Download as pdf or txt
Download as pdf or txt
You are on page 1of 22

BIT Numer Math

DOI 10.1007/s10543-014-0541-x

Order conditions for G-symplectic methods

John C. Butcher · Gulshad Imran

Received: 24 June 2014 / Accepted: 4 December 2014


© Springer Science+Business Media Dordrecht 2015

Abstract General linear methods for the solution of ordinary differential equations
are both multivalue and multistage. The order conditions will be stated and analyzed
using a B-series approach. However, imposing the G-symplectic structure modifies the
nature of the order conditions considerably. For Runge–Kutta methods, rooted trees
belonging to the same tree have equivalent order conditions; if the trees are superfluous,
they are automatically satisfied and can be ignored. For G-symplectic methods, similar
results apply but with a more general interpretation. In the multivalue case, starting
conditions are a natural aspect of the meaning of order; unlike the Runge–Kutta case
for which “effective order” or “processing” or “conjugacy” has to be seen as having
an artificial meaning. It is shown that G-symplectic methods with order 4 can be
constructed with relatively few stages, s = 3, and with only r = 2 inputs to a step.

Keywords G-symplectic methods · Order conditions · Conformability · Trees ·


Rooted trees

Mathematics Subject Classification 65L05 · 65L06

Communicated by Anne Kværnø.

J. C. Butcher (B) · G. Imran


Department of Mathematics, University of Auckland, Auckland, New Zealand
e-mail: butcher@math.auckland.ac.nz
G. Imran
e-mail: ggul005@aucklanduni.ac.nz

123
J. C. Butcher, G. Imran

1 Introduction

General linear methods are multistage and multivalue and are designed to give numer-
ical solutions to the initial value problem
y  (x) = f (y(x)), y(x0 ) = y0 , f : RN → RN , y0 ∈ R N . (1.1)

At the start of step number n, r vectors yi[n−1] , i = 1, 2, . . . , r are used as input and at
the end of the step the output is yi[n] , i = 1, 2, . . . , r . Like a Runge–Kutta method, there
are s stage values Yi , i = 1, 2, . . . , s, to be computed as well as the corresponding
stage derivatives Fi = f (Yi ), i = 1, 2, . . . , s.
These are interrelated by the equations (see [2])

s 
r
Yi = h ai j F j + u i j y [n−1]
j , i = 1, 2, . . . , s,
j=1 j=1
(1.2)

s 
r
yi[n] = h bi j F j + vi j y [n−1]
j , i = 1, 2, . . . , r,
j=1 j=1

where the coefficients ai j , u i j , bi j , vi j comprise the (s +r )×(s +r ) partitioned matrix


 
AU
.
B V

If we write
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
y1[n−1] y1[n] Y1 F1
⎢ [n−1] ⎥ ⎢ [n] ⎥ ⎢
⎢ y2 ⎥ ⎢ y2 ⎥ Y 2⎥ ⎢ F2 ⎥
y [n−1] =⎢ ⎥ y [n] =⎢ ⎥, Y = ⎢ ⎥ ⎢ ⎥
⎢ .. ⎥ , .
⎢ . ⎥ ⎢

.
. ⎥, F = ⎢ . ⎥,
⎣ .. ⎦
⎣ . ⎦ ⎣ . ⎦ . ⎦
yr[n−1] yr[n] Ys Fs

then (1.2) can be rewritten in a compact form


Y = h(A ⊗ I N )F + (U ⊗ I N )y [n−1] ,
y [n] = h(B ⊗ I N )F + (V ⊗ I N )y [n−1] .

This paper is concerned with a special class of “G-symplectic” general linear meth-
ods. These have similar properties to symplectic Runge–Kutta methods and, although
quadratic invariants cannot be conserved, as in the Runge–Kutta case, they conserve
a more general quantity related to the invariant.
Symplectic, or canonical, Runge–Kutta methods were introduced in [10]. They are
the subject of two important expository works [11] and [13]. Of special interest in the
present paper is the study of the order conditions for these methods published in [12].
The generalization to multivalue G-symplectic methods was introduced in [7] and [3].
The significance and management of parasitism is the subject of [5]; a preliminary
discussion of the question of parasitism appeared in [4].

123
Order conditions for G-symplectic methods

Let T denote the set of all rooted trees and T # the set containing the members of
T together with the empty tree, which will be written as ∅. Following the notation of
[3], let G denote the set of mappings T # → R. The subsets of G in which ∅ → 0 (or
1) will be denoted by G 0 (or G 1 respectively).
According to the “B-series” criterion for order, a method (A, U, B, V ) has order p
if there exists ξ ∈ G r and η ∈ G s1 such that
η = A(ηD) + U ξ , (1.3)
Eξ = B(ηD) + V ξ , (1.4)

where (1.4) has to hold only for trees of order p or less. A brief discussion of B-series,
including the use of E ∈ G 1 and D ∈ G 0 will be included in Sect. 3.
In practice, y [0] is computed by a formula of the form y [0] = Sh y(x0 ), where the
“starting procedure” Sh : R N → (R N )r is used to initiate the process defined by (1.2).
In the order criterion given by (1.3) and (1.4), ξ is the B-series representative of Sh
and will be referred to as the “starting method”.
By recursively evaluating the order conditions (1.3) and (1.4), tree by tree, it is
possible to obtain the order conditions up to any required order. However, just as for
Runge–Kutta methods, the canonical conditions have an influence on the results. Also
the results need to be understood relative to the starting method. In other words, the
value of ξ (t) needs to be taken into account for |t| ≤ p. It is found that a generalization
of the results in [12] apply in the multi-value case and this will be one of the principal
conclusions of the paper.
In Sect. 2, the definition of the G-symplectic property is given and it is shown how
this can be transformed into a possibly complex form, in which V becomes a diagonal
matrix.
This is followed in Sect. 3 by a review of the roles played by trees and rooted trees
in questions concerning numerical order. The main Sect. 4 is devoted to an analysis of
the order conditions for G-symplectic methods. The particular case of methods with
order 4 is considered in Sect. 5, where sample methods are derived and analysed.
In particular, starting and finishing methods are constructed for these specific cases.
Numerical simulations are presented in Sect. 6. These confirm the conservative nature
of the new methods and add to the more extensive range of numerical tests given
in [5].

