Adaptive regularization strategies for nonlinear PDE Coarse mesh solvers for quasilinear PDE Sara Pollock Texas A&M Department of Mathematics Finite Element Circus, UMass Dartmouth, October 16-17, 2015 S. Pollock Texas A&M Mathematics 1/18 Adaptive nonlinear solver FE Circus October 16-17, 2015 MOTIVATION: Convergence problem!! (FE Circus, Oct. 2014) Z κ(u)∇u · ∇v = Z f v, κ(u) = 1 + 1 10−2 + (u − a)2 The standard Newton iterations for solving the nonlinear problem: hF(U), vh i = 0 : Starting with initial guess U = U0 , iterate: Find: w such that hDF(U)w, vi = −hF(U), vi Update: U = U + w Check: if kF(U)k < tolerance Sensitivity to initial data: This method did not converge for a ∈ (0.155, 1.35). In fact , starting the Newton method with the solution U0 = sin(πx) sin(πy) converged in 3 iterations, but failed to converge with a perturbation of more than 2 × 10−2 . For the results here a damped Newton method is implemented where the step size is decreased by a factor of 1/2 until the residual decreases (line search method). A better method by means of Tikhonov regularization with appropriate penalty parameters is currently under investigation. S. Pollock Texas A&M Mathematics 1/18 Adaptive nonlinear solver FE Circus October 16-17, 2015 Recent results: −div(κ(u, |∇u|)∇u) + b(u) · ∇u = f Recent papers: SP, Stabilized and inexact adaptive methods for capturing internal layers in quasilinear PDE. Available as arXiv:1507.06965 [math.NA], 2015 SP, An improved method for solving quasilinear convection diffusion problems on a coarse mesh. Available as arXiv:1502.02629 [math.NA], 2015 SP, A regularized Newton-like method for nonlinear PDE. Numerical Functional Analysis and Optimization, Accepted. Available as arXiv:1412.6487 [math.NA], 2014 Right: Log-scale visualization of diffusion coefficient, meshsize: 1/48. κ(u) = 1 + {ε + (u − 0.5)2 }−1 , S. Pollock Texas A&M Mathematics ε = 10−2 , 10−3 , 10−4 , 10−5 . 2/18 Adaptive nonlinear solver FE Circus October 16-17, 2015 Recent results: −div(κ(u, |∇u|)∇u) + b(u) · ∇u = f Recent papers: SP, Stabilized and inexact adaptive methods for capturing internal layers in quasilinear PDE. Available as arXiv:1507.06965 [math.NA], 2015 SP, An improved method for solving quasilinear convection diffusion problems on a coarse mesh. Available as arXiv:1502.02629 [math.NA], 2015 SP, A regularized Newton-like method for nonlinear PDE. Numerical Functional Analysis and Optimization, Accepted. Available as arXiv:1412.6487 [math.NA], 2014 Right: Log-scale visualization of diffusion coefficient, meshsize: 1/24. κ(u) = 1 + {ε + (u − 0.5)2 }−1 , S. Pollock Texas A&M Mathematics ε = 10−2 , 10−3 , 10−4 , 10−5 . 2/18 Adaptive nonlinear solver FE Circus October 16-17, 2015 Recent results: −div(κ(u, |∇u|)∇u) + b(u) · ∇u = f Recent papers: SP, Stabilized and inexact adaptive methods for capturing internal layers in quasilinear PDE. Available as arXiv:1507.06965 [math.NA], 2015 SP, An improved method for solving quasilinear convection diffusion problems on a coarse mesh. Available as arXiv:1502.02629 [math.NA], 2015 SP, A regularized Newton-like method for nonlinear PDE. Numerical Functional Analysis and Optimization, Accepted. Available as arXiv:1412.6487 [math.NA], 2014 Right: Log-scale visualization of diffusion coefficient, meshsize: 1/12. κ(u) = 1 + {ε + (u − 0.5)2 }−1 , S. Pollock Texas A&M Mathematics ε = 10−2 , 10−3 , 10−4 , 10−5 . 