Vous êtes sur la page 1sur 17

Appendix H

Runge-Kutta Methods
In this appendix we will analyze the conditions on the coecients of
an explicit Runge-Kutta Method that are necessary and sucient to
guarantee convergence with accuracy of order . In particular, we
will establish the connection between these conditions and the set of
rooted trees with no more than nodes. As a consequence, we will
be able to show that there are -stage methods of order for 4
but not for > 4.
We begin by briey considering more general one-step methods,
+1 = ( , , , ), for approximating solutions of the scalar ODE
= (, ()).
The local truncation error at is the quantity dened by
(+1 ) := ( , ( ), , ) + .
From our discussion following the convergence analysis of Eulers
Method in the body of the text, we can show that a 0-stable one-step
method will converge to a solution of the ODE +1 [ , + ]
with global order of accuracy if +1 for some > 0
depending only on max( , + ) ( +1) ().
One approach to constructing methods satisfying such an esti-
mate is to dene them using Taylors Theorem with Remainder by
()
letting ( , , , ) = =0 ( )/! be the Taylor polyno-
mial of degree for () centered at and evaluated at +1 . To
implement this idea, we must be able to express () ( ) in terms
of and its derivatives evaluated at ( , ). The resulting one-step
methods are known as Taylor Methods. Taylor Methods are an option

263
264 H. Runge-Kutta Methods

if the vector eld that denes the ODE is given in a form that can be
dierentiated symbolically, which is not always the case.
To demonstrate how this would be carried out, and for later use,
we examine the expressions for the rst few derivatives of in terms
of and its derivatives. We ignore the dierential equation at rst
and dierentiate () = (, ()) and use multi-index notation for
mixed partial derivatives,
+
, = , (, ) = .

In this form, we can distinguish terms arising from dierentiating
from those that arise by dierentiating factors of coming from the
chain ruleterms that we will eventually also write in terms of .
Because of equality of mixed partial derivatives, these terms exhibit
a binomial pattern,
= ,
= 1,0 + 0,1 , (H.1)
2,0 1,1 2 0,2 1,0 0,1 0,1
= [ + 2 + ] + [( + ) ].

Even when this procedure is possible, by hand or with automatic


symbolic dierentiation, the number of terms required to carry the ex-
pansion to high order can yield diminishing returns with the growing
cost of evaluation.
An alternate approach originally proposed and developed by
Runge and Kutta only requires evaluation of at arbitrary (, ) val-
ues to match the terms of Taylor polynomial above to order . Runge-
Kutta Methods approximate (y(+1 ) y( ))/ using a weighted
average of samples of the vector eld (, ) that denes an ODE.
For the method to be explicit, locations of the samples must be cho-
sen based upon information obtained in previous samples. Because of
this, the general form of an explicit one-stage Runge-Kutta Method is
+1 = + 0 ( , ). For the right-hand side to match the rst-
order terms of the Taylor expansion above, we must have 0 = 1. This
tells us that Eulers Method is the unique explicit one-stage Runge-
Kutta Method that is convergent. No higher-order terms occur when
a one-stage method is used.
H. Runge-Kutta Methods 265

The general form of an explicit two-stage Runge-Kutta Method


is

,1 = ( , ),

,2 = ( + 21 , + 21 ,1 ), (H.2)

+1 = + (1 ,1
+ 2 ,2 ).
Two of the example methods in the text t this pattern, the midpoint
method (12 = 1/2, 1 = 0, 2 = 1) and Heuns Method (12 =
1, 1 = 2 = 1/2). Both solved the second-order accuracy model
problem exactly and also appeared to converge to the solution of the
absolute stability model problem with second-order accuracy.
To estimate the local truncation error of these methods, we per-
form Taylor expansions of the terms of the general explicit 2-stage

Runge-Kutta Methods. Substituting ,1 in the denition of ,2 ,

,2 = ( + 21 , + 21 ).
Then by Taylor expanding in powers of the perturbations (to rst
order to obtain terms to second order),

,2 = + 21 ( 1,0 + 0,1 ) + (2 ).
When this is inserted in the expression for +1 , we nd
2
+1 = + (1 + 2 ) + 22 21 ( 1,0 + 0,1 ) + (3 ).
2
Comparing this with (H.1), the conditions for this expansion to match
the rst two terms of the Taylor series
2
(+1 ) = ( ) + + + (3 )
2
are
1 + 2 = 1,
22 21 = 1.
We may use 2 to parametrize a family, 1 = 1 2 , 21 = 1/(22 ),
of solutions of these equations. It is straightforward to check that the
midpoint method and Heuns Method satisfy these conditions.
266 H. Runge-Kutta Methods

The parameters of a Runge-Kutta Method are often displayed in


the form of a so-called Butcher tableau:

2 21
.. .. ..
. . .

