P UBLICATIONS MATHÉMATIQUES ET INFORMATIQUES DE R ENNES

G ILBERT S TRANG
The Finite Element Method–Linear and Nonlinear Applications
Publications des séminaires de mathématiques et informatique de Rennes, 1974, fasci-
cule S4
« Journées éléments finis », , p. 1-11
<http://www.numdam.org/item?id=PSMIR_1974___S4_A9_0>

© Département de mathématiques et informatique, université de Rennes,

1974, tous droits réservés.
L’accès aux archives de la série « Publications mathématiques et informa-
tiques de Rennes » implique l’accord avec les conditions générales d’utili-
sation (http://www.numdam.org/conditions). Toute utilisation commerciale ou
impression systématique est constitutive d’une infraction pénale. Toute copie
ou impression de ce fichier doit contenir la présente mention de copyright.

Article numérisé dans le cadre du programme

Numérisation de documents anciens mathématiques
http://www.numdam.org/
 THE FINITE ELEMENT ÎIET$OD—LIHEAR AND liONLINEAR APPLICATIONS

Gilbert Strang, M.I.T.

Numerical analysis is a crazy mixture of pure and applied mathe-

ratics. It asks us to *o trio thingc at once, and on the surface they

do appear complementary: Ci) to propose a good algorithm; and (ii)

to analyze it. In principle, the analysis sho'uldrevsal what makes

the algorithm good, and suggest how to Frake it better• For some preb-

lems--cornputin8 the eigenvalues of a large matrix, for example, v;hich

used to be a hopeless mess—-this cembination of invention and analysis

has actually succeeded. But for partial dlfferential équations, v/hich

corne to us in such terrible variety, there seems to ha a lonz way to 50.

Ve want to speak about an algorithm wbich, at ieast in lts rapicly

developing extensions to nonlinear problems, is still new and flexible

enough to be improved by analysis. It is knovm as the finite élément

method, and vas créâted to solve the équations of elasticity and pia-.-

tîclty. In this instance, the "minerical analysts" were ail engAm?ers.

They needed a botter technique than finite différences, especjally for

complicated Systems on irregular domains; and they found one. Thcir

metîiod falls into the framev/ork of the Ritz-Galsrkin technique^ which

opérâtes with problems in "variational forni" —startinr; cither from an

extremun principle, or from the v/cak fern of the dlfferential equati.cn,

. vhich is the engineer's équation of virtual work. The key^ idea v/hich

has nade th1s ciasslcal approach a suneess is to use pieeewise poly-

1
nom'jals as trial funetions in the variational problem.

'The most important applications are still to structural problems, but

no lonr.er to the deaifrn of airplancs; that has now bcen superseded by
the safety of nuclear reactors«
 - 2 -

We plan to begin by describing thc mcthod as it applies to linear

problems. Because the baslc idea is mathematlcally sound, convergence

can be proved and the error can be estimated. This theory has been

developed by a great many numerical analysts, and we can sumrcarize only

a few of the most essential points—the conditions v/hich guarantee con-

vergence, and which govern its speed. This* linear analysis .has left

eve^yone happier, and some divergent éléments .have been thrown out, but

the method itcelf has not been enormously changed. For nonlinear prob-

lems the situation is entirely dj fferent. It seems to me that numerical

analysts, especially those in optimization and nonlinear syster.s, can

still make a najor contribution. The time 1s actually a little short,

because the large scale programs for plasticity, buckling, and nonlinear

oîasticity are already bein/r written. BUt ever.yone is agreed that they

• trcmendously expensive, and that new ideas are noeded.