2/18 Adaptive nonlinear solver FE Circus October 16-17, 2015 Main idea: inexact solves of regularized problems Coarse mesh regime: Extract information from regularized problem for mesh refinement. I I Partial solves of inexact, regularized problems as the internal layers are uncovered. Goal of regularization: stability not accuracy Preasymptotic regime: Build sequence of initial guesses to regularized problems, and continue to refine the mesh adaptively. I I I I As the mesh is refined, reduce the regularization to increase the accuracy of each solve, constructing a better initial guess for the subsequent refinement. Update the regularization parameters adaptively to increase both accuracy and efficiency through the preasymptotic regime. For convergence of the method: Stopping criteria for inexact nonlinear solves. Update criteria for regularization parameters. Asymptotic regime: Solve the discrete problem, and refine mesh adaptively. I Reduction to a standard Newton method. S. Pollock Texas A&M Mathematics 3/18 Adaptive nonlinear solver FE Circus October 16-17, 2015 Target problem class Problem 1: quasilinear diffusion and convection-diffusion: F(x, u, ∇u) = −div(κ(x, u, |∇u|)∇u) + b(u) · ∇u − f = 0 in Ω, u = 0 on ∂Ω. Diffusion: b(u) = 0. Convection-diffusion: b(u) 6= 0. 1 Existence and uniqueness results for F(x, u, ∇u) = 0 carry over to the discrete problem assuming the mesh is fine enough. 2 Local uniqueness and approximation properties of the finite element solution hold assuming the mesh is fine enough. 3 Existence an (local) uniqueness of F(x, u, ∇u) = 0 does not necessarily carry over to the coarse mesh problem. 4 The coarse mesh problem may be considered noisy and ill-posed. 5 Coarse mesh Jacobians are ill-conditioned and sometimes indefinite. Adaptive method: Assume the initial mesh is not fine enough. (1) Construct a sequence of efficient adaptive partitions over which the problem is well-resolved and stable. (2) Construct sequence of transitional solutions to determine mesh refinement and to start a Newton-like iteration on the next partition. S. Pollock Texas A&M Mathematics 4/18 Adaptive nonlinear solver FE Circus October 16-17, 2015 Newton-like udpates: Tikhonov and Ψtc Common idea: stabilize the Jacobian by penalizing the update. Newton update: linearize F(u, x) = 0 ⇒ F 0 (x, un )wn = −F(x, un ), un+1 = un + wn . Pseudo-transient continuation: Discretize ∂u/∂t + F(x, u) = 0 with backward Euler: (∆tn−1 I + F 0 (x, un ))wn = −F(x, un ), un+1 = un + wn . Tikhonov regularization of the linerization with positive semidefinite R: minimize Gα (w) = kF 0 (x, un )w + F(x, un )k2 + αkRwk2 , yields iteration (αn RT R + F 0 (x, un )T F 0 (x, un ))wn = −F 0 (x, un )T F(x, un ), un+1 = un + wn , Without the normal equations: Larentiev regularization (αn R + F 0 (x, un ))wn = −F(x, un ), un+1 = un + wn . Coincides with ψtc stabilization of ∂(Ru)/∂t + F(x, u) = 0 with α = 1/∆t . Important idea from regularization: Generality : positive semidefinite R chosen to regularize only bad dof. Set R as a selectively-nullified Laplacian stiffness matrix: I dof in regions of low curvature in null (R) I Predict curvature with the jump-term in residual-based indicator. S. Pollock Texas A&M Mathematics 5/18 Adaptive nonlinear solver FE Circus October 16-17, 2015 Improved method: continuation point of view 1 Replace backward Euler discretization with Newmark update un+1 − un = ∆tn {(1 − γ)u̇n + γ u̇n+1 }, u̇ = ∂u/∂t. Reason: for γ > 1, controllable increase in numerical dissipation. Problem: Resulting iteration still requires use of normal equations to stabilize ill-conditioned indefinite Jacobians. 2 Stabilize the Newmark update by σ-splitting un+1 − un = ∆tn {σ(1 − γ)u̇n + (1 − σ)(1 − γ)u̇n + σγ u̇n+1 + (1 − σ)γ u̇n+1 }, and linearize the (1 − σ) fraction by R(u̇n+1 − u̇n ) = F(x, un+1 ) − F(x, un ) = F 0 (x, ū)wn for fixed ū. Resulting iteration: αn R + γ (1 − σ)F 0 (x, ū) + σF 0 (un , x) wn = −F(x, un ), un+1 = un + wn , Good news: Stabilizes iteration without use of normal equations. Convergence rate: Newmark update decreases rate of convergence from quadratic to q-linear with rate q = 1 − 1/γ. S. Pollock Texas A&M Mathematics 6/18 Adaptive nonlinear solver FE Circus October 16-17, 2015 Inexact iteration for quasilinear diffusion Previous iteration: αn R + γ (1 − σ)F 0 (x, ū) + σF 0 (x, un ) wn = −F(x, un ), un+1 = un + wn , Observation 1: γ introduces numerical dissipation by scaling down residual. The residual: F(x, u) = −g(u) + f (x) . Observation 2: g(u) is not the problem!! ⇒ The oscillations are caused by the