Stability, Consistency, and Convergence

:
A 21st Century Viewpoint

Douglas N. Arnold

School of Mathematics, University of Minnesota 1920–1993
Society for Industrial and Applied Mathematics
Feng’s significance for the scientific development of China cannot be
Feng Kang Distinguished Lecture exaggerated. He not only put China on the map of applied and
computational mathematics, through his own research and that of his
Institute of Computational Mathematics and Scientific/Engineering Computing students, but he also saw to it that the needed resources were made
Chinese Academy of Sciences available. . . .
April 7, 2009 Many remember his small figure at international conferences, his eyes
and mobile face radiating energy and intelligence. He will be greatly
missed by the mathematical sciences and by his numerous friends.
– Peter Lax, writing in SIAM News, 1993

The failure of the Sleipner A offshore platform Convergence, consistency, and stability of discretizations

L : X → Y bounded linear operator on Banach spaces.
Continuous problem: Given f ∈ Y find u ∈ X such that Lu = f .
Assume it is well-posed: ∀f ∃! u s.t. Lu = f , f 7→ u is continuous
Discrete problem: Lh : Xh → Yh operator on finite dimensional spaces,
fh ∈ Yh . Find uh ∈ Xh such that Lh uh = fh .
The discretization is convergent if uh is sufficiently near u.
The discretization is consistent if Lh and fh are sufficiently near
L and f .
The discretization is stable if the discrete problem is well-posed.
6m

“Fundamental metatheorem of numerical analysis”
A discretization which is consistent and stable is convergent.

$700,000,000 Richter magnitude 3
3 / 44 4 / 44

367879 7 0. u (x .394 consistency error = O (k ) + O (h2 ) 4 0. v iXh∗ ×Xh = ∇u · ∇v dx ∀u . k . The method is convergent if it tends to 0 as h → 0.460. To quantify consistency we use a norm in the space Yh . . 2 0. 0) = u0 (x ) x ∈Ω Discretization: n γn n γn n γn 1 0.Xh ) kLh rh u − fh kYh hfh .684800 uh ( x . t = 0.5328 k 3 0. The Finite difference discretization: Xh = Yh = grid fns consistency error is then kLh rh u − fh kYh . consistency. .127120 12 -17.460. x ∈ grid. v ∈ Xh ΩZ krh u − uh kXh ≤ kL− 1 h kL(Yh .632121 6 0. Lu := −∆u = f in Ω. and stability To quantify convergence we use Continuous problem: 1 A restriction operator rh : X → Xh . A simple example: Dirichlet problem for Poisson’s equation Measuring convergence. t + k ) − uh ( x .Xh ) . t ) = 0 x ∈ ∂Ω.118720 14 -2747.6 ??? 0 7 / 44 8 / 44 . t ) = 0 x ∈ Ω . T ] 0 u (x .110160 13 211. Given f ∈ L2 (Ω) find u ∈ H01 (Ω) such that 2 A norm in the space Xh . fh = f |grid pts −∆h u (X ) S In this context the fundamental metatheorem is a theorem: Finite element discretization: Z Xh ⊂ H01 (Ω). t ∈ [ 0.6 Z 1 x 14 ex −1 dx = 38. The method is stable if it X remains bounded as h → 0. t ) − ∆h uh (x . 4u (X ) − u (N ) − u (S ) − u (E ) − u (W ) h Lh u (X ) := h2 W E The stability constant is kL−1 h kL(Yh . The discretization error is then krh u − uh kXh .264242 8 0. . . t ) − ∆u (x . .170904 10 -0.12 5 0. Yh = Xh∗ Lh uh = fh =⇒ Lh rh u − Lh uh = Lh rh u − fh =⇒ rh u − uh = Lh−1 (Lh rh u − fh ) hLh u .207274 9 0. T ] Z 1 ∂t γn+1 = x n ex −1 dx γn+1 = 1 − n γn γ1 = 1−e−1 = 0.632121 . 2k . The method is consistent if it N tends to 0. t ∈ [0. v iXh∗ ×Xh = f v dx Ω 5 / 44 6 / 44 An elementary example of instability Finite differences for the heat equation Z 1 Initial value/boundary value problem: Compute x 14 ex −1 dx 0 ∂u (x . t ) = 0.068480 15 38.145480 11 1.