Honlinear probiens présent a new challenge also to the analyst v:ho

concerned with error estimâtes. The main aim of this paper is to de-

scribe some very fragmentary•results (Section III) and several open

questions (Section I V ) . We are primarily interested in those non-

lincàrities which arise, in an otherwise linear problem, v/hen the solu-

tion is required to satisfy an lnequality cohstraint. This is typical

of the problems in plasticity. The solution is still determined by a

variational principle, but the class of admissible functions becomes a

convèx set instead of a subspace. In other words, the équation of

Virtual work becomes a variational ineauality %

At the end we look in still a différent direction, at linear r>ro-

prammlng constrained by dlffcrcntial équations. Here we need not only

r.-cd algorithme and a proper numerical analysis, but also air-vcrc to

 - 3 -

the more fondamental questions of existence, uniqucness, and regularity.

II • Llnear Equations, The finite élément mcthod appliçs above ail to

elliptic boundary value probiens, which we-write in the followlns form:

Find u in the space of admissible functions V such that

:
(1) a(u v)-i(V)
f for ail v in V.

Standard example: / / ( u v + u v ) d x d y « // fy.dxdy

x x y y for ail v in tfj(ft).
This ls the weak form of Poisson*s équation -Au - f. Becauae the ex-

pression a(u,v) is in this case symmetric and positive definite, the

probien is équivalent, to: Minimize J(v) = a(v,v) - 21(v) over the

admissible space V. The "strain energy" a(v,v) is the natural norni

in v/hich to estinate the error.

The error cornes from changing to a finite-dimensional problem:

Find u. in S. such thât

h h
(2) a (u ,v ) « 4 (v )
h h h h h for ail v h in S . h

It is this problÊM v/hich the-computer actually solves, once it is glven

f ô r
a basis the space S^. Very briefly, it has to f a n
a , a n d t h e l o â d
the stiffness natrix « h^*i *J^ vector F^r» ^(•j)*
solve the linear sy-'stem KQ = F, and print out the approxinate- aolution

u. = I Q,*,. That sounds straightf orv/ard, but it is néarly impossible

A
•* J J

unless the basis functions $j are extrcmely simple, and nearly use-

less unless t h e y c a n provide a good approximation to the true solution

2
u. The finite élément method manages to combine both propertles.

*Y/e shall have to réfer to the bool: [1] and to lts bibliography, bpth
for the construction of piecev/ise polynomials and for the proof of their
approximation properties. Perhaps the favorites, when dérivâtives of
order m appear in the energy a(v,v), are the polynomials of degrée
m+1.
 - 4 -

Our plan in thls section is to su.nmarize four of the main points'

in the thcory of convergence Each of them is concernée! with the chance

in solution when there is a change in the problem—when the admissible

•space V is replaced by S , or the given

h a and l are approximated

by a h and l^. . To give sorne kind of order to the discussion, v;e for-

mulate ail four as applications of the "fundainental theorem of nunerical

analysis":

CONSISTENCY + STABILITY/^ CONVERGENCE.

1 # The classiez! Ritz-Galerkin case: The energy a(v,v) is symmetric

positive definite; is a subspace of V; a h « a and £ h - l.

Since every is an admissible v, we may compare (1) and ( 2 ) :

a u , v T h i s m e a n s t h a t i n t 5 e t e n e r
a(u,v ) = ^ h h h^* ? - E y inner product,"

h i s the projection of u onto the subspace S^. In other words, the

positive definiteness of a(v,v) implies tv/o properties at once

Ll, p. 4 0 ] : the projection u h is no larger than u itseïf,

a u > u a u u
(3) ^ h h ^ - < > )>

and at the same time u. is as close as possible to u:

a
C'O a.(u-u ,u-u ) < a ( u - v , u - v ) for ail v
h h In h h h S .
h

Property (3) represents stabllity; the approximations are uniformly

bounded. Given that u can be approximated by the subspace —in

this Ritz-Galerkin context, consistency is the same as approxinability—

convergence follows imraediately from ( 4 ) .

2. The indefinite case: u is only a stationary point of the functional

J(v). This corresponds te the use of Lagrange multipliers in optimlza-

tion; the form a(v,v) can take cither sign, and v may include two
différent types of unknowns—both déplacements and stresses, in the

"mixed method" and "hybrid method."

Consistency reduces as before to approximation by polynonials.

