Вы находитесь на странице: 1из 30

Geometric Integration

Grady Lemoine
AMATH 575
May 31, 2007
Motivation
• Most systems of ODEs are too complex to solve
analytically
– Must numerically integrate
• But they also have structure we would like to
preserve
– Symmetries
– Volume preservation
– First integrals or Lyapunov functions
• Problem: A generic numerical integrator doesn’t
preserve these
• Solution: Construct a special numerical
integrator that does!
Outline
• Introduction and Preliminaries
• Survey of Integrators that Preserve
Structure
– Volume preservation
– First integrals and Lyapunov functions
• Example of Geometric Integration
• Conclusion
• References
Geometric Integration
• Called “geometric” because it preserves
aspects of the geometry of phase space
• Has been used successfully in
– Orbital mechanics
– Molecular dynamics
– Quantum mechanics
• Applied primarily to autonomous systems
Advantages of Geometric
Integration
• Usable for larger time steps and longer
integration times than generic methods
• Guaranteed to be qualitatively correct
even for chaotic systems
– Can preserve invariant tori
• More accurate for some or all time scales
• Sometimes cheaper computationally
Disadvantages of Geometric
Integration
• Structure to be preserved must be known
in advance
• Not generic – integrator must be specially
constructed for each problem
• Can typically only preserve one aspect of
structure (volume preservation, one first
integral…)
• Sometimes more expensive
computationally
Digression: Order Doesn’t Matter*
Theorem (Yoshida, Qin, and Zhu):
Let ψ be an integrator with order of accuracy n > 0. Then the
integrator χ defined by
1
χτ = ψ γτ ψ (1− 2γ )τ ψ γτ , where γ = (2 − 2 n +1 −1
)

is of order n+1. If ψ is self-adjoint (ψ-τ-1= ψτ), then so is χ, and χ


has order n+2 (and ψ had even order). Note that 1−2γ < 0.
Practical meaning: Many of the integrators we come up with
will be low-order. But if we have any integrator that
preserves a property, we can bootstrap it to arbitrary order.
*Caveat: This requires taking a “backward” step, so we can’t use it if backward
steps aren’t allowed (e.g. if we’re preserving a Lyapunov function).
Volume Preservation
• Flows: x′ = f(x) is volume-preserving if
m
∂f i
∇⋅ f = ∑ =0
i =1 ∂xi

• Integrator ψτ is volume-preserving if
⎛ ∂ψ τ ,i ⎞
det⎜ ⎟ =1
⎜ ∂x ⎟
⎝ j ⎠

• Two ways to construct volume-preserving


integrators
– Splitting method
– Correction method
Splitting Method
• Consider an m-dimensional volume-
preserving flow
m
∂f i
xi ' = f i ( x), ∑ =0
i = 0 ∂xi

• Note that
∂f m m −1
∂f i m −1
∂f i
= −∑ ⇒ f m = −∑ ∫ dxm
∂xm i =1 ∂xi i =1 ∂xi
• We can use this to split the original system
into m-1 Hamiltonian systems
Splitting Method
• For each j from 1 to m-1, get a 2D Hamiltonian
system:
∂H j
xj '=
∂xm
where H j = ∫ f j ( x)dxm
∂H j
xm ' = −
∂x j
• Solve each of these in sequence by your method
of choice
– May be simple enough to solve exactly
– If not, any symplectic integrator will do (symplectic
integrator = integrator that is a symplectic map, e.g.
Verlet method)
Correction Method
• Directly adjusts integrator to maintain
constant phase space volume
• Simplest case (based on Euler method):
x1,n +1 = x1,n + τ f1 ( ~
x)
⎛ ∂xi ,n +1 ⎞
J = det⎜ ⎟, i, j = 1, … , m − 1
~ where ⎜ ∂x ⎟
xm −1,n +1 = xm −1,n + τ f m −1 ( x ) ⎝ j ,n ⎠
~
x = ( x1,n , … , xm −1,n , xm ,n +1 )
xm ,n+1
~
∫ J ( x ) dx m , n +1 = x m ,n

