You are on page 1of 64

LINEARIZATION METHODS FOR VARIATIONAL

INTEGRATORS AND EULER-LAGRANGE


EQUATIONS

Thesis by
Anil N. Hirani
Advisors
James R. Arvo, Jerrold E. Marsden

In Partial Ful llment of the Requirements


for the Degree of
Master of Science in Computer Science

California Institute of Technology


Pasadena, California

April 2000
ii

c
 April 2000
Anil N. Hirani
All rights Reserved
iii
Acknowledgements
I'd like to thank my advisors Prof. Jerrold E. Marsden and Prof. Jim Arvo for their invalu-
able help and advice. The comments made by the members of my candidacy committee
were also very helpful. Besides my two advisors, the committee members were Professors
Peter Schroder, Alan Barr and Tom Hou. I am also thankful to Matthew West and Mathieu
Desbrun for some very helpful suggestions. In the text I've cited the speci c result where
Matthew's suggestions were most important. Mathieu was also generous with his advice
and gave me some very good suggestions. Once, he gave me a table of coecients for nite
di erences and that really helped precipitate the completion of this work. Discussions with
Antonio Hernandez, Dong-Eui Chang and Sameer Jalnapurkar on some geometric mechanics
derivations were also helpful.
iv
v
LINEARIZATION METHODS FOR VARIATIONAL
INTEGRATORS AND EULER-LAGRANGE EQUATIONS
by
Anil N. Hirani
In Partial Ful llment of the
Requirements for the Degree of
Master of Science in Computer Science

Abstract
Hamiltonian systems arise in a wide variety of idealized physical systems and Hamilton's
equations often must be solved numerically. In general, traditional nite di erence methods
for numerically integrating ordinary di erential equations do not take into account the
special structure of Hamilton's equations. One would like the numerical solution to preserve
some or all of the properties of the continuous time solution. Variational integrators, also
called discrete Euler-Lagrange equations in this thesis, provide a way to do that.
This thesis is about di erentiating, i.e linearizing Euler-Lagrange equations and discrete
Euler-Lagrange equations by di erentiating the Lagrangian or discrete Lagrangian. The
original motivation for di erentiating the Lagrangian was to develop higher order varia-
tional integrators. However it was found that the resulting intgerators are not higher order.
Nonetheless, the pursuit of this goal turned up several interesting results :
i. Di erentiating a Lagrangian along the acceleration, i.e along (q; _ q ) results in the
Euler-Lagrange equation 0 = 0 (Lemma 2, page 20).
ii. Di erentiating a Lagrangian along a general direction (u; u_ ), determining the Euler-
Lagrange equation and then setting u = q_ results in an Euler-Lagrange equation that
is the time derivative of the original Euler-Lagrange equation corresponding to the
original Lagrangian (Lemma 3, page 21).
iii. A similar operation on the discrete Lagrangian gives a \discrete time derivative" of
the discrete Euler-Lagrange equation corresponding to the original discrete Lagrangian
(Lemma 4, page 24).
iv. The operations described in the previous two items are part of a commutative diagram
(Theorem 1, page 28).
v. Numerical measurements and analytical calculation of order of accuracy for speci c
examples using this approach show that higher order methods are not obtained in
general (Chapter 5, especially Section 5.4 and 5.5).
vi. The manifold consisting of triples that satisfy a discrete Euler Lagrange equation is a
smooth manifold (Proposition 4, page 36).
vii. The algorithm on TQ × TQ is symplectic and the algorithm that updates triples of
points on Q is presymplectic (Theorem 2, page 37).
vi
I also give a derivation of a symplectic form on TQ × TQ where Q is the con guration
manifold. Most proofs given in this thesis are done using local coordinates. A possible
direction for further study is to give intrinsic proofs in order to gain a deeper understanding
of the geometric meaning of the operations of this thesis.
vii

Table of Contents

Acknowledgments iii
Abstract v
1 Introduction 1
1.1 Variational Integrators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Preliminaries 7
2.1 Lagrangian Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Hamiltonian Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.4 Legendre Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.5 Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3 Variational Integrators 13
3.1 Discrete Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2 Invariance Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4 Linearization 15
4.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.2 Di erentiating Lagrangian along (q; _ q ) . . . . . . . . . . . . . . . . . . . . . . 16
4.3 Continuous Time Linearization . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.4 Discrete Time Linearization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.5 Discrete and Continuous . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.6 Linearization Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.7 Invariance Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.7.1 Symplectic form on TQ × TQ . . . . . . . . . . . . . . . . . . . . . . . 31
4.7.2 Presymplectic algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 34
5 Examples and Conclusion 39
5.1 Falling Particle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.1.1 Continuous . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.1.2 Discrete . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.2 Simple Harmonic Oscillator . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
5.2.1 Continuous . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
5.2.2 Discrete . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.3 Dung Oscillator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5.3.1 Continuous . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
viii
5.3.2 Discrete . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.4 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.5 Accuracy Calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.6 Using more accurate uk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
Bibliography 55
1

Chapter 1

Introduction

The traditional method of numerically solving the initial value problem (IVP) for ordinary
di erential equations (ODEs) is to discretize derivatives by nite di erences. For example
for q ∈ R and the equation
dq = f(q; t); (1.1)
dt
one can choose a step size h for stepping the time discretely and replace the left hand side
of (1.1) by
qk+1 - qk :
h
Much of the development of numerical methods for IVPs has concerned itself with what to
do with the right hand side of equation (1.1) in order to discretize the equation completely.
The development has resulted in a well developed theory of truncation errors (accuracy),
stability, and convergence. The books by Hairer et.al [6], Lambert [11], Gear [3] and Iserles
[8] are excellent sources of information about numerical methods for ODEs. The table of
coecients for nite di erence approximations for various order derivatives on pages 16 and
17 of Fornberg's book [1] is also quite handy.
But consider now the following simple system of equations, given a scalar function
H : R × R ! R such that
q_ @H (1.2)
=
@p
p_ @H :
= -
@q
This is an example of a Hamiltonian system of equations, which is the form that many
idealized equations in physical sciences take. Some equations that may not at rst appear
to be in this form are also actually Hamiltonian in a sense de ned in Chapter 2 and in
Chapter 5 of the book by Marsden and Ratiu [15].
Some examples of equations that are Hamiltonian are : equation of a particle in a
potential, rigid body equations, rigid body with a xed point in gravitational eld, Euler
equations for an ideal, incompressible uid, Maxwell-Vlasov equations for a collisionless
plasma, Maxwell's equations and many others. The study and solution of Hamiltonian
equations is thus a subject with a long tradition.
2
It can be shown that for such an equation like (1.2) the function H is constant along
solutions,i.e H_ (q(t); p(t)) = 0 for all time t if (q(t); p(t)) is a solution of the equations 1.2.
If equation (1.2) is discretized using an arbitrary method there is no guarantee that the
discrete solutions will also satisfy the property of keeping H constant along solutions. Even
worse, there may be a systematic biased error introduced which will make the computed
solution drift from the true solution. Since H is often the energy of the system this is not
acceptable for many situations. Another important property of Hamiltonian equations is
that the ow is symplectic, i.e the ow of the vector eld de ned by the equations preserves
a certain di erential form (this will be de ned in Chapter 2). One consequence of symplectic
ow is that areas in phase space are conserved. An arbitrary discretization of equation (1.2)
may not yield an algorithm that is symplectic.
However, one would like a numerical solution to have some of these properties like
energy conservation or symplecticness. For example, Figures 1.1 and 1.2, which are taken
from Iserles and Zanna [7] show how initial conditions evolve under the forward Euler and
implicit midpoint methods. The smiley face shown is a cloud of initial conditions which
are followed over time. The equation being solved is that of a simple harmonic oscillator,
namely q + q = 0 and so the smiley face should just rotate rigidly through phase space. But
in Figure 1.1 the smiley face enlarges. Figure 1.2 shows the ideal behavior. The particular
implicit method used in Figure 1.2 is symplectic, so it conserves phase space area as one
steps through time.
To see a comparison of the momentum and energy behavior of fourth order non sym-
plectic Runge-Kutta versus that of a symplectic algorithm please refer to Figure 13 on page
18 of Marsden et.al [12]. It is shown there that the non symplectic Runge-Kutta method
does not preserve momentum while the symplectic method does. Furthermore, the energy
computed by symplectic methods appears to stay near the true value while that computed
by the Runge-Kutta method does not. The behavior displayed by the symplectic integrators
in Figure 13 of Marsden et. al [12] is in some sense the best possible because of a theorem
proved in Ge and Marsden [2], that a constant time stepping integrator cannot preserve
the symplectic form, momentum, and energy simultaneously. However, Kane et.al [9] show
that by using adaptive time stepping and an appropriate de nition of symplecticness one
can indeed get symplectic, momentum and energy preserving algorithms.
In developing symplectic integrators early researchers used generating functions (which
are symplectic transformations) and Hamilton-Jacobi equation. A more recent approach is
to create variational integrators using discrete Euler-Lagrange equations, a method which
is described in the next section.

1.1 Variational Integrators


If one works with the Hamiltonian, symplectic integrators for Hamiltonian equations can
be created using generating functions. In general this is more complicated than the method
discovered by Veselov [19]. This method, which uses the Lagrangian instead of the Hamil-
tonian, can be characterized as the creation of integrators based on discrete variational
principles. It is described and further developed by Wendlandt and Marsden [20]. Given
any Hamiltonian system of equations, under some general conditions a corresponding La-
grangian exists and the Hamiltonian equations are equivalent to some variational principle.
This will be made clearer in Chapter 3. The method is to discretize the continuous La-
grangian and use a discrete variational principle.
3

Figure 1.1: Phase ow for harmonic oscillator as approximated by forward Euler method.
Horizontal axis is position and vertical axis is momentum. The upright (rightmost) smiley
face is a collection of initial conditions and others are the location of the collection at various
times. From Iserles and Zanna [7].

At least this was the current wisdom when I started working on this problem. In recent
work Kane et.al [10] and West [17] have shown that the discrete Lagrangian should be
considered to be an approximation for the action integral over one time step. But this
point of view will not be explored in this thesis.
Roughly speaking, the method used in this thesis is to replace the position variable by
some position and the velocity variable in the Lagrangian by some nite di erence. Finally
one replaces a variational principle in which an integral appears by one in which a sum
appears. Extremizing according to the principle yields a discrete Euler-Lagrange equations
which form the basis for the algorithm. It can be proved that the algorithm is symplectic
and momentum preserving.

1.2 Results
In this thesis I present some results regarding di erentiating, i.e linearizing Lagrangians
and discrete Lagrangians, the resulting Euler-Lagrange and discrete Euler-Lagrange equa-
tions and the connections between all these. The original motivation for di erentiating the
Lagrangian was to develop higher order variational integrators. However it was found that
the resulting intgerators are not, in general, higher order. Thus the discrete side results
4

Figure 1.2: Phase ow for harmonic oscillator as approximated by implicit midpoint rule.
Horizontal axis is position and vertical axis is momentum. The upright (rightmost) smiley
face is a collection of initial conditions and others are the location of the collection at various
times. From Iserles and Zanna [7].

described below are not computationally useful. Nonetheless, the pursuit of this goal turned
up several interesting results :
i. Di erentiating a Lagrangian along the acceleration, i.e along (q; _ q ) results in the
Euler-Lagrange equation 0 = 0 (Lemma 2, page 20).
ii. Di erentiating a Lagrangian along a general direction (u; u_ ), determining the Euler-
Lagrange equation and then setting u = q_ results in an Euler-Lagrange equation that
is the time derivative of the original Euler-Lagrange equation corresponding to the
original Lagrangian (Lemma 3, page 21).
iii. A similar operation on the discrete Lagrangian gives a \discrete time derivative" of
the discrete Euler-Lagrange equation corresponding to the original discrete Lagrangian
(Lemma 4, page 24).
iv. The operations described in the previous two items are part of a commutative diagram
(Theorem 1, page 28).
v. Numerical measurements and analytical calculation of order of accuracy for speci c
examples using this approach show that higher order methods are not obtained in
general (Chapter 5, especially Section 5.4 and 5.5).
5
vi. The manifold consisting of triples that satisfy a discrete Euler Lagrange equation is a
smooth manifold (Proposition 4, page 36).
vii. The algorithm on TQ × TQ is symplectic and the algorithm that updates triples of
points on Q is presymplectic (Theorem 2, page 37).
I also give a derivation of a symplectic form on TQ × TQ where Q is the con guration
manifold. The structure of the rest of this thesis is as follows. Chapter 2 is a quick review of
continuous time Lagrangian and Hamiltonian mechanics. Chapter 3 introduces variational
integrators and Chapter 4 describes the linearization results mentioned above. Chapter 5
gives examples of linearization on both the continuous and the discrete side. In this chapter
I also do some derivations of order of accuracy calculations. These are accompanied by
results from numerical experiments for measuring order of accuracy. Both, the numerical
experiments and the analytical calculations show that our method does not, in general, give
higher order symplectic integrators.
6
7

Chapter 2

Preliminaries

Before developing the theory and describing the linearization method I will give some nec-
essary de nitions, standard results and examples. This section is adapted primarily from
Lecture 5 of Marsden [13] and from Chapters 2 and 5 of the book by Marsden and Ratiu
[15].

