Numerical Analysis of Partial Differential Equations (PDFDrive)

Jacques Louis Lions ( E d.
Numerical Analysis of Partial Differential Equations
Lectures given at a Summer School of the

Centro Internazionale Matematico Estivo (C.I.M.E.),
held in Ispra (Varese), Italy,
July 3-11, 1967
C.I.M.E. Foundation
c/o Dipartimento di Matematica “U. Dini”
Viale Morgagni n. 67/a
50134 Firenze
Italy
cime@math.unifi.it
ISBN 978-3-642-11056-6 e-ISBN: 978-3-642-11057-3

DOI:10.1007/978-3-642-11057-3
Springer Heidelberg Dordrecht London New York
©Springer-Verlag Berlin Heidelberg 2010

st
Reprint of the 1 Ed. C.I.M.E., Ed. Cremonese, Roma 1968
With kind permission of C.I.M.E.
Printed on acid-free paper
Springer.com
CENTRO IfJTER!~JAZIONALE PIATEMATICO CSTIVO
- S. ALBERTONI -
I' ALCUNI METODI D I CALCOLO NELLA TEORIA DELLA D I F F U S I O H C D E I
1JEUTRONI"
Corso tenuto ad Ispra d a l 3-11 Luglio 1967

ALCUIII METDDI DI CALCOLO IJELLA TCORIA A MULTIGRIJPPI DELLA DIF-
FUSIONF: D C I I.JCUTROEI1
- S , Albertoni -
PARTE la - Teoria s t a z i o n a r i a a multif!ruppi, (I,?)
5 1 - F,lotazioni e Problemi
Sia ncRn un aperto l i m i t a t o e xr ( x l , x 2 , . .xn) t R . . FIel
seguito considereremo g l i s p a z i ~ ' ( 0 1 , H~ (Dl, H A ( R ) , e sa g &
i n t e r 0 >o, ( L ~ ) , ( ) ~ , ( saranno
) i g - prodotti diretti
d i L 2 , H I , Hi.
Volendo considerare principalmente problemi d i "trasmissione"
r
supporremo = U
1
",
e s i a n o t = 30, r i = a n 1
. rispettivanente
l e frontiere d i R , ni: yrs - a n r n ans sono l e " i n t e r f a c c e " , e
cio& l e p a r t i d i frontiers conuni a ?r,zs,
I n f i n e vr s i a l a normale, d i r e t
t a verso l ' e s t e r n o , d i aQr. Pre
messo c i b supponiamo assegnate
l e funzioni r e a l i : D~ ( x l , tli(x),
?q
ci(x) c o ~~(x),fl(xl.(~,q=
.
= 1 , 2 , . , n ; i = 1 , 2 , , , ,g)
I problemi da r i s o l v e r e sono due,
e cio8:
Problema A - Trovare l a soluzione ~ ( x d) e l sistema:
soddisfacente a u[, = 0, e a l l e condizioni d i "trasrnissioneu
valide per x c y P s , essendo d l a d e r i v a t a conormale a as2 e

r
ur, us l e r e s t r i z i o n i d i u a nr e QS rispettivanente.
Osservazione I - Dal punto d i v i s t a f i s i c o l e (1) forniscono l a

d i s t r i b u z i o n e s t a z i o n a r i a d e l f l u s s o neutronico e n t r o un r e a t t z
r e R (composto da r e g i o n i d i m a t e r i a l i d i v e r s i R i ) ripartito
i n g gruppi d i energia decrescente ( u i 5 il f l u s s o d e i neutro-
n i d i iE e n e r g i a ) ove f i sono Le s o r g e n t i neutroniche e s t e r n e
e:
Di un "coef f i c i e n t e " d i d i f f usione
Ai un "coef f i c i e n t e " d i assorbirnento
~i un " c o e f f i c i e n t e " d i rirnozione d a l gruppo i-1 a 1 gruppo i
(Ccco perch6 cl:O)
B~ un " c o e f f i c i e n t e " d i f i s s i o n e .
-
Osservazione I1 E ' u t i l e per il s e g u i t o dare pi2 e s p l i c i t a m e ~
t e l a s t r u t t u r a formale d e l sistema (1):
Ipotesi sul problema A:
1) Positivita dei coefficienti D ~ A, ~ ,gi, Ci (supposti misura-

bili a limitati in R) e cioB: ~~(x))a>o, ~ ~ ( x ) s b > o~~(x))c>o
,
(i=2,3...),D i (x)2d>o; xi, v sono coefficienti numerici >o.
Pq
Inoltre si suppone C 1 (x) : 0.
2) Ellitticit; uniforme e cio8: 4 {A ; p=1,2.. .n, A 6 Rl,x c T )

P P
esistono m, H>o tali che:
per ogni i=1,2...g. Inoltre a sari abbastanza regolare da

garantire che tutte le operazioni di traccia abbiano un sen-
SO.
-
Problema B Trovare il massimo autovalore >o, lo, e la corri-
spondente autosoluzione positiva l+,(x)>o, per il Problema A omo-
geneo, e cioB per fl=o e ulr=o; tal caso fisicamente corrispon-
de ad una ripartizione di neutroni autosostenentesi (reattore
critic01 .
§ 2 - Soluzione del problema A

Cercheremo u in IIo1 (n) (u 1 r=o): le condizioni di trasmissione
appariranno come "condizioni naturali", automaticamente soddi-
sfatte nella formulazione variazionale, come B ben noto nel caso
d'una equazione sola (g-1). Per g>l il Pb non 6 autoaggiunto.
Si k allora prcferito seguire un ragionamento elementare basato
sul fatto che c 1(x)~o. Introdotte le forne lineari:
a(u,v)=
I n
Ei
1 Q
(cpq 'D
1
- - + Aiu iv i)dn
. axp axq
P9
siano A , B, C i corrispondenti operatori (matriciali) penerati

(nel senso delle distribuzionil, Per le ipotesi fatte sui coef-
ficienti essi risultano def initi su tutto ( H : ) ~ con codominio
c ( ~ Gli ) elenenti d i natrice corrispondenti sono:
ove :
Introdotti A, 3 , C il Problema A si riconduce alla.risoluzione

dell'equazione in u:
I1 procedimento esistenziale k allora il seguente:

Se u=o allora la:
(6) Aou=(A-C)u=f 5 costituita (si osservi la struttura delle

( 2 1) da g equazioni disaccoppiate (cl~ 0 del
) tipo:
Assumendo fe(f1-I)g (o pin senplicemente ( L Z ) ~ )per le ipotesi

1) e 2 ) ) esiste allora un operatore G: di Green(') che ci forni-
sce u1 = G: f 1 ( 6 : b un isomorfisrno di H-'(O) su H:(Q)). Trova-
to u l , u2 b fornita da: u2=G~(~'u1+f2) = G:(C~G:~~+~), e in ge-
nerale si ha: ui= C io ( C iGo

i-2fi-1
Ne consegue llesistenza (u=o) di un operatore matriciale di

Green, Go, per il nostro problema, che realizza un isomorfismo
di ( H " ) ~ su (i~:)~, e u=Gof r ( H : ) ~ soddisfa "naturalnente" alle
condizioni di trasmissione.
I1 caso p # o si tratta riducendo la (5) nella forma:
Essendo Go un operatore limitato da ( L ~ ) ~ +(H:)~, Bij moltipli-

2
catori in L (01, e llimmersione di (I~:)~+(L~)~compatta, allo-
ra T b compatto da (II:)~+(H:)~, e pertanto il Problema A 2 ri-
condotto ad un classic0 Problema di Riesz-Fredholm. Lo spettro
puntuale { p : } amnette llunico punto di accumulazione all'infini
to, e i l i i L- si accumulano solo verso zero.
LJ :
5 3 - Soluzione del Problema I3
a) In tal caso si ha:

Pep risolvere la (10) si 6 trovata una rappresentazione a nucleo
(funzione di Green) dell'operatore L-l. Se g.1 la funzione di
1
Green G (x,y) 6 stata trovata, nel caso dei coefficienti discon-
tinui, e per la la volta, da ~tam~acchia(') come soluzione di:
(11) A'G'(X,~) = 6 (6x,y misara di Dirac in (xiy))

x IY
Allora
u = 1n
~'(x,y) f(y)dy ? soluzione di A 1 u=f e si ha
Osservazione -
Questo fatto di Q'CQ non ci permette (bench?
plausibile) di concLudere che G' t L' (QxQ) (e i suoi iterati).
Allbra sfruttando il fatto che per l'operatore L si pos-
sono (vedi (6)) fare dei ragionamenti per singole equazioni (ez
sendo queste disaccopiabili) si pu6 costruire subito una matrice
"fornale" d i Green di elenenti ti;:
(O) Si osservi che Habetler e Martino (8) gii nel 1958 avevano
considerato il Problema B assumendo perb formalmente l'esisteq
za della funzione di Green.
Ad esempio:
se i Gi fossero t L'(QXR),
-
Si verifica poi subito che ~ - ~ z 1{ t2 ~tale
~ ) che
e ~( H : ) ~ ,
L-~LUEU, 5 ~la~matrice
e pertanto L - ~ I { ~ } di Green del
problema Lu = 4 .
-
Osservazione Per le ipotesi di positivits fatte sulle ci l'opee
tore L-' lascia invariato il cono, in (L*)~,dei vettori >o, come
pure, per le ipotesi sulle B ~ l'operatore
, L-~B. (compatto in
(L2)E).
Ne consegue subito, per noti risultati di Krein-Rutman, (l),
che esiste un autovalore massimo (dominante e semplice) X o del
roue
problena, . . = L - ~ B ~positivo
, e maggiore del valore assoluto
di ogni aitro autovalore, a1 quale corrisponde un'autosoluzione
uo pure >o in Q.
b) - Determinazione iterativa di A,.
Richiamiamo il Teorema di I. Marek, (6)

Se II,K sono due operatori lineari con DH, D K C K , spazio di
Hilbert, e se K B limitato e B 1 esiste limitato da X-DH, e H - ~ K
2 limitato con autovalore dominante X, (e autosoluzione x,) all2
ra il process0 iterative di Rayleigh-Kellog descritto dalle:
u (0) = x(~) (approssimazione zero!
converge:
Per avere una determinazione i t e r a t i v a d i X o basta a p p l i c a r e t a l e
Teorema n e l 7ostr0 caso K=B, H=L, X ; ( L ~ ) ~ .
S i o s s e r v i ora che il calcolo d e l l ' i t e r a z i o n e m + l
s i deve r i s o l v e r e un'equazione d e l t i p o :
e l e (131, come a 1 s o l i t o disaccoppiandosi, permettono d i t r o v a r e

.
l e u(m+1)9i una per v o l t a ( i = 1 , 2 , . .g) risolvendo problemi d e l
tipo:
A ciascuna i t e r a z i o n e s i devono p e r c i s r i s o l v . r e problemi d i

t i p o ben noto. Qcesto consente p e r c i s d i applicare tecniche sva-
i
r i a t e s t u d i a t e per problemi d e l t i p o A u = f .
Ad esempio l a u ("' ) 9i puh e s s e r e t r o v a t a , (9) , rnininizzando
il funzionale:
'L
essendo f i nota d a l l a precedente i t e r a z i o n e .
Ne consegue che adattando questo metodo s i ha un c i c l o doppiamente
iterative-variazionale per trovare uo, X o .
Evidentemente l a u ) 9 i pus e s s e r e t r o v a t a anche c o l meto-
do d e l l e d i f f e r e n z e f i n i t e , e appunto s i sono f a t t e d e l l e espe-
r i e n z e numeriche comparative a 1 riguardo.
5 4 - Esperienze numeriche
a ) Risoluzione d e l Problema ( 1 5 ) a t t r a v e r s o il metodo d i Riesz.
Se s i assune una'lbase" f i n i t a F v ( x ) ; v = 1 , 2 , . . . M, e s e s i rappre-
i cone
M
sentano l e ( = i a F
1
, l e condizioni d i ninino per
il funzionale (15) diventano:
i = 1 , 2 , ...g
- a= oI
aa.
;
1,v v = 1,2,.,.M
Queste c i danno un sistema algebrico l i n e a r e d e l t i p o ( p e r ogni i ) :
R ia (m+l),i-B(ra),i
- ...* aN(m+ 1 1 ,i ; (i=1,2,...~)
essendo R,S,Q m a t r i c i NXN d i elementi n o t i .

Ad esempio:
L'inversione d e l l e m a t r i c i R 2 s t a t a f a t t a con l'algoritmo di

Gauss.
b) Confronti t r a il metodo iterative variazionale d e s c r i t t o e

quello d e l l e d i f f e r e n z e f i n i t e (assunto cone elemento d i confron-
to).
Caso monodimensionale: g=2, r = 5 , D,A,3,C funzioni c o s t a n t i a
t r a t t i ; 1.1=20.
X o (con d i f f e r e n z e f i n i t e ) = Xdf
F sono l e autosoluzioni d e l l ' o p e r a t o r e d i

v
Laplace nonodin?ensionale
1 (con il n o s t r o metodo) = X N
0
Ecco una t a b e l l a r i f e r e n t e s i a i v a r i c a s i :
L'errore % 2 a 1 pih a t t o r n o a110
0,1%. Circa l'andamento d e l l e
s o l u z i o n i u l , u2 n e i c a s i s p e r i -
rnentati s i ha un accord0 d e l no-
s t r o metodo con q u e l l o a l l e d i f -
ferenze f i n i t e s i n o a '3 c i f r e s i
gnif i c a t i v e n e l l e zone c e n t r a l i ,
e uno meno buono (2 c i f r e ) n e l l e
a l t r e zone.
Caso bidimensionale -
R i s u l t a t i analoghi a i precedenti, perch2 l ' e y
r o r e 5 (!I eguale s i a FET l ' a s s e xl che per l ' a s s e x 2 ) n e i n o s t r i
e s p e r i n e n t i non ha n a i superato l o 0,38. (Fv sono p r o d o t t i d i auto-
soluzione d e l l f o p e r a t o r e precedente).
Caso tridimensionale: E - 2 , r=3; D , A , B , C c o s t a n t i a p e z z i , simmetria

r i s p e t t o a i p i a n i x , = o , x2=0. (Fv p r o d o t t i d i autosoluzione come
prima).
X): L
Risultati
a ) NZ = M = 1, I1 = 3 e cio2 N t o t a l e abbastanza piccolo
Y X
()I=FJ t1 N 1: con 6 i t e r a z i o n i (30'' IB?? 7090) A H approssima X
X Y Z df
e n t r o il 3,3%.
b ) aurnentando II da 3 a 10 l ' e b s i r i d u c e a 1 3 4 (1'47" d i rnacchi
Y
na),
In generale X = XI,[ 6 d i t i p o monotono (crescente i n N) e i n 15 it: ,
r a z i o n i a 1 p i h , n e i c a s i c o n s i d e r a t i , s i ha l ' a u t o v a l o r e can l a ap-
prossimazione c e r c a t a (1%) mentre 6 ben noto che con il netodo del-
l e d i f f e r e n z e f i n i t e il nunero d e l l e i t e r a z i o n i s a l e , i n genere, a l
meno a c i r c a 50460.
u l ,u2 ( s u l l e r e t t e y=6, Z=8 i n f i g u r a l sono i n buon accord0 con i
valori ottenuti a differenze f i n i t e ,
tranne n e l l e i n t e r f a c c e ove l o scar-
t o 4 ?. 2 , 3 % ,
Osservazione - I1 netodo i t e r a t i v o variazionale f o r n i s c e l'autovalore

massimo i n modo a s s a i soddisfacente, s i a p e r precisione che pep tern-
p i d i c a l c o l a t o r e ( r i s p e t t o a 1 metodo d e l l e d i f f e r e n z e f i n i t e ) , ma
invece 4 i n f e r i o r e a quest'ultimo metodo p e r l a precisione d e l l a t i
bulazione d e l l a soluzione s p e c i e per quel che riguarda l'andamento
d e l l a u2 che pu6 presentare d e i "picchi" n e l l e zone non c e n t r a l i
( e vicino a l l e i n t e r f a c c e ) , picco a v o l t e a s s a i ma1 d e s c r i v i b i l e
c o l metodo variazionale.
PARTE 2a - Teoria a multigruppi dipendente dal tenpo
5 -
1 E' ben noto che nella teoria della diffusione dei neutroni
nell'approssimazione a pih gruppi g di velocith, che supporremo
due per sempliciti, l'evoluzione del tempo dei flussi veloce e
lento, rappresentati da u l , u2 e della concentrazione dei cosi
detti "neutroni ritardati" rappresentata da C, 8 retta dal seeue;
te sistema:
essendo assegnate:
1) i coefficienti (funzioni nisurabili e limitate essenzialmente
>O) e le funzioni di "sorgente" f1,2,3 '
2) le condizioni (Dirichlet) per le ul, uz a1 contorno ~QxJo,TL ,
O<t<T, ncR, ,
3) le condizioni di "trasmissione" relative ad u l , u2 rispetto
una interfaccia y (una sola per semplicita) che rappresenta
la frontiera comune a due subregioni nl,n2
4) la condizione iniziale di Cauchy per tutte le incognite:
I1 sistema stazionario associato ad (11 8 ellittioo second0

Mirenberg-Douglis, ma non secbndo Petrowsky, e pertanto essenzial-
mente diverso, ad esempio, del sistema ellittico del tipo elastici
ti, Infatti, l'operatore natriciale associato 6 del tipo:
da cui si vede subito che sulla diagonale principale l'ordine degli

operatori sono rispettivanente 2, 2, 0.
Risulta quindi naturale assegnare le conclizioni a1 contorno (e di
trasmissione) solo per u2, u2.
Inoltre essendo il sistema a coefficienti discontinui i risul-
tati di Agmon-Douglis-Plirenberg per i sistemi ellittici (che esigc
no la continuit; dei coefficienti) non si applicano. Appare percib
naturale, pensare ad una impostazione di tipo variazionale.
Per questo vedi un lavoro di prossima pubblicazione di Albertoni-
Daneri-Geymonat. (5)
Non presenteremo, pertanto, qui risultati teorici circa (1)
essendo le ricerche in corso, ma considerereno solo il problem
discretizzato che si ottiene da (1) mediante il metodo delle linee
(discretizzazione solo sulle variabili spaziali) ottenendo:
essendo ( = vettore di R 2 H + N l ove !J 5 il numero dei punti del re-

tic010 spaziale (esclusa la frontiera) e !Il k il numero dei punti
ove h(xIf0. Inoltre 2:
v;' 0 v;' matrice diagonale NXN
matrice diagonale EJXN
E identiti in R N
1
essendo:
L1 + DIA una matrice pentadiagonale NXN ,
L2 + D2A una matrice pentadiagonale NXN ,
All + all una matrice diagonale NXN a c o e f f i c i e n t i >o ,
A12 + a12 una matrice diagonale NXN a c o e f f i c i e n t i > o ,
A2] + a 2 ~ una matrice diagonale NXN a c o e f f i c i e n t i >o ,
A22+ a22 una matrice diagonale NXN a c o e f f i c i e n t i >o ,
H + h una matrice tIIXN a c o e f f i c i e n t i >o ,
El + X una matrice NX?I1 a c o e f f i c i e n t i >o .
Osservazione I -
Lo s t u d i o d e l problerna d i s c r e t o ( 2 ) $ egualmente
n o l t o importante i n a n a l i s i numerica, perch$, a d i f f e r e n z a d i
quanto accade n e i c a s i p a r a b o l i c i , c ~ n c e r n e n t ii n generale p i c o
meno l a d i f f u s i o n e d e l c a l o r e , l a matrice pub avere a u t o v a l o r i > O ,
il che cornporta a v o l t e una crescenza n o l t o r a p i d a d i $ cosa sem-
pre d e l i c a t a da con-:rollare d a l punto d i v i s t a numerico,
Inoltre l e v ( i n generale c o s t a n t i ) sono 5 l o 6 , mentre
192
g l i a l t r i c o e f f i c i e n t i sono 2 1 , e p e r t a n t o d a l punto d i v i s t a
numerico s i incontrano d i f f i c o l t a s i m i l i a q u e l l e che s i hanno
n e i prohlemi d i "boundary layer" connessi con equazioni d i f f e r e n -
z i a l i contenenti p i c c o l i parametri n e l l e d e r i v a t e p i h a l t e ,
Usando schemi i m p l i c i t i ( p e r r a g i o n i d i s t a b i l i t s ) s i generano
da ( 2 ) "grossi" s i s t e m i l i n e a r i per i q u a l i occorrono netodi it:
r a t i v i l a cui convergenza, che e r a da indagare, B s t a t a v e r i f i c z
t a i n , (3)
5 2 - Proprieth del sistema (2).
In ( 3 ) sono s t a t i o t t e n u t i i seguenti r i s u l t a t i :
1) Q 5 irriducibile;
2) Q 6 essenzialmente >o;
3) Q possiede un autovalore wo>-A cui corrisponde un autovettore
v>o, t a l e che se ai 5 un qualsiasi a l t r o autovalore 8:
R a.<w
1 0;
4) d a l l e 11, 2 ) 31, seguendo Birhoff-Varga (7), s i dimostra (t-1:
((t)= K eYoiv + 0 (ept) con R a i < v w
0'
5 3 - Metodi d i risoluzione d i ( 2 1 ,
a ) Metodo Esplicito: + ( t )= (I+At Q ) ((t-At). Tale metodo 8 s t a t o
s c a r t a t o nei n o s t r i c a s i perch8 ha una soglia d i s t a b i l i t a t r o ~
po bassa (At troppo piccolo).
b) Metodo Implicito: 4 ( t ) = ( I - ~ t~ ) " + ( t - A t ) ,
Ad ogni passo temperale c ' 6 da risolvere un sistema del tipo:
Se A t w o c l a l l o r a , i n base a l l a p r o p r i e d 5 del 5 2 , i metodi

d i Jacobi e Gauss-Seidel r e l a t i v i sono convergenti.
-
Osservazione Quando s i ha un "transiente" molto rapido s i 6 tro-
vat0 che anche il metodo implicit0 (e pure q u e l l i d i Crank-Flicolson
e s i Saulyev ( 4 ) r i s u l t a n o molto imprecisi, Pertanto s i pone il
problema d i trovare qualche metodo meno imprecise. La valutazio-
ne ( 4 ) ha f o r n i t o l ' i d e a base per il seguente metodo che chiame-
reno metodo U.
Nel netodo w s t ; pensato d i esprimere l a soluzione n e l l a forma:
Allora l a ( 2 ) s i trasforrna i n :
In ogni caso per6 c t & il problema d i determinare w o che non

B determinabile con procedimento t i p o "metodo d e l l e potenze Ray-
leigh-Kellog" non essendo l t a u t o v a l o r e q u e l l o d i modulo massimo.
Questa questione 6 abbastanza d i f f i c i l e d a l punto d i v i s t a nume-
r i c o , perch; & vero che s i pu6 t e n t a r e d i " t r a n s l a r e " l o s p e t t r o
a1 f i n e d i condursi ad un problema d i autovalore d i massimo mod2
l o , ma c o s i facendo, dovendo poi s o t t r a r r e il passo d i t r a s l a z i c
ne, s e questo k molto grande s i pu6 perdere ogni s i g n i f i c a t o .
Allora posto B=M-N :
c i s i r i d u c e , come equazione a g l i a u t o v a l o r i per w (essendo w o

0
autovalore d i -VB) a l l a ( 5 ) : (M+uo v-' x = Mx.
I n t r o d o t t o un parametro f i t t i z i o v s i dimostra che l t e q u a z i o n e
(M+ w ~ - ' ) x - 1
v
Nx possiede un autovalore d i massimo modulo c u i
corrisponde un a u t o v e t t o r e > o o t t e n i b i l e c o l metodo iterative
d e l l e potenze. Q u e l l o che s i dimostra k che p = p ( w ) & monotona
decrescente, e che v = l individua w,.
I1 sistema ( 4 ) & poi r i s o l t o con il metodo i m p l i c i t o Crank-
Nicolson ed i r e l a t i v i metodi i t e r a t i v i r i s u l t a n o convergenti. ( 3 )
§ 4 - Metodo dei passi frazionari
Recentemente, (5) , abbiamo pure speri~entatoil metodo di
Narchuck ed altri per la (1) (g= 1) assurr.endo C =o e come deconposi-
zione dellfoperatore A una del tipo (vedi § 5 ) :
(7 1 a a)
Al. ax (D ax + ib
aY
-
A2= a (D -1
a
aY
+ ;b
Lo schema alternato per il passaggio da tn+ tntlltr[tn,tntl[

2 il seguente:
n+i
-v - tdt
AIU - 0
fornente u(tn+,) = untl (tntl). La discretizzazione della ( 8 ) ci

da poi sistemi del tipo:
"-1 &
dt
=
a1U ; V-' matrice diagonale;
(9)
v-' &
dt a2U ; al,a2 matrici tridiagonali
-
Osservazione Un primo vantaggio del metodo k che abbiamo ora a
che fare con matrici tridiagonali (invece di pentadiagonali) in-
vertibili anche con metodi diretti.
L1w0 c a l c o l a t o come i n d i c a t o a l l a f i n e d e l § 3 r i s u l t a oo = 63,69,
mentre il valore e s a t t o B 63,21, (Le d i f f e r e n z e f i n i t e sovrastima-
no l ' a u t o v a l o r e e , a n o s t r a conoscenza, c i sono r i s u l t a t i t e o r i c i
per questa stima s o l o n e l caso d i c o e f f i c i e n t i c o n t i n u i i n T I ,
I r i s u l t a t i numerici sono r i p o r t a t i i n t a b e l l a dove ;; & il valore
nedio su 0 d e l l a soluzione e s a t t a , h l a soluzione approssima-
t a o t t e n u t a con l a trasformazione w ed il metodo d i Marchuck, 3,
D
l'analoga soluzione senza trasformazione w , u quello o t t e n u t o
I Mu
con il metodo i m p l i c i t o che f a seguito a l l a trasformazione w , me5
tre la cIM 6 q u e l l a che s i r i f e r i s c e a 1 metodo i m p l i c i t o d i r e t t o .
Come s i vede i n ogni caso l a trasforrnazione w , conunque s i a asso-
c i a t a ad a l t r e tecniche, d2 i r i s u l t a t i m i g l i o r i , ed il metodo d e i
p a s s i f r a z i o n a r i s i B r i v e l a t o superiore, n e i c a s i f a t t i , a 1 meto-
do i m p l i c i t o .
t S o l u z i o n e esatta
-u - - -
- Mu U~ U ~ ~ u U~~
Bibliograf i a
(1) S.Albertoni: "Metodi v a r i a z i o n a l i per c e r t i s i s t e m i d i equa-

zioni a derivate parziali" - 1st. Lombardo Scienze e L e t t e r e
Vol. 100-1966.
(2) S .Albertoni - M.Lunel1i -

G.Mangioni: "Metodi i t e r a t i v i va-
r i a z i o n a l i per problemi e l l i t t i c i n e l l a t e o r i a d e i r e a t t o r i
nucleari" A t t i Seminario Mat. e Fis. Univ,Modena Vol.XIV ,1965.
(3) A.Daneri - A.Daneri - 1 , G a l l i ~ a n i : "A numerical approach t o

t h e time dependent neutron d i f f u s i o n equations" EUR 3742e -
1968.
(4) 1,Galligani: "Numerical s o l u t i o n s of t h e time dependent d i f -

fusion equations using t h e a l t e r n a t i v e method of Saul'yev" -
Calcolo Vo1.2, Suppl. 1, 111 - 1965,
(5) S.Albertoni - A.Daneri - G.Geymonat: "Existence and approxi-
mation theory f o r general d i f f e r e n t i a l equations of t h e multi-
group d i f f u s i o n r e a c t o r theory" ( i n corso d i pubblicazione),
(6) 1,Marek: " I t e r a t i o n s of l i n e a r unbounded o p e r a t o r i n non

s e l f - a d j o i n t eigenvalores problems and Kellog i t e r a t i o n
processes" (Cech.Meth.Journa1 1 2 (1962)).
(7) G.Birkoff - R,S.Varna: "Reactor C r i t i c a l i t y and non negative

matrices" J.Soc.Indust.App1.Math. 6 , 354-377 (19581.
(8) G. I. Habetler - M.A. Martino: "The multigroup d i f f u s i o n

equations of r e a c t o r physics, KAPL, 1886, J u l y 28, 1958
(9) G.Stampacchia: "Su un problema r e l a t i v o a l l e equazioni d i ti-

po e l l i t t i c o d e l 2 O ordine" Ricerche d i Mat., Vol. V (19561,
CENTRO INTERNAZIONALE MATEMATICO ESTIVO
(C. I. M. E. )
"PROBLEMS OF OPTIMIZATION AND NUMERICAL STABILITY'

IN COMPUTATIONSfI
Corso tenuto ad Ispra dal 3-11 Luglio 1967

PROBLEMS OF OPTIMIZATION AND NUMERICAL STABILITY
IN COMPUTATIONS )
by
I. ~ a b u g k a(Praga)
Computer Science is a new scientific discipline. An important part

of this discipline is the numerical mathematics. The "Art of Computationn
i s becoming science ; new questions and problems become important.
A typical problem i s the problem of the creation of numerical me-
thods, the determination of their "worthn and, in general, the choice of
the most suitable method for the given purpose.
For example, the program-library in a computing centre centains
mostly many algorithms for solving single mathematical problerfis. Opinions
on the expedience of these algorithms a r e usually quite different and sub-
jective. This statement i s still more apparent when a method of applied
mathematics is to be appreciated, especially in the field of scientific-
technical computations. These scientific - technical computations a r e that
part of the computer science in which I have some experience.
My paper will deal with questions which are more o r l e s s asso-
ciated with this kind of computations.
I think that these computations may be characterized a s a mathematical
and constructive way of processing (transformation) of the given informa-
t.on to the required one ) , I am sure that in scientific-techical
computations it i s necessary to emphasize the knowledge of information
which we may collect and the appreciation of its reliability. Further it is
In this paper some results obtained recently in Prague will be given

2 ) ~ e n r i c i23 defines numerical analysis a s the theory of constructive
methods in mathematical analysis (with emphasis on the word
llconstructiven) .
necessary to formulate clearly the required information on the given pro-
blem . The necessity of a mathematical and constructive way cf this pro-
cessing i s obvious here .
The "clarity fl of the given and required information is an important
part for a successful solution of a technical problem. Numerical mathe-
matics a r e the rudiments of this constructive processing of information.
Numerical method generates (in a constructive manner) a mapping, from
the class (space) of the given information to the class of the required
one. It i s important that this mapping i s defined on the entire class
- of
information. This class will be the domain of definition of the given me-
thod (mapping) .
Numerical process is an exact constructive law (prescription) of
creation of the given mapping.
.Computation is a concrete realisation of the numerical process in the
given case. We shall talk about - realisation
exact when we compute
without round-off e r r o r s and about a realization ( o r disturbed realisa-
tion) in a real computation.
It i s obvious that there a r e many different manners of a construc-
tive creation of one given mapping, i. e. many processes exist which
transform the given information to the requested one and solve the s a m e
mathematical problem. It is evident that the question of choosing a pro-
cess is very important.
It i s c l e a r that the choice and every optimization must necessarily be
relative to the given information. This does not mean, however, that so-
me methods might not be advantageous in a certain generality.
The manner in which we appreciate the method i s of great importan-
ce. My experience is that, from the practical point of view, it is very
important to respect an incredulity of the given information . This incre-
dulity can be of different kinds . Some of them will be shown in the next
v
I. Babuska
part of the paper . It i s essential that the method -and in general all con-
clusion - be stable with respect to these incredulities. I think that this
stability i s one of the most important points when choosing a method in
practice.
In the next part I shall point out some aspects of these questions.
2 . T h e p r o b l e m of q u a d r a t u r e f o r m u l a s 1 )
In this section I shall show some aspects of ideas, which I men-

tioned previously, in a simple case of quadrature formulas.
Let our task be to determine numerically
Jo
We shall suppose that we know the following about 'the integrated
function f(x) :
1. The function f(x) i s a continuous periodic function with

the period 2v .
2. We can evaluate only the function f(x) (i. e. compute the va-
lues of f(x) ) .
In this case, the simpliest quadrature formula T (f) i s mostly used in
n
practice, with
This formula i s the well known trapezoid formula.

I will now analyse the question, if there a r e any reasons for selectins
') In this part we a r e not dealing with the problems of the round-off
errors.
the trapezoid formula ; we may ask e. g. why the Simpson-formula isn't
better than the formula previously mentioned. Some arguments for choosing the
trapezoid formula (in this case of integration of a periodic function) a r e
included in some papers, e. g. Milne (25) , Davis (18) and others.
The e r r o r bounds for the trapezoid formula a r e studied in many
papers. See (4) (5), (21), (24)/ and others. We will now analyse the pro-
F
blem of the choice of the quadrature formula according to the information
we mentioned previously. In our considerations we shall confine the class of
possible formulae to the linear one,
The choice of the quadrature formula means, in our case, to determi-
ne of the sequence of linear functionals I in the form
n
with the requirement that Jn(f) + J(f) (weak) for all functions f(x)
of the given class of functions.
We shall measure the amount of work in using a formula by the num-
ber of evaluations of the integrated function.
Let us now assume that B is a Banach space. Then we can de-
fine
(2.4) w (n, B) = inf sup (2 a j f (yj) - J(f) 1
and
w(n, B) is the minimal possible e r r o r under the assumption that we know

only that I\ f \\ B 4 1. 3 (11,B) has analogous meaning when we confine our-
selves to use equidistant points in the quadrature formula.
We shall further introduce
(2.6) /\(n,B) = sup

" ~ ~ 6 4
A (n, B) i s evidently the error-bound of the trapezoid formula in
the space B. An objective measure of convenience of the given formula i s
given here by the comparison of A (n, B) with w(p, 8).s ( n , 8 ) reSP. This
appreciation is obviously relative to the space B.
The choice of the space B is very problematical in practice .
In majority of cases there is a large incredulity a s to whetherit is convenient
to take the integrated function a s an element of a certain space - B . If the
conclusion on the suitability of a formula is strongly dependent on the choi-
ce of B , then the conclusion is not "stablett and it is not advantegeous
to use that formula in practive. Further we shall see that this ttunstabilitytt
will appear in the Fase of the optimal formula, i.e. when we use the
formula whose error equals 9 (n, B) o r w (n, B) then the results
will strongly depend on the space B . Conversely , a formula will be
advantageous in practice if its e r r o r is nearly equal to 3 (n,B) or
w(n, B) but more o r l e s s independent of the space B.
Later we shall see that only the trapezoid formula which i s not
an optimal one, has this property. We now introduce a class of
Banach spaces of periodic functions.
Definition 2.1 The Hilbert space H (over complex numbers) will be said
to be periodic if:
1) Every f E H is a 2w periodic, continuous function.
2) Let 11 f signify the norm in the space C, then
3) If f h H , then g(x) = f(x+c)e H for every real c and !I f I! =lip H H .

The space H will be said to be strongly periodic if it is periodic
and if :
1) e
ikx
eH, k = ..., - 1, 0, 1, ... and IIe
ikx
(IH= -ikx
Oe
\IH .
5) If ( jl & ( k l , then ~ ( e ~ ~ ~H ~ ~ l ~ e ~ ~ ~ l l
for O606d2
and D does not depend on n.
At t h e beginning of t h i s section it w a s s a i d that f(x) i s a perio-
dic function. It i s obvious that t h i s information is insufficient. However,
I think it is convenient t o a s s u m e that the function f(x) i s a n element
of a periodic o r strongly periodic s p a c e H .
It is evident that now too we have a l a r g e incredulity a s
r e g a r d s the concrete selection of the s p a c e H. The importance of this in-
credulity is well seen i n t h e next t h e o r e m and example.
Theorem 2.1 Let H be a strongly periodic s p a c e with the n o r m
20
2
(2.9) \I~I\'- JY-l2
+ A \fl\ dx, A > 0.
Then the error-bound of t h e formula
where
i s equal to Q
2
(n, H) .
The t h e o r e m 2..1 affirms that the f o r m u l a (2.10) is an optimal one
if we a r e using t h e equidistant net. Now we s h a l l introduce the follo-
wing example
Wsin x
Example 2 . 1 Let f(x) = e , = 3,lO.
Then (f)=
1
1 2r
f(x) dx = 4,88079258586502208.. .
resp . 2815,71662846625447.
In Tab. 2.1 we show t h e r e s u l t obtained by the trapezoid formula
R(A' f o r A = 1 . F r o m this table we see that a n optimal formula used in a n
n
inconvenient s p a c e may give bad r e s u l t s . We s e e that the conclusion of the
convenience of the optimal formula i s v e r y llunstablell with r e s p e c t to the choi-
c e of H) . From this table we a l s o s e e that t h e trapezoid formula ( C = 1 ) gives
v e r y good r e s u l t s ; however, the following t h e o r e m is t r u e :
Theorem 2.2 F o r e v e r y periodic s p a c e H
(2.11) A (n. H) > g (n, H)

This theorem shows that the trapezoid formula cannot b e optimal in a
periodic space. Nevertheless t h i s formula i s v e r y advantageous in prac-
tice. The explanation of t h i s fact c a n be s e e n in the following statement:
T h e o r e m 2.3 . Let H be a periodic space. Then
No o t h e r formula has the p r o p e r t y that the left-hand s i d e of (2. 12) i s boun-

ded for all periodic s p a c e s (except f o r a finite number of indices of
n)
This t h e o r e m shows that the e f f i c ~ e n c yof the trapezoid formula i s
I
Nunber
O< sln x
of Tn(f) , f = e R ~ " ) (f) , f = eWsin x
points
n 0<= 3 q =1 0 ( X = 3 Cy = 1 0
8 4,88241999058958100 3047,90959481962441 4, 646334604 ... 2900, 3 5 0 3 0 . . .

16 4, 8 8 0 7 9 2 5 8 5 9 3 6 6 6 1 7 3 2815,77672896656761 4, 8 1 9 0 2 2 2 3 ... 2780, 1 4 C 8 1 . . .
24 4, 8 8 0 7 9 2 5 8 5 8 6 5 0 2 4 0 8 ' 2815, 7 1 6 6 2 8 9 7 9 0 3 7 5 8 4,85310536 .. . 2799, 7 4 3 9 4 . . .
Exact
4,88079258586502405 2815,71662846625447
value
roughly the same a s that of an optimal formula. This statement i s now
"stabletf with respect to the choice of H . The asymptotic optimality (in
the sense of (2.12) ) of the trapezoid formula i s valid only when the equi-
distant net i s used, i. e. when we compare A (n, H) with 9 (n, H) .
The situation becomes more complicated when we compare A (n, H)
with w(n, H) . What happens will be seen from the following theorems:
Theorem 2.4 F o r every sequence 5 2, . . ,f i,2 0 there exists a perio-

dic space H such that
(2.13) lim sup A (n, W = 00
n e w w (n. H) f n
Theorem 2.5 . Let H be a strongly periodic space. Then we have
(2.14) lim sup HI < rn

n + w w (n, H) G
Theorems 2.4 and 2.5 show that it is reasonable to demand the
universal efficiency relative to the s e t of strongly periodic spaces. The
necessity to confine Ourselves to some kind of incredulity is natural. This
is a general statement when dealing with every kind of incredulity. I
think , however, that the previously mentioned theorems show well the
role of the trapezoid and optimal formulas and the role of the incredulity.
We have been dealing with the analysis of the convenience of the
trapezoid formula
n
.
T (f) Now let us mention the Brror of T ( f ) .
n
Practically all the known e r r o r estimations a r e based on the choice
of the space H (or on the more general space B) .
The choice may be carried out a priori i.e. before the computa-
tion or a posteriori i. e. the choice is made with respect to the
results obtained during the computation.
The e r r o r estimate is then
The norm U f \( has to be estimated (a priori o r a posteriori) .

There a r e many papers dealing with the estimation of A ( n , H) for
different spaces, e. g. (5),(2) (24), (6), and others. Many results have
been gathered in special books . See e. g. (26), (34) and others. With
a suitable choice of H (resp B ) we may obtain the estimates in
the ,"derivative-freet1 estimates (22) ; for the estimation of
n' 1
9
w(n, H) and (n, H) see e. g. (5), (8). (21), (36) Moore's important and .
principal results (30), (31), (32) may also be understood a s an a posteriori choice
of B and an a posteriori estimation of the norm )( fll
B
.This a posteriori choice
is made here by the computer. I want to emphasize that a choice of an a priori
given class of possible spaces is given here. The problem of the optimal choice
of a space in connection with the e r r o r estimation will now be discussed.
Let f(x) be a 2r - periodic continuous function and let 4#
# xf) = E (H ; H periodic, fE H) .
Then the following theorem is valid :
Theorem 2.6. Let f(x) = 2

k= -m
ak e
ikx
Then
There is often a simple way of obtaining this "best" estimation :

7,. -
vv t: illiroduce
Definition 2.2. The function g(x) = & k=-a,
b eikx will be called an
k
00
overfunction to the function f(n) = x a eikxif bk
k -I
> a
kl
..
, k=. -1, O., 1.. .
k=-a,
This will be denoted by g@ f.
The overfunction can often be simply constructed. E.g .
sin x 2 1 cos x
e sin x e ( l t c o s 2x) . With this knowledge of the overfunction
it is possible to get an estimation of (f) in (2.16) .
Theorem 2.7 Let g@ f and g"' be a continuous function. Then
The efficiency of this estimation will be shown in the following exam-

ple :
Msinx Kcos x
Example 2.2. Let f(x) = e ,0(> 0 . Then f(x)@ e . In
table 2.2 we introduce the value 7n(f) and 6n =IJ(f) - ~ ~ 1( f )
for 0( = 10, = 50. From table 2.2 a good accordance between the
estimation and the real error may be seen. This error estimation
is closely related to the ideas, of Dshlquist (19), (20) .
We dealt with the analysis of the computation of .(2.1) . Similar
ideas can be used for the computationof the Fourier coefficients
We obtain a principal new problem when we want to compute simulta-

neously the values
P.
j = 1,. .. .
k Obviously the simplest way is to
J
J
compute these values independently. There is a question if it is possible
to gain something when we make the computations simultaneously.
I will show it in the simplest case. Let us assume that
we will compute both values J and J simultaneously. Put
0 1
-- .
Numbe:. 10 s i n x
f = e
0f
'points
n Tn ( f ) Jn (f) - Tn(f) '?n (f)
8 3047,909594819624415 232, 192966353369944 232,3732719565787845
16 2815,776728966567611 0, 0601005G03131402 0,0601005003 142606
24 2815,716628979037584 0,0000005127831 140 0,0000005127831 167
32 2815,716628466254842 0, 0000000000003720 0,0000000000003743
50 s i n x
f = e
8 0,6480887567505754520 +21 0,3548333783656418203 +21 0,5063081633888883995 +21

16 0,3384555456320188415 +21 0,0452001672473682524 +21 0,0452789214681998668 +21
24 0,2951999264551136014 +21 0,0019445480701799697 +21 0,0019445486761790910 +21
32 0,2932816292532110631 +21 0,0000262508682774314+21 0,0000262508682775479 +21
40 0,2932554985285131181 +21 0,0000001201435794864 +21 0,0030001201435794864 +21
48 0, 2932554985285131 1 8 1 + 21 0, 0050000002019997072 +21 0,0000000002019997072 +21
.
Table 2. 2
I. Babuska
Q(") i s apparently the minimal possible e r r o r in a simultaneous com-

0,l
putation.
We shall analyse what ca,l be gained by this kind of computation.
Let
((2.20) 8'")
o, 1 (H) = max ( h (n. H), 1 J1 1 )
This is the e r r o r if we compute J with the trapezoid formula and

0
if we put Jl(f) = 0 .
The following theorem may be proved :
Theorem 2.8 . Let H be strongly periodic. Then
(n)
' 0,l
K
lim sup cfi
n
dn)
0, 1
(H)
The theorem shows that we can gain practically nothing while performing
a simultaneous computation. Theorem 2.8 i s a special c a s e of theorems
which have been proved by P. ~ f i k r y l (33) .
We analysed the c a s e if only the function values were used in com-
puting. All I said can be done if we use also the values of k deriva-
tives. Here we shall assume besides (2.7) the the following :
(2.7) ~ \ f ( S ) ~ ~ c ~ C s ( H ) ~s ~= fOl , \...,

H k
and
and D does not depend on n .

In this case the space H will be said k-periodic o r k-strongly pe-
riodic.
I. Babuska
Analogously to (2.5) we now have

k n
(2.22) r k ( n , B) = inf sup 17
s=Oj = l
ajs) I(') (5i)- J(f) 1
j=l,... , n
s=o, . . ., k
NOW I shall mention a special result of K. Segeth (see (35) ) who studied
this field of problems. One of the problems here is roughly speaking,
the following :
Is it better to use more values of a functions in the quadrature
o r i s it better to compute and use the values of the derivatives?
An answer to this is given by the comparison between yoand yk . It

can be shown that for 2-periodic spaces
2
(2.23) %(n, H)=inf suplx 2 (s) (6) 29- .
a. f
s=O j = l J
(-
n
J ) -J (f)1
Let us assume that the amount of work needed for the evaluation
of f(x) i s equal to 1 and that for the derivative i s & . Then the whole
work with the use of n points will be n(l +d) . This value will be
the measure of the "work" when using the given formula with n
points .
9, (n1 H)
(2.24) S(o(, H) = lim sup
.n w y ([n (ltool , H I
gives now the required answer (relatively to the space H) .
Thus, for example, the following theorem is true :
Theorem 2.9 .Let H be a 2-strongly periodic space. ~ e t O ( & 1 .
Let ~leinxll; = g(n2) where g i s a? entire function. Then S ( Q , H)>l. If g is
not a polynomial then S ( O ( . H) = w for@ > 1 and S(1, H) = 3.
I. Babuska
Theorem 2.9 shows more o r l e s s that if the amount of work needed for the eva-
luation of derivatives i s not l e s s than that needed for the evaluation of the fun-
ction, it i s not advantegeous to use the formula with derivatives.
Previously in this section we dealt with the trapezoid formula T
n
. An ana-
logous role is played here by the formula
n n
(2.25) T(2) (f) =
n n k=l
1 Iz
f(
n
k) .+
n3
a I k=l I" (
n
9
k)
There i s also a theorem analogous to Theorem 2.9 for the use of (2.25), given
more exactly and in detail more in (25). As an illustration I shall give the fol-
lowing example: 2 11
1 sin x
Example 2 . 3 . Compute also J(f) =
2v
- f(x) dn for ffx) = e e ' ,P
= 10,50.
Let us assume O(= 1 . 0

In table 2.3 we s e e the e r r o r when using the formulas
T and T; 2) In . dependence on the amount of work (i.e. on n resp. n ( l t C ( ) . We
n
s e e that the computation without the sue of derivatives i s more advantageous.
This agrees fully with the theoretic investigations. In accordance with the theo-
r e m the e r r o r of the formula with derivatives is nearly three times l a r g e r than
that of the formula without derivatives. All we said was connected with the com-
putation of .(2.1) , and (2.18) respectively.
Now I shall briefly speak about the computation of
212
(2.26)
0
We shall not analyse all the problems associated with this computation. All
can be done analogously. The formula which plays the same role here a s the i
trapezoid formula i s the following (see(6)) :

(2.27) (
(g) 1 - In f
1 n(3
n j) f(?g
j)
where
r;J b(n) eikx
Sn(x) = C k
k= -C"]Z t 1
b(n) = b for
k k
ir < l;]
I
10 s i n x
f (x) = e f (x) = e 5 0 sin x
Amount E r r o r of t h e E r r o r of the E r r o r of the E r r o r of the
of work f o r m u l a without f o r m u l a with f o r m u l a without f o r m u l a with
derivatives derivatives derivatives dirivatlves
(T,)
n ( ~ ~) ( ~ 1 (T,) ( ~ ~ ( ~ 1 )
16 0,60100 -1 0, 18030 0 0,45200 +20 0, 15147+21

0,37 -12 0,11200 - 1 1 0,26250+17 0, 78754+17
0,20199+12 0,60599+12
1
Table 2. 3
v
I. Babuska
The e r r o r estimation by a n overfunction can be made. As an illustra-

stion I shall show
Example 2.4 . Compute
(2.28) .=I
*
t-'ir
--
2
dsin x
e cos x dx, c ! = 1,5
Apparently this integral may be written like this :
In the table 2.4 we have shown the e r r o r s of (2.27) in compari-

son with the Qombergs integration (see (16) (17) ) . In the table
there i s also shown the e r r o r obtained by the use of the overfun-
COS X
ction e cos x. The computation was made with computer ICT
1905 in double precision .
In a simple case of quadrature we have shown some aspects
of incredulity with respect to the given information and the meaning
of "stabilityn of a conclusion. It is possible to generalize these ideas
in different ways .A possibility of a generalization can be seen in
(28) . I shall, however, not deal with it here.
3. Boundary-value Problems for Ordinary Differential Equations
In section 2 we showed one kind of incredulity a s re.gards the

given information and how to deal with it. I s h a l l now mention so-
me other aspects of incredulity. A simple problem will again be anali-
sed. Let us solve the following boundary-value problems
r
:-
;%.
6.- -
Od'a
3 3
m1 a1 mI
P;$
w +ar- e m m
e t
;;
@a
' x g CON I 3
d dI d do- 0"
;E I I I
W w
k X
UY
X a,
a, II
"
II "
-0
a, c
o m m *
1 + +
I I
.+a-a
CONOW
M ~ C O N
l g 0-66 dddS
x
Y
5
r:
"
.r(
a,
UY
a,
II
II
W h
r- w
N
@am
. N C O * C O
-@a 1 3 3
w . I l l
e 3 O d N C -
O Z mt- . + m a d
b . oo,o (0 C\1
. . n
~ N V F :
dddb'
W
u Z a ~t - m m
-+cum
a t - * m
P I C 3 N m
E;?
2 > w
Q O r :
with the boundary conditions
We assume that p(x) , g(x) and f(x) a r e sufficiently smooth and
P (4>#dl> qt (XI )) 0
The functions p, $ ,f have a physical meaning. Nevertheless,
we know them only approximately in practice .
Let the poss\ible disturbances (incredulities) of p , q , f be
cr,9 , respectively. F r o m the physical point of view these perturban-
c s s a r e small i n a certain sense (norm). They may also have further
properties. Such perturbances will be called admissible disturbancies. We
shall assume that small admissible disturbances result in a small change
in the solution .
It i s well known that a numerical process cannot be realized with an
absolute exactness. Every realization of a process by computation i s distur-
bed (by round-off e r r o r s ) . We can mostly imagine, however, this distur-
-
bed realization a s an exact one (without disturbance) but with thc distur-
bed given information. We shall speak about replaced disturbances (of
information) in this case I ) . It i s reasonable to speak about a suitable
numerical process if the replaced disturbances
a ) a r e admissible
b) the order of disturbances i s the same a s the order of e r r o r in the indivi-

dual operations.
he method of replaced disturbances (backword-method) was used with

l a r g e success by Wilkinson. See (I&),
( 45).
Bauer (13) (14) (15) used a similar a p p ~ o a c hin his investigations of
I 1
numerical processes in algebraic problems.
T h e r e . a r e suitable and non suitable processes
0-
. I shall show them
by the process of solving (3.1) and (3.2) .
Example 3.1. The method of combination of solutions leads to a non suita-
bid process. This method, a s known , consists in solving two ini-
tial-value problems for the initial conditions y(0) = 0, y(0) = 1 1,2
and the required solution (3.1) (3.2) i s determined by a suitable
combination. Let p(x) = (1 + x) , q(x) = 500, f(x) = cos x - (500 +
2
t 9 (1 t x)) sin 7 x, L = 1,a =b = 0 . We solve the initial problem
by the Runge-Kutta-Gill method of the 4th order for step h = 0,025
( computer LGP 30) See (12) . The results obtained a r e given in Table
Example 3.2. The factorization methods leads to a suitable process.

By this method (see e.g. (12) ) we solve the following system
Let us solve the same problem a s in Example 3 . 1 by this method.

The initial problems a r e also solved with Runge-Kutta-Gill method with
h = 0,025 . We obtain the results mentioned in Table 3. 1 .
I have said that we can mostly consider the disturbed realization
of a process a s an exact realization with the disturbed input (i.e. given)
information. In this case in the method of factorization the replaced
disturbances a r e small in the following norms : 6,? in C norm
and 7 in the norm ~= lfldx j x ~\I c . It may be seen that
0
x y ( x ) by m e t h o d of Exact solution y ( x ) by m e t h o d of
combination factorization
0,100 0, 3090103 0,3090170 0,3090018
0,400 0, 95 10075 0, 9510565 0,9510461
0,500 1,005C31 1,000000 0,9999897
0,700 0,8577343 0,809017 0,8090081
0,750 1, 374171 0,7071068 0,7070985
0,800 0,0000000 0,5877852 0,5877778
0,900 9,700032 0, 3090110 0 , 3090119
Table 3. 1
these disturbances a r e admissible.
It i s obvious that the questions of existence of a suitable numerical
process for the solution of the given problem i s very important. The
method of factorization may be generalized to a general boundary ( o r
multipoint) problem for the system
x'(s) - A(s) x (x) = f(s).
J. Taufer , s e e (38), (39) , has investigated in detail the replaced

disturbances for a concrete kind of factorization and has shown that
his factorization method is suitable in the previously mentioned sen-
s e , Another kind of factorization method, sometimes called method of
the transfer of boundary conditions, was investigated in recent years,
f o r example, by Abramov (I), (2) who also briefly mentioned the
possibility of showing the suitability of this process for the general
case (-3.4). See (3) .
In (7) and (12) the stability of the differential equations of the
factorization method in special c a s e s has been studied.
Example 3.3
As an example I shall show the computation of a continuous
beam of 20 fields built in at the end and constantly loaded. In practice,
the method of transfer of matrices which i s very similar to the method
of combination of solutions i s very often used. See e.g. (45) .
In the following table 3.2 there a r e shown the moments at some
supports computed by the usual method a s well a s by Taufer s facto-
rization method. The previously mentioned factorization method can
also be used in solving the eigenvalue problem. See (40) .
4. Stability of numerical processes.
In the previous sections we dealt with some aspects of increduli-
ty a s to the choice of a numerical process.
J .$
U
*:
+
w .rl
c L m m m m
:22 m
c
m
1 1
m d '
n o o
o o c
1
m
m
n
1
m p o m
h m 0 m
a J2 m o o m
c n o o m
m o o c n
U
lu
::
P)
:E
U
m m 4 4
I l l 1
9: m
m m
m ln d'
m C U ( 0 4
Ex$
a 0
m N c m
c n c n m o
m d ' r l m
Q a
w
a E m m a ,
a
a m a, $ 4 ; ;
g a l $
65 z
:;E:
a
m
1
m
1
m
1
m
1
z?
* al
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
5 0
0
.
0
0
0
. . .
0
0
0
w" 2 l n l n l n l n
w
0
$ x
ti 7 N
4
m
4
O
N
1
d
I. Babuska
In this section we shall deal with a quantitative characterization of the

numerical stability of a given numerical process. See (9), (lo), ( l l ) ,
(12) . In computations of problems of mathematical analysis, the exi-
stence of a subscript -n (e. g. number steps) i s typical s o that we
obtain the required result only for n + a . In section I we
introduced a numerical process. Here we shall define it more exac-
tly .
. ..
. Let there be given a sequence of normed vector spaces
and a sequence of continuous operators :
mapping the Cartesian product
x(") x x x . . . x xln)
1
into x(n) .
it1
-'n -pn+l
Further let the s e t s
)M: C )x: tor R = - p n, -pn+ 1 , . . ., 0

be given. Then the sequence of equations
I. Babuska
will be called a numerical p r o c e s s . The s e t M$) will be called the

set of input data and t h e e l e m e n t s x") k = 1 , 2
k n
.. .
b e called the solution corresponding to input elements x(n!k=-pn,
k
. . ., 0.
In p r a c t i c e , the n u m e r i c a l p r o c e s s e s a s by Definition 4 . 1
cannot b e solved exactly by the computer (round-off e r r o r s ) . Hence
we introduce the following definition :
Definition 4.2.
Let t h e r e b e given a n u m e r i c a l p r o c e s s in the s e n s e of Defini-

tion 4 . 1
k = -p
n'
.
. ..
, 0 and a sequence of n u m b e r s {- asn)
L e t t h e r e be given the input e l e m e n t s x (n)
n
k
=
'
(n) f
j = - pn, . . ., Nn ; n = 1,2,. ., and denote x/ )E X!n b . ., 0.1,. . .
. .. , N n the e l e m e n t s satisfying t h e eqwatiops;
The solution of the given numerical p r o c e s s corresponding to

input e l e m e n t s x 0')
k '
k = -pn, . . ., 0 and to the sequence f ( n ) will be called
ps - solution if
1
(4.4) lim sup & (irP (n) s u p \ ;("I- xin)( 4 c ns
~4 o 14, 16 ai L , i=-pn' ' " Nn .
and C does not depend on n .
We will s p e a k about Bs -solution if So = inf s .
The investigations of concrete given processes have been done in the
previously mentioned way in many cases. See e.g. (12), (27)) (32))
(41)) (44)) and others.
I shall now give some examples explaining the meaning of the
previous definitions. Let us solve the initial problem for an ordinary
differential equation
The Runge-Kutta method can be written as follows
A slight change will be made to simplify the notation in Definitions

4.1 and 4 . 2 . We shall investigate two processes
Let further
v
I. Babuska
The spaces x ! ~ ) ,Y[P)z[~!

1
dn)a r e spaces of real numbers with
the norm ) x \ . The meaning of the previously mentioned mapping
is evident in this ease . The numerical process i s also clear. The
disturbed process is as follows
I1 . The equations (4.9) (4.11) (4. 12) remain unchanged.

(4. 10) now has the following form
We can compute these processes in a different manner.

These computations differ in disturbances. The following mathematical
models can be assumed
a ) Fixed point computation
b) Scaled point ifloating point) computation

v
I. Babuska
bli) Normalized floazing point cornputatinn
c ) Normalized floating point computation with computation (4.9) in t h e

process I1 in double precision
The sequence f (n) is obvious and I shall not d e s c r i b e it.

The following theorem may be proved.
T h e o r e m 4.1. The previously mentioned p r o c e s s e s a r e
a ) B1 solution,
I
b),b) B solutinn,
1
c) Bo solution.
I shall now show the meaning of t h i s t h e o r e m by m e a n s of the
following example :
Example 4. 1. We s h a l l solve the initial problem f o r the equation
with the standard Runge-Kutta method of the 4th d e g r e e .

1 . 1
y(a t -) 1s t o be solved. H e r e we obviously have C = - . Our t a s k is
2 -
to e s t i m a t e 1 y (n) - y (n)
1 independence on
2
n. Since t h e solution of
n n
( 4 . 1 4 ) a n d (4.15) is y(x) = - xm and we do not know
'i
("1
I. Babuska
1 1
we shall use y%)+ y(a+;) and put =)~:)-~(a+~)]
In the following figures there a r e the outcomes of computations. In

fig. 4.1 a a r e the results for the process I. b, a = 0 , 5 , obtained
with MINSK 22. The parameter n has been selected a s a deca-
did value.
It i s interesting to ask what happens i f w e investigate
(n) = -0') - (-(n)
x n )I .
It may be shown that this i s also a
7 n 'Yn
.
B solution In fig. 4.1 b we see the results .
1
A further interesting question i s what happens if we use n
diadic. In fig. 4.2 we the results & (nl f o r n diadic.&'e
also see
n
s e e that this computation has a different character. This i s more o r
l e s s an accident.
In computation 1I.b there is no difference between diadic and
dedacid n as mentioned above.
F o r a diadic n exactly the same results for computation 1.b and
11. b . a r e obtained. In fig. 4 . 3 and 4.4 we have the results for I. b
n dedadic and 1I.b .
In fig. 4.5 we have the results for 11. c and a = 300 .
From the mentioned example we can clearly s e e that in
computations there a r e different kinds of importance with respect to
the stability. E.g. we have seen that the floating point makes the round-
off smaller but 'the results remain unchanged. Such a kind of consi-
derations may be very valuable in practice, yet we cannot deal
with it here.
It i s obvious that the knowledge of stability, especially the
B stability i s an important factor in a suitable choice of method.
S
"This computer i s a diadic one.