∆s and ∆t must be small in comparison to the space and time scales of physically relevant motions. The circumstances are rst. then kGkL(l∞ . uh (x . that the initial value problem be properly posed. 11 / 44 12 / 44 . nk )]  ∂t 2 ∂x 2 h with centered differences and timestep k proportional to h. 9 / 44 10 / 44 Milestones: stability Milestones: the equivalence theorem Von Neumann realized the importance of numerical stability. They went on to analyze stability for evolution equations via discrete Fourier analysis. on numerical weather prediction with Charney & Fjörtoft: capstone paper was the 1956 Survey of the stability of linear finite difference equations by Lax and Richtmeyer.. consistency. and Lewy were first to realize that a consistent discretization of a well-posed problem need not converge. which is necessary for convergence. they considered the wave eq. so uh ( · . But this does not alone insure The term stability refers to a property of the. If k > h2 /4. . (n + 1)k = (I + k ∆h )uh ( · . . t k ∂2u ∂2u k = λh (x . For consistent finite difference discretizations of well-posed linear initial-value problems. We shall give a distortion may possibly be amplied in the course of the computation denition of stability in terms of the uniform boundedness of a certain to such an extent that they will totally obscure the signicant large. and second. sequence of nite accuracy. the method cannot be convergent. . stability is necessary and sucient for convergence. . If the nite dierence solution is to approximate closely the continuous solution. 0) 7→ uh is bounded in l∞ true domain of dependence and uniformly in h and k and we obtain stability. the numerical domain of dependence does not cover the If k ≤ h2 /4 then kGv kl∞ ≤ kv kl∞ .Stability for the discretized heat equation Milestones: consistency does not imply convergence Courant. giving birth to von Neumann stability analysis for a finite Lax equivalence theorem difference discretization of an evolution equation. . t ) = (x . the small-scale motions for which there is inevitably a large dierence equations with increasingly ner mesh. (n + 1)k ) − uh (x . . This gave us the CFL condition.l∞ ) > 1. and the method is unstable. set of operators and then show that under suitable circumstances. x 4k k Gv (x ) = (1 − )v (x ) + [v (N ) + v (S ) + v (E ) + v (W )] h2 h2 If λ > 1. An important then difference methods for PDEs in another landmark paper in 1950. nk ) =: G[uh ( · . . for scale motions. nk ) − ∆h uh (x . linear initial value problems. and problems was developed by many researchers in 1950s. Friedrichs. . nk ) = 0 In their classic 1928 paper. studying first The convergence theory of finite difference discretizations for evolution linear algebra problems in his landmark paper with Goldstine in 1947. t ) uh · . stability ⇐⇒ convergence.

W Hilbert spaces. div σ = f 1 1 Z Z Z Z B (σ. This is easy if Bh is coercive: polynomial spaces of a certain sort. w i := B (v . If Vh and Wh are piecewise So we need to bound γh below. the stability constant can be Find u ∈ V such that B (u . Bh : Vh × Wh → R bilinear. u . For simplicity we henceforth assume V = W . w ) Vh . fh ∈ Wh∗ . Wh ⊂ W . it is a finite element method. or a conforming FEM. w ). γh := inf sup . 06=v ∈Vh 06=w ∈Vh kv kV kw kV Find uh ∈ Vh such that Bh (uh . fh = f |Wh . 1 k L−1 h kL(Vh . v ) ≥ γkv k2 . this is a true Bh (v . B : V × W → R bdd bilinear form. hf . This is a generalized Galerkin method. τ. but otherwise can be quite difficult. Vh = Wh ⊂ V (although both Vh 6= Wh and Vh 6⊂ V are of interest). w ) = hfh .Vh ) = ∗ Ex: V = W = H01 (Ω). w ∈ Wh . Bh = B |Vh ×Vh . Wh of equal finite dimension. w i. w i. w ) = Ω ∇v · ∇w dx. v i = Ω f v dx R R γh Discrete problem: where Bh (v . w ) = hf .Finite element methods Stability of finite element methods Continuous problem: V . v ∈ L2 −1 −1 Ω Ω P1 –P0 RT0 –P0 P1 -P1 (20 elts) P1 -P1 (40 elts) P1 -P0 (40 elts) Raviart-Thomas 1976 15 / 44 16 / 44 . B (v . w ∈ W. 1) σ + grad u = 0. σ =f on (−1. Ex: If Vh ⊂ V . v ) := 0 0 (στ −τ u +σ v ) dx = fv dx ∀τ ∈ H . v ∈ L 1 2 (σ · τ − div τ u + div σ v ) dx = fv dx ∀τ ∈ H (div). expressed as the reciprocal of the inf-sup constant: ⇐⇒ Lu = f where hLv . 13 / 44 14 / 44 Mixed Laplacian in 1D Mixed Laplacian in 2D Babuška–Narasimhan 0 0 σ + u = 0. f ∈ W∗ For a generalized Galerkin method. Galerkin method.

