Sci Cade 03

Exponential Time Differencing and Lie-Group
Methods for Stiff Problems

SciCADE 2003, June 30 - July 4
Borislav V. Minchev
Borko.Minchev@ii.uib.no
http://www.ii.uib.no/borko
Department of computer science
University of Bergen, Norway
Exponential Time Differencing and Lie-Group Methods for Stiff Problems p.1/29
Contents
Introduction
Exponential Time Differencing (ETD),

connection with Lie-group integrators
Numerical schemes of 3
th
and 4
th
order in
time
Stability
Numerical experiments
Conclusions
Introduction
Consider partial differential equation of the form
u
t
= Lu +N(u, t)
where L is a higher-order linear term and N is low-order
nonlinear term.
After discretization in space we obtain a systems of ODEs
u = Lu +N(u, t)
Introduction
Consider partial differential equation of the form
u
t
= Lu +N(u, t)
where L is a higher-order linear term and N is low-order
nonlinear term.
After discretization in space we obtain a systems of ODEs
u = Lu +N(u, t)
Basic methods
LI = Linearly Implicit Use an explicit multi-step formula to advance the

nonlinear part and implicit scheme to advance the linear part (Ascher, Ruuth and
Wetton 95).
SS = Split Step Write the solution in the form

u(t) exp(c
1
tL)F(d
1
tN)exp(c
2
tL)F(d
2
tN) exp(c
k
tL)F(d
1
kN)u(0),
where c
i
andd
i
are constants (Sanz-Serna and Calvo 94, Boyd 01).
SL = Sliders Extent ion of LI, split the linear part into three regions (low,
medium and high), (Fornberg and Driscoll 99)
Basic methods

Wetton 95).

u(t) exp(c
1
tL)F(d
1
tN)exp(c
2
tL)F(d
2
tN) exp(c
k
tL)F(d
1
kN)u(0),
where c
i
andd
i
Basic methods

Wetton 95).

u(t) exp(c
1
tL)F(d
1
tN)exp(c
2
tL)F(d
2
tN) exp(c
k
tL)F(d
1
kN)u(0),
where c
i
andd
i
Basic methods
IF = Integrating Factor Make a change of variable then solve exactly for the
linear part and use the numerical scheme for the transformed nonlinear equation
(Mayday, Patera, Rnquist 90, Milewski and Tabak 99, Trefethen 00, Krogstad
03).
v = exp(Lt)u
exp(Lt)u
t
Lu
. .
v
t
= exp(Lt)N(u)
v
t
= exp(Lt)N(exp(Lt)v)
ETD = Exponential Time Differencing

Basic methods
03).
v = exp(Lt)u
exp(Lt)u
t
Lu
. .
v
t
= exp(Lt)N(u)
v
t

Basic methods
03).
v = exp(Lt)u
exp(Lt)u
t
Lu
. .
v
t
= exp(Lt)N(u)
v
t

Exponential Time Differencing
(1)
u = Lu +N(u(t), t)
u(t
0
) = u
0
Use the same integrating factor like in IF
d
dt
(exp(Lt)u) = exp(Lt)N(u(t))
Integrate over a single time step of length h
u
n+1
= exp(Lh)u
n
+exp(Lh)
_
h
0
exp(L)N(u(t
n
+), t
n
+ )d
Exponential Time Differencing
(1)
u = Lu +N(u(t), t)
u(t
0
) = u
0
Use the same integrating factor like in IF
d
dt
(exp(Lt)u) = exp(Lt)N(u(t))
Integrate over a single time step of length h
u
n+1
= exp(Lh)u
n
+exp(Lh)
_
h
0
exp(L)N(u(t
n
+), t
n
+ )d
ETD (1/2)
N(u(t), t) = constant
ETD1: u
n+1
= exp(Lh)u
n
+L
1
N
n
(exp(Lh) I), where N
n
= N(u
n
, t
n
)
N(u(), ) = N
n
+