- 56 - 9
I. Babuska
There i s a question if it exist a B0 solution for computation

in simple precision normalized fleating point. The answer i s positive.
For the sake of simplicity , I will show one of them on the example of the
k
quadrature formula
n
TLet . nm2 . Let the process of the computa-
tion of T be the following
n
=k +P , k =O.
kn+l n k-n o
Then the computation is a B solution provided that we compute

0
with normalized floating point and simple precision.
We shall show an other interesting example. Let us solve the ini
tial problem for the differential equation
The usual difference method leads to the following formula

1 2
(4.9) Ynt3 -
YntZ .I-yntl =E (13fn+2'2fn+l fn)
This formula can be written in the following form

(4.10) z - z =hf
ntl n n
Ynt2 - Yn+l = * ( I 3 'nt2 - "n+l '

and the following theorem i s true :
Theorem 4.2. The numerical process based on (4.9) resp. (4.10) is a B
2'
B
1
prccess respectively for 4 (n) = [ 1, I, . . .I
This example (see (12) (41) ) shows the possibilities of getting a better
stability through simple changes in the method. The question when it
is possible to write a formula in a forma having a botter stability is sol-
vedin (41).
As a further example I shall show the numerical stability of the
numerical process of overrelaxation for usual finite-differential equations
(see (32) ) . We put N = co in the definition
n
4. 1 . Let us measure the
error of the result in the norm
where &
i s the e r r o r in one point of the
J
net; m i s the numer of the net-points and h = - i s the step.
n
Let the matrix A of finite-difference equations have the form
where I is the unit matrix, Then the following theorem holds.

Theorem 4.3. Let the previous assumptions hold. Then the numerical
process is
2
a B
process if 0 < - < w -< 2-C .h, C >0 .
Evidently a special case of theorem 4.3 is when w is indepen-
dent of h o r w is the optimal overrelaxation parameter.
We shall introduce an example.
Example 4.2. Let us solve the one dimensional proble y n = 1, y(0) =y(l)=O
with the finite-difference method and overrelaxation. Because of the
round-off e r r o r the iterations do not, in general converge to the requi-
red solution. They will uquasi-converge" in a more o r l e s s well known
1. .
sense. In Fig. 4.6 we see n = - ~n dependence on h.
h
We s e e a good agreement with theorem 4.3. It i s possible to
formulate the theorem 4.3 for a 0 < w -< 1 in a more general form.
See (12) .
Further processes have also been investigated. I shall mention
here the stability of the Kellog process for the determination of eigen-
value (see (27) ) and a numerical process for solving a problem of
the theory of reactors (44) and the process for computation of con-
form mapping . (see also (12) ) .
I have shown a few different aspects of incredulity with regard to
the given information which appear in computations. I think that this
kind of investigations is very important when choosing an algorithm in
general.
- 59 -
Fig. 4. 1. a.
I. b decadic
Fig. 4.1. b
10
(0.5, 1 )
I . h decadic
n -
Fig. 4 . 2 .
( 0.5 , 1 >
II. b
Fig. L. 3.
Fig. 4 . 4 .
(300, 300.5)
II. b
I. ~ a b u i k a
References
----------
I11 A.A. A ~ ~ ~ M o6 BI I:e p e H O C e PPaHMllAbIX Y C J I O B M ~AJlR CMCTeM J ~ H H ~ ~ H H x
O ~ H K I ~ O B ~ H H ~UM @ @ e ? e H ~ Y I t v r b H bYl ~~ B B H ~ H M BblrI.MaT.

~ ~ I4 MaT.
1$14a. 1 ~ 0 1 ,1; 542-545.

L2] A.A.. A6pa~oe: B a p a a ~Mf2TOAa
~ IIpOrOHKEI. , i t . ~ h l r l . ~ B T . YI MBT. 9143.
1961, 1, 349-351
131 A.A. Abramov : Transfer of boundary conditions for system of
ordinary linear differential equations. Proc. of IFIP Congress 65,
p. 420
[5] I. Babuska : Uber die optirnale Berechnung d e r Fourierischen Koeffi-

zienten. Apl. Mat. 11, 1966, 113-122.
[61 1. Babuska : U'ber universelloptimale Quadraturformeln. Apl. Mat. 1968
17J I. Babuska , M. P r a g e r : Numerisch stabile Methoden zur ~ a s u n g

von Randwertaufgaben. ZAMM 1961 , H. 4-6
181 i1.&36yd~a, C .J.C06one~: ~ ~ T H ; ~ I M ~ ~9MCJIeHHblX

u M R MeTOAOB.
Apl.mat. 10, 1965, 96-129
[91 I. Babuska, M. ~ r a ' ~ e E.
r , .VitBsek; Numerische ~ t a b i l i t 2 t von
Rechenprozessen. Wiss. Z. Techn. Hochsch. Dresden 1963, 12,
101-110.
1101 I. Babuska, M. Prager, E. VitBsek; Numerick6 r*es'eni differenciil-

nich rovnic 1964 , SNTL
0 11 I. Babuska, M.PrBger, E. VitBsek: Stability of Numerical Processes,
Proc. of IFIP 65 , 602-603 .
1121 I. ~ a b u g k a ,M. PrBger, E. VitBsek: Numerical Process in Differen-
tial Equations. Interscience Publishers 1966
(131 F. L. Bauer: Numerische ~ b s c h z t z u nund
~ Berechnung von Eigen-
werten nichtsymmetrischen Matrizen. Apl. Mat. 10, 1965,
178-189.
1141 F. L. Bauer et al. Moderne Rechenanlagen , Stuttgart 1965, p. 64.
[151 F.L. Bauer: Genauigkeitsfragen bei der Losung linearer Gleichungssy-

steme. ZAMM 46, 1966, 409-421.
b63 F.L. Bauer, H. Rutishauser, E. Stiefel : New Aspect in Numerical

Quadrature. Proc. of Symp. in Appl. Mat. 1963 , XV, 199-218.
I. Babuska
1171 R. Bauman : Algol Manual d e r Alcor-Gruppe, Sonderdruck a u s

Elektronischen Rechenanlagen H 516 (1961) H 2 (1962) R. 01-
denburg ,Munchen.
1181 P . J.Davis: On the Numerical Integration of Periodic Analytic
Functions. Proceedings of Symposium Madison 1959.
[191 G. G. Dahlquist : On Rigorous E r r o r Bounds in the Numerical Solution
of Ordinary Differential Equations. Numerical Solution of Nonlinear
Differential Equations. Wiley 1966, 89-96 .
b0] G. G. Dahlquist: P r i v a t e communication.
[21] H. Ehlich : Untersuchungen z u r numerischen F o u r i e r analyse. Math.
Zeitschr. 91 (1966) , 380-420 .
1221 G. Hammerlin: Uber ableitungsfreie Schranken f:r Quadraturfehler.
Numerische Mathematik 5, 1963 , 226-233; 7, 1965 , 232-237.
1231 P. Henrici: Elements of numerical Analysis. J . Wiley Sons, Inc
New York-London-Sydney, 1964.
1241 D. J a g e r m a n : Investigation of Modified Mid-Point Quadrature F o r m u ?
l a , Math. of Comp. 20 1966 , 78-89.
I,
1251 G. Kowallewski : Interpolation und genahreQuadratur. Leipzig

1 9 3 0 , p. 130
1271 I. Marek: Numerische ~ t a b i l i t z td e r P r o z e s s e vom Keloggschen Typus.

Liblice 1967 Apl. Mat. 13, 1968.
[28] J. Milota ; Universal Almost Optimal F o r m u l a e Solutions of Boundary -
Value P r o b l e m s f o r Ordinary Differential Equations. Liblice
1967 , Apl. Mat. 13, 1968.
1291 R. E. Moore: The automatic Analysis and Control of E r r o r in Digital Com
putation. Vol. 1, 61-130 . Proceedings of a s e m i n a r University
of Wisconsin, Madison Octobre 5-7, 1964.
1301 R. E. Moore; Interval Analysis. P r e n t i c e Hall 1966 .
f31] R. E. Moore: P r a c t i c a l Aspect of Interval Computation. Liblice 1967.
Apl. Mat. 13, 1968 .
[321 M. P r a g e r : Numerical Stability of t h e Method of Overrelaxation.

Liblice 1967. Apl. Mat. 13, 1968.
\331 P . P r i k r y l : On Computation of F o u r i e r Coefficients in Strongly Periodic

Spaces. Liblice 1967, Apl. Mat. 13, 1968.
I. Babuska
A. Sard : L i n e a r Approximation. Providence 1963.

K. Segeth : On Universally Optimal Quadrature F o r m u l a e Involving
Values of Derivatives of Integrand. Liblice 1967 , Apl. Mat. 13, 1968.
H. J. Stetter : Numerical Approximation of F o u r i e r - T r a n s f o r m .

Num. Math. 8 , 1966 , 235 - 249.
J. Taufer: On Factorization Method. Apl. Mat. 11, 1966,427-452 .
I
J. Taufer : Faktorisierungsmethode f u r ein Randwertproblem e i n e s l i -
nearen Systems von Differentialgleichungen, Liblice 1967 . Apl.
Mat. 13, 1968.
I
J. Taufer: Faktorisierungsmethode fur ein Eigenwertproblem eines

l i n e a r e n Systems von Differentialgleichungep , Liblice 1967
E. VitBsek: Numerical. Stability in Solution of ordinary Differential
Equations of Higher O r d e r , Liblice 1967, Apl. Mat. 13, 1968.
J. H. Wilkinson : Rounding e r r o r s in algebraic p r o c e s s e s . London
H. M. S. 0. 1963
J. H. Wilkinson : A. Survey of E r r o r s Analysis of Matrix Algorithms.
Liblice 1967, Apl. Mat. 13, 1968.
1
R. Zezula : Numerische ~ t a b i l i t L t eines Algorithmus z u r Berechnung d e s
E i g e n p a r a m e t e r s e i n e s Matrizenoperator mit Hilfe d e r Reduktions-
methode und d e r Banachschen Iterationene. Liblice 1967, Apl. Mat.
13, 1968.
[45] R. Zurmuhl : Matrizen und i h r e technische Anwendungen. Berlin 1964.

(C. I. M. E. )
J. H. BRAMBLE
ERROR ESTIMATES IN ELLIPTIC BOUNDARY VALUE PROBLEMS
C o r s o t e n u t o a d I s p r a d a l 3 - 11 L u g l i o 1967
INTRODUCTION
Error Estimates in Elliptic Boundary Value Problems
J. H. Bramble (University of Maryland)
In these lectures, I will discuss some methods of obtaining error
estimates for finite difference approximations to solutions of elliptic
differential boundary value problems. Because of the limited time, I shall
restrict my attention to the Dirichlet problem, although some of the
methods are easily carried over to other boundary conditions. The first
part will be devoted to second order problems. In fact, in order to
illustrate the methods, I will restrict my attention to the Dirichlet
problem for ~oisson'sequation, with "zero boundary values" and the
classical Dirichlet problem for Laplace's equation. The last part will
be devoted to some results qn higher order elliptic equations. In most
cases, I will not give more than a sketch of the proof, indicating the
details. Instead of choosing to discuss a general class of operators
and a corresponding general class of difference schemes, I shall tonsider
a specific operator and certain specific difference schemes so as not to
obscure the essential points of the method of analysis.
I will not, during these talks, make extensive references to related
work but shall include in the bibliography a number of closely related
papers. I shall restrict the references to the specific results under
discussion.
I choose to formulate the first problem in a weak form. Let R be
a bounded open set in E with boundary aR and let A be the Laplace
n
operator
We need the following class of functions V defined on R :
where C~ is the class of infinitely differentiable functions with support

0
contained in R and 'H is the Hilbert space obtained by completing cm

0 0
with respect to the norm
All functions for the present will be assumed to be real valued. We now
state problem I.
Problem I: Given F E (L,)' , find u E Lp , 1 6 p < ~ _ 2 such that
R
I uA$= <$,F> ,\d$ E V .
Here (L,)' is the space of continuous linear functionals defined on L,
The quantity <4,F> is the value of the functional F at the point 4 .

For example, if "F c L1" then <Q , F> = F .
The space L is the
P
completion of the cm functions in R with respect to the norm
We will discuss this problem later but let me first remark that when F
and aR are sufficiently smooth Problem I is just the classical problem
The second problem is the classical Dirichlet problem
Problem 11: Given f E c'(~R) , find u E c'(R) such that Au = 0 in R .

We will refer to well known results (eg. on regularity) for this problem
when needed. We can immediately state the following theorems.
Theorem 1: There exists a unique solution to problem I.
Theorem 2: If aR is such that at each point of aR there exists
a barrier then problem I1 has a unique solution.
Theorem 2 is classical. As we will see Theorem 1 (existence part)
can be proved by means of a difference method.
Let us now formulate corresponding difference problems and investigate
their properties.
Let ENh be the set of mesh points in E

N
, i.e., points of the
form (ilh, ', iNh) , for h > 0 and il, . ., iN integers.
We make some definitions.
a) Kh = R r) ENh
( ei is the vector with 1 in the i s position and 0 in the others )

N
We can immeidately state the following.
Lemmal: Suppose - A V 3 0 in R h , V 2 0 in EN,,

h
- R,, . Then V 5 0
in E
Nh . From this it immediately follows that the discrete problem
Ah V(x) = F(x) , x E R,,
has always one and only one solution for any given F and' g . Thus we
can introduce the discrete Green's function Gh(x,y) defined as

We can now make some statements about G .
Lemma2: Gh(x,y)30 .
Lemma 3: F or any V defined on \
This follows from uniqueness in the discrete problem,
We next introduce the following function. Let
L
Then we can prove
Lemma 4: For suitably chosen yN , a, and do (independent of h )
Gh(x,y) .' V(x-y) , Y E Rh .

The proof of lemma 4 consists i n showing t h a t it i s possible t o choose
yN ,a and d
0
i n such a way that f o r y E
-N
A ~(x-y)=h , x = y
h,x
A
n>x
V(x-y)aO , X # Y
V(x-y) 0 all x .
Then i t follows from lemma 1 'applied t o V(x-y) - Gh(x,y) t h a t lemma 4 is
true From lemma 4 we can prove
Lemma 5: Let 1$ p < ~ _ 2

N
. Then there i s a constant C
P
independent
of h and x such that
After using lemma 4 the sum i s estimated by comparing the sum with corresponding
analogou~i n t e g r a l s using the f a c t that x - y is subharmonic, 6 1 N-2,
X # Yo . For the case N 2 3 , for example, we can obtain the estimate
where S i s a s u f f i c i e n t l y l a r g e sphere containing R and having center
a t an a r b i t r a r y point y
o
E R . The integral i s convergent i f p < -N
N-2 '
By taking V
Lemma 6:
- 1 in lema 3 we obtain
Gh(x,y) = 1 •
YE^\
Now it is a simple matter to give a convergence theorem. First we
define the discrete problem.
Problem 111:
A u (x) 0
h h
=
xERh
Of course u,, exists and is unique. Now we have
Theorem 3: Let u be the solution of problem I1 and suppose that
u r c2m . Then u,, f u uniformly as h + 0 . We apply lemma 3 to
Hence by lemmas 5 and 6,
Clearly the right hand side tends to zero as h + 0 .

Immediately from (7) we can deduce
Theorem 4 : If u E c*'~(K) 0 < a ,< 1 , then
a
The proof is obvious. Since u is harmonic, inu ul 6 Ch in and
lu(y) - u r n ( E Ch .
Now no matter how smooth u is in the closure of R the best possible
result is with a = 1 . The last term on the right of (7) prevents us
from obtaining a higher order estimate. As we will see, however, theorems
3 and 4 can both be improved in the sense that if we place some restrictions
on aR then we can obtain similar theorems with u less regular up to
the boundary.
Thesrem 5: Suppose aR is such that u can be approximated uniformly
by a sequence of functions each of which is harmonic in R

- . Then i -t u
h
uniformly as h + 0 . ( Gh is the extension of uh as a constant on each
cube Ch(x) = IY [xi - h/2 < yi ,< xi + h/2 , i = 1 , ... , N} )
Remark. Conditions on aR have been given by several authors,(c.f.
Brelot [8] and Walsa [ll] such that every function continuous
on aR can be approximated uniformly by functions harmonic in K .

This is not necessary for our theorem as we can see by the following example.
Let R be star-shaped with respect to some point which, without loss, we
call the origin. Then if u exists, it can be uniformly approximated by a

-
sequence each member of which is harmonic in R . For, define
where pn f 1 as n + 0 . Now Au(pnx) = 0 in if . Clearly
since u r c09;) ,cn -+ u uniformly in K . Thus starehapedness is a
sufficient condition for the convergence of uh to u whenever u exists.
The proof of theorem 5 is immediate. We have only to estimate
u, - un for h small and n large. But from lennna 3
which we can clearly make as small as we wish by first taking n large
and then h small. We also can prove the following.
Theorem 6: Let aR E cL (piecewise, with no reentrant cusps if
N = 2 ). Suppose that u is Hijlder continuous with exponent 0 < X 6 1 .

Then, for every E > 0
sup lu- sh1 L K(E)

R
where K(E) is a constant which depends on E, R and u but not on h .

The proof of this is based on the following two lemmas.
Lemma 7: Let aR E cL and u be harmonic in R and h - B6lder continuous

in 'ii . Let d(x) be the distance from the point x E R to the boundary
(which is well defined in a strip S6 of fixed width 6 ) . Then for
every E > 0 there is a K(E) such that
for x E
Rhfl Ss .
This lemma follows from the mean value theorem for harmonic functions.
One obtains an estimate for the HBlder continuity of the second derivatives
which depends on the distance to the boundary.
The next lema is crucial.
Lemma 8: For every E > 0 there exists K(E) such that, if aR E 'C
The proof of this.lemma is tedeous, long and involved, but the motivation
is the following. We consider formally
where G is the continuous Green's function and we have assumed that d
has been suitably extended to R . Then, formally,
It is not difficult to see using the maximum principle that for aR E C2
$(XI a c dE(x)
Hence the procedure for proving lemma 8 is based on constructing a suitable
comparison function and using the maximum principle, lemma 1.
We shall now turn our attention, for a time to problem I. First I
will briefly sketch a proof of Theorem 1. For uniqueness it suffices to
show that if IVA~= 0 , V ~ E V then v = 0 . But one can show that if

cm
$ E CQ then there exists a 4 E V such that A4 = $ . Uniqueness follows.
The existence follows once the inequality
is established. (One used the Hahn-Banach theorem.)
Since we shall be interested in the convergence properties of related
difference schemes, we shall sketch the proof of (8) by the difference method.
For any integrable function f we define
where C (x) is the cube with center x

h
, side h and sides parallel to
axes, defined previously.
We let 4 be the solution of

h
Then it follows from lemmas 3 and 5 and HBlder's inequality that
where
h
is the extension of Oh as a constant in each cube Ch '
What we would like to show is
Lemma 9: Let 4 E V . Then mh + 4 as h + 0 , weakly in L

P ,
l<pcm .
This can be done by using (9) and the fact that Lp ,1 < p <
is reflexive and hence bounded sets are weakly compact. We then choose
a sequence {hn] , hn + 0 as n + such that mh

+$EL
n P '
weakly as n + . We then show that $ = 4 and that every sequence
converges to 4 .
Clearly (8) follows from lemma 9 and Theorem 1 is then proved.
We next want to pose a discrete problem analogous to problem I .

Problem I1 :
where \(x) is the function of y for fixed x defined by

Note that i f F E L1 then
It i s now rather easy t o prove
Theorem 7: Let F E L1 and u and uh be t h e solutions t o problems
I and I1 , respectively. Then ii + u weakly i n L

h P
, 1f p < -
N-2
as .
h+O .
We consider the following i d e n t i t y for a r b i t r a r y @ E V :
Again, using lemma 5 we can show t h a t
so that again we can take a sequence ii + ii (some element of L )

hn P
weakly i n L 3 1 < p < -.
Clearly the l e f t hand s i d e of (10) converges
to
P
ti^$ a s n +
N-2
NOW. 'hn
+ $ weakly i n L
P
for 1 < p < -
and by (8) and (9) $ and @ a r e bounded. Thus it is a simple matter
hn
t o conclude that
Hence ii is the solution of problem I o r ii = u . But every sub sequence
converges t o u so that ii
h
-r u .
In order t o show the strong convergence we introduce a sequence
Fn E c O a ( ~ ) f o r each u and such that

lim
n*
IF-F = o . We also
R
introduce the corresponding functions u and u nh
n
. Now we have
Now the f i r s t and l a s t terms a r e bounded by C where C does IF-F 11

L1
not depend on h . The middle term can be treated rather e a s i l y d i r e c t l y
a
but we can observe t h a t since Fn E C (R)cL2 f o r each n
0
a s a special
case of t h e work of CBa [9] we conclude that
I l ~ ~ ~ -L2
u ~ l 0l
-+ as h + 0 f o r fixed n . But since it i s easy to
see that u
n
E V and hence i s bounded so t h a t i n fact llii
1
n d U n Lp
+ 0
as h + 0 for a l l p < . Thus taking n large and then h small we
have proved.
Theorem 8: Let F E L1 and uh and u be the solutions t o problem I
and I respectively.
1
Then iih + u strongly i n L , 1 I p <
P N-2
as h + ~ . -
The next two theorems give information when F is not necessarily i n L
1 '
Theorem 9: Let F E (L,)' and l e t C be the c l a s s of piecewise
P
continuous functions i n R . Suppose t h a t F , when considered a s a
functional on C
P
n
La has support Q contained i n R (i.e. 0 i s
closed and i f 9 E C nLa and 4 = 0 on Q then <$ ,F> = 0 ). Then

P
iih + u weakly i n L , 1 $ p < - a s h + 0
P N-2
.
We e a s i l y obtain iih A9 = <qh,F> and llchll 6 C Il~ll .
L
P
(Loo) '
1t then must be shown that <qh,F> + <$,F> , as h -+ 0 .
The additional hypothesis is imposed now since F can not necessarily
be approximated (strongly) by nice functions. However by using the results
of Bramble and Hubbard [5] we can show uniform convergence of 6h

to ( on compact subsets of R .
Instead of puting a condition on F we could impose one on aR .
Theorem 10: Let F E (L,)' and aR E cL . Then 6
h
-+ u weakly
in L , 1 $ p <
P N-2
- .
as h - + ~ .
,
If aR E c2 then in fact each ( E V will belong to C1 (K) so
that theorem 6 with X = 1 can be applied to show the uniform convergence
of $h to $ in K . In fact using the estimate of theorem 6 (slightly
modified) we obtain a kind of estimate for the rate of weak convergence.
Theorem 11: Let F E (L,)' and OR E c2 . Then for any J, E

m
Co
where C depends on J, but not on h .

We next observe that if F is the "delta function" when restricted
to V for an arbitrary point x then the solution u of problem I will
be the Green's function with singular point x . That is we have the
existence and uniqueness of G(x,y) such that
(11) ((XI = -J G(X,Y) ~ ( y )dy , v 4 E v .

R
Now it is possible to prove the following.

Lemma 10: Let F E L1 . Then
is the solution to problem I.
This lemma is proved by first showing that G(x,y) is symmetric
and then applying the basic relation (11) enjoyed by G . The only
difficulty arrises in showing that the theorem of Fubini-Tonelli on inter-
change of integration can be applied. This can be done with the aid of
the difference approximations.
Having the representation of lema 10 we can show
Theorem 12: Let F c L

q
, for some q > N/2 . Then ti,
n
+ u
pointwise in R as h + 0 .
Clearly from lemma 3
and hence
Combining this with (12) we have

But from theorem 9 and the continuity of $ in R it follows that for
each fixed x E R ch(xy.)

N
Gfx,.) weakly in Lp , 1 1 p < N-2 '
-+ -
Since F E L , for some q > N/2 the theorem follows,
P
In this same spirit we can prove
Theorem 13: Let F E L

9
for some q > N/2 and aR E c2 .
Then 6 + u uniformly as h -+ 0
h
.
To show this we simply approximate F strongly in L by a sequence
9
F with Fn E c~(R) for each n Then .
so that
where C does not depend on x and $n E V . For large n the second
term is small and by the previous remarks

Onh n' uniformly as h -+ O .
Finally, if F is smooth and aR is smooth we have the analog of
theorem 6.
Theorem 14: Suppose F E coy' and aR E c2 . Then for problem I
we have for every E > 0
where K(E) is a constant independent of h .

This is similar to theorem 6 for 11. Take u such that
1
Aul - F ,
u
1
E "'c , Then set u2
= - U1 . Theorem 14 then follows from
theorems 4 and 6.
In order to obtain rate of convergence estimates which are of higher
order it is clearly necessary to modify the difference scheme near the
boundary. We will for the time being still be considering problem I but
with various assumptions on aR and F .

We shall now redefine Rt, and 3% .
a) Nh(x) is the set of "neighbors" of x with respect to

ah a
c) 3% is the set of points on aR which lie on "mesh lines".
For V defined on aRh we define

where uih is the distance between adjacent points of in the xi
direction 0 < ai <1 .
Now everything said so far remains true with this definition , with
the minor redefinition of (F)h and \(x) regarding problem I1.
Essentially, if one defines F as "zero" outside R then werything
is the same. However we can give now an additional estimate for Gh
which allows us to obtain second order convergence for smooth solutions.
Lemma 11:
This is trivially obtained by taking for V in lemma 3, V = 0 on
3% and V = 1 in R,, . Then
Let us call the analog of problem 11, problem 12. Then we get
immediately
Theoren 15: Let u and uh be th; s~lutionsof I and I respectively

2
and u E (m .
C4 Then
We simply note that
so that puting V = uh -u in lemma 3 we have the result. In fact we
note that even though the approximation locally near the boundary is of
order h the contribution to the error is a term of order h3 . This,
we will see, will lead to higher order approximations.'
Before mentioning other approximations we shall see what can be said
if certain more specific information is known about F . Thus we shall
consider briefly the case in which F is very smooth except at one point
and 3R is also smooth. Again in this case the Green's function technique
is fruitful.
The following theorem gives information about the error in this case.
4
Theorem 16: Let aR be smooth and u E C m-0) where the origin 0
is an arbitrary point of (for convenience always chosen to be the center
of a mesh cube.) Suppose that m +X > 2 -N where m is an integer and
O < X f l and
where k is a multi-index kl = (kl - . , k.J ) , kl, ...,k.J non-negative
integers, lkl = 1 ki and
The fur u and uh the solutions of I and I2 respectively
In this case since we have quite specific knowledge of the behavior
of the solution at the origin we obtain an estimate for the error which
is point dependent. Thus even if the solution u has the form
2-Nt6
~ ( x )= 1x1 t regular function , 6> 0 ,
we get from the theorem that uh + u uniformly on every compact subset
not containing the origin.
Also note that under the assumption (13) if u E ,6 > 0 (m=Z)
then the convergence is second order. This shows clearly that the usual
4
sufficient condition that u E C (K) is far from necessary for 0(h2)
convergence.
h
Note also that when u E C (m=0) we obtain a uniform rate of hh
1-E
or h ,
As might be imagined the theorem is proved by using the representation
lemma 3 and estimating the resulting expressions. The details are long and
technical and are found in [7] but I want to point out the crucial
points. First of all the essential ingredient is the majorant (lemma 4)
for Gh . This tells us that, as might be expected Gh behaves quite
like the fundamental solution for Laplace's equation in the neighborhood
of the singularity. Having this we are led to proving a sum relation
analogous to a well known integral expression
Lemma 12: If -N O , V y ER,, then
The proof is done by showing that the sum is majorized by
C I
R
Ix-yIP Iz-ylQdy
For this formulation, however, we can get a sharper theorem than
either theorem 6 or 15.
Theorem 17: Let aR E C* (piecewise, with no reentrant cusps if

N = 2). Suppose that u is the solution of I1 and uh is the solution
of 1I2 (the analog of I1 for the reformulation).

1
Let u E cPyA(K) .
Then for E > 0
The proof is similar to that of theorem 6. Note that with
u E c2"(F) the theorem 4 would show only a rate of hA and for
cl", A < 1 we would conclude nothing. However, we get second order

-
convergence when the second derivatives are HBlder continuous in R ,
Clearly these methods are not restricted to these particular difference
formulations for the Dirichlet problem for Poisson's equation, One can
treat
a) More general operators (second order).
b) Various boundary conditions.
c) Eigenvalue problems.
d) Various difference approximations.
I shall discuss an example of the last extension, since it brings
out the fact that in the transition from the interior to a curved boundary
one can (in the Dirichlet problem) take approximations which are of the
order of accuracy (locally) worse by a factor of hL and still obtain as a
global error that of the interior.
For this example I choose N = 2 and consider the nine point
approximation
so that Ah is now locally 4th order. Now we take A to be defined by
h
(14) in R,, at points of Rh (say $ ) where only I$, points are
involved in (14). At $ = R,, - % we take a second order approximation
to A . aRh we take as in the second formulation. One can then show
by appropriate modifications of the previous Green's function method that
if u,, is the solution of our new problem I1 (for Au = 0 ) then we have

3
Theorem 18: Let u be the solution of I1 and u E C6 (n . Then if
is the solution of 113

uh
6
But in fact we can lessen considerably the requirement that u E C (R)
and obtain by the methods of theorem 17
Theorem 19: Let aR r c2 and u and u the solution of I1 and 113

h
respectively. Then if u E CP" (T)
So far the approximations mentioned have all possessed a common property,
i.e., that of being "of positive type." This means that if ~~j is the
matrix of coefficients of the linear system then

t h e second condition possibly f a i l i n g near the boundary. In f a c t ,
it is t h i s condition, together with
t h a t makes lemma 1 t r i v i a l .
To show t h a t t h i s is j u s t a convenience I wish t o give another

4
example, Suppose instead of (15) we used the 9-point O(h ) approximation
4 6
It i s possible t o show t h a t the e r r o r is of the order h when u E C @)
and probably a theorem l i k e 19 i s also true. Since the properties (15) and (16)
a r e not possessed by the resulting system the analysis is much more d i f f i c u l t .
It i s i n t e r e s t i n g t o note, however, that a corresponding d i s c r e t e Green's
function w i l l s t i l l be positive, a f a c t which i s no longer completely t r i v i a l .
As i s evident the preceding discussion is i n many respects special f o r
second order equations, since much use was made of the maximum principle o r ,
what is the same, the p o s i t i v i t y of the Green's function. Thus i t appears
t h a t , i n attempting t o t r e a t higher order equations, we should work more with
norms other than the maximum norm.

I would like first to sketch some results of Thornge[lO ]
on higher order equations and difference approximations.
Consider the differential operator
where B and y
n
..
are multi-indices, i.e. B=(B1,. ,BN) , where the B 's
j
1
are non-negative integers, ( 6 ( =
j=l
Bj and similarly for y .
The a
BY
are real constants and
We assume that L is elliptic, i.e. for real 5 =
The Dirichlet problem (111)
has a smooth solution provided F and aR are sufficiently smooth.

Consider approximations of the form
where u = u(Sh) and the C 's are complex numbers defined for all a
5 a
but zero except for a finite number of a's A point ([+a)h will be.
called a neighbor of [h if Ca f 0 .
This time Rh will be defined as those points of RflENh whose
neighbors also lie in R . Define - \= a\ . The characteristic
polynomial of L is defined as the trigonometric polynomial

h
where 0 = (O1,...,ON) , (a,0) = 1 aj 0j . Because of periodicity 0 can
be taken in the set
We say that Lh is consistent with L if at an a&trary point (which we
take as the origin)
Lh uo = L u(0) + o(1) when h + 0 .

( + O(hk) - consistant of order k )
Now it can be shown

Lemma.13: Lh is consistent with L if and only if
p(B) = L(0) + o ( [elzm) when 0 + 0 .
Now we shall denote the set of complex valued mesh functions on
R, by Dh and
The sum will always be finite since all functions considered will vanish out-
side some bounded set.
Define
and
Now we call the difference operator Lh elliptic if p(9) satisfies
p(9) > 0 for 0 # 0 E S .
In particular if Lh is elliptic p(9) is real so that C-a = ra .

With this definition, Thom6e then gives two a priori inequalities
which we state as the next two theorems.
Theorem 20: Let Lh be consistent with L . Then Lh is elliptic
if and only if there is a constant C independent of u and h such that
The main tool in the proof is the "Fourier transform." For this reason
only constant coefficients are treated.
If we define R;1 that part of R,, whose neighbors are in R,, and
< = Rh -4 and also
Then it can be shown

Theorem '21: Let aR be sufficiently smooth and L consistent
h
with L and elliptic. Then
This theorem will show that the difference approximation can be cruder
near the boundary.
Let uhg be the solution of problem 111.
From theorem 20, I11 has one and only one solution. We have the following
1
convergence estimate of ThomCe .
Theorem 22: Let u and u be the solutions of I11 and 1111
h
,
2mtl
respectively. Suppose Lh consistent with L and elliptic and u E C 0.
Then if e
h
= u - uh in Rh and 0 outside Rh we have
As interesting examples ThomCe gives a number of applications of his
theorem to special cases. I want to discuss two of them since in these two
one can obtain an additional inequality which together with theorem 22 shows
that
(A) Let
and take,
..
(We remark that in this case we can also take ulJ to be variable and treat
the self adjoint operator
The matrix aij is assumed symmetric and to satisfy
for real 5 .
In obtaining (17) the reason for the low power of h is that near the
boundary the approximation gave rise to a lower order error term. Thus we want
to try to estimate the lm-1) L2 norm in such a way that the approximations
near the boundary are not so important. For motivation we consider a very
simple way of obtaining an L2 inequality in the continuous problem. We
suppose that aR is smooth and let J, be the smooth function satisfying

Then by the maximum principle ( :: 0 in R . Clearly since u and $
are zero on aR
Thus since ( $ 0 in R
and hence
Now the importance of this estimate is that $ = 0 on aR and hence Lu
is not so influential near the boundary. It is just this type of estimate
that would give us something for the difference problem. Unfortunately (18)
does not hold in the case of the present difference approximation. However
what can be shown in the following. For any V such that V = 0 in

E~ -%
where C does not depend on h . The f a c t t o note is that although U8)
holds pointwise it was only used i n the aean. Now one expects, because of
consistency, that an expression f o r the difference operator, analogous t o
(18) w i l l hold t o within higher erder terms. It turns out that these terms
can be estimated i n the means, hence giving us (20). Thus we conclude
Theorem 23: In the case of example (A) we have
The one thing that must be used here i s an inequality given by Thomge.
That is t h a t
(B) For the second example i f we take (N=2)
and
then we have the f i r s t boundary value problem f o r the biharmonic equation
Thomee's r e s u l t is the f i r s t e r r o r estimate i n t h i s problem for a general
domain. He obtains according t o theorem 22 an estimate of the order of

h1I2 f o r the m = 2 norm. We want t o look a t the m - 1= 1 norm and
obtain an order h estimate f o r t h e e r r o r i n t h i s norm.

Again we are motivated by a pointwise differential inequality.
Miranda, in obtaining a maximum principle,made use of the fact that
(Unfortunately this is special for N = 2 ).
Again we can obtain the analogus expression for the difference operator,
which is
where the subscript x and denote the usual forward and backward divided
differences. Once again we can introduce a function 4 such that
Then it is (almost) clear that the last term
can be estimated in the mean by

Thus we are led to the discrete a priori inequality
This together with theorem 22 leads to the estimate llehll 5 C h .

h,1
Although it is not true that the maximum norm can be estimated by the
Dirichlet integral in the continuous case, we can obtain a meaningful
estimate in the discrete case, N = 2 ,
Lemma14: Let, V = O in E N - R h . Then
max 112
Rn
IvI I clln hl IlvI1 +
h,1
This can be obtained by using the discrete Green's function G for Ah .

From the representation lemma 3 and partial summation we have
and by Schwarz's inequality
But it is easy to see that
Hence we obtain the estimate lehl I clln hl 112 h

Concerning second order approximations i n t h i s problem (B) I have
given one i n Bramble [ 4 ] and ~lkl,

i n 1131 has given one.
To my knowledge no second order approximation has been proved i n general
for (A).
As regards the second order approximation given by Zl6mal f o r problem (B)
I wish to mention t h a t it i s simplier than t h e one given by me and also t h a t
the technique used by Zl6mal holds f o r more general 4& order equations but
for N = 2 . He shows, e s s e n t i a l l y , t h a t a c e r t a i n second order interpolation
near the boundary i s s u f f i c i e n t t o increase the r a t e of convergence from
h1I2 f o r the m = 2 - norm t o h3I2 and i n the case of example (B) from
h t o h2 f o r the m - 1= 1- norm.
BIBLIOGRAPHY
~abuzka,I., Prgger, M., and Vit6s$t,E. Numerical Processes in

Differential Equations. Interscience publishers, New York (1966).
Bramble, J. H. "On the convergence of difference schemes for classical

and weak solutions of the Dirichlet problem." To appear in the pro-
ceedings on Differential Equations and Their Applications 11, Bratislava,
Czechoslovakia (1966).
Bramble, J. H. (editor) Numerical Solution of Partial Differential

Equations. Academic Press, New York (1966) .
Bramble, J. H. "A second order finite difference analog of the first
biharmonic boundary value problem" Numerische Mathematik 9, 236-249
(1966).
Bramble, J. H. and Hubbard, B. E. "Approximation of derivatives by

finite difference methods in elliptic boundary value problems."
Contributions to Differential Equations, Vol. 111, No. 4 (1964).
Bramble, J. H., Hubbard, B. E. "Discretization error in the classical

Dirichlet problem for Laplace's equation by finite difference methods."
Univ. of Md. Tech. Note BN-484 (1967) (to appear, SIAM Series B).
Bramble, J. H., Hubbard, B. E., and Zlgmal, M. "Discrete analogs of

the Dirichlet problem with isolated singularities." Univ. of Md.
Tech. Note BN-475 (1966) (in print).
~rglot,M, "Sur l'approximation et la convergence dans la theorie des

fonctions harmoniques ou holomorphes .I1 Bull. Soc. Math. France 73,
55-70 (1945).
9. Cga, J. "Sur l'approximation des problemes aux limites elljptiques .I1

Compte rendus 254, 1729-1731 (1962) .
10. ~homge,V. "Elliptic difference operators and Dirichlet's problem."
Contributions to Differential Equations, Vol. 111, No. 3 (1964).
11. Walsh, J. L. "The approximation of harmonic functions by harmonic

polynomials and by harmonic rational functions."
12. Zlgmal, M. "Asymptotic error estimates in solving elliptic equations

of the fourth order by the method of finite differences." SIAM
Series B2, 337-344 (1965).
13. Zlhal, M. "Discretization and error estimates for elliptic boundary

value problems of the fourth order." (in print).
CENTRO INTERNAZIONALE MATEMATIC0 ESTIVO
(C. I. NI. E. )
G. CAPRIZ
THE NUMERICAL APPROACH TO HYDRODYNAMIC PROBLEMS

THE NUMERICAL APPROACB TO 1IYDRODYNAMIC PROBLUIS
by
G. Capriz
(Centro Studi Calcolatrici Elettroniche

iiel CIJR presso l'Universit2 d i Fisa,
Pisa, I t a l y )
1. Introduction
I n t e r e s t i n the numerical solution of hydrodynamic problems

has been ? l i v e for a long time; the book of r e f . A, for instance,
bears the date 1322. The reasons of tile i n t e r e s t are obvious: so few
e x p l i c i t solutions of the equations of hydrodynamics a r e known and
g r e a t gaps s t i l l e x i s t i n the knowledge on the q u a l i t a t i v e behaviour
of general solutions [A, 21 . A s r e f . 4 already shows, attempts a t
numerical integration were made before the age of computers: some
references t o t h i s e a r l i e r work can be found i n the textbooks of
Allen [il and Thom-Apelt , among others.
Von ilcumann called a t t e n t i o n repeatedly t o t h i s f i e l d of
research [e. g . - 6 , p. 2361 , ~ u g g e s t i n gt h a t computers would be the
r i g h t t o o l f o r iilquiry. Attention was devoted a t f i r s t t o studies of
conpressible flows [6- A; - 7 , vol. 4 1 ; sometimes through the integration
of reciuced equations of the boundary layer type [ I , vol. 3; 8, and
the papers quoted there] .
Pile h e u r i s t i c i n t e r e s t of tlie chlculations was soon pointed
out and anong thc f i r s t problems tackled were those for which the
c u r i o s i t y of thc experimenter had not y e t been s a t i s f i e d by the
r e s u l t s of the theoretician [9, 103 .
Incompressible flows a r c studied now with g r e a t zest; there is
i n t e r e s t i l l such flows for analysis of motions with a f r e e surface
, 12, g ] , and of motions of natural convection l;, 5 , for
wheathcr analysis and prediction k ] p c r h a p s using a "shallow fluid"
approximation [s,
l8] , for forecasts of flood waves i n r i v e r s b]
and many other questions.
G. Capriz
Perhaps the moot ambitious goal is pursued by tliose r e s e a r ~ h

workers who t r y t o decide, by a thorough numerical study, as t o what
extent the Navier-Stokes equations ( i n a f i n i t e difference form) a r e
able t o describe phenomena of fluid flow i n s t a b i l i t y and even of
t r a n s i t i o n t o turbulence. Interesting r e s u l t s havo been already
obtained i n the description of the formation of Karman vortices
behind an obstacle E l4 , of the spike and bubble i n the Rayleigh-
-Taylor form of i n s t a b i l i t y f o r superposed f l u i d s E l-J 21 of the
Taylor vortices a t high lieynolds number i n the Couette flow E l 231 .
The calculations a r e so precise t h a t they can be used t o deduce
values of the functionals of flow (suc;i a s heat transfer coefficients
and viscous drag c o e f f i c i e n t s ) much nore s a t i s f a c t o r i l y than by other
approximate means. The wealth of r e s u l t s obtained by IIarlow and
collaborators a t the Los Alamos Laboratory a r e so spectacular t h a t
they have found ,space i n S c i e n t i f i c A r i ~ e r i c a n u , Science ,
Datamation , Sciences k].
Attempts have been made t o follow, i n a f l u i d flow, the pro-
duction of s c ~ a l le d a e s from larger ones i n three dimensions but the
work was hampered by the occurrence of numerical i n s t a b i l i t y
, .
292 Silrilar and otner d i f f i c u l t i e s have limited the range of
r e s u l t s obtained i n studies on the development of perturbations i n
laminar plane Poiseuille flow @, 21, 2?] . Nore d e t a i l e d are the
conclusions of another analysis of t r a n s i t i o n from laminar t o
turbulent flow (for a flow over a f l a t p l a t e ) ; proposals have
a l s o been made for a d i r b c t numerical study of turbulent flows.
G. Capriz
2 . F i n i t e d i f f e r e n c e approximations f o r t h e Uavier-Stokes equations
I n almost a l l t h e r e s e a r c i ~ e squoted i n Section 1 the authors

have r e s o r t e d t o f i n i t e d i f f e r e n c e analogues of t h e Navier-Stokes
equations. So we introduce now those equations r e s t r i c t i n g our
consideration t o t h e incompressible case
& = - (grad Y ) . 'I&-

grad Y t 9 ,
here % i s speed, $ i s t h e r a t i o of pressure over (constant) d e n s i t y ,

3 kinematic v i s c o s i t y and g applied f o r c e per u n i t mass.
To eqns ( l ) , (2) t h e appropriate boundary conditions must
be added, perhaps on unknown boundaries (flows with a f r e e s u r f a c e ) .
For t h e purposes of a numerical study, d i s c r e t c equivalents
t o ( I ) , (2) and t h e boundary conditions can be used, which a r e based
on a n e t of p o i n t s where t h e r e l e v a n t q u a n t i t i e s must be determined.
The d i s c r e t c e q u i v a l e n t must have a form which suggests
f e a s i b l e numerical alqorithms ; they must be s u f f i c i e n t l y accurate
without leading t o cul. ,crsome computations and not be s u b j e c t t o
numerical i n s t a b i l i t y . Although t h e requirements a r e numerous and
s t r i n g e n t t h e r e i s a v a r i e t y of procedures t h a t meets them; t h e choice
depends on a n o t w e l l defined c r i t e r i o n of economy.
To o b t a i n convenient numerical algorithms t h e d i f f e r e n t i a l
eqns ( I ) , (2) a r e not t h e b e s t s t a r t i n g p o i n t s , f o r a number of
reasons; f i r s t of a l l one must t r y t o s e p a r a t e t h e unknowns% and
'S
I f t h e boundary conditions do not involve Y , t h i s unknown can be
eliminated a l t o g e t h e r from (1) using (2). I n f a c t , t h i s equation
s t a t e s t h a t 3 i s solenoidal; hence it can be expressed a s t h e c u r l
of a vector p o t e n t i a l y which i s i t s e l f s o l e n o i d a l ; a t t h e same
t i m e y c a n be eliminated from (1) by taking t h e c u r l of both
G. Capriz
sides :
X = c u r l 4- ,
$=curl 3 ,
%=curl y ,
at + c u r l (grad 2 , ) = - V curl X t curl

-g ; (4
here account was taken of the identity: c u r l c u r l _v = grad div 1 2.-

Thus, using eqns ( 3 ) ( 4 ) the conservation of mass is exactly verified.
I f one wants t o work i n terms of the variables 5 and 9 d i r e c t l y , one
can s u b s t i t u t e ( 2 ) with a consequence of ( l ) ,i n whose derivation ( 2 )
plays a r61e. Here d i f f i c u l t i e s are net because the very important
eqn ( 2 ) would thus intervene only i n d i r e c t l y ; i n practice one finds
t h a t g r e a t care must be taken i n the computation i f the approximate
values of %have t o correspond t o values of D which a r e s u f f i c i e n t l y
small t o be accepted. In theory one could r e l y on the following
consequence of (1)
AJ= d i v -g - div (grad 3 2.) . (5)
In practice one finds t h a t the use of d i s c r e t e equivalents of ( I ) , (5)

leads t o rapid accumulation of e r r o r s and t o large values of D, a t
l e a s t where the d i s c r e t i z a t i o n i s based on a r e l a t i v e l y coarse net.
I t i s more convenient to s u b s t i t u t e (1) with the equation
e t d i v
at (y @ y ) = - g r a d p - Y curl curl y g
t -- (6)
and (5) by t h i s consequence of ( 6 ) :

G. Capriz
This equation implics t h a t D ( t + A t ) vanishes though t h e " s t a r t i n g

value" D ( t ) may be d i f f e r e n t from zero; by such a device e r r o r s
introduced a t one s t a g e tend t o be reduced i n the next ( k ) .
The form given t o t h e non-linear term i n ( 6 ) is more convenient
than t h e form of t h e corresponding term i n (1) f o r our purposes.
I n f a c t , onc aims a t trasforming t h e d i f f e r e n t i a l equation i n t o a
d i f f e r e n c e equation (spacewise)through t h e following s t e p s :
i ) i n t e g r a t e over a mesh-elenent V , transforming a l l volume i n t e g r a l s
containing space d e r i v a t i v e s i n t o s u r f a c e i n t e g r a l s over t h e
boundary S of V.
i i ) approximate surface i n t e g r a l s using only t h e values of t h e
functions a t t h e meshpoints.
For s t e p (1) eqn ( 6 ) is d i r e c t l y f i t so, of course, a r e eqns (3),
( l ) ] ; 2rccisoi.y irc get (n, u n i t vector of t h e e x t c r i o r normal)
S i m i l a r l y fro^. ( 7 ) it follows
---
(*) The "penalty method" or the "mctilod of a r t i f i c i a l d e r i v a t i v e s "
described by Professor Lions could a l s o have been used.
G. Capriz
We quote here a l s o t h e i n t e g r a t e d v e r s i o n s of eqns ( 3 ) , ( 4 )
& dd v t Js ( 3 - grad &1x 2 dS =
tiumerical quadratures must now be introduced t o approximate the

i n t e g r a l s i n ( 8 ) , ( 9 ) o r ( l o ) , (11). For t h e sake of s i m p l i c i t y we
consider h e r e only t h e case of a r e g u l a r cubic mesh. I t i s easy t o
r e a l i z e then, (although we n o t e n t e r h e r e i n t o d e t a i l s ) t h a t , f o r
the simplest and r e l a t i v e l y most p r e c i s e approximation of eqns
(S), ... ( l l ) ,one n u s t introduce a cubic l a t t i c e with t h e following
condition: I f f o r i n s t a n c e i s supposed t o be known on one l a t t i c e
p o i n t Po , then %must be known on t h e s i x n e a r e s t p o i n t s P1... 6'
and convcrscly.
A s a vcry s i n ~ p l eexample, consider t h e second eqn (10) : we
have (h, mcsh-size)
6
Then, t h e s t r u c t u r e of t h e system (8) , ( 9 ) and (lo), (11) i s such

t h a t , t o achieve b e s t approximations, it i s convenient t o take V
successively as coinciZcnt with different b u t overlapping cubic
G. Capriz
c e l l s . For instance, with referencc t o eqnc (101, (11) notice t h a t

must be known i n the centres of the faccs of the f i r s t c e l l V1
Y
-
whereasumust be known a t the centre of t!le c e l l i t s e l f . The c e l l s
of type V2 must be such t h a t + i s known a t the ccntres of the faces,
whereas f i s known a t the centre of the c e l l i t s e l f , and so on.
The procedure can thus be arganizcd so t h a t no interpolation
i s required except f o r the approxinate cxpression of the non-linear
terns. We must introduce now a d i s c r e t i z a t i o n i n the :ime variable.
Leaving without a superscript the values a t the end \ of the k-th
time s t e p and using the superscript k t i for values a t the end tktl
of the (kt1)-th step, tile simplest f i n i t e difference approximations
t o (8), (9) are
6(,1 -n . grad rg t [div @ -s] dS +
These approximations a r c very rough, but have the g r e a t advantage of

leading t o exnLicit formulae. 'Ilie care taken i n writing the condition
which implies conservation of mass and a l s o the special form given t o
the term measuring the diffusion of momentum i s j u s t i f i e d now: it
allows the acceptance of the rough formulae above.
Leaving aside f o r the moment the question of the boundary conditions,
the process t o follow is t h i s .
Assume t h a t the i n i t i a l values of 3 be given a t t=O. The corrcsponding
d i s t r i b u t i o n of [P i s determined through eqn (13). This s t e p can be
G, Capriz
acconlplished through one of the many methods available for the

integration of Laplace equation, for instance through an i t e r a t i v e
.
overrelaxation procedure 121 Successively, the right-hand side
of eqn (12) is computed and new values of 2 are determined. The
process is then repeated.
Attention was confined so f a r t o time dependent flows. There
i s i n t e r e s t of course a l s o i n the study of steady flows; f o r such a
study some of the remarks s t i l l apply. The d i f f i c u l t i e s i n respect
t o diffusion of d i l a t a t i o n cio not occur; we find instead problems of
convergence i n the schcnes of successive approximation t h a t must be
introCuceci t o deal with tile non-linear terms.
-
3. Boundary conditions
ijcundary concitions for the approximate analysis of our

problems must not be l i g h t l y stated. For instance: i s the usual
condition of no s l i p a t a wall always justifiecl?
Only a reference t o physical circumstances allows one t o
give a s a t i s f a c t o r y answer t o t h i s question. Whether or not slippage
i s t o be alloweci depends upon the thic1:ncss of the boundary layer
t h a t one would expect t o develop i n the t r u e fluid. I f t k i s is much
l e s s than the ~ l i ~ e n s i o nofs a l a t t i c e c e l l and one i s not interested
i n the u e t a i l of tile bounaary flow t!len a f r e e s l i p condition i s
appropriate; i f the boundary layer i s much larger than one c c l l , then
a no-slip condition i s required. For i n t e r r ~ c u i a t ecases, the proper
condition t o use depends upon the exact circumstances, and i n some
cases i t i s appropriate t o t r y both ways and compare the r e s u l t s .
Another point one must emphasize: sometimes it is convenient
for computational purposes t o introduce f i c t i t i a s mesh-points ogt-
side the boundary. I f such a device i s used, one must be sure t h a t
the f i n i t e difference approximation t o D vanishes a l s o a t the exterior
G. Capriz
f i c t i t i o u s . c e l l s so t h a t no diffusion o£ D inside the boundary

occurs.
A l l these warnings a r e of course of an experimental
character and a r c connected not t o any inadequacy i n principle of
the f i n i t e difference approximations, but rather t o the need t o
operate with a r e l a t i v e l y small number of c e l l s .
Even more d e l i c a t e is the question of writing adequate
approximations t o conditions a t a free surface 1 .
Over 1 conditions on s t r e s s components must be s t a t e d ; for
instance i f the externally applied s t r e s s is a pressure Ya we
should use, i n a system of cartesian coordinates, the conditions
-
where N i s the exterior normal t o -
and T ( I ), T ( * ) a r e two
orthogonal tangential vectors. These conditions arc very d i f f i c u l t
t o s e t up s a t i s f a c t o r i l y on a computer and workers i n the f i e l d
have resorted t o conditions such as D=O, 49 = (f t o balance equations
a
and unknowns. The f i r s t choice i s j u s t i f i e d on the grounds t h a t the
gravest source of e r r o r s i s diffusion of d i l a t a t i o n D throuqh the
boundaries. The second choice i s notivatcd by the renark t h a t often
viscous e f f e c t s a r e small when compared with a d i r c c t l b imposed
s t r e s s . On the other hand the l o c a l orientation of the surface can
be usually determined only very roughly, so t h a t a more precise use
of eqns ( 1 4 ) i s not j u s t i f i e d .
I t remains t o follow the changes of the free surfacc w i t h
time. This is accomplished by introducing marker p a r t i c l e s on the
f r e e surface (actually i n the marker-and-cell method the rnarker
p a r t i c l e s a r e distributed throughout the f l u i d , though, f o r analytical
G. Capriz
purposes, they are e s s e n t i a l only a t the boundary).