2 G-symplectic methods in diagonal form

G-symplectic methods are a special class of general linear methods defined by the
condition (see [2]), that the partitioned matrix M in (2.1) below is zero, where D is a
real diagonal matrix and G is a non-singular symmetric matrix
   
M11 M12 D A + AT D − B T G B DU − B T GV
M= = . (2.1)
M21 M22 U T D − V TG B G − V T GV

It will always be assumed that the eigenvalues of V are distinct. For practical compu-
tations each of the matrices A, U , B and V should have real elements. However, for

123
J. C. Butcher, G. Imran

analysis, it is convenient for V to be a diagonal matrix. We will assume that V is a


diagonal matrix
V = diag(z 1 , z 2 , . . . , zr ),
and that z 1 = 1. This assumption can always be satisfied by introducing a similarity
transformation to V , which has distinct eigenalues, and making corresponding changes
to B and U .
If r is odd, the remaining z values occur in conjugate pairs and, if r is even, one of
the z values is −1 and the remainder occur in conjugate pairs. It follows from M22 = 0
that G = diag(g1 , g2 , . . . , gr ), with real diagonals. If z i = z j , we will assume that
gi = g j , U ei = U e j and ei B = e j B. We will always assume that g1 = 1, This does
not lead to a loss of generality because the elements of U can be rescaled if necessary.
By consistency combined with M12 = 0, it follows that
U e1 = 1, D = diag(b), where e1T B = bT .
This enables us to write
 T

B∗G
b , = V .
B= , U = 1U DU
B
We will denote the “principal component” of ξ by ζ and the remainder of this vector
by
ξ . Hence (1.4) takes the form

Eζ = bT (ηD) + ζ, (2.2)
ξ=
E B(ηD) + V ξ. (2.3)

The conditions associated with (2.3) do not impose conditions on the coefficients of
the method and should be thought of as conditions on ξ . As we will see in Subsection
4.3, we can assume that they are always satisfied. Hence, the difficulty of constructing
a method of specific order is to choose the coefficients of the method along with ζ so
that (2.2) is satisfied.
For methods with r ≥ 3, complex numbers will arise in the diagonal form repre-
sentation of a method. For practical use, these methods can be transformed into real
form by a change of basis operation
U → U T, B → T −1 B, V → T −1 V T.
The matrix G, introduced in M undergoes a related transformation
G → T ∗ GT.
For reference purposes we restate the upper-triangular blocks of the Hermitian matrix
equation M = 0, given by (2.1) in the case that the transformation to diagonal form
has taken place.
G = V ∗ GV, (2.4)
DU = B ∗ GV, (2.5)
D A + AT D = B ∗ G B. (2.6)

123
Order conditions for G-symplectic methods

3 Trees, rooted trees and order conditions

As for general linear methods, as well as for Runge–Kutta methods, order conditions
are defined in terms of Taylor series, or B-series as they are referred to in this context,
based on expansions in terms of elementary differentials.
We recall the general form of a B-series expansion about the point (x0 , y0 ) on the
trajectory of a solution to the initial value problem (1.1).

 α(t)h |t|
B(α, y0 ) = α(t0 )y0 + F(t)(y0 ). (3.1)
σ (t)
t∈T

The coefficient of y0 is denoted by α(t0 ), where t0 is the “empty tree”, and T denotes
the set of all rooted trees. Special names will be given for the rooted trees of order not
exceeding 4. These are

t1 = , t2 = , t3 = , t4 = , t5 = , t6 = , t7 = , t8 = .

The notation t0 , t1 , . . . , is in contrast to t, t1 , t2 , which will denote arbitrary trees,


Two special cases ot B-series are D and E, corresponding to
D(t0 ) = 0,
D(t1 ) = 1,
D(t) = 0, |t| > 1,
E(t0 ) = 1,
1
E(t) = , t ∈ T.
t!
The significance of these series is that B(D, y0 ) = h f (y0 ), B(E, y(x0 )) = y(x0 + h).
By using compositions, these enable the evaluation of series for

h f (B(α, y0 )) = B(αD, y0 ), α(t0 ) = 1,


B(α, y(x0 + h)) = B(Eα, y(x0 )).

If t1 , t2 are arbitrary rooted trees with positive order we can construct an asymmetrical
product t1 t2 by attaching the roots and designating the root of t1 to be the root of t1 t2 .
This will mean that, although t1 t2 and t2 t1 are not in general the same rooted tree, they
are equivalent in the sense that they have the same underlying tree even though the
roots may be different. Following the definitions in [12] we state
Definition 3.1 Two rooted trees are equivalent if they have the same underlying tree.
We will regard the underlying tree as an equivalence class for the corresponding
rooted trees.
Definition 3.2 A tree is superfluous if it contains a rooted tree of the form tt. A rooted
tree is superfluous if it belongs to a superfluous tree. A tree (respectively rooted tree)
is non-superfluous if it is not superfluous.

123
J. C. Butcher, G. Imran

Table 1 Trees and corresponding rooted trees including two superfluous trees

Tree Rooted trees

t1

(superfluous) t2 = t1 t1

t3 = t2 t1 t4 = t1 t2

t5 = t3 t1 t7 = t1 t3

(superfluous) t6 = t4 t1 = t2 t2 t8 = t1 t4

t9 = t5 t1 t14 = t1 t5

t10 = t3 t2 = t6 t1 t11 = t2 t3 = t7 t1 t15 = t1 t6 t16 = t1 t7

t12 = t8 t1 = t2 t4 t13 = t4 t2 t17 = t1 t8

A list of trees and the corresponding rooted trees showing, in particular, which of these
are superfluous, is given in Table 1.

4 Order conditions for G-symplectic methods

4.1 Overview

The procedure we will adopt to derive the order conditions is to construct the vectors
η and ηD recursively using the asymmetrical product to express formulae for high
order rooted trees in terms of lower order. From (ηD)(t), we will then construct the
order condition related to each t. We will also use the canonical conditions (that is
G-symplectic conditions) to interrelate conditions for t1 t2 and t2 t1 as is done in [12].
To motivate the systematic study of order, we will consider the special case of
p = 3 in Sect. 4.2. One of the conclusions arising from this preliminary investigation
is that the order conditions associated with the non-principal components can always
be satisfied for suitable choices of the non-principal starting series ξ which depend
directly on the values of
B(ηD). These special choices will be considered in the general
case in Sect. 4.3.

123
Order conditions for G-symplectic methods

A second observation arising from the low order case is that for equivalent rooted
trees, such as t3 and t4 , the corresponding order conditions can only be satisfied
together if a so-called “conformability condition” is satisfied. The role of conforma-
bility conditions and their equivalence to what will be called “weak conformability
conditions” will be discussed in Sect. 4.4.
The culmination of this section consists of Theorems 4.5 and 4.6 which will be
presented in 4.5. Finally, in Sect. 4.6, the conditions to order 5 will be summarized.

