You are on page 1of 52

Acta Numerica (2003), pp.

399–450 c Cambridge University Press, 2003
DOI: 10.1017/S0962492902000144 Printed in the United Kingdom
Geometric numerical integration
illustrated by the St¨ ormer–Verlet method
Ernst Hairer
Section de Math´ematiques,
Universit´e de Gen`eve, Switzerland
E-mail: Ernst.Hairer@math.unige.ch
Christian Lubich
Mathematisches Institut,
Universit¨ at T¨ ubingen, Germany
E-mail: Lubich@na.uni-tuebingen.de
Gerhard Wanner
Section de Math´ematiques,
Universit´e de Gen`eve, Switzerland
E-mail: Gerhard.Wanner@math.unige.ch
The subject of geometric numerical integration deals with numerical integra-
tors that preserve geometric properties of the flow of a differential equation,
and it explains how structure preservation leads to improved long-time be-
haviour. This article illustrates concepts and results of geometric numerical
integration on the important example of the St¨ ormer–Verlet method. It thus
presents a cross-section of the recent monograph by the authors, enriched by
some additional material.
After an introduction to the Newton–St¨ ormer–Verlet–leapfrog method and
its various interpretations, there follows a discussion of geometric properties:
reversibility, symplecticity, volume preservation, and conservation of first in-
tegrals. The extension to Hamiltonian systems on manifolds is also described.
The theoretical foundation relies on a backward error analysis, which trans-
lates the geometric properties of the method into the structure of a modified
differential equation, whose flow is nearly identical to the numerical method.
Combined with results from perturbation theory, this explains the excellent
long-time behaviour of the method: long-time energy conservation, linear
error growth and preservation of invariant tori in near-integrable systems,
a discrete virial theorem, and preservation of adiabatic invariants.
400 E. Hairer, Ch. Lubich and G. Wanner
CONTENTS
1 The Newton–St¨ ormer–Verlet–leapfrog method 400
2 Geometric properties 408
3 Conservation of first integrals 416
4 Backward error analysis 419
5 Long-time behaviour of numerical solutions 425
6 Constrained Hamiltonian systems 440
7 Geometric integration beyond St¨ ormer–Verlet 447
References 447
1. The Newton–St¨ ormer–Verlet–leapfrog method
We start by considering systems of second-order differential equations
¨ q = f(q), (1.1)
where the right-hand side f(q) does not depend on ˙ q. Many problems in
astronomy, molecular dynamics, and other areas of physics are of this form.
1.1. Two-step formulation
If we choose a step size h and grid points t
n
= t
0
+ nh, the most natural
discretization of (1.1) is
q
n+1
−2q
n
+q
n−1
= h
2
f(q
n
), (1.2)
which determines q
n+1
whenever q
n−1
and q
n
are known. Geometrically, this
amounts to determining an interpolating parabola which, in the mid-point,
assumes the second derivative prescribed by equation (1.1); see Figure 1.1,
left.
q
n−1
q
n
q
n+1
t
n−1
t
n t
n+1
h h
q
n−1
q
n
q
n+1
q
n−
1
2
q
n+
1
2
v
n−
1
2
v
n
v
n+
1
2
t
n−1
t
n t
n+1 t
n−
1
2
t
n+
1
2
Figure 1.1. Method (1.2): two-step formulation (left);
one-step formulations (right).
Geometric numerical integration 401
1.2. One-step formulations
Introducing the velocity ˙ q = v turns equation (1.1) into a first-order system
of doubled dimension
˙ q = v, ˙ v = f(q), (1.3)
an equation in the so-called phase space. In analogy to this, we introduce
discrete approximations of v and q as follows:
v
n
=
q
n+1
−q
n−1
2h
, v
n−
1
2
=
q
n
−q
n−1
h
, q
n−
1
2
=
q
n
+q
n−1
2
, (1.4)
where some derivatives, in order to preserve second-order and symmetry,
are evaluated on the staggered grid t
n−
1
2
, t
n+
1
2
, . . .; see Figure 1.1, right.
Inserting these expressions into the method (or simply looking at the picture)
we see that method (1.2) can now be interpreted as a one-step method Φ
h
:
(q
n
, v
n
) →(q
n+1
, v
n+1
), given by
(A)
v
n+
1
2
= v
n
+
h
2
f(q
n
),
q
n+1
= q
n
+hv
n+
1
2
,
v
n+1
= v
n+
1
2
+
h
2
f(q
n+1
).
q
n
q
n+1
q
n+
1
2
v
n
v
n+
1
2 v
n+1
(1.5)
There is a dual variant of the method on the staggered grid (v
n−
1
2
, q
n−
1
2
) →
(v
n+
1
2
, q
n+
1
2
) as follows:
(B)
q
n
= q
n−
1
2
+
h
2
v
n−
1
2
,
v
n+
1
2
= v
n−
1
2
+hf(q
n
),
q
n+
1
2
= q
n
+
h
2
v
n+
1
2
.
q
n
q
n−
1
2
q
n+
1
2
v
n−
1
2
v
n
v
n+
1
2
(1.6)
For both arrays (A) and (B), we can concatenate, in the actual step-by-
step procedure, the last line of the previous step with the first line of the
subsequent step. Both schemes then turn into the same method, where the
q-values are evaluated on the original grid, and the v-values are evaluated
on the staggered grid:
v
n+
1
2
= v
n−
1
2
+hf(q
n
), (1.7)
q
n+1
= q
n
+hv
n+
1
2
.
This is the computationally most economic implementation, and numerically
more stable than (1.2); see Hairer, Nørsett and Wanner (1993, p. 472).
402 E. Hairer, Ch. Lubich and G. Wanner
1.3. Historical remarks
Isn’t that ingenious? I borrowed it straight from Newton. It comes right out of
the Principia, diagram and all. R. Feynman (1965, p. 43)
The above schemes are known in the literature under various names. In
particular, in molecular dynamics they are often called the Verlet method
(Verlet 1967) and have become by far the most widely used integration
scheme in this field.
Another name for this method is the St¨ormer method, since C. St¨ ormer,
in 1907, used higher-order variants of it for his computations of the motion
of ionized particles in the earth’s magnetic field (aurora borealis); see, e.g.,
Hairer et al. (1993, Section III.10). Sometimes it is also called the Encke
method, because J. F. Encke, around 1860, did extensive calculations for the
perturbation terms of planetary orbits, which obey systems of second-order
differential equations of precisely the form (1.1). Mainly in the context of
partial differential equations of wave propagation, this method is called the
leapfrog method. In yet another context, this formula is the basic method for
the GBS extrapolation scheme, as it was proposed, for the case of equation
(1.1), by Gragg in 1965; see Hairer et al. (1993, p. 294f.). Furthermore, the
scheme (1.7) is equivalent to Nystr¨om’s method of order 2; see Hairer et al.
(1993, p. 362, formula (III.1.13

)).
A curious fact is that Professor Loup Verlet, who later became inter-
ested in the history of science, discovered precisely ‘his’ method in several
places in the classical literature, for example, in the calculations of loga-
rithms and astronomical tables by J. B. Delambre in 1792: this paper was
translated and discussed in McLachlan and Quispel (2002, Appendix C).
Even more spectacular is the finding that the ‘Verlet method’ was used in
Newton’s Principia from 1687 to prove Kepler’s second law. An especially
clear account can be found in Feynman’s Messenger Lecture from 1964; see
Feynman (1965, p. 41), from which we reproduce with pleasure
1
two of
Feynman’s original hand drawings.
The argument is as follows: if there are no forces, the body advances
with uniform speed, and the radius vector covers equal areas in equal times,
simply because the two triangles Sun-1-2 and Sun-2-3 have the same base
and common altitudes (see Figure 1.2). If the gravitational force acts at the
midpoint, the planet is deviated in such a way that the top of the second
triangle moves parallel to the sun ray (see Figure 1.3). Hence, the triangle
Sun-2-4 also has the same area. The whole procedure (uniform motion on
half the interval, then a ‘kick’ to the velocity in the direction of the Sun,
and another uniform motion on the second half) is precisely variant (B) of
the St¨ ormer–Verlet scheme.
1
. . . and with permission of the publisher
Geometric numerical integration 403
Figure 1.2. Uniform motion of a planet
(drawing by R. Feynman).
Figure 1.3. Gravitation acting at mid-point
(drawing by R. Feynman).
1.4. Interpretation as composition method (symplectic Euler)
We can go a step further and split the formulae in the middle of the schemes
(1.5) and (1.6). We then arrive at the schemes (v
n
, q
n
) → (v
n+
1
2
, q
n+
1
2
)
given by
(SE1)
v
n+
1
2
= v
n
+
h
2
f(q
n
),
q
n+
1
2
= q
n
+
h
2
v
n+
1
2
,
q
n
q
n+
1
2
v
n
v
n+
1
2
(1.8)
as well as the adjoint scheme (v
n+
1
2
, q
n+
1
2
) → (v
n+1
, q
n+1
) obtained by for-
mally replacing the subscript n by n + 1 and h by −h,
(SE2)
q
n+1
= q
n+
1
2
+
h
2
v
n+
1
2
,
v
n+1
= v
n+
1
2
+
h
2
f(q
n+1
).
q
n
q
n+1
q
n−
1
2
q
n+
1
2
v
n−
1
2
v
n
v
n+
1
2
v
n+1
(1.9)
404 E. Hairer, Ch. Lubich and G. Wanner
Both these schemes, in which one variable is used at the old value and the
other at the new value, are called the symplectic Euler method.
We thus see that the above scheme (A) is the composition of the symplectic
Euler schemes (SE1) with (SE2), while the scheme (B) is the composition
of method (SE2) followed by (SE1).
1.5. Interpretation as splitting method
We consider the vector field (v, f(q)) of (1.3) ‘split’ as the sum of two vector
fields (v, 0) and (0, f(q)), as indicated in Figure 1.4. The exact flows ϕ
[1]
t
and
ϕ
[2]
t
of these two vector fields, which both have a constant time derivative,
are easily obtained:
ϕ
[1]
t
:
_
q
1
= q
0
+t v
0
v
1
= v
0
and ϕ
[2]
t
:
_
q
1
= q
0
v
1
= v
0
+t f(q
0
).
(1.10)
These formulae are precisely those which build up the formulae (SE1) and
(SE2) above:
(SE2) = ϕ
[2]
h/2
◦ ϕ
[1]
h/2
,
(SE1) = ϕ
[1]
h/2
◦ ϕ
[2]
h/2
.
(SE2)
y
0
y
1
ϕ
[1]
h/2
ϕ
[2]
h/2
(SE1)
y
0
y
1
ϕ
[2]
h/2
ϕ
[1]
h/2
(1.11)
For the two versions of the St¨ ormer–Verlet method we thus obtain the dia-
grams
(A) = (SE2) ◦ (SE1),
(B) = (SE1) ◦ (SE2),
(A)
y
0
y
1
ϕ
[2]
h/2 ϕ
[1]
h
ϕ
[2]
h/2 (B)
y
0
y
1
ϕ
[1]
h/2
ϕ
[2]
h
ϕ
[1]
h/2
(1.12)
˙ q = v
˙ v = f(q)
q
v
=
˙ q = v
˙ v = 0
q
v
+
˙ q = 0
˙ v = f(q)
q
v
Figure 1.4. The phase space vector field split into two fields.
Geometric numerical integration 405
or, in explicit formulae,
Φ
(A)
h
= ϕ
[2]
h/2
◦ ϕ
[1]
h
◦ ϕ
[2]
h/2
,
Φ
(B)
h
= ϕ
[1]
h/2
◦ ϕ
[2]
h
◦ ϕ
[1]
h/2
. (1.13)
This way of composing the flows of split vector fields is often referred to
as Strang splitting, after Strang (1968). For a careful survey of splitting
methods we refer to McLachlan and Quispel (2002).
1.6. Interpretation as variational integrator
A further approach to the St¨ ormer–Verlet method is obtained by discretizing
Hamilton’s principle. This variational principle states that the motion of a
mechanical system between any two positions q(t
0
) = q
0
and q(t
N
) = q
N
is
such that the action integral
_
t
N
t
0
L
_
q(t), ˙ q(t)
_
dt is minimized, (1.14)
where L(q, v) is the Lagrangian of the system. Typically, it is the difference
between the kinetic and the potential energy, that is,
L(q, v) =
1
2
v
T
Mv −U(q), (1.15)
with a symmetric positive definite mass matrix M. When M does not
depend on q, the Euler–Lagrange equations of this variational problem,
d
dt
∂L
∂v
=
∂L
∂q
, reduce to the second-order differential equation M¨ q = −∇U(q).
We now approximate q(t) by a piecewise linear function, interpolating grid
values (t
n
, q
n
) for n = 0, 1, . . . , N, and the action integral by the trapezoidal
rule. We then require that q
1
, . . . , q
N−1
be such that, instead of (1.14),
N−1

n=0
S
h
(q
n
, q
n+1
) is minimized, (1.16)
where
S
h
(q
n
, q
n+1
) =
h
2
L
_
q
n
,
q
n+1
−q
n
h
_
+
h
2
L
_
q
n+1
,
q
n+1
−q
n
h
_
. (1.17)
The requirement that the gradient with respect to q
n
be zero, yields the
discrete Euler–Lagrange equations

Q
S
h
(q
n−1
, q
n
) +∇
q
S
h
(q
n
, q
n+1
) = 0
for n = 1, . . . , N − 1, where the partial gradients ∇
q
, ∇
Q
refer to S
h
=
S
h
(q, Q). In the case of the Lagrangian (1.15) these equations reduce to
M(q
n+1
−2q
n
+q
n−1
) +h
2
∇U(q
n
) = 0, (1.18)
406 E. Hairer, Ch. Lubich and G. Wanner
which is just the two-step formulation (1.2) of the St¨ ormer–Verlet method,
with f(q) = −M
−1
∇U(q).
This variational interpretation of the St¨ ormer–Verlet method was given
by MacKay (1992). A comprehensive survey of variational integrators can
be found in Marsden and West (2001).
1.7. Numerical example
We choose the Kepler problem
¨ q
1
= −
q
1
(q
2
1
+q
2
2
)
3/2
, ¨ q
2
= −
q
2
(q
2
1
+q
2
2
)
3/2
. (1.19)
As initial values we take
q
1
(0) = 1 −e, q
2
(0) = 0, ˙ q
1
(0) = 0, ˙ q
2
(0) =
_
1 +e
1 −e
, (1.20)
with e = 0.6. The period of the exact solution is 2π. Figure 1.5 presents
the numerical values of the St¨ ormer–Verlet method for two different step
sizes. These solutions are compared to those of the explicit midpoint rule in
Runge’s one-step formulation; see Hairer, Lubich and Wanner (2002, p. 24,
Figure 1.2. and equation (1.3)). This second method is of the same order and
for the first steps it behaves very similarly to the St¨ ormer–Verlet scheme (the
first step is even identical!), but it deteriorates significantly as the integration
interval increases. The explanation of this strange difference is the subject
of the theories below.
1.8. Extension to general partitioned problems
For the extension of the above formulae to the more general system
˙ q = g(q, v), ˙ v = f(q, v), (1.21)
we follow the ideas of De Vogelaere (1956). This is a marvellous paper,
short, clear, elegant, written in one week, submitted for publication – and
never published. We first extend the formulae (1.8) and (1.9), by taking
over the missing arguments from one equation to the other. This gives
(SE1)
v
n+
1
2
= v
n
+
h
2
f
_
q
n
, v
n+
1
2
_
,
q
n+
1
2
= q
n
+
h
2
g
_
q
n
, v
n+
1
2
_
,
(1.22)
and
(SE2)
q
n+1
= q
n+
1
2
+
h
2
g
_
q
n+1
, v
n+
1
2
_
,
v
n+1
= v
n+
1
2
+
h
2
f
_
q
n+1
, v
n+
1
2
_
.
(1.23)
Geometric numerical integration 407
−1
1
−1
1
−1
1
−1
1
58 steps h = 0.10 300 steps h = 0.10
116 steps h = 0.05 1200 steps h = 0.05
... St¨ ormer–Verlet scheme ... Runge’s method
Figure 1.5. Kepler problem: the dashed line is the exact solution.
In each of these algorithms the derivative evaluations of both formulae are
taken at the same point. The extensions of the St¨ ormer–Verlet schemes are
now obtained by composition, in the same way as in Section 1.4:
(A) = (SE2)◦(SE1)
v
n+
1
2
= v
n
+
h
2
f
_
q
n
, v
n+
1
2
_
,
q
n+1
= q
n
+
h
2
_
g
_
q
n
, v
n+
1
2
_
+g
_
q
n+1
, v
n+
1
2
_
_
,
v
n+1
= v
n+
1
2
+
h
2
f
_
q
n+1
, v
n+
1
2
_
,
(1.24)
and, for the dual version,
(B) = (SE1)◦(SE2)
q
n
= q
n−
1
2
+
h
2
g
_
q
n
, v
n−
1
2
_
,
v
n+
1
2
= v
n−
1
2
+
h
2
_
f
_
q
n
, v
n−
1
2
_
+f
_
q
n
, v
n+
1
2
_
_
,
q
n+
1
2
= q
n
+
h
2
g
_
q
n
, v
n+
1
2
_
.
(1.25)
408 E. Hairer, Ch. Lubich and G. Wanner
exact
(q
n
, v
n
)
(q
n
, v
n+
1
2
)
(q
n+1
, v
n+
1
2
)
(q
n+1
, v
n+1
)
St¨ ormer–Verlet (A)
exact
(q
n−
1
2
, v
n−
1
2
)
(q
n
, v
n−
1
2
)
(q
n+
1
2
, v
n+
1
2
) (q
n
, v
n+
1
2
)
St¨ ormer–Verlet (B)
Figure 1.6. St¨ ormer–Verlet methods for ˙ q = v, ˙ v = −sin q −v
2
/5,
initial values (−1.8, 0.3), step size h = 1.5. Black points indicate
where the vector field is evaluated.
For illustrations see Figure 1.6. The first equation of (1.24) is now an implicit
formula for v
n+
1
2
, the second one for q
n+1
, while only the last one is explicit.
Such implicit methods were not common in the 1950s and might then not
have delighted journal editors – or programmers:
No detailed example or discussion is given. This will best be done by those
working on these problems in the Brookhaven, Harwell, MURA or CERN group.
De Vogelaere (1956)
2. Geometric properties
We study geometric properties of the flow of differential equations which
are preserved by the St¨ ormer–Verlet method. The properties discussed are
reversibility, symplecticity, and volume preservation.
2.1. Symmetry and reversibility
The St¨ ormer–Verlet method is symmetric with respect to changing the di-
rection of time: in its one-step formulation (1.5), replacing h by −h and
exchanging the subscripts n ↔ n + 1 (i.e., reflecting time at the centre
t
n+1/2
) gives the same method again. Similarly, the replacements h ↔ −h
and n −
1
2
↔ n +
1
2
leave the formulation (1.6) unchanged. In terms of the
numerical one-step map Φ
h
: (q
n
, v
n
) →(q
n+1
, v
n+1
), this symmetry can be
stated more formally as
Φ
h
= Φ
−1
−h
. (2.1)
Such a relation does not hold for the symplectic Euler methods (1.8) and
(1.9), where the above time-reflection transforms (SE1) to (SE2) and vice
versa.
Geometric numerical integration 409
−2 −1 0 1 2
(´ q, −´ v)
(q, −v)
ρ ρ
ϕ
t
ϕ
t
(q, v) (´ q, ´ v)
exact
−2 0 2
(q, v)
(´ q, ´ v)
(´ q, −´ v)
(q, −v)
(A)
ρ
ρ
Figure 2.1. A reversible system (left) and the symmetric St¨ ormer–
Verlet method (right); the same equation as in Figure 1.6.
The time-symmetry of the St¨ormer–Verlet method implies an important
geometric property of the numerical map in the phase space, namely re-
versibility, to which we turn next. The importance of this property in nu-
merical analysis was first emphasized by Stoffer (1988).
The system (1.3) has the property that inverting the direction of the initial
velocity does not change the solution trajectory, it just inverts the direction
of motion. The flow ϕ
t
thus satisfies that
ϕ
t
(q, v) = (´ q, ´ v) implies ϕ
t
(´ q, −´ v) = (q, −v), (2.2)
and we call it reversible with respect to the reflection ρ : (q, v) → (q, −v).
This property is illustrated in Figure 2.1, left. The numerical one-step map
Φ
h
of the St¨ ormer–Verlet method satisfies similarly
Φ
h
(q, v) = (´ q, ´ v) implies Φ
h
(´ q, −´ v) = (q, −v), (2.3)
for all q, v and all h; see Figure 2.1, right. This holds because practically all
numerical methods for (1.3), and in particular the St¨ ormer–Verlet method
and the symplectic Euler methods, are such that
Φ
h
(q, v) = (´ q, ´ v) implies Φ
−h
(q, −v) = (´ q, −´ v), (2.4)
as is readily seen from the defining formulae such as (1.5). The symmetry
(2.1) of the St¨ ormer–Verlet method is therefore equivalent to the reversibility
(2.3). Let us summarize these considerations.
Theorem 2.1. The St¨ ormer–Verlet method applied to the second-order
differential equation (1.1) is both symmetric and reversible, i.e., its one-step
map satisfies (2.1) and (2.3).
410 E. Hairer, Ch. Lubich and G. Wanner
In some situations, the flow is ρ-reversible with respect to involutions ρ
other than (q, v) →(q, −v), that is, it satisfies
ρ ◦ ϕ
t
= ϕ
−1
t
◦ ρ. (2.5)
For example, the flow of the Kepler problem (1.19) is ρ-reversible also with
respect to ρ : (q
1
, q
2
, v
1
, v
2
) → (q
1
, −q
2
, −v
1
, v
2
). In general, the flow of a
differential equation ˙ y = F(y) is ρ-reversible if and only if the vector field
satisfies ρ ◦ F = −F ◦ ρ. We then call the differential equation ρ-reversible.
By the same argument as above, the St¨ormer–Verlet method is then also
ρ-reversible for ρ of the form ρ(q, v) = (ρ
1
(q), ρ
2
(v)), that is,
ρ ◦ Φ
h
= Φ
−1
h
◦ ρ. (2.6)
2.2. Hamiltonian systems and symplecticity
We now turn to the important class of Hamiltonian systems
˙ p = −∇
q
H(p, q), ˙ q = ∇
p
H(p, q), (2.7)
where H(p, q) is an arbitrary scalar function of the variables (p, q). When
the Hamiltonian is of the form
H(p, q) =
1
2
p
T
M
−1
p +U(q), (2.8)
with a positive definite mass matrix M and a potential U(q), then the
system (2.7) turns into the second-order differential equation (1.3) upon
expressing the momenta p = Mv in terms of the velocities and setting
f(q) = −M
−1
∇U(q). Equation (2.8) expresses the total energy H as the
sum of kinetic and potential energy.
A characteristic geometric property of Hamiltonian systems is that the
flow ϕ
t
is symplectic, that is, the derivative ϕ

t
= ∂ϕ
t
/∂(p, q) of the flow
satisfies, for all (p, q) and t where ϕ
t
(p, q) is defined,
ϕ

t
(p, q)
T
J ϕ

t
(p, q) = J with J =
_
0 I
−I 0
_
, (2.9)
where I is the identity matrix of the dimension of p or q; see, e.g., Arnold
(1989, p. 204) or Hairer et al. (2002, p. 172).
The relation (2.9) is formally similar to orthogonality (which it would be
if J were replaced by the identity matrix), but, unlike orthogonality, it is not
related to the conservation of lengths but of areas in phase space. In fact, for
systems with one degree of freedom (i.e., p, q ∈ R), equation (2.9) expresses
that the flow preserves the area of sets of initial values in the (p, q)-plane; see
Figure 2.2, left. For higher-dimensional systems, symplecticity (2.9) means
that the flow preserves the sum of the oriented areas of the projections of
ϕ
t
(A) onto the (p
i
, q
i
)-coordinate planes, for any two-dimensional bounded
Geometric numerical integration 411
manifold of initial values A; see, e.g., Hairer et al. (2002, p. 171f.) for a
justification of this interpretation.
The St¨ ormer–Verlet method (1.24) applied to the Hamiltonian system
(2.7) reads
(A)
p
n+
1
2
= p
n

h
2

q
H
_
p
n+
1
2
, q
n
_
,
q
n+1
= q
n
+
h
2
_

p
H
_
p
n+
1
2
, q
n
_
+∇
p
H
_
p
n+
1
2
, q
n+1
_
_
,
p
n+1
= p
n+
1
2

h
2

q
H
_
p
n+
1
2
, q
n+1
_
,
(2.10)
and a similar formula for variant (B). In the particular case of the Hamil-
tonian (2.8), the method reduces to the St¨ ormer–Verlet method (1.5) with
f(q) = −M
−1
∇U(q), upon setting p
n
= Mv
n
.
A numerical method is called symplectic if, for Hamiltonian systems (2.7),
the Jacobian of the numerical flow Φ
h
: (p
n
, q
n
) → (p
n+1
, q
n+1
) satisfies
condition (2.9), that is, if
Φ

h
(p, q)
T
J Φ

h
(p, q) = J (2.11)
for all (p, q) and all step sizes h.
Symplecticity of numerical methods was first considered by De Vogelaere
(1956), but was not followed up until Ruth (1983) and Feng (1985). In
the late 1980s, the results of Lasagni (1988), Sanz-Serna (1988), and Suris
(1988) started off an avalanche of papers on symplectic numerical methods.
Sanz-Serna and Calvo (1994) was the first book dealing with this subject.
Theorem 2.2. The St¨ ormer–Verlet method applied to a Hamiltonian sys-
tem is symplectic.
ϕ
t
Figure 2.2. Symplecticity of the St¨ ormer–Verlet
method for a separable Hamiltonian.
412 E. Hairer, Ch. Lubich and G. Wanner
We give four different proofs of this result, which all correspond to differ-
ent interpretations of the method: as a composition method, as a splitting
method, as a variational integrator, and using generating functions. Each
of these interpretations lends itself to generalizations to other symplectic
integrators, of higher order and/or for constrained Hamiltonian systems.
Yet another proof is based on the preservation of quadratic invariants and
will be mentioned in Section 3 below. The second proof applies only to
Hamiltonians of the special form (2.8), the third proof is formulated for
such Hamiltonians for convenience.
The historically first proof, due to De Vogelaere (1956), uses the interpre-
tation of the St¨ ormer–Verlet method as the composition of the symplectic
Euler method
(SE1)
p
n+
1
2
= p
n

h
2

q
H
_
p
n+
1
2
, q
n
_
,
q
n+
1
2
= q
n
+
h
2

p
H
_
p
n+
1
2
, q
n
_
,
(2.12)
and its adjoint
(SE2)
q
n+1
= q
n+
1
2
+
h
2

p
H
_
p
n+
1
2
, q
n+1
_
,
p
n+1
= p
n+
1
2

h
2

q
H
_
p
n+
1
2
, q
n+1
_
.
(2.13)
The method (SE1) is indeed symplectic, as is seen by direct verification of
the symplecticity condition
_
∂(p
n+1/2
, q
n+1/2
)
∂(p
n
, q
n
)
_
T
J
_
∂(p
n+1/2
, q
n+1/2
)
∂(p
n
, q
n
)
_
= J.
The matrix of partial derivatives is obtained from differentiating equation
(2.12):
_
I +hH
T
qp
0
−hH
pp
I
__
∂(p
n+1/2
, q
n+1/2
)
∂(p
n
, q
n
)
_
=
_
I −hH
qq
0 I +hH
qp
_
,
where all the submatrices of the Hessian, H
qp
, H
pp
, etc., are evaluated at
(p
n+1/2
, q
n
). In the same way, (SE2) is seen to be symplectic. Hence their
composition (2.10) is also symplectic.
The second proof is the most elegant one, but it applies only to the case
of separable Hamiltonians H(p, q) = T(p) + U(q). It is based on the inter-
pretation of the St¨ ormer–Verlet method as a splitting method. As in (1.13),
we have for variant (A)
Φ
h
= ϕ
U
h/2
◦ ϕ
T
h
◦ ϕ
U
h/2
, (2.14)
where ϕ
T
t
and ϕ
U
t
are the exact flows of the Hamiltonian systems with
Hamiltonian T(p) =
1
2
p
T
M
−1
p and U(q), i.e., ˙ p = 0, ˙ q = M
−1
p and
˙ p = −∇U(q), ˙ q = 0, respectively, corresponding to the splitting H(p, q) =
Geometric numerical integration 413
T(p)+U(q) of the Hamiltonian (2.8) into kinetic and potential energy. Since
the flows of Hamiltonian systems are symplectic, so is their composition
(2.14). This is illustrated in Figure 2.2, right. Variant (B) has the flows of
T and U interchanged in (2.14), and is thus likewise symplectic.
The third proof uses the interpretation of the St¨ ormer–Verlet method
as a variational integrator (see Section 1.6). The symplecticity of varia-
tional integrators derives from non-numerical work by Maeda (1980) and
Veselov (1991). Using (1.4) and the first line of (1.5), we have for S
h
(q, Q)
of (1.17), in the case of the Lagrangian (1.15) which corresponds to the
Hamiltonian (2.8),
−∇
q
S
h
(q
n
, q
n+1
) = M
q
n+1
−q
n
h
+
h
2
∇U(q
n
) = Mv
n
= p
n
(2.15)
and similarly