The speed of the p a r t i c l e s is determined by interpolation or
extrapolation from nearest mesh points; f i n a l l y t h e i r movement
i s followed s t e p by step.
A p r a c t i c a l procedure is a s follows.
One builds up i n the computer a picture of the f l u i d s e t
i n a wider f i e l d of c e l l s where the f r e e surface can impinge. There
are markers t o show which c e l l s are occupied ( a t l e a s t i n p a r t )
and which a r e free. Pressure and velocity f i e l d s a r e determined over
a l l the occupied c e l l s , boundary conditions intervening i n the
boundary c e l l s . To avoid ambiguities ( i . e . a wrong labeling of
i n t e r n a l c e l l s a s empty) a t l e a s t four marker p a r t i c l e s for c e l l
are d i s t r i b u t e d a t time t = O i n a l l occupied c e l l s with further
provisions f o r exceptional cases.
G, Capriz
4. 14umerical i n s t a b i l i t y ; accuracy
Phenomena of numerical i n s t a b i l i t y have been mentioned

already; i t is well known t h a t t h e i r onset depends c r i t i c a l l y on
meshsize, and time-step size. It i s a l s o comon experience t h a t
e x p l i c i t algorithms such a s t h a t embodied i n formulae (12), (13)
a r e usually nore prone t o the disease than more complex implicit
algorithms.
I f we reduce eqns (12), (13) t o a non-dimensional form
by introducing a typical velocity U a typical body force per
u n i t mass G , thc time s t e p 5 and the meshsize h (assuming
for simplicity t h a t the l a t t i c e is cubic though phenometia of in-
s t a b i l i t y may be yucnchcd sometimes by introducing meshes with
appropriatc side-ratio 1311 ) we see t h a t the solution of the
i i n i t c difference equations Jepends locally on the following
parameters
For those who a r e physically inclined we remark t h a t d can be

1
construed as a Strouhal n u d e r of the flow based on the numerical
time s t e p and meshsize. ~ i r c i l a r l ? d and d can be conlbined
with d l t o express nufi!erical Froude and Reynolds nunbers
Conditioils of numerical s t a b i l i t y can then be expressed throug!~

limitatiorls on rl .
, Fll , Ri! The choice of the ralcvant values
of Nand C will depend on the problem i n hand, of course.
In the study of flows with a free surface U ' can be taken bs the
G. Capriz
speed of surface waves: using shallow f l u i d theory

1
Z
%= ( 'i;t a n h k H ) ,
k , wave number; H, depth of f l u i d .

I n t h e experiments of ilarlow and c o l l a b o r a t o r s t y p i c a l
s t a b i l i t y conditions were found t o be, experimentally, [g, p. 28)
I n o t h e r cases t h e l o c a l v e l o c i t y intervenes [i, p. 1371

i n s t u d i e s on t h e behaviour of a p e r t u r b a t i o n i n a steady flow
t h e excess speed due t o t h e p e r t u r b a t i o n seems t o have relevance.
I n a l l cases it was found by experiment o r was suggestcd by heuri-
s t i c arguments t h a t RN must be of ordcr of unity i f i n s t a b i l i t y
has t o be avoided. Although t h e value o f u t h a t must be used i s
n o t known e x a c t l y i n advance, rough evaluations a r e u s u a l l y
possible. Then t h e condition j u s t mentioned can give an i d e a of
t h e s i z e of t h e problem i n hand from a computational p o i n t of view.
I f conditions such a s (15) a r e s a t i s f i e d t h e r e s u l t s of a
computation a r e l i k e l y t o look reasonable, i . e . not wildly wrong,
but they may s t i l l be f a r from accurate. I t would be nice t o have
some t e s t s f o r accuracy. A check on tile value of D must always
be kept with automatic s t o p when D reaches an unacceptaL-lc l e v e l .
I f t h e condition of incompressibility i s s a t i s f a c t o r i l y appro::inated
t h e measure of t h e domain occupied by t h e f l u i d ( a s shown by nar1,;er
p a r t i c l e s ) must be constant. I n t h e marker-and-cell method a chccl;
can t h e r e f o r e be nade by comparing t h e a w ~ b e rof cell:; Ij sontaining
a t l e a s t one p a r t i c l e with t h e n u ~ b e rof boundary c e l l s (since these
c e l l s a r e constantly c o n t r o l l e d i n a program, t h e check i s simple).
G: Capriz
The experimental value of 5 can be compared with t h e o r e t i c a l

estimates
One such estimate -
pq
f o r plane problems is
where A i s the constant area of a cross-section of the region

occupied by the f l u i d , P is the length of the boundary of the cross-
section and the r a t i o of p a r t i c l e spacing t o c e l l size.
Further checks a r e sometimes made on the basis of evaluations
of t ~ t a ki i n e t i c energy.
3. rlumerical analysis of hydrodynamic s t a b i l i t y of steady flows_,
I mentioned already t h a t a good deal of research e f f o r t i s

applied t o the numerical study of s t a b i l i t y of c e r t a i n c l a s s i c a l
flows: the Poiseuille flow, the Couette flow, the flow over a f l a t
p l a t e , etc.
In these cases the boundaries are fixed and one can make use
conveniently of eqn. ( 3 ) , ( 4 ) ; the time-independent functions
describing a fundamental flow a r e supposed t o be known:
X- - -X0 ' - 4; 4- do,

w V
and d e t a i l s on the behaviour of perturbations X s - -

N
2 :3 - a r e required.
The equations a r e :
A) &
-9
N
-
N V
- = curl , - = curl , ~ = c u r l y r ,
q=
4.
at t curl [ g a d -
N
'U 7 t o t 5) t grad U.- - -9 curl -
X
G. Capriz
&
Eecause t h e f i e l d of can be taken t o be solenoidal, t h i s

equation can be w r i t t e n i n t e r n of - only
-
aaY - c u r l
a t
Lgrad c u r l
- *( 3+ curl
ry
- ) t grad t e e c u r l y]=
N
I t i s convenient t o w r i t e inunediately t h i s equation a l s o

i n a non-dimensional form using a t y p i c a l v e l o c i t y U and a t y p i c a l
dimension of the domain L , introducing t h e notation
and a physical Reynolds n u d e r
and presuming now t h a t t h e operators A and c u r l a c t over

non-dinensional space v a r i a b l e s
-h A I*-
az
R curl [grad curl P .(? + c u r l I*)-
-
9
t grad s - curl r']
-
=
Usually one wants t o know t h e s o l u t i o n of eqn (16) f o r a s u f f i c i e n t l y

ample i n t e r v a l of time and over a domain f o r the space v a r i a b l e s
$
which i s not bounded, though sometimes t h e expected phenomenon is

periodic i n one o r more space v a r i a b l e s and a reduction t o a bounded
domain ensues.
On the p a r t of t h e boundary t h a t represents walls
N
(fixed or i n steady notion) ??. = 0 ; o f t e n one can conclude from

t h i s t h a t a l l components of \y and t h e i r normal d e r i v a t i v e s vanish.
There may be conditions a t i n f i n i t y and, on o t h e r p a r t s of
G. Capriz
the boundary, periodicity conditions may apply; besides the i n i t i a l

conditions must be known.
When one is interested i n the decay of an istantaneaus
disturbance or i n the s p a d i n g of s e l f amplified perturbation,
these a r e the only conditions t h a t apply. In other cases pertur-
bations may be Continuously fed from outside; then y and -
derivatives a r e assigned on portions of the boundary as known
functions of time .
-
Because the choice ?' k' = 0 corresponds t o the fundamental
solutions of eqn (16) the i n t e r e s t centres a t f i r s t on the small
perturbations. Although a precise statement can be made only i n
one special case, i t is generally presuned t h a t the behaviour of
a perturbation of small amplitude can be q u a l i t a t i v e l y decided on
the basis of the linearized equation
' r I*) -
3
R c u r l [grad c u r l 1.
- 5 + grad? curl I*]=
We come thus t o a rather complex linear diffusion problem; i n the
plane case, where Y * has only one non-vanishing component eqn
( 3 ) has been the object of many c l a s s i c a l studies, f o r instance
those r e l a t e d t o the s t a b i l i t y of Poiseuille flo~v,o r the flow on
a f l a t plate.
Because the c o e f f i c i e n t s of eqn ( 1 7 ) a r e independent of
time, the solutions can be written as l i n e a r combinations of
functions of the type
where A i s a function of the space variables only and 1; i s a

complex constant. The equation t h a t follows f o r _A , from (17),
G. Capriz
kA& -R c u r l [grad c u r l & . 5 t grad 5 . c u r l ;] =
and the associated boundary conditions add t o an eigenvalue

problem depending on the positive parameter Rc . I t i s essential
t o decide which is the infimum Rc of the s e t of values of R
f o r which one eigenvalue K(R) a t l e a s t has a positive r e a l part.
I n t h i s f i e l d the early work of Thomas must be quoted [ g ] .
I t happens sometimes t h a t the value of k corresponding
t o Rc vanishes; t h i s analytical f a c t i s r e l a t e d t o the physical
existence of non-trivial steady flows. I n such cases the eigenvaluc
problem ( l C ) i s further simplified. Cesides, a search for non-trivial
solutions of the non l i n e a r problen:
*
- c u r l [grad c u r l Y * ( +~ c u r l y*) +
Ah
- + grad
J = R
2 . c u r l \Y
J
e
(19)
*
with the associated boundary conditions, can be attempted.
For a special case of t h i s problem we have d e f i n i t e r e s u l t s due
n
t o Velte, Kirchgassner and others research vorkcrs a t Freiburg
- .
c36, 37 3 8 1 The special case i s examiiled i n some d e t a i l l a t e r .
Mention nust be b r i e f l y nade here of 'the numerical techniques used
t o tuckle eqns (16) , (2); (g), (2)with the associated boundary
conditions.
A t r i v i a l e x p l i c i t method can be used i n connection with
eqn ( 1 6 ) , ( 1 7 ) ; but more often, t o lessen phenomena of numerical
i n s t a b i l i t y , it is more convenient t o evaluate the term under the
biharmonic operator as the average of tile values a t t i m e s t a n d
'2 t h Z , mantaining f o r the other terms the evalutation a t time
I f such technique is adopted a matrix representing the

G. Capriz
d i s c r e t e equivalent of a l i n e a r combination of the operators A A

and 4 must be inverted. Even when use i s made of an e x p l i c i t
method a matrix inversion (although simpler) i s required. Techniques
of d i r e c t inversion or i t e r a t i v e methods must t e called for.
Direct inversion though cumbersone may bc a t t r a c t i v e because
it i s needed only once for a l l time steps. The economy of the
procedure i s much enhanced i n cases where the solution i s periodic
i n one or more space variables because tile matrices involved are
then c i r c u l a n t i n submatrices which may even be c i r c u l a n t i n t h e i r
turn. Formulae f o r the inversion of c i r c u l a n t or block c i r c u l a n t
matrices a r e quoted i n the next section [39, 403.
For the solution of the problem ( 1 9 ) with the associated
boundary conditions an i t e r a t i v e procedure is always called f o r ,
t o deal with the non-linear terms. Starting with a reasonable guess,
one can make use of tlie d i s c r e t e equivalent of the i t e r a t i o n
A A Y * ( ~ )-= R
- curl ' (k-l). (5+ c u r l y (k-11
-
curl y s
t grad 2
-
Here again, i f the boundary conditions express periodicity a t l e a s t
i n one variable, techniques of inversion of c i r c u l a n t matrices may be
of use. Both i n the analysis of the time-dependent case and during
the i t e r a t i o n (20) phenomena of numerical i n s t a b i l i t y may occur.
A word of warning is liecessary here; a mild form of numerical i n s t a b i l i t y
i n diffusion problems may be wrongly taken sometimes as indicative of
hydrodynamic i n s t a b i l i t y . The study of the same problem with two d i f -
f e r e n t meshsizes (one rectangular and one square for instance, i n the
plane case) is recommended. "Nmerical" eddies change then wavelength
so as t o cover the same number of c e l l s (the typical wavelength of
"numerical" eddies i s ten c e l l s ) .
We have mentioned tliat i n t:ic n m e r i c a l solution of our
problems under pericdicity conditions bloc]:-circulant natriccr; appcar.
To show t h a t sirnple deviccs can savc a t t i n e s a l o t of k~orl;, thc
property of these matrices i s rccallcd hcrc, t h a t a l l o ~ ~ans cacicr
inversion.
Let ii0 f&-]- .
. . . ?-.-..,
I ...nn-1 ) =
-*io
n = no , nl , .
L J , , ~
. ... .
L
nl
*.. .Ao
A '.
be a block-circulant matrix, wilcrc the Ai arc blocl;s of ordcr n.
Let I be the i d e n t i t y matrix of ordcr n and $,
the m-th roots of unity and put
bl , .bn-l . ..
I. ..
v =
Then
I :.-I . . a, I
lo
1.
v-l = - . ... .
;
El & "-1
1 m-lI. . *bm-l1
and
#(do) 0 ... 0
f
Y'AV=A =
0
:. : .
.,
#( b-m-l
C. Capriz
where
I£ the matrices (4 ) a r e not singular a l s o A is son singular

and
A-l= "A -1 .
% A-' is a l s o block-circulant; precisely
Therefore the inversion of the matrix A of ordcr man is reduced t o

the invcrsion of n rnatrices of orclcr n .
I f , besides, the matrix A
i s block-syrinetric (A1 = $-1 , A2 = , ...
) , then we need i n v e r t
only [n/2] ~ a t r i c c s ;the nuccessivc algebraic manipulations a r e also
simpler then because the inverse matrix i s a l s o bloc]:-synmetric.
7. A simple analytical scheme f o r the study of the s t a b i l i t y of

Couette flow.
To i l l u s t r a t e with one example the analytical and numerical
problems t h a t a r i s e i n the study of thc s t a b i l i t y 02 a steady flow,
wc examine now i n some d e t a i l the behaviour of perturbations intro-
duced i n the circumferential flow between two concentric coaxial
r o t a t i n g cylinders (Couette flow). An analysis of t h i s problem has
i n t e r e s t for many reasons:
(i) the Couette flow i s one of the very few steady flows of a viscous
fluid for which one has a precise a n a l y t i c a l description.
(ii) the s t a b i l i t y of the Couette flow can be studied i n the laboratory
through r e l a t i v e l y sir:ple cxperirients. There have Lcen precise
G, Capriz
experimental studies of Taylor, Conelly and more recently

k, 21 .
of coles 42.
(iii) tlie Couette flow is subject t o a form of hydrodynamic insta-
b i l i t y t h a t lends i t s e l f t o an analytical treatment, through
l i n c a r i s a t i o n of tlic perturbation equation, w i t i i forecasts
anply con£irned Ly experiments [sf441 .
(iv) tile special kype of hydrodynamic i n s t a b i l i t y lends i t s e l f
t o a rigorous a n a l y t i c a l treatment also through a study of
ti.e corilplete non-linear equations [37] .
I.,e fu1ldar:cntal reason f ~ (r i i ) ,(iii),( i v ) i s tilt f a c t t h a t
i n s t a l i l i t - i s iri r.0s.l; cases due t o t r a n s i t i o n t o other forms of
steady flor, (Taylor vortices - for which a x i a l symmetry s t i l l
holds - or Coles vavy v o r t i c e s ) rathcr than t o t r a n s i t i o n towards
turbulence (a:, ilapycns inctcac for Poiseuille fiow).
Apart fror, i t s analytical-el:pcrimental i n t e r e s t , tlie Taylor
vortcs £10;: ;;as ir;i~ortai~cei n prncticc a t l o a s t f o r tvo rcasons:
the flcbi i i ~plair, Learirigs of larcjc r o t o r s (turLines, a l t e r n a t o r s )
i s roug,~l;. a Cosettc floi,, suLject a t high spceds t o Taylor in-
s t a l i l i t j r ; billci~Prylor vortices appear tlie viscous losses i n the
lubricarlt Lecom c!~ciliiiglmr than is forecast on the assumption of
Couctte flov; 1,encc the i n t e r e s t of a precisc understanding of t h e
Taylor i n s t a L i l i t y . Therc have been also a t t e a p t s t o use the Taylor
vorticcs a s s e a l s i l l Lcaringr;. Secondlj, boundary layers along
concave rralls a r c subject t o i n s t a b i l i t y of a similar kind, t h a t
can I*! s t u i i e i Ly s i ~ ! : i l a rmeans (Gdrtler vortices) (4 .
The s t a r t i n g point f o r a nunerical anal-rjis of tile ( a x i s y m e t r i c )
Taylor vortices a r e equation (16) anci tilc appropriate boundary
coniitions. iraturally tile p e c u l i a r i t i e s of tile problem allow
C
binplizication; 2 i s i n the circunferential direction, y i s
assumed not t o iepcnd on 6 but only on t h c r a d i a l and G i a l
coordinates r and z. i:e can use the gap r Z - rl between cvli?ler::
G. Capriz
a s typical dinension and put 5 = (r -

rl) / (r2 - rl) ,
= Z/ (r2 - rl) CI
together
with T = v t / (r2 - rllL.
Then the r a d i a l and a x i a l components of spccc: are r e l a t e d t o tile
*
derivatives of the transverse cornpoilent of )Y and it would Le un-
economical t o introduce the other two compo~entsof y*simplyt o dg
f i n e the transverse component of speed ; t h e simplest sc;leme der ivcs
P
-
from the use of second component y of y and the transverse conpollent
of speeu. Reasons of simplicity connectcc! ~ i i t i ispecial features of
the ,problem (such a s tile a x i a l ayzmetry) sucjcjcst a s l i g h t modification
of the usual formulae and the adoption of t!ie following ones, which
a r e s e l f explanatory (al, fi2angular speeds of i n t e r n a l and
external cylinder respectively , 1 = / n2 nl)
In writing the two scalar equations a formal complication

follows from the use of cylinqrical coordinates. To make t h i s ex-
position a s simple a s possible reference is made here only t o an
asymptotic oase: t h a t of small clearance (r2 rl) / r 2 & 1 - .
The equations valid i n t h a t case a r e
(see 2 f o r the derivation); here T i s a mean R e y n o l d s ' n d e r

G. Capriz
The boundary conditions express: vanishing of mean a x i a l flow,

periodicity i n the a x i a l d i r e c t i o n (with a r b i t r a r y period 2 q),
vanishing of the perturbation on tile cylinders
A f u r t i ~ c rsimplification can be obtained i n (21) by choosing t o

put 2 = 1 ; it may seem t h a t the s i r p i i f i c a t i o n denies physical
significance t o the resulting prohlcn. In f a c t i t i s found t h a t
the problem i s i n t e r e s t i n g and c e r t a i n conscquenccs (such as tile
c r i t i c a l Taylor number) deduced i n the special case can bc applied
with good approx'ination a l s o for 1 i n the closed i n t e r v a l ( 0 , l ) .
\Je consider then i n the following sections t h i s problem: find i n
5
the rectangle (06 Q 1 , - q i j 5 q , t 3 0 ) a v c c t d r ( Y / , V )
s a t i s f y i n g the equations
the boundary conditions (22) and given i n i t i a l conditions.

C. Capriz
3. -S
. o ~ er e s u l t s regarding thc d i f f e r e n t i a l problem.
'ule Lcgin the study of our example with an analysis of

properties of Lounhry and i n i t i a l value problems related t o
p r o l l e r ~(22), (23). To begin with, it is convenient t o consider
from a partly formal point of view tile solution of the l i n e a r
system
with the boundary conditions (22). Obviously i f ( , ) , 5 5

Iv (5, 5 ) i s n solution of this problem so i s a l s o )Ir ( J t b) 5,
5 3
Y ( , t L) (b, any constant) ; we require of t o be odd and V
t o be even i n .
Then we separate variables, looking f o r
solutions ynl Vn of tlle type
The functions An , Bn s a t i s f y the system of equations (an = n ~ / q )
and the boundary conditions
A
Supposing t h a t an eigenvalue Tn and a corresponding eigenvector
( ,) e x i s t f o r problem (25). (26) , then multiplying both
s i d e s of the f i r s t eqn (25) by the complex conjugate A*
An of A
G. Capriz
and i n t e g r a t i n g over ( 0 , l ) one g e t s
where
Similarly from the second eqn (25) one g e t s
with
Cecause an i s a r e a l p o s i t i v e number, such a r e a l s o I1 ,

ana I2 ; these two numbers vanish only on t h e t r i v i a l s o l u t i o n
of (25), (26).
lience I = I i s r e a l negative (from eqn ( 2 8 ) ) , and t h e eigenvalue
A
Tn
i s r e a l p o s i t i v e (from eqn ( 2 7 ) ) .
As a consequence tihe associated cigenvector can be taken t o have
r e a l components. From t h c formulae above it follows t h a t , i f
A
An , En
do not vanish,
On t h e other hand
G. Capriz
and f i n a l l y
The equation, which the eigenvalucs s a t i s f y , i s e a s i l y found.

A A
IIotice t h a t An and Cn a r e both s o l u t i o n of Lke equation i n
d2 2 2
[ ~ ~ - a , + )a n T n ] y = O .
d 5
4 A
IIence, Loth An ancl En can be expressed a s l i n e a r combinations of
functions e zc
with
l e t u s i n d i c a t e with z1 . .
-zl , z2 -z2 , z3 , -zg t h e s i x d i s t i n c t
dctcrmination; of z [ ~ o t i c et h a t i n e q u a l i t y (29) excludes t h e
occurrcnce of multiple r o o t s ] .
By imposing t h e boundary conditions t h e equation f o r t h e
eigenvalues can be found. I t i s expressed by p u t t i n g equal t o zero
t h e determinant of a 6x6 matrix whose f i r s t t h r e e l i n e s a r e
G. Capriz
anci the o t h e r tlilrcc are formed witiil thc sane co1ur:ns i n thc
ordcr 2, 1, 4, 3, G , 5.
I t can be cliccked Elat Cic determinant A i s equal t o the
differencc of the squarcs of two sums S1 and S2 , where S1 is
tile sum of the determinants of thc matrices of order 3 obtained
by extracting the columns 1, 3, 5; 2, 3, 6; 2, 4 , 5; 1, 4 , 6
of the matrix (30) and S2 by a similar sum t h e r e the columns
1, 3, 6; 1, 4, 5; 2, 3, 5; 2, 4 , 6 a r c involved. Easy develop-
ments lead t o tile r e s u l t
A A
from which an i n p l i c i t multivalued function Tn = Tn (an) can be
computed. On one Lranch of t h i s function tklc r e l a t i o n T

h
- 4
+_ zit n - an
i s s a t i s f i e d (because then the functions e , i = 1,2,3 a r e
not independent), but t h a t Lranch i s w i t l ~ o u ti n t e r e s t f o r
conputing eigenvalues i n view of inequality (29).
A graphical rcpresentation of tlic two branches over which
..
A
ti12 valuc of 'I corrzspo~dingt o a given of an i s the lowest

n.
and next lowest a r e shown i n tile figure.
G. Capriz
m e numerical experiments then show t h a t t h e r e a r e r e a l p o s i t i v e

eigenvalues of our problem and t h a t appropriate values of q can
be found such t h a t t o an = x / q t h e r e corresponds an eigenvalue
A
T1 which is not an eigenvalue when a is chosen equal t o j n / q
(j = 2, 3,...). To such T t h e r e corresponds then only one
A A
eigenvector of t h e type sought i n t h i s Section: T1 { T~ ( j = 2 ,3.. .I .
These r e s u l t s , based here simply on numerical evidence,
can be reached without recourse t o experiments [sce 361 , -
througl~reference t o p r o p e r t i e s of s o l u t i o n s of a v a r i a t i o n a l
problem based on t h e e q u a l i t y
G. Capriz
which follows from ( 2 4 ) and t h e boundary conditions ( 2 2 ) .
We r e c a l l now some r e s u l t s regarding the l i n e a r d i f f u s i o n

problem
with t h e boundary conditions ( 2 2 ) . Ile do n o t quote here p r o p e r t i e s

of s p e c i a l s o l u t i o n s corresponding t o p a r t i c u l a r i n i t i a l d a t a b u t
r a t h e r s t a t e t h e general behaviour of s o l u t i o n s depending on T .
*'P1 ,s .
Remark t h a t the general s o l u t i o n of ( 3 2 ) , ( 2 2 ) i s a l i n e a r
combination of functions of thc type y =c
V=c
J t
Ul (1,s ) where y i s a r e a l nuculcr and )V
the d i f f e r e n t i a l system
(f
. ) ,
V1 s a t i s f y
with the usual boundary conditions.

G. Capriz
By reference t o t h i s eigenvalue problem it i s possible t o prove

s e e b t P p . l ~ b - l l i J that:
h
( i ) when T is smaller than the smallest eigenvalue T1 of the
problem of Sect. 8, # i s necessarily negative;
h A
(ii)when i s i n an appropriate i n t e r v a l T1 , T~ t6(6 > o)
T
there are solutions of ( 2 2 ) , ( 3 2 ) exponentially increasing
with time.
G. Capriz
9. The non-linear problem.
We consider here t h e non-linear problem, whose s o l u t i o n

represents t h e Taylor v o r t i c e s within our approximation :
with tile boundary conditions ( 2 2 ) .

We give p r e c i s e sense t o t h i s problem by s t a t i n g t h e ~ c t
where we seek a n o n - t r i v i a l solution: i t i s a s u b s e t a o f a Sanach
space E of v e c t o r s (ly ,V ) obtained thus.
Consider t h e s e t 3, 5
of functions 'f ( I 5 ) cleflined i n
a s t r i p S1 l a r g e r than t h e s t r i p S :06 $ 1, of c l a s s C" i n S1 ,
5
periodic i n f with period 2q; introduce i n tile norm
Y m
-
and l e t 5 be t h e closure of 3 with reference t o t h i s norm.
- Then E-is t h e Banach space of v e c t o r s ( ) V I 'V) with ?y
in , v i n y l and t h e norm
We consider a l s o t h e s e t of a l l functions f ( 5 , y ) , C m
i n S , p e r i o d i c i n S w i t h p e r i d 2q, which vanish i n a s t r i p along
t h e boundary of G. The space obtained by closure of t h e s e t with
G. Capriz
reference t o t h e norm
Then
e
11 yIIE,
will bc indicated with
El'
i s t h e s e t of v e c t o r s (Y),V ) of C :iitii y 6 H2
1: .
and V t iil ; i n f a c t vectors such t h a t y 6 V L H ~s a t i s f y , i n
G2,
a generalized scnsc, t h e boundary conditions at = 0, = 1. 7 3
\re w i l l look thcn f o r s o l u t i o n s of our 2roblcrn (33) i n
7 .
I t i s p o s s i b l c t o silov f i r s t of a l l t h a t t l ~ e r ca r e no solutioils
n
e::cept t h e t r i v i a l one, f o r T < T1 ( f o r a proof, scc 9 , p p .
59-60).
- , pp. 4-51 ; s e e a l s o f o r some
I t i s p o s s i b l e t o ailow f u r t h e r [37
c o ~ n e n t sthe rlZIIrcss [46]- t h a t t h e t r a n s formation a 1
(y ,v ) 4( y ,v )
I
defined throug!l tile ilon l i n e a r pro:~lcm i n Y' ,If:
~ ~ l = r a? ( ~
~ , .v ) _
2 5 +&y.5, I I
and Ly requiring ti:at ( y , V ) Be i n E , ar.d (y , V' ) !JC i n 2
i s a complctc f u n c t i c n a l t r a n s f o r n a t i o n of tilt s: acc C i n t o i t s c l f .
Its fixed p o i n t r a r c t h c s o l u t i o n s of our probler?,
Siriilary t h c t r a n s f o r r a t i o n d :
Elrourjh t h e liilcar problem
(y .
v)+
I I
(y
, v- ) Bef ineci
Azr = -3 Y
13
and by requiring t h a t ( :V; he i n S, and (y
I
,P
I
;,e i n (k- is
a l s o a c o ~ p l c t cl i n e a r t r a n s f o r ~ a t i o nof L i n t o i t s e l f . Tile fixeC
p o i n t s a r c t h c cigcnfunctions of the proj;lcr: of s c c t . 2.
-.
;ioT.r i t can l..c proved tililt his
tbc FrCclict i i f f c r c n t i o l of
tile t r m s : o r r - . ~ t i o ~ hi a t t h e p o i n t ( 0 . 0 ) of npocc 3.
A l l ve Iinvc s a i d rcr-ains t r u c i f tre suLnLitutc t h c s p ~ c ci;
with t h c r:ubsyacc R
1
of t k e v c c t ~ r s(Y,'Lr) of L r,uch t:;at Y is
odd a n d t l i s cvcn iil j. Thc advantarjc of considcrincj o c r problcr: i:i
A1 i s t h i s , tli':t (as rcrnilrkcG i n Scct. 3 ) t h c r c exist: c::oiccs GT c.;
G. Capriz
A
such t h a t t o t h e associated cigenvalue T1 t h e r e correspons only
one eigenvector:
h
T1 has m u l t i p l i c i t y 1.
h
Then, f o r a theoren of Leray-Shauder, Tl i s a branching
p o i n t f o r t h e s o l u t i o n s of t h e problem ( 3 3 ) , (22): a n o n - t r i v i a l
s o l u t i o n of our problem must e x i s t i f t h e value of 'I' i s chosen
4 h
w i t h i n a s u f f i c i e n t l y small i n t e r v a l ( T1, TI t d l d)O 1.
10. liw,erical study of t h e non-lincar e l l i p t i c probleri
The a n a l y t i c a l developmnts of S e c t s 2, 9 assure us of t h e

existence of n o n - t r i v i a l steady s t a t e s o l u t i o n s of t h e non-linear
N
p e r t u r b a t i o n equations ( 2 2 ) , (32) f o r T ) P1, hence of a Sranching
of t h e fundanental solution. Tiley allow us a l s o t o c a l c u l a t e
approxirateiy t h e value of tile Taylor nupber tiiat c l l a r a c t e r i ~ c st h e
t r a n s i t i o n . I n p r a c t i c e onc vould l i k e t o know t h c a ~ p l i t u d eof t h c
p e r t u r b a t i o n a s a function of T Lcyond t h e c r i t i c a l value T1
A
.
i;nov~lcJge of t h a t arplitufic lcadr; f o r i n s t a n c c t o an evaluation of
t h c couples a c t i n g on t h e r o t a t i n g c:?linCers, norc p r e c i s e l y of t h e
excess of those couples beyond t h e valuc t h a t would be prcclictcd f o r
Couctte flow. For such an evaluation a rccourse t o n u ~ c r i c a lrethods
i s c s r ; e l t i a l . One can pursue e i t h e r the r.ur:,crical i n t e g r a t i o n of t h e
stcady s t a t c equations ( 2 2 ) , ( 2 3 ) tilrougli a process of d i c c r c t i z a t i o n
and rucccssivc approrir:ations / 101 o r a numcricai i n t e g r a t i o n of
tllc d i f f u s i o n equations ( 2 2 ) , (23) u n t i l n s t a t c i s rcaci~cdsufficientl;.
near the s t c a c y s t a t c [ 2-2 ] .
We rjive hcrc i i r s t of a l l soi:.c d e t a i l s of t h e f i r s t process
O F d i s c r e t i z a t i o n i n a s p e c i a l casc considcrinq thc f i n i t c -
diffcrciicc problcn:, which Ccrivcs from ( 2 2 ) , (33) f o r tile clioicc
C. Capriz
q = 1, when t h e n e t points a r e chosen t o have coordinates
r=mh, j=? (p 1
b v i t h h = . ( n t l ) -1 ,K = 1, 2 , ....n,
p = 1, 2, .... .
r.tl The boundary c;,nditions f o r a t ?= (i,
5 = 1 c a l l f o r the use of f i c t i t i o u s e x t e r n a l p o i n t s , idicrcas

t h e conditions a t S = 5 1 inply p r o p e r t i e s of. the operating
natrices.
P r e c i s e l y , t h e f i n i t e d i f f e r e n c e problem can be v r i t t c n a s follows
where y , a r c two vectors vitlr i n 2 + 2n corzponents, each of wllich

gives t h e approximate value of t h e functions , v over tire mesh
p o i n t s ordered from l e f t t o r i g h t and f r o n t o p t o Lottor?. ;
Ul , u2 , Dl a r e block-circulant ant! bloclc-syrmetric n a t r i c c s of
order 2n t 2 i n subnatrices of order n :
J1 =(or -1, 0, . . . , 0, I,],

with 0 n u l l matrix of order n ; I i d e n t i t y matrix of order n ;
A, G , C syrmetric n a t r i c e s of order n, of which t h c f i r s t i s
pentadiagonal and t h e o t h e r two t r i d i a g o n a l :
G. Capriz
M2 a r e non-linear operators a c t i n g t h e f i r s t on t h e vector

Ml,
-
Y
and the second on t h e cornpunci vector (
- I
If 1.
As i n t h e case of Lhe d i f f e r e n t i a l problem, it i s possible
t o show t h a t t h e values of T , for which t h e associated l i n e a r
homogeneous systerr,
has n o n - t r i v i a l s o l u t i o n s , i - e - t h e values of T which a r e r o o t s of

t h e a l g e b r a i c equation of degree 2n 2 t 2n
has s o l u t i o n s , a r e p o s i t i v e (we w i l l r e f e ~
t o these values a s t h e
-\Y -
h
eigenvalues of t h e problem). I n f a c t , i f , V i s a s o l u t i o n of
(35) corresponding t o t h e eigenvalue T , then
xT
but - u22 ) a r e p o o i i i v e d e f i n i t e q u a d r a t i c Poimc
U1 --x , - (5T
x . I t follows t h a t
h
i n t h e components of - must be p o s i t i v e ;
P
f,
A
i t follo\rs a l s o t h a t -
'v can be talien t o have r e a l corponents.
Again a s i n t h e case of che d i f f e r e n t i a l problen i t i s founii
t h a t nor. t r i v i a l s o l u t i o n s of ti:e non- l i n e a r problem ( 3 4 ) niay e x i s t
n
only f o r values of T cjrcatcr than t h e lowest eigenvaluc T of ( 3 5 ) .
C
Po reach a proof of a c t u a l e x i s t e n c e of a s o l u t i o n unclcr t h e contiition
h
T > Tc , some preliminary r e s u l t s a r e required.
F i r s t l y we remark t h a t eigensolutions of ( 2 5 ) can Le w r i t t e n a s folloyiis
irllcrc k. = sin
1,
[ (2i-1) r 4 ], ,,(=_n_
2nt2 ' l$r$n ;
G. Capriz
and
-)4 , a r c n-vectors which s a t i s f y t h e equations
3
( 1
1 + 2cos 2r d l3 + 2cos 4 r d I )
- = h T sin 2 r l c
(37)
( c+ 2cos 2 r d I )g = - h s i n 2rA '-f .
~ 1 i r r . i n a t i n g one o b t a i n s t h e equation i n
-'P
with
= - ( c + 2 cosrA I ) ( A t 2 cos 2 r d L ; + 2 cos 4 ro( I ) .
'-r
hence s o l u t i o n s of our problem (35) e x i s t provided t h a t i14? sin'(2r.O
the
coinciclcs with one of,cigenvalues of t i ~ cmatrix Cr- (r = 1, ... n) .
one elementary uevclopments show C47J
--- t h a t t h e matrices
(A + 2 cos 2r 4 E t 2 cos $1 a -'
~ ) ( r = 1, ... n) a r c p o s i t i v e ; on tile
o t h e r hand t h c rcatrices - ( C t 2 cos 2 r d I ) ( r = 1, ... n) are
i r r c d u c i b l e , d i a g o n a l l y dominant n a t r i c e s with pos,itive i i a g o n a l
c l c n e n t s and non-positive off-diagonal c l e n e n t ; s o t n a t t h e matrices
- (C + 2 cos 2 r 4 arc positive.
I ) - ~
iicnce t h e C
r a r e positive; nore precisely
C n ( Cn-l < ... < C1 . (39)

I f vc C ~ ~ O G fCo r T
4
a value T s o t h a t h4 ? s i n 2 2r 4 i s an cigenvalue
of Cr , we can c o n s t r u c t t;irough (38) t h e corresponding eigenvector
-f and, s u c c e s s i v e l y , througii t h e second eqn (37) and formulae (36)

t h e ciqcnvcctor ( Y-/ , y ) of our o r i g i n a l problem ( 3 5 ) .
G. Capriz
Equivalentlywe could say t h a t t h e system
admits of s o l u t i o n s .
2
Thus, we can determine n eigenva!.ues of M
( i f each i s counted with t h e appr@& m u l t i p l i c i t y ) . Actually
correspond two eigenvectors ( - ,

it can be checked t h a t t o each eigenvalue s o determined t h e r e
, 1 ; t h e f i r s t i s of t h e
foyn ( 3 6 ) t h e second has a s i m i l a r s t r u c t u r e but t h e r 6 l e of t h e
trigonometric functions i n t h e d e f i n i n g formulae i s reversed.
2
We have accounted s o f a r f o r 2n eigenvalues of Pl ; t h e
remaining 2n a r e zero; i n f a c t M i s a s i n g u l a r matrix.
1
Let us consider now t h e s p e c t r a l r a d i u s ( C r ) of C r ; a
theorem of Perron-Frobenius and t h e i n e q u a l i t i e s ( 3 9 ) assure us
that (Cr) < 'j (Cn-l)<
p ...
4 'j (C1) , and t h a t 'j (C1) i s a
simple eigenvalue of C1 s o t h a t 7 (C1) i s n o t an cigenvalue of
C2, . . . Cn.
I n conclusion
is t h e minimum value of T f o r which t:.e problem (35) has solution.

To t h i s value of T t h e r e corresponds a unique s o l u t i o n ( a p a r t
from a constant f a c t o r ) of t h e type
whereyl ,- a r e n(nt1)-vectors.
Consider now t h e vector space Ii of t h e 4n(n+l)-vectors
( $' I - Y
fl ) wit11 ) , of t h e type ( 4 1 ) ; l e t E be norned (, I f Lie .
choose any v e c t o r ( l/)
-1
y 1 i n D and c a l c u l a t e t h e v e c t o r s
G. Capriz
these belong a l s o t o B .
Iience we can consider t h e eigenvectors
of (35) and t h e n o n - - t r i v i a l s o l u t i o n s of (34) r e s p e c t i v e l y a s fixed
p o i n t s of t h e following compact nappings of E i n t o i t s e l f
and
I
IJotr 2 i s thn Fr6chet d i f f e r e n t i a l of c a l c u l a t e d over
t h e n u l l clement of B . I f we see'- s o l u t i o n of our problems ( 3 4 ) , (35)
A
exclusively within I3 then we f i n d t h a t f o r T = Tc t h e r e corzesponds
a simple eigcnvaluc of (35) .
A theorem of Leray Schauder assures u s f t!len,of t h e existence
of a n o n - t r i v i a l s o l u t i o n of (34) f o r each choice of T i n an appropriate
A 4
i n t e r v a l ( T ? td ) , d ) C ; i n other vords
cf C - c is a branching p o i n t
f o r thc s o l u t i c n s of ( 3 4 ) .
G. Capriz
11. Notes on t h e numerical experiments
We have d e a l t s o f a r with fundamental questions r e l a t e d t o

t h e system (34) ; we comment now b r i e f l y on problems connected
w i t h t h e planning of a c t u a l numerical experiments.
Because of t h e non-linear nature of system ( 3 4 ) , i t s
p r a c t i c a l s o l u t i o n c a l k f o r an i t e r a t i v e procedure of t h e Y;pe
envisaged i n eqn (20) ; f o r s i m p l i c i t y we nake r e f e r e n c t h9re t o
t h e scheme
although t h e a l t e r n a t i v e scheme
seems t o be f a s t e r .
I t i s found t h a t , i f t h e s t e p 11 i s chosen t o be small
n u l l vector when T C
-% .
enough, t h e v e c t o r (Y(") y ( n ' ) tends with increasing n t o tile
, whereas i t converges toiiards t h e
A
n o n - t r i v i a l s o l u t i o n of (34) when T > Tc.

A
lJotice t h a t t h i s happens al.though f o r T > Tc , eqn ( 3 4 )
admits always a t r i v i a l s o l u t i o n . A p r e c i s e a n a l y s i s of t h i s
Lehaviour i s n o t a v a i l a b l e ; we can add hcrc only a h c u r i s t i c
argument which i n d i c a t e s a bound on h f o r t h e s t a b i l i t y of t h e
process (42). This bound was v e r i f i e d c l o s e l y i n p r a c t i c e ; it i s
of t h e type mentioned i n Sect. 4.
G. Capriz
Let US call g?) , the e r r o r s i n y ("I an6

respectively . Then, from (42) ve get
A
where now y , *stand f o r the solution of (34) .
We accept the approximate e q u a l i t i e s
so t h a t i t follows from system (43) t h a t
This equality implies t h a t the e r r o r decreases only i f the spectral

radius f ( * of the matrix
G. Capriz
A 4
does not exceed unity. g ( W depends on h , T and a l s o on Y , ; -
b u t these two l a s t vectors a r e unknown t o s t a r t with : a reasonable
guess f o r t h e s o l u t i o n i s required i n p r a c t i c e f o r an evaluation
of t h e conditions of convergence; such conditions w i l l put then
r e s t r i c t i o n s on h depending on t h e value of T. However t h e c a l c u l a t i o n
of t h e s p e c t r a l r a d i u s oflvllis n o t an easy matter; a s a consequence
one i s forced t o r e l y on rougher estimates, such as' t h e following one.
& (k-1)
Assume t h a t i n t h e v e c t o r -
E (k-l)= ( -'
t;$k-l)
) a l l but
one component vanish, f o r i n s t a n c e t h e e r r o r component r e l a t i v e t o

t h e value of _v over a c e r t a i n mesh p o i n t P. Then we may take, a s an
approximation, t h a t only a few components of - E ( k ) a r e d i f f e r e n t from
zero, p r e c i s e l y those r e l a t i v e t o values of over meshpoints adjoining
P.
I f P i s s u f f i c i e n t l y f a r from t h e boundary t h e components
taken t o be non-null a r e those of order m - n, m - 1, m + 1 , m + n .
During t h e next i t e r a t i o n , leading t o - E , there i s a
"backfire" e f f e c t of t h e spread e r r o r over t h e m-th component 5
m
(kt11
A s a rough estimate of t h e condition of s t a b i l i t y it i s required t h a t
C. Capriz
Notice i n c i d e n t a l l y t h a t , i f t h i s criterion i s adapted t o

A
apply t o t h e h e a t - t r a n s f e r equation -3 u = d--T i n the f i n i t e

3t a,
n
d i f f e r e n c e form untl
I
- unj = l d ~ t /( A X ) 1 [u yl-2 u j t u j l l I
j
,
it leads t o t h e s t a b i l i t y r u l e [ d ~ t / ( & x ) * ] 5 2/3 a s can

be e a s i l y checked; llerc t h e notation i s obvious. A s i s wc '.l known
.
a more appropriate a n a l y s i s i n t h i s case suggcsts thc upp limit -
1/2 r a t h e r 2/3 f o r t h e r a t i o [ d h t / ( ~x ) ~ ] , a t l e a s t i n t h e case
of simple boundary conditions.
Sirnilary, i f t h e c r i t e r i o n i s adapted t o apply t o t h e wave
equation a2u - = c &- i n t h e f i n i t c - d i f f e r e n c e form

3t2 a x2
it leads t o t h e r u l e (c A t / A x ) 6 1 .
Returning ncw t o our problem, we a r c - i n t e r c s t e d i n t h e
s o l u t i o n of a l i n e a r system extracted from t h e system
A
i n f a c t we have supposeil that'j-y be n u l l .

Decause a l l components of E$-"
but one a r c a l s o n u l l ,
we lntend t o examinc t h e approximation \?here a l l b u t f i v e components
of &$I vanish. Thcse components s - t l s f y t h e following reduced
system
wEiich can be e a s i l y . s o l v e d . I n t h e sccoiid s t e p leading t o 5
linaar
s y s t c n i i i t h 13 unknoi~ns i s involved; we leave o u t d e t a i l s t o ricofc
t;ic r e s u l t
Fror? a r1ur:crical p o i n t of v i a ; i n c q s a l i t y ( 4 4 ) I?;I'I Lc

iiiterprctcd i~oi,a; a c o c s t r a i , l t inponeri upon tile change of)Uovcr
two nlcilcsh s t e p s ; i l l f a c t ( 1 4 ) i ~ p l i c s ,i l l view of (/15),
-
An a l t e r n a t i v e , physically sigriilicant, ititerpretation

05 (45) i n possible; consider t h e noZulus S of t h e p r o j e c t i o n i n
the (r, 2)-plane of t!le v e l o c i t y of t h e f l u i d and i n d i c a t e with RI!
tile Xeynolds nunbcr based on S , tile physical s i z e ( r 2 - rl) h
of t h e mesh and t h e v i c c o o i t y V .T:';len (45) can be w r i t t e n
C. Capriz
These c r i t e r i a of s t a b i l i t y , though rough, have

proved t o be very u s e f u l i n t h e preparation of computer programs.
For samples of r e s u l t s of numerical work we cake r e f e r e n c e , f o r
instance, t o paper [?J .
G. Capriz
1 L. F. Richardson, Weather p r e d i c t i o n by n m e r i c a l proccss.

C d r i d y e Univ. Press., London 1922.
k] R. BsrXer, I n t g g r a t i o n des Equations du mouv2mcnt d'un f l u i d c

visqucux incompressible. Encyclopedia of Physics, vo1.8/2,
Springer (1966).
13) R. Finn, S t a t i o n a r y s o l u t i o n s of t h e TJavier-Stokes equations.

Proc. Symposia Appl. Math., _1_Z (1965) , 121-153.
D.N. de Allen, Relaxation Flethod~ i n Engineering and Science,

Mc Graw-IIill, 1954.
[5] A. Thoni, C . J . Apelt, F i e l d Computations i n Engineering and

Physics. Van Costrand, 1961,
b] J. von Meumann, Collected Works, vol. 5. Perganon P r e s s , 1963.
[6 d i d . , vol. 6
[7] B. Aldcr, S.Fernbach, It. Rotenberg, ed. , ikthods i n

computational physics. Academic Press. 1, S t a t i s t i c a l
Physics (1962) ; 2, Quantum ilcchanics (1763) ; 3, Punda-
mental Methods ixIIydrodynanics (1364) ; 4 , ~ ~ p l i c a t i o n s
i n IIyclrodynamics (13G5) ; -5, Fluclear ~ a r t x c l eKinematics
(1366) ; -6, lruclcar Physics.
181 F.II. IIarlotr, The p a r t i c l e -in - c e l l nethods f o r numerical

s o l u t i o n of problems i n f l u i d dynamics. Proc. Symposia
Appl. Math., 15 (19631, 263-288.
b]
- J.R. P a s t a , S. Ulam, I I e u r i s t i c numerical work i n some problems
of hydrodynaniics. IIath. Tables Other Aids Conp., 13
(1953), 1-12.
[ld A. B l a i r , IT. 1Ictropolis, J , von Ileunann, A.11. Taub, $1. Tsingou,
A study of a numerical s o l u t i o n t o a two-dimensional
hydrodynamical problem. 1"ltt. Tables Other Aids Comp.,
13
- (1353), 145-124.
G. Capriz
[11] J.C. Welch, F.H. IIarlow, J . P . Fhannon, B.J. Daly, The MAC
method, a computing technique f o r solving viscous,
incompressible, t r a n s i e n t fluid-flow problems involving
f r e e surfaces. Los Alamos Scient. Lab., LA - 3425.
[id F. I!. Ilarlow, J.E. Welch, rrumerical c a l c u l a t i o n of tine-depez

dent viscous inconpressible flow of f l u i d with f r e e
-
surface. Phys. Fluids 8 (1965), 2132-2139.
1131 F.11. IIarlov, J.E. tlclch, Numerical study of large-amplitude

free-surface motions. Phys. F l u i d s , 9 (19661, 042-351.
[14] J.D. Bellun;~, S.W. Churchill, Computation of n a t u r a l

convection by f i n i t e d i f f e r e n c e rncthods. Proc. I n t .
Conference on neat Transfer, I n s t . flech. Cng.,
London (1361).
[15] 1.1.R. Abbott, A numerical nethod f o r solving t h e equations of

n a t u r a l convection . n a narrow concentric c y l i n d r i c a l
annulus with a horizmtal a x i s . Quat. Journ. Flech.
Appl. Kath., 17 (19641, 471-431.
[16] C.C.
in [h .
L e i t h , l m e r i c a l simulation of t h e e a r t h ' s atmosphere
vol. - 4 . 1-29.
[17] A. ::asai~ara, E. Isaacson, J.J. Stoker, EJumerical s t u d i e s

of f r o n t a l motion i n t h e atmosphere,
T e l l u s , 2 (1965), 1.
114 C. Isaacson, Fluid dynamical c a l c u l a t i o n s i n EJumerical

Solution of P a r t i a l D i f f e r e n t i a l Cquations, J.11.
Cramblc, ed.,Acadcruic Press. Mew York (1366) , 35-43.
1 J.C. Fronw,P.H. Iiarlou, IIumerical s o l u t i o n of the problem of

vortex s t r e e t devclopnent. P:?ys. F l u i d s , 5 (1963) ,
975-332.
G. C a p r i z
P. I?. I!arlotl, J . C . F r o m , Dynamics and h e a t t r a n s f e r i n t h e

von ~ d r n ~ hwake
n of a r e c t a n g u l a r c y l i n d e r .
Phyc. F l u i d s , -
7 ( 1 3 6 4 ) , 1147-1156.
Z.J. Daly, A n u n e r i c a l s t u d y of two f l u i d Rayleigh-Taylor

i n s t a b i l i t y . The P h y s i c s of F l u i d s , 10 ( 1 9 6 7 ) , 297.
A. L. K r i l o v , E.1;. P r o i z v o l o v a , 1Jurr.crical a n a l y s i s of t h e f l u i d
f l o ~ ?between two r o t a t i n g c y l i n d e r s . P r o c c e d i n g s
(CGOPIIWK PABOT) Computing C c n t r c !loscow Univ. , 2 -
( 1 9 6 3 ) , 174-181.
G . C a p r i z , C.Ghelardoni, C.Lo%bardi, PJumerical s t u d y o f t h e

s t a b i l i t y problem f 4 r C o u e t t c flow.
Phys. P l y i d s , - 9 ( 1 9 6 6 ) , 1934-1936.
.
F. A . Ilarlo\?, J .E F r o m , Computer e x p e r i m e n t s i n f l u i d
dynamics.
S c i e n t i f i c American, 212 ( 1 9 6 5 ) , 104-110.
F.H. Harlow, J . P . Shannon, J.E. :Jelch, L i q u i d waves by computer.

S c i e n c e , 149 (1965) , 1092-1093.
J .E. b ~ e l c h , Computer s i m u l a t i o n o f w a t e r waves,

-
Datamation 12 (1966), 41.
F.H. IIarloti, J.P. Shannon, J.E. Welch, Un c a l c u l a t e u r q u i f a i t

-
d e s vagues. S c i e n c e s , 7 ( 1 9 6 6 ) , 14.
D. Greenspan, P.C. J a i n , R. Manohar, E. Ebble, A . S a b u r a i ,

Numerical s t u d i e s o f t h e Navicr-Stokes e q u a t i o n s .
Math. Res. C e n t e r , Techn. Summary Rept. 402 (1964).
P.C. J a i n , Numerical s t u d y of t h e EJavier-Stokes e q u a t i o n s f o r

t h e p r o d u c t i o n o f s m a l l e d d i e s from l a r g e o n e s .
Math. Res. C e n t e r . Techn. Sunmary Rept. 491 ( 1 9 6 4 ) .
G. Capriz
E. De Luca, Numerical s t u d i e s of p o i n t p e r t u r b a t i o n s i n
laminar plane P o i s e u i l l e n o t i o n .
Army Material Res. Agency, Tech. Rcpt. APIRA TR 63-10.
M. Capovani, G.Capriz, G.Lombardi, Studio numeric0 d e l l a

s t a b i l i t a d e l mot0 d i un f l u i d o viscoso i n un canale.
Calcolo - 2, Suppl. 1 (1965), 33-49.
E. Dellorno, A numerical program f o r dealing with f i n i t e - m p l i t u d e

distcrbance i n plane p a r a l l e l laminar flows.
!!eccanica, 2 (1967), 95-108.
D.F. De Santo, H . 0 . K e l l c r , Numerical s t u d i e s of t r a n s i t i o n

from laminar t o t u r b u l e n t flow over a f l a t p l a t e .
J . Soc. Ind. Appl. !lath. , 2 (1962), 569-595.
J.A.T. Bye, Obtaining s o l u t i o n s of t h e 1Javier-Stokes equation

by r e l a x a t i o n processes.
Comp. J., 8 (1965-66), 53-56.
L.H. Thomas, The s t a b i l i t y of plane P o i s e u i l l e flow.
Phys. Rev. -91 (1953), 780-733.
W. V e l t e , S t a b i l i t a t s v e r h a l t e n und Vcrzweigung s t a t i o n s r e r
LGsungen d e r tlavier-Sto1:csschen Cleichungcn.
Arch. Rat. Mch. Anal., j& (1964), 97-125.
W. Velte, S t a b i l i t S t und V e r z v c i g u n ~S t a t i o n S r e r LSsungcn d e r

1Javicr-Stokesschen Glcichungcn beim Taylor Problem.
Arch. Rat. Iiech. Anal., 22 (1366), 1 - 1 4 .
I<. i;irchgYssner, Die I n s t a b i l i t a t d e r StrSrnung zwischcn zwci
roticrcndcn Zylindern gcqeniibcr Taylor-Wirbeln f u r
belicbige Spaltbreitcn.
Z.A.M.P. -
12 (1961), 14-30.
C.Chelardoni, Qucstioni connesse coll'impostazione a n a l i t i c a
e l a s o l u z i o r e numerica d i un problema d i idrodinarnica.
Calcolo, 2 , Suppl. 1 (1965), 51-66.
G. Capriz
bd G .Ghelardoni .
, G Lombardi , Soluzionc numerica d i un problema
d i s t a b i l i t a idrodinamica.
Calcolo, 2, Suppl. 1 (1965), 67-80.
[43 G.I. Taylor, S t a b i l i t y of a viscous l i q u i d contained between

two r o t a t i n g c y l i n d e r s . P h i l , Trans. Roy. Soc
A 223 (19231, 283-343.
. (London)
R.J. Donnelly, D. F u l t z , Experiments on t h e s t a b i l i t y of s p i r a l
flow between r o t a t i n g c y l i n d e r s .
Proc. tlat. Acad. S c i . , 46 (1960) , 1150-1154.
D.Coles, T r a n s i t i o n i n c i r c u l a r Coucttc flow.

.
Navy Dept Rept. , Cambridge, 1Iass.
S. Cfiandrasel:har, Hydrodynamic and hydromagnetic s t a b i l i t y .

Oxford Univ. Prcss (1961) .
II. I J i t t i n g , Uber den E i n f l u s s d e r S t r g m l i n i e n - ~ r i i m u nauf
~
d i e S t a b i l i t a t laminarcr S t r h u n g e n .
Arch. Rat. Ilech. Anal., 2 (1958), 243-283.
G.Prodi, Problemi d i diramazione per equazioni funzionali.

.
A t t i V I I I Congresso U.M.1 , T r i e s t e (1967) , Under press.
C.Ghelardoni, Considcrazioni numeriche s u un problema d i

idrodinamica.
Calcolo, Under p r e s s .
(C. I. M. E. )
A. DOU
ENERGY INEQUALITIES IN AN ELASTIC CYLINDER
Corso tenuto ad Ispra dal 3 - 7 luglio 1967

ENERGY INEQUALITIES IN AN ELASTIC CYLINDER
by
A. DOU
(University of Madrid)
1 The nrincinle of Saint - Venant.

More than a centrary ago, in 1855 and 1856, B.de Saint-Venant
[ 1 stated a principle that has been applied steadily in the calculus

of beams, but has not yet been proved. F r o m the point of view of
Numerical Analysis the importance of the Principle of Saint-Venant is
obvious and great.
We consider an homogeneous, isotropic and perfectly elastic cylin-
der R
C '
F o r simplicity we assume also
Q = ~~,Y)(IxI<II,IYI < ~ I ) c R2
We assume throughout the paper that there a r e no body forces
and no lateral surface forces; and also that the cylinder i s in statical
equilibrium.
We a r e given the surface forces T(x, y) and T-(x, y) acting
upon the bases of the cylinder z = e and z = - respectively, and
we may assume without loss. of generality that they a r e either
odd
- and write ~ ( l ),
for = -t:T+) =\ -T;(')(=, y) , - T~(1)(x. Y) , T~

(1)(x, Y)] , x. .
Y ~ Q
o r that they a r e even and - write T (2),

A. Dou
for z = I: T( 2) = I
T( 2) (x, y), T;?)(X, y), T~(2)(x, y )
1
1.
f o r z = - I , : ~ - ( ~ ) = { ~ y ) ( x ,~y(2 . (2)
~) ,I x . y -) T ~ (x.Y)} .x.Y(Q.
When t h e r e i s no need to specify, we shall write T ( ~ ,) g = 1, 2, o r

simply T if no confusion i s possible. The specifications r e f e r to the
component r(g) of the corresponding s t r e s s t e n s o r with r e s p e c t to the
33
.
independent variable z T h e s e corresponding s t r e s s t e n s o r will be cal-
led, respectively, odd and even ; obviously a l l t h e i r components a r e
odd o r even functions of z.
We shall consider only those s u r f a c e f o r c e s
(2) ~ ( ~ T' =5( T 1(x,Y), T ~ ( x Y ) T, ~ ( xY)}

, , x,Y~B.
which s a t i s f y the following conditions :
(34 T~cc(G), ~,Jc(G), iZ1.2,3, d=l,2,
In (3a) and throughout the paper the subindex k a f t e r t h e com-

m a means p a r t i a l derivative with r e s p e c t to x , ( x l , X2, x3) 3 (x, Y, z).
The conditions (3b) e x p r e s s that the s u r f a c e f o r c e s a r e self-equilibra-
ted a t each base, and (3c) i n s u r e s that the corresponding s t r e s s t e n s o r
be contilluous in
-R .
Surface f o r c e s satisfying conditions (3) s h a l l be called p e r m i s s i b l e
s u r f a c e f o r c e s (psf) , and the corresponding boundary conditions shall
b e called p e r m i s s i b l e boundary conditions (pbc) . The set of a l l psf
i s obvious by a vector s p a c e o v e r the r e a l s .
A. Dou
Let 11s 1) be the L ~ ( Q ) n o r m of S(x, y ) and let

11 (x)(( be the euclidean norm of t h e s t r e s s t e n s o r Z (g) ,
-x p (x, y, z ) G -R . Then the principle of Saint-Venant c a n b e s t a t e d thus :
Except in the neighborhood of the b a s e s of the cylinder
"C
I/t (z) (1 is s m a l l compared w i t h 1. = 1 ~ 11~ . In p r a c t i c a l applica-
tions l l n e i g h b ~ r h o o dmeans,
~~ f o r instance, that the point x 6 5, is
s u c h that z a'
>, -g- , / >l0.t ; and llsmallll m e a n s that g (3
can be i:~plected altogether, say \If (:)I[ 6 ('110) m a r IITill .
i = l , 2, 3
for 1z1< 4 ( 1 5 , !> 10 h . Today the b e s t justification of t h i s prin-

ciple i s i t s successfuly steady application f o r m o r e than hundred
y e a r s . Mathematically speaking I do not think that this principle,
as stated, is true, although t h e r e must be some very general
conditions, under which it will be t r u e .
The application of the principle to t h e calculus of b e a m s is
indeed both, obvious and useful . In g e n e r a l b e a m s a r e bouded
with forces F(x, y ) , which s a t i s f y conditions (3a) , (3c) and the
first of (3 b 1 ) f o r i = 1, 2; but one h a s
ydxdy = M I ,
2'
Q
The principle allows to dispense with the functions F

1' F2' F3
and take into account only the resultant R, the bending moments
M1' 2
M and the torsion moment M
3
.
T o t h e s e quantities corresponds
in R a unique e l e m e n t a r y solution (x, y, z ) of Saint-Venantls type,
i . e . independent of z, and therefore i s a l i n e a r combination of
- 166 -
A, Dou
uniform traction o r compression, p u r e bending and p u r e torsion.

We r e m a r k the following interesting c o r o l l a r y :
The principle of Saint-Venant implies that t k only bounded solutions
f o r the s t r e s s t e n s o r t ( x , y, z ) in the infinite cylinder R a r e those
WJ
of Saint Venant's type, i. e . independent of z.
In the r e m a i n d e r of t h i s s e m i n a r I s h a l l give t h r e e inequalities
that b e a r on the principle of Saint-Venant and proved by R.A. Toupin
1 2 1 , J. J. Roseman 131 and myself, 14,51 and outline the proof of the
third. Finally I shall comment on related questions.
2 E n e r g y inequalities.
The f i r s t two inequalities due to Toupin and Roseman a s s u m e

that one end, z = f , is loaded with psf T(x, y) and t h e o t h e r and,
z = -e , is free of f o r c e s . This i s achieved in o u r presentation s e t -
(1) +
ting ~ ( l= )T(') = (112) T (x, y ) and loading the cylinder with psf T
+ T (2) .
The r e s u l t of Toupin a s s e r t s that t h e total e l a s t i c energy of
that part of the cylinder beyond a distance s f r o m the loaded
end, U(s) , s a t i s f i e s the inequality
U(s) < U(O) . exp {- s - h

).
where 0 <h <s and c(h) is a c h a r a c t e r i s t i c decay length depen-
ding on e l a s t i c constants and on t h e s m a i l e s t a h a r a c t e r i s t i c frequency
of f r e e vibration of Rh . Substantially i t s a y s that , if we move
away f r o m t h e loaded end, then the total e l a s t i c energy in t h e r e m a i -
ning part of t h e cylinder d e c r e a s e s exponentially.
The r e s u l t of Roseman a s s e r t s that, if P(s) is a point of the
cylinder whose distance t o the loaded end i s s , then
A. Dou
where the f i r s t member and U have already been defined, K is a constant depen-
ding on elastic constants of the body and 0 and a a r e positive constants depending
on the geometry of the cross-section. This result i s similar to the previous poin-
twise estimates of J. H. Bramble and L. E. Payne 161, but this one i s good up to
the boundary.
The third inequality i s contained in the following theorem :
Assume that z i s the s t r e s s tensor in R t corresponding to any psf T =
= ( T I T2, T 3 ) E&,(R) be the total elastic ootential energv pf 2 in and let
-2-
0 st.
< ln Then there is a constant K depending only on 0
such that
(4) " IT^,

< $ ~ ~ P , P ~ + u T J I ~ , + ~ T ~ ~+~I I~T ~ , 2~1 ) -
Now I shall outline the proof of this theorem given in 61. We need the
constituent equations relating the s t r e s s e s with the displacements in an homoge-
neous, isotropic and elastic body, the equations of Elasticity governing
the dispalcements and the theorem of Castigliano.
The s t r e s s tensor <= (( f ..)), i, j = I , 2, 3, analytic in R and continuous
13
in corresponding to the psf T (g)= (g) (g), ~ ( g ) ), g = 1 . 2 , i s given in t e r m s
E, (TI ,T2
of the first derivatives of the displacements u(r, y, z)=(ul,u2 us), (x,y,z)tR, by
A> 0 , p>O 0 <C= 2()itp)

A
'11 ' l i j Kronec ker
symbol.
The displacements u are given by the unique solution of the
Navier equations of elasticity
satisfying the following permissible boundary conditions (pbc) :

A. Dou
which express that no l a t e r a l s u r f a c e f o r c e s a r e present, and
corresponding to the
A tensor
psf in the b a s e s of the cylinder.
0(x, y, z ) g ((0. .)) , i, j = 1, 2 , 3 , continuous in R i s

1.l
-
called a yrirtual s t r e s s tensor f o r z in R , if it s a t i s f i e s the
same pbc (7) and (8) a s 7 and m o r e o v e r i s a solution of the
equilibrium equations
Let
be the positive definite quadratic f o r m of the e l a s t i c energy density

due to the t e n s o r P = (( e i j ) ) . Let t be a s t r e s s t e n s o r in R
1
and 0 a virtual s t r e s s tensor for T. Integrating W7 o v e r 0,
one obtains the total e l a s t i c energy &,(R) due t o the s t r e s s tensor

7 . Then , the t h e o r e m of Castiglano, o r the minimum s t r a i n
energy t h e o r e m s a y s that
The main idea of t h e proof is to construct for any psf, o r

at least f o r s o m e types of psf, a virtual s t r e s s t e n s o r such,
that its virtual energy o r energy integral i s bounded , even when
A. Dou
Now, the equilibrium equations (9) a r e automatically satisfied

by any t e n s o r O(x, y, z) of the f o r m
5 1
= d
11
M
,22
(x, y ) . Nu ( z )
provided that the r e a l n u m b e r s d . .satisfy

11
It is easy to conjecture and verify, that the function N(z)

must i n c r e a s e exponentially, and in the p r e s e n t proof the functions
sink pz for g = 1 and cosh z f o r g = 2 have been taken.
Although t h e boundary conditions ( 7 ) , (8) m u s t a l s o be satisfied,
there is also abundant r o o m to try t o take c a r e of e v e r y condition.
Now we may outline the proof in t h r e e s t e p s . F i r s t to get
sufficient types of psf such that meet the two following requirements:
1) every set of psf can be decomposed in a l i n e a r combination
of these types ;
2) it will be possible f o r each type to construct a virtual s t r e s s
t e n s o r in ae.It turns out that, for a cross-section like . Q, a l -
r e a d y five types of psf a r e sufficient, each type depending on one a r b i -
t r a r y function and possibly s o m e r e a l constants. The s t r o n g e s t condition that
each type of $1 = T 3 ) , where the
( f 1 T I , Ij2T2, f
A. Dou
f1 , f2 , t3 a r e the three a r b i t r a r y constants, must satisfy ,

is that
as can be conjectured from (12 ) .