4.2 Criteria for order 3

In this subsection, ηi will denote η(ti ) for i = 0, 1, 2, 3, 4, with similar meanings for
ξ i , ζi ,
ξi and (ηD)i . Without loss of generality, set ζ1 = 0.
For order 1, we have
(Dη)1 = η0 = 1
(E ξ1 =
ξ )1 =
B(ηD)1 + V ξ1 =
B1 + V ξ1 , (4.1)
T T
(Eζ )1 = 1 = b (ηD)1 = b 1. (4.2)

From (4.1), it follows that )−1


ξ1 = (I − V B1 and from (4.2), it follows that bT 1 = 1.
For order 2, we have

(Dη)2 = η1 = A(Dη1 ) + U ξ 1 = A1 + U ξ1 =: c, (4.3)

(Eξ )2 = ξ2 + ξ1 = B(ηD)2 + V ξ2 = Bc + V
ξ2 , (4.4)
(Eζ )2 = 1
2 + ζ2 = bT (ηD)2 + ζ2 = bT c + ζ2 . (4.5)

The vector c is defined in (4.3), extending the notation for Runge–Kutta methods.
From (4.4), it follows that
ξ2 = (I − V )−1 (
Bc − ξ1 ) and from (4.5), it follows that
bT c = 21 . Because the canonical (G-symplectic) conditions hold, we can conclude that
this last equation is automatically satisfied. Evaluate the quadratic form 1T Q1 where
Q is given from the two sides of (2.6). This gives 2bT A1 = (B1)∗ G B1 which can
be rewritten using (4.3) as 2bT (c − U ξ 1 ) = (B1)∗ G B1. Hence, making use of (2.4)
and (2.5), and the commutativity of the diagonal matrices G and V , we have
2bT c = 2bT U ξ + (B1)∗ G B1
1
T
= 21 (DU )ξ 1 + (B1) ∗ G B1
= 21T (B ∗ GV )ξ + (B1) ∗ G B1
1
B∗G
= 21T ( V )
ξ1 + 1 + ((I − V ξ1 )∗ G(I
) −V )
ξ1
= 1 + ((I − V ) ∗ V
ξ1 ) G(2 + (I − V
)) ξ1
= 1 + ξ ∗ (I − V
1
)∗ G(I
+V ) ξ 1
= 1.

This means that the order 2 condition is automatically satisfied for a G-symplectic
method with order 1.

123
J. C. Butcher, G. Imran

For order 3 we have


(Dη)3 = η12 = c2 ,
(E
ξ )3 =
ξ3 + 2 ξ1 =
ξ2 +
B(ηD)3 + V ξ3 =
Bc2 + V ξ3 , (4.6)
T T 2
(Eζ )3 = + ζ3 + 2ζ2 = b (ηD)3 + ζ3= b c + ζ3
1
3 (4.7)
(Dη)4 = η2 = A(ηD)2 + U ξ 2 = Ac + U ξ 2 ,
(E
ξ )4 = ξ1 =
ξ2 + 1
ξ4 +
B(ηD)4 + V
2 ξ4 =
B(Ac + U ξ 2 ) + V ξ4 , (4.8)
(Eζ )4 = 1
6 + ζ4 + ζ2 = bT (ηD)4 + ζ4 = bT (Ac + U ξ 2 ) + ζ4 . (4.9)

To satisfy (4.6) and (4.8) exactly, it is necessary and sufficient that


)−1 (
ξ3 = (I − V Bc2 − 2
ξ2 − ξ1 ),
) (
ξ4 = (I − V −1
B(Ac + U ξ 2 ) − ξ2 − 21
ξ1 ).

From (4.7) and (4.9), we find using (2.5) and some manipulation,
bT c2 + bT Ac = + 3ζ2 − bT U ξ 2 = 21 + 3ζ2 − 1T DU ξ 2
1
2
= 1
2 + 3ζ2 − 1 B ∗ GV ξ 2 = 21 + 2ζ2 + ξ 1∗ G(I − V )ξ 2 .
T

We can obtain a second formula for this quantity by evaluating the bilinear form 1T Qc,
where Q is given from the two sides of (2.6). The result is

bT Ac + bT c2 = 1D Ac + cT D(A1 + U ξ 1 ) = 1T (D A + AT D)c + cT (DU )ξ 1


= 1T (B ∗ G B)c + cT (B ∗ GV )ξ 1 = 21 + ξ 1∗ G(I − V )ξ 2 + ξ 1∗ Gξ 1 .

The two expressions for bT Ac + bT c2 are identical if and only if


2ζ = ξ ∗ Gξ2 1 1 (4.10)

and this is interpreted as a condition on the starting method which allows order 2
to be compatible with the G-symplectic conditions. Conditions of this type will be
referred to as conformability conditions. The family of these and the closely related
weak conformability conditions are discussed in Sect. 4.4. We will preceed this with
Sect. 4.3, where we will study the relation of the quantities
ξ (t) to the order conditions
for the non-principal components.

4.3 The value of


ξ

As noted in Sect. 2, the order conditions (2.2), (2.3) apply in the diagonal case. In
particular (2.3) are a condition on the non-principal starting method coefficients ξ.
Consider a particular row of B denoted by β T with the corresponding diagonal element
denoted by z and the corresponding component of
of V ξ denoted by ξ . Then these
quantities satisfy the equation
Eξ = β T (ηD) + zξ. (4.11)

123
Order conditions for G-symplectic methods

Because ξ(t0 ) = 0, the operation E is linear acting on a vector made up from ξ(t1 ),
ξ(t2 ), . . . , with as many terms as there are rooted trees to consider. In the case of order
4, this vector and the corresponding vector of β T (ηD) values for these rooted trees
are, shown below, together with the matrix E  representing the linear operator that has
been referred to.
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
ξ(t1 ) βT1 1 0 0 0 0 0 0 0
⎢ ξ(t2 ) ⎥ ⎢ β T η(t ) ⎥ ⎢ 1 1 0 0 0 0 0 0⎥
⎢ ⎥ ⎢ 1 ⎥ ⎢ ⎥
⎢ ξ(t3 ) ⎥ ⎢ β T η(t )2 ⎥ ⎢ 1 2 1 0 0 0 0 0⎥
⎢ ⎥ ⎢ 1 ⎥ ⎢ ⎥
⎢ ξ(t4 ) ⎥ ⎢ ⎥ ⎢ 1 1 0 1 0 0 0 0⎥
 ⎢ ⎥  ⎢ β T η(t2 ) ⎥  ⎢ ⎥.
ξ =⎢ ⎥, β=⎢ ⎥, E =⎢ 2
⎢ ξ(t5 ) ⎥ ⎢ β T η(t1 )3 ⎥ ⎢ 1 3 3 0 1 0 0 0⎥ ⎥
⎢ ξ(t6 ) ⎥ ⎢ T ⎥ ⎢ 1 3 1 1 0 1 0 0⎥
⎢ ⎥ ⎢ β η(t1 )η(t2 ) ⎥ ⎢ ⎥
⎣ ξ(t7 ) ⎦ ⎢ ⎥ 2 2
⎣ 1 1 0 2 0 0 1 0⎦
⎣ β T η(t3 ) ⎦ 3
ξ(t8 ) β T η(t4 )
1 1
6 2 0 1 0 0 0 1