h
(N
n
N
n1
)
ETD2:
u
n+1
= exp(Lh)u
n
+ h
1
L
2
N
n
(exp(Lh)(I +Lh) 2Lh I)
+h
1
L
2
N
n1
(exp(Lh) +Lh +I)
In general
ETDs: u
n+1
= exp(Lh)u
n
+h
s1
m=0
g
m
m
k=0
(1)
k
_
m
k
_
N
nk
, where g
m
are given by
Lhg
0
= exp(Lh) I
Lhg
m+1
+I = g
m
+
1
2
g
m1
+
1
3
g
m2
+ +
g
0
m+1
, m 0.
ETD (1/2)
ETD1: u
n+1
= exp(Lh)u
n
+L
1
N
n
n
= N(u
n
, t
n
)
N(u(), ) = N
n
+

h
(N
n
N
n1
)
ETD2:
u
n+1
= exp(Lh)u
n
+ h
1
L
2
N
n
+h
1
L
2
N
n1
(exp(Lh) +Lh +I)
In general
ETDs: u
n+1
= exp(Lh)u
n
+h
s1
m=0
g
m
m
k=0
(1)
k
_
m
k
_
N
nk
, where g
m
are given by
Lhg
0
= exp(Lh) I
Lhg
m+1
+I = g
m
+
1
2
g
m1
+
1
3
g
m2
+ +
g
0
m+1
, m 0.
ETD (1/2)
ETD1: u
n+1
= exp(Lh)u
n
+L
1
N
n
n
= N(u
n
, t
n
)
N(u(), ) = N
n
+

h
(N
n
N
n1
)
ETD2:
u
n+1
= exp(Lh)u
n
+ h
1
L
2
N
n
+h
1
L
2
N
n1
(exp(Lh) +Lh +I)
In general
ETDs: u
n+1
= exp(Lh)u
n
+h
s1
m=0
g
m
m
k=0
(1)
k
_
m
k
_
N
nk
, where g
m
are given by
Lhg
0
= exp(Lh) I
Lhg
m+1
+I = g
m
+
1
2
g
m1
+
1
3
g
m2
+ +
g
0
m+1
, m 0.
ETD (2/2)
Another way of producing the same formulas (Krogstad 03)
Lemma 1 Let N
k
=
d
k
dt
k
t = t
0
N(u(t), t) then the solution u(t) of (1) is
u(t
0
+ h) = exp(Lh)u
0
+
k=0
t
k+1
k
(tL)N
k
,
where
0
(z) =
exp(z) 1
z
,
k+1
(z) =

k
(z)
k
(0)
z
, k = 1, 2, . . .
Lie-group Methods
Crouch and Grossman Methods
Munthe Kaas Methods
Commutator free Lie Group Methods

Algorithm (CF)
for r = 1 : s do
Y
r
= Exp(
k

k
r,J
F
k
) Exp(
k

k
r,1
F
k
)(p)
F
r
= hF
Y
r
= h
i
f
i
(Y
r
)E
i
end
y
1
= Exp(
k

k
J
F
k
) Exp(
k

k
1
F
k
)p
Lie-group Methods
Munthe Kaas Methods

Algorithm (CF)
for r = 1 : s do
Y
r
= Exp(
k

k
r,J
F
k
) Exp(
k

k
r,1
F
k
)(p)
F
r
= hF
Y
r
= h
i
f
i
(Y
r
)E
i
end
y
1
= Exp(
k

k
J
F
k
) Exp(
k

k
1
F
k
)p
Lie-group Methods
Munthe Kaas Methods

Algorithm (CF)
for r = 1 : s do
Y
r
= Exp(
k

k
r,J
F
k
) Exp(
k

k
r,1
F
k
)(p)
F
r
= hF
Y
r
= h
i
f
i
(Y
r
)E
i
end
y
1
= Exp(
k

k
J
F
k
) Exp(
k

k
1
F
k
)p
ETD1 and Lie-group Integrators
We can rewrite the equation (1) in the form
u = (L, N(u, t)).u = F

u,t
(u)
u(0) = u
0
where . represents the Lie algebra action
(A, a).u = Au + a
Let F
u,
t
(u) be the Frozen Vector Field at the point ( u,
t)
F
u,
t
(u) = (L, N( u,
t)).u = Lu +N( u,
t)
The ow of such vector eld is the solution of u = F
u,
t
(u), u(0) = u
0
Exp(tF
u,
t
).u
0
= exp(tL)u
0
+t
0
(tL)N( u,
t)
u = (L, N(u, t)).u = F