In the second s t e p we consider e a c h type of psf T ( ~ ) . F o r
each T ( ~ )t h e r e is a corresponding stress tensor % (g) in R(
and for each 2 ( g ) we construct a virtual s t r e s s tensor o(~)
such , that its virtual energy s a t i s f i e s f o r all 8 the following
inequality
where C is a constant depending on a lower bound lo

of f? , t?0 < e, if and only if g = 1 , that i s only if T ( ~ )is
3
odd; Wdll is any n o r m of a=( i 3 )6 R ; y isone
Lame constant ; and T (x, y) is the function of the third compo-
3
nent of the psf T ( ~ )= ( Y1 g 2 T 2 ~b. 3T3) .
In the t h i r d and l a s t step of the proof we decompose any given
psf T in a linear combination of at m o s t sixteen s u m m a n d s ,
each one of t h e m being of the five types defined in the f i r s t step.
Because of t h e linearity of the Navier and of equilibrium equations,
and of conditions (3) and (7) , the s t r e s s t e n s o r . corresponding to
7
the psf T may be decomposed a l s o the s a m e l i n e a r combination of a t -
m o s t sixteen corresponding s t r e s s t e n s o r s .
Then we apply the second s t e p of the proof to each summand
of the decomposition and t h e r e f o r e we get un upper bound of the total
e l a s t i c energy of the cylinder Re in t e r m s of the s q u a r e s of the
n o r m s of the third component of each summand. It turns out, that the
s q u a r e of the n o r m of t h e third component c a n b e always estimated
A. Dou
by m e a n s of a l i n e a r combination of the s q u a r e s of the n o r m s that

a p p e a r in the inequality (4) . Therefore, t h e inequality is proved.
3. Related ,questions
a) When we stated the inequality ( 1 4 ) , we said, that, if and on-

l y if the psf T are even , that is of the f o r m T ( ~ ), then
the constant C of (14) does not depend on a lower bound of e.
On the c o n t r a r y , if the given psf T a r e odd in (4) , then C
in (14) o r K in (4) does depend on a lower bound
0
of e.
(See the l a s t formula of page 91 in [5] ) .
I call the attention on the physical interpretation of this r e s u l t
and how intuitive it looks.
b) F r o m inequality (4) i t follows that if Z is a s t r e s s ten-
s o r in the infinite cylinder R , without e x t e r n a l f o r c e s , and if
03
is bounded in R , then 5 is an e l e m e n t a r y soiution of Saint-

0
Venantts type. T h i s c o r o l l a r y of (4) may be called a weak principle

of Saint-Venant. I recall that this r e s u l t is c l o s e t o the above men-
tioned c o r o l l a r y of the principle of Saint-Venant.
It would b e interesting to know
if t h e r e e x i s t s in R a coun-
2 m
t e r e x a m p l e of t h e principle , allowing
1 3 4 , d ( x y, z )
C 1 t o diverge
oc.1 3
when ) z ( + m , in such a way that a t the s a m e t i m e Z Y, ")I
i=l
r e m a i n s bounded.
c ) The mentioned r e s u l t s of Toupin and Roseman, and a l s o the
proof of (4) , imply that in the cylinder R the s t r e s s e s decay
e
A. Dou
exponentially .This assertion about the I1decayingN i s not n e c e s s a -

rily t r u e in the neighborhood of the b a s e s of the cylinder, no m a t t e r
how long the cylinder may be.
However, it s e e m s , that f o r any given psf T it i s possible to
aproximate T a s much a s d e s i r e d by psf T
* in. such a way, that
Y
the corresponding stress tensor t does take the maximum value in
the b a s e s of the cylinder Re , provided that Re is sufficient-
ly long.
d ) Dividing the Navier equations (6) by /U , they take the f o r m
grad div
All the stated r e s u l t s a r e valid for t h e Poisson's r a t i o C ,

such that - 1 < U < 112 , although I do not know of any p r a c t i c a l
elastic m a t e r i a l with ct negative.
One consequence of the proved weak principle of Saint-Venant

i s that in R t h e r e a r e only e l e m e n t a r y solutions of Saint-Venantls
M
type f o r Q such that -1 < U < 112 .

But it has been proved by t h e a u t h o r , [7] , that if
6 = -1 - E , E positive and s m a l l , then t h e r e a r e , in the infinite cylin-
der with a c i r c l e for c r o s s - s e c t i o n and without e x t e r n a l forces,
periodic in z solutions of (16) , s o that the principle of Saint-
Venant does not hold for Q < -1, in s p i t e of the fact that the
Navier equations r e m a i n elliptic.
de Saint-Venant, B . , M e m o i r e s de l l A c a d e m i e d e s Sciences d e s
[I;
s a v a n t s & r a n g e r s , 14(1855) , 233-560 and Memoire s u r l a flexion
d e s p r i s m e s , Journal de Liouville, S e r . 2, (1856) , 89-189.
Toupin, R. A. , Saint-Venantls principle, A r c h . Rational

[ 21 Mech. Anal. 18(1965) , 83-96
Roseman, J. J . , A pointwise e s t i m a t e for the s t r e s s in a cylinder

[3)
and i t s application to Saint-Venant's principle, - Arch. Rational
Mech. Anal . 2 l(1965) , 23-48 .
(41 Dou, A. , On the P r i n c i p l e of Saint-Venant, MRC Thecnical

Summary Report f 472, May 1964.
DOU, A. , Upper E s t i m a t e of the Potential Elastic E n e r g y of a

151
Cylinder, Comm. P u r e Appl. Math. 19(1966), 83-93.
Bramble, J . H. , and Payne, L. E . , *-Pointwise bounds in th_e

[ 61
f i r s t biharmonic boundary value problem, J. Math. and P h y s .42
--
(1963) , 278-286 .
[ i] in Dou, A., Bounded solutions of the Elasticity equations in the -

the unbounded cylinder, Proceedings of the Inter. C o n g r e s s of
Math. , Moscow, 1966. A b s t r a c t s of s h o r t co~nmunications.
(C. I. M. E. )
TODDDUPONT
ON THE EXISTENCE OF AN ITERATIVE METHOD FOR THE S9LUTION OF

ELLIPTIC DIFFERENCE EQUATIONS WITH AN IMPROVED WORK ESTIMATE
Corso tenuto ad Ispra 3-11 luglio 1967

On the Existence of an I t e r a t i v e Method f o r the Solution of
E l l i p t i c Difference Equations with an Improved Work Estimate
Todd Dupont
Rice University, Houston
1. Introduction
The i t e r a t i v e s o l u t i o n of the d i f f e r e n c e equations associated
with V + ( a ( x ) ~ u )= f i s a problem which has received a g r e a t d e a l of
a t t e n t i o n i n the l i t e r a t u r e . I n t h i s paper i t w i l l be shown t h a t
f o r r a t h e r g e n e r a l domains i n the plane t h e r e e x i s t s an i t e r a t i v e
method with work e s t i m a t e s which a r e b e t t e r than those now a v a i l a b l e .
This i s an a b s t r a c t existence theorem, not a complete s p e c i f i c a t i o n
of an improved i t e r a t i v e procedure.
On a r e c t a n g l e with a ( x ) constant the Peaceman-Rachford [ 5 1
procedure gives a work e s t i m a t e of ~ ( h - ' l o g h-'1og c - l ) f o r reduction
2
of t h e L -norm of the e r r o r by a f a c t o r ;. S t i l l on a r e c t a n g l e but
f o r more general equations the procedure of Gunn [31 g i v e s the same
work estimate f o r the reduction of the analogue of the D i r i c h l e t
i n t e g r a l of the e r r o r by a f a c t o r c . Gunn's procedure a l s o gives
an estimate of O(h- 2 (log h- 1) 2log c - l ) f o r reduction of t h e uniform
norm of t h e e r r o r by a f a c t o r c again on a r e c t a n g l e . We s h a l l show
t h a t , i f the domain i s the union of r e c t a n g l e s with s i d e s p a r a l l e l
t o the coordinate axes and i f a ( x ) is twice continuously d i f f e r e n t i a b l e
i n the c l o s u r e of the domain, a work estimate of the form
O(h 2 ( l o g h")'log c - l ) can be obtained f o r the reduction of t h e
uniform norm of the e r r o r by a f a c t o r s .
I n s e c t i o n s 2 and 3 we d e f i n e t h e problem and the i t e r a t i o n
and give a n a n a l y s i s of t h e work required t o produce t h e s o l u t i o n .
This a n a l y s i s i s done assuming the main lemma of t h e paper, Lemma 1,
which i s proved i n s e c t i o n s 4 and 5.
2. D e f i n i t i o n of t h e Problem
We w i l l work with a domain D = R1 J R2, where each Ri is a
2
r e c t a n g l e i n R with s i d e s p a r a l l e l t o t h e coordinate a x i s . &T
choice of
--- two r e c t a n g l e s i s made f o r s i m p l i c i t y , though i t i s c l e a r
--
t h a t what follows holds f o r any f i n i t e union of such r e c t a n g l e s .
Assume t h a t we have a square g r i d of mesh s i z e h covering R1 and R2
and t h a t t h e s i d e s of R1 and R2 l i e on g r i d l i n e s . de a l s o assume
t h a t 3R1 n aR2 c a D . Let Dh denote the g r i d p o i n t s i n t h e i n t e r i o r
of D, 3Dh t h e g r i d p o i n t s on aD and Dh = Dh ;aDh. Similarly, define
de a r e attempting t o solve
f o r u,, a f u n c t i o n defined on Bh, where Lhuh i s any one of the f i v e

point dpproximations (4.4) t o c . ( a h ) generated by (4. I ) , ( 4 . 2 ) , and
(4.3) and f and g a r e given f u n c t i o n s . We s h a l l assume a t o be a
2
p o s i t i v e C f u n c t i o n i n 6. I n what follows we a r e i n t e r e s t e d only
i n f i n d i n g t h e s o l u t i o n t o the a l g e b r a i c equations (although we
s h a l l use t h e f a c t t h a t t h i s i s an approximation t o t h e s o l u t i o n of
t h e d i f f e r e n t i a l problem).
3. D e f i n i t i o n and ~ n a l y s i sof the I t e r a t i o n
The i t e r a t i o n t h a t w i l l be examined here i s a nested one. The
o u t e r i t e r a t i o n c o n s i s t s of a Schwarz a l t e r n a t i o n , and the inner
i t e r a t i o n c o n s i s t s of a Gunn i t e r a t i o n . The outer i t e r a t i o n i s i m -
p l i c i t and the equations involved i n i t w i l l be approximately solved
by the Gunn procedure. The o u t e r i t e r a t i o n i s defined a s follows.
Let uiyh be t h e s o l u t i o n of
where j = 1 i f i i s odd and j = 2 i f i i s even. L i s the r e s t r i c t i o n

j ,h
t o R j , h of Lo. As we intend t o solve f o r u ~ by , a Gunn
~ iteration,
we need t o s p e c i f y an i n i t i a l guess on R which we tdke t o be

j ,h
u i f i 2 2 and u ~ i f , i =~ 1.
i-2, h
We s h a l l assume the following lemma f o r the present and proceed
with the a n a l y s i s of the i t e r a t i o n .
Lemma 1. There e x i s t s q depending only on R1,R2 and a such t h a t , if
then maxw h r q h s q c 1 , where j ' j .

The Domain The Boundary Values
For example, i f D i s a s shown i n t h e f i g u r e and wh i s d e f i n e d

on K l , h 9 L l , h w h ' O On R l , h , and t h e boundary v a l u e s of wn a r e a s
shown, t h e n wh r q < 1 on t h a t p a r t of bR i n s i d e R1.
2,h
Let w ~ be, t h~e c a l c u l a t e d approximation t o u ~ , ~Then.
L e t t i n g uh be the s o l u t i o n of ( 2 . 1 ) and qi,h = w ~ - ,uh, ~ we obtain
It should be noted t h a t Y ~ , ~ being

, , t h e d i f f e r e n c e between the t r u e
s o l u t i o n and t h e approximation t o i t , i s e q u a l t o zero on the bound-
ary. Let us adopt t h e f o l l o w i n g n o t a t i o n :
Zi = max lzi,hl 9
Rj,h'Rj ' , h
where j = 1 i f i i s odd and j = 2 i f i i s even and j # j ' . By the

maximum p r i n c i p l e and Lemma 1 i t follows t h a t
Theref ore,
Now i f we neglect round-off e r r o r , l e t z denote the norm (with

r e s p e c t t o t h e uniform norm) of the e r r o r propagator i n the inner
i t e r a t i o n , and l e t wFrUe denote the exact s o l u t i o n of ( 3 . 1 ) , we
l,h
obtain
\. s 3 max I twr u e~ -, ~
R i-2,h1 '
j, h
Thus,
v r s man / w ~ , T~i , h+ - h' h' - Wi-2,hI
R +
j ,h
(3.8)
< ;(xi + Xi_2 + vi) .
Therefore,
(3.9) vi 6
0
=(xi + .
Since
it follows t h a t
Now, (3.6) and (3.9) imply t h a t
Therefore,
This implies t h a t xi converges t o zero i f a l l t h e r o o t s of
a r e l e s s thdn one i n a b s o l u t e value. This i s the case i f , say,

0 < : < (1-q)/5.
Thus, i f z i s s u f f i c i e n t l y small (independently of h ) , t h e r e
e x i s t s v such t a t 0 r v < 1 and xn < ~ v ~ ~ ~ [ x ~ f x ~ f x ~ + x ~ ~ .
Hence, it r e q u i r e s O(1og c - l ) o u t e r i t e r a t i o n s t o reduce the uniform
norm of t h e e r r o r by a f a c t o r c .
Thus, i t remains only t o analyze t h e work required t o reduce
t h e norm of t h e e r r o r by a f a c t o r I: a t each s t e p . I n Gunn's paper
i t i s shown t h a t t h e work t o reduce t h e A-norm of t h e e r r o r on a
rectangle by a f a c t o r 6 i s ~ ( h - ~ ( l oh-g 1) l o g b-'), where the a-norm
2
is I L U A ~ U bh
= h ) ~ being
, t h e standard f i v e point approxi-
R
mation t o t h e ~ a e l a c eoperator. It i s easy t o see t h a t t h e r e
e x i s t s Co and C 1' independent of h, such t h a t
Thus, it s u f f i c e s t o take 6 = c ~ - ~ c Hence,

I. t h e work a t each s t e p i s
-1 2
Consequently, the estimate f o r the t o t a l work i s ~ ( h - ~ ( hl o )~ log . - I ) .
Theorem. Let R 1 and R2 be two-dimensional r e c t a n g l e s with boundaries

aRi c o n s i s t i n g of segments p a r a l l e l t o the coordinate axes and l e t
D = R1 U R2. Let t h e r e e x i s t a sequence (h,), 0 < hm - 0 , such t h a t
aD f a l l s on a square g r i d of mesh hm. Let aR1 ? aR2 c aD. Let
a ( x ) E c 2 ( 6 ) , a ( x ) rao>O. Then, t h e r e e x i s t s an i t e r a t i v e procedure
f o r obtaining the s o l u t i o n of (2.1) such t h a t t h e number of a r i t h m e t i c
operations required t o reduce the uniform norm of the e r r o r i n the
2
approximate s o l u t i o n by a f a c t o r c i s no more than 0 ( h i 2 ( l o g hm) log F
- 1) .
4. Proof of Lemma 1
I n t h i s s e c t i o n we s h a l l prove a s p e c i a l case of Lemma 1 which
i s s u f f i c i e n t t o imply Lemma 1 when used with the maximum p r i n c i p l e
Let R = ((x1,x2): Wxicdi). Let 0 s c < dl and suppose dl,d2 and
c a r e a l l i n t e g r a l multiples of ho > 0. Let hm = h02". If 3 c R2,
let nh =
-n fl ((ph,qh): p,q i n t e g e r s ) . Let anh be t h e p o i n t s (x1,x2)
of Eh such t h a t (xlih,x2) o r (xl,x2ih) i s not i n Eh. -
Let Oh = ih'\azh.
Let 7 1 -
U ( X ) = h- ( ~ ( x + e ~ , ~u)( x ) ) and V- u(x) = h- 1( u ( ~ ) - u ( x - e ~ , ~ ) )
Xi,h 'i, h 3
L
f o r i = 1,2, where e l = ( 1 , 0 ) , e 2 = ( 0 ' 1 ) . Let Ahu(x) = 1 Vx 7- u(x).

i=l i , h 'i, h
f o r i = 1 , 2 , d e f i n e L i , h y Li (, h~Y) L(') as foltows:
i,h
Thus, ),L! ) :L and )L: a r e t h e t h r e e most common f i v e p o i n t approxi-

mations t o ~ . ( a v u ) .
Lemma 1'. I f a i c 2 ( ~ )t n e r e e x i s t s q sucn t h a t 0 < q c 1 and such

t n a t , i f u h i s the s o l u t i o n of
on aRh r T(x1,x2): 0 :x1 < c j ,

on hRh [ ( x1 , x 2) : xl r ' c ) ,
f o r Lh = L$), j = 1 , 2 , o r 3, and h = hm some m, t h e n max u h ( c , x 2 ) ' 4.

The proof of Lemma 1' goes a s f o l l o w s .
1) For s u f f i c i e n t l y s m a l l s and f o r a l l h = hm t h e r e e x i s t s q
such t h a t U ~ ( C ,<~ q) < 1 f o r y = kh and y i0,il b [d2-:,d21.
2) For edch i > 0 there exists q 1 such t h a t f o r h s u f f i c i e n t l y
small u (c,)') < q f o r y = kh E ( ~ , d ~ - d .

h
3) Take c f o r 1 ) and h < b f o r 2 ) . Note t h a t t h e r e a r e o n l y
a f i n i t e number of h n t s 2 6 and t h a t we can a p p l y t h e s t r o n g maximum
p r i n c i p l e t o each of t h e s e t o g e t a q < 1 f o r each one. Then t a k e
q t o be t h e l d r g e s t one of t h e f i n i t e s e t g i v e n by 11, 2) and t h e

hnts 2 6.
I n c a r r y i n g out the proof of 1 ) and 2) we use t h e following:
Lemma 2. I f Ahvh = -C i n Rh, c 2 0 and vh = f and aRh, where
f E C(aR) and i s twice continuously d i f f e r e n t i a b l e on t h e closure

of each edge of R, then
where
K = [max f - min £11 m i n d . + [c + max(lDx2 f l , l D 2 £111 max d .
i=1,2 J aR 1 X2 i=1,2j
(Naturally IDxq£1 i s used on s i d e s of aR which a r e p a r a l l e l t o the
2
1
x1 a x i s and s i m i l a r l y f o r Dx2 f . )
2
Proof: The proof w i l l be c a r r i e d out f o r ox hvh. F i r s t extend
1'
*
vh t o R i and Ri, where R ' = ( ( x1' x 2 ): -d 1 s x 1 <- 0 < x2 s d2} and
R" = f ( x1 , x 2) : dl< x1 s 2dl, 0 5 x2 s d 2 ] . This extension i s made

s u b j e c t t o the c o n s t r a i n t Ahvh = -C . This extension i s c l e a r l y not
unique, but take any such extension and note t h a t ~ ~ ( . . ~ ~ ,= ~0 ? ~ ~
Thus, gax I P v I . On the
- -
inRh. ,hvhl = rnax ) L

1 1 xl, h ' x l , h h
Rh aRh
s i d e s of Rh given by x2 = 0 and x2 = d2' I v ~ l , h ' x l , vh h 1 < max IDx2 1f l .
On the s i d e s xl = 0 and xl = d l ' 'xl, hixl, hvh = - c - 'r2, h7'x2,hvh
Thus, I\l,h:xl,h~hl s c + max ID2 f 1 . Hence, t h e t o t a l change in

X2
2 2
1'; x l , h v h 1 i s n o m o r e t h a n [ c + m a x { / D x f l , l ~f l ) ] d l .
1 X2
Foreachx
2
t h e r e e x i s t xl and x i such t h a t vx h ~ h ( x , l , x 2 ) ";'(man f - min £1

1'
and vX h v h ( ~ ; , ~ 2 L) - dil(max f - min £1. So, t h e conclusion of
1'
Lemma 2 holds f o r Vx
hvh .
1'
The proof of 1 ) goes a s follows. It i s c l e a r l y s u f f i c i e n t
t o show 1 ) f o r y < s only. This i s done by f i r s t making t h e change
I
of v a r i a b l e s vh(x,y) = a ( ~ , ~ ) ' u ~ ( x , y )This
. gives nhvh = f n , wnere
3.
I f h [ < C. C i s independent of h, since f h = (nha2)uh + a - ) i ( ~ i 2 ) - ~ h ) u h .
~ 1 ~ 0 ,
for x < c on dRh ,
for x r c on aR
h '
where g i s continuous on aR, g ( c , 0 ) = 0, and f o r x < c , g ( ~ , ~ )

2 a(x,y)k - ~ ( c , o ) ' , and f o r x L c , g ( ~ , r~ 0.
)
1
It was shown by M i l l e r [47 t h a t Iv ( c , ~ I) r a ( c , ~ ) ~ a x ( ~ , c d ; ' ]
l,h
f o r a l l y. We know by Lemma 2 t h a t I V ~ , ~ ( C1. r, ~Ky.) F i n a l l y , by
Nalsh and Young [71 v tends uniformly t o v2 a s h tends t o zero,
2,h
..
where v2 i s the s o l u t i o n t o Av = 0 i n R , v = g on 3. Thus f o r h
small and y small Iv ( c , ~ 1) i s small. Hence, given 6 > 0, we can
2,h
1
take y and h s u f f i c i e n t l y small such t h a t u ( c , ~ )r max(-,cd-l) t i .
h 2 1
I f we i n s i s t t h a t the bound on y be smaller than t h e above bound on
h, we g e t t h i s r e l a t i o n f o r a l l h. The proof of 1 ) i s complete.
The proof of 2) i s c a r r i e d out a s follows. The boundary d a t a
a r e increased t o a Cm f u n c t i o n g which i s s t i l l not g r e a t e r than one
and which i s s t i l l zero f o r xl = d. Now i f we l e t Bh denote the
s o l u t i o n t o the new problem, then uh s iih and ihtends uniformly t o
i, where 6 i s the s o l u t i o n t o t h e problem V-(avu) = 0 i n R and
u = g on bR, by the r e s u l t s of s e c t i o n 5. We know by the strong
maximumprinciplethat max b(c,y)<l. Thus3q<lsuch
Y E [ €,d2-€1
t h a t f o r h s u f f i : i e n t l y smdll max uh(c,y) 2 q < 1.
YE[E ,d2-E1
The only t h i n g l e f t t o prove i s t h e convergence r e s u l t used
above f o r uh. This i s very s i m i l a r t o many such convergence
r e s u l t s but w i l l be included f o r completeness.
5. Bound on the D i s c r e t i z a t i o n Error

rilthough t h i s i s properly p a r t of the proof of Lemma l ' , we
s h a l l s t a t e i t s e p a r a t e l y a s it has some small amount of independent
interest.
Lemma 3. Let R = {(x1,x2): O c n l c d l , O c x 2 < d2) . Let a F c2(R
a > 0, and l e t g be a continuous function on bR. Let u be the
s o l u t i o n of ?.(a:u) = 0 in R such t h a t u = g on bR. Let uh be
-
defined on Rh by Lhuh = 0 i n Rh and uh = g on aRh where
Lh = L6) 6) Then max Iuh-uI
h , h, o r Lh. - 0 a s h - 0. (Note no r a t e
can be given without more assumptipns; see Walsh and Young [ 7 1 . If

g i s assumed L i p s c h i t z continuous, then an e r r o r bound of 0(h 217)
could be obtained. )
This proof i s modeled on t h a t of Pucci :61 which i s q u i t e
s i m i l a r t o t h a t of Bers [ I ] .
Proof of Lemma 3:
1) Note t h a t i f Lhvh 2 0 (or 5 0 ) i n Rh, then pax vh = ma* vh
h aRh
(or min vh = min v h ) .
8h aRh
2) I f M ~ m a x ( m a x ( I a I t l a I j . m a x a f (max(Ia l ' + l a x 2 ~ ' ) ) )
R X1 X2 R R X1
and B 2 M +2+ (rnin a)-1, then t h e r e e x i s t s

R r
ho such t h a t 0 < h ho
imp 1i e s
Odl Bx Odl Bxl
Lh(e - e l) = L
1 ,h
(e - e ) r - 1 f o r L~ =
,)!,'L : :L o r L0)
~ .
3x
Proof: ~ : , ~ ( e')(x1,x2).= a(x ,x ) V V- eBX1 + 42[ a x1 ( 5 1 , x 2) o ~ ~3X1,
Xl,hXl,h
+ ax (e2,x2)yx e 3Xl,
1 l,h
0 (xl-h)
z (min a ) [ e (a2 - 2 I
R
f o r h small. L and LO) can be t r e a t e d s i m i l a r l y .
l,h l,h
3) There e x i s t s h such t h a t , i f Rh c
* \, -
h < ho, and Lhvh = f
*
i n (Rh) , then mgx /vhl 2 mag lv,l +K mgx [ £ I .
\ aRh (5-2
Proof: Let wh=rnag lvhl + (e .e ( 71 . Then
aRh
t * --
Lh(whTvn) i 0 i n ( R ~ )and wh*vh 2 0 on 3% . Therefore, lvh 1 5 lwhl .

Thus, 3) holds i f we take K = e - 1.
4) Let R: = Rh n R 6 , "X ((x ,x ) c R: d i s t ( ( x l , n 2 ) , a R ) 2 5).

1 2
* h1/5
Let Rh = Rh . Then Lh(u-uh) = 0(h1I5) i n Rh and the constant i s
t
independent of h.
Proof: Schauder i n t e r i o r e s t i m a t e s (see [21) imply t h a t t h e r e e x i s t s

such t h a t lux (x1,x2) I + Iux2(x1,x2) I < Kh-lI5 and
1
if I(x1,x2) - (xi,xi)l 6 2h and ( x 1 , x 2 ) , ( x i x i ) E %* . Now l e t t i n g

hL
a(xl i h,x2) = a(x1,x2) i ha (x1,x2) + -a (4*,x2), we get
1 *lXl
= ( I ) t (11) t (111) #
1 2
(I): Using u(xlih,x2) = u(xl,x2) i huX1 (xl ' x2 ) + ~ uxh 1 1 .*' 2 ) '
( ? t
we see t h a t vx o-
1,h X l , h
u(x1,x2) =
l'xlxl
(x1,x2) t 0(h1l5) i n RE .
(11): Using the above expression for u(xl*h,x2), we see that
u(xl+h,x2) - u(xl-h,x2) = 2h ux(x1,x2) + O(h2t115) i n R~
-k
.
(111): These terms a r e 0(h4I5) i n <.
Hence Lh(1) U(X1, X 2) l,h
+
2,h
,X
1 2
) = 0(h1I5) + 7.(apu) = 0(h1/5)
Jc
i n Rh. Now, (a(x1,x2) -( X1t 1'x 2 )'1)2 = a(x1,x2) - 2(a(xl,x2)a(xl+h,x~)'l
5 chL . Using t h i s r e l a t i o n and the Schauder estimates, we get
( 2 ) u = LL1)u + 0(h4I5)
*
= 0(h1I5) i n Rh , Next using a(xl
h
+T,~2)
Lh
- a(x1,x2 + a(xl+h,x2) -f O(h2 ) exactly as above, we get Lh( 3 ) i

2
Proof: For (x1,x2) c Rh\Rh
* 0 0
l e t (x1,x2) denote t h e point on
0 0
aR which i s c l o s e s t t o (xl,x2). Note t h a t (x1,x2) E aRh and a l s o
U ( X ~ X- ~uh)(X1 , x 2)I Iu(x1)x2) - 0

u(x1)x2)
0
I ' Iu(x1,x2)0 0
- uh(x1,x2) I
0 0
and u(xl,x2) ' uh(xlYx2)
0 0
T ~ s ,lu(xl>x2) - u ~ ( ~ ~ ) x ~ ) I
i Iu(x1,x2) - 0
I + Iuh(xl)x2)
u(x1,x2)
0 0 0
- I . TO estimate the
f i r s t term n o t i c e t h a t I (x1,x2) - x x 1 h t h 'I5 2h1I5, h < 1,
o o 1/51 where a i s the modulus
and t h e r e f o r e lu(xl,x2) - u(x1,x2)I w(2h L
0 0
of c o n t i n u i t y of u. I n o r d e r t o estimate / U ~ ( X ~ )- Xuh(x1)x2)
~) I we
again make t h e change of v a r i a b l e vh(x1,x2) = ~ ( x ~ , xh (X
~ 1), X
~2 )u.
3-
Then, a s b e f o r e , Ahvh = f h , where f h = (nha2)uh - (a-')(,Li)- L ~ ~ ) .) u ~
Now, e s t i m a t i n g crudely, we g e t I f b [ i C , C independent of h. Next,

-
h' - 'l,h t V 2 , h 9 where
+ X ~ )0 0
I V ~ , ~ ( X ~- ,v2,kx1,x2) I I. By Lemma 2 the second term i s bounded by
ch1I5. Thus, i t remains only t o e s t i m a t e I Y ~ , ~ ( X ~ , vX~ ~, ~) ( x ~ , x1 . ; ) -

This i s done by noting t h a t v converges uniformly t o the s o l u t i o n
l,h
5
v1 of Av = 0 i n RO and v = a 2 g on 3R which i s continuous i n R . dence,
0 0
we get 1vl,h(x1)x2) - 'l,h (x 1, x 2 )I + 0. This proves 5 ) .
Combining 31, 4 ) and 5) we g e t max lu-uhl

6) - 0 as h - 0.
Rh
This completes the proof of Lemma 3 .
6. Remarks
1) It seems q u i t e probable t h a t t h e arguments used here can

be extended t o n dimensions and somewhat more general equations
1 2
with a work estimate of O(h-"(log h- ) log c - I ) .
2) I n the proof of Lemma 1 we were only a b l e t o prove the

existence of the q and not estimate i t . It i s , of course, t h i s
f a c t which makes t h i s procedure only an a b s t r a c t existence theorem.
It i s s u f f i c i e n t , by arguments i n M i l l e r ' s paper, t o f i n d q f o r
c = 1ply and i n t h i s case we can g e t a good estimate f o r q depending
only on the constant of e l l i p t i c i t y f o r what seems t o be the worse
possible case. That i s , of course, the case where a(xl,x2) = max a
f o r xl *
c , a ( x x ) = min a f o r x 2 c . I n t h i s case
1' 2
-
q = (max a + min a ) 1max a . However, I have been unable t o show
t h a t t h i s i s t h e worst case and it seems t o be a q u i t e i n t e r e s t i n g
problem.
3) I g r a t e f u l l y acknowledge the encouragement and a s s i s t a n c e

of J i m Douglas, J r .
References
1. L. Bers, On mildly nonlinear p a r t i a l d i f f e r e n c e equations

of e l l i p t i c type, J . Res. Nat. Bur. S t . , u ( 1 9 5 3 ) ,
229-236.
2. A. Friedman, P a r t i a l D i f f e r e n t i a l Equations of Parabolic Type,
P r e n t i c e Hall, New York, 1964. (Chapter 3, Section 8 ) .
3. J. E , Gunn, The s o l u t i o n of e l l i p t i c d i f f e r e n c e equations by

semi-explicit techniques, J . Soc. Indust. ~ p p l .Math.,
Numerical x n a l y s i s , Ser . B, 2(1964), 24-45.
4. K , M i l l e r , Numerical analogues t o t h e Schwarz a l t e r n a t i n g

procedure, Numerische Mathematik , l(1965 1, 91-103.
5. D . W. Peaceman and H. H. Rachford, The numerical s o l u t i o n of

parabolic and e l l i p t i c d i f f e r e n t i a l equations, Journal
of the Society of I n d u s t r i a l and Applied Mathematics,
-
3(1955), 28-41.
6. C . Pucci, Some Topics i n Parabolic and E l l i p t i c Equations,
University of Maryland Lecture S e r i e s , No. 36, 1958.
7. J . L . Walsh and D. Young, On the degree of convergence of

s o l u t i o n s t o d i f f e r e n c e equations t o the s o l u t i o n of
t h e D i r i c h l e t problem, J . of Math. and Phys., 2 ( 1 9 5 4 ) ,
80-93.
(C. I. M. E. )
J. R. CANNON and J I M DOUGLAS
T H E APPROXIMATION O F HARMONIC AND PARABOLIC FUNCTIONS ON

HALF-SPACES FROM INTERIOR DATA
C o r s o t e n u t o a d Ispra d a l 3-11 L u g l i o 1967

The Approximation of Harmonic and Parabolic Functions on Half-Spaces
from I n t e r i o r Data
J. R. Cannon (Minneapolis) and J i m Douglas, J r . , (Chicago)
1. I n t r o d u c t i o n . The o b j e c t of t h i s paper i s t o d i s c u s s the

numerical approximation of functions e i t h e r harmonic o r parabolic
i n a half -space given t h e approximate values of the functions on
some subset of t h e open half-space. Two t h i n g s a r e immediately
obvious. One i s t h a t , without some knowledge of the g l o b a l be-
havior of the f u n c t i o n s , nothing can be s a i d about the functions i n
t h e whole half-space. The second i s t h a t , i f t h e r e e x i s t s one
f u n c t i o n extending t h e d a t a t o the half-space, i n general t h e r e
a r e i n £ i n i t e l y many. Throughout t h e paper t h e n o t a t i o n i s based
on approximating any one of t h e s o l u t i o n s ; i t i s c l e a r t h a t t h e
e r r o r e s t i m a t e s apply e q u a l l y w e l l t o every s o l u t i o n .
The g l o b a l c o n s t r a i n t imposed on harmonic functions i s a con-
venient g e n e r a l i z a t i o n of boundedness (with a known bound) on the
h a l f - s p a c e ; i n t h e p a r a b o l i c case the s o l u t i o n s a r e assumed non-
negative. These d i f f e r e n t assumptions lead t o minor d i f f e r e n c e s
i n t h e arguments and r e s u l t s ; however, teach assumption could be
employed i n e i t h e r c a s e . Two cases a r e t r e a t e d f o r t h e measurement
of t h e d a t a . F i r s t , i t i s assumed t h a t t h e f u n c t i o n i s measured with
a known accuracy on an e n t i r e hyperplane p a r a l l e l t o t h e boundary
of t h e h a l f - s p a c e . Later t h e d a t a a r e measured with a prescribed
accuracy on a r e c t a n g l e on t h e same hyperplane.

- - - - - - - --
This r e s e a r c h was supported i n p a r t by t h e National Science

Foundation and t h e Air Force O f f i c e of S c i e n t i f i c Research.
When the d a t a a r e given on a l l of t h e hyperplane, i t i s
proved i n both the harmonic and p a r a b o l i c cases t h a t an approxi-
mate a n a l y t i c continuation can be found i n a p r a c t i c a l manner and
t h a t the accuracy of the approximation on any s p e c i f i e d rectangle
i n the open half-space depends i n a Holder-fashion on the accuracy
of the measurement and the amount of computation involved i n the
d i s c r e t e procedure. When the data a r e given on a compact subset
of the hyperplane p a r a l l e l t o the boundary of the half-space, the
s i t u a t i o n i s somewhat d i f f e r e n t . I n the harmonic case the dependence
of the approximate s o l u t i o n on the d a t a i s very weak; i n f a c t , the
a p r i o r i estimate of the e r r o r derived h e r e i n i s of such a nature
a s t o preclude p r a c t i c a l c a l c u l a t i o n . We do not know whether the
estimate can be improved; however, we do show t h a t , i f t h e d a t a
surface i s r o t a t e d ever so t r i v i a l l y , no a p r i o r i e s t i m a t e i s
feasible. Indeed, even uniqueness f a i l s . In the parabolic case
f o r d a t a on a compact subset of the hyperplane, our a p r i o r i e s t i -
mate of t h e e r r o r i s again weaker than t h a t f o r d a t a on a l l the
hyperplane, but the e s t i m a t e does i n d i c a t e t h a t t h e computing problem
i s practical. No example i s known t o us t o i n d i c a t e t h a t t r i v i a l
r o t a t i o n of t h e d a t a s u r f a c e should lead t o f a i l u r e ; t n e autnors
disagree a s t o the l i k e l i h o o d of such a f a i l u r e f o r parabolic functions
Harmonic f u n c t i o n s i n two v a r i a b l e s a r e t r e a t e d a t considerable
length i n s e c t i o n 2 and t h e r e s u l t s f o r two v a r i a b l e s a r e extended
r a t h e r e a s i l y t o s e v e r a l v a r i a b l e s i n s e c t i o n 3. Solutions of t h e
heat equation a r e t r e a t e d analogously i n s e c t i o n 4.
A number of the r e s u l t s i n t h i s paper have been presented
[ 2 , 3 , 4 1 i n s e v e r a l e a r l i e r l e c t u r e s by one of the a u t h o r s ; however,

no proofs have been given and the theorems here s t a t e d a r e stronger
than those given e a r l i e r . I n p a r t i c u l a r , the r e s u l t s f o r d a t a
given on a compact s e t a r e much more p r e c i s e here. Some r e l a t e d
numerical experiments were reported i n [I].
2. Harmonic Functions i n Two Variables. Let u be a harmonic
function i n t h e half-plane (y>O] having t h e r e p r e s e n t a t i o n
where
-1 2 2 -1
(2.2) P(x,y)= n yCxty1
and u i s a signed measure such t h a t
where p = ut - u- i s t h e decomposition of u i n t o t h e d i f f e r e n c e of
nonnegative measures. The c o n s t r a i n t .(2.3) i s t h e global
r e s t r i c t i o n imposed upon u t o f o r c e continuous dependence of u
on t h e d a t a t o be s p e c i f i e d l a t e r .
Assume f i r s t t h a t u i s known approximately on t h e l i n e
where g(x) i s continuous. L a t e r t h i s condition w i l l b e modified.

It i s c l e a r t h a t
t h u s , only t h e values of u on the s t r i p (O<y<Y) a r e of i n t e r e s t .

The f i r s t question t o be discussed w i l l be t h e approximation of
u on t h e s t r i p {6sy<Y), b>O.
Note t h a t , i f msx+Xsmtl,
mtl kt1
I
xtX
P(x-r,Y)dp(~)t 1
k=mtl
P(x-~,y)du(~) I
< ~ ~ - ' Y M [ x I - ' ,~ 2 2 ,
by (2.3) . Thus,
where C denotes a constant depending on Y. Let xi = F~ = ih, i=O,

*1, + 2 , . . . , where h = n-l ,n an i n t e g e r . I f m and q a r e integers
such t h a t msxi<m+l and qsxi+X<q+l, then
xi+x mtl q-1 k+l
, P i FY ) ) = ,[ P(xi- F ,Y)du(?)+ L ,f P(ximF, Y ) ~ w ( ? )
X k w l k
i Xi
(2 .a) xi+x
+ .r P(xi-F , Y ) d v ( ~ .)
4
Then, from (2.3) it follows t h a t
mtl
I ,r P ( x ~ - F , Y ) ~ w (-F )i P(x.-F. , Y ) ~ ( (, FF ~~ + ~1 ) )
1 J
X x. se . < e l
i 1 3
(2.9)
s max a P (xi-E ,Y) I.

ks~sktl
- 200 -
Since Px(x ,y)= - 2n"xy(x 2+y 2)- 2

,
L
Hence, it follows from (2.7)- (2.11) t h a t
since
xi-^ ,Y) ( LC.

m
(2.13) max IT3 P (xi-? ,Y) I+ C mar I-3P
xi?;~5mtl k-1 kscrk+l "
L e t A denote t h e c l a s s of a l l sequences [(ai ,bi)) = a
such t h a t
C (a.+b )sM, n = 0 , + l ,k2, ....

rise .<n+l J j
J
For a € A s e t
The sequence (ai] r e p r e s e n t s t h e d i s c r e t i z a t i o n of t h e unknown

measure u+ and (bi) t h a t of p ; n o t e t h a t , a s a consequence
of (2.7) , t h e Poisson i n t e g r a l can be truncated w i t h a
predictable e f f e c t on t h e s o l u t i o n . Now, f i t u(x,y;a) t o t h e
data on ( y = Y) a s follows.
- 201 -
Set
(2 :16) e(a) = sup l u ( x i , ~ ; u ) - g(xi) I ,a € A.

-w<i<w
For a E A t h i s supremum i s f i n i t e , s i n c e both u(x,Y;a) and

g(x) a r e bounded a s a r e s u l t of (2.3) and i t s d i s c r e t i z a t i o n
(2.14). Let
since +
i ( P ( r F j , c j + l ) ) , i ( [ , F~ ~ ~+ ~ ) ) E) )A , (2.4) and (2.12)
imply t h a t
What i s wanted i s a s o l u t i o n of t h e minimizdtion problem given

by t h e equation e(a) = [ f o r some a E A ; b u t , s i n c e t h i s i s an
infinite-dimensional l i n e a r programming problem, i t i s not a t a l l
clear that a solution exists. However, (2.12) implies t h e existence
of a ' € A such t h a t
although i t i s not obvious how one would o b t a i n such an a ' .

For t h e moment assume t h a t an a ' has been found and s e t
The function v i s then t h e approximation t o u i n the s t r i p

{O<~<Y
; ]consider now t h e question of how good an approximation
i t is t o u. Note t h a t s i n c e terms appear and disappear from
v a s x moves from -at0 -,consequently, v i s n e i t h e r continuous
nor harmonic. L e t
(2.21) z(x,y) = c P(x-4 ,y)(a.-b.) , { ( a . b . ) ) = a ' ;
-a<?, <a j J J J' J
J
z i s both continuous and harmonic i n {y>07. Moreover, i t
follows from (2.6) t h a t
(2.22) Iz(xi,Y)-v(xi,Y) 11 CM[XI-l,

i = 0 1 , *2,. ..
From ( 2 . 4 ) , (2.19), and (2.22),
The g l o b a l c o n s t r a i n t s (2.3) and (2.14) imply t h a t ux and zx
a r e bounded on {y = Y] by a m u l t i p l e of M depending only on Y .

Hence,
The bounding of z-u i n t h e s t r i p { O < y < ~is

) facilitated
by t h e following lemma.
Lemma 1: Let w(x,y) be harmonic f o r y>O and l e t w have the
r e p r e s e n t a t i o n (2.1). Let 1 1 ~ ([n,n+l))IM,
1 n = 0,*1, k2,. .. . If
I w(x,Y) I < E and rp0, t h e r e e x i s t 8
0
>O and a constant C such t h a t
f o r 0< c < cO.
Proof: It i s easy t o s e e t h a t i t i s s u f f i c i e n t t o bound w(0,y).

Let O<b<Y. Since
- 204 -
Then, g i s a harmonic function t h a t i s nonnegative on t h e
segments { ( x , ~ ) : - ASXs8,y a 6 1 and E(x,y): x = * 0 , bsysY+b), and
(2.34) g(x ,Y+b) 2- l o g
A harmonic minorant f o r g i s given by t h e harmonic function vanishing
on t h e t h r e e lower s i d e s of t h e r e c t a n g l e and equal t o y c o s 1 ~ x / 2 ~

on t h e upper s i d e . Thus,
(2.35) g(x , Y ) > -

y ~s i n~h 9
~ E, 9
si n h g
and
s i n h n(y- 6) 128
-1 -1
(2.36) i f (iy) 1 a [11+4n 6 8)r ~ i n hnY/20
K
1 2 2 2
Since sioh qxlsinhqy = xy-l[1+6- q (x -y )+. . .I,
provided t h a t
Choose 9 s o t h a t t h e second i n e q u a l i t y holds. Then choose
6 c min(q,kYq). F i n a l l y cnoose co so t n a t Beo M. Tnen,

K s c0nst.M dnd i t follows t h a t
1-.pl $7
(2.39) Iw(O,y) I s l f ( i y ) Is CM , T ~ Y ~ Y ,
and t n e lemma nas been proved.
The r e s t r i c t i o n t o small c i s n o t important, but i t is o n l y

i n that case t h a t the r e s u l t is useful. The t r i v i a l example
- 205 -
u = e - Y s i n x shows t h a t t h e lemma i s c l o s e t o b e s t
possible.
The a p p l i c a t i o n of Lemma 1 t o t h e f u n c t i o n z-u i s
immediate, s i n c e
It f o l l o w s from (2.24) and (2.40) t h a t
R e c a l l t h a t z d i f f e r s from v ( x , y ) = u ( x , y ; a l ) , t h e approximate
harmonic c o n t i n u a t i o n , by n o t more t h a n c M [ ~ 1 - ' a l o n ~t h e l i n e {y=Y).
It i s e a s i l y s e e n t h a t
Thus, t h e f o l l o w i n g theorem h a s been proved.

Theorem 1. L e t a ' E A be a n y s o l u t i o n of (2.29) and d e f i n e
~ ( x , ~ ; a by
' ) (2.15). I f q > 0 and i f c , xel, and h a r e
s u f f i c i e n t l y s m a l l , t h e r e e x i s t s a c o n s t a n t C such t h a t
provided t h a t u has t h e r e p r e s e n t a t i o n (2 . l ) and ( 2 . 3 ) and (2.4)

hold.
Note t h a t t h e e r r o r e s t i m a t e i s i n terms of t h e g l o b a l
constraint M , t h e measurement e r r o r L : , and t h e parameters h and X
of t h e approximation. Obviously, t h e d a t a need have been known
o n l y a t t h e p o i n t s (xi,Y), b u t t h i s i s a t r i v i a l r e d u c t i o n .
As was remarked on e a r l i e r , t h e infinite-dimensional l i n e a r
programming problem (2.17) imposes s e r i o u s p r a c t i c a l l i m i t a t i o n s .
I f an approximation t o t h e s o l u t i o n u i s d e s i r e d on t h e r e c t a n g l e
R = { / x 1 <XI, v<ygY] , then i t seems i n t u i t i v e l y c l e a r t h a t the
values of ai and bi assigned a t p o i n t s (xi,O) a t a g r e a t d i s t a n c e
from R should have a n e g l i g i b l e e f f e c t on t h e approximation i n
R and, consequently, could be s e t equal t o zero. A f i n i t e linear
programming problem would r e s u l t . The i n t u i t i o n i s c o r r e c t , but
t h e proof seems n o n t r i v i a l . Two arguments, both complicated, w i l l
be presented. The f i r s t method w i l l g i v e a b e t t e r e r r o r e s t i m a t e ,
but t h e harmonic conjugate w i l l be introduced, l i m i t i n g t h e
argument t o two-dimensional problems. The second argument a l s o
i s based s t r o n g l y on complex a n a l y s i s , but the a n a l y t i c function
a r i s e s from extending an independent v a r i a b l e t o t h e complex
domain. This method of a t t a c k can be a p p l i e d t o harmonic functions
i n s e v e r a l v a r i a b l e s and t o s o l u t i o n s o f t h e h e a t equation. One
pays f o r t h e g e n e r a l i t y by obtaining weaker error estimates.
During t h e argument t o be given below ' q u a n t i t i e s X2 ,X3 ,X4,
and X 5 , i n a d d i t i o n t o t h e X1 of t h e d e f i n i t i o n of R , w i l l be
introduced. 'Xi ,i = 2 , . ..,5, w i l l tend t o i n f i n i t y i n obtaining
t h e e s t i m a t e s , and t h e following r e l a t i o n s w i l l hold: X1< X5< X4<
X3 = X4 + X2. More p r e c i s e requirements f o r t h e s e terms w i l l
appear l a t e r .
Retain t h e r e p r e s e n t a t i o n (2.1) , (2.3) and t h e measurement
(2.4). The e s t i m a t e (2.12) c a r r i e s over i n t h e form
where X2 > 0. Let X3 = X2 + X4, where X4 > X1. Let A now
r e p r e s e n t t h e sequences a = (ai ,bi) 1, where
Set
(2.45)
and l e t
(2.46) e(a) = max lu(xi ,Y;u)-g(xi) 1,a E A.

Ixi l cx4
t h e determmation of an a ' E A such t h a t e ( a l ) = 5 i s a standard
f i n i t e - dimensional l i n e a r programming problem. It follows

again t h a t
s i n c e t h e e r r o r i s t r e a t e d only f o r ( x i 1 X4 and the weights a r e

not s e t t o zero u n t i l [ x i [ > X4 + X2. Again s e t v(x,y) = ~ ( x , ~ ; a ' )
f o r some s o l u t i o n a ' of t h e mimimization problem and s e t
(2.49) z(x,y) = c P(x-F y ) ( a . - b . ) , [(a. b.)] = a'.
jy J J J' J
I F j la3
As b e f o r e , i f w = u-z,
C l e a r l y , w i s harmonic i n (y>O) and
W(X,Y) = . ~ P ( XY- ~Y ) ~ v ( F ) ,
(2.51) I v 1 ([n ,n+l)) 5 2M.
Since Iwx(x ,Y) I ICM, (2.50) holds f o r a l l x E [-x4,x4j with a

different C.
The remainder of t h e a n a l y s i s of w i s s i m i l a r t o t h a t given
i n Lemma 1 , but t h e l i m i t a t i o n of t h e bound i n (2.50) t o
{ l x l < ~ leads
~) to certain additional features. First, it i s clear
that
(2.52) IW(XYY)
1s CM, I x k x4.
Now l e t X1< X5< X4. Then, i f 6 > O ,
Similarly,
(2.54)
-1
J w ~ ( x , Y + ~46) I ~yl +-2CM =
- y3. Ixlsx5 .
X4 X5
C l e a r l y , Iwx(x,y) 1s C(6)M, yr6>0. The estimation of w i s reduced
t o e x a c t l y t h e same problem a s was t r e a t e d i n Lemma 1, except
t h a t i t is not s u f f i c i e n t t o look a t w (0 ,y) . The r e c t a n g l e used
t o estimate t h e harmonic function g introduced i n (2.33) should
be centered i n x a t an a r b i t r a r y x
0
E [-X4 ,X41, t h e harmonic
*
conjugate w should. vanish a t (xo ,Y+6), and 13 should b e replaced
by X5-X1. The constant K of (2.32) should be replaced by
Choose 6 s rnin(n,%Yq) and X5 such t h a t
s inhn (y- 6) /2 (X5-X1)

(2.56) -
s innrrY/2 (X5-X1) 1
Y - q , y26.
Then, choose Xi and Xi s u f f i c i e n t l y l a r g e and h ' s u f f i c i e n t l y

small t h a t
For X 2 s Xi, X4r Xi, and hsh' , i t follows t h a t

1-f + q I y -,
(2.58) Iw(x :Y) 1s CM '3 Y2T, lxIs xl.
I t i s easy t o s e e t h a t
Theorem 2 . Let a ' E A be a s o l u t i o n of t h e l i n e a r programming

problem (2.47) and d e f i n e u ( x , ~ ; a ' ) by (2.45). Let . q > 0.
Then t h e r e e x i s t X5 > Xl and Xi > 0 , Xi > X 5 and h ' > 0 such
if X2 2 xi, x4 2 x;, and h r h'. Moreover, v3 can be bounded a s

i n (2.59). In p a r t i c u l a r , t h e e r r o r i n t h e approximation depends i n a
Holder-continuous fashion on t h e measurement e r r o r s and t h e
amount of c a l c u l a t i o n performed, a s indicated by h , X 2 , X4, and
X5'
A second deviation of an e r r o r estimate can be based
on t h e extension of t h e v a r i a b l e y t o t h e complex domain,
(2.60) a = y + iy.
*
This a n a l y s i s begins w i t h t h e r e l a t i o n s (2.50) , (2.51) , and
(2.52). Note f i r s t t h a t
Consider t h e i n t e g r a l
Since
~ ( x - F ) ~ += u( x~ - ~F )~~ + (y2+y* 2)2 + 2(x-E)2 (y2- y *2 )
and 1 (x-F) 2 + y2 I2 2 y4, t h e denominator of P(x- F ,a) i s bounded

away from z e r o f o r any a such t h a t Re u > 0. Thus, t h e i n t e g r a l
converges f o r any x and any o such t h a t Re D > 0 , and w(x ,o) i s ,
f o r each x , a holomorphic function of a i n t h e half-plane ( ~ ue > 01.

- 211 -
r;'
Moreover, f o r y > 0 ,
Hold x f i x e d i n t h e i n t e r v a l [-XI ,Xll and consider t h e r e c t a n g l e Q =

?; J(
(Y r y 5 Y1, 0 < y < Y1 1 i n t h e a-plane. Then, i t follows from
(2.61) and (2.63) t h a t
= 0, u E ?Q
(2.64)
E 2Q ,C =
*
C(Y,Y1,Y1 ) .
I t then follows from (2.31)- (2.36) t h a t
where
* * *
s inh n (Y1-y ) / (Y1-Y) * *
(2.66) a(y) = O s y <Y1.
sinhfl;/(yl-y)
1
Now, consider t h e r e c t a n g l e Q ' = bay9(YtYl) ,/ y* I rY2)
* , where
0 < 5;
* *
and 0 < Y2 < Y1 and x remains f i x e d . The same r e s u l t holds
q
f o r y < 0. Then,
(CM)
1-a a 1
~4 , Y = 7 (Y+Y1) , a E "',a = a (Y2),
*
(2.67) Iw(x,o) 1s
1
C(6,1(YtY1), Y2)M, a i Q' .
*
Hence,
(2.68) (w(x,y) 1 1 [ ( c M ) ~ - ~ Y ,~lxIsx1 1
I ~ ( ~ ,) 6syq(YtYl) ,
where
*
s i n h n (y- 8) /2y2
(2.69) P(Y) i
s i n h ~ ( + ( Y + Y ~6))12~:
-
Since R = { ( x1s X1, q r y r Y) , there e x i s t a constant C and a

positive number 8 such t h a t
The remainder of the argument leading t o a theorem of the same type

a s Theorem 2 i s the same a s before, although i t i s c l e a r t h a t the exponen
R i s not so l a r g e a s before. However, the advantage of t h i s approach
w i l l be seen immediately below, a s well a s l a t e r i n several other
app 1ica t ions.
So f a r it has been assumed t h a t the data g(x) has been known
f o r a l l x. Clearly, i t i s not p r a c t i c a l t o measure g f o r d l 1 x ;
consequently, it i s both p r a c t i c a l l y useful and mathematically
i n t e r e s t i n g t o l i m i t the data t o the following:
the choice of the subscript being made t o coincide i n the argument

below with the previous X4. Again l e t us approximate u on the
rectangle R , but i t i s not assumed t h a t X1 < X4. I f the approximation
i s made a s before, r e l a t i o n s (2.43)- (2.52) remain v a l i d , although X4 i s
fixed now and not a parameter t o be chosen.
The estimation of the e r r o r w w i l l involve f i r s t the extension
of the x-variable t o the complex domain and then the y-variable.
Let .g = x+ix
* and consider the i n t e g r a l
which can e a s i l y be seen t o converge and t o be holomorphic in
x for
*
Ix I < y. In fact, i f
*
Ix I s pY, O < p < 1, then
and
Thus, t h e holomorphic function w (y ,Y) s a t i s f i e s t h e following bounds :
By t h e argument leading t o (2.36) ,

1-a(x*) a(x*) .
I W ( ~ * , Y1s) (C(P)M) Yl '
By symmetry t h e above r e s u l t holds with y replticed by 1 y f o r 1 y / r P Y.

Thus, w (.c ,Y) s a t i s f i e s t h e fallowing c o n s t r a i n t s on t h e boundary of the
half-strip i / Imr r yo<cY, R e y 2 01:
Let x be fixed so t h a t x > X4, and l e t xl > x. Then, by the
same drgument a s above,
(2.78)
n(xl-x)
@(x,xl) = sinh / sinh -
rrxl ,
2yo 2yo
The r a t i o e(x,xl) i s an increasing f u n c t i o n of xl and has the l i m i t
a s x 1 tends t o i n f i n i t y . Hence.
The argument leading from (2.50) t o Theorem 2 can be repeated

*
with the symbol X4 replclcing the parameter X4 i n the e a r l i e r argu-
?t
ment. Let X be chosen so t h a t (2.56) i s v a l i d and Let l l < X 5 < X 4 .
5
It follows t h a t the q u a n t i t i e s vl,v2, dnd v defined i n (2.501, (2.531,
3
and (2.54), r e s p e c t i v e l y , can be replaced a s follows:
The r e l a t i o n (2.58) becomes

Note t h a t t h e estimate (2.83) contains i n p a r t i c u l a r two parameters,
>t
X4 and X5, t h a t were introduced i n t o the a n a l y s i s f o r convenience
but which a r e not involved i n the c a l c u l a t i o n . Obviously, these

parameters should be s e l e c t e d so a s t o minimize the bound. The
minimum i s c l o s e l y approximated by equating the two a d d i t i v e terms
i n the bracket:
As P + xi1 + h tends t o zero,
* a(y0)
-
nX4 - log l o g ( € + xi1 + h)
2yo
1
2yo
-7 log
* - X5)
l0g(X4
nx4
Consequently,
The constant C depends on M,p, 6,yo, and 7 . The i n e q u a l i t y (2.87)

implies the following r e s u l t .
Theorem 3. Let a ' E A be a solution of the minimization problem

(2.44) - (2.47), where X4 is defined by (2.71). Let u(x,g;a? be
defined by (2.45). Then, there e x i s t positive constants C,Y, and 6
depending on tne rectangle R,X4, and M such that
provided t h a t u i s a harmonic function i n the upper half-plane having

the representation (2.1) - (2.3).
The e r r o r estimate above is not adequate f o r practical usage,
although it does e s t a b l i s h a quite weak continuous dependence
relationship. It i s perhaps of some theoretical i n t e r e s t t o consider
the asymptotic form of the estimate a s X4 tends t o zero. As X4
tends t o zero,
and becomes small quite rapidly. It should a l s o be noted t h a t

the method of analysis used t o derive (2.87) could be extended to
give a unique continuation tneorem f o r harmonic functions i n a
half -plane.
One consequence of Theorem 3 is a u~~iquenees
theorem for harmonic
functions. Let be a l i n e segment of length 2X4 centered a t (0,Y)
and making an angle 9 with the x-axis. Theorem 3 implies that a
narmonic function t h a t is bounded. on l;y>O] and is constant on r0
(i.e., Q=O) is identically constant. This i s f a l s e f o r 9 # 0. For
example, l e t La be tne l i n e determined by ra and l e t xa be i t s i n t e r -
section with the x-axis. Let
L i s a l e v e l curve of u This i n d i c a t e s t h a t t h e weak continuity
8 8'
r e s u l t i n g from Theorem 3 i s t o be expected. It a l s o implies t h a t
no a p r i o r i estimate can be obtained from data given on Tg .
3. W i c Functions i n More than Two v a r i a b l e s . The r e s u l t s
of the l a s t s e c t i o n can r a t h e r e a s i l y be generalized t o harmonic
Let x = (xl, , . , ,xn) E R
functions i n more than two v a r i a b l e s .
1 n
and y E R1. Let ~ ( x , ~ ) = c ~ ~ be( the
~ nPoisson
~ ~ tkernel~ ~ f)o r~ ~
the half -space {y>O]. Let u be a harmonic function i n the h a l f -
space {y>O] having the r e p r e s e n t a t i o n
Denote t h e cube { sicx i<sit l ) , s i an i n t e g e r , by S. Assume t h a t
(3.2) I ~ I ( s )r M , all S .
Also, assume t h a t , f o r some Y r 0 ,
Clearly, (3.1)-(3.3) a r e t h e d i r e c t analogues of (2.1)-(2.4).

Again l e t
be the r e c t a n g l e on which i t i s d e s i r e d t o approximate u. The

approximation w i l l be accomplished i n p r e c i s e l y the same fashion
a s before and the method of proof w i l l be a straight-forward gen-
e r a l i z a t i o n of the e a r l i e r proof based on extending y t o the complex
variable a = y + i y* .
Let I = ( i l , . . . , i n ) be a multi-index of i n t e g e r s , x i = i h , and
1
xI = (X , . . . ,x: ) . s e t zI = (x! . Then i t i s
j
< xJ <
il n j J
easy t o see t h a t
Let A denote the sequences a = { ( a I , b I ) ) such t h a t
aI,bI 2 0 , all I ,
a I = bI = 0 , i f m a x l ~ iI > x,, = x2. + x 4 ,
j
c (aI+bI) s M , all S .
zI ~ S
For a E A, s e t
(3.7) u(x,y;a) = C P(x-51,~)(aI-bI) ,Y > 0 ,

I'iJ - x j ICX,
j
and l e t
(3.8) e(a) = max lu(xI,y;a) - g(xI)j .

