Carpenter Kennedy 1994
Carpenter Kennedy 1994
Carpenter Kennedy 1994
--o 211
NASA
Technical
Memorandum
109112
Runge-
Center,
Hampton,
Virginia
Christopher University
A. Kennedy of CMifornia,
San
Diego,
La JoHa,
California
N94-32950
Uncl
a s
June
1994
G3/02
0009987
FOURTH-ORDER
2N-STORAGE
RUNGE-KUTTA
SCHEMES
Mark
It. Carpenter
_ Christopher
A. Kennedy
Abstract
fourth-order
Runge-Kutta scheme
schemes that
these
schemes
require
efficiency
characteristics with
problems. is considerably
scheme more
is competitive and
fourth-order third-order
method low-storage
efficient
accurate
Section
1: Introduction
Many coupled
result The
in the
numerical
approximation
of systems depends
of
technique
efficiency,
robustness,
ODE's
tile direct
simulation
consideration memory
often
availability and
of fast
memory.
technique (RK)
minimizes
storage
is essential
format, exceptional
N is the that
cases
specific be put
of the
fourth-order
RK
schemes RK
2N-storage
format.
four-stage reduced
scllelIle$.
fourth-order accuracy,
schemes
locations.
of the RK
for many
apl)lications
tile 2N-storage
third-order
the
principal also
motivation determine
of this the
work
schemes scheme.
both
the
elliciency scheme
practical small
Branch, University
a fourthvalue.
VA 23681.
2N-storage
with
a vanishingly
Flow Physics Sciences,
stability
NASA
is of little
Center,
Tile
*Aerospace tApplied
Engineer, Mecha.ics
of California,
La Jolla,
92093.
efficiency number number absolute eighth-order [4] showed increased stages stage.
is determined
relative
to the
Typically, of accuracy
because
demonstrated scheme.
Similarly,
Sharp
seventh-order order
RK schemes The
by including
additional the
additional of each
are used
to optimize
increase
efficiency
fourth-order largest
that
satisfies
constraint lowest
is sought.
stability and
envelope the
truncation
2, we describe complete
conventional
2N-storage (3,3)
set of three-stage
third-order
RK schemes (4,3)
We then
introduce
third-order fourth-order
RK schemes
(5,4)
that and
require then
2N draw
accuracy,
Section
2: Runge-Kutta
Nomenclature
We are concerned
with
the numerical dU
solution
of the initial
value
problem
U(to) with
approximation
is made by
an M-stage
over a time kl ki
h is accomplished F(to,
F
U ) to
+ cih, M V 0 @ h E ai,jkj
i = 2,...,M /
j=l
v
where U = U(to) formula. and
= v + bjkj
j=l
U_ =
U(to + h) and
the fixed
scalars
the coefficients
of the RK
Butcher
[5] developed
a succinct
nomenclature
for writing
all RK
schemes
in terms
of the fixed
scalar
coefficients
general,
c1 g2,1
M-stage,
explicit
ILK scheme
can be expressed
as
CM
aM,1
aM,M_
bl The are entries in the first column (ci, i = 1, M) of each RK stage are the i.
bM
intermediate
time ci =
levels.
The
coefficients cl
ai,j
= 0 for the
explicit
last row
weights exploit
of the final stage. the onto technique the same of leaving register useful without
Each
stage
is written becomes
the previous
M-stage +
AjdUj_I
uj =
So that results the algorithm is self-starting, algorithm. in a 2N-storage
vj_l +
A1 = 0.
j= dU and
RK scheme between
and
Specifically, Bj and
the 2N-storage
Aj and
conventional
RK variables
Bj
BM
=
_--
aj+l,j
bM
(j # M)
(j _ 1,bj # 0) (j # 1, bj =0) (1)
Aj Aj
= -"
The
precise
values
of Aj
and RK
Bj
that
are are
to yield order
a higher
order
scheme the
remain
to be of
Third-order is nontrivial.
schemes
schemes
determination 2N storage
We begin
by demonstrating
the procedure
high-order
case.