Q
S
h
(q
n
, q
n+1
) = M
q
n+1
−q
n
h

h
2
∇U(q
n+1
) = Mv
n+1
= p
n+1
. (2.16)
Given (p
n
, q
n
), the first of the above two equations determines q
n+1
, and
the second one p
n+1
. The one-step map Φ
h
: (p
n
, q
n
) → (p
n+1
, q
n+1
) of the
St¨ ormer–Verlet method is thus generated by the scalar-valued function S
h
via (2.15) and (2.16). The desired result then follows from the fact that a
map (p, q) →(P, Q) generated by
−∇
q
S(q, Q) = p, ∇
Q
S(q, Q) = P,
is symplectic for any function S. This is verified by directly checking the
symplecticity condition. Differentiation of the above equations gives the
following relations for the matrices of partial derivatives P
p
, P
q
, Q
p
, Q
q
:
S
qq
+S
qQ
Q
q
= 0, S
qQ
Q
p
= I,
S
Qq
+S
QQ
Q
q
= P
q
, S
QQ
Q
p
= P
p
.
These equations yield
_
P
p
P
q
Q
p
Q
q
_
T
_
0 I
−I 0
__
P
p
P
q
Q
p
Q
q
_
=
_
0 I
−I 0
_
after multiplying out, as is required for symplecticity. This completes the
third proof of symplecticity of the St¨ ormer–Verlet method.
A fourth proof of the symplecticity is based on ideas of Lasagni (1988). A
step of the St¨ ormer–Verlet method can be generated by a function
´
S
h
(p
1
, q
0
)
in the same way as the symplectic Euler method:
p
1
= p
0
−∇
q
´
S
h
(p
1
, q
0
), (2.17)
q
1
= q
0
+∇
p
´
S
h
(p
1
, q
0
).
As we have seen in the first proof, such maps are symplectic. The generating
414 E. Hairer, Ch. Lubich and G. Wanner
function is simply
´
S
h
= hH for the symplectic Euler method. For the
St¨ ormer–Verlet method
´
S
h
is obtained as
´
S
h
(p
1
, q
0
) =
h
2
_
H(p
1/2
, q
0
) +H(p
1/2
, q
1
)
_
(2.18)

h
2
4

q
H(p
1/2
, q
1
)
T
_

p
H(p
1/2
, q
0
) +∇
p
H(p
1/2
, q
1
)
_
,
where q
1
and p
1/2
are defined by the St¨ ormer–Verlet formulae and are now
considered as functions of (p
1
, q
0
). We do not give the computational details,
which can be found in Hairer et al. (2002, Section VI.5) for a more general
class of symplectic integrators.
2.3. Volume preservation
The flow ϕ
t
of a system of differential equations ˙ y = F(y) with divergence-
free vector field (div F(y) = 0 for all y) satisfies det ϕ

t
(y) = 1 for all y. It
therefore preserves volume in phase space: for every bounded open set Ω,
and for every t for which ϕ
t
(y) exists for all y ∈ Ω,
vol (ϕ
t
(Ω)) = vol (Ω).
The vector field (v, f(q)) of a second-order differential equation (1.3), written
as a first-order system, is divergence-free. The same is true for Hamiltonian
vector fields (−∇
q
H(p, q), ∇
p
H(p, q)).
The St¨ ormer–Verlet method preserves volume,
vol (Φ
h
(Ω)) = vol (Ω),
in the following two situations.
For the method (2.10), applied to a Hamiltonian system (2.7), this follows
from its symplecticity (2.11), which implies det Φ

h
(p, q) = 1 for all (p, q).
For partitioned differential equations of the form
˙ q = g(v), ˙ v = f(q), (2.19)
the method (1.24) can be interpreted as the splitting (1.13), where ϕ
[1]
t
and
ϕ
[2]
t
are the exact flows of ˙ q = g(v), ˙ v = 0 and ˙ q = 0, ˙ v = f(q), respectively.
Since the vector fields of these flows are divergence-free, they are volume-
preserving and so is their composition.
The same idea allows us to extend the St¨ormer–Verlet method to a volume-
preserving algorithm for systems partitioned into the three equations
˙ x = a(y, z), ˙ y = b(x, z), ˙ z = c(x, y), (2.20)
for which the diagonal blocks of the Jacobian are zero. We split them sym-
metrically, giving
ϕ
[1]
h/2
◦ ϕ
[2]
h/2
◦ ϕ
[3]
h
◦ ϕ
[2]
h/2
◦ ϕ
[1]
h/2
, (2.21)
Geometric numerical integration 415
x
y
z
t = 0
t = 1.6
x
y
z
t = 0
t = 1.6
Figure 2.3. Volume-preserving deformation of the ball of
radius 0.9, centred at the origin, by the ABC flow (left)
and by method (2.22) (right).
where ϕ
[1]
t
is the (volume-preserving) flow of ˙ x = a(y, z), ˙ y = 0, ˙ z = 0 and
similarly for ϕ
[2]
t
and ϕ
[3]
t
. Written out, this becomes
x
n+
1
2
= x
n
+
h
2
a(y
n
, z
n
),
y
n+
1
2
= y
n
+
h
2
b
_
x
n+
1
2
, z
n
_
,
z
n+1
= z
n
+hc
_
x
n+
1
2
, y
n+
1
2
_
, (2.22)
y
n+1
= y
n+
1
2
+
h
2
b
_
x
n+
1
2
, z
n+1
_
,
x
n+1
= x
n+
1
2
+
h
2
a(y
n+1
, z
n+1
).
An illustration of this algorithm, applied to the ABC-flow
˙ x = Asin z +C cos y,
˙ y = Bsin x +Acos z,
˙ z = C sin y +Bcos x,
is presented in Figure 2.3 for A = 1/2, B = C = 1.
More ingenuity is necessary if the system is divergence-free with nonzero
elements on the diagonal of the Jacobian. Feng and Shang (1995) give a
volume-preserving extension of the above scheme to the general case.
416 E. Hairer, Ch. Lubich and G. Wanner
3. Conservation of first integrals
A non-constant function I(y) is a first integral (or conserved quantity, or
constant of motion, or invariant) of the differential equation ˙ y = F(y) if
I(y(t)) is constant along every solution, or equivalently, if
I

(y)F(y) = 0 for all y. (3.1)
The latter condition says that the gradient ∇I(y) is orthogonal to the vector
field F(y) in every point of the phase space.
The foremost example is the Hamiltonian H(p, q) of a Hamiltonian system
(2.7): since H

= (∇
p
H
T
, ∇
q
H
T
) and ∇
p
H
T
(−∇
q
H)+∇
q
H
T

p
H = 0, the
total energy H is a first integral. Apart from very exceptional cases, H is
not constant along numerical solutions computed with the St¨ ormer–Verlet
method. Later we will see, however, that H is conserved up to O(h
2
) over
extremely long time intervals.
Example 3.1. The Kepler problem (1.19) is Hamiltonian with H(p, q) =
1
2
(p
2
1
+ p
2
2
) − 1/
_
q
2
1
+q
2
2
. In addition to the Hamiltonian, this system has
the following conserved quantities, as can be easily checked: the angular
momentum L = q
1
p
2
− q
2
p
1
, and the nonzero components of the Runge–
Lenz–Pauli vector
_
_
A
1
A
2
0
_
_
=
_
_
p
1
p
2
0
_
_

_
_
0
0
q
1
p
2
−q
2
p
1
_
_

1
_
q
2
1
+q
2
2
_
_
q
1
q
2
0
_
_
.
Figure 3.1 shows the behaviour of these quantities along a numerical so-
lution of the St¨ ormer–Verlet method. The method preserves the angular
momentum exactly (see Section 1.3), and there are only small errors in the
Hamiltonian along the numerical solution, but no drift. There is, however, a
linear drift in the Runge–Lenz–Pauli vector. In contrast, for explicit Runge–
Kutta methods, none of the first integrals is preserved, and there is a drift
away from the constant value for all of them.
10 20
−.006
−.004
−.002
.000
.002
H
A
2
Figure 3.1. The Hamiltonian and the second component of
the Runge–Lenz–Pauli vector along the numerical solution
of the St¨ ormer–Verlet method with step size h = 0.02.
Geometric numerical integration 417
Example 3.2. (Conservation of total linear and angular momentum
of N-body systems) A system of N particles interacting pairwise, with
potential forces depending on the distances of the particles, is formulated as
a Hamiltonian system with total energy
H(p, q) =
1
2
N

i=1
1
m
i
p
T
i
p
i
+
N

i=2
i−1

j=1
V
ij
_
|q
i
−q
j
|
_
. (3.2)
Here q
i
, p
i
∈ R
3
represent the position and momentum of the ith particle of
mass m
i
, and V
ij
(r) (i > j) is the interaction potential between the ith and
jth particle. The equations of motion read
˙ q
i
=
1
m
i
p
i
, ˙ p
i
=
N

j=1
ν
ij
(q
i
−q
j
)
where, for i > j, we have ν
ij
= ν
ji
= −V

ij
(r
ij
)/r
ij
with r
ij
= |q
i
−q
j
|, and
ν
ii
is arbitrary, say ν
ii
= 0. The conservation of the total linear momen-
tum P =

N
i=1
p
i
and the total angular momentum L =

N
i=1
q
i
p
i
is a
consequence of the symmetry relation ν
ij
= ν
ji
:
d
dt
N

i=1
p
i
=
N

i=1
N

j=1
ν
ij
(q
i
−q
j
) = 0,
d
dt
N

i=1
q
i
p
i
=
N

i=1
1
m
i
p
i
p
i
+
N

i=1
N

j=1
q
i
ν
ij
(q
i
−q
j
) = 0.
The exact preservation of linear first integrals, such as the total linear
momentum, is common to most numerical integrators.
Theorem 3.3. The St¨ ormer–Verlet method preserves linear first integrals.
Proof. Let the linear first integral be I(q, v) = b
T
q + c
T
v, so that b
T
v +
c
T
f(q) = 0 for all q, v. Necessarily then, c
T
f(q) = 0 for all q, and b = 0.
Multiplying the formulae for v in (1.5) by c
T
thus yields c
T
v
1
= c
T
v
0
.
Quadratic first integrals are not generally preserved by the St¨ ormer–Verlet
method, as the following example shows.
Example 3.4. Consider the harmonic oscillator, which has the Hamil-
tonian H(p, q) =
1
2
p
2
+
1
2
ω
2
q
2
(p, q ∈ R). Applying the St¨ ormer–Verlet
method gives
_
p
n+1
ωq
n+1
_
= A(hω)
_
p
n
ωq
n
_
(3.3)
418 E. Hairer, Ch. Lubich and G. Wanner
with the propagation matrix
A(hω) =
_
1 −
h
2
ω
2
2


2
_
1 −
h
2
ω
2
4
_

2
1 −
h
2
ω
2
2
_
. (3.4)
Since A(hω) is not an orthogonal matrix, H(p, q) is not preserved along
numerical solutions. Notice, however, that the characteristic polynomial is
λ
2
−(2−h
2
ω
2
)λ+1, so that the eigenvalues are of modulus one if (and only
if) [hω[ ≤ 2. The matrix V of eigenvectors is close to the identity for small
hω, and the norm of V
−1
(p
n
, ωq
n
)
T
is conserved.
The St¨ ormer–Verlet method does, however, preserve an important sub-
class of quadratic first integrals, and in particular the total angular momen-
tum of N-body systems. As we have seen in Section 1.3, Newton was aware
that the method preserves angular momentum in the Kepler problem and
used this fact to prove Kepler’s second law. In the following result C is a
constant square matrix and c a constant vector.
Theorem 3.5. The St¨ ormer–Verlet method preserves quadratic first in-
tegrals of the form I(q, v) = v
T
(Cq + c) (or I(p, q) = p
T
(Bq + b) in the
Hamiltonian case).
Proof. By (3.1), f(q)
T
(Cq+c)+v
T
Cv = 0 for all q, v. Writing the St¨ ormer–
Verlet method as the composition of the two symplectic Euler methods (1.8)
and (1.9), we obtain for the first half-step
v
T
n+1/2
(Cq
n+1/2
+c) = v
T
n
(Cq
n
+c)
+
h
2
_
f(q
n
)
T
(Cq
n
+c) +v
T
n+1/2
Cv
n+1/2
_
,
where we notice that the term in the second line vanishes. For the second
half-step we obtain in the same way v
T
n+1
(Cq
n+1
+c) = v
T
n+1/2
(Cq
n+1/2
+c),
and the result follows.
The most important source of first integrals of Hamiltonian systems is
Noether’s theorem, which states that continuous symmetries yield first in-
tegrals: if the associated Lagrangian is invariant under the flow α
s
of the
vector field a(q), that is, L(α
s
(q), α

s
(q)v) = L(q, v) for all real s near 0 and
all (q, v), then I(p, q) = p
T
a(q) is a first integral; see, e.g., Arnold (1989,
p. 88). For Hamiltonian systems of the form (2.8), where the associated La-
grangian is (1.15), it can be shown that a(q) must be linear: a(q) = Bq +b,
with MB skew-symmetric. Hence, for Hamiltonian systems (2.7) with a
Hamiltonian of the form (2.8), all first integrals originating from Noether’s
theorem are preserved by the St¨ ormer–Verlet method.
Theorem 3.5 yields yet another proof (and further insight) of the symplec-
ticity of the St¨ ormer–Verlet method, following an argument by Bochev and
Geometric numerical integration 419
Scovel (1994): consider the Hamiltonian system ˙ p = −∇U(q), ˙ q = M
−1
p
together with its variational equation
˙
Y =
_
0 −∇
2
U(q)
M
−1
0
_
Y with Y =
_
P
p
P
q
Q
p
Q
q
_
.
The derivative of the flow is then ϕ

t
(p, q) = Y (t) corresponding to the
initial conditions p, q and Y (0) = I. The derivative Φ

h
(p, q) of the numeri-
cal solution with respect to the initial values equals the result Y
1
obtained
by applying the method to the combined system of the Hamiltonian sys-
tem together with its variational equation, partitioned into (p, P
p
, P
q
) and
(q, Q
p
, Q
q
). Symplecticity means that the components of Y
T
JY are first
integrals. Since they are of the mixed quadratic type considered above,
Theorem 3.5 shows that they are preserved by the St¨ormer–Verlet method:
Y
T
1
JY
1
= Y
T
0
JY
0
, which is just the symplecticity Φ

h
T

h
= J.
4. Backward error analysis
The theoretical foundation of geometric integrators is mainly based on a
backward interpretation which considers the numerical approximation as
the exact solution of a modified problem. Such an interpretation has been
intuitively used in the physics literature, e.g., Ruth (1983). A rigorous for-
mulation evolved around 1990, beginning with the papers by Feng (1991),
McLachlan and Atela (1992), Sanz-Serna (1992) and Yoshida (1993). Expo-
nentially small error bounds and applications of backward error analysis to
explain the long-time behaviour of numerical integrators were subsequently
given by Benettin and Giorgilli (1994), Hairer and Lubich (1997), and Re-
ich (1999a). We explain the essential ideas and we illustrate them for the
St¨ ormer–Verlet method.
4.1. Construction of the modified equation
The idea of backward error analysis applies to general ordinary differential
equations and to general numerical integrators, and a restriction to special
methods for second-order problems would hide the essentials. We therefore
consider the differential equation
˙ y = F(y), (4.1)
and a numerical one-step method y
n+1
= Φ
h
(y
n
). The idea consists in
searching and studying a modified differential equation
˙ y = F(y) +hF
2
(y) +h
2
F
3
(y) + , (4.2)
such that the exact time-h flow ¯ ϕ
h
(y) of (4.2) is equal to the numerical flow
Φ
h
(y). Unfortunately, the series in (4.2) cannot be expected to converge in
general, and the precise statement has to be formulated as follows.
420 E. Hairer, Ch. Lubich and G. Wanner
Theorem 4.1. Consider (4.1) with an infinitely differentiable vector field
F(y), and assume that the numerical method admits a Taylor series expan-
sion of the form
Φ
h
(y) = y +hF(y) +h
2
D
2
(y) +h
3
D
3
(y) + (4.3)
with smooth D
j
(y). Then there exist unique vector fields F
j
(y) such that,
for any N ≥ 1,
Φ
h
(y) = ¯ ϕ
h,N
(y) +O(h
N+1
),
where ¯ ϕ
t,N
is the exact flow of the truncated modified equation
˙ y = F(y) +hF
2
(y) + +h
N−1
F
N
(y). (4.4)
Proof. Disregarding convergence issues, we expand the exact flow of (4.2)
into a Taylor series (using the notation ¯ y(t) = ¯ ϕ
t
(y))
¯ ϕ
h
(y) = y +h
˙
¯ y (0) +
h
2
2!
¨
¯ y (0) +
h
3
3!
¯ y
(3)
(0) +
= y +h
_
F(y) +hF
2
(y) +h
2
F
3
(y) +
_
(4.5)
+
h
2
2!
_
F

(y) +hF

2
(y) +
__
F(y) +hF
2
(y) +
_
+
and we compare like powers of h in the expressions (4.5) and (4.3). This
yields recurrence relations for the functions F
j
(y), namely,
F
2
(y) = D
2
(y) −
1
2!
F

F(y), (4.6)
F
3
(y) = D
3
(y) −
1
3!
_
F

(F, F)(y) +F

F

F(y)
_

1
2!
_
F

F
2
(y) +F

2
F(y)
_
,
and uniquely defines the functions F
j
(y) in a constructive manner.
4.2. Modified equation of the St¨ormer–Verlet method
Putting y = (q, v)
T
and F(y) = (v, f(q))
T
, the differential equation (1.3) is
of the form (4.1). For the St¨ ormer–Verlet scheme (1.5) we have
Φ
h
(q, v) =
_
q +hv +
h
2
2
f(q)
v +
h
2
f(q) +
h
2
f
_
q +hv +
h
2
2
f(q)
_
_
. (4.7)
Expanding this function into a Taylor series we get (4.3) with
D
2
(q, v) =
1
2
_
f(q)
f

(q)v
_
, D
3
(q, v) =
1
4
_
0
f

(q)f(q) +f

(q)(v, v)
_
, . . .
and the functions F
j
(q, v) can be computed as in the proof of Theorem 4.1.
Since the St¨ ormer–Verlet method is of second order, the function D
2
(q, v)
Geometric numerical integration 421
−2 0 2 4
−2
2
−2 0 2 4
−2
2
−2 0 2 4
−2
2
−2 0 2 4
−2
2
St¨ ormer–Verlet (A) St¨ ormer–Verlet (B)
Figure 4.1. Numerical solution with step size h = 0.9 for the two
versions of the St¨ ormer–Verlet method compared to the exact
flow of their modified differential equations truncated after the
O(h
2
) term.
has to coincide with the h
2
-coefficient of the exact solution and we have
F
2
(q, v) = 0. We then get
F
3
(q, v) =
1
12
_
−2 f

(q)v
f

(q)f(q) +f

(q)(v, v)
_
, (4.8)
and for the next function we obtain F
4
(q, v) = 0. The vanishing of this
function follows from the symmetry of the method (cf. Section 4.3). For
larger (odd) j the functions F
j
(q, v) become more and more complicated,
and higher derivatives of f(q) are involved. The explicit formula for F
3
(q, v)
also shows that the modified differential equation (4.2) is no longer a second-
order equation like (1.3).
A similar computation for the version (B) of the St¨ ormer–Verlet method
(see (1.6)) gives
F
3
(q, v) =
1
24
_
2 f

(q)v
−4 f

(q)f(q) −f

(q)(v, v)
_
, (4.9)
and, obviously, also F
2
(q, v) = F
4
(q, v) = 0.
As a concrete example consider the pendulum equation for which f(q) =
−sin q. The two pictures of Figure 4.1 show the exact flow of the modi-
fied differential equations (truncated after the O(h
2
) term) corresponding
to the two versions (1.5) and (1.6) of the St¨ ormer–Verlet scheme together
with the numerical solution for the initial value (p
0
, q
0
) = (1.0, −1.2). The
shade of the numerical approximations (dark grey to light grey) indicates
the increasing time. We observe a surprisingly good agreement.
In both cases the solutions of the modified equation are periodic, and the
numerical approximation lies near a closed curve, so that correct qualitative
behaviour is obtained. This is explained by the fact that for f(q) = −∇U(q)
422 E. Hairer, Ch. Lubich and G. Wanner
the vector fields (4.8) and (4.9) are Hamiltonian with
H
3
(p, q) =
1
12

2
U(q)(p, p) +
1
24
∇U(q)
T
∇U(q) and
H
3
(p, q) = −
1
24

2
U(q)(p, p) −
1
6
∇U(q)
T
∇U(q),
respectively. Consequently, the exact solutions of the truncated modified
equation stay on the level curves of
¯
H(p, q) = H(p, q) + h
2
H
3
(p, q) which
are drawn in Figure 4.1.
4.3. Properties of the modified differential equation
In Section 4.4 we shall see that the numerical solution is extremely close to
the exact solution of a truncated modified equation. To study properties
of the numerical solution it is therefore justified to investigate instead the
corresponding properties of the modified differential equation.
It follows from the definition of the modified equation that for methods
of order r, that is, Φ
h
(y) = ϕ
h
(y) +O(h
r+1
), we have
F
j
(y) = 0 for j = 2, . . . , r.
Furthermore, if the leading term of the local truncation error is E
r+1
(y),
that is, Φ
h
(y) = ϕ
h
(y) +h
r+1
E
r+1
(y) +O(h
r+2
), then
F
r+1
(y) = E
r+1
(y).
By Theorem 2.1 the St¨ ormer–Verlet method is symmetric. For such meth-
ods the modified equation has an expansion in even powers of h, that is,
F
2j
(y) = 0 for j = 1, 2, . . . . (4.10)
This can be proved as follows: to indicate the h-dependence of the vector
field (4.2), we let ¯ ϕ
t,h
(y) denote the (formal) flow of (4.2). Backward error
analysis tells us that Φ
h
(y) = ¯ ϕ
h,h
(y). We thus have Φ
−h
(y) = ¯ ϕ
−h,−h
(y)
and, by the group property of the exact flow, Φ
−1
−h
(y) = ¯ ϕ
h,−h
(y). The
symmetry condition (2.1) thus implies that ¯ ϕ
t,h
(y) = ¯ ϕ
t,−h
(y) for t = h,
and the computation of (4.5) shows that this is only possible if (4.10) holds.
Geometric properties of a numerical method have their counterparts in
the modified equation. Let us explain this for the properties discussed in
Sections 2 and 3.
Theorem 4.2. (Reversible systems) If the St¨ ormer–Verlet method (1.5)
is applied to a differential equation (1.3), then every truncation of the
modified differential equation is reversible with respect to the reflection
ρ(q, v) = (q, −v).
Theorem 4.3. (Hamiltonian systems) If the St¨ ormer–Verlet method
(2.10) is applied to a Hamiltonian system, then every truncation of the
modified differential equation is Hamiltonian.
Geometric numerical integration 423
Theorem 4.4. (Divergence-free systems) If the St¨ ormer–Verlet method
(1.24) is applied to a divergence-free system of the form (2.19), then every
truncation of the modified differential equation is divergence-free.
Theorem 4.5. (First integrals) If the St¨ ormer–Verlet method (1.24) is
applied to a differential equation with a first integral of the form I(q, v) =
v
T
(Cq + c), then every truncation of the modified differential equation has
I(q, v) as a first integral.
The proofs are based on an induction argument. Since they are all very
similar (see Hairer et al. (2002, Chapter IX)), we only present the proof
of Theorem 4.3, for the case where the Hamiltonian H(p, q) is defined on
a simply connected domain. This proof was first given by Benettin and
Giorgilli (1994) and Tang (1994), and its ideas can be traced back to Moser
(1968).
Proof. With y = (p, q), the Hamiltonian system (2.7) is written more com-
pactly as ˙ y = J
−1
∇H(y) with J of (2.9). We will show that all the coefficient
functions of the modified equation can be written as
F
j
(y) = J
−1
∇H
j
(y). (4.11)
Assume, by induction, that (4.11) holds for j = 1, 2, . . . , N (this is satisfied
for N = 1, because F
1
(y) = F(y) = J
−1
∇H(y)). We have to prove the
existence of a Hamiltonian H
N+1
(y). The idea is to consider the truncated
modified equation (4.4), which is then a Hamiltonian system with Hamil-
tonian H(y) + hH
2
(y) + + h
N−1
H
N
(y). Its flow ϕ
N,t
(y), compared to
that of (4.2) and thus to the one-step map Φ
h
of the St¨ ormer–Verlet method,
satisfies
Φ
h
(y) = ϕ
N,h
(y) +h
N+1
F
N+1
(y) +O(h
N+2
),
and also
Φ

h
(y) = ϕ

N,h
(y) +h
N+1
F

N+1
(y) +O(h
N+2
).
By Theorem 2.2 and by the induction hypothesis, Φ
h
and ϕ
N,h
are sym-
plectic transformations. This, together with ϕ

