Discontinuous Galerkin Method for Gravitational ... · between adjacent elements K. Fan, W. Cai, X....
Transcript of Discontinuous Galerkin Method for Gravitational ... · between adjacent elements K. Fan, W. Cai, X....
Discontinuous Galerkin Method for Computing Perturbations of a
Schwarzschild Spacetime
Scott FieldDepartment of Physics
Brown Universitywith Jan Hesthaven (adviser) and Stephen Lau
Organization of PresentationOrganization of Presentation
• Overview of Perturbation Treatment of the Extreme Mass Ratio Binary (EMRB) Problem
• Discontinuous Galerkin Method • Numerical Scheme for Perturbation
Equations• Results• Conclusion and Future Work
Extreme Mass Ratio Binary
Physical RelevancePhysical Relevance
• EMRBs expected to be primary source of gravitational waves detected by LISA– “Small” star of mass orbiting a super-
massive black hole of where • Motivation for considering new high-order
(in this particular case spectral) numerical methods: Construction of quality gravitational waveforms in the shortest amount of time.
pm1/ < <= MmpµM
Background and AssumptionsBackground and Assumptions
• Point particle in stable orbit of SMBH– Schwarzschild– Particle follows a time-like geodesic in equatorial plane– Stress energy tensor
• Particle causes small metric perturbations– Metric– Perturbations can be combined into master-functions
which evolve according to
)]([4 τδτ βαα β zxuug
dmTSchp −
−= ∫
222122 Ω++−= − drdrffdtds rMf /21−=
α βh⇒+= α βα βα β hgg Sch
α βα β πδ ThGLin 8)( =
)(),()(),(),())(( ////22 rprrtFrprrtGtxrV rPAPAPAPA
xt −∂+−=−∂+− ∂ δδψ
Equations and QuantitiesEquations and Quantities
• In the distant wave-zone and at BH horizon:
• Metric perturbations can be reconstructed everywhere in space
2||)!2()!2(
641
lmlm llE ψ
π
−+= lmlmlm l
limL ψψπ
*
)!2()!2(
64 −+=
lmAlm
Plm
mlx Yi
ll
rihh 2
,2
)!2()!2(
21
−≥
+ +−+=− ∑ ψψ
Discontinuous Galerkin Method
Numerical Simulations of EMRBsNumerical Simulations of EMRBs• Numerous proposed methods, such as…
– Frequency-domain• Fourier decomposition• Becomes numerically expensive when one must sum
over many modes – e.g. eccentric orbits– Time-domain
• Typically spatial derivatives are approximated by finite differences
– Errors fall off as a power– Long time integrations can produce phase errors in low order
methods• Distributional source term can be troublesome• Large computational domain-size
– Waveform/BC issues
DG: A Hybrid of MethodsDG: A Hybrid of Methods
• Finite element method: if Dk = Ω and basis functions are “coupled” (ie span multiple elements) we get a finite element method
•F
• Legendre collocation, “Pseudospectral,” on each element
• When basis functions are constants, dG formally becomes a finite volume method
Figure created by Oleg Alexandrov
Spectral and Finite Difference ErrorsSpectral and Finite Difference Errors• 1D advection
equation with speed 2pi, on a domain 0 to 2pi. Integration with RK4
)2sin()( txexu π−=
C
Spectral Methods for Time Dependent Problems by Jan Hesthaven & David Gottlieb
Advection: Phase ErrorsAdvection: Phase Errors
• When evolving hyperbolic PDEs for long times, phase errors can become significant
• An example with the advection equation
Numerical solution propagate at wrong speed
)exp()0,( :
0
ikxxICx
ct
=
=∂
∂+∂
∂
ψ
ψψ
)](exp[),( ctxiktx −=→ ψ
)exp( :
0
jj
jNumj