Section
3: Third-Order
Runge-Kutta
Methods
Runge-Kutta three-stage
scheme, explicit
at least scheme
three
stages as
are required.
The
general
Butcher
is written
C2 c3
I bi
b2
b3
For third-order
accuracy,
seven
constraints
between
the
coefficients
c/must
be satisfied
[5]:
3
i=1
and
b, = l (1)
3 _bic,=-
1 (2)
_=1
i=l
1(3) bic_ = -
3 Y_ biai:jcj
i,j=l
1 (3) = (2)
ci = _aij
j=l
i=
1,3
(3)
The
superscripts
on each
that
the
constraint equations
algebraic
result,
of which
a two-parameter then
c3 are defined
free parameters
(Cl =
0 for explicit
RK schemes),
the following
for case
1 (for which c3 2
c2 # 0,2/3,
or c3 when - 2
c3 # 0): 1 ; points bl= X-b2-b3 from These 1 ; a3,2 = nonzero 1 ; a3,2 = non-zero 4b-_ number). Equations (6) (4) through _3 number), and (5) ; a3.2 = 6b3c 2 (4) solutions by
_
-- C2) ; b3 --
b2 Two exist
2c2(c3
3 c2 2c3(c3-c2)
additional when
solutions
account
excluded case).
are governed
c2 =
2, c3 = 0 and 1 b2 = _;
b3 is an arbitrary 3
bl --
_--b3
c2 = c3 = set of (3,3)
32and
b3 is an arbitrary
RK schemes.
The
linear
stability
envelope
for the
(3,3)
RK schemes
is determined
by
G = 1+ where Z is a complex
bi
Z + that
bici
Z 2 qi,j=l
biaijcj
Z3 later
= and
1 2t- Z
-)t-
_Z
71"
Z 3
number
in detail
G is the
amplification (2);
of G are have
precisely identical
equations
defined
in equation accuracy,
envelopes.
however,
will be problem
Numerous [6].
different
methods
for choosing
criteria
are summarized
elsewhere
The (3).
the
2N-storage
scheme.
The
original
Aj and
Bj to obtain
C2
c3
o I
a2,1
B1 a3a b2 b3 =
a3a bl
BI + B2(A2
0
B3,
A3,
B1
+ I)
B2 [A3B3 + B2] Ba
defined of B1,
in equations B2,
By definition, is not
be satisfied.
By adjusting
solutions,
this
final
For case
c3 # 0, a one-parameter
family
of solutions
If we define
= = = = = =
_/i36c24
+ 36c23 -
135c_
84c2-
12)
2c_ + c2-2 12c_36c 4 69c23 34c_ by 18c_ + 36c 3 + 62c_ 18c_13c] 11c2 8c2 + 8 13c2 + 2 (7) + 2 4
+ 28c2 34c]
46c23 +
is given
C2
B2
12c2(c2 -
1)(3z2-
Zl)
(3z21) 2 1) 2
Zl) 2
144c2(3c2-24(3c2B3
2)(c22)(c2-
(3z2A s =
z,) 2 -(6c_
12c2(c2+
1) (3z2 - z,) 3z3 - 1) 2 3(2c_ l)zs (8) that both eliminates the
- 4c2 + 1)zx -
(2c2 + 1)Zl
A 3 =
-z4z,
108(2c2-
4 + 72c2z6
72c_(2c2-13) A scheme
denominators
vanishes.
square
roots
and
is close to the 0
optimum
truncation
error
was identified
by Williamson
[1] as
A1
B1
0 = _:
9 153 128
3 16
15 16
==_
A2 A3
B2 B3
:s
16 8 15
(9)
1 6
3 10
8 15
This
scheme,
which
is widely
used,
will be referred
to as the
(3,3)
2N-storage
scheme
of Williamson
in the remainder
of this paper.
c2 = ], c3 = 0, and results 0
2 3
b3 is an arbitrary
nonzero
number,
we find a 2N-storage
that
in the scheme
0
1
2
3
3
4
3
4
=_
-:
9 2
-7
1 3
r.