But stability is no longer automatic; even the simplest indefinite

v v v:hlch h a s a
form J(v) * i 2— . unique stationary point at the origin,

if* V is the plane R * — w i l i collapse on the one-ainensional subspace

given by v 2 « 0. Therefore, for each finite élément space S h and

each functional J ( v ) , it has to be proyed that a degeneracy of this

kind does not occur.

 - 5 -

The proper stability condition is due to Babuska and Brezzi:

(5) sup |a(v,w)| > c||w||.

I|v||-1
Brezzi.has succceded in verifylng this condition for several important

hybrid éléments. For other applications the vérification is still in-

complète, and the convergence of stationary points—v;hich is critical

to the v/hole theory of optimization—remains much harder t o p r o v e than

the convergence of miniraa.

'3. The modifled Clalerkin method: a and l are changed to a K and

t. (numerical intégration of the stiffness matrix and load vector),

n
and v^ may lie outslde V (non-conforming éléments)'.

The effect on u can be estimated by combining (1) and ( 2 ) :

(6) a U
h ~ h u , u
~ h^u e
^h^JCu^u-Uj^)-^^*)^^^).

Stability, in this situation, means a lov/er bound for the left side:

a ( u u u u ) c a u u u u ) #
(7) h ~ h* ~ h - ( - h> " h

Consistency is translated into an upper bound for the right side, and

it is checked by applylng the patch test: v/henever the solution is in

a "state of constant strain"—the highêst derivatives in a(u,u) are

SJ.1 constant—then u h must'coincide v/ith 3

u.

The patch test applies especially to honconforming éléments, for

s
v/hich a(v ,v ) h h the derivatives of v h introduce delta-functions,

which are simply ignored in the approximate energy a .

h This is ex-

tremely illégal, but sti.M the test is sometimes passed and the approxi-

mation is consistent. Convergence was estabîished by the author fer

one such élément, a,nd Raviart, Ciarlet, Crouzeix, and Lesaint have re-

cently made the list much more complète.

3
T h e patch test is also an idéal way to check that a finite élément
program is actually working.
 - 6 -
H. Superconvergence : Extra accuracy of the finlte élément approxi-

mation at certain points-of the domain. It was recognized very early

fl ,lfl
that in semé spécial cases—u -= f v/ith linear éléments, or u = f

v/ith c u b i c s — t h e conputed u^ is exactly correct at the nodes. (The

Ornen's function lies in S^,) And even earlier there arosc the dif-

culty of interpreting the finite élément butput in a more gênerai

problem; and its dérivâtives can be évaluâted at any point in the

domain, but which points do we chooso? This question is as important

as.ever to the engineers.

u u
In many problems the error ~ h oscillâtes v:ithin each élément,

and there must be points of exceptional accuracy. Thomee discovered

superconvergence at the nodes of a regular mesh,. for u^ * u x x ,

and his analysis has been extended by Douglas, Dupont, Bramble, and

Wendroff. It is.not usually carried out in our context of consistency

and stabillty, but perhaps it could be: consistency is checked by a

patch test at the superconvergence points, to see which polynomial

solutions and which derivatives are.correctly reproduced, and stabillty

needs to be established in the pointv/ise sensé.*

III.. Variational Inequallties. What happens when a constraint like

v \J> is enforced on the admissible functions v, so that the func-

tional J(v) is minimized only over a convex subset K of the original

space V? This occurs haturally in plasticity theory, vjhen v rep-

resents the stress; wherever the yield lirait is reached, the dif-

ferential équation (Kooke's lav/) is replaced by plastic flow. For

1
the minimising u, the "free boundary ' which marks out this plastic
s
région u « $ is not known in advance. Since such a solution u

lies on t h e e d g e of the convex set K, J(u) £ J(u+e(v-u)). is guaran-

teed only for c > 0. This translates into the yarlational ineouality

which détermines v:

and Scott
'•Convergence in wlll be discussed by Bramble^ at this Congress; it
is one of the outstandîng problems in the linear theory.
s
This boundary cannot be found by solving the original linear problem
and then replaçing u by nln(u,^)!
 - 7 -

(8) a(u,v-u) >_ l(v-u) for ail v in K.