1 (1)

1 2
The Butcher tableau for the midpoint method is

0
1 1
2 2

0 1
The modied trapezoidal method is displayed in this format as

0
1 1
1 1
2 2


If we expanded ,2 to higher order, we would discover that three
parameters do not provide enough freedom to obtain a method of
order 3. In order to satisfy the two additional 3 conditions appearing
in square brackets in the expression (H.1) for , another stage is
needed.
The form of an explicit Runge-Kutta Method with = 3 stages
is

,1 = ( , ),

,2 = ( + 21 , + 21 ,1 ),

,3 = ( + (31 + 32 ), + (31 ,1 + 32 ,2 )),

+1 = + (1 ,1 + 2 ,2 + 3 ,3 ).
H. Runge-Kutta Methods 267

4
The coecient of 4! in the Taylor expansion of ( + ) in terms of
and its derivatives is
(4) = [ 3,0 + 3 2,1 + 3 2 1,2 + 3 0,3 ]
+ [3( 1,0 + 0,1 )( 1,1 + 0,2 )]
+ [( 2,0 + 2 1,1 + 2 0,2 + ( 1,0 + 0,1 ) 0,1 ) 0,1 ].

It clearly becomes worthwhile to nd a framework to simplify the


development and comparison of the Taylor and Runge-Kutta sides
of these expansions to higher orders. The autonomous scalar case
is exceptional, as can be seen by setting all -derivatives to zero in
the expressions above. For greater generality, we shift our setting
and notation and now consider an R vector-valued f (y) and y()
that is a solution of y = f (y). The nonautonomous case can be put
into this form using the standard device of replacing by additional

dependent variables +1 satisfying +1 = 1. In this setting, the
general -stage explicit Runge-Kutta Method takes the form



y+1 = y + y, , (H.3)
=1

where
1


y, = f (y, ), with y, = y + , y, . (H.4)
=1

An elegant formalism for organizing, visualizing, and understand-


ing both the Taylor expansion of the solution y( +) and the Runge-
Kutta expansion of y+1 obtained by Taylor expanding the terms in
(H.3) and (H.4) has been developed and advocated by Butcher [BJ],
following on the work of Gill [GS] and Merson [MRH]. This approach
associates terms in both expansions with rooted trees. To motivate
it, we begin by reviewing the formal Taylor expansion to degree 5 for
a function y() : R R satisfying y = f (y), where f is a smooth
function from R to R ,
2
y( + ) = y() + y () + y () + + y() () + ,
2! !
268 H. Runge-Kutta Methods

and then give the representation of its terms using rooted trees. We
wish to represent the derivatives y() () in terms of f and its deriva-
tives. Here, in the column on the left, we list successive derivatives
of the function g() = f (y()) where again at rst we ignore the dif-
ferential equation:
g() = f (y()) = f , y , 1,
g () = fy y , y , 1,
g () = fyy y2 + fy y , y , 1 + 1 = 2,
g () = fyyy y3 + 3fyy y y + fy y , y(4) , 1 + 1 + 2 = 4,
g(4) ()= fyyyy y4 + 6fyyy y2 y
+3fyy y y
+4fyy y y + fy y(4) , y(5) , 1 + 1 + 1 + 2 + 4 = 9.
In the column on the right, we list the correspondence between each
derivative of g and the next higher derivative of . We also list the
number of terms that each row represents as a sum. The terms of
the sum refer recursively to terms from previous rows that appear
in subsequent rows and the numbers of terms they represent. Recall
that the th derivative of f with respect to y is a symmetric -linear
function from (R ) R . For > 1, the symmetry is nontrivial
and decreases the number of its independent coecients with respect
to a basis from +1 accordingly. For example, when = 2 there are
(( + 1))/2 independent components.
Next we expand the rows recursively to write the Taylor expan-
sion as a linear combination of elementary dierentials. These are
multilinear operator compositions that express y() in terms of f and
its derivatives evaluated at . Since fy...y is an operator with ar-
guments, we will use a naturally related notation for lists that will
be familiar to those who have encountered the articial intelligence
programming language LISP. For our purposes, a list begins with an
open parentheses and the rst element, the th partial derivative of
with respect to for some 0, followed by sublists, then a close
parentheses. If a sublist has zero sublists, we omit its parentheses,
and we can close all open parentheses with a right square bracket.
(Our convention will be to do so when more than two are open.)
H. Runge-Kutta Methods 269

