Stability, Consistency, and Convergence:

A 21st Century Viewpoint

Douglas N. Arnold

School of Mathematics, University of Minnesota 19201993

Society for Industrial and Applied Mathematics
Fengs 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
 A simple example: Dirichlet problem for Poissons equation Measuring convergence, consistency, and stability
To quantify convergence we use
Continuous problem: 1 A restriction operator rh : X Xh .
Given f L2 () find u H01 () such that 2 A norm in the space Xh .
Lu := u = f in . The discretization error is then krh u uh kXh . The method is
convergent if it tends to 0 as h 0.

To quantify consistency we use a norm in the space Yh . The

Finite difference discretization: Xh = Yh = grid fns
4u (X ) u (N ) u (S ) u (E ) u (W )
Lh u (X ) :=
h2
fh = f |grid pts
Finite element discretization:
Z Xh H01 (), Yh = Xh
consistency error is then kLh rh u fh kYh . The method is consistent if it
N tends to 0.
h
W E The stability constant is kL1
h kL(Yh ,Xh ) . The method is stable if it
X
remains bounded as h 0.
h u (X ) S
In this context the fundamental metatheorem is a theorem:
Lh uh = fh = Lh rh u Lh uh = Lh rh u fh = rh u uh = Lh1 (Lh rh u fh )

hLh u , v iXh Xh = u v dx u , v Xh
Z krh u uh kXh kL 1
h kL(Yh ,Xh ) kLh rh u fh kYh
hfh , v iXh Xh = f v dx
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 ) u (x , t ) = 0 x , t [ 0, T ]
Z 1 t
n+1 = x n ex 1 dx n+1 = 1 n n 1 = 1e1 = 0.632121 . . . u (x , t ) = 0 x , t [0, T ]
0
u (x , 0) = u0 (x ) x

Discretization:
n n n n n n
1 0.632121 6 0.145480 11 1.684800 uh ( x , t + k ) uh ( x , t )
h uh (x , t ) = 0, x grid, t = 0, k , 2k , . . .
2 0.367879 7 0.127120 12 -17.5328 k
3 0.264242 8 0.110160 13 211.394 consistency error = O (k ) + O (h2 )
4 0.207274 9 0.118720 14 -2747.12
5 0.170904 10 -0.068480 15 38,460.6

Z 1
x 14 ex 1 dx = 38,460.6 ???
0

7 / 44 8 / 44
Stability for the discretized heat equation Milestones: consistency does not imply convergence
Courant, Friedrichs, and Lewy were first to realize that a consistent
discretization of a well-posed problem need not converge.
uh (x , (n + 1)k ) uh (x , nk )
h uh (x , nk ) = 0 In their classic 1928 paper, they considered the wave eq. t
k
2u 2u k = h
(x , t ) = (x , t )
uh , (n + 1)k = (I + k h )uh ( , nk ) =: G[uh ( , nk )]

t 2 x 2 h
with centered differences and timestep k proportional to h. x

4k k
Gv (x ) = (1 )v (x ) + [v (N ) + v (S ) + v (E ) + v (W )]
h2 h2
If > 1, the numerical domain of
dependence does not cover the
If k h2 /4 then kGv kl kv kl , so uh ( , 0) 7 uh is bounded in l
true domain of dependence and
uniformly in h and k and we obtain stability.
the method cannot be convergent.
If k > h2 /4, then kGkL(l ,l ) > 1, and the method is unstable.

This gave us the CFL condition, which is necessary for convergence.

9 / 44 10 / 44

Milestones: stability Milestones: the equivalence theorem

Von Neumann realized the importance of numerical stability, studying first
linear algebra problems in his landmark paper with Goldstine in 1947, and
then difference methods for PDEs in another landmark paper in 1950, on
numerical weather prediction with Charney & Fjortoft:
If the nite dierence solution is to approximate closely the continuous
solution, s and t must be small in comparison to the space and time
scales of physically relevant motions. But this does not alone insure
accuracy; the small-scale motions for which there is inevitably a large
distortion may possibly be amplied in the course of the computation
to such an extent that they will totally obscure the signicant large-
scale motions.
The convergence theory of finite difference discretizations for evolution
problems was developed by many researchers in 1950s. An important
capstone paper was the 1956 Survey of the stability of linear finite
difference equations by Lax and Richtmeyer.
The term stability refers to a property of the. . . sequence of nite
dierence equations with increasingly ner mesh. We shall give a
denition of stability in terms of the uniform boundedness of a certain
set of operators and then show that under suitable circumstances, for
linear initial value problems, stability is necessary and sucient for
convergence.. . . The circumstances are rst. . . consistency. . . and second,
that the initial value problem be properly posed.

They went on to analyze stability for evolution

equations via discrete Fourier analysis, giving
birth to von Neumann stability analysis for a finite
Lax equivalence theorem
difference discretization of an evolution equation. For consistent finite difference discretizations of well-posed linear
initial-value problems, stability convergence.