[ x i I <x,
j
Then,
and again a s o l u t i o n a ' E A of the minimization problem can be

obtained by l i n e a r programming. Set v ( x , y ) = u ( x , y ; a t ) f o r some
s o l u t i o n a ' and s e t
Then, v i s the approximation t o t h e harmoaic function u. As before,

- 220 -
v i s not harmonic, but z i s and d i f f e r s from v on R by cMXil. Thus,
i t i s s u f f i c i e n t t o estimate w (x,y) = u(x,y) - z ( x , y ) on R.
The harmonic function w s a t i s f i e s the following r e l a t i o n s :
Iv(S) I s 2M , all S ,
(3.11)
I W ( X , Y )I~ 2c t CM(X;'+~) r yl , I X ~ I r x4 ,
I g CM
IW(X,Y) , all x .
Repeating the argument beginning with (2.60) with o a s before and
Y 2 y sY1, we find t h a t
The i n t e g r a l
converges and i s holomorphic f o r Re a > 0, and I w(x, a ) 1 CM i f
6 i y < Y1. Hold x fixed and s e t Q = {YsyaYl,'Osy <Y1].

* * The
*
r e l a t i o n s (2.64) and (2.65) again hold with a ( y ) s t i l l given by
(2.66). Define Q ' a s follows (2.66). The remainder of the argument
i s unaltered and
where f~ > 0. Cleariy, X4 can be chosen so t h a t
maxlw) s C(M,R)(C+X;1t h ) B .
R
In any event, the e s t i m a t e s i n the two v a r i a b l e case c a r r y over
e s s e n t i a l l y unaltered t o the higher dimensional case.
The argument given i n the two v a r i a b l e case f o r the d a t a
known only on an i n t e r v a l can be c a r r i e d over, but t h e dependence
on the data and parameters remains weak.

4. Parabolic Functions. The approximation of s o l u t i o n s of the
heat equation from i n t e r i o r d a t a can be accomplished i n a fashion
very s i m i l a r t o the method applied t o harmonic f u n c t i o n s . It i s
s u f f i c i e n t ro consider t h e case of a s i n g l e space v a r i a b l e ; t h e
extension t o s e v e r a l space v a r i a b l e s i s a s immediate a s i n the
harmonic c a s e .
The most n a t u r a l g l o b a l c o n s t r a i n t t o apply t o a s o l u t i o n u
of the heat equation i s t h a t i t be nonnegative, s i n c e temperatures
a r e n a t u r a l l y bounded on one s i d e along with a l l the o t h e r common
physical q u a n t i t i e s s a t i s f y i n g parabolic equations. In addition,
l e t us assume t h a t u i s bounded on each half-space Tt r to > 0)
by M(to), where M(T) = 1 and M(t) i s otherwise unknown:
Assume a l s o t h a t
It follows from a theorem of Widder [51 t h a t
where
$ -xL/4t
(4.4) K ( x , t ) .= ( 4 n t ) - e
and u =
+. Now,
ntl
1 2 u(nG,'f) 2 ~(nG-t,~)dp(~)
n
Hence,
Thus, the constant M assumed i n the harmonic problem has been

c a l c u l a t e d under the n a t u r a l assumptions above. ( I f the harmonic
function had been assumed nonnegative, the same c a l c u l a t i o n would
have been v a l i d . )
It i s easy t o see t h a t
-x;/~T
(4.7) 0 s u(x,T) - , K(x-5,T)dW(r) 5 CMe
x-X2
Also,
xi+x2
(4.8) IJ ~(x~-:,~)d~(e) r - ) u ( [ F ~C,j + l ) ) I
K ( x ~ -. FT ~
xi-X2 Ix.-5.
1 3
I<x 2
s CMh.
Let A denote the s e t of sequences a = f a . ) such t h a t

3
a 2 0 , all j
j
(4.10)
aj = O , 1x.l
3
>x3=x2tx4,
and s e t
Note t h a t t h e c o n s t r a i n t (4.6) does n o t appear i n a d i s c r e t i z e d
form i n t h e d e f i n i t i o n of t h e admissible sequences A; only the
nonnegativity i s r e q u i r e d . Let
It follows from (4.9) t h a t

2
-X2/4T
(4.13) = inf e ( a ) < E t CM(h + e 1.
Let a ' = (a.'] be a s o l u t i o n of t h e optimization problem (4.13) and

J
let v(x,t) = u(x,t;al). I n d i s t i n c t i o n from the previous harmonic
c a s e s , v ( x , t ) i s not t h e d e s i r e d approximation t o u. Again, l e t us
attempt t o approximate u on t h e r e c t a n g l e R = (1x1 r XI, 6 r t 5 Ti,
noting t h a t i t i s again t r i v i a l t o e s t i m a t e u f o r t > T.
It follows from (4.13) t h a t
"
Consequently, i f X2 :
! 1,
1I s 1h; then,
Choose x 1. s o t h a t [xi-n- '2 '2
2
-X2/4T
(4.16) 1 a' r ~(#,~)-~[1+2ctC~(h+e 13 2
ns! . < n t l j
J
1 1 ; (4.16) i s t h e d i s c r e t e analogue of
i f [ n , n + l ) c ( 1x1 s X4+Z-7h)
(4.6) and i s a consequence of t h e procedure f o r 1x1 Ix44-:h. This
is the reason t h a t some v e r s i o n of (4.6) was not imposed i n t o
the d e f i n i t i o n of A . Since X2 grows r a t h e r slowly t o enable the
2
CM exp(-X2/4T) term t o become small, it i s not too r e s t r i c t i v e t o
require that
Set
(4.18)
"1 = , otherwise ,
and l e t
(4.19)
The function V i s the approximate s o l u t i o n of the continuation

problem. It follows t h a t
2
-X2/4T
(4.20) (v(x,T) - v(x,t) 1 s CMe , 1x1 g X5 .
and w(x,t) = z ( x , t ) - u ( x , t ) , then
(4.22) 0 s v ( [ n , n t l ) ) r 3M , xi1 andh s u f f i c i e n t l y small,

-x;/~T
/ w ( x , ~ ) Ir y1 = 2c + CM(e t h ) , 1x1 r X 5 '
Iw(x,T) I < CM , all x .

The e s t i m a t i o n of w on R w i l l follow the harmonic argument based
on extending y t o the complex domain. F i r s t , note t h a t
i f T < t s T1, T1 - T < ~ ( x ~ - x and
~ ) ~1x1, I X1. Let o = t + it* .
Then,
-2 *2
(4.24) IK(x,o)I s ~ l + i t * / t ~ ' ~ ( x , t ( l + tt )),
and
(4.25) w(x,o) = / K(x-!,a)dv(~) , Re 0 > 0
i s holomorphic i n o i n Re a > 0 f o r each fixed x. It follows from
(4.22) and (4.24) t h a t
* *
Let x be f i x e d , 1x1 r XI, and l e t Q = (Tit<Tl,Ort d l ) . Then,
* *
I W ( X , ~ ( T + T ~ ))I+ I I~ (CM) 1-a(t*) Y2
a(t ) ,
* * * )/(TI-T)
s i n h n(T1-t
a(t ) = 9
s i n h n ~ i (TI-T)
/
1
1 t ~ )sT2),
by the proof of Lemma 1. Let Q ' = ( + ~ S ~ ~ ( T + T , T2<T1
*I * * *.
Then, *
1-o ( T ~ a) (T;) 1
Y2 , a€a~'n(ty(TtT~))
(4.28) Iw(x,a)l s
c ,, ocaQ1 .
Thus, by another a p p l i c a t i o n of the proof of Lemma 1,
1 *
sinh ~ ~ ( t - ~ 6 ) / 2 T ~
B(t) =
s i n h n(T+T1-6)/4T2
*
In particular ,
f o r some a > 0 .
Theorem 4. Let a ' E A be a s o l u t i o n of t h e minimization problem
(4. l o ) , (4.12), (4.13), and l e t u(x, t ; a ') be defined by (4.11). Then,

t h e r e e x i s t p o s i t i v e c o n s t a n t s a , c and C depending on the rectangle
R and M such t h a t P,
The approximation of a parabolic function u from d a t a on a

f i n i t e interval,
can be c a r r i e d out i n a f a s h i o n q u i t e s i m i l a r t o t h a t f o r a harmonic
f u n c t i o n ; however, t h e dependence of the s o l u t i o n on a r e c t a n g l e

R = {brtsT, l x l r ~ on
~ ) t h e d a t a i s somewhat stronger i n the p a r a b o l i
case. Again l e t x = x + i x* and set
Since IK(x, t ) 1 < ~ ( xt ), e x *214t, it follows t h a t w ( ~t, ) i s e n t i r e
in x f o r each t > 0 and t h a t l ~ ( ~ , t )s I CM i f 0 c 4 r t , lx

*I i X1
*.
The r e l a t i o n analogous t o (2.76) i s
* * *
Iw(ix , T ) [ r (CM) l - a ( x I y1
a(x ) ,
(4.33) * *
* s i n h n(X1-x )/2X4
a(x ) = 9
s inh nxf/2x4
since (4.22) implies t h a t w i s bounded by yl on (t=T, 1x1s x 4 ) . If

0 < X2
k; *
< X. i t again follows t h a t
1'
*
Since X1 i s not r e s t r i c t e d , then the bound (4.34) can be put i n the
form
f o r any 7 > 0. The argument leading £rod (4.23) t o (4.30) can be

repeated with yl replaced by the e s t i m a t e (4.35). Hence,
a = a(R) > 0. If the q u a n t i t y i n the brackets i s approximately

minimized by equating t h e two terms, i t can be seen t h a t
Hence,
where c and C a r e p o s i t i v e . Note t h a t t h i s implies, i n p a r t i c u l a r ,

that
-1 -M
maxlwl 5 C(logyl , any M ,
R
i f y1 i s s u f f i c i e n t l y small.
Theorem 5. If a ' E A i s a s o l u t i o n of t h e minimization problem

(4.10), (4.12),(4.13) f o r f i x e d X > 0 and i f u ( x , t ; a l ) i s defined
4
by (4.11), t h e r e e x i s t p o s i t i v e constants c and C depending on
R, M, and X4 such t h a t
While the e r r o r bound derived i n Theorem 5 i s not q u i t e a s strong

a s Holder c o n t i n u i t y , i t i s nevertheless s u f f i c i e n t l y strong, t o be
of p r a c t i c a l use.
It i s c l e a r t h a t the r e s u l t s above extend t o the case of s e v e r a l
space v a r i a b l e s .
References
1. Cannon, J. R . , Some numerical r e s u l t s f o r the s o l u t i o n of the
2.
+
heat equation backwards i n time, Numerical Solutions of Non-
l i n e a r D i f f e r e n t i a l E uations, e d i t t e d by D . Greenspan, John
w a n d Sons, Inc. , ew York, 1966.
Douglas, J r . , J r . , Approximate harmonic continuation, A t t i d e l
Conve no su & e uazioni _alle d e r i v a t e p a r z i a l e , Nervi,
2 h b ~ a i bo z l o n l Crernonese , Roma .
3. , Approximate continuation of harmonic and
parabolic i u n c t i o n s , Numerical Solution of P a r t i a l D i f f e r e n t i a l
E u a t i o n s , e d i t t e d by J. H. ~ r a m a z m i Press,
c I n c . , New
-6.
4. , The approximate s o l u t i o n of an unstable
physical problem s u b j e c t t o c o n s t r a i n t s , Functional Analysis
and 0 t i m i z a t i o n e d i t t e d by E. R. C a i a n i e l l o , ~ c a d e m l cP r e s s ,
~ . h i 9 6 6 .
5. Widder, D. V . , P o s i t i v e temperatures on an i n f i n i t e rod,
Transactions of the American Mathematical Society 2 (1944)
85-95.
(C. I. M.E. )
B. E. HUBBARD
"ERROR ESTIMATES IN THE FIXED MEMBRANE PROBLEM
C o r s o tenuto ad I s p r a dal 3-11 Luglio 1 9 6 7

ERROR ESTIMATES IN THE FIXED MEMBRANE PROBLEM
B. E. HUBBARD 1)
E s t i m a t e s of the discretization e r r o r in the fixed membrane pro-

blem appear to be considerably m o r e difficult to obtain than for the Diri-
chlet problem. F o r example, the detailed e s t i m a t e s by Ssulev 6 will illu-
s t r a t e t h i s point . I shall give h e r e a particular d i s c r e t e analogue which i s
a natural one to considcr and yet one whose e r r o r analysis i s easily r e l a -
ted to that f o r the Dirichlet problem.At the s a m e t i m e we shall considera-
bly rc,ciucc the usual regularity assumptions in o u r e r r o r analyses. The r e -
s u l t s presented h e r e a r e d r a u n , f o r the m o s t p a r t , f r o m a p a p e r 2
with J . H . Bramble.
Let R be a bounded domain in E

n
. We shall approximate the
problem
with eigenvalues O < A 1 < A2 < ... and eigenfunct~ons u l , u 2 , . . . by tilose

of the d i s c r e t e problem
where R = R i U R and
h I1
a Rh , a r e those defined by Bramble 1 .
2
The o p e r a t o r s Ah is the usual O(h ) approximation a t polnts
x q ; .
For x E R I 1 we define
I
where I. =
I
0 . 110, .. 0 is the vector of 1en;:th h i n the x
direction and 0<a . pj < 1, to admit tile possibility that x + L Y . ~ I,.

1 1 1
l ) ~ l l i v e r s l t y of Maryland, supported in p a r t b,; tlie P'orscl~ungsinstitut f u r

Mathematik, E. T. I-I. Zurich.
- 234 -
B. E. Hubbard
x - P . h . a r e points where the mesh l i n e s cut

1 1
aR . If v(x) i s sufficiently
smooth then we note that
s o that, in general, the o p e r a t o r A h

defined by
i s only a n O(1) approximation to the Laplace o p e r a t o r f o r X ERh .

For V, W functions defined a t m e s h points of E , having com-
n
pact support we define
(V, WIh = hn 1 V(x)W(x)

X
where the s u m s a r e taken o v e r grid p o l ~ t t sin E:

n
.
It i s easily s e e n that p1 s a t i s f i e s the variational principle
11-2 2
Dh(v, V )+ I1 b(x) V (x)
x E R'
II
p1 = min -
V(x) = 0 (V, VIh
@ Rh
where b(x) i s defined by
T h e r e f o r e we s e e that the m l t r i x of o u r d i s c r e t e problem i s s y m m e t r i c ,

positive definite and hence, p ,U exist and 0 <p < p 2 ( .. . .
We a s s u m e the usual n o r m a l i z a t i o ~ l s
B. E. Hubbard
I uk (x)2 dx = ( U k Uk)h = 1 .
R
Multiplying the equationfor u by U. and suniming we see that

k J
Likewise, multiplication of the equatio.1 for U. by uk and summing

J
yields the second equation
- h k
U.)
J h
= -(uk. A h
U.) = p j ( u k UjIh
~h
.
while t h e f i r s t i s the d i s c r e t e analogue of Green's second identity applied
to u U. vanishing outside of Rh , Adding t h e s e equations yields
k' J
If we w e r e t o s e t j =k t h i s equation relates (Pk-d k ) to the

local e r r o r , (Auk - .
A huk) If t h i s local e r r o r w e r e uniformly
2
O(h ) then , upon proving that f o r h ho(k) <
2
we would have established that ( p - A k) = O(h ) under t h e usual
k
4
assumptions that u E C (R) .
Since the l o c a l e r r o r at points of
R' is only O(1) however, we take the following additional s t e p s . We
h
introduce t h e Green's function f o r the o p e r a t o r d h on R , which,
h
f o r p a r a m e t r i c values y E Rh , s a t i s f i e s t h e d i s c r e t e equation
B. E. Hubbard
Many properties of G (x,y) a r e discussed by Bramble

h
1 . For
example, the discrete representation formula applied to U. is
J
after using the equation for U. and the symmetry of G over the s e t
J h
Rh
Thus
(3) ( -
A Uk A hUk. UjIh = p j(Q)h. U .)
Jh
where a) is defined by
k
But this is the mapping A u - A u k - + e , of the local e r r o r to the

h k
discretization error, e(x) , in the Dirichlet problem. Just a s in the
2
Dirichlet problem we obtain the estimate Q) = O(h ) when.
4
u E C (R) from the local estimate
k
and from the inequalities

B. E. Hubbard
The number M , , is a function of the derivations of u of order

J k
l e s s than o r equal to j .
To ?rove (2) and thus complete the classical e r r o r estimate for
the eigenvalues we will make two assumptions
a ) pj dj as h-0. j=1,2, ...

(b) Ak i s simple.
Both a r e made for convenience since, (a) convergence theorems

exist for much more general domains than those which shall be conside-
red here and (b) the corresponding e r r o r estimates can be obtained for
multiple eigenvalues 2 , but with somewhat greater effort.
It follows from (1) and ( 3 ) that for given k
If j# k our assumption that pj+dj as h+O implies that
and consequently
Since the matrix of our problem i s symmetric, and positive definite the
eigenvectors {u] span the space of our discrete problem and we
have from Parsevalls identity
Then (6) implies that

E.B. Hubbard
The inequality (2) follows immediately in t h e classical c a s e s i n c e

(uk, uklh+ $ 2
.
uhdx = 1 'This fact can be shown m o r e directly by writing
R
where
s o that 8k + O as h + 0 . In fact, if we choose the sign of

uk
s o that (u U ) > 0 , then 8 k c a n be defined by
k' k h
Setting j = k in (5) l e a d s to the f i r s t inequality of the following

theorem
Theorem 1 : If is a simple eigenvalue then

k
where lut - Uk 4- m aRh

x luk(x) - Uk(x) 1 , and K
k
i s a bounded
quantity.
Proof : T o prove the second inequality we u s e (7) and (8) in the equation
- 239 -
E.B. Hubbard
T o prove the t h i r d inequality we substitute u (x) - U (x) into the di-

k k
s c r e t e Green's formula,
A s w a s mentioned in the l e c t u r e s by Bramble, G (x, y) h a s the following

h
majorant
where p ( (x = (x l 2 + a h and y n ,a,

do a r e geometrical constants.
Thus if n = 2, 3 we s e e that the r e s u l t of apllying Schwarzl inequality
to the second and third t e r m s of the above identity yields
- 240 -
E. B. Hubbard
from which the final inequality in the theorem follows. If n > 4 the
result can still be obtained by an iteration process that begins by noting
that
where j i s large enough s o that Schwarzl inequality will give bounded

quantities. Having done this we have an inequality of the form
where Q i s a bounded quantity, computable in t e r m s of known quantities.

By repeatedly substituting this inequality into itself we will obtain t e r m s
to which we can apply Schwarzt inequality. Thus the theorem i s proved.
Ah analogous theorem holds if Ak i s a multiple eigenvalue.
We shall now give some specific e r r o r estimates a s corollaries

of this theorem . The first of these gives an e r r o r estimate of the clas-
sical type.
4
Corollary 1 : If A < c and u Z C (R) then , for ilk simple,
Proof : Just a s in the Dirichlet problem it follows from (4) and esti-
I
2
mates for the discrete Green's function that (a)
k h
= O(h ) .
It i s
2
easily seen that. E (u ) = O(h ) which completes the proof.
h k
E. B. Hubbard
In the remaining two c o r o l l a r i e s we s h a l l r e l a t e the e r r o r estima-

t e s directly to the smoothness of the boundary.
Corollary
-- 2 : If 3 R € H1(2, A, y ) , i. e. if the functions which give the
equation of the boundary in local coordinates have two derivatives and
the second derivatives satisfy a Holder condition with exponent y > 0

and A then
where K i s a computable constant.
~ At, 7 I) , the c l a s s of
P r o o f : It i s known, Gunter 3 that u k H(2,
2
u € C (R) whose second derivatives satisfy a Holder condition with ex-
ponent /' E ( 0 , ) , and constant At . Using this information
and the mean value t h e o r e m for u it i s e a s i l y established that
k
where d(y) = m i n x-y . In view of the inequality

X E R
2 2
we s e e that ( @ gKh
k I . Once again E (u ) = O(h ) s o that the con-
h k
clusion follows.
Corollary 3 : L e t R be a bounded domain in E 2 whose boundary

i s composed of p analytic a r c s . L e t
n
- ,i = 1, ...,p , the interior
ai
E. B. Hubbard
angles at the corners, x. ,

1
be such that " < 2Q
- < 27d. Then
i
Proof : The work of S. Lehman 5 implies that
if we require that x. a r e greater than a distance ch from the mesh

1
lines. We can investigate the behaviour of (x) by means of the
k
following inequalities
where 0 < /3 <I$'< a.

1'
and p sufficiently small. We note that K( p) ,w
as p+ a .
1
. A direct calculation then yields for x E Sp (xi)
E. B. Hubbard
This estimate is consistent with those obtained by P. Laasonen [4] for

the Dirichlet problem and in certain respects gives improved estimates.
Clearly, [lak11 will have the form indicated in the theorem .

Likewise we can show that [ k
E (u ) < K ( E ) h 2 i u - r h2] so that
h Ic
the corollary follows. We see from (9) that luk(x) Uk(x) - Isatisfies
a point dependent estimate of the type (10) .
- 244 -
REFERENCES
------------
1 J. Brabmle : (C. I. M.E. Lectures, this volume)
2 J. Bramble and
B. Hubbard : Effects of boundary regularity on the discretization
e r r o r in the fixed membrane eigenvalue problem.
(To appear)
3 Gunter : Die Potentialtheorie und ihre Anwendung auf

Grundlagen der Mathematischen Physik B. G.
B. G. Teubner, Leipzig (1957) .
4 P.Laasonen: On the degree of convergence of discrete approxi-
mations for the solutions of the Dirichlet problem.
Ann. Acad. S i. Fenn. Series A. 246, 1-19 (1957) .
5 S. Lehman : Developments at an analytic corner of solutions of
partial differential equations, J. Math. Mech .8, 727-760
(1959) .
6 V.K. Saulev : On the solution of the problem of eigenvalues by
the method of finite differences, Translation in
English in A. M. S. Translations (2) Vol. 8, 277-287
(C. I. M. E. f
K. JORGENS
CALCULATION OF THE SPECTRUM OF A SCHRODINGER OPERATOR

- 247 -
CALCULATION O F THE SPECTRUM OF A SCHRODINGER OPERATOR
by
Let T denote a selfadjoint linear operator in a complex Hilbert

space H , U (T) the spectrum of T, u d(T) the discrete spectrum of
T (that is the set of all isolated eigenvalues of T of finite multiplicity)
and Q ~ ( T=) U (T) - ad(T) the essential spectrum of T . The opera-
tor T is said to be semibounded (below) , if (T (T) has a finite lo-
wer bound; in this case we denote by v (T) and (T) the greatest
lower bound of U(T) and a ( T ) respectively, s o that V (T) ,< p (T)
e
holds for every semibounded operator T, and v (T) </A (T) holds if
and only if V (T) belongs to u d(T) .
Semibounded operators play an important role in the quantum theo-
r y of systems with a finite number of degrees of freedom. The operator
T of total energy of a N-particle s'ystem in the Schrddinger representa-
tion is a selfadjoint semibounded differential operator of second order
in the Hilbert space L ~ ( R (with
~ ) m = 3N) . An eigenvalue .a < IIL (T)
is the total energy corresponding to a stable state of the system; in
particular V (T) (if this number is smaller than p (T) ) i s the enrgy
of the ground state. The number p (T) is the smallest possible energy
corresponding to an unstable state of the system and ,U (T) - u (T)is
the minimal energy required to lift the system from its ground state to
an unstable state.
Considerable effort in numerical analysis has been devoted to the

problem of computing V (T) and other eigenvalues of T below p (T).
There a r e the well-known variational procedures (Rayleigh-Ritz o r
Galerkin) which give upper bounds to these numbers, and more recently
the methods of Weinstein, Aronszajn and Bazley (see [ 2 ) and the litera
ture quoted there ) for obtaining lower bounds, such that it i s now possi-
ble to compute these eigenvalues up to a given e r r o r with a reasonable
K. Jorgens
amount of work. On the other hand little has been done to find numerical
approximations for ,U (T) although this number is clearly a s important
as v (T) . We discuss here, for Schr6dinger operators T , two me-
thods which make it possible to compute ,U (T) in many cases :
The perturbation method. Let T be representable a s a sum T = T +A,

0
where T i s selfadjoint and semibounded and A is T -compact (that
0 0
is the domain D(A) of A contains D(T ) and every sequence
0
{un) C D(To) with 1

llun -k 1< C contains a subsequence
J
such that { A U ; ~ , )converges strongly in H) . Then it follows that
J
e (T) = ue(To) ,
and in particular p (T) =,U(T )
o
. The method consists
in finding a representation of this sort such that ,U (T ) is known o r
0
at least is computable by known methods .
N
The method of splitting. Let m =
j=1 mj ,
m.
with x.J R ' for x R
m
. Let
m
where T. is a Schrijdinger operator in L2(R j ) acting on functions
J -
of the variable x . and p . the operator of multiplication by a real func-
J ~k
tion p .
~k J k
(x., .
) Now consider the operators
acting on functions in L ~ ( R ~ for

- ~ i ~= 1.2,.
) .., N . Then under certain
hypotheses we have the formula
p(T) = min U (Si) + ,U (Ti) )

(i)
K. Jorgens
which reduces the problem to the (supposedly easier) computation of the

numbers ,u (T.) and to the well-solved problem of computing the numbers
1
($1 .
Schrtldinger operators. We consider operators of the form
satisfying the following conditions :

2
a . bnd
J
p are real , '
(aj) L2, ~ O C( R ~ ,)
(4) m in the sense of

.(Rm) ,
L2, loc
1
j=1
3 ja J. = 0
Then the operator T i s defined and symmetric on C: (R"') and

0
can also be written a s
We now use a uliqueness theorem which is a slight extension of a

special case of a theorem of Ikebe and Kato 14 1 .
Theorem 0 (Ikebe-Kato). Let the coefficients a . and p satisfy
J
(4) and
1 m
a; E C (R ) and
J
for some a > 0 , C > 0 and for all x E Ftm , Then T (defined in
0
m
C: (R ) ) has a unique selfadjoint extension (also denoted by T )with
0
domain
Furthermore T
0
i s semibounded below .
The perturbation method. We consider perturbations of T of the form
0
m
2
(7) T = c i a . + a . t b.) + p+q
j=l J J J
where the coefficients a . and p of T are a s in theorem 0

J 0
and the perturbed coefficients a . t b . and p+q satisfy (4) (when
J J
substituted for a . and
J
p respectively) Then T . i s defined and sym-
metric in C: (Rm) and can be written as T * T0 t A with
Theorem -
1. Let a . and p satisfy the assumptions of theorem 0 and
J
assume that
(i) the derivatives a J.ak a r e uniformly bounded
m -
(ii) the functions B (bj)\nd satisfy condition (6)
j=1
(when substituted for p ) with some a >O and C>0 .
Then T has a unique selfadjoint extension with domain D(T) = D(To)
which i s semibounded. If in addition the function
tends to zero for ( x + co , then A i s T 0-compact.
This theorem and some similar ones a r e contained in [5 ] . The most

remarkable feature of these results i s that the coefficients of the unper-
turbed operator T need not be constant o r even bounded. F o r opera-
0
tors T
0
with constant coefficients Birman [3 1 and Balslev [I lhave
K. Jorgens
obtained very general perturbation theorems using the theory of quadra-

tic forms in Hilbert space .
The proof of theorem 1 r e s t s on an identity of the form
f o r functions u r H ; ~ ~ ( R ~, ) where K p and H for e v e r y p ~ ( O . 1

a r e integral operators with kernels K p (x, y) and Hp (x, y) identically
zero for I x-y I > p and satisfying the inequalities
I K p ( ~ Y)1 ( 5 c 1 x-y 1 2-m
for I X - ~ (-< p with a constant C independent of x, y and P .

Example 1. T =
0
-A (that i s a . = p = 0 ). It i s well known that
J
u (T
0
)=
= U (T ) = ( 0 , m) in this case. By theorem 1 the operator
e o
m
T = C (i a J. + bJ. ) 2 + q has the essential spectrunl u (T) = (0, co ) and
e
j=1
therefore p (T) = 0 .
1
Example 2. Let m = 3, x = (x x
1' 2"3)
and a (x) = - - x2 , a2(x) =
1 2
p
=5
1
p
x l , a (x) = 0 and p(x) = 0 The operator
3
.
T with these coeffi-
0
cients describes the motion of a charged particle under the influence of
a homogeneous magnetic field of strength /? in the x -direction. The
3
spectrum of T
0
can be found by a separation of variables (see 151 );
we get u ( T ) = oe(T0) =(lpl,
0
) . By theorem 1 any perturbation A
satisfying the assumptions of the theorem does not change the essential
spectrum. In particular we can take b. = 0 for j = 1 , 2 , 3 and
J
q(x) = - y ~ x I -1 ; the p e r t u r b e d o p e r a t o r T= T
0
f q now c o r r e s p o n d s t o
the so-called Zeeman-effect in quantum t h e o r y . According to t h e o r e m 1
we have (Te(T) = (IPI, 0) ; f u r t h e r m o r e it i s possible to show (in the
case where y> 0 ) by a Ritz-type variational argument, that the d i s c r e t e

s p e c t r u m of T c o n s i s t s of infinitely mahy eigenvalues below
p ( T ) = I p l ( s e e 15)) .
The method of splitting. We s t a r t with the following example :
Example 3. The S c h m d i n g e r o p e r a t o r f o r a s y s t e m of N p a r t i c l e s of
equal c h a r g e and m a s s in a c e n t r a l Coulomb field. H e r e we have m=3N and
- - ..,n )
x = (xl, x2,. where ,; E R ' ~ i s the position of t h e j-th particle.
N J
The o p e r a t o r i s
with positive constants p and p . Theorem 1 applied to t h i s o p e r a t o r

(with T = -A ) only gives that T h a s a unique selfadjoint and s e m i -
0
bounded extension. Condition (iii) is not satisfied, and in fact we shall
show that p (T) can b e computed with the help of formula (3) and i s
found to be s m a l l e r than p ( T0 ) =0 for N -> 2 .
T h e o r e m 2. L e t T. f o r j = 1 , 2 ,
J
...,
N be o p e r a t o r s of the f o r m
m j
( 7 ) in L2(R ) with coefficients satisfying the assumptions of t h e o r e m 1.
N
Denote by G . the v a r i a b l e s in T . and put m = m , and
J J i=l J
- -
..- .
x = (xl, x 2 , , ,xN) L e t p . be non-negative functions defined on R J
m,+m
k
~k
satisfying (6) with m = m.+m for s o m e positive constants and C,
J k
and such that the functions
tend to zero a s 1 ijI-ca and as ; k
for fixed
k
;
k
r I
n for fired ;.. Then the
J
m-m.
operators (1) and (2) in c ~ ( R
and~ C) ~ ( R 1 ) respectively have unique selfa-
0 0
djoint and semibounded extensions and formula (3) i s valid. If ae (T.)=I,U(T,),
1 1
ca)
for the index i giving the minimum in (3), then we also have 0 (T) =(p(T),oo).
e
A theorem of this type has for the first time beer. proved by
M.G. ZIslin [8 1 for an operator somewhat more general than (8) . Theo-
rem 2 above i s contained in [63 . Related results can be found in the
papers of C.van Winter [7] ; here the interaction t e r m s
a r e not
'jk
assumed to be non-negative and consequently formula (3) has to be re-
placed by a more complicated one. Finally Zislin and Sigalov [9] have
refined the theorem for the operator (8) by taking into account its symme-
t r y with respect to permutation of the variables zJ. . They prove a formula
similar to (3) for the restriction of T to the invariant subspace of
L ~ ( R ~corresponding
) to an irreducible representation of the symmetric
group S
N
.
A summary of this work and of related results has been gi-
ven by Sigalov in his recent article [lo1 .
Theorem 2 clearly applies to example 3. Denote the operator in que-
stion by T~ ; then T = T = A
1 j
1
1 .
(x -
-1 .
-p 1 3
in L (R ) and
2
ue(T ) = ( 0, WJ) by theorem 1. T is the well-known operator for a hy-
1
drogen-like atom; its lowest eigenvalue i s v ( T ) =
4
-lp2 .
The opera-
N
tors S. corresponding to T a r e all identical (except for the numbering
1
N- 1
of variables) and equal to T .
Therefore by theorem 2 we have
N N- 1 N N- 1 N- 1
ue(T ) = ( V ( T ), o o ) . It followsthat ,U ( T ) = V ( T ) <p(T )<
-I -'p2
4
for N > 2 . I f / ? ) N y Z i s l i n [8]has shown that
N N Mi-1 M
v (T ) < p (T ) ; in this case we have p (T ) p ( T ) for
M=l,2 ,..., N .
REFERENCES
------------
Balslev, E. : The singular spectrum of elliptic differential operators
in LP(R )
n
. Math. Scand. -
19 (1966), 193-210.
Birman, M.
(russianj. S.Math.OnSbornik
: the spectrum of singular boundary value problems
97 (1961) , 125-174.
-
Bazeley, N. W. : Lower bounds for eigenvalues. J.Math. and Mech.
10 (1961), 289-308
- .
Ikebe, T. and T. Kato: Uniqueness of the self-adjoint extension of singu-
l a r elliptic differential operators. Arch. Rat.Mech.Analysis -
9 (1962),
77-92.
Jorgens, K. : Zur Spektraltheorie d e r Schrodinger-Operatoren. Math.
Zeitschr. -
96 (1967) , 355-372,
Jorgens , K. : Uber das wesentliche Spektrum elliptischer Differential-
operatoren vom Schrodinger-Typ, Report, Inst. f.Angew. Math. Univ.
Heidelberg (1965) .
van Winter, C. : Theory of finite systems of particles; I the Green
function, I1 scattering theory. Math.-Phys. Publications of the Danish
Royal Academy of Sciences, Volume 2 (1965).
Zislin, M. G. : On the spectrum of many-particle SchrCfdinger operators
. -
(russian) Trudy Mosk. Mat. Obsc. 9 (1960), 81-120 .
Zislin, M. G. and A.G. Sigalov : On the spectrum of the energy opera-
tor of atoms on subspaces corresponding to irreducible representations
.
of the permutation group (russian) Izvestija Akad. Nauk SSSR -29
(1965), 835-860 .
Sigalov, A.G. : On an important mathematical problem in the theory
.
of atomic spectra. (russian) Uspechi Mat. Nauk - 22 (1967), 3-20 .
(C. I. M. E. )
A. LASOTA
CONTINGENT EQUATIONS AND BOUNDARY VALUE PROBLEMS
C o r s o tenuto ad I s p r a dal 3-11 Luglio 1967

CONTINTENT EQUATIONS AND BOUNDARY VALUE PROBLEMS
by
A. LASOTA (Krakow)
It i s known that the theory of differential equations with multi-

valued right-hand sides (sometimes also called contingent equations) is
closely related to the optimal control theory. Our aim i s to show that the
technique of the contingent equations is useful in the theory of continuous
and discrete boundary value problems too. In section 1 we s t a r t wtih
a new topological method which permits to establish the existence of
solutions provided a criterion of uniqueness i s fulfilled. Section 2 con-
tains a result of A.PliS concerning contingert nquations [5] . Ir section
3 and 4 existence theorems for continuous and discrete boundary value
problems a r e given. The main result, an approximation theorem, is
stated in section 5 .
1. Fixed point theorem . Let E a Banach space and let n(E)
denote the family of all nonempty subsets of E. F o r a set A in n(E) ,
a mapping H:A-2 n(E) i s called upper semicontinuous if its graph
i s closed in A xE . The map H is said to be compact if, for any

bounded subset B of A, the closure of the set
i s compact in E. The map H is called completely continuous if it

is upper semicontinuous and compact.
F o r a single valued mapping h: A3E the upper semicontinuity
of the mapping H:x+ h(x) i s equivalent to the continuity of h, the
compactness of H IS equivalent to that of h and the complete continui-
ty of H means the complete continuity of h.
The following theorem i s due to A.Lasota and Z.Opia1 [4] .
A. Lasota
.Theorem 1. Let U be a neighborhood of 0 in the space E

and let H: U - - , n(E) be a completely continuous mapping such that
Then for any mapping h: E-+E, the condition
(1.3) h(x) - h(y) H ( ~ - Y ) for x-y eU

implies that the equation
has exactly one solution.

The special case of theorem 1 was presented at the second con-
ference on differential equations (Equadiff 11) in Prague 1966.
The proof of theorem 1 is based on the old idea of Hadamard to
use the monodromy principle.
If h is a mapping of the form h(x) = Ax + b where b c E and
A is a linear completely continuous operator (A: E - t E ) , theorem
1 yiews the essential part of the first theorem of Fredholm. It suffi-
ces to set H(x) = {AX} and to observe that condition (1.1) means
the uniqueness of solution of the linear homogeneous equation x =Ax.
Note that in the Fredholm theorem the map A may be non-contrac-
tive in general.
2 . Plis lemma. Let 2=

[a, b] be a compact interval of the real
n
line R and let cf(R ) denote the family of all nonempty, convex, clo-
n
sed subsets of the n-dimensional real space R The space of all .
n
continuous mappings of 2
into. R with the usual norm
A. Lasota
( 1. 1 stands f o r the Euclidean n o r m in Rn) will b e denoted cn. By

P (p, A) (p 6 Rn , A n(Rn)) we denote the Euclidean distance of the
point p to the set A .
L e m m a 1. Let G be a mapping of 9
into cf (Rn) and l e t
n
a sequence of v k { c C of absolutely continuous functions satisfy
t h e following conditions
(2.1) lirn vk(t) = ~ ( t ) for t e 3 almost everywhere

k + m (a. e. )
(2.3) lim (v;( (t) , G(t) ) = 0 a. e.

ii -+a
Then the function v is absolutely continuous and
3. Continuous boundary p r o b l e m s
Now p a s s to the differential equations. Consider a map
F: JXRL
cf(Rn) , a map f: JXRL
Rn and L: cn+ Rn s u c h that :
(i) F(t, x) i s homogeneous in x ( F ( t , 9 x) = AF(t, x) for real a)
and completely continuous , i.e. the set
{ ( t , x, Y ) : Y € F ( t , x) , 1x 1 = 1j
i s compact in R~~ ;
(ii) f(t, x) i s continuous in (t, x) and s a t i s f i e s the condition
(3.1) f ( t , 4 - f ( t , y ) 6 F(t, x-y) (t, x, y ) e &RZn ;
(iii) L is l i n e a r and continuous .

Given the mapping F, f, L and a point re R n , we consider t h e contia-
gent equation
A. Lasota
with the homogeneous linear condition
and the o r d i n a r y differential equation
with the l i n e a r condition
n
An absolutely continuous function x t C will be called a solution
(in Caratheodory sense)' of equation (3.2) if it s a t i s f i e s condi-
tion (3.2) almost e v e r y a h e r e on 3.
From the t h e o r e m 1 we can d e r i v e the following ( s e e [2), 133 ) .
0 0 0
T h e o r e m 2. If the functions F , f, L s a t i s f y conditions 1 ,2 ,3
and if x = 0 i s the unique solution of equation (3.2) satisfying
(3.3) then there exists one and only one solution of the equa-
tion ( 3 . 4 ) satisfying (3.5).
Note that the t h e o r e m 2 is not true if t o mean solution of
(3.2) in the usual (non Caratheodory) sense.
In
o r d e r t o prove t h e o r e m 2 c o n s i d e r the mapping H of
n
E = c n~ R~ into cf(R ) s u c h that for e v e r y point (x, p) i t s image
H(x,p) is a s e t of a l l p a i r s ( y , q ) given by the formulae
for u(s) F(s,x)) and the mapping h of E into itself s u c h that

for e v e r y point (x, p) its image h(x, p) i s a pair (y, q) given by
the formulae
A. L a s o t a
It i s e a s y to s e e that the maps H, h satisfy conditions (1. 1) , (1.2),

(1.3) and that the existence of solutions of problem ( 3 . 4 ) , ( 3 . 5 ) is
equivalent to the existence of solutions of the functional equation (1.4) .
Map H is evidently compact. Thus to end the proof it i s sufficient
to show that H is upper semicontinuous. Suppose that
Then we have
k k k
(3.6) yk(t) = uk(s) d s +p , u ( t ) g F ( t , x (t)) ,
a
and consequently
k /
(3.8) (Y (t))E F(t. Xk(t))
From (3.8) and the upper semicontinuity of F we obtain
lim p ((Yk ( t ) ) ' , F ( t , x(t)) = 0 .

k -tco
Thus from the l e m m a 1 it follows immediately that
In addition, upor1 passing to the l i m i t in (3.6) (for t = 0)

and (3.7) we have
From (3.9), (3. 10) and (3. 11) it follows that (y, q) H(x, q) and we
a r e done.
4.Discrete boundary problems. In order to obtain a discrete analogue of
theorem 2 consider a finite sequence
a = t < t <.. . < t = b
o 1 n
.
Then we can replace the contingent equation (3.2) by the multivalued diffe-
rence equation
A Xi
(4.1) € F(ti, xui))
I
and the differential equation (3.4) by the difference equation
Axi
(4.2) TA = f(ti , x(ti)) i = 0, ..., n-1 .
where A x . = xitl-xi
1
and A t , = t . - t .
1 1+1 1
.
We assume that solutions of equations (4.1) and (4.2) a r e continuous in
0and linear in each interval [ti , titl] .
Theoreme 3. If the mappings F, f, L satisfy conditions (i) , (ii) ,
(iii) and if x=0 is the unique solution of equation (4. 1) satisfying
(3.3) , then there exists one and only one solution of (4.2) satisfying
(3 .5)
More exstensive theorems and some applicaticns in the theory of
discrete boundary value problems a r e due to F?H. Szafraniec [6 3.
The proof of theorem 3 is based on the same idea a s the proof of theo-
r e m 2. But in this case the lemma 1 is not needed.
5. Approximation theorem. The development of the theory of differential

inequalities allows to obtain precise uniqueness theorems for a general
class of differential boundary value problems. It is interesting that on the
same way we can establish the existence of solutions and the convergence
of finite difference approximations. Namely setting %= (to, . . . ,t n) and
f ( ' L ) = man titl - ti : i = 0,. . . . n-1 1 we have the following
Theorem 4. If the mappings F, f, L satisfy (i) , (ii) , (iii) and if

x = 0 is the unique solution of the contingent boundary vqlue problem
A. Lasota
( 3 . 2 ) , ( 3 . 3 ) then :
0
1) there exists a unique solution x of problem ( 3 . 4 ) , ( 3 . 5 ) ,
2) for J(z) sufficiently small there exists a unique solution
2
x of problem (3.5) , (4.2) ,
3) lim bxr-xO\\= 0.
S(%)
+o
The statement lo i s an immediate1 consequence of theorem 2 .
0
In order to prove 2 it is sufficient to show that for sufficiently
small S (@ problem ( 4 . 1 ) , ( 3 . 3 ) has only the trivial solution x =0
and to use theorem 3 . To this end suppose that for each integer k
there exists a sequence
and a corresponding nontrivial solution xk of problem (4. 1) , ( 3 . 3 ) , i.e.
k k
Since xk is linear in each interval (t., t . ) we can write ( 5 . 1) in
1 1+1
the form
and since problem ( 5 . 1 ) , ( 5 . 2 ) i s homogeneous we can assume that
(5.4)
From this and ( 5 . 3 ) it follows that the functions xk a r e equiconti-
nuous on J . Upon passing to suitable subsequence, if necessary,

we may assume that the sequence { xk 1 convergences to a function
x e c n . Thus passing to the limit in ( 5 . 2 ) , ( 5 . 3 ) and ( 5 . 4 ) we
obtain (see lemma 1 )
A. Lasota
which is impossible.
k.
To prove statement 3' consider a sequence ) + such that 1
k
( ) - P O and the corresponding sequence {xk] of so1utio.n~ of
( 4 . 2 ) , (3.5) . We have
and on other hand

0
(xO(t)) = f(t. xO(t)) , Lx = r .
From this and (3. 1) it follows that
(uk(t)) e F(tik k k
, u (t. )) t
1
ek(t) for tE(ti>
k
Lu = 0
k k o
where u = x - x and
k
Now suppose that 1) u I\+0 . Then upon passing to the sultable sub.
k
sequence we may assume that )lu I\ . + c (C E (0, t o 3 1 . Setting
k
v = uk 1 uk we obtaln
The functions vk a r e evidently equicoi~tinuous and we may a l s o a s s u -

m e that the sequence i vk { converges lo a function v C" . The
lemma 1 yields
v' (t) F ( t , v(t))
and obviously Lv = 0 , (I v 11=1 which i s impossible. Thus J J u (\-,0

k .
and we a r e do~:e.
A. Lasota
In o r d e r to illustrate theorem 4 by an, example, consider for the

equation ( 3 . 4 ) the following aperiodic boundary value condition
We s t a r t with a lemma concerning differential inequality
Lemma 2 . If the functions y ( t ) i s nonnegative and such that
then x = 0 is the unique solution of the inequality (5. 6) satisfying

the boundary value condition
Note that the estimation ( 5 . 7 ) i s the best possible of this type.

The lemma 2 i s due to S. Kasprzyk and .J. Myjak [I].
From theorem 4 and lemma 2 it follows immediately
Theorem 5 . If the function f(t,x) i s continuous and if
where the continuous nonnegative function 9 satisfies inequality

(5.7) then
l o there exists a unique solution of ( 3 . 4 ) , ( 5 . 5 ) ;
2' for sufficiently small $(%) there exists a unique solution x
of ( 4 . 2 ) , ( 5 . 5 ) ;
A. Lasota
REFERENCES
1 S.Kasprzyk, J. Myjak, On the existence and uniqueness of solutions

of F'loquet boundary value problem , Zeszyty Naukowe U. J., P r a c e
Matematyczne (to appear) .
2 A.Lasota , Une generalisation du premier theorem de Fredholm et
s e s application & la theorie des equations differentielles ordi-
naires, Ann. Polon. Math. 18(1966) , 65-77.
3 A. Lasota, Z. Opial, On the existence of solutions of linear pro-

blems for ordinary differential equation^ , Bull. Acad. Polon. Sci.,
Ser. s c i . math. , astronom. et phys. , 14(1966) , 371-376.
4 A. Lasota, Z.Opia1, On the existence and uniqueness of solutions

of nonlinear functional equations, ibidem (to appear) .
5 A. Plis, Mesurable orientor fields, ibidem, 13(1965) , 565-569.
6 F. H. Szafraniec, Existence theorems for discrete boundary problem,

Ann. Polon. Math. (to appear) .
(C. I. M. E. )
REDUCTION A DES PROBLEMES DU TYPE CAUCHY-KOWALESKA
torso t e n u t o a d I s p r a d a l 3 - 1 1 L u g l i o 1967
REDUCTION A DES PROBLEMES DU TYPE CAUCHY-KOWALESKA
Par
J. L. Lions
Universite de P a r i s
Introduction.
I1 intervient assez souvent en pratique des systemes differen-

tiels . qui ne sont pas d~1. lype de c a u c h y - ~ o w i l e s k a: en hydrodynamique
l e s dquations de Navier Stokes , l e s equations de l a magneto-hydrodyna-
mique, en meteorologie (Marchouk 1 2 ) ) etc.. .. .
On donne ici deux procedes , de portke semble-t-il a s s e z gdne-
rale, pour approcher de tels probl6mes par des problemes du type
Cauchy - Kowaleska qui sont eux memes ensuite approches par des me-
thodes numeriaues II~tandard~~.
Le p r e m i e r procede, dit de penalisation , est une variante dtun
procede introduit p a r Courant [l] en calcul des variations , es
dl;sage maintenant frequent en contrble optimal ; nous donnons au N, 1
l a methode s u r lfexemple des equations de Navier-Stokes, suivant
Temam [1],[27 .
Une idee vojsine avait deja ete introduite pour l e s equations
lineaires dans Bramble-Payne [l ] . Une variante a ete utilisee dans
Lattes - Lions [I] pour llapproximation numerique de problemes ma1
poses .
Le deuxihme procede, dit des derivees artificielles, e s t expose s u r
deux exemples : au n. 2 s u r ltexemple des equations de Navier
Stokes et au n. 3 s u r un probl6me lineaire quelque peu "singulier" -
mais qui intervient dans l e s applications (cf. Garipo-r [I] un un
problgme relevant de techniques analogues dans Douglas- Peceman-
Rachford rll ) .
J. L. Lions
1. Equations de Navier-Stokes-; Methode de penalisation.
1.1 - Position du probleme (1)
Soit R un ouvert borne de R~ , n = 2 ou 3 de frontiere

assez reguliere .
On c h e r c h e une fonction u = { ul, ..., u e t une fonction p
definies dans l e cylindre Q = Rx 1 o, Tr telles que
n
3u t grad p f dansQ, (9>0)
i=l 1
(1.2) div u =0 dans Q ,
(1.4) u(x, 0) = uo (x) dans R
ob f = f f l , ..., f n ( e t u
o
sont d o n n e e s d a n s QetR .
Naturellement il faut p r e c i s e r h quel s e n s on c h e r c h e une solution
de ( 1 . 1 ) - (1.4) ; mais il e s t de
toutes facons c l a i r que, B cause
de ( 1 . 2 ) (qui ne contient p a s de derivee
3 P ) l e problhme mixte
(1.1) . . . (1.4) , n t e s t p a s du type de Cauchy -Kowaleska. Cela conduit
d e s difficultes numeriques : l o r s q u e l l a n p a s s e p a r ex. aux differences
finies, l e s conditions ( 1 . 2 ) introduisant d e s ncontraintesltcompliquant
s e r i e u s e m e n t l a m i s e en oeuvre d e s calculs.
Notre objet e s t donc ici d l a p p r o c h e r l e s y s t e m e (1.1). . .(1.4)
p a r un s y s t 6 m e de Cauchy-Kowaleska.
Cela peut s e f a i r e g r o s s o mod0 d e deux facons ( 2 ) ..
(i) pait "suppression" d e la "contrainten ( 1 . 2 ) ;
(ii) p a r introduction dlup " t e r m e artificiell1E 3~

A '
3 t '
' ) ~ f . a u s s i l e s conferences de C. Capriz, c e Volume.

2) A n o t r e connaissance . I1 y a t r e s probablement beaucoup d l a u t r e s
precedes.
.I. L. Lions
La llmethode" (i) fait llobjet de ce N la methode (ii) 1)

fait llobjet du n. 2 suivant.
1.2. Penalisation.
L1idee generale est de supprimer (1.2) et d'ajouter dans (1.1)

un t e r m e de l a forme
-- grad (div u ) ,
&
> 0 I1petitl1- ce qui conduit 5 une fonction u ne verifiant plus
(1.2) mais 5 l a place de (1.2) ?i une condition
Pcecisons cela - Introduisons dans ce but l e s outils suivants :
1 n
V = j v l v;(H0(R:) , div v = 0 } , (2)
)",)
H= I f 1 fb(~~(R div f = O ] ;
V et H sont des espaces de Hilbent sur R (3), pour l e s

produits scalaires
dx 9
i, j = l R J
dont l e s normes correspondantes sont notees 11 u (1 , I f 1 ,
2)~A(~) est defini dans l e s konferences de Raviart, ce vol.

"Toutes l e s fonctions sont 5 valeurs reelles
J. L. Lions
Pour u, v, w & V, on pose
Le probleme (1.1) . . .(1.4) peut alors s e f o ~ m u l e r de la fa-

2
Con suivante ( 2 ) : trouver u t L (0, T;Y))'( tel que
Remarque 1.1.
La condition "div u = 0" est contenue dans l'appartenance A

V et la pression p n'a "disparufl que parce que l'on a p r i s
l e produit scalaire avec v de divergence nulle (et nu1 au bord).
On montre ceci (Leray El] , Ladyrenskaya [I] , Lions [I],
Lions-Prodi
a) il existe une solution u de (1.7) (1.8), verifiant en outre
2
b) s- i n= 2, il y a unicite de u dans L ( 0, ;V) verifiant (1.7)
(5
(1.8)
On va maintenant approcher u (dans un sens convenable) par
des fonctions u ne verifiant plus (1. 2)
Aprks ce qu'on a dit au debut de cette section, il est nature1
On defiliit ainsi une forme tri-lineaire continue s u r ' v .

Suivant l e travail classique de Leray [I]
3) Espace defini dans Raviart, ce vol.
2
4, On montre, Lions r d , q u e s i u t L (0, f ; v) et satisfait - (1.7)
alors (1.8) a un sens.
5 , Dans le cas 'In = 3. on ignore s'il y a unicite ou non.

'J. L. Lions
1
(1.10) W = (Ho ( 9 ) )"
(1.11) K = ( ~ ~ ()" 9 )
l e s produits scalaires dans W et K etant "les m^emestt que dans

V et H, puis de chercher u verifiant
Mais il y a 18 une difficult@ : on verifie sans peine que s u r

V , o n a
ce qui conduit aux inegalites de ltenergie pour (1.7) et & llexistence
dtune solution globale (i. e pour T quelconque) de (1.7) (1.8) 1) ,.
par contre (1.13) nlest pas vrai pour u6W de sorte que tres
probablement 2, l e probleme (1.12) e s t ma1 pose . On doit done
le modifier de l e facon suivant (cf. Temam [I] [2] ) : on introduit
forme trilineaire qui Cvidemment coincide sur V avec b(u, v, w).

Mais V ut (H: (62) )n = W on a :
n,
(1.5) b (u,u,u) =2I 1, u.1 3
3x.
(u2)dx +:
J
div u S u 2 ) d x = 0
J
1, J 1 j
1) Naturellement (1. 13) a lieu "8 cause de laphysique"

2) Mais cela ntest pas demontre .
- 274 - J . L. Lions
et on peut maintenant introduire le probleme penalise suivant
trouver &'(o,T ;W) verifiant
(t(tl .V)+ ( ( ~ ( ( t )v)) +b

9 (U (t),
1
(t) . v) +-(div UL( ' ).di~~)=
1.3. Resultats
On montre alors (Temam [I] p] ) :
Theoreme 1.1. Le probleme (1.16) admet une solution ; il y a unicite

-- s i
n = 2,.
En outre
--
Theoreme 1.2. . Lorsque 1 + 0 on a
-
(i) pl n = 3 , on peut extraire
de facon que
2 2
(1.17) u +u dans L (0,T ;W) faible et L (o,T;K) fort 1)
I'
(ii) -
si n = ' 2 , o n l a : :
(1.18) u
I
-+ u dans
2
L ( o , ~ ; w ) faible et L
i (0, T;K)fort .
En outre
Remarque 1.2
Faisant v =u dans (1. 16) on en deduit aussitot

t
" ~ x t r a c t i o n inutile s i l e probleme de Navier Stokes admet une solution

unique.
"Quotient de llespace des distributions s u r Q
par R (avec extraction
de suite s i n=3) .On peut preciser ce resultat.
J. L. Lions
1.4 Applications.
Du point de vue numerique l'approximation s e fait maintenant en

deux temps :
1.4.1 . On choisit au moyen de (1.19) ;
1.4.2 On integre numeriquement (1. 16) par l e s differences

finies (ou Galerkin) ; il s1agit maintenant dlun probleme non lineaire
mais ou les methodes standard slappliguent. Cf. T e m a p c]
Remarque 1 . 3
R. Temam a egalement adopte dans Temarn [2] 2 (1.16)

la methode des pas fractionnaires (variante des directions alternees):
cfr. Douglas (ce volume), Marchouk [I] , Yanenko [I] , Temam [3],
Lieutaud [I] .
2 Equations de Navier-Stokes ; methode des derivees artificielles.
2.1 Idee generale de la hethode .

L1idee est grosso mod0 de "remplacer" (1,2) p a r
O P + div u = O
a?-
mais cela conduit des difficult& avu l e s integrales dlenergie . Pour
I1recuperers l e s integrales dlenergie on modifie llequation (1.1) comme
suit : on cherche. 1)
avec
l ) ~ e notations
s sont celles du n. 1 .
J. L. Lions
(2.3) a qtL + div u E

L Q = O (6)O)
(2.4) \ Cr ( x , t ) = ~s i x e , ~F]o.T[,
(2.5) 0
2
q0 choisi quelconque dans L (52) .
On peut montrer (par le m&me genre de methodes que dans

Lions [I] , Lions-Prodi [l])les resultats suivants :
Theoreme 2.1. . Le probleme- ((2.1) . . . (2.5) admet une solution
-I1 y a unicite si n= 2
Lorsque on oeut extraire une sous suite
encore notee u L , q b , telle que
(2.6) ut + 2
u dans L (0, r,W) faible et
2
L (0, T, K) fort
En outre
grad (qt- lu' 1 2)+ grad p dans (Q) .

2.3. Remarques et applications
Si l'on compare l e s methodes de penalisation (N. 1) et des derivees

artificielles on peut faire l e s observations suivantes :
(i) l a methode de penalisation copduit k une mise en oeu-
vre numerique plus simple (notamment puisqu'll y a une equation de
moins) ;
(ii) l e choix de t est plus simple dans la methode de penali-
sation ;
(iii) p a r contre la methode des d e r i d e s artificielles est de por-
J . L. Lions
tee peut et1.e plus generale comme l e montre l'exemple donne au N.

ci a p r e s , oh une methode de penalisation ne semble pas commode.
3 . Equations avec conditions d'Cvolution au bord : hethode des derive-

e s artificielles.
Soit R C R~ ouvert borne de frontiere r reguliere .

On cherche une fonction u=u (x, t ) solution de
(3.1) A u = 0 pour xeR , t > o
(3.2) oh
A?= -1-
3 xi
(
1J
(
3 xj
X
avec
(3.3)
(oh
3u
--=
av
3u
i t a i j 3 7 ; cos (n, xi) , n = normale 5 r eit6rieure 5 R)
J
Donc l e s d e r i d e s en : nlapparaissent que s u r 2 et 2 cause

de (3.1) le probleme n'est pas du type de Cauchy - Kowaleski.
' ) ~ t done A ne contienl pas de derivation en t-
2' Pour un peu simplifier on prend des fonetions P valeurs reelles .

,I. L. Lions
3 . 2 Solution "fontionnelle" du probleme

-- (3. 1)(3.3) ( 3 . 4 )
On utilise ici, s a n s l e s r a p p e l e r , quelques notations et r e s u l t a t s
de Lions-Magenes [I] [2] .
Soit gbHli2 (r); soit w l a solution de
A w = 0 dans R
(3.5)
wl = g
r
Alors
est defini et g H -112 (,-)

F --
dtoh un o p e r a t e u r
avec
Cet operateur B est coercitif au s e n s
Le probleme (3. 1) (3.3) (3.4) equivaut au suivant

s i w designe l a t r a c e de u s u r
- , alors il faut trouver
w solution de
On peut a l o r s appliquer les metilodes

--- standard
- pour la r e s o -
lution de (3. 10) (3. 1 1 ) . (cf. p a r ex. Lions-Magelles
r21 , chap 3) .
Du point de .ue numerique, h o r m i s d e s c a s p a r t i c u l i e r ~ ob
lton peut expliciter l t o p 6--
r a t e-
u r integro
-- differentiel singulier
--- R ( et oh
.J. L. Lions
l'on peut a l o r s appliquer l e s methodes de c h e r r u a u l t [g . chap. 51,

l'integration de (3. 10) ne s e m b l e pas t r e s facile.
C1oil l'inter'et (eventuel) de l a methode d e s d e r i v e e s artificielles.
-
3.3. Methode d e s d e r i v e e s artificielles.
On c h e r c h e solution de
1
(3.12) & "~7
'+As t = 0 , C > O ,
oh
= prolonpment de u 3 R(ou suppose que u C H (r),d>o) .
Id
0 0 0
On peut a l o r s m o n t r e r l e s r e s u l t a t s suivants :
Theoreme
--- 3. 1. ----
L e probleme (3.12) (3. 13) (3. 14) -admet une solution
--
-Tlleor.&me
.... 3.2.
-- Lorsque
- --+ 0 on a
u ,
u dans L 2 ( 0 , ~ H1(fl)
; ) .
t
3.4 Applications.
On peut nlaintenant d i s c r e t i s e r (3. 12) (3. 13) (3.14) . I1 slagit

desormais d'un problPme standard. Cette inethode a e t e donnee dans
Lions
r21.
3.5 -R--e m a r q u e s
Si A e s t s e l f adjoint, ou peut c o n s i d e r e r l e probleme
J. L. Lions
Au = 0 dans O X [t >OJ ,
Alors le mCthode d e s d e r L v e e s artificielles conduit au proble-

m e suivant : on cherche u solution d e
t
(3'21) oh ui = prolongement de u 1, ,
On a d e s r e s u l t a t s analogues 3 ceux de 3.2 .