2.1 Lagrangian Mechanics


Let Q be a con guration manifold of a mechanical system and L : TQ ! R be a function
(called a Lagrangian). Let q = (q1 ; : : : ; qn ) be local coordinates for Q and let c : [a; b] ⊂
R ! TQ be a smooth curve on TQ: Using standard abuse of notation I will write

c(t) ≡ (q(t); dqdt(t) ) ≡ (q(t); q_ (t))


for t ∈ [a; b]. The action integral S is de ned as
Zb
S= L(q(t); q_ (t))dt:
a
Then the variational principle called Hamilton’s Principle states that
d @L - @L = 0 for i = 1; : : : ; n:
S = 0 , dt (2.1)
@q_ i @qi
Definition 1 The evolution equations
d @L @L
dt @q_ i - @qi = 0 (2.2)
are called the Euler-Lagrange equations.

2.2 Geometry
Definition 2 If f : M ! R is a smooth function on a di erentiable manifold M then it can
be di erentiated at any point m ∈ M to obtain a map Tmf : TmM ! Tf(m)R. The tangent
space of R can be identi ed with itself since R is a vector space. This gives us a linear map
df(m) : TmM ! R. Thus df(m) ∈ Tm ∗
M, the dual of the vector space TmM. This map df
8
is called the differential of f. The directional derivative in the direction v ∈ TmM is
de ned to be df(m) · v.
Definition 3 A symplectic form
on a vector space Z is a nondegenerate skew sym-
metric, bilinear form on Z. The pair (Z;
) is called a symplectic vector space. The
form
is said to be (weakly) nondegenerate if
(z1 ; z2 ) = 0 for all z2 ∈ Z implies
z1 = 0.
It can be shown that there is a basis for Z in which

(u; v) = uT J v
for all u; v ∈ Z and where
 
0 I
J=
-I 0

where 0 and I are n × n zero and identity matrices for a symplectic vector space Z of dimen-
sion 2n. Thus nite dimensional symplectic vector spaces always have an even dimension.
For an example of a symplectic vector space consider Z = W × W ∗ where W is a vector
space and W ∗ is its dual. The canonical symplectic form
on Z is de ned as

((w1 ; 1 ); (w2 ; 2 )) = 2 (w1 ) - 1 (w2 )
where w1 ; w2 ∈ W and 1 ; 2 ∈ W ∗.
Definition 4 A symplectic manifold is a pair (P;
), where P is a manifold and
is a
closed (weakly) non degenerate two-form on P.
An example of a symplectic manifold is the cotangent bundle of the con gurations space.
Let the manifold Q be the con guration space of a mechanical system. Then TQ is the
disjoint union (over all points of Q) of the tangent spaces and is called the tangent bundle.
At each point p ∈ Q the tangent space TpQ is a vector space and hence there is an associated
dual vector space Tp∗Q. The disjoint union of all these over all the points of the manifold is
the cotangent bundle T ∗Q and it is a symplectic manifold.
Definition 5 Let (P1 ;
1 ) and (P2;
2 ) be symplectic manifolds. A C1 -map ' : P1 ! P2
is called symplectic or canonical if
'∗ (
2 ) =
1 :
This pull-back notation means that for all q ∈ P1, and all v; w ∈ TqP1

1p (v; w) =
2'(p) (Tp' · v; Tp' · w) (2.3)
where
1p means
1 evaluated at point p and Tp' is the derivative of ' at p. An important
result (see e.g Chapter 5 of Marsden and Ratiu [15]) is
Proposition 1 A smooth canonical transformation between symplectic manifolds of
the same dimension is volume preserving and is a local di eomorphism.
9
2.3 Hamiltonian Mechanics
The motivation for these de nitions above concerning symplectic manifolds and maps is
the following proposition which is proved by Marsden and Ratiu in Chapter 2 of their book
[15]. But rst we need a de nition of Hamiltonian vector eld on a symplectic manifold.
Definition 6 Let (P;
) be a symplectic manifold. A vector eld X on P is called Hamil-
tonian if there is a function H : P ! R such that for all v ∈ TzP


z (X(z); v) = dH(z) · v:
In this case one writes XH for X. Hamilton’s equations are the evolution equations
z_ = XH (z).
Proposition 2 The ow 't of a Hamiltonian vector eld XH consists of canonical
(i.e symplectic) transformations. Conversely if the ow of a vector eld X consists of
canonical transformations, then it is Hamiltonian.

2.4 Legendre Transform


A transform that allows us to move between Hamiltonian and Lagrangian mechanics is the
Legendre Transform de ned below.

Definition 7 Given a Lagrangian L : TQ ! R de ne a map FL : TQ ! T ∗Q called the


fiber derivative by

FL(v) · w =
d  L(v + sw)
ds s=0
where v; w ∈ TqQ.
Thus FL(v) · w is the derivative of L at v along the ber TqQ in the direction w. In a
local chart U × E for TQ where U is open in the model space E for Q,
FL(u; e) =( q; D2 L(u; e))
where D2 L is the partial derivative of L with respect to its second argument. For a nite
dimensional manifold Q with (qi ) the coordinate for Q, and (qi ; q_ i ) the induced coordinates
for TQ the ber derivative has the local expression
 
F L(qi ; q_ i ) = qi ; @@L
q_ i
Notation : In the above equation FL(qi ; q_ i ) should be read
F L(q1 ; : : : ; qn ; q_ 1 ; : : : ; q_ n ):
A similar convention will be used throughout.
10
Thus FL is given by
pi = @@L
q_ i · (2.4)
This transformation given by
FL :(qi ; q_ i ) ! (qi ; pi )
is also called the Legendre transform.
In the Euler-Lagrange equations (2.2) the chain rule can be used to to get
Xn  @2 L 
q
_ j + @2 L q j - @L = 0; i = 1; : : : ; n (2.5)
j i @q_ j @q_ i @qi
j=1 @q @q_
Definition 8 To write q j in equation (2.5) as a function of q and q_ it is required that the
matrix
@2 L
@q_ j @q_ i
be invertible. A Lagrangian for which this is invertible is called a regular Lagrangian.
If L is regular then one can do the change of variables (qi ; q_ i ) ! (qi ; pi ) at least locally due
to the implicit function theorem. The Legendre transform allows us to go back and forth
between Hamiltonian and Lagrangian mechanics due the following proposition :
Proposition 3 Let L(qi ; q_ i ) be a regular Lagrangian. De ne a new function
 n 
X
H(qi ; pi ) = pi q_ i - L(qi ; q_ i )
i=1
where
pi = @@Lq_ i ·
With this de nition the Euler-Lagrange equations are equivalent to the Hamilton's
equations, i.e
 i @H
d @L - @L = 0 , dq dt = @pi ;
i
dt @q_ @q i dp i @H
dt =- @qi

2.5 Algorithms
Definition 9 An algorithm on a phase space P is a collection of maps Fh : P ! P, usually
depending smoothly on h ∈ R for small h > 0. Advancing the solution zk by one time step
can be written as zk+1 = Fh (zk ). This means advancing time by h which is also called the
step size. Here all zk ∈ P.

Definition 10 An algorithm Fh is
11
i. a symplectic integrator if each Fh is symplectic,
ii. an energy integrator if H ◦ Fh = H,
iii. a momentum integrator if J◦ Fh = J, where J is the momentum map for the action
of a Lie group G.
Remark 1 It has been shown by Ge and Marsden [2] that a xed step size numerical
integrator can only be symplectic and momentum preserving or energy and momentum
preserving. In the remainder of this thesis I will refer to symplectic-momentum integrators
as simply symplectic integrators.
Remark 2 In this thesis I will not address the momentum conservation issue and I will not
give the de nitions for the terms used in the de nition of a momentum integrator above.
These can be found in Wendlandt and Marsden [20] and in the references contained therein.
12
13

Chapter 3

Variational Integrators

3.1 Discrete Mechanics


As described by Veselov [19] and Wendlandt and Marsden [20] one can develop a discrete
picture corresponding to the continuous mechanics described in Section 2.1. Instead of a
Lagrangian L : TQ ! R and the variational principle
Zb
S= L(q(t); q_ (t))dt
a
one de nes a \discrete" Lagrangian Ld : Q × Q ! R: The action integral above is replaced
by an action sum Sd : QN+1 ! R de ned as
NX
-1
Sd = Ld (qk ; qk+1 )
k=0
where qk ∈ Q. The discrete variational principle states that the evolution equations extrem-
ize the action sum given xed end points, q0 and qN . Extremizing Sd over q1 ; : : : ; qN-1
leads to the discrete Euler-Lagrange equations :
D1 Ld (qk ; qk+1 ) + D2 Ld (qk-1 ; qk ) = 0 for k = 1; : : : ; N - 1 (3.1)
where Di is the partial derivative with respect to the i-th argument. These equations can
also be written as
D1 Ld ◦  + D2 Ld = 0 (3.2)
where  : Q × Q ! Q × Q is de ned implicitly by (qk-1 ; qk ) = (qk ; qk+1 ). If D1 Ld is
invertible then equation (3.2) de nes the discrete map  which ows the system forward in
discrete time. Then one can show various properties like symplecticness, momentum map
preservation of  as done by Wendlandt and Marsden [20]. In order to obtain an algorithm
in terms of the original continuous Lagrangian one must relate L with Ld and in Wendlandt
and Marsden [20] this is done by de ning
 
y +x y-x
Ld (x; y) = L 2 ; h (3.3)
where h is the step size.
14
For examples of variational integrators constructed according to this scheme please con-
sult Section 4 of Wendlandt and Marsden [20]. See Chapter 5 for some additional examples.
These examples show the construction for variational integrators by the scheme described
above as well as integrators obtained by di erentiating the discrete Lagrangian according
the main ideas of this thesis.

3.2 Invariance Properties


To talk about the symplecticness of the algorithm  : Q × Q ! Q × Q one has to show
that some symplectic form is being preserved i.e that  is a symplectic map. But which
symplectic form ? The most natural choice is to pull back the canonical symplectic form
from the cotangent bundle T ∗Q. This is what was done by Veselov [19] and which is
explained by Wendlandt and Marsden [20]. To do this, we de ne the ber derivative
F Ld : Q × Q ! T ∗Q (3.4)
such that
F Ld :(qk ; qk+1 ) ! (qk ; D1Ld (qk ; qk+1 )):
Now one can de ne the symplectic 2-form on Q × Q by pulling back the canonical 2-form
from T ∗Q by FLd . Thus if
CAN is the canonical 2-form on T ∗Q then ! on Q × Q is
! = FL∗d (
CAN ) (3.5)
Please refer to Sections 3.1 and 3.2 of Wendlandt and Marsden [20] which give a proof
that the algorithm  preserves the symplectic form, i.e ∗ ! = !. There one can also nd
a local coordinate expression for !.
15

Chapter 4

Linearization

In this chapter I will describe our method for di erentiating, i.e linearizing Euler-Lagrange
equations and variational integrators by manipulating the Lagrangian and the discrete La-
grangian. This manipulation is such that in the continuous case, applying the appropriate
variational principle yields a continuous time derivative of the Euler-Lagrange equation
corresponding to the original Lagrangian. In the discrete case, the discrete Lagrangian is
manipulated and the discrete variational principle yields a \discrete time derivative" of the
discrete Euler-Lagrange equations corresponding to the original discrete Lagrangian. If the
original discrete Euler-Lagrange equation is linear, the discrete time derivative is simply a
nite di erence.
Of course, to linearize continuous (or discrete) Euler-Lagrange equations one could just
as well start with the equation and take its derivative (or nite di erence). But as men-
tioned in Section 1.2, my original aim which led to the present results, was to modify the
Lagrangian and then examine the resulting Euler-Lagrange equations. This is because the
discretization of the Euler-Lagrange equations is generated from the Lagrangian and not
from the continuous time equations.

