Mathematisches Forschungsinstitut Oberwolfach

Report No. 14/2006

Geometric Numerical Integration

Organised by
Ernst Hairer (Geneve )
Marlis Hochbruck (Düsseldorf )
Arieh Iserles (Cambridge )
Christian Lubich (Tübingen )

March 19th – March 25th, 2006

Abstract. The subject of this workshop was numerical methods that pre-
serve geometric properties of the flow of an ordinary or partial differential
equation. This was complemented by the question as to how preserving the
geometry affects the dynamical behaviour.

Mathematics Subject Classification (2000): 65xx.

Introduction by the Organisers

The subject of this workshop was numerical methods that preserve geometric
properties of the flow of an ordinary or partial differential equation: symplectic
and multisymplectic integrators for Hamiltonian systems, symmetric integrators
for reversible systems, methods preserving first integrals and numerical methods
on manifolds, inclusive of Lie group methods and integrators for constrained me-
chanical systems, and methods for problems with highly oscillatory solutions. The
unifying theme was structure preservation: not just the “how?” but also “why?”,
“where?” and “what for?”.
The motivation for developing structure-preserving algorithms for special classes
of problems originates independently in such diverse areas of research as astronomy,
molecular dynamics, mechanics, control theory, theoretical physics and numerical
analysis, with important contributions from other areas of both applied and pure
mathematics. Moreover, it turns out that preservation of geometric properties
of the flow not only produces an improved qualitative behaviour, but also allows
for a significantly more accurate long-time integration than with general-purpose
methods.
In addition to the construction of geometric integrators, an important aspect of
geometric integration is the light it sheds on the relationship between geometric
806 Oberwolfach Report 14/2006

properties of a numerical method and favourable error propagation in long-time

integration. A very successful organising principle is backward error analysis,
whereby the numerical one-step map is interpreted as (almost) the flow of a mod-
ified differential equation. In this way, geometric properties of the numerical inte-
grator translate seamlessly into structure preservation on the level of the modified
equation. Much insight and rigourous error estimates over long time intervals can
then be obtained by combining backward error analysis with the KAM theory and
related perturbation theories for Hamiltonian and reversible systems. While this
approach has been very successful for ordinary differential equations, much less
is currently known about highly oscillatory systems and geometric integrators for
partial differential equations.
Geometric numerical integration has been an active interdisciplinary research
area since the last decade. Although the subject is in a lively phase of intensive
development, the results so far are substantive and they impact on a wide range
of application areas and on our understanding of core issues in computational
mathematics. This is evidenced by the monographs [1, 2].

References
[1] E. Hairer, Ch. Lubich, G. Wanner, Geometric Numerical Integration. Structure-Preserving
Algorithms. Springer, Berlin, 2002, 2nd edition 2006.
[2] B. Leimkuhler, S. Reich, Simulating Hamiltonian Dynamics. Cambridge Univ. Press, 2004.
Geometric Numerical Integration 807

Workshop: Geometric Numerical Integration

Table of Contents

Benedict Leimkuhler
Geometric Integrators for Thermostatted Molecular Dynamics . . . . . . . . . . 811
Ilan Degani (joint with David J. Tannor)
Calculating Multidimensional Discrete Variable Representations From
Cubature Formulae . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 813
Erwan Faou (joint with Guillaume Dujardin)
Sobolev estimates for splitting schemes applied to the linear Schrödinger
equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 814
Chris Budd (joint with John F Williams)
Parabolic Monge-Ampère Methods for Blow-up problems in several spatial
dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 817
Tudor S. Ratiu
Semidirect Products and Their Relation to Integrable Systems . . . . . . . . . . . 819
Debra Lewis (joint with Peter Olver)
Doing what comes naturally (learning optimal control from the experts) . . 822
Melvin Leok (joint with Anthony M. Bloch, Islam I. Hussein, Taeyoung
Lee, N. Harris McClamroch, Amit K. Sanyal)
Lie Group Variational Integrators and its Applications to Geometric
Control Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 825
Sebastian Reich (joint with Jason Frank, Nigel Wood, Andrew Staniforth)
Semi-implicit semi-Lagrangian time-stepping methods and regularized
fluid equations in numerical weather prediction . . . . . . . . . . . . . . . . . . . . . . . 828
Marcel Oliver (joint with Onno Bokhove)
Parcel Eulerian–Lagrangian fluid dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . 829
Brynjulf Owren (joint with Håvard Berland, Bård Skaflestad)
Properties of exponential integrators for nonlinear wave equations . . . . . . . 832
Assyr Abdulle
Numerical Methods for Multiscale Problems . . . . . . . . . . . . . . . . . . . . . . . . . . 834
Moody T. Chu (joint with Nicoletta Del Buono)
Structural Preserving Isospectral Flows for Quadratic Pencils . . . . . . . . . . . 837
Philippe Chartier (joint with Erwan Faou)
A numerical method for Hamiltonian systems based on piecewise smooth
space-approximations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 841
808 Oberwolfach Report 14/2006

G.R.W. Quispel (joint with D.I. McLaren)

Integral–preserving Integrators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 844
Sergio Blanes (joint with Fernando Casas, Ander Murua)
Splitting Methods in Geometric Numerical Integration . . . . . . . . . . . . . . . . . 846
Ander Murua (joint with Sergio Blanes, Fernando Casas)
Splitting methods for the harmonic oscillator . . . . . . . . . . . . . . . . . . . . . . . . . 849
Laurent O. Jay
Using additivity in numerical integration of DAEs . . . . . . . . . . . . . . . . . . . . . 850
Antonella Zanna (joint with Robert I. McLachlan, Hans Z. Munthe-Kaas
and G. R. W. Quispel)
Explicit, volume preserving splitting methods for divergence-free
polynomial vector fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 853
Zaijiu Shang
Stability analysis of symplectic methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 854
Simon J.A. Malham (joint with Gabriel Lord, Anke Wiese)
Efficient strong integration of linear stochastic systems . . . . . . . . . . . . . . . . . 855
Elena Celledoni
Semi-Lagrangian methods and new integrators for convection dominated
problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 857
Begoña Cano
Conserved quantities of some Hamiltonian wave equations after full
discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 860
Yuri B. Suris
Euler-Poincaré integrators: variational construction and integrability . . . . 861
Francesco Fassò
Stability of rigid body motions (and numerical integrations) . . . . . . . . . . . . . 864
Anthony Bloch
Geodesic Flows on Manifolds and their Discretizations . . . . . . . . . . . . . . . . . 867
Volker Grimm (joint with Marlis Hochbruck)
Exponential integrators for highly oscillatory differential equations . . . . . . . 868
David Cohen
Highly oscillatory Hamiltonian systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 869
Katina Lorenz (joint with Christian Lubich)
Adiabatic integrators for highly oscillatory differential equations . . . . . . . . . 870
Syvert P. Nørsett (joint with Arieh Iserles)
Highly oscillatory quadrature, the one dimensional case . . . . . . . . . . . . . . . . 871
Arieh Iserles (joint with Syvert P. Nørsett)
Multivariate highly oscillatory quadrature . . . . . . . . . . . . . . . . . . . . . . . . . . . . 872
Geometric Numerical Integration 809

Robert I. McLachlan
Integration and applications of generalized Euler equations . . . . . . . . . . . . . 875
Some open problems in Geometric Numerical Integration
Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 877
Geometric Numerical Integration 811

Abstracts

Geometric Integrators for Thermostatted Molecular Dynamics

Benedict Leimkuhler

Nosé dynamics is a popular and effective device for simulating molecular systems
in the canonical ensemble. Let a closed Hamiltonian system be given with energy
function
1
H(q, p) = pT M −1 p + V (q),
2
with q and p positions and momenta, respectively, M a positive definite symmetric
mass matrix, and V the potential energy function. Nosé dynamics is derived from
the extended phase space Hamiltonian,
π2
HN = H(q, p̃/σ) + + gkB T ln σ,
2µ
where g is the number of degrees of freedom, kB the Boltzmann constant, and
T is the target temperature at which sampling is desired. µ is a parameter that
effectively allows the strength of dynamic coupling to be adjusted. The momen-
tum appearing in HN should be treated as canonical to q, whereas the physical
momentum is related to p̃ by the change of variables
p̃
p= .
σ
It was shown by Nosé that canonical sampling can be obtained along (assumed
ergodic) trajectories of HN via the relation
Z Z Z
 0

. . . δ HN − HN dp̃dσdπ = exp(−βH(q, p))dp,

where the integration is performed over the physically accessible phase space of
the thermostatting variables, (σ, π) ∈ (0, ∞) × R.
While useful for understanding the concept of Nosé dynamics, HN is not usu-
ally recommended for simulation because, on the one hand, computation of certain
types of averages (e.g. autocorrelation functions) requires data at equally spaced
points in time, and, more importantly, the equations of motion corresponding
to HN are poorly scaled for σ → 0. This is the motivation for the well known
Nosé-Hoover reformulation of Nosé dynamics which forms the basis for most nu-
merical treatments (see Section 2). An alternative formulation incorporating time-
transformation in the Hamiltonian setting can be used as the basis for symplectic
numerical methods ([1]), based on an alternative time-transformation technique
(Poincaré transformation). Simulation with the Nosé-Poincaré method is based
on the Hamiltonian
0
(1) HN P = σ(HN − HN )
0
where the constant HN must be chosen so that HN vanishes at the initial value,
and hence for all time along Hamiltonian dynamics in the extended phase space.
812 Oberwolfach Report 14/2006

The primary advantage anticipated from the Nosé-Poincaré approach is improved

stability in long-term simulations and/or at large timestep size, and some evidence
for this (in certain situations) was presented in [1] and in a subsequent paper on
constant temperature/pressure dynamics [2].
A more subtle problem with the Nosé/Nosé-Hoover/Nosé-Poincaré formulation
is that poor ergodicity may be observed, particularly for systems at low temper-
ature (i.e. near-crystalline) or systems with strong harmonic components such as
arise in modelling chemical bonds in organic molecules. A number of devices, e.g.
Nosé-Hoover chains have been introduced to facilitate more effective sampling of
phase space. It is natural to attempt to extend the Hamiltonian framework to
allow symplectic treatment with chains of thermostats. Recent papers have in-
troduced Nosé-Poincaré chains [3], in direct analogy to Nosé-Hoover chains, and
Recursive Multiple Thermostats [4]. In the most general form, these thermostats
can be described by a Hamiltonian of the form
HGN = H(q, p̃/Πα {σα }) + HG (σ1 , σ2 , . . . , σm , π1 , π2 , . . . , πm )
where the ”bath” HG is chosen so that the canonical density can be obtained
through integration over the thermostatting variables:
Z Z Z
 0

. . . (δ HGN − HGN dp̃)dσ1 dσ2 . . . dσm dπ1 dπ2 . . . dπm = exp(−βH(q, p))dp

Examples of this type of thermostatting bath are discussed below. While these
methods can enable thermostatting which is quite rapid in relation to the simula-
tion time interval, the numerical methods appear in some cases to require small
timesteps compared to unthermostatted molecular dynamics and even compared
to alternatives such as Nosé-Hoover chains [5]. Until now, the cause of this has
not been explained. The numerical difficulties may actually become worse as the
numerical method achieves a better sampling, further complicating the issue. As
we show below, the instability is in part due to the fact that the Nosé-Poincaré
form enables propagation of the momenta and thermostat variables separately, in
such a way that the simple relationship between the physical and scaled Nosé mo-
menta (p̃) may be violated within a timestep. This numerical difficulty can render
inconsequential any advantage that might be obtained from the conservation of
symplectic structure. Moreover, the way in which typical symplectic methods are
constructed is based on a succession of stages. It is observed that within these
stages, the thermostatting variables may be driven into strong oscillations which
can have detrimental effects both for accuracy and stability of methods.
In this article we show that we can greatly enhance the stability of integrators
for Nosé chains (and also for Nosé dynamics proper) by using a simple general
principle to construct the balanced numerical method wherein, even at each sub-
stage of a numerical timestep, the thermal variables are retained near mechanical
equilibrium, provided the physical variables are in thermal equilibrium. We demon-
strate the concept by constructing several integration methods and applying them
to model problems including both a harmonic oscillator and a 3-body molecu-
lar model problem. We show that the balanced symplectic Nosé-Poincaré based
Geometric Numerical Integration 813

method and its chain counterpart have very strong advantages over both Nosé-
Hoover as well as other symplectic but not balanced integrators.

References
[1] S. Bond, B. Laird and B. Leimkuhler The Nosé-Poincaré method for constant temperature
molecular dynamics, Journal of Computational Physics 151 (1999), 114–134.
[2] J. Sturgeon and B. Laird, Symplectic algorithm for constant-pressure molecular dynamics
using a Nose-Poincare thermostat, Journal of Chemical Physics 112 (2000), 3474–3482.
[3] B. Leimkuhler and C. Sweet, The canonical ensemble via symplectic integrators using Nosé
and Nosé-Poincaré chains, Journal of Chemical Physics 121 (2004), 108–116.
[4] B. Leimkuhler and C. Sweet, A Hamiltonian formulation for recursive multiple thermostats
in a common timescale, SIAM J Applied Dynamical Systems 4 (2005), 187–216.
[5] E. Barth, B. Leimkuhler and C. Sweet, Approach to thermal equilibrium in biomolecular
simulation, in New Algorithms for Macromolecular Simulation, Springer Lecture Notes in
Computational Science and Engineering 49 (2005).

Calculating Multidimensional Discrete Variable Representations From

Cubature Formulae
Ilan Degani
(joint work with David J. Tannor)

Discrete Variable Representations (DVRs) are heavily used in ”real life” high di-
mensional problems in computational quantum mechanics. First, to semi-discretize
the Hamiltonian operator a finite dimensional function space S is chosen. The ba-
sic idea of DVR is to then find a basis of δ-like functions (DVR basis functions),
each localized above its DVR point, which have several convenient properties for
computation:
(1) The projection of any f ∈ S onto a DVR basis function is done simply by
evaluating f at the corresponding DVR point.
(2) In the DVR basis the quantum mechanical potential operator is approxi-
mated simply by a diagonal matrix obtained by evaluating the potential
function on the DVR points.
(3) Orthogonal projection of functions to subspaces spanned by a subset of
DVR basis functions is easy to compute using item 1. This is needed if we
know that the wavefunctions of interest are localized in a subset of config-
uration space (view this as the subset of Rd over which our wavefunctions
live).
In this talk I describe the basics of DVR through the familiar example of S =
span{e−iqx , . . . , eiqx } and the discrete Fourier transform. Then I quote the general
definition of a DVR set given by Littlejohn et. al. in [1] as a general framework
for multidimensional DVRs - not necesarily with cartesian product grids, and not
necesarily with Fourier function spaces. However, until very recently almost all
known multidimensional DVRs were of the product type, where the dimension
of S grows rapidly with increasing dimension of configuration space. Because
814 Oberwolfach Report 14/2006

of this problem the search for non product DVRs is one of the major themes
in computational quantum mechanics. The talk ends by stating the following
problem whose solution gives DVR sets:
Problem. Denote the configuration space by Ω, and the space of multivari-
able degree q polynomials over Ω by PqΩ . Suppose we are given: (a) A de-
gree 2q or 2q + 1 cubature formula for the region Ω and weight function w(x),
with nodes λ1 , . . . , λN ∈ Ω and positive weights ω1 , . . .R, ωN . (b) An n1 dimen-
sional function space B with the inner product hg|hiw = Omega w(x)g ∗ (x)h(x)dx,
and such that PqΩ ⊆ B. Find, or prove that there do not exist, N functions
u1 , . . . , uN ∈ mathcalB such that
(1) uα (λβ ) = δαβ √1ωα , α, β = 1, . . . , N .
(2) huα |uβ iw = δαβ , α, β = 1, . . . , N .
(3) PqΩ ⊆ span{u1 , . . . , uN }.
This problem is solved in [2], thus associating families of non product DVRs, whose
function space is span{u1 , . . . , uN }, with any member in the menagerie of existing
(positive weight) cubature formulae.

References
[1] R.G. Littlejohn, M. Cargo, T. Carrington Jr., K.A. Mitchell, B. Poirier, A general framework
for discrete variable representation basis sets, J.Chem.Phys. 116 (2002), 8691.
[2] I. Degani, D.J. Tannor, Calculating multidimensional discrete variable representations from
cubature formulae, to appear in the Journal of Physical Chemistry, John Light festschrift
issue.

Sobolev estimates for splitting schemes applied to the linear

Schrödinger equation
Erwan Faou
(joint work with Guillaume Dujardin)

The aim of this work is to try to understand the long time behavior of splitting
methods applied to the linear Schrödinger equation. Let us consider the equation
∂ψ
(1) i = Hψ, ψ(0, x) = ψ0 (x),
∂t
where H = −∆+V is the Hamiltonian of the problem. The function ψ = ψ(t, x) is
a function depending on the time t and the space variable x in the one-dimensional
torus T. The potential V (x) is a real function. We consider the approximation
method:
(2) ψ1 = exp(i(δt)∆) exp(−i(δt)V )ψ0 ,
where δt is the stepsize, and where no space approximation is made. What is the
long time behavior of the corresponding “numerical” solution?
The eigenvalues of the Laplace operator being the n2 , n ∈ Z, we expect that the
splitting method (2) exhibits resonances (at least this is the case after a spectral
discretization). Moreover, even if these resonances are avoided, backward error
Geometric Numerical Integration 815

analysis cannot be used as in the finite dimensional case. Formally, the modified
energy in the case of the scheme (2) can be written
iδt
(3) H̃ = −∆ + V + [−∆, V ] + · · · + δtn Hn+1 + · · ·
2
where Hn+1 is a symmetric operator of order n. We arrive at the conclusion that
Sobolev estimates are the key to analyze the long time behavior of (2) and the
possible conservation of the modified energy (3) (see for instance [1]).
We make the following assumption for the potential function:
(4) ∀ n ≥ 0, kDn V k ∞ ≤ λn! εn ,
where D = −i∂x . The numbers λ and ε measure the “analytic size” of V : If these
parameters are small, V is close to 0 in analytical norm. In the following, we
denote by k · k the L2 -norm on T.
Before showing anything concerning the solution of the splitting method (2),
we first show the following result for the exact solution:
Theorem 1. Let ψ(t, x) be the solution of (1) and assume that ψ0 satisfies
(5) ∀ n ≥ 0, kDn ψ0 k ≤ M0 n! r0−n
for constants M0 and r0 . Assume moreover that the potential V satisfies (4) for
sufficiently small λ and ε. Then ψ is uniformly L2 -analytic, in the sense that there
exist constants M and R such that
∀ t ≥ 0, ∀ n ∈ N, kDn ψ(t, ·)k ≤ M n!R−n .
The proof of this theorem relies on the fact that for all s, H s commutes with
H and hence the quantities kH n/2 ψ(t, ·)k are constants along the solution of (1).
The key is to estimate the difference between the operator H n/2 and the derivative
Dn for all n. More precisely, we define the formal series:
X ρn X ρn
(6) Sρ (ψ) = kDn ψk and Zρ (ψ) = kH n/2 ψk .
n! n!
n≥0 n≥0

Assume that V satisfies (4) with λ ≤ ε2 ≤ ε0 . Then there exist positive constants
c1 , C1 depending on ε0 such that for all analytic function ψ, we have
(7) ∀ σ, 0 ≤ σ ≤ r0 , c1 Sσ (ψ) ≤ Zσ (ψ) ≤ C1 Sσ (ψ).
This relation implies the result.
We consider now the splitting method (2). In the following, δt is considered as
a fixed parameter and we introduce the familly of schemes:
Lλ := exp(i(δt)∆) exp(−iλ(δt)V ).
For λ = 0, we have L0 := exp(i(δt)∆). In particular, we have for all δt, [Ds , L0 ] = 0
for all s ∈ R. Let H0 be an operator such that [H0 , L0 ] = 0. We seek a modified
operator H(λ) such that
(8) [H(λ), Lλ ] = 0.
816 Oberwolfach Report 14/2006

P
Setting the formal series H(λ) = H0 + λH1 + n≥2 λn Hn , the equation (8) is
equivalent to the formal series equations
n−1
X
iδt∆ (−iδt)n−p
(9) ∀ n ≥ 0, [Hn , e ]=− [Hp , eiδt∆ V n−p ].
p=0
(n − p)!

Let us write the equation (9) for n = 1. We get

[H1 , eiδt∆ ] = iδteiδt∆ [H0 , V ].

After a Fourier transormation in T, we denote by (H1 )kℓ the coefficients of the

operator H1 acting on CZ . To solve this equation, we use the following non-
resonance condition (see also [1, 2]): There exist positive numbers γ > 0 and
ν > 1 such that

1 − eikδt
(10) ∀ k ∈ Z, ≥ γ|k|−ν .
δt

We thus get the bounds on the coefficients, ∀ k, ℓ ∈ Z, |k| =

6 |ℓ|,

δt 1 2
|(H1 )kℓ | ≤ iδt(ℓ2 −k2 )
([H0 , V ])kℓ ≤ |k − ℓ2 |ν ([H0 , V ])kℓ .
1−e γ

Under smoothness assumptions on V , this shows that H1 is an analytic operator.

Note that we can add to H1 a diagonal operator without affecting the equation
(8), so that we can assume that H1 is positive. By considering the operator
H̃(λ) := H0 + λH1 , we obtain [H̃(λ), Lλ ] = O(λ2 ). Continuing the process, our
aim is to obtain a modified operator commuting with Lλ up to a high order in λ
and to obtain analytic estimates for it. This work is still in progress.
To conclude, we show a numerical experiment with
3
V (x) = and ψ0 (x) = sin(x)
5 − 4 cos(x)

2π
and the stepsizes h = = 0.196 and h = 0.2. We plot the corresponding
62
− 22
numerical solutions after 105 iterations, and the relative energy errors. We observe
a drift in the resonant case, while the solution remains quasiperiodic in the case
of a non-resonant stepsize.

References
1. E. Hairer, C. Lubich, G. Wanner, Geometric Numerical Integration. Structure-preserving
Algorithms for Ordinary Differential Equations. Springer, Berlin, 2002.
2. Z. Shang, Resonant and Diophantine step sizes in computing invariant tori of Hamiltonian
systems. Nonlinearity 13 (2000), 299–308.
Geometric Numerical Integration 817

2 2

1 1

Re(ψ(t,x))

Re(ψ(t,x))
0 0

−1 −1

−2 −2

0 1 2 3 4 5 6 0 1 2 3 4 5 6
x ∈ (0,2π) x ∈ (0,2π)

Figure 1. Plot the numerical solution after 105 time steps for a
non-resonnant stepsize (left) and a resonnant stepsize (right).

0.05 25

0.04 20
Energy error

Energy error
0.03 15

0.02 10

0.01 5

0 0
0 0.5 1 1.5 2 0 0.5 1 1.5 2
Time 4 Time 4
x 10 x 10

Figure 2. Relative energy error for a non-resonnant stepsize

(left) and a resonnant stepsize (right).