u,t
(u)
u(0) = u
0
(A, a).u = Au + a
Let F
u,
t
t)
F
u,
t
(u) = (L, N( u,
t)).u = Lu +N( u,
t)
u,
t
(u), u(0) = u
0
Exp(tF
u,
t
).u
0
= exp(tL)u
0
+t
0
(tL)N( u,
t)
u = (L, N(u, t)).u = F

u,t
(u)
u(0) = u
0
(A, a).u = Au + a
Let F
u,
t
t)
F
u,
t
(u) = (L, N( u,
t)).u = Lu +N( u,
t)
u,
t
(u), u(0) = u
0
Exp(tF
u,
t
).u
0
= exp(tL)u
0
+t
0
(tL)N( u,
t)
ETD2 and Lie-group Integrators (1/3)
y = (A
u,t
, a).y =

F
u,t
(y)
y(0) = y
0
where
A
u,t
=
_
_
_
_
L
N(u,t)N
0
t
0 0
_
_
_
_
, a =
_
_
_
_
N
0
1
_
_
_
_
,
y =
_
_
_
_
u
v
_
_
_
_
, y
0
=
_
_
_
_
u
0
0
_
_
_
_
, N
0
= N(u
0
, 0)
Let

F
u,
t
(y) be the Frozen Vector Field at the point ( u,
t)
F
u,
t
(u) = (A
u,
t
, a).y = A
u,
t
y +a,
where
A
u,
t
=
_
_
_
_
L

NN
0
t
0 0
_
_
_
_
,

N = N( u,
t)
The ow of such vector eld is the solution of y =

F
u,
t
(y), y(0) = y
0
Exp(t
F
u,
t
).y
0
=
_
_
exp(tL)u
0
+t
0
(tL)N
0
+ t
2
1
(tL)

NN
0
t
t
_
_
Let

F
u,
t
(y) be the Frozen Vector Field at the point ( u,
t)
F
u,
t
(u) = (A
u,
t
, a).y = A
u,
t
y +a,
where
A
u,
t
=
_
_
_
_
L

NN
0
t
0 0
_
_
_
_
,

N = N( u,
t)
The ow of such vector eld is the solution of y =

F
u,
t
(y), y(0) = y
0
Exp(t
F
u,
t
).y
0
=
_
_
exp(tL)u
0
+t
0
(tL)N
0
+ t
2
1
(tL)

NN
0
t
t
_
_
The commutator between two vector elds is
[(A, a).u , (B, b).u] = (C, c).u,
where C = [A, B], c = (Ab Ba)
3
th
order RK schemes
Classical third-order RK method:

0
1
2
1
2
1 - 1 2
1
6
2
3
1
6
ETDRK3: (Cox & Matthews)

0
1
2
1
2
0
1 -
0
2
0
4
2
3
1
+
0
8
2
+ 4
1
4
2
1
3
th
order RK schemes
Classical third-order RK method:

0
1
2
1
2
1 - 1 2
1
6
2
3
1
6
ETDRK3: (Cox & Matthews)

0
1
2
1
2
0
1 -
0
2
0
4
2
3
1
+
0
8
2
+ 4
1
4
2
1
3
th
order CF schemes
CF3: (Celledoni, Martinsen, Owren)

0
1
3
1
3
2
3
0
2
3
1
3
1
12
3
4
ETDCF3:
0
1
3
1
3
0
2
3
0
2
3
0
1
3
1
12
0
3
4
0
3
th
order CF schemes
CF3: (Celledoni, Martinsen, Owren)