w ) − hfh .. w ∈Vh kw kV kuh − rh u kV ≤ γh−1 kB k inf ku − v kV . w )| sup = sup ≤ kB k inf ku −v kV . v ∈Vh By the fundamental theorem. For nonconforming methods we get some extra terms coming from the z }| { z }| { |Bh (u . For Lagrange Pr elements infv ∈Vh ku − u kH 1 = O (hr ) (as long as u is smooth and the simplices don’t degenerate). and x 2y 2 compose to get the shape functions. For approximating the space On cubes. the shape functions and degrees of freedom. for a conforming FEM we have |Bh (rh u .e. (Strang 1972. w i| |B (rh u − u . P1 P2 P3 For distorted quads or 1 1 1 x y x y x y hexes. (Cea 1964 in the coercive case. consisting of all continuous functions which elements Pr ⊂ Q0r ⊂ Qr . also give O (hr ) belong to Pr on each simplex. coercive case) consistency error ≤ kBh k inf ku − v k + sup v ∈Vh v ∈Vh kw k 17 / 44 18 / 44 Bounds on the approximation error Quadrilaterals and hexahedra Finite element spaces must be constructible from local information. fh = f |Vh . v ∈Vh The simplest (and most important) case is a conforming FEM. bounded by the approximation error times kB k. i. Babuška 1972) conformity error In the general case. the most common finite elements are the Qr = ni=1 Pr (I ). we use a xy x 2 xy y 2 x 2 xy y 2 multilinear map from a x 2 y xy 2 x 2 y xy 2 reference cube. w i| conformity error. w ) − hf . w ) − hfh . this becomes |B (rh u . w ∈Vh kw kV w ∈Vh kw kV |v ∈Vh {z ku − uh kV ≤ (1 + γh−1 kB k) inf ku − v kV v ∈Vh } For a conforming method the consistency error is Stable conforming finite elements are quasioptimal. Then the consistency error is With the triangle inequality. Bh = B |Vh ×Vh . But the clever serendipity N Lagrange elements. 19 / 44 20 / 44 .Consistency of finite element methods Convergence of finite element methods For rh take the V -orthogonal projection into Vh : rh u = arg min ku − v kV . approximation. instead of Pr the natural choice is H 1 using simplicial meshes. w i| consistency error := kLh rh u − fh kVh∗ = sup .