The internal representation of such a list in a LISP interpreter is


in the form of a rooted tree data structure, the same algebraic struc-
ture that has been used to visualize and organize the terms of Taylor
and Runge-Kutta expansions and their relationship. This suggests
that LISP may be convenient for performing calculations involved in
the derivation and analysis of Runge-Kutta Methods. A rooted tree
is a set of nodes connected by edges oriented away from a distin-
guished node called the root, so it is a connected simple graph that
contains no cycles, i.e., a tree. Graphically, we represent the lists
associated with an elementary dierential by a root node labeled by
fy...y ( partial derivatives) attached to subtrees corresponding to
those sublists. Lists with no sublists are leavesterminal nodes with
no edges leaving them:
y () = (f ) = f ,
y () = (fy f ),
y () = (fyy f f ) + (fy (fy f )),
y(4) () = (fyyy f f f ) + 3(fyy (fy f ) f ) + (fy (fyy f f )) + (fy (fy (fy f ],
y(5) () = (fyyyy f f f f ) + 6(fyyy (fy f ) f f ) + 3(fyy (fy f )(fy f ))
+ 4(fyy (fyy f f ) f ) + 4(fyy (fy (fy f )) f ) + (fy (fyyy f f f ))
+ 3(fy (fyy (fy f ) f )) + (fy (fy (fyy f f ] + (fy (fy (fy (fy f ].
We implicitly evaluate f and its derivatives at y(). We will refer to
the th term of the th-order (row) formal Taylor expansion above as
. Observe that each dierent term of a particular order arises from
terms of the previous order from the vector-valued Leibniz rule and
chain rule, through the addition of one derivative to each operator
factor (we consider arguments f as preceded by a 0th-order identity
operator) and adding a corresponding argument y = f . In terms of
rooted trees, this corresponds to the process of constructing all rooted
trees with nodes by attaching a new edge and leaf to each node (one
at a time) of each rooted tree with 1 nodes. The coecients in the
equations above represent the number of distinct ways each such tree
can be built in this manner. Instead of expanding existing trees with
new leaves, we will see that the new rooted trees that occur at the th
stage of a Runge-Kutta expansion are built by joining any number of
270 H. Runge-Kutta Methods

trees built at the ( 1)st stage to a new root node. By considering


the multiplicities of ways the trees are built in both models and the
coecients that arise from the Runge-Kutta weighting coecients,
we will obtain the matching conditions that are necessary to achieve
a certain order of accuracy.
Below, we exhibit the rooted trees corresponding to each term
in the Taylor expansion above, along with their associated coef-
cients in that expansion. The coecients in the numerators that
represent multiplicities of the various terms in the expansion can be
interpreted and determined directly in terms of the number of ways
the corresponding trees can be constructed by repeatedly attaching
an edge and leaf to any node of smaller trees, starting from an initial
root node. For example, the factor 3 associated with 24 corresponds
to the fact that it can be obtained by attaching an edge and leaf to
either of the two leaves of 13 or by attaching an edge and leaf to
the root node of 23 . Similarly, the factor 6 associated with 25 cor-
responds to the fact that it can be obtained either by attaching an
edge and leaf to any of the three leaves of 14 or by attaching an edge
and leaf to the root node of 24 that itself has multiplicity 3.

f
fyy fy fy f
fy f f
f
1 1 1 1
11 : (f ), 1!
12 : (fy f ), 2!
13 : (fyy f f ), 3!
23 : (fy (fy f )), 3!

