 : Jv = 0}
and

N (J T) = {w ∈ RN : J Tw = 0}.

• Assume that

dim N (J) = 2 and dim N (J T) = 1.


Let q1 and q2 span N (J) and ϕ span N (J T).
Then

v = β1q1 + β2q2, w = αϕ,


where (β1, β2) ∈ R2, α ∈ R.

7
• Suppose we have a solution curve x = x(s)
passing through the branching point p = 0:
x(0) = 0, ẋ(0) = v.

• By differentiating F (x(s)) = 0 twice with re-


spect to s at s = 0, taking the scalar product
with ϕ, and using J Tϕ = 0, one proves:

Lemma 2 Any tangent vector v ∈ RN +1 to


M at p = 0 satisfies the equation

hϕ, B(v, v)i = 0.

• Substituting here v = β1q1 + β2q2, we obtain


the Algebraic Branching Equation:

b11β12 + 2b12β1β2 + b22β22 = 0,


where bij = hϕ, B(qi , qj )i, i, j = 1, 2.

8
• Def. 5 A branching point, for which
(a) dim N (J) = 2 and dim N (J T) = 1;
(b) b2
12 − b11b22 > 0,
is called a simple branching point.

x(2)
x(1)

v (2)

0
v (1)
N (J)

• Suppose that one solution curve x = x(1)(s)


passing through a simple branch point p = 0
is known and v (1) = ẋ(1)(0) = q1, so that
(1) (1)
β1 = 1, β2 = 0.
(2) (2)
Thus, b11 = 0 and v (2) = β1 q1 + β2 q2
tangent to the second solution curve x =
x(2)(s) satisfies
(2) (2)
2b12β1 + b22β2 =0
or
(2) b22 (2)
β1 = − β .
2b12 2

9
Lemma 3 Consider the (N +1)×(N +1)-matrix
(1)
 
Fx(x (s))
D(s) =  h iT  .
ẋ(1)(s)
Its determinant ψ(s) = det D(s) has a regular
zero at the simple branching point.

Indeed, let q2 ∈ N (J) be a vector orthogonal to


q1 = v (1). Then

D(0)q2 = 0,
so D(0) is singular and has eigenvalue λ(0) = 0.

Moreover, one can show that this eigenvalue is


simple and its smooth continuation λ = λ(s) for
D(s) satisfies
hϕ, B(q1 , q2)i b
λ̇(0) = = 12 6= 0,
hp, q2i hp, q2i
!
ϕ
where DT(0)p = 0 with p = .
0

Thus ψ̇(0) 6= 0.

10
3. Moore-Penrose numerical continuation
• Numerical solution of the ALCP means com-
puting a sequence of points
x(1), x(2), x(3), . . .
approximating the curve M with desired ac-
curacy, given an initial point x(0) that is
sufficiently close to x0.
• Predictor-corrector method:

– Tangent prediction: X 0 = x(i) + hiv (i).


X0

x(i+1)
v (i)

x(i)

– Newton-Moore-Penrose corrections towards


M:
(X k , V k ), k = 1, 2, 3, . . . .

– Adaptive step-size control.

11
• Def. 6 Let J be an N × (N + 1) matrix with
rank J = N . Its Moore-Penrose inverse is
J + = J T(JJ T)−1.

• To compute J +b efficiently, set up the sys-


tem for x ∈ RN +1:
(
Jx = b,
v Tx = 0,
where b ∈ RN and v ∈ RN +1, Jv = 0, kvk = 1.
Then x = J +b is a solution to this system,
since

JJ +b = b, v TJ +b = (Jv)T[(JJ T)−1b] = 0.

• Let x(i) ∈ RN +1 be a regular point on the


curve

F (x) = 0, f : RN +1 → RN ,
and v (i) ∈ RN +1 be the tangent vector to
this curve at x(i) such that

Fx(x(i))v (i) = 0, kv (i)k = 1.

12
For the next point x(i+1) ∈ RN on the curve,
solve the optimization problem

min{kx − X 0k | F (x) = 0},


x
i.e. look for a point x ∈ M which is nearest to
X 0:

X0

x
v (i) v
(i+1)
x
x(i)

This is equivalent to solving the system


(
F (x) = 0,
v T(x − X 0) = 0,
where v ∈ RN satisfies Fx(x)v = 0 with kvk = 1
and X 0 is the prediction.