ikxIC
cDt
=
=+∂
∂
ψ
ψψ
)](exp[)( tcxikt numjj −=→ ψ
Advection: Phase ErrorsAdvection: Phase Errors
• If using second order central difference
we have a phase error of
– Typically phase errors scale as – To achieve small phase error one must:
• Run for short times• Use a fine grid• High order method
)(211
xD jj
jNum ∆−
= −+ ψψψ
6)(sin1)exp()exp(
2xkktcxk
xkktctikcikct num∆=
∆∆−=−−−
orderxkt )( ∆∝
Advection: Phase ErrorsAdvection: Phase Errors
• Using a 4th order Runge-Kutta time integrator, evolve a Gaussian packet using first and second order approximations to
• In each example we have chosen x∂
∂
5.~x∆
11stst Order Upwind Differencing Order Upwind Differencing
22ndnd Order Central Differencing Order Central Differencing
Discontinuous GalerkinDiscontinuous Galerkin
• Domain Ω approx. by local elements Dk
• Local solution is a linear combination of basis functions
– Form a basis for space of polynomials in Dk of degree at most N
• Global solution is a direct sum of local solutions
nP
)( ),()()(ˆ),( :00
xltxuxPtutxuDx ki
N
ii
khn
N
n
kn
kh
k ∑∑==
==∈
),(),(),(1
txutxutxu kh
K
kh =⊕=≈
K
k
kh D
1=
=Ω≈Ω
)(0: kNn DPNnPSpan =≤≤
Discontinuous GalerkinDiscontinuous Galerkin• Given an operator L such that can be
cast in integral form on elements Dk
– Require of numerical solution uh
– Galerkin condition: test and basis functions taken to be the same
To resolve: coupling between elements?
∫ ≤≤=kD
kn
kh NndxvLu 0 ;0)(
0=Lu0)( =∫ vdxLu
kD
N
Nodal Discontinuous Galerkin Methods by Jan Hesthaven and Tim Warburton
Numerical FluxNumerical Flux
• To couple elements first perform IBPs
• Numerical flux – and are the exterior and interior solutions– passes information between elements, ensures
stability of scheme, implements boundary conditions– Choice of related to dynamics of PDE system
• Consistency condition:
• Example:– Central flux
vdxufndxvufuvvdxufu hDD
xkk
tD
kx
kt
kkhh
khh
))((ˆ))((0))(( *⋅−=∂−∂→=∂+∂ ∫∫∫∂
),()( ** −+=hh
uufuf h+h
u −h
u
)(*huf
),()( *hhh uufuf =
2)()()(*
+− += hhh
ufufuf
Example: Advection with a Distributional Source Term
Problem StatementProblem Statement
• Each sector of the perturbation equations similar in spirit to 2 copies of advection
• Equation:
• Domain:• BC: • Initial data:• Jump condition:• Sends information G(t) off to the right at speed c
)/()(),( 00 cxtGxtx −Θ=ψ
0)()(1 =−∂
∂+∂
∂ xtGxtc
δψψ
)()0()0( tG=− −+ ψψ
baba <<=Ω 0 s.t. ],[0),( =taψ
) , ,( paritym
Numerical SchemeNumerical Scheme
• Basis space: Lagrange interpolating polynomials of order N on defined on each element
– Legendre-Gauss-Lobatto nodal points η • Integrate with fourth order Runge-Kutta• Numerical flux:
– PDE shifts solution to the right, so use an “upwinded” form
∏≠= −
−=
N
ijj ji
ji
xxl
0
)(ηηη
)2ˆ
2ˆ
()(*−−−+++ ⋅++⋅+= ψψψψψ nncf h
A Generalized DG (GDG) MethodA Generalized DG (GDG) Method• GDG extends DG to solutions (analytically)
discontinuous at an interface• Idea: treat the δ function in the PDE as an
additional numerical flux term in an appropriate way, ie mimic the physics of the PDE– Require usual delta property over Ω
– Freedom to choose how to “split” the delta function between adjacent elements
K. Fan, W. Cai, X. Ji, A generalized discontinuous Galerkin (GDG) method for Schrodinger equations with nonsmooth solutions, J. Comp. Phys., 227 (2008) 2387-2410.