Parabolic Monge-Ampère Methods for Blow-up problems in several

spatial dimensions
Chris Budd
(joint work with John F Williams)

One of the key geometric features associated with the solution of partial differential
equations is the evolution of structures on several different length scales. Indeed,
the analysis of multi-scale phenomenon is now a major source of research activity.
In certain problems, for example porous media flows, the length scales are widely
separated and these problems can be studied using, for example, multi-scale finite
element methods. However, there are many other problems in which length scales
are not widely separated and indeed we see a continuum of length scales. These
are generally driven by some nonlinearity in the problem and are tightly coupled
in an essentially geometric manner. Scaling laws are almost universal in many
partial differential equations and often describe asymptotic features of a solution
such as the formation of singularities and interfaces, when the effects of bound-
ary and initial conditions are less important. Examples arise in fluid mechanics,
magneto hydrodynamics, solid mechanics, nonlinear optics and electrostatics. It
818 Oberwolfach Report 14/2006

is clear that numerical methods which can exploit the scaling structures in a par-
tial differential equation are likely to perform well. To construct such methods
we need to understand the scaling symmetries that lie behind the various scaling
structures. Such symmetries are generally described by Abelian, diagonalizable Lie
groups, and thus have a simpler structure than many other problems studied by
geometric methods. However, the universality of scaling laws makes it appropriate
to develop methods just for them. The sort of problems that we are interested in
studying have length scales L(t) driven by nonlinear effects, which are vanishingly
small in some limit. Typically there is some time T for which L → 0 as t → T .
Such problems pose a challenge for any numerical method, and cannot be studied
by using a method with a fixed grid of size h when L(t) ≪ h. In such cases it is
appropriate to use an adaptive method, in which the spatial mesh can adapt itself
to the natural scale of the solution.
In this talk we will examine a class of r-adaptive moving mesh methods for
approximating solutions of the parabolic equation

(1) ut = ∆u + f (u), x ∈ Ω ⊂ Rd , u|∂Ω = 0.

We will consider geometrically based adaptive methods for both slowly evolving
solutions of (1) and also solutions be which become singular (blow-up) in a finite
time T . We will concentrate in our calculations on the case of d = 2, although the
methods can in principle work in higher dimensions. Whilst the solution of (1)
in one-dimension by using (scale invariant) r-adaptive methods is now fairly well
understood, less progress has been made in extending these methods to higher
dimensions where there are significant new difficulties. In this talk we describe
a new method for solving (1) in spatial dimensions greater than one which is
well adapted to exploiting emergent scaling structures. These methods will be
based on the Parabolic Monge-Ampère method. In this method a fixed mesh
in computational coordinates ξ is mapped to a moving mesh x in the physical
coordinates. The map from computational to physical coordinates is determined
in terms of the gradient of a mesh potential Q which satisfies a fully nonlinear
parabolic partial differential equation of the form

(2) (I − γ∆)Qt = M (u)H(Q)1/d

where H(Q) is the determinant of the Hessian of Q, d is the spatial dimension and
M (u) is a monitor function which depends upon the solution. In two dimensions

H(Q) = Qξξ Qηη − Q2ξη .

The equation (2) is solved in parallel with (1) and is constructed (through a care-
ful choice of the function M to have the same scaling symmetries. We show that
any scaling structure present in the solution u is inherited by the mesh poten-
tial function and derive from this a series of properties of the resulting mesh.
We demonstrate through examples, that (when careful attention is paid to mesh
Geometric Numerical Integration 819

regularity and smoothing) the Parabolic Monge-Ampère method is effective for

computing the singular solutions of (1) in two spatial dimensions.

Semidirect Products and Their Relation to Integrable Systems

Tudor S. Ratiu

This talk is a progress report on several topics that are interrelated: semidirect
product reduction and the discretization of integrable systems.
The classical theory of reduction for semidirect products of Lie groups with
vector spaces goes back to the late seventies ([1], [4], [5], [6]) and took its definitive
form in [2] and [3]. Due to lack of space the semidirect product theory will not be
reviewed here and we refer to the quoted papers. So only the discretization part
of the talk will be summarized below.
Next we shall give an interesting example of an integrable system on the semi-
direct product SE(3), namely the Kowalevski top, and will discretize it according
to the Hirota method. This is joint work with my student Murat Turhan.
The Kowalevski top is the third integrable case of rigid body dynamics:

Π̇ = Π × Ω + M glΓ × χ
Γ̇ = Γ × Ω
where the principal moments of inertia are I1 = I2 = 2I3 and the unit vector on
the line connecting the fixed point with the center of mass is χ = (1, 0, 0); M is
the total mass of the body, and l is the length of the segment that connects the
fixed point to the center of mass. Explicitly, this is
dΩ1 dΩ2 dΩ3
2 = Ω2 Ω 3 , 2 = −Ω1 Ω3 + x0 Γ3 , = −x0 Γ2 ,
dt dt dt
dΓ1 dΓ2 dΓ3
= Ω3 Γ 2 − Ω 2 Γ 3 , = Ω1 Γ 3 − Ω 3 Γ 1 , = Ω2 Γ 1 − Ω 1 Γ 2 ,
dt dt dt
where x0 := 2MglI1 . The Kowalevski equations have four conserved quantities. The
first three are standard for any heavy top:
(1) Hamiltonian: H1 = I21 (2Ω21 + 2Ω22 + Ω23 ) − x0 Γ1
(2) Angular momentum (Casimir): H2 = I1 (2Ω1 Γ1 + 2Ω2 Γ2 + Ω3 Γ3 )
(3) Spatial gravity vector (Casimir): H3 = Γ21 + Γ22 + Γ23
To prove integrability, one more conserved quantity is needed. Here is how Kowa-
levski found it. Let z = Ω1 + iΩ2 and ξ = Γ1 + iΓ2 . Then the equations of motion
imply :
2ż = −iΩ3 z + ix0 Γ3 and ξ˙ = iΓ3 z − iΩ3 ξ.
We can eliminate Γ3 by considering the combination z 2 − x0 ξ:
d 2
(z − x0 ξ) = 2z ż − x0 ξ˙
dt
= z(−iΩ3 z + ix0 Γ3 ) − x0 (iΓ3 z − iΩ3 ξ)
= −iΩ3 (z 2 − x0 ξ).
820 Oberwolfach Report 14/2006

By complex conjugation we get

d 2
(z̄ − x0 ξ̄) = iΩ3 (z̄ 2 − x0 ξ̄).
dt

This implies that |z 2 − x0 ξ|2 is conserved. In terms of the original variables, we

obtain the Kowalevski integral

H4 = |z 2 − x0 ξ|2 = |(Ω1 + iΩ2 )2 − x0 (Γ1 + iΓ2 )|2

2
= (Ω21 − Ω22 − x0 Γ1 ) + (2Ω1 Ω2 − x0 Γ2 )2 .

It is still a mystery where this comes from. If one writes in matrices what
Kowalevski did, one passes from R3 to so(3) to su(2) using the Pauli spin ma-
trices. But why this isomorphism of Lie algebras should give one more integrable
case is not known.
The discretization of the Kowalevski top according to the Hirota formulation
proceeds in the following way. Impose the dependent variable transformation:
g1 g2 g3
Ω1 = , Ω2 = , Ω3 = ,
f f f
g4 g5 g6
Γ1 = , Γ2 = , Γ3 =
f f f

and introduce the Hirota operator

dgj df
Dt g j · f = f − gj , j = 1, . . . , 6.
dt dt
The equations of motion become


 2Dt g1 · f = g2 g3



 2Dt g2 · f = −g1 g3 + x0 g6 f

Dt g3 · f = −x0 g5 f

 Dt g 4 · f = g 3 g 5 − g 2 g 6



 Dt g 5 · f = g 1 g 6 − g 3 g 4

Dt g 6 · f = g 2 g 4 − g 1 g 5

This is invariant under the gauge transformation

f (t) 7−→ f (t)h(t), gj (t) 7−→ gj (t)h(t)

where h(t) is an arbitrary function of t.

Replace the bilinear differential operator Dt gj · f by the corresponding bilinear
difference operator

Dt gj · f −→ [gj (t + δ)f (t) − gj (t)f (t + δ)]/(2δ)

Geometric Numerical Integration 821

where δ is a time interval. Using gauge invariance and time reversibility, we obtain
the following bilinear equations :


 2(g1t+1 f t − g1t f t+1 )/(2δ) = (g2t+1 g3t + g2t g3t+1 )/2



 2(g2t+1 f t − g2t f t+1 )/(2δ) = −(g1t+1 g3t + g1t g3t+1 )/2






 + x0 (g6t+1 g3t + g6t g3t+1 )/2



 I1 (g3t+1 f t − g3t f t+1 )/(2δ) = −x0 (g5t+1 f t + g5t f t+1 )/2




 (g t+1 f t − g t f t+1 )/(2δ) = (g t+1 g t + g t g t+1 )/2
4 4 3 5 3 5


 − (g2t+1 g6t + g2t g6t+1 )/2



 (g5t+1 f t − g5t f t+1 )/(2δ) = (g1t+1 g6t + g1t g6t+1 )/2





 − (g3t+1 g4t + g3t g4t+1 )/2






 (g6t+1 f t − g6t f t+1 )/(2δ) = (g2t+1 g4t + g2t g4t+1 )/2


− (g1t+1 g5t + g1t g5t+1 )/2

where f t+1 := f (t + δ).

The product f t f t+1 cancels so we get the algorithm:


 2(Ωt+1
1 − Ωt1 )/δ = (Ωt+1 t t t+1
2 Ω 3 + Ω2 Ω 3 )

 2(Ωt+1 − Ωt2 )/δ = −(Ωt+1 t t t+1 t+1
+ Γt3 )

 2 1 Ω3 + Ω1 Ω3 ) + x0 (Γ3

(Ωt+1
3 − Ωt3 )/δ = −x0 (Γt+1
2 + Γt2 )

 (Γ1t+1 − Γt1 )/δ = (Ωt+1 t t t+1 t+1 t t t+1
3 Γ2 + Ω3 Γ2 ) − (Ω2 Γ3 + Ω2 Γ3 )



 (Γ2t+1 − Γt2 )/δ = (Ωt+1 t t t+1 t+1 t t t+1
1 Γ3 + Ω1 Γ3 ) − (Ω3 Γ1 + Ω3 Γ1 )

(Γ3t+1 − Γt3 )/δ = (Ωt+1 t t t+1 t+1 t t t+1
2 Γ1 + Ω2 Γ1 ) − (Ω1 Γ2 + Ω1 Γ2 )

gt gt gt gt gt gt
where Ωt1 = f1t , Ωt2 = f2t , Ωt3 = f3t , Γt1 = f4t , Γt2 = f5t , Γt3 = f6t . These are
the “Discrete Kowalevski Equations (DKEs)”. One can explicitly solve for the
variables at time t + 1, so these equations are “explicit” and “time-reversible”
equations.
Discrete Conserved Quantities.

1. Hamiltonian: H10 = (Ωt3 )2 − H11 (Ωt1 )2 + (Ωt2 )2 − H12 (Γt1 )2
2. Angular momentum: H20 = Ωt1 Γt1 + Ωt3 Γt3 − H21 (Ωt2 Γt2 ) − H22 (Γt1 )2
3. Spatial gravity vector: H30 = (Γt3 )2 − H31 (Γt1 )2 − H32 (Γt2 )2

2
4. Kowalevski integral: H40 = (Ωt1 )2 + (Ωt2 )2 + (Γt1 )2 + (Γt2 )2 − H41 (Ωt2 Γt2 ) −
H42 (Γt1 )2 .
Hik (i = 1, 2, 3, 4; k = 1, 2) are quantities to be determined by the requirement
that the integrals are indeed invariant. For example, let’s find H32 . Since we want
H30 , H31 and H32 to be invariant, we must have:

H30 = (Γt+1 2 1 t+1 2 2 t+1 2

3 ) − H3 (Γ1 ) − H3 (Γ2 ) ,
H30 = (Γt3 )2 − H31 (Γt1 )2 − H32 (Γt2 )2 ,
H30 = (Γt−1 2 1 t−1 2 2 t−1 2
3 ) − H3 (Γ1 ) − H3 (Γ1 ) .
822 Oberwolfach Report 14/2006

Under the condition Γt+m

k 6= 0 (k = 1, 2, 3; m = −1, 0, 1), these equations
uniquely determine H3 , H3 and H32 .
0 1

H32 can be written of the form N32 /D32 in terms of Γt+1 t t−1
j , Γj , Γj :

t 2

N32 = (Γt1 )2 − (Γt−1
1 )
2
(Γ3 ) − (Γt+1
3 )
2

t+1 2

− (Γt3 )2 − (Γt−1
3 )2
(Γ 1 ) − (Γ t 2
1 )
t−1 2

t+1 2

D3 = (Γ1 ) − (Γ1 ) (Γ2 ) − (Γt2 )2
2 t 2

t+1 2

− (Γt2 )2 − (Γt−1
2 )
2
(Γ1 ) − (Γt1 )2 .

Use DKEs to eliminate Γt+1 i and Γt−1

i (i = 1, 2, 3) and express H32 in terms of
Ωti , Γti (i = 1, 2, 3) using only information at time t. This is possible and one gets
H32 = h32 (Ωt1 , Ωt2 , Ωt3 , Γt1 , Γt2 , Γt3 , x0 , δ)
In this way we can express H10 , H11 , H12 , H20 , H21 , H22 , H30 , H31 , H40 , H41 and H42 .
It is shown that H32 in the expression above is a conserved quantity of the DKEs.
This is done finding a Gröbner basis on the formula manipulation software
Risa/ASIR. This shows that H32 (t + 1) − H32 (t) = 0.
Now we found many discrete integrals. Actually, after obtaining the generators
of the ideal I, we solve those polynomials with using software Maple and we get
four independent conserved quantities H10 , H22 , H31 and H41 , the rest of Hik ’s are
dependent.

References
[1] V. Guillemin and S. Sternberg, The moment map and collective motion, Ann. of Phys. 1278
(1980), 220–253.
[2] J.E. Marsden, T.S. Ratiu, and A. Weinstein, Semidirect products and reduction in mechan-
ics, Trans. Amer. Math. Soc. 281 (1984), 147–177.
[3] J.E. Marsden, T.S. Ratiu, and A. Weinstein, Reduction and Hamiltonian structures on duals
of semidirect product Lie Algebras, Contemp. Math., Amer. Math. Soc. 28 (1984), 55–100.
[4] T.S. Ratiu, The Euler-Poisson Equations and Integrability, Ph.D. Thesis, U.C. Berkeley
(1990).
[5] T.S. Ratiu, Euler-Poisson equations on Lie algebras and the N -dimensional heavy rigid
body, Proc. Natl. Acad. Sci., USA 78 (1981), 1327–1328.
[6] T.S. Ratiu, Euler-Poisson equations on Lie algebras and the N -dimensional heavy rigid
body, Amer. J. Math. 104 (1982), 409–448, 1337.

Doing what comes naturally (learning optimal control from the

experts)
Debra Lewis
(joint work with Peter Olver)

Optimal control problems involve minimization of a cost function C subject to

constraints: if state variables x ∈ Rn satisfy the control law ẋ = v(x, u, t) for
Geometric Numerical Integration 823

some function v : Rn × Rk × R → Rn and controls u ∈ Rk , then we can seek a

control curve u : I → Rk minimizing
Z


L(γ) = C(x(t), u(t), t) + p(t)T (ẋ(t) − v(x(t), u(t), t) dt.
I
(Appropriate boundary conditions must be imposed.) The Lagrange multiplier
p ∈ Rn serves to enforce the evolution equation ẋ = v. This variational formu-
lation extends naturally to nonlinear manifolds, with the vector-valued Lagrange
multiplier p being replaced by a one-form on the nonlinear state space. However,
practical methods for determination of the extremals of such problems typically
make use of some additional structure on the manifold. We focus on the case in
which the state space is a Lie group G.
The global trivialization of the tangent and cotangent bundles of Lie groups
significantly simplifies the analysis of optimal control problems. If we right or
left trivialize the vector field for the evolution of the state variables, then the
Lagrange multipliers used in the formulation of the variational problem naturally
lie in the dual of the Lie algebra. We consider the case in which the state space
G and control space H are both Lie groups (both G and H may be vector spaces,
products of groups, etc), with evolution equation triv ġ = ω(g, h, t) for some map
ω : G×H ×R → g; here triv : T G → g denotes the algebra component of the right
(left) trivialization of the tangent vector ġ. Given a cost function C : G×H ×R →
R we define
Z
L(γ) = (C(g, h, t) + p(t) · (triv ġ(t) − ω(g, h, t))) dt,
I
where γ = (g, h, p) : I → G × H × g∗ . If we define the ‘functional derivatives’
δC ∗ δC ∗
δg : G → g and δh : H → h by
δC δC ∂C
dC(g, h, t) · (δg, δh, δt) = δg (g, h, t) · δg + δh (g, h, t) · δh + ∂t (g, h, t)δt
for all δg ∈ Tg G, g ∈ G, δh ∈ Th H, h ∈ H, δt ∈ R, t ∈ I, then
Z   
∗
dL(γ) · δγ = δp · (triv ġ − ω) + δCδg ± adω p − ṗ ·ξ
I
  
δC δω ∗ ∂C
+ δh − δh p · η + ∂t δt dt.
Here ξ : I → g and η : I → h denote the trivialized variations of g and h; the sign
of the infinitesimal coadjoint action term is determined by the choice of right or left
trivialization. If the functional derivative δC
δh is nondegenerate with respect to h,
δω ∗
then the control h for a critical curve γ is determined by the equation δCδh = δh p.
If δC
δh is degenerate, e.g. if C does not depend on the control, then optimal control
curves must be sought by other methods.
Our motivating application is the ‘falling cat’: it is well known that many an-
imals have the ability to right themselves in free-fall, following a trajectory with
trivial angular momentum. Etienne-Jules Marey, a physiologist and innovator in
high-speed photography, studied the maneuvers used by cats and other quadrupeds
to right themselves while falling. Marey (1894) published a brief and apparently
824 Oberwolfach Report 14/2006

controversial article describing a zero angular momentum cat flip achieved us-
ing internal torques and changes of the moments of inertia of the front and back
portions of the body. Kane and Scher (1969) constructed a simple, elegant math-
ematical model of the falling cat, modeling the front and back halves of the cat
as rigid bodies coupled by a joint satisfying a ‘no-twist’ condition and undergo-
ing relatively little backward bending. Montgomery (1993) translated their model
into modern geometric language, formulating the evolution equations in terms of
connections on an appropriate principal bundle and showing that the Kane-Scher
motions are the solutions of an optimal control problem closely related to geodesic
flow.
The Kane-Scher model captures some features of the flip, but the twisting and
changes of the relative moments described by Marey and others are explicitly
ruled out in their model. As a starting point for a more realistic model, we have
constructed an alternative simple flip: the cat is modeled as two axisymmetric
bodies coupled by a universal joint; the bodies are rigid during each phase of the
maneuver, but the moments of inertia differ from phase to phase. The associated
equations have solutions consisting of a series of spins about the axes of symmetry
(with corresponding counter-rotation of the other half of the body), followed in
some cases by a folding motion about the perpendicular axis; the axis of symmetry
of the front half remains in the initial vertical plane throughout the flip, and the
axes of symmetry of the front and back half are at right angles until the final phase
of the flip. Specifically:
• Start with the two halves ‘belly up’, at right angles and centered about
the vertical
• Spin the front half, counter-rotating the back half until its axis of symme-
try is perpendicular to the vertical plane
• Spin the back, counter-rotating the front downward
• Spin the front, counter-rotating the back until it returns to the vertical
plane
• If necessary, rotate both halves about the perpendicular to achieve an
overall rotation through π about the horizontal axis in the initial vertical
plane.
The relative speeds of the spins and counter-rotations are determined by the cur-
rent relative moments of inertia of the two halves.
These motions form two families, which we label odd and even (this labeling is
derived from the integers parameter determining each family):
‘Odd’ flips: The head twist overshoots in first phase and swings back in the third
phase. At the end of the first three phases, the axes of symmetry have returned
to their original positions, forming a ‘V’, with backward spinal bend and the belly
facing downward; the two halves of the body now fold inward—each half rotates
through π2 in the vertical plane. The maximum relative rotation is through 3π 2 .
‘Even’ flips: The head twist undershoots in first phase and rotates in the same
direction in the third phase. At the end of the three spin/counter-rotation phases
Geometric Numerical Integration 825

the ‘cat’ has completed the flip—no folding is necessary. The maximum relative
rotation is through 2π.
The Kane-Scher model seems too rigid—it shows just how simple a zero mo-
mentum flip can be, but it doesn’t look much like what cats actually do—while the
alternative model seems too acrobatic—the transitions are unrealistically abrupt
and the maximum relative rotations are severe. We intend to construct relatively
simple motions between these extremes, using an optimal control formulation with
several contributing costs. We soften the Kane-Scher constraints and blend the
stages of the alternative method using time-minimization augmented by penalties
for implausible behavior:
• Time: Elapsed time for flip is weighted heavily.
• Relative motion of front and back halves: Severe spinal bending and twist-
ing are penalized, but small-to-moderate relative motions are low-cost.
• Change of shape of body components: Variations of the moments of inertia
are relatively low-cost for moderate changes, expensive for large changes.
• Working blind: Inability to see the ground is penalized, but the penalty
for being moderately off the final head position is low.

References
[1] T. Kane and M. Scher, A dynamical explanation of the falling cat phenomenon, Int. J. Solids
Structures5 (1969), 663–670.
[2] E.-J. Marey, Méchanique animale: Des mouvements que certains animaux exécutent pour
retomber sur leurs pieds lorsqu’ils sont précipités d’un lieu élevé, La Nature 10 (1894), 369–370.
[3] R. Montgomery, Gauge theory of the falling cat, Fields Inst. Commun. 1 (1993), 193–218.

Lie Group Variational Integrators and its Applications to Geometric

Control Theory
Melvin Leok
(joint work with Anthony M. Bloch, Islam I. Hussein, Taeyoung Lee, N. Harris
McClamroch, Amit K. Sanyal)

The geometric approach to mechanics serves as the theoretical underpinning of

innovative control methodologies in geometric control theory. These techniques
allow the attitude of satellites to be controlled using changes in its shape, as
opposed to chemical propulsion, and are the basis for understanding the ability
of a falling cat to always land on its feet, even when released in an inverted
orientation.
Curiously, while the geometric structure of mechanical systems plays a critical
role in the construction of geometric control algorithms, these algorithms have
typically been implemented using numerical schemes that ignore the underlying
geometry. We introduce new geometric integration techniques that are applica-
ble to the construction of efficient implementations of optimal control algorithms
arising from geometric control theory.
826 Oberwolfach Report 14/2006

Generalized Galerkin Variational Integrators [7]. The most common crit-

icism leveled against the geometric integration community is that the preserva-
tion of geometric structure alone is not a sufficient criterion for a good numerical
method. Variational integrators are obtained by discretizing the action integral in
two stages, by choosing a finite-dimensional function space to discretize sections of
the configuration bundle over space-time, and by approximating the integral using
numerical quadrature. Important considerations in numerical analysis, such as
adaptivity, approximability, and accuracy can be incorporated into the context of
discrete mechanics through an appropriate choice of function space and numerical
quadrature, as discussed in, Generalized Galerkin Variational Integrators [7].