13
The linearization of the system about X 0 is
F (X 0) + Fx(X 0)(X − X 0) = 0,
(

(V 0)T(X − X 0) = 0,
or
Fx(X 0)(X − X 0) = −F (X 0),
(

(V 0)T(X − X 0) = 0,
where Fx(X 0)V 0 = 0 with kV 0k = 1. Thus

X = X 0 − Fx+(X 0)F (X 0 )
leading to the Moore-Penrose corrections:

X k+1 = X k − Fx+ (X k )F (X k ), k = 0, 1, 2, . . . ,

V0
X0
X1
V1
x
v (i) v
(i+1)
x
x(i)

where V k ∈ RN +1 such that Fx(X k )V k = 0 with


kV k k = 1 should be used to compute Fx+ (X k ).
14
Approximate V k : Fx(X k−1)V k = 0.

V0
X0

X1
V1
X2
v (i)
V2
x(i+1)
x(i) v (i+1)

Implementation: Iterate for k = 0, 1, 2, . . .


!
k J
J = Fx(X ), B= T ,
Vk
! !
JV k F (X k )
R= , Q= ,
0 0
k −1 k+1 W
W =V −B R, V =
kW k
X k+1 = X k − B −1Q.
If kF (X k )k < ε0 and kX k+1 − X k k < ε1 then

x(i+1) = X k+1, v (i+1) = V k+1.

15
4. Limit cycles of autonomous ODEs

• Assume, the ODE system

u̇ = f (u, α), u ∈ Rn, α ∈ R,


has at α0 an isolated periodic orbit (limit
cycle) L0. Let u0(t+T0) = u0(t) denote the
corresponding periodic solution with minimal
period T0

• Introduce the matrix

A(t) = fu (u0(t), α0), A(t + T0) = A(t),


and consider a matrix M (t) which satisfies

Ṁ = A(t)M, M (0) = In,


wthere In is the identity n × n matrix.

Def. 7 The eigenvalues µ1, µ2, . . . , µn = 1 of


the monodromy matrix M (T0) are called
the multipliers.

If |µ| < 1 for each multiplier of M (T0) except


µn = 1, L0 is stable.
If |µ| > 1 for at least one multiplier of M (T0),
L0 is unstable.

16
• Let L0 be a cycle of period T0 at α0 and
u0(t) its corresponding solution.

L0

u(0)

• Consider a periodic boundary-value prob-


lem on [0, 1]:
(
ẇ − T0f (w, α) = 0,
w(0) − w(1) = 0.
Clearly, w(τ ) = u0(T0τ + σ0), α = α0 is a
solution to this BVP for any phase shift σ0.

• Let v(τ ) be a smooth period-1 function. To


fix σ0, impose the integral phase condi-
tion:
Z 1
Ψ[w] = hw(τ ), v̇(τ )idτ = 0
0

17
Lemma 4 The condition
Z 1
hw(τ ), v̇(τ )idτ = 0
0
is a necessary condition for the L2-distance
Z 1
ρ(σ) = kw(τ + σ) − v(τ )k2dτ
0
between 1-periodic smooth functions w and v to
achieve a local minimum with respect to possible
shifts σ at σ = 0.

Since kwk2 = hw, wi,


Z 1
1
ρ̇(0) = hw(τ + σ) − v(τ ), ẇ(τ + σ)idτ
2 0 σ=0
Z 1
= hw(τ ) − v(τ ), ẇ(τ )idτ
0
Z 1 Z 1
= hw(τ ), ẇ(τ )idτ − hv(τ ), ẇ(τ )idτ
0 0
1 1
Z Z 1
= dkw(τ )k2 − hv(τ ), ẇ(τ )idτ
2 0
Z 1
0

= hw(τ ), v̇(τ )idτ .


0

18
5. Boundary-value continuation problems

• Def. 8 BVCP: Find a branch of solutions


(u(τ ), β) to the following boundary-value prob-
lem with integral constraints



 u̇(τ ) − H(u(τ ), β) = 0, τ ∈ [0, 1],
B(u(0), u(1), β) = 0,

Z 1



 C(u(τ ), β) dτ = 0,
0
starting from a given solution (u0(τ ), β0).
Here u ∈ Rnu , β ∈ Rnβ and

H : Rnu × Rnβ → Rnu ,


B : Rnu × Rnu × Rnβ → Rnb ,
C : Rnu × Rnβ → Rnc
are smooth functions.

