A Posteriori Error Estimates For Discontinuous ...linlin/presentations/201503_APost.pdf ·...

45
A Posteriori Error Estimates For Discontinuous Galerkin Methods Using Non-polynomial Basis Functions Lin Lin Department of Mathematics, UC Berkeley; Computational Research Division, LBNL Joint work with Benjamin Stamm Dimension Reduction: Mathematical Methods and Applications, Penn State University, March, 2015 Supported by DOE SciDAC Program and CAMERA Program 1 Lin Lin A Posteriori DG using Non-Polynomial Basis

Transcript of A Posteriori Error Estimates For Discontinuous ...linlin/presentations/201503_APost.pdf ·...

  • A Posteriori Error Estimates For Discontinuous Galerkin Methods Using

    Non-polynomial Basis Functions

    Lin LinDepartment of Mathematics, UC Berkeley; Computational Research Division, LBNL

    Joint work with Benjamin Stamm

    Dimension Reduction: Mathematical Methods and Applications, Penn State University, March, 2015

    Supported by DOE SciDAC Program and CAMERA Program

    1Lin Lin A Posteriori DG using Non-Polynomial Basis

  • Outline• Introduction: Adaptive local basis functions

    • Computable upper bound for Poisson’s equation

    • Computable upper / lower bound for indefinite equations

    • Numerical examples

    • Conclusion and future work

    Lin Lin 2A Posteriori DG using Non-Polynomial Basis

  • Motivation• Spatially inhomogeneous quantum systems

    Ω

    3Lin Lin A Posteriori DG using Non-Polynomial Basis

  • Kohn-Sham density functional theory

    • Efficient: Always solve an equation in 𝑅𝑅3, regardless of the number of electrons 𝑁𝑁.

    • Accurate: Exact ground state energy for exact 𝑉𝑉𝑥𝑥𝑥𝑥[𝜌𝜌], [Hohenberg-Kohn,1964], [Kohn-Sham, 1965]

    • Best compromise between efficiency and accuracy. Most widely used electronic structure theory for condensed matter systems and molecules

    • Nobel Prize in Chemistry, 1998

    4

    𝐻𝐻 𝜌𝜌 𝜓𝜓𝑖𝑖 𝑥𝑥 = −12Δ + 𝑣𝑣𝑒𝑒𝑥𝑥𝑒𝑒 𝑥𝑥 + ∫ 𝑑𝑑𝑥𝑥′

    𝜌𝜌 𝑥𝑥′

    𝑥𝑥 − 𝑥𝑥′+ 𝑉𝑉𝑥𝑥𝑥𝑥 𝜌𝜌 𝜓𝜓𝑖𝑖 𝑥𝑥 = 𝜀𝜀𝑖𝑖𝜓𝜓𝑖𝑖 𝑥𝑥

    𝜌𝜌 𝑥𝑥 = 2�𝑖𝑖=1

    𝑁𝑁/2

    𝜓𝜓𝑖𝑖 𝑥𝑥 2 , ∫ 𝑑𝑑𝑥𝑥 𝜓𝜓𝑖𝑖∗ 𝑥𝑥 𝜓𝜓𝑗𝑗 𝑥𝑥 = 𝛿𝛿𝑖𝑖𝑗𝑗 , 𝜀𝜀1 ≤ 𝜀𝜀2 ≤ ⋯

    Lin Lin A Posteriori DG using Non-Polynomial Basis

  • Discretization costBasis Example DOF / atom Construction

    Uniform basis PlanewaveFinite differenceFinite element

    500~10000 or more

    Simple and systematic

    Quantum chemistry basis

    Gaussian orbitals Atomic orbitals

    4~100 Fine tuning

    Non-systematic convergence

    Q: Combine the advantage of both?

    5Lin Lin A Posteriori DG using Non-Polynomial Basis

  • Adaptive local basis functions• Idea: Use local eigenfunctions as basis functions

    • How to patch the basis functions together?

    6Lin Lin A Posteriori DG using Non-Polynomial Basis

  • Discontinuous Galerkin method

    Kohn-Sham

    New terms

    • [LL-Lu-Ying-E, J. Comput. Phys. 231, 2140 (2012)] • Interior penalty method [Arnold, 1982]

    7Lin Lin A Posteriori DG using Non-Polynomial Basis

  • Why a posteriori error estimator• Measuring the accuracy of eigenvalues and densities

    without performing an expensive converged calculation, or benchmarking with another code.

    • Optimal allocation of basis functions for inhomogeneous systems.

    8Lin Lin A Posteriori DG using Non-Polynomial Basis

  • Residual based a posteriori error estimatorVast literature for second order PDE and eigenvalue problems

    • Polynomial basis functions, finite element:

    [Verfürth,1996] [Larson, 2000] [Durán-Padra-Rodríguez, 2003] [Chen-He-Zhou, 2011]...