N,h
(y) = I + O(h), therefore
implies
J = Φ

h
(y)
T

h
(y) = J +h
N+1
_
F

N+1
(y)
T
J +JF

N+1
(y)
_
+O(h
N+2
).
Consequently, the matrix JF

N+1
(y) is symmetric. The function JF
N+1
(y) is
therefore the gradient of some scalar function H
N+1
(y), which proves (4.11)
for j = N + 1.
The last argument of this proof requires that the domain be simply con-
nected. For general domains, we have to use the representation (2.17) with
the help of the generating function (2.18). We refer to Section IX.3.2 of
Hairer et al. (2002) for details of the proof.
424 E. Hairer, Ch. Lubich and G. Wanner
4.4. Exponentially small error estimates
Theorem 4.1 proves a statement that is valid for all N ≥ 1, and it is natural
to ask which choice of N gives the best estimate.
Example 4.6. Consider the simple differential equation ¨ q = f(t) (which
becomes autonomous after adding
¨
t = 0). If we try to compute the modified
equation for the St¨ ormer–Verlet method, we are readily convinced that its
q-component is of the form
¨ q(t) = f(t) +h
2
b
2
¨
f(t) +h
4
b
4
f
(4)
(t) +h
6
b
6
f
(6)
(t) + . (4.12)
Putting f(t) = e
t
, the solution of this modified equation is
¯ q(t) = C
1
+tC
2
+ (1 +b
2
h
2
+b
4
h
4
+b
6
h
6
+ ) e
t
,
and inserted into (1.2) we obtain
(1 +b
2
h
2
+b
4
h
4
+b
6
h
6
+ )(e
−h
−2 + e
h
) = h
2
. (4.13)
This shows that 1+b
2
h
2
+b
4
h
4
+ is analytic in a disc of radius 2π centred
at the origin. Consequently, the coefficients behave like b
2k
≈ Const (2π)
−2k
for k →∞.
Now consider functions f(t) whose derivatives grow like f
(k)
(t)≈k! M R
−k
.
This is the case for analytic f(t) with finite poles. The individual terms of
the modified equation (4.12) then behave like
h
2k
b
2k
f
(2k)
(t) ≈ Const
h
2k
(2k)!
(R 2π)
2k
≈ Const

4πk
_
h 2k
R 2π e
_
2k
(4.14)
(using Stirling’s formula). Even for very small step sizes h this expression is
unbounded for k →∞, so that the series (4.12) cannot converge. However,
formula (4.14) tells us that the terms of the series decrease until 2k ap-
proaches the value 2πR/h, and then they tend rapidly to ∞. It is therefore
natural to truncate the modified equation after N terms, where N ≈ 2πR/h.
To find a reasonably good truncation index N for general differential
equations, we have to know estimates for all derivatives of F(y) and of the
coefficient functions D
j
(y) of the Taylor expansion of the numerical flow.
One convenient way of doing this is to assume analyticity of these functions.
Exponentially small error bounds were first derived by Benettin and Gior-
gilli (1994). The following estimates are from Hairer et al. (2002, p. 306).
Theorem 4.7. Let F(y) be analytic in B
2R
(y
0
), let the coefficients D
j
(y)
of the method (4.3) be analytic in B
R
(y
0
), and assume that
|F(y)| ≤ M and |D
j
(y)| ≤ µM
_
2κM
R
_
j−1
(4.15)
hold for y ∈ B
2R
(y
0
) and y ∈ B
R
(y
0
), respectively. If h ≤ h
0
/4 with
Geometric numerical integration 425
h
0
= R/(eηM) and η = 2 max
_
κ, µ/(2 ln 2−1)
_
, then there exists N = N(h)
(namely N equal to the largest integer satisfying hN ≤ h
0
) such that the
difference between the numerical solution y
1
= Φ
h
(y
0
) and the exact solution
¯ ϕ
N,t
(y
0
) of the truncated modified equation (4.4) satisfies

h
(y
0
) − ¯ ϕ
N,h
(y
0
)| ≤ hγMe
−h
0
/h
,
where γ = e(2 + 1.65η +µ) depends only on the method.
The proof of this theorem is technical and long; see Hairer et al. (2002,
Section IX.7) for details. We just explain how the assumptions can be
checked for the St¨ ormer–Verlet method (4.7).
We let y = (q, v)
T
, F(y) = (v, f(q))
T
, and we consider the scaled norm
|y| = |q| +h|v|. The quantities R and M are then given by the problem.
The computation of the beginning of Section 4.2 shows that the functions
D
j
(q, v) are composed of derivatives of f(q) so that they are analytic on
the same domain as f(q). To find the constants µ and κ in (4.15), we use
|F(y)| = |v|+h|f(q)| ≤ M for |q−q
0
|+h|v−v
0
| ≤ 2R, and we estimate
_
_
_
_
Φ
h
(q, v) −
_
q +hv
v +
h
2
f(q)
__
_
_
_
=
_
_
_
_
_
h
2
2
f(q)
h
2
f(q +hv +
h
2
2
f(q))
_
_
_
_
_
≤ hM
for |q − q
0
| + h|v − v
0
| ≤ R and for hM ≤ R. This follows from the fact
that the argument of f satisfies |q + hv +
h
2
2
f(q) − q
0
| ≤ R + hM ≤ 2R.
Considered as a function of h, Φ
h
(q, v) is analytic in the complex disc [h[ ≤
R/M. Cauchy’s estimate therefore yields
|D
j
(q, v)| =
1
j!
_
_
_
_
d
j
dh
j
_
Φ
h
(q, v) −
_
q +hv
v +
h
2
f(q)
__¸
¸
¸
¸
h=0
_
_
_
_
≤ M
_
M
R
_
j−1
for j ≥ 2. This proves the estimates (4.15) with µ = 1 and κ = 1/2.
5. Long-time behaviour of numerical solutions
In this section we show how the geometric properties of Section 2 turn into
favourable long-term behaviour. Most of the results are obtained with the
help of backward error analysis.
5.1. Energy conservation
We have seen in Example 3.4 that the total energy H(p, q) of a Hamiltonian
system is not preserved exactly by the St¨ormer–Verlet method. In that ex-
ample it is, however, approximately preserved. Also for the Kepler problem,
Figure 3.1 indicates no drift in the energy. As the following theorem shows,
the Hamiltonian is in fact approximately preserved over very long times for
general Hamiltonian systems.
426 E. Hairer, Ch. Lubich and G. Wanner
Theorem 5.1. The total energy along a numerical solution (p
n
, q
n
) of the
St¨ ormer–Verlet method satisfies
[H(p
n
, q
n
) −H(p
0
, q
0
)[ ≤ Ch
2
+C
N
h
N
t for 0 ≤ t = nh ≤ h
−N
for arbitrary positive integer N. The constants C and C
N
are independent
of t and h. C
N
depends on bounds of derivatives of H up to (N +1)th order
in a region that contains the numerical solution values (p
n
, q
n
).
We give two different proofs of this result, the first one based on the
symplecticity, the second one on the symmetry of the method. When the
Hamiltonian is analytic, both proofs can be refined to yield an estimate
Ch
2
+ C
0
e
−c/h
t over exponentially long times t ≤ e
c/h
, with c proportional
to 1/Ω, where Ω is an upper bound of |M
−1/2

2
U(q)M
−1/2
|
1/2
, i.e., of the
highest frequency in the linearized system.
The first proof uses the symplecticity of the St¨ormer–Verlet method via
backward error analysis, in an argument due to Benettin and Giorgilli (1994).
It applies to general symplectic methods for general (smooth) Hamiltonian
systems (2.7). We know from Theorem 4.3 that the modified differential
equation, truncated after N terms, is again Hamiltonian, with a modified
Hamiltonian
¯
H that is O(h
2
) close to the original Hamiltonian H in a neigh-
bourhood of the numerical solution values. Consider now
¯
H along the nu-
merical solution. We write the deviation of
¯
H as a telescoping sum
¯
H(p
n
, q
n
) −
¯
H(p
0
, q
0
) =
n−1

j=0
_
¯
H(p
j+1
, q
j+1
) −
¯
H(p
j
, q
j
)
_
.
By construction of the modified equation, we have for its flow ¯ ϕ
h
(p
j
, q
j
) =
(p
j+1
, q
j+1
) +O(h
N+1
). On the other hand, the flow ¯ ϕ
t
preserves the mod-
ified Hamiltonian, and hence
¯
H(p
j+1
, q
j+1
) −
¯
H(p
j
, q
j
) =
¯
H(p
j+1
, q
j+1
) −
¯
H( ¯ ϕ
h
(p
j
, q
j
)) = O(h
N+1
).
Inserting this estimate in the above sum yields the result.
The second proof uses only the symmetry of the St¨ ormer–Verlet method.
It was given in Hairer and Lubich (2000b) because its arguments extend
to numerical energy conservation in oscillatory systems when the product
of the step size with the highest frequencies is bounded away from 0 (see
Section 5.4). Backward error analysis, or the asymptotic h
2
-expansion of
the numerical solution, shows that there exists, for every n, a function q
n
(t)
with q
n
(0) = q
n
and q
n
(−h) = q
n−1
+O(h
N+1
) satisfying
q
n
(t +h) −2q
n
(t) +q
n
(t −h) = h
2
f(q
n
(t)) +O(h
N+2
) (5.1)
for t in some fixed interval around 0. The functions q
n
(t + h) and q
n+1
(t)
agree up to O(h
N+1
), as do their kth derivatives multiplied with h
k
,
Geometric numerical integration 427
for k ≤ N. By Taylor expansion in (5.1),
N/2

l=1
2
(2l)!
d
2l
q
n
dt
2l
(t) h
2l−2
= f(q
n
(t)) +O(h
N
). (5.2)
Because of the symmetry of the method, only even-order derivatives of q
n
(t)
(and even powers of the step size) are present in (5.2).
We multiply (5.2) with ˙ q
n
(t)
T
M and integrate over t. The key observation
is now that the product of ˙ q
n
(t) with an even-order derivative of q
n
(t) is a
total differential (we omit the superscript n in the following formula):
˙ q
T
Mq
(2l)
=
d
dt
/
l
[q]
with
/
l
[q] =
_
˙ q
T
Mq
(2l−1)
−¨ q
T
Mq
(2l−2)
+ ∓(q
(l−1)
)
T
Mq
(l+1)
±
1
2
(q
(l)
)
T
Mq
(l)
_
.
In particular, /
1
[q] =
1
2
˙ q
T
M ˙ q. Moreover, for f(q) = −M
−1
∇U(q) we
clearly have ˙ q
T
Mf(q) = −( d/dt)U(q). For the energy functional
H[q](t) =
N/2

l=1
2
(2l)!
/
l
[q](t) h
2l−2
+U(q(t))
we thus obtain ( d/dt)H[q
n
](t) = O(h
N
), and hence
H[q
n
](h) −H[q
n
](0) = O(h
N+1
). (5.3)
Since the functions q
n
(t+h) and q
n+1
(t), together with their kth derivatives
scaled by h
k
(k ≤ N), are equal up to O(h
N+1
), we further have
H[q
n+1
](0) −H[q
n
](h) = O(h
N+1
). (5.4)
Moreover, with p
n
(t) = M ˙ q
n
(t) we have
H[q
n
](0) = H(p
n
(0), q
n
(0)) +O(h
2
) = H(p
n
, q
n
) +O(h
2
), (5.5)
where the last equation follows by noting
p
n
= M
q
n+1
−q
n−1
2h
= p
n
(0) +O(h
2
).
Hence, from (5.3)–(5.5),
H(p
n
, q
n
) −H(p
0
, q
0
) = H[q
n
](0) −H[q
0
](0) +O(h
2
)
= O(nh
N+1
) +O(h
2
),
which completes the proof.
428 E. Hairer, Ch. Lubich and G. Wanner
5.2. Linear error growth for integrable systems
General Hamiltonian systems may have extremely complicated dynamics,
and little can be said about the long-time behaviour of their discretizations
apart from the long-time near-conservation of the total energy considered
above. At the other end, the simplest conceivable dynamics – uniform mo-
tion on a Cartesian product of circles – appears in integrable Hamiltonian
systems. Their practical interest lies in the fact that many physical sys-
tems are perturbations of integrable systems, with planetary motion as the
classical example and historical driving force.
A Hamiltonian system (2.7) is integrable if there exists a symplectic trans-
formation
(p, q) = ψ(a, θ) (5.6)
to action-angle variables (a, θ), defined for actions a = (a
1
, . . . , a
d
) in some
open set of R
d
and for angles θ on the whole d-dimensional torus
T
d
= R
d
/(2πZ
d
) = ¦(θ
1
, . . . , θ
d
); θ
i
∈ R mod 2π¦,
such that the Hamiltonian in these variables depends only on the actions
H(p, q) = H(ψ(a, θ)) = K(a). (5.7)
In the action-angle variables, the equations of motion are simply
˙ a = 0,
˙
θ = ω(a), (5.8)
with the frequencies ω = (ω
1
, . . . , ω
d
)
T
= ∇
a
K (note that ∇
θ
K = 0). This
has a quasi-periodic (or possibly periodic) flow:
ϕ
t
(a, θ) = (a, θ +ω(a)t). (5.9)
For every a, the torus ¦(a, θ) : θ ∈ T
d
¦ is thus invariant under the flow. We
express the actions and angles in terms of the original variables (p, q) via
the inverse transform of (5.6) as
(a, θ) = (I(p, q), Θ(p, q)),
and note that the components of I = (I
1
, . . . , I
d
) are first integrals of the
integrable system.
Integrability of a Hamiltonian system is an exceptional property: the
system has d independent first integrals I
1
, . . . , I
d
whose Poisson brackets
vanish pairwise, that is,
¦I
i
, I
j
¦ = ∇
q
I
T
i

p
I
j
−∇
p
I
T
i

q
I
j
= 0 for all i, j.
The solution trajectories of the Hamiltonian systems with Hamiltonian I
i
exist for all time (in the action-angle variables, their flow is simply ϕ
[i]
t
(a, θ) =
Geometric numerical integration 429
−2
−1
0
1
2
0
1
2
q
p
ψ
−1
π
2
π 3π
2

θ
a
Figure 5.1. Transformation to action-angle variables.
(a, θ + te
i
) with e
i
denoting the ith unit vector of R
d
), and the level sets
of I are compact (the invariant tori ¦a = Const θ ∈ T
d
¦). Conversely, the
Arnold–Liouville theorem (Arnold 1963) states that every Hamiltonian sys-
tem having d first integrals with the above properties can be transformed to
action-angle variables with a Hamiltonian depending only on the actions.
Example 5.2. The harmonic oscillator H(p, q) =
1
2
p
2
+
1
2
q
2
is integrable,
with the transformation to action-angle coordinates given by
_
p
q
_
=
_√
2a cos θ

2a sin θ
_
,
with a = H(p, q): see Figure 5.1. Here, the action-angle coordinates are
symplectic polar coordinates.
Example 5.3. The Kepler problem, with H(p, q) =
1
2
(p
2
1
+p
2
2
)−(q
2
1
+q
2
2
)

1
2
in the range H < 0, is integrable with actions a
1
= 1/

−2H and a
2
= L (the
angular momentum, L = q
1
p
2
−q
2
p
1
). The frequencies are ω
1
= ω
2
= 2π/T,
where T = 2π/(−2H)
3/2
is the period of a trajectory with total energy H.
Example 5.4. A further celebrated example of an integrable system is
the Toda lattice (Toda 1970, Flaschka 1974), which describes a system of
particles on a line interacting with exponential forces. The Hamiltonian is
H(p, q) =
d

k=1
_
1
2
p
2
k
+ exp(q
k
−q
k+1
)
_
430 E. Hairer, Ch. Lubich and G. Wanner
1.94
1.96
1.94
1.96
.64
.66
.64
.66
10 20 30 40 50
−2.62
−2.60
−2.58
10 20 30 40 50
−2.62
−2.60
−2.58
λ
h = 0.1
λ
1
h = 0.05
λ
1
λ
2
λ
2
Runge 2
St¨ ormer–Verlet
t
λ
3
t
λ
3
Figure 5.2. Toda eigenvalues along the numerical solution.
with periodic extension q
d+1
= q
1
. The eigenvalues of the matrix
L =
_
_
_
_
_
_
_
a
1
b
1
b
d
b
1
a
2
b
2
0
b
2
.
.
.
.
.
.
0
.
.
. a
d−1
b
d−1
b
d
b
d−1
a
d
_
_
_
_
_
_
_
,
a
k
= −
1
2
p
k
,
b
k
=
1
2
exp
_
1
2
(q
k
−q
k+1
)
_
are first integrals whose Poisson brackets vanish pairwise.
We consider the case d = 3 and choose initial values q
0
= (1, 2, −1)
T
and
p
0
= (−1.5, 1, 0.5)
T
. Figure 5.2 shows the eigenvalues of L along the nu-
merical solution of the St¨ ormer–Verlet and the second-order Runge method
obtained with step sizes h = 0.1 (left) and h = 0.05 (right) on the interval
0 ≤ t ≤ 50. Not only the Hamiltonian (Theorem 5.1), but all d first integrals
of the integrable system are well approximated over long times with an error
of size O(h
2
). This is explained by Theorem 5.5 below.
The global error in (p, q) is plotted in Figure 5.3. We observe a linear
error growth for the St¨ ormer–Verlet method, in contrast to a quadratic error
growth for the second-order Runge method.
The study of the error behaviour of the numerical method combines
backward error analysis, by means of which the numerical map is inter-
preted as being essentially the time-h flow of a modified Hamiltonian sys-
tem, and the perturbation theory of integrable systems, a rich mathematical
theory originally developed for problems of celestial mechanics (Poincar´e
1892/1893/1899, Siegel and Moser 1971). The effect of a small perturbation
Geometric numerical integration 431
50 100
.0
.2
.4
.6
.8 global error
St¨ ormer–Verlet, h = 0.02
Runge 2, h = 0.02
global error
St¨ ormer–Verlet, h = 0.02
Runge 2, h = 0.02
Figure 5.3. Global error of the St¨ ormer–Verlet and
the second-order Runge method on the Toda lattice.
of an integrable system is well under control in subsets of the phase space
where the frequencies ω satisfy Siegel’s diophantine condition:
[k ω[ ≥ γ[k[
−ν
for all k ∈ Z
d
(5.10)
for some positive constants γ and ν, with [k[ =

i
k
i
. For ν > d−1, almost
all frequencies (in the sense of the Lebesgue measure) satisfy (5.10) for some
γ > 0. For any choice of γ and ν the complementary set is, however, open
and dense in R
d
.
For general numerical integrators applied to integrable systems (or per-
turbations thereof) the error grows quadratically with time, and there is a
linear drift away from the first integrals I
i
. For symplectic methods such
as the St¨ ormer–Verlet method there is linear error growth and long-time
near-preservation of the first integrals I
i
, as is shown by the following result
from Hairer et al. (2002, Section X.3).
Theorem 5.5. Consider applying the St¨ ormer–Verlet method to an inte-
grable system (2.7) with real-analytic Hamiltonian. Suppose that ω

∈ R
d
satisfies the diophantine condition (5.10). Then there exist positive con-
stants C, c and h
0
such that the following holds for all step sizes h ≤ h
0
:
every numerical solution (p
n
, q
n
) starting with frequencies ω
0
= ω(I(p
0
, q
0
))
such that |ω
0
−ω

| ≤ c[ log h[
−ν−1
, satisfies
|(p
n
, q
n
) −(p(t), q(t))| ≤ C t h
2
,
|I(p
n
, q
n
) −I(p
0
, q
0
)| ≤ C h
2
,
for t = nh ≤ h
−2
.
The constants h
0
, c, C depend on d, γ, ν and on bounds of the Hamiltonian.
The basic steps of the proof are summarized in Figure 5.4. By backward
error analysis, the numerical method coincides, up to arbitrary order in h,
with the flow of the modified differential equation, which is a Hamiltonian
432 E. Hairer, Ch. Lubich and G. Wanner
backward error
analysis
integrable
Hamiltonian
H(p, q)
numerical
solution
(p
n
, q
n
)
modified
Hamiltonian
¯
H(p, q)
Hamiltonian
K(a)
action-angle
variables
(p, q) = ψ(a, θ)
modified Hamiltonian
K(a) +ε
¯
K
1
(a, θ)
with ε = h
p
Lindstedt–Poincar´e
series
(a, θ) = χ(b, ϕ)
modified Hamiltonian
K(b) +εK
1
(b) + +ε
N
K
N
(b)

N+1
R(b, ϕ)
Figure 5.4. Transformations in the proof of Theorem 5.5.
perturbation of size ε = h
2
of the original, integrable system. We are thus in
the realm of classical perturbation theory. In addition to the transformation
to the action-angle variables (a, θ), which gives the modified Hamiltonian in
the form K(a) +ε
¯
K
1
(a, θ), we use a further symplectic coordinate transfor-
mation (a, θ) = χ(b, ϕ) which eliminates, up to high-order terms in ε, the
dependence on the angles in the modified Hamiltonian. This transformation
is O(ε) close to the identity. It is constructed as
b = a −∇
θ
S(b, θ), ϕ = θ +∇
b
S(b, θ),
where the generating function S(b, θ) is given by a Lindstedt–Poincar´e per-
turbation series,
S(b, θ) = εS
1
(b, θ) +ε
2
S
2
(b, θ) + +ε
N
S
N
(b, θ).
The error propagation is then studied in the (b, ϕ)-variables, with the result
|b(t) −b
0
| ≤ C t ε
N+1
,
|ϕ(t) −ϕ
0
−ω
ε
(b
0
) t| ≤ C (t +t
2
) ε
N
,
for t
2
≤ 1/ε
N
,
with ω
ε
(b) = ω(b) +O(ε). Transforming back to the original variables (p, q)
finally yields the stated result.
Theorem 5.5 admits extensions in several directions. It is just one of a
series of results on the long-time behaviour of geometric integrators.
Geometric numerical integration 433
• The theorem does not apply directly to the Kepler problem, which
has two identical frequencies ω
1
= ω
2
= (−2H)
3/2
. However, since
the angular momentum a
2
= L is preserved exactly by the St¨ ormer–
Verlet method, it turns out that the modified Hamiltonian written
in the action-angle variables of the Kepler problem is independent of
the angle θ
2
. Only the angle θ
1
must therefore be eliminated via the
perturbation series, and this involves only the single frequency ω
1
for
which the diophantine condition is trivially satisfied. The proof and
result of Theorem 5.5 thus extend to the Kepler problem.
• The linear error growth remains intact when the method is applied
to perturbed integrable systems H(p, q) + εG(p, q) with a perturbation
parameter of size ε = O(h
α
) for some positive exponent α.
• Under stronger conditions on the initial values or on the system, the
near-preservation of the action variables along the numerical solution
holds for times that are exponentially long in a negative power of the
step size (Hairer and Lubich 1997, Moan 2002). For a Cantor set of
initial values and a Cantor set of step sizes this holds even perpetually,
in view of the existence of invariant tori of the numerical integrator
close to the invariant tori of the integrable system (Shang 1999, 2000).
• Perturbed integrable systems have KAM tori, i.e., deformations of the
invariant tori of the integrable system corresponding to diophantine fre-
quencies ω, which are invariant under the flow of the perturbed system.
If the method is applied to such a perturbed integrable system, then the
numerical method has tori which are near-invariant over exponentially
long times (Hairer and Lubich 1997). For a Cantor set of non-resonant
step sizes there are even truly invariant tori on which the numerical
one-step map reduces to rotation by hω in suitable coordinates (Hairer
et al. 2002, p. 371).
• There is a completely analogous theory for integrable reversible systems
(Hairer et al. 2002, Chapter XI). These are differential equations with
reversible flow (2.2), which are transformed to the form (5.8) by a trans-
formation (q, v) = (µ(a, θ), ν(a, θ)) that preserves reversibility, i.e., µ
is odd in θ and ν is even in θ. In that theory, only the reversibility
of the numerical method comes into play, not the symplecticity. There
is again linear error growth, long-time near-preservation of the action
variables, and an abundance of invariant tori.
• For dissipatively perturbed integrable systems, where only one torus
survives the perturbation and becomes weakly attractive, the existence
of a nearby invariant torus of the numerical method is shown under
weak assumptions on the step size in Stoffer (1998) and Hairer and
Lubich (1999).
434 E. Hairer, Ch. Lubich and G. Wanner
5.3. Statistical behaviour
The equation of motion of a system of 864 particles interacting through a Lennard-
Jones potential has been integrated for various values of the temperature and
density, relative, generally, to a fluid state. The equilibrium properties have been
calculated and are shown to agree very well with the corresponding properties of
argon. L. Verlet (1967)
In molecular dynamics, it is the computation of statistical or thermodynamic
quantities, such as temperature, which is of interest, rather than single tra-
jectories. The success of the St¨ormer–Verlet method in this field lies in the
observation that the method is apparently able to reproduce the correct
statistical behaviour over long times. Since Verlet (1967), this has been
confirmed in countless computational experiments. Backward error analysis
gives indications as to why this might be so, but to our knowledge there
are as yet no rigorous mathematical results in the literature explaining the
favourable statistical behaviour.
In the following we derive a result which is a discrete analogue of the
virial theorem of statistical mechanics; cf. Abraham and Marsden (1978,
p. 243) and Gallavotti (1999, p. 129). It comes as a consequence of the long-
time near-conservation of energy. Consider the Poisson bracket ¦F, H¦ =

q
F
T

p
H −∇
p
F
T

q
H of an arbitrary differentiable function F(p, q) with
the Hamiltonian. Along every solution (p(t), q(t)) of the Hamiltonian system
we have
¦F, H¦(p(t), q(t)) =
d
dt
F(p(t), q(t)),
and hence the time average of the Poisson bracket along a solution is
1
T
_
T
0
¦F, H¦(p(t), q(t)) dt =
1
T
_
F(p(T), q(T)) −F(p(0), q(0))
_
.
If F is bounded along the solution, this shows that the average is of size
O(1/T) as T → ∞. In particular, this condition is satisfied if the energy
level set ¦(p, q) : H(p, q) = H(p(0), q(0))¦ is compact.
Example 5.6. For a separable Hamiltonian (2.8) the choice F(p, q) = p
T
q
yields the virial theorem of Clausius (Gallavotti 1999, p. 129),
lim
T→∞
1
T
_
T
0
p(t)
T
M
−1
p(t) dt = lim
T→∞
1
T
_
T
0
q(t)
T
∇U(q(t)) dt,
i.e., the time average of twice the kinetic energy equals that of the virial
function q
T
∇U(q).
Geometric numerical integration 435
For the numerical discretization there is the following result.
Theorem 5.7. Let H(p, q) be a real-analytic Hamiltonian for which
K
δ
= ¦(p, q) : [H(p, q) −H
0
[ ≤ δ¦ is compact
for some δ > 0. Let F(p, q) be any smooth real-valued function, bounded
by µ on K
δ
. Then the numerical solution (p
n
, q
n
) obtained by the St¨ ormer–
Verlet method satisfies
¸
¸
¸
¸
1
N
N

n=0

¦F, H¦(p
n
, q
n
)
¸
¸
¸
¸


Nh
+Ch
2
for Nh ≤ e
c/h
and h ≤ h
0
,
(5.11)
where the prime on the sum indicates that the first and last term are taken
with weight
1
2
. The constants C, c, h
0
> 0 depend on bounds of H on a
complex neighbourhood of K
δ
and on bounds of the first three derivatives
of F on K
δ
, but they are independent of h and (p
0
, q
0
) ∈ K
δ/2
.
In particular, the left-hand side of (5.11) is O(h
2
) for h
−2
≤ Nh ≤ e
c/h
.
Proof. By Theorem 5.1 and the remark thereafter, we know that
y
n
:= (p
n
, q
n
) ∈ K
δ
for nh ≤ e
c/h
.
Since y
n+1
= ϕ
h
(y
n
) +O(h
3
) and ¦F, H¦(ϕ
t
(y)) =
d
dt
F(ϕ
t
(y)), we have
h
2
¦F, H¦(y
n
) +
h
2
¦F, H¦(y
n+1
) =
_
h
0
¦F, H¦
_
ϕ
t
(y
n
)
_
dt +O(h
3
)
= F(y
n+1
) −F(y
n
) +O(h
3
).
Hence
1
N
N

n=0

¦F, H¦(y
n
) =
1
Nh
_
F(y
N
) −F(y
0
)
_
+O(h
2
),
which yields the stated estimate.
Example 5.8. We give a numerical experiment with a small-scale version
of Verlet’s argon model. It considers N
A
atoms interacting by the Lennard-
Jones potential
V (r) = 4 ε
__
σ
r
_
12