Lie Group Variational Integrators [4, 5]. Lie group variational integrators
preserve the Lie group structure of the configuration space without the use of lo-
cal charts, reprojection, or constraints. Instead, the discrete solution is updated
using the exponential of a Lie algebra element which satisfies a discrete variational
principle. These yield highly efficient geometric integration schemes for rigid body
dynamics that automatically remain on the rotation group. By representing the
attitude as a rotation matrix, we avoid coordinate singularities associated with
local charts such as Euler angles, and obtain a global representation of the con-
figuration that is particularly important in efficiently simulating chaotic orbital
motion.
These ideas were introduced in the paper, A Lie Group Variational Integrator
for the Attitude Dynamics of a Rigid Body with Applications to the 3D Pendu-
lum [4]. This is applied to a system of extended rigid bodies interacting under
their mutual gravitational potential in the paper, Lie Group Variational Integra-
tors for the Full Body Problem [5], wherein symmetry reduction to a relative frame
is also addressed.

Discrete Optimal Control on Lie Groups [1, 6, 3]. Our approach to discretiz-
ing the optimal control problem is in contrast to traditional techniques such as
collocation, wherein the continuous equations of motion are imposed as constraints
at a set of collocation points. In our approach, modeled after [2], the discrete equa-
tions of motion are derived from a discrete variational principle, and this induces
constraints on the configuration at each discrete timestep. The discrete dynamics
are more faithful to the continuous equations of motion, and consequently more
accurate solutions to the optimal control problem are obtained.
For the purpose of numerical simulation, the corresponding discrete optimal
control problem is posed on the discrete state space as a two stage discrete vari-
ational problem. In the first step, the discrete dynamics for the the rigid body
is derived from the discrete Lagrange-d’Alembert principle, as well as methods
introduced in the work on Lie group variational integrators [4]. These discrete
equations are then imposed as constraints to be satisfied by the extremal solutions
to the discrete optimal control problem, and necessary conditions for the discrete
Geometric Numerical Integration 827

extremal solution are obtained in terms of the given terminal states. This ap-
proach is described in the paper, A Discrete Variational Integrator for Optimal
Control Problems on SO(3) [1].
The paper, Attitude Maneuvers of a Rigid Spacecraft in a Circular Orbit [6],
discusses optimal control problems whereby the forces are only applied at the start
and end of the maneuver, but where the relative frame is prescribed by a nominal
trajectory, like a circular orbit. The problem is posed as a discrete optimal control
problem with constraints, which is solved directly using a sequential quadratic
programming method.
An adjoint formulation can also be adopted, wherein the forward computation
computes the trajectories, and the backward computation yields the sensitivity of
the cost functional in terms of the design parameters. This yields an efficient and
robust optimal control algorithm based on multiple shooting, as documented in
Optimal Attitude Control of a Rigid Body using Geometrically Exact Computations
on SO(3) [3].
The typical reason for numerical instability in shooting based optimization al-
gorithms is due to the difficulty in computing the sensitivities of the terminal
boundary conditions on the initial controls in a stable fashion. We exploit the
structure preserving properties of variational integrators to compute the sensitiv-
ities in a robust manner.
These ideas are applicable to the efficient construction of quantum gates, since
a quantum gate is realized by generating a trajectory in the space of unitary
operators from the identity to the desired operator through the use of external
controls, and can therefore be formulated in terms of optimal control on Lie groups.

References
[1] I. I. Hussein, M. Leok, A. K. Sanyal, and A. M. Bloch, A discrete variational integrator
for optimal control problems in SO(3), Proc. IEEE Conf. on Decision and Control (2006),
arXiv:math.OC/0509536, (submitted).
[2] O. Junge, J. E. Marsden, and S. Ober-Blöbaum, Geometric mechanics and optimal control,
Proc. of the 16th IFAC World Congress (2005), (to appear).
[3] T. Lee, M. Leok, and N. H. McClamroch, Optimal attitude control of a rigid body
using geometrically exact computations on SO(3), J. Optim. Theory Appl. (2006),
arXiv:math.OC/0601424, (submitted).
[4] T. Lee, N. H. McClamroch, and M. Leok, A Lie group variational integrator for the attitude
dynamics of a rigid body with applications to the 3D pendulum, Proc. IEEE Conf. on Control
Applications (2005), 962–967.
[5] , Lie group variational integrators for the full body problem, Comp. Meth. Appl. Mech.
Eng. (2005), arXiv:math.NA/0508365, (submitted).
[6] , Attitude maneuvers of a rigid spacecraft in a circular orbit, Proc. American Control
Conf. (2006), arXiv:math.OC/0509299, (accepted).
[7] M. Leok, Generalized Galerkin variational integrators: Lie group, multiscale, and
pseudospectral methods, (preprint), 2004, arXiv:math.NA/0508360.
828 Oberwolfach Report 14/2006

Semi-implicit semi-Lagrangian time-stepping methods and regularized

fluid equations in numerical weather prediction
Sebastian Reich
(joint work with Jason Frank, Nigel Wood, Andrew Staniforth)

The fundamental components of a numerical weather prediction (NWP) code are

provided by (i) data assimilation, (ii) dynamic core, and (iii) parameterization of
unresolved phenomena such as precipitation. The talk was primarily concerned
with numerical aspects of the dynamic core, which is provided by the inviscid
Euler equations of stratified and rotating ideal fluid dynamics. I first discussed
geometric and conservation aspects of the equations of motion and their poten-
tial preservation under numerical discretizations in space and time. I also briefly
summarized the multi-scale nature of atmospheric fluid dynamics.
In the second part of my talk, I reviewed the widely used semi-implicit semi-
Lagrangian (SISL) time-stepping method. The SISL method is, for example, used
by the UK Met Office to overcome the severe step-size restrictions due to unre-
solved waves as well as strong advection in their non-hydrostatic Unified Model.
The idea of the Unified Model is to only use unapproximated Euler equations
for the dynamic core and to have the spatial and temporal approximations se-
lect the desired spatial and temporal resolution. The practical implementation of
the Unified Model methodology poses challenging questions to the practitioners
and theoreticians alike. Our own current work focuses on an interpretation of the
semi-implicit method as a regularization of the unapproximated Euler equations
and the implementation of such a regularization within an explicit time-stepping
method.
In the final part of my talk I reported about first results on an explicit time-
staggered semi-Lagrangian method. The time-staggered formulation is based on
the regularized Störmer-Verlet time-stepping method for the Hamiltonian particle-
mesh (HPM) method as proposed in [1, 2]. It has been demonstrated that the
explicit Sörmer-Verlet time-stepping of the regularized equations is, on a linear
equation level, equivalent to the effect of semi-implicit time-stepping of the unreg-
ularized shallow-water equations [3, 4]. Numerical results were presented for the
rotating shallow water equations and a non-hydrostatic vertical slice model.
The ultimate goal is to implement the regularized time-staggered semi-Lagrang-
ian method for the three-dimensional Euler equations and to test the methods
within the Unified Model framework.

References
[1] J. Frank, G. Gottwald, and S. Reich. The Hamiltonian particle-mesh method. In M. Griebel
and M.A. Schweitzer, editors, Meshfree Methods for Partial Differential Equations, vol-
ume 26 of Lect. Notes Comput. Sci. Eng., pages 131–142, Berlin Heidelberg, 2002. Springer-
Verlag.
[2] J. Frank and S. Reich. The Hamiltonian particle-mesh method for the spherical shallow
water equations. Atmos. Sci. Let., 5:89–95, 2004.
Geometric Numerical Integration 829

[3] J. Frank, S. Reich, A. Staniforth, A. White, and N. Wood. Analysis of a regularized, time-
staggered discretization method and its link to the semi-implicit method. Atmos. Sci. Let.,
6:97–104, 2005.
[4] N. Wood, A. Staniforth and S. Reich. Improved regularization for the time-staggered dis-
cretization and its link to the semi-implicit method. Atmos. Sci. Let., in press.

Parcel Eulerian–Lagrangian fluid dynamics

Marcel Oliver
(joint work with Onno Bokhove)

1. Introduction
Conservation laws play an important role in geophysical fluid mechanics. In
the absence of forcing and dissipation, conservations laws can be derived system-
atically from the variational or Hamiltonian structure of the equations of fluid
motion. Conservation or near-conservation of mass, energy, and vorticity in the
underlying idealized model and its numerical treatment is considered desirable to
enhance stability and accurate ensemble forecasting, even though the dynamics
of the atmosphere and oceans are ultimately driven by forcing on the large scales
and subject to viscosity on the small scales.
Particle methods for fluids [6] have favorable conservation properties, as can be
shown by interpreting the dynamics of particles as a non-autonomous Hamiltonian
system with particle position and particle velocity as variables; parcels interact via
the transport of mass or vorticity [4]. Generally, the Hamiltonian associated with
a particle representation consists of a type of Bernoulli function, that is, the sum
of the Lagrangian kinetic energy plus an Eulerian potential function depending
on space, evaluated at the position of the particle, and time. It turns out that in
continuum fluid dynamics a single fluid parcel satisfies the same non-autonomous
Hamiltonian formulation, which is finite dimensional for the one distinguished fluid
parcel. The Eulerian potential is now transported by the flow generated by the
parcel velocities, thereby coupling the parcels to a continuum. We call this the
parcel Eulerian–Lagrangian formulation.
The question we ask is how such parcel Eulerian–Lagrangian formulations relate
to the well-known Lagrangian or Eulerian continuum Hamiltonian mechanics. The
answer turns out to be natural, but is not immediately obvious: The restriction
from continuum to parcel variational principle is done by choosing special varia-
tions which are concentrated on points in label space, and which are constrained
by the natural local conservation laws of the continuum formulation. Vice versa,
the continuum formulation can be recovered from the parcel formulation by a
procedure which essentially amounts to integration over label space.
We can therefore, on the one hand, derive parcel formulations, and thus natural
particle schemes, for equations of continuum mechanics by a general procedure.
On the other hand, we can translate results from one formulation to the other.
In particular, the continuum Poisson bracket inherits the Jacobi identity by con-
struction from the parcel formulation, where it is more easily verified.
830 Oberwolfach Report 14/2006

2. Continuum vs. parcel formulation for shallow water

The rotating shallow water equation are known to be Hamiltonian with
Z


(1) H= 2 1
h |u|2 + g (h + b)2 dx

and possesses the Poisson formulation dF /dt = {F , H} with bracket

Z Z  
δF ⊥ δG δF δG δG δF
(2) {F, G} h dx = q · − ·∇ + ·∇ dx ≡ {F, G} ,
δu δu δu δh δu δh
where q = (f + ∇⊥ · u)/h denotes the potential vorticity. Inserting appropriate
test functionals, it is easy to derive the Eulerian equations of motion,
(3a) ∂t u + u · ∇u + f u⊥ + g ∇(h + b) = 0 ,
(3b) ∂t h + ∇ · (h u) = 0 .
In the parcel formulation, on the other hand, each “fluid parcel” is treated as an
independent, two degree of freedom nonautonomous Hamiltonian system, where
1

(4) H(X, U , t) = |U |2 + g h(X, t) + b(X) .
2
and Poisson formulation dF/dt = {F, H} with bracket
⊥
(5) {F, G} = f ∇U F · ∇U G + ∇X F · ∇U G − ∇X G · ∇U F .
The resulting parcel equations of motion are
dX
(6a) = ∇U H = U ,
dt
dU ⊥
(6b) = −f ∇U H − ∇X H = −f U ⊥ − g ∇(h + b) ,
dt
where the dynamics of the parcels is linked through the mass transport equation
Z


(6c) h(X, t) = h0 (a) δ X − χ(a, t) da ,
D
which is not part of the parcel variational structure, but rather plays the role
of an external potential. This parcel formulation readily yields the equations of
smoothed particle hydrodynamics [6] upon replacing h0 by a finite sum of point
masses and the δ-distribution in (6c) by a mollifier or blob function. Note that the
continuum Hamiltonian H is not the parcel Hamiltonian H integrated over label
space—there is no factor 21 in front of the parcel potential energy.
To pass to the parcel formulation, we start with the continuum Hamilton prin-
ciple and note that the continuity equation acts as a constraint. (The formalism
could be cast into the abstract framework of Euler–Poincaré reduction for semi-
direct products [5].) Letting η denote the flow map, so that χ̇ = u ◦ χ, we must
take variations of the shallow water action with respect to the flow map. This cor-
responds to an Eulerian variation w via δχ = w ◦ χ, and the continuity equation
implies that δh + ∇ · (wh) = 0.
Geometric Numerical Integration 831

The key idea is to restrict, under the above constraint, the class of variations
to
(7) δχ(a, t) = δ(a − A) δX ,
where δ( · ) denotes the Dirac mass. In a second step, we identify the restricted
variation of the full action integral as the total variation of a per-parcel action
integral, which then yields the parcel Hamiltonian structure. We can also reverse
this process by integrating the parcel Poisson formulation over label space, which
allows us, for example, to prove the Jacobi identity for the continuum Poisson
bracket by merely verifying the parcel Poisson bracket; for details, see [1].

3. Continuum vs. parcel formulation for ideal fluids

The procedure applies to any fluid equation with an advected scalar, such as
the compressible Euler equations in any dimension [1]. Of particular interest are
the incompressible Euler equations in two dimensions, where vorticity plays the
role of the advected scalar. The continuum Hamiltonian is the well known kinetic
energy Hamiltonian,
Z
1
(8) H= |u|2 dx ,
2
the vorticity Poisson bracket reads
Z    
⊥ δF δG
(9) {F , G} = ω ∇ ·∇ dx ,
δω δω
and the vorticity equations of motion are
(10a) ∂t ω + u · ∇ω = 0 ,
(10b) ω = ∇⊥ · u ,
(10c) ∇·u = 0.
The parcel formulation, on the other hand, has the nonautonomous Hamiltonian
(11) H(X, t) = −ψ(X, t) ,
Poisson bracket
(12) {F, G} = ∇⊥ F · ∇G ,
and corresponding parcel equations of motion
dX
(13a) = −∇⊥ H = ∇⊥ ψ ,
dt
(13b) ∆ψ = ω ,
Z


(13c) ω(X, t) = ω0 (a, t) δ X − χ(a, t) da .
D
Replacing ω0 by a finite sum of delta distributions yields the familiar point vortex
dynamics, and additional smoothing of the Dirac delta in the transport integral
will lead to a numerical vortex method [2].
832 Oberwolfach Report 14/2006

In this case, naively restricting the Hamilton principle to the vorticity transport
law results in a tautology. The key observation which allows us to restrict the
continuum variational principle in a nontrivial way is that, when starting from
an extended Lagrangian formulation, the particle relabeling symmetry yields a
vorticity-like conservation law without assuming a bijection between momenta and
velocities. This conservation law is then used to restrict the extended Lagrangian
and the construction can proceed as for the shallow water equations; for details
see [1].
Our procedure provides, in particular, a correspondence between the Hamil-
tonian formulation for ideal fluids, and the Hamiltonian formulation of 2D point
vortex dynamics.

References
[1] O. Bokhove and M. Oliver, Parcel Eulerian–Lagrangian fluid dynamics of rotating geophysical
flows, Proc. R. Soc. Lond. A, in press.
[2] A. Chorin, Numerical study of slightly viscous flow, J. Fluid Mech. 57 (1973), 785–796.
[3] J. Frank, G. Gottwald, and S. Reich, A Hamiltonian Particle-Mesh Method for the Rotating
Shallow-Water Equations, Lecture Notes in Computational Science and Engineering, Vol. 26
(2002), 131–142.
[4] J. Frank and S. Reich, Conservation properties of smoothed particle hydrodynamics applied
to shallow water equations, BIT 43 (2003), 40–54.
[5] D.D. Holm, J.E. Marsden, and T.S. Ratiu, Euler-Poincaré equations and semidirect products
with applications to continuum theories, Adv. in Math. 137 (1998), 1–81.
[6] J.J. Monaghan, Smoothed Particle Hydrodynamics, Annu. Rev. Astron. Astrophys. 30 (1992),
543–574.

Properties of exponential integrators for nonlinear wave equations

Brynjulf Owren
(joint work with Håvard Berland, Bård Skaflestad)

Semilinear PDEs have the form

(1) ut = Lu + N (u)
where L is a linear unbounded operator and N is a nonlinear map. We consider
exponential integrators of the form
r−1
X
Un,r = ecr hL un + h ajr (hL)N (Un,j ), r = 1, . . . , s
j=1
s
X
un+1 = ehL un + h br (hL)N (Un,r )
r=1

The coefficient functions arj (z) and br (z) are assumed to be

(1) entire functions in the complex plane
(2) bounded on the imaginary axis
Geometric Numerical Integration 833

Sometimes we also required that φ(iy) ≤ C|y|−q for y ∈ R. Rigorous conditions

for order of convergence of exponential integrators applied to semilinear PDEs
have been derived by Hochbruck and Ostermann [4]
The cubic Schrödinger equation in 1D on a periodic domain is
(2) iut + uxx = (λ|u|2 + V (x))u
We treat separately the case when V (x) ≡ 0. We consider the Sobolev space H s
as the functions f (x) such that
X Z π
1
(1 + k ) |f (k)| < ∞ where fˆ(k) =
2 s ˆ 2
f (x) e−ikx dx
2π −π
k∈Z

It follows from the a paper by Bourgain [2] that the problem (2) with V (x) ≡ 0 is
well-posed for every s ≥ 0. Moreover, as pointed out by Christ et al. [3], for non-
negative integers k, one has that supt ku(t)kH k is bounded by a constant which
depends only on k and the H k -norm of the initial datum.
One easily sees that for s > 12 , u ∈ H s implies that |u|2 u ∈ H s . From this fact,
we observe that any exponential integrator as defined above will map un ∈ H s to
un+1 ∈ H s whenever s > 12 . For a nonvanishing V (x) one may have the situation
that the corresponding N (u) in (1) maps H s → H σ for σ < s. It then follows
that the exponential integrators maps H s to H s if s − σ < min{1, 2q} if all the
coefficient functions have decay rate q.
Finally, one may consider nonlinear wave equations of the form
wtt = Lw + f (w)
which may be transformed into a system of the form (1) by using the vector
variable u = (w, wt ). One then finds that if f : H s → H s , then the exponential
integrator maps elements of H s × H s−1 to H s × H s−1 .
In general it may be hard to prove that the numerical approximation obtained
by an exponential integrator is bounded in H s independently of the stepsize, even
for finite time intervals. However, such bounds have been proved for linearised
versions of the equations, and numerical evidence seems to indicate that this may
also be valid for the 1D nonlinear Schrödinger equation. Given that such a bound
exists, one may combine the above mentioned smoothness properties with the
nonstiff order conditions derived in [1] to prove order of convergence results for
the exponential integrators expressed in terms if smoothness of the data.
A Matlab toolbox which implements a wide range of exponential integrators
for many examples of PDEs is available for free at
http://www/math.ntnu.no/num/expint/
This package also facilitates easy addition of new problems and integrators.

References
[1] H. Berland, B. Owren, B. Skaflestad B-series and order conditions for exponential integra-
tors SIAM J. Numer. Anal. 43 (2005), 1715–1727.
834 Oberwolfach Report 14/2006

[2] J. Bourgain, Fourier Transform Restriction Phenomena for Certain Lattice Subsets and Ap-
plications to Nonlinear Evolution Equations, Geometric and Functional Analysis 3 (1993),
107–156.
[3] M. Christ, J. Colliander and T. Tao, Instability of the periodic nonlinear Schrödinger equa-
tion, Preprint, 2003.
[4] M. Hochbruck, A. Ostermann, Explicit exponential Runge-Kutta methods for semilinear
parabolic problems SIAM J. Numer. Anal. 43 (2005), 1069–1090,

Numerical Methods for Multiscale Problems

Assyr Abdulle

Consider uε , the solution of a (partial) differential equation

(1) Lε uε = f ε ,
where Lε is a linear differential operator for which ε indicates the small scales
in the coefficients. Due to the oscillations in these coefficients, uε will typically
exhibit small scale oscillations. The scale gap between the length of the oscillation
and the phenomena of interest (large scale behavior) makes direct simulation of
(1) with a standard numerical method often impossible. A common approach is
to compute an effective or homogenized operator L̄, in which the small scales have
been averaged out and to solve the corresponding homogenized equation (see [9]
and references therein)
(2) L̄ū = f¯.
The problem (2) can be solved with standard methods. However, this procedure
has several drawbacks. First, restrictive assumptions on the data (periodicity, ho-
mogeneity) are needed to derive explicit equations for the homogenized problem 2 .
Second, the coefficients of the homogenized equation have usually to be computed
numerically so that a control of the overall procedure (i.e. the numerical dis-
cretization of the homogenized equations with numerically computed coefficients)
is difficult. Third, the fine scale behavior, i.e. the oscillations of the solution, is
lost in the homogenization process.
In this report we discuss several numerical methods for multiscale problems con-
structed within the framework of the Heterogeneous Multiscale Methods (HMM)
[1], [2], [3], [4], [5], [6], [7], [8]. These methods discretize the physical problem
directly by a “macroscopic numerical model” with a macroscopic discretization.
The input coefficients of the macroscopic numerical model are unknown, since
the macroscopic model is not supposed to be known (we do not precompute an
averaged equation as (2)). These coefficients are recovered on the fly by solv-
ing a “microscopic numerical model” on sampling domains within the macro dis-
cretization. A variety of micro-macro approaches based on iterative schemes, have
been proposed in the literature, mainly for elasticity problems, including nonlinear
problems (see [5] for a discussion and references). The iterative procedure in these
approaches increases the cost of the methods and makes a full error control of
these micro-macro schemes difficult even for linear problems.
Geometric Numerical Integration 835