source f (x). ⇒ Solution: Scale down f (x) until problem stabilizes. New inexact iteration: αn R + γ (1 − σ)F 0 (x, ū) + σF 0 (x, un ) wn = −g(un ) + δ f (x), un+1 = un + wn , Now require: δ → 1 from below: consistency. And want: γ → 1 from above: efficiency. Easier but important: σ → 1 From below, α → 0 with kF(x, u)k S. Pollock Texas A&M Mathematics 7/18 Adaptive nonlinear solver FE Circus October 16-17, 2015 γ → 1 from above: assumptions and update criteria Assumption: Upper bound on γ. Given a rate tolerance εT , γnk is bounded above for all n, k by γMAX , satisfying γnk ≤ γMAX < 1/εT . Update criteria for γ Given a user set tolerance εT , and γn < γMAX , and a minimum number of iterations between updates Imin , the dissipation parameter γn is updated after each iteration n on which all the following criteria hold. 1 γn > 1. 2 The observed rate of convergence is within tolerance of the predicted value n+1 kr k 1 krn k − 1 − γn < εT . 3 The observed rate of convergence is sufficiently stable. n+1 kr k krn k − < εT . krn k krn−1 k 4 At least Imin ≥ 2 iterations have passed since the last update of γ. S. Pollock Texas A&M Mathematics 8/18 Adaptive nonlinear solver FE Circus October 16-17, 2015 γ → 1 from above: definition and monotonicity Definition Given an initial γ0 > 1, and decrease parameter 0 < q < 1, for refinement k > 0 γ̃n+1 = q · hrn , rn i hrn , gn+1 − gn i , γn+1 = max 1 , γ̃n+1 . Stability: q · γn q · γn n+1 < γ̃ < . γn (2 + εT ) − 1 1 − εT γn Monotonic decrease of γ: Assuming the stronger upper bound: γnk 1 ≤ γMONO = εT q 1− , q̄ with 1 < q̄ < q, then for γn+1 = γ̃n+1 > 1, the sequence satisfies γn+1 < q̄γn . Key estimate: n 2 kr k S. Pollock Texas A&M Mathematics 1 − εT γn n n+1 < hr , g 1 − g i < kr k 2 − n + εT . γ n n 2 9/18 Adaptive nonlinear solver FE Circus October 16-17, 2015 δ → 1 from below: assumptions and update criteria Update criteria for δ: Given a user set rate tolerance εT and a convergence tolerance εCON , the iterations are terminated and δ is updated after iteration n of refinement k, when conditions (1-4) are all met, or when conditions (1) and (5) are satisfied. (1) δk < 1. If δk = 1, it is not updated. (2) Sufficient decrease in the sequence of residuals krk k krkn k < min{krk0 k, krk−1 k}. (3) The observed rate of convergence is beneath the accepted rate (1 − 1/2γ) krn+1 k 1 < 1− n. n kr k 2γ (4) The observed rate is sufficiently stable krn+1 k εT krn k + > . krn k 2 krn−1 k (5) krn k < εCON . The iteration reached convergence. S. Pollock Texas A&M Mathematics 10/18 Adaptive nonlinear solver FE Circus October 16-17, 2015 δ → 1 from below: definition Definition Given an initial δ0 < 1, for k > 0 h f , γσ(gn+1 − gn ) + (γ(1 − σ)g0 (un ) + αR)wn + gn i δ˜ k+1 = , qk k f k2 δk+1 = min{1 , δ˜ k+1 }, with R = Rk , α = αn , γ = γn and σ = σn . In terms of the one-step linearization error L e (un ) := gn+1 − gn − g0 (un )wn , gn+1 = g(un + wn ), the recursive relation for δ: 1 δk+1 = qk γσ e n δk + h f , L (u )i . k f k2 Effective bound: δk+1 > (1/q̄)δk , 1 < q̄ ≤ qk whenever qk k f k h fˆ, L e (un )i > −δk 1 − . q̄ γnk σnk Convergence of δ to 1 is also established. S. Pollock Texas A&M Mathematics 11/18 Adaptive nonlinear solver FE Circus October 16-17, 2015 Adaptive algorithm for nonlinear diffusion Set the parameters qγ , εT , γMAX , εCON , σ0 and δMIN . Let ū = 0. Start with initial u0 , γ0 , δ0 . On partition Tk , k = 0, 1, 2, . . . 1 Compute Rk , and g0 (ū). 