f
fyyy fy f f
f fyy fy fyy
f f f

1 3 1
14 : (fyyy f f f ), 4!
24 : (fyy (fy f ) f ), 4!
34 : (fy (fyy f f )), 4

f
fy f
f
fyyyy fyyy
f f
fy fy fy f
f f

1 1 6
44 : (fy (fy (fy f ], 4!
15 : (fyyyy f f f f ), 5!
25 : (fyyy (fy f ) f f ), 5!
H. Runge-Kutta Methods 271

f
fy f fyy fy fy
fyy f f
fyy fyy
fy f f f

3 4 4
35 : (fyy (fy f )(fy f )), 5!
45 : (fyy (fyy f f ) f ), 5!
55 : (fyy (fy (fy f )) f ), 5!

f
fy fyyy fy f
f fy fyy
f f

1 3
65 : (fy (fyyy f f f )), 5!
75 : (fy (fyy (fy f ) f )), 5!

f
fy fy fyy fy fy fy fy f
f

1 1
85 : (fy (fy (fyy f f ], 5!
95 : (fy (fy (fy (fy f ], 5!

Below, we will perform Taylor expansions of the terms in the


Runge-Kutta samples (H.4) for = 4 stages through order 3 . When
we form their weighted sum (H.3), this yields terms up to order 4 .
These Runge-Kutta expansions very quickly become horrendous, but
when they are interpreted in terms of rooted trees, another surpris-
ingly simple pattern describing the terms present at each stage and
their coecients quickly emerges, just as we saw for Taylor expansion
of y( + ). Before we wade through the formulas, we preview the al-
gebraic and analytical basis for this pattern and its consequences for
determining the order of an -stage Runge-Kutta Method. In (H.4)
we have used y, to denote the arguments of the sample of f that

denes the y, . The rst stage of any explicit Runge-Kutta Method
simply samples the vector eld at the current time-step, y,1 = y .
Then for any method other than Eulers Method, another stage sam-

ples f at y,2 = y + (21 y,1 ), and we can formally expand f (y,2 )
fyy
about y in a Taylor series of the form =0 ( ! (21 f ) ). If a

third stage is used, it samples f at y,3 = y + (31 y,1 + 32 y,2 ).
Expanding f (y,3 ) in powers of involves substituting the prior ex-

pansion of y,2 , combining like terms with y,1 in the perturbation
of y in the argument (here just f ), and then expanding in powers
272 H. Runge-Kutta Methods

of the resulting power series. The th power of this series results in


f
terms of the form yy
! operating on -fold products of its terms. Any
terms whose orders in sum to a particular order contribute to the
overall result at that order, much like the convolution of coecients
that gives the coecient of a certain order in a polynomial product.
Any subsequent stage can be expanded in the same manner. The
description of this process in terms of rooted trees is simply that new
trees are built by attaching any number of trees obtained at the prior
stage to a new root node. In the th stage we expand the evalua-

tion of f at y, = + (1 y,1 + + (1) y,1 ). To do so, we
rst collect like terms in the expansions we have already obtained for


y,1 , . . . , y,1 to obtain a single expansion y, = y + ( ) .
This simply involves summing the parameters for the current stage
times the corresponding coecients obtained at the previous stage.
f
Then the ( yy ! ( ) ) term of the Taylor expansion of f (y, )
is comprised of terms of the form (fyy 1 ). This is repre-
sented as a rooted tree whose root node is attached to the trees
corresponding to the terms 1 , . . . , in the prior stage of the ex-
pansion. Now here is the expansion for = 4.

y,1 = f (y ) = f ,


y,2 = f (y + (21 y,1 )) = f (y + (21 f ))
= f + (fy (21 f ))
fyy fyyy
+( (21 f )2 ) + ( (21 f )3 ) +
2! 3!
2 2 3 3
= f + 21 (fy f ) + 21 (fyy f f ) + 21 (fyyy f f f ) + ,
2! 3!


y,3 = f (y + (31 y,1 + 32 y,2 ))
= f (y + (31 + 32 )f
3
+ 2 32 21 (fy f ) + 2
32 21 (fyy f f ) + )
2!
H. Runge-Kutta Methods 273