_
σ
r
_
6
_
.
The Hamiltonian of the system is (3.2) with V
ij
= V . We choose N
A
= 7
436 E. Hairer, Ch. Lubich and G. Wanner
−1280
−1260
−1240
0
20
40
−5000
0
5000
total energy
temperature
virial function
Figure 5.5. Computed total energy, temperature and virial function
of the argon crystal, 10 000 steps of size h = 40 [fsec].
and the data of Biesiadecki and Skeel (1993); see also Hairer et al. (2002,
p. 15). Figure 5.5 shows the Hamiltonian, the temperature
T(p) =
1
N
A
k
B
1
2 m
N
A

i=1
|p
i
|
2
(k
B
is Boltzmann’s constant), and the virial function
C(q) =
N
A

i=2
i−1

j=1
V

(r
ij
)r
ij
(with r
ij
= |q
i
− q
j
|) over an interval of length 4 10
5
[fsec], obtained by
the St¨ ormer–Verlet method with step size h = 40 [fsec]. The units in the
figure are such that k
B
= 1. The size of the oscillations in the Hamiltonian
is proportional to h
2
, whereas that in the temperature and in the virial
function is independent of h. At the end of the integration (after 10 000
steps) the averages of twice the kinetic energy and of the virial function are
217.1 and 217.8, respectively.
5.4. Oscillatory differential equations
Nonlinear mass-spring models have traditionally been very useful in explain-
ing various phenomena of more complicated ‘real’ physical systems. We
have already mentioned the Toda lattice. An equally famous problem is the
Fermi–Pasta–Ulam model (Fermi, Pasta and Ulam 1955, Ford 1992), where
Geometric numerical integration 437
a nonlinear perturbation to a primarily linear problem is studied over long
times. Here we use a variant of this problem for gaining insight into the long-
time energy behaviour of the St¨ ormer–Verlet method applied to oscillatory
systems with multiple time scales. We are interested in using step sizes h for
which the product with the highest frequency ω in the system is bounded
away from zero. (Values of hω ≈ 1/2 are routinely used in molecular dynam-
ics.) In this situation, backward error analysis is no longer applicable, since
the ‘exponentially small’ error terms are then of size O(e
−c/hω
) = O(1).
x
1
x
2
x
2m−1 x
2m
stiff
harmonic
soft
nonlinear
Figure 5.6. Chain with alternating soft nonlinear
and stiff linear springs.
Example 5.9. Consider a chain of 2m mass points, connected with alter-
nating soft nonlinear and stiff linear springs, and fixed at the end points; see
Galgani, Giorgilli, Martinoli and Vanzini (1992) and Figure 5.6. The vari-
ables x
1
, . . . , x
2m
stand for the displacements of the mass points. In terms
of the new variables
q
i
= (x
2i
+x
2i−1
)/

2, q
m+i
= (x
2i
−x
2i−1
)/

2
(which represent a scaled displacement and a scaled expansion/compression
of the ith stiff spring) and the momenta p
i
= ˙ q
i
, the motion is described by
a Hamiltonian system with
H(p, q) =
1
2
2m

i=1
p
2
i
+
ω
2
2
m

i=1
q
2
m+i
+
1
4
_
(q
1
−q
m+1
)
4
+
m−1

i=1
(q
i+1
−q
m+i+1
−q
i
−q
m+i
)
4
+ (q
m
+q
2m
)
4
_
,
where ω ¸1 is a large parameter. Here we assume cubic nonlinear springs,
but the special form of the nonlinearity is not important.
For an illustration we consider m = 3 and choose ω = 30. In Figure 5.7 we
have plotted the following quantities as functions of time: the Hamiltonian
H (actually we plot H −0.8 for graphical reasons), the oscillatory energy I
defined as
I = I
1
+I
2
+I
3
with I
j
=
1
2
p
2
m+j
+
1
2
ω
2
q
2
m+j
,
438 E. Hairer, Ch. Lubich and G. Wanner
.04 .08
0
1
1 2
0
1
50 100 150
0
1
2500 5000 7500
0
1
H
I
T
2
H
I
T
1
T
2
H
I
I
1
I
2
I
3
H
I
h = 0.5/ω
Figure 5.7. Different time scales in a Fermi–Pasta–Ulam problem,
and energy conservation of the St¨ ormer–Verlet method (last picture).
and the kinetic energies of the mass centre motion and of the relative motion
of masses joined by a stiff spring,
T
1
=
1
2
_
p
2
1
+p
2
2
+p
2
3
_
, T
2
=
1
2
_
p
2
4
+p
2
5
+p
2
6
_
.
The system has different dynamics on several time scales: on the fast scale
ω
−1
the motion is nearly harmonic in the stiff springs, on scale ω
0
there is
the motion of the soft springs driven by the nonlinearity, and on the slow
scale ω there is an energy exchange between the stiff linear springs. For the
first three pictures the solutions were computed with high accuracy.
In the last picture we show the results obtained by the St¨ ormer–Verlet
method with step size h = 0.5/ω. We note that both H and I are approxi-
mately conserved over long times. For fixed ω the size of the oscillations in
H is proportional to h
2
. However, the oscillations remain of the same size if
h decreases and ω increases such that hω remains constant. The oscillations
in I are of size O(ω
−1
) uniformly for h →0.
The equations of motion for the above example are of the form
¨ q = −Ω
2
q −∇U(q) with Ω =
_
0 0
0 ωI
_
, (5.12)
with a single high frequency ω ¸1 and with a smooth potential U(q) whose
derivatives are bounded independently of ω. In addition to the total energy
Geometric numerical integration 439
as a conserved quantity,
H(p, q) =
1
2
p
T
p +
1
2
q
T

2
q +U(q),
the system has an adiabatic invariant: over times exponentially long in ω,
the oscillatory energy
I(p, q) =
1
2
p
T
_
0 0
0 I
_
p +
ω
2
2
q
T
_
0 0
0 I
_
q (5.13)
is preserved up to O(ω
−1
). This holds uniformly for all initial values for
which the total energy is bounded by a constant independent of ω, i.e., for
bounded (p, q) with (0 I)q = O(ω
−1
).
Now consider applying the St¨ ormer–Verlet method to such a system. The
step size is then restricted to hω < 2 for linear stability, as Example 3.4
shows. The Hamiltonian H(p
n
, q
n
) and the oscillatory energy I(p
n
, q
n
) of
(5.13) oscillate rapidly, but stay within an O((hω)
2
) band over long times.
The oscillations do not become smaller when h is decreased but ω is increased
such that their product hω is kept fixed. Nevertheless, the following result
shows that the time averages of the total and oscillatory energies
H
n
=
h
T

|jh|≤T/2
H(p
n+j
, q
n+j
),
I
n
=
h
T

|jh|≤T/2
I(p
n+j
, q
n+j
),
for an arbitrary fixed T > 0, remain constant up to O(h) over long times
even when hω is bounded away from zero, but within the range of linear
stability.
Theorem 5.10. Let the St¨ormer–Verlet method be applied to the problem
(5.12) with a step size h for which 0 < c
0
≤ hω ≤ c
1
< 2. Let ¯ ω be
defined by the relation sin(
1
2
h¯ ω) =
1
2
hω and suppose [ sin(
1
2
kh¯ ω)[ ≥ c

h for
k = 1, . . . , N for some N ≥ 2 and c > 0. Suppose further that the total
energy at the initial value (p
0
, q
0
) is bounded independently of ω, and that
the numerical solution values q
n
stay in a region where all derivatives of
the potential U are bounded. Then, the time averages of the total and the
oscillatory energy along the numerical solution satisfy
H
n
= H
0
+O(h),
I
n
= I
0
+O(h),
for 0 ≤ nh ≤ h
−N+1
.
The constants symbolized by O are independent of n, h, ω with the above
conditions.
440 E. Hairer, Ch. Lubich and G. Wanner
It should, however, be noted that the time averages H
n
and I
n
do not, in
general, remain O(h) close to the initial values H(p
0
, q
0
) and I(p
0
, q
0
).
The estimates of Theorem 5.10 can be improved to O(h
2
) if a weighted
time average is taken, replacing the characteristic function of the interval
[−T/2, T/2] by a smooth windowing function with bounded support, and
if the oscillatory energy I is replaced by J(p, q) = I(p, q) + q
T
_
0 0
0 I
_
∇U(q),
which is preserved up to O(ω
−2
) over exponentially long time intervals.
For hω → 0, the long-time near-preservation of the adiabatic invariant I
can be shown using backward error analysis (Reich 1999b), but this argu-
ment breaks down for hω bounded away from zero as in Theorem 5.10.
We comment only briefly on the proof of Theorem 5.10; see Hairer et al.
(2002, Chapter XIII) for the full proof. It is based on representing the
numerical solution locally (on bounded intervals) by a modulated Fourier
expansion
q
n
=

|k|<N
z
k
(t)e
ik¯ ωt
+O(h
N
) for t = nh,
where the coefficients z
k
(t) together with all their derivatives (up to some
arbitrarily fixed order) are bounded by O(¯ ω
−|k|
). A similar representation
holds for p
n
. The expansion coefficients y
k
(t) = z
k
(t)e
ik¯ ωt
satisfy a system
of equations similar in structure to (5.2) (but of higher dimension). This
permits us to use similar arguments to the second proof of Theorem 5.1
to infer the existence of certain modified energies H

and I

, which the
numerical method preserves up to O(h) over times h
−N+1
. Finally, the time
averages H
n
and I
n
can be expressed, up to O(h), in terms of these modified
energies.
6. Constrained Hamiltonian systems
A minimal set of coordinates of a mechanical system is often difficult to
find. The minimal coordinates may be defined only implicitly, or frequent
changes of charts are necessary along a solution of the system. In this
situation it is favourable to formulate the problem as a Hamiltonian system
with constraints.
6.1. Formulation as differential-algebraic equations
We consider a mechanical system with coordinates q ∈ R
d
that are subject
to constraints g(q) = 0. The equations of motion are then of the form
˙ p = −∇
q
H(p, q) −∇
q
g(q)λ, (6.1)
˙ q = ∇
p
H(p, q), 0 = g(q),
Geometric numerical integration 441
where the Hamiltonian H(p, q) is usually given by (2.8). Here, p and q are
vectors in R
d
, g(q) =
_
g
1
(q), . . . , g
m
(q)
_
T
is the vector of constraints, and

q
g =
_

q
g
1
, . . . , ∇
q
g
m
_
is the transposed Jacobian matrix of g(q).
To compute the Lagrange multiplier λ, we differentiate the constraint
0 = g
_
q(t)
_
with respect to time. This yields the so-called hidden constraint
0 = ∇
q
g(q)
T

p
H(p, q), (6.2)
which is an invariant of the flow of (6.1). A further differentiation gives
0 =

∂q
_

q
g(q)
T

p
H(p, q)
_

p
H(p, q)
−∇
q
g(q)
T

2
p
H(p, q)
_

q
H(p, q) +∇
q
g(q)λ
_
, (6.3)
which allows us to express λ in terms of (p, q), if the matrix

q
g(q)
T

2
p
H(p, q)∇
q
g(q) is invertible (6.4)
(∇
2
p
H denotes the Hessian matrix of H). Inserting the so-obtained function
λ(p, q) into (6.1) gives the ordinary differential equation
˙ p = −∇
q
H(p, q) −∇
q
g(q)λ(p, q),
˙ q = ∇
p
H(p, q) (6.5)
for (p, q), which is well defined on the domain where H(p, q) and g(q) are
defined, and not only for g(q) = 0. The standard theory for ordinary dif-
ferential equations can be used to deduce existence and uniqueness of the
solution. Important properties of the system (6.1) are the following.
• Whenever the initial values satisfy (p
0
, q
0
) ∈ / with
/=
_
(p, q) : g(q) = 0, ∇
q
g(q)
T

p
H(p, q) = 0
_
, (6.6)
the solution stays on the manifold / for all t; hence, the flow of (6.1)
is a mapping ϕ
t
: /→/.
• The flow ϕ
t
is a symplectic transformation on /, which means that

t
(p, q)ξ)
T
J ϕ

t
(p, q)η = ξ
T
J η for ξ, η ∈ T
(p,q)
/. (6.7)
Here, T
(p,q)
/ denotes the tangent space of / at (p, q) ∈ /, and the
product ϕ

t
(p, q)ξ has to be interpreted as the directional derivative on
the manifold.
• For Hamiltonians satisfying
H(−p, q) = H(p, q),
the flow ϕ
t
is ρ-reversible for ρ(p, q) = (−p, q) in the sense that (2.5)
holds for (p, q) ∈ /.
442 E. Hairer, Ch. Lubich and G. Wanner
The first of these properties follows from the definition of λ(p, q). For
(p
0
, q
0
) ∈ /, a first integration of (6.3) gives (6.2) and a second integration
yields g(q) = 0 along the solution of (6.5).
To prove the symplecticity, we consider the (unconstrained) Hamiltonian
system with K(p, q) = H(p, q) + g(q)
T
λ(p, q). Its flow is symplectic and
coincides with that of (6.5) on the manifold /.
The reversibility is a consequence of the fact that H(−p, q) = H(p, q)
implies λ(−p, q) = λ(p, q). The flow of (6.5) and hence also its restriction
onto / is thus ρ-reversible.
Example 6.1. (Kepler and two-body problems on the sphere)
Following Kozlov and Harin (1992), we consider a particle moving on the
unit sphere attracted by a fixed point a on the sphere. The potential is given
as a fundamental solution of the Laplace–Beltrami equation on the sphere:
U(q, a) = −
cos ϑ
sin ϑ
, cos ϑ = ¸q, a). (6.8)
The Kepler problem on the sphere is then of the form (6.1) with
H(p, q) =
1
2
p
T
p +U(q, a), g(q) = q
T
q −1.
The left picture of Figure 6.1 shows the solution corresponding to the point
a = (0.3

2, 0.3

2, 0.8)
T
and to initial values given in spherical coordinates
by ϕ
0
= 1, θ
0
= 1.1 and ˙ ϕ
0
= 1.2,
˙
θ
0
= −1.1. The point a and the initial
value are indicated by a larger symbol in Figure 6.1.
Figure 6.1. Solutions of the Kepler problem (left)
and the two-body problem on the sphere (right).
Geometric numerical integration 443
Whereas in Euclidean space the two-body problem reduces to the Kepler
problem, this is not the case on the sphere. For the two-body problem the
Hamiltonian is
H(p
1
, p
2
, q
1
, q
2
) =
1
2
p
T
1
p
1
+
1
2
p
T
2
p
2
+U(q
1
, q
2
)
with U(q
1
, q
2
) given by (6.8). The constraints are g
i
(q
1
, q
2
) = q
T
i
q
i
− 1 for
i = 1, 2. The solution with initial values ϕ
10
= −0.3, θ
10
= 1.1, ϕ
20
=
−0.8, θ
20
= 0.6 and ˙ ϕ
10
= 0.9,
˙
θ
10
= −0.5, ˙ ϕ
20
= 0.3,
˙
θ
20
= −0.1 is plotted
in Figure 6.1 (right).
Example 6.2. (Rigid body) The motion of a rigid body with a fixed
point chosen at the origin can be described by an orthogonal matrix Q(t).
Letting I
1
, I
2
, I
3
denote the moments of inertia of the body, its kinetic en-
ergy is
T =
1
2
_
I
1

2
1
+I
2

2
2
+I
3

2
3
_
,
where the angular velocity Ω = (Ω
1
, Ω
2
, Ω
3
)
T
of the body is defined by
´
Ω =
_
_
0 −Ω
3

2

3
0 −Ω
1
−Ω
2

1
0
_
_
= Q
T
˙
Q,
(Arnold 1989, Chapter 6). In terms of Q, the kinetic energy on the manifold
O(3) = ¦Q[ Q
T
Q = I¦ becomes
T =
1
2
trace (
´
ΩD
´

T
) =
1
2
trace (Q
T
˙
QD
˙
Q
T
Q) =
1
2
trace (
˙
QD
˙
Q
T
),
where D = diag (d
1
, d
2
, d
3
) is given by the relations I
1
= d
2
+d
3
, I
2
= d
3
+d
1
,
and I
3
= d
1
+d
2
. With P = ∂T/∂
˙
Q =
˙
QD, we are thus concerned with
H(P, Q) =
1
2
trace (PD
−1
P
T
) +U(Q),
and the constrained Hamiltonian system becomes
˙
P = −∇
Q
U(Q) −QΛ,
˙
Q = PD
−1
, 0 = Q
T
Q−I, (6.9)
where Λ is a symmetric matrix consisting of Lagrange multipliers. This is
of the form (6.1) and satisfies the regularity condition (6.4).
6.2. Development of the Rattle algorithm
The most important numerical algorithm for the solution of constrained
Hamiltonian systems is an adaptation of the St¨ ormer–Verlet method. Its
historical development is in three main steps.
444 E. Hairer, Ch. Lubich and G. Wanner
First step. For Hamiltonians H(p, q) =
1
2
p
T
M
−1
p + U(q) with constant
mass matrix M (cf. Section 2.2), the problem is a second-order differential
equation M¨ q = −∇
q
U(q) − ∇
q
g(q)λ with constraint g(q) = 0. The most
natural extension of (1.2) is
q
n+1
−2q
n
+q
n−1
= −h
2
M
−1
_

q
U(q
n
) +∇
q
g(q
n

n
_
,
0 = g(q
n+1
). (6.10)
This algorithm (called Shake) was originally proposed by Ryckaert, Ciccotti
and Berendsen (1977) for computations in molecular dynamics. The p-
components, not used in the recursion, are approximated by p
n
= M(q
n+1

q
n−1
)/2h.
Second step. A one-step formulation of this method, obtained by a formal
analogy to formula (1.5), reads
p
n+1/2
= p
n

h
2
_

q
U(q
n
) +∇
q
g(q
n

n
_
,
q
n+1
= q
n
+hM
−1
p
n+1/2
, 0 = g(q
n+1
), (6.11)
p
n+1
= p
n+1/2

h
2
_

q
U(q
n+1
) +∇
q
g(q
n+1

n+1
_
.
This formula cannot be implemented, because λ
n+1
is not yet available at
this step (it is computed together with q
n+2
). As a remedy, Andersen (1983)
suggests replacing the last line in (6.11) with the projection step
p
n+1
= p
n+1/2

h
2
_

q
U(q
n+1
) +∇
q
g(q
n+1

n
_
,
0 = ∇
q
g(q
n+1
)
T
M
−1
p
n+1
. (6.12)
This modification, called Rattle, is motivated by the fact that the numerical
approximation (p
n+1
, q
n+1
) lies on the solution manifold /.
Third step. Jay (1994) and Reich (1993) observed independently that the
Rattle method can be interpreted as a partitioned Runge–Kutta method
and thus allows the extension to general Hamiltonians
p
n+1/2
= p
n

h
2
_

q
H(p
n+1/2
, q
n
) +∇
q
g(q
n

n
_
,
q
n+1
= q
n
+
h
2
_

p
H(p
n+1/2
, q
n
) +∇
p
H(p
n+1/2
, q
n+1
)
_
,
0 = g(q
n+1
), (6.13)
p
n+1
= p
n+1/2

h
2
_

q
H(p
n+1/2
, q
n+1
) +∇
q
g(q
n+1

n
_
,
0 = ∇
q
g(q
n+1
)
T

p
H(p
n+1
, q
n+1
)
whenever (p
n
, q
n
) ∈ /. The first three equations of (6.13) determine
(p
n+1/2
, q
n+1
, λ
n
), whereas the remaining two are equations for (p
n+1
, µ
n
).
Geometric numerical integration 445
For a sufficiently small step size, these equations have a locally unique solu-
tion (Hairer et al. 2002, p. 214).
Example 6.3. (Kepler problem on the sphere) We apply the Rattle
method with a large step size h = 0.07 to the problem of Example 6.1.
The numerical solution, plotted in Figure 6.2, shows a precession as it ap-
pears in computations with symplectic integrators for the Kepler problem
in Euclidean space; see Figure 1.5. We remark that the value of the Hamil-
tonian along the numerical solution oscillates around the correct value and
the energy error remains bounded by 0.114 on very long time intervals.
Since the constraint g(q) is quadratic and the Hamiltonian is separable,
the formulae (6.13) are explicit with exception of the computation of λ
n
, for
which a scalar quadratic equation needs to be solved.
Example 6.4. (Rigid body) The Rattle method (6.13) applied to (6.9)
yields
P
1/2
= P
0

h
2

Q
V (Q
0
) −
h
2
Q
0
Λ
1
,
Q
1
= Q
0
+hP
1/2
D
−1
, Q
T
1
Q
1
= I, (6.14)
P
1
= P
1/2

h
2

Q
V (Q
1
) −
h
2
Q
1
Λ
2
, D
−1
P
T
1
Q
1
+Q
T
1
P
1
D
−1
= 0,
where both Λ
1
and Λ
2
are symmetric matrices. For consistent initial values,
Q
0
is orthogonal and Q
T
0
P
0
D
−1
=
´

0
is skew-symmetric. Working with
´

0
= Q
T
0
˙
Q
0
= Q
T
0
P
0
D
−1
,
´

1/2
= Q
T
0
P
1/2
D
−1
,
´

1
= Q
T
1
P
1
D
−1
,
Figure 6.2. Numerical solution of the Kepler
problem on the sphere, obtained with the
Rattle method using step size h = 0.07.
446 E. Hairer, Ch. Lubich and G. Wanner
instead of P
0
, P
1/2
, P
1
, the equations (6.14) become the following integrator
(Q
0
,
´

0
) →(Q
1
,
´

1
) :
• find an orthogonal matrix I +h
´

1/2
such that
´

1/2
=
´

0

h
2
Q
T
0

Q
V (Q
0
)D
−1

h
2
Λ
1
D
−1
holds with a symmetric matrix Λ
1
;
• compute Q
1
= Q
0
(I +h
´

1/2
);
• compute a skew-symmetric matrix
´

1
such that
´

1
=
´

1/2

h
2
Q
T
1

Q
V (Q
1
)D
−1
−(
´

1/2
+
´

T
1/2
) −
h
2
Λ
2
D
−1
holds with a symmetric matrix Λ
2
.
This algorithm for the simulation of the heavy top is proposed in McLachlan
and Scovel (1995). An efficient implementation uses the representation of
the appearing orthogonal matrices by quaternions (Hairer 2003).
6.3. Geometric properties of Rattle
For consistent initial values (p
n
, q
n
) ∈ /, the Rattle method (6.13) yields an
approximation (p
n+1
, q
n+1
) which is again on /. We thus have a numerical
flow Φ
h
: /→/. The geometric properties of Section 2 for the St¨ ormer–
Verlet method extend to this algorithm.
Theorem 6.5. The Rattle method is symmetric, that is, Φ
h
= Φ
−1
−h
on /.
For Hamiltonians satisfying H(−p, q) = H(p, q), the method is reversible
with respect to the reflection ρ(p, q) = (−p, q), that is, it satisfies ρ ◦ Φ
h
=
Φ
−1
h
◦ ρ on /.
The proof is by straightforward verification, as for the St¨ ormer–Verlet
method.
Theorem 6.6. The Rattle method is symplectic, that is,

h
(p, q)ξ)
T
J Φ

h
(p, q)η = ξ
T
J η for ξ, η ∈ T
(p,q)
/. (6.15)
This result was first proved by Leimkuhler and Skeel (1994) for the method
(6.11)–(6.12), and by Jay (1994) and Reich (1993) for the general case (6.13).
One proof of Theorem 6.6 is by computing Φ