    • Polynomial basis functions, discontinuous Galerkin:

    [Karakashian-Pascal, 2003], [Houston-Schötzau-Wihler, 2007], [Schötzau-Zhu, 2009], [Giani-Hall, 2012] ...

    9Lin Lin A Posteriori DG using Non-Polynomial Basis

  • Difficulty• A posteriori error analysis relies on the detailed

    knowledge of asymptotic approximation properties of the basis set

    • Difficult for “equation-aware” basis functions Adaptive local basis functions Heterogeneous multiscale method (HMM) [E-Engquist

    2003] Multiscale finite element [Hou-Wu 1997] Multiscale discontinuous Galerkin [Wang-Guzmán-

    Shu, 2011] etc

    Lin Lin 10A Posteriori DG using Non-Polynomial Basis

  • Outline• Introduction: Adaptive local basis functions

    • Computable upper bound for Poisson’s equation

    • Computable upper / lower bound for indefinite equations

    • Numerical examples

    • Conclusion and future work

    Lin Lin 11A Posteriori DG using Non-Polynomial Basis

  • Model problem

    Lin Lin 12

    Discontinuous space (broken Sobolev space)

    𝜅𝜅Ω

    𝕍𝕍𝑁𝑁 ⊂ 𝐻𝐻2(𝒦𝒦)

    𝐹𝐹

    A Posteriori DG using Non-Polynomial Basis

    Piecewise constant function belongs to 𝕍𝕍𝑁𝑁

  • DG discretizationBilinear form (𝜃𝜃 = 1 corresponds to the symmetric form)

    Define the inner products

    Average and jump operators

    Lin Lin 13

    ⋅ ⋅

    A Posteriori DG using Non-Polynomial Basis

  • Error quantificationDG approximation

    Error in the broken energy norm

    Goal: Find a sharp upper bound for

    Lin Lin 14A Posteriori DG using Non-Polynomial Basis

  • Upper bound of errorTheorem ([LL-Stamm 2015]). Let 𝑢𝑢 ∈ 𝐻𝐻#1 Ω ⋂𝐻𝐻2(𝒦𝒦) be the true solution and 𝑢𝑢𝑁𝑁 ∈ 𝕍𝕍𝑁𝑁 the DG-approximation. Then

    where

    The key is to find the dependence of 𝑎𝑎𝜅𝜅 , 𝑏𝑏𝜅𝜅 , 𝑐𝑐𝜅𝜅 w.r.t. 𝕍𝕍𝑁𝑁.

    Lin Lin 15

    ResidualJump of gradientJump of function

    A Posteriori DG using Non-Polynomial Basis

  • Projection operator𝐿𝐿2 𝜅𝜅 -projection operatorInner product

    Projection operator onto basis space

    Therefore

    Lin Lin 16A Posteriori DG using Non-Polynomial Basis

    Similar to𝐻𝐻1(𝜅𝜅) norm

  • Estimating constants Define

    ⊥ is in the sense of the inner product ⋅,⋅ ∗,𝜅𝜅Lemma. Let 𝜅𝜅 ∈ 𝒦𝒦, 𝑣𝑣 ∈ 𝐻𝐻1 𝜅𝜅 . Then

    Proof:

    Similar for 𝑏𝑏𝑘𝑘

    Lin Lin A Posteriori DG using Non-Polynomial Basis 17

  • Numerical procedure for computing the constants• Basic idea: estimate the constants by iteratively solving

    generalized eigenvalue problems on an infinite dimensional space

    • 1D demonstration, generalizable to any d-dimension. • Consider 𝜅𝜅 = 0,ℎ , spectral discretization with Legendre-

    Gauss-Lobatto (LGL) quadrature:

    Integration points 𝑦𝑦𝑗𝑗 𝑗𝑗=1𝑁𝑁𝑔𝑔 , integration weights 𝜔𝜔𝑗𝑗 𝑗𝑗=1

    𝑁𝑁𝑔𝑔

    Lin Lin 18A Posteriori DG using Non-Polynomial Basis

    0 ℎ𝑦𝑦𝑗𝑗

  • Numerical representation of inner productLGL grid points defines associated Lagrange polynomials of degree 𝑁𝑁𝑔𝑔 − 1

    Approximate any 𝑣𝑣 ∈ 𝐻𝐻1 𝜅𝜅

    Define

    Inner product

    Lin Lin A Posteriori DG using Non-Polynomial Basis 19

  • Numerical representation of inner product‖ ⋅ ‖∗,𝜅𝜅 requires differentiation matrix

    Differentiation becomes matrix-vector multiplication

    Lin Lin A Posteriori DG using Non-Polynomial Basis 20

  • Numerical representation of inner productProjection onto constant

    In sum

    Lin Lin A Posteriori DG using Non-Polynomial Basis 21

  • Estimating 𝑎𝑎𝑘𝑘

    Lin Lin A Posteriori DG using Non-Polynomial Basis 22

    Here

    Handling the orthogonal constraint by projection𝑄𝑄 = 𝐼𝐼 − Π𝑁𝑁𝜅𝜅

  • Estimating 𝑎𝑎𝑘𝑘

    This is a generalized eigenvalue problem

    Solve with iterative method, such as the Locally Optimal Block Preconditioned Conjugate Gradient (LOBPCG) method [Knyazev 2001]

    Only require matrix-vector multiplication.