Heterogeneous multiscale methods for variational problems. We consider

the following elliptic model problem in the domain Ω ⊂ Rd
(3) −∇ · (aε ∇uε ) = f in Ω, uε = 0 on ∂Ω,
where we assume that the tensor aε (x) = a(x, xε ) = a(x, y) is symmetric, coercive
and periodic with respect to each component of y in the unit cube Y = (0, 1)d . We
further assume that f ∈ L2 (Ω), aij (x, ·) ∈ L∞ (Rd ), that x → aij (x, ·) is smooth
from Ω̄ → L∞ (Rd ) and that Ω is a convex polygon. The FE-HMM for the elliptic
homogenization problem, based on the macro finite element (FE) space S01 (Ω, TH )
(TH is the macro triangulation) is defined by a modified macro bilinear form [3],[8],
X |K| Z
H H
(4) B(u , v ) = ∇uh a(xk , x/ε)(∇v h )T dx,
|Kε | Kε
K∈TH

where Kε = xk +ε[−1/2, 1/2]d is a sampling sub-domain centered at the barycenter

xk of the triangle K, where |K|, |Kε | denote the measure of K and Kε , respectively,
and where uh is the solution of the following micro problem: find uh such that
(uh − uH ) ∈ Wper
1
(Kε ) and
Z
(5) ∇uh a(xk , x/ε)(∇z h )T dx = 0 ∀z h ∈ Sper
1
(Kε , Th ).
Kε
1 1
For Kε ⊂ K ∈ TH , we consider the micro FE space Sper (Kε , Th ) ⊂ Hper (Kε )/R
of piecewise linear polynomials on the micro triangulation Th , periodic on the
boundary ∂Kε . The meshsizes of the macro and micro spaces are denoted by H
and h, respectively. The following convergence estimates have been obtained in
[3] for the fully discrete FE-HMM
 h
2   h
2 
0 H 0 H 2
(6) ku − u kH 1 (Ω) ≤ C H + , ku − u kL2 (Ω) ≤ C H + ,
ε ε
 h
ε ε
(7) ku − up kH̄ 1 (Ω) ≤ C H + ε + ,
ε
where uε is the solution of problem (3), u0 is the solution of the homogenized
problem corresponding to (3) and uεp is a reconstructed solution obtained from
uH with fine scale solution (u − uH ) periodically extended on each element K
(see [8],[1]). For the estimate (7), the norm H̄ is mesh dependent since uεp can be
discontinuous across the macro elements K. Similar fully discrete estimates have
also be derived for the effective velocity in transport problems [4] and for elasticity
problems [5].
Heterogeneous multiscale methods for dynamic problems. Consider the
parabolic homogenization problem
∂uε
(8) = ∇ · (aε ∇uε ) in (0, T ) × Ω
∂t
(9) uε = 0 on (0, T ) × ∂Ω, uε (0, x) = g(x) on Ω,
where uε = uε (t, x), Ω ⊂ Rd is a bounded domain, and aε (x) is symmetric, uni-
formly coercive and bounded. We further assume that aε (x) and g(x) are regular
836 Oberwolfach Report 14/2006

enough in order to have a smooth solution of the above problem. In the sequel we
describe the algorithm in one dimension for simplicity (see [1] for generalizations).
A coarse model for (8) is given by ∂U 0 0
∂t = ∇ · (a ∇U ), where a (the homogenized
tensor) reflects the large scale impact of aε . As for the previous problem, we do not
assume that a0 is known. The finite difference heterogeneous multiscale method
(FD-HMM) [1] is defined by a macroscopic scheme on a coarse grid {xi }N i=0 of Ω
k k
Ji+1/2 + Ji−1/2
(10) Uik+1 = Uik + ∆t
∆x
with meshsize ∆x Rand evolved with a coarse time step ∆t. The unknown fluxes
Ji±1/2 = |K ε±1 | K ε aε ∇ûε dx are given by the average of micro solutions ûε
xi±1/2 xi±1/2

obtained by solving small scale problems on sampling domains Kxεi±1/2 of size ε

centered around xi±1/2 for a small time [tk , tk + δ]
∂ ûε
(11) = ∇ · (aε ∇ûε ) (t, x) ∈ (tk , tk + δ) × Kxεi±1/2 ,
∂t
(12) ûε − U k (x) ε−periodic on (tk , tk + δ) × ∂Kxεi±1/2 , û(tk , x) = U k (x),
where δ is a relaxation time (see [1]). The initial values at the cells Kxεi±1/2 are
obtained by linear reconstruction of the known solution {Uik }N k
i=0 at time t .
For the convergence analysis [1] we have to introduce an intermediate problem.
Let {Uik }N k
i=0 be the solution of the FD-HMM at time t = t0 + k∆t and let
k N
{Ūi }i=0 be the solution of a finite difference method similar to (10) but with
fluxes J¯i±1/2 = (a0 (Ūi±1/2 − Ūi ))/∆x. Then it can be shown [1]
ε
max |Uik − Ūik | ≤ CT ∀tk ∈ [0, T ].
0≤i≤N ∆x
{Ūik }N
i=0 can be seen as a standard FD method for the homogenized problem, for
which standard convergence results, provided enough smoothness of the solution,
give maxi |U (tk , xi ) − Ūik | ≤ CT (∆t + (∆x)2 ) ∀tk ∈ [0, T ]. Combining both re-
sults, we obtain the following error estimate for the FD-HMM when compared to
the homogenized solution of problem (8)
 ε 
max |U (tk , xi ) − Uik | ≤ CT ∆t + (∆x)2 + ∀tk ∈ [0, T ].
0≤i≤N ∆x
Here, the exact solution for the micro problem (11) has been assumed. For a fully
discrete analysis, the discretization of this problem with a micro spatial mesh with
meshsize δξ has to be considered. The term δξ ε is expected to appear [6].

References
[1] A. Abdulle and W. E, Finite difference HMM for homogenization problem, J. Comput.
Phys. 191 (2003), pp. 18–39.
[2] A. Abdulle and C. Schwab, Heterogeneous multiscale FEM for diffusion problem on rough
surfaces , SIAM Multiscale Model. Simul. 3 (2005), pp. 195–220.
[3] A. Abdulle, On a-priori error analysis of fully discrete heterogeneous multiscale FEM, SIAM
Multiscale Model. Simul. 4 (2005), pp. 447-459.
Geometric Numerical Integration 837

[4] A. Abdulle, Multiscale methods for advection- diffusion problems, Discrete Contin. Dyn.
Syst. B, Suppl. Vol. (2005), pp. 11-21.
[5] A. Abdulle, Analysis of a heterogeneous multiscale FEM for problems in elasticity, Math.
Mod. Meth. Appl. Sci. (M3AS) 16 (2006), pp.1-21.
[6] A. Abdulle, Finite volume heterogeneous multiscale method, in preparation.
[7] W. E and B. Engquist, The heterogeneous multiscale methods, Comm. Math. Sci. 1 (2003)
pp. 87-132.
[8] W. E, P. Ming and P. Zhang, Analysis of the heterogeneous multi-scale method for elliptic
homogenization problems, J. Amer. Math. Soc. 18 (2004), pp. 121-156.
[9] V.V. Jikov, S.M. Kozlov and O.A. Oleinik, Homogenization of differential operators and
integral functionals, Springer-Verlag, Berlin, Heidelberg, 1994.

Structural Preserving Isospectral Flows for Quadratic Pencils

Moody T. Chu
(joint work with Nicoletta Del Buono)

The eigeninformation (λ, u) of the quadratic pencil,

(1) Q(λ) := Q(λ; M0 , C0 , K0 ) = λ2 M0 + λC0 + K0 ,
is critical to the understanding of the dynamical system
(2) M0 ẍ + C0 ẋ + K0 x = f (t),
which arises frequently in many important applications, including applied mechan-
ics, electrical oscillations, vibro-acoustics, fluid mechanics, and signal processing.
It is easy to see that the linear pencil,
   
C0 M0 K0 0
(3) L(λ) := L(λ; M0 , C0 , K0 ) = λ+ ,
M0 0 0 −M0
in the so called Lancaster structure, is equivalent to Q(λ). Recently, it has been
shown that for almost all quadratic pencils there exists real-valued 2n × 2n real
matrices Πℓ and Πr such that
   
⊤ CD MD KD 0
(4) Πℓ L(λ)Πr = L(λ; MD , CD , KD ) = λ+ ,
MD 0 0 −MD
where MD , CD , KD are all real-valued n × n diagonal matrices [2, 3, 4]. Such a
transformation is significant in that it links the dynamical behavior of a multiple-
degree-of-freedom system directly to that of a system consisting of n independent
single-degree-of-freedom subsystems. It breaks down the interlocking connectiv-
ity in the original system into totally disconnected subsystems while preserving
the entire spectral properties. Thus it will be of great value in practice if the
transformations Πℓ and Πr can be found from any given pencil. The theory of
existence of Πℓ and Πr in [2, 3] was established on the basis of the complete spec-
tral information of L(λ). To construct Πℓ and Πr from the availability of spectral
information certainly is impractical. The emphasis of this talk is to construct Πℓ
and Πr numerically by structure preserving isospectral flows without knowing the
spectral information.
838 Oberwolfach Report 14/2006

Denote
   
ℓ11 ℓ12 r11 r12
(5) Πℓ = , Πr = ,
ℓ21 ℓ22 r21 r22
where each ℓij or rij is an n × n matrices. In order to maintain the Lancaster
structure in the transformation Π⊤
ℓ L(λ)Πr , it is necessary that the following five
equations hold:
−ℓ⊤ ⊤
11 K0 r12 + ℓ21 M0 r22 = 0,
−ℓ⊤ ⊤
12 K0 r11 + ℓ22 M0 r21 = 0,
(6) ℓ⊤
12 C0 r12 + ℓ⊤
22 M0 r12 + ℓ⊤
12 M0 r22 = 0,
ℓ⊤
11 C0 r12 + ℓ⊤
21 M0 r12 + ℓ⊤
11 M0 r22 = ℓ⊤ ⊤ ⊤
12 C0 r11 + ℓ22 M0 r11 + ℓ12 M0 r21
= −ℓ⊤ ⊤
12 K0 r12 + ℓ22 M0 r22 .

Additionally, the matrices Πℓ and Πr must be such that the left-hand sides of the
following three expressions,
−ℓ⊤ ⊤
12 K0 r12 + ℓ22 M0 r22 = MD ,
(7) ℓ⊤
11 C0 r11 + ℓ⊤
21 M0 r11 + ℓ⊤
11 M0 r21 = CD ,
ℓ⊤
11 K0 r11 − ℓ⊤
21 M0 r21 = KD ,
are diagonal matrices. The conditions (6) and (7) together constitute a homoge-
neous second-degree polynomial system of 8n2 − 3n equations in 8n2 unknowns. It
is not obvious how the system could be solved analytically. The underdetermined
system does suggest, however, that there is plenty of leeway to choose the transfor-
mation matrices Πℓ and Πr . In particular, the “orbit” of L(λ) under (Lancaster)
structure preserving equivalence transformations is a nontrivial manifold on which
perhaps a smooth path connecting (M0 , C0 , K0 ) to (MD , CD , KD ) can be defined.
To characterize the orbit, denote the Lancaster pair in (3) by (A0 , B0 ) where
   
K0 0 C0 M0
(8) A0 = , B0 = .
0 −M0 M0 0
We are interested in developing two one-parameter families of structured preserv-
ing transformations TL (t), TR (t) ∈ R2n×2n , with TL (0) = TR (0) = I2n . Consider
the actions of these families of transformations on (A0 , B0 ) by
(9) A(t) = TL⊤ (t)A0 TR (t), B(t) = TL⊤ (t)B0 TR (t),
respectively. Clearly, regardless how TL (t) and TR (t) are defined, (A(t), B(t)) is
isospectral to (A0 , B0 ) for any t. It appears sufficient to limit ourselves to a special
class of transformations where matrices TL (t) and TR (t) satisfy, respectively,
 
dTL (t) L11 (t) L12 (t)
(10) = TL (t)L(t) = TL (t) ,
dt L21 (t) L22 (t)
 
dTR (t) R11 (t) R12 (t)
(11) = TR (t)R(t) = TR (t) ,
dt R21 (t) R22 (t)
Geometric Numerical Integration 839

where each Lij (t) or Rij (t), i, j = 1, 2, is a n × n real one-parameter matrix yet to
be defined. Upon substitution, we observe from (9) that
dA
= ṪL⊤ A0 TR + TL A0 ṪR = L⊤ A + AR,
dt
dB
= ṪL⊤ B0 TR + TL B0 ṪR = L⊤ B + BR.
dt
It is interesting to note that these differential equations are similar to those dis-
cussed in [1] which leads to a Lie-Poisson system. By insisting that (A(t), B(t))
maintains the Lancaster structure, that is,
   
K(t) 0 C(t) M (t)
(12) A(t) = , B(t) = ,
0 −M (t) M (t) 0
we see that the entries of L(t) and R(t) should satisfy:
(13) R12 = −DM,
(14) R21 = DK,
(15) L12 = D⊤ M ⊤ ,
(16) L21 = −D⊤ K ⊤ ,
(17) L11 − L22 = D⊤ C ⊤ ,
(18) R11 − R22 = −DC,
where D ∈ Rn×n is an arbitrary matrix parameter. Also hidden implicitly in (17)
and (18) are the other two free matrix parameters. There are several possible ways
to choose the parameters and to arrange the diagonal blocks of L(t) and R(t). For
instance, corresponding to the choice,
 ⊤  " C⊤ ⊤
# 
⊤

D 0 2 M N L 0
(19) L = ⊤ + ,
0 D⊤ −K ⊤ − C2 0 NL⊤
  C   
D 0 − 2 −M NR 0
(20) R = C + ,
0 D K 2
0 NR
an isospectral flow of the triplet (M (t), C(t), K(t)) can be defined by the au-
tonomous system:
1
K̇ = (CDK − KDC) + NL⊤ K + KNR ,
2
(21) Ċ = (M DK − KDM ) + NL⊤ C + CNR ,
1
Ṁ = (M DC − CDM ) + NL⊤ M + M NR .
2
Furthermore, by assuming NR (t) = NL (t), the symmetry specified for the matrix
parameter D in Table 1 will preserve the symmetry for the flow (M (t), K(t), C(t))
defined by the dynamical system (21).
The remaining task is to control the free matrix parameters in such a way that
the structure preserving isospectral flow (A(t), B(t)) is also a gradient flow for a
certain properly selected objective function. Our idea is based on the scenario that
840 Oberwolfach Report 14/2006

D(t) M (t) C(t) K(t)

skew-symmetric symmetric symmetric symmetric
symmetric symmetric skew-symmetric symmetric
symmetric skew-symmetric skew-symmetric skew-symmetric
skew-symmetric skew-symmetric symmetric skew-symmetric
Table 1. Preserving symmetries of (M (t), C(t), K(t)) by D(t),
if NR (t) = NL (t).

given a sufficiently smooth objection function f : Rn → R whose state variable

x ∈ Rn is constrained to the integral curve of ẋ = g(x)u starting with x(0) = x0
where u(t) is the control, perhaps one way we can do is to choose the control u so
that the vector ẋ is as close to −∇f (x) as possible. This amounts to the selection
of the least squares solution u defined by
(22) u(t) = −g(x(t))† ∇f (x(t)),
where g(x)† stands for the Moore-Penrose generalized inverse of g(x). In this way,
the closed-loop dynamical system,
(23) ẋ = −g(x)g(x)† ∇f (x),
defines a descent flow x(t) for the objective function f (x). This scenario fits us
perfectly if we seek matrix parameters NR , NL and D to minimize the function,
1
f (K, C, M ) := koffdiag(M )k2F + koffdiag(C)k2F + koffdiag(K)k2F
2
(24) + δh(diag(M ), diag(C), diag(K)),
subject to the condition that (M (t), C(t), K(t)) is governed by the differential
system (21). Our purpose in designing the objective function is to minimize the
off-diagonal entries of (M, C, K) while using the function h to monitor the behavior
of the diagonal entries by a factor of δ. The free matrix parameters D, NL and
NR are used as controls to direct the flow and can be obtained as the least squares
solution to the equation:
1    ∂h

2 (K ⊗ C − C ⊗ K) K ⊗ I I ⊗ K vec(D) vec(−offdiag(K) − δ ∂K )
 K ⊗ M − M ⊗ K C ⊗ I I ⊗ C  vec(NL⊤ )= vec(−offdiag(C) − δ ∂h ) 
∂C
1 ∂h
2 (C ⊗ M − M ⊗ C) M ⊗ I I ⊗ M vec(N R ) vec(−offdiag(M ) − δ ∂M )
Since the gradient flow can be tracked by available ODE integrator, it is feasible
for numerical computation. Computer simulations seem to suggest the working of
this approach.

References
[1] A. M. Bloch and A. Iserles, On an isospectral Lie-Poisson system and its Lie algebra,
DAMTP Tech. Rep. 2005/NA01, Found. Comp. Math., to appear.
[2] M. T. Chu and N. Del Buono, Total decoupling of a general quadratic pencil, Part I: Theory,
preprint, 2005.
Geometric Numerical Integration 841

[3] S. D. Garvey, M. I. Friswell, and U. Prells, Co-ordinate tranfromations for second order
systems, I: General transformations, J. Sound Vibration, 258(2002), 885–909.
[4] S. D. Garvey, M. I. Friswell, and U. Prells, Co-ordinate transformations for second or-
der systems, II: Elementary structure-preserving transformations, J. Sound Vibration, 258
(2002), 911–930.
[5] I. Gohberg, P. Lancaster, and L. Rodman, Matrix Polynomials, Computer Science and
Applied Mathematics. Academic Press, Inc., New York-London, 1982.

A numerical method for Hamiltonian systems based on piecewise

smooth space-approximations
Philippe Chartier
(joint work with Erwan Faou)

Consider a Hamiltonian system


q̇ = ∇p H(q, p),
(1)
ṗ = −∇q H(q, p),

where (q, p) ∈ Rd × Rd , and with a separable Hamiltonian H of the form

1 T
H(q, p) = p p + V (q),
2
where V (q) is the potential function. In many applications, such as for instance
molecular dynamics, it is of importance that the numerical flow used to compute
the solution of (1) preserves the volume form and the Hamiltonian. However, it is
generally admitted that no standard method can satisfy both requirements, apart
from exceptional situations such as for instance a quadratic Hamiltonian. In the
case of multi-quadratic potentials as considered by R. Quispel and R.I. McLachlan
in [5], a possible approach consists in solving in sequence the d Hamiltonian systems
with Hamiltonians
1 2 1X T
H [i] (qi , pi ) = pi + V [i] (qi ) + p̄j p̄j ,
2 2
j6=i
[i]
V (qi ) = V (q̄1 , . . . , q̄i−1 , qi , q̄i+1 , . . . , q̄d ) ,
obtained by freezing all components (denoted with a bar) except the two conjugate
coordinates qi and pi . If each subsystem is then solved exactly and the same
step-size is used for all, the resulting “numerical” method preserves the desired
quantities, since each sub-step is symplectic and preserves H [i] (and thus H).
In order to retain the possibility of solving exactly each sub-system and at the
same time to cover more general situations, we give up the requirement of exact
Hamiltonian preservation and we consider a multi-quadratic piecewise approxima-
tion of H. If instead of (1) we now solve

q̇ = ∇p H τ (q, p),
(2)
ṗ = −∇q H τ (q, p),
842 Oberwolfach Report 14/2006

where H τ (q, p) = 12 pT p + V τ (q) is a C 1,1 multi-quadratic approximation of H,

the aforementioned procedure applied with exact solution of the sub-systems gives
a first-order method which preserves H τ exactly as well as the volume form. If
supK |H − H τ | ≤ CK τ 2 for a compact subset K of Rd × Rd containing the
numerical solution, then H is conserved up to an error of size τ 2 over arbitrarily
long intervals of integration (including infinite ones).
In this talk, we describe the implementation of the method with quadratic
B-splines only for the case of separable Hamiltonians and prove the following
properties of the flow of piecewise polynomial Hamiltonian systems:
Theorem 1. Let H be a continuously differentiable scalar function defined on R2d
such that f = J −1 ∇H is Lipschitz over the whole space R2d and consider the flow
ϕt associated with f . Then, for a fixed t ∈ R, ϕt satisfies the following properties:
• (iii) for any y ∈ R2d , H(ϕt (y)) = H(y), that is to say ϕt is Hamiltonian-
preserving.
• (iv) ϕt is a.e. differentiable on R2d .
• (vi) (ϕ′t )T Jϕ′t = J a.e. on R2d .
• (vii) |det(ϕ′t )| = 1 a.e. on R2d .
An explicit expression of the exact solution is given for the two-dimensional
case that serves as a basis for higher dimensions and the numerical scheme used
here is shown to be of order 1 and becomes an order 2 method when composed
with its adjoint, though in a slightly weaker sense than usual:
Theorem 2. Let ϕt be the exact flow of the system (1) and ϕτi,t the exact flow
of the Hamiltonian system with Hamiltonian H [i,τ ] . The numerical flow Φτh as
defined above with stepsize h > 0 and space discretization parameter τ is of the
form Φτh = ϕτ1,h ◦ . . . ϕτd,h and satisfies the following estimate for all Lipschitz
function g with compact support:
(3) |hϕh − Φτh |gi| ≤ C(hτ + h2 kgkL1 )
for a constant C depending on V , and for sufficiently small h and τ .
If the systems for the H [i,τ ] are solved for i= 1, .. . , d and then for i = d, . . . , 1
∗
in reverse order, the resulting method Φτh/2 ◦ Φτh/2 is symmetric and
 ∗
(4) |hϕh − Φτh/2 ◦ Φτh/2 |gi| ≤ C(hτ + h3 kg ′ kL1 ).

Eventually, we present numerical results for three different test problems, for
which the usual behaviour of symplectic integrators is exhibited and no resonances
occur: the Kepler problem, Fermi-Pasta-Ulam mechanical system and the Sine-
Gordon equation.

References
[1] C. Le Bris and P.L. Lions, Renormalized solutions of some transport equations with partially
w 1,1 velocities and applications, Ann. Mat. Pura Appl. 1 (2004), 97–130.
Geometric Numerical Integration 843

[2] P. Chartier, E. Faou, and A. Murua, An algebraic approach to invariant preserving in-
tegrators: The case of quadratic and hamiltonian invariants, to appear in Numer. Math.
(2005).
[3] L. C. Evans and R.F. Gariepy, Measure theory and fine properties of functions, CRC Press,
1992.
[4] E. Hairer, C. Lubich, and G. Wanner, Geometric numerical integration. structure-preserving
algorithms for ordinary differential equations, Springer Series in Computational Mathemat-
ics 31, Springer, Berlin, 2002.
[5] R. I. McLachlan and G. R. W. Quispel, Geometric integration of conservative polynomial
ODEs, Applied Numerical Mathematics 45 (2003), 411–418.
[6] R.J. Di Perna and P.L. Lions, Ordinary differential equations, transport theory and Sobolev
spaces, Invent. Math. 3 (1995), 511–547.
844 Oberwolfach Report 14/2006

Integral–preserving Integrators
G.R.W. Quispel1,2
(joint work with D.I. McLaren1 )

1 Department of Mathematics, La Trobe University, Melbourne, Victoria 3086, Australia

2 ARC Centre of Excellence for Mathematics and Statistics of Complex Systems
Many systems of differential equations possess so–called geometric properties, for
example one or more first integrals, symplectic structure, volume preservation,
and others. Numerical solution of such systems is usually best effected using a
geometric integrator, i.e. one that can preserve this property exactly (i.e. up to
round–off accuracy) [1, 2, 3].
In this extended abstract we are particulary interested in integral–preserving inte-
grators (IPIs) which are designed to provide exact preservation of any first integral
possessed by a system of ordinary differential equations (ODEs), e.g. momentum,
angular momentum, energy, etc. We concentrate here on IPIs using so–called
discrete gradients [4, 5, 6, 7, 8].
Consider thus an ODE
dx
(1) = f (x) x ∈ Rn
dt
which possesses a first integral I(x), i.e.
dI(x)
(2) =0
dt
provided x satisfies (1).
Under conditions that are generally satisfied, (1) and (2) are equivalent to the
existence of a skew matrix S(x) such that (1) can be rewritten∗
dx
(3) = S(x)∇I(x).
dt
An IPI for (3) is given by
x′ − x ¯
(4) = S̄(x, x′ )∇I(x, x′ ).
h
Here h denotes the time step, and
x := x(nh), x′ := x((n + 1)h),
¯ denote a “discrete skew matrix” resp a “discrete gradient”.
and S̄ and ∇
Discrete gradients must satisfy
¯
I(x′ ) − I(x) =: ∇I(x, x′ ) · (x′ − x)
Discrete skew matrices must be skew, otherwise the only conditions on S̄ and ∇ ¯
are that they must be consistent, i.e. in the limit h → 0, i.e. x′ → x, they must
go to S resp. ∇.
∗
Note that in general S(x) does not satisfy the Jacobi identity.
Geometric Numerical Integration 845

¯ are not at all unique. Many examples of discrete gradients

It follows that S̄ and ∇
have been constructed [5, 6], here we give two in R2 :
 I(x′ ,x2 )−I(x1 ,x2 ) 
1
x′1 −x1
¯ 1 I(x, x′ ) := 
∇ 


I(x′1 ,x′2 )−I(x′1 ,x2 )
x′2 −x2

Here x =: (x1 , x2 ), x′ =: (x′1 , x′2 ).