0
1
3
1
3
2
3
0
2
3
1
3
1
12
3
4
ETDCF3:
0
1
3
1
3
0
2
3
0
2
3
0
1
3
1
12
0
3
4
0
3
th
order schemes and ETD2
ETD2RK3:
0
1
2
1
2
0
1
0
4
1
4
1
4
2
3
1
+
0
8
2
+ 4
1
4
2
ETD2CF3:
0
1
3
1
3
0
2
3
2
3
4
3
1
4
3
1
9
2
9
2
1
+
0
18
2
+ 6
1
9
2
3
2
1
3
th
order schemes and ETD2
ETD2RK3:
0
1
2
1
2
0
1
0
4
1
4
1
4
2
3
1
+
0
8
2
+ 4
1
4
2
ETD2CF3:
0
1
3
1
3
0
2
3
2
3
4
3
1
4
3
1
9
2
9
2
1
+
0
18
2
+ 6
1
9
2
3
2
1
4
th
order schemes (1/4)
CF4:
(Celledoni, Martinsen, Owren)
0
1
2
1
2
1
2
0
1
2
1
2
1
2
0 0
1
2
-
1
2
0 1
1
4
1
6
1
6
-
1
12
-
1
12
1
6
1
6
1
4
4
th
order schemes (2/4)
ETDRK4:
(Krogstad)
0
1
2
1
2
0
1
2
0
1
2
0
1
2
1
2
0
0 0
1
2
-
1
2
0
0
0
4
2
3
1
+
0
4
2
+ 2
1
4
2
+ 2
1
4
2
1
4
th
order schemes (3/4)
ETDRK4B:
(Krogstad)
0
1
2
1
2
0
1
2
1
2
1

0
1
0
2
1
0 2
1
4
2
3
1
+
0
4
2
+ 2
1
4
2
+ 2
1
4
2
1
4
th
order schemes (3/4)
ETDCF4:
0
1
2
1
2
0
1
2
0
1
2
0
1
2
1
2
0
0 0
1
2
-
1
2
0
0
0
1
4
0
1
6
0
1
6
0
-
1
12
0
-
1
12
0
1
6
0
1
6
0
1
4
0
Stability
We consider the model equation
u = cu +u
u(0) = 1
where c and are scalars
For computing the
i
functions we use the approach suggested by Kassam and
Trefethen
i
(L) =
1
2i
_
i
(t)(tI L)
1
dt,
where is a contour well separated from 0
Stability
We consider the model equation
u = cu +u
u(0) = 1
where c and are scalars
For computing the
i
functions we use the approach suggested by Kassam and
Trefethen
i
(L) =
1
2i
_
i
(t)(tI L)
1
dt,
where is a contour well separated from 0
Stability regions for 3
th
order schemes
5 4 3 2 1 0 1 2 3 4 5
5
4
3
2
1
0
1
2
3
4
5
Stability region for ETDRK3
5 4 3 2 1 0 1 2 3 4 5
5
4
3
2
1
0
1
2
3
4
5
Stability region for ETD2RK3
Stability regions for 4
th
order schemes
5 4 3 2 1 0 1 2 3 4 5
5
4
3
2
1
0
1
2
3
4
5
Stability region for ETDRK4
5 4 3 2 1 0 1 2 3 4 5
5
4
3
2
1
0
1
2
3
4
5
Stability region for ETDRK4B
Example 1: Test equation
u = 1.5u + 2t
2
t + u
u(0) = 2
L = 1.5 and N = 2t
2
t +u
Test equation
10
3
10
2
10
1
10
0
10
12
10
10
10
8
10
6
10
4
10
2
10
0
10
2
Stepsize h
E
r
r
o
r

a
t

t
i
m
e

=

6
5
Test equation
ETDCF3
ETDRK3
ETD2CF3
ETD2RK3
10
3
10
2
10
1
10
0
10
12
10
10
10
8
10
6
10
4
10
2
10
0
Stepsize h
E
r
r
o
r