)0()()( vdrxvx =∫Ω
δ
)0()0()0()()()()()()(11
vbvavdrxvxdrxvxdrxvxiiii DDDD
=+=+= ∫∫∫++
δδδ
Treating an Advected Treating an Advected δδ Source Source
• Upwind the source:
• GDG principle suggests we add a flux term to the existing flux term on the right hand side of the interface
δ*,,1)( LicG +
0*)]()()[( =∫ dxxtcGxliD
δ
)0()()]()()[(1
ltcGdrxtcGxliD
=∫
+
δ
*)],([ txcψ
Comparison With Analytic SolutionComparison With Analytic Solution
• Analytic solution is– Choose c = 2, G(t) = cos(t), final time = 15
)()(),( cxtGxtx −Θ=ψ
Exponential convergenceExponential convergence • Final time = 30• 4 domains on [-5,5]• Time-step set below CFL
requirement of largest N
44thth order RK order RK • Final time = 30• 30 Domains on [-5,25]• Polynomial Order = 10
A Numerical Schemefor the Perturbation Equations
Issues and ResolutionsIssues and Resolutions
• Particle must be located at an interface for spectral accuracy to be maintained– Coordinate transformation
• The derivative of a delta function term differentiates the test functions– Rewrite equations in full first order
• Non-trivial boundary conditions– Exact outgoing boundary conditions
• Computing waveform at null infinity– Use flat space extraction (Schwarzschild
solution asymptotically flat)
Rewriting the Equations in First Rewriting the Equations in First Order SystemOrder System
• As first order system, and specializing to
• Using properties of distributions, one can show this is the first order form of the original equation
)()(
)()()(
pxt
pxt
t
xxtJxxtJrV
−+Π− ∂=Φ∂
−+−Φ− ∂=Π∂Π−=∂
Π
Φ
δδψ
ψ
(t)J |][ (t)J |][
)(),()(),())(( 22
Φ=Π= =∂=− ∂
−∂+−=−∂+− ∂
pp xxxxxt
prpxt rrrtFrrrtGrVψψ
δδψ
pp rtr =)(
Semidiscrete Form Semidiscrete Form
upwind numerical flux (generalization of advection example)
−−+
−
Π−+
ΠΦ− ∂=
ΦΠ∂
Π
Φ
)(),()(),(
0
0
0
rprtrJrprtrJVxt
δδψ
ψ
)(),(),(0
xltxutxu ki
N
ii
kh
kh ∑
=
=
dxuHxldxJxldxufxlndxxlufdxtxuxlkkkkk D
hiD
iD
hiD
ixhD
hti ∫∫∫∫∫ ++⋅−∂=∂∂
)()()()()(ˆ)()(),()( *
JuHufu xt
++− ∂=∂ )()(
)(*huf
LGL Nodal Points +
Treating the Source TermsTreating the Source Terms
• Diagonalize the sourced sector of the system
• GDG suggests adding to on the RHS of the interface and to on the LHS of the interface
• Numerical flux at particle interface is changed as
λωλω
−=Π+=Φ
2/)()(
/2)()(
pxt
pxt
xxJJxxJJ
−−=∂−∂
−+=∂+∂
ΦΠ
ΦΠ
δλλδωω
δδ ω *,*, )())(5(. RRJJ =+ ΦΠ*)(ω
δδ λ *,*, )())(5(. LLJJ =− ΦΠ*)(λ
δδ λω *,*,** )()( LR ++Φ→Φ δδ λω *,*,** )()( LR −+Π→Π
Boundary ConditionsBoundary Conditions
• Non-reflecting BC– Reduces domain size, especially useful for eccentric
orbits where many periods required– Sommerfeld BC works well near BH horizon as– At the right boundary Sommerfeld fails as
instead
– depends on form of potential – Approximated using known techniques to satisfy a
specifiable error tolerance, here
Stephen Lau. Rapid Evaluation of Radiation Boundary Kernels for Time-domain Wave Propagation on Blackholes. gr-qc/0401001
MxeV 2/~2/1~ rV
),,()( ψψ bxt rtF=∂+∂
1010−<ε
),,( ψbrtF
Sketch of BC in Flat 3+1 SpaceSketch of BC in Flat 3+1 Space
• For given ell, write down an arbitrary outgoing solution in the Laplace frequency domain
• Apply Sommerfeld operator• Rewrite result in form
• Taking the inverse Laplace transform
• Theorem: In the time-domain, the kernel is a sum of exponentials