h
(p, q)ξ using implicit differ-
entiation, and by verifying the identity (6.15). Further proofs are based on
the interpretation as a variational integrator (Marsden and West 2001), and
on explicit formulae of a generating function as in (2.18); see Hairer (2003).
Geometric numerical integration 447
7. Geometric integration beyond St¨ ormer–Verlet
In this article we deliberately considered only the St¨ ormer–Verlet method
and a few selected geometric properties. Even within the class of ordinary
differential equations, we have not mentioned important topics of geometric
integration, such as
• higher-order methods, for instance, symmetric composition, partitioned
Runge–Kutta, and linear multistep methods,
• the structure-preserving use of variable step sizes,
• differential equations with further geometric properties such as differ-
ential equations on Lie groups, problems with multiple time scales, etc.
The reader will find more on these topics in the monographs by Sanz-
Serna and Calvo (1994) and Hairer et al. (2002), in the special journal issue
Budd and Iserles (1999), and in the survey articles by Iserles, Munthe-Kaas,
Nørsett and Zanna (2000), Marsden and West (2001) and McLachlan and
Quispel (2002).
REFERENCES
R. Abraham and J. E. Marsden (1978), Foundations of Mechanics, 2nd edn, Ben-
jamin/Cummings, Reading, MA.
H. C. Andersen (1983), Rattle: A ‘velocity’ version of the Shake algorithm for
molecular dynamics calculations, J. Comput. Phys. 52, 24–34.
V. I. Arnold (1963), Small denominators and problems of stability of motion in
classical and celestial mechanics, Russian Math. Surveys 18, 85–191.
V. I. Arnold (1989), Mathematical Methods of Classical Mechanics, 2nd edn,
Springer, New York.
G. Benettin and A. Giorgilli (1994), On the Hamiltonian interpolation of near to
the identity symplectic mappings with application to symplectic integration
algorithms, J. Statist. Phys. 74, 1117–1143.
J. J. Biesiadecki and R. D. Skeel (1993), Dangers of multiple time step methods,
J. Comput. Phys. 109, 318–328.
P. B. Bochev and C. Scovel (1994), On quadratic invariants and symplectic struc-
ture, BIT 34, 337–345.
C. J. Budd and A. Iserles, eds (1999), Geometric Integration: Numerical Solution
of Differential Equations on Manifolds, special issue of R. Soc. Lond. Philos.
Trans. Ser. A 357(1754).
R. De Vogelaere (1956), Methods of integration which preserve the contact trans-
formation property of the Hamiltonian equations, Report No. 4, Department
of Mathematics, University of Notre Dame, IN.
K. Feng (1985), On difference schemes and symplectic geometry, in Proc. Fifth In-
tern. Symposium on Differential Geometry and Differential Equations, August
1984, Beijing, pp. 42–58.
448 E. Hairer, Ch. Lubich and G. Wanner
K. Feng (1991), Formal power series and numerical algorithms for dynamical sys-
tems, in Proc. International Conference on Scientific Computation, Hangzhou,
China (T, Chan and Z.-C. Shi, eds), Vol. 1 of Series on Appl. Math., pp. 28–
35.
K. Feng and Z. Shang (1995), Volume-preserving algorithms for source-free dynam-
ical systems, Numer. Math. 71, 451–463.
E. Fermi, J. Pasta and S. Ulam (1955), Studies of nonlinear problems, Los Alamos
Report No. LA-1940. Later published in E. Fermi: Collected Papers, Vol. II,
Chicago University Press (1965), pp. 978–988.
R. Feynman (1965), The Character of Physical Law, first published by the BBC
(1965). MIT Press (1967).
H. Flaschka (1974), The Toda lattice, II: Existence of Integrals, Phys. Rev. B 9,
1924–1925.
J. Ford (1992), The Fermi–Pasta–Ulam problem: Paradox turns discovery, Physics
Reports 213, 271–310.
L. Galgani, A. Giorgilli, A. Martinoli and S. Vanzini (1992), On the problem of
energy equipartition for large systems of the Fermi–Pasta–Ulam type: Ana-
lytical and numerical estimates, Physica D 59, 334–348.
G. Gallavotti (1999), Statistical Mechanics: A Short Treatise, Springer, Berlin.
E. Hairer (2003), Global modified Hamiltonian for constrained symplectic integra-
tors, Numer. Math. To appear.
E. Hairer and Ch. Lubich (1997), The life-span of backward error analysis for
numerical integrators, Numer. Math. 76, 441–462.
E. Hairer and Ch. Lubich (1999), Invariant tori of dissipatively perturbed Hamil-
tonian systems under symplectic discretization, Appl. Numer. Math. 29, 57–
71.
E. Hairer and Ch. Lubich (2000a), Long-time energy conservation of numerical
methods for oscillatory differential equations, SIAM J. Numer. Anal. 38, 414–
441.
E. Hairer and Ch. Lubich (2000b), Energy conservation by St¨ ormer-type numerical
integrators, in Numerical Analysis 1999 (G. F. Griffiths and G. A. Watson,
eds), CRC Press LLC, pp. 169–190.
E. Hairer, S. P. Nørsett and G. Wanner (1993), Solving Ordinary Differential Equa-
tions I: Nonstiff Problems, Springer, Heidelberg.
E. Hairer, Ch. Lubich and G. Wanner (2002), Geometric Numerical Integra-
tion: Structure-Preserving Algorithms for Ordinary Differential Equations,
Springer, Berlin.
A. Iserles, H. Z. Munthe-Kaas, S. P. Nørsett and A. Zanna (2000), Lie-group meth-
ods, in Acta Numerica, Vol. 9, Cambridge University Press, pp. 215–365.
L. Jay (1994), Runge–Kutta type methods for index three differential-algebraic
equations with applications to Hamiltonian systems, Thesis No. 2658, Uni-
versity of Gen`eve.
V. V. Kozlov and A. O. Harin (1992), Kepler’s problem in constant curvature
spaces, Celestial Mech. Dynam. Astronom. 54, 393–399.
F. M. Lasagni (1988), Canonical Runge–Kutta methods, ZAMP 39, 952–953.
B. J. Leimkuhler and R. D. Skeel (1994), Symplectic numerical integrators in con-
strained Hamiltonian systems, J. Comput. Phys. 112, 117–125.
Geometric numerical integration 449
R. MacKay (1992), Some aspects of the dynamics of Hamiltonian systems, in The
Dynamics of Numerics and the Numerics of Dynamics (D. S. Broomhead and
A. Iserles, eds), Clarendon Press, Oxford, pp. 137–193.
S. Maeda (1980), Canonical structure and symmetries for discrete systems, Math.
Japonica 25, 405–420.
J. E. Marsden and M. West (2001), Discrete mechanics and variational integrators,
in Acta Numerica, Vol. 10, Cambridge University Press, pp. 357–514.
R. I. McLachlan and P. Atela (1992), The accuracy of symplectic integrators, Non-
linearity 5, 541–562.
R. I. McLachlan and G. R. W. Quispel (2002), Splitting methods, in Acta Numerica,
Vol. 11, Cambridge University Press, pp. 341–434.
R. I. McLachlan and C. Scovel (1995), Equivariant constrained symplectic integra-
tion, J. Nonlinear Sci. 5, 233–256.
P. C. Moan (2002), On backward error analysis and Nekhoroshev stability in the
numerical analysis of conservative systems of ODEs, PhD thesis, University
of Cambridge.
J. Moser (1968), Lectures on Hamiltonian systems, Mem. Amer. Math. Soc. 81,
1–60.
H. Poincar´e (1892/1893/1899), Les M´ethodes Nouvelles de la M´ecanique C´eleste,
Tome I–III, Gauthier-Villars, Paris.
S. Reich (1993), Symplectic integration of constrained Hamiltonian systems by
Runge–Kutta methods, Technical Report 93-13, Department of Computer
Science, University of British Columbia.
S. Reich (1999a), Backward error analysis for numerical integrators, SIAM J. Nu-
mer. Anal. 36, 1549–1570.
S. Reich (1999b), Preservation of adiabatic invariants under symplectic discretiza-
tion, Appl. Numer. Math. 29, 45–56.
R. D. Ruth (1983), A canonical integration technique, IEEE Trans. Nuclear Science
NS-30, 2669–2671.
J.-P. Ryckaert, G. Ciccotti and H. J. C. Berendsen (1977), Numerical integration
of the cartesian equations of motion of a system with constraints: Molecular
dynamics of n-alkanes, J. Comput. Phys. 23, 327–341.
J. M. Sanz-Serna (1988), Runge–Kutta schemes for Hamiltonian systems, BIT 28,
877–883.
J. M. Sanz-Serna (1992), Symplectic integrators for Hamiltonian problems: An
overview, in Acta Numerica, Vol. 1, Cambridge University Press, pp. 243–
286.
J. M. Sanz-Serna and M. P. Calvo (1994), Numerical Hamiltonian Problems, Chap-
man and Hall, London.
Z. Shang (1999), KAM theorem of symplectic algorithms for Hamiltonian systems,
Numer. Math. 83, 477–496.
Z. Shang (2000), Resonant and diophantine step sizes in computing invariant tori
of Hamiltonian systems, Nonlinearity 13, 299–308.
C. L. Siegel and J. K. Moser (1971), Lectures on Celestial Mechanics, Vol. 187 of
Grundlehren der Mathematischen Wissenschaft, Springer, Heidelberg.
D. Stoffer (1988), On reversible and canonical integration methods, SAM-Report
No. 88-05, ETH Z¨ urich.
450 E. Hairer, Ch. Lubich and G. Wanner
D. Stoffer (1998), On the qualitative behaviour of symplectic integrators, III: Per-
turbed integrable systems, J. Math. Anal. Appl. 217, 521–545.
G. Strang (1968), On the construction and comparison of difference schemes, SIAM
J. Numer. Anal. 5, 506–517.
Y. B. Suris (1988), On the conservation of the symplectic structure in the numerical
solution of Hamiltonian systems, in Numerical Solution of Ordinary Differen-
tial Equations (S. S. Filippov, ed.), Keldysh Institute of Applied Mathematics,
USSR Academy of Sciences, Moscow, pp. 148–160. (In Russian.)
Y.-F. Tang (1994), Formal energy of a symplectic scheme for Hamiltonian systems
and its applications, I, Comput. Math. Appl. 27, 31–39.
M. Toda (1970), Waves in nonlinear lattice, Progr. Theor. Phys. Suppl. 45, 174–200.
L. Verlet (1967), Computer ‘experiments’ on classical fluids, I: Thermodynamical
properties of Lennard-Jones molecules, Phys. Rev. 159, 98–103.
A. P. Veselov (1991), Integrable maps, Russ. Math. Surv. 46, 1–51.
H. Yoshida (1993), Recent progress in the theory and application of symplectic
integrators, Celestial Mech. Dynam. Astronom. 56, 27–43.

400

E. Hairer, Ch. Lubich and G. Wanner

CONTENTS
1 The Newton–St¨rmer–Verlet–leapfrog method o 2 Geometric properties 3 Conservation of first integrals 4 Backward error analysis 5 Long-time behaviour of numerical solutions 6 Constrained Hamiltonian systems 7 Geometric integration beyond St¨rmer–Verlet o References 400 408 416 419 425 440 447 447

1. The Newton–St¨rmer–Verlet–leapfrog method o
We start by considering systems of second-order differential equations q = f (q), ¨ (1.1)

where the right-hand side f (q) does not depend on q. Many problems in ˙ astronomy, molecular dynamics, and other areas of physics are of this form. 1.1. Two-step formulation If we choose a step size h and grid points tn = t0 + nh, the most natural discretization of (1.1) is qn+1 − 2qn + qn−1 = h2 f (qn ), (1.2)

which determines qn+1 whenever qn−1 and qn are known. Geometrically, this amounts to determining an interpolating parabola which, in the mid-point, assumes the second derivative prescribed by equation (1.1); see Figure 1.1, left.
qn+1 qn−1 qn h tn−1 tn h tn+1 tn−1 tn− 1 2 tn qn−1 qn− 1 2 qn vn tn+ 1 2 tn+1 vn− 1 2 vn+ 1 2 qn+1 qn+ 1 2

Figure 1.1. Method (1.2): two-step formulation (left); one-step formulations (right).

Geometric numerical integration

401

1.2. One-step formulations Introducing the velocity q = v turns equation (1.1) into a first-order system ˙ of doubled dimension q = v, ˙ v = f (q), ˙ (1.3)

an equation in the so-called phase space. In analogy to this, we introduce discrete approximations of v and q as follows: vn = qn+1 − qn−1 , 2h vn− 1 =
2

qn − qn−1 , h

qn− 1 =
2

qn + qn−1 , 2

(1.4)

where some derivatives, in order to preserve second-order and symmetry, are evaluated on the staggered grid tn− 1 , tn+ 1 , . . .; see Figure 1.1, right. 2 2 Inserting these expressions into the method (or simply looking at the picture) we see that method (1.2) can now be interpreted as a one-step method Φh : (qn , vn ) → (qn+1 , vn+1 ), given by vn+ 1 = vn +
2

h 2

f (qn ),
2

vn+ 1
2

qn+1 vn+1 qn+ 1
2

(A)

qn+1 = qn + h vn+ 1 , vn+1 = vn+ 1 +
2

(1.5)

h 2

f (qn+1 ).

qn

vn

There is a dual variant of the method on the staggered grid (vn− 1 , qn− 1 ) → 2 2 (vn+ 1 , qn+ 1 ) as follows:
2 2

qn = qn− 1 +
2

h 2

vn− 1 ,
2

vn− 1
2

vn+ 1 qn+ 1
2 2

(B)

vn+ 1 = vn− 1 + h f (qn ),
2 2 2

qn+ 1 = qn +

h 2

vn+ 1 .
2

qn− 1
2

(1.6)

qn

vn

For both arrays (A) and (B), we can concatenate, in the actual step-bystep procedure, the last line of the previous step with the first line of the subsequent step. Both schemes then turn into the same method, where the q-values are evaluated on the original grid, and the v-values are evaluated on the staggered grid: vn+ 1 = vn− 1 + h f (qn ),
2 2

(1.7)

qn+1 = qn + h vn+ 1 .
2

This is the computationally most economic implementation, and numerically more stable than (1.2); see Hairer, Nørsett and Wanner (1993, p. 472).

402

E. Hairer, Ch. Lubich and G. Wanner

1.3. Historical remarks
Isn’t that ingenious? I borrowed it straight from Newton. It comes right out of the Principia, diagram and all. R. Feynman (1965, p. 43)

The above schemes are known in the literature under various names. In particular, in molecular dynamics they are often called the Verlet method (Verlet 1967) and have become by far the most widely used integration scheme in this field. Another name for this method is the St¨rmer method, since C. St¨rmer, o o in 1907, used higher-order variants of it for his computations of the motion of ionized particles in the earth’s magnetic field (aurora borealis); see, e.g., Hairer et al. (1993, Section III.10). Sometimes it is also called the Encke method, because J. F. Encke, around 1860, did extensive calculations for the perturbation terms of planetary orbits, which obey systems of second-order differential equations of precisely the form (1.1). Mainly in the context of partial differential equations of wave propagation, this method is called the leapfrog method. In yet another context, this formula is the basic method for the GBS extrapolation scheme, as it was proposed, for the case of equation (1.1), by Gragg in 1965; see Hairer et al. (1993, p. 294f.). Furthermore, the scheme (1.7) is equivalent to Nystr¨m’s method of order 2; see Hairer et al. o (1993, p. 362, formula (III.1.13 )). A curious fact is that Professor Loup Verlet, who later became interested in the history of science, discovered precisely ‘his’ method in several places in the classical literature, for example, in the calculations of logarithms and astronomical tables by J. B. Delambre in 1792: this paper was translated and discussed in McLachlan and Quispel (2002, Appendix C). Even more spectacular is the finding that the ‘Verlet method’ was used in Newton’s Principia from 1687 to prove Kepler’s second law. An especially clear account can be found in Feynman’s Messenger Lecture from 1964; see Feynman (1965, p. 41), from which we reproduce with pleasure1 two of Feynman’s original hand drawings. The argument is as follows: if there are no forces, the body advances with uniform speed, and the radius vector covers equal areas in equal times, simply because the two triangles Sun-1-2 and Sun-2-3 have the same base and common altitudes (see Figure 1.2). If the gravitational force acts at the midpoint, the planet is deviated in such a way that the top of the second triangle moves parallel to the sun ray (see Figure 1.3). Hence, the triangle Sun-2-4 also has the same area. The whole procedure (uniform motion on half the interval, then a ‘kick’ to the velocity in the direction of the Sun, and another uniform motion on the second half) is precisely variant (B) of the St¨rmer–Verlet scheme. o
1

. . . and with permission of the publisher

Geometric numerical integration

403

Figure 1.2. Uniform motion of a planet (drawing by R. Feynman).

Figure 1.3. Gravitation acting at mid-point (drawing by R. Feynman).

1.4. Interpretation as composition method (symplectic Euler ) We can go a step further and split the formulae in the middle of the schemes (1.5) and (1.6). We then arrive at the schemes (vn , qn ) → (vn+ 1 , qn+ 1 ) 2 2 given by (SE1) vn+ 1 = vn +
2 2

qn+ 1 = qn +

h 2 f (qn ), h 1 2 vn+ 2 ,

vn+ 1 qn+ 1
2 2

(1.8)

qn

vn

as well as the adjoint scheme (vn+ 1 , qn+ 1 ) → (vn+1 , qn+1 ) obtained by for2 2 mally replacing the subscript n by n + 1 and h by −h, qn+1 vn+1 vn− 1 2 qn+1 = qn+ 1 + h vn+ 1 , vn+ 1 2 2 2 2 (1.9) (SE2) h qn+ 1 vn+1 = vn+ 1 + 2 f (qn+1 ). 2 2 qn− 1 2 qn vn

404 E. [2] ϕh/2 [2] ϕh/2 (SE1) (1. [2] ϕh/2 ϕh (B) (1.5. f (q)) of (1. which both have a constant time derivative. Lubich and G. are called the symplectic Euler method. Wanner Both these schemes. . f (q)). while the scheme (B) is the composition of method (SE2) followed by (SE1). The exact flows ϕt and [2] ϕt of these two vector fields. y0 [1] y0 ϕh/2 For the two versions of the St¨rmer–Verlet method we thus obtain the diao grams [1] ϕh/2 y1 y1 (A) [2] (A) = (SE2) ◦ (SE1). 1. (1. are easily obtained: ϕt [1] : q 1 = q 0 + t · v0 v1 = v0 and ϕt [2] : q1 = q0 v1 = v0 + t · f (q0 ). The phase space vector field split into two fields. in which one variable is used at the old value and the other at the new value.11) [1] [2] (SE1) = ϕh/2 ◦ ϕh/2 . Interpretation as splitting method We consider the vector field (v.3) ‘split’ as the sum of two vector [1] fields (v. We thus see that the above scheme (A) is the composition of the symplectic Euler schemes (SE1) with (SE2). Hairer.4. as indicated in Figure 1. 0) and (0.12) [2] [1] ϕh/2 (B) = (SE1) ◦ (SE2).10) These formulae are precisely those which build up the formulae (SE1) and (SE2) above: [1] ϕh/2 y1 [2] [1] y1 (SE2) (SE2) = ϕh/2 ◦ ϕh/2 .4. Ch. ϕh y0 [1] y0 ϕh/2 v v v q=v ˙ v = f (q) ˙ = q q=v ˙ v=0 ˙ + q q=0 ˙ v = f (q) ˙ q Figure 1.

For a careful survey of splitting methods we refer to McLachlan and Quispel (2002). . . reduce to the second-order differential equation M q = −∇U (q). . Φh Φh (A) = ϕh/2 ◦ ϕh ◦ ϕh/2 . This variational principle states that the motion of a mechanical system between any two positions q(t0 ) = q0 and q(tN ) = qN is such that the action integral tN L q(t). ∇Q refer to Sh = Sh (q. We then require that q1 . that is. (1. N −1 Sh (qn . = ϕh/2 ◦ ϕh ◦ ϕh/2 . 1. .14) where L(q.15) these equations reduce to M (qn+1 − 2qn + qn−1 ) + h2 ∇U (qn ) = 0. When M does not depend on q. qn+1 ) = h qn+1 − qn L qn . (1. instead of (1. . 2 h (1. it is the difference between the kinetic and the potential energy. Typically. L(q. . interpolating grid values (tn . qn+1 ) is minimized. We now approximate q(t) by a piecewise linear function.15) with a symmetric positive definite mass matrix M .16) where Sh (qn . . v) is the Lagrangian of the system. n=0 (1.Geometric numerical integration 405 or. . 2 h + h qn+1 − qn L qn+1 . [1] [2] [1] [2] [1] [2] (B) (1. 1. qn ) for n = 0. where the partial gradients ∇q . . In the case of the Lagrangian (1. N − 1. qn+1 ) = 0 for n = 1. d ∂L ∂L ¨ dt ∂v = ∂q . 1 2 (1. q(t) dt ˙ t0 is minimized. . N . qN −1 be such that. Interpretation as variational integrator A further approach to the St¨rmer–Verlet method is obtained by discretizing o Hamilton’s principle. Q). v) = v T M v − U (q). .13) This way of composing the flows of split vector fields is often referred to as Strang splitting.17) The requirement that the gradient with respect to qn be zero. qn ) + ∇q Sh (qn . yields the discrete Euler–Lagrange equations ∇Q Sh (qn−1 . and the action integral by the trapezoidal rule. . after Strang (1968).14). in explicit formulae. .6. the Euler–Lagrange equations of this variational problem.18) .

2 + q2 )3/2 q2 = − ¨ 2 (q1 (1.9).406 E. 1−e (1.5 presents the numerical values of the St¨rmer–Verlet method for two different step o sizes. and equation (1. o with f (q) = −M −1 ∇U (q). This gives (SE1) and (SE2) qn+1 = qn+ 1 + 2 vn+ 1 = vn + 2 2 qn+ 1 = qn + h 2 h 2 h 2 h 2 f qn . ˙ (1. ˙ v = f (q. Ch. by taking over the missing arguments from one equation to the other. submitted for publication – and never published. A comprehensive survey of variational integrators can be found in Marsden and West (2001). Extension to general partitioned problems For the extension of the above formulae to the more general system q = g(q. We first extend the formulae (1.23) . short. q2 (0) = 0.19) with e = 0.8) and (1. This variational interpretation of the St¨rmer–Verlet method was given o by MacKay (1992). These solutions are compared to those of the explicit midpoint rule in Runge’s one-step formulation.2. vn+ 1 .6.8.22) g qn+1 . 1. This second method is of the same order and for the first steps it behaves very similarly to the St¨rmer–Verlet scheme (the o first step is even identical!).20) q2 . Figure 1. clear.21) we follow the ideas of De Vogelaere (1956). Wanner which is just the two-step formulation (1. v). The period of the exact solution is 2π. 2 g qn .3)). but it deteriorates significantly as the integration interval increases. vn+ 1 . 2 (1. see Hairer. Hairer. vn+ 1 . Figure 1. The explanation of this strange difference is the subject of the theories below. q1 = − 2 ¨ 2 (q1 + q2 )3/2 As initial values we take q1 (0) = 1 − e. Lubich and Wanner (2002. ˙ q2 (0) = ˙ 1+e . p. 2 (1. Numerical example We choose the Kepler problem q1 . This is a marvellous paper. written in one week. Lubich and G. 2 vn+1 = vn+ 1 + 2 f qn+1 .2) of the St¨rmer–Verlet method. 1. v). vn+ 1 . 24. elegant. q1 (0) = 0.7.

2 (A) = (SE2)◦(SE1) qn+1 = qn + 2 g qn .. 2 h 2 h 2 g qn .10 1 58 steps h = 0.Geometric numerical integration 407 h = 0. (1..4: vn+ 1 = vn + 2 h 2 h 2 f qn . Kepler problem: the dashed line is the exact solution. vn− 1 + f qn . vn+ 1 . 2 (B) = (SE1)◦(SE2) vn+ 1 = vn− 1 + 2 2 2 f qn .10 300 steps 1 −1 −1 h = 0..5. The extensions of the St¨rmer–Verlet schemes are o now obtained by composition. vn+ 1 2 2 2 . qn = qn− 1 + 2 h 2 f qn+1 . In each of these algorithms the derivative evaluations of both formulae are taken at the same point..25) qn+ 1 = qn + h 2 g qn .24) vn+1 = vn+ 1 + and. Runge’s method Figure 1. vn+ 1 . St¨rmer–Verlet scheme o . .05 1200 steps 1 −1 −1 . vn+ 1 + g qn+1 . for the dual version.05 116 steps 1 h = 0. in the same way as in Section 1. (1. vn− 1 . vn+ 1 2 2 . vn+ 1 .

while only the last one is explicit.6) unchanged. the second one for qn+1 .8. vn+ 1 ) 2 exact (qn . Hairer. vn+1 ) (qn+ 1 . reflecting time at the centre tn+1/2 ) gives the same method again. vn+ 1 ) 2 2 exact (qn . replacing h by −h and exchanging the subscripts n ↔ n + 1 (i. Black points indicate where the vector field is evaluated.e. vn− 1 ) 2 (qn− 1 . De Vogelaere (1956) 2. v = − sin q − v 2 /5. Geometric properties We study geometric properties of the flow of differential equations which are preserved by the St¨rmer–Verlet method. 2 Such implicit methods were not common in the 1950s and might then not have delighted journal editors – or programmers: No detailed example or discussion is given. vn+ 1 ) 2 (qn+1 . Lubich and G. the replacements h ↔ −h and n − 1 ↔ n + 1 leave the formulation (1. Wanner (qn .408 E. where the above time-reflection transforms (SE1) to (SE2) and vice versa. this symmetry can be stated more formally as (2.6. o ˙ ˙ initial values (−1. MURA or CERN group.1. Harwell.. symplecticity.3). vn ) → (qn+1 . vn− 1 ) 2 2 St¨rmer–Verlet (B) o Figure 1. . For illustrations see Figure 1. vn+ 1 ) 2 (qn+1 . Similarly.1) Φh = Φ−1 .6. Symmetry and reversibility The St¨rmer–Verlet method is symmetric with respect to changing the dio rection of time: in its one-step formulation (1.5. Ch.8) and (1.9). In terms of the 2 2 numerical one-step map Φh : (qn . 0.5). step size h = 1. The first equation of (1. vn ) St¨rmer–Verlet (A) o (qn . This will best be done by those working on these problems in the Brookhaven. The properties discussed are o reversibility.24) is now an implicit formula for vn+ 1 . St¨rmer–Verlet methods for q = v. −h Such a relation does not hold for the symplectic Euler methods (1. and volume preservation. vn+1 ). 2.

−v) exact (q. (2.1) and (2. −v) = (q. −v). v) = (q. right. v) implies ϕt (q. v) (q. v) implies Φ−h (q. v) ρ 1 2 −2 (q. The importance of this property in numerical analysis was first emphasized by Stoffer (1988).4) as is readily seen from the defining formulae such as (1. (2. v) implies Φh (q. −v) = (q.2) and we call it reversible with respect to the reflection ρ : (q. the same equation as in Figure 1.Geometric numerical integration 409 ϕt (q. (2.1. The system (1. A reversible system (left) and the symmetric St¨rmer– o Verlet method (right). −v) ϕt (q. The numerical one-step map o Φh of the St¨rmer–Verlet method satisfies similarly Φh (q. The time-symmetry of the St¨rmer–Verlet method implies an important o geometric property of the numerical map in the phase space.1. namely reversibility. v) = (q.6.1) is both symmetric and reversible.1. Theorem 2.3) has the property that inverting the direction of the initial velocity does not change the solution trajectory. i. −v). This holds because practically all numerical methods for (1.. v) ρ 0 ρ (q. left. .3) for all q. are such that Φh (q.3). see Figure 2. The flow ϕt thus satisfies that ϕt (q. Let us summarize these considerations. v) = (q. −v) 2 (q. −v). The symmetry (2. and in particular the St¨rmer–Verlet method o and the symplectic Euler methods. −v). The St¨rmer–Verlet method applied to the second-order o differential equation (1. This property is illustrated in Figure 2. v) ρ −2 −1 0 (q. it just inverts the direction of motion. −v) = (q.3).3).1) of the St¨rmer–Verlet method is therefore equivalent to the reversibility o (2.1. its one-step map satisfies (2. v and all h. v) → (q. to which we turn next.e.5). −v) (A) Figure 2.

v2 ). A characteristic geometric property of Hamiltonian systems is that the flow ϕt is symplectic. q) is defined. q) = J with J = 0 I . q) = pT M −1 p + U (q). p.9) where I is the identity matrix of the dimension of p or q. v2 ) → (q1 . Lubich and G. the flow is ρ-reversible with respect to involutions ρ other than (q. q) and t where ϕt (p.7) turns into the second-order differential equation (1. it is not related to the conservation of lengths but of areas in phase space. e. ˙ (2. that is. q) is an arbitrary scalar function of the variables (p. symplecticity (2. −q2 . t (2. ˙ q = ∇p H(p. qi )-coordinate planes. q)T J ϕt (p. left. see Figure 2.. ϕt (p. h 2.2. We then call the differential equation ρ-reversible. 204) or Hairer et al.19) is ρ-reversible also with respect to ρ : (q1 . that is. but.5) For example. that is. By the same argument as above. Ch. see. q) of the flow satisfies.3) upon expressing the momenta p = M v in terms of the velocities and setting f (q) = −M −1 ∇U (q). q)-plane. it satisfies ρ ◦ ϕt = ϕ−1 ◦ ρ. Wanner In some situations. The relation (2. equation (2. unlike orthogonality. the St¨rmer–Verlet method is then also o ρ-reversible for ρ of the form ρ(q. v1 . the flow of a differential equation y = F (y) is ρ-reversible if and only if the vector field ˙ satisfies ρ ◦ F = −F ◦ ρ. (2.9) is formally similar to orthogonality (which it would be if J were replaced by the identity matrix). ρ ◦ Φh = Φ−1 ◦ ρ. q).8) 2 with a positive definite mass matrix M and a potential U (q). for all (p. q ∈ R).g. −v1 . In fact. ρ2 (v)). For higher-dimensional systems. q). for any two-dimensional bounded . Equation (2. When the Hamiltonian is of the form 1 H(p. for systems with one degree of freedom (i. then the system (2. v) → (q. p.9) means that the flow preserves the sum of the oriented areas of the projections of ϕt (A) onto the (pi . (2002. Hamiltonian systems and symplecticity We now turn to the important class of Hamiltonian systems p = −∇q H(p.9) expresses that the flow preserves the area of sets of initial values in the (p. Hairer. q2 . −v). the flow of the Kepler problem (1.6) where H(p. q). In general.410 E. −I 0 (2. v) = (ρ1 (q). 172).2.7) (2. the derivative ϕt = ∂ϕt /∂(p. Arnold (1989.8) expresses the total energy H as the sum of kinetic and potential energy.e.. p.

for Hamiltonian systems (2. Symplecticity of the St¨rmer–Verlet o method for a separable Hamiltonian. Sanz-Serna (1988).7) reads pn+ 1 = pn − 2 h 2 h 2 ∇q H pn+ 1 . that is.) for a justification of this interpretation. In the late 1980s. 171f. the method reduces to the St¨rmer–Verlet method (1.8). (2. 2 and a similar formula for variant (B).11) for all (p.9). A numerical method is called symplectic if. The St¨rmer–Verlet method applied to a Hamiltonian syso tem is symplectic.. Symplecticity of numerical methods was first considered by De Vogelaere (1956). qn + ∇p H pn+ 1 .24) applied to the Hamiltonian system o (2.Geometric numerical integration 411 manifold of initial values A. (2002. q) and all step sizes h. p. the Jacobian of the numerical flow Φh : (pn . Hairer et al.2. 2 (A) qn+1 = qn + 2 ∇p H pn+ 1 .5) with o f (q) = −M −1 ∇U (q). but was not followed up until Ruth (1983) and Feng (1985). In the particular case of the Hamiltonian (2. see. upon setting pn = M vn . q)T J Φh (p. e. if Φh (p. qn ) → (pn+1 . and Suris (1988) started off an avalanche of papers on symplectic numerical methods. q) = J (2.2.7). qn+1 .g. the results of Lasagni (1988). qn . .10) pn+1 = pn+ 1 − h 2 ∇q H pn+ 1 . Theorem 2. qn+1 ) satisfies condition (2. qn+1 2 2 . Sanz-Serna and Calvo (1994) was the first book dealing with this subject. The St¨rmer–Verlet method (1. ϕt Figure 2.