12
3_
4
_I
3
= results
c3 = in the 0
2 and :
b3 is an arbitrary
nonzero
number,
we find a 2N-storage
scheme
0 -1 -1
1 4 5 12 1 3
2
3
1 12
3
4
=_
3_
4
1
3
Because identical
the
are a subset
of the general
(3,3) RK scheme
solutions,
they
too have
linear
stability stage
RK schemes, (c4 =
stages seven
nonlinear in fourteen
equations variables,
and
adds
which
equations additional
is a six parameter
If 2N-storage 6
is required,
equations third-order
must
be satisfied,
which that
eliminates involves
three
The
general
2N-storage by enforcing
as a three-parameter constraint.
be generated
fourth-order
If third-order time
step is increased
by one
By en_0rcing
the
constraint,
increased
stability envelope of the (4,3) R.K schemes mpre than offsets this additional work. The (4,3) schemes are more efficient per unit time, which is consistent with the work of Sharp et al. [4]. An extensive study of the stability domains of the (4,3) R.K schemes shows that a nearly optimal scheme can be obtained by setting
G----I+
(i__lbi)
Z+
(_=lbici)Z2-t
(_i,j=lblaijcj)Z3-4
(i,j,k=l _
biaijajkck)Z
l+Z+ schemes
Z2+
Z3+=-:Z accurate schemes for the nonlinear that satisfy problem but fourth-order constraint accurate
problem.
the 2N-storage
in four stages
a nearly
optimal
stability 0 z_
3 1 3
envelope
!
3 5 3
0 -1 =_ 2
3
12
4
4 i
12
(10) -i -1
1 3
5 12
1 4
0 0
5 II
4 II 9 or 18 II 1 12
19 36 3 4 3
19 36 51 76 19 27 19 n 2
19 36 27 19 2 9
11 36 419 396
11 9 16 9 18 II
205 243
:=_
II 6 182 II
=_
243 38
_2
9 13 57 27 76 1 6
I
4
1 11
3 4
17 66
1 12
If we demand 0 1
9 4 9 6
that
the intermediate
time
levels
cj be monitonically
increasing,
then
1
9 11 36 1 12 3 4 7 20 2 5
0
5 9
1
9 3 4 or
3_
9 5 9
3
9
0
II 15
1
3 5 6 3 5
S 18 4"1 90
5_. 6 1 6 3 5
=:_ -1
33 25
2_
5 5 4
8_
9
5 3
-1
3 20 i 4 7 20 1 4
L
4
-I For problems
2 in which
5
4
s
4
third-order larger
temporal stability
accuracy envelope.
is sufficient,
these
new
schemes
should
be
considered
because
of their
Section
4: Fourth-Order
l:tunge-Kutta
Twelve
algebraic
equations Five
in fourteen general
the
solution
to the four-stage
solutions covering
special that
Butcher
could
in general, equations
format.
2N-storage solutions.
additional
be
by any of the
five general
of stages, 2N-storage
degrees
of freedom The
that
equations. form as
fourth-order
in Butcher
el C2 a2,1 a3,1 a4,1 as,1
and
At A2
a3,2 a4,2 a5,2 a4,3 a5,3 .a5,4
B1 B2 /33 /34 Bs
C3
C4 C5
:,
A3 A4 As
b2 cl =
b3 0, and
b4 A1
bs = 0. For any
schemes, are
(5,4)
RK scheme,
the
nonlinear
b,=
; 5 _i,j=t
EL, b,c,=
biciai,jcj = 1 g (4)
5 _i=1 )"]i,j=l
bic_ = i_ (3)
5 _i,j=l ;
blaijcj
1(3)
s )-]4=1
blc 3
biaijc --
1(4)
5 _i,j,k=l
biai,jaj,kck
_ --
--
and
5 ci -_ j=l ai,j
i = 1,
Again,
the
(11)
indicate
the
from
which (4,3)
the
constraint
was obtained.
to those
(3,3) and
condition. represents
These
13 constraint
equations
which
at least
a seven-parameter
relationships
between
the original
RK variables o Bi
Bi + B2(A2
2N-storage
variables
A j, B
: = = = = = = = = :
+ 1)
+ 1) + 1] + 1}
+ B4) + B3
---- B4,
B5
If the
(5,4)
RK scheme Specifically,
to a 2N-storage defined
format,
equations
are produced.
relationships
bl are
determined Again,
constraint
equations
immediately
as,3.