4.1 Motivation
Consider the discrete Euler-Lagrange equation which gives the symplectic algorithm as
described in Chapter 3. This is repeated here for convenience :
D1 Ld ◦  + D2 Ld = 0 (4.1)
where  : Q × Q ! Q × Q is de ned implicitly by the above equation and (qk-1 ; qk ) =
(qk ; qk+1 ). Since Ld : Q × Q ! R, the resulting algorithm map  takes a pair of points
(qk-1 ; qk ) ∈ Q and produces a pair (qk ; qk+1 ). It is reasonable to conjecture that an
algorithm that is able to use more points, say 3 points, might result in a better truncation
error since 2 points allow one to approximate derivative and 3 will allow one to approximate
acceleration as well. This was the motivation for our subsequent ideas that led to the
linearization methods of this thesis. Thus, what is desired is an algorithm map (2) :
Q × Q × Q ! Q × Q × Q.
16

4.2 Differentiating Lagrangian along (q̇; q̈)


With a view to introducing accelerations into the framework, consider what happens when
the Lagrangian is a function of q; q_ and q , i.e the Lagrangian is some function L(2) (q; q;
_ q ).
Hamilton's principle (2.1) and simple variational calculus (see for example Chapter 2, page
42, equation 22 of Gelfand and Fomin [4]) ensures that the solutions q(t) satisfy the fourth
order Euler-Lagrange equations
d2 @L(2) - d @L(2) + @L(2) = 0: (4.2)
dt2 @q i dt @q_ i @qi
This is fourth order because in general the term
d2 @L(2) (q; q;_ q ) (4.3)
dt2 @q i
involves a fourth derivative term. Thus in going from L(q; q_ ) to L(2) (q; q;
_ q ) one goes from a
second order Euler-Lagrange equation to a fourth order Euler-Lagrange equation. This can
be represented by Figure 4.1. On the left something as yet unspeci ed is done to introduce

Hamilton's Principle
L(q; q_ ) −−−−−−−−−−−−−! EL(L) is second order
 
 
? ?
L(2) (q; q;_ q ) −−−−−−−−−−−−−!
Hamilton's Principle
EL(L(2) ) is fourth order

Figure 4.1: On the left q has been introduced into the Lagrangian but on the right the
equations go from being second to fourth order.
acceleration into the Lagrangian. On the right one sees that the Euler-Lagrange equations
skip an order and go from being second order to fourth order. The question marks indicate
an as yet unspeci ed operation. What is the most obvious operation one can perform on
L(q; q_ ) to introduce acceleration? The answer is di erentiation. Thus one might expect the
question mark on the right down arrow to perhaps represent di erentiation also.
However, as it turns out, when the second order Euler-Lagrange equations for L(q; q_ )
namely equation (2.2) are di erentiated one gets a third order equation as shown by the
following calculation (using implied summation convention where necessary). First, by
di erentiating the Euler-Lagrange equations for L(q; q_ ) we get :
d  d @L (q; q_ ) - @L (q; q_ ) = 0: (4.4)
dt dt @q_ i @qi
But the left hand side of the above equation is
d  @2 L q_ j + @2 L q j  - @2 L q_ j - @2 L q j ; (4.5)
dt @qj @q_ i @q_ j @q_ i @qj @qi @q_ j @qi
17
which is
@2 L qj + q_ j  @3 L q_ k + @3 L qk  +
=
@qj @q_ i @qk @qj @q_ i @q_ k @qj @q_ i

@2 L ...q j + q j @3 L q_ k + @3 L qk  -
@q_ j @q_ i @qk @q_ j @q_ i @q_ k @q_ j @q_ i
@2 L q_ j - @2 L q j
@qj @qi @q_ j @qi

@2 L qj + @3 L q_ k q_ j + @3 L qk q_ j +
=
@qj @q_ i @qk @qj @q_ i @q_ k @qj @q_ i
@ L ...q j + @3 L q_ k q j + @3 L q k q j -
2
@q_ j @q_ i @qk @q_ j @q_ i @q_ k @q_ j @q_ i
@2 L q_ j - @2 L q j
@qj @qi @q_ j @qi
Therefore the di erentiation of the Euler-Lagrange equation for L(q; q_ ) gives the following
equation :
@2 L qj + @3 L q_ k q_ j + @3 L qk q_ j +
@qj @q_ i @qk @qj @q_ i @q_ k @qj @q_ i
@2 L ...q j + @3 L q_ k q j + @3 L q k q j -
@q_ j @q_ i @qk @q_ j @q_ i @q_ k @q_ j @q_ i
@2 L q_ j - @2 L q j = 0: (4.6)
@qj @qi @q_ j @qi
This equation (4.6) is third order. Thus the diagram in Fig. 4.1 will not commute if down
arrows are thought of as di erentiation. On the right one needs to go to a third order
Euler-Lagrange equation.
However if L(2) (q; q;
_ q ) was only linear in q then the term
@L(2)
@q i
in equation (4.2) will be a function only of q and q_ as a result of which
d2 @L(2)
dt2 @q i
will only be a third order term. The Euler-Lagrange for such a Lagrangian will be third
order. As the left down arrow was being thought of as di erentiation this might work out.
This suggests getting L(2) (q; q;_ q ) from L(q; q_ ) by simply formally di erentiating L(q; q_ )
with respect to time. If L(q; q_ ) is simply formally di erentiated wrt t then one gets
@L q_ i + @L q i
L(2) (q; q;_ q) = @q (4.7)
i @q_ i
Note that now L(2) (q; q;
_ q ) is linear in q , as desired. Now the next step ought to be to try
18

Hamilton's Principle
L(q; q_ ) −−−−−−−−−−−−−! EL(L) is second order
 
di erentiate


di erentiate
L(2) (q; q;_ q )
−−−−−−−−−−−−−! EL(L(2) ) is third order
linear inq Hamilton's Principle

Figure 4.2: On the left q has been introduced into the Lagrangian by di erentiating L.
Since L(2) is linear in q , EL(L(2) ) should be third order. But there is a major problem as
is explained in the text.

and see if the diagram in Figure 4.2 commutes. But that will not be worthwhile because
of the two Lemmas that follow. Lemma 1 below shows that for Lagrangians L of a certain
type, when L(2) is obtained as in equation (4.7) and Hamilton's principle is applied using
Lagrangian L(2) one gets the trivial identity 0 = 0 for the Euler-Lagrange equations. The
next Lemma 2 removes the restriction on the form of Lagrangian. 1
Lemma 1 Let the L be of the form
L(q; q_ ) = LT (q_ ) + LV (q) (4.8)
(The superscripts T and V are reminders that the term LT (q_ ) is the kinetic energy
and the term LV (q) is minus the potential energy term). Construct a new Lagrangian
L(2) (q; q;_ q ) as
@L q_ i + @L q i
L(2) (q; q;_ q) = @q (4.9)
i @q_ i
For such a Lagrangian L(2) the Hamilton's principle gives no information about the
evolution of state, i.e the Euler-Lagrange equation for L(2) is 0 = 0.
Proof When the Lagrangian of the form (4.8) is di erentiated to get L(2) de ned by
equation (4.9)
T i @LV i
L(2) (q; q;_ q ) = @L
@q_ i q + @qi q_ (4.10)
is obtained since the other terms are 0: De ne
L(T )(q;_ q ) ≡
@LT q j (4.11)
@q_ j
L(V )(q; q_ ) ≡
@LV q_ j
@qj
1 The more general result in Lemma 2 was suggested by Matthew West when I showed him Lemma 1.
19
Now the Euler-Lagrange equation for L(2) (q; q;
_ q ), namely equation (4.2) is

@L(T ) - d @L(T ) + d2 @L(T )  +
@qi dt @q_ i dt2 @q i (4.12)
 (V )
@L - d @L(V ) + d2 @L(V )  = 0
@qi dt @q_ i dt2 @q i
Substitute for L(T ) and L(V ) from equation (4.11) into equation (4.12) and consider the L(T )
and L(V ) parts separately. The L(T ) part of equation (4.12) becomes
d @L(T ) + d2 @L(T )
0 - dt @q_ i  dt2 @q i
d @ @LT q j + d2 @  @LT q j 
=-
dt @q_ i @q_ j dt2 @q i @q_ j
 2 T 
d @ L q j + d2 @LT
=-
dt @q_ i @q_ j dt2 @q_ i
 2 T   2 T 
d @ L j d @ L j
=-
dt @q_ i @q_ j q + dt @q_ j @q_ i q
=0

where the last step follows from the equality of mixed partials. Similarly the L(V ) part of
equation (4.12) becomes
@L(V ) - d @L(V ) + 0
@qi dt @q_ i
@ @LV q_ j - d @  @LV q_ j 
=
@qi @qj dt @q_ i @qj
@ L q_ j - d @LV
2 V
=
@qi @qj dt @qi
@ L q_ j - @LV q_ j
2 V
=
@qi @qj @qj @qi
= 0
where again the last step follows from the equality of mixed partials. Thus equation (4.12)
becomes the trivial identity 0 = 0: 

Remark 3 The type of Lagrangian L assumed in Lemma 1 arises commonly. Some common
examples would be : a particle in, say, a cubic potential eld where
2
L(q; q_ ) = m q_2 + q3
or a simple pendulum of length l for which
2 _2
L(; _ ) = l 2 + gl cos :
20
The restriction on the form of Lagrangian can be removed as is shown by the following
more general result.
Lemma 2 Let L be any suciently smooth function of q and q_ . Construct a new
Lagrangian L(2) (q; q;_ q ) as
@L q_ i + @L q i
L(2) (q; q;_ q) = @q (4.13)
i @q_ i
For this Lagrangian L(2) the Hamilton's principle gives no information about the evo-
lution of state, i.e the Euler-Lagrange equation for L(2) is 0 = 0.
Proof The Euler-Lagrange equation corresponding to the Lagrangian L(2) is
d2 @L(2) - d @L(2) + @L(2) = 0 (4.14)
dt2 @q i dt @q_ i @qi
for i = 1; : : : ; n. The rst term of this equation is (summation over repeated indices)
d2 @L(2) = d2 @L
dt2 @q i dt2@q_ i
d @ 2L j @2L j
=
dt @qj @q_ i q_ + @q_ j @q_ i q
@2 L q j +  d @2 L  q_ j + @2 L ...q j +  d @2 L  q j (4.15)
=
@qj @q_ i dt @qj @q_ i @q_ j @q_ i dt @q_ j @q_ i
The second term of equation (4.14) is
d @L(2) = d  @2 L q_ j + @L + @2 L q j 
dt @q_ i dt @q_ i @qj @qi @q_ i @q_ j
@2 L qj + d @2 L  q_ j + @2 L q_ j + @2 L q j

=
@q_ i @qj dt @q_ i @qj @qj @qi @q_ j @qi
 
@2 L ...q j + d @2 L qj (4.16)
+
@q_ i @q_ j dt @q_ i @q_ j
and the last term of equation (4.14) is
@L(2) = @2 L q_ j + @2 L q j (4.17)
@qi @qi @qj @qi @q_ j
Substituting the expressions (4.15), (4.16) and (4.17) in equation (4.14) and using equality
of mixed partials one nds that the left hand side of equation (4.14) is 0 thus completing
the proof. 
Thus di erentiating the Lagrangian along (q; q_ ) apparently does not make the diagram
4.2 commute. Section 4.3 below shows how this can be remedied by a di erent approach
to make it work. This results in continuous time linearization of the Euler-Lagrange equa-
tions. Section 4.4 shows a similar discrete time linearization of the discrete Euler-Lagrange
equations.
21
4.3 Continuous Time Linearization
In Section 4.2 the Lagrangian L was di erentiated along the acceleration vector (q; _ q) to
get a new Lagrangian L(2) which was linear in q . I showed that this new Lagrangian L(2)
yields an Euler-Lagrange equation which is a trivial identity. Now I will generalize this
idea of di erentiating L to get a new Lagrangian but this time the new Lagrangian does
yield a nontrivial Euler-Lagrange equation. Then I will describe and prove the relationship
between di erentiating the Lagrangian and the corresponding Euler-Lagrange equations.
Given L : TQ ! R, and q ∈ Q the di erential of L at (q; q_ ) ∈ TqQ is the map dL(q; q_ ) :
T(q;q_ ) TQ ! T R = R: See Section 2.2, De nition 2 for details. In Section 4.2, I wrote
@L (q; q_ ) · q_ i + @L (q; q_ ) · q i
L(2) (q; q;_ q) = dL(q; q_ ) · (q;_ q ) = @q (4.18)
i @q_ i
and found that the Euler-Lagrange equation for L(2) is 0 = 0. In equation (4.18) the
di erentiation is along the acceleration vector (q;_ q ) which is in TTQ at (q; q_ ). (Actually
to be correct one should really write such an element of TTQ as (q; q; _ q ) or even better, as
(q; q;
_ q;
_ q )). The obvious generalization of this di erentiation is to di erentiate along an
arbitrary vector in TTQ. With this view let vq ∈ TqQ and let wvq ∈ TvqTQ: De ne L(2) by
L(2) (wvq ) ≡ dL(vq ) · wvq (4.19)
In coordinates one writes vq = (q; q_ ) and wvq = (q; q; _ u; u_ ) where u ∈ TqQ and u is not
necessarily equal to vq = (q; q_ ). Thus in local coordinates equation (4.19) is
@L (q; q_ ) · ui + @L (q; q_ ) · u_ i
L(2) (q; q;_ u; u_ ) ≡ @q (4.20)
i @q_ i
We now introduce the notation EL(L) to represent the Euler-Lagrange equations corre-
sponding to a Lagrangian L. The main result of this section is that di erentiating EL(L)
results in the same equations that are obtained by rst di erentiating L to get L(2) , nding
EL(L(2) ) and setting u = q_ . This is summarized in the following Lemma.
Lemma 3 The following diagram commutes
EL(L) −−−−! EL(L(2) )ju=q_
 