• The BVCP is (formally) well posed if

nβ = nb + nc − nu + 1.

19
6. Discretization via orthogonal collocation

• Mesh points: 0 = τ0 < τ1 < . . . < τN = 1.

• Basis points:
j
τi,j = τi + (τi+1 − τi),
m
where i = 0, 1, . . . , N − 1, j = 0, 1, . . . , m.

• Approximation:
m
u(i)(τ ) = ui,j li,j (τ ),
X
τ ∈ [τi, τi+1],
j=0
where li,j (τ ) are the Lagrange basis poly-
nomials
m
Y τ − τi,k
li,j (τ ) =
τ − τi,k
k=0,k6=j i,j

and ui,m = ui+1,0.

20
• Orthogonal collocation:
 P 
m u i,j l ′ (ζ ) − H(
Pm
u i,j l (ζ ), β) = 0,


 j=0 i,j i,k j=0 i,j i,k
F : B(u 0,0 , uN −1,m, β) = 0,
 PN −1 Pm i,j
j=0 ωi,j C(u , β) = 0,


i=0
where ζi,k , k = 1, 2, . . . , m, are the Gauss
points (roots of the Legendre polynomials
relative to the interval [τi, τi+1]), and ωi,j are
the Lagrange quadrature coefficients.

• Approximation error: Introduce

h= max |τi − τi−1|


i=1,2,...,N

– in the basis points:

ku(τi,j ) − ui,j k = O(hm)

– in the mesh points:

ku(τi) − ui,0k = O(h2m)

21
• BVCP for the limit cycle branch with α ∈ R:

 ẇ(τ ) − T f (w(τ ), α) = 0, τ ∈ [0, 1],

w(0) − w(1) = 0,
R1
0 hw(τ ), v̇(τ )i dτ = 0.

• Corresponding HUGE ALCP:

F (x) = 0, x = ({wj,k }, T, α) ∈ RmnN +n+2


where j = 0, 1, . . . , N − 1, k = 0, 1, . . . , m.

• The linearization of BVCP with respect to


(w, T, α):
 
D − T fw (w, α) −f (w, α) −T fα(w, α)
δ0 − δ1 0 0
 
 
Intv̇ 0 0
produces a sparse Jacobian matrix Fx:
 
w0,0 w0,1 w1,0 w1,1 w2,0 w2,1 w3,0 T α
• • • • • • • •
• • • • • • • •
 
 
 • • • • • • • • 

 • • • • • • • • 

 • • • • • • • • 
 • • • • • • • • 
• • • • • • • •
 
 
 • • • • • • • • 
 • • • • • • • • 
• • • • • • • •
 
 
 • • • • • • • • 
 • • • • • • • • 
• • • •
 
 
• • • •
• • • • • • • • • • • • • •

22
Computation of the multipliers

• After Gauss elimination:

w0,0 w0,1 w1,0 w1,1 w2,0 w2,1 w3,0


 
T α
• • • • • • • •
• • ◦ • • • • •
 
 
 • • ◦ ◦ • • • • 
 • • ◦ ◦ ◦ • • • 
• • • • • • • •
 
 
 • • ◦ • • • • • 
 • • ◦ ◦ ◦ ◦ • • • • 
• • ◦ ◦ ◦ ◦ ◦ • • •
 
 
 • • • • • • • • 
 • • ◦ • • • • • 
∗ ∗ ◦ ◦ ◦ ◦ • •
 
 ⋆ ⋆ 
 ∗ ∗ ◦ ◦ ◦ ◦ ⋆ ⋆ • • 

 • • • • 

 • • • • 
• • • • • • • • • • • • • •
• • • • • • • • • • • • • • • •

• Let P0 be the matrix block marked by ∗’s


and P1 the matrix block marked by ⋆’s. We
have w0,0 = w(0), wN,0 = w(1) implying

P0w(0) + P1w(1) = P0u(0) + P1u(T ) = 0,


so that M = −P1−1P0 is the numerical ap-
proximation of the monodromy matrix M (T )
and its eigenvalues are the numerical approx-
imations of the cycle multipliers.

23

You might also like