A second example is given by
¯ ′ ¯ ′
¯ 2 I(x, x′ ) := ∇1 I(x, x ) + ∇1 I(x , x)
∇
2
In general, the above will give a numerical integration method of first–order accu-
racy.
There are various ways to obtain IPIs of higher order of accuracy. One method that
we have developed constructs a method of order n + 1 from a method of order n.
For the sake of simplicity we here only give the construction of a second–order IPI
from a first–order one:
x′ − x ¯
Let = S1 ∇I(x, x′ ) be a first–order IPI, with S1 := S = S(x), then a
h
second–order IPI is given by
x′ − x ¯
= S2 ∇I(x, x′ ),
h
where  
1
S2 := S + h SQS + C .
2
Here C and Q are skew matrices given by:
∂S ij ml ∂I
C ij := S
∂xm ∂xl
and
1 T
Q := (B − B),
2
where B is defined by


¯ =: ∇I + B(x) · (x′ − x) + O kx′ − xk2
∇
Similarly one obtains higher–order IPIs.
Concluding, we remark that
(1) Multiple integrals are preserved similarly.
(2) We are currently studying N –dimensional systems with N − 1 integrals.
(3) Some generalizations of the above ideas to PDEs exist.
Acknowledgement: This work was supported by the Australian Research Council.
846 Oberwolfach Report 14/2006

References
[1] E. Hairer, C. Lubich and G. Wanner, Geometric Numerical Integration, 2nd edition, Springer
(2006).
[2] B. Leimkuhler and S. Reich, Simulating Hamiltonian Dynamics, CUP, (2004).
[3] R.I. McLachlan and G.R.W Quispel (eds), Special Issue on Geometric Integration, J.Phys.A.
(2006) in press.
[4] G.R.W. Quispel and H.W. Capel, Solving ODEs numerically while preserving a first integral,
Phys.Lett. 218A (1996), 223–228.
[5] G.R.W. Quispel and G.S. Turner, Discrete–gradient methods for solving ODEs numerically
while preserving a first integral, J.Phys. A29 (1996) L341–L349.
[6] R.I. McLachlan, G.R.W. Quispel and N. Robidoux, Geometric integration using discrete
gradients, Phil.Trans.Roy.Soc.A 357 (1999) 1021–1045.
[7] D.I. McLaren and G.R.W. Quispel, Integral–preserving integrators, J.Phys.A 37 (2004)
L489–L495.
[8] D.I. McLaren and G.R.W. Quispel, Bootstrapping the order of discrete–gradient integral–
preserving integrators, preprint, (2006).

Splitting Methods in Geometric Numerical Integration

Sergio Blanes
(joint work with Fernando Casas, Ander Murua)

A widely used technique in the geometric numerical integration of differential

equations is the splitting idea: if the differential equation can be split into two or
more parts that are either solvable or simpler to integrate than the original system,
it is also possible to build integrators by composition of the corresponding flows.
They are frequently used in celestial mechanics, quantum mechanics, molecular
dynamics, accelerator physics and, in general, for solving numerically Hamiltonian
systems, as well as Poisson systems and reversible differential equations [6, 7, 8, 9].
They can be adapted to many different problems and frequently can be combined
in a natural way with other techniques to improve their performances. In addition,
many problems can be split in many different ways leading to systems differing
in the following aspects: (i) the geometric properties which are preserved; (ii)
the cost for evaluating each part; (iii) the magnitude of the error when using a
given method; (iv) the alternative techniques to combine with splitting (modified
vector fields, processing, etc.); (v) the families of splitting methods which can be
used. Then, the knowledge of the algebraic structure a given problem in many
cases helps to choose highly efficient methods from the literature, or to build new
methods tailored for a given family of problems. This is well illustrated by the
time-dependent Schrödinger equation
 
∂ 1 2
(1) i ψ(x, t) = − ∇ + V (x) ψ(x, t).
∂t 2µ
After a semidiscretization, we have to solve the following linear system of ODEs
d
(2) i c(t) = Hc(t), c(0) = c0 ∈ CN
dt
Geometric Numerical Integration 847

with formal solution c(t) = e−itH c0 and H = T + V ∈ RN ×N is an Hermit-

ian matrix. Fourier methods are frequently used to approximate the solution
because (Vc)i = V (xi )ci is computed with N products and Tc = F −1 DT F c,
where F is the fast Fourier transform (FFT) and DT is a diagonal

matrix, can be
τV τ V (xk )
computed with O(N log N ) operations. Notice that e c k = e ck , k =
0, 1, . . . , N − 1 where the exponentials are computed only once at the beginning of
the integration and they are stored, so their cost can be safely neglected. Similarly,
for the kinetic part we have that eτ T c = F −1 eτ DT F c. Then, standard splitting
methods, for a time step, τ = −i△t, and preserving unitarity read as follows
m
Y
(3) Un (τ ) = F −1 ebj τ DT F eaj τ V .
j=1

In addition, for this family of problems we can also consider that the commutator,
[V, [T, V]], is a diagonal matrix. Since the cost of the algorithm (3) is dominated
by the FFTs, we can add these terms to the potentials, without increasing the
cost, to improve the performance. Higher order terms can also be considered as
well as new terms can be added to the kinetic energy in case the potential is a
low degree polynomial. In addition, the processing technique is very easy to use
in this problem. It decompose the method Un (τ ) as follows: Un = UP UK U−1 P
so, for p steps Upn = UP UpK U−1 P , where U K is the kernel and U P is the processor
or corrector (see [1, 2] and references therein).
Alternatively, Eq. (2) can also be split in two parts with a completely different
structure, allowing to use different techniques and methods. Consider c = q + ip,
then Eq. (2) can be rewritten as a classical Hamiltonian system with Hamiltonian:
H = 12 pT Hp+ 12 qT Hq = A+B. This splitting has a very particular property with
respect to the Poisson brackets: {A, {A, {A, B}}} = {B, {B, {B, A}}} = 0. This
property drastically reduces the number of order conditions allowing to reach high
order methods with a relatively small number of stages, preserving symplecticity.
This problem illustrates how flexible the splitting methods are and the many
tools which can be used to obtain efficient methods. A large number of methods
have recently appeared in the literature, and tailored for many different families
of problems. At this point, we make a review about the last fifteen years for
some of the most popular families of methods: (i) Composition of second order
symmetric methods; (ii) Composition of a first order methods with its adjoint,
which is equivalent to splitting in two general parts; (iii) splitting Nyström methods
when the kinetic energy is quadratic in momenta; (iv) splitting Nyström methods
using modified potentials; (v) Near-integrable systems; and (vi) Linear separable
oscillatory systems with the property {A, {A, {A, B}}} = {B, {B, {B, A}}} = 0.
A great effort in the search of methods has been done by many different authors
leading to highly efficient methods for all these families of problems. Most reviewed
methods can also be found in [2, 4, 5, 6, 7, 8, 9].
If we return to the classical representation of the Schrödinger equation, we find
that it fits in case (vi) (as well as many other important oscillatory linear systems
like e.g. the Maxwell equations). The best methods from the previous review
848 Oberwolfach Report 14/2006

appear mainly in [5]. However, using the processing technique and due to the
structure of this problem, we noticed that it is very easy to build methods at any
order. In addition, it is possible to consider additional stages in the composition to
improve simultaneously both accuracy and stability [3, 4], a fact very unusual for
general problems. Following these procedures we build methods up to order twenty
which outperform the previous schemes. We also build a new family of second order
methods whose coefficients are chosen in order to get both large stability domain
and high accuracy (making use of Chevishev techniques), and this occurs even
for very large time step. Some of these methods reach round off accuracy nearly
for all time steps inside the stability domain. These new methods have nearly
the same stability as the symmetric second order decomposition (which attain the
maximum stability among splitting methods of this family) and outperform all
other schemes for all time steps and accuracies desired.
The analysis is carried out for the harmonic oscillator, and this toy problem
allowed us to build highly efficient methods for non trivial problems like the
Schrödinger equation. These methods are tested in the one-dimensional problem
(1) with a Morse potential showing the performance expected from the theoretical
analysis.
As a result, we conclude: (I) Splitting methods are powerful methods for many
problems; (II) It is very important to use the most appropriate methods for each
problem; and (III) In some cases we can also build methods tailored for particular
problems which outperform the existing methods from the literature.

References
[1] S. Blanes, F. Casas, and A. Murua, On the numerical integration of ordinary differential
equations by processed methods, SIAM J. Numer. Anal. 42 (2004), pp. 531–552.
[2] S. Blanes, F. Casas, and A. Murua, Composition methods for differential equations with
processing, SIAM J. Sci. Comp. 27 (2006), 1817–1843.
[3] S. Blanes, F. Casas, and A. Murua, Symplectic splitting operator methods for the time-
dependent Schrödinger equation, J. Chem. Phys. (2006). In press.
[4] S. Blanes, F. Casas, and A. Murua, Linear stability of splitting methods (2006). In prepara-
tion.
[5] S. Gray and D.E. Manolopoulos, Symplectic integrators tailored to the time-dependent
Schrödinger equation, J. Chem. Phys. 104 (1996), 7099–7112.
[6] E. Hairer, C. Lubich, and G. Wanner, Geometric Numerical Integration. Structure-
Preserving Algorithms for Ordinary Differential Equations, Springer Ser. Comput. Math.
31, Springer-Verlag, Berlin (2002).
[7] B. Leimkuhler and S. Reich, Simulating Hamiltonian systems, Cambridge University Press,
2004.
[8] R.I. McLachlan and R. Quispel, Splitting methods, Acta Numerica 11 (2002), 341–434.
[9] J.M. Sanz-Serna and M.P. Calvo, Numerical Hamiltonian Problems, Chapman & Hall, Lon-
don (1994).
Geometric Numerical Integration 849

Splitting methods for the harmonic oscillator

Ander Murua
(joint work with Sergio Blanes, Fernando Casas)

We consider the application of splitting methods to linear problems of the form

(1) q̇ = M p, ṗ = N q.
Such linear systems arise when space discretization of different linear partial differ-
ential equations, for instance, the time-dependent Schrödinguer equation [2, 4, 1].
Studying the behaviour of splitting methods applied to such linear systems is also
of interest as a mean to understand the stability of splitting methods applied to
non-linear systems of the form
(2) q̇ = g(p), ṗ = f (q).
Note that a linear system of the form (1) is obtained after linearization around a
stationary point of (2). In this sense, it is of interest the construction of splitting
methods with extended stability properties [3].
Application of a splitting method to the system (1) is equivalent to approxi-
mating the exponential  
0 hM
exp
−hN 0
by means of a product of the form
     
I 0 I ham M I 0 I ha1 M
K(h) = ··· ,
−hbm N I 0 I −hb1 N I 0 I
with appropriately chosen values of the real parameters ai , bi .
The stability of K(h) is related to the stability polynomial p(x) of the splitting
method [3, 2, 4], which is a polynomial in even powers of x that approximates
cos(x).
We show that, for a given even polynomial p(x), there exists a finite number
of symmetric splitting methods (that is, a finite number of different choices of the
parameters {ai , bi : 1 ≤ i ≤ m}) that have p(x) as a stability polynomial, and
we give a step by step algorithm that explicitly give all such methods.
We construct families of stability polynomials with extended stability intervals
and very good accuracy properties that allow, together with the aforementioned
algorithm, the construction of very stable and efficient splitting methods with
processing for the numerical solution of linear systems of the form (1).

References
[1] S. Blanes, F. Casas, and A. Murua, Symplectic splitting operator methods tailored for the
time-dependent Schrödinger equation, to appear in J. Chem. Phys. (2006)
[2] S. Gray and D.E. Manolopoulos, Symplectic integrators tailored to the time-dependent
Schrödinger equation, J. Chem. Phys. 104 (1996), pp. 7099–7112.
[3] M.A. López-Marcos, J.M. Sanz-Serna, and R.D. Skeel, An explicit symplectic integrators
with maximal stability interval, in: D.F. Griffiths and G.A. Watson, eds., Numerical Analy-
sis, A.R. Mitchel 75th Birthday Volume (World Scietific, Singapore, 1996) 163–176.
850 Oberwolfach Report 14/2006

[4] R.I. McLachlan and S.K. Gray, Optimal stability polynomials for splitting methods, with ap-
plications to the time-dependent Schrödinger equation, Appl. Numer. Math. 25, 275 (1997).

Using additivity in numerical integration of DAEs

Laurent O. Jay

We report on extensions of Gauss methods and the Hilber-Hughes-Taylor (HHT)

method for differential-algebraic equations (DAEs) arising in mechanics.
Gauss methods have not been considered as having much practical interest for
the numerical solution of DAEs. This has been mainly due to poor convergence
properties of these methods when applied in a standard direct way. Stiffly accurate
methods have generally better convergence properties to solve DAEs [1, 3, 5].
However, in the context of geometric numerical integration of ODEs, e.g., for
Hamiltonian or Lagrangian systems, Gauss methods have excellent theoretical
properties [2]. It was an open problem and a question of interest to know if there
was any way of generalizing and developing methods based on Gauss coefficients
that would lead to methods of high order and quality for DAEs while preserving
geometric properties of DAEs. It was shown recently that such generalizations are
indeed possible for DAEs of index 2 [6, 10] and index 3 [7, 8].
Let us consider DAEs of the form

(1a) y ′ = v(t, y, z),

(1b) (p(t, y, z))′ = f (t, y, z) + r(t, y, ψ),
(1c) 0 = g(t, y),

where in mechanics the quantities y, v, p, f , and r represent respectively general-

ized coordinates, generalized velocities, generalized momenta, generalized forces,
and reaction forces due to the holonomic constraints g(t, y) = 0 (r(t, y, ψ) =
−gyT (t, y)ψ). Differentiating (1c) once one obtains additional velocity constraints

(1d) 0 = gt (t, y) + gy (t, y)v(t, y, z).

The matrices pz (t, y, z) and gy (t, y)vz (t, y, z)p−1

z (t, y, z)rψ (t, y, ψ) are assumed to
be invertible. The above formulation generalizes both Hamiltonian (p = z) and
Lagrangian systems (v = z) with holonomic constraints. Applied to the system (1)
one step with stepsize h of the standard 1-stage Gauss RK method is divergent in
general even when p(t, y, z) = z and r(t, y, ψ) is linear in the algebraic variables ψ.
Geometric Numerical Integration 851

The modified method based on the implicit midpoint rule that we propose reads

h
Y1 = y0 + v(T1 , Y1 , Z1 ),
2
h h
p(T1 , Y1 , Z1 ) = p(y0 , z0 ) + f (T1 , Y1 , Z1 ) + r(t0 , y0 , Ψ0 ),
2 2
y1 = y0 + hv(T1 , Y1 , Z1 ),
0 = g(t1 , y1 ),
h h
p(t1 , y1 , z1 ) = p(t0 , y0 , z0 ) + hf (T1 , Y1 , Z1 ) + r(t0 , y0 , Ψ0 ) + r(t1 , y1 , Ψ1 ),
2 2
0 = gy (t1 , y1 )v(t1 , y1 , z1 ),

where h is the stepsize, T1 := t0 + h/2, and t1 := t0 + h. We call this method

the (1, 1)-Gauss-Lobatto specialized partitioned additive Runge-Kutta (SPARK)
method. It mixes coefficients from the midpoint rule with those from the trape-
zoidal rule to treat holonomic constraints properly. It makes use of the additivity
of the differential equations (1b). Note that the quantity Ψ0 above is an internal
algebraic variable, it is not an initial condition or a value ψ0 coming from the pre-
vious step. We can generalize these ideas to obtain higher order schemes based on
Gauss coefficients mixed with Lobatto coefficients. The family of Gauss-Lobatto
SPARK methods makes great use of the partitioning and additive structure of the
equations (1). We summarize our findings in the following theorem [7, 8]:
Theorem 1: For the overdetermined system of DAEs (1) the (s, s)-Gauss-
Lobatto SPARK methods are constraint-preserving, symmetric, and of maximal
order 2s. For holonomically constrained Hamiltonian systems and Lagrangian
systems these methods are also symplectic and variational.
The HHT method is widely used in structural dynamics [4]. The HHT method
for y ′′ = f (t, y, y ′ ) or equivalently for

y ′ = z, z ′ = f (t, y, z),

can be expressed as an implicit non-standard one-step method as follows

h2
y1 = y0 + hz0 + ((1 − 2β)a0 + 2βa1 ) ,
2
z1 = z0 + h ((1 − γ)a0 + γa1 ) ,
a1 = (1 + α)f (t1 , y1 , z1 ) − αf (t0 , y0 , z0 ),

where a0 and a1 are approximations to the acceleration a(t) := f (t, y(t), z(t)) at
t0 + αh and t1 + αh respectively. The coefficients α, β, γ are taken according to
α ∈ [−1/3, 0], β = (1 − α)2 /4, γ = 1/2 − α. We have extended the HHT method
to the DAEs (1) with v = p = z, i.e., to DAEs of the form

(2) y ′ = z, z ′ = f (t, y, z) + r(t, y, ψ), 0 = g(t, y), 0 = gt (t, y) + gy (t, y)z.

852 Oberwolfach Report 14/2006

Given (y0 , z0 , a0 ) we define the extended HHT method for (2) as follows
h2 h2
(3a) y1 = y0 + hz0 + ((1 − 2β)a0 + 2βa1 ) + ((1 − b)R0 + bR1 ),
2 2
h
(3b) z1 = z0 + h ((1 − γ)a0 + γa1 ) + (R0 + R1 ) ,
2
(3c) a1 = (1 + α)f (t1 , y1 , z1 ) − αf (t0 , y0 , z0 ),
where b 6= 1/2 is a new free coefficient,
(3d) R0 = r(t0 , y0 , Ψ0 ), R1 = r(t1 , y1 , Ψ1 ),
and Ψ0 is not a value ψ0 coming from the previous step or an initial condition,
but Ψ0 and Ψ1 are internal algebraic variables determined by the two sets of
constraints
(3e) 0 = g(t1 , y1 ), 0 = gt (t1 , y1 ) + gy (t1 , y1 )z1 .
Once again we make use of the additivity of the differential equations for z ′ in
(2). To make the method less implicit, one can replace R1 by r(t1 , y0 + hz0 , Ψ1 )
in (3d). Theorem 2 below remains valid in this situation. One can show global
convergence of order 2 of the extended HHT method [9]:
Theorem 2: Consider the overdetermined system of DAEs (2) with initial
conditions (y0 , z0 , a0 ) at t0 satisfying
g(t0 , y0 ) = 0, gt (t0 , y0 ) + gy (t0 , y0 )z0 = 0, a0 − a(t0 + αh) = O(h).
Then the numerical solution (yn , zn , an ) at tn to the system of equations (3) sat-
isfies for 0 ≤ h ≤ h0 and tn − t0 = nh ≤ Const
yn − y(tn ) = O(h2 ), zn − z(tn ) = O(h2 ), an − a(tn + αh) = O(h2 ),
where (y(t), z(t)) is the exact solution to (1) at t passing through (y0 , z0 ) at t0 .

References
[1] E. Hairer, Ch. Lubich, and M. Roche, The numerical solution of differential-algebraic sys-
tems by Runge-Kutta methods, Lect. Notes in Math. 1409, Springer, Berlin, (1989).
[2] E. Hairer, Ch. Lubich, and G. Wanner, Geometric numerical integration, Springer, Berlin,
Comput. Math. 31, (2006).
[3] E. Hairer and G. Wanner, Solving ordinary differential equations II. Stiff and differential-
algebraic problems, Springer, Berlin, Comput. Math., Vol. 14, (1996).
[4] H. M. Hilber, T. J. R. Hughes, and R. L. Taylor, Improved numerical dissipation for time in-
tegration algorithms in structural dynamics, Earthquake Eng. and Struct. Dynamics, Math.
Comput., 5, (1977), 283–292.
[5] L. O. Jay, Symplectic partitioned Runge-Kutta methods for constrained Hamiltonian sys-
tems, SIAM J. Numer. Anal. 33, (1996), 368–387.
[6] L. O. Jay, Specialized Runge-Kutta methods for index 2 differential algebraic equations,
Math. Comput., 75, (2006), 641–654.
[7] L. O. Jay, Beyond conventional Runge-Kutta methods in numerical integration of ODEs and
DAEs by use of structures and local models, J. Comput. Appl. Math., (2006), to appear.
[8] L. O. Jay, Symplectic specialized partitioned additive Runge-Kutta methods for conservative
systems with holonomic constraints, Technical report, Dept. of Math., Univ. of Iowa, USA,
(2006).
Geometric Numerical Integration 853

[9] L. O. Jay and D. Negrut, Extensions of the HHT method to differential-algebraic equations
in mechanics, Technical report, Dept. of Math., Univ. of Iowa, USA, (2006).
[10] A. Murua, Partitioned Runge-Kutta methods for semi-explicit differential-algebraic systems
of index 2, Technical report, EHU-KZAA-IKT-196, Univ. of the Basque country, Spain,
(1996).

Explicit, volume preserving splitting methods for divergence-free

polynomial vector fields
Antonella Zanna
(joint work with Robert I. McLachlan, Hans Z. Munthe-Kaas and G. R. W.
Quispel)

In this talk we address the problem of the numerical integration of divergence-

free vector fields by volume-preserving methods. It is well known that devising
methods which preserve volume is quite a hard task. To-date the general tech-
niques consist in splitting the given vector field into the sum of two-dimensional
volume-preserving systems and solve those by symplectic methods (Feng Kang)
or by solving for n − 1 variables and then correct for the last one in order to ob-
tain a volume-preserving method (Shang Zai-jiu, 1994, Quispel 1995). However,
these methods are generally implicit and expensive. Explicit methods exist for
particular problems, for instance trigonometric ones.
We address the case of polynomial vector fields.
The main idea is to split the given divergence free vector field into the sum of
pieces that
• Can be integrated exactly easily, or
• Can be integrated in a volume preserving manner by simple explicit meth-
ods (a.k.a. Forward Euler) and whose adjoint is also explicit.
Then, the basic split terms can be combined to obtain higher order integrator,
either by Yoshida’s technique, or by other symmetric composition methods.
Several new methods are presented for linear and quadratic problems. Roughly,
these can be divided into two classes: i) methods that distinguish the diagonal part
(all the terms in equation i that include the variable xi , for i = 1, 2 . . . , n), and ii)
methods that do not distinguish the diagonal part. The diagonal part is generally
more difficult to treat as its coefficient are interconnected as a result of volume-
preservation, however it is computationally less expensive, as the off-diagonal part
requires computations of a order of n higher.
Among the methods for the diagonal part, we mention the splitting in d shears,
where d is the degree of the polynomial vector field, and exponentiation. As for the
off-diagonal part, we consider splitting in strictly lower triangular systems (as these
can be integrated in a volume-preserving manner by any Runge–Kutta method) by
permutations, as well as splitting in n natural shears, which are integrated exactly
by a step of Forward Euler. As for methods that do not distinguish the diagonal
part, we consider a splitting in n + d shears. The splitting in shears has been
854 Oberwolfach Report 14/2006

successfully used in the context of Hamiltonian integration in particle accelerator

physics.