• Implicit method
• Can be based on another method – could
obtain higher order or other properties too
Preserving a First Integral
• Many systems have first integrals
– Hamiltonian systems are the obvious example
• Preserving a first integral can provide
guarantees about system behavior
– In general: Level surfaces of the integral are invariant
sets of the system; preserving an integral means the
numerical solution also has those invariant sets
– e.g. orbital mechanics: Preserving energy
guarantees the orbiting body won’t erroneously fall
into the Sun or leave the solar system
How to Preserve a First Integral
• Given: System x′ = f(x), first integral I(x)
• Construct equivalent “skew gradient” system:

x' = S ∇I ( x), S T = − S
• Define skew discrete gradient system:
~
xn+1 − xn =τ S ∇I (xn , xn+1)
• Show that I(xn+1) = I(xn)
Finding the Matrix S
• Sometimes a matrix S is already known
– Hamiltonian systems: S = J if we’re preserving H
• Otherwise it’s easy to construct:
f (∇I )T − (∇I ) f T
S=
| ∇I | 2
• S is not unique
• If the critical points of I are nondegenerate, there
exists an S that is as smooth as f and the
derivatives of I
Finding the Discrete Gradient
• The discrete gradient must satisfy two
properties:
I ( xn +1 ) − I ( xn ) = ( ∇ I ( xn , xn +1 )) ⋅ ( xn +1 − xn )
∇ I ( xn , xn +1 ) = ∇I ( xn ) + O( xn +1 − xn )

• Given these properties, the integral is preserved:


~
I ( xn +1 ) − I ( xn ) = ( ∇ I ) ⋅ ( xn +1 − xn ) = τ ( ∇ I ) S ( ∇ I ) = 0
T

• The discrete S used here may be any consistent


antisymmetric matrix (e.g. S((xn+1 + xn)/2)
Example Discrete Gradients
• The discrete gradient isn’t unique
• Itoh and Abe:
I ( x (i ) ) − I ( x ( i −1) )
∇ I ( xn , xn +1 ) i =
xn +1,i − xn ,i
where x ( i ) = [ xn +1,1 , … , xn +1,i , xn ,i +1 , … , xn ,m ]T
• Harten, Lax, and Van Leer:
1
∇ I ( xn , xn +1 ) = ∫ ∇I ( xn + ξ ( xn +1 − xn ))dξ
0

• HLL discrete gradient is symmetric in xn and xn+1,


so it produces a self-adjoint integrator
Preserving a Lyapunov Function
• Lyapunov functions occur for:
– Gradient systems
– Systems with dissipation
– Asymptotically stable fixed points
• Preserving a Lyapunov function is very
similar to preserving a first integral
How to Preserve a Lyapunov
Function
• Given: System x′ = f(x), Lyapunov function
V(x)
• Construct equivalent “linear-gradient”
system:
x' = L ∇V ( x), L negative semidefinite
• Define linear discrete gradient system:
~
xn+1 − xn =τ L ∇V (xn , xn+1)
• Show that V(xn+1) ≤ V(xn)
Building the Integrator
• L is not unique; one possible solution
(I = identity matrix):
f (∇V )T − (∇V ) f T + ( f T ∇V ) I
L=
| ∇V | 2
• Behavior of L near critical points of V is
similar to behavior of S for first integrals
• Same method may be used to construct
the discrete gradient
• Discrete L matrix handled the same way
as discrete S matrix
Geometric Integration Example:
Comet Hale-Bopp
• Bright comet visible in the
Spring of 1997
• Highly eccentric orbit
(e = 0.9951) – hard to
integrate
• Dynamics are Hamiltonian
• Good candidate for
geometric integration