    Lin Lin A Posteriori DG using Non-Polynomial Basis 23

  • Estimating 𝑏𝑏𝑘𝑘

    How to estimate 𝑢𝑢, 𝑣𝑣 𝜕𝜕𝜅𝜅. Importance of Lobatto grid

    Here 𝑀𝑀𝑏𝑏 = �𝑊𝑊

    Lin Lin A Posteriori DG using Non-Polynomial Basis 24

  • Generalize to high dimensionsTensor product LGL grid ⇒ Tensor product Lagrange polynomials

    Lin Lin A Posteriori DG using Non-Polynomial Basis 25

    𝜕𝜕𝑙𝑙

  • Compare with asymptotic results for polynomial basis functionsFor polynomial basis functions [e.g. Houston-Schötzau-Wihler, 2007]

    Lin Lin A Posteriori DG using Non-Polynomial Basis 26

    ℎ = 1

  • Penalty parameterParameter {𝛾𝛾𝜅𝜅}• Large enough for coercivity of the bilinear form• “magic parameter” in interior penalty method [Arnold

    1982]

    Define

    Lemma. If 𝛾𝛾𝜅𝜅 ≥12

    1 + 𝜃𝜃 2𝑑𝑑𝜅𝜅2, then the bilinear form is coercive

    Automatic guarantee of stability

    Lin Lin A Posteriori DG using Non-Polynomial Basis 27

  • Penalty parameterComputation of 𝑑𝑑𝜅𝜅 through eigenvalue problem

    By setting 𝑣𝑣𝑁𝑁 = Φ𝑐𝑐, span Φ = 𝕍𝕍𝑁𝑁 𝜅𝜅 . Can be solved with direct method

    Lin Lin A Posteriori DG using Non-Polynomial Basis 28

  • Upper bound estimatorThe last constant

    𝑑𝑑𝜅𝜅𝑢𝑢(𝑢𝑢𝑁𝑁) involves the true solution 𝑢𝑢 and therefore is the only constant that cannot be explicitly computed.

    However, numerical result shows that 𝑑𝑑𝜅𝜅𝑢𝑢 𝑢𝑢𝑁𝑁 ≈ 𝑑𝑑𝜅𝜅

    is a good approximation.

    Lin Lin A Posteriori DG using Non-Polynomial Basis 29

  • Outline• Introduction: Adaptive local basis functions

    • Computable upper bound for Poisson’s equation

    • Computable upper / lower bound for indefinite equations

    • Numerical examples

    • Conclusion and future work

    Lin Lin 30A Posteriori DG using Non-Polynomial Basis

  • Model problemIndefinite equation

    𝑉𝑉 ∈ 𝐿𝐿∞ Ω and −Δ + 𝑉𝑉 has no zero eigenvalue.

    Bilinear form

    DG approximation

    Lin Lin 31A Posteriori DG using Non-Polynomial Basis

  • Computable upper bound

    Lin Lin A Posteriori DG using Non-Polynomial Basis 32

    Energy norm

    Theorem ([LL-Stamm 2015]). Let 𝑢𝑢 ∈ 𝐻𝐻#1 Ω ⋂𝐻𝐻2(𝒦𝒦) be the true solution and 𝑢𝑢𝑁𝑁 ∈ 𝕍𝕍𝑁𝑁 the DG-approximation. Then

  • Computable lower boundTheorem ([LL-Stamm 2015]). Let 𝑢𝑢 ∈ 𝐻𝐻#1 Ω ⋂𝐻𝐻2(𝒦𝒦) be the true solution and 𝑢𝑢𝑁𝑁 ∈ 𝕍𝕍𝑁𝑁 the DG-approximation. Then

    where

    Lin Lin 33A Posteriori DG using Non-Polynomial Basis

    All constants other than 𝑑𝑑𝜅𝜅𝑢𝑢are computable

  • Computable lower boundBubble function 𝑏𝑏𝜅𝜅

    For instance, 𝑏𝑏𝜅𝜅 𝑥𝑥 = 4 𝑥𝑥 1 − 𝑥𝑥 , 𝜅𝜅 = 1

    Lemma.

    where

    Lin Lin A Posteriori DG using Non-Polynomial Basis 34

  • Outline• Introduction: Adaptive local basis functions

    • Computable upper bound for Poisson’s equation

    • Computable upper / lower bound for indefinite equations

    • Numerical examples

    • Conclusion and future work

    Lin Lin 35A Posteriori DG using Non-Polynomial Basis

  • 1D Poisson equation−Δ𝑢𝑢 𝑥𝑥 = sin 6𝑥𝑥

    Adaptive local basis functions with 11 basis per element.