Voir probleme de c e type dans Garipov [l] .
BIBLIOGRAPHIE
---------------
Bramble-Payne [ll cf. Payne, dans Numerical Solution.. . Bramble ed. Acad.
P r e s s , 1967.
Y. Cherruault 111 Approximation d l o p e r a t e u r s l i n e a i r e s e t applications.
R. Courant [I] Variational methods for the solution of problems of equilibrium
and vibrations- Bull. A. M. S. 49 (1943), p. 1-23.
J. Douglas et P e a c e m a n - Rachford [I] SIAM, 1955, p. 42-65 e t 28-42.
R.M. Garipov [I] On the l i n e a r theory of gravity waves. Archive Rat. Mech.
Anal. 24, 1967, p. 352-362.
O.A. Ladyzenskaya [I] Equations de l'hydrodynamique.. . Moscow 1962
R. L a t t e s e t J. L. Lions [l] Methode de quasi r e v e r s i b i l i t e e t Applications.
Dunod 1967.
J. L e r a y [l] J . Math. P u r e s e t Appl. 12(1933), p. -82 ; 13(1934), p. 33-418;
Acta Math. 63(1934), p. 193-248.
J. Lieutaud [I] These. P a r i s 1968.
J . L . Lions [I] Equations differentielles
[2] Cours CEA-EDF - Paris. 1965
J. L. Lions e t E . Magenes [I] Probli?mes aux l i m i t e s non homogPnes (11)
Annales Inst. F o u r i e r , 11(1961), p. 137- 178.
121 ProblPmes aux l i m i t e s non homogenes e t applications-vol. 1
e t 2 Dunod. 1968.
J . L . Lions e t G. P r o d i [I]c.R. Acad. Sc. P a r i s , 248(1959), p. 3519-3521
G:I. Marchuk [I] Methodes numeriques.. Novosibirsk 1965 .
121 P r o b l & m e s de Metheorologie . Leningrad 1967. (en r u s s e )
Traduction F r a n c a i s e , Armand Colin, P a r i s , 1968.
R. Teman [I] C.R. Acad. Sc. P a r i s , t. 263, p. 241-244.
[2] Article s u r Navier S okes B P a r a i t r e
[3] T h e s e , P a r i s , 1967.
N. N. Yanenho [iJ Metilode deb pas Ir*ctioll~~dii-es. . . . (sir r u s s e )
Novosibirsk. 1966 - Traduction F r a n c a i s e , Armand Colin, P a r i s ,
1968.
(C. I. M. E. )
J . L . LIONS
"PROBLEMES AUX LIMITES NON HOMOGENES A DONNEES

IRREGULIERES ; UNE METHODE DfAPPROXIMATION I'
Corso tenuto a d I s p r a dal 3-11 Luglio 1967

PROBLEMES AUX LIMITES NON HOMOGENES A DONNEES IRREGU-
LIERES ; UNE METHODE D'APPROXIMATION
Par
J . L. Lions
Universite de P a r i s
Introduction.
eme
Pour l e s operateurs elliptiques du 2 ordre, une methode d'appro-
ximation pou l e s probl6mes aux limites homog6nes ?I donnes irregulieres
(le 2eme membre etant p a r ex. une rnasse de Dirac, la solutiun c o r r e -
spondante etant alors le noyau de Green) est donnee darls Bramble [I]
et pour l e s donnees frontigres irregulieres (la solution correspondante
&ant alors p a r exemple le noyau de Poisson)
- une autre methode est
donnee dans Jamet 1
Wous introduisons ici une methode differente valable pour l e s opera-
t e u r s elleptiques d'ordre quelconque et des donnees au bord arbitraire-
ment irreguligres et valable aussi pour l e s equations d'evolution. Mais
supposant (A l a difference des travaux cites de Bramble et Jamet) que
le frontigre et l e s coefficients dp, y I1tr&s reguli5rs". Les details te-
chiniques sont dans le cas general fort longs et utilisent l e s resultats
de Lions-Magenes [21 ; nous expliquons ici l a mCthode s u r des exem-
ples simples.
5 1. Cas elliptique
1. Example 1
1.1- Position du probleme
Soit R un ouvert borne de de frontiere 7 a s s e e regu-

libre . Dans R ou s e donne un operateur A de la forme
J. L. Lions
a , a , . c*(R) , avec
0 1J
On considere l e probleme de Dirichlet
oh l1on suppose que
On montre (cf. Lions-Magenes El] , [z] chap. 2) que l e --

pro-
bleme (1. 3) (1.4) --
admet, s o u s l e s conditions (1.5) (1.6) , une solution
2
unique qui appartient & L (Q) .
Notre objet e s t de t r o u v e r une approximation de c e problbme

(dans L'(R) ) .
1 . 2 . IdCe generale de l a methode.
On va changer l a nature du probleme (1. 3) (1.4) (du type l t ~ i r i c h l e t " )

en l l a ~ o r o c h a n tp a r d e s o r o b l e m e s de Neumann.
Soit
b
ij3 l ' o p e r a t e u r de derivee conormale a s s o c i e A.
a3
a
On r ~ m p l a c e a l o r s (1.4) p a r 6, 2 t u = g sur r , >0 .
On e s t ainsi conduit au probleme
- (du type Neumann) suivant
trouver (pour > 0 donne) u solution de
€
I ) Tout c e qui suit s t e t e n d aux o p e r a t e u r s elliptiques

conque (on n'ulilise janiais de principe du maximum).
- -
dlordre q ~ e l -
J. L. Lions
On
E a g
verifie
t
fadement
u6 = .
que
sur p.
ce probleme admet une solution , m i -
Clue dans H 1(R) l ) .

-
Reste ensuite faire E 0 .
1.3. -
Resultats de convergence lorsque f 4 0 .
2
Theoreme 1.1. Lorsque
-- -4 0, u
E
-,u dans L (R) .
Principe de la demonstration.
1
On deduit de 18 ltexistence et ltunicite de u
€
dans H (R) .
2) Introduisons solution dans H'(R) de
"E.
A% v = u dans 7
(1.10)
t 2 . t y~
L= 0 sur F,
3
Puisque ul 6 ~ ' ( 0 ) , on a; v E H (0) (cf. p a r ex. Lions-
6
Magenes , chap. 2 t s i lton prend dans (1.9) v = vE e t que
[2]
l t o n integre p a r parties , on obtient 2)
')l?ormellement, (1.1) (1.8) est un probleme voisin et regularis6 de

.T. L. Lions
3) Mais en reprenant l a demonstration de la regularit@ au bord,

on montre que
(1.12)
I1 1I <C , 1
(uE , C, independant de & ,
A1Ors \Ia%+v~ 1 HI/^(^) 6 C2 IuI ( et
Mais alors (1.11) donne
lutl
2
6 c211gllH-~ IUtItC1 lfl la l
(f1
d'oh
4) D'apres (1.13) on peut extraire une suite encore notee u

C
telle que
2
(1.14) u --, w dans L (R) faible ,
t
Mais alors (Lions-Magenes 12 J , chap. 2 ) , puisque AuE =
= f , u F 4 w J p dans H - ' / ~ ( P) UL+?-W-
dans H -3/2(r)
r a av
et (1.8) donne wlr = g . Comme Aw = f , on a w = u et
donc
0
(1.15) u dans ~ ' ( 0 ) faible.

u~ +
5) Alors v t t v dans H2(R) faible , oh A'V = u,

2
= 0 et luh -u12 =lqf +Iu( - 2Re (uL ,u)
1
=(g,~ v')~ t(f, V
E
) t \uI2 - 2Re ( u L , u ) :
1
- - - 3- v + --3 v dans H-lI2 (r) faible et donc
€ vt - i)u* 3u*
2 3v 2
(uL-ul -+ -(g.-- )
r f v ) - 1 = 0 .
1.4. Applications.
On peut effectuer une approximation "en deux temps" : on ctloisit

dlabora puis on approche (1.7) (1.8) par les metholies generales
de Aubin 111 , Cea el] .
2.1. Position du probleme . Methode.

On considere maintenant encore (1.3) (1.4) aver f 6 L'(R) mais
avec cette fois
La solution u correspondante est cette fois dans H"(Q)

(Lions - Magenes [I] 2 ) .
Cette fois l a brCgularisationll (1.7) (1.8) ntest plus suffisante et
il faut une double regularisation,
On considere sur ~ ~ ( une
0 )forme b(u, u) sesqullineaire coer-
cive choisie Ifla plus simple possiblev , par ex. b(u, v) = produit
scalaire sur ~"$2) .
On definit alors u (oh =itl, t2] ) . comme la solution
--
2 2)
t
dans H (R) de
On peut alors montrer le
Theoreme 2.1 . Lorsque €40 , on a
(2.3) u
€
-, u dans
-- H-'(R).
..
Cela sgffira pour comprendre comment resoudre le cas general :
& H-' (7) , s > 0 quelconque,
Comparer J. P. Aubin . J. L. Lions [l] (ce volume) .
- 290 -
J. L. Lions
La demonstration consiste A choisir dans (2.2) :
v = v solution de
C
(t l ~ ' t ~ f ) v E= ( - ~ t l ) - ' Lu 08 (-At11
- 1 e s t Itinverse de ltiaomor-
1
phisme (-A + l ) de H (0) s u r H-I (R) , et avec des conditions aux limites
0
convenables.
2.2. Application.
On a (theoreme de Sobolev fractionnaire; cf .J. Peetre 111 )
Donc :
Si n 6 3 on peut prendre dans (2.1) g =

- x ) ,x E r;
S (x - - 0 0
le solution u correspondante est le noyau de Poisson .
Cf. pour cela une autre methode (utilisant l e principe du maximum)
dans P . Jamet fi] .
2.3 Remarque
P a r l e genre de mt'hode donnee ici, on peut construire ltapproxi-
mation (Iten deux t e m p s r t ) des problemes genCraux -
non
-- variationnels
A elliptique d t o r d r e 2m, B . recouvrant A

J
les .
I1 devait e t r e possible de trouver une approximation directe de
ces problemes , mais cela n r e s t pas fait B notre connaissance.
4 2. Cas parabolique.
1. Nous nous bornons A un cas particulier t r e s simple.
Dans le cylindre Q=R x ] O,T [ , on considere l e probleme
- 291 -
J. L. Lions
oh A est donne comme au 91 , f tL

2
(Q) . avec
et avec
Z
Alors la solution u est en particulier dans L (Q)
(cf. Lions-Magenes [2], chap. 4) .
On t t a p ~ r o c h e tce
- ' probl&me de la facon suivante : on designe par
la solution de
'3 Ut
(1.5) + A u = f ,
37 t
On peut alors montrer le

Theoreme 1.1. Lorsque
(1.8) u!
t
- &
u
0, on
dans
a
2
L (Q) .
On peut , apres avoir I1regularisell l e problhme initial en

(1.5) (1.6) (1.7) . utiliser l e s methodes de Raviart
PI .
2 Remarque.
P a r double regularisation (analogue au N. 2 , 4 1) on pourra atteo-
d r e le c a s oh g g H-' (r), s > 0 quelconque .
'I Les coefficients de A peuvent dependre de t et tout ce qui

suit s1etend aux operateurs d l o r d r e quelconque.
BIBLIOGRAPHIE
---------------
J . P . Aubin [I] Approximation . . . Bull. S. M. F. MBmoire N. 12, 1967.
J. P. Aubin - J. L. Lions [I]Ce volume.
J.H. Bramble [l] Ce Volume.
J. Cea [I] Approximation variationnelle des problgmes aux limites.

Annales Inst. Fourier 14 (1964), p. 345-444.
P. Jamet [I] A paraitre

J . L. Lions-E. Magenes Ll] Probl6mes aux limites non homog6nes (11).
Annales Inst. Fourier, 11(1961), p. 137-178.
-
[2] Problgmes aux limites non homog6nes et applications. Vol. 1
e t 2, P a r i s , Dunod 1968.
J. Peetre [I]Espaces d'interpolation et theorgme de Sobolev. Annales Inst.

Fourier, 16(1966), p. 279-317.
P. A. Raviart [I]Ce Volume.
CENTRO INTERNAZIONALE MATEMATIC0 ESTIVO
(C. I. M. E. )
J. P. AUBIN et J. L. LIONS
REMARQUES SUR LIAPPROXIMATION REGULARICEE DE PROBLEMES

AUX LIMITES
C o r s o t e n u t o a d I s p r a d a l 3-11 Luglio 1 9 6 7
REMARQUES SUR LIAPPROXIMATION REGULARISCE DE PROBLEMES
AUX LIMITES
Par
J.P.Aubin et J . L . Lions
Introduction
-----------
On donne ici un procede assez general permettant d'approcher dans
la topologie l a plus fine possible ( I ) l a solution de certains problemes aux li-
mites variationnels lineaires de nature elliptique, par des solutions de proble-
mes aux differences finies (ou dtautres problemes approches) .
La methode proposee utilise essentiellement :
(i) une Vegularisation" des operateurs differentiels :

(ii) des operateurs de prolongement et restriction ( p r ) .
h' h
Une technique du type (i) a dejA ete utilisee (mais non assortie de
discretisation) dans J. L. LIONS -G. STAMPACCHIA [I] (cf. (3.14) de ce dernier
travail) .
La technique (ii) est utilisee dans J. P. AUBIN [l] .
[z]
Ltusage simultane de ces deux techniques conduit aux resultats de cette
note; comme cons4quence on obtient une estimation Sde l l e r r e u r dans les
espaces tlusuels" de fonctions differentiables.
Le meme genre de methodes peut slappliquer aux problkmes non
lineaires ou unilateraux dans l e s cas ( helas r a r e s ) oh l1on conduit un theo-
r e m e de regularite.
P o s i t i o n du p r o b l e m e
1 . ....................
I . 1. Soit $ un espace de Hilbert s u r ( , de norme (I I( , et

soit V1 son anti-dual, ( f ,v) design2r.t !o forme shsquilineaire mettant Vt et
V en anti-dualite .
I. e. la topologie raisonnable de ltespace oh est l a solution du problgme que

lton approche, sans perte de regulariete. Pour dlautres resultats dans ce sens, cf.
l e s conterences de BRAMBLE (ce Volume) .
- 296 - J. P. Aubin et J. L. Lions
Soit A~ f((V :vl) I ) , tel que

2
(1.1) Re ( A v , v ) 34\\v\l , d > 0 , v v € v .
L1adjoint A* de A verifie alors Re (Rv , v ) >oflvV bl v h V et
on sait que A et A$ sont des isomorphismes de V surv'.
On suppose maintenant connu un "theoreme de regulariten (cf.
section 1.2 c i apres pour des exemples) : on suppose connus deux espaces de
Hilbert W et F avec
W C V ,FCV1
l'injection de chaque espace dans la suivant etant continue (mais W -resp.

F - ntetant pas necessairement dense d a n s v - r e s p y ' ) , de sorte que
la
- solution u &s V de fl Au = f " appartient
a w.
(mais A ntest pas necessairement un isomorphisme de W s u r F , sf.
I
aussi No 5).
1.2. Example .
Si R e s t un ouvert borne de fk , de frontiere reguliere , on
prend
espace de Sobolev d t o r d r e m muni de la structure hilbertienne
On considere ltoperateur A defini par
De facon generale, $ (X, Y) designe ltespace des applications lineaires con-

tinues de X dans Y .
- 297 -
J. P. Aubin e t J. L. Lions
oh l e s a sont des fonctions donnees dans R et l'on suppose

Pq
que (1. 1) a lieu .
L1espace V1 n'est pas un espace de distributions s u r R. Si
k
l'on prend F = H (R) , a l o r s l a solution u dans V de Au= f
verifie
des distributions dans R, A = I ( - 1 ) IpI Dp (apqDq),
(le4)
I A u. = f au sens
et m conditions aux limites "du type NeurnannM s u r r (1)
et sous des hypotheses de regularite convenables on a a l o r s

,
u6 (R)
kt2m
et s i donc l1on prend W =H (R) alors (1.2) a lieu
1.3. Le probleme e s t maintenant l e suivant : s i f est donne dans

V1 , on connait une methode systematique d'approximation dans
- V de
l a solution de l1 Au =f1I (cf. J. P. Aubin [l] ) ; s i maintenant f est
donne dans
- F , peut-on approcher u dans
- W ? Nous allons donner
un procede assez general ayant cette propriete .
2 .................................
Schemas variationnels regularises.
Soit h un parametre de (Rn destine tendre v e r s 0. De facon

analogue 3 Aubin [l] , pour chaque h on suppose connus
(i) un espace V de dimension finie

h
(2.1) (ii) un operateur p 'injeotif de V dans W
h h-
(iii) un operateur r (lineaire ou non ) de V dans V
h h
tels que , ((1 111 designant. l a norme dans W :
(') Pour details, cf.par ex. Lions-Magenes

PI , Chap. 2 .
- 298 -
J. P. Aubin et J. L. Lions
(2.2) (ii) d(h)+o s i h d o,

(iii) p h r h v 4 v dans V (resp W ) Y V a V (resp. W) lorsque h - r o .
P l a ~ o n snous dans l a situation de 1.2.

On peut alors construire cf. J. P.Aubin [2] ph et rh satisfaisant
2 (2.1) (2.2) et avec
kt2m
(2.3) $(h)=O (Ihlmtk) (si W = H ( R ) ) (1)
2.3. Schema variationnel regularise .

Remarque preliminaire.
Une fois en possession des operateurs p l'approximation

h"hJ
ttnaturellel' de l1Cquation
(ou encore, en posant (A u,v) = a(u, v) :
est :
1
Uhevh
Mais il nly a pas de raison pour que, sans hypothese supplementai-
r e , on ait alors ph u + U dans W lorsque f est dans F.
h
La methode proposCe consiste B regulariser l e schema (2.4) .
La forme b(n, v)
On choisit '2' b(u, v) forme sesquilineaire continue s u r W, telle
I).
hn) designe alors l a maille du reseau s u r lequel on d i s c r e i s e
le ~ r o b l e m eaux limites
2)0n aura interet t~prendre b(u, v) l e plus simple possible , p a r exemple

produit scalaire l e plus simple possible dans W.
- 299 - J. P. Aubin et J . L. Lions
Re b(v, v ) > t l\lvl\l V v C W , J3 > 0 .

On definit alors u ( Vh comme la ( I ) solution de
h
(2. 6, ~ ( hb(ph
) uh, ~ ~ v ~ ) + a ( ~ =~(I.uphvh) ~ r c~vh
~ . . ~Jvh ) .
oh f ( h ) > 0 et E( h ) 3 0 lorsque h3 0 (2)
Ctest l e schema variationnel regularise qui a la propriete recher-

chee comme le montre le theoreme suivant, qui s e r a etabli au N. 3 :
Th@or+me 2.1. On suppose que A a l e s proprietes (1.1) (1.2) ,

que
7
ph satisfait B (2.1) (2.2), et que b satisfait A ( 2 . 5 ) . Si u
h
est
-
la solution du schema (2.6) on a :
(i) -
si f 6 Vt alors u 3 u p dans V lorque h-)o (3).,
h h 7
(ii) -
si f e E' et
- s-
i E (h) est choisi de s o r t e que
(2.7) +J&4 0 lorsque h3 0 ,
alors p u C ) u dans
-W lorsque h a 0 ,
h h
Estimation de lterreur
En utilisant la theorie de ltinterpol;tion (4) on peut obtenir l1esti-
mation de l t e r r e u r suivante ; de facon generale, soit [w, V] l'kspace de
Hilbert intermediaire (J. L. LIONS [l] ) de parametre 6 . 0 <8< 1 (5),
Alors
-
("11 es! immediat de verifier que (2.6) admet une solution unique.
n faudra p r e c i s e r ce point - cf. (2.7) c i dessous .
( 3 ) Cette propriete est verifiee (comme on le rnontre sans peine)pour le schema
(2.4)
(4)0n utilise ici seulement la theorie hilbertienne de. ltinterpolation, comme in-
troduite dans Lions [I] . Un expose detail16 en est donne dans Lions-Magenes
[I), Chap. 1.
J . P. Aubin et J. L. Lions
Sous l e s hypoth6ses du Theor6me 2. I . , o n :
Le theorkme e s t egalement demontre au N. suivant.
-Remarque.
--- 2.1.
La theorie de l'interpolation est un outil commode pour ltobten-

tiorl tlc \'estimation de l ' e r r e u r . Pour un autre exemple dvapplication (utili-
sant cette fois l'interpolation d'espaces de Banach, comme dans Lions-Peetre
il)) donne
P S ~ dans Peetre-Thomee
3 . .......................................
1)emonstration d e s T h e o r e m e s 2. 1. et 2 . 2
3.1. Le c a s (i) du Theoreme 2.1.
On fait v = u dans (2.6) et on prend l e s parties reelles des deux

h h
membres. I1 vient
On peut alors supposer, p a r extraction dvune sous suite encore

notee h , que
Ph Uh -
Mais comme ( f , p
"+
k h
dans V faible.
) = a( u , phvh ), on deduit de (2.6)
(3.2) R h ) b(phuh; phvh) + a(phuh-u, PhVh ) = 0
Si Iron prend dans (3.2) v = r v , et puisque p r v 4 v dans V

h h h h
fort on peut p a s s e r la limite :
--
1) L e s c designent des constantes diverses.

- 301 -
donc u
+ = u . Enfin, prenant v =
h
u
h
dans ( 3 . 2 ) on obtient
I(h)
b ( ~ h u hphuh)
, + a ( ~ h u h - uPhuh
' -u)' C(Ph uh -u, U)3 0
7
tlloh ( i ) (et en o u t r e \[(h) p h u h - + O dans W) .
3 . 2 --------
L e c a s ( i i ) du T h e o r e m e 2.1.
Evidemment l e s estimations du c a s 3. 1 sont valables.

P ~ . e n o n sdans ( 3 . 2 ) v = u - r u e t posons
h h h
Sh = PhUh - Ph rh U.
I1 vient
d'ob
t(h)b(phuh.~ ~ u ~ ) - + a ( $ ~ t(h)b(phuh.
>&)= ~ ~ r ~ u ) + a (hur h- u,p h d
et comme u W on peut u t i l i s e r (2.2) (i) . Donc
<
- ,
1 2
[ (h) \\\ph~hJI1
1
!lShll 2 'c ( f ( h ) 'l ( h )2)
On peut done d l a p r @ s(3.5) supposer p a r e x t r a c t ~ o nque phUh conver-
ge -
dans W faible e t d'aprhs ( 3 . 1 ) v e r s u, donc
- 302 - J . P. Aubin e t J. L.Lions
(3.7) phUh "j u dans W faible (1)
Mais on deduit de (3.3) que :
d'oh
b( 6 h. $h *. [-
1
(h)
a ( g; Shl5 Ib( p h r h ~ih)
J +
( (h)
6.
et grace A (2.7) , b ( d , J )d0 d'oh l e theoreme, c a r
h h
Ph Uh
- u = Jh+(phrb u-u) .
On deduit de (3.6) st de (2.7) que
Ilu - p u 11 I C
h h
(cas oh 9 = 1) e t p a r ailleurs, dans l e c a s oh 9 = 0 : (\(u- p u Ill < c
h h -
.
On interpole entre ces deux majorations: on en deduit l e resultat.
4. Exemple
4.1. On s e place dans l e cadre de 1 . 2 e t 2.2. On considere donc un

probleme de Neumann
avec A operateur elliptique d'ordre 2m .

kt2m
Si F = H 40) , a l o r s W = H ( 0 ) de s o r t e qu'A b( u , v )
correspond un operateur differentiel (soit B) d'ordre 2(kt2m).
Si on utilise l e s operateurs p introduits .dam J. P. Aubin
h et ClJ
et 4
qhap I1 , 2 et 63. , conduisant A (2.3) , on doit a l b r s choislr dans
(2.6)
(4.21
avec
(1) Cela, sous lthypoth&se , ~ ( h ) / < c. L1hypoth&se plus restrictive (2.7) intervient
pour l a convergence forte -
Le schema variationnel regularise (2.6) e s t alors un schema aux

differences finies, qui peut e t r e explicit6 . (cf. J. P.Aubin, [I;] , chap 11 .
6 4 4
1.
Si 0 < 0 j 1 , ltespace [w,V] est ltespace
Mais si
r -
C (R) designe Itespace des fonctions r fois continament
differentiables dans 5 on a, dtapr&s le theoreme de Sobolev fractionnai-
r e , (S. L. Sobolev [I] . J. P e e t r e [I] )
(R) C cr(5) ~i
(4.5) n
k+2m(l-0) > -2 + r.
n
Soit k donne tel que kt2m > - + r Alors s i l1on choisit 0 de
2
.
n 1
facon que 0 < (k - - - r) + 1 , on deduit du Theoreme 2.2 et de
2
(4.5) que
0
((phuh-u (1
5 c I ~ T.
p verifiant (4.3) .
c (Q)
P a r consequent
5 - --------------
Transposition
Supposons maintenant que
II i) W e s t dense dans
ii) F e s t dense dana Vt

V
et que
k
(5.2) A et A est un isomorphisme de W s u r F 1)
Alors le transpose 'A* est un isomorphisme de F 1 s u r W' qui

prolonge A donc :
(5.3) A est un isomorphisme de F' s u r W1 .
Soit f un Clement de W1. Considerons l a solution u dans F' de
llCquat ion
(5.4) Au = f
Sans les hypotheses du theoreme 2.1, l e schema variationnel regularise

(2.6) a encore un sens et l'on s e pose le probleme de savoir s i p u
h h
converge vers u dans l'espace F' .
5-2. Exemple.
011 s e place dans le cadre de 1.2. D1apres J. L. LionszE. Magenes
[l] , chap. 2 , on peut choisir des espaces W et F tels gue llequation
(5.4) conduise $Ides problemes aux limites non homoghes pour llopera-
IPI P
teur A = - ) D ( apq Dq ) .
Sous les hypotheses (5. I ) , (5-2) et celles du theoremes 2-1, s i

fE W 1, 3lors p u converge fortement vers
h h
0 dans F1 .
Demonstration
Soit pi (W1. Vh) le transpose
h
de p
et Ah lloperateur
de s u r lui meme defini par la forme sesquilineaire :
Vh
" ~ f e s ~ a c eW contient donc maintenant l e s conditions aux limites , I

oh (fh , vh ) est l a forme sesquilineaire s u r V mettant en antidualite

h
Vh et Vi .
La solution u de (2.6) est alors egale B
h
-1
(5-6) t hi = A h Ph f
et l e theoreme 2.1 affirme que s i f appartient h F, a l o r s

-1
(5.7) u - p h uh = Thf =(A -ph%lp\ )f
converge v e r s 0 dans W .
On montre de m&mequten remplapant A par A?
converge v e r s 0 dans W pour tout f de F

Les operateurs T* forment done un ensemble equicontinue de
h
aP(F, w ) .
-1 -1
T = T I = (A -p A p t )
P a r transposition, l e s operateurs
h h.
forment egalement un ensemble equicontinu de d ( w t ; F t )
h h h
.
Pour que l e s operateurs T converge6 simplement v e r s o dans
h
( W F ) , il suffit a l o r s que 11 Th f /IFl
converge vers o pour tout
element f dlun ensemble dense dans Wt. Prenons a l o r s Vt : Vt est
dense dans Wt et, dtapr&s l a premigre partie du theoreme 2-1 , on sait que
\I ~ h (IFf Q I ( T ~f 11 = J U -phuhJJ:end vers 0 .

On en deduit donc que si f 6 Vt
tend v e r s o avec h.
BIBLIOGRAPHIE
J.P. Aubin 611 Approximation des espaces de distributions et des opera-

teurs diffdrentiels .
Bull. s. M. F. Memoire N. 12, 1967
[2] Evaluation des e r r e u r s de troncature des approximations
des espaces de Sobolev. A paraitre dans Journal of Math. Analysis
and Applications.
Bramble [I] Conferences de ce Volume et Bibliographie de ce travail.
J. L. Lions [I) Espaces intermediaires entre espaces hilbertiens ;appli-

cations. Bull. Math. R.P.R. t.2 , (1958) , P. 419-492
J . L. Lions-E. Magenes b] Problemes aux limites non homog&nes et appli-
cations Vol. 1, 1968 .
J . L. Lions-J. Peetre ( 1 1 Sur une classe dlespaces dfinterpolation.
.
Institut des Hautes Etudes Scientifiques N. 19 - 1964
J. L. Lions-G. stamPacchia[1] Variational Inequalities - Comm. Pure Applied
Math .XX (1967), p. 493-519.
J. Peetre [I] Espaces dfinterpolation et theoreme de Soboleff. Annales
Institut Fourier, 16, 1 (1966), p. 279-317 .
J. Peetre-V. Thomee [I] .
On the rate of convergence for discrete initial
value problems. A paraitre.
S. L. Sobolev [l] Applications de lfanalyse Fonetionnelle en Physique
Mathdmatique - Leningrad. 1950 ,
(C. I. M. E. )
W. V. PETRYSHYN
ON THE APPROXIMATION-SOLVABILITY OF NONLINEAR FUNCTIONAL

EQUATIONS IN NORMED LINEAR SPACES
C o r s o tenuto a d Ispra. dal 3-11 Luglio 1967

ITERATIVE CONSTRUCTION OF FIXED POINTS OF CONTRACTIVE TYPE
MAPPINGS IN BANACH SPACES
by
W. V. PETRYSHYN
(University- Chicago)
Introductory remarks. It i s known that the problem of finding solutions

of nonlinear functional equations (e. g. , nonlinear integral equations, bounda-
r y value problems for nonlinear ordinary and partial differential equations,
etc.) can be formulated in t e r m s of finding fixed points of a given nonlinear
mapping defined on some subset of an infinite dimensional functional space.
F o r compact mappings a general existence theory of fixed points based
upon topological arguments has been constructed over a number of deca-
des Lassociated with the names of Brouwer, Poincare, Lefschetz, Schauder,
Leray, Tichonoff, Rothe, Krasnoselsky, Altman, and others). More recen-
tly, there has begum a systematic study of fixed points (their existence
and actual construction) of various classes of noncompact mappings (Brodski-
Milman
- , De M a r r p 8 1 , Browder p a , Sb, 5 , 6, 4 1 , Kirk 23 ,
Kachurovsky Bg , Edelstein E l , 12,131 , Belluce-Kirk - - , Gohde
rl]
[19,20], De P r i m a D 0-3 , Lions and Stampacchia D7]
- , Browder-Petryshyn
[8,9] , Kaniel p 2 ] , Shinbrot PO] , Petryshyn [31,32,33, 35, 3 6 1 ,
Opial i30) , Lees-Schultz [26], de Figueiredo L15, 161,
- Browder-de Figuei-
redo 1 7 1 , Petryshyn-Tucker [37 1 , and others) .

The purpose of this part of my talk is to survey, unify and extend a
number of recent results concerning the iterative construction- of fixed
points of noncompact contractive and strictly pseudocontractive mappings
acting in a Banach o r a Hilbert space. We first devote a section to the de-
finitions of various concepts used in this paper and to the historical develop-
ment of the iterative method in the construction of fixed points of contrac-
tive mappings; then we devote a section to the complete proofs of the
most recent results which, a s will be shown below, include all those men%-
W. V. Petryshyn
ned in t h e s u r v e y section. In the l a s t section we discuss t h e convergence of

fixed points of s t r i c t l y contractive mappings kA studied in [6], where
k is a r e a l number such that < 1. It is worth noting that o u r
0 <k
proofs, which use the a r g u m e n t s analogous t o those in (Is , 8 , 9, 3 0-1 a r e
s u r p r i s i n g l y simple.
1.1. Historical r e m a r k s and s t a t e m e n t s of r e s u l t s and definitions.
Let X b e a Banach space, C a closed convex s u b s e t of X and
A a (possibly) nonlinear mapping of C into X which i s contractive
( o r nonexpansive), i. e.,
(1.1) 1 Ax-Ay I( 5 IJx-y I( for a l l x and y in C.
The study of the iterative construction of fixed points of contractive map-

pings i s an extension of the c l a s s i c a l theory of nthe method of successive
approximationsn f o r s t r i c t l y contractive mappings developed by Picard,
Banach, Cacciopoli and o t h e r s , where A is s t r i c t l y contractive on C if
(1.2) 1 x-y \(for a l l x, y C and s o m e kc1
It i s known 24 that if a s t r i c t l y contractive o p e r a t o r A maps C into C,

then the P i c a r d sequence { x ) given by
ntl
x
ntl
= Ax
n
n+ 1
= A x (x given in C; n = 0, 1,2,
0 0
...)
converges t o the unique fixed point of A in C.
In the c a s e of contractive mappings A of C into C, the P i c a r d s e -
quence need not converge, t h e r e need not b e any fixed points, nor need the
fixed point b e unique if it does exist. T o obtain s o m e positive r e s u l t s on
the P i c a r d sequence f o r contractive mappings, we need to impose further
r e s t r i c t i o n s on the Banach s p a c e s X , the s u b s e t s C and o r the operator
A . F o r the s a k e of completeness and c l a r i t y we r e c a l l the following de-
finitions: X i s uniformly convex if f o r any 6 > 0 there exists a 6 ( 6 )>O
such that 1 X-y 1 1 r for IIx \I< 1 and
IIy II-< 1 implies that l(xty tyl<2(1-6 (t));
W.V. Petryshyn
X is sti-ictly convex if )( txt(1-t)y 1 <1 for all t e (0, 1) and all

x, y e H such that l(x 1 y (1 = 1; a mapping
I= P of X into X i s comple-
tely continuous if P i s continuous and compact; P i s weakly continuous
if x 2 x implies P x n Px, where nLn and fl?n denote the
n
weak and strong convergence, in X, respectively; P is strongly continuous
if x3 x implies P x +Px. Using Schauderls Existence Theorem 1 [38],
n n
Krasnoselsky [251proved the following positive result for completely conti-
nuous contractive mappings.
Proposition 1.1. If X is uniformly convex, C a closed bounded con-
vex subset of X and A a completely continuous contractive mapping of
C into C, then the sequence
- {xntl} determined by
(1.4) x n t 1 = 1/2Ax n t 1/2xn (x0 given in C; n = 0,1,2, ...)

converges to a fixed point of A in f2 .
In 39 Schaefer extended Proposition 1.1 in two directions by proving
the following.
Proposition 1.2. -
If X, C -
and A satisfy the conditions of Proposition
1.1 and A i s a fixed number such that 0 < a < 1, then the sequence
{xnt 1)
determined bv
converges to a fixed point of -

A in C.
Proposition 1.3. If
- X i s a real Hilbert space H, C a closed boun-
ded convex subset of H and A a weakly continuous contractive map of C
into C then x determined by (1.5) converges weakly to a fixed point of
n t1
A - C.
in
Let us add that the proofs in [25,39 1were essentially based on the
assertion that when X i s uniformly convex and A i s a contractive map
of C into
ntl
lar, i. e., (AX x
0
x )
X O
0 as -
n -+a -
C, then the mapping AX=A A + (1-A )I is asymptotically regu-
for any given x in C
0
- 312 - W. V. Petryshyn
(see Browder-Petryshyn [8] ). In 14 Edelstein extended the validity of

Proposition 1.2 to strictly convex Banach spaces X. In his proof Edel-
stein used Mazurls Theorem (see [29] ) since in case X i s strictly con-
vex we cannot conclude that AX i s asymptotically regular without first
proving the strong convergence of the Picard sequence { A: XI.
As was already noted in Introductory remarks, the study of the existen-
ce and of the iterative construction of fixed points has been recently carried
further for non-compact and contractive type mappings. Namely, suppos'e '
r > 0 in a real o r complex

C = B (0) denotes the closed ball of radius
r
Hilbert space H, S the boundary of B and A a contractive mapping
r r
of B into H which satisfies on S the Leray-Schauder condition
r r
(1.6) Ax - p x { 0 for all x in Sr and a n y p > l .
Applying the recent theory of monotone operators Browder [ 3a] (see also De
P r i m a [ l o ] ) proved that A has a fixed point in
r
B .
Using the above
existence theorem the writer 1 3 5 1 has derived the following result for de-
micompact mappings A ( a mapping P i s said to be demicompact if
whenever { un\ i s a bounded sequence and { un - Pun 1 i s strongly con-
vergent then there exists a strongly convergent subsequence { u n1. \ L
Proposition 1.4. Let -A be a demicompact contractive mapping of
B
r
into H which satisfies (1.6) . Then for any x o Br
~ and any
E ( 0 , l ) the sequence determined by the retraction-iteration method
i'n+l\
converges to a fixed point of A in Br. -

Proposition 1.4 extends in two directions the iterative construction
of fixed points of contractive mappings. First, we replace the assumption of
compactness of A by demicompactness; second, for such mappings, instead
of assuming that A maps B into B we assume only that A
r r
satisfies the Leray-Schauder condition on S
In our view the latter result
r'
for demicompact operators is the main advancement of the theory. Note
that if A is compact, then it is clearly demicompact but the converse is
not true. For example, it is easy to see that A = . T+ S, where T is
strictly contractive and S is compact, is demicompact but it is certainly
not compact. Clearly, if A maps B
into B then r = 1 for each
r r n
n and (1.7) reduces to the method (1.5). In this case, Proposition 1.4.
i s valid for an arbitrary closed bounded convex subset C without any chan-
ge in its proof not only in Hilbert spaces but also in uniformly convex
Banach spaces. Thus, in this case, Propositions 1.1. and 1.2 follow from our
Proposition 1.4. Let us add that (1.6) i s implied by the following conditions:
for all x in S
r
Remark 1.1. It was shown by de Figueiredo-Karlovitz F8 ]that if
the retraction R C of X onto C = B (0) i s contractive for a Banach spa-
1
ce X of dimension > 2 , then X i s a Hilbert space. This fact indicates
that the retraction-iteration method (1.7) i s essentially restricted to Hilbert
spaces.
Further progress in the theory of the iterative construction of fixed
points was made in Browder-Petryshyn [8]where a number of results
concerning the Picard sequence (A"x\ was established. We mention here
only two results from [8 ]which a r e directly relevant to our discussion.
We first reaall that a mapping P of X into -
X is said to be strongly
closed' *I if for any sequence tun/ in X with u - - \ u and Pu -\v we
n n
have Pu = v.
(*)some authors refer to the above property a s ndemiclosedness. U We prefer

the tirill ?strongly closedn(employed alsi, ill pg)
since it i s more in harmony
with the concept of wstrong'continuityuused, especially, by Soviet mathematicians
On the other hand, the therm ndemiclosednessn of P was used in [41 , 24 ,
361 to desci-ibe the situationu u +un and Pu --Iv implies that Pu = v"
n n
which seems to be more in harmony with the concept of ndemocontinuityn(i.
U n + ~ implies Pun-Pu
J )introduced by Browder and accepted by the ma-
thematical community.
- 314 -
W. V. Petrhyshyn
Proposition 1.5. Let

- A be an asympotically regular. (i.
- e.
n
A"" x-A x + 0 a s - n ,w for every x -
in X) Contractive mapping of X
into X
- such that I-A i s strongly closed. Suppose that the s e t F(A) of
-
fixed points of A is not empty. Then for any xn € X the weak limit of
any weakly convergent subsequence 1 A"(') x0 \ -of the Picard sequence
{ inx0 /-l i e s in F(A) . In particular, -if X - i s reflexive and F consists of
n
a single point y,then
- A x0 -y .
We r e m a r k that Propostion. 1.5 i s valid, in particular, when X is
a Hilbert space o r a Banach space having a weakly continuous duality map-
ping ( * ) J. In this case it has been shown in [ 36 ] (see also 15.91 ) using
the theory of monotone and J-monotone operators that if A i2 contracti-
ve then (I-A) i s strongly closed. Hence in this case we drop in Proposi-
tion 1.5 the assumption that (I-A) i s strongly closed (see Theorem 4 in
Remark 1.2. Proposition 1.5 was stated for mappings A of X into

X instead of C into C in order to employ the theory of monotone'and
J-monotone~operators in the proof of strong closedness of (I-A). This i s
not a restriction in case of a Hilbert space since a contractive mapping
defined on C can be extended to a contractive mapping defined on all of
H. But it is a restriction in case of a Banach space since, in general,
such an extension theorem i s not available.
Proposition 1.6. Let -A be an asymptotically regular contractive
mapping of X into
- X with a nonempty s e t F(A) of fixed points. Suppose
further that A satisfies the followinn condition.
(m) (I-A) maps bounded closed sets into closed sets.
Then for each x in X the Picard sequence convergence to a fixed point
0
(*)L% X be the dual space of X and (f, x) the value of the linear functional f
in X? a s the element x of X. Let y be a continuous strictly increasing real va-
.
lues function defined on R+ = {t:t > 01 with p (0) = 0 A mapping J of X
into (see ) i s called a duaiity mapping with a gauge function (r if
1
(Jx, x) = JX ( x 1 and 11 J x I(=
r ( 11 x (1 ) for every x in X.
W. V. Petrhyshyn
in F(A ).
-
It was shown in [8,39] that for any fixed A E(0,l) the mapping
A' = A I t (1-A ) A (or AX = A + (1-A )I) i s a contractive asymptotical-
ly regular mapping with F ( A i ) = F(A) (or F(Ah ) = F(A)) provided that A
i s a contractive mapping of a uniformly convex Banach space X into X
with a nonempty s e t F(A) of fixed points. Since, furthermore, forJ€(O, 1)
we have I - Ata =(I-A)(I-A) ( o r I-AX =A(I-A)) we s e e that A satisfies
condition ( a )if and only if A\ (or AA ) does also. Using the above re-
m a r k s we have the following corrollary of Proposition 1.6.
Proposition 1.7. Let A be a contractive mapping of a uniformly con-
-
vex Banach space X into X with a nonempty s e t F(A) of fixed points.
Suppose that A -
satisfies condition (0).
Then the sequence
(xn.1)
determi-
-
ned by (1.5) o r by
converges to a fixed point of A.

It will be shown in the next section that Proposition 1.7 remains valid
when A i s assumed only to be a contractive mapping of a closed bounded
convex subset C of a uniformly convex Banach space X into C. In this
c a s e we drop the assumption that A has a nonempty set F(A) of fixed
points since in this case the existence of fixed points of A in C follows
from the recent existence theory proved independently by Browder [4],
Kirk 1231 and Gohde 1191 . Since, a s i s not hard to see, the demicompac-
tness (and, in particular, compactness) of A implies the validity of con-
dibon (a)(*),
it follows the modified Proposition 1.7 i s an extension of Pro-
position 1.4 for the case when A map C into C.
(*)1t was stated in [ 8 1 that condition ( a ) i s equivalent to the demicompact-

ness. This i s c l e a r 6 not the case since the mapping A=I satisfies condition
( t r ) but certainly is not demicompact. The authors of 181 caught this inac-
curacy too late for the correction to be inserted in the paper 8
- 316 - W.V. Petrhyshyn
When in Proposition 1.5 the space X is a Banach space with a wea-

kly continuous duality mapping (and, in particular, a Hilbert space) Opial
/30] strengthened the result of Proposition 1.5 (or rather of Theorem 4
in 8 ) by showing that the entire sequence Anx0) is necessarily weakly
convergent (see also Theorems 7 and 8 in [9) ) . More specifically, the
following was provetl in [30] .
Proposition
---- 1.8. --
Let C be a closed set in a uniformly convex Banach
-
space X -
having a weakly continuous duality mapping and let A be a con-
tractive mapping of C into C with at least one fixed point in
-- C . Then for
any
- x in C and any A( ( 0 , l ) the Picard sequence
0
convergent to a fixed point of A in C. -
We recall that if C is bounded, then the existence of at least one
fixed point of A in C follows from the results in @, 19,233. Let us add
that Opial did not use the theory of monotone o r J-monotone operators to
prove the strong closedness of (I-A). His arguments were similar to
Schaeferfs [39] . We recall that for weakly continuous contractions in real
Hilbert spaces the weak convergence of [A; xO}was first proved by
Schaefer 1391 - . The extension of this result to general contractions was
thus carried out in two stages, the proof of Proposition 1.5 by Browder-
Petryshyn [8], and the proof of Proposition 1.8 by Opial PO] .
Remark 1.3. It was already noted in [9] that Gohdels assertion in
1 1 4 that the weak convergence result in E9] follows already from the
existence theorem i s inaccurate since the proof by Schaefer uses two facts,
the existence of !it least one fixed point and the fact that if a subsequence
of {A; xo) converges -ueakly to y, then y is a fixed point of A. This
latter fact follows a s in the proof of Theorem 7 i n p ] and Opial has uti-
lized this argument together with a simplified form of'schaefer's proof.
Let us add that under the stronger assumption that the strong limit
set of the iterates i s nonempty, convergence results for contractions have
been given by Edelstein E l , 12,131 and Gohde LO]. However, there seems
W. V. Petryshyn
to be no explicit way of extracting convergent subsequence and therefore the

methods a r e not really constructive.
In [9] Browder and the writer studied various classes of nonlinear
operators (contractive, strictly pseudocontractive, psedncontractive) mapping
a closed bounded convex subset C of a real Hilbert space H into C
o r into H. F o r the purposes of our present discussion we shall mention
only those results from [9] which bear direct relation to the iterative con-
struction of fixed points by means of the processes (1.5) o r (1.7) .
As a first example we state the following proposition which for a
Hilbert spdce i s an extension of Proposition 1.8 .
Proposition 1.9. -- A be a contractive mapping of
Let - - C into- H. Suppose
that if u 6 C and
-if u = R Au, then u i s a fixed point of A. Then
- A has-
-
fixed points in
C
C and the sequence kntl) determined by the retraction-
iteration method
converges weakly to a fixed point of A in C.

-
C = B (0) then our second assumption reduces tp the
Note that if
r
Lerady-Schauder condition (1.6) and in this case the retraction-iteration method
is practically constructible.
Following b] we say that a bounded closed convex s e t -
C is unifor-
mly smooth with smoothing constant R > 0 if and only if for each boundary
point x in C
1
1) C has only one supporting hyperplane at x (i.e., (xl, V) = c
1 0
and (x u) c for all u in C)
1' ->
I
0
2) There exists
0
in H such that u -x
u
0 1
= R and u -u < R
0
I 1 11-
for all u(C (i.e., CCB (u ) and x 6 ~ B ( U )).
R 0 1 0
Our next result i s the following extension of Proposition 1.4.
Proposition 1.10. Suppose that in addition to the conditions of Proposi-
- 318 -
W. V. Petryshyn
tiion 1.9 we a,ssume that A i s a demicompact map of C into - H, where

is uniformly smooth with smoothing constant R . Then the sequence
{xntl) determined by the retraction-iteration method (1.7A) converges to a
fixed point of A in C.
-
Let us note that the main difficulty in proving Proposition 1.10 l i e s in
the proof of the demicompactness of R A. The r a t h e r tricky proof of this
C
in [9] utilizes the idea (which, essentially, was used for the first time
by the w r i t e r in [35] ) that outside some narrow band about the surface of
C the retraction is strictly contractive. Needless to say the chief difficul-
ty in the application of Propositions 1.9 and 1.10 to actual problems l i e s in
the fact that unless C i s a ball there i s no efficient way of construction
the retraction mappings R,
L
.
Following p]we s a y that the mapping A of C into H is -
stric-
tly pseudo-contractive if there exists a constant k < 1 such that
2
IIAX-A~ 112 51x-yll t kll (I-A)X - (I-A)~II f o r all x, y c C.
It has been shown in 1 9 3 that for every fixed t such that 0 < t -< 1-k the
mapping A = tA + (1-t)I i s contractive, A has the s a m e fixed points a s
t t
A in C and (A ) = i I + ( I - ) ~ ) A =~ A where
tA .
7
= 1-(I- )t with Y
t < 1-k for any fixedl\€(0,1) if and only if > k. Consequently we have
T
the following extensions of Propositions (1.9) and (1.10) for A mapping
C into C .
Proposition 1.11. Let A be a strictly pseudocontractive mapping of
C into
- C. Then , for any x 6 C and any
0
yI such that k <
converges weakly to a fixed point of A in C. If additionally we assume that
-
A i s demicompact, then
{"
A
d
x converges strongly.
Similar result holds for the retraction-iteration method (1.7) when it i s
applied to strictly pseudocontractive mappings A of B (0) into H
r
which on S satisfy the Leray-Schauder condition (1.6) ,
r
Remark 1.4. In [ g l t h e authors also discuss the construction of fixed
W, V. Petryshyn
points of a pseudocontractive mapping A of B (0) into H with A satisfy-

r
ing the Leray-Schauder condition on S , where A is s a i d to be pseudocon-
--
r
t r a c t i v e if
1) Ax - ny[ 2 <IIx - Yll +I](I-A)x - (I - A ) ~ I for all x, y 6 C.
The mothod used in the construction of fixed points of s u c h mappings i s e s -

sentially based on t h e s t r i c t contraction mapping principle combined with
the radial retraction R. The idea is that for each fixed s with 0 < s < 1
we construct the fixed points
s r
u
in B (0) of the mapping A = sA
s
+
+ ( 1 - s ) uo f o r any uo in B r (0) by the method
n
(1.10) u = l i m (u ) = l i m
' n n
{ R(l - ( 1 ~ ~ ) w) , wo 6 Br(o)
w h e r e the r e a l number /A l i e s in a c e r t a i n interval ( s e e [9] and then

show that u +u,
S
where u is a fixed point of A closest to u
0
. We
will not investigate the method (1.10) s i n c e this is outside t h e scope of o u r
discussion.
Let X be a r e a l s e p a r a b l e Banach space. A p a i r of sequence ( X ,
{ p n ) ) , w h e r e each X is a finite dime1:sional subspace of X and P
n) {
n n
i s a l i n e a r projection of X onto Xn , is s a i d t o be a projectionally
cqmplete systeln if P x +x
n
f o r each x in X ( s e e [34] ) .
Definition. A nonlinear mapping A of D(A) ( C X) into X will
be called gy -compact
-- if P A is continuous in X f o r a l l sufficien-
n n
tly l a r g e n and if t h e r e e x i s t s a constant Y -
> 0 such that f o r any p
dominating y (i. e., p 2 7 if )c > 0 and p > y if = 0) and any bounded
sequence {xn} with x$ Xn n D(A) the s t r o n g convergence of the sequence

{P Ax - pxn) implies the existence of a strongly convergent subsequence
bn) :noanelement x in D(A) s u c h that x +x
n.1
and
W. V. Petryshyn
Note that our present definition of PY -compactness is a generaliza-

tion of the concept of P-compactness introduced and studied by the writer
in [31,32,33J . Indeed, when Y = 0 . D(A) = X and the sequences
{xJ
and P a r e such that
n
X n C Xn+l(n = 1.2.3 ,...). nU Xn = Xand 1 P n1 -< K , K>- 1, then the

Po-compactness i s just the P-compactness. In what follows we shall retain the
the term P-compactness whenever IX, = 0 , Let us remark that upon a clo-
s e r examination of the proofs of fixed point theorems in L31,32,33J it be-
comes obvious that in those proofs we have only used the properties of
A and X which a r e embodied in the requirement that A be P -corn-
1
pact (according to our extended definition) a s a mapping from B-(0)
1-
into X. However, in the applications of fixed point theorems in [32,33]

to monotone operators we still have to use the definition a s given in 31,32 .
It turns out that our present extended definition of Py -compactness is
more suitable for theoretical and applicational purposes and, particularly,
for the purpose of unification of various results.
Returning to our discussion of iterative methods we would like to
finish this section by stating the following result proved in [37] for
P-compact mappings.
Proposition 1.12. Let - C be a closed bounded convex subset of a real
uniformlv convex Banach space having a projectionallg complete system.
If A is a contractive P -compact mapping of C into C, then the -
1
Picard sequence determined by (1.5) converges to a fixed point
('nt~ 1
2.1 Convergence proofs. In this section we prove three theorems
concerning the strong and weak convergence of the Picard sequences
( A; x \ under certain conditions on A and X, where A is a contrac-
tive o r a strictly pseudocontractive mapping of C into C. We have seen
in the previous section that in each case the strong o r the weak convergen-
W. V. Petryshyn
ce of ( A x~) depended upon the existence of fixed points of A in C. It

turns out that in c a s e A is a contractive mapping (which is the case
that we a r e interested in this discussion) the properties and the additional
conditions on A which insure either the strong o r the weak conver-
gence of the iterative method (1.5) admit at the same time a simple proof
of an existence theorem which, a s will be seen below, i s general enough
to cover all c a s e s surveyed in Section 1.1. Thus, there i s no necessity to
r e f e r the reader to the m o r e general existence theorems proved in
[4, 19, 233 by rather complicated arguments .
Unless explicitly stated otherwise in this section the set C will al-
ways be assumed to be a closed bounded convex subset of a given space
Theorem 1. -
Let be a general Banach space and let be a con-
tractive mapping of a subset C -
of X -
into C such that (I-A)(C) --
is a
closed s e t in X. Then
-A has a fixed point in C.
-
Proof. Assume without l o s s of generality that the origin 0 ( C.
Let r be a sequence of numbers such that 0 < r < 1 for each n
n n
and r -1
a s n+m. It i s obvious that A = r A i s a strictly contrac-
n n n
tive mapping of C into C. Hence, by the s t r i c t contraction mapping
principle, for each n there exists a unique point in C u such that
n
A u = u F o r the sequence {un
n n n'
} thus determined we have
u
n
- Au n = rn Aun -Au
n
= (r
n
- 1 ) A un- 0 a s n+m
since r -1
n
and {AUA C C is bounded. Hence 0 l i e s in the closu-
re of (I - A)(C) . Since, by assumption, (I-A)(C) i s closed, 0 t (I-A(C).
and therefore A has a fixed point in C.
Lemma 1. If in addition to conditions of Theorem 1 we assume that
X is strictly convex, then the s e t F(A) of fixed points of A in
- C-
is
a closed convex set.
W. V. Petryshyn
-
Proof. By Theorem 1, F(A) # 6 . Suppose that x and y a r e in
F(A) and z = tx t (1-t)y for t 6 (0,l). Then, by the contractivity of A
and convexity of C, z C and
Since X i s strictly convex the above inequality implies that Ax - Az =

= a(Az - Ay) f o r some real number a. Therefore, the vector A z is on
the line through Ax and Ay and, hence, on the line through x and y.
But we ;!so have that 1) AZ- AX^ -< 1 z-xl and I < ~ I Z - ~ ~.I Hence
A Z - A ~ ~
we must have that Az = z . The closedness of F(A) is obvious .

Remark 2.1. Lemma 1 i s valid under the assumption that X is
strictly convex and the contractive mapping A has a nonempty set F(A)
of fixed points.
Lemma 2. Let - X be -a uniformly convex Banach space and A
a contractive mapping of C (cX) into C. Then for any fixed A€ (0, 1)
the mapping AA = A l t ( 1 - ))A i s asymptotically regular (i.e.,
A?' n
xO - A> xo-O a s n +m - -
for x
0
C) and
- A> has the same fixed
points a s A.
Proof. It i s obvious that F(A) = F(AA ) and that AA maps C
into C and i s contractive. Now let u be a fixed point of A in C
and, for any given x in C, let be the Picard sequence. Then
0
x0 - u 1. 1 1 ~ : ~ ~ xo-~uA xo- ul for all n .

Thus 1 An xO - u I
converges to some d > 0. If d = 0, then
0- 0
n
Altl
Xo
-
A x --r 0 and the proof is finished. Suppose now that d > 0.
0 0
Since Au = u and
- 323 -
W. V. Petryshyn
and since ( An t 1 x I
-u +d O , ~ I ~xO-u
; J+dO and P A n x0-uI 5 1 A ~ X ~ - U I it
/
follows from the uniform convexity of X that
since A- 1 = ( 1 - ( I - A ) ; hence the proof of Lemma 2 i s complete.

We now prove the first main theorem (compare it with Proposition
1.6) of this section by the simple arguments analogous to those used in r8).
Theorem 2. Let -- X be a uniformly convex Banach space and A -
a
contractive mapping of the subset --
C of X -
into C. Suppose further that
(I-A) satisfies condition (d),i.-

e., (I-A) -.maps every (bounded) closed
subset
- -of C --- into a closed subset. Then the set F(A) of fixed points .-
--of A -
in C - i s a nonempty closed convex set and -
for
- each x0 in-- C
and 1( ( 0 , l ) the sequence
- -- 1
An xO} converges (strongly)
------ to a fixed
point of A --
in C.
Proof. Since, by condition
- (4)
, (I-A) (C) is a closed set it follows
from Theorem 1 that F(A) # fl . Since uniform convexity of X implies
its strict convexity, Lemma 1 implies that F(A) i s a closed convex set.
If u is a fixed point of A, and thfrcfore of A the sequence
1 i s decreasing .
It suffices therefore to show that there exists a subsequence of
IA; xO 1 which converges strongly to a fixed point of A in C. Let G
be the strong closure of the s e t [ A; xO} . BY ~ e r n r n a2,
n
(I-AA ) (AA xo) -+0 a s n m . Hence 0 lies in the strong closure of
(I-Al )(C) . But , since for any AE (0, I), A ( 1 ( A ) we s e e that
(1-Al ) satisfies condition ( 4 )if and only if (I-A) does also. Hence, by
condition (4)
, the set (I-AA )(C) is closed since G is closed (and
necessarily bounded) and therefore 0 lies in (1-AA )(C) . Hencc there
exists a strongly convergent subsequence of [ A; xu} which converges to
an element y in G such that (I-A ) = 0, i.e. , y i s a fixed point
Y
We shall now prove that in addition to completely continuous o p e r a t o r s

the two c l a s s e s of mappings which satisfy condition (4) a r e demicompact
and P -compact mappings studied by the w r i t e r .
1
L e m m a 3. L e t - X b e a general Banach s p a c e and A a continuous
-demicompact
- mapping of D(A)(C X) -
into X . Then fi s a t i s f i e s condition-
Proof. L e t
-.-
Q be a bounded closed s e t in D(A) and let / uk]
.
C J
be a sequence in Q such that (I-A)u +v a s n+m. Since
n
u is
a bounded sequence, and 1, (I-A)un
n
) i s strongly convergent, by demicompac-
I \
tness of A, t h e r e e x i s t s a strongly convergent subsequence
IUni}
Hence the closedness of Q and the continuity of A imply that u n . + u
1
for some u in Q with (I-A)u,,-+ (I-A)u = v and L e m m a 3 is proved.