2 64 0.01 H (curl) and H (div).2829 2.0 16 1.87159 2. It is conforming.2670 2.0 analysis. but not in were very difficult to H (curl) ∩ H (div) (which contains H 1 as a closed −2 10 simulate.58 2 51.0 64 0. Z Z Standard finite elts converge. .2616 exact: π 2 = 9. v ) Why? The method is clearly stable in the space Ω Ω H (curl) ∩ H (div).2629 2. . from O (h) to O (1) on non-parallel quadrilaterals and hexahedra. showed a striking consequence in this computation of n % err. and std elts −1 10 It was found that they give good approximation in H 1 .0 8 4. This θ −4 10 w t=1 was phemenon was 2 θ variable. R2 ) × H01 (Ω) s.1 n λh rate n λh rate From The Sleipner accident and its causes. named locking.3474 8 9.0 (a) Unfavorable geometrical shaping of some nite elements in the global 32 10. but only Qr achieves O (hr ) on In 2005 A-B-F showed the problem is even worse for vectorial general hexahedra.6 16 0. In conjunction with the subsequent post-processing of the 64 10.0 32 0. because subspace).Reduced convergence for distorted serendipity elements Reduced convergence for vectorial elements Folklore had it that Qr and Q0r both afford O (hr ) approximation. 0 Q2 Q2 Bermudez et al. 21 / 44 22 / 44 Vector Laplacian Thin elastic plates curl curl u − grad div u = f in Ω An elastic plate is a flat structural element u · n = 0. For Q0r on general hexahedra. .87754 2.0 16 9.05 2. curl u × n = 0 on ∂Ω whose thickness t is much less than the Z Z dimensions of its midsurface Ω ⊂ R2 .84 1.08 2.86906 (b) Inadequate design of the haunches at the cell joints.75 2. 1998: 8 10. extrapolated: 10.90136 2.02 2. It turns out (A-Boffi-Falk 2002). −1 0 10 10 23 / 44 24 / 44 . (curl u · curl v + div u div v ) dx = f · v dx ∀v Reissner–Mindlin plate equations (simplified): Ω Ω Find (θ.t. the convergence rate is elements. this led to underestimation of the shear forces at the wall supports by some 45%.89 1.0 32 9.72 1. So this method is inconsistent! standard finite elements −3 10 lead to large errors How can we solve this problem? Introduce σ = curl u as a new when t is small.21 the fundamental eigenvalue of a square acoustic cavity: 4 12.19 2.84 1. rate 2 48.0 4 14.0 analysis results. . but not to the soln! ∇S θ · ∇S ψ + t −2 (∇w − θ) · (∇v − ψ) dx = fv dx ∀(ψ. and solve the resulting system using appropriate elements for w t = . B Jakobsen. w ) ∈ H01 (Ω.8 8 3. which support the tricell walls. . reduced load bearing capacity were: 16 10.1 64 9.4 32 0. rate n % err. The popular lowest order Raviart–Thomas elements go reduced from r to br /2c. this is true on parallelpids. .40 1.99708 The reasons for the.

We 10 −3 How can we solve the problem? Introduce ζ = t −2 (∇w − θ). the method is conforming and linear elements give additional DOF for θ and good approximation in H 1 . ζ. exactly in order to achieve stability. but lose stability and convergence as h → 0 with t fixed. w . .01 w w −1 0 −1 0 −1 0 computational engineering. so we still have nonconformity for w 10 −2 a large consistency error.01 Z it turns out that this is w B (θ.01 t = . θ w On the basis of experiments they declared this element locking-free. v ) = ∇S θ · ∇S ψ + t −2 (∇w − θ) · (∇v − ψ) dx Ω θ w ζ Where does the error come from? The bilinear form is H 1 coercive We introduced the and 10 −1 (hence stability). but kB k behaves like t −2 . 10 10 10 10 10 10 P1 -P1 AF OZF 27 / 44 28 / 44 . ψ. 25 / 44 26 / 44 An attempt at another simple locking free elt Thin elastic shells An elastic shell is a curved structural Oñate-Zarate-Flores (1994): element whose thickness is much less ζ than the dimensions of its midsurface. although coercive. and it is tricky to find stable elements. and despite tremendous efforts no one has −2 −2 −2 10 10 10 succeeded to develop a numerical method −3 10 10 −3 −3 10 for shells which is certifiably convergent θ θ θ over a wide range of conditions. Shells are the primadonna of structures: Again the nonconformity error is controllable. none quite as simple.01 10 2 θ 10 2 θ This stands as a major challenge in w t = . and so 10 −1 0 10 Ω we proved that the AF element converges as h → 0. −4 w t=1 −4 w t=1 −4 w t=1 10 2 θ t = . The difficulty is that this form is not Additional locking-free elements have since been devised. which doesn’t blow-up as t → 0. The causes and cure of locking The simplest locking free plate elements Z The AF element (1989): B (θ. v . but 10 2 θ t = . A careful analysis (A–Falk 1997) showed that the inf-sup but occasionally experience catastrophic constant behaves like min(1. Thus we have stability if collapse. The locking problems encountered for −1 10 10 −1 −1 10 plates are much greater for shells. η) = ∇S θ·∇S ψ+ζ·(∇v −ψ)−(∇w −θ)·η+t ζ·η dx 2 easy to estimate. ψ. However stability is they can deliver remarkable performance. uniformly in t. the have to bound the θ shear stress. w . subtle. The new bilinear form is −4 w t=1 nonconformity error. t = O (h). . h2 /t 2 ).