Assuming answer
a solution
of solutions.
is not forthcoming
complexity
determine equations
specific
solutions by
19 equations special
Frequently,
we can The
considerably
ci or Ai.
simplifies
considerably
if we assume but
of freedom
in the general
solution
produces
Butcher
this casebecomes
_-+ _" + 22/-A 2+
3 2 ._ 4_/2 422/3 41 1 ._ 2213 42213 4_ _ 1 1 1 2_" 3 22/3 3
3-4-
--"-C2213 4-
3 -- _ 1
3
+ --c 2&
22/3 6
2213
----]---_
22/3
__ 4- 221__A3 + I_
i'_ j 3-_._3_. 6 -22132 12
____
2213
--_-I_ 4.
3 -6 --
_
12
_/_ + 2213
___ __ _____
2213 6 3 6 6 6
2_3
12 1 + 3 _
2213 x-T
The
2N-storage
array
A1
becomes
B1
0 -1 = -1/3 +
--'_22/3
2/3+_ +
2_3+ 1/6
3 -1/3_ 22/3
A2 A3
A4
B2 B3
B4
2 2/3 -- 2
113- _
2212
A5
Bs
-1+_
= --(__1 4--_4- --"_'J, the linear 72 22/3_
1/3+&+ _2_3 6
stability envelope is found by considering
Because
_i,j,k,l=l 5
biaijajkaklcl
G=l+Z+sZ
where
+_z 3 +_z 1 1
analytic array 0 -I
=
--
(7_
+-if+ "_/'2
2_--_)Z5
_i,i,k,l=a
blaijajkamct
_.
- 1
A second 2N-storage
B1
solution
has been
obtained
by setting
all values
of the Aj = -1
for j = 2, 5. The
A1
A2
A3
B2
B3
-I -I -i
B4 B5 is
B4 4_.__=t '
12 B4
Z3+_-Z
+\
because _
of the small
The the
coefficient
B4 is defined
5 4/X2_ envelope,
'
with
,s_ _
13824
,m
768 "
Because
_ _,
_m,,i_ic_tio. b_om_,_ _+Z+_Z' _Z_ _Z"+ (_)Z'.Neithor two m_trix _ + + ofthe =_'y_ic
solutions the second has a desirable stability small. numerical as we shall they see later. value The first is of marginal they provide value, good and is extremely However, roots. are of some because initial guesses for determining
In general, ultimate
we must
resort that
solution stability 10
of the envelope.
19 equations
The of
goal is a scheme
We can constrain
freedomin the generalsystemby demandingthat the resulting scheme optimal, which produces be a systemwith 20equationsand 20 unknowns. To definethe optimal constraint,we beginwith the amplificationmatrix for the (5,4) RK scheme, which is givenby
\i,j----1
) (
as being optimal is as large is problem integrated. and
i,j,k--I
) (
when its stability for a particular and depends partial differential A model
i,j,k,l=l
)
encompasses of the complex of the a (PDE's), that = for
scheme
can
envelope choice
as possible dependent
on the structure
eigenvalues
of the system
specifically 2N-storage
parabolic
equations equation
produces If we
mechanics equation
is Ut aU, reduces
aUxx.
U** terms,
to a system
complex
matrix
to represent study
combinations of the
stability of the
stability
number
amplification upper inviscid could details.) stability 1 <a< 28--'6 -maximizes reveals at c_ _ that
toward
the sixth-order
reached
a Second-order
or a Fourier
5
parameter
a is defined
By adjusting a more
envelopes
refined
Note inviscid
all curves
structure (real at _ =
(imaginary inviscid _
axis) 3' 1
Further stability
the maximum
stability provides _
c_ = values
a compromise because
Caution
stability changes
values
dramatically the
with
small
RK scheme, a. Figure
(4,4) RK scheme,
and
three
schemes
presented
do not
characteristics. a =
stability analytic
case a _
_8 is about inviscid
a vanishingly
small
The
optimal
value
a =
is used
to provide
the 20th 11
nonlinear
equation
in 20 unknowns.