 (4.21)
L −−−−! L(2)
Proof First I will describe how to read the commutative diagram above. The up arrows
mean apply Hamilton's principle and the result is the Euler-Lagrange equations correspond-
ing to the Lagrangian one starts with at the bottom. The right arrows indicate di erentia-
tion. The top right arrow is di erentiation wrt t. The bottom right arrow is di erentiating
as in (4.20). The notation ju=q_ in the top right corner only makes sense if one arrives at
that corner from the bottom.
To be speci c here is how the various paths through the diagram should be read. Start
at bottom left corner. Going up and then right means starting with the Lagrangian L nd
the Euler-Lagrange equation corresponding to it and then di erentiate that equation wrt t.
22
On the other hand the path going right and then up means, di erentiate L as in (4.20)
which gives an expression for L(2) in terms of L. Then nd the Euler-Lagrange equation
for this Lagrangian. Finally substitute u = q_ . The resulting equations will be the same as
those obtained by the up and right path.
I will now give a local coordinate proof of the commutativity of (4.21). I have already dif-
ferentiated EL(L) and the result is given in equation (4.6). Thus the result of di erentiating
EL(L) can be considered to be the EL(L) itself and the linearized EL(L) in equation (4.6).
Let us now see what one gets by writing EL(L(2) ) in terms of the original Lagrangian L .
Since L(2) (q; q;
_ u; u_ ) has 2 independent variables q and u it follows that EL(L(2) ) consists
of the two equations
d @L(2) - @L(2) = 0 (4.22)
dt @q_ i @qi
d @L(2) - @L(2) = 0 (4.23)
dt @u_ i @ui
The left hand side of equation (4.22) when written in terms of L is equivalent to
d @  @L uj + @L u_ j  - @  @L uj + @L u_ j  (4.24)
dt @q_ i @qj @q_ j @qi @qj @q_ j
d @ L uj + @ L u_ j - @ L uj + @2 L u_ j 
 2 2   2
(4.25)
=
dt @q_ i @qj @q_ i @q_ j @qi @qj @qi @q_ j
@ L u_ j + uj d @ L + @ L u j + u_ j d @2 L -
2 2 2
=
@q_ i @qj dt @q_ i @qj @q_ i @q_ j dt @q_ i @q_ j
2 2
@ L uj - @ L u_ j :
@qi @qj @qi @q_ j (4.26)
Now
d @2 L = @3 L q_ k + @3 L q k (4.27)
dt @q_ i @qj @qk @q_ i @qj @q_ k @q_ i @qj
d @2 L = @3 L q_ k + @3 L q k (4.28)
dt @q_ i @q_ j @qk @q_ i @q_ j @q_ k @q_ i @q_ j
Thus one equation of EL(L(2) ) obtained by substituting (4.27) and (4.28) in (4.26) is
@2 L u_ j + @3 L uj q_ k + @3 L uj qk + (4.29)
@q_ i @qj @qk @q_ i @qj @q_ k @q_ i @qj
@2 L u j + @3 L u_ j q_ k + @3 L u_ j q k -
@q_ i @q_ j @qk @q_ i @q_ j @q_ k @q_ i @q_ j
@2 L uj - @2 L u_ j = 0
@qi @qj @qi @q_ j
23
When one sets u = q_ one equation of EL(L(2) )ju=q_ in terms of L is obtained as
@2 L q j + @3 L q_ j q_ k + @3 L q_ j qk + (4.30)
@q_ i @qj @qk @q_ i @qj @q_ k @q_ i @qj
@2 L ...q j + @3 L qj q_ k + @3 L q j qk -
@q_ i @q_ j @qk @q_ i @q_ j @q_ k @q_ i @q_ j
@2 L q_ j - @2 L qj = 0
@qi @qj @qi @q_ j
which except for a rearrangement of terms is the same as the linearized EL(L) in equation
(4.6). The other equation of EL(L(2) ) is equation (4.23) which in terms of L is equivalent
to
d @  @L uj + @L u_ j  - @  @L uj + @L u_ j  = 0; (4.31)
dt @u_ i @qj @q_ j @ui @qj @q_ j
which is equivalent to
d @L @L
dt @q_ i - @qi = 0; (4.32)
which is the same as the original EL(L). This completes the proof of commutativity of
(4.21). 
Thus I have shown that the Euler-Lagrange equation corresponding to L, namely EL(L)
can be di erentiated by di erentiating L along (u; u_ ) rst to get a new Lagrangian L(2) and
then deriving the Euler-Lagrange equations EL(L(2) ) corresponding to this Lagrangian and
nally setting u = q_ .

4.4 Discrete Time Linearization


With a view of linearizing the discrete Euler-Lagrange equations I will mimic on the discrete
side what was done on the continuous side. Thus I rst obtain L(d2) from Ld by di erenti-
ating. Let Ld : Q × Q ! R be the discrete Lagrangian. By analogy with the continuous
case one rst di erentiates this to get dLd (q; r) : TQ × TQ ! R where (q; r) ∈ Q × Q and
(vq ; vr ) ∈ T(q;r)(Q × Q) =
 TqQ × TrQ. Then one de nes

L(d2) (vq ; vr ) ≡ dLd (q; r) · (vq ; vr ) (4.33)


In local coordinates
L(d2) (qk ; uk ; qk+1 ; uk+1 ) = dLd (qk ; qk+1 ) · (uk ; uk+1 ) (4.34)
The use of subscripts k and k + 1 is suggestive of the fact that qk+1 is the position
after one discrete time step, starting from qk . However, the above de nition of L(d2) is valid
for any q1 ,u1 ,q2 ,u2 in the appropriate spaces. Note also that here qk etc. are vectors, i.e
qk = (q1k ; : : : ; qnk ) etc. I will assume, as done by Wendlandt and Marsden [20] that Q ⊂ V
where V is a vector space. This is essential if one is going to be averaging and subtracting
points in Q.
24
Now I will prove a result on the discrete side analogous to what was done in Lemma
3 on the continuous side. Let DEL(Ld ) represent the Discrete Euler-Lagrange equations
corresponding to the discrete Lagrangian Ld as in (3.1). Then the following is true
Lemma 4 The following diagram commutes :
Ld −−−−! L(d2)
 
 
  (4.35)

( 2) 
DEL(Ld ) −−−−! DEL(Ld ) uk= qk+1 -qk h

Proof First I must describe how to read the diagram. The top right arrow represents
di erentiation de ned in (4.34). The down arrows represent using the discrete variational
principle to get the discrete Euler-Lagrange equations. The bottom right arrow has to be
given a meaning. I will de ne an operation of discrete di erentiation as described below.
In the case of linear equations this is the same as nite di erencing.
First let's check carefully what DEL(L(d2) ) is. This is obtained by extremizing the discrete
action sum, i.e by extremizing
NX
-1
S(d2) = L(d2) (qk ; uk ; qk+1 ; uk+1 ) (4.36)
k=0
One extremizes S(d2) over qk and uk for k = 1; : : : ; N - 1 keeping q0 ,qN ,u0 and uN xed
and by setting
@S(d2) = 0; @S(d2) = 0
@qjk @ujk
for j = 1; : : : ; n. This yields the two sets of equations that comprise the discrete Euler-
Lagrange equations for L(d2) . These equations, which I will call DEL(L(d2) ) are
@L(d2) (q ; u ; q ; u ) + @L(d2) (q ; u ; q ; u ) = 0 (4.37)
@qjk k-1 k-1 k k @qjk k k k+1 k+1
@L(d2) (q ; u ; q ; u ) + @L(d2) (q ; u ; q ; u ) = 0 (4.38)
@ujk k-1 k-1 k k @ujk k k k+1 k+1
for k = 1; : : : ; N - 1 and j = 1; : : : ; n.
Now I can start the proof of the commutativity of (4.35). To do this one must write
DEL(L(d2) ) in terms of Ld . Consider the rst set of equations represented by (4.37). Substi-
tute for L(d2) in terms of Ld using (4.34). With this substitution, equation (4.37) becomes
@ ui @Ld (q q ) + ui @Ld (q ; q ) +
@qjk k-1 @qik-1 k-1; k  k @qik k-1 k
@ ui @Ld (q q ) + ui @Ld (q ; q ) = 0; (4.39)
@qjk k @qik k; k+1 k+1 @qi
k+1
k k+1
25
which implies
2 2
uik-1 @j Ldi (qk-1 ; qk ) + uik @j Ld i (qk-1 ; qk ) +
@qk @qk-1 @qk @qk
2 2
uik @j Ld i (qk ; qk+1 ) + uik+1 @j Ldi (qk ; qk+1 ) = 0: (4.40)
@qk @qk @qk @qk+1
Now let's substitute
i i
uik = qk+1h- qk
etc. to get
qik - qik-1 @2 Ld (q ; q ) + qik+1 - qik @2 Ld (q ; q ) +
h @qjk @qik-1 k-1 k h @qjk @qik k-1 k
qik+1 - qik @2 Ld (q ; q ) + qik+2 - qik+1 @2 Ld (q ; q ) = 0 (4.41)
h @qjk @qik k k+1 h @qjk @qik+1 k k+1
The above equation represents one equation of

DEL(L(d2) ) uk= qk+1h-qk :
In terms of Ld the other equation, i.e (4.38) becomes
@ ui @Ld (q q ) + ui @Ld (q ; q ) +
@ujk k-1 @qik-1 k-1; k  k @qik k-1 k
@ ui @Ld (q q ) + ui @Ld (q ; q ) = 0 (4.42)
@ujk k @qik k; k+1 k+1 @qik+1 k k+1
which is
@Ld (q ; q ) + @Ld (q q ) = 0 (4.43)
@qjk k-1 k @qjk k; k+1
Now DEL(Ld ) is
@Ld (q ; q ) + @Ld (q ; q ) = 0 (4.44)
@qjk k-1 k @qjk k k+1
which is the same as (4.43). The discrete time derivative of this is same as (4.41). This
derivative is taken just like the usual derivative except that instead of q_ k-1 one writes
qk - qk-1
h
If an expression is linear in its variables, its discrete time derivative is the same as its
nite di erence. Consider for example an expression f(q) linear in q. In particular let
f(q) = a1 q + a0 . Then by the above description of discrete time derivative, the discrete
26
time derivative of f(q) is

df  q1 - q0
dq q0 h
which is
a1 q1 -h q0 :
Note that this is the same as
f(q1 ) - f(q0 ) ;
h
the nite di erence of the expression f(q).
Thus we have shown that

DEL(L(d2) )  qk+1h-qk
consists of two sets of equations. The rst set, equation (4.37), after substituting
uk = qk -hqk-1
is the discrete time derivative of DEL(Ld ). The second set, equation (4.38) is the same as
DEL(Ld ) which shows that the diagram in equation (4.35) commutes. 
Thus the situation is exactly the same as the continuous case in which
EL(L(2) ) ju=q_
consisted of two sets of equations the rst of which was the derivative of EL(L) and the
second of which was the same as EL(L).

Linearized Algorithm The linearized algorithm is generated by extremizing the action


sum corresponding to L(d2) : This yields two discrete Euler-Lagrange Equations, the linearized
and the original, just like for the continuous Lagrangian L(2) : I call the two discrete Euler-
Lagrange equations collectively as DEL(L(d2) ): It was seen that the DEL(L(d2) ) are given by
(4.37) and (4.38). After making the substitutions uik = (qik+1 - qik )=h in these equations,
one gets what I called

DEL(L(d2) )  qk+1h-qk
namely equation (4.41). Notice that (4.41) can be solved for qk+2 assuming invertibility of
the matrix
@2 Ld (q ; q ) (4.45)
@qjk @qik+1 k k+1
27
appearing in the equation. For a regular Lagrangian, by de nition, the matrix
@2 L (4.46)
@q_ j @q_ i
is invertible. It is known [16] that at least for Lagrangians of the form \Kinetic Energy -
Potential Energy", the invertibility of (4.46) implies the invertibility of (4.45) for small step
size. Presumably the same might be true for a general Lagrangian.
Thus given a triple of points (qk-1 ; qk ; qk+1 ) one can solve (4.41) for qk+2 . Thus an
algorithm that takes the triple
(qk-1 ; qk ; qk+1 )
to the triple
(qk ; qk+1 ; qk+2 )
has been de ned. One can also write the algorithm in terms of the original continuous
Lagrangian but I will not do that.

4.5 Discrete and Continuous


Now the continuous and discrete Lagrangians must be related in some way and this is
summarized in the following result.
Lemma 5 The following diagram commutes
L −−−−! L(2)
 
 
  (4.47)
Ld −−−−! L(d2)
Proof The vertical arrows represent implementation decisions, i.e it is up to us to choose
a discretization that will have the right properties. Starting from the top left corner of
(4.47) and going down, i.e L ! Ld should be read as writing Ld in terms of L: This is
provided by equation (3.3) written here as
 
Ld (qk ; qk+1 ) = L qk+12+ qk ; qk+1h- qk (4.48)
where, in L(q; q_ ) the position q has been approximated by (qk+1 + qk )=2 and q_ by (qk+1 -
qk )=h . A similar implementation decision must be made in going from L(2) (q; q;_ u; u_ ) to
L(d2) (qk ; uk ; qk+1 ; uk+1 ) i.e on going down the right arrow. This is done by setting
q = (qk+1 + qk )=2
q_ = (qk+1 - qk )=h
u = (uk+1 + uk )=2
u_ = (uk+1 - uk )=h
28
in L(2) . Now
L(d2) (qk ; uk ; qk+1 ; uk+1 ) =
 
L(2) qk+12+ qk ; qk+1h- qk ; uk+12+ uk ; uk+1h- uk (4.49)

i i @L  qk+1 + qk qk+1 - qk 
L(d2) (qk ; uk ; qk+1 ; uk+1 ) = uk+12+ uk @q i 2 ; h
uik+1 - uik @L  qk+1 + qk qk+1 - qk 
+
h @q_ i 2 ; h (4.50)
where the last equality follows from (4.20). Equation (4.50) represents starting from top
left, going right and then bottom. This must be compared with the other path of going
down from top left and then going right. To do this one must use equation (4.34). Here,
dLd (qk ; qk+1 ) =

1 @L - 1 @L ; : : : ; 1 @L - 1 @L ; 1 @L + 1 @L ; : : : ; 1 @L + 1 @L
(4.51)
2 @q1 h @q_ 1 2 @qn h @q_ n 2 @q1 h @q_ 1 2 @qn h @q_ n
Thus
dLd (qk ; qk+1 ) · (uk ; uk+1 ) =
   
uik 12 @q
@L
i -
1 @L i 1 @L 1 @L
h @q_ i + uk+1 2 @qi + h @q_ i (4.52)
where i = 1; : : : ; n which implies that
i i @L uik+1 - uik @L
L(d2) (qk ; uk ; qk+1 ; uk+1 ) = uk+12+ uk @q i+ h @q_ i (4.53)
where the partials are evaluated at ((qk+1 + qk )=2; (qk+1 - qk )=h). But (4.53) is the same
as (4.50) which completes the proof of commutativity of (4.47). 
The previous three lemmas can be summarized in one diagram as given by the following
theorem.
29
Theorem 1 The following diagram commutes

EL(L) −−−−! EL(L(2) )u=q_
 
 
 
L −−−−! L(2)
 
 
  (4.54)
Ld −−−−! Ld2
( )

 
 
 

DEL(Ld ) −−−−! DEL(L(d2) )
uk= qk+1h-qk
Proof This follows from the previous three lemmas by \pasting" the corresponding com-
mutative diagrams together. 