11 / 44 12 / 44
Finite element methods Stability of finite element methods
Continuous problem:
V , W Hilbert spaces, B : V W R bdd bilinear form, f W
For a generalized Galerkin method, the stability constant can be
Find u V such that B (u , w ) = hf , w i, w W. expressed as the reciprocal of the inf-sup constant:

Lu = f where hLv , w i := B (v , w ). 1
k L1
h kL(Vh ,Vh ) =

Ex: V = W = H01 (), B (v , w ) = v w dx, hf , v i = f v dx

R R h

Discrete problem: where

Bh (v , w )
Vh , Wh of equal finite dimension, Bh : Vh Wh R bilinear, fh Wh . h := inf sup .
06=v Vh 06=w Vh kv kV kw kV
Find uh Vh such that Bh (uh , w ) = hfh , w i, w Wh .

This is a generalized Galerkin method. If Vh and Wh are piecewise So we need to bound h below. This is easy if Bh is coercive:
polynomial spaces of a certain sort, it is a finite element method.
Ex: If Vh V , Wh W , Bh = B |Vh Vh , fh = f |Wh , this is a true Bh (v , v ) kv k2 ,
Galerkin method, or a conforming FEM.
but otherwise can be quite difficult.
For simplicity we henceforth assume V = W , Vh = Wh V
(although both Vh 6= Wh and Vh 6 V are of interest).
13 / 44 14 / 44

Mixed Laplacian in 1D Mixed Laplacian in 2D

BabuskaNarasimhan
0 0
+ u = 0, =f on (1, 1) + grad u = 0, div = f

1 1
Z Z Z Z
B (, u ; , v ) := 0 0
( u + v ) dx = fv dx H , v L
1 2
( div u + div v ) dx = fv dx H (div), v L2
1 1

P1 P0 RT0 P0

P1 -P1 (20 elts) P1 -P1 (40 elts) P1 -P0 (40 elts) Raviart-Thomas 1976
15 / 44 16 / 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 .
v Vh
By the fundamental theorem, for a conforming FEM we have
|Bh (rh u , w ) hfh , w i|
consistency error := kLh rh u fh kVh = sup .
w Vh kw kV kuh rh u kV h1 kB k inf ku v kV .
v Vh
The simplest (and most important) case is a conforming FEM, i.e.,
Bh = B |Vh Vh , fh = f |Vh . Then the consistency error is With the triangle inequality, this becomes
|B (rh u , w ) hf , w i| |B (rh u u , w )|
sup = sup kB k inf ku v kV .
w Vh kw kV w Vh kw kV |v Vh {z ku uh kV (1 + h1 kB k) inf ku v kV
v Vh
}
For a conforming method the consistency error is
Stable conforming finite elements are quasioptimal.
bounded by the approximation error times kB k.
(Cea 1964 in the coercive case, Babuska 1972)
conformity error
In the general case,
For nonconforming methods we get some extra terms coming from the
z }| {
z
}| {
|Bh (u , w ) hfh , w i| conformity error. (Strang 1972, 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, the
shape functions and degrees of freedom. For approximating the space On cubes, instead of Pr the natural choice is
H 1 using simplicial meshes, the most common finite elements are the Qr = ni=1 Pr (I ). But the clever serendipity
N
Lagrange elements, consisting of all continuous functions which elements Pr Q0r Qr , also give O (hr )
belong to Pr on each simplex. approximation.

P1 P2 P3
For distorted quads or
1 1 1
x y x y x y hexes, 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, and
x 2y 2
compose to get the shape
functions.

For Lagrange Pr elements infv Vh ku u kH 1 = O (hr )

(as long as u is smooth and the simplices dont degenerate).
19 / 44 20 / 44
Reduced convergence for distorted serendipity elements Reduced convergence for vectorial elements
Folklore had it that Qr and Q0r both afford O (hr ) approximation. It turns out
(A-Boffi-Falk 2002), this is true on parallelpids, but only Qr achieves O (hr ) on In 2005 A-B-F showed the problem is even worse for vectorial
general hexahedra. For Q0r on general hexahedra, the convergence rate is elements. The popular lowest order RaviartThomas elements go
reduced from r to br /2c. from O (h) to O (1) on non-parallel quadrilaterals and hexahedra.
0
Q2 Q2
Bermudez et al. showed a striking consequence in this computation of
n % err. rate n % err. rate
2 48.58 2 51.21 the fundamental eigenvalue of a square acoustic cavity:
4 12.08 2.0 4 14.72 1.8
8 3.02 2.0 8 4.84 1.6
16 0.75 2.0 16 1.89 1.4
32 0.19 2.0 32 0.84 1.2
64 0.05 2.0 64 0.40 1.1

n h rate n h rate
From The Sleipner accident and its causes, B Jakobsen, 1998:
8 10.3474 8 9.99708
The reasons for the. . . reduced load bearing capacity were:
16 10.2829 2.0 16 9.90136 2.0
(a) Unfavorable geometrical shaping of some nite elements in the global 32 10.2670 2.0 32 9.87754 2.0
analysis. In conjunction with the subsequent post-processing of the 64 10.2629 2.1 64 9.87159 2.0
analysis results, this led to underestimation of the shear forces at
the wall supports by some 45%. extrapolated: 10.2616 exact: 2 = 9.86906
(b) Inadequate design of the haunches at the cell joints, which support
the tricell walls.