In the finite élément method, vie minimize an approximate functional

c
J (v)
h a (v,v) - 2* (v)
h h over a finite-dimensional convex set K^.

Por example, the piecewise polynomials may be constrained by v h £

at the nodes of the triangulation., Agairi the rainimizirg u ft is de-

termined by a variatlonal inequality,

now a polygonal free boundary is to be

The practical problem is to carry out this minimization, and co:r.-

pute u ; we are in exactly the situation described in the introduction,

h

with many proposed algorithms and a difficult task of compariscn and

analysis. The theorctical problem, which assumes that u h has somehow

been found, is to estimate its distance fron the true solution u. We

want to report on thislatter problem, and it ic natural t o a s k the sane

four questions about convergence v;hich were ansv/ered in the linear case.

The easiest way is to take the questions in reverse order;

4. Superconvergence is almost certainly destroyed by the error in

n
determining the free boundary. Even in one dimension with u - 1.
2
u differs from u^ by 0(h ),

3. The approximation of a and £, by a^ and 1^ lead to no

difficultles; the identity ( 6 ) simply beccmes an inequality, if v:e

combine (8) and ( 9 ) , and the patch test ir> still décisive. The saaie is

true for nonconformins éléments, and the extra term A in the error

estimâtes [1, p. 178] is exactly conied from the linear cane.

2. It is an open problem, both for K and for the discrète K^,

to show how stability can compensate for the indefiriteness cf a(v,v).

1« This is the basic question in the nonlinear Rltz-Galerkin

method: if the trial functions in K h can approximate u to a certain

#
accuracy, how close is the particular choice u^? It is no longer ex-

actly optimal, because it is ne longer the projection of u. But we

2
hope to prove, in the natural norm I|v|| - a ( v , v ) , that

l l ^ h | | < c min||u-v ||. h

Pirst, we ask how large this minimum is, choosing v^ to be the

piecewise polynomial u^ in which interpolâtes u at the finite

 - 8 -

on the regularity of u. For our obstacle problem, v/ith -Au » f in

the clastio part and u «.# in the plastic part, it is JIOW knovm that
2 »

u lies in W ' # (Brezis and Kinderlehrer may announce this long-

sought resuit in Vancouver.) At the free boundary there is a jump in

the second derivative of u—v/hich absolutely l.'mits the accuracy of

the interpolation. Courant's linear approximation, on triangles of

size h, is still ôf order ||u-Uj|| = Q(h). But for polynomials of

higher degrec, and a smooth free boundary, this is improved only to

0(h )—and no cléments can do better. There are Ofl/h) triangles

in v/hich the gradient is in error by 0(h). .Therefore there is no

justification for usine cubic polynomials, and the question is whether

quadratics are worthvrhile; v:e believe so.

u u
To-prove that the actuel error ~ h is of the same order as

u - U j , we àcpônd on an a priori estimatc of Palk [ 2 ] . It resembles (**),

but the change in <8) and (9) frem équations, to inequalities produc•..-s

a new term:
a
(10) ||u-u ||* < ||u-v ||
h h + 2/(f+Au)(u-v +u -v). h h

V/e may choose any v^ in and any v in K—and for simpliciuy

s 2
we have specialized to l » « / fv and a = a h /|Vv| . The new

term is autcmatically zéro in the elastic part, v/here -Au » f f but

eîsewhere f + Au > 0.

To estimate (10), we take v =* and v ^ » U j — w h i c h lies in

breause it cannot exceed i> at the nodes, v/here it agrées with u.