4.6 Linearization Summary


I summarize the linearization process by giving the various Euler-Lagrange equations in one
place here. In the following equations of this section, i; j = 1; : : : ; n and k = 1; : : : ; N - 1,
where n is the dimension of the con guration manifold and 0; : : : ; N represents discrete
time.
Euler Lagrange equations EL(L) for L(q; q_ ) :
d @L @L
dt @q_ i - @qi = 0 (4.55)

Euler Lagrange equations EL(L(2) ) for L(2) (q; q;


_ u; u_ ) :
d @L(2) - @L(2) = 0 (4.56)
dt @q_ i @qi
d @L(2) - @L(2) = 0 (4.57)
dt @u_ i @ui
30
EL(L(2) )ju=q_ in terms of L:
@2 L q j + @3 L q_ j q_ k + @3 L q_ j qk + (4.58)
@q_ i @qj @qk @q_ i @qj @q_ k @q_ i @qj
@2 L ...q j + @3 L qj q_ k + @3 L q j qk -
@q_ i @q_ j @qk @q_ i @q_ j @q_ k @q_ i @q_ j
@2 L q_ j - @2 L qj = 0
@qi @qj @qi @q_ j
d @L @L
dt @q_ i - @qi = 0 (4.59)
Equation (4.58) is the derivative of equation (4.55).
Discrete Euler-Lagrange equations DEL(Ld ) for Ld (qk ; qk+1 ) :
@Ld (q ; q ) + @Ld (q ; q ) = 0 (4.60)
@qjk k-1 k @qjk k k+1

Discrete Euler-Lagrange equations DEL(L(d2) ) for L(d2) (qk ; uk ; qk+1 ; uk+1 )


@L(d2) (q ; u ; q ; u ) + @L(d2) (q ; u ; q ; u ) = 0 (4.61)
@qjk k-1 k-1 k k @qjk k k k+1 k+1
@L(d2) (q ; u ; q ; u ) + @L(d2) (q ; u ; q ; u ) = 0 (4.62)
@ujk k-1 k-1 k k @ujk k k k+1 k+1

DEL(L(d2) )  qk+1h-qk in terms of Ld :
qik - qik-1 @2 Ld (q ; q ) + qik+1 - qik @2 Ld (q ; q )+ (4.63)
h @qjk @qik-1 k-1 k h @qjk @qik k-1 k
qik+1 - qik @2 Ld (q ; q ) + qik+2 - qik+1 @2 Ld (q ; q ) = 0
h @qjk @qik k k+1 h @qjk @qik+1 k k+1
@Ld (q ; q ) + @Ld (q ; q ) = 0 (4.64)
@qjk k-1 k @qjk k k+1
Equation (4.63) is the discrete time derivative of equation (4.60).
31
4.7 Invariance Properties
To see why the proofs of invariance properties given by Wendlandt and Marsden [20] will
still work, it is helpful to consider the following picture.
L : TQ ! R L(2) : TTQ ! R
Ld : Q × Q ! R L(2) : TQ × TQ ! R

The left column shows the usual continuous Lagrangian and the discrete Lagrangians used
by Wendlandt and Marsden [20]. It was shown by Wendlandt and Marsden [20] that the
algorithm obtained from Ld is a symplectic-momentum algorithm. On the right column
are the Lagrangian and discrete Lagrangian I have obtained. Note that the right column is
the same as the left one with Q replaced by TQ. Thus all the invariance proofs of the left
column still apply.
This means that EL(L(2) ) has symplectic ow and DEL(L(d2) ) in equations (4.37) and
(4.38) is a symplectic integrator. I derive the the symplectic form on TQ × TQ in subsection
4.7.1 below. One can check explicitly that this symplectic form, given as equation (4.68)
below, is preserved by the algorithm DEL(L(d2) ) though this also follows without checking
from the proof in Wendlandt and Marsden [20] as indicated above.
But we must check that similar invariance properties remain true after we set u = q_ in
EL(L(2) ) or uk = (qk+1 - qk )=h in DEL(L(d2) ). This will be done in subsection 4.7.2 for
DEL(L(d2) ).

4.7.1 Symplectic form on TQ × TQ


I will de ne the symplectic form on TQ × TQ by pulling back the canonical 2-form from
T ∗TQ. The pull back will be done under a ber derivative of L(d2) where, as de ned earlier,
L(d2) : TQ × TQ ! R. This ber derivative will be de ned by analogy to the ber derivative
for Ld : Q × Q ! R, de ned by Wendlandt and Marsden in [20]. That ber derivative can
be de ned by taking a parameterized curve (q(t); q1 ) where the second component is xed
and where q(0) = q0 and q_ (0) = u0 . This is a curve in Q × fq1 g ⊂ Q × Q and can be
identi ed with the curve q(t). Thus it represents the vector (q0 ; u0 ) ∈ Tq0Q . The ber
derivative FLd : Q × Q ! T ∗Q is then de ned as

FLd (q0 ; q1 ) · (q0 ; u0 ) =
d 
dt t=0 Ld (q(t); q1 ) (4.65)

I will de ne the ber derivative FL(d2) : TQ × TQ ! T ∗TQ by analogy to this. But rst I
note that the following steps are needed to de ne the symplectic form on TQ × TQ :
i. De ne the ber derivative FL(d2) : TQ × TQ ! T ∗TQ,
ii. Compute the tangent map of FL(d2) , namely
T((q0;u0 );(q1 ;u1)) FL(d2) :
T(q0;u0 ) TQ × T(q1;u1 ) TQ ! TFL(d2)((q0 ;u0 );(q1;u1 )) T ∗TQ
32
iii. Write the local coordinate expression for the canonical one form  on T ∗TQ,
iv. De ne the symplectic form on TQ × TQ as
= -d(FLd(2)∗ ) where FL(d2)∗ means the
pullback under the ber derivative FL(d2) .
Step 1 : Since L(d2) maps TQ × TQ ! R, the ber derivative FL(d2) maps TQ × TQ ! T ∗TQ.
Note that FL(d2) evaluated at a point in TQ × TQ is a point in T ∗TQ so FL(d2) can be de ned
by de ning how it acts on a vector in TTQ. By analogy to the de nition of ber derivative
F Ld given above, let ((q(t); u(t)); (q1 ; u1 )) be a parameterized curve from an open interval
of R including 0 into TQ × f(q1 ; u1 )g ⊂ TQ × TQ such that
q(0) = q0
u(0) = u0
q_ (0) = e
u_ (0) = f
The curve ((q(t); u(t)); (q1 ; u1 )) can be identi ed with the curve (q(t); u(t)) and this curve
then represents the vector ((q0 ; u0 ); e; f) ∈ T(q0;u0 ) TQ. The ber derivative FL(d2) evalu-
ated at ((q0 ; u0 ); (q1 ; u1 )) ∈ TQ × TQ can be de ned by how its acts on ((q0 ; u0 ); e; f) ∈
T(q0;u0 ) TQ as follows:

d  L(2) ((q(t); u(t)); (q ; u ))
FLd2 ((q0; u0 ); (q1 ; u1 )) · ((q0 ; u0 ); e; f) = dt
( )
 d 1 1
t=0
Note the analogy with the de nition in equation(4.65). When the above calculation is
carried out, in matrix notation it results in the following de nition
2 ((q ; u ); (q ; u )) =
F Ld
( )
0 0 1 1
(2) (2)
((q0 ; u0 ); D1 Ld ((q0 ; u0 ); (q1 ; u1 )); D2 Ld ((q0 ; u0 ); (q1 ; u1 ))) (4.66)

Note that FL(d2) maps the point ((q0 ; u0 ); (q1 ; u1 )) ∈ TQ×TQ to a covector in T(∗q0;u0) TQ
based at (q0 ; u0 ) so the rst (q0 ; u0 ) that appears on the right hand side of equation(4.66)
is the base point of the covector.
Step 2 : To compute the expression for the tangent map of FL(d2) consider a smooth
parameterized curve ((q0 (t); u0 (t)); (q1 (t); u1 (t))) in TQ × TQ such that
q0 (0) = q0 q_ 0 (0) = e0
u0 (0) = u0 u_ 0 (0) = f0
q1 (0) = q1 q_ 1 (0) = e1
u1 (0) = u1 u_ 1 (0) = f1
Then the tangent map at ((q0 ; u0 ); (q1 ; u1 )) acting on the vector
(((q0 ; u0 ); e0 ; f0 ); ((q1 ; u1 ); e1 ; f1 )) ∈ T(q0;u0 ) TQ × T(q1;u1) TQ
33
is de ned by
T((q0;u0);(q1 ;u1 )) FL(d2) · (((q0 ; u0 ); e0 ; f0 ); ((q1 ; u1 ); e1 ; f1 )) =

d  FL(2) ((q (t); u (t)); (q (t); u (t)))
dt t=0 d 0 0 1 1

In matrix notation this works out to be


((q0 ; u0 ); D1 L(d2) ((q0 ; u0 ); (q1 ; u1 )); D2 L(d2) ((q0 ; u0 ); (q1 ; u1 ));
e0 ; f 0 ;
(2) (2) (2) (2)
D1 D1 Ld · e0 + D2 D1 Ld · f0 + D3 D1 Ld · e1 + D4 D1 Ld · f1 ;
(2) (2) (2) (2)
D1 D2 Ld · e0 + D2 D2 Ld · f0 + D3 D2 Ld · e1 + D4 D1 Ld · f1 )

with the base point ((q0 ; u0 ); D1 L(d2) ; D2 L(d2) ). In the expression above, the notation D1 ,
D2 , D3 , D4 refers to the partial derivative with respect to the rst, second, third and
fourth arguments respectively, namely partial derivatives with respect to q0 , u0 , q1 , u1
respectively. The entire expression above is not really needed for the actual derivation of
the symplectic form
as will be seen in steps 3 and 4 below. It is given above just for
completeness.
Step 3 : In invariant notation, the canonical one form on T ∗TQ is
 (v) =
; T TQ · v
where ∈ T ∗TQ, v ∈ T (T ∗TQ) and TQ : T ∗TQ ! TQ is the canonical projection. To
write this in local coordinates, let E be the Euclidean space that models Q locally and
let U ⊂ E be open. The base point of in local coordinates is some (q; u) ∈ U × E
and so = ((q; u); ( 1 ; 2 )) for some ( 1 ; 2 ) ∈ E∗ × E∗ . The vector v ∈ T (T ∗TQ) has
the local coordinate expression v = ((q; u); ( 1 ; 2 ); e; f; 1 ; 2 ) for some (e; f; 1 ; 2 ) ∈
E × E × E∗ × E∗ . Thus in local coordinates
TQ(q; u; 1 ; 2 ) = (q; u)
and so the tangent map acting on vector v, namely T TQ · v is given in local coordinates
by
T TQ · v = ((q; u); e; f) (4.67)
This gives us the expression for  (v) in local coordinates as follows
 (v) = ((q;u);( 1; 2 )) ((q; u); ( 1 ; 2 ); e; f; 1 ; 2 )
=
( 1 ; 2 ); (e; f)
= 1 (e) + 2 (f)
34
Step 4 : Now I can de ne

= -d(FL(d2)∗ )
and derive a local coordinate expression for this. I will rst derive the local coordinate
expression for the pullback part of the expression above, namely for FL(d2)∗  as follows.
L(d2)∗ )((q0 ;u0);(q1 ;u1 ))(((q0 ; u0 ); e0 ; f0 ); ((q1 ; u1 ); e1 ; f1 ))
(F

= FL(d2) ((q0 ;u0 );(q1 ;u1)) (T((q0;u0 );(q1 ;u1)) FL(d2)