Image credit: A. Dimai, R. Volcan, A. Zardini


(Col Druscie Obs., ACC)
Orbital Mechanics
• For simplicity, just look at the two-body problem
• Newton’s Law: r ' ' = − µ r , µ = GM
Sun
| r |3
• After a rescaling to get µ = 1 and some algebra,
get a Hamiltonian setup in polar coordinates.
Note: H = total energy.
q1 = r p1 = r '
q2 = θ p2 = r 2θ '
(r ' ) 2 + (rθ ' ) 2 1 p12 p22 1
H= − = + 2−
2 r 2 2q1 q1
• Notice that H has no dependence on q2, so p2 is
constant. This simplifies matters considerably.
Energy-Preserving Integrator
• Need antisymmetric matrix S and discrete
gradient of first integral (energy)
• System is Hamiltonian, so S = J
• Use HLL discrete gradient in order to get self-
adjoint integrator
– Guaranteed 2nd-order accuracy, 4th-order after
forming composition from earlier theorem
• HLL discrete gradient (holding p2 constant):
⎛ 2q1,n q1,n +1 − p22 (q1,n + q1,n +1 ) ⎞
⎜ ⎟
⎜ 2 2
2q1,n q1,n +1 ⎟
⎜ 0 ⎟
∇ H ( xn , xn +1 ) = ∫ ∇H ( xn + ξ ( xn +1 − xn ))dξ = ⎜ ⎟
1
( p1,n + p1,n +1 )
0 ⎜ ⎟
⎜ 2 ⎟
⎜ p2 ⎟
⎜ q1,n q1,n +1 ⎟
⎝ ⎠
Energy-Preserving Integrator
• Form the integrator step:
~
xn+1 − xn =τ S ∇I (xn , xn+1)
• In this case we get:
⎛ τ ( p1,n +1 + p1,n ) 2 ⎞
⎛ q1,n +1 − q1,n ⎞ ⎜ ⎟
⎜ ⎟ ⎜ τ p2 (q1,n +1q1,n ) ⎟
q − q
⎜ 2,n +1 2,n ⎟ ⎜
= − ⎟
2 ( q1, n +1 + q1, n )
2
⎜p ⎟ 2 q q p
− p1,n
⎟ ⎜ −τ
1, n +1 1, n
⎜ ⎟
⎜ 1,n +1 2 2
2q1,n +1q1,n ⎟
⎜p − p ⎟
⎝ 2,n +1 2,n ⎠ ⎜ 0 ⎟
⎝ ⎠

• This is an implicit 2nd-order integrator that


preserves energy exactly
Numerical Results
• Compared geometric integrator with standard
Runge-Kutta
– Geometric integrator was made 4th-order using
composition
– Results compared against 4th-order Runge-Kutta
– Both fixed-step
• Simulation parameters:
– Work done in “canonical” units
• Distance measured in AU (1 AU = 1.5e8 km = mean Earth-Sun
distance)
• Time measured in TU (1 TU = (1/2π) year)
– Time span was 5 orbits (80,353 TU or 12,789 years)
– Time step was 0.2 TU (~12 days)
– Coarse time step chosen to investigate nature of errors
What’s Going On?
• Geometric integrator
– Makes small, cumulative errors in θ
– Preserves energy as designed (error ~10-14)
– Preserving energy and angular momentum (the p2
coordinate from the Hamiltonian) means the size and
shape of the orbit are fixed for all time
• Runge-Kutta
– Makes cumulative errors in energy
– Orbit period depends on energy!
– Runge-Kutta falls out of phase with exact solution
• Geometric integration preserves a vital aspect of
the system’s structure
Conclusion
• Geometric integration offers a way to
preserve essential structure of a system
during numerical solution
• It can preserve many properties:
– Symmetries
– Volume preservation
– First integrals and Lyapunov functions
• This can lead to a substantial increase in
accuracy
References
• R. I. McLachlan and G. R. W. Quispel, “Six Lectures on the
Geometric Integration of ODEs,” MSRI Preprint No. 1998-
061
• C.J. Budd and M.D. Piggott, “Geometric integration and its
applications,” Preprint Dept. of Math. Sc., Univ. of Bath
(2000).
http://citeseer.ist.psu.edu/article/budd00geometric.html
• G. R. W. Quispel, “Volume-preserving integrators,” Physics
Letters 206A (1995), 26-30
• R. R. Bate, D. D. Mueller, and J. E. White, Fundamentals of
Astrodynamics, 1971
• Hale-Bopp orbital elements:
http://www2.jpl.nasa.gov/comet/ephemjpl8.html
• Code available on request (email
gl@amath.washington.edu)

Вам также может понравиться