The
family
and,
because
of the nonlinearity
in the equations,
biaijaLkak,lct
(13)
the partial
linearity A1
to reduce to The
procedure (11)
is used
to obtain with
roots
were found
as an initial have
random
initial
real roots
identified
all nine
equations.
formulations roots
is optimal
We can, several
however, theoretical
arguments that
certain
should
in determining
are relevant
to this work
intermediate changing
time
be in the
interval
[0,1] to control
the effect
of
derivatives. bj,j = 1,5 should rational be positive. numbers requiring a small number of digits.
weights
of the should
Coefficients
incorporate
nine roots
satisfy
condition
conditions numbers.
I and
able
to
in terms
of "convenient"
1 shows
the four
roots
which values
satisfy
the
condition
that
they
monotonically
increasing
of ci. Solution
3 satisfies
all bj > O.
12
Table COEF Ax A2 A3 A4 As B1 B2
1. Four Sets of Coefficients SOLUTION 0 -0.4812317431372 -1.049562606709 -1.602529574275 -1.778267193916 9.7618354692056E - 2 0.4122532929155 0.4402169639311 1.426311463224 0.1978760537318 0 9.7618354692056E - 2 0.3114822768438 0.5120100121666 0.8971360011895 1
For optimal SOLUTION 0 -0.4801594388478 -1.4042471952 -2.016477077503 -1.056444269767 0.1028639988105 0.7408540575767 0.7426530946684 0.4694937902358 0.1881733382888 0 0.1028639988105 0.487989987833 0.6885177231562 0.9023816453077
(5,4) 2N-Storage 2 SOLUTION 0 -0.4178904745 -1.192151694643 -1.697784692471 -1.514183444257 0.1496590219993 0.3792103129999 0.8229550293869 0.6994504559488 0.1530572479681 0 0.1496590219993 0.3704009573644 0.6222557631345 0.9582821306748
RK Schemes 3 SOLUTION 0 -0.7274361725534 -1.906288083353 -1.444507585809 -1.365489400418 4.1717869324523E - 2 1.232835518522 0.5242444514624 0.7212913223969 0.2570977031703 0 4.1717869324523E 0.377744236865 0.6295990426348 0.8503409780005 - 2 4
B3
B4 B5
c1 C2 c3 (:4 C5
values satisfy
of the
parameter III.
a produce
similar
results,
although
able
to find
condition
the
The floating point numbers can be expressed rational form of the coefficients with 26-digit
A1 = 0; A2 --_ 567301805773. 1357537059087, A3 =
numbers.
For solution
3,
_2404267990393. 2016746695238_
3550918686646. 2091501179385_
As
n=
--_
BI
1432997174477. 9575080441755,
B2
--
B3
--
B4
-"
Cl
"0;
C2 -_
C3:
C4 _
and errors
can be used
on machines
that
have
up to 26 significant
digits
without
appreciable
roundoff
in the coefficients.
Section
5: Accuracy
and
Efficiency
of (5,4)
RK Schemes
The
four
new RK (9)),
schemes 2N-storage
given RK
in Table scheme
with
the
classical
(4,4)
storage equation
new
(4,3) 2N-storage
RK scheme
ODE by y'
used =
by Dormand y(0)
y cos(x),
_< x
20, with
solution
4 shows approach
study solution 13
schemes
classical
(5,4)
approach
the exact
solution
a slope the
of -4. Note
that
the
(4,3) scheme
of error)
and
schemes (5,4)
the
difference over
four
of the
(5,4)
to establish
problem-dependent
by y' = y n sin n-l(x)cos(x), = expSin"(*), The the where the exponent rates (4,4) If the 3 and
= 1 on the interval
5 shows schemes
for all seven the theoretical -5. This the rate two
convergence conventional
with
approaches
generally problems,
four solution
between second
problems,
No clear
is evident
any particular
(5,4) 2N-storage
verifies test
developed is more
schemes requires
in both
evaluations in terms
scheme,
was made
to optimize
of error. and
The more
portion errors
problem.
to be more
adversely
conventional
scheme.