2 Set r0 = −g(u0 ) + δ0 f0 , and set α0 = kr0 k. 3 While the Exit Criteria are not met on iteration n − 1 : n krn k o . (a) Set σ = max σ0 , 1 − K 0 n n n 0 n (b) Solve (α Rk + γ {σ g (u ) + (1 − σ)g0 (ū)})wn = rn , for wn . (c) Update un+1 = un + wn , and rn+1 = −g(un+1 ) + δk fk . (d) If Criteria to update γ are satisfied, update γn+1 . (e) Update αn ≤ krn k. 4 5 If Criteria to update δ are satisfied, update δk+1 for partition Tk+1 , with qδ = (qγ )P+1 , where P counts the number of times γ was updated on refinement k. Compute the error indicators to determine the next mesh refinement. S. Pollock Texas A&M Mathematics 12/18 Adaptive nonlinear solver FE Circus October 16-17, 2015 Exit strategy Given a user set tolerance εCON , and an accepted rate of convergence given by qACC = 1 − 1/(2γ), and a baseline number of allowed iterations IBASE , set the maximum number of iterations IMAX by & ' log (krk−1 k) − log krk0 k + 1, IMAX = max{IACC , IBASE }. IACC = log(qACC ) Exit criteria: Exit the solver on partition Tk after calculating residual krn+1 k when one of the following conditions holds. 1 The iteration reached convergence. 2 Conditions (2-4) of δ-update Criteria: the iteration converges an at acceptable rate and is stopped before convergence. 3 Maximum number of iterations IMAX completed: reset γ0k+1 , δk+1 , and u0k+1 . S. Pollock Texas A&M Mathematics 13/18 Adaptive nonlinear solver FE Circus October 16-17, 2015 Model problem with steep layers Quasilinear diffusion problem on Ω = (0, 1)2 . −div(κ(u)∇u) = f in Ω , u = 0 on ∂Ω. 1 ε = 6 × 10−5 , a = 0.5, and k = 1. κ(s) = k + ((ε + (s − a)2 ) f (x, y) chosen so the exact solution u(x, y) = sin(πx) sin(πy). Parameter a controls the position of the internal layers. The initial mesh has 144 elements. εT = 5 × 10−3 , q = 0.9. 106 103 γ 105 δ 4 102 101 100 10-1 10-2 10-3 5 10 15 20 25 Refinements 30 10 103 102 101 100 10-1 10-2 10-3 -4 3510 |u−uk |1 |u−uk |0 ηk n−1/2 103 104 Degrees of freedom, n Figure: Left: Parameters γ, δ → 1. Center: error. Right: mesh from Level 20. S. Pollock Texas A&M Mathematics 14/18 Adaptive nonlinear solver FE Circus October 16-17, 2015 Figure: Level 18 updates. S. Pollock Texas A&M Mathematics 15/18 Adaptive nonlinear solver FE Circus October 16-17, 2015 Model problem with unstable linearization Quasilinear diffusion problem on Ω = (0, 1)2 . −div(κ(u)∇u) = f in Ω , u = 0 on ∂Ω. κ(s) = 10 + sin(s/ε) + arctan(s/ε), 2 ε = 6 × 10−3 . 2 f (x, y) = (1 − x)(1 − y)(e8x − 1)(e8y − 1). 102 γ δ 101 100 10-1 10-2 5 10 15 Refinements 20 Figure: Left: Parameters γ, δ → 1. Right: Solution, Level 20. S. Pollock Texas A&M Mathematics 16/18 Adaptive nonlinear solver FE Circus October 16-17, 2015 Updates from Levels 5, 10 and 20 ··· ··· Figure: Level 5 updates: 1, 11 and 22. ··· ··· Figure: Level 10 updates: 1, 9 and 18. Figure: Level 20 updates. S. Pollock Texas A&M Mathematics 17/18 Adaptive nonlinear solver FE Circus October 16-17, 2015 Current directions Finite element solution Current directions Analysis of parameter choice for convection-diffusion problems (M. Ebrahimi, SP) and more general boundary conditions (SP). Generalization to inner iteration with multigrid method for inexact linear solves (Y. Zhu, SP). Finite element solution Accelerated explicit pseudo-time method for low-order discretization of Monge-Ampère type equations (G. Awanou, SP). Goal-oriented adaptivity for multiscale methods (E. Chung, Y. Efendiev, W.T. Leung, SP). S. Pollock Texas A&M Mathematics 18/18 Finite element solution Adaptive nonlinear solver FE Circus October 16-17, 2015 Main references Engl, H.W., Hanke, M. and Neubauer, A., Regularization of Inverse Problems, Springer, 1996. Holst, M., Adaptive Numerical treatment of elliptic systems on manifolds, Adv. Comput Math., 15, 1-4, 139-191, 2001. Kelley, C.T, and Keyes, D, Convergence Analysis of Pseudo-Transient Continuation, SIAM J. Numer. Anal., 35(2), 508-523, 1998. N. M. Newmark, A method of computation for structural dynamics, J. Eng. Mech.-ASCE, 85, EM3, 67-94, 1959. SP, A regularized Newton-like method for nonlinear PDE, Accepted: Numer. Func. Anal. Opt., arXiv:1412.6487 [math.NA], 2014. SP, An improved method for solving quasilinear convection diffusion problems on a coarse mesh, Available as arXiv:1502.02629 [math.NA], 2015. SP, Stabilized and inexact adaptive methods for capturing internal layers in quasilinear PDE, Available as arXiv:1507.06965 [math.NA], 2015. S. Pollock Texas A&M Mathematics 18/18 Adaptive nonlinear solver FE Circus October 16-17, 2015