At least in the case of order 4, (4.11) can be rewritten in the form

 − z I )
(E .
ξ =β (4.12)

In the important case r = 2, z = −1, which is known to be sufficient for the con-
struction of methods up to order 4, (4.11) can be written   −1 β
ξ = (I + E)  and, for
convenience, we have
⎡ 1

2 0 0 0 0 0 0 0
⎢ 1 ⎥
⎢−4 1
0 0 0 0 0 0⎥
⎢ 2 ⎥
⎢ 0 − 21 1
0 0 0 0 0⎥
⎢ 2 ⎥
⎢ ⎥
⎢ 0 − 41 0 1
0 0 0 0⎥
 −1
(I + E) =⎢
⎢ 1
2 ⎥.

⎢ 8 0 − 43 0 1
2 0 0 0⎥
⎢ 1 ⎥
⎢ 0 − 41 − 41 0 21 0 0 ⎥
⎢ 16 ⎥
⎢ 1 ⎥
⎣ 24 0 0 − 21 0 0 21 0 ⎦
1
48 0 0 − 41 0 0 0 21

4.4 Conformability

Definition 4.1 The starting method (ζ, ξ ) = ξ is “conformable of order p”, if for
t1 , t2 ∈ T , such that |t1 | + |t2 | ≤ p − 1, it holds that
 ∗
ζ (t1 t2 ) + ζ (t2 t1 ) − ξ (t1 ) Gξ (t2 ) = 0 (4.13)

or, equivalently, that


 ∗
ζ (t1 t2 ) + ζ (t2 t1 ) − ζ (t1 )ζ (t2 ) −
ξ (t1 ) G
ξ (t2 ) = 0. (4.14)

123
J. C. Butcher, G. Imran

Definition 4.2 The starting method (ζ, ξ ) = ξ is “weakly conformable of order p”,
if for t1 , t2 ∈ T , such that |t1 | + |t2 | ≤ p, it holds that
 ∗  ∗
Eζ (t1 t2 ) + Eζ (t2 t1 ) − Eξ (t1 ) GEξ (t2 ) = ζ (t1 t2 ) + ζ (t2 t1 ) − ξ (t1 ) Gξ (t2 )
(4.15)

or, equivalently, that


 ∗
Eζ (t1 t2 ) + Eζ (t2 t1 ) − Eζ (t1 )Eζ (t2 ) − E ξ (t1 ) GE ξ (t2 )
 ∗

= ζ (t1 t2 ) + ζ (t2 t1 ) − ζ (t1 )ζ (t2 ) − ξ (t1 ) G ξ (t2 ). (4.16)

We will show in Theorem 4.3 that starting methods are conformable if and only if they
are weakly conformable.
However, our first task is to find a recursive expression for the order conditions
which we will rewrite in a modified form.
A G-symplectic method has order p if and only if

bT 1 = 1,

BηD(t) = E ξ (t) −
ξ (t), |t| ≤ p,
η(t) = AηD(t) + U ξ (t), |t| ≤ p−1,
T ∗
b ηD(t1 )AηD(t2 ) + (BηD(t1 )) GV ξ (t2 ) = Eζ (t1 t2 ) − ζ (t1 t2 ), |t1 | + |t2 | ≤ p.
(4.17)

Just as for Runge–Kutta methods, these conditions are not independent.


Theorem 4.1 If a G-symplectic method has order p then its starting method is weakly
conformable of order p.

Proof From (4.17) and the same equation with t1 and t2 interchanged we have

Eζ (t1 t2 ) − ζ (t1 t2 ) + Eζ (t2 t1 ) − ζ (t2 t1 )


= bT ηD(t1 )AηD(t2 ) + bT ηD(t2 )AηD(t1 )
+ (BηD(t1 ))∗ GV ξ (t2 ) + (BηD(t2 ))∗ GV ξ (t1 )
= (ηD(t1 ))T (D A + AT D)ηD(t2 )
 
+ (BηD(t1 ))∗ G Eξ (t2 ) − BηD(t2 ) + (V ξ (t1 ))∗ G(Eξ (t2 ) − V ξ (t2 ))
= (ηD(t1 ))T (D A + AT D − B ∗ G B)ηD(t2 )
 
+ (BηD(t1 ))∗ G Eξ (t2 ) + (V ξ (t1 ))∗ G(Eξ (t2 ) − V ξ (t2 ))
 
= (Eξ (t1 ) − V ξ (t1 ))∗ G Eξ (t2 ) + (V ξ (t1 ))∗ G(Eξ (t2 ) − V ξ (t2 ))
 
= (Eξ (t1 ))∗ G Eξ (t2 ) − ξ (t1 )∗ Gξ (t2 ),

which is equivalent to (4.13).


123
Order conditions for G-symplectic methods

For the remainder of this subsection, t1 , t2 , u 1 , u 2 will denote rooted trees. Let

(t1 , t2 ) = Eζ (t1 t2 ) + Eζ (t2 t1 ) − ζ (t1 t2 )


 ∗  ∗
− ζ (t2 t1 ) − Eξ (t1 ) GEξ (t2 ) + ξ (t1 ) Gξ (t2 ),
ϑ(u , u ) = ζ (u u ) + ζ (u u ) − ξ (u )∗ Gξ (u ).
1 2 1 2 2 1 1 2

According to Definitions 4.1 and 4.2, a starting method is weakly conformable of


order p if if (t1 , t2 ) = 0 whenever |t1 | + |t2 | ≤ p and is conformable of order p if
ϑ(u 1 , u 2 ) = 0 whenever |u 1 | + |u 2 | ≤ p − 1. Before proving the equivalence of these
two concepts, we introduce a preliminary Lemma.
Lemma 4.2

(t1 , t2 ) = C(t1 , t2 , u 1 , u 2 )ϑ(u 1 , u 2 ), (4.18)
u 1 ≤t1 ,u 2 ≤t2

where C(t1 , t2 , u 1 , u 2 ) = 0 when u 1 = t1 , u 2 = t2 and is non-negative otherwise.