Our
second
problem
of the linear Ou Ox = = O,
hyperbolic
equation
defined
by
(14)
sin27r(-t), sin2_'(x),
(15) (16)
u(x,t) and is a model Equations problems; for the class of problems (14) through (16)
sin2_r(x-t), that
are solved
four RK schemes
and the (4,4) and (5,4) RK schemes. scheme developed by Carpenter is imposed et
used
sixth-order accurate. 14
to be formally
sixth-order
physical
boundary
condition
by solving the differentiated boundary condition on the boundary with the RK procedure. This techniquewasshownby Carpenteret al. [9] to yield a fourth-order temporally accurateprocedure. Specifically,the boundary condition is d3u(O,t)/dt 3. = gin(t), where g is the physical boundary
condition function compares with at the of the inflow temporal plane. The CFL and that the governs spatial the stability of the used. hyperbolic Table and problem is a advancement and discretization limits 1 in the appendix (5,4) RK schemes
CFL's operators,
the viscous
stability
(4,4),
various
including
the sixth-order
compact
operator.
grid
refinement
with
a vanishingly grid,
small
recover fourth-order
the
theoretical temporal
accuracy.
temporal
2 shows
with a CFL
of _ (about
theoretical Table
maximum 2. Grid
refinement
at CFL
2N-storage
RK schemes
The error
new
scheme (5,4)
accurate
linear digit
problem.
The
In addition, of CFLmax
stable
greater
the
theoretical
stability
appendix.)
relative The
efficiency advantage
of each in terms
scheme
is not
addressed
in any schemes
of the
previously
discussed
test
is partially and
The
wave equation; is defined RK scheme as the was were and time study stable,
of times
spatial
terms
evaluated
at a CFL of a both (4,4) RK schemes _, were run at a CFL of 4 and the (5,4) R.K schemes at a CFL of 5 _., where 4 < k < 16. The resulting schemes are all numerically stable, errors vary identical. by a factor of 256 for the were run fourth-order to a physical case. The effective cost per unit
temporal
was, therefore,
All schemes
to determine
the order
of accuracy. 15
sixth-ordercompactspatial operator [7] wasusedwith the boundaryconditionsthat werepreviously described. Tables3(a) and 3(b) showa grid refinementstudy at two different CFL numbers,the first of which is nearthe CFLmax the second which is at half the valueof the first. Only the (4,4) RK and of resultsare presentedherebecause (3,4) and the (4,4) RK schemes the producethe sameresults to five significant digits. The resultsfor the (5,4) scheme comefrom solution 3. (This study is a linear advectionproblem for which the four-stagethird-order scheme performsto fourth-order accuracy.) Table 3(a). Grid-RefinementStudy for Three RK Schemes on
Ux = 0, Run at Equal Cost Near Maximum CFL Linear Advection Equation Ut +
(4,4)
(5,4)
Table
3(b).
Study Half
RK Schemes CFL
on Linear
Advection
Equation
Ut +
U, = 0, Run
at Equal
(3,3)(2N) Grid 41 81 161 321 641 logL2 -3.6257 -4.6231 -5.5515 -6.4611 -7.3659 Rate 3.31 3.08 3.02 3.01
(4,4)
(5,4)
(5,4)
and
the
(4,4)
schemes
are
nearly errors,
identical
when
compared when
at equal the
cost.
The error
(5,4) is near
susceptible
to roundoff
as can be identified
absolute
schemes PDE's.
ODE's
overriding
concern clearly
of integrator Williamson's
the newly
developed and
outperform
of accuracy
accuracyper unit cost is comparablewith that of the standard four-stagefourth-order (4,4) RK scheme.