= (q0 ;u0 );D1 L(d2) ;D2 L(d2) (T((q0;u0);(q1 ;u1 ))FL(d2)
D1 Ld
2 ((q ; u ); (q ; u )) · e + D L(2) ((q ; u ); (q ; u )) · f
( )
= 0 0 1 1 0 2 d 0 0 1 1 0
where FL(d2) means
F Ld
2
( )
((( q0 ; u0 ); e0 ; f0 ); ((q1 ; u1 ); e1 ; f1 ))):
Thus the pullback FLd(2)∗  at ((q0 ; u0 ); (q1 ; u1 )) ∈ TQ × TQ can also be written in another
notation (with summation over repeated indices) as
@L(d2) dqi + @L(d2) dui
@qi0 0 @ui0 0
where the partials and the dqi0 and dui0 are all functions of ((q0 ; u0 ); (q1 ; u1 )). This leads
nally, to the expression for the symplectic form
= -d(FL(d2)∗ ) on TQ × TQ as
@2 L(d2) dqi ^ dqj + @2 L(d2) dqi ^ duj +
@qj1 @qi0 0 1 @uj @qi 0
1 0
1
@2 L(d2) dui ^ dqj + @2 L(d2) dui ^ duj (4.68)
@qj1 @ui0 0 1 @uj @ui 0
1 0
1

4.7.2 Presymplectic algorithm


In Section 4.4, I derived the algorithm on TQ × TQ and it was implicitly de ned by
the discrete Euler-Lagrange equation corresponding to L(d2) , for which I used the short-
hand notation DEL(L(d2) ). I'll call this algorithm on TQ × TQ the algorithm  where
 : ((q0 ; u0 ); (q1 ; u1 )) ! ((q1 ; u1 ); (q2 ; u2 )). The equations corresponding to this algo-
rithm are summarized in Section 4.6 as equations (4.61) and (4.62). This algorithm is
clearly symplectic and has all the other invariance properties associated with discrete Euler-
Lagrange equations as developed by Veselov [19] and Wendlandt and Marsden [20]. This
is because in DEL(L(d2) ) I am just using a discrete Lagrangian on the manifold TQ × TQ
instead of using a discrete Lagrangian on the manifold Q × Q.
Presymplecticness. For the algorithm to be useful it was necessary to derive an algo-
rithm that updates points on the con guration space rather than its tangent bundle. For
35
this I introduced the algorithm that is implicit in

DEL(L(d2) ) uk= qk+1h-qk
which is the discrete Euler-Lagrange equation corresponding to L(d2) but with the u variables
replaced by the linear combination of q variables shown above. As mentioned earlier, h is
the step size. I showed in Section 4.4 that this de nes an algorithm on Q × Q × Q and I will
call this algorithm  where  : (q0 ; q1 ; q2 ) ! (q1 ; q2 ; q3 ). The corresponding equations
are summarized in Section 4.6 in equations (4.63) and (4.64).
Since this is an algorithm on a space of dimension 3n where dim(Q) = n, clearly this
algorithm has no hope of being symplectic unless n is even. This is because one cannot
de ne a nondegenerate 2-form on an odd dimensional manifold. But placing the restriction
of even dimension is too restrictive because it would exclude for example, the important
case of a rigid body. Instead I will show that the algorithm is presymplectic, i.e preserves
a closed, skew-symmetric di erential 2-form which is allowed to be degenerate.
What follows in this paragraph and the next illustrates the strategy one would use to
prove presymplecticness. It does not work because of reasons described in the paragraphs
titled Linearity and Submanifold below. A modi cation of the strategy however does work
and is described following the awed strategy of these two paragraphs. Nevertheless this
awed strategy is given for illustrative purposes rst. In this strategy, to prove presymplec-
ticness we rst try to show show that the following diagram commutes :
Q×Q×Q −−−−!
i TQ × TQ
 



 (4.69)
Q × Q × Q −−−i−! TQ × TQ
In the diagram above, the function i is inclusion; we rst show that it induces a subman-
ifold structure on Q × Q × Q. By showing this and that the diagram commutes, we obtain
an easy proof of presymplecticness as follows. De ne ! = i∗
. Then ∗ ! = ∗ i∗
=
(i ◦ )∗
= ( ◦ i)∗
= i∗ ∗
= i∗
= !. As indicated above, I'll show that the diagram
above has two problems that force us to change this strategy for proving presymplecticness.
Linearity. The rst problem is that in general, i may not be the inclusion function. This
is because
i : (q0 ; q1 ; q2 ) ! ((q0 ; u0 ); (q1 ; u1 ))
is de ned by
u0 q1 - q0 (4.70)
=
h
u1 = q2 -h q1
and there is no guarantee that this implies (q0 ; u0 ) and (q1 ; u1 ) are in TQ. Note that as
mentioned in Section 4.4 we have assumed that Q is embedded in a vector space so that
we can perform operations like the one above in equation (4.70). But, for our proof of
presymplecticness we need an even stronger assumption, namely that Q itself is linear. All
36
the examples in Chapter 5 involve a linear Q. For anyone wanting to prove for nonlinear
manifolds, the invariance properties for the discrete side of the theory outlined in this thesis,
the rst step would be to redo the proofs that appear in this subsection.
Submanifold. The second problem is as follows. For the commutativity of the diagram
(4.69) one requires that
i((q0 ; q1 ; q2 )) = (i(q0 ; q1 ; q2 ))
for every (q0 ; q1 ; q2 ) ∈ Q × Q × Q. Let (q0 ; q1 ; q2 ) be arbitrary. Now,
i((q0 ; q1 ; q2 )) = i(q1 ; q2 ; q3 ) = ((q1 ; u1 ); (q2 ; u2 ))
where q1 and q2 are from the original triple and q3 is obtained by solving equation (4.63).
On the other hand,
(i(q0 ; q1 ; q2 )) = ((q0 ; u0 ); (q1 ; u1 )) = ((q1 ; u1 ); (q~ 2 ; u~ 2 )):
Here q~2 is obtained from q0 and q1 by solving equation (4.60) which is DEL(Ld ). Thus to
show that (4.69) commutes, we have to show that
((q1 ; u1 ); (q2 ; u2 )) = ((q1 ; u1 ); (q~ 2 ; u~ 2 )) (4.71)
for all (q0 ; q1 ; q2 ) ∈ Q × Q × Q. But since q2 is arbitrary it may not be equal to q~2 . This
obstacle to proving presymplecticness on Q × Q × Q is removed by proving it instead for
the appropriate subset of Q × Q × Q as follows. De ne a set S ⊂ Q × Q × Q by
S = f(q0 ; q1 ; q2 ) ∈ Q × Q × Q j D1 Ld (q1 ; q2 ) + D2 Ld (q0 ; q1 ) = 0g: (4.72)
From the preceding description it should be clear that the algorithm
:Q×Q×Q!Q×Q×Q
is actually an algorithm on S, i.e  : S ! S.
Now we just need to show that S is a smooth submanifold of Q × Q × Q and the proof
outlined above for Q × Q × Q can then be extended to work for S. The proof that S is a
smooth manifold is done below in Proposition 4 and the proof of presymplecticness of the
algorithm on S is done in Theorem 2 below.
Proposition 4 S as de ned above is a smooth submanifold of Q × Q × Q and hence
of TQ × TQ.
Proof: Suppose dim(Q) = n. De ne a map F : Q × Q × Q ! Rn by (q0 ; q1 ; q2 ) !
D1 Ld (q1 ; q2 )+ D2 Ld (q0; q1 ). Then by de nition, S = F-1 (0), where 0 ∈Rn . The derivative
of F evaluated at (q0 ; q1 ; q2 ) is
DF(q0 ; q1 ; q2 ) =
[D 1 D2 Ld (q0 ; q1 ); D1 D1 Ld (q1 ; q2 ) + D2 D2 Ld (q0 ; q1 ); D2 D1 Ld (q1 ; q2 )]
This is an n × 3n matrix where each of the comma separated submatrices is an n × n matrix.
37
Recall that the discrete Euler-Lagrange equations DEL(Ld ) are
D1 Ld (q1 ; q2 ) + D2 Ld (q0 ; q1 ) = 0 (4.73)
For these to be uniquely solvable for q2 as a function of q0 and q1 one requires, by the
implicit function theorem, a nondegeneracy condition on Ld , namely that D2 D1 Ld (q1; q2 )
is a nonsingular matrix. (It is sucient to assume this for q1 = q2 by continuity). But then
DF above is maximal rank for all (q0 ; q1 ; q2 ) ∈ Q × Q × Q satisfying equation (4.73). Thus
by the submersion theorem, S = F-1 (0) is a closed submanifold of Q × Q × Q. Now since
Q×Q×Q is a submanifold of TQ×TQ it follows that S is a smooth submanifold of TQ×TQ. 
Now we can give the proof that the algorithm is presymplectic by rst pulling back
the symplectic form on TQ × TQ to a presymplectic form on S and then showing that this
presymplectic form is preserved. This is the content of the following theorem. Note that
by presymplectic form we mean a form which has all the other properties of a symplectic
form except that it may be degenerate.
Theorem 2 The diagram below commutes
S i
−−−−! TQ × TQ
 



 (4.74)
S −−−i−! TQ × TQ
In addition, the algorithm implicit in

DEL(L(d2) ) uk = qk+1h-qk
is presymplectic on S.
Proof: The commutativity of the diagram follows from construction of the algorithm
implicit in equation (4.63), i.e in

DEL(L(d2) ) uk = qk+1h-qk
and from the de nition of S given in (4.72). To show presymplecticness on S let ! = i∗
.
Then ∗ ! = ∗ i∗
= (i ◦ )∗
= ( ◦ i)∗
= i∗ ∗
= i∗
= !. 
38
39

Chapter 5

Examples and Conclusion

In this chapter I will derive the equations for three speci c examples. The rst is a particle
falling from rest under gravity. I will rst show how di erentiating the Lagrangian yields
the time derivative of the Euler-Lagrange equation. Then I will derive the discrete Euler-
Lagrange equations from the original discrete Lagrangian and the di erentiated discrete
Lagrangian. The latter is a nite di erenced version of the former, i.e is the discrete time
derivative of the original discrete Euler-Lagrange equation. The solution is a quadratic
polynomial and both the discrete equations solve the equation of motion exactly.
To see an example in which the discrete equation is not an exact solution of the con-
tinuous equations of motion I repeat the whole exercise for a simple system for which the
solution is not a polynomial. The system I use is a simple harmonic oscillator. Finally I
give an example in which the resulting discrete Euler-Lagrange equation is not linear in the
variable that is being solved for. For this I will use the example of an undamped Dung
oscillator in which the position of interest, qk+1 appears in a cubic equation.
In all these examples the con guration space Q is linear. The theory in Chapter 4 was
developed for a general, possibly nonlinear Q by using coordinate charts. In practice, for
nonlinear Q, one would use Lagrange multipliers to enforce the constraint of evolution on
the nonlinear Q rather than using coordinate charts. This is done for example by Wendlandt
and Marsden [20].

5.1 Falling Particle


5.1.1 Continuous
Consider a particle of mass 1 falling under gravity. The Lagrangian is
L(q; q_ ) = 21 q_ 2 + gq
where positive q direction is downwards. The Euler-Lagrange equation is q = g which
implies, if q(0) = 0 and q_ = 0 that
q(t) = g2 t2
40
Now I derive the linearized Euler-Lagrange using the method developed in this thesis. As
de ned in Section 4.3 in equation (4.20) let
L(2) (q; q;_ u; u_ ) = dL(q; q_ ) · (u; u_ )
@L (q; q_ ) · u + @L (q; q_ ) · u_
@q =
@q_
For the case of the given Lagrangian L(q; q_ ) = 1=2q_ 2 + gq this means
L(2) (q; q;_ u; u_ ) = gu + q_ u_
The resulting Euler-Lagrange equations for this Lagrangian are
u = 0
q = g
Then when one sets u = q_ in the rst equation, one obtains
...
q = 0
which is the time derivative of the original Euler-Lagrange equation
q = g
corresponding to the original Lagrangian L(q; q_ ) = 1=2q_ 2 + gq. This shows a speci c
example of the steps for linearization of Euler-Lagrange equations that were developed in
Section 4.3.

