adaptivity and symmetry for odes and pdes chris budd
TRANSCRIPT
Adaptivity and symmetry for ODEs and PDEs
Chris Budd
Talk will look at
• variable step size adaptive methods for
ODES• scale invariant adaptive methods for PDES
Basic Philosophy …..
• ODES and PDEs develop structures
on many time and length scales
• Structures may be uncoupled (eg. Gravity waves
and slow weather evolution) and need multi-scale
methods
• Or they may be coupled, typically through (scaling)
symmetries and can be resolved using adaptive
methods
Conserved quantities:
Symmetries: Rotation, Reflexion, Time reversal, Scaling
Kepler's Third Law
Hamiltonian Angular Momentum
The need for adaptivity: the Kepler problem
2/3222
2
2/3222
2
)(,
)( yx
y
dt
yd
yx
x
dt
xd
yuxvLyx
vuH
,)(
)(2
12/122
22
),(),(),,(),(, 3/13/2 vuvuyxyxtt
Kepler orbits
Stormer
Verlet
Symplectic
Euler
Forward
Euler
Global error
H error
FE
SV
t
Main error
Larger error at close approaches
Kepler’s third law is not respected
Adaptive time steps are highly desirable for
accuracy and symmetry
But …
Adaptivity can destroy the symplectic shadowing
structure [Calvo+Sanz-Serna]
Adaptive methods may not be efficient as a splitting
method
AIM: To construct efficient, adaptive, symplectic
methods EASY which respect symmetries
H error
t
The Sundman transform introduces a continuous
adaptive time step.
IDEA: Introduce a fictive computational time
),( qpgd
dt
qp Hdt
dpH
dt
dq ,
Hamiltonian ODE system:
),( qpg SMALL if solution requires small time-steps
qp Hqpgd
dpHqpg
d
dq),(,),(
))0(),0((, 0 qpHHqtp tt
)),()(,(),,,( ttt qqpHqpgqqppK
),( qpgd
dt
Can make Hamiltonian via the Poincare Transform
New variables
Hamiltonian
Rescaled system for p,q and t
Now solve using a Symplectric ODE solver
Choice of the scaling function g(q)
Performance of the method is highly dependent on
the choice of the scaling function g.
Approach: insist that the performance of the
numerical method when using the computational
variable should be independent of the scale of the
solution and that the method should respect the
symmetries of the ODE
),...,( 1 Nii uuf
dt
du
0,, uutt i
)()( tutu iii
The differential equation system
Is invariant under scaling if it is unchanged by the symmetry
It generically admits particular self-similar solutions
satisfying
eg. Kepler’s third law relating planetary orbits
),...,(),...,( 111
NN uuguug N
Theorem [B, Leimkuhler,Piggott] If the scaling function
satisfies the functional equation
Then
Two different solutions of the original ODE mapped
onto each other by the scaling transformation are
the same solution of the rescaled system scale
invariant
A discretisation of the rescaled system admits a
discrete self-similar solution which uniformly
approximates the true self-similar solution for all
time
Example: Kepler problem in radial coordinates
A planet moving with angular momentum
with radial coordinate r = q and with dr/dt = p
satisfies a Hamiltonian ODE with Hamiltonian
2
2 1
2 qq
pH
ppqqtt 3/13/2 ,, If symmetry
2/33/2 )()()( qqgqgqg
0
Numerical scheme is scale-invariant if
If there are periodic solutions with close
approaches
Hard to integrate with a non-adaptive scheme
10
q
t
2
2/3
1
0
)(
qqg
Consider calculating them using the scaling
No scaling
Levi-Civita scaling
Scale-invariant
Constant angle change
H Error
nopt 2
1
2
3
Method order
Surprisingly
sharp!!!
Scale invariant methods for PDES
These methods extend naturally to PDES with
scaling and other symmetries
uuyxyxtt
uuufu xxt
),,(),(,
,...),,,(
.),(,0),0(,)(
,)(
,),.(
,0
,
,
2
3
3
vttsutuvquu
uuu
vuvvvuuu
uuuiu
uuu
uuu
xm
t
xxm
t
tt
t
xxxxt
xxt
Examples
Parabolic blow-up
High-order blow-up
NLS
Chemotaxis
PME
Rainfall
Need to continuously adapt in time and space
Introduce spatial analogue of the fictive time
Adapt spatially by mapping a uniform mesh from a computational domain into a physical domain
Use a strategy for computing the mesh which takes symmetries into account
C P
),( C ),( yxP
Introduce a mesh potential ),,( tQ
,..),(),..)(),(( QQQtytx
DQQQ
DQ
QQyxQH
2
1
det),(
),()(
2
,..),,( yx uuuM
Geometric scaling
Control scaling via a measure
dt QHQMQI /1)()(
Spatial smoothing
(Invert operator using a spectral method)
Averaged measure
Ensures right-hand-side scales like P in d-dimensions to give global existence
Parabolic Monge-Ampere equation PMA
(PMA)
Evolve mesh by solving a MK based PDE
Because PMA is based on a geometric approach, it has natural symmetries
1. System is invariant under translations and rotations
2. For appropriate choices of M the system is invariant under scaling symmetries
LQQyxLyx ),(),(
ddtt QHQLMLLQHQLMQ
T
LQ /1/1 ))()(())()((,
PMA is scale invariant provided that
ddd tyxuMTTtyxLUuMQLM /11/1/1 )),,(())),,((()(
UuuyxLyxTtt ),,(),(,
2/12/12/1 )log(,),( TTLTUttT
),(),(, ** yxyxttu
3uuuu yyxxt
XdtYXutYXutYXM ddd 22 ),,(),,(),,(
Example: Parabolic blow-up in d dimensions
ddd uuMuMTuTM 2/11/12/1 )()()(
Scale:
Regularise:
uuyxyxtttt 2/12/1 ),,(),(),()(
Basic approach
• Discretise PDE and PMA in the computational domain
• Solve the coupled mesh and PDE system either
(i) As one large system (stiff!) or
(ii) By alternating between PDE and mesh
Method admits exact discrete self-similar solutions
solve PMA simultaneously with the PDE3uuu Xt
Mesh:
Solution:
XY
10 10^5
Solution in the computational domain
10^5
12
Same approach works well for the Chemotaxis eqns, Nonlinear Schrodinger eqn, Higher order PDEsNow extending it to CFD problems: Eady, Bousinessq