    Lin Lin A Posteriori DG using Non-Polynomial Basis 36

  • Effectiveness of upper/lower estimtaorMeasure local effectiveness (𝐶𝐶𝜂𝜂 ≥ 1, 𝐶𝐶𝜉𝜉 ≤ 1)

    Lin Lin A Posteriori DG using Non-Polynomial Basis 37

  • 1D indefinite−Δ𝑢𝑢 𝑥𝑥 + 𝑉𝑉 𝑥𝑥 𝑢𝑢(𝑥𝑥) = sin 6𝑥𝑥

    Adaptive local basis functions with 11 basis per element.

    Lin Lin A Posteriori DG using Non-Polynomial Basis 38

  • Effectiveness of upper/lower estimtaor

    Lin Lin A Posteriori DG using Non-Polynomial Basis 39

  • 2D Helmholtz−Δ𝑢𝑢 + 𝑉𝑉𝑢𝑢 = 𝑓𝑓,

    𝑉𝑉 = −16.5, 𝑓𝑓 𝑥𝑥,𝑦𝑦 = 𝑒𝑒−2 𝑥𝑥−𝜋𝜋 2−2 𝑦𝑦−𝜋𝜋 2

    Adaptive local basis functions with 31 basis per element.

    40Lin Lin A Posteriori DG using Non-Polynomial Basis

  • Effectiveness for upper/lower bound

    Lin Lin A Posteriori DG using Non-Polynomial Basis 41

  • Validate the approximation for 𝑑𝑑𝜅𝜅𝑢𝑢

    Note that

    Although 𝑑𝑑𝜅𝜅𝑢𝑢 is not known, it is only sufficient to have

    𝑑𝑑𝜅𝜅𝑢𝑢 ≈ 𝑑𝑑𝜅𝜅 or 𝑑𝑑𝜅𝜅𝑢𝑢 ≪ 𝑏𝑏𝜅𝜅𝛾𝛾𝜅𝜅

    1D:

    Lin Lin A Posteriori DG using Non-Polynomial Basis 42

  • Validate the approximation for 𝑑𝑑𝜅𝜅𝑢𝑢

    2D indefinite

    Lin Lin A Posteriori DG using Non-Polynomial Basis 43

  • Conclusion• Systematic derivation of a posteriori error estimation for

    general non-polynomial basis function

    • Explicitly computable constants for upper/lower estimator.

    • The only one non-computable constant can be reasonably estimated by known ones.

    Lin Lin 44A Posteriori DG using Non-Polynomial Basis

  • Future work• Eigenvalue problem

    • Nonlinearity, atomic force, linear response properties

    • Implementation in DGDFT

    • Other basis functions, including MsFEM, HMM, MsDG etc.

    Ref:LL and B. Stamm, A posteriori error estimates for discontinuous Galerkin methods using non-polynomial basis functions. Part I: Second order linear PDE, arXiv:1502.01738

    Thank you for your attention!

    Lin Lin 45A Posteriori DG using Non-Polynomial Basis

    A Posteriori Error Estimates For �Discontinuous Galerkin Methods Using �Non-polynomial Basis FunctionsOutlineMotivationKohn-Sham density functional theoryDiscretization costAdaptive local basis functionsDiscontinuous Galerkin methodWhy a posteriori error estimatorResidual based a posteriori error estimatorDifficultyOutlineModel problemDG discretizationError quantificationUpper bound of errorProjection operatorEstimating constants Numerical procedure for computing the constantsNumerical representation of inner productNumerical representation of inner productNumerical representation of inner productEstimating 𝑎 𝑘 Estimating 𝑎 𝑘 Estimating 𝑏 𝑘 Generalize to high dimensionsCompare with asymptotic results for polynomial basis functionsPenalty parameterPenalty parameterUpper bound estimatorOutlineModel problemComputable upper boundComputable lower boundComputable lower boundOutline1D Poisson equationEffectiveness of upper/lower estimtaor1D indefiniteEffectiveness of upper/lower estimtaor2D HelmholtzEffectiveness for upper/lower boundValidate the approximation for 𝑑 𝜅 𝑢 Validate the approximation for 𝑑 𝜅 𝑢 ConclusionFuture work