1
-L e m m a 4. -
Let X be a r e a l Banach s p a c e with a projectionally com-
plete
- ([xn \, pn}). L A A be a Lipschitrian - P -compact mapping of
1
D(A)CX --
into 2 . Then
- A is demicornpact and hence s a t i s f i e s condition
P.
Proof. L e t
- Q b e a bounded closed s e t in D(A) and l e t
be a sequence in Q s u c h that (I-A)uk - + ~ a s k -c w . Since the
system ( 1 ~ , ~{Pn))
) is projectionally complete in X , f o r each inte-
ger k and = 1 t h e r e e x i s t s a n integer n(k) (which we c a n
k
and shall a s s u m e that n(k) > k) s u c h that Ilu k - P .(kIUk 1 < 'k ' This
and t h e relation (I-A)u -+ v imply that with w - in X
k n(k) - Pn(k)Uk n(k)
we have
1 Pn(kfWn(k)-Wn(k) + v 1 -< 1Pn(k) Awn(k)-Pn(k) Auk 1
Since { P n ) is uniformly bounded, say, by K-

> 1 and A is
Lipschitzian, s a y with constant L > 0, it follows f r o m the above inequa-
W. V. Petryshyn
Thus, since A is
P -compact, there exists a subsequence
1 ~n(j)I /
IWn(k) 1
P . Aw .*Au
and an element u in D(A) such that w .*u
a s n(j) -co . n(J)
Consequently, u.-uI/( -W
and
1 IJu
~ ( J I "(J) J j n(j) II +
+ H wn(j) ,-u II- t o as n(j) --coo. Since { uj ] , being a subsequence of

lies in Q and Q is closed we must have that u e Q and Au = v,
i.e., A i s demicompact. The last assertion of Lemma 4 follows from
the above proof and Lemma 3.
Remark 2.2. In virtue of Lemmas 3 and 4 and the observations fol-
lowing Proposition 1.4, Theorem 2 implies the validity of all propositions
of the previous section asserting the (strong) convergence of the sequence
{A; xO) provided that in these propositions A maps a closed bounded
convex subset C of a uniformly convex Banach space X into C.
To derive the strong convergence of the retraction-iteration method,
Remark 1.1 implies that we need consider only Hilbert spaces H. In rip-
tue of Lemma 3, the strong convergence of the retraction-iteration me-
thod (1.7) o r (1.7A) will follow from the following lemma.
Lemma 5. -
Let A be a contractive demicompact mapping of a closed
bounded convex subset C of r e a l space H into H, where C is unifor-
-
mly smooth with smothing constant R. If
- R denotes the retraction of
C
H onto- C, then the mapping RC A is demicompact.
-
Note. In case C = B (0),
r
Lemma 5 was first proved by the writer
in[35] . For C, which i s uniformly smooth with smoothing constant R,
Lemma 5 was proved in [9] by the arguments which we use here. It is
essentially based on the following lemma whose rather tricky proof is gi-
ven in [9 1 .
Lemma 6. If - C i s ,uniformly smooth with smoothing constant R and
A& (0, 1). t < 1-k if and only if > k. Indeed, if t < 1-k and 1 > Y > 0
i s to hold for any 1E (0, 1) it follows that
I must satisfy the inequali-
ty r = 1-t(1- ) > 1 - (1-k)(l-A) > k . On the other hand, if k<

= 1-t(1-l\ ) is to hold for any) C (0.1) the parameter t must satisfy
l=
the inequality k < 1-t, i.e., t < 1-k. Now, since A -1 = (1-f)(A-I) , it
follows that if A is demicompact then s o is Ab, and consequently, by
Theorem 2, the Picard sequence { (A~); x O j= {A" xO) converges to a
fixed point of A in C,
To justify the obeervation following Proposition 1.11, concerning the
convergence of the retraction-iteration method for a strictly pseudocontrac-
tive mapping of B (0) into H, by the second part of Proposition 1.11 and
'I
Lemma 5, it suffices to show that
tion on S i. e., A] x -j'$
r
A satisfies the Leray-Schauder condi-
x # 0 for all x in S and any > 1.
r ' r
Suppose, to the cbntrary, that AI xO- F o x O = 0 for some x in S
0 r
and some
A~ =-
0 1- r
> 1 , since Po
> 1 , in contradiction to (1.6) .
Suppose now that we drop the additional assumption that A satisfies
condition (d) . Without this condition, thus far it was only possible (see
I.
[39, 8, 30 , 9 ) to prove the weak convergence of ( xo) . In discus-
sing the weak convergence of {A; x o j we will use the arguments simi-
l a r to those applied in [a, 30,9] . The two succeeding lemmas were proved
in a somewhat different form in [30] .
Lemma 7 . Let
-X be a uniformly convex Banach space having a wea-
kly continuous duality mapping J of
- X into X*
- . Let {xn] be a se-
w
quence in X such that xn4xo . Then there exists a subsequence

/ xmi 01 [xn] such that f o r any -
in X
with equality holding in -

(2.2) if and only if x =x
0'
W. V. Petryshyn
Proof. It is obvious that there exists a subsequence xm) I

[xn)
such that 1 xm-xOll doand 1 xm-x 1 +d a s m + m, where d0 and
d a r e some real numbers. Hence, since x -x 0 implies
m 0
-x ) A0, the definition of J and the passage to the limit in the ine.
J(xm o
quality
imply that /1 (d0)d0 -< P (d )d, from which (2.2) follows. To complete the
0
proof suppose d =d
o .
in (2.2) Since llxm (tx+(l-t)xo) ) 5 t Jkm-x +
- 1
t (I-t) l1xm - xO 1 for any t in it follows from the first part of
Lemma 7 and our assumption d = d that
0
I Xm-X~I + Ilxm-x Jdd0

and [x -(txt(l-t)xg)ll
m
+d
0'
Hence, by uniform convexity of X, it follows that 11 ( x ~ - x ~ ) - ( x ~ - xo>) ~ ~ +

i.e., x=x
0'
An immediate consequence of Lemma 7 is the following result which
is essential in our discussion.
Lemma 8. Let - X be a uniformly convex Banach space with a weakly
continuous duality mapping into
- is a contrac-
tive mapping of a closed convex s e t C of -
X into X, then the map-
- (I-A) is strongly closed.

ping
Proof. Let {xn 1 be a sequence in C which converges weakly
to some x
0
in X and l e t {xn-Axn ] converge strongly to some y
0
in X. Since C i s closed and convex and hence weakly closed, x
0
lies in C. Hence it suffiees to show that (I-A)x - . Let [ xm be
a subsequence of (x
1 n
1 such that lim
0 - Yo
1 xm -x 0 11 exists. Since
x -Ax -+ yo, there
m m
exists a sequence z +0
m
such that Ax =
m
= xm
-
yo t zm and
1) lim
- 11-
which implies that > lm
im x , It follows from
Xm-x~ m m-y~-Axo~ f
this and Lemma 7 that x - AX
O-Yo 0'
Remark 2.4. As was already neted, Lemma 8 was first proved in a
more general setting in
[51 using the theory of J-monotone operators under
the assumption that A is defined on all of X. In case of a Hilbert spa-
ce a somewhat different proof i s given in
191 .
Remark 2.5. Using Lemma 8 it is easy to show that if X has pro-
perties assumed in Lemma 8 and A is a contractive .mapping of a clo-
sed bounded convex subset C of X into X , then the s e t (I.-A)(C)
i s closed . Indeed, let ( un\ be a sequence in C with (I-A)u +v
n n
as
n --, co. We need to show that
0
v
lies in (I-A)(C) . Since X, being uni-
formly convex, i s reflexive, we rnay replace un by a subsequence, [ which
we again denote by (un 1, u AU
n
for some u in X. Since
such that
0 0
C is closed and convex and hence weakly closed, u lies in C. Hence,
0
by Lemma 8, (I-A)uO = v Consequently by Theorem 1, if A maps C
0'
into C, then A has fixed points in C.
We now prove the second main theorem (compare it with Proposition
1.8) in this section by the arguments similar to those used in [39,30,1
in the case of Hilbert space.
Theorem 3. Let - C be a closed bounded convex subset of a uniformly
convex Banach space having a weakly continuous duality mapping J of X
-
into X* -
- and let A be a contractive mapping of into C. Then for
C -
any
- x in C and any),E ( 0 , l ) the sequence {A" xO) converges weakly
0 -
-C.
to a fixed point of A in
Proof. By Theorem 1, Remark 2.4 and Lemma 1, A has a nonem-
pty closed convex s e t F(A) of fixed points in C and F(A) = F(A ) .
Since, for any u in F(A), the sequence [1 xO - ]
UJ is decreasing,
we can define the nonnegative r e a l valued function g(u) from F(A) into
+
- 330 - W.V. Petryshyn
It is easy to s e e that g(u) i s a continuous convex function and, thus,

weakly lower semicontinuous on F(A). Since F(A) i s weakly compact, the-
r e exists an element u in F(A) such that
0
Furthermore, u
is unique. Indeed, suppose there exists another point
0
v in F(A) such that d = g(v ). Then, by the convexity of g(u), it fol-
0 0 0
lows that, for any t ~ [ 0 ,11 , g(tuo t (1-t)v ) < tg(uo)t(l-t)g(vo) = do. Hence,
0 -
a s n -+ m , we have
11Anx0-u0 1 +do, I~A; x0-vOl(+dO and IA; xo-(tu 0t(1-t)v 0 1+d 0
from which, on account of uniform convexity of X , it follows that
1 ( A ~x0-u 0)-(Anxo-vo)ll +o, i. e., uo = vo .

We now prove that
n
A x ~u
A 0 0'
If not, then since {An x
0
is ]
bounded and X i s reflexive, there exists a subsequence [Ani xO) and
i
an element u in C such that A x \u and u # u Since, by
I 0 O' ni
Lemma 2, At\ i s asymptoticallji regular it follows that (I-AXA x d -,0
"i
and A x a uGC. Hence, by Lemma 8, (I-A)u = 0, i.e. , u 6 F(A).
0
Finally, by Lemma 7 ,
This yields the contradiction of the definition of d

Hence we must ha-
0'
ve that u = u, i. e., every weakly convergent subsequence of
0 xol I4
converges weakly to u
0'
Hence
01 n.
A x converges weakly to u
0 '
Remark 2.6. F o r a somewhat different proof of Theorem 3 in c a s e
X is a Hilbert space s e e [91 . We r e m a r k that the class of Banach
spaces X having weakly continuous duality mappings J includes a l l
Hilbert spaces a s well a s the sequence spaces 1 for 1 < p < a but,
a s was shown in (71 , it does not include any of the P
L spaces except
W, V. Petryshyn
',I , As was already noted, when F(A) has just one point, Theorem 3
was first proved in [8] by using the theory of J-monotone operators. In
its present form it was proved in [30] .
We remark that the weak convergence of the retraction-iteration me-
thod (1.7A) (i. e., Proposition 1.9) and of the sequence {A; (i. e. , pro-
position 1.11) follows from Theorem 3 and the observations following Re-
mark 2 . 3 since, a s is not hard to prove the mappings
a r e contractive from C to C and satisfy the needed conditions on the
RCA and
AY
boundary of C.
For the sake of completeness we end this section by proving Edel-
stein's extension of Proposition 1.2 to strictly convex Banach spaces (see
Theorem 4. If A is a completely continuous contractive mapping

-
of a subset C of a strictly convex Banach spaces into C, then for
X - P
a fixed point of A in C. -
Proof. It suffices to show that there exists a subsequence of
[A; x s i c h converges to a fixed point of A in C. F i r s t note that,
by Schauderts fixed point principle o r by Theorem 1 above, A has nonem-
pty set F(A) of fixed points in C which is closed, convex and
F(A) = F(AA ). Next, let M be the convex closure of the set A(C)u{x 1.
0
rn'
By Masurts he or em [2d M is compact. Since {A: x o ) ~f i , there
exists a subsequence
If
AA
I
xO and u in C such that
u F(A) , then Theorem 4 follows. Suppose, to the contrary, that
A xj u .
o
/implies .thatThenfor, some

u F(A) for each x in F(A) , the strict convexity of X
a > 0 , depending on x, a kiJ)-xl - I I A ~ u-x 1 >0.
n.1
Since A x +u, it follows that, for all suft'iciently large n. ,
0 1
W. V. Petryshyn
Since XCF(A ) and AA is contractive, the latter two inequalities and the
strict convexity of X imply that I I A ~ xo -xll< Xu-x(- a/2 for all suffi-
i
ciently large p which contradicts the convergence A A x -+u.
0
3.1. Convergence of fixed points of strictly contractive mappings kA.
Let A be a contractive mapping of a subset C (closed bounded and
convex) of a Banach space X into C and let kn] be a sequence of real
numbers such that k - c l (n+m) and O < k <I. Let v be an arbitrary
n n 0
.
point of C Then for each n the mappings A (x)=k Ax + (1-k )v is stric-
n n n 0
tly contractive mapping of C into C. Hence, for each n, there exists a unique
point u in C such that A u =u Our problem is to discuss the conditions
n n n n'
under which, a s n-cao, the sequence (u,] converges to a fixed point of A in
C. In case X is a uniformly convex Banach space (in particular, a Hilbert spa-
ce) with X* strictly convex and with a weakly continuous duality mapping
J of X into X* and A is a contractive mapping of X into X such that
A maps C into C , this problem was completely settled in an interesting
paper by Browder [63 . Browderts discussion rests heavily upon the theory
of J-monotone operators introduced in [53 and further developed in [7]
since, a s shown by Browder, there exists a nice nonnection between such
operators and contractive mappings defined on the whole space. X .
The purpose of this section is to rederive ~ r o w d d r f sresults [6]
for a contractive mapping A which is defined only onthe subset C , i.e.,
A maps C into C. It is worth noting that one cannot use the theory of
J-monotone operators when A i s defined only on C . Our proofs utilize the
assertions of Lemmas 7 and 8 . Let us add that our proofs thus become qui-
te simple. At the end of this section we consider a practically useful choi-
/
ce of the sequence k suggested by de Figueiredo [17]
nl
gle iteration process converges.
for which a sin-
We now reformulate and prove somewhat generalized versions of Brow-

derfs results in 161 .
W, V, Petryshyn
-T h e o r e m 5. Let X be a uniformly convex Banach space having a

-
- X into X*
weakly continuous duality mapping J of
-- - and l e t A be a con-
-X
t r a c t i v e mapping of a subset C of into
- C.-L e t v 0
be an a r b i t r a r y point
of C. F o r each kn -
-- in (0, 1) A be the s t r i c t l y contractive mapping of
n
C into C defined by A x = k Ax+(l-k )v f o r a l l x in C. L e t u be the
- n n n 0 -----
n - - -
unique fixed point of A in C.
n -
Them f r o m each sequence n(j)+co we can
e x t r a c t a strongly convergent subsequence converging to a fixed point of A
in C.
-
-
Proof. It suffices t o a s s u m e f o r a given sequence k . = k . - t l
J ~(JI
as j-cco , that v . = u AVand to prove that v.-+v. Since C is weakly
J n(j) J
closed, v l i e s in C and s i n c e v . i s the fixed point of A . =
J n(l)
.-
= k At(1-k )v = k.A + (l-k.)v in C we have
n(j) n(j) 0 J J 0
a s j+co since
k,+l and Av. is bounded. Hence, by L e m m a 8, v i s
J J j
a fixed point of A in C , i. e. , vgF(A) , Now f o r each j we have
and, since v is a fixed point of A in C,
(3.2) ( 1 - kJ . ) ~t k j (v-AV) = (l-kj)v.
Subtracting (3.2) f r o m (3.1) and taking its inner product with J(v.-v) we get
J
Since v., v E C and A i s contractive on C, i t follows from this and the

J
definition of J that
W. V. Petryshyn
Thus, it follows from (3.3) and the definition of J that
Cancelling the positive factor (1-k ), we obtain

1
Since v . 2 ~and J i s weakly continuous, J(vj-v) L O . Hence it follows

J
f r o m (3.4) that v.-v --to, i,e., v+v
J 1 1 J
.
Theorem 6'. If in addition to conditions of Theorem 5 we assume that
there exists a point u in F(A) such that
0 -
(3.5) 0 0 ,J(u0-v)) -
( V -U > 0 for all v in F(A) ,
-
then u -+
n
u as k
0 - n
-k 1 (n ca).
Proof. be a point of F(A) such that (3.5) holds for all v

Let u
0
in F(A) . Then, since u is a fixed point of A in C, by repeating a part
0
of the proof of Theorem 5 we get
Since v.-u -v
J 0
u
0'
- and J i s weakly continous , ~ ( v ~ - u ~ ) - J ( v - u
0) '
Hence, by (3.5), (vO-uO,J(v.-u ))--c (vO-uO,J(v-u )) < 0, Consequently, (3.6)
J 0 0 -
implies that v.+u
J 0
a s j +ca . }
Since { v j was any weakly convergent s e -
quence gotten from [ u n i , it follows that u +u
n 0'
Note. Observe that we did not require for X to be strictly convex.
-
-A be a contractive mapping of a subset C of-a
Corollary 1. Let
Hilbert space H - into C. F o r any fixed point v in C and any
0- -
let A be the mapping of C into
kn6(0, 1) -
n - C defined by A
n
(x)=k
n d
Axt(1-k vo
for
--
u +u
n
all
0 -
x in
as n
C.
-
Let
-n u
a,
be the unique fixed point of A
n -
in
, where u 0 is the fixed point of A -
C .-Then
in C nearest to
W. V. Petryshyn
Proof. To obtain
- Corollary 1 from Theorem 6, first note that in case
X i s a Hilbert space H we can identify its dual with H by the inner pro-
duct and the simplest weakly continuous duality mapping i s the identity map-
ping I . Thus, all that remains to prove i s that to any point v
0
in C
there exists a (unique) point u in F(A) nearest to v such that
0 0
0 0 0
- ~ -
.
(v -u , u -v) > 0 for all v in F(A) This fact was proved by Browder [6]
-
(see also [9,27] ). But, for the sake of completeness we reproduce this proof
here.
It i s well known that, since F(A) i s a closed bounded convex set in H,
to each point v in C o r in H there exists a unique point u in F(A)
0 0
nearest to v If, for any other point v in F(A) , ' we put u = (1-t)u +tv
0' t 0
for 0- - then by definition of u 0 we have the relation
< t<l,
The above inequality implies that for any v in F(A)
On the other hand, suppose there exists another point u in F(A) such that
1
for all v in F(A)
(3.8) (vo-U1' ul-v)-> 0.
Setting v = u in (3.7) and v = u in (3.8) and adding we get

1 0
i , e. , uO=U , i s uniquely characterized by (3.7).

and, therefore, u
0
Special choice of the sequence
-(knJ
Let kn be a sequence of real numbers such thah k -+ 1 a s
n
n + a, and O < k (1. The partial disadvantage of Theorem 5 and Corolla-
n
r y 1, from the computational point of view, i s that to obtain a fixed point
u of A in C we need to go through two limiting processes, one to obtain
0
the fixed point u of A = k A in C (we assume heer, a s in 671, that OtC
n n n
and take v =0) for each given k the second to construct u a s the limit of
0 n' 0
u a s n 4 m (i.e.,k -+I).
n n
In [17] de Figueiredo, using Browderts Theorem 1 in [6] ( a s s e r -
ting the strong convergence of u to u in Hilbert space H) shows that by
n 0
choosing k = n/n+l and, for each y in C , determining the sequence
IYn\.
. n
by the process
2
0
n
(3.9) Y, ' An (n=1,2, ...,
An = n / n t l A)
we get the strong convergence of to u Using our Theorem 5 we now

y
n 0'
prove the convergence of the method (3.9) for mappings A acting in certain
Banach spaces.
Theorem 7. Let X be a Banrch space having a weakly continuous duality
-
m a p p i x J of
- X into
- X* -
and .- let A be a c o n t ~ a c t i v emapping of
C(CX) -
into C. Assume that 0 C and that there exists u (F(A) such that
0
-
(uo, J(uo-v))< 0 for all v in F(A) A =k A(k = n / n t l ; n = l . 2 , 3 , . .),
n n n
.
then for each y in C the sequence determined by the process (3.9)
0-
converges to a fixed point of A in C . -
Proof. Let
n
x .
be the fixed point of A in C Then since A i s
n n
strictly contractive for any initial approximation x in C we have the
0
e r r o r estimate
( 3 k nk
1 o 011-
(3.10)
7A X - x
H ~ ~ x ~ -ix ~ l l < m
1 -
ntl ( ~ t n ) ~ - ~ ~
where m i s the diameter of C. Since, by Theorem 5, x --t u and
n 0
u EF(A) it i s easy to s e e that y n3 u0' Indeed, it follows from (3.9) and
0
(3.10) that
W. V. Petryshyn
ria -1
Since, a s i s not hard to verify, n n a / ( l t n ) -+O as n.-+ (a for any a >1
and x--+ u it follows from the above inequality that y -+ u
n 0 n 0'
-
Note. In case X is a Hilbert space, in virtue of Corollary 1, Theorem
7 reduces to the result contained in 1171 .
W. V. Petryshyn
References
----------
1. L . P . Belluce and W.A. Kirk, Fixed point t h e o r e m s for families of contrac-
tion mappings, Pac. J. Math. 18(1966), 213-217.
2. M. S. Brodsky and D.P. Milma, On the c e n t e r of a convex s-
e t , Dokl. Akad.
Nauk SSSR 59 (1948), 837-840.
-of
3a. F. E. Browder, Existence of periodic solutions f o r nonlinear equations
evolution , P r o c . Nat. Acad. Sci, USA , 5 3 (1965), 1100-1103.
3b. --
, Fixed point t h e o r e m s f o r noncompact mappings in Hilbert
space,
- P r o c . Nat. Acad. Sci., USA, 53 (1965), 337-342.
4. -- Proc.
, Nonexpansive nonlinear o p e r a t o r s in Banach space,
Nat. Acad. Sci., USA, 54 (1966), 1041-1044.
5. ---m a p -
, Fixed point t h e o r e m s f o r nonlinear semicontractive
pings in Banach spaces, Arch. Rat. hlech. And Anal. , 21 (1966), 259-269.
6. -
, Convergence of approximants to fixed points of nonexpan-
sive nonlinear mappings in Banach spaces, Arch. Rat. Mech. and Anal. (1967),
82-90.
7 . F. E. Browder and D. G. B e Figueiredo, J-monotone nonlinear o p e r a t o r s in
Banach s p a c e s , Konkl. Nederl. Akad. Wetcnsch, 69 (1906) , 412-420.
8. F, E. Browder wna W. V. Petryshyn, The solution by iterattion of nonlinear

functional equations in Banach spaces, Bull.Amer. Math.Soc., 72 (1966), 571-575.
9. , Construction of fixed points of non-

l i n e a r mappings in Hilbert space, (to a p p e a r i n J. Math. Anal. and Appl.).
10. C.R. De P r i m a , Nonexpansive mappings in convex l i n e a r topological spa-

c-e s (to appear).
11. M. Edelstein, An extension of Banachfs contraction principle, P r o c . A m e r .
Math., 12 (1961), 7-10.
12. , On predominantly contractiv'e mappings, J. London Math. Soc.,
37(1962), 74-79.
13. , On nonexpansive mappings in Banach s p a c e s , Proc. Cambrid-
ge Phil. Soc. , 60 (1964), 439-447.
14. , A r e m a r k on a t h e o r e m of M.A. Krasnoselsky, Amer. Math.
Monthly, 13 (1966), 509-510.
15. D. G. De Figueiredo, Fixed point t h e o r e m s ,for weakly
- continuous
-- map-
pings, Math.Res. Center, Univ. of Wisconsin, Technical Rep. No. 638 (1966).
16. ---- , Fixed point theorems---for nonlinear operators and
-
Galerkin approximations, J. of Differenzial Equations, 3(1967), 271-281.
-
17. - , Topics in nonlinear functional analysis, Inst. for
fluid Dynamics and Appl. Math. , Univ. of. Maryland, Lecture Notes Nn 48
18. --- , and L.A. Karlovitz, On the radial projection in nor-

med
----- spaces, Bull. Amer. Math. Soc. 73 (1967), 364-368.
19. D. Gohtie Zum Prinsip d -

-- e--
r kontraktiven Abbildun~, Math. Nachr. 30
(1965), 251-258.
20. ----- --
,Uber Flxpunkte bei steligen Selbstabbildungen rnit kompakten
Itericrten. Math. Nachr. 28 119641. 45-55.
21. R. 1. Kachurovsky, On some fixed point principles, Usen Zap. Moskov
Reg. Ped. lnst. 96(1960),-215-219.
22. S. Kaniel, Quasi-compact
---- nonlinear operators in Banach space and appli--
cations,
-- Arch. Rat. Mech. and Anal., 21(1966), 259-278.
23. W.A. Kirk, A----fixed

-- point theorem for mappings which do not increase
distance,
-- Amer, Math. Monthly, 72(1965), 100471006.
24. A. N. Lolmo~oroffand S. V. F d r m i n , Elements
-- of the theory of functiors-
and functional
------ analysis,
- Graylock P r e s s , 1957.
25. M.A. Krasnoselsky, Two ~ b s e r v a t i o n sabout the method-
of successive ap-
proximations, Usp. Math. Fauk 10(1955), 123-127.
26. M. Lees and M. Ii. Schultz, A Leray-Schauder principle for A-Compact
mappings and the numerical
- solution of----nonlinear two-point boundary value problenis,
-------
John Wiley and Sons, Ioc. , New York 1966, 167-179.
27. J. L. Lions and G. Stampacchia, Variational inequalities, (to appear).
28. R.De Marr, Common fixed points for commuting - -- mappings,
contraction
Pat. J . Math. 13(1963), 1139-1141.
29. S, Mazur, Uber
------die
- kleinste konvex Menge, die eine-----
gegebene kompakte,
Menge enthalt,
---- Studia Math. 2(1930), 7-9.
30. --
Z . Opial, Weak convergence of the sequence of successive approxin:ations
-
for nonexpansivekappings (to appear).
31. W. V. Petryshyn, On a fixed point theorem for nonlinear P-compact ope-
--- -
r--
a t o r s in Banach space,
- Bull. Amer. Math. Soc. 72(1966), 329-334.
32. ------ - ---
, On nonl.inear P-Compact operators in Banach space
with
--- applications to constructive fixed point theorems, J. Math. Aaal. Appl.
--
15(1966), 228-242.
33. --- , Further r e m a r k s on nonlinear
------ P-compact operatop-
in Banach space, P r o c . Nat. Acad. Sci. USA, 55(1966), 684-687 (extended v e r -
sion in J. Math. Anal. Appl. 16(1966), 243-253).
34. , On the extension and the solution of nonlinear o p e r a t o r equa-
tions, Illinois J. Math, , 10(1966), 255-274.
-
35. , Construction of fixed points of demicompact mappings in Hil-
b e r t space, J. Math. Anal. Appl. 14(1966), 276-284.
36. , R e m a r k s on fixed point t h e o r e m s and t h e i r extensions, T r a n s .
Amer. Math. Soc. 126(1967), 43-43.
37. , and T. S. Tucker, On functional equations involving P -compact
o p e r a t o r s (to appear).
38. J . Schauder, D e r Fixpunktsatz in Funktionalraumen, Studia Math. 2(1930),
171-180.
39. H. Schaefer, Uber die Methode s u k z e s s i v e r Approximationen, J a h r e s b e r .
Deutsch. Math. Verein. 59(19579, 131-140.
40. M. Shinbrot, A fixed point theorem and s o m e applications, Anch. Rat.
Mech. Anal. 17(1964), 255-271.
41. E, Zarantonello, The c l o s u r e of the n u m e r i c a l range contains the spectrum,
Bull. A m e r . Math. Soc. 1 0 ( 1 9 6 4 ) , 781-787.
(C. I. M. E. )
W. V. PETXYSHYN
ITERATIVE CONSTRUCTION O F FIXED POINTS OF CONTRACTIVE

TYPE MAPPINGS IN BANACH SPACES
C o r s o tenuto ad I s p r a dal 3-11 Luglio 1967

ON THE APPROXIMATION-SOLVABILITY O F NONLINEAR FUNCTIONAL
EQUATIONS IN NORMED LINEAR SPACES
by
W, V. Petryshyn
1. 1. Introduction. The purpose of this part of my talk i s to outline some

of the recent results concerning the projection-and the approximation-solvabili-
ty of nonlinear functional equations in normed linear spaces. We first state
a general theorem on projectional-solvability proved in [15] which at the s a -
me time unites the e a r l i e r results on linear equations obtained in Fl, 18, 10,
6, 131 (*) with the recent results on nonlinear monotone o p e r a e r s obtained
in [7,12,2,3, 19, 141 .
Let X be a separable Banach space, X a finite dimensional subspa-
n
ce of X and P a linear projection of
n
X onto X Let --+n and
n
.
na "denote respectively the strong and the weak convergence in X. The
-
sequence {xn) i s called projectionally complete in X if and only i f
P_x+x for each x in X. Let A be a (possibly) nonlinear mapping of
I,
X into X and { A ~ ]a sequence of mappings defined by

In [ld the author considered the problem of giving a constructive proof
of the existence and the uniqueness of solutions of the equation
a s strong limits of solutions x in X of approximate equations

n n
a complete and self-contained theory of the projection method s e e 1151 ;

for a more complete l i s t of references and contributions to the theory of pro-
jection and Galerkin methods s e e rl 1, 18, 15, 51
-1
.
(+)1n [IS] the author also conside:ed the problem of constructing solutions
of Eq. (1) a s wkak limits of solutions x._ of Eq. (2) in case the operator A
I1
is demicontinuous o r weakly continuous and monotone o r K-monotone.
W. V. Petryshyn
Definition 1. Eq. (1) i s s a i d t o be strongly projectionally-solvable if

and only if t h e r e e x i s t s an integer N > 0 s u c h that f o r each n >- N and
each f in X Eq. (2) h a s a unique solution x in X such that x +x
n n n
and x is t h e unique solution of E q . ( l ) .
One of the r e s u l t s obtained in 15 w a s t h e following t h e o r e m .
T h e o r e m I. Suppose t h e r e e x i s t s N > 0 and a continuously s t r i c t l y incr?
t
sing function c( ( r ) of R - 0 )into R -
with d (0) = 0 -
and
l i m d ( r ) = co s u c h that A i s continuous in X for n
r+ m n n- -> N , A x ---c Ax
n
f o r each x -
in X -
and
(3) IIA X-A y .II n -

> q ( x-y
I ) f o r each n
Then Eq. (1) is strongly projectionally-solvable provided A s a t i s f i e s

-> N- and
- -a l -l x and y in X
n'
the following condition (c): -

If {x,] i s any subsequence
-- of
is a n a r b i t r a r y sequence in X- with x Xm such that x 2 x -
and
m m
A x + g, - then Ax = g.
m m
We omit h e r e t h e proof of T h e o r e m 1 s i n c e below we s h a l l prove a
much m o r e general t h e o r e m . Theorem I w a s then used in(15] to e s t a b l i s h
t h e s t r o n g projectional-solvability of various c l a s s e s of l i n e a r and nonlinear
equations. T o illustrate the generality of T h e o r e m 1 we now deduce f r o m it the
following two important c o r o l l a r i e s .
Corollary 1. - If X i s a Hilbert space H - and A = T t S, where T and - -
S a r e l i n e a r bounded o p e r a t o r s s u c h that S is completely continuous and
T is definite. i. e.. T is such that
then Eq. (1) i s strongly projectionally-solvable provided that N(A) = 0 ,

where N(A) i s the null s p a c e of A.
Proof.
- Since A s a t i s f i e s condition (c), A
i s continuous in X
P n
and A x - t A x , it suffices to show that 'A s a t i s f i e s (3), i. e., that t h e r e e x i s t s
n
N>O and (in fact, @ (P) -< o r ) such that
w.v. W. V. Petryshyn
(3') x d (x ) ( cx ) for all x in X and n -

> N.
n
0
Assume, to the contrary, that (3 ) does not hold. Then we can find a s e -
quence { x m ] c H with l l ~ ~= !1 such that x -x
m
and A x -0
m m
as
m -+a . First, the complete continuity of S implies that S x -+ Sx
m m
and
But, T x Tx. Hence, Tx = -Sx o r Ax = 0. Now, since T i s defi-

m m
nite, we have
+ - (SX,x) - (Tx, X ) - (Tx, X) + (Tx, x) = 0
s i ce Ax = 0. Thus, x = 1 and Ax = 0 in contradiction to N(A) = (0) .

Remark. Corollary 1 was proved in [6] .An independent proof for
T satisfying slightly weaker conditions was given in [lo]. gee also [lq .
F o r other applications of Theorem I to linear bounded and unbounded opera-
t o r s s e e 1151 .
Corollary 2. -
If X = H and A is continuous nonlinear complex mono-
- - -
tone operator, i. e.,
-then Eq. (1) i s strongly projectionally-solvable.
Proof. The continuity of A implies that A i s continuous in Xand

n n
that A x-+ Ax. Since P x=x for all x in X the Buniakovski-Schwartz
n n n'
inequality, when applied to (5) shows that (3) is satisfied with d ( r ) = cr.
Hence by Theorem I, it suffices to show that A satisfies condition (c).
Suppose { xm] C H
with x 6 X
m m and I = 1 such that x 2 x and
m
Amx,+ g. Then, by (5) ,
W. V. Petryshyn
- (g, x) + (Ax, x) = 0.
Hence, x
m
---c x and A x + Ax = i. e., A satisfies condition (c)
m m
.
Remark. Corollary 2 i s certainly true when A is strongly monotone,
i. e. , Re(Ax-Ay, x-y) -> c 1 x - ~112 . When A=I-F , where F i s monotone,
the result was first proved by Minty [lq (see also Kachurowski [7]) . F o r
general continuous A, it was proved by Browder [2]. For complex monoto-
ne, operator A with the additional assumption that A i s bounded, Corol-
lary 2 was proved by Zarantonello [19] and by Browder [3] with the boun-
dedness assumption dropped. All the above proofs a r e nonconstructive and
rather complicated.
In his forthcoming paper [4] Browder notes that for certain classes of
nonlinear opertaros (e. g. J-monotone operators) a direct verification of con-
dition (c) seems sometimes to require additional hypothesis, which thus
might in general restrict the class of equations to which Theorem I is applica-
ble. He then shows that if instead of condition (c) we assume that Eq. (1) pos-
sesses a solution, then somewhat weaker assumptions ensure that Eq. (1) i s
strongly projectionally-solvable. In his subsequent paper [57 Browder
exploits further the solvability conditions (s) by deriving general results on
the relation between solvability condition and the approximation-solvability
(to be defined below) of Eq. (1). In [16] the speaker showed that at least
when X is a reflexive Banach space then, under the assumption of the other
hypotheses of Theorem I, the condition (c) and the solvability condition (5)
a r e equivalent. In his subsequent paper 1173 the speaker showed the equiva-
lence between the solvability condition (S) and the modifed condition (c)
(the so-called, condition (H)) without the assumption that the underlying
W,V, Petryshyn
s p a c e s a r e Banach spaces. Some of the r e s u l t s mentioned above a r e given in

the next section.
2. Approximation-solvability.
-.-- Let X and Y b e two normed l i n e a r
s p a c e s and l e t [XAC X n
and {Y ICY
be two sequences of finite dimensional
subspaces. L e t P be a mapping of X into X and Q a mapping of Y into
n n n
.
Yn L e t A be a sequence of nonlinear mappings of X into Y defined by
n n n
A = QnA , where A is a nonlinear map of X into Y .
we give a constructive proof of t h e existence and uniqueness
of solutions of the equation
a s s t r o n g l i m i t s of solutions x CX of equations
n n
To make the nature of o u r problem m o r e p r e c i s e we need the following defini-

tions ( s e e a l s o F 6 , 1 7 3 and Browder [4,5] ) . .
Definition 2. A quadruple of sequences ( X
--- {
nJ / ~ n ~j p n '(Qnf
f I
called a n approximation s c h e m e f o r Eq. (6) if and only if dim X =dim Y
n n
for each n , P and Q a r e continuous, and P x j x f o r each x in X and
n - n n
Q n y + y --
for y -
in Y.
Definition 3. Eq. (6) i s s a i d t o be uniquely approximation-solvable with
r e s p e c t to the given s c h e m e = (x , { PQ ) if t h e r e e x i s t s an
-integer N >0 such that f o r each n>N and each f in Y , Eq. (7) h a s a unique
solution x in X such that in X and x is the unique solution of Eq.
x -+ x -
n - n -- n
--. We emphasize that t h e mappings Pn and Qn a r e not a s s u m e d to

Note.
be projections o r even l i n e a r though f o r most of t h e direct methods (7) f o r the
approximate solution of Eq. (6) they a r e l i n e a r mappings induced by any one
of the the following groups of methods: the projection o r Galerkin type m e -
thods, and the finite difference methodsp)

- (assuming, a s in Lax-Richtmyer
[ 83, that exact and approximate solutions belong to the s a m e space).
The main result of this section i s the following theorem.
Theorem 1. Let X and Y be two normed linear spaces and let rn - be an
approximation scheme for Eq. (6). Suppose.there exists an integer N>O --
and a
i s continuous in X for each
n
>N
n- , Q A P x -t Ax for each x -
in X -
and
n n
(8) n>N -
and all
-x and
- y in Xn .
Then Eq. (6) is uniquely approximation-solvable if and only if A sati-
-
sfies the following condition (H): If- rm i s any subscheme of the approxima-
tion scheme rn for Eq. (6) and is any bounded sequence in X -
with
such that for some g , then there exists a
oftm] and ana element-x in -X such that x and A x
-+ x -
m.
1
m. m ,
1 1
+
-
Proof. (Uniqueness. ) Suppose that uf v and Au=Av. Because, for
each u and v in X, P u and P v lie in X (8) implies that for each
n n n'
Since P U-HI, P V-+V and Q AP x -+Ax for each x in X, the passage

n n n n
-
to the limit in the above inequality implies that 0 >o( (llu-vlb. The latter
inequality contradicts the properties of& ( r ) and shows that A i s one-to-
one.
' '1f we do not assume that X, and Yn are subspaces of X and Y respectively
(as is usually the case when Eq. (7) i s a finite difference analog of Eq. ( 6 ) ) ,
then all results outlined in this section remain valid without any substantial
chahge in the proof provided that, a s in Aubin [I] (see also the papers discus-
sed during this meeting by Lions, Reviart) we introduce the mappings
1' :>,+X , s :Y+Y and pn:X-X

n n
subject to certain conditions (for details s e e
and b e paper of Raviart) guaranteeing that the operators A uefiiled by
A = s Ap have the needed properties.
n
n n n
W. V, Petryshyn
(Existence.) By (8) and our assumptions, A i s a one-to-one continuous

n
mapping of X into Y for each n>N,
- and he,9ce, by the Brouwer Theorem on
n n
Invariance of Domain, the range R(A ) i s an open set in Y . furthermore, it
n n'
follows from (8) and the continuity that R(A ) is also a closed s e t in Y for
n n
each n>N .
- Indeed, if { Y ~ ) C H ( Aand ~) y
m
in Y then there exists a
n'
sequence {x,} C X such that y = A x and, by (8) flyl 2
> (I~x
-+ .
-tkJI).
n m n m
1 f p ( r ) denotes the continuous strictly increasing function of
R lnto R with! (0) = 0 which i s the inverse of d (r), then since y m is
a convergent sequence
(lxt - Xk 1 5 p ( IIy - ~ ~ l-OJl ) (1,k - ~ )
Thus there exists an element x in X such that x + x and A x +A x=

n m n m n
y R(A ) by the continuity of A for n> N, i.e., R(An) is closed. Since
-
n n
R(A ) i s a nonempty s e t in Y , which is both open and closed in Y n , it
n n
follows that R(A ) = Y
n n
.
Hence for each n - > N and for each given f in Y there exists a unique
element x
n
in X such that A x = Q (I). F o r the sequence
n n n n
xn thus {
determined, (8) and the fact that Qn A ( O ) ( ~KO I(
< for some constant K imply
0
that
whence, in view of ( 7 ) and the fact that 11 ~ ~ , ( f )-( ( K< for some constanl K
(11 xn 11 ) 5 KtK
0
which implies lllat I( xll 1 z p ( K O t ~,)
i. e. , { xl, is a boundpd seque~lce with x , ~ X I , and such that
n n
A x =
= Qn(f)+ f, a s n+m. Ilence, conditio~)(11) implies that there exists a subse-

quence {xn f
of [xi1 and a.) element x in X sucll that x x
n.
and
I I
A
n.
x
n.
+ Ax, a s n.+
I
ce . Because A x = Q (f)+C
n. n. n.
as n-
I
a, we
I I 1
. Sincc, a s
I 1
:>htai~ithat Ax = f, i . ~ . , x is a solution of Eq. (6) was shown
W. V. Petryshyn
above, x i s the unique solution of Eq. (6) for a given f in Y we conclude

a posteriori that the selection of the subsequence was not necessary. Conse-
quently, the entlre sequence ( a n / converges strongly to x, i. e., Eq. (6) i s
kmj
uniquely approximation-solvable.
Converse . Let I'm= ({xmI , Y(), ,(P,], ) be an a r b i t r a r y
sut;schenle for Eq, (6) and let x be any bounded sequence in X with
m
x in X such that
m m
for some g in Y. To prove that A satisfies condition (H) note f i r s t that,

by our conditions on Q
m'
Since, by hypothesis, Eq. (6) i s uniquely approximation solvable, for every

m-
> N and every g in Y, there exists a unique element y in X such
m m
that
(11) Amyrn = Qm(g)
Ym-t Y in X, a s m+w, and y i s the unique solution of the equation

Ay = g. Substracting (9) from (11) we obtain the equation
whence, in virtue of (8), for all > N we derive the inequality

m-
This, (10) and the properties of

P ( r ) imply that
W. V. Petryshyn
Since y
m
-
y, (12) implies that x
m
-
y, a s m d m , and Ay =g. Consequen-
tly, (9) implies that A x + Ay and thus shows that A satisfies condition
m rn
(HI .
The following theorem gives the relation between the solvability condi-
tion (S) and condition (H) .
Theorem 2. Under the hypothesis of Theorem 1 the following equivalent as-
sertions a r e valid:
(A1) A satisfies condition (H)
(A 2) Eq. (6) i s uniquely approximation-solvable.
(A3) F o r any given f in - X.

Y, Eq. (6) has a solution x in
Remark. Looking over the proofs of Theorem 1 and 2 we note that nowhe-
r e did we use the continuity property of the mappings P and Q assumed in
n n
Definition 2. Consequently, Theorems 1 and 2 remain valid without the assum-
ption that P and Q a r e continuous. This fact may on occasion prove to
n n
be important and useful particularly in the application of these theorems to
the collocation and finite difference methods.
Let us r e m a r k in passing that combining the results of Theorem 2 with
those of Theorem 2 in [16] we have the following useful theorem .
Theorem 3. If additionally we assume that X and
- Y a r e separable
reflexive Banach spaces, then under the hypotheses of Theorem 1 the follo-
wing equivalent assertions a r e valid;
(A1) A satisfies condition (H).
(A2) Eq. (6) is uniquely approximation-solvable.

-Y, Eq. (6) has a solution
(A3) F o r any given f in x in X.
(A4) A satisfies condition (C) : rmis any subscheme of rn -
and
is an arbitrary sequence in with x in such that
1- m - Xm
and A x + g, then Ax g.
- - .=
m m
Remark. Theorem I follows a s a corollary (sufficiency part) of Theo-
rem 3 (A4) .
W. V. Petryshyn
An immediate and practively useful c o r r o l l a r y of T h e o r e m 2 is the fol-

lowing theorem.
T h e o r e m 4. L
- and Y -b e two n o r m e d linear s p a c e s and l e t
e t X --
be a n approximation s c h e m e f o r Eq. (6): Suppose that A is a continuous
mapping of X -
into Y and that
Q (y) converges to y strongly in Y -
n
and
uniformly
--- for y in a compact subset of Y. Suppose f u r t h e r that t h e r e e x i s t s
--
an integer N>O and a function ( r ) a s in T h e o r e m 1 s u c h that the inequality
(8) holds. Then-- , under the above hypothesis,
--- a s s e r t i o n s (A ), (A ) and (A3)
1 2 -
-
of T h e o r e m 2 a--
r e equivalent.
Remark. Under the hypotheses of T h e o r e m 4 with d ( r ) satisfying sli-
ghtly weaker conditions, the equivalence of a s s e r t i o n s (A ) and (A ) was f i r s t
2 3
obtained by Browder in [5] .
3. P e r t u r b e d problems. O u r next t h e o r e m s deal with the problem of
constructive solutions of p e r t u r b e d equations
a s s t r o n g l i m i t s of solutions xn€ Xn of approximate equations
where B i s a completely continuous mapping of X into Y(i. e., B i s conti-

nuous and m a p s bounded s e t s in X into relatively compact s e t s in Y) and
A s a t i s f i e s c e r t a i n conditions t o be specified below.
T-
h-
e o-
r e m 5. Suppose that and Y
X- a r e normed l i n e a r s p a c-
e s , Qn(y)
-converges
--- y strongly
tu----- in Y -and ---- uniformly f o r y in a compact s u b s e t of
Y , B i--------
s completely c- o n t i n u o--~and
~ ~ A s-a t i s-
fies-
condition (11). -
Suppose
-further
------- that f o r a given f -in Y -
and-each n N the solutions-x n - of Eq.
-
1%
- exist and a r e uniformly bounded by a-constant
--- independent of n.
If f o r any given
------ in
- Y Eq.-(3
f - x in- X,
h a s a t most one solution
--
then Eq. (8) i s approximation-solvable
- - (i.e., Eq. (14) h a s a solution x in
n -
W. V. Petryshyn
X for each n > N such that x + x, a-

s n*, and x is
-- -the unique so-
n- - n
lution of Eq. (141.
Let us remark that though the second hypothesis of Theorem 5 i s rather
restrictive, nevertheless it is not artificial and often can be verified in appli-
cations. Condition (a) which was introduced by Browder
L51is essentially
of the same strength. Furthermore, if X is a reflexive Banach space,
%
Y = X* and P and Q = P* a r e linear projections in X and X , re-
n n n
spectively, then the second hypothcsis of Theorem 5 i s implied by the coer-
civeness condition
which has .been successfully used by a number of authors (see, for exam-
ple, Browder [5, 3 3 and Lions [93 j .
Remark.
---- For the proofs of Theorem 2 to 5 a s well a s for other r e -
sults and the relation of condition (H) to the concept of P-compactness see
References
----.-
1. J. P. Aubin, Nonlinear stability-appro~imation~of

nonlinear operational equa-
tions (to appear).,
-
'2. F.E. Browder, On the solvability of nonlinear functional equations, Duke
Math. J. 69 (1963), 557-566.
3. , R e m a r k s on nonlinear functional equations 11, Illinois J.

Math. 7 (1965), 608-618.
4. , Nonlinear a c c r e t i v e o p e r a t o r s in Banach s p a c e (to a p p e a r ) .
5. , Approximation-solvability of nonlinear functional equa-
tions in n o r m e d l i n e a r s p a c e s (to appear in Archive Rat. Mech. Anal.).
6. S. Hiltlebrandt and E, Wienholtz, Constructive proofs of representation theo-
r e m s in Hilbert space, Comm. P u r e Appl. Math. 17 (1964) , 369-373.
-
3. R.I. Kachurovski, Monotone nonlinear o p e r a t o r s in Banach spaces, Dokl.
Akad. Nauk. SSR, (1965), 679-698.
8. P. D. L a x and R.D. Richtmyer, Survey of stability of l i n e a r finite difference
equations, Comm. P u r e Appl. Math. 9(1956), 267-293.
9. J. -L. Lions, Sur c e r t a i n e s equations paraboliques non l i n e a i r e s , Bull. Soc.
Math. F r a n c e , 93(1965), 155-175.
10. V. E. Medvedev. On the convergence of the Bubnov-Galerkin method, P r i k l .
Mat. Mekh. 17(1963), 1148-1151.
11. S. G. Mikhlin, Variationsmethoden d e r Mathematischen Physik, Akademie-
Verlag, Berlin, 1962.
12, G. J. Minty, Monotone (nonlinear) o p e r a t o r s in Hilbert space, Duke Math.

J., 29 (1962), 344-346.
13. W. V. Petryshyn, On a c l a s s of K-p. d. and non-K-p. d. o p e r a t o r s and ope-
r a t o r equations, J. Math. Anal. Appl. 10(1965), 1-24.
14. , On the extension and the solution of nonlinear o p e r a t o r
equations, Illinois J. Math. 10 (1966), 255-274.
15. , Projection methods in nonlinear n u m e r i c a l functional
analysis, (to a p p e a r in J. Math. Mech.)
16. , R e m a r k s on t h e approximation-solvability of nonlinear
functional equations (to a p p e a r in Archive Rat. Mech. Anal. )
17. , On t h e approximation-solvability of nonlinear equations
(to a p p e a r in Math. Ann. )
18. N. I. Polsky, Projection methods
----- in applied mathematics, Dokl. Akad.
Nauk, SSSR 143 (1962), 787-790.
19. E. H. Zarantonello, The closure of the numerical range contains the spec-
trum, Bull. Amer. Math. Soc. 7 0 (1964), 781-787.
--
(C. I. M. E. )
P.A. RAVIART
NAPPROXIMATION DES EQUATIONS DIEVOLUTION PAR DES METHODES

VARIATIONNELLES"
C o r s o t e n u t o a d I s p r a d a l 3 - 1 1 L u g l i o 1967
359 - -
APPROXIMATION DES EQUATIONS DtEVOLUTION PAR DES METHODES
VARIATIONNELLES
Par
P..A. Raviart
(Universite de Rennes)
Introduction .
Le present cours e s t une introduction A ltapproximation p a r l e s dif-
ferences finies des solutions des equations aux derivees partielles dtCvo-
lution. On y traite des equations lineaires du l e r ordre et du 26me o r -
d r e en t A ltaide des methodes variationnnelles : cette etude reprend avec
quelques ameliorations techniques un certain nombre de resultats de 1101.
Dans chaque cas, aprks un bref rappel theorique, on examine deux sche-
m a s classiques, l'un implicite et ltautre explicite, et on donne des theo-
rgmes de stabilite et de convergence. En appliquant c e s resultats a un c e r -
tain nombre dtexemples, on obtient en particulier des resultats de stabili-
te pour des equations paraboliques et hyperboliques 3 coefficients m e s u r a ~
bles et bornes dans des domaines cylindriques.
L e s methodes utilisees peuvent s e generaliser A ltCtude de ltapproxi-
mation dtbquations dtevolution couplees 17 1 , dtequations B coefficients
non bornes [8] , ou dlequations p r i s e s dans des domaines non cylindri-
ques [9] . Elles permettent egalement dtetudier dans des cas generaux
l e s methodes de directions alternCes et B pas fractionnaires 16 , 3 .
Enfin , elles peuvent stappliquer avec succks B certaines equations non
1 .---
E q u a t i o n s d t e v o l u t i o n d u l e r ----
- o r d r e en t.
1.1. --
Formulation abstraite.
--
Soient V et H deux espaces de Hilbert s u r C. On suppose que V
est separable, que VC A avec injection continue et que V est dense dans
- 360 - P. A. Raviart
H. On designe p a r 11 11 l a norme dans V, p a r ( , ) et I respecti-

venent le produit scalaire et l a norme dans H . Soit V1 llantidual
fort de V, de norme 11 11* s i on identifie
; H A son antidual Ht , on
a V C H C Vt avec injections continues. Si f & Vt et v 6 V, (f, v) e s t le
produit scalaire qui met Vt et V en antidualite. Si f H, on retrou-
ve l e produit scalaire dans H .
Soit u, v -+a(t;u, v) une famille de formes sesquilineaires continues
sur VXV dependant du parametre t 6 [o, TI, T < a, , avec l e s pro-
prietes suivantes :
V u, v 6 V , l a fonction t -+ a(t;u, v) e s t mesurable et

(lo { a(t;u, v) -
< h i,uII IIv,,, K = constante independante de t ;
il existe deux constantes reellles et d > 0 telles que

( I *2,
1 2
~e a(t;v, v) + /\lv/ ->Qt/,vl
2
, V V E v pep. en t. .
Soit ~ ( t ) ( E [ ( v ,V1) ltoperateur defini pour presque tout t par:
De la separabilite de ltespace V et de lthypothese (1.1) , on deduit

2
que l'operateur A(.) (t--cA(t)) e s t lineaire continu de L (0, T ; V)
2
dans L (0, T;V1) PI
Sous l e s hypothBses precedentes et pour f donne dans ~ ' ( 0 ,T;V'),

u donne dans H , il existe une fonction u et une seule verifiant
0
Si X e t Y sont deux espaces vectoriels topologiques, (X, Y) e s i 4

ltespace des applications lineaires continues de X dans Y .
( 2 ) ~ iX est un espace de Banach de norme 11 I! LP(0. T;X) , 1 < p < o desi-
gne llespace des (classes de) fonctions t --ru(tf fortement mesuFables s u r
(0,T) pour l a mesure de Lebesgue dt B valeurs dans X et telles
T I/P
x 4 0
I
U ( ~ ) I I dt )
X
.
< to Modification usuelle s i p = o .
P. A. Raviart
Pour la demonstration de ce theoreme , voir 1 4 3 , [5] .

Remaraue 1.1.
Dtapr&s[5] , toute fonction u satisfaisant i (1.4) e s t p.p. 6ga-
l e 2 une fonction continue de LO, TJ dans H et l a condition (1.6) a

un sens. Notons egalement que s i u et v sont deux fonctions satisfai-
sant 2 (l.4), on a l a formule de Green .
Dans la suite, on decomposers l a forme a(t;u, v) en
oh a (t;u, v) est une forme sesquilineaire continue s u r V X V verifiant

0
vu, v v , la fonction t -t a (t;u, v)

0
est rnesurable et
< KO 11 ulllvll
ao(t;u, v) - , K 0 = constante independante de t ,
(3) $!designe la derivee de u au sens des distributions sur10, T[ i valeurs

.
dans V1 (cf. [15] ) Autrement dit, Ithypothese g & ~ ~
(0, T;V1) signifie qutil
du 2
eniste une fonetion - = g E L (0, T;V1) telle que
dt
[
(g(t), v ) r ( t ) dt =
P. A. Raviart
tandis que a (t;u, v) e s t une farme sesquilineaire contlnue s u r V %H

1
avec
v u E: V , 'dv € H , la fonctlon t a (t;u, v) e s t mesurable
1
-t
1
(1. et a (t;u, v) ( Kl J J u I/ I VJ , K2 = constante indipendante
de t l,
ll est clair qutune telle decomposition (1.8) est toujours possible et l e s

proprietes ( 11 , (1.2 et (1.8) , ( 1 1 , (1,2)0 , ( 1 1 sont equi-
valentes .
1.2. Exemple I.
Slit R un ouvert de l'espace euclidien R~ de f r o n t i e r e y = 3 R.

L
On designe p a r L (R) l'espace des (classes d e ) fonctions definies (p.p)
sur R 5 valeurs complexes et de c a r r e sommable s u r R pour la mesure
de Lebesgue dx. On munit LL(R) de l a norme hilbertienne
IIu 1 = ( j' (x), dxl1I2 .

~~(f-4
1
On designe p a r H (R) l'espace de Sobolev des (classes de) fonctions
2 311
u € L (R) dont l e s derivees premieres i ) ,~i = 1,. , n , sont dans ..
2
L (R) , c e s derivdes &ant prises au sens dek distributions s u r R (cf. [14] ).
1
On munit H (R) de l a norme hilbertienne
Si la irontiere r e s t .assel r6guli6ren , on peut definir l a trace

r o "
de u sur rlorsque u & ~ ' ( 0 ): cgest un element de L 2 ( r ) , espace
des (classes de) fonctions complexes de c a r r e sommable s u r pour
la mesure superficielle d6 . En outre l'application ---lo est
1
tinue de H (R) dans L2(r) .
P. A. Raviart
1 1
On note H (0) 11adh6rence dans H (R) de c W ( R ) , sous espace
0 0
des fonctions indefiniment differentiables B support compact dans R Si .
1
l a frontigre r
e s t nassez r6guli$ren , H (0) e s t exactement lfespace des
1
0
. 1 -
fonctions u 6 H (R) telles que F o u = 0 On definit H (R) comme &ant
ltantidual fort de ,HI (R)
; il e s t aisC de voir (& ltaide du theorgme de
lo -
Hahn Banach) que H (R) s e compose des distributions T s u r R de l a
for me
Ceci etant pose, on prend
H = L~(R)
(1.12) 1 1 1
V = sous espace ferme de H ( 0 ) avec H (R) C V C H (0).
0
La famille de formes sesquilineaires a(t;u, v) est donnee p a r
0s les
a3
a . ., ai, a. € L (QT)
1J
, QT ~ $0, T1 avec
P.P. dans Q
T*
On considgre l a decomposition suivante de la forme a(t;u,v) :
+J;ao(x. t) - a ) u(x) i ( x ) dx .
I1 est clair que a (t;u, v) verifie les hypoth&ses (1.1) et (1.2) avec
0 0 0
oh 6 (x, t) est la norme euclidienne de la matrice (a..(x, t))

i<i,
,tandis
q - j<n
-
que a (t;u, v) verifie llhypoth&se (1. 1)
1 1'
Nous sommes dans l e s conditions dlapplication du theorkme 1.1. Il
suffit maintenant dlinterprCter le probleme resolu par ce theoreme . Exa-
minons deux choix simples pour l'espace V.
2 2
On prend f E L (0, T;H-'(R) ) , u E L ( 0 ) . Soit 96 c:(Q~) une fonc-
0
tion complexe indefiniment differentiable B support compact dans Q
T'
Si u est la solution de (1.4) , (1.5) et (1.6) , on a
Ainsi au sens des distributions s u r QT , l a solution u verifie
La condition
2 1
u L (0, T;H (R) ) signifie que u,
3 u (i = 1,. ..,n) appar-
5;;:
0
2
tient A L (QT) et que
ro u = 0 p.p. en t
1
s i la frontiere r de R est
aassez reguli&re\ En resume la solution u = u(x, t) de (1.4) , (1.5) , (1.6)
verifie
(i) llequation (1.17) au sens des distributions s u r Q
T'
(ii) la condition initiale u(xj 0) = uo(x) ,
(iii) l a condition aux limites u(x, t ) = 0 , x & E pep. en t .
P. A. Raviart
On a ainsi resolu l e problgme de Cauchy avec conditions aux limites de

Dirichlet pour lloperateur
1
b) V = H (R).
2 2
On prend ici f C L (QT) , uoE L (R) . Comme au c a s precedent, la
solution u = u(x, t ) de (1.4) , (1.5) , (1.6) vkrifie l'equation (1.17) au sens
des distributions s u r QT . On en deduit que pour tout
1
V EH ( 0 )
=i f(x.t)T(x) dx pep. e n t ;
et ceci doit e t r e egal A
1
Formellement, on en deduit que pour tout v H (R) et p.p, en t
dloh en utilisant --
toujours formellement la formule de Green
n
(1.19)
3
- (x, t) = 2
1~
a . .(x,t) cos (n, x , ) 3u
- (x, t ) = 0 ,
24 ) i, j = l
1J 1 3 x.J
x E ~ , P . P .en t
oh n desipi~e l a normale extdrieure en x P et cos (n, x.) designe

1
P. A. Raviart
Erne
le i cosinus directeur de cette normale. En resume, dans ce cas, l a
solution u = u(x, t) de (1.4) , (1.5) et (1.6) vCrifie
(i) ltequation (1.17) au sens des distributions s u r
QT
(ii) l a condition initiale u(x, 0) = u (x) ,
0
(iii) l a condition aux limites formelle (1.19)--- .
Ceci resoud l e probleme de Cauchy avec conditions aux limites de Neumann
pour ltopCrateur (1.18) .
Signalons que l e s raisonnements formels effectuCs peuvent &tre justi-
fies avec des hypoth&ses de regularites (cf. [6] ) .
I1
1.3. Exemple -
2
On dCsigne par H(b ;R) l'espace des fonctions u E L (R) telles que
11
2
u = $4 L (0) . On m;!:iit H(b ;R) de l a norme hilbertienne
i = l 3x.
On prend alors par exemple
H
2
= L (R)
I'
, V = H(L\ ; n ) .
oh a = a(x, t ) 6 L*(Q T) avec
(1.22) Re a(x, t ) -> d >0 p.p. dans QT .

Les formes a (t;u, v) et a (t;u, v) dCfinies p a r
0 1
verifient l e s hypotheses (1.1) , (1.2) et (I. 1) avkc en particulier

0 0 1
On interprete l e problGme resolu p a r le theoreme 1.1 de la meme facon

2 2
que precedemment. Soient f E L (Q ) et u EL (R) ; l a solution u = u(x, t)
T 0
(i) (1.25 %+A ( a d u) = f au sens des distributions s u r

Q~
(ii) l a conditioq initiale u(x, 0) = u (x)

0
(iii) l e s conditions aux limites ---
formelles
(1.26) Au(x,t)=o. a u ) ( t ) x6.f.

5;;
Oh
a designe l a derivee pormale B .
On pourrait bien entendu multiplier l e s exemples (cf. f4] ) .
2. F o r m u l a t i-
o-
n-a-b s t r a i t e d e l ' a p-p r o x i m a-----------
tion des equations
du l e r
--- o r d r e e n t.
--a
soit une partie bornee de ItP - 101 dont ltadherence rontient

lrorigine ; on notera li l e point generique dea& . Dans la suite, h
s e r a un parametre destine 5 tendre vers 0.
Sgit V un espace de Hilbert s u r C dependant du parametre
h
h (4) . On note ( , )h l e produit scalaire dans Vh et 1 I h la norme
correspondante. On s e donne une autre norme sur ' 11 11 h '
Cquivalente B l a norme I J ; elle verifie done
C(h) = constante (5)
----------
(4)En pratique V s e r a une espace de dimension finie tel que dim V +CQ
lorsque h +O
h
. 11
(5) Le .as interessant e s t celuj oh C(h) -+W lorsque h -0 .