Proof If u ≤ t1 t2 then either u = u 1 u 2 for u 1 ≤ t1 , u 2 ≤ t2 or u = u 1 ≤ t1 . Hence,

(Eζ )(t1 t2 ) = ζ (u 1 u 2 )E(t1 − u 1 )E(t2 − u 2 )
u 1 ≤t1 ,u 2 ≤t2

+ ζ (u 1 )E(t1 − u 1 )E(t2 ) + E(t1 t2 ) (4.19)
u 1 ≤t1

Also
 ∗
Eξ (t1 ) GEξ (t2 )
 ∗  
 
= ξ (u 1 )E(t1 − u 1 ) + e1 E(t1 ) G ξ (u 2 )E(t2 − u 2 ) + e2 E(t2 )
u 1 ≤t1 u 2 ≤t2
 ∗  
 
= ξ (u 1 )E(t1 − u 1 ) G ξ (u 2 )E(t2 − u 2 ) + E(t1 )E(t2 )
u 1 ≤t1 u 2 ≤t2
 
+ ζ (u 2 )E(t2 − u 2 )E(t1 ) + ζ (u 1 )E(t1 − u 1 )E(t2 ).
u 2 ≤t2 u 1 ≤t1

Subtract this result from (4.19) plus (4.19) with t1 and t2 interchanged. The result is

(t2 , t2 ) = E(t1 − u 1 )E(t2 − u 2 )ϑ(u 1 , u 2 )
u 1 ≤t1 ,u 2 ≤t2
 
+ E(t1 t2 ) + E(t2 t1 ) − E(t1 )E(t2 ) .

We obtain the required result with C(t1 , t2 , u 1 , u 2 ) = E(t1 − u 1 )E(t2 − u 2 ), where


we note that E(t1 t2 ) + E(t2 t1 ) = E(t1 )E(t2 ).

123
J. C. Butcher, G. Imran

Theorem 4.3 A starting method is conformable with order p if and only if it is weakly
conformable with order p.

Proof The “only if” part of the result follows from Lemma 4.2. To show that con-
formability follows from weak conformability, we can assume that |t1 t2 | = p in (4.18)
and we can assume that ϑ(u 1 , u 2 ) = 0 unless |u 1 u 2 | = p − 1, because lower values
of |u 1 u 2 | would have already been covered by lower values of p by the induction
principle. For a given u 1 , u 2 pair written as

u 1 = [u 11 u 12 · · · τ m 1 ],
u 2 = [u 21 u 22 · · · τ m 2 ],

where each of u 11 , u 12 , u 21 , u 22 , · · · is not equal to τ , define the function g(u 1 , u 2 ) =


max(m 1 , m 2 ). If g(u 1 , u 2 ) = p−3, the highest possible value, then necessarily u 1 = τ
and u 2 = [τ p−3 ], or vice versa. In this case we can choose t1 = τ, t2 = [τ p−2 ]. In
this case,
(τ, [τ p−2 ]) = ( p − 2)ϑ(τ, [τ p−3 ]).

We now carry out induction on decreasing values of g(u 1 , u 2 ). Choose t1 , t2 so that


g(t1 , t2 ) = 1+ g(u 1 , u 2 ), |t1 t2 | = 1+|u 1 u 2 | with t1 = u 1 , t2 = u 2 τ or t1 = u 1 τ, t2 =
u 2 . It is found that
(t1 , t2 ) = g(t1 , t2 )ϑ(u 1 , u 2 ),

where terms, which are zero by the induction hypothesis, have been omitted. This
completes the “if” part of the proof.

Table 2 illustrates the selection of t1 and t2 required for the proof of ϑ(u 1 , u 2 ) = 0.

Table 2 Illustrating the choice of (t1 , t2 ) for given (u 1 , u 2 ) in the proof of Theorem 4.3
g(u1 ,u2 ) u1 u2 g(t1 ,t2 ) t1 t2
0 1
1 2
2 3
1 2
0 1
3 4
2 3
1 2

1 2

0 1

0 1

123
Order conditions for G-symplectic methods

4.5 The main results

The principal result is that if and only if conformability holds, the order conditions
for t1 t2 and t2 t1 are both satisfied or neither is satisfied. There is a related result in
the case for the superfluous rooted tree t1 t1 . In addition, we will present a convenient
expression in Lemma 4.7, for the order condition associated with a rooted tree written
in the form t1 t2 .
Before the results on the relationship between conformability and order we establish
a preliminary lemma.
Lemma 4.4 For a G-symplectic method with a specific starting method and given t1
and t2 ,
bT (ηD)(t t ) + bT (ηD)(t t ) = (Eξ )(t )∗ G(Eξ )(t ) − ξ(t )∗ Gξ(t ). (4.20)
1 2 2 1 1 2 1 2

Proof We have
bT (ηD)(t1 t2 ) = bT (ηD)(t1 )η(t2 )
= bT (ηD)(t1 )A(ηD)(t2 ) + bT (ηD)(t1 )U ξ(t2 )
= (ηD)(t1 )∗ D A(ηD)(t2 ) + (ηD)(t1 )∗ DU ξ(t2 ).

Add a similar result for bT (ηD)(t1 t2 ) and use the facts that D A + AT D = B ∗ G B and
DU = B ∗ GV and we find
bT (ηD)(t1 t2 ) + bT (ηD)(t2 t1 )
= (B(ηD)(t1 ))∗ G B(ηD)(t2 ) + (BηD)(t1 ))∗ V ξ(t2 ) + (V ξ(t1 ))∗ G BηD)(t2 )
 ∗  
= (Eξ )(t1 ) − V ξ(t1 ) G (Eξ )(t2 ) − V ξ(t2 )
 ∗  
+ (Eξ )(t ) − V ξ(t ) GV ξ(t ) + V ξ(t )∗ G (Eξ )(t ) − (V ξ(t ))
1 1 2 1 2 2
= (Eξ )(t1 )∗ G(Eξ )(t2 ) − (V ξ(t1 ))∗ GV ξ(t2 )
= (Eξ )(t )∗ G(Eξ )(t ) − ξ(t )∗ Gξ(t ).
1 2 1 2

Theorem 4.5 A G-symplectic general linear method has order p with respect to a
given starting method only if the starting method is conformable of order p.

Proof Assume |t1 t2 | ≤ p add the order conditions associated with t1 t2 and t2 t1
(bT (ηD)(t1 t2 )−(Eζ )(t1 t2 ) + ζ (t1 t2 ))+(bT (ηD)(t2 t1 )−(Eζ )(t2 t1 )+ζ (t2 t1 )) = 0.
(4.21)

Substitute from (4.20) and the weak conformability, and hence the conformability,
condition of order p follows.

Theorem 4.6 If a G-symplectic method, with a given starting method, has order at
least p − 1 and the starting method is conformable of order p, then it satisfies the

123
J. C. Butcher, G. Imran

order condition for t1 t2 , where |t1 t2 | = p if and only if it satifies the order condition
for t2 t1 .