a
t

t
i
m
e

=

6
5
Test equation
ETDRK4
ETDCF4
ETDRK4B
Example 2: Kuramoto-Sivashinsky equation
u
t
= uu
x
u
xx
u
xxxx
x [0, 32]
u(x, 0) = cos(
x
16
)(1 +sin(
x
16
))
We discretise the spatial part using Fourier spectral method. The transformed equation
in the Fourier space is
u
t
=
ik
a
u
2
+ (k
2
k
)
u,
(L u)(k) = (k
2
k
4
) u(k) and N( u, t) =
ik
2
(F((F
1
( u))
2
))
Example 2: Kuramoto-Sivashinsky equation
u
t
= uu
x
u
xx
u
xxxx
x [0, 32]
u(x, 0) = cos(
x
16
)(1 +sin(
x
16
))
We discretise the spatial part using Fourier spectral method. The transformed equation
in the Fourier space is
u
t
=
ik
a
u
2
+ (k
2
k
)
u,
(L u)(k) = (k
2
k
4
) u(k) and N( u, t) =
ik
2
(F((F
1
( u))
2
))
Kuramoto-Sivashinsky equation
10
2
10
1
10
0
10
5
10
4
10
3
10
2
10
1
10
0
Stepsize h
R
e
l
a
t
i
v
e

e
r
r
o
r

a
t

t
i
m
e

=

6
5
KuramotoSivashinsky equation
ETDRK3
ETD2RK3
ETD2CF3
10
2
10
1
10
0
10
7
10
6
10
5
10
4
10
3
10
2
10
1
10
0
Stepsize h
R
e
l
a
t
i
v
e

e
r
r
o
r

a
t

t
i
m
e

=

6
5
KuramotoSivashinsky equation
ETDRK4
ETDCF4
ETDRK4B
Conclusion
We present a constructive way of generating exponential time differencing

Runge-Kutta (ETDRK) methods from Lie group methods. In particular we show
how to choose the algebra action in order to recover the classical ETD and ETDRK
methods.
We discuss how to choose the frozen vector eld in order to receive better
approximation of the nonlinear part.
Better stability properties
The above approach can be successfully applied when the nonlinear part of the
system is periodic.
Conclusion

methods.
system is periodic.
Conclusion

methods.
system is periodic.
Conclusion

methods.
system is periodic.
References
S.M. Cox, P.C. Matthews, Exponential time differencing for stiff systems, J. Comp.
Phys. 176, 430-455 (2002)
A.K. Kassam, L.N. Trefethen, Fourth-order time stepping for stiff PDEs, Submitted
to SIAM J. Sci. Comput. (2002)
E. Celledoni, A. Martinsen, B. Owren, Commutator-free lie group methods, FGCS

19(3), 341-352 (2003)
S. Krogstad, Generalized integrating factor methods for stiff PDEs, available at:
http://www.ii.uib.no/stein

Sci Cade 03

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Sci Cade 03

Uploaded by

Copyright:

Available Formats

Exponential Time Differencing and Lie-Group

Methods for Stiff Problems

Exponential Time Differencing (ETD),

LI = Linearly Implicit Use an explicit multi-step formula to advance the

SS = Split Step Write the solution in the form

LI = Linearly Implicit Use an explicit multi-step formula to advance the

SS = Split Step Write the solution in the form

LI = Linearly Implicit Use an explicit multi-step formula to advance the

SS = Split Step Write the solution in the form

ETD = Exponential Time Differencing

ETD = Exponential Time Differencing

ETD = Exponential Time Differencing

Crouch and Grossman Methods

Munthe Kaas Methods

Commutator free Lie Group Methods

Crouch and Grossman Methods

Munthe Kaas Methods

Commutator free Lie Group Methods

Crouch and Grossman Methods

Munthe Kaas Methods

Commutator free Lie Group Methods

u = (L, N(u, t)).u = F

u = (L, N(u, t)).u = F

u = (L, N(u, t)).u = F

Classical third-order RK method:

ETDRK3: (Cox & Matthews)

Classical third-order RK method:

ETDRK3: (Cox & Matthews)

CF3: (Celledoni, Martinsen, Owren)

CF3: (Celledoni, Martinsen, Owren)

We present a constructive way of generating exponential time differencing

Better stability properties

We present a constructive way of generating exponential time differencing

Better stability properties

We present a constructive way of generating exponential time differencing

Better stability properties

We present a constructive way of generating exponential time differencing

Better stability properties

E. Celledoni, A. Martinsen, B. Owren, Commutator-free lie group methods, FGCS

You might also like