Stability analysis of symplectic methods

Zaijiu Shang

We study linear stability of numerical methods for Hamiltonian systems. For ana-
lytic methods, the linear stability can be analyzed in the framework of Dahlquist,
and a well known result is that symmetric analytic methods are I-stable. Note that
an analytic symplectic method is always I-stable because it is symmetric when ap-
plying to linear Hamiltonian systems. As a consequence, Symplectic Runge-Kutta
methods are I-stable. For non-analytic methods (e.g., partitioned methods), the
linear stability was only analyzed for very few symplectic integrators of lower order
in literature [1, 2, 6].
We give a careful analysis of the linear stability of symplectic Euler methods of
arbitrarily high order. We prove that the stability interval of a symplectic Euler
method shrinks as the order increases. The limit of the interval, as the order
approaches infinity, is [− π2 , π2 ], which is just the connected component containing
the origin of the set on which the cosine function takes non-negative real values.
The first order symplectic Euler method has stability interval [−2, 2], which is
biggest in the class of symplectic Euler methods.
We also analyze the linear stability of other partitioned symplectic methods. For
Lobatto IIIA-IIIB Pairs, we find that the stability set is bounded and is expanding
as the order increases. The number of connected components of the stability set
of the Lobatto IIIA-IIIB Pairs is finite and equal to m − 1, where m is the order
of the methods. The limit of the set is the whole real line. The Lobatto IIIA-IIIB
Pairs have much better stability than symplectic Euler methods.
We also explore some relationships between linear stability and nonlinear sta-
bility of symplectic methods for stable Hamiltonian systems.
The detailed results about this topic are referred to [5] and [7]. For nonlinear
stability analysis of symplectic methods, see [3, 4].

References
[1] E. Hairer, C. Lubich and G. Wanner, Geometric Numerical Integration—Structure-
Preserving Algorithms for Ordinary Differential Equations, Springer Series in Computational
Mathematics 31, Sprnger-Verlag, New York, 2002.
[2] M. Z. Qin and M. Q. Zhang, Multi-stage symplectic schemes of two kinds of Hamiltonian
system for wave equation, Computers Math. Applic. 19 (1990), 51-62.
[3] Z. J. Shang, KAM theorem of symplectic algorithms for Hamiltonian systems, Numer. Math.
83, (1999), 477-496.
[4] Z. J. Shang, Resonant and Diophantine step sizes in computing invariant tori of Hamilton-
ian systems,Nonlinearity 13, (2000), 299-308.
[5] Z. J. Shang, stability analysis of symplectic methods, preprint, AMSS, CAS, (2006).
[6] D. L. Wang, Some aspects of Hamiltonian systems and symplectic difference methods, Phys-
ica D 73, (1994),1-16.
Geometric Numerical Integration 855

[7] Y. Wang, Pendulum Equation and Small Twist Version of the KAM Theorem, Master thesis
(in Chinese), AMSS, CAS, (2004).

Efficient strong integration of linear stochastic systems

Simon J.A. Malham
(joint work with Gabriel Lord, Anke Wiese)

We are interested in designing efficient numerical schemes for the strong ap-
proximation of linear Stratonovich stochastic differential equations driven by two
Wiener processes and with non-commutative vector fields. One solution to such
systems is the Neumann series obtained by successive Picard iteration. The log-
arithm of the Neumann expansion is known as the Magnus expansion, and this
provides an alternative solution ansatz. For ease of exposition, we also specialize
to the still non-trivial case when the coefficient matrices are constant.
Classical numerical schemes such as the Euler-Maruyama and Milstein methods
correspond to truncating the stochastic Taylor expansion to generate global strong
order 1/2 and order 1 schemes, respectively. Numerical schemes based on deter-
ministic Runge–Kutta methods have also been derived—see Kloeden & Platen
(1999) and Talay (1995). At the linear level, the Neumann, stochastic Taylor and
Runge–Kutta type methods are equivalent. In the stochastic context, Magnus
integrators have been explored by Castell & Gaines (1995), Burrage (1999) and
Misawa (2001).
We will present numerical schemes based on truncated Neumann and Magnus
expansions. Higher order multiple Stratonovich integrals are approximated across
each time-step by their expectations conditioned on the increments of the Wiener
processes on suitable subdivisions (see Gaines & Lyons 1997).
We prove that for linear SDEs driven by two Wiener processes:
(1): Superior accuracy is provided for a given stepsize by a modified class of
order one Magnus integrators.
(2): Accuracy of all stochastic integrators asymptotically scales like the
square-root of the computational cost for small stepsizes.
Our first statement reflects that the exponential of the Magnus series is a natural
solution ansatz for linear SDEs. Magnus integrators should therefore be preferable
to classical stochastic numerical schemes or Neumann integrators provided the cost
of computing the matrix exponential is not significant.
Our second statement naturally arises in the time-ordered integration of infor-
mation generated at infinitesimally small scales by the two driving Wiener signals.
In particular, for small stepsizes the accurate representation of the Lévy area
(chordal area process) between the two Wiener processes dominates the compu-
tational cost for all methods of order one and higher. Coincidentally, half-order
methods, which do not require the Lévy area, also obey this square-root scaling.
856 Oberwolfach Report 14/2006

There are several potential sources of cost contributing to the overall computa-
tional effort of a stochastic numerical integration scheme. The main ones are the
efforts associated with:
(A): Quadrature: the accurate representation of multiple integrals;
(B): Evaluation: computing and combining the individual terms and special
functions such as the matrix exponential.
In the fixed time-step scenario, stochastic numerical schemes of order 1/2 do not
invoke any quadrature effort because they only involve increments of the Wiener
processes across each time-step. When two or more Wiener processes are present,
to obtain a higher order stochastic integrator we need to include the Lévy area. In
the variable time-step scenario, order 1/2 integrators do not necessarily converge
to the correct solution (see Gaines & Lyons 1997) and a successful integrator must
include the Lévy area (see Lyons 1998, who proves that the solution is continuously
controlled by the driving processes and the Lévy area).
Our research/presentation is outlined as follows. We start by reviewing conver-
gence results for the Neumann and Magnus expansions. Then we present the Neu-
mann and Magnus integrators for linear systems. In particular we show that high
order Magnus stochastic integrators are simple and concise, and that both Neu-
mann and Magnus integrators can be expressed using a computationally favourable
basis of high order stochastic multi-dimensional integrals. We compute the global
truncation errors associated with both types of methods, and use these to reduce
each method to their canonical forms. We compare the local truncation errors
for the Magnus and Neumann integrators, and the non-trivial extension to the
corresponding global comparison. We prove that the order 1/2 Magnus integrator
is globally more accurate than the order 1/2 Neumann integrator for the same
stepsize. Further, we prove that a given modified order 1 Magnus integrator is
globally more accurate than the corresponding Neumann scheme.
In the second part, we show how to approximate multiple Stratonovich integrals
by their conditional expectations (Gaines & Lyons 1997). For numerical methods
up to global strong order 2, the multiple Stratonovich integrals approximated in
this way are only single sums and therefore relatively cheap to compute. We show
that the accurate representation of the Lévy area dominates the quadrature effort
resulting in the square-root scaling law between the global error and computational
effort. Then we present numerical experiments that reflect our theoretical predic-
tions, and in particular the superior accuracy of Magnus methods. Also we apply
both the Neumann and Magnus integrators to a stochastic Riccati differential sys-
tem that can be linearized. Lastly we discuss the implications of our conclusions
on Neumann and Magnus integrators to dynamic programming applications such
as filtering problems.

References
[1] P.M. Burrage, Runge–Kutta methods for stochastic differential equations, Ph.D. thesis, Uni-
versity of Queensland (1999).
Geometric Numerical Integration 857

[2] F. Castell & J. Gaines, An efficient approximation method for stochastic differential equa-
tions by means of the exponential Lie series, Mathematics and computers in simulation 38
(1995), 13–19.
[3] J.G. Gaines & T.J. Lyons, Variable step size control in the numerical solution of stochastic
differential equations, SIAM J. Appl. Math. 57(5) (1997), 1455–1484.
[4] P.E. Kloeden & E. Platen, Numerical solution of stochastic differential equations, Springer
(1999).
[5] T. Lyons, Differential equations driven by rough signals, Rev. Mat. Iberoamericana 14(2)
(1998), 215–310.
[6] T. Misawa, A Lie algebraic approach to numerical integration of stochastic differential
equations, SIAM J. Sci. Comput. 23(3) (2001), 866–890.
[7] D. Talay, Simulation and numerical analysis of stochastic differential systems, In Proba-
bilistic Methods in Applied Physics (ed. P. Kree and W. Wedig). Lecture Notes in Physics,
vol. 451, chap. 3 (1995), pp. 63–106.

Semi-Lagrangian methods and new integrators for convection

dominated problems
Elena Celledoni

In this talk we consider nonlinear convection diffusion problems with a domi-

nating convection term. For example of the type,
∂
(1) u(x, t) + V · ∇u(x, t) = ν∇2 u + f (x),
∂t
with x ∈ Ω ⊂ Rd and V : Rd × [0, T ] → R is a given vector field, u : Rd × [0, T ] →
Rd , and u(x, 0) = u0 (x). We can distinguish between two important special cases
of the above equation. The first case arises when u is a conserved passive scalar,
f (x) = 0, u can represent temperature, or salt concentration in the water, V is
known a priori. The second case is given by the Navier–Stokes equations where u
is a vector field, V = u, and f (x) = −∇p, with the constraint ∇ · u = 0.
The case when the parameter ν tends to zero is particularly interesting and
very challenging from the numerical point of view.
If we semidiscretize (1) in space with a finite element or finite difference method
we obtain a system of ordinary differential equations of the type
(2) yt − C(y)y = Ay + f, y(0) = y0 .
Here C is the discretized convection operator, A corresponds to the linear diffusion
term, often negative definite.
A typical approach for solving numerically the semidiscretized equations is to
treat convection and diffusion separately, the diffusion with an implicit approach
and the convection with an explicit integrator possibly of higher order, see for
example [1], [3].
We propose here a new class of integration methods particularly suited for
convection diffusion problems. These methods are exponential integrators and
their peculiarity is that they allow for the computation of exponentials of the
linearized convection term of (2). This means that terms of the form W (h) · g =
858 Oberwolfach Report 14/2006

exp(γhC(w)) · g, where w and g are fixed vectors in Rn , and γ is a parameter of

the integration method, appear as building blocks for the methods.
A simple example is the following first order integrator for (2),
(3) yn+1 = exp(hC(yn ))yn + hAyn+1 + hf.
The goal here is to present and analyze higher order methods which share some
of the features of (3). The main reason for developing this type of methods is that
as it turns out they can be applied to the numerical integration of the considered
PDEs in a semi-Lagrangian fashion. The main challenge in the numerical approx-
imation of convection dominated phenomena is to avoid the occurence of spurious
oscillations in the numerical solution, (numerical dispersion), without adding too
much artificial diffusion. This task is achieved nicely by semi-Lagrangian meth-
ods, [Pi], [HE], [RM], [2]. In these methods linear convective terms are integrated
exactly by computing first the characteristics corresponding to the grid points of
the adopted discretization, and then producing the numerical approximation via
a simple although expensive interpolation procedure.
The exponential exp(γhC(w)) · g is the solution of the semidiscretized equation
(4) v ′ = C(w)v, v(0) = g, in [0, h],
which corresponds to the pure convection problem
γt + V · ∇γ = 0, γ(xi , 0) = gi , in [0, h] × Ω, i.e.
(5) Dγ
Dt = 0, γ(xi , 0) = gi , in [0, h] × Ω,
D·
where xi are the nodes of the space discretization, V = V(xi ) = wi , and Dt is the
total derivative. If we apply (3) directly to the problem (1) before discretizing in
space, and we make the exponential exp(hC(yn ))yn correspond to the exact inte-
gration of a pure convection problem, we obtain the following transport-diffusion
algorithm studied in [Pi],
Dun+ 1
Dt
2
= 0, un+ 12 (x, tn ) = un (x), on [tn , tn + h]
(6) un+ 12 (x) = un+ 12 (x, tn + h)

un+1 = un+ 21 + hν∇2 un+1 + hf,

the convecting vector field is V(x) = un (x). The exact integration of the pure
convection problem in (6) can be obtained by introducing characteristics, we get
un+ 21 (x) = un+ 12 (x, tn + h) = un (X(tn ))
(7) dX
dτ = un (X(τ )), X(tn + h) = x,
and the equation for the characteristics X(τ ) must by integrated backwards in
time, either exactly or with a suitable numerical integrator. Now substituting the
result in the second equation of (6) we have
dX
dτ = un (X(τ )), X(tn + h) = x,
(8)
un+1 = un (X(tn )) + hν∇2 un+1 + hf.
Geometric Numerical Integration 859

The outlined correspondence between (3) and (8) motivates the use of exponen-
tials of the linearized semidiscrete convection operator of (2) in the integrators.
This work is also related to the methods presented in [MPR], and [XK] for the
disretization of the Navier-Stokes equations. Previous preliminary work has also
been presented in [5].
If we consider the semidiscretized ordinary differential equation

(9) ẏ − C(y)y = Ay, y(0) = y0 ,

where for the sake of simplicity, but without loss of generality we have dropped the
forcing term f compared to (2). The general format for the considered methods is
for i = 1 : s do
Pi
Yi = ϕi yn + h j=1 ai,j ϕi ϕ−1
j AYj
P k P
ϕi = exp(h k αiJ C(Yk )) · · · exp(h k αki1 C(Yk ))
end P
yn+1 = ϕn+1 yn + h si=1 bi ϕn+1 ϕ−1
i AYi
P k P
ϕn+1 = exp(h k βJ C(Yk )) · · · exp(h k β1k C(Yk ))
These methods can be directly interpreted as transport-diffusion algorithms for
convection diffusion equations.

References
[1] U. M. Ascher, S. J. Ruuth and B. T. R. Wetton, Implicit-explicit methods for time-dependent
partial differential equations, SIAM J. Num. Anal., 32 (1995), 797–823.
[2] M.J. Baines, Moving finite elements , Monographs on Numerical Analysis, Clarendon Press,
Oxford.
[3] C. Canuto, M. Y. Hussaini, A. Quarteroni and T. A. Zang, Spectral methods in Fluid
Dynamics, Springer-Verlag, Berlin, (1988).
[4] E. Celledoni, A. Marthinsen and B. Owren, Commutator-free Lie group methods, FCGS,
19 (2003), 341–352.
[5] E. Celledoni, Eulerian and semi-Lagrangian commutator-free exponential integrators, CRM
Proceedings and Lecture Notes, 39 (2005), 77–90.
[HE] R.W. Hockney and J.W. Eastwood Computer simulations using particles , McGraw-Hill,
New York, (1981).
[MPR] Y. Maday, A. T. Patera and E. M. Rønquist, An operator integration factor splitting
method for time dependent problems: Application to incompressible fluid flows, J. of Sci.
Comp., 5 (1990), 263–292.
[Pi] O. Pirroneau, On the transport-diffusion algorithm and its applications to the Navier-Stokes
equations , Numer. Math. 38 (1982), 309–332.
[RM] M.D. Rees and K.W. Morton, Moving point, particle and free Lagrange methods for
convection-diffusion equations, SIAM J. Sci. Statist. Comput. 12 (1991), 547-572.
[XK] D. Xiu and G.E. Karniadakis, A semi-Lagrangian high-order method for Navier-Stokes
equations, J. of Comput. Phys. 172 (2001), 658-684.
860 Oberwolfach Report 14/2006

Conserved quantities of some Hamiltonian wave equations after full

discretization
Begoña Cano

It is well known that some differential systems have some invariants which would be
good to conserve after numerical integration. More precisely, I have concentrated
on some Hamiltonian wave equations, like the nonlinear wave and Schrödinger
ones. Let us assume that the standard method of lines is the one used to integrate
these systems, first discretizing in space and then in time.
I have studied how a symmetric space discretization makes that the space dis-
cretized system also has some invariants or ‘nearly’ invariants which well approx-
imate the continuous ones. When an even number of derivatives turn up in each
of the terms of an invariant quantity, the natural discretization of the invariant,
which is the one which uses the same space discretization as that used to construct
the space discretized system, is also an invariant of the latter. However, in other
case, the approximation of the invariant must be done through a pseudospectral
discretization. Some results were already obtained in the literature concerning the
momentum for the nonlinear wave equation and for a particular difference scheme
and using variational techniques [4]. As distinct here, we arrive at similar conclu-
sions, but for any symmetric space discretization, with more direct techniques, and
the conclusions obtained are also valid for other quantities in nonlinear Schrödinger
equation.
Then, integrating in time, the question which arises is whether geometric inte-
grators (which behave so well in the numerical conservation of invariants for ODEs
[2, 3, 5]) also behave in the same advantageous way in the numerical integration
of these space discretized systems, which become arbitrarily stiff when the space
grid is refined. Numerical experiment seem to suggest so. Then, some conjectures
are given (difficult to prove precisely but justified heuristically and corroborated
numerically) such that, taking them as hypotheses, some theorems are obtained on
a proper behaviour with time of the discretized invariants after at least moderate
times.
A more detailed description of the results can be seen in [1].

References
[1] B. Cano, C onserved quantities of some Hamiltonian wave equations after full discretization,
published online in Numerische Mathematik.
[2] E. Hairer, C. Lubich and G. Wanner, Geometric Numerical Integration. Structure-
Preserving Algorithms for Ordinary Differential Equations, Springer-Verlag, Berlin Heidel-
berg New York, 2002.
[3] E. Hairer and C. Lubich, S ymmetric multistep methods over long times, Numer. Math. 97
(2004), 699-723.
[4] M. Oliver, M. West and C. Wulff, Approximate momentum conservation for spatial semi-
discretization of semilinear wave equations, Numer. Math. 97 (2004), 493–535.
[5] J. M. Sanz-Serna and M. P. Calvo, N umerical Hamiltonian Problems, Chapman & Hall,
1994.
Geometric Numerical Integration 861

Euler-Poincaré integrators: variational construction and integrability

Yuri B. Suris

After recalling the most famous example of Euler–Poincaré equations on a semi-

direct product Lie algebra, – the equations of motion of a d-dimensional heavy top
(rigid body with a fixed point), two questions have been posed:
(1) Why do semi-direct products appear in this context? In particular, what
has E(d) to do with rotation around a fixed point (one would expect
SO(d))?
(2) How can one find symplectic (Poisson) discretizations of Euler-Poincaré
equations on semi-direct products?
The most elegant solutions to both problems are claimed to be based on the La-
grangian mechanics (with reduction): with continuous time (on T G) for the first
problem, and with discrete time (on G × G) for the second one. The main ingredi-
ents for the Lagrangian reduction procedure, leading to Euler-Poincaré equations
on semi-direct products, are:
• G – Lie group (configuration space), with the Lie algebra g.
• Φ : G × V → V – representation of G on a vector space V :
Φ(g) · v f ür g ∈ G, v ∈ V.
• φ : g × V → V – corresp. representation of g on V :
d ǫξ

φ(ξ) · v = Φ(e ) · v for ξ ∈ g, v ∈ V.
dǫ ǫ=0
• Bilinear map ⋄ : V ∗ × V → g∗ : for v ∈ V , y ∈ V ∗
hy ⋄ v, ξi = −hy, φ(ξ) · vi ∀ξ ∈ g.
• Fix p ∈ V and let G[p] be its isotropy subgroup:
G[p] = {h : Φ(h) · p = p} ⊂ G.
• Continuous time Lagrange function L : T G → R (discrete time Lagrange
function L : G × G → R), invariant w.r.t. left multiplication of G[p] :
L(g, ġ) = L(l) (Ω, P ), g) = Λ(l) (W, P ),
L(g, b
Ω = g −1 ġ, W = g −1 gb, P = Φ(g −1 ) · p.
• Reduced configuration space: g × Op , resp. G × Op , where Op ⊂ V is the
G-orbit of p ∈ V under Φ.
Theorem 1. [1] Reduced Euler-Lagrange equations take the Euler–Poincaré form:

Ṁ = ad∗ Ω · M + dP L(l) (Ω, P ) ⋄ P,
Ṗ = −φ(Ω) · P,
where
M = dΩ L(l) (Ω, P ) ∈ g∗
862 Oberwolfach Report 14/2006

(this is the Legendre transformation g × Op → g∗ × Op , (Ω, P ) 7→ (M, P )). The

above system is Hamiltonian w.r.t. Poisson bracket
{F1 , F2 }(l) (M, P ) = hM, [dM F1 , dM F2 ] i + hP, φ∗ (dM F2 ) · dP F1 − φ∗ (dM F1 ) · dP F2 i
(Lie-Poisson bracket of the semi-direct product g ⋉ V ∗ for the representation −φ∗
of g on V ∗ ), with the Hamilton function
H (l) (M, P ) = hM, Ωi − L(l) (Ω, P ).
Theorem 2. [2] Reduced discrete Euler-Lagrange equations take the discrete Euler-
Poincaré form:
(  
∗ (l)
Mn+1 = Ad Wn · Mn + dP Λ (Wn , Pn ) ⋄ Pn ,
Pn+1 = Φ(Wn−1 ) · Pn ,
where
′
Mn = D W Λ(l) (Wn−1 , Pn−1 ) ∈ g∗
(discrete Legendre transformation G × Op → g∗ × Op , (Wn−1 , Pn−1 ) 7→ (Mn , Pn )).
Evolution (Mn , Pn ) 7→ (Mn+1 , Pn+1 ) is Poisson w.r.t. Lie-Poisson bracket of the
semi-direct product g ⋉ V ∗ .
These results lead to a recipe for constructing Euler-Poincaré integrators:

◮ For a given system of Euler-Poincaré equations, find a Lagrangian formu-

lation with a G[p] -invariant continuous time Lagrange function on a Lie group
G,
L(g, ġ) = L(l) (Ω, P ),
with Ω = g −1 ġ, P = Φ(g −1 ) · p.
◮ Take any G[p] -invariant discrete time Lagrange function on G,
L(gn , gn+1 ) = Λ(l) (Wn , Pn ),
with Wn = gn−1 gn+1 , Pn = Φ(gn−1 ) · p , and such that
Λ(l) (W, P ) = hL(l) (Ω, P ) + O(h2 )
as W = I + hΩ + O(h2 ) (approximation).
◮ Compute corresponding discrete Euler-Poincaré equations.

These recipe has been illustrated by two integrable examples [3].