(/ii the case of quadratic polynomials, some members of will go

• b.-vf the yield limit within the triangles, but we have-Jfco be

enough to permit that; it docsn't hurt the error estimâte, and

^ y it is only constraints on v^ at nodal "checkpoiato" v/hich can

be enforced in practice.) With this choice of v and v , the term s

h

in (10) are
1
(1) With Courant s linear finite éléments:
2 2
llu-UjH* * h ; /(f+AuXu-Uj) h ; /(f+ûu) (u -#> < 0. h

(ii) With quadratic finite cléments:

2 3 3
Hu-Ujlj ~ h*; /(f+Au)(u-u ) * h ; I /(f+Aunu^) h .
 - 9 -

(The next-to-last intégral is split into a part coupletely within

3
the plastic région, vjiere u-Uj * h , and a part formed from those

triangles which cross the free boundary. This transition région has
2
ârea 0 ( h ) , and the integrand u-Uj is 0(h ).) Substituting back
2
into ( 1 0 ) , the rates of convergence are h and h^ in the two c a s e s —

thèse rates are confirned by expérimenta

IV t Open Problems, True plasticity theory is a deeper mathematical

problem than the m o d e ! we have used above. The rcason is that the

history of the loading f has to be taken into account; a part, of the

domain can go from elastic to plastic and back again, as the external

loads are increased. Therefore incrémental theory introduces a time pa-

rameter, and a rate of loading f in the functional J — a n d it computes

the stress rate c. In other words, as Kaier and C^purso have shovm, we

have a time-dependent variational m c q u a l l t y ;

(li) min J(v) = J ( u ) , with K = {ve/f',v<0 where u(t)»*}.

veK(t) u -

Notice.that <rt each instant the convex set dépends on the current state

u. In a practical problem the state is actually a vector of stresses

and plastic multipliers, but we hope that this quasl-static obstacle

probien will serve as a reasonable model. We also hope that the new

results on regularlty can be extendcd to u(t). But eveh on this as-

sumption, there remain three new problems in humerical analysis:

(i) Keeplng time continuous, to prove convergence of the finite

élément approximations. The difficulty is that the convex Set K, and

therefore the minimizing u, dépend discontinuously on the current

state u; therefore it is not true that is close tô u v/henever

u. is close to u.
n
(îi) To admit finite différence approximations in time, and to dé-

termine the stability liraits on the lnterval Lt.

(iii) To find a qulck way of solving, with adéquate accuracy, the

obstacle problem which arises at each time step.

We bellevé that thèse are among the nost important questions in

nonlinear finite élément analysis, and that answers can be found.

 - 10 -

A second class of problems, of an entirely différent type, aricos

if .c a K ini-eresiert onl:* in the nultiplc X of the load f which will

induce plastic collapse. This is known as limit analysis, and no longer

requires us to follow the loading history. In place of minimizirg a

quadratic functional, the problem falls into the'framewerk of lnfiritc-

dimcnsicnal linear programminç. Here is a typical example, v/ith un-

•known stresses -c^(x y) $ and multiplier X: Maxlmize X, subject to

Equiiibriun: £ — * Xfj in SI, Z a j JL « Xgj on and

3xi

Piecewise Linear
a
Yield Conditions: Z b ^ ^ ^ < c in ft, 1 < a < M.

Suppose we make this problem finite-dimensional^ by assuming that

the stresses (and also the'displacements, which are the unknowns in the

dual program) belong to piecewise polynomial spaces S^. The contjnu-

ous linear programming problem is thon approximated, in a compietely

natural way, by a discrète one [3]. . But we know nothing about the rate
;

of.convergence.
 - 11 -

REFERENCES

1. G. Strang and 6. Fix, An Analysis of the Finite Element Method,

Prentice-Hall (1973).
2. R. Falk, Error estimates for the approximation of a class of
variational inequalities, Mathematics of Comoutation, to appear.
3. E. Anderbeggen and H. Kn$$pfel, Finite element limit analysis
using linear programming, Int. J. o f Solids and Structures 8 (1972),
$$-1431.

This paper will appear in the Proceedings

of the International Congress of Mathematicians,

Vancouver, Canada, 1974. I am very grateful for

the support of the National Science Foundation

(P22928).