= f + (fy ((31 + 32 )f
3
+ 2 32 21 (fy f ) + 2
32 21 (fyy f f ) + ))
2!
fyy
+( ((31 + 32 )f + 2 32 21 (fy f ) + )2 )
2!
fyyy
+( ((31 + 32 )f + )3 ) +
3!
= f + (31 + 32 )(fy f ) + 2 32 21 (fy (fy f ))
3 2
+ 32 21 (fy (fyy f f )) +
2!
2
+ (31 + 32 )2 (fyy f f )
2!
23
+ (31 + 32 )32 21 (fyy (fy f ) f )
2!
3

+ (31 + 32 )3 (fyyy f f f ) + ,
3!


y,4 = f (y + (41 y,1 + 42 y,2 + 43 y,3 ))
= f (y +(41 +42 +43 )f + 2 (42 21 + 43 (31 + 32 ))(fy f )
3
+ (243 32 21 (fy (fy f ))
2!
2
+ (42 21 + 43 (31 + 32 )2 )(fyy f f )) + )
= f + (fy ((41 +42 +43 )f +2 (42 21 +43 (31 +32 ))(fy f )
3
+ (243 32 21 (fy (fy f ))
2!
2
+ (42 21 + 43 (31 + 32 )2 )(fyy f f )) + ))
fyy
+( ((41 + 42 + 43 )f
2!
+ 2 (42 21 + 43 (31 + 32 ))(fy f ) + ))2
fyyy
+( ((41 + 42 + 43 )f + ))3 +
3!
274 H. Runge-Kutta Methods

= f + (41 + 42 + 43 )(fy f )
+ 2 (42 21 + 43 (31 + 32 ))(fy (fy f ))
3
+ (243 32 21 (fy (fy (fy f )))
2!
2
+ (42 21 + 43 (31 + 32 )2 )(fy (fyy f f ) + ))
2
+ (41 + 42 + 43 )2 (fyy f f )
2!
23
+ (41 +42 +43 )(42 21 +43 (31 +32 ))(fyy (fy f ) f )+
2!
3
+ ((41 + 42 + 43 )3 (fyyy f f f ) + ).
3!
Below is a summary of the elementary dierential terms that ap-
peared in the expansion, according to their order in . The notation

, identies the order , the order of the leading derivative of f ,
and the index among such terms. Next to this is the of the cor-
responding term in the Taylor expansion, followed by the equation
of coecients from the respective expansions. Recall the notation

= 1 =1 for expressions that appear repeatedly in the expan-
sion.

,
Elementary coecient
Dierential
, = coecient

1
1 (f ) 1
1,1 11 1 + 2 + 3 + 4 = 1!
1
2 (fy f ) 2
1,1 12 2 2 + 3 3 + 4 4 = 2!
1 1
3 (fyy f f ) 2
2,1 13 ( 2 + 3 23 + 4 24 )
2! 2 2
= 3!
1
3 (fy (fy f )) 3
1,1 23 3 32 2 + 4 (42 2 + 43 3 ) = 3!
1 1
4 (fyyy f f f ) 2
3,1 14 ( 3 + 3 33 + 4 34 )
3! 2 2
= 4!
1 1
4 (fy (fyy f f )) 3
1,2 34 ( + 4 (42 22 + 43 23 ))
2! 3 32 2
2 = 4!
3
4 (fyy (fy f ) f ) 3
2,1 34 3 3 32 2 + 4 4 (42 2 + 43 3 ) = 4!
1
4 (fy (fy (fy f ] 4
1,1 44 4 43 32 2 = 4!


The rooted trees , corresponding to elementary dierentials
up to order = 4 are shown below in order of their occurrence in
stages = 1, . . . , 4 of the Runge-Kutta approximation. Within a
H. Runge-Kutta Methods 275

stage, we have listed trees by the order of the -fold product that
produces them, i.e., by how many previously existing trees are at-
tached to a new root node in order to construct the tree.