2 (2.e. In the same way. Hqp .12): T I + hHqp 0 −hHpp I ∂(pn+1/2 . The matrix of partial derivatives is obtained from differentiating equation (2. The historically first proof. of higher order and/or for constrained Hamiltonian systems. 2 ∇q H pn+ 1 . are evaluated at (pn+1/2 . Hpp . qn . Wanner We give four different proofs of this result. qn ). h h/2 h/2 (2. Ch. which all correspond to different interpretations of the method: as a composition method.14) where ϕT and ϕU are the exact flows of the Hamiltonian systems with t t ˙ ˙ Hamiltonian T (p) = 1 pT M −1 p and U (q). q = 0. due to De Vogelaere (1956).. 2 (2.13) The method (SE1) is indeed symplectic. and using generating functions. corresponding to the splitting H(p. q = M −1 p and 2 p = −∇U (q). qn+1 . uses the interpretation of the St¨rmer–Verlet method as the composition of the symplectic o Euler method (SE1) and its adjoint (SE2) qn+1 = qn+ 1 + 2 pn+ 1 = pn − 2 2 qn+ 1 = qn + h 2 h 2 ∇q H pn+ 1 . etc. qn+1/2 ) ∂(pn . (SE2) is seen to be symplectic. but it applies only to the case of separable Hamiltonians H(p. qn ) = I −hHqq .8). qn+1 . As in (1. 0 I + hHqp where all the submatrices of the Hessian. Hairer. respectively.412 E. Yet another proof is based on the preservation of quadratic invariants and will be mentioned in Section 3 below. as a variational integrator. q) = ˙ ˙ . p = 0. Each of these interpretations lends itself to generalizations to other symplectic integrators. The second proof applies only to Hamiltonians of the special form (2. o we have for variant (A) Φh = ϕU ◦ ϕT ◦ ϕU . 2 ∇p H pn+ 1 . qn+1/2 ) ∂(pn . as is seen by direct verification of the symplecticity condition ∂(pn+1/2 ..12) pn+1 = pn+ 1 − 2 h 2 h 2 ∇p H pn+ 1 . as a splitting method. qn+1/2 ) ∂(pn . The second proof is the most elegant one. the third proof is formulated for such Hamiltonians for convenience. q) = T (p) + U (q). i. It is based on the interpretation of the St¨rmer–Verlet method as a splitting method.13). qn ) T J ∂(pn+1/2 . Hence their composition (2.10) is also symplectic. Lubich and G. qn . qn ) = J.

15) SqQ Qp = I.16).2. Qp . in the case of the Lagrangian (1. 0 I −I 0 Pp Pq Qp Qq = 0 I −I 0 after multiplying out. qn+1 ) of the St¨rmer–Verlet method is thus generated by the scalar-valued function Sh o via (2. The generating (2. q1 = q0 + ∇p Sh (p1 . the first of the above two equations determines qn+1 . is symplectic for any function S. q0 ) o in the same way as the symplectic Euler method: p1 = p0 − ∇q Sh (p1 . Q) = p. q) → (P. such maps are symplectic. These equations yield Pp Pq Qp Qq T qn+1 − qn h + ∇U (qn ) = M vn = pn h 2 (2.15) and (2. Variant (B) has the flows of T and U interchanged in (2. As we have seen in the first proof. This completes the third proof of symplecticity of the St¨rmer–Verlet method. and is thus likewise symplectic. as is required for symplecticity.8) into kinetic and potential energy. Differentiation of the above equations gives the following relations for the matrices of partial derivatives Pp . The third proof uses the interpretation of the St¨rmer–Verlet method o as a variational integrator (see Section 1.4) and the first line of (1. qn ) → (pn+1 . right.17).14). q0 ).15) which corresponds to the Hamiltonian (2. Q) of (1. Q) generated by ∇Q Sh (qn . Qq : Sqq + SqQ Qq = 0.8). SQQ Qp = Pp . ∇Q S(q.16) h 2 Given (pn . Pq . −∇q Sh (qn .17) . qn+1 ) = M and similarly qn+1 − qn h − ∇U (qn+1 ) = M vn+1 = pn+1 . The one-step map Φh : (pn . The symplecticity of variational integrators derives from non-numerical work by Maeda (1980) and Veselov (1991). A step of the St¨rmer–Verlet method can be generated by a function Sh (p1 . qn ). Using (1. (2. The desired result then follows from the fact that a map (p. o A fourth proof of the symplecticity is based on ideas of Lasagni (1988).14). qn+1 ) = M −∇q S(q.6). Since the flows of Hamiltonian systems are symplectic. Q) = P. SQq + SQQ Qq = Pq . and the second one pn+1 . so is their composition (2.Geometric numerical integration 413 T (p)+U (q) of the Hamiltonian (2. q0 ).5). This is illustrated in Figure 2. we have for Sh (q. This is verified by directly checking the symplecticity condition.

20) for which the diagonal blocks of the Jacobian are zero. v = f (q). which implies det Φh (p.3). z).7). ∇p H(p. written as a first-order system.19) [1] the method (1. y). this follows from its symplecticity (2.18) 2 h2 − ∇q H(p1/2 .24) can be interpreted as the splitting (1. f (q)) of a second-order differential equation (1. q) = 1 for all (p. ˙ y = b(x. v = 0 and q = 0. q1 )T ∇p H(p1/2 .10). q1 ) . q). they are volumepreserving and so is their composition. and for every t for which ϕt (y) exists for all y ∈ Ω. The same idea allows us to extend the St¨rmer–Verlet method to a volumeo preserving algorithm for systems partitioned into the three equations x = a(y. ˙ z = c(x.3. The same is true for Hamiltonian vector fields (−∇q H(p. q).11). We do not give the computational details. The St¨rmer–Verlet method preserves volume. Ch. o vol (Φh (Ω)) = vol (Ω). For the method (2. vol (ϕt (Ω)) = vol (Ω). is divergence-free. Wanner function is simply Sh = hH for the symplectic Euler method. 4 o where q1 and p1/2 are defined by the St¨rmer–Verlet formulae and are now considered as functions of (p1 . respectively. (2002.5) for a more general class of symplectic integrators. ˙ v = f (q). Hairer. which can be found in Hairer et al. q)). applied to a Hamiltonian system (2. [1] [2] [3] [2] [1] (2. For partitioned differential equations of the form q = g(v). ˙ (2. Section VI. q0 ) = 2. q0 ).13). ˙ ˙ ˙ ˙ Since the vector fields of these flows are divergence-free. where ϕt and [2] ϕt are the exact flows of q = g(v). in the following two situations. q1 ) (2.414 E. The vector field (v. It therefore preserves volume in phase space: for every bounded open set Ω. Sh (p1 . giving ϕh/2 ◦ ϕh/2 ◦ ϕh ◦ ϕh/2 ◦ ϕh/2 . We split them symmetrically. q0 ) + H(p1/2 . Volume preservation ˙ The flow ϕt of a system of differential equations y = F (y) with divergencefree vector field (div F (y) = 0 for all y) satisfies det ϕt (y) = 1 for all y. Lubich and G.21) . ˙ (2. For the St¨rmer–Verlet method Sh is obtained as o h H(p1/2 . z). q0 ) + ∇p H(p1/2 .

applied to the ABC-flow yn+1 = yn+ 1 + x = A sin z + C cos y. 2 2 = zn + h c xn+ 1 . centred at the origin.9. y = 0. Volume-preserving deformation of the ball of radius 0. z).22) h b xn+ 1 . Feng and Shang (1995) give a volume-preserving extension of the above scheme to the general case. B = C = 1. yn+ 1 .6 t = 1. this becomes xn+ 1 = xn + 2 [1] yn+ 1 2 zn+1 h a(yn .3 for A = 1/2. 2 2 2 (2. 2 h = yn + b xn+ 1 . 2 2 h xn+1 = xn+ 1 + a(yn+1 .3. zn+1 . zn . ˙ is presented in Figure 2. 2 2 An illustration of this algorithm. where ϕt is the (volume-preserving) flow of x = a(y.Geometric numerical integration 415 t = 1. . More ingenuity is necessary if the system is divergence-free with nonzero elements on the diagonal of the Jacobian. by the ABC flow (left) and by method (2.6 y z y z x x t=0 t=0 Figure 2. z = 0 and ˙ ˙ ˙ [2] [3] similarly for ϕt and ϕt . ˙ y = B sin x + A cos z. Written out. zn+1 ). ˙ z = C sin y + B cos x.22) (right). zn ).

however. . Hairer.004 −. however.3). q) of a Hamiltonian system (2. and the nonzero components of the Runge– Lenz–Pauli vector         p1 0 q1 A1 1 − q2 .7): since H = (∇p H T .1.006 H 10 20 A2 Figure 3.002 . In contrast.19) is Hamiltonian with H(p. In addition to the Hamiltonian. Apart from very exceptional cases. and there is a drift away from the constant value for all of them. this system has the following conserved quantities. and there are only small errors in the Hamiltonian along the numerical solution. A2  = p2  ×  0 2 2 q1 + q2 0 0 0 q1 p2 − q2 p1 Figure 3. or invariant) of the differential equation y = F (y) if ˙ I(y(t)) is constant along every solution.1. for explicit Runge– Kutta methods. The Hamiltonian and the second component of the Runge–Lenz–Pauli vector along the numerical solution of the St¨rmer–Verlet method with step size h = 0. if I (y)F (y) = 0 for all y.1 shows the behaviour of these quantities along a numerical solution of the St¨rmer–Verlet method. o . Wanner 3.002 −. or constant of motion. that H is conserved up to O(h2 ) over extremely long time intervals. Example 3. none of the first integrals is preserved. The foremost example is the Hamiltonian H(p.1) The latter condition says that the gradient ∇I(y) is orthogonal to the vector field F (y) in every point of the phase space. Ch.000 −. H is not constant along numerical solutions computed with the St¨rmer–Verlet o method.02. The method preserves the angular o momentum exactly (see Section 1. a linear drift in the Runge–Lenz–Pauli vector. but no drift. (3.416 E. the total energy H is a first integral. The Kepler problem (1. or equivalently. ∇q H T ) and ∇p H T (−∇q H)+∇q H T ∇p H = 0. q) = 1 2 2 2 2 2 (p1 + p2 ) − 1/ q1 + q2 . There is. Conservation of first integrals A non-constant function I(y) is a first integral (or conserved quantity. as can be easily checked: the angular momentum L = q1 p2 − q2 p1 . Lubich and G. Later we will see.

Applying the St¨rmer–Verlet 2 2 method gives pn+1 ωqn+1 = A(hω) pn ωqn (3. q) = 1 2 N i=1 1 T p pi + mi i N i−1 Vij i=2 j=1 qi − qj . and Vij (r) (i > j) is the interaction potential between the ith and jth particle. Let the linear first integral be I(q. (3. such as the total linear momentum.2) Here qi . for i > j. Necessarily then.4. mi N pi = ˙ j=1 νij (qi − qj ) where. we have νij = νji = −Vij (rij )/rij with rij = qi − qj .3) . is formulated as a Hamiltonian system with total energy H(p. say νii = 0. i=1 j=1 The exact preservation of linear first integrals. o Proof. (Conservation of total linear and angular momentum of N -body systems) A system of N particles interacting pairwise.3. as the following example shows. v) = bT q + cT v. and b = 0. Theorem 3. Quadratic first integrals are not generally preserved by the St¨rmer–Verlet o method.5) by cT thus yields cT v1 = cT v0 . 1 pi × pi + mi N N q i × pi = i=1 i=1 qi × νij (qi − qj ) = 0.2. which has the Hamilo tonian H(p. Example 3. pi ∈ R3 represent the position and momentum of the ith particle of mass mi . so that bT v + cT f (q) = 0 for all q. The equations of motion read qi = ˙ 1 pi . The St¨rmer–Verlet method preserves linear first integrals.Geometric numerical integration 417 Example 3. with potential forces depending on the distances of the particles. and νii is arbitrary. Consider the harmonic oscillator. cT f (q) = 0 for all q. Multiplying the formulae for v in (1. q) = 1 p2 + 1 ω 2 q 2 (p. q ∈ R). The conservation of the total linear momentum P = N pi and the total angular momentum L = N qi × pi is a i=1 i=1 consequence of the symmetry relation νij = νji : d dt d dt N N N N pi = i=1 i=1 j=1 N νij (qi − qj ) = 0. v. is common to most numerical integrators.

with M B skew-symmetric. that is. and in particular the total angular momentum of N -body systems. The matrix V of eigenvectors is close to the identity for small hω. q) = pT (Bq + b) in the Hamiltonian case). q) = pT a(q) is a first integral. H(p. 2 where we notice that the term in the second line vanishes. e.8) and (1. Ch.5. Notice. and the norm of V −1 (pn . 88). As we have seen in Section 1.8). however. v. o Theorem 3. see. o Proof. Writing the St¨rmer– Verlet method as the composition of the two symplectic Euler methods (1.15). Hairer. and the result follows. By (3. p. then I(p.418 E. For the second T T half-step we obtain in the same way vn+1 (Cqn+1 + c) = vn+1/2 (Cqn+1/2 + c). that the characteristic polynomial is λ2 − (2 − h2 ω 2 )λ + 1.9). For Hamiltonian systems of the form (2. where the associated Lagrangian is (1. we obtain for the first half-step T T vn+1/2 (Cqn+1/2 + c) = vn (Cqn + c) h T f (qn )T (Cqn + c) + vn+1/2 Cvn+1/2 .7) with a Hamiltonian of the form (2. The St¨rmer–Verlet method does. preserve an important subo class of quadratic first integrals. αs (q)v) = L(q.1). Newton was aware that the method preserves angular momentum in the Kepler problem and used this fact to prove Kepler’s second law. following an argument by Bochev and o . In the following result C is a constant square matrix and c a constant vector. The St¨rmer–Verlet method preserves quadratic first ino tegrals of the form I(q. Theorem 3.4) Since A(hω) is not an orthogonal matrix.3. L(αs (q). ωqn )T is conserved. + The most important source of first integrals of Hamiltonian systems is Noether’s theorem. q) is not preserved along numerical solutions. v) for all real s near 0 and all (q. Wanner with the propagation matrix A(hω) = 1− h2 ω 2 2 hω 2 − hω 1 − 2 1− h2 ω 2 4 h2 ω 2 2 . (3.g. f (q)T (Cq+c)+v T Cv = 0 for all q. all first integrals originating from Noether’s theorem are preserved by the St¨rmer–Verlet method. so that the eigenvalues are of modulus one if (and only if) |hω| ≤ 2. it can be shown that a(q) must be linear: a(q) = Bq + b.8). Arnold (1989.. Lubich and G. for Hamiltonian systems (2. v) = v T (Cq + c) (or I(p. v). Hence. which states that continuous symmetries yield first integrals: if the associated Lagrangian is invariant under the flow αs of the vector field a(q).5 yields yet another proof (and further insight) of the symplecticity of the St¨rmer–Verlet method. however.

The idea consists in searching and studying a modified differential equation y = F (y) + hF2 (y) + h2 F3 (y) + · · · . ˙ (4. q = M −1 p ˙ ˙ together with its variational equation ˙ Y = 0 −∇2 U (q) Y −1 0 M with Y = Pp Pq .1) and a numerical one-step method yn+1 = Φh (yn ). the series in (4. We therefore consider the differential equation y = F (y).2) such that the exact time-h flow ϕh (y) of (4. beginning with the papers by Feng (1991). Exponentially small error bounds and applications of backward error analysis to explain the long-time behaviour of numerical integrators were subsequently given by Benettin and Giorgilli (1994). Theorem 3. o 4. A rigorous formulation evolved around 1990.5 shows that they are preserved by the St¨rmer–Verlet method: o T JY = Y T JY .1. Ruth (1983). Y1 1 0 0 h h 4. Qp Qq The derivative of the flow is then ϕt (p. e. which is just the symplecticity Φ T JΦ = J. q) of the numerical solution with respect to the initial values equals the result Y1 obtained by applying the method to the combined system of the Hamiltonian system together with its variational equation. Unfortunately. Such an interpretation has been intuitively used in the physics literature. Construction of the modified equation The idea of backward error analysis applies to general ordinary differential equations and to general numerical integrators.2) cannot be expected to converge in general. Pq ) and (q. ˙ (4. Since they are of the mixed quadratic type considered above. Backward error analysis The theoretical foundation of geometric integrators is mainly based on a backward interpretation which considers the numerical approximation as the exact solution of a modified problem. Pp . q and Y (0) = I.g. and a restriction to special methods for second-order problems would hide the essentials. Qp . and the precise statement has to be formulated as follows. We explain the essential ideas and we illustrate them for the St¨rmer–Verlet method.Geometric numerical integration 419 Scovel (1994): consider the Hamiltonian system p = −∇U (q). and Reich (1999a). Sanz-Serna (1992) and Yoshida (1993). The derivative Φh (p. Qq ).2) is equal to the numerical flow Φh (y). Hairer and Lubich (1997). partitioned into (p. Symplecticity means that the components of Y T JY are first integrals. McLachlan and Atela (1992). q) = Y (t) corresponding to the initial conditions p.. .

6) F F (y) − 1 2! F F2 (y) + F2 F (y) . F )(y) + F 3! (4. This yields recurrence relations for the functions Fj (y).1).3). v) o .7) Expanding this function into a Taylor series we get (4. For the St¨rmer–Verlet scheme (1. Then there exist unique vector fields Fj (y) such that. v) .1. we expand the exact flow of (4. namely.5) we have o Φh (q.4) Proof. for any N ≥ 1.1. Modified equation of the St¨rmer–Verlet method o Putting y = (q.5) F (y) + hF2 (y) + · · · + · · · and we compare like powers of h in the expressions (4. and assume that the numerical method admits a Taylor series expansion of the form Φh (y) = y + hF (y) + h2 D2 (y) + h3 D3 (y) + · · · (4. where ϕt.3) with smooth Dj (y). v) = q + hv + h f (q) 2 v + h f (q) + h f q + hv + 2 2 2 h2 2 f (q) . Hairer. Lubich and G. v) = 1 4 0 . v)T and F (y) = (v. f (q)f (q) + f (q)(v.2) into a Taylor series (using the notation y(t) = ϕt (y)) h3 (3) h2 ¨ ˙ y (0) + y (0) + · · · ϕh (y) = y + h y (0) + 2! 3! = y + h F (y) + hF2 (y) + h2 F3 (y) + · · · + h2 F (y) + hF2 (y) + · · · 2! (4.N is the exact flow of the truncated modified equation y = F (y) + hF2 (y) + · · · + hN −1 FN (y).420 E. 4. ˙ (4. v) can be computed as in the proof of Theorem 4.. 2! 1 F (F. Consider (4. Φh (y) = ϕh. (4. f (q))T .1) with an infinitely differentiable vector field F (y). v) = 1 2 f (q) . Ch.2. Disregarding convergence issues. the differential equation (1.N (y) + O(hN +1 ). the function D2 (q. and the functions Fj (q.3) with D2 (q.. Wanner Theorem 4.5) and (4. f (q)v D3 (q. F2 (y) = D2 (y) − F3 (y) = D3 (y) − 1 F F (y). Since the St¨rmer–Verlet method is of second order.3) is of the form (4. and uniquely defines the functions Fj (y) in a constructive manner.

6) of the St¨rmer–Verlet scheme together o with the numerical solution for the initial value (p0 .1 show the exact flow of the modified differential equations (truncated after the O(h2 ) term) corresponding to the two versions (1.2) is no longer a secondorder equation like (1. v) = 0. also F2 (q. v) (4. −1. The shade of the numerical approximations (dark grey to light grey) indicates the increasing time. obviously. has to coincide with the h2 -coefficient of the exact solution and we have F2 (q. For larger (odd) j the functions Fj (q. This is explained by the fact that for f (q) = −∇U (q) . A similar computation for the version (B) of the St¨rmer–Verlet method o (see (1. q0 ) = (1. The explicit formula for F3 (q. v) become more and more complicated.5) and (1.1. We then get F3 (q. As a concrete example consider the pendulum equation for which f (q) = − sin q. v) = 0. so that correct qualitative behaviour is obtained.2). The vanishing of this function follows from the symmetry of the method (cf.3). We observe a surprisingly good agreement. f (q)f (q) + f (q)(v. v) = F4 (q.Geometric numerical integration 421 2 2 −2 0 2 4 −2 0 2 4 −2 St¨rmer–Verlet (A) o −2 St¨rmer–Verlet (B) o Figure 4.6)) gives F3 (q.8) and for the next function we obtain F4 (q.9 for the two versions of the St¨rmer–Verlet method compared to the exact o flow of their modified differential equations truncated after the O(h2 ) term. v) = 1 24 2 f (q)v .3). In both cases the solutions of the modified equation are periodic.9) and. v) = 1 12 −2 f (q)v . and higher derivatives of f (q) are involved. Numerical solution with step size h = 0. The two pictures of Figure 4. v) also shows that the modified differential equation (4. Section 4. v) (4. −4 f (q)f (q) − f (q)(v. and the numerical approximation lies near a closed curve. v) = 0.0.

Geometric properties of a numerical method have their counterparts in the modified equation. Lubich and G. .1) thus implies that ϕt. F2j (y) = 0 for j = 1. .1. by the group property of the exact flow. if the leading term of the local truncation error is Er+1 (y).2). The −h symmetry condition (2.2). . that is. then Fr+1 (y) = Er+1 (y). q) which are drawn in Figure 4. Let us explain this for the properties discussed in Sections 2 and 3. we let ϕt.9) are Hamiltonian with H3 (p. Backward error analysis tells us that Φh (y) = ϕh. q) = 1 ∇U (q)T ∇U (q) 24 1 1 − ∇2 U (q)(p.−h (y) for t = h. Φh (y) = ϕh (y) + hr+1 Er+1 (y) + O(hr+2 ). p) + and respectively.8) and (4. 2. 24 6 1 12 ∇2 U (q)(p. 4. r.5) o is applied to a differential equation (1. q) = H(p.10) holds. Theorem 4. .3). p) − ∇U (q)T ∇U (q).h (y) = ϕt. . It follows from the definition of the modified equation that for methods of order r. and the computation of (4. then every truncation of the modified differential equation is reversible with respect to the reflection ρ(q. . that is. Properties of the modified differential equation In Section 4.10) This can be proved as follows: to indicate the h-dependence of the vector field (4.3. Theorem 4. −v). then every truncation of the modified differential equation is Hamiltonian.3.−h (y) and. (Reversible systems) If the St¨rmer–Verlet method (1. Consequently.422 E. (4.10) is applied to a Hamiltonian system.h (y) denote the (formal) flow of (4. Furthermore. We thus have Φ−h (y) = ϕ−h. Φh (y) = ϕh (y) + O(hr+1 ). we have Fj (y) = 0 for j = 2. v) = (q. By Theorem 2. q) = H3 (p.4 we shall see that the numerical solution is extremely close to the exact solution of a truncated modified equation.h (y). . that is. Ch. Wanner the vector fields (4. q) + h2 H3 (p. Hairer.5) shows that this is only possible if (4. For such metho ods the modified equation has an expansion in even powers of h. To study properties of the numerical solution it is therefore justified to investigate instead the corresponding properties of the modified differential equation. (Hamiltonian systems) If the St¨rmer–Verlet method o (2. . .2. Φ−1 (y) = ϕh. the exact solutions of the truncated modified equation stay on the level curves of H(p.−h (y).1 the St¨rmer–Verlet method is symmetric.

17) with the help of the generating function (2. v) = v T (Cq + c).11) holds for j = 1. By Theorem 2. This proof was first given by Benettin and Giorgilli (1994) and Tang (1994). together with ϕN. and its ideas can be traced back to Moser (1968). Consequently. The function JFN +1 (y) is therefore the gradient of some scalar function HN +1 (y). then every truncation of the modified differential equation has I(q. v) as a first integral. (2002) for details of the proof. satisfies Φh (y) = ϕN. Chapter IX)). we only present the proof of Theorem 4. which proves (4.11) for j = N + 1. We have to prove the existence of a Hamiltonian HN +1 (y). N (this is satisfied for N = 1.3. compared to o that of (4.h (y) + hN +1 FN +1 (y) + O(hN +2 ).h (y) + hN +1 FN +1 (y) + O(hN +2 ).t (y).2) and thus to the one-step map Φh of the St¨rmer–Verlet method. we have to use the representation (2. This.2 of Hairer et al. We will show that all the coefficient ˙ functions of the modified equation can be written as Fj (y) = J −1 ∇Hj (y).18).h are symplectic transformations. because F1 (y) = F (y) = J −1 ∇H(y)). . Theorem 4. . (2002. by induction. We refer to Section IX. The proofs are based on an induction argument. For general domains.4). . which is then a Hamiltonian system with Hamiltonian H(y) + hH2 (y) + · · · + hN −1 HN (y).h (y) = I + O(h). the Hamiltonian system (2. .11) Assume.2 and by the induction hypothesis.7) is written more compactly as y = J −1 ∇H(y) with J of (2. 2. that (4. The last argument of this proof requires that the domain be simply connected.19). then every truncation of the modified differential equation is divergence-free. With y = (p.3. . Its flow ϕN.4. Φh and ϕN. and also Φh (y) = ϕN. (4. the matrix JFN +1 (y) is symmetric. q). (Divergence-free systems) If the St¨rmer–Verlet method o (1.24) is o applied to a differential equation with a first integral of the form I(q. Since they are all very similar (see Hairer et al.24) is applied to a divergence-free system of the form (2. for the case where the Hamiltonian H(p. (First integrals) If the St¨rmer–Verlet method (1.5.Geometric numerical integration 423 Theorem 4. The idea is to consider the truncated modified equation (4.9). q) is defined on a simply connected domain. Proof. therefore implies J = Φh (y)T JΦh (y) = J + hN +1 FN +1 (y)T J + JFN +1 (y) + O(hN +2 ).