5.1.2 Discrete
The same exercise can now be repeated on the discrete side. Here, rst I give the derivation
of the discrete Euler-Lagrange equation using the method given by Wendlandt and Marsden
[20]. The discrete Lagrangian Ld is
 2  
1
Ld (qk-1 ; qk ) = 2 q k - qk-1
+g
q k -1 + qk
(5.1)
h 2
The discrete Euler-Lagrange equation is
@Ld @Ld
@qk (qk-1 ; qk ) + @qk (qk ; qk+1 ) = 0 (5.2)
Doing this computation for the the discrete Lagrangian above one gets the discrete Euler-
Lagrange
qk+1 - 2qk + qk-1 = g (5.3)
h2
It is worth noting that this is the discrete version of the equation q = g where the left
hand side in (5.3) is the O(h2 ), centered nite di erence approximation for q at qk . This
equation (5.3) gives an algorithm for advancing qk . One can solve for qk+1 in terms of the
values of q at the previous two time steps, namely qk and qk-1 .
41
Now I will derive the linearization on the discrete side using the method of this thesis.
According to this, as shown in Section 4.4 equation (4.34) one writes
L(d2) (qk ; uk ; qk+1 ; uk+1 ) = dLd (qk ; qk+1 ) · (uk ; uk+1 )
@Ld u + @Ld u
=
@q k @q k+1
 k k+1 
qk+1 - qk + g u +  qk+1 - qk + g  u
= -
h2 2 k h2 2 k+1
Similarly
   
Ld2 (qk-1 ; uk-1 ; qk ; uk ) = q k - qk-1 g q k - qk-1 g
h2 + 2 uk-1 + h2 + 2 uk (5.4)
( )
-

The two discrete Euler Lagrange equations corresponding to L(d2) are


@L(d2) (q ; u ; q ; u ) + @L(d2) (q ; u ; q ; u ) = 0 (5.5)
@qk k-1 k-1 k k @qk k k k+1 k+1
@L(d2) (q ; u ; q ; u ) + @L(d2) (q ; u ; q ; u ) = 0 (5.6)
@uk k-1 k-1 k k @uk k k k+1 k+1
For the speci c Lagrangian L(d2) that is being considered these become, respectively
uk+1 - 2uk + uk-1 = 0 (5.7)
h2
qk+1 - 2qk + qk-1 = g (5.8)
h2
The second equation above is the original discrete Euler-Lagrange equation and when one
substitutes
uk = qk+1h- qk
in the rst equation above one gets the linearized discrete Euler-Lagrange equation
qk+2 - 3qk+1 + 3qk - qk-1 = 0 (5.9)
h3
...
It is again worth noting that this is the discrete approximation of the equation q = 0 which
is the linearized continuous Euler-Lagrange equation. Indeed the left hand side of equation
(5.9) is the O(h) one sided nite di erence approximation of the third derivative at qk-1 .
Also note that this equation (5.9) can be obtained by a nite di erence of the equation
(5.3) as follows.
1  qk+2 - 2qk+1 + qk - qk+1 - 2qk + qk-1  = g - g = 0
h h2 h2
Note that the discrete Euler-Lagrange equation could have been derived in another way.
As summarized in Section 4.6 in equation (4.63), one can write the DEL(L(d2) ) in terms of
42
Ld as
2 2
uk-1 @q@@qLd (qk-1 ; qk ) + uk @@qL2d (qk-1 ; qk ) +
k k-1 k
2 Ld 2
@
uk @q2 (qk ; qk+1 ) + uk+1 @q@@qLd (qk ; qk+1 ) = 0 (5.10)
k k k+1
For the speci c Ld in this example, given above in 5.1 one nds that
@2 Ld (q ; q ) = - 1
@qk @qk-1 k-1 k h2
2
@ Ld (q ; q ) = 1
@q2k k-1 k h2
@2 Ld (q ; q ) = 1
@q2k k k+1 h2
@2 Ld (q ; q ) = - 1
@q @q k k+1 h2
k k+1

Thus equation (5.10) becomes


- uk-1 + uk + uk - uk+1 = 0
h2
which is what was obtained above in equation (5.7). Thus in this example I have shown a
speci c case of the commutativity of diagram (4.54).

5.2 Simple Harmonic Oscillator


5.2.1 Continuous
Consider an undamped mass spring system with mass m and spring constant K. The
Lagrangian is
L(q; q_ ) = 12 mq_ 2 - 12 Kq2
The Euler-Lagrange equation corresponding to this is
mq + kq = 0 (5.11)
If one takes the initial conditions q(0) = 1 and q_ = 0 the solution to the above equation is
q(t) = cos( K=m)t. To linearize equation (5.11) using the method of this thesis let
L(2) (q; q;_ u; u_ ) @L (q; q_ ) · u + @L (q; q_ ) · u_
=
@q @q_
= -Kqu + mq _ u_
43
The Euler-Lagrange equations corresponding to this Lagrangian are
d @L(2) - @L(2) 0
dt @u_ @u =

d @L(2) - @L(2) 0
dt @q_ @q =

The rst of these yields


mq + kq = 0
which is the Euler-Lagrange equation corresponding to the original Lagrangian L(q; q_ ) and
the second of these yields
mu + Ku = 0
which after the substitution u = q_ becomes
...
q
m + Kq_ = 0 (5.12)
Note that equation 5.12 above is the time derivative of the original Euler-Lagrange equation
as expected.

5.2.2 Discrete
The discrete Lagrangian is
 2  2
Ld (qk-1 ; qk ) = 21 m qk -hqk-1 - 12 K qk-12+ qk
 2  2
1
Ld (qk ; qk+1 ) = 2 m q k +1 - qk 1
- K
q k + qk+1
h 2 2
Hence
@Ld m K
@qk (qk-1 ; qk ) = h2 (qk - qk-1 ) - 4 (qk-1 + qk )
@Ld (q ; q ) = - m (q - q ) - K (q + q )
@qk k k+1 h2 k+1 k 4 k k+1
which implies that the discrete Euler-Lagrange equation (5.2) is
 
m qk+1 - 2qk + qk-1 + K (q + 2q + q ) = 0
h2 4 k-1 k k+1 (5.13)
Comparing this to the continuous time Euler-Lagrange equation (5.11) one nds that the
acceleration q (t) of the continuous time equation has been replaced by the O(h2 ) centered
nite di erence discrete approximation of q at qk . Also, the position q(t) in the continuous
time equation has been replaced by a weighted average of the positions at qk-1 ; qk and
qk+1 .
In order to linearize the discrete Euler-Lagrange equation (5.13) using the method of
44
this thesis one de nes
L(d2) (qk-1 ; uk-1 ; qk ; uk ) = dLd (qk-1 ; qk ) · (uk-1 ; uk )
L(d2) (qk ; uk ; qk+1 ; uk+1 ) = dLd (qk ; qk+1 ) · (uk ; uk+1 )
which means

L(d2) (qk-1 ; uk-1 ; qk ; uk ) = uk-1 m (q - q ) - K (q + q )

-
h2 k k-1 4 k-1  k
+ uk
m (q - q ) - K (q + q )
h2 k k-1 4 k-1 k
 
L(d2) (qk ; uk ; qk+1 ; uk+1 ) = uk - hm2 (qk+1 - qk ) - K4 (qk + qk+1 )

+ uk+1
m (q - q ) - K (q + q )
h2 k+1 k 4 k k+1
Then the discrete Euler-Lagrange equations (5.5) and (5.6) are
 
m uk+1 - 2uk + uk-1 + K (u + 2u + u ) = 0
h2 4 k-1 k k+1 (5.14)
 
m qk+1 - 2q
h2
k + qk-1 + K (q + 2q + q ) = 0
4 k-1 k k+1 (5.15)
The second of these is the same as the original discrete Euler-Lagrange equation and when
one substitutes
uk = qk+1h- qk
in the rst equation, namely in (5.14), one gets the linearized version of the original discrete
Euler-Lagrange equation. The linearized equation that one gets with the above substitution
is
 
m q k +2 - 3qk+1 + 3qk - qk-1
h3  
K q k - qk-1
+2
q k +1 - qk q k +2 - qk+1
= 0 (5.16)
+
4 h h +
h
Note that the equation (5.16) is the discrete version of the continuous equation (5.12)
m...q + Kq_ = 0
...
The q of the continuous equation has been replaced by the O(h) one sided nite di erence
approximation at qk-1 and the velocity q_ appearing in the equation has been replaced by
a weighted average of the O(h) approximations of the velocities at several time steps. Also
note that equation(5.16) can be obtained from the original discrete Euler-Lagrange equation
(5.13) by nite di erencing as follows
1 ((5.13) at(q ; q ; q ) - (5.13) at(q ; q ; q ))
h k+2 k+1 k k+1 k k-1
45
since that gives
m  qk+2 - 2qk+1 + qk  + K (q + 2q + q )
h h2 
4h k+2 k+1 k

m q k +1 - 2qk + qk-1 K
-
h h2 -
4h (qk-1 + 2qk + qk+1 ) = 0 (5.17)
which is the same as equation (5.16). As explained earlier, nite di erencing is the same as
taking the discrete time derivative in such cases where the equation is linear in its variables.

5.3 Duffing Oscillator


A prototypical example of a nonlinear oscillator is the Dung oscillator. I will consider
the unforced, undamped Dung oscillator. The equation of motion for this oscillator is the
nonlinear di erential equation q - q + q3 = 0. The vector eld of this equation has 3 xed
points, one of which is at the origin which is a saddle point. The other two xed points are
centers with periodic orbits surrounding each of them. For more on this oscillator and for
its phase plot see Example 1.1.7 on page 29 of the book by Wiggins [21]. In what follows, I
will derive the equation of motion for the undamped, unforced Dung oscillator from the
Lagrangian. Then I will demonstrate the linearization by the method of this thesis and
repeat that for the discrete version.

5.3.1 Continuous
Let the Lagrangian be
2 4
L(q; q_ ) = 21 q_ 2 + q2 - q4
Then the Euler-Lagrange equation is
q - q + q3 = 0 (5.18)
On di erentiating the Lagrangian L(q; q_ ) to get L(2) (q; q;
_ u; u_ ) one gets
L(2) (q; q;_ u; u_ ) = qu - q3 u + q_ u_
and the corresponding Euler-Lagrange equations are
u - u + 3q2 u = 0 (5.19)
q - q + q3 = 0 (5.20)
As expected, the second one is the Euler-Lagrange equation corresponding to the original
Lagrangian and the rst one on substituting u = q_ becomes the derivative of the original
equation, i.e it becomes
...
q - q_ + 3q2 q_ = 0 (5.21)
46
5.3.2 Discrete
The discrete Lagrangian is
 2  2  4
Ld (qk-1 ; qk ) = 12 qk -hqk-1 + 12 qk-12+ qk - 14 qk-12+ qk
and similarly for Ld (qk ; qk+1 ). The discrete Euler-Lagrange equation DEL(Ld ) is
qk+1 - 2qk + qk-1 - 1 (q + 2q + q ) +
h2 4 k-1 k k+1
 
1  qk-1 + qk 3 +  qk + qk+1 3 = 0 (5.22)
2 2 2
Di erentiating the discrete Lagrangian Ld to get L(d2) according to the method of this thesis
one gets
L(d2) (qk-1 ; uk-1 ; qk ; uk ) =
(q - q ) (q + q ) (q + q )
uk-1 - k h2 k-1 + k-14 k - k-116 k
( q - q ) (q + q ) (q + q )
+ uk
k k-1 + k-1 k - k-1 k (5.23)
h2 4 16
and similarly for L(d2) (qk ; uk ; qk+1 ; uk+1 ). There are two corresponding discrete Euler-
Lagrange equations one of which is the same as the discrete Euler-Lagrange equation cor-
responding to the original discrete Lagrangian, i.e it is the same as equation (5.22). The
other equation is the discrete time derivative of that equation and in terms of uk etc. it is
given by
uk+1 - 2uk + uk-1 - uk-1 + 2uk + uk+1
h2  
4 2  2
1 q k + qk+1 q k + qk+1
+
4 3uk+1 2 + 3uk
2

+ 3uk
qk-1 + qk 2 + 3u  qk-1 + qk 2

0 (5.24)
2 k-1 2 =

Substituting
uk = qk+1h- qk
47
in the above equation gives the discrete time derivative of the original discrete Euler-
Lagrange equation. The equation one gets with this substitution is
qk+2 - 3qk+1 + 3qk - qk-1
h3      
1 q k - qk-1 q k +1 - qk q k +2 - qk+1
-
4 h + 2 h +
h
  2
1 qk+2 - qk+1 qk + qk+1
+
4 3 h 2
  2
+3
q k+1 - qk q k + qk+1
h 2
  2
+3
q k+1 - qk q k-1 + qk
h 2
  2 
+ 3
qk +1 - qk q k -1 + qk
= 0 (5.25)
h 2
It is worth noting some features of the equation above. Compare this equation with the
continuous Euler-Lagrange equation (5.21) and note that equation (5.25) is the discrete
version of the continuous equation.
... The rst term in (5.25) is the O(h) one sided nite
q
di erence approximation of . The next term in the discrete time equation is a weighted
_ 2 at various
average of q_ at various points and the last big term is a weighted average of 3qq
2
points where the q term in q is itself an average position. All this is as one would expect
from the commutative diagram (4.54) of Chapter 4. Note that equation (5.25) cannot be
obtained by taking a a nite di erence from the original discrete Euler-Lagrange equation
(5.22) since the original equation is not linear. However, equation (5.25) can be obtained by
taking the discrete time derivative of the original discrete Euler-Lagrange equation (5.22).