1
2
1 2
1
1,1 : = 1 1,1 : = 2 2,1 : 2 = 3

2
1
1
1
3,1 : 3 = 3
3
1,1 : = 6
3
1,2 : 2 = 12




3
1 4
1
2,1 : = 8 1,1 : = 24

Just as for the rst-order Taylor expansion of y( + ), there is


only one tree at the rst stage of the Runge-Kutta expansion, the
tree corresponding to f itself. However, at the second stage, there is
already an innite family of trees corresponding to the innite series of
terms (fyy f f ) with ys and operands f for = 0, 1, . . . . So while
the third-order tree corresponding to (fyy f f ) appears at the second
stage, the other third-order tree corresponding to (fy (fy f )) does
not. This tree only appears in the rst-order term of the expansion
of this series at the third stage when (fy is added in front of existing
terms including (fy f ), or in tree form, the tree corresponding to
(fy f ) is attached to a new root node. Carrying this further, we
can see that the tree with nodes having depth 1 corresponding
to the elementary dierential of the form (fy (fy . . . (fy f ] does not
occur until the th stage of a Runge-Kutta expansion. Note that the
tree corresponding to (fyy f f ), and in fact every tree that occurs at
the second stage, also recurs as the rst term in each order at every
subsequent stage.
276 H. Runge-Kutta Methods

Conversely, since every tree with nodes arises by attaching some


number of trees with strictly fewer nodes to its root, every tree with
nodes does occur by the th stage. This shows that stages are
necessary for a Runge-Kutta expansion to match a Taylor expansion
to order , because at least one term is missing with fewer stages. It
also shows that all of the terms necessary for matching are present at
the th stage, but suciency depends upon the relation between the
number of parameters, ( + 1)/2, that dene an stage method and
the number of coecient equations corresponding to the elementary
dierential (or rooted tree) up to a given order. In particular, we have
seen that there is one parameter for one-stage methods and one tree
for rst-order agreement, and one method, Eulers Method, satises
the matching conditions. There are three parameters for two-stage
methods, and only two trees of order two or less, resulting in a one-
parameter family of explicit two-stage Runge-Kutta Methods of order
two. There are three more parameters for three-stage methods, and
with two more trees at order three with conditions to match, we
reach a two-parameter family of three-stage methods of order three.
With four more parameters for a four-stage method, but also four
additional trees with four nodes, there will again be a two-parameter
family of four-stage methods of order four. But since there are nine
rooted trees having ve nodes, two free fourth-stage parameters plus
ve new fth-stage parameters are still decient by two. Six stages
are required to achieve a fth-order method.
The conditions for matching the Runge-Kutta and Taylor ex-
pansion terms involving an elementary dierential can be obtained
directly from the structure of the corresponding rooted tree. For this
purpose, we relabel the nonleaf nodes with index symbols for summa-
tion. Note that the earlier labeling with y-derivatives of f was helpful
but not actually necessary to recover the elementary dierential, and
the same holds here for recovering coecients. The coecients devel-
oped in successive stages arose from summing over the parameters
of the method times corresponding coecients of previous elementary
dierential terms. Therefore, each time we attach an existing node to
a new root node, we contribute a sum of the corresponding coe-
cients to the coecient corresponding to the resulting tree. Each leaf
H. Runge-Kutta Methods 277

node contributes = for the stage it represents and therefore


does not need to be indexed. Though the purposes are slightly dier-
ent, the form in which coecients appear is the same as that of the
coecients. The s are used to construct evaluation points for f ,

, = f (y, ), with y, = y + ,1 y,1 + + , y,1 . When we
have obtained suciently many (i.e., ) of these evaluations, the s

are used to obtain y+1 = y + ,1 y,1 + + , y, . Because of
this, the form of the coecient formula associated with a particular
tree is the same regardless of the number of stages of the method in
which it appears. In other words, the coecient formulas and corre-
sponding matching conditions for a four-stage method reduce to those
for a three-stage method simply by eliminating all terms involving co-
ecients whose rst index is 4. If we only retain terms involving
coecients whose rst index is 2, we recover the two conditions we
found for a two-stage method to be second-order.
The rooted trees corresponding to exactness of the solution of
the polynomial accuracy model problems = ( ) are the depth-
2
one trees with nodes, 1,1 = 1 , that occur at the second stage.
These are the only rooted trees of order 2. Therefore, exactness
on equations whose solutions are polynomials of degree is necessary
and sucient for general th-order accuracy of an explicit Runge-
Kutta Method when 2. Even though these trees occur at the
second stage, for > 2, stages are required to match the Taylor
coecient. The rooted trees corresponding to th-order accuracy for
the absolute stability model problem = are the maximal depth