A. Dynamically symmetric top is a heavy top with an isotropic tensor of

inertia J = 21 I and with the Lagrange function
1
L(g, ġ) = L(l) (Ω, P ) =
hΩ, Ωi − hP, Ai.
2
Legendre transformation: M = Ω. Equations of motion:

Ṁ = A ∧ P,
Ṗ = −M P.
Geometric Numerical Integration 863

Hamilton function: H(M, P ) = 21 hM, M i + hP, Ai. This system is completely

integrable, due to a Lax representation L̇(λ) = [U (λ), L(λ)], with the Lax matrix
 
M λP − λ−1 A
L(λ) = .
λP T − λ−1 AT 0
An integrable discretization of the symmetric top is generated by the following
discrete Lagrange function:
1
L(gn , gn+1 ) = Λ(l) (Wn , Pn ) = − tr log(2I + Wn + Wn−1 ) − hhPn , Ai.
h
One should note a non-trivial discretization of the kinetic energy. Legendre trans-
formation (coming to replace M = Ω):
2 Wn − I I + (h/2)Mn+1
Mn+1 = · ⇔ Wn =
h Wn + I I − (h/2)Mn+1
Theorem 3. Discrete Euler-Lagrange equations:


 M = Mn + hA ∧ Pn ,
 n+1

 I − (h/2)Mn+1
 Pn+1 = Pn .
I + (h/2)Mn+1
This explicit map is Poisson w.r.t. Lie-Poisson bracket on e(d)∗ , and is completely
integrable, due to Lax representation Ln+1 (λ) = Vn (λ)Ln (λ)Vn−1 (λ), with
 
Mn λQn − λ−1 A
Ln (λ) = ,
λQTn −λ
−1 T
A 0
h h2
where Qn = (I + 2 Mn )Pn + 4 A. One of integrals of motion:
1 h
Hh (M, P ) = hM, M i + hP, Ai + hM, P ∧ Ai
2 2
(deformation of the original Hamiltonian).
B. Clebsch case of the rigid body motion in an ideal fluid
One (simplest) flow of the hierarchy (B – a diagonal d × d matrix):
1 1
L(g, ġ) = L(l) (Ω, P ) =
hΩ, Ωi + hP, BP i.
2 2
Legendre transformation: M = Ω. Equations of motion:

Ṁ = P ∧ (BP ),
Ṗ = −M P.
Hamilton function: H(M, P ) = 21 hM, M i − 12 hP, BP i. Completely integrable, due
to Lax representation L̇(λ) = [U (λ), L(λ)], with
L(λ) = λB + M + λ−1 P P T .
864 Oberwolfach Report 14/2006

An integrable discretization of the Clebsch case is generated by the following dis-

crete Lagrange function:
L(gn , gn+1 ) = Λ(l) (Wn , Pn )
1 4  h2 
−1
= − tr log(2I + Wn + Wn ) + log 1 + hPn , BPn i .
h h 4
Note a non-trivial discretization of both the kinetic and the potential energy.
Legendre transformation (coming to replace M = Ω):
2 Wn − I I + (h/2)Mn+1
Mn+1 = · ⇔ Wn =
h Wn + I I − (h/2)Mn+1
Theorem 4. Discrete Euler-Lagrange equations:


 Pn ∧ (BPn )

 Mn+1 = Mn + h ,
 1 + (h2 /4)hPn , BPn i

 I − (h/2)Mn+1


 Pn+1 = Pn .
I + (h/2)Mn+1
This explicit map is Poisson w.r.t. Lie-Poisson bracket on e(d)∗ , and is completely
integrable, due to Lax representation Ln+1 (λ) = Vn (λ)Ln (λ)Vn−1 (λ), with
 h2 −1
Ln (λ) = I + B (λB + Mn + λ−1 Pn ),
4
where Pn = (I+ h2 Mn )Pn PnT (I− h2 Mn ). Integrals of motion are O(h)-deformations
of the original ones.

References
[1] H. Cendra, D.D. Holm, J.E. Marsden, T. Ratiu. Lagrangian reduction, the Euler-Poincaré
equations, and semi-direct products. – In: Geometry of differential equations, Amer. Math.
Soc. Transl. 186 (1998), 1–25.
[2] A. Bobenko, Yu. Suris, Discrete Lagrangian reduction, discrete Euler-Poincaré equations,
and semi-direct products. Lett. Math. Phys. 49 (1999), 79–93.
[3] Yu. Suris, The problem of integrable discretization: Hamiltonian approach. Basel:
Birkhäuser, 2003. 1070 pp. (Progress in Mathematics, 219).

Stability of rigid body motions (and numerical integrations)

Francesco Fassò

1. Introduction. It is well known that the proper rotations of an Euler top (a

rigid body with a fixed point and no forces acting on it) about the minimum and
maximum axes of inertia are stable. This talk describes some recent results on the
stability of these special motions for a rigid body with a fixed point subject to small
(∼ ǫ) conservative forces [4]. This study is rooted in an asymptotic theory of the
long–term dynamics of small perturbations of the Euler top based on Nekhoroshev
theory [3]. At variance from KAM theory, Nekhoroshev theory [11] permits the
study of resonant motions, even though its results are valid only for a finite, but
Geometric Numerical Integration 865

very long, time scale Tǫ = exp(1/ǫ• ) (the bullet stands for some positive constant).
The problem has geometric, dynamical and numerical aspects.
2. Geometry. The unperturbed system, the Euler top, is a a superintegrable
system: it has three degrees of freedom, but four integrals of motion. Accordingly,
motions are quasi–periodic on two–dimensional tori. The symplectic geometry of
the fibration by these tori is of interest in itself and important for the comprehen-
sion and the study of the perturbed dynamics. For a review, see [8].
In short, the existence of the four integrals of motion produces a doubly fibered
invariant structure of the phase space. The first fibration is given by the Hamil-
tonian K and by the norm G of the angular momentum vector, or equivalently, by
the frequencies of the quasi–periodic motions. Each fiber FKG of this fibration is a
four–dimensional manifold which contains all tori with frequencies determined by
K and G; in view of the rotational invariance of the system, the tori in FKL are
determined by the direction µ of the angular momentum vector in space. In fact,
FKG is fibered by the two–tori, with base the two–dimensional sphere |µ| = 1.
The (symplectic) geometric properties of this structure play a key role in the
formulation of Nekhoroshev theorem for superintegrable systems [7], which in the
present case ensures that K and G are (nearly) constant and that the motion is
approximately quasi–periodic on the time scale Tǫ . In other words, the motion
is approximately an Euler top motion around the istantaneous direction of the
angular momentum vector, which however moves in space.
Thus, all the interesting dynamics takes place on the base S 2 of the toric fibra-
tion of FKG . Accordingly, the theory focusses on the properties of the long–term
motion of µ on the unit sphere—which is clearly relevant in applications.
3. Dynamics. The motion of µ, on the time scale Tǫ , can be studied via the
normal forms constructed within the Nekhoroshev approach. As it turns out, the
properties of the motion of µ depend crucially on the resonance properties of the
frequencies. As shown in [3], for any initial data (not too close to the unstable
manifold of the rotation around the middle axis of inertia) and for times Tǫ :
(i) If the frequencies are nonresonant (up to a certain order), then the motion
of µ is regular. In particular, µ stays close to a level curve of a differentiable
function on S 2 , and the Lyapunov exponents are exponentially small.
(ii) If the frequencies are in a (low order) resonance, instead, then µ can undergo
wide chaotic movements and wander unpredictably through an extended,
essentially two–dimensional, region of S 2 . The maximal Lyapunov exponent
is positive, though small (∼ ǫ• ).
Numerical computations in sample cases provide a clear evidence of the existence
of these chaotic movements in resonance [2].
Assume now that, at the initial time, the angular momentum vector is nearly
aligned with the maximum or minimum axis of inertia. Then, for times Tǫ , nonres-
onant motions are still of the type (i), but resonant motions are more stable than
in (ii): even though the motion remains chaotic, with positive Lyapunov exponent,
µ stays close to a level curve of a differentiable function, as in the nonresonant
866 Oberwolfach Report 14/2006

case [4]. Thus, the stability of proper rotations consists in the spatial localization
of chaos.
4. Numerics. This research has relayed over extensive numerical investigations,
both to test the optimality of the theory and to be guided in the formulation of
a precise conjecture, subsequently proved, about the stability of proper rotations,
which was unknown [2, 5]. Numerical integrations used a Strang splitting ”kinetic
+ potential” and fourth order compositions [9]. Within such an approach it is
important, particularly for perturbation theory, that the flow of the integrable
part be integrated exactly [10]. The exact integration of the flow of an Euler top
is easy in the case of a kinetically symmetric body (two equal moments of inertia),
because the flow involves only elementary functions [6, 1]. Correspondingly, [2, 5]
dealt only with that case. Current work is focussing on the, significantly more
difficult, exact integration of the flow of an Euler top with three distinct moments
of inertia. It is known that the flow can be expressed in terms of theta functions,
but explicit, workable expressions of the flow seem not to be available in the
classical literature.

References
[1] G. Benettin, A.M. Cherubini and F. Fassò, A “changing chart” symplectic algorithm for
rigid bodies and other Hamiltonian systems on manifold, SIAM Journal on Scientific Com-
puting 23 (2001), 1189-1203.
[2] G. Benettin, A.M. Cherubini and F. Fassò, Regular and chaotic motions of the fast rotating
rigid body: a numerical study, Discrete and Continuous Dynamical Systems, Series B 2
(2002), 521-540.
[3] G. Benettin and F. Fassò, Fast rotations of the rigid body: A study by Hamiltonian per-
turbation theory. Part I., Nonlinearity 9 (1996), 137-186.
[4] G. Benettin, F. Fassò and M. Guzzo, Long–term stability of proper rotations of the perturbed
Euler rigid body, Communications in Mathematical Physics 250 (2004), 133-160.
[5] G. Benettin, F. Fassò and M. Guzzo, Long term stability of proper rotations and local chaotic
motions in the perturbed Euler rigid body, Regular and Chaotic Mechanics 11 (2006), to
appear.
[6] A. Dullweber, B. Leimkuhler and R. McLachlan, Symplectic splitting methods for rigid body
molecular dynamics. Journal Chemical Physics 107 (1997), 5840–5851.
[7] F. Fassò, Hamiltonian perturbation theory on a manifold, Celestial Mechanics and Dynam-
ical Astronomy 62 (1995), 43-69.
[8] F. Fassò, Superintegrable Hamiltonian systems: geometry and perturbations, Acta Appli-
candae Mathematicae 87 (2005), 93-121.
[9] E. Hairer, C. Lubich and G. Wanner, Geometric Numerical Integration (Springer–Verlag,
2002, Berlin).
[10] R. McLachlan, Composition methods in the presence of small parameters, BIT–Numerical
Mathematics 35 (1995), 258–268.
[11] N.N. Nekhoroshev, An exponential estimate of the time of stability of nearly integrable
hamiltonian systems. Russian Mathematical Surveys 32 (1977), 1-65.
Geometric Numerical Integration 867

Geodesic Flows on Manifolds and their Discretizations

Anthony Bloch

I describe here related pieces of work on the geometry and dynamics of geodesic
flows on various manifolds of interest. This is joint work with P. Crouch, A.
Iserles, J. Marsden, T. Ratiu and and A. Sanyal. In particular we are interested
in representations of integrable flows and their discretizations.
Some of this work concerns continuous and discrete versions of the generalized
rigid body equations and the role of these equations in numerical analysis, op-
timal control and integrable Hamiltonian systems. In particular, we consider a
symmetric representation of the rigid body equations on the Cartesian product
SO(n) × SO(n) and study its associated symplectic structure. We describe the
relationship of these ideas with the Moser-Veselov theory of discrete integrable
systems and with the theory of variational symplectic integrators.
In their paper on discrete analogues of some classical systems such as the rigid
body and the geodesic flow on an ellipsoid, Moser and Veselov introduced their
analysis in the general context of flows on Stiefel manifolds. Recently we have an-
alyzed a general class of continuous time, quadratic cost, optimal control problems
on Stiefel manifolds, which in the extreme dimensions again yield these classical
physical geodesic flows. In this work we extend the symmetric representation to
the geodesic flow on the ellipsoid and the more general Stiefel manifold case. The
metric we choose on the Stiefel manifolds is the same as that used in the symmet-
ric representation of the rigid body flow, and that used by Moser and Veselov. In
the extreme cases of both the ellipsoid and the rigid body, the geodesic flows are
known to be integrable. We obtain the extremal flows using both variational and
optimal control approaches, and elucidate the structure of the flows on general
Stiefel manifolds. We also discuss discrete representations of these flows.
In some related work we show that the left-invariant geodesic flow on the sym-
plectic group with metric given by the Frobenius norm is an integrable system
that is not contained in the Mishchenko-Fomenko class of integrable systems. We
show that this system may be expressed as a flow on symmetric matrices and that
the system is bi-Hamiltonian. We analyze the Poisson structure of this system
and list the Casimirs. This flow on symmetric matrices is in some sense dual to
the flow of the generalized rigid body equations which flow on the space of skew
symmetric matrices. It has a Lax pair structure which differs by a sign from the
double bracket equations which define another integrable system on symmetric
matrices, the Toda lattice.

References
[Bloch, Crouch, Marsden, and Ratiu(2002)] Bloch, A. M., P. Crouch, J. E. Marsden, and T. S.
Ratiu [2002], The symmetric representation of the rigid body equations and their discretiza-
tion, Nonlinearity 15, 1309–1341.
[Bloch and Iserles(2005)] Bloch, A.M. and A. Iserles [2005], On an isospectral Lie-Poisson system
and its Lie algebra, to appear in Foundations of Computational Mathematics.
868 Oberwolfach Report 14/2006

[Bloch, Iserles, Marsden, and Ratiu(2005)] Bloch, A. M., A. Iserles, J. E. Marsden, and T. S.
Ratiu [2005] A class of integrable geodesic flows on the symplectic group, preprint.
[Bloch, Courch and Sanyal(2005)] Bloch, A. M., P.E. Crouch and A. Sanyal [2005] A variational
problem on Stiefel manifolds, preprint.

Exponential integrators for highly oscillatory differential equations

Volker Grimm
(joint work with Marlis Hochbruck)

The subject of my talk is the study of differential equations of type

(1) y ′′ = −Ay + g(y), y(t0 ) = y0 , y ′ (t0 ) = y0′ ,
where A is a symmetric and positive semi-definite real matrix of arbitrarily large
norm. The large norm of A introduces an oscillatory solution and therefore differ-
ential equations of this type are called oscillatory or highly oscillatory. Oscillatory
differential equations of this type arise in many different applications, for exam-
ple in semidiscretisations of wave equations, and are currently a subject of high
interest.
The nature of oscillatory solutions is such that one needs to apply standard
methods with a step size smaller than the inverse of the largest frequency, which
is the square root of the largest eigenvalue of A. The idea to overcome this step-
size restriction is to preserve the oscillations intrinsic to the equation by solving
exactly linear problems and allowing approximations only to the smoother function
g. Since this involves the evaluation of a matrix exponential, these methods are
called exponential integrators.
The superior performance of exponential integrators for oscillatory differential
equations could be theoretically confirmed by error bounds that are not dependent
on the norm of A or higher derivatives of the solution. This is especially important
for semidiscretisations of partial differential equations. For example, the Gautschi-
type exponential integrator possesses error bounds independent of the frequencies
(cf. [5]) and of the dimension of the system to be solved (cf. [2]), that is, the
accuracy of the Gautschi-type method in time is independent of the grid chosen
in space. Another example is the mollified impulse method, introduced in [1]. No
standard methods are known to share this favourable property.
In this talk, joint work with Marlis Hochbruck (cf. [3]) is presented that unifies
the error analysis known for some exponential integrators to the whole family of
exponential integrators proposed in [4].

References
[1] B. Garcı́a-Archilla, J. M. Sanz-Serna and R. Skeel, Long-time-step methods for oscillatory
differential equations, SIAM J. Sci. Comput. 30 (1998), 930–963.
[2] V. Grimm, A note on the Gautschi-type method for oscillatory second-order differential
equations, Numer. Math. 102 (2005), 61–66.
[3] V. Grimm and M. Hochbruck, Error analysis of exponential integrators for oscillatory
second–order differential equations, J. Phys. A: Math. Gen. 39 (2006), 5495–5507.
Geometric Numerical Integration 869

[4] E. Hairer, Ch. Lubich and G. Wanner, Geometric Numerical Integration, Springer-Verlag
(2002).
[5] M. Hochbruck and Ch. Lubich, A Gautschi-type method for oscillatory second-order differ-
ential equations, Numer. Math. 83 (1999), 403–426.

Highly oscillatory Hamiltonian systems

David Cohen

We consider Hamiltonian systems where high-frequency oscillations are generated

by a linear part. In the first part of the talk, we give a short review of the results
obtained so far for Hamiltonian function of the form:
 
1 T 1 T 0 0
(1) H(p, q) = p p + 2 q q + U (q),
2 2ε 0 I
where 0 < ε ≪ 1 and U is a smooth potential with bounded derivatives. As a
classical model problem, we consider the modified Fermi-Pasta-Ulam problem (see
[7]).
We present a class of numerical methods (the trigonometric methods see [6], [8],
[5] and more generally [7]) designed for such problems (explicit, use of large step
size). Moreover, these methods nearly conserve the total energy of the system and
the oscillatory energy
1 1
(2) I(p, q) = ||p2 ||2 + 2 ||q2 ||2 ,
2 2ε
with partitions p = (p1 , p2 ) and q = (q1 , q2 ) (according to the partition of the
matrix in (1)) over long time intervals.
It is now time to enlarge the class of problems studied, namely by considering
the following Hamiltonian function
1 1 1 1
(3) H(p, q) = pT1 M1 (q)−1 p1 + pT R(q)p + pT2 p2 + 2 q2T q2 + U (q),
2 2 2 2ε
with a symmetric positive definite matrix M1 (q) and a symmetric matrix R(q)
with R(q1 , 0) = 0. Surprising applications, such as the planar elastic dumbbell
spacecraft of [9] or the motion of a triatomic molecule (the water molecule, for
example) can now be considered.
For such Hamiltonian problems, we develop new numerical methods (based on
the trigonometric methods). We present a frequency expansion of the numerical
solution: the modulated Fourier expansion. We also discuss two invariants of the
system that determines its coefficients. These invariants are related to the total
energy and the oscillatory energy of the original system. This allows us to prove
the near-conservation of the total and the oscillatory energy over very long time
intervals.
We finally give some words on the application of the trigonometric methods for
some nonlinear wave equation.
870 Oberwolfach Report 14/2006

References
[1] D. Cohen, Analysis and numerical treatment of highly oscillatory differential equations,
Doctoral Thesis, Univ. Geneva, 2004.
[2] D. Cohen, Conservation properties of numerical integrators for highly oscillatory Hamil-
tonian systems, IMA J. Numer. Anal. 26 (2006) 34–59.
[3] D. Cohen, E. Hairer & Ch. Lubich, Modulated Fourier expansions of highly oscillatory
differential equations, Found. Comput. Math. 3 (2003) 327–345.
[4] D. Cohen, E. Hairer & Ch. Lubich, Numerical energy conservation for multi-frequency
oscillatory differential equations, BIT 45 (2005) 287–305.
[5] B. Garcia-Archilla, J.M. Sanz-Serna & R.D. Skeel, Long-time-step methods for oscillatory
differential equations, SIAM J. Sci. Comput. 20 (1999) 930–963.
[6] E. Hairer & C. Lubich, Long-time energy conservation of numerical methods for oscillatory
differential equations. SIAM J. Numer. Anal. 38 (2000), 414–441.
[7] E. Hairer, C. Lubich & G. Wanner, Geometric Numerical Integration. Structure-Preserving
Algorithms for Ordinary Differential Equations. Springer Series in Computational Mathe-
matics 31, 2002. Second ed., 2006.
[8] M. Hochbruck & C. Lubich, A Gautschi-type method for oscillatory second-order differential
equations, Numer. Math. 83 (1999) 403–426.
[9] A.K. Sanyal, J. Shen & N.H. McClamroch, Variational Integrators for Mechanical Systems
with Configuration Dependent Inertia, Submitted to Internat. J. Numer. Methods Engrg.

Adiabatic integrators for highly oscillatory differential equations

Katina Lorenz
(joint work with Christian Lubich)

The aim of my research work is to develop integrators for highly oscillatory differ-
ential equations in classical mechanics, with the general case of Hamiltonians
1 1
(1) H(p, q) = pT M (q)−1 p + U (q) + 2 V (q),
2 ǫ
1
where a constraining potential ǫ2 V (q) penalizes some directions of motion.
In this talk, we present an approach in which we first consider the case of a time-
dependent Hamiltonian
1 1
(2) H(p, q, t) = pT M (t)−1 p + U (q, t) + 2 q T A(t)q,
2 2ǫ
with a symmetric positive definite matrix M (t), a symmetric positive semidefinite
matrix A(t) and a sufficiently smooth Potential U (q, t). The matrices and their
derivatives are supposed to be bounded independently of ǫ.
In this time-dependent case, we found some useful transformations to reach an
almost-separation of our problem into slow and fast movements [1]. After again
transforming the fast subsystem to smoother, so-called adiabatic variables, we
develop adiabatic integrators showing order two in the positions and one in some
momenta, and another two numerical integrators of global second order, called the
adiabatic midpoint rule and the adiabatic Magnus method.
In [2], we prove that these integrators can be used with significantly larger step
sizes than those required by traditional schemes and we illustrate this in the talk by
Geometric Numerical Integration 871

applying them to some testproblems, where an almost-crossing of eigenfrequencies

takes place. We also present an adaptive step size control succesfully used in
the case of almost-crossings and nonadiabatic transitions. Stating the good long-
time-step behaviour of the adiabatic integrators, we shortly look at the frequency-
dependent case (1). We expect a similar accuracy of the integrators, but in the
case of an almost-crossing of frequencies, the problem shows a chaotic behaviour
(the so-called Takens chaos) and will surely raise new aspects.

References
[1] E. Hairer, C. Lubich, und G. Wanner, Geometric numerical integration, Springer Verlag,
Berlin, 2006, ch. VIII and XIV.
[2] K. Lorenz, T. Jahnke, C. Lubich, Adiabatic integrators for highly oscillatory second order
linear differential equations with time-varying eigendecomposition, BIT Numer. Math. 45
(2005), pp. 91-115.

Highly oscillatory quadrature, the one dimensional case

Syvert P. Nørsett
(joint work with Arieh Iserles)

The quadrature of highly oscillatory problems has been a numerical challenge for
years. Already in 1928 Filon [1] gave the clue on how it could be handled. Later
on little serious were done with such problems, except for using Filon’s method in
a naive fashion. Only Håvie [2] did work using the ideas of Filon. A lot of papers
were written, but with the wrong ideas: trying to banish oscillation by sampling
on a finer grid, rather than using oscillation to improve computation. In 2003–4
Iserles and Nørsett started to look into this challenging problem.
The present talk will focus on the problem:

Z 1
(1) I[f ] = f (x)expiωg(x) dx
0

Here ω is a real positive number and G is a given real function.

The multidimensional case will be taken care of by Iserles in the following talk.
We start our talk by giving an example of how bad it becomes if one uses Gauss
quadrature on the whole interval [0,1]. For example with Gauss quadrature of
order 40 we get an error of O(1) for ω ≥ 60 .
Next we give an asymptotic expansion for I[f], leading to an asymptotic method.
The Filon methods from 1928 exhibit excellent results for all positive ω . Filon
used a quadratic interpolation for f in (1). The general idea is then to interpolate
f with a polynomial p based on the interpolation points c1 , . . . , cν ∈ [0, 1].
We will show that the best result is obtained obtain by using c1 = 0 and cν = 1.
In other words we should use Filon–Lobatto quadrature. The asymptotic error is
then equal to O(ω −2 )
872 Oberwolfach Report 14/2006

From the asymptotic expansion we also see that using Hermite interpolation at
both at 0 and 1 for the function and its derivatives give us an error asymptotic
equal to O(ω −r ) where r is related to the number of derivatives at 0 and 1.
Next we study which effect the interpolation points have on the asymptotic
error. We will show that by using interpolation just to f , but letting interpolation
points depend on ω, we obtain an asymptotic error as if we had used derivatives.
The error in the Filon-method can be estimated by using interpolation for f by
a polynomial of higher order than in the method itself.
To get the Filon-method to work we have assumed that the moments can be
found exactly. This can be a shortcoming of the Filon method. However, based on
some ideas of Levin [3] and Olver [4] we will indicate how to rectify this problem.
A third approach has recently been given by Huybrechs and Vandewalle[5].
They solve (1) by finding an integration path in the complex plane. We will
indicate their approach.
Material for the background of this talk can be found in Iserles and Nørsett
[6],[7].