424 E. It is therefore natural to truncate the modified equation after N terms. Exponentially small error bounds were first derived by Benettin and Giorgilli (1994).2) we obtain (1 + b2 h2 + b4 h4 + b6 h6 + · · · )(e−h − 2 + eh ) = h2 . respectively. The following estimates are from Hairer et al. However. This is the case for analytic f (t) with finite poles. 306). Wanner 4.14) tells us that the terms of the series decrease until 2k approaches the value 2πR/h. If we try to compute the modified becomes autonomous after adding t equation for the St¨rmer–Verlet method.15) hold for y ∈ B2R (y0 ) and y ∈ BR (y0 ).6. Hairer.12) then behave like h2k b2k f (2k) (t) ≈ Const √ h2k (2k)! ≈ Const 4πk 2k (R · 2π) h · 2k R · 2π e 2k (4. ¨ Putting f (t) = et . p. and inserted into (1. let the coefficients Dj (y) of the method (4. Theorem 4. we are readily convinced that its o q-component is of the form ¨ (4. Consequently. The individual terms of the modified equation (4. and assume that F (y) ≤ M and Dj (y) ≤ µM 2κM R j−1 (4.1 proves a statement that is valid for all N ≥ 1.3) be analytic in BR (y0 ).14) (using Stirling’s formula).4. (4. and it is natural to ask which choice of N gives the best estimate. If h ≤ h0 /4 with . To find a reasonably good truncation index N for general differential equations. Even for very small step sizes h this expression is unbounded for k → ∞. the solution of this modified equation is q(t) = C1 + tC2 + (1 + b2 h2 + b4 h4 + b6 h6 + · · · ) et . Let F (y) be analytic in B2R (y0 ).13) This shows that 1+b2 h2 +b4 h4 +· · · is analytic in a disc of radius 2π centred at the origin.12) q (t) = f (t) + h2 b2 f (t) + h4 b4 f (4) (t) + h6 b6 f (6) (t) + · · · . formula (4. Ch. so that the series (4. Exponentially small error estimates Theorem 4. Now consider functions f (t) whose derivatives grow like f (k) (t) ≈ k! M R−k. Lubich and G. we have to know estimates for all derivatives of F (y) and of the coefficient functions Dj (y) of the Taylor expansion of the numerical flow. and then they tend rapidly to ∞. One convenient way of doing this is to assume analyticity of these functions.12) cannot converge. (2002. Consider the simple differential equation q = f (t) (which ¨ ¨ = 0).7. where N ≈ 2πR/h. Example 4. the coefficients behave like b2k ≈ Const (2π)−2k for k → ∞.

To find the constants µ and κ in (4.65η + µ) depends only on the method. Figure 3.Geometric numerical integration 425 h0 = R/(eηM ) and η = 2 max κ. v) = 1 j! dj q + hv Φh (q. We just explain how the assumptions can be checked for the St¨rmer–Verlet method (4. Section IX. Φh (q. 5. v) is analytic in the complex disc |h| ≤ R/M . The quantities R and M are then given by the problem. . Cauchy’s estimate therefore yields Dj (q. f (q))T . v)T .7).7) for details.1 indicates no drift in the energy. v) are composed of derivatives of f (q) so that they are analytic on the same domain as f (q). The proof of this theorem is technical and long.2 shows that the functions Dj (q.1.4 that the total energy H(p. where γ = e(2 + 1. Long-time behaviour of numerical solutions In this section we show how the geometric properties of Section 2 turn into favourable long-term behaviour. we use F (y) = v +h f (q) ≤ M for q −q0 +h v −v0 ≤ 2R. q) of a Hamiltonian system is not preserved exactly by the St¨rmer–Verlet method. F (y) = (v.4) satisfies Φh (y0 ) − ϕN. however. see Hairer et al.t (y0 ) of the truncated modified equation (4. The computation of the beginning of Section 4. v) − v + h f (q) 2 = h 2 f (q) 2 f (q + hv + h f (q)) 2 h2 2 ≤ hM for q − q0 + h v − v0 ≤ R and for h M ≤ R.15). and we consider the scaled norm y = q + h v . Also for the Kepler problem. 5. v) − j v + h f (q) dh 2 ≤M h=0 M R j−1 for j ≥ 2. and we estimate q + hv Φh (q. This proves the estimates (4. approximately preserved.h (y0 ) ≤ hγM e−h0 /h . As the following theorem shows. µ/(2 ln 2−1) . In that exo ample it is. the Hamiltonian is in fact approximately preserved over very long times for general Hamiltonian systems. Energy conservation We have seen in Example 3. o We let y = (q. (2002. Most of the results are obtained with the help of backward error analysis. 2 Considered as a function of h.15) with µ = 1 and κ = 1/2. This follows from the fact 2 that the argument of f satisfies q + hv + h f (q) − q0 ≤ R + hM ≤ 2R. then there exists N = N (h) (namely N equal to the largest integer satisfying hN ≤ h0 ) such that the difference between the numerical solution y1 = Φh (y0 ) and the exact solution ϕN.

The second proof uses only the symmetry of the St¨rmer–Verlet method. qn ) − H(p0 .7). with a modified Hamiltonian H that is O(h2 ) close to the original Hamiltonian H in a neighbourhood of the numerical solution values. Inserting this estimate in the above sum yields the result. the first one based on the symplecticity. i. in an argument due to Benettin and Giorgilli (1994). is again Hamiltonian. qj+1 ) − H(pj . a function q n (t) with q n (0) = qn and q n (−h) = qn−1 + O(hN +1 ) satisfying q n (t + h) − 2q n (t) + q n (t − h) = h2 f (q n (t)) + O(hN +2 ) (5. and hence H(pj+1 .1) for t in some fixed interval around 0. with c proportional to 1/Ω. qn ). qn ) − H(p0 . The functions q n (t + h) and q n+1 (t) agree up to O(hN +1 ). qj+1 ) − H(pj . of the highest frequency in the linearized system. truncated after N terms. Hairer. Lubich and G. Ch. The total energy along a numerical solution (pn . where Ω is an upper bound of M −1/2 ∇2 U (q)M −1/2 1/2 . qj ) . Consider now H along the numerical solution. as do their kth derivatives multiplied with hk . qn ) of the St¨rmer–Verlet method satisfies o |H(pn . Backward error analysis. qj ) = H(pj+1 . . qj+1 ) + O(hN +1 ). the second one on the symmetry of the method. It applies to general symplectic methods for general (smooth) Hamiltonian systems (2. o It was given in Hairer and Lubich (2000b) because its arguments extend to numerical energy conservation in oscillatory systems when the product of the step size with the highest frequencies is bounded away from 0 (see Section 5.426 E. we have for its flow ϕh (pj . q0 ) = j=0 H(pj+1 . for every n. We write the deviation of H as a telescoping sum n−1 H(pn .. On the other hand. shows that there exists. Wanner Theorem 5.3 that the modified differential equation. By construction of the modified equation. q0 )| ≤ Ch2 + CN hN t for 0 ≤ t = nh ≤ h−N for arbitrary positive integer N . the flow ϕt preserves the modified Hamiltonian. The first proof uses the symplecticity of the St¨rmer–Verlet method via o backward error analysis. When the Hamiltonian is analytic. qj ) = (pj+1 . CN depends on bounds of derivatives of H up to (N + 1)th order in a region that contains the numerical solution values (pn . The constants C and CN are independent of t and h.e. or the asymptotic h2 -expansion of the numerical solution. We know from Theorem 4. We give two different proofs of this result. both proofs can be refined to yield an estimate Ch2 + C0 e−c/h t over exponentially long times t ≤ ec/h . qj+1 ) − H(ϕh (pj .4).1. qj )) = O(hN +1 ).

˙ Moreover. where the last equation follows by noting pn = M Hence. together with their kth derivatives scaled by hk (k ≤ N ).4) .2) with q n (t)T M and integrate over t.2). The key observation ˙ is now that the product of q n (t) with an even-order derivative of q n (t) is a ˙ total differential (we omit the superscript n in the following formula): q T M q (2l) = ˙ with Al [q] = q T M q (2l−1) − q T M q (2l−2) +· · ·∓(q (l−1) )T M q (l+1) ± (q (l) )T M q (l) . Moreover.5) (5. q n (0)) + O(h2 ) = H(pn . qn ) − H(p0 . for f (q) = −M −1 ∇U (q) we 2 ˙ T M f (q) = −( d/dt)U (q). A1 [q] = 1 q T M q. are equal up to O(hN +1 ). only even-order derivatives of q n (t) (and even powers of the step size) are present in (5. ˙ ¨ ˙ In particular. and hence H[q n ](h) − H[q n ](0) = O(hN +1 ). qn+1 − qn−1 = pn (0) + O(h2 ). which completes the proof. qn ) + O(h2 ). By Taylor expansion in (5. q0 ) = H[q n ](0) − H[q 0 ](0) + O(h2 ) = O(nhN +1 ) + O(h2 ). we further have H[q n+1 ](0) − H[q n ](h) = O(hN +1 ).3)–(5. 2h (5. We multiply (5. with pn (t) = M q n (t) we have H[q n ](0) = H(pn (0).3) Since the functions q n (t + h) and q n+1 (t). (5.Geometric numerical integration 427 for k ≤ N .1). For the energy functional clearly have q ˙ N/2 d Al [q] dt 1 2 H[q](t) = l=1 2 Al [q](t) h2l−2 + U (q(t)) (2l)! we thus obtain ( d/dt)H[q n ](t) = O(hN ).5). from (5. H(pn . N/2 l=1 2 d2l q n (t) h2l−2 = f (q n (t)) + O(hN ).2) Because of the symmetry of the method. (2l)! dt2l (5.

q) = ψ(a. q) via the inverse transform of (5. such that the Hamiltonian in these variables depends only on the actions H(p. .7) with the frequencies ω = (ω1 . Ij } = ∇q IiT ∇p Ij − ∇p IiT ∇q Ij = 0 for all i. . . (5. and note that the components of I = (I1 . . with planetary motion as the classical example and historical driving force. . the equations of motion are simply a = 0. . . At the other end. q). . the simplest conceivable dynamics – uniform motion on a Cartesian product of circles – appears in integrable Hamiltonian systems. θ) : θ ∈ Td } is thus invariant under the flow.9) For every a. θ) = (I(p. Wanner 5. θ)) = K(a).7) is integrable if there exists a symplectic transformation (p. θ) = . Lubich and G. This has a quasi-periodic (or possibly periodic) flow: ϕt (a. A Hamiltonian system (2. . θi ∈ R mod 2π}. The solution trajectories of the Hamiltonian systems with Hamiltonian Ii [i] exist for all time (in the action-angle variables. q) = H(ψ(a. θ). j. . Linear error growth for integrable systems General Hamiltonian systems may have extremely complicated dynamics. θ + ω(a)t). ˙ ˙ θ = ω(a). Θ(p. . {Ii . ad ) in some open set of Rd and for angles θ on the whole d-dimensional torus Td = Rd /(2πZd ) = {(θ1 . We express the actions and angles in terms of the original variables (p. . θ) = (a. θd ). . ωd )T = ∇a K (note that ∇θ K = 0). Hairer. Their practical interest lies in the fact that many physical systems are perturbations of integrable systems. θ) (5. .2. . . the torus {(a. Ch. that is. In the action-angle variables. Integrability of a Hamiltonian system is an exceptional property: the system has d independent first integrals I1 . defined for actions a = (a1 .8) (5.6) as (a. (5. . Id whose Poisson brackets vanish pairwise. q)). .6) to action-angle variables (a. their flow is simply ϕt (a. . .428 E. Id ) are first integrals of the integrable system. and little can be said about the long-time behaviour of their discretizations apart from the long-time near-conservation of the total energy considered above.

the Arnold–Liouville theorem (Arnold 1963) states that every Hamiltonian system having d first integrals with the above properties can be transformed to action-angle variables with a Hamiltonian depending only on the actions. The frequencies are ω1 = ω2 = 2π/T . and the level sets of I are compact (the invariant tori {a = Const θ ∈ Td }). q): see Figure 5. the action-angle coordinates are symplectic polar coordinates. 1 Example 5. θ + tei ) with ei denoting the ith unit vector of Rd ). L = q1 p2 − q2 p1 ). Example 5. is integrable with actions a1 = 1/ −2H and a2 = L (the angular momentum. where T = 2π/(−2H)3/2 is the period of a trajectory with total energy H. 2 2 with the transformation to action-angle coordinates given by √ 2a cos θ p . q) = 1 p2 + 1 q 2 is integrable. 2 2 Example 5. Here. Transformation to action-angle variables.1. The Kepler problem. (a. Flaschka 1974). The Hamiltonian is d H(p. q) = k=1 1 2 p + exp(qk − qk+1 ) 2 k . which describes a system of particles on a line interacting with exponential forces. The harmonic oscillator H(p. with H(p. q) = 1 (p2 +p2 )−(q1 +q2 )− 2 √2 1 2 in the range H < 0.1. Conversely. A further celebrated example of an integrable system is the Toda lattice (Toda 1970.Geometric numerical integration p 2 1 429 a ψ −1 2 1 0 π 2 0 −1 −2 q π 3π 2 θ 2π Figure 5.2.4. = √ q 2a sin θ with a = H(p.3.

and the perturbation theory of integrable systems.66 . 0. We observe a linear error growth for the St¨rmer–Verlet method.60 −2. The effect of a small perturbation . −1)T and p0 = (−1. Siegel and Moser 1971).5)T .1 (left) and h = 0. . by means of which the numerical map is interpreted as being essentially the time-h flow of a modified Hamiltonian system.96 1.3. The eigenvalues of the matrix  b1 bd a2 b2 0   ak = − 1 pk ..2 shows the eigenvalues of L along the numerical solution of the St¨rmer–Verlet and the second-order Runge method o obtained with step sizes h = 0. a rich mathematical theory originally developed for problems of celestial mechanics (Poincar´ e 1892/1893/1899. q) is plotted in Figure 5.  2 .64 −2.60 λ3 t 10 20 30 40 50 t 50 −2. b2 .94 h = 0.2. with periodic  a1  b1   L =    bd extension qd+1 = q1 ..64 λ1 λ2 .94 h = 0. Lubich and G. . The global error in (p. 2 2 .58 −2. Ch.05 (right) on the interval 0 ≤ t ≤ 50. 1.58 −2.5. Hairer. Wanner λ 1.96 1. in contrast to a quadratic error o growth for the second-order Runge method. ad−1 bd−1  0 bd−1 ad are first integrals whose Poisson brackets vanish pairwise.  bk = 1 exp 1 (qk − qk+1 ) . The study of the error behaviour of the numerical method combines backward error analysis. We consider the case d = 3 and choose initial values q0 = (1.430 E..62 10 20 30 40 λ2 St¨rmer–Verlet o λ3 −2. . but all d first integrals of the integrable system are well approximated over long times with an error of size O(h2 ). Toda eigenvalues along the numerical solution. This is explained by Theorem 5.1). Not only the Hamiltonian (Theorem 5.62 Figure 5.66 .5 below. 2.1 1. Figure 5.05 λ1 Runge 2 .

of an integrable system is well under control in subsets of the phase space where the frequencies ω satisfy Siegel’s diophantine condition: |k · ω| ≥ γ|k|−ν for all k ∈ Zd (5. q(t)) ≤ C t h2 . h = 0. By backward error analysis.10). For general numerical integrators applied to integrable systems (or perturbations thereof) the error grows quadratically with time. with the flow of the modified differential equation. Section X.10) for some positive constants γ and ν. c. the numerical method coincides. which is a Hamiltonian . h = 0. qn ) starting with frequencies ω0 = ω(I(p0 .02 o 50 100 Figure 5. C depend on d.0 global error Runge 2. Theorem 5. as is shown by the following result from Hairer et al.02 St¨rmer–Verlet. For any choice of γ and ν the complementary set is.3. γ. satisfies (pn . for t = nh ≤ h−2 . almost all frequencies (in the sense of the Lebesgue measure) satisfy (5. ν and on bounds of the Hamiltonian. The basic steps of the proof are summarized in Figure 5.4.Geometric numerical integration 431 .8 . up to arbitrary order in h. For symplectic methods such as the St¨rmer–Verlet method there is linear error growth and long-time o near-preservation of the first integrals Ii .2 .3). c and h0 such that the following holds for all step sizes h ≤ h0 : every numerical solution (pn .5. Consider applying the St¨rmer–Verlet method to an inteo grable system (2. and there is a linear drift away from the first integrals Ii . however.4 . q0 ) ≤ C h2 . The constants h0 .7) with real-analytic Hamiltonian. Suppose that ω ∗ ∈ Rd satisfies the diophantine condition (5. qn ) − I(p0 . qn ) − (p(t). q0 )) such that ω0 − ω ∗ ≤ c| log h|−ν−1 . with |k| = i ki . I(pn . Global error of the St¨rmer–Verlet and o the second-order Runge method on the Toda lattice.10) for some γ > 0. open and dense in Rd . (2002.6 . Then there exist positive constants C. For ν > d − 1.

with the result b(t) − b0 ≤ C t εN +1 . qn ) action-angle variables (p. The error propagation is then studied in the (b. θ) modified Hamiltonian H(p.4. θ). which gives the modified Hamiltonian in the form K(a) + εK1 (a. θ) is given by a Lindstedt–Poincar´ pere turbation series. q) Hamiltonian K(a) Lindstedt–Poincar´ e series (a. where the generating function S(b. for t2 ≤ 1/εN . We are thus in the realm of classical perturbation theory. ϕ) Figure 5. θ) = χ(b. This transformation is O(ε) close to the identity.5. S(b. θ) = εS1 (b. θ) + ε2 S2 (b.5 admits extensions in several directions. ϕ = θ + ∇b S(b. Transformations in the proof of Theorem 5. the dependence on the angles in the modified Hamiltonian. ϕ) which eliminates. Transforming back to the original variables (p. q) = ψ(a. ϕ(t) − ϕ0 − ωε (b0 ) t ≤ C (t + t2 ) εN . perturbation of size ε = h2 of the original. θ). In addition to the transformation to the action-angle variables (a. Wanner backward error analysis integrable Hamiltonian H(p. q) finally yields the stated result. q) numerical solution (pn . Ch. It is constructed as b = a − ∇θ S(b. integrable system. . It is just one of a series of results on the long-time behaviour of geometric integrators. Theorem 5. θ). θ).432 E. we use a further symplectic coordinate transformation (a. Hairer. θ). Lubich and G. up to high-order terms in ε. θ) = χ(b. ϕ) modified Hamiltonian K(a) + εK1 (a. θ) + · · · + εN SN (b. θ) with ε = hp modified Hamiltonian K(b) + εK1 (b) + · · · + εN KN (b) +εN +1 R(b. with ωε (b) = ω(b) + O(ε). ϕ)-variables.

q) with a perturbation parameter of size ε = O(hα ) for some positive exponent α.5 thus extend to the Kepler problem. v) = (µ(a. θ)) that preserves reversibility.. 2002. only the reversibility of the numerical method comes into play. θ).. and an abundance of invariant tori. Chapter XI). p. 371). the near-preservation of the action variables along the numerical solution holds for times that are exponentially long in a negative power of the step size (Hairer and Lubich 1997. which has two identical frequencies ω1 = ω2 = (−2H)3/2 . . • The linear error growth remains intact when the method is applied to perturbed integrable systems H(p. which are invariant under the flow of the perturbed system. since o the angular momentum a2 = L is preserved exactly by the St¨rmer– Verlet method. i. then the numerical method has tori which are near-invariant over exponentially long times (Hairer and Lubich 1997). µ is odd in θ and ν is even in θ. deformations of the invariant tori of the integrable system corresponding to diophantine frequencies ω. • For dissipatively perturbed integrable systems.Geometric numerical integration 433 • The theorem does not apply directly to the Kepler problem.e. in view of the existence of invariant tori of the numerical integrator close to the invariant tori of the integrable system (Shang 1999. If the method is applied to such a perturbed integrable system. In that theory. • Under stronger conditions on the initial values or on the system. which are transformed to the form (5.e. However. These are differential equations with reversible flow (2. it turns out that the modified Hamiltonian written in the action-angle variables of the Kepler problem is independent of the angle θ2 . • Perturbed integrable systems have KAM tori. For a Cantor set of non-resonant step sizes there are even truly invariant tori on which the numerical one-step map reduces to rotation by hω in suitable coordinates (Hairer et al. Only the angle θ1 must therefore be eliminated via the perturbation series. 2002. and this involves only the single frequency ω1 for which the diophantine condition is trivially satisfied. ν(a. Moan 2002). where only one torus survives the perturbation and becomes weakly attractive. There is again linear error growth. long-time near-preservation of the action variables. • There is a completely analogous theory for integrable reversible systems (Hairer et al. i. The proof and result of Theorem 5. the existence of a nearby invariant torus of the numerical method is shown under weak assumptions on the step size in Stoffer (1998) and Hairer and Lubich (1999).2). not the symplecticity.8) by a transformation (q. For a Cantor set of initial values and a Cantor set of step sizes this holds even perpetually. q) + εG(p. 2000).

L. Hairer. . 1 T →∞ T lim T 0 p(t)T M −1 p(t) dt = lim 1 T →∞ T T 0 q(t)T ∇U (q(t)) dt. to a fluid state. Example 5. Consider the Poisson bracket {F. q(0)) . this condition is satisfied if the energy level set {(p. p. this has been confirmed in countless computational experiments. Statistical behaviour The equation of motion of a system of 864 particles interacting through a LennardJones potential has been integrated for various values of the temperature and density. this shows that the average is of size O(1/T ) as T → ∞.434 E. it is the computation of statistical or thermodynamic quantities. For a separable Hamiltonian (2. the time average of twice the kinetic energy equals that of the virial function q T ∇U (q).3. q) = H(p(0). i. H}(p(t). q(t)) = dt and hence the time average of the Poisson bracket along a solution is 1 T T 0 {F.e. Verlet (1967) In molecular dynamics. q(t)). q(T )) − F (p(0). q) with the Hamiltonian. H}(p(t). The equilibrium properties have been calculated and are shown to agree very well with the corresponding properties of argon. q(0))} is compact. p. 243) and Gallavotti (1999.8) the choice F (p. Abraham and Marsden (1978. q(t)) dt = 1 F (p(T ). It comes as a consequence of the longtime near-conservation of energy. Since Verlet (1967). The success of the St¨rmer–Verlet method in this field lies in the o observation that the method is apparently able to reproduce the correct statistical behaviour over long times. 129). q) : H(p. relative. Wanner 5. q) = pT q yields the virial theorem of Clausius (Gallavotti 1999. In particular. p. rather than single trajectories. Ch. {F. but to our knowledge there are as yet no rigorous mathematical results in the literature explaining the favourable statistical behaviour. cf. In the following we derive a result which is a discrete analogue of the virial theorem of statistical mechanics. 129). Backward error analysis gives indications as to why this might be so. Lubich and G. generally.. H} = ∇q F T ∇p H − ∇p F T ∇q H of an arbitrary differentiable function F (p. Along every solution (p(t). such as temperature. which is of interest. q(t)) of the Hamiltonian system we have d F (p(t). T If F is bounded along the solution.6.

H}(pn . H}(yn ) + {F. we have = 0 {F. the left-hand side of (5. H}(yn ) = n=0 1 F (yN ) − F (y0 ) + O(h2 ). In particular. Let H(p. q0 ) ∈ Kδ/2 . The Hamiltonian of the system is (3. h0 > 0 depend on bounds of H on a 2 complex neighbourhood of Kδ and on bounds of the first three derivatives of F on Kδ . By Theorem 5. c. It considers NA atoms interacting by the LennardJones potential V (r) = 4 ε σ r 12 − σ r 6 . H}(ϕt (y)) = h h {F. Theorem 5. Nh which yields the stated estimate. The constants C. q) : |H(p. we know that yn := (pn . bounded o by µ on Kδ . qn ) obtained by the St¨rmer– Verlet method satisfies 1 N N {F. Then the numerical solution (pn . d dt Since yn+1 = ϕh (yn ) + O(h3 ) and {F. Hence 1 N N {F. qn ) ≤ n=0 2µ + Ch2 Nh for N h ≤ ec/h and h ≤ h0 . H} ϕt (yn ) dt + O(h3 ) = F (yn+1 ) − F (yn ) + O(h3 ). (5. Let F (p.Geometric numerical integration 435 For the numerical discretization there is the following result. We choose NA = 7 . q) − H0 | ≤ δ} is compact for some δ > 0. H}(yn+1 ) 2 2 h F (ϕt (y)).11) where the prime on the sum indicates that the first and last term are taken with weight 1 . Proof. but they are independent of h and (p0 .1 and the remark thereafter. We give a numerical experiment with a small-scale version of Verlet’s argon model. q) be any smooth real-valued function.11) is O(h2 ) for h−2 ≤ N h ≤ ec/h . q) be a real-analytic Hamiltonian for which Kδ = {(p.2) with Vij = V .7. Example 5.8. qn ) ∈ Kδ for nh ≤ ec/h .

15). temperature and virial function of the argon crystal. Ford 1992).4. The size of the oscillations in the Hamiltonian is proportional to h2 . whereas that in the temperature and in the virial function is independent of h. Pasta and Ulam 1955. and the data of Biesiadecki and Skeel (1993). Figure 5. 5. Wanner total energy temperature virial function Figure 5. We have already mentioned the Toda lattice. and the virial function NA i−1 C(q) = i=2 j=1 V (rij )rij (with rij = qi − qj ) over an interval of length 4 · 105 [fsec]. An equally famous problem is the Fermi–Pasta–Ulam model (Fermi. obtained by the St¨rmer–Verlet method with step size h = 40 [fsec]. Lubich and G.5.436 −1240 −1260 −1280 40 20 0 5000 0 −5000 E. where . Hairer. Oscillatory differential equations Nonlinear mass-spring models have traditionally been very useful in explaining various phenomena of more complicated ‘real’ physical systems. the temperature 1 1 T (p) = NA kB 2 m NA pi i=1 2 (kB is Boltzmann’s constant).5 shows the Hamiltonian. respectively.8. At the end of the integration (after 10 000 steps) the averages of twice the kinetic energy and of the virial function are 217. (2002. The units in the o figure are such that kB = 1. see also Hairer et al.1 and 217. Computed total energy. 10 000 steps of size h = 40 [fsec]. Ch. p.

x1 x2 ··· x2m−1 x2m soft nonlinear stiff harmonic Figure 5. the oscillatory energy I defined as I = I1 + I2 + I3 2 2 with Ij = p2 m+j + ω qm+j .7 we have plotted the following quantities as functions of time: the Hamiltonian H (actually we plot H − 0. backward error analysis is no longer applicable. We are interested in using step sizes h for which the product with the highest frequency ω in the system is bounded away from zero. but the special form of the nonlinearity is not important. Consider a chain of 2m mass points. In terms of the new variables √ √ qm+i = (x2i − x2i−1 )/ 2 qi = (x2i + x2i−1 )/ 2. where ω 1 is a large parameter. For an illustration we consider m = 3 and choose ω = 30. Example 5. Here we assume cubic nonlinear springs. Chain with alternating soft nonlinear and stiff linear springs. since the ‘exponentially small’ error terms are then of size O(e−c/hω ) = O(1). 1 2 1 2 . .) In this situation. Giorgilli. (Values of hω ≈ 1/2 are routinely used in molecular dynamics. The variables x1 . . x2m stand for the displacements of the mass points. (which represent a scaled displacement and a scaled expansion/compression ˙ of the ith stiff spring) and the momenta pi = qi .Geometric numerical integration 437 a nonlinear perturbation to a primarily linear problem is studied over long times. Here we use a variant of this problem for gaining insight into the longtime energy behaviour of the St¨rmer–Verlet method applied to oscillatory o systems with multiple time scales. .8 for graphical reasons).9. q) = 1 2 2m p2 + i i=1 m−1 ω2 2 m 2 qm+i + i=1 1 (q1 − qm+1 )4 4 + i=1 (qi+1 − qm+i+1 − qi − qm+i )4 + (qm + q2m )4 . the motion is described by a Hamiltonian system with H(p.6. .6. Martinoli and Vanzini (1992) and Figure 5. see Galgani. and fixed at the end points. connected with alternating soft nonlinear and stiff linear springs. In Figure 5.