trees with nodes, 1,1 = , i.e., the trees with one edge leaving
every node except the leaf. For this problem, all derivatives of f (y)
beyond the rst are zero. This tree does not occur until the th stage.
Therefore, stages are necessary for general th-order accuracy.
Below the trees representing eight elementary dierentials of or-
der 4, we have collected the factors from the two sides of the match-
ing conditions in the table into one of the form 1/( ). Here ( )
arises as the ratio of the multiplicity of ways the tree can be con-
structed by addition of edges and leaves on the Taylor side and a
combination of th-order expansion factorials and multinomial coe-
cients on the Runge-Kutta side. These factors can also be computed
278 H. Runge-Kutta Methods

directly from their trees using the following simple algorithm. The
density of any leaf, a rooted tree of order one arising from evaluat-
ing f in the rst stage, is 1. At every subsequent stage, at which we
attach one or more trees to a new root node, the density of the result-
ing trees is the product of the densities of the trees being attached,
3
times the order of the resulting tree. For example, the tree 2,1 is a
rooted tree with four nodes corresponding to the elementary dier-
ential (fyy (fy f ) f ). It is rst obtained at the = 3rd stage of the
Runge-Kutta expansion in the = 2nd-order term of the expansion,
by joining the trees corresponding to (fy f ) and f . The former has
density 1 2 and the latter has density 1, so since the resulting tree
has order 4, its density is 1 2 4 = 8.
For = 4 stages, the eight fourth-order matching conditions are
1 + 2 + 3 + 4 = 1, 2 2 + 3 3 + 4 4 = 12 ,
2 22 + 3 32 + 4 42 = 13 , 3 32 2 + 4 (42 2 + 43 3 ) = 16 ,
2 23 + 3 33 + 4 43 = 14 , 1
3 32 22 + 4 (42 22 + 43 32 ) = 12 ,
3 3 32 2
+4 4 (42 2 + 43 3 ) = 18 , 4 43 32 2 = 1
24 .

The recommended procedure for solving the equations is to choose



2 , . . . , and then solve the equations, =1 , = 0, . . . ,
1, for 1 , . . . , . Next, solve for the that are determined by linear
equations. In this case, the fourth, sixth, and seventh equations above
allow us to solve for 32 , 42 , and 43 . The classical fourth-order
Runge-Kutta Method corresponds to the solution obtained by setting
2 = 3 = 12 , 4 = 1, given in tableau form as

0
1 1
2 2
1 1
2 0 2
1 0 0 1
1 2 2 1
6 6 6 6

For the scalar ODE with = 0, i.e., = (), this method reduces
to the Simpson-parabolic quadrature method.
H. Runge-Kutta Methods 279

Among the methods satisfying the matching equations of a given


order, optimal methods can be obtained by minimizing local trun-
cation error bounds. A well-known example of this is the two-stage
method of order 2 known as Ralstons Method, given in tableau form
as
0
3 3
4 4
1 2
3 3

Pairs of closely related Runge-Kutta Methods can be used for au-


tomatic step-size control in the same manner that pairs of multistep
methods are used for error estimation and step-size modication. A
well-known example of this technique is the Runge-Kutta-Fehlberg
pair consisting of a ve-stage and six-stage method of orders 4 and 5,
respectively. Further details on these and other topics, including spe-
cial cases for scalar, autonomous, and constant coecient systems of
equations, methods based on extrapolation, methods to treat second-
and higher-order equations directly, etc., can be found in [BJ].
The region of absolute stability for an explicit -stage method is
determined by one step of the method applied to the absolute stability
model problem, = . For 4, we know that the coecients
can be chosen so that the method has order of accuracy = 4. In
this case the region of absolute stability is { C () 1}

where () = =0 /!, the truncation to degree of the exact
exponential series solution of the model problem (Figure 5.14). For
> 4, we must replace by some polynomial of degree that
depends on the specics of the method.
Implicit Runge-Kutta Methods can be employed if larger stability
regions are required. See [IA1] for a discussion of these methods and
their relation to Gauss-Legendre quadrature and collocation methods.
Of particular relevance to the topic of this appendix are several pub-
lications on Runge-Kutta Methods for Hamiltonian systems. Explicit
symplectic Runge-Kutta Methods only exist for general Hamiltoni-
ans that are separable, but the implicit Gauss-Legendre Runge-Kutta
Methods are symplectic and they are optimal for general Hamiltoni-
ans. See [IA2], [SJM], [CS], [HLW], [YH], [CP].

Vous aimerez peut-être aussi