21 / 44 22 / 44

Vector Laplacian Thin elastic plates

curl curl u grad div u = f in
u n = 0, curl u n = 0 on
Z Z
(curl u curl v + div u div v ) dx = f v dx
Standard finite elts converge. . . but not to the soln!
Why? The method is clearly stable in the space
An elastic plate is a flat structural element
whose thickness t is much less than the
dimensions of its midsurface R2 .
v ReissnerMindlin plate equations (simplified):
Find (, w ) H01 (, R2 ) H01 () s.t.
Z Z
S S + t 2 (w ) (v ) dx = fv dx (, v )

H (curl) H (div). It is conforming, and std elts 1

10
It was found that they
give good approximation in H 1 . . . but not in
were very difficult to
H (curl) H (div) (which contains H 1 as a closed 2
10
simulate, because
subspace). So this method is inconsistent!
standard finite elements 3
10
lead to large errors
How can we solve this problem? Introduce = curl u as a new when t is small. This
4
10 w t=1
was phemenon was 2
variable, and solve the resulting system using appropriate elements for w
t = .01

H (curl) and H (div). named locking.

1 0
10 10
23 / 44 24 / 44
 The causes and cure of locking The simplest locking free plate elements

Z The AF element (1989):

B (, w ; , v ) = S S + t 2 (w ) (v ) dx
w

Where does the error come from? The bilinear form is H 1 coercive We introduced the and
10
1

(hence stability), the method is conforming and linear elements give additional DOF for and
good approximation in H 1 . . . but kB k behaves like t 2 , so we still have nonconformity for w 10
2

a large consistency error. exactly in order to

achieve stability. We 10
3

How can we solve the problem? Introduce = t 2 (w ), the have to bound the
shear stress. The new bilinear form is
4
w t=1
nonconformity error, but 10
2
t = .01
Z it turns out that this is w

B (, w , ; , v , ) = S S +(v )(w )+t dx 2

easy to estimate, and so 10
1 0
10

we proved that the AF element converges as h 0, uniformly in t.

which doesnt blow-up as t 0. The difficulty is that this form is not Additional locking-free elements have since been devised, although
coercive, and it is tricky to find stable elements. none quite as simple.
25 / 44 26 / 44

An attempt at another simple locking free elt Thin elastic shells

An elastic shell is a curved structural

Onate-Zarate-Flores (1994): element whose thickness is much less
than the dimensions of its midsurface.
w
On the basis of experiments they declared this element locking-free. Shells are the primadonna of structures:
Again the nonconformity error is controllable. However stability is they can deliver remarkable performance,
subtle. A careful analysis (AFalk 1997) showed that the inf-sup but occasionally experience catastrophic
constant behaves like min(1, h2 /t 2 ). Thus we have stability if collapse.
t = O (h), but lose stability and convergence as h 0 with t fixed.
The locking problems encountered for
1
10 10
1 1
10
plates are much greater for shells, 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

over a wide range of conditions.
4
w t=1 4
w t=1 4
w t=1
10
2
t = .01
10
2
10
2
This stands as a major challenge in
w t = .01 t = .01
w w
1 0 1 0 1 0
computational engineering.
10 10 10 10 10 10

P1 -P1 AF OZF
27 / 44 28 / 44
The impact of instability on an eigenvalue problem Finite element exterior calculus
For many electromagnetic problems, 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 = u v dx v ,  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.
systems of partial dierential equations. This approach brings to
On a square the positive eigenvalues are 1, 1, 2, 4, 4, 5, 5, 8, . . . bear tools from dierential geometry, algebraic topology, and

10
homological algebra to develop discretizations which are compatible
9
with the geometric, topological, and algebraic structures which
On an unstructured 8
7 underlie well-posedness of the PDE problem being solved.
mesh, the discrete 6
 Finite element exterior calculus, homological techniques, and applications,
5
spectrum looks nothing 4
Arnold, Falk & Winther, Acta Numerica 2006, pp. 1155.
3
like the true spectrum. 2
1
0