3 like the true spectrum. Ω) − 1 −→ L2 (Ω) → 0 div Given f ∈ L2 Λk (Ω). Acta Numerica 2006. the resulting problem is always well-posed. 0 29 / 44 30 / 44 The Hodge Laplacian Mixed formulation of the Hodge Laplacian To get an impression of FEEC consider the Hodge Laplacian. algebraic topology. . find σ ∈ H Λk −1 . 3 but also to a sequence 2 1 of spurious ones. We should not discretizing a differential 5 to the true eigenvalues. via the inf-sup condition. 4. 31 / 44 32 / 44 . and 10 homological algebra to develop discretizations which are compatible 9 with the geometric. If we also account for the harmonic forms. systems of partial dierential equations. . τ i − hd τ. This approach brings to On a square the positive eigenvalues are 1. mesh. pp.  Finite element exterior calculus is an approach to the design and R understanding of nite element discretizations for a wide variety of solved using standard P1 finite elements. p ∈ Hk : hσ. u i = 0 ∀τ ∈ H Λk −1 Hodge Laplacian: Given a k -form f find a k -form u such that hd σ. ω ⊥ Zk ker d range d So we need to capture these in our discretization. q i = 0 ∀q ∈ Hk (dd ∗ + d ∗ d )u = f The proof of well-posedness. Falk & Winther. v i + hdu . This is Introducing σ = d ∗ u. 2. v i = hf .The impact of instability on an eigenvalue problem Finite element exterior calculus For many electromagnetic problems. and applications. k k −1 Poincaré inequality: kωkL2 ≤ c kd ωkL2 . with dimension equal to the k th Betti number. ω ∈ H Λk . . dv i+hp. 2 1 0 On this structured 10 9 mesh the discrete 8 From the point of view of FEEC. 1. 5. v i ∀ v ∈ H Λk hu . the discrete 6  Finite element exterior calculus. we get the mixed formulation of Hodge the PDE most closely associated to the de Rham complex Laplacian. bear tools from dierential geometry. 4. 5 spectrum looks nothing 4 Arnold. Lagrange elements don’t seem at all 7 eigenvalues converge 6 natural for curl curl problems. topological. stable finite elements are elusive. A striking example of what can goRwrong with unstable elements is the curl-curl eigenvalue problem: curl u · curl v dx = λ u v dx ∀v . 5. 1155. 8. homological techniques. d d d 0 → H Λ0 (Ω) − −→ H Λ1 (Ω) −−→ · · · −−→ H Λn (Ω) → 0 Continuous problem: grad curl 0 → H (Ω) −−→ H (curl. and algebraic structures which On an unstructured 8 7 underlie well-posedness of the PDE problem being solved. relies on two fundamental results: The degree of well-posedness is determined by the harmonic forms Hodge decomposition: H Λk = Bk ⊕ Hk ⊕ Zk ⊥ Hk := Zk ∩ (Bk )⊥ . Ω) −−→ H (div. u ∈ H Λk . 4 1-form with elements devised for 0-forms.