Proof Assume that one and only one of the two order conditions is satisfied. Then the
equality (4.21), becomes inequality and conformability is contradicted.

Finally in this subsection, we establish a convenient form for the order condition
associated with t1 t2 .

Lemma 4.7 Assuming that the order conditions up to order p − 1 are satisfied, the
order condition

bT (ηD)(t1 t2 ) = (Eζ )(t1 t2 ) − ζ (t1 t2 ), |t1 t2 | = p,

is equivalent to

bT (ηD)(t1 )A(ηD)(t2 ) = (Eζ )(t1 t2 ) − ζ (t1 t2 ) − (Eζ (t1 ) − ζ (t1 ))ζ (t2 )
+
ξ (t1 )G ξ (t2 ) − (E ξ (t1 ))∗ G
V
ξ (t2 ). (4.22)

Proof We have

bT (ηD)(t1 t2 ) = bT (ηD)(t1 )η(t2 )


= bT (ηD)(t1 )(A(ηD)(t2 ) + U ξ (t2 ))
= bT ηD)(t1 )A(ηD)(t2 ) + ηD)(t1 )∗ DU ξ (t2 )
= bT ηD)(t1 )A(ηD)(t2 ) + ηD)(t1 )∗ B ∗ GV ξ (t2 )
= bT ηD)(t )A(ηD)(t ) + (B(ηD)(t )∗ GV ξ (t )
1 2 1 2

= b ηD)(t1 )A(ηD)(t2 ) + (Eξ (t1 ) − V ξ (t1 ))∗ GV ξ (t2 ),


T

from which (4.22) follows.


4.6 Summary of conditions to order 5

ξ (ti )∗ G
It will be convenient to define X i j = ξ (t j ). Hence the conformability condi-
tions can be written in the form

ζ (ti t j ) + ζ (t j ti ) = ζ (ti )ζ (t j ) + X i j .

It will also be convenient to define X i j = ξ (ti )∗ GV


ξ (t j ). Table 3 gives a summary of
the conditions to order 5, together with some explanatory notes. Note that rooted trees
of order 5 are designated as t9 = t5 t1 , t10 = t6 t1 , t11 = t7 t1 , t12 = t8 t1 , t13 = t4 t2 ,
t14 = t1 t5 , t15 = t1 t6 , t16 = t1 t7 , t17 = t1 t8 .

123
Order conditions for G-symplectic methods

Table 3 Summary of conditions to order 5

p Condition (n) n Notes

1, 2 bT 1 = 1 1 This is identical with the consistency


condition for the method
3 ζ2 = 21 X 11 2 This is the conformability condition
ϑ(t1 , t1 ) = 0
bT c2 = 13 + X 11 3 The order condition bT (ηD)3 = (Eζ )3 − ζ3
with (2) substituted
4 ζ3 = 13 bT c3 − 12
1 − 1X
2 11 4 From bT (ηD)5 = (Eζ )5 − ζ5
ζ4 = X 12 − ζ3 5 From ϑ(t1 , t2 ) = 0
5 ζ5 = 14 bT c4 − 20
1 − 1 X − 3ζ
2 11 2 3 6 From bT (ηD)9 = (Eζ )9 − ζ9
ζ7 = X 13 − ζ5 7 From ϑ(t1 , t3 ) = 0
ζ6 = 18 X 11
2 + 1X
2 22 8 From ϑ(t2 , t2 ) = 0
ζ8 = X 14 − 18 X 11
2 − 1X
2 22 9 From ϑ(t1 , t4 ) = 0
b c Ac = 10 + 6 X 11 − 41 X 11
T 2 1 5 2 + 3ζ + X
2 3 12 10 by eliminating ζ5 from (6) and (10) a single
 − 2X  − X 
+ζ5 + X 22 + X 23 − X 12 order condition is found giving a weighted
22 23
sum of bT c4 and bT c2 Ac in terms of other
quantities
1 + 5 X +X +X +X
bT (Ac)2 = 20 11 From bT (ηD)13 = (Eζ )13 − ζ13
12 11 12 22 24
 − X − X
− 21 X 12 22 24

5 Derivation of methods with order 4

5.1 The coefficient matrices

We will restrict ourselves to methods with pqr s = 4123, because it was shown in
[8] that methods with this choice exist. Furthermore, we will write the method in
diagonal form V = diag(1, −1). For convenience of implementation we will require
A to be lower triangular. The matrix G in the definition of the G-symplectic property
will be chosen to be G = diag(1, −1). The fact that g22 is negative will give the
opportunity to force at least one of the diagonal elements of A to vanish and this will
be an additional cost advantage in implementations. Note also that G being a diagonal
matrix is required for consistency with (2.4).
Write the elements in the second column of U as x1 , x2 , x3 and the conditions (2.5)
and (2.6) make it possible to write the full (A, U, B, V ) matrix as

⎡ ⎤
2 b1 (1 − x 1 )
1 2 0 0 1 x1
  ⎢ b1 (1 − x1 x2 ) 1 b2 (1 − x 2 ) 0 1 x2 ⎥
AU ⎢ 2 2 ⎥
= ⎢ b1 (1 − x1 x3 ) b2 (1 − x2 x3 ) 1 b3 (1 − x 2 ) x3 ⎥
B V ⎢ 2 3 1 ⎥.
⎣ b1 b2 b3 1 0 ⎦
b1 x1 b2 x2 b3 x3 0 −1

123
J. C. Butcher, G. Imran

The values of
ξ1 , ζ2 and the vector c are found as


ξ1 = 21 bT x,
ζ2 = − 1 ξ12 ,
2
c = A1 +
ξ1 x.

Treating
ξ1 as a parameter, the vector b is required to satisfy the equations

bT 1 = 1, (5.1)
bT x = 2
ξ1 , (5.2)
T 2
b x = 0, (5.3)
bT c2 = 1
3 −
ξ12 . (5.4)

Equations (5.1), (5.2) and (5.4) are order conditions and can be read from Table 3 while
(5.3) is the condition for zero parasitism growth factors. When the components of b are
solved from the linear Eqs. (5.1), (5.2) and (5.3) and the solution substituted into (5.4),
a complicated equation in x1 , x2 , x3 ,
ξ1 is found. An exhaustive search was carried
out to identify suitable choices of the x components which gave a rational solution
for
ξ1 . The rules on the x components was that when any of these had a rational value

Table 4 Methods based on selected solutions to (5.1), (5.2), (5.3) and (5.4)