rs ∂+
),(),(ˆ1),()( rsrsr
rss outoutr ΨΩ=Ψ∂+
'),'(),'()(0
dTrTrTTC b
T
brt ∫ Ψ−Ω=Ψ∂+∂
( )∑=
≅Ωd
kkk t
1,, exp βγ
Boundary ConditionsBoundary Conditions
Long-time bound on error due to boundary conditions. Final-Time = 100M
Waveform MatchingWaveform Matching
• Want to “read off” at null infinity • Recording at introduces an error • suggests that• At right boundary we match waveforms as ( )
• This is an ODE sourced by and approximates the waveform at infinity – Reduces the error to
)( 3SpaceFlat A/P −+= rOVV SpaceFlat A/P ~ψψ
)( 2−xO
)(3)(3)(~ 2)1()2(SpaceFlat
2A/P
2 bb
bb
b xtfx
xtfx
xtf −+−+−=== ψψ
A/PψA/Pψ bx )( 1−
bxO
2=
)()2( xtf −A/P2=ψ
Initial DataInitial Data
• Typically trivial initial data supplied, but this is inconsistent with the PDEs
• “Switch-on” the source terms by multiplying them by a function which smoothly interpolates from 0 to 1– We used [ ]1))((
21 +− τδ terf
Initial Data: 1+1 Wave Equation Initial Data: 1+1 Wave Equation
• Solution to 1+1 wave equation with distribution source term– Have analytic
solution• Without smoothing 4th
order temporal convergence is abruptly lost (red). It is recovered with a smoother (blue)
Features of ImplementationFeatures of Implementation
• Spectral global accuracy, even at particle– Potentially useful for incorporating a self-force– Excellent phase resolution for long integrations
• Reduced computational domain without sacrificing accuracy of waveform or introducing reflection– Very useful for problems without periodicity
• Particle trajectory need not be a priori• No length scale associated with particle
Results
Results: Circular OrbitResults: Circular Orbit0)0,( ,]1000 ,200[* ,9456.7 ,1 ,)2,2(),( =−=Ω=== xurMm pBH
Energy and Angular Momentum Energy and Angular Momentum LuminosityLuminosity
lmmlm llimL ψψ
π
)!2()!2(
64 −+=
2||)!2()!2(
641
lmlm llE ψ
π
−+=
Results: Eccentric OrbitResults: Eccentric Orbit
2436.4887153.22
0682.52873.8
63517.
max
min
==
==
=−=
ϕTTrr
latussemityeccentrici
r
0)0,( ,]1000 ,200[* ,1 ,)2,2(),( =−=Ω== xuMm BH
Eccentric Orbit WaveformsEccentric Orbit Waveforms )2,2(),( =m
Eccentric Orbit QuantitiesEccentric Orbit Quantities
Selected energy and angular momentum flux calculated at null infinity
• Averages computed according to
r
T
Tfm
fm TTTdtE
TTE
f
4 10
0 0
=−−
> =< ∫
What's Next?What's Next?• Short term:
– Calculate self force• Done for and when orbit quasi-circular• For quasi-circular inspiral orbits would like to calculate
waveforms due to dissipation• Long term:
– Exact BC for left side of computational domain • Arbitrarily short domain size
– Could calculate up-to ell=25 for 200 orbits of particle initially at r=7.9M in ~3 weeks
» On Dell laptop, coded in Matlab» Optimistically on a GPU with python code one might
achieve a speed up factor of 50 or more
– Waveform extraction• Weak link, suggestions?
lmE lmL
Summary and KerrSummary and Kerr• Purposed a high order DG method for EMRBs
– Equations as a first order system, exact outgoing radiative BCs, waveform matching, and smooth startup
– Results agree with literature to stated accuracy– Global spectral accuracy
• EMRB problem for Kerr solution– 2+1 dimensional– 2D problems with distributional sources have been
considered in the problem by Fan et al