Conclusions
fourth-order
(RK)
schemes that
that simple
require
only two
coefficients A particular
neither
scheme
is identified
equations)
desirable In the
efficiency and
and 2N-
(boundary) has
value
problems.
inviscid step
scheme the
accuracy
size and
allowable (5,4)
stability RK scheme
(3,3) the
RK scheme (4,4)
with
to work these
ratio results
and
is nearly
in an absolute differential
on nonlinear
equations
equations.
Acknowledgments
The
second
author
would Center,
financial Physics
support Branch,
provided under
while
in residence
at NASA
Langley
Research
contract
NAG-l-l193.
References
[1] Williamson, [2] Fyfe, [3] Verner, SIAM [4] Sharp, than D.J., J.H., J.H., "Low-storage Runge-Kutta schemes," J. Comp. Math. of the Phys., 35, 48 (1980). 20, 392 (1966). Truncation Error,"
formula," Estimates
Comp., Local
Pairs
with
One
More
Derivative
Evaluation
SIAM Numerical
14, 2 (1993),
Analysis Wiley _
of Ordinary Sons,
Methods,
John Ways
Chichester Error,"
R.D.,
"Thirteen
to Estimate
Global 17
Math.,
48,
(1986),
pp. 1-20.
[7] Carpenter,M.H., Gottlieb, D., and Abarbanel,S., "The Stability of NumericalBoundaryTreatments for CompactHigh-OrderFinite-DifferenceSchemes," Comp. Phy, 108, 2, (1993), pp. J.
272-295.
J.R.,
and
Prince,
P.J.,
"A Family
of embedded
Runge-Kutta
formulae,"
J. Comp.
Math.,
pp. 19-26. D., Abarbanel, for the S. and Initial ICASE Don, W.-S. " The Theoretical Accuracy of
Discretizations
Boundary Report
Value
Problem: Dec.
A Careful
Study to
Error,"
NASA-CR-191561, Analysis.
No. 93-83,
1993. Submitted
of Numerical
Appendix
Table au**,
A1 shows
the
inviscid
and spatial
viscous
stability
limits techniques.
obtained The
from inviscid
the
equation
ut + au,
= case the
discretized
with various
and temporal
a = 0; the viscous operator. for compact pentadiagonal hyperbolic (4,4), (4,4), and and
limit
is the special
first
bandwidth signify
and
The
presented
stability stability
inviscid
case with of _ =
spatial (5,4)
operator are
(6T), v_ 4 =
the 0.354
(3,3) and
scheme
has
an
CFL/stage
schemes
sense, (4,4),
RK scheme RK schemes
efficient.
Relative
has been
assigned
viscous
spatial
is 0.s3.__.._o 3 followed
respectively.
efficiencies,
if we assume
an absolute
outperforms
proposed
In the
viscous
scheme
dramatically
conventional as efficient
In the inviscid
RK scheme
is competitive
as the
(4,4) RK scheme. 18
Table AI: Inviscid and ViscousCFL limits Invis Stab Limit Vis Stab 4 5 3 4 Number of stages 3 4 4 3 4 Order of accuracy 3
Limit 5 4 4.65 2.47 1.55 1.85 1.17 1.22 1.55 1.02 0.95 1.37 0.93 0.86 0.47
2nd(E)
4th (E) 4th (T) 6th (E)
v_
1.26 l 1.09 v_ 2 0.88 1.00 0.81 0.78 0.94 0.77 0.74 v_
7f
2v_
2.06 _ 1.78 v/_ 1.44 1.63 1.32 1.28 1.53 1.26 1.21 2v_
3.34
2.43 1.92 2.10 1.67 1.71 1.93 1.56 1.51 1.81 1.49 1.43 1.07
2.51 1.33 0.83 0.99 0.63 0.65 0.83 0.55 0.51 0.74 0.50 0.46 0.25
2.78 1.47 0.92 1.10 0.70 0.73 0.92 0.61 0.57 0.82 O.56 0.51 O.28
6th (T)
6th (P) 8th 8th 10th (E) (P) (T)
8th (T)
10th (E) 10th (P) Fourier
19
;;h
.1'-4 o,'-t
,.c
u'J
,-i
r.n
c_
0
u_
u.)
r.o
FIGURE
2. Refined parametric
study
(in variable
a) of stability
envelope
.5
_,V
1.0
1.5
21
FIGURE schemes.