1 2 3 T2 = 1 2 p2 + p2 + p2 . 0 ωI (5. In the last picture we show the results obtained by the St¨rmer–Verlet o method with step size h = 0. and on the slow scale ω there is an energy exchange between the stiff linear springs. the oscillations remain of the same size if h decreases and ω increases such that hω remains constant. Wanner H 1 H I 1 I T2 T1 T2 0 .04 . However. The oscillations in I are of size O(ω −1 ) uniformly for h → 0.5/ω 2 H I I I1 I2 I3 1 0 50 100 150 0 2500 5000 7500 Figure 5. and energy conservation of the St¨rmer–Verlet method (last picture). For fixed ω the size of the oscillations in H is proportional to h2 . T1 = 1 2 p2 + p2 + p2 . Ch. Lubich and G. Hairer.438 E.5/ω. o and the kinetic energies of the mass centre motion and of the relative motion of masses joined by a stiff spring. We note that both H and I are approximately conserved over long times. on scale ω 0 there is the motion of the soft springs driven by the nonlinearity.7. For the first three pictures the solutions were computed with high accuracy.08 0 H 1 1 h = 0. 4 5 6 The system has different dynamics on several time scales: on the fast scale ω −1 the motion is nearly harmonic in the stiff springs. The equations of motion for the above example are of the form q = −Ω2 q − ∇U (q) ¨ with Ω = 0 0 . Different time scales in a Fermi–Pasta–Ulam problem.12) with a single high frequency ω 1 and with a smooth potential U (q) whose derivatives are bounded independently of ω. In addition to the total energy .

Let ω be √ defined by the relation sin( 1 hω) = 1 hω and suppose | sin( 1 khω)| ≥ c h for 2 2 2 k = 1. Then.12) with a step size h for which 0 < c0 ≤ hω ≤ c1 < 2. 2 2 the system has an adiabatic invariant: over times exponentially long in ω. The oscillations do not become smaller when h is decreased but ω is increased such that their product hω is kept fixed. Now consider applying the St¨rmer–Verlet method to such a system. for 0 ≤ nh ≤ h−N +1 . H(p. q) = 1 T p 2 ω2 T 0 0 p+ q 0 I 2 0 0 q 0 I (5. . I n = I 0 + O(h). remain constant up to O(h) over long times even when hω is bounded away from zero. but stay within an O((hω)2 ) band over long times. q) with (0 I)q = O(ω −1 ). h. The o step size is then restricted to hω < 2 for linear stability.13) is preserved up to O(ω −1 ). This holds uniformly for all initial values for which the total energy is bounded by a constant independent of ω. Suppose further that the total energy at the initial value (p0 . i. Let the St¨rmer–Verlet method be applied to the problem o (5. Nevertheless. |jh|≤T /2 for an arbitrary fixed T > 0. the oscillatory energy I(p. qn+j ). the following result shows that the time averages of the total and oscillatory energies Hn = In = h T h T H(pn+j . for bounded (p. . Theorem 5. but within the range of linear stability. q) = 1 1 T p p + q T Ω2 q + U (q). and that the numerical solution values qn stay in a region where all derivatives of the potential U are bounded.13) oscillate rapidly. . .10. the time averages of the total and the oscillatory energy along the numerical solution satisfy H n = H 0 + O(h). qn ) of (5.4 shows.. q0 ) is bounded independently of ω. . qn+j ). ω with the above conditions.Geometric numerical integration 439 as a conserved quantity. The constants symbolized by O are independent of n. N for some N ≥ 2 and c > 0.e. qn ) and the oscillatory energy I(pn . The Hamiltonian H(pn . |jh|≤T /2 I(pn+j . as Example 3.

but this argument breaks down for hω bounded away from zero as in Theorem 5. the long-time near-preservation of the adiabatic invariant I can be shown using backward error analysis (Reich 1999b). see Hairer et al.10 can be improved to O(h2 ) if a weighted time average is taken. where the coefficients zk (t) together with all their derivatives (up to some arbitrarily fixed order) are bounded by O(ω −|k| ). q) + q T 0 I ∇U (q). however. remain O(h) close to the initial values H(p0 .440 E. q) = I(p. Wanner It should. q = ∇p H(p. in terms of these modified energies. This permits us to use similar arguments to the second proof of Theorem 5.1 to infer the existence of certain modified energies H ∗ and I ∗ . in general. Chapter XIII) for the full proof. −2 ) over exponentially long time intervals. q0 ) and I(p0 . It is based on representing the numerical solution locally (on bounded intervals) by a modulated Fourier expansion qn = |k|<N zk (t)eikωt + O(hN ) for t = nh.2) (but of higher dimension).10. 6. be noted that the time averages H n and I n do not. The estimates of Theorem 5. Lubich and G. In this situation it is favourable to formulate the problem as a Hamiltonian system with constraints. Finally. Constrained Hamiltonian systems A minimal set of coordinates of a mechanical system is often difficult to find. up to O(h). The expansion coefficients yk (t) = zk (t)eikωt satisfy a system of equations similar in structure to (5. which is preserved up to O(ω For hω → 0. replacing the characteristic function of the interval [−T /2.1) . the time averages H n and I n can be expressed. and 0 0 if the oscillatory energy I is replaced by J(p. ˙ (6. We comment only briefly on the proof of Theorem 5. Formulation as differential-algebraic equations We consider a mechanical system with coordinates q ∈ Rd that are subject to constraints g(q) = 0. q0 ). or frequent changes of charts are necessary along a solution of the system. T /2] by a smooth windowing function with bounded support. q). The minimal coordinates may be defined only implicitly.1. which the numerical method preserves up to O(h) over times h−N +1 . (2002. The equations of motion are then of the form p = −∇q H(p. Ch. 6.10. A similar representation holds for pn . ˙ 0 = g(q). q) − ∇q g(q)λ. Hairer.

q) ∂q − ∇q g(q)T ∇2 H(p. ∇q g(q)T ∇p H(p. q). q) = 0 . hence. ˙ q = ∇p H(p. and not only for g(q) = 0. q). if the matrix is invertible (6. q) = H(p. .1) gives the ordinary differential equation p = −∇q H(p.q) M denotes the tangent space of M at (p.8). . q).5) for (p. T(p. • For Hamiltonians satisfying H(−p. q0 ) ∈ M with M = (p. q) = (−p.6) the solution stays on the manifold M for all t. q)ξ)T J ϕt (p. the flow ϕt is ρ-reversible for ρ(p. . q) ∇p H(p. q) in the sense that (2. q)η = ξ T J η for ξ. q) into (6. and the product ϕt (p.q) M. . ∇q gm is the transposed Jacobian matrix of g(q). and ∇q g = ∇q g1 . q) and g(q) are defined. q) − ∇q g(q)λ(p.1) is a mapping ϕt : M → M.1). . Here. q) + ∇q g(q)λ .4) (∇2 H denotes the Hessian matrix of H). which means that (ϕt (p. (6. q) is usually given by (2. q) ˙ (6. ∂ ∇q g(q)T ∇p H(p. Inserting the so-obtained function p λ(p. the flow of (6. we differentiate the constraint 0 = g q(t) with respect to time. q)ξ has to be interpreted as the directional derivative on the manifold. The standard theory for ordinary differential equations can be used to deduce existence and uniqueness of the solution. To compute the Lagrange multiplier λ.5) holds for (p. • The flow ϕt is a symplectic transformation on M. q)∇q g(q) p (6. q). (6. .2) which is an invariant of the flow of (6. which is well defined on the domain where H(p. gm (q) is the vector of constraints. • Whenever the initial values satisfy (p0 . . q) : g(q) = 0. . p and q are T vectors in Rd .7) Here. This yields the so-called hidden constraint 0 = ∇q g(q)T ∇p H(p. η ∈ T(p. A further differentiation gives 0= (6.3) which allows us to express λ in terms of (p.1) are the following. q) ∈ M. p ∇q g(q)T ∇2 H(p. . q) ∇q H(p. g(q) = g1 (q). q) ∈ M. q).Geometric numerical integration 441 where the Hamiltonian H(p. Important properties of the system (6.

a). The reversibility is a consequence of the fact that H(−p. Ch.5) on the manifold M. θ0 = 1. θ0 = −1. Figure 6. a .1. q) = H(p.1 and ϕ0 = 1. q) implies λ(−p.1) with H(p. q) + g(q)T λ(p. a first integration of (6.3 2.3 2. 0.2) and a second integration yields g(q) = 0 along the solution of (6. To prove the symplecticity. Lubich and G. (6. we consider a particle moving on the unit sphere attracted by a fixed point a on the sphere. q). q). For (p0 . The potential is given as a fundamental solution of the Laplace–Beltrami equation on the sphere: U (q. Hairer. 0. q) = H(p. q0 ) ∈ M.8) The Kepler problem on the sphere is then of the form (6. Its flow is symplectic and coincides with that of (6. q) = 1 T p p + U (q. (Kepler and two-body problems on the sphere) Following Kozlov and Harin (1992). we consider the (unconstrained) Hamiltonian system with K(p. .8)T and to initial values given in spherical coordinates ˙ ˙ by ϕ0 = 1. Wanner The first of these properties follows from the definition of λ(p. q).3) gives (6. The flow of (6. q) = λ(p. Solutions of the Kepler problem (left) and the two-body problem on the sphere (right). 2 g(q) = q T q − 1. sin ϑ cos ϑ = q.1.1.1.1 shows the solution corresponding to the point √ a = (0. The left √ picture of Figure 6.5).2. Example 6.5) and hence also its restriction onto M is thus ρ-reversible.442 E. a) = − cos ϑ . The point a and the initial value are indicated by a larger symbol in Figure 6.

This is of the form (6.9. In terms of Q. 1 2 3 where the angular velocity Ω = (Ω1 . d3 ) is given by the relations I1 = d2 +d3 . ϕ20 = 0.2. (Rigid body) The motion of a rigid body with a fixed point chosen at the origin can be described by an orthogonal matrix Q(t).4).1.1) and satisfies the regularity condition (6. its kinetic energy is T = 1 2 I1 Ω2 + I2 Ω2 + I3 Ω2 . I2 = d3 +d1 .6 and ϕ10 = 0. Ω =  Ω3 −Ω2 Ω1 0 (Arnold 1989. ϕ20 = ˙ ˙ ˙ ˙ −0. Ω2 .3. Development of the Rattle algorithm The most important numerical algorithm for the solution of constrained Hamiltonian systems is an adaptation of the St¨rmer–Verlet method. The solution with initial values ϕ10 = −0. q1 . θ10 = 1. With P = ∂T /∂ Q = QD. the kinetic energy on the manifold O(3) = {Q | QT Q = I} becomes T = 1 2 trace (ΩDΩT ) = 1 2 ˙ ˙ trace (QT QDQT Q) = 1 2 ˙ ˙ trace (QDQT ). q2 ) = qi qi − 1 for i = 1. I3 denote the moments of inertia of the body.5.2.8.9) where Λ is a symmetric matrix consisting of Lagrange multipliers. we are thus concerned with H(P. Letting I1 . The constraints are gi (q1 . Q = P D−1 . Chapter 6). θ20 = 0. q2 ) given by (6. θ20 = −0. q2 ) T with U (q1 .1 is plotted in Figure 6. d2 . ˙ ˙ and I3 = d1 + d2 . . ˙ 0 = QT Q − I. q2 ) = 1 T p p 2 1 1 + 1 T p p 2 2 2 + U (q1 . For the two-body problem the Hamiltonian is H(p1 . Ω3 )T of the body is defined by   0 −Ω3 Ω2 ˙ 0 −Ω1  = QT Q. 6.1 (right).8). Example 6. 2. Its o historical development is in three main steps. Q) = 1 2 trace (P D−1 P T ) + U (Q). θ10 = −0. (6.Geometric numerical integration 443 Whereas in Euclidean space the two-body problem reduces to the Kepler problem. I2 . this is not the case on the sphere.3. and the constrained Hamiltonian system becomes ˙ P = −∇Q U (Q) − QΛ. where D = diag (d1 . p2 .

The pcomponents. ∇p H(pn+1/2 . Lubich and G. whereas the remaining two are equations for (pn+1 .444 E. the problem is a second-order differential equation M q = −∇q U (q) − ∇q g(q)λ with constraint g(q) = 0. pn+1 = pn+1/2 − ∇q U (qn+1 ) + ∇q g(qn+1 )λn+1 .10) This algorithm (called Shake) was originally proposed by Ryckaert. Ch.11) h 2 qn+1 = qn + hM −1 pn+1/2 . As a remedy. qn+1 ) lies on the solution manifold M.12) 0 = ∇q g(qn+1 )T M −1 pn+1 . qn+1 ) . (6.13) h 2 0 = g(qn+1 ). qn ) + ∇p H(pn+1/2 . µn ). obtained by a formal analogy to formula (1. For Hamiltonians H(p.5). (6. qn+1 . λn ). Jay (1994) and Reich (1993) observed independently that the Rattle method can be interpreted as a partitioned Runge–Kutta method and thus allows the extension to general Hamiltonians pn+1/2 = pn − qn+1 = qn + h 2 h 2 ∇q H(pn+1/2 . qn ) ∈ M. Andersen (1983) suggests replacing the last line in (6. The first three equations of (6. not used in the recursion. qn+1 ) + ∇q g(qn+1 )µn . Ciccotti and Berendsen (1977) for computations in molecular dynamics. 0 = g(qn+1 ).2) is qn+1 − 2qn + qn−1 = −h2 M −1 ∇q U (qn ) + ∇q g(qn )λn .13) determine (pn+1/2 . pn+1 = pn+1/2 − ∇q H(pn+1/2 . reads pn+1/2 = pn − h 2 ∇q U (qn ) + ∇q g(qn )λn .2). . Hairer. Third step. because λn+1 is not yet available at this step (it is computed together with qn+2 ). 0 = ∇q g(qn+1 )T ∇p H(pn+1 . This modification. Second step. qn+1 ) whenever (pn . This formula cannot be implemented. is motivated by the fact that the numerical approximation (pn+1 . are approximated by pn = M (qn+1 − qn−1 )/2h. 0 = g(qn+1 ). (6. The most ¨ natural extension of (1. qn ) + ∇q g(qn )λn . Section 2. called Rattle. (6. Wanner First step. q) = 1 pT M −1 p + U (q) with constant 2 mass matrix M (cf.11) with the projection step pn+1 = pn+1/2 − h 2 ∇q U (qn+1 ) + ∇q g(qn+1 )µn . A one-step formulation of this method.

these equations have a locally unique solution (Hairer et al.9) P1/2 = P0 − ∇Q V (Q0 ) − Q0 Λ1 . 1 Figure 6. For consistent initial values. 1 h 2 (6. obtained with the Rattle method using step size h = 0.114 on very long time intervals.3. Since the constraint g(q) is quadratic and the Hamiltonian is separable. see Figure 1.2. the formulae (6.Geometric numerical integration 445 For a sufficiently small step size. Numerical solution of the Kepler problem on the sphere.1. (Kepler problem on the sphere) We apply the Rattle method with a large step size h = 0. 2002. h 2 QT Q1 = I. Example 6.07 to the problem of Example 6. (Rigid body) yields h 2 h 2 The Rattle method (6. shows a precession as it appears in computations with symplectic integrators for the Kepler problem in Euclidean space. p.13) applied to (6. Example 6. We remark that the value of the Hamiltonian along the numerical solution oscillates around the correct value and the energy error remains bounded by 0. plotted in Figure 6. 214).4. 1 where both Λ1 and Λ2 are symmetric matrices.2. 0 Ω1 = QT P1 D−1 . T D−1 P1 Q1 + QT P1 D−1 = 0. Q0 is orthogonal and QT P0 D−1 = Ω0 is skew-symmetric. for which a scalar quadratic equation needs to be solved. Q1 = Q0 + hP1/2 D−1 . 0 0 Ω1/2 = QT P1/2 D−1 . The numerical solution.14) P1 = P1/2 − ∇Q V (Q1 ) − Q1 Λ2 .13) are explicit with exception of the computation of λn .5. . Working with 0 ˙ Ω0 = QT Q0 = QT P0 D−1 .07.

see Hairer (2003).6 is by computing Φh (p. Ω0 ) → (Q1 . h The proof is by straightforward verification. Geometric properties of Rattle For consistent initial values (pn . Theorem 6.11)–(6.q) M. The Rattle method is symplectic. that is. and by verifying the identity (6. Theorem 6. 6. qn ) ∈ M. . Further proofs are based on the interpretation as a variational integrator (Marsden and West 2001). Hairer.15). q). the method is reversible with respect to the reflection ρ(p. it satisfies ρ ◦ Φh = Φ−1 ◦ ρ on M. η ∈ T(p. This algorithm for the simulation of the heavy top is proposed in McLachlan and Scovel (1995).18). • compute a skew-symmetric matrix Ω1 such that Ω1 = Ω1/2 − QT ∇Q V (Q1 )D−1 − (Ω1/2 + ΩT ) − Λ2 D−1 1 1/2 holds with a symmetric matrix Λ2 . We thus have a numerical o flow Φh : M → M. P1 . One proof of Theorem 6. q)ξ using implicit differentiation. q)η = ξ T J η This result was first proved by Leimkuhler and Skeel (1994) for the method (6. as for the St¨rmer–Verlet o method.446 E.15) h 2 h 2 h 2 h 2 (Φh (p. The geometric properties of Section 2 for the St¨rmer– Verlet method extend to this algorithm. −h For Hamiltonians satisfying H(−p.6. The Rattle method is symmetric. and on explicit formulae of a generating function as in (2. the Rattle method (6. that is. (6. Φh = Φ−1 on M. qn+1 ) which is again on M.3. Lubich and G.5. Wanner instead of P0 . P1/2 .12). • compute Q1 = Q0 (I + hΩ1/2 ). q). q) = (−p. Ω1 ) : • find an orthogonal matrix I + hΩ1/2 such that Ω1/2 = Ω0 − QT ∇Q V (Q0 )D−1 − Λ1 D−1 0 holds with a symmetric matrix Λ1 .13). that is. An efficient implementation uses the representation of the appearing orthogonal matrices by quaternions (Hairer 2003).13) yields an approximation (pn+1 .14) become the following integrator (Q0 . q)ξ)T J Φh (p. Ch. for ξ. and by Jay (1994) and Reich (1993) for the general case (6. q) = H(p. the equations (6.

Dangers of multiple time step methods. V. • the structure-preserving use of variable step sizes. Springer. P. G. Reading. Marsden and West (2001) and McLachlan and Quispel (2002). 74. B. Phys. Marsden (1978). Geometric integration beyond St¨rmer–Verlet o In this article we deliberately considered only the St¨rmer–Verlet method o and a few selected geometric properties. C. 1117–1143. Surveys 18. Trans. I. and in the survey articles by Iserles. etc. Symposium on Differential Geometry and Differential Equations. Arnold (1963). 337–345. symmetric composition. D.Geometric numerical integration 447 7. Budd and A. Geometric Integration: Numerical Solution of Differential Equations on Manifolds. problems with multiple time scales. J. V. MA. Comput. H. J. August 1984. E. such as • higher-order methods. Philos. in the special journal issue Budd and Iserles (1999). Mathematical Methods of Classical Mechanics. J. Nørsett and Zanna (2000). Lond. Ser. J. Beijing. Russian Math. 109. 4. and linear multistep methods. 318–328. Giorgilli (1994). in Proc. partitioned Runge–Kutta. • differential equations with further geometric properties such as differential equations on Lie groups. Iserles. Scovel (1994). The reader will find more on these topics in the monographs by SanzSerna and Calvo (1994) and Hairer et al. Department of Mathematics. On quadratic invariants and symplectic structure. pp. BIT 34. De Vogelaere (1956). Phys. Report No. 42–58. Statist. K. On difference schemes and symplectic geometry. 2nd edn. C. Benjamin/Cummings. New York. J. (2002). Munthe-Kaas. J. Foundations of Mechanics. 2nd edn. University of Notre Dame. for instance. I. A 357(1754). . 85–191. Bochev and C. IN. Rattle: A ‘velocity’ version of the Shake algorithm for molecular dynamics calculations. On the Hamiltonian interpolation of near to the identity symplectic mappings with application to symplectic integration algorithms. Abraham and J. Methods of integration which preserve the contact transformation property of the Hamiltonian equations. Even within the class of ordinary differential equations. eds (1999). we have not mentioned important topics of geometric integration. REFERENCES R. special issue of R. Skeel (1993). Soc. 24–34. R. Feng (1985). Comput. Phys. Biesiadecki and R. Fifth Intern. Arnold (1989). 52. Andersen (1983). Benettin and A. Small denominators and problems of stability of motion in classical and celestial mechanics.

CRC Press LLC. . Ford (1992). 38. 76. University of Gen`ve. H. P. 57– 71. Lie-group methods. M. Feng (1991). Wanner (1993). 952–953. ZAMP 39. Long-time energy conservation of numerical methods for oscillatory differential equations. 2658. Chan and Z. E. Nørsett and G.448 E. Numer. Anal. Volume-preserving algorithms for source-free dynamical systems. J. Giorgilli. Berlin. Runge–Kutta type methods for index three differential-algebraic equations with applications to Hamiltonian systems. Phys. J. Solving Ordinary Differential Equations I: Nonstiff Problems. 112. Geometric Numerical Integration: Structure-Preserving Algorithms for Ordinary Differential Equations. International Conference on Scientific Computation. Nørsett and A. pp. Hairer and Ch. F. pp. Rev. 29. 54. B. Gallavotti (1999). Jay (1994). in Numerical Analysis 1999 (G. Global modified Hamiltonian for constrained symplectic integrators. L. Zanna (2000). Heidelberg. Lubich and G. Celestial Mech. The Fermi–Pasta–Ulam problem: Paradox turns discovery. A. 271–310. Lasagni (1988). in Acta Numerica. 28– 35. Hairer and Ch. A. Physica D 59. Chicago University Press (1965). Flaschka (1974). Kozlov and A. 215–365. Iserles. V. Thesis No. MIT Press (1967). Fermi: Collected Papers. 71. Griffiths and G. Shang (1995). Leimkuhler and R. Comput. pp. The Toda lattice. eds). LA-1940. 334–348. E. 393–399. 9. J. Feng and Z. R. Galgani. Vol. 441–462. Hairer and Ch. Munthe-Kaas. Phys. Shi.. Math. K. Math. 414– 441. A. P.-C. Dynam. E. Fermi. To appear. 117–125. Los Alamos Report No. Pasta and S. Feynman (1965). L. pp. Symplectic numerical integrators in constrained Hamiltonian systems. Hairer. E. Watson. The Character of Physical Law . Hangzhou. Springer. eds). Lubich (2000b). On the problem of energy equipartition for large systems of the Fermi–Pasta–Ulam type: Analytical and numerical estimates. Vanzini (1992). A. Cambridge University Press. Z. Hairer. Appl. e V. Hairer. The life-span of backward error analysis for numerical integrators. Later published in E. Vol. Harin (1992). Ulam (1955). Math. 169–190. B 9. Springer. 451–463. F. Hairer and Ch. Ch. SIAM J. Lubich (1999). China (T. Lubich (1997). 978–988. Astronom. first published by the BBC (1965). J. Springer. O. E. 1 of Series on Appl. II: Existence of Integrals. Martinoli and S. S. 1924–1925. D. Lubich (2000a). Numer. Ch. G. Berlin. S. Invariant tori of dissipatively perturbed Hamiltonian systems under symplectic discretization. Statistical Mechanics: A Short Treatise. Canonical Runge–Kutta methods. Energy conservation by St¨rmer-type numerical o integrators. Studies of nonlinear problems. Skeel (1994). E. Physics Reports 213. Math. E. Numer. Vol. Formal power series and numerical algorithms for dynamical systems. Numer. Numer. Math. in Proc. H. E. Lubich and G. Kepler’s problem in constant curvature spaces. Wanner (2002). II. Wanner K. Hairer (2003).

Backward error analysis for numerical integrators. Splitting methods. Marsden and M. 477–496. 45–56. Numerical Hamiltonian Problems. Shang (2000). 11. J. u . Iserles. I. Japonica 25. L. University of British Columbia. 243– 286. Lectures on Hamiltonian systems. R. McLachlan and G. McLachlan and C. C. J. Numer. M.-P. Z. Sanz-Serna (1988). Paris. Stoffer (1988). 341–434. Math. Symplectic integrators for Hamiltonian problems: An overview. K. R. University of Cambridge. Nonlinearity 13. eds). R. W. Vol. Ryckaert. Some aspects of the dynamics of Hamiltonian systems. 23. D. BIT 28. Equivariant constrained symplectic integration. C. 83. Vol. 10. E. 1. in Acta Numerica. Quispel (2002). J. PhD thesis. pp. 187 of Grundlehren der Mathematischen Wissenschaft. Maeda (1980). R.Geometric numerical integration 449 R. IEEE Trans. SAM-Report No. Clarendon Press. Ciccotti and H. Department of Computer Science. Berendsen (1977). Sanz-Serna (1992). J. 405–420. Scovel (1995). C. Mem. Ruth (1983). Amer. 29. 233–256. J. G. Poincar´ (1892/1893/1899). Z. 36. 299–308. S. R. Numer. Reich (1999a). Les M´thodes Nouvelles de la M´canique C´leste. M. Springer. Atela (1992). The accuracy of symplectic integrators. J. On reversible and canonical integration methods. Soc. Discrete mechanics and variational integrators. Calvo (1994). On backward error analysis and Nekhoroshev stability in the numerical analysis of conservative systems of ODEs. 327–341. 1549–1570. S. S. in Acta Numerica. pp. 81. ETH Z¨rich. Resonant and diophantine step sizes in computing invariant tori of Hamiltonian systems. Cambridge University Press. Canonical structure and symmetries for discrete systems. MacKay (1992). KAM theorem of symplectic algorithms for Hamiltonian systems. Math. Oxford. SIAM J. Nuclear Science NS-30. J. Moser (1971). Vol. in The Dynamics of Numerics and the Numerics of Dynamics (D. 137–193. e e e e Tome I–III. H. London. Math. Siegel and J. 1–60. Appl. Nonlinearity 5. Nonlinear Sci. I. Math. in Acta Numerica. D. Runge–Kutta schemes for Hamiltonian systems. M. 88-05. Symplectic integration of constrained Hamiltonian systems by Runge–Kutta methods. Heidelberg. Cambridge University Press. 541–562. S. Gauthier-Villars. 5. A canonical integration technique. P. Shang (1999). Reich (1993). Comput. Preservation of adiabatic invariants under symplectic discretization. Lectures on Celestial Mechanics. Reich (1999b). Moser (1968). Sanz-Serna and M. pp. Anal. J. Numerical integration of the cartesian equations of motion of a system with constraints: Molecular dynamics of n-alkanes. Technical Report 93-13. 357–514. Broomhead and A. McLachlan and P. J. Chapman and Hall. Cambridge University Press. Moan (2002). S. 877–883. West (2001). I. Numer. pp. Phys. P. Vol. 2669–2671.

148–160. in Numerical Solution of Ordinary Differential Equations (S. L. Moscow. Astronom. 56. Anal. Verlet (1967). Yoshida (1993). 217. J. Theor. 159. 98–103. H. Toda (1970). Integrable maps. M. (In Russian. Comput. Filippov. Waves in nonlinear lattice. Recent progress in the theory and application of symplectic integrators. Veselov (1991). 521–545. Math. Celestial Mech. On the qualitative behaviour of symplectic integrators. Phys. ed. SIAM J. Numer. P. Dynam. Phys. 5. Anal. USSR Academy of Sciences. Math. 1–51.-F. G. Keldysh Institute of Applied Mathematics. Stoffer (1998).). Suppl. 31–39. Math. I. Ch. Surv. Computer ‘experiments’ on classical fluids. Appl. B. S. Tang (1994). Formal energy of a symplectic scheme for Hamiltonian systems and its applications.) Y. 45. 506–517. pp. 27. A. I: Thermodynamical properties of Lennard-Jones molecules. III: Perturbed integrable systems. Rev. Hairer. 174–200. Lubich and G. Russ. Suris (1988).450 E. Wanner D. 46. Strang (1968). On the conservation of the symplectic structure in the numerical solution of Hamiltonian systems. 27–43. . Progr. Appl. Y. On the construction and comparison of difference schemes.