On this structured 10
9
mesh the discrete 8
From the point of view of FEEC, Lagrange elements dont seem at all
7
eigenvalues converge 6
natural for curl curl problems. We should not discretizing a differential
5
to the true eigenvalues, 4
1-form with elements devised for 0-forms.
3
but also to a sequence 2
1
of spurious ones. 0

29 / 44 30 / 44

The Hodge Laplacian Mixed formulation of the Hodge Laplacian

To get an impression of FEEC consider the Hodge Laplacian. This is Introducing = d u, we get the mixed formulation of Hodge
the PDE most closely associated to the de Rham complex Laplacian. If we also account for the harmonic forms, the resulting
problem is always well-posed.
d d d
0 H 0 ()
H 1 () H n () 0
Continuous problem:
grad curl
0 H () H (curl, ) H (div, )
1
L2 () 0
div Given f L2 k (), find H k 1 , u H k , p Hk :
h, i hd , u i = 0 H k 1
Hodge Laplacian: Given a k -form f find a k -form u such that hd , v i + hdu , dv i+hp, v i = hf , v i v H k
hu , q i = 0 q Hk
(dd + d d )u = f
The proof of well-posedness, via the inf-sup condition, relies on two
fundamental results:
The degree of well-posedness is determined by the harmonic forms
Hodge decomposition: H k = Bk Hk Zk
Hk := Zk (Bk ) , with dimension equal to the k th Betti number.
k k 1
Poincare inequality: kkL2 c kd kL2 , H k , Zk
ker d range d
So we need to capture these in our discretization.
31 / 44 32 / 44
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.
dimensional subspaces Vhk 1 H k 1 , Vhk H k .
d k 1
H k 1 H k
We require the subcomplex property dVhk 1 Vhk in order to obtain hk bounded
k 1 k
the yh y h hk a projection
Discrete Hodge decomposition: Vhk = Bkh Hkh Zkh d k 1
Vhk 1 Vhk hk d k 1 = d k 1 hk 1
where Hkh := (Bkh ) Zkh .
Theorem
Generalized Galerkin method:
Find h Vhk 1 , uh Vhk , ph Hkh : If kv hk v k < kv k v Hk , then the induced map on
cohomology is an isomorphism.
hh , i hd , uh i = 0 Vhk 1
gap Hk , Hkh sup kv hk v k


hd h , v i + hduh , dv i+hph , v i = hf , v i v Vhk v Hk
huh , q i = 0 q Hkh kv k=1
The discrete Poincare inequality holds uniformly in h.
The generalized Galerkin method is stable and convergent.
When is this discretization stable, consistent, 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. We will construct the Vhk as finite
element spaces assembled on simplicial meshes, and so must specify On a simplex T , the space Pr k (T ) has a natural set of DOFs
polynomial shape functions and DOFs. given in terms of moments weighted by Ps l (f ) on a faces f .

A key tool is the Koszul complex Similarly Pr k has DOFs based on Ps l .

0 Pr 0 Pr 1 1 Pr n n 0 In this way we obtain the two primary families of finite element
with ()x := x yx. It satisfies the homotopy property homog
eneous differential forms.
polys.

(d + d ) = (r + k ) Hr k The DOFs also determine cochain projections. They are not

bounded, but they can be modified to become so.
It turns out that there are precisely two natural families of polynomial
differential forms to use as shape functions on simplices:
the spaces Pr k and the spaces Pr k := Pr 1 k + Pr 1 k +1

35 / 44 36 / 44
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 0 (T
r
0
) = Pr (T ) H 1
Lagrange elts Pr k spaces of constant degree r :
d d d
P n (T n 2
) = Pr 1 (T ) L discontinuous elts 0 Pr 0 (T )
Pr 1 (T ) Pr n (T ) 0
r
grad curl div
n = 2: P 1 (T ) H (curl) RaviartThomas elts 76 0 0
r

1
n = 2: Pr (T ) H (curl)
n = 3: Pr 1 (T ) H (curl)
n = 3: Pr 1 (T ) H (curl)
n = 3: Pr 2 (T ) H (div)
2 DemkowiczMonkVardapetyanRachowicz 2000
n = 3: Pr (T ) H (div)
BrezziDouglasMarini elts 85 Whitney 1957, Bossavit 1988
Pr k spaces with decreasing degree:
Nedelec 1st kind edge elts 80
d d d
0 Pr 0 (T )
Pr 1 1 (T ) Pr n n (T ) 0
Nedelec 2nd kind edge elts 86
grad curl div
0 0
Nedelec 1st kind face elts 80
Nedelec 2nd kind face elts 86
These are extreme cases. For every r 2n1 such FEdR subcomplexes.
37 / 44 38 / 44

Applications of FEEC

Stable finite elements for the Hodge Laplacian, curl-curl

problems, div-curl problems, . . .

Stable finite elements for Maxwells 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, which had been sought by engineers and
numerical analysts for four decades. But thats another story . . .

39 / 44