5.4 Numerical Results


For the simple harmonic oscillator example I implemented (in MATLAB) the discrete Euler-
Lagrange equation and the linearized discrete Euler-Lagrange equation which were both
derived in Section 5.2. After rearranging terms in equation (5.13) and solving for the
position at next time step in terms of the previous two one gets the algorithm
2 2
qk+1 = - -48 ++ !2!2 hh2 qk - qk-1 (5.26)

where ! = K=m. Similarly in the linearized discrete Euler-Lagrange equation (5.16) after
rearranging and solving for the position at next time step in terms of the previous three
positions one gets
2 2 2 2
qk+2 = 124 +-!!2 hh2 qk+1 + -412++!!2 hh2 qk + qk-1 (5.27)
The rst two positions q0 and q1 are required to start the rst integrator and the rst three
positions q0 , q1 and q2 are required to start the second integrator. These starting values
were calculated from the analytical solution.
48
The global error was measured as described below. Let  be the computed ow in one
time step so that qk+1 = (qk ). Then de ne global error as a function of the time step h
and of time at which error is measured as
 
e(h; T ) = (T=h)(q0 ) - q(T ) (5.28)
where q(t) is the analytic solution and n means n applications of the algorithm. Thus
e(h; 1) is the the error at time 1 for step size h.
The time at which the error is measured is not important for the estimation of the order
of the algorithm. For example Figures 5.1 and 5.2 give the errors measured at times 1 and
5 respectively. The algorithm called linearized discrete EL seems to better than discrete EL
at time 1 but does worse at time 5. But when the order of the algorithms is estimated it is
seen that this does not matter. For reference I give the point wise error for step size 0.001
for both the algorithms in Figure 5.3. From this it is clear that the error at time 1 is less
for the linearized discrete EL but at time 5 that error is more.
To estimate the order of the algorithms I computed e(h; 1) for h = h0 , 2h0 , 4h0 , 8h0 and
16h0 with h0 = 0:001. Then I computed the ratios e(2h0 ; 1)=e(h0 ; 1), e(4h0 ; 1)=e(2h0 ; 1)
etc. Since these ratio was nearly 4 for both the algorithms one can conclude that the
order of both the algorithm is O(h2 ). The tables 5.4 and 5.4 give some evidence that for
estimating the order of the algorithm the time at which global error measurement is done
is not important as long as the same time is used for all comparisons.

5.5 Accuracy Calculations


In this section I will show the calculation of the local error (or order) of the algorithms
for the simple harmonic oscillator in detail. The result for the Dung oscillator will also
be mentioned without the details of the derivation. We will use the following de nition
and method to compute our local error. This method entails computing how much the
exact solution fails to satisfy the di erence method. Speci cally, the calculation is done by
replacing q(k + 1) in the algorithm by q(t + h) and doing a Taylor expansion for q(t + h)
around t. Similarly, q(k - 1) is replaced by q(t - h) and q(k + 2) by q(t + 2h) etc. After
cancellations, whatever function of h remains gives what we will call the order of accuracy
or error of the method.
Consider the discrete Euler-Lagrange equation (i.e the algorithm) for computing the
solution for a simple harmonic oscillator that was derived in Section 5.2. This is given here
for convenience.
 
m qk+1 - 2q
h2
k + qk-1 + K (q + 2q + q ) = 0
4 k-1 k k+1

In the above equation let qk+1 = q(t + h) etc. in the left hand side. As a result one obtains
that the left hand side is
m K
h2 (q(t + h) - 2q(t) + q(t - h)) + 4 (q(t + h) + 2q(t) + q(t - h)) :
49

−5 Global Error at time 1


x 10
1.8

Discrete EL
1.6 Linearized Discrete EL

1.4
|Computed−Analytic| at time 1

1.2

0.8

0.6

0.4

0.2

0
0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016
Step size h

Figure 5.1: Global error at Time 1 for various step sizes

−4 Global Error at time 5


x 10
1.4

Discrete EL
Linearized Discrete EL
1.2

1
|Computed−Analytic| at time 5

0.8

0.6

0.4

0.2

0
0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016
Step size h

Figure 5.2: Global error at Time 5 for various step sizes


50

−7 Pointwise error for h=0.001


x 10
8

Discrete EL
6 Linearized discrete EL

2
Computed − Analytic

−2

−4

−6

−8
0 1 2 3 4 5 6 7 8 9 10
Time

Figure 5.3: Point wise error for step size h = 0:001 as a function of time

Algorithm e(0:002;1) e(0:004;1) e(0:008;1) e(0:016;1)


e(0:001;1) e(0:002;1) e(0:004;1) e(0:008;1)
Equation(5.26) 3.9924 3.9788 3.9574 3.9673
Equation(5.27) 3.9633 3.9671 3.9369 3.9946
Table 5.1: Ratio, on doubling step size of global error at time 1

Algorithm e(0:002;1) e(0:004;1) e(0:008;1) e(0:016;1)


e(0:001;1) e(0:002;1) e(0:004;1) e(0:008;1)
Equation(5.26) 4.0025 3.9991 3.9982 3.9935
Equation(5.27) 4.0053 4.0026 4.0060 3.9999
Table 5.2: Ratio, on doubling step size of global error at time 5
51
On substituting the Taylor expansions for various terms one gets
m q(t) + hq  (t) + h2 q  (t) + h3 q(3) (t) + h4 q(4) (t) + h5 q(5) (t) + O(h6 )
h2 2 6 24 120
m
-
h2 2q

(t)

m h 2  h 3 (3) h 4 (4) h 5 (5) 6
h2 q(t) - hq (t) + 2 q (t) - 6 q (t) + 24 q (t) - 120 q (t) + O(h )

+

K h 2  h 3 (3) h4 (4) h 5 (5)


6
4 q(t) - hq (t) + 2 q (t) - 6 q (t) + 24 q (t) - 120 q (t) + O(h )

+

K
+ 2q(t)
4
K q (t) + hq (t) +
 h2 q  (t) + h3 q(3) (t) + h4 q(4) (t) + h5 q(5) (t) + O(h6 ) :
+
4 2 6 24 120
Carrying out the various cancellations above yields
m h2 q  (t) + h4 q(4) (t) + O(h6 )
h2 12

K 4q 2 
(t) + h q (t) +
h4 q(4) (t) + O(h6 )
+
4 12
 2   2 
= m q  (t) + h12 q(4) (t) + O(h4 ) + K q(t) + h4 q  (t) + O(h4 )
= mq  (t) + Kq(t) + O(h2 )
= O(h2 )
since mq  (t) + Kq(t) = 0. A similar calculation for equation (5.16) yields
 
LHS = (mq 3 (t) + Kq  (t)) + 1 mq(4) (t) + 1 Kq  (t) h
( )
2 2

1 mq(5) (t) + 5 Kq(3) (t) h2 + O(h3 )
+
4 12
2
= O(h )

since mq(3) (t) + Kq  (t) = 0 and mq(4) (t) + Kq  (t) = 0. Thus according to our de nition of
local error the error of both the discrete Euler-Lagrange equation and that of the linearized
version is O(h2 ). The equality of order is also seen in the numerical global error comparison
in the previous section.
For the Dung oscillator the local error computed was O(h2 ) for the discrete Euler-
Lagrange equation and O(h) for the linearized version. This shows that the error can be
worse for the linearized integrator.

5.6 Using more accurate uk

One idea to try to improve the order of accuracy of the linearized discrete Euler-Lagrange
equation is to use a better approximation for uk . But when one uses the O(h2 ) centered
52
nite di erence approximation of the rst derivative by setting
uk = qk+1 2h
- qk-1
(5.29)
instead of the O(h) one sided nite di erence approximation as in
uk = qk+1h- qk
one still gets a method that is O(h2 ) for the case of the simple harmonic oscillator. The
linearized discrete Euler-Lagrange equation that one gets for the harmonic oscillator if the
more accurate uk is used is
m  -qk-2 + q + q + qk+2 
h3 2 k-1 k+1 2
   
K qk - qk-2 q k +1 - qk-1 q k +2 - qk
+
4 2h + 2 2h +
2h = 0 (5.30)

This is an expected linearization of the continuous equation in that the rst term is the
centered di erence O(h2 ) approximation of the third derivative and the other term is the
average of the O(h2 ) approximation of the rst derivative at several points.

5.7 Conclusion
I have shown that Euler-Lagrange equations and discrete Euler-Lagrange equations can be
linearized, i.e di erentiated, by di erentiating the Lagrangian or discrete Lagrangian. The
original motivation for di erentiating the Lagrangian was to develop higher order varia-
tional integrators. However it was found that the resulting intgerators are not higher order.
Nonetheless the pursuit of this goal turned up several interesting results :
i. Di erentiating a Lagrangian along the acceleration, i.e along (q; _ q ) results in the
Euler-Lagrange equation 0 = 0 (Lemma 2, page 20).
ii. Di erentiating a Lagrangian along a general direction (u; u_ ), determining the Euler-
Lagrange equation and then setting u = q_ results in an Euler-Lagrange equation that
is the time derivative of the original Euler-Lagrange equation corresponding to the
original Lagrangian (Lemma 3, page 21).
iii. A similar operation on the discrete Lagrangian gives a \discrete time derivative" of
the discrete Euler-Lagrange equation corresponding to the original discrete Lagrangian
(Lemma 4, page 24).
iv. The operations described in the previous two items are part of a commutative diagram
(Theorem 1, page 28).
v. Numerical measurements and analytical calculation of order of accuracy for speci c
examples using this approach show that higher order methods are not obtained in
general (Chapter 5, especially Section 5.4 and 5.5).
vi. The manifold consisting of triples that satisfy a discrete Euler Lagrange equation is a
smooth manifold (Proposition 4, page 36).
53
vii. The algorithm on TQ × TQ is symplectic and the algorithm that updates triples of
points on Q is presymplectic (Theorem 2, page 37).
I also gave a derivation of a symplectic form on TQ × TQ where Q is the con guration
manifold. Most proofs given in this thesis were done using local coordinates. A possible
direction for further study is to give intrinsic proofs in order to gain a deeper understanding
of the geometric meaning of the operations of this thesis.
Consider for example the operation of setting u = q_ after the variational principle has
been applied. The calculus equivalent of this is to nd stationary points in a problem of
nding extrema and then imposing constraints after the stationary points have been found.
The setting of u = q_ after applying the variational principle is the calculus of variations
equivalent of this problem. This apparently \wrong" operation leads to Euler-Lagrange
equations that have some structure, namely that they are the derivative of the original
equations.
Similar comments apply on the discrete side. However, on the discrete side, the lin-
earization described in this thesis may be a simple consequence of commuting of mixed
partials.
54
55

Bibliography

[1] Fornberg, Bengt. A Practical Guide to Pseudospectral Methods, Cambridge University


Press, 1996.
[2] Ge, Zhong and Marsden, Jerrold E. Lie-Poisson Hamilton-Jacobi Theory and Lie-
Poisson Intgerators, Physics Letters A, volume 133, 3, 134-139, 1988.
[3] Gear, C. William. Numerical Initial Value Problems in Ordinary Di erential Equations.
Prentice-Hall, 1971.
[4] Gelfand, I.M and Fomin, S.V. Calculus of Variations. Prentice-Hall, 1963.
[5] Gonzalez, Oscar. Design and analysis of conserving integrators for nonlinear Hamil-
tonian systems with symmetry. Ph.D thesis, Stanford University, Department of Me-
chanical Engineering, 1996.
[6] Hairer, E., Norsett, S.P., Wanner, G. Solving Ordinary Di erential Equations I, Nonsti
Problems, Second Revised Edition, Springer-Verlag, 1993.
[7] Iserles, Arieh and Zanna, A. Qualitative numerical analysis of ordinary di erential
equations. DAMTP preprint DAMTP/NA5, Cambridge University, September 1995.
[8] Iserles, Arieh. A rst course in the numerical analysis of di erential equations, Cor-
rected Printing, Cambridge University Press, Cambridge, UK, 1996.
[9] Kane, C., Marsden, J.E., Ortiz, M. Symplectic- Energy-Momentum Preserving Vari-
ational Integrators. Journal of Mathematical Physics, 40, 1999. Available from
www.cds.caltech.edu.
[10] Kane, C., Marsden, J.E., Ortiz, M., West, M. Variational Integrators and the New-
mark Algorithm for Conservative and Dissipative Mechanical Systems. CDS Technical
Report, Caltech, 1999. Available from www.cds.caltech.edu.
[11] Lambert, J.D. Computational Methods in Ordinary Di erential Equations, Wiley,
1973.
[12] Marsden, J.E., O'Reilly, O.M., Wicklin, F.J and Zombro,B.W. Symmetry, Stability,
Geometric Phases and Mechanical Integrators (Part II). Nonlinear Science Today,
volume 1, number 2, 1991.
[13] Marsden, Jerrold E. Park City Lectures on Mechanics, Dynamics and Symmetry, Avail-
able from \http://www.cds.caltech.edu/ marsden/lecturenotes/lecturenotes.html",
July 1997.
56
[14] Marsden, J.E. Personal communication.
[15] Marsden, Jerrold E. and Ratiu, Tudor S. Introduction to mechanics and symmetry : a
basic exposition of classical mechanical systems, Second Edition, 1999.
[16] Marsden, Jerrold E. Personal Communication, March 2000.
[17] West, Matthew. Personal communication, 1999 (publication in preparation).
[18] Strang, Gilbert. Introduction to Applied Mathematics, Wellesley-Cambridge Press,
MA, 1986.
[19] Veselov, A.P. Integrable discrete-time integrators and di erence operators. Funkts.
Anal. Prilozhen, volume 22(2), 1-13, 1988.
[20] Wendlandt, Je rey M. and Marsden, Jerrold E. Mechanical integrators derived from a
discrete variational principle. Physica D, volume 106, 223-246, 1997.
[21] Wiggins, S. Introduction to Applied Nonlinear Dynamical Systems and Chaos,
Springer, 1990.

You might also like