3.
Comparison
of optimized
RK[5,4]
schemes
with
conventional
RK[3,3]
and
RK[4,4]
2"0 /_
a = RK (3, 3) b = RK (4, 4) 1 c = RK (5, 4); a I -._-_. d = RK (5, 4); av= 1 24_ 100
I 0 .5 1.5 2.0
22
FIGURE 4. Grid convergence study comparing RK[3,3] schemes. Test problem is y' = y cos(x).
several
RK[5,4] schemes
\ \
-2
\ \
:,, ,,"',\
-4 _k-, _,.,
O
k.
y'=
y cos(x)
"_,_\ "\,.'-,,\\
.. ".. \\
-6
{D O O
_.
-8....
",.
"\\
-. ,_,,,,
" . ,\\
-10
.... _"
.......... f (4,4) (5,4) I 2.0 Ioglo f ,3N 2N I 3.0 (steps) I I 4.0
",:
-12
-14 1.0
"""1 5.0
23
0_
,_
0 IF)
/
/
/,
,/o
/,
/o"
/ /.
/#
4
O_ 0 o c_ 0"_ 0
/.
/"
/o /,
/.
/,
t_ t_
o_._
0
I 0 C'M _
I cO
I 09 0
I 'q
0 "_t'--
'
'
_
(JoJJ_)
'
L6ol
_e3
REPORT
DOCUMENTATION
PAGE
Publicreporting burden forthiscollection ofinformationis estimated to average1 hourper response, ncluding i thetime forreviewing instructions, searchingexistingdata sources, gathering and maintaining data needed,andcompleting end reviewing collectionof information. Sendcomments the the regarding this burden estimateor any otheraspectofthis collectionof information, including suggestions reducing for thisburden,toWashington Headquarters Services. irectorate Information D for Operations and Reports,1215JeffersonDavis Highway, uite1204, Arlington, 22202-4302,andtothe OfficeofManagement ndBudget,Paperwork S VA a Reduction Project(0704-0188),Washington, C 20503. D 2. REPORT DATE 1. AGENCY USE ONLY (Leave blank)
June 1994
4. TITLE AND SUBTITLE
Fourth-Order
2N-Storage
Runge-Kutta
Schemes
6. AUTHOR(S)
Mark H. Carpenter Christopher A. Kennedy 7. PERFORMING ORGANIZATION NAME(S) ANDADDRESS(ES) NASA Langley Research Center Hampton, VA 23681-0001
8. PERFORMING ORGANIZATION REPORT NUMBER
9. SPONSORING I MONITORINGGENCY A NAME(S) ANDADDRESS(ES) National Aeronautics and Space Administration Langley Research Center Hampton, VA 23681
NASA TM-109112
11. SUPPLEMENTARY
NOTES
Carpenter: Langley Research Center, Hampton, VA. Kennedy: University of California at San Diego, LaJolla, CA.
12a. DISTRIBUTION / AVAILABILITY STATEMENT 12b. DISTRIBUTION CODE
13. ABSTRACT
A family of five-stage fourth-order Runge-Kutta schemes is derived; these schemes required only two storage locations. A particular scheme is identified that has desirable efficiency characteristics for hyperbolic and parabolic initial (boundary) value problems. This scheme is competitive with the classical fourth-order method (high-storage) and is considerably more efficient and accurate than existing third-order low-storage schemes.
Runge-Kutta,
low storage,
fourth-order
25
16. PRICE CODE
A03
17. SECURITY CLASSIFICATION OF REPORT 18. SECURITY CLASSIFICATION OF THIS PAGE 19. SECURITY CLASSIFICATION OF ABSTRACT 20. LIMITATION OF ABSTRACT
Unclassified
NSN 7540-01-280-5500
Unclassified
Unclassified
Unlimited
Standard Form 298 (Rev. 2-89) Prescribed ANSI Std. Z39-18 by 298-102