A U B c
⎡ ⎤ ⎡ ⎤
0 0 0 ⎡ ⎤   1
1 1 1 − 38 25 ⎢ 45 ⎥
⎢2 0 0⎥ ⎣ 1 −1 ⎦ 3 24 ⎢ ⎥
4123A ⎣3 ⎦ 1 3 5 ⎣ 12 ⎦
2 3 1 1 − 15 − 24
5 − 10 2
3 8 11
⎡ ⎤ ⎡ 20 ⎤
1 0 0 ⎡ 1
⎤   9
⎢ 2 ⎥ 1 25 ⎢ 20 ⎥
5 − 38 1
4123B ⎢ 5 0 0⎥ ⎣1 1⎦ 24 3 ⎢ 7 ⎥
⎣ 6 ⎦ 5 ⎣ 12 ⎦
5 1 −1 24 − 38 − 13 3
4 − 43 0 4
⎡ ⎤
1 0 0 ⎡ 1⎤   ⎡1⎤
⎢ 3 ⎥ 1 3 3 3 − 81 3
4123C ⎢ 5 1 0⎥ ⎣1 −1 ⎦ 4 8 ⎣1⎦
⎣ 6 6 ⎦ 3 1 − 81 − 81
1 1 0 1 1 4 1
2 2
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
0 0 0 1 1   0
⎢ −1 − 81 3 3
4123D ⎣ 6
1
6 0⎥⎦
⎢1 −1 ⎥
⎣ 3⎦
8 4 ⎣0⎦
1 − 81 − 81 1 2
−1 5 1 1 3
4 3
⎡ 12 12 3
⎤ ⎡ ⎤
1 0 0 ⎡ 1⎤   1
⎢ 41 ⎥ 1 2 2 − 31 2 ⎢ 6 ⎥
4123E ⎢ 0 0⎥ ⎣1 1⎦ 3 3 ⎢ 1 ⎥
⎣3 ⎦ 1 − 31 − 31 ⎣ 6 ⎦
5 − 21 1 1 − 21 3 2
6 4 3
⎡ ⎤
⎡1 ⎤ ⎡ ⎤   1
4 0 0 1 − 21 2 − 31 2 ⎢ 3 ⎥
4123F ⎣1 0 0⎦ ⎣1 1⎦ 3 3 ⎢ 5 ⎥
− 31 − 31 1 ⎣ 6 ⎦
5 − 16 1 1 1 3
6 4 2 5
6

123
Order conditions for G-symplectic methods

n/d it was regarded as acceptable only if max(|n|, |d|) ≤ 8. Furthermore, one of the
x components was required to be equal to 1, leading to at least one zero diagonal in
A. A selection of the solutions satisfying these requirments are listed in Table 4. The
solutions and the full methods based on these are given names such as 4123A, 4123B,
etc.
These six methods are selected from the many choices available based on the relative
simplicity of the rational entries in the defining matrices. However, methods 4123A
and 4123B are interesting in that, in each case, the matrix A has only a single non-
zero element on the diagonal. If the methods give acceptable numerical results then
they could possibly have efficiency advantages. The value of ξ1 is zero in the case of
4123C and 4123D and this will possibly give a simplicity in the structure of starting
and finishing methods. We will continue this section by discussing the B-series for the
principal component.

5.2 The values of ζ2 , ζ3 and ζ4

We recall the covenient assumption ζ1 = 0. From the value of ξ1 , we can infer


ζ2 = − 21 ξ12 . We are now in a position to calculate
ξ2 using the results in Sect. 4.3. The
first two components of the vector β  are bT x and bT xc and hence, using the first two

rows and columns of E, we find

    

ξ1 1
0 bT x
= 2
ξ2 − 41 1
2 bT xc

and hence ξ2 = 21 bT x(c − 21 1). Now calculate ζ3 and ζ4 from conditions 4 and 5 in
Table 3. The results are

ζ3 = 13 bT c3 − 12
1
+
ξ12 /2,
ζ4 = −ζ3 − ξ1
ξ2 .

The values of ζ (ti ), i = 5, 6, 7, 8 do not have any roles in the fourth order conditions
and hence, there values are arbitrary. The final step in the specification of methods
will be the determination of the values of ξ (ti ), i = 3, 4, 5, 6, 7, 8.

5.3 The values of


ξ (t) for |t| ≤ 4

Using the known values of ζ2 and


ξ2 we compute

η2 = Ac + ζ2 1 +
ξ2 x. (5.5)

We now know the first four entries in β  and we deduce the third and fourth entries in

ξ . Having evaluated ξ3 and ξ4 , we find

123
J. C. Butcher, G. Imran

η3 = Ac2 + ζ3 1 +
ξ3 x,
η4 = Aη2 + ζ4 1 +
ξ4 x.

Finally, all components of β  + I )−1 gives 


 can be found and a multiplication by ( E ξ.

5.4 Starters and finishers for sample methods

From the methods in Table 4, we will select 4123A and 4123C for further analysis.
For method 4123A, it is found that


ζ1 ζ2 ζ3 ζ4 = 0 − 32
1
− 4320
7
8640 ,
149

ξ1
ξ2
ξ3
ξ4
ξ5
ξ6
ξ7
ξ8 = 41 − 16
1
− 960
49
− 384
13 2543 193 619 163 .
57600 7680 34560 69120

We will construct a Runge–Kutta finishing method which, to order 3, gives a B-series


ζ −1 . That is, the Runge–Kutta finisher will satisfy the order conditions bT 1 = 0, bT c =
−ζ2 , bT c2 = −ζ3 , bT Ac = −ζ4 and a possible tableau is

0
1 1
2 2
. (5.6)
1 − 121
28 149
121
− 4320
391 16
135 − 4320
121

The series ζ −1 and ζ are now found to order 4:

1 7
ζ1−1 ζ2−1 ζ3−1 ζ4−1 ζ5−1 ζ6−1 ζ7−1 ζ8−1 = 0 32 4320 − 8640 − 1440 − 8640 − 17280 0 ,
149 19 149 149

ζ1 ζ2 ζ3 ζ4 ζ5 ζ6 ζ7 ζ8 = 0 − 32 1
− 4320
7 149 19 2519 149 1
8640 1440 138240 17280 1024
.

A possible starter for the principal component is given by the Runge–Kutta method

0 0 0 0 0
1 1
3 3 0 0 0
2 1643 135
3 3072 1024 0 0 , (5.7)
1 − 2194976 111943776 3498243 548744
812303 118383361 1001284 14445

47
480 − 15
1
− 15
1 17
480

found by solving order conditions

(ti ) = ζi , i = 1, 2, . . . , 8.