κ κ κ 0 ←−− Pr Λ0 ←−− Pr −1 Λ1 ←−− · · · ←−− Pr −n Λn ←−− 0 In this way we obtain the two primary families of finite element with (κω)x := ωx yx. The generalized Galerkin method is stable and convergent. It turns out that there are precisely two “natural” families of polynomial differential forms to use as shape functions on simplices: the spaces Pr Λk and the spaces Pr− Λk := Pr −1 Λk + κPr −1 Λk +1 35 / 44 36 / 44 . v i ∀v ∈ Vhk v ∈Hk huh . q i = 0 ∀q ∈ Hkh kv k=1 The discrete Poincaré inequality holds uniformly in h. We will construct the Vhk as finite element spaces assembled on simplicial meshes. They are not bounded. dv i+hph . Theorem Generalized Galerkin method: Find σh ∈ Vhk −1 . uh ∈ Vhk . Vhk ⊂ H Λk . hσh . consistent. Hkh ≤ sup kv − πhk v k  hd σh . (d κ + κd )ω = (r + k )ω ∀ω ∈ Hr Λk The DOFs also determine cochain projections. ph ∈ Hkh : If kv − πhk v k < kv k ∀v ∈ Hk . When is this discretization stable. It satisfies the homotopy property homog eneous differential forms. A key tool is the Koszul complex Similarly Pr− Λk has DOFs based on Ps Λl . τ i − hd τ. then the induced map on cohomology is an isomorphism. and so must specify On a simplex T . but they can be modified to become so.Discretization Bounded cochain projections To discretize we use a generalized Galerkin method based on finite The key property is that there must exist a bounded cochain projection. and convergent? 33 / 44 34 / 44 Finite element differential forms: shape functions Finite element differential forms: DOFs Thus we require finite dimensional de Rham subcomplexes which admit bounded cochain projections. v i = hf . dimensional subspaces Vhk −1 ⊂ H Λk −1 . the space Pr Λk (T ) has a natural set of DOFs polynomial shape functions and DOFs. v i + hduh . uh i = 0 ∀τ ∈ Vhk −1 gap Hk . polys. given in terms of moments weighted by Ps− Λl (f ) on a faces f . d k −1 · · · −−→ H Λk −1 −−−→ H Λk −−→ · · · We require the subcomplex property dVhk −1 ⊂ Vhk in order to obtain   πhk bounded  k −1 πk the yπh y h πhk a projection Discrete Hodge decomposition: Vhk = Bkh ⊕ Hkh ⊕ Zkh ⊥ d k −1 · · · −−→ Vhk −1 −−−→ Vhk −−→ · · · πhk d k −1 = d k −1 πhk −1 where Hkh := (Bkh )⊥ ∩ Zkh .

. Stable finite elements for Maxwell’s equations and related EM problems Stable approximation of mixed eigenvalue problems Preconditioning and multigrid A posteriori error estimation The biggest success to date is the construction of stable mixed finite elements for elasticity. 39 / 44 . 37 / 44 38 / 44 Applications of FEEC Stable finite elements for the Hodge Laplacian. . . div-curl problems. . But that’s another story . which had been sought by engineers and numerical analysts for four decades. For every r ∃ 2n−1 such FEdR subcomplexes. . Bossavit 1988 Pr Λk spaces with decreasing degree: n = 3: Pr− Λ1 (T ) ⊂ H (curl) Nedelec 1st kind edge elts ’80 d d d 0 → Pr Λ0 (T ) − −→ Pr −1 Λ1 (T ) −−→ · · · −−→ Pr −n Λn (T ) → 0 n = 3: Pr Λ1 (T ) ⊂ H (curl) Nedelec 2nd kind edge elts ’86 grad curl div 0→ −−→ −−→ −−→ →0 n = 3: Pr− Λ2 (T ) ⊂ H (div) Nedelec 1st kind face elts ’80 2 Demkowicz–Monk–Vardapetyan–Rachowicz 2000 n = 3: Pr Λ (T ) ⊂ H (div) Nedelec 2nd kind face elts ’86 These are extreme cases. curl-curl problems.Finite element differential forms/Mixed FEM Finite element de Rham subcomplexes r =1 Each of the families gives a family of finite element de Rham subcomplexes which admit bounded cochain projections. P − Λ0 (T r 0 ) = Pr Λ (T ) ⊂ H 1 Lagrange elts Pr− Λk spaces of constant degree r : d d d P − Λn (T n 2 ) = Pr −1 Λ (T ) ⊂ L discontinuous elts 0 → Pr− Λ0 (T ) − −→ Pr− Λ1 (T ) −−→ · · · −−→ Pr− Λn (T ) → 0 r grad curl div n = 2: P − Λ1 (T ) ⊂ H (curl) Raviart–Thomas elts ’76 0→ −−→ −−→ −−→ →0 r 1 n = 2: Pr Λ (T ) ⊂ H (curl) Brezzi–Douglas–Marini elts ’85 Whitney 1957.