References
[1] Filon, L. N. G. , On a quadrature formula for trigonometric integrals, Proc. R. Soc. Edin-
bourgh 49 (1928), 38–47.
[2] Håvie, T., Remarks on an expansion for integrals of rapidlyoscillation functions, BIT13
(1973), 16–29.
[3] Levin, D., Fast integration of rapidly oscillatory functions , J.Comput. Appl. Math.67
(1973), 95–101.
[4] Olver S., Moment-free numerical integration of highly oscillatory functions , Technical Re-
port NA2005.04 (2005), DAMTP, University of Cambridge.
[5] Huybrechs D. and Vandewalle S., On the evaluation of highly oscillatory integrals by analytic
continuation , Report TW.431 (2005)
[6] Iserles A. and Nørsett S.P(a)., Efficient quadrature of highly oscillatory integrals using
derivatives, Proceedings of Royal Soc. A.461 (2005), 1383–1399.
[7] Iserles A. and Nørsett S.P.(b) , On quadrature methods for highly oscillatory integrals and
their implememtation, BIT.44 (2004), 755-772.

Multivariate highly oscillatory quadrature

Arieh Iserles
(joint work with Syvert P. Nørsett)

We commence where Syvert Nørsett’s talk left: the quadrature of univariate

integrals of the form
Z b
I[f, (a, b)] = f (x)eiωg(x) dx,
a
where f and g are real, non-oscillatory, sufficiently smooth functions and ω ≫ 1.
The challenge being to ‘transplant’ the theory into a multivariate setting, we let
Geometric Numerical Integration 873

Ω ⊂ Rd be a bounded, open, nonempty domain with piecewise-smooth boundary,

let f and g be smooth cl Ω → R functions and consider
Z
I[f, Ω] = f (x)eiωg(x) dV.
Ω
Exactly like in the univariate case, the key to understanding highly oscillatory
quadrature is the asymptotics of I for ω ≫ 1. The behaviour of I for large ω
depends on the function f and its derivatives at three types of critical points [5]:
(1) Stationary points x0 ∈ cl Ω, where ∇g(x0 ) = 0;
(2) Resonance points x0 ∈ ∂Ω, where ∇g(x0 ) is normal to the boundary;
and
(3) Vertices x0 ∈ cl Ω, where the boundary has a discontinuity and cone of
outward normals is of full diemnsion.
Commencing from the asymptotic expansion
X∞ Z
1 fm
I[f, Ω] ∼ − m+1
n⊤ (x)∇g(x) 2
eiωg(x) dS,
m=0
−iω) ∂Ω k∇g(x)k
where  
⊤ fm−1
f0 = f, fm = ∇ ∇g , m ∈ N.
k∇gk2
Pr
Suppose that all the critical points are ξ1 , . . . , ξr . We represent f = k=0 fk ,
where each fk for k = 1, . . . , r is a C ∞ bump function, equalling 1 in a neighbour-
hood of ξk and 0 outside a larger neighbourhood. Since I[f0 ] decays exponentially
fast with ω ≫ 1, it is enough to consider the contribution of each ξk individually
and to sum them up [3].
(1) Stationary
  points: Let ξj be a nondegenerate point: ∇g(ξj ) = 0 and
∂ 2 g(ξj )
det ∂xk ∂xl 6= 0. Then, according to a theorem of Hörmander,

X∞
iωg(ξj ) am [f ]
I[f, Ω] ∼ e m+d/2
,
m=0
ω
where am depends on derivatives of f at ξj .
(2) Resonance points: Essentially, a resonance pojnt ξj ∈ ∂Ω is nothing else
but a stationary point in the (d − 1)-dimensional manifold ∂Ω, therefore
∞
X
iωg(ξj ) bm [f ]
I[f, Ω] ∼ e m+(d−1)/2
,
m=0
ω
where the bm s depend again on derivatives of f at ξj .
(3) Vertices: Here the one-dimensional framework scales up immediately,
∞
X cm [f ]
I[f, Ω] ∼ ,
m=0
(−iω)d+1
where cm depends on f and its derivatives at the vertex.
874 Oberwolfach Report 14/2006

Pr
The full asymptotic expansion now follows from I[f, Ω] = k=1 I[fk , Ω]. Note
that (although this is possible in principle, using a theorem by Fedoryuk) we do not
seek an explicit asymptotic expansion, since it is exceedingly complex and difficult
to implement in practice. What matter to us is not the precise form of expansion
coefficients but the pattern and nature of information, in terms of functions and
derivative values, that they require. Our approach to approximating I[f, Ω] can be
summed up in the Filon paradigm, paying tribute to the pioneering contribution
of Louis Napoleon Georges Filon [2, 3]. Thus, we seek a quadrature formula Q
such that Q[f, Ω] = I[f, Ω] + O(ω −s ) for some s ≥ 1 and ω ≫ 1. To this end
(1) Given Ω and g, identify the nature of critical points ξ1 , . . . , ξr ∈ cl Ω.
(2) At each ξk identify how many derivatives of f are required, once the in-
tegral is applied to the relevant bump function, for error O(ω −s ) in the
asymptotic expansion. Note that we do not care what the asymptotic ex-
pansion is, only how many derivatives it requires!
(3) Find a function (typically, a polynomial) φ that interpolates (in a Hermite,
or perhaps Birkhoff–Hermite sense) f and its derivatives, up to the order
that we have identified in (2), at each ξk . Note that φ might interpolate f
at further points: although this does not change s, typically this practice
results in much reduced error.
(4) Set Q[f, Ω] = I[φ, Ω]. Therefore,
Q[f, Ω] − I[f, Ω] = I[φ − f, Ω] = O(ω −s ), ω ≫ 1.
The above approach works exceedingly well in practice, provided that I[φ, Ω]
can be computed exactly. If φ is a polynomial, this is equivalent to the requirement
that the moments I[xi , Ω], where xi = xi11 xi22 · · · xidd , can be computed exactly
for relevant multi-indices i. This is often, but not always, the case. Thus, it is
only fair to mention two competing approaches, both available in a multivariate
setting. Firstly, the Levin-type methods of Sheehan Olver, that, in place of inter-
polation, collocate the function f and its derivatives at the critical points [4]; and
secondly, the numerical stationary phase technique of Daan Huybrechs and Ste-
fan Vandewalle, that uses Laguerre-type quadrature along non-oscillatory paths in
the complex plane [1]. Each of the three techniques – Filon-type, Levin-type and
numerical stationary phase – has its own advantages and disadvantages and work
is in progress to clarify further these issues.

References
[1] D. Huybrechs & S. Vandewalle, ”The construction of cubature rules for multivariate highly
oscillatory integrals”, Tech, Rep. (2005).
[2] A. Iserles & S.P. Nørsett, ”Quadrature methods for multivariate highly oscillatory integrals
using derivatives”, to appear in Maths Comp. (2006).
[3] A. iserles & S.P. Nørsett, ”A general approach to the quadrature of highly oscillatory inte-
grals”, to appear (2006).
[4] S.S. Olver, ”On the quadrature of multivariate highly oscillatory integrals over non-polytope
domains”, to appear in Num. Math. (2006).
[5] R. Wong, Approximation of Integrals, Academic Press, Boston (1989).
Geometric Numerical Integration 875

Integration and applications of generalized Euler equations

Robert I. McLachlan

Many Hamiltonian partial differential equations, for example the Camassa–Holm

equations, the Landau–Lifshitz equation, and (Arnold’s celebrated discovery of
1966) the Euler fluid equations, have the form of so-called generalized Euler equa-
tions, the (reduced) equations of geodesics on an infinite-dimensional group with
respect to an invariant metric. The talk explored some analytic and numeri-
cal features of the Euler equations associated with the group of all diffeomor-
phisms of the plane. The metric, typically an H k metric of the form hu, vi =
RR
hu, (1 − ∇2 )k vi dx dy, is allowed to vary. The equations of motion are usually
written in Lie–Poisson form as
ṁ + u · ∇m + ∇uT · m + m(∇ · u) = 0
where the momentum m is the Legendre transform of the velocity u; in the example
above, m = (1 − ∇2 )k u. A striking feature of Euler equations in Rn is that they
admit singular solutions in which the momentum is concentrated on delta functions
on submanifolds of lower dimension m. In the Euler fluid case these correspond
when (m, n) = (0, 2) to point vortices; (0, 3) to vortex sticks; (1, 2) or (2, 3) to
vortex sheets; and (1, 3) to vortex filaments. (In the codimension 2 and 3 cases
the evolution is singular and has to be regularized.) These are used both to
understand fluid motion and to simulate general fluid flows, as in the point vortex
method. For our equation singular solutions take the form
Z
m(x, t) = p(s)δ(x − q(s, t)) ds
M

where the submanifold M is either a set of curves or points. They are Hamiltonian
on T ∗ Q (the cotangent bundle of the embeddings Q = Emb(M, R2 ) of the points
or curves in the plane) with Hamiltonian
ZZ
H= p(s)G(kq(s) − q(u)k)p(u) ds du,
M×M

the integrals reducing to sums in the case of point particles. Here G is the Green’s
function of the metric, i.e. u = G ∗ m. Numerical evidence of Holm and Staley
suggests that smooth initial momentum distributions with compact support evolve
towards a sum of such momentum sheets. They appear to be attracting, at least in
a neighbourhood. Therefore one should first understand the motion of momentum
sheets themselves, and the talk described a first look at the stability and well-
posedness of straight and circular momentum sheets. Unlike the parallel (and
ill-posed) case of vortex sheets, which undergo the Kelvin–Helmholtz instability
ending in a finite-time singularity, momentum sheets are well-posed but generally
vulnerable to a high-frequency instability. This is the opposite to what one might
expect, because the metric is chosen to specifically filter out the high frequencies
in the momentum. Even more remarkably, an important special case, of a straight
876 Oberwolfach Report 14/2006

momentum sheet moving normal to itself under the H 1 (Helmholtz) metric, is in

fact linearly stable. Circular sheets are always weakly (algebraically) unstable [2].
Generalized Euler equations can be solved numerically in Eulerian or Lagrangian
form. The first is a Lie–Poisson PDE on diff (R2 )∗ and has no known Hamilton-
ian discretizations; Holm and Staley use a mimetic finite difference scheme. The
second is a canonical Hamiltonian PDE on T ∗ Diff (R2 ) and can be discretized
by the point particle solutions described above. These can be used to simulate
smooth solutions (many particles are needed, but the calculation can be speeded
up by the Marker-and-Cell method), momentum sheets (place particles on the
sheets), or just the point particle solutions themselves, as is done in the appli-
cation to image registration.
PP In all cases one has a Hamiltonian system with
Hamiltonian H = pi G(kqi − qj k)pj , for which only implicit symplectic inte-
grators are available. However, the equations of motion do have the feature that,
once kqi − qj | and G have been calculated, the Jacobian derivative of the vector
field can be calculated essentially for free. (G is a Gaussian or Bessel function.)
This can be exploited in the iterative solver. For ẋ = f (x) and the midpoint rule
x0 7→ x1 := x0 + ∆tz, we have to solve g(z) := z − f (x0 + 21 ∆tz) = 0. The standard
iteration is z k+1 = z k −g(z k ) while Newton’s method is z k+1 = z k −(I −A)−1 g(z k ),
PM
A = 12 ∆tf ′ (z k ). The new iteration is z k+1 = z k − ( i=0 Ai )g(z k ), which uses one
vector field and Jacobian evaluation and M Jacobian–vector multiplies. In prac-
tice M is determined dynamically using a suitable termination criterion, and z 0 is
given by polynomial extrapolation from previous time steps, the order determined
dynamically. The iteration is advantageous both for N -particle and for lattice (e.g.
semidiscrete PDE) systems. In practice only a few (typically 1–3) evaluations of
the vector field are needed per time step [1]. Open problems in the momentum
sheet application include finding a high-order regularization of singular Green’s
functions that does not introduce a temporal instability, conserving the so-called
potential vorticity, and integrating the normal motion of straight sheets; these are
degenerate, having a double zero eigenvalue in each Fourier mode, which is split
by the particle method. The last two problems can be fixed by first reducing by
the particle-relabelling symmetry G = Diff(R) to obtain a Lie–Poisson system on
T ∗ (Q/G) ⊕ g∗ , but a Hamiltonian discretization on this space is more difficult.
A second application is to image registration (warping, morphing) [4]. One is
given two images f and g and seeks a diffeomorphism ϕ that minimizes kf − g ◦
ϕk. ϕ is induced by the motion of point particles. This gives an optimization
problem with respect to the initial momenta (and positions, if desired) of the
particles, constrained by the equations of motion. Constrained optimization relies
on accurate satisfaction of the constraints, indicating the use of a shooting method,
where all our knowledge of fast solution of initial value problems can be applied.
The iteration described above works even better on optimization, control, and
boundary value problems, because excellent initial guesses are available; this again
favours implicit over explicit methods. The numerical results are encouraging but
a major open problem is to find a direct link between the symplecticity of the
integrator and the behaviour of the optimization problem.
Geometric Numerical Integration 877

References
[1] R I McLachlan, A new implementation of symplectic Runge–Kutta methods, preprint.
[2] R I McLachlan and S Marsland, The Kelvin–Helmholtz instability of momentum sheets in
the Euler equations for planar diffeomorphisms, preprint.
[3] R I McLachlan and S Marsland, N -particle dynamics of the Euler equations for planar dif-
feomorphisms, preprint.
[4] R I McLachlan and S Marsland, Discrete mechanics and optimal control for image registra-
tion, preprint.
[5] R I McLachlan and G R W Quispel, Geometric integrators for ODEs, J Phys A, to appear.
[6] R I McLachlan and G R W Quispel, eds., Special issue on geometric integration, J Phys A,
to appear.

Discussion
Some open problems in Geometric Numerical Integration

• The implicit midpoint rule applied to y ′ = J∇H(y) is a Poisson map.

For more general Poisson systems y ′ = J(y)∇H(y), does there exist any
nonstandard generalization of the implicit midpoint rule that is at least
e
conjugate to a Poisson map for a perturbed matrix J(y)? One can use for
example the multiplicative structure of the equations to obtain nonstan-
dard generalizations of the implicit midpoint rule, e.g.,
 
h yn + yn+1
yn+1 = yn + (J(yn ) + J(yn+1 ))∇H .
2 2

[Laurent O. Jay]
• To give a lower bound on the complexity of approximating the solution
of a differential equation, where (in contrast to standard complexity the-
ory) the algorithm can depend on the differential equation. Hence divide
differential equations into natural complexity classes. [Robert McLachlan]
• The use of fixed-point iterations enlarges the class of functions that can
be computed quickly and also enlarges the class of numerical methods, for
example to include implicit Runge-Kutta methods. Fixed point iterations
with degenerate fixed points, for example (x, y) 7→ (f (x, y), g(x, y)) where
(x, x) 7→ (x, x), define a function by mapping an initial condition (x0 , y0 )
to a limit (x∗ , x∗ ) which depends on the initial condition. Examples are
the AGM (Arithmetic-Geometric Mean) used in computing elliptic func-
tions and an algorithm for the matrix sign function. The problem asks
whether such iterations can be used in numerical integration of differen-
tial equations. [Robert McLachlan]
878 Oberwolfach Report 14/2006

• Construct a volume-preserving integrator that preserves all symmetries

and reversing symmetries of the AAA flow:
dx
= A sin z + A cos y
dt
dy
= A sin x + A cos z
dt
dz
= A sin y + A cos x
dt
or prove that such a construction is impossible. [G. Reinout W. Quispel]
• What (if any) are the benefits of symplectic in space and time integrators
for Hamiltonian wave equations. [Sebastian Reich]
• Backward error analysis for ODEs is well-developed and has been used very
successfully in the analysis of numerical methods, for example to explain
long time approximate energy conservation of symplectic discretizations
of Hamiltonian systems. Recently there has been some progress in the de-
velopment of a backward error analysis for multisymplectic discretizations
of semilinear Hamiltonian PDEs and approximate momentum and energy
conservation of symplectic discretiztions could be explained. However a
backward error analysis for fully nonlinear PDEs - modelling for example
problems from continuum mechanics like fluid dynamics and elastodynam-
ics - is still missing, and this poses a challenging problem for the Geometric
Integration community. [Claudia Wulff ]

Reporter: Volker Grimm

Geometric Numerical Integration 879

Participants

Prof. Dr. Assyr Abdulle Prof. Dr. Fernando Casas

Mathematisches Institut Universitat Jaume I
Universität Basel Departament de Matematiques
Rheinsprung 21 E-12071 Castellon
CH-4051 Basel
Prof. Dr. Elena Celledoni
Dr. Sergio Blanes Dept. of Mathematical Sciences
Universidad Politecnia de Valencia Norwegian University of Science
Instituto de Matematica and Technology
Multidisciplinar A. Getz vei 1
Edificio 8-6 N-7491 Trondheim
E-46022 Valencia
Dr. Philippe Chartier
Prof. Dr. Anthony M. Bloch Universite de Rennes 1
Department of Mathematics IRISA/INRIA
University of Michigan Campus de Beaulieu
4844 East Hall, 530 Church Street F-35042 Rennes Cedex
Ann Arbor, MI 48109-1043
USA
Prof. Dr. Moody T. Chu
Department of Mathematics
Prof. Dr. Chris J. Budd North Carolina State University
School of Mathematical Sciences Campus Box 8205
University of Bath Raleigh, NC 27695-8205
Claverton Down USA
GB-Bath Somerset BA2 7AY
Dr. David Cohen
Dr. Maria Paz Calvo Mathematisches Institut
Departamento de Matematica Aplicada Universität Tübingen
Facultad de Ciencias Auf der Morgenstelle 10
Universidad de Valladolid 72076 Tübingen
Prado de la Magdalena s/n
E-47011 Valladolid
Dr. Ilan Degani
Chemical Physics Department
Prof. Dr. Begona Cano The Weizmann Institute of Science
Departamento de Matematica Aplicada 76100 Rehovot
Facultad de Ciencias ISRAEL
Universidad de Valladolid
Prado de la Magdalena s/n
Dr. Erwan Faou
E-47011 Valladolid
INRIA Rennes
Campus Beaulieu
F-35042 Rennes
880 Oberwolfach Report 14/2006

Prof. Dr. Francesco Fasso Prof. Dr. Laurent Jay

Dipartimento di Matematica Pura Department of Mathematics
ed Applicata College of Liberal Arts a. Sciences
Universita di Padova University of Iowa
Via Belzoni, 7 14 MLH
I-35131 Padova Iowa City, IA 52242
USA
Dr. Jason Frank
CWI - Centrum voor Wiskunde en Prof. Dr. Ben Leimkuhler
Informatica Department of Mathematics
Postbus 94079 University of Leicester
NL-1090 GB Amsterdam University Road
GB-Leicester, LE1 7RH
Dr. Volker Grimm
Mathematisches Institut Dr. Melvin Leok
Universität Düsseldorf Department of Mathematics
Universitätsstr. 1 University of Michigan
40225 Düsseldorf 2074 East Hall
530 Church Street
Prof. Dr. Ernst Hairer Ann Arbor, MI 48109-1043
Departement de Mathematiques USA
Universite de Geneve
Case Postale 64 Prof. Dr. Debra Lewis
2-4 rue du Lievre IMA
CH-1211 Geneve 4 400 Lino Hall
207 Church St.
Prof. Dr. Marlis Hochbruck Minneapolis, MN 55455
Mathematisches Institut USA
Universität Düsseldorf
Universitätsstr. 1 Katina Lorenz
40225 Düsseldorf Mathematisches Institut
Universität Tübingen
Prof. Dr. Arieh Iserles Auf der Morgenstelle 10
Dept. of Applied Mathematics and 72076 Tübingen
Theoretical Physics
University of Cambridge Prof. Dr. Christian Lubich
Wilberforce Rd Mathematisches Institut
GB-Cambridge CB3 OWA Universität Tübingen
Auf der Morgenstelle 10
Dr. Tobias Jahnke 72076 Tübingen
Freie Universität Berlin
Mathematisches Institut Dr. Simon J.A. Malham
Arnimallee 14 Dept. of Mathematics
14195 Berlin Heriot-Watt University
Riccarton
GB-Edinburgh, EH14 4AS
Geometric Numerical Integration 881

Prof. Dr. Robert McLachlan Prof. Dr. Brynjulf Owren

Institute of Fundamental Sciences Department of Mathematics
College of Sciences NTNU
Massey University N-7491 Trondheim
Private Bag 11-222
Palmerston North 5300
Prof. Dr. Reinout Quispel
NEW ZEALAND
Department of Mathematics
La Trobe University
Prof. Dr. Hans Z. Munthe-Kaas Victoria, 3086
Dept. of Mathematics Australia
University of Bergen
Johs. Brunsgate 12
Prof. Dr. Tudor S. Ratiu
N-5008 Bergen
Departement de Mathematiques
Ecole Polytechnique Federale
Prof. Dr. Ander Murua de Lausanne
Informatika Fakultatea CH-1015 Lausanne
EHU/UPV
Donostia Prof. Dr. Sebastian Reich
E-San Sebastian Institut für Mathematik
Universität Potsdam
Prof. Dr. Syvert P. Norsett Postfach 601553
Department of Mathematics 14415 Potsdam
NTNU
N-7491 Trondheim Prof. Dr. J.M. Sanz-Serna
Departamento de Matematica Aplicada
Prof. Dr. Marcel Oliver Facultad de Ciencias
School of Engineering and Science Universidad de Valladolid
International University Bremen Prado de la Magdalena s/n
Campus Ring 12 E-47011 Valladolid
28759 Bremen
Julia Schweitzer
Sheehan Olver Mathematisches Institut
Dept. of Applied Mathematics and Heinrich-Heine-Universität
Theoretical Physics Gebäude 25.22
University of Cambridge Universitätsstraße 1
Wilberforce Rd 40225 Düsseldorf
GB-Cambridge CB3 OWA
Prof. Dr. Zaijiu Shang
Prof. Dr. Alexander Ostermann Academy of Mathematics and
Institut für Mathematik System Sciences
Universität Innsbruck Chinese Academy of Sciences
Technikerstr. 25 Bejing 100080
A-6020 Innsbruck P.R. of China
882 Oberwolfach Report 14/2006

Dr. Daniel Stoffer Dr. Claudia Wulff

Departement Mathematik Department of Mathematics and
ETH-Zentrum Statistics
Rämistr. 101 University of Surrey
CH-8092 Zürich GB-Guildford Surrey, GU2 7XH

Dr. Yuri B. Suris Prof. Dr. Haruo Yoshida

Zentrum Mathematik National Astronomical Observatory
Technische Universität München of Japan
Boltzmannstr. 3 Mitaka
85747 Garching bei München Tokyo 181-8588
JAPAN
Gilles Vilmart
Universite de Rennes 1 Dr. Antonella Zanna
IRISA/INRIA Dept. of Mathematics
Campus de Beaulieu University of Bergen
F-35042 Rennes Cedex Johs. Brunsgate 12
N-5008 Bergen
Prof. Dr. Gerhard Wanner
Section de Mathematiques
Universite de Geneve
Case postale 240
CH-1211 Geneve 24