123
Order conditions for G-symplectic methods

For the non-principal component, instead use the order conditions (ti ) =
ξi , and a
possible tableau is

0 0 0 0 0
1 1
3 3 0 0 0
2 3731 2265
3 15789 5263 0 0 . (5.8)
1 − 83170080
80098847 324490781
83170080 − 256765981
83170080
61207
53280
0 − 12800
823 15129
12800 − 12800
15789 4683
12800

For method 4123C, it is found that


ζ1 ζ2 ζ3 ζ4 = 0 0 108
1
− 108
1
,

ξ1
ξ2
ξ3
ξ4
ξ5
ξ6
ξ7
ξ8 = 0 − 12 − 36 − 72
1 1 1 5 1 1 1 .
108 48 108 216

A suitable starter for the principal component, together with a finisher to be applied
to the principal component, are respectively

0 0
1 1 1 1
2
1 2 −1
2 , 1
2 2
2 −1 . (5.9)

54 − 27 − 54 − 54
1 1 1 1 1 1
54 27

The starting method for the 4123C method is found in a similar way to the 4123A
method. A suitable tableau is

0 0 0 0 0
1 1
3 3 0 0 0
2 1 5
3 4 12 0 0 . (5.10)
1 − 18 35
24 − 58 7
24
0 0 2
1
− 43 4
1

6 Numerical simulations

The experiments reported here are based on the Kepler problem. A comparison is
made between the performance of methods 4123A and 4123C compared with the
symplectic Runge–Kutta method, [6] and [13, pp 102, 166]:

123
J. C. Butcher, G. Imran

4123A 4123B R

Fig. 1 Simulation of Kepler problem for methods 4123A, 4123C and R respectively with e = 0.1, h =
2π/256 after 0, 3,000, 6,000 and 9,000 orbits

0.0002 0.0002 0.002

0 0 0

−0.0002 −0.0002 −0.002

0 2π 0 2π 0 2π

Fig. 2 Deviations from exact orbits for methods 4123A, 4123C and R respectively with e = 0.1, h =
2π/256 after 3,000 (dotted lines), 6,000 (dashed lines) and 9,000 (full lines) orbits

θ θ
2 2 0 0
1
θ 21 − θ 0
2
θ ,
1− 2 θ 1 − 2θ θ2
θ 1 − 2θ θ

where θ = (2 − 21/3 )−1 .


The first experiment is the simulation of a large number of orbits using eccentricity
e = 0.1 and stepsize h = 2π/256. The results for this experiment are shown in Fig. 1.
It is difficult to judge the accuracy of the three methods from this figure, because
of the almost overlapping of the orbits. Hence, to express more detailed information,
Fig. 2 is presented. This shows the deviation from the exact orbit written in polar
coordinates.
To appreciate the close adherence of the Hamiltonian to its initial value, Fig. 3 is
shown. For the same problems and with stepsize h = 2π/256 and eccentriciy e = 0.1,
the deviation of the Hamiltonian is shown after 1,000 orbits. The calculations were
also carried out for 3,000, 6,000 and 9,000 orbits but the results are almost identical.
This remark applies to each of the three methods considered.

123
Order conditions for G-symplectic methods

3 × 10−8 3 × 10−8 3 × 10−8

2 × 10−8 2 × 10−8 2 × 10−8

10−8 10−8 10−8

0 0 0

−10−8 −10−8 −10−8

0 2π 0 2π 0 2π

Fig. 3 Deviations of the Hamiltonian from its initial value for methods 4123A, 4123C and R respectively,
for the Kepler problem with e = 0.1, h = 2π/256 after 9,000 orbits

2 × 10−8 2 × 10−8

0 0

−2 × 10−8 −2 × 10−8

0 2π 0 2π

Fig. 4 Deviations of angular momentum from its initial value for methods 4123A and 4123C respectively
for the Kepler problem with e = 0.1, h = 2π/256 after 1,000 orbits

Finally, we consider the deviation of the angular momentum from its initial value,
and the results are presented in Fig. 4. For the same problem and with stepsize h =
2π/256 and eccentriciy e = 0.1, the deviation of the angular momentum is shown after
1,000 orbits. As for the Hamiltonian calculations, the results were almost identical for
3,000, 6,000 and 9,000 orbits. For the Runge–Kutta method, angular momentum is
exactly conserved but these results are not included in the figure because the deviations
are down to the level of round-off error.

7 Conclusions

The order requirements for parasitism-free G-symplectic methods lead to a set of


conditions related to individual rooted trees together with a set of conformability
conditions which interrelate equivalent rooted trees. The results are summarized up to
order 5 and specific methods are found for order 4. These give satisfactory results in
numerical simulations.

123
J. C. Butcher, G. Imran

Acknowledgments The authors acknowledge the support of the Marsden Fund. Furthermore, they thank
the referees for their valuable comments.

References

1. Butcher, J.C.: The order of numerical methods for ordinary differential equations. Math. Comp. 27,
793–806 (1973)
2. Butcher, J.C.: General linear methods. Acta Numer. 15, 157–256 (2006)
3. Butcher, J.C.: Numerical methods for ordinary differential equations, 2nd edn. Wiley, New York (2008)
4. Butcher, J.C.: Dealing with parasitic behaviour in G-symplectic integrators, in recent developments in
the numerics of nonlinear hyperbolic conservation laws. Springer, Heidelberg (2013)
5. Butcher, J.C., Habib, Y., Hill, A.T., Norton, T.J.T.: The control of parasitism in G-symplectic methods.
SIAM J. Numer. Anal. 52, 2440–2465 (2014)
6. de Frutos, J., Sanz-Serna, J.M.: An easily implementable fourth-order method for the time integration
of wave problems. J. Comput. Phys. 103, 160–168 (1992)
7. Hairer, E., Lubich, C., Wanner, G.: Geometric numerical integration, 1st edn. Structure-preserving
algorithms for ordinary differential equations. Springer, Berlin (2003)
8. Imran, G.: Accurate and efficient methods for differential systems with special structures. PhD. Thesis,
University of Auckland, NZ
9. Lopez-Marcos, M., Sanz-Serna, J.M., Skeel, R.D.: Cheap enhancement of symplectic integrators. In:
Griffiths, D.F., Watson G.A. (eds.) Numerical Analysis, pp.107–122 (1996)
10. Sanz-Serna, J.M.: Runge–Kutta schemes for Hamiltonian systems. BIT 28, 877–883 (1988)
11. Sanz-Serna, J.M.: Symplectic integrators for Hamiltonian problems. Acta Numer. 1, 243–286 (1992)
12. Sanz-Serna, J.M., Abia, L.: Order conditions for canonical Runge–Kutta schemes. SIAM J Numer.
Anal. 28, 1081–1096 (1991)
13. Sanz-Serna, J.M., Calvo, M.P.: Numerical Hamiltonian problems, 1st edn. Chapman and Hall, London
(1994)

123

You might also like