P. A. Raviart
On designe p a r 1) 1)
* l a n o r m e duale de )I )Ih , i.e.
On a trivialement
(2.3) lfhlh 5 C(h) IJfhl/fh VfhcVh

Soit S un param6tre entier > 0 destine & tendre v e r s t c o e t
soit k = (ko;kl,. ..,kS-l ) R' satisfaisant B
max k = 0
S
On va c o n s i d e r e r deux s c h e m a s type dtapproximation de
et nous degagerons s u r c e s deux s c h e m a s l e s m6thodes g e n e r a l e s dt6tude

de lfapproximation .
Pour chaque couple { h, k ) on s e donne
P. A. Raviart
On considgre l e schema implicite 5 un pas de temps variable. Trouver

(s)
Uh,k=(Uh,k(Vh ; S = o , l , S ..., I
solution de
Comme deuxiBme schema, on considgre l e schema explicite egalement a un

pas de temps variable. Trouver u ; s = 0,1,. . ., S
tion de
On va etudier successivement l a stabilite de c e s deux schemas et l a

convergence des solutions u vers l a solution u de (1.4) , (1.5) ,
h, k
(1.6) dans des s e n s qui seront precises. Bien entendu, on peut conside-
r e r des schemas dtapproximation plus elabores. (cf. [ll] p a r exemple) .
3.1. Quelques lemmes.
Donnons dlabord deux l e m m e s faciles qui seront d'un usage constant

dans la suite.
Lemme 3.1.
Soit
- X un espace de Hilbert s u r C et soit B (X;X)
d'adjoint B* tel que
Alors pour tout x 6 X , on

- a
P. A. Raviart
B ~ B * -112
Demonstration. Dtapr$s (3.1) ltop6rateur (-) existe et appar-
2
tient d .
(X;X) On a done
d'ob l e resultat puisque
Lemme 3.2. (Lemme de Gronwall discret) .

Si l e s 7 , s = 0, 1, ..., S, sont des nombres -> 0 qui verifient
(3.3)
7 s
<c+
-
6 2 * k r y r , r = I,...,s .? -< c 0
r =o
-
avec C > O , g > O , ona -
Demonstration. On montre par recurrence que

s-1
'fs
<c+
- 2 k
r
< C
r -
(I* 6kr) .
r=o r=o
On obtient l e resultat en remarquant que l t $kr -< exp ( S k r ) ,
3.2. Stabilite du schema implicite.
Pour Itetude de l a stabilite, on va faire s u r les operateurs A")

h, k
l e s hypothkses suivantes :
P. A. Raviart
il existe une constante fi > 0 independante de h, k, s telle

{
I
2
(3.6) que Re (A V v ) , >
h,k,o h' h h -@lIvhflh ' VVh6'h '
( il existe une constante P 3 0 independante de h, k, s telle que
Ltoperateur A peut e t r e considere comme la .partie principalen de

h, k, 0
A(')
h, k
.
Considerons dtabord l e schema implicite (2.6) .
Sous l e s hypothgses precedentes, l e schema (2.6) admet une solution
-
unique uh, E v r l lorsque
max k
s s
G .
< p2
Demonstration. Connaissant u(s)

h, Vh , le schema (2.6) donne pour
u Itequation
h, k
Pour que cette equation admette une solution unique, il suffit dtapr&sl e
theoreme de Lax-Milgram que
Mais en vertu de (3.6) et (3.7)
11 suffit donc que l'on ait pour s = 0, 1, ..., S- 1

P.A. Raviart
dloh l e resultat.
On fait l e s hypotheses precedentes, Pour tout choix des constantes E

5Y telles que 0 <E -
< 2 f i , 0 < )1 < 1, l a solution
u
h, k
du schema
implicite (2.6) vkrifie une i d g a l i t 6 de llbnergie discrgte lorsque k est-
?DL
essez petit (1 - -6
Yl
"2"
ks 1 7 )
Le schema implicite (2.6) est donc inconditionnellement stable au sens de

(3.9) .
Demonstration. En multipliant scalairement (2.6) par u + ) (et en
h, k
supprimant l e s indices h et k pour alleger 1'Ccriture) , on obtient
On prend deux fois la partie reelle de l'equation obtenue et on remarque
On en dkduit donc que pour s = 0,1, ..., S- 1
dloh pour tout & >0

P. A. Raviart
En sommant (3.10) de 0 & s-1 , on trouve
Si k e s t choisi de facon que
2P2
-
s-1
x kr/u
1
(r+l) 2
-x
< 2P
2 s-2 rtl) 2
I I .
s) 2
r=o
- t r=o k r I u ( +(l-?)lU(
Llinegalite (3.11) devient a l o r s e n divisant p a r Y
pour s = 1 , ..., S .
I1 e s t a l o r s possible d'appliquer l e l e m m e de Gronwall d i s c r e t : l e resultat
sten deduit trivialement.
3.3. Stabilite du schema explicite .

ConsidCrons maintenant l e schema explicite (2.7) . Nous allons etudier
l a stabilitC sous l e s m e m e s hypotheses que precedemment. Fixons dlabord
P. A. Raviart
une notation. Si B E (V ;V )
h h h
d , nous designons p a r ) Bh 1 l a norme
suivante de 110p6rateur B
h
On fait l e s hypothhses (3.5), (3.6) e t (3.7) Sous l e s conditions de
ot~ > 0 e s t une constante a r b i t r a i r e m e n t petite e t independante de h, k, s ,

-7
(3.15) kSc(hl2 ( f , s = 0.1, .... S-1 ,
& fl > 0 e s t une constante a r b i t r a i r e m e n t grande e t independante de

h, k, s, l a solution u du s c h e m a explicite (2.7) verifie llinegalite de 1'6-
h, k
nergie d i s c r g t e
~[r~t\I:C s=l
t L r=o k~Ifh,kIIh
(r) 2
eup (LP'
s-1
2
r =o kr)
avec
L e s c h e m a expiicite e s t donc conditionnellement stable.
Demonstration. On multiplie s c a l a i r e m e n t (2.7) p a r 4:; ; i l vient
pour s = 0,1,. .., S-1 ( a p r 6 s suppression d e s indices h e t k)

E n prenant deux fois l a p a r t i e r e e l l e de cette equation e t e n remarquant
que
2Re (u
on obtient
Mais d t a p r & s ( 3 . 7 )
*
1% (s)u(s), u(s)) + ( J s ) , u(S)) I 1 +\I
2(p u(s)l f(s)ll .(~)n <
-
-<pL!JU(S)(I tk 4
(PI U(')l t llf(~)II )2
pour tout L. > 0 , c e qui donne e n utilisant (3.6)
On obtient a i n s i pour s = 0, 1, ...,S- I

(3.18) u(stl)-u(d)I '+ (2-6)ks Re(A(s)u(s)
0 u(s)j -
<
1
-< p--t k
s
1
(P 1 ~ ' ~+ )l l f ( s ) ~ ~ * ) 2,
I1 faut rnaintenant r n a j o r e r l u - u ( ~l2) . D f a p r e s (2.7) on a

1 u ( ~ + ' ) - u(') J -< kS ( 1(0) ~ ~ I JA(;) U ( ~I )+If(') I
+
d ' o t ~ e n dtilisant l e l e m m e 3. 1 avec B = A( s )

0
P. A. Raviart
Remarquons maintenant que d t a p r & s (2.1) e t (2.3)
De (3.19) e t (3.20) on deduit que pour s = 0,1,. . ., S-1
oh 3 est une constante >0 a r b i t r a i r e . Reportant l a majoration (3.21) d e
1 dans llinegslite (3.18) , on trouve que
(s) *
lu(
stl) 2
1
. I U ( ~ ) 2t(2-,5
) - ( i t b ) k s IA:)(
A(~)+(A, 1
O )-1'212)ks R~(A?)U'), u"')
Sous l e s conditions de stabilite (3.14) e t (3.15) e t en choisissant p a r exem-

ple 5 =r,t=r 2
, on a pour s = O , l , S-1 ...,
En s o m m a n t (3.22) de 0 s-1, < S) , on obtient

(s -
I1 suffit dlappliquer l e l e m m e de Gronwall d i s c r e t pour obtenir l e resultat.'

P. A. Raviart
R e n a r q u e 3.1.
-
Les conditions de stabilitd ne dependent que des parties principales
is:, des operateurs A ( s )
h, k
.
Corollaire 3.1.
Sous l e s hypoth&ses (3.5), (3.6) , (3.7) et lorsque pour tout s

=
*
) (i. e.
l a partie principale de chaque operateur (s)
Eh, k, 0(Ah, k, 0 %,k
auto-adjointe), l e schema explicite (2.7) est stable sous l e s conditions
Demonstration. Il suffit de r e m a r q u e r que lorsque est un opdrateur

%, k, 0
auto-adjoint, on a
avec
Faisons maintenant deux hypoth&ses .dont l a p r e m i & r e ne fait que

p r e c i s e r llhypoth&se (3.6)
Pour tout s = 0, 1 , . . .,S-1 , il existe une constante

(3.26)
i telle que
vh'vh)h ?P~II~~~~E Yvh6Vh

ps >O
pour tout s = O,1,. ..,S- I, il existe une constante Ms >0
'h' 'hih M~ /l"hlh ,,vh,/h ' v"h h' Vh '

P. A. Raviart
Bien entendu, on lorsque (3.6) et (3.26) sont verifies.
Corollaire 3.2.
-S ~ u sl e s hypoth&ses (3.5) , (3.6)' (3.7) , (3.26) et (3.27) ,Je schema

explicite (2.7) est stable s i
-
2 2Ps
(3.28) ks C(h) < 2
M
(1-1)<? , s = O , . . .,S-1
S
Si de plus A -
*
) pour tout S, la stabilite est assuree s i
-7-
h, k, o - (Ah, k, o -
Demonstration.
- - .
On part de l a condition de ~ t a b i l i t e ~ ( 3 . 1 4 ) On deduit de
llhypoth&se (3.26) que
et de llhypoth&se (3.27) que
Les conditions de stabilite (3.14) et (3.15) sont donc entralnees par (3.28).
jc
Lorsque A(') = ) pour tout S, l a condition (3.29) entralne
h, k, 0 (Ah, k, 0
(3.24) et (3.25) .
-Remarque 3.2.
En pratique, on aura toujours
2
2Ps/Ms < f (resp. 2 P
/ ~ ~ )<pour P
assez grand independant de h, k, s.
P. A. Raviart
Nous allons appliquer l e s resultats precedents A ltetude de l1approxi-

mation par des methodes de diffkrences finies de problemes evoques aux
n .
0
1.2 et 1.3. Precisons dtabord quelques notations. On choisit
et on considere l e reseau 9 des points M de la forme
A tout point ~ b $on ~associe l e pave wM et llensemble 6 definis

h
Par
On designe par oMh ,M

la fonction cararteristique du pave w
h
. Enfin on intro-
v .
A
duit lloperateur aux differences fillies

hi
P. A. Raviart
avec
h h
x - i --
+T (xl,. .., Xi-l ' Xi+
-y i
' Xitl"..'xn) '
4.1. Exemple I-a.

0
On considere llexemple a du N- 1.2. (correspondant aux conditions
aux limites de Dirichlet) . On pose
et on designe p a r '(Rh) llespace des suites vh = { v h ( ~ 6) C ; M 6 flh f

de c a r r e sommable s u r Rh . On choisit alors
muni du produit scalaire
On definit ensuite l'operateur q h e ( ( ~ h ; ~ 2 ( R)) par
11 est clair que
On considere enfin
Calculons l a constante C(h) intervenant dans (2.1) . On a

P. A. Raviart
puisque l e s fonctions x 4 q v (x
h h
7 h.12
1
) ont. l e u r s supports contenus
dans R . 0,, e n deduit que
Definissons maintenant l e s o p e r a t e u r s A e t A(') . ~i l e s

h, k, 0 h, k, 1
S
, s = O,l,. . ., S-1, sont d e s n o m b r e s > 0 t e l s que
v fi6c, p.p. dans R%(t t

s' s t 1 )
on pose
P. A. Raviart
Il est clair que toutes l e s hypothCses (3.5) ,(3.6), (3.7) , (3.26) et (3.27) sont
verifiees avec p = d et
Q(x, t) designant toujours l a norme euclidienne de la matriee (a. ( x , t))

-
l<i,j<n -
Notons aussi que l e s operateurs A sont auto-adjoints s i l a matrice
h, k, 0
precedente est hermitienne p.p. dans Q
T*
Il nous reste idefinir l e s elements f(S) et u!~)
h, k h, k
.
Pour cela, consi-
2
derons 110p6rateur r € (L (R) ; Vh) :
h
11
n f
2 -1
Si f 6 L (0, T;H (R)) est donne sous la forrne f=f t i- avec
0 x.
i=l 1
2
fi 6 L (QT) pour i = 0, 1, ...,n, on choisit
s=O,l,. ..,s-1.
On prend enfin
- 383 -
P.A. Raviart
0
Tous l e s rCsultats obtenus au N 3 peuvent alors stappliquer. En
particulier, en utilisant l e corollaire 3.2, il y a stabilite du schema expli-
cite pour l e s k assez petits lorsque :
S
ob est (comme ) une constante > 3 arbitrairement petite independante

7
de h, k, s. Lorsque la matrice (a. .(x, t)) est hermitienne pep. dans
l-
1.l< i , j<- n
QT, on 0btier.t l a condition de stabiiite pour k assez petit
S
Remarque 4.1.
Si R est borne, on peut prendre au lieu de (4.9)
L e s normes (4.9) et (4.9)' sont alors equivalentes -uniformehent

-- etl h.
n
( h= ( ) I 2 et l e s conditions de stabilite (4.21) et (4.22) ont lieu sans
i=l h
la restriction nks assez petit".
4.2. Exemole I. b.
0
Passons maintenant i llexemple b du N 1.2. (correspondant aux
conditions aux limites de Neumann). On pose cette lois
On choisit toujours V,, = / 2 ( ~ , 1 ) muni du produit scalaire

P.A. Raviart
On pose encore
La relation (4.8) ne subsiste plus mais on a
Lemme 4.1.
L8application vh+]Jvh 11 definie par
est une norme SIJr

vh
.
Demonstration. Le seul point non trivial est de verifier que vh h=O
# -
entralne v = 0. En effet JR lqhvhIs = 0 irnplique q v = 0 s u r R de sor-
h h h
~ , n i l # fi) n ne reste
te que vh(M) = 0 pour tout M t ROh = ( M ( M L ~ iy .
plus qu% examiner l e cas oh M 6 R -RO Pour un tel M , on a
h h
.
fir fl R # et
-M
w fl R = 6 ; il existe alors un i tel que -M+(hiI
w
h
21flQ./
f yl ou que w
h R# .
Choisissons par exemple l a premik-
r e 6veLltualit6 : alors
- ~ + ( h 2)
w
~/
n
R # 6 et M + ~ R~; E dloh v ( M + ~ ~ ) = o .
h
Puisque
A P-(hi/2) p+(hi/2) I
-Oh
) = 0 s u r R, -
1 P E R, hi
11
M+(hi 12)
nous en deduisons que l e coefficient de , cfest-&dire
Oh
-1
(V (Mthl)
hi h
- vh(M)) , est nu1 dtoh vkM) = 0 .
Il est facile de verifier que l'expression (4.10) de C(h) reste valable.
0
On definit ensuite l e s operateurs A exactement comme au N 4.1.
h, k, i
Les elements f(S) et u(O) seront donnes comme suit. On considbre
h, k h, k
lloperateur rh&&(~2(R), Vh)
1
( rh v )(M ) =-hl. :.hn
6, ;(x) dx. MC Rh
P. A. Raviart
ob
5
v est le prolongement de v par

2
0 dans Q(h)-Q . Alors si f 2
L (91)
et u 6 L (Q), on pose
0 1
0
11 est clair que les resultats de stabilite du N 4.1. restent valables
(inegalites (4.21) et (4.22) pour k assez petit), la remarque 4.1. etant
S
evidemment mise A part.
4.3. Exemple I1 .
0
On considere ltexemple du N 1.3. On pose
On choisit
puis Vh
2
=e
'(Rh) avec l e produit scalaire habitue1 . On definit
qh
&(v 11 ;L (R)) de la maniere standard et on prend
P.A. Raviart
avec
On verifie aisement, comme pour le lemme 4.1, que (4.32) definit une nor-
me s u r V .D'autre part la constante C(h) e s t donnee par
h
i
On definit l e s operateurs A(') en posant
h, k, i
svec
puis
(4.38)
S
2 2
Si f 6 L (Q ) et s i u 6 L ( Q ) , on definit
T 0
f ( s ) et
h, k
do)
h, k
comme en (4.21),
(4.28) , (4.29) .
I1 y a stabilite du schema explicite dans l e cas general oh a est com-
plexe lorsque
- 387 -
P.A. Raviart
inf ess. Re a(x, t)
s = O,l, ...,s-1,
pour l e s k a s s e z petit. Dans l e c a s oh a e s t reel, l a stabilite a lieu lors-
S
que
toujours pour k a s s e z petit.

S
On comparera l e s resultats obtenus ici avec ceux de [is] . On pour-
rait ici a u s s i multiplier l e s exernples. Nous renvoyons & 1 2 3 pour l a di-
scrdtisation de formes a(t;u,v) plus gendrales et & [ 11 pour un point
de vue un peu different.
5. Etude de l a convergence.
Nous allons donner seulement un resultat de convergence faible pour

simplifier. Nous nous restreindrons au c a s du schema explicite, l a metho-
de et l e resultat s e generalisant trivialement au cas du schema implicite.
Soit X un espace de Hilbert s u r C tel que H soit un sous espa-
ce ferme de X et soit h lloperateur de projection orthogonale de X
s u r H. On considere ensuite une famille dloperateurs p J(v;'~ ;
2 h, k
L (0, T;H)) tels que
'h, k = 5
'h, k (V:~;L~(O, T;H)) avec l e s p r o p r i e
(5' 2, 11 qh, k 'h, k 11 f ( 0 , < C2 max ('1 (

Ivh,k h '
T;H) s=o, ...
,S
pour tout vh, = {vk

l: 6 Vh ; s = 0, I, , ,s)6vh .. St 1
. l e s constantes C1 et
C2 &ant independantes de h et k. Soit w un operateur d e d (V;X).
Soit un s o u s espace dense d e V et soit p une application l i n e a i r e
de 19 dans V
h
. Nous faisons l e s hypothPses suivantes :
-
H. 1 Si une famille [wh, k ) e s t telle que
(5.3)
on a
Ph, k Wh,k - 2
W dans L (0, T;X) faible l o r s q u e h, k 0,
2
a) (5.4) w=?TWCL(O,T;V), W=ww,
1
b) pour tout v C o e t toute fonction .f 6 C (0, T )
-
H. 2. L e s donnees f(') e t u (O)
h, k h, k
satisfont
oh l e s constantes C et C sont independantes de h e t k. De plus, on a

3 4
1
pour tout v 6 g e t toute fonctlon C (0, T)
Sous l e s conditions d'application du

-- theorhme de stabilite (3.3) et s o u s
-----
l e s hypoth&ses de consistance 1-1.1 e t 11.2 , l a solution
explicite a l e s p r o p r i e t e-------
--
s de convergence suivantes ---
u
h, k
du s c h e m a
lorsque h -et k 0 -
P.A. Raviart
-
L
(5.11) dans L (0, T;X) faible,
-+ wu -
Ph, k Uh, k
(5.12) qh, k Uh,k u -

dans LCO(O,T;H)-. faible.
Demonstration. u Soit
h, k
la solution du schema explicite (2.7) .
Lorsque l e s conditions de stabilite (3.11) et (3.15) sont verifiees, on a en
vertu du theoreme 3.3 et des hypotheses (5.7) et (5.8)
oh C est une constante independante de h et k. Ensuite, on deduit de

5
2
(5.1) et (5.2) que p reste dans un borne de L (0, T;X) tandis
h, k Uh, k
que qh, kUh,k
reste dans un borne de LCO(O,T;H) On peut donc extrai- .
r e de l a famille {h, k] une sous famille, notee toujours h, k telle que
2
Uh, k+ ,U dans L (0, T;X) faible,
Ph,
(5'. 13)
m
uh, k-t%=XU* dans L (0, T;H) faible.
qh,
2
D1apr&s H. 1, L (0, T;V) et 4 = wu
.%'
11 suffit maintenant de demontrer
que uy est la solution de (1.4) , (1.5) et (1.6) .
Soit v r A t soit 6 (C: 0, T ) . Nous deduisons de (2.7) (en suppri-
mant l e s indices h et k)
Mais pour k assez petit
puisque y ( 0 ) =0 et que (tS,l)= 0 . On a donc pour k assez petit

En passant $I l a limite gr'ace B H. 1 et H.2 , on trouve
pour tout v( fret toute fonetion ye cW(0, T 0

). Comme 29 e s t dense dans
V , on obtient (5.15) pour tout v6V. I1 e n resulte que, au s e n s des distri-
butions s u r 10, TIA valeurs dans V1 , on a
L
Mais puisque A(. ) % (.) et f(. ) 6 L (0, T;V1) , on dCduit de (5.16)
2
(5.17)
I
I1 r e s t e 3 montrer que
~ $ L6 (0, T;V9) ,
u*( t )
$(O)
+ A(t) %(t) = f(t) dans V1 , p.p. en t.
=u
0
. Soit 'f) e C1(0, T) nulle dans un
voisinage de T. Alors pour k a s s e z petit e t pour v g on a
En passant B l a limite, on obtient pour tout vEV
L16quation (5.18) et l a formule de Green (1.7) entralnent

P. A. Raviart
De la densite de V dans H, on dCduit que u+ (0) = u

0
.
Enfin de l'unicite de la solution u du problhme continu, il resulte
que clest toute la famille u qui converge v e r s u au sens indique.
h, k
Remarque 5.1.
Sous des hypothhses raisonnables, on peut obtenir L p a r t i r du theo-
2
r e m e precedent des resultats de convergence forte dans L (0, T;X) et
m
L (0. T;H) ; nous renvoyons pour ce point & 1167 et 131 Notons que .
2
dans [lo] on demontre la convergence forte dans L (0, T;H) p a r un argu-
ment de compacite : ce genre de raisonnement est utile pour l e s proble-
m e s non lineaires (cf. [i~]) .
Remaraue 5 . 2 .
On peut utiliser differemment llinCgalitC de stabilite (3.16) pour pas-
s e r & l a limite. Dans [I] , on construit L p a r t i r de u une approxi-
h, k
mation de la solution u dans l'espace
Ph, k Uh, k
muni de la norme hilbertienne naturelle et on obtient directement la conver-
gence de PhJ Uh,k v e r s u dans cet espace @.
6. Retour s u r l e s exemples.
0
Nous reprenons maintenant l e s exemples du N 4 et nous allons
montrer que l e s hypothCses H.1 et H.2 sont verifiees dans tous l e s cas.
0
6.1. Exemple du N 4.1.
On choisit
P. A. Raviart
2
(6.1) X = ( L ( Q ) ) ~ " muni d e l a s t r u c t u r e hilbertienne produit.
2 2
A tout F = (f, f , f
1 2"""
f n ) ~( L ( R ) ) Y 1 , on a s s o c i e t F = f C L (R) . On
2 ntl
definit l t o p e r a t e u r U~&(H;(R) ; (L (R)) ) p a r
w = (v, gl,...,z) 3v
n
,
1
v6Ho(R).
On prend ensuite
(6.3)
(6.4) (Phv)(M) = v(M) . V M € Rh , v d c:(R) .

Definissons l a famille d10p6rateurs Dtabord s i v eV on
Ph, k ' h h'
pose
puis si v
h, k
={v(;lk€vh ; s = 0, 1, . .., s 1, on pose
(t) = Ph (s) , t t <t s=O,l, ..., s-1.
(6.6) Ph, k Vh, k
Vh,
S -< st1 '
I1 e s t a l o r s c l a i r que l e s hypothbses (5.1) e t (5.2) sont satisfaites.
P a s s o n s maintenant l a verification de llhypoth&se H. 1. a . Si

2
W = (w,w w ) e s t l i m i t e faible dans L (0, T;X) d e p mon-
1' " " n h, k Wh, kJ
3w
t r o n s d t a b o r d que wi = .. a0
, i = 1 , . ,n. Soit 9 6 S (QT) ; on a pour
h a s s e z petit
A
J Q A i 'h, k Wh,k ( x , t ) . y ( x , t ) . d x d t = - JQT q h , k W h , k(x, t). v h p .t ) dx dt

A
2
Mais q
h,k
w
h,k
(resp Oh, qh, Wh, ) *W ( r e s p . w.) dans L (QT)
1
1 n
2
faible e t i l e s t facile de v e r i f i e r que v hi?+ 2
3 x. dans 1
L (Q ) fort.
T
00
En passant 2 l a limite, on voit donc que pour tout 96 Co (QT ) , on a
P. A. Raviart
w.(x, t).y(x, t). dx dt =+

1
-(
QT QT
ce qui signifie que w =e
au s e n s des distributions s u r QT. On en
2 1 3xi 2
deduit que w C L (0, T;H](R)). Pour montrer que W E L (0, T ; H ~ ( R ) ) il
,
suffit de remarquer que prolongement
dans R
n
Ph, k k 3, de ph, k Wh,k p a r
(0, T) converge vers W prolongement de W par 0 dans
o
(0, T ) dans
R ~ X ( L ~ (0,~ TX) ) ) ~ " faible. On voit alors comme precedem-
- 2 1 n
ment que w L (0, T;H (R ))). Si R est "assez reguliern, ce que nous
supposons, il en resulte que r o w = o d1oh W ~ 2L(0, T;H;(R)).
Consacrons nous maintenant A la verification des hypotheses H. 1.b.
1
Introduisons dtabord une notation. Si (4 6 C (0, T) , posons
b
2
On voit aisement que
fk +? et vkgk dans L (0, T) fort. On
remarque alors que
- 394 -
P. A. Raviart
Puisque
h' 4 h
-
v
2
wv dans (L (R))
ntl
conclusion resulte de l a convergence de p

fort (verification immediate), la
v e r s ww dans
h, k Wh,k
( L ~ ( Q ~ ) ) ~faible.
*'
Il ne nous r e s t e plus qutB examiner l e s hypoth&ses H. 2. Nous allons

pour cela demontrer l e s deux lemmes suivants.
Lemme 6.1.
2
Sgit
- -
uo€ L (0); alors
Demonstration. On a
Dtautre part
(rh U ~ hV)h qhf
' =~ @lo> hv)
et (6.12) resulte de l a convergence de qhfh V vers v dans
2
L (R) fort,
Lemme 6.2.
Si
- f
2
L (0, T;H
- 1(R) e s t donne sous l a forme f = f +C - afi
0 i=l 3 ~ :
- -
395
P. A. Raviart
O r il est immediat de voir que

M
f . ( x , t ) d x Z - / f,(x,t)V Oh(x)dx
JGw
Gh.
1 1 R 1 hi
de s o r t e que
A A
I1 en resulte que
On en deduit que
S- 1
s-0
<
- 2/
i = o QT
1fi(x. t)12 dx dt .
D'autre part
- 396 -
P.A. Raviart
2 nil
Puisque p wv dans (L (R)) fort, on trouve que
h PhV
Mais la derniere expression n'est autre que

e s t donc demontre.
l (f(t),v)? (t) dt. Le lemme
On a ainsi verifie les hypotheses H. 2. On peut donc appliquer l e theo-

r e m e 5.1. Si u est la solution d'un schema stable, on a
h, k
2 2
--tu dans L (QT) faible et dans L*(o, T;L (R)) faiblt,
qh, k Uh, k
(6*16)
hi
qh, k uh, k+g dans L ~ ( Q T faible,
) i = 1, ..., n.
Reniarque 6.1.
Conformement B la remarque 5.1, on peut demontrer que dans (6.16)
toutes l e s convergences sont en fait des,oonvergences fortes.
0
6.2. Exemple du N 4.2.
0
Ce qui i ete fait au N 6.1. s e transpose triGialement dans ce cas.
6.3. Exemple du NO 4.3.
On choisit ici
et on definit l'operateur ~ $ ( H ( A ; Q ) ; ( L ~ ( R ) p) ~a r)
On prend ensuite
qui est bien dense dans H(A ;R) lorsque R e s t "assez reguliern. La verifi-
cation des hypotheses de consistance s e fait comme au NO 6.1.
P. A. Raviart
7. Equations devolution du 2eme o r d r e en t. Formulation abstraite.
I
On considere l e triplet d'espaces V, H, V'] comme au NO I. 1. Soit
u, v s a(t;u, v) une famille de formes sesquilinkair-es continues s u r
V% V dependant du parametre t ([o, T] ,T < m , avec l e s proprietes suivan-
tes :
(7.1) a(t;u, v) = ao(t;u, v) + a1 (t, u, v) ,
a) l e s formes u, v + a (t;u, v) sont sesquilineaires continues s u r VX V et

0
t / u , v 6 V chaque fonction t -c ao(t;u, v) est une fois continhment differen-
tiable dans 1 0 , T] ,
(7.2) Vu,veV,
b) l e s formes u, v -a (t;u, v) sont sesquilinkaires continues s u r Vx H

1
et u, v c V chaque fonction t -c a (t;u, v) e s t mesurable,
1
On designe par A(~)&c((v;v~)

llopkrateur defini par a(t, u, v) .
On a alors l e theoreme suivant :
Theoreme 7.1.
2
Sous l e s hypoth6ses precedentes et pour f donne dans L (0, T;H),
u donne dans V, u, donne dans H, il existe une------fonction u vkrifiant
-0 I
Pour la demonstration de ce theoreme, nous renvoyons L [4] , i 5 ] .

On trouvera egalement dans 151 des proprietes supplernentaires de regulari-
t6 de la solution u. Ltunicite de l a solution u ntest connue qutavec des
hypotheses supplementaires, par exemple
l a fonction t +a l (t;u, v) est une fois continiment differentiable

(7'9) danslo,~], \/UCV ,VVEH.
Les exemples s'obtiennent de la meme facon que pour l e s equations du

l e r ordre : nous renvoyons [I] .
Dans la suite on supposera pour simplifier un peu que A(t) = A ne
depend pas du parametre t mais nos methodes sladaptent aisement au cas
general (cf. [lo]) .
8. Formulation abstraite de ltapproximation des Bquations d16volution
du 2&me ordre en t. Resultats de stabilite.
8.1. Les schemas dtapproximation.

On introduit comme au 2 l e s espaces de Hilbert Vh On choisit
NO .
T
ensuite un pas de temps k constant (pour simplifier): k = ,- avec S m.
U
Pour chaque couple { h, k] , on s e donne
A h6$(Vh;vh) (independant de k) ,
On definit alors un schema implicite. Trouver (s)

uh, k ={uh, Vh ; s =O, ...,S}
verifiant :
(8.2)
(stl)
2(uh,k
- d Sh, )k h, k
-
t U ( s - l \ + ~ hUh,(s+l) (s)
-fh, .
, s = I,. , ,S-1 ,
De la memefacon, on considere l e schema explicite : trouver u =

h, k
(s) Vh ; s = 0, 1,
= juh,
. ..
,S) verifiant
- 399 - P . A . Raviart
On va etudier maintenant l a stabilite de c e s deux schemas.
8.2. Stabilite du schema implicite.
On va faire s u r les operateurs A

h l e s hypotheses suivantes :
il existe une constante

P> 0 independante de h telle que
il existe une consta::te P > 0 independante de h telle que
Sous l e s hypothese precedentes, l e schema implicite (8.2) admet une

solution unique uh, ~i?,~'
lorsque k< .
Demonstration analogue ?I celle du theoreme 3.1.
Sous l e s hypotheses precedentes et lorsque k e s t assez petit, il exi-

s t e une constante K > 0 independante de h. k. s. telle que l a solution u
h, k
du schema implicite (8.2) verifie ltinegalitC de 18&nergiediscrete pour
P.A. Raviart
Le schema implicite (8.2) est donc inconditionnellement stable au sens de
DCmonstration. Introduisons dlabord une notation : on pose
Lie schema implicite (8.2) peut alors stCcrire (en supprimant l e s indices
h et k ) p o u r r = 1, ...,S-1
(8.10)
$ 2 U ( r + l )+ A ,(r+l) = f( r ) .
Multiplions scalairement (8.10) par
-v u( ) et prenons deux fois la partie
reelle de l'equation obtenue ; nous obtenons
On remarque que
ce qui entraine
En majorant l e second membre de cette inegalite, on trouve
+ (Ao U ,
En multipiant par < S) , on en de-

k et en sommant de r = 1 ?I r = s-1 (s -
duit compte tenu de 11in6galitC de coercivitC (8.6)
- 401 -
P.A. Raviart
S
+ 2k ( 1 +
i ~ ( ~ )p2k 2 llu(r)~/ ,
r=2 r=2
dfoh pour s = 1 , . ..,S
I1 suffit maintenant dlappliquer l e lemme de Gronwall discret P (8.11) (ce qui

est loisible s i k assez petit) pour obtenir llinegalite de stabilite p. 8) .
8.3. Stabilite du schema explicite.
Passons maintenant B 116tude de la stabilite du schema explicite sous

l e s m&mes hypothhses que precedemment.
On fait l e s hypothBses (8; 4), (8.5), (8.6) et (8.7) . Sous l e condition de

stabilite
o~ y > 0 est une constante arbitrairement petite et independante.de h, k, la

solution u du schema explicite (8.3) verifie ltinCgalit6 de llenergie discrete
h, k
Dour s = 0.1. . . .. S-1 et Dour k assez wetit
oh K est une constante > 0 independante de h, k. s et /Y

L
.
P.A. Raviart
Demonstration. L e schema explicite (8.3) s'ecrit avec l e s notations

(8.9) et en supprimant l e s indices h et k
(8.14)
vu(') + A u(r)= f ( r ) = 1,. ..,S-1 .
-
>
Multiplions scalairement (8.14) p a r v u( r ) t v urr) ; on obtient, en prenant
deux fois la partie reelle de lt6quation obtenue et en remarquant que
2Rs( V V u('), fu(r) t. V I ( ~ ? ) = ~ v l ~ ~ ( ~, ) /
+ 2 Re (A u"),
1
PU(~)) .
On en deduit que (8.15) peut s'ecrire
2 1 t P ( A u(r),
0
t V(A
0
u(r), U(r))- k2 p ( v ~~( ~ vu('))=
~) ,
= 2Re (f , p ~ ( ~~u ()' )t) - 2Re ( A ~ U ( ' ) , fu('' + vu(r))
ce qui donne en majorant le second membre de cette equation de la fapon

habituelle
- 403 -
P.A. Raviart
Multiplions llinCgalitC (8.16) p a r k et s o m m o n s d e r = 1 B s ; on trouve

pour s = l , . .., S-1
si bien que (8.17) peut s l C c r i r e
D1apr&s l e l e m m e 3. I., on a
2Re (A0u(') , VU")) -

< 2 / A J / ~ (Ao u ' ~ ) , u ' ~ ) ) " ~~ q u ( 1~ -<)
Dans c e s conditions 11in6galitC (8.18) devient

P . A . Raviart
Sous l a condition de stabilite (8.12) 11inCgalit6 (8.19) e n t r a i n e pour k assez

petit de facon que 2k 5 , 2p2k c p,~
ou K1 e s t m e constante independante de h, k, s. I1 suffit maintenant d1appli-

quer l e l e m m e 3.2. Qemme d e Gronwall d i s c r e t ) pour obtenir l e resultat.
Remaraue 8.1.
La condition de stabilite (8.12) ne depend que de A p a r t i e princi-

h, 0
pale de l l o p e r a t e u r A
h'
F a i s o n s maintenant llhypoth&se
, il existe une constante M > 0 telle que
Corollaire 8.1.
-
Sous l e s hypothhses (8.4), (8.5) , (8.6) , (8.7) et (8.21) , l e s c h e m a ex-
-
plicite e s t s t a b l e (au s e n s d e--
8.13) s i
La demonstration e s t immetliate.
P.A. Raviart
On pourrait, comme pour l e s equations du l e r ordre, illustrer l a

theorie faite p a r de nombreux exemples. Ce point ntoffre aucune nouvelle
difficult6 et nous renvoyons, comme au NO 4, i [ 2 1 , [lo] , [I] pour la di-
scretisation de ltop6rateur A(t) . L16tude de l a convergence s e fait comme
au N
0
5 . On obtient avec des hypoth&ses convenables des r6sultats de con-
vergence faible dans ~ ~ (T;V)
0 , (cf. [lo] ) . Pour des resultats de conver-
gence forte nous renvoyons A [IO] , i161, 131.
1 J. -
p. AUBIN Approximation des espaces de distributions et des opera-
teurs differentiels, Memoires Soc. Math. France, 12, 1967.
2 J. CEA - Approximation variationnelle des problemes aux limites,
Ann. Inst. Fourier (1964), 14, p. 345-444.
3 J. LIEUTAUD - These (a paraftre),
4 J, L. LIONS -
Equations differentielles operationnelles e t problemes aux
limites, Springer Verlag, Berlin (1961) .
5 J. L. LIONS - Equations differentielles operationnelles dans l e s espaces
de Hilbert, Cours CIME (1963), Editions Cremonese, Rome.
6 - -
J. L. LIONS E. MAGENES Problemes aux limites non homogenes (11),
Ann. Inst. Fourier (1961), 11, p. 137-178.
7 J. L. LIONS - P. A. RAVIART - Remarques s u r la resolution et l'approxi-
mation dtequations dtCvolution couplkes, I. C. C. Bull. (1966), 5,
p. 1-20 .
8 - -
J. L. LIONS P.A. RAVIART Remarques s u r l a resolution , exacte et
approchee , d'equations d~evolutionparaboliques a coefficients non
born& ,Calcolo (1967), 4, 2, p. 221-234.
9 A. MIGNOT - Methodes dlapproximation des solutions de certains problems
aux limites lineaires, These, P a r i s (1967).
10 P. A. RAVIART - Sur llapproximation de certaines equation dlevolution lineaires
et non lineaires, Jour. Math. pur. et appl. (1967), 46, p. 11-183.
-
11 P.A.RAVIART On the approximation of weak solutions of linear parabo-
lic equations by a class of multistep difference methods Technical
report CS 31, Stanford University (1965) .
-
12 P.A. RAVIART Sur la resolution et ltapproximation de certaines equa-
tions paraboliques non lineaires dkgenerees, Arch. Rat. Mech.
Anal. (1967), 25, p. 64-80.
13 R.D. RICHTMYER - Difference methods for initial value problems, Inter-
science New-York (1957) .
-
14 L. SCHWARTZ Theorie des distributions, Hermann, Pariz, t. I (1951) ,
t. 11 (1957) .
-
15 L. SCHWARTZ ThCorie des distributions 2 valeurs vectorielles, E r e
partie Ann. Inst. Fourier, (1957), 7, p. 1-139.
16 R. TEMAN - Sur la stabilite et la convergence de la methade des pas frac-
tionnaires, Th&se P a r i s (1967) .
(C. I. M. E. )
H. BREZIS et M. SIBONY
" METHODES D'APPROXIMATION E T DtITERATION POUR LES OPERATEURS

MONOTONES1t
C o r s o tenuto a d I s p r a dal 3-11 Luglio 1967

INTRODUCTION
Soit V un Banach sur iR ; V' son dual. On s e propose de r6soudre
numQriquement certaines Qquations de l a forme
(1) Au = f p u r f do~~ic!dans V'

03 A e s t un op6rateur monotone non ngcessairement l i n e a i r e de V dans V ' ,
Nous donnons un thbrsme d'existence e t d'unicitb.
On c o n s i d b e ensuite l16quation :
(2) AuX + A BuX = f
avec B monotone born6 de V dans V'. On montre, sous certaines hypothzses
que uX= u dans V f o r t 03 u e s t l a solution de (1).
A l'espace V, nous associons un espace V de dimension finie. Dans

h
V n m s avons a l o r s 1'6quation discr6tis6e :
h
(3) $ Uh .- fh.
Sous certaines hypothzses l a solution B de ( 3 ) converge fortement
dans V vers l a solution u de (1).
Enfin nous donnons une mdthode i t e r a t i v e permettant de rgsoudre ex-
plicitement (3).
Le plan e s t l e suivant :
Z - un theorerne d'existence e t d'unicit6,

2 - h o p r i 6 t 6 s de 1'opi;rateur A
-1
3 - PropriEtS de l a solution de l'equation AuA+ BuX=f
4 - Applications
5 - dthodes dlapproximtions numeriques
6 - Application aux famillrs dqappmximtion { v1 ~ ' ~Vh,, ph,
i rh)
7 - Resolution du probldme discretise : dthode iterative

8 - Applications & l a r6solution nm6rique de certaines equations aux
d6riv6es partielles
I - Un th6or;me d'existence e t d1unicit6
~ 6i nf i t ion 1.1
On d i t qu'un espace n o d V e s t uniform6ment convexe si, Y E t e l que
o < E < 2 ,3 6 ( ~ )> 0 t e l que l e s r e l a t i o n s :
IIuII:1 9 IvI121 et Ilu-VILE

-- IIu+vII<2-6 , Vu,vtV
Soit V un Banach sur lR uniformkment convexe de nome )I 1). Soit V1 son
dual. Soit A un op6rateur non nkcessairement l i n 6 a i r e de V .dam V1. On
cherche u c V verifiant :
(1.1) Au = f" pour f donn6 dans V'.
Th6or$me 1.1.
S i l1op6rateur A v 6 r i f i e l e s conditions
pour tout u e t v B V, O; y e s t une applicatior, strictement croissante de
I\.+ IR t e l l e que
(1.3) Les r e s t r i c t i o n s de k aux segments de V sont continues
dans V' faible.
&or$ f E V' , i' u c V unique t e l que Au = f.

Exists~cc!
L3. condition \u strictement croissante entraine :

(1.4) -
(Au Av , u-v) 2 0
D'autre part, si on fait v = 0 dans (1.2) il vient
(~u,u)2 (A(O) ,u) + -

(~((lul() ~(0))llull
et l'on sait qu'avec les hypoth2ses (1.3) (1.4) et (1.5) le problsme
(1.1) admct une solution. cf. 131 et [43.

Avant de dgmontrer 1 'unicit6, nous avons besoin de deux lemmes
Leme 1.1.
On suppose (1.2) et (1.3). Aiors l'ccsemble 15 des solutions de 116qua-
tion Au = f est un convexe fern6 de V.
DEmonstration : L'Cquation ?.u = f est gquivalente 2 la relation
En effet : Au = f + (Au - f , V-u) 0 et

(Av-f,v-u) = (6u-f,v-u) + (AV-AU,V-u) 0
d'aprss 1s monotonic de A.
RCciproquement pasons v = u + t w , t; > 0 : (1.6) donne V Y EV
(:,(u+t ?)-f ,'f) 2 0 et l'on fait t -4 0, E'OU
Au = f.
S I
v = {U (avo-f,v0-u) 2 O! est un dcmi enpsce fern6 de V.
0
Donc S = ensemble des solutions de Au = f e s t un convexe fern6 de V, c z
Lemme 1.2.
L'ensemble des solutions de Au = f e s t s i t u d sur une sphe're de V,
!@nnstration : Soient u e t u2 6eux solutions de Iq6quation Au = f .

1
En reportant dans (1.2) il vient :
- L f ( ~ ~ u 2( ~I U) )~ -I II I U ~ I=I o)
(~(nq)
~ I o i i: llulll = lb2i
-
UnicitE
L'ensemble S des solutions e s t un convexe ferm6 de V s i t u 6 sur une
sphbre. Comme V e s t uniform6ment convexe, s est r6duit 2 un point, dl06
llunicit&.
Leme 2.1,
Soit r c 8, e t r A6 8+ , V 2 0. Soit y une application strictement
croissante de R+ -+8 t c l l e que

(9(ri) =, ~ ( r ) ()r i =- r) -3
A o 0
Alors rA-r quanC X - 0.
J!+h,o"!+h~_r$j~o : On montre facilement que r e s t borne. Supposons main-

X
tenant quc r A-rj r. U o r s on pourrait e x t r a i r e r
u
,
p # r e t V(rll) +$r)
Dew cas :
1) S i p < r
S o i t r t e l que ;; < r < r & p a r t i r dsun c e r t a i n rang r < r ce qui

1 1 u- 1
donne :
y ( r p ) 5 't'(rl) < q(r)
e t & l a limite :
ce qui e s t absurde.
2)Sip>r
Soit rl t e l que r < rl r ce rlui

u- 1
entraine :
y>(rp)2 y(r1) > k ( r )
e t 5 l a limite
ce qui e s t absurde.
C.Q.F.1).
On suppose (1.2) c t (1.3): !!I-ors , e s t nonotone, born; ( c ' e s t 2

'31
d i r e transforme l e s e~sembiesborn& en des ensembles bornks) e t continu
de V' f o r t dms V f o r t .
Dgmonstration : I1 e s t 6vident que A-'

-u-"-.s-------
c s t ~ronotone, Soit Au = f crvec
I l f l 11.
~ On a
(nu - A(o),u) = (f t . n(o),uI , ( Y ( ~ U I J )~ ( 0 ) l)l ~ l l

-
D ' O ~: .p(/lu(J)2 a + lla(o)ll + d o )
Ceci montre que *I-le s t borne.
Montrons que A-' e s t continu de V' f o r t clans V f o r t . I1 s u f f i t dc
m n t r e r que s i f n = Aun e t s i f n --+ f = Au dans V1 f o r t , a l o r s un -+ u
dans V f o r t .
La s u i t e f e s t born&; donc u aussi. Suivant un u l t r a f i l t r e plus f i n

n n
un -j 5 dans V f a i b l e , Aun f dans V' f o r t , Par s u i t e A{ = f = Au.
Dl03 u = 5 d'aprss l t u n i c i t 8 . I1 en r g s u l t e que un --+u dans V f a i b l e .
Dl autre part
(Aun - t~u J Un =* U) 2 ('f'(lunll) - ~ ( I I u I I ) ) ( I I u-~ ~[ ~ i l ) O
et
(Aun - AU , un -- U) = (Aun f , un - U) -4 0
Donc
(~(llunll)" ?( llull) (llunll " 11211) +0

DVapr6sl e lemme 2.1., [IUh! --+ [(uII. D'ou un --+u dans V f o r t .
Th6orbme 3 $1.
Soit A un op6rateur de V dans V v ir6rifiant (1.2) e t (1.3) e t s o i t 73
un op6rateur monotone borne de V - - + V h t l que l e s r e s t r i c t i o n s Ftia see-
ments de V s o i e ~ tcontinues dans V v fzible. Llors
1) v f c- V' , V A 2 0 l'eqmtion
admet une solution unique uA .
2) uA4 u d m s V f o r t quand A 0 02 u e s t l a solution de lq6qua-
t i o n Au = f.
Gmocstration : 1 ) Pol,r l'existcnce e t 11unicit6 on applique l e thko-

----"-..--"I-
rsme 1.1. h 1'Qquation (3.1).
2) bntro:lo quc ul - u d m s V f o r t , quand h +0

e) 3 N > o tel que ! u A l 11,
~ POUP t o u t A ~ S S O Zp e t i t . NOU. avon.
en e f f e t r (Au1 - ~ ( 0,uX)
) 1 (Y( I(uA(( ~ ( 0 ) [uA(I "
or AuA = f - A BuA
( f - h BuA - > i ( ~ ) , ~ A )('f'(((uA(1)- 'do)) nuA[(
Dqautre part
-
(BuA B ( O ) , U ~ 2
) 0
=+ (f - ~ ( 3,uA)
) 2 (Y(!U,,(() .- ~ ( 0 )CAI(
) k(B(O1 ,uA) +
I~r(llu,ll) - r ( o ) ! lu,ll 2 llf ,- A ~ ( o.-l A ~(o)ll,, IluAl,

D'oG
(3.2) V ( ! ~ U ~c! )Y ( 0 ) .+ [If - n(0) - A B(O)~,,,'
ce qui entrainc Ilu A [ bornE, car sinon il e x i s t e r a i t u LI t e l l e que
ll~+,ll -, +a,
; e t on aurnit d o r s l ~ ( J l u ~-+
J ) ) +a, ce qui e s t contraire 2
(3.2)-
b) u h + L u faible~.eat.
h effet CO~BLL' 3 C S born6,

~ 1 ah()2 ;! e n t r d n e I(BuhII( W' . Donc quand
t,uh -
Soit % un u l t r s f i l t r e plus f i n . Suivalt ?A, uh V-
.
-
faiblement e t
f fortement I1 en r 6 s d t e que Av = f = Au. Donc v = u. En e f f e t ,
on va montrer que s i u
X v faiblement (iu,ll ( !I)
e t ,4ui -3 f f ortene~lt
9Av = f.
-
Nous Rvons (AuX A w , uh - v) 2 0 v v 6 V
e t & l a l i m i t e ( f - Av , v - v ) 2 O d v L V,
or ceci e s t 6quivalent ?liv = f. Donc uh - > u faitlemcnt.
c ) Montrons que u u fortemcnt , On a

X - r F -+
(AUA - fiu , uA - U) 2 ( Y ( I ( u ~ I-( ) 'f'(bD)(IIuAll - IIuII)
( f - h Buh - Au,u X - u) 2 (Y'(l(uXll- v(lull))(llu,ll - !UII).
ce qui donne
- X ( B U ~ , U ~ -LU() d / l u X [ j.= Q ( I U I ( ) I ( I I U- ~II:1II)

~~
Come u -u e t Pu, sont bornks ; quand 1 -+ 0 l e l e r mcmbre tend vers 0.

X
Donc :
11 rgsdte d.u lle&me 2.1. que [ u

h
1 4 IIu 1 e t come V e s t unifom6ment
convexe on en dgduit que uA ---t, u d m s V fcrt .

IV - Applications
Exemple 4.1.
$2 &ant un ouvert born6 fie Soit Y = \ i 9 P ( ~ () I ) muni de l a norme
V est un Banach unifom6ment convexe pour c e t t e norme. L10p6rateur
AU = - i=l
11
rs D. 1
( I D ~ U ~D P~ -U~applique
) <'P(iI) dans W - ' ? ~ ( Q )avec -P1 + -Y1 = 1.
Montrons que A v 6 r i f i e l e s hyyothsses du th6or;me 1.1. avec'
~ ( r =) F1, r 2 0. En e f f e t :
( I ) On d6signe par d9'(9),

W p c W llespace des (classes des) fonctions
u LP (a)t e l l e s que D e w = s C ~ P ( ~ ) , 11s dd$riv6es C t a n t prisen au sen3
1 axi l.tlv(fi) e s t ltadh6renke de 3 ( R ) lane
des d i s t r i b u t i o n s sur S.
$yp(Gi). ~ ' e i ~ j uc -c' ' ~ ( Q ) , l/p +;/,I = ! "fsignc l e dual f o r t de I&"(").
0
-
Donc l a condition (1.2) e s t vgrifike. I1 e s t evident que l'application
t (A(uttv),w) e s t coiltinue v u,v,w E V, dl05 (1.3).
Par consequent pour f donne dans w-~"(G), il e x i s t e u c d0t P ( n )
unique t e l que
ku = f.
Exemple 4 - 2 .
Plus &nGralement l e s problsmes
admettent une solution unique u e w~'I'(R)

pour f donn6 dans ~ ~ ~ ' ~ ( 0 ) .
0
Soit V un espace de Banach de norme I/ 1, e t s o i t h = (hl,. ..,hn) e

n
Wt
h # 0 un p a r a ~ ? t r sdestin6 2 tendre vers 0.
D6finition 5.1. (cf. [ l l )
Nous appellerons h-approxireat ion de 1'espace V , l e quadruplet
{V,Vh,ph,rh} d e f i n i par l a donnee de
1) Un espace V de dimension f i n i e e t de normc

11
11 1) h'
2 ) Une aynlication p C- Z(7 ,V.) appelge prolongement de Vh dans V.
h h
3) Un, ;iplication r $(V,V!,) ap?elge r e s t r i c t i o n de V dans Vh.
li
Dgfinition 5,2,
La f x n i l l e {V9ilh,phsrh}e s t d i t e consistante s i dans V f o r t
Soit V,' l c uual de Vii. IOUS

notons par ( , )h l a dualit6 entre Ul c t VA.
V"h e s t muni dc l a norme duale
Dgsignons par r[ l'adjoint de ph e t par p*h l ' a d j o i n t de rh. Nous svons
ucf'nition 5.3.
On d i t que f,
1.
c V k ?onverpe discrdtemcnt vers f 6 V' s i
Soit V un Emach sur il?, w i f o n r C ~ c n tconvcxe, de norme 11 1. Soit V'
son dual. Soit I: un opirateur non nc:ct:ssoirement l i n k a i r e de V d:ms V'. On
se donne uno f m i l l c d7apyoximatior. tV,V:+,pk.r h ) de l'espace V *t I'on
f a i t l e s h j ~ o t r ~ c s esuivantes
r :
(11 1 ) .? c s t ur, op6rateur born6 v g r i f i a n t (1.2) e t (1.3).

H 2 ) La familie {V,Vh,ph9rh} e s t consistante e t ph e s t injective.
On pose 2% = r: A ph : Vh -9 Vi.
Nous avons a l o r s l e sch6ma suivant :
On munit Vh de l a norme llUhlh= Ilphylb.11 en r 6 s d t e que
D1 autre p a r t s i f
li
converge discrstemcrit vers f a l o r s 1 I
h h
1h
e s t born6.
En eeffet :
Th6orSme 5 v l .
Pour t o u t f E VA, .i.1 e x i s t e uh E Vh unique t e l que

h
De plus si f converge discretement vers f 6 V P e l o r s phu, converge dans

h
V fort vers 13 solution u de (1,~).
Lemme 5.1.
~ ' 6 q u a t i o n (5.1) admet we solution unique.
Dkmonstration
..-"-------"-- : Apnliquons l c thgorime 1.1. 2 190p8rateur i\h'
1 ) Montrons que Vh mi Ee l a norme I(u,llh= I ( "Uh I( e s t uniform&
memnt convexe. En effet :
t/6 > O , O c E < 2 3 6(c) t e l que s i
d ' a p s s l'tmiforne convexitk de V, e t ceci cntreinc
3) L'applicizt ion t -,(:,,,(tu,,+v5) ,wh)h = (-lph(tuh+vh),phwh)
est continue d1spr8s (1.3). 2roG lakxi~tuncee t lq*micit63.e % solution

de (5.1).
Leme
--- 5,2.
S i fh converge discrGtement vers T dors
ce qui donne :
(fh , - ~ ( 0 ) Il%llh
- % ( o ) , % ) ~ (~(Ily,Hh) )
D'ot .i(lly,l'? 2 ~ ( 0 +) Ifh .. ~ ' 0 , I l ; ,

La convergence discrbte de f h vers f + IlfdlL born6 e t
@r. a done ~ ( l l t j l / , L~ )cte. i e c i entraine que l vhll h e s t born6 car sinon, il
existerait une suite cxtraite 3: t t e l e que l l ~ ~ - l,l ~+co e t on aurait

ukI k) .-+ +a ce qui e ~ inpossible.
t
Leme 5.3.
Si f 11 converge discrGtement vers f : alors ph uh converge vers u dans

V faible e t Aph y, converge vers f dans V' faible.
D6monstration : D'qrbs l e lemme prdc6dent

.s..---"-<*-".,m-
Il p h %(I 2 ~ t e .Done sui-
vant un u l t r a f i l t r e ; ph y,-3 5 &ins V faible e t comme Aph 5 e s t born6,

Aph 5 + II dans V' faible.
Iiontrons quc rl :z Au = f. Soit v 6 V
D'autre part (fh,rhvlh -+ i f ,v). En eff5t :

I(fh-rlfYrhv)hll Ifh - $fl: Ilrhvllh
I (fh-$f 9rhv)h 1 2 8 fh-rif/l,*
IIphrhvl1 O
(d1apr8s l a convergence discr?te dc f ver7 f ) .

h
et (r;f,rhvlh = (fy$rhv) -+ ( f , v )
., IOU v = v vv C v
e t par cons6quznt : 0 = f = Au.
~ % ) = (fh-rip,uh)h + (r;f ,uJh

Par x i l i e u r s : ( L ~ ~ , =~ (%,uhlh
et
(fh-rif ,%Ih -
) 0
( x i f > u h ) h= ( f ,ph%) (f
D'oii
Rappelans cP. k] qw si un f i l t r e x. converge vers x dm? V f a i b l e ,

1
Axi - 3 y dans V' f s i b l e e t

linl sup (LK;,x~) 2 (Y,x)
I1 ?n rdsultc qu* A( = f = fi:d e t donc dtnpr8s l'unicitk de l a solu-
tion u = 6. Cl'0i.i l e lcjme.

(fh,uJh = (f "#f,y.,)h + (f,ph~)
h h
on en d6duit que
l i m ( A ~ ~ - A U ~ ~ ~ I=+ 0, - U )
h+o -
D oii
l i m (9
h+b
(!P<+# 1 - ~ ( l l ~ l l ) X I l ~ ~ ~=~0l l - l l ~ I I )
c t par consEquent lP,lhP --311ul.
On en d64uit que ph\ -3 u clans V fort.
VI - Application aux f k l l e s d ~ a p p r o x i m t i o n{\Ii1",vh i

,Ph ,rh1
soit ( 9 ~ = ( v ~ . ~ ~ L P,( Div

Q) E LWI.Q h t un ouvert born6
do iR" -
e t ~ ~= ' f prig? au sens des distributions.
axi
Choi;~2e Qh
Nous dksignons par % l e r6senu dcs points it de l a forme

\ n
M = (yhl,.,.,m h )
n n
, mi c Z , i = 1,n , !I = (hi ,...,hn) , h 6% ,h # 0.
Nous posons :
1 = (qi)l<i<n
--
, qi entier p o s i t i f lq ( - ELi=:lqi
r$ = 0 1 1 11 e % , ,>; Q}
hfin :
Choix de Vh
Vh = espace des s u i t e s de nonbres r b l s
s o i t qi$.,.On pose ( y l ( x ) =
>I
I+, s i x a t,oVh . e s t un espace de dimension
f i n i e , que nous mettons en dualit6 evec V;1 par ? e ~ r o d u i tsciiiaire :
Posons h.
g(x + T
.1
- hi
- )-q ~
vi4(x) = h.
1
Choix de d
Yotons
m
0 = fonction caraetgristique de [(mi - B)hi
1
, (mi 1
+ i)hi]
hi
,t e0
Xhi hi hi
n m.
f = fonction caract6ristique de {,o = i=l8hi'
@
((1) Want liopCrateur de ;onvolution), e s t un prolongement de Vh dans <7P
On a i d e n t i f i g % l a fonction $ =
-7
4I' %.
-" V e s t muni de l a norme
Choix dc r
M
Soit une s u i t e de fonctions y
h
6 dymsupport dans lh = &-$
I;Oh
tendant vers 1 dans Posons a l o r s
rh u(U) = <yhy 9 yh> 1 , Y 6 d 1' P

On =it ( e l . [I]) que l e a a,proximztions (1;"
1
Vh ( a ) , q,i , ,ah) sont
consistantes.
Application 6.1.
s o i t A : \ i r P ( n ) - 9 ~ - ~ ~ ' ( ! ? ) ddfini psr
ktant un ouvert born6 de ZZn Be fronti6rc r "trss rggulisre . Au probl&ne

de Dirichlet :
(6.1) A U = E ulr=o y o u r f r ~ ~ ( ~ )
nous nssocions l a formulation variationnelle
avec
avec
D'apr;s l a dgfinition ie 1%' nnus avons :
e t f h converge discr2tcment vcrs f .
D'd~r&il e thboramc 5.1, il e x i r t e % c: Vh unique vgrifinnt (6.18) e t

'i
1'on 3 ph% u dnns d*P(Cl)
1
fort ; u i t i n t l a solution du problzifie (6 .I).
Appl.ication 6.2.
On cherche u c $yp(,2j v6rifiant

0
(6. 5 Au - SAU = f s20 f don& dens L ~ ( ~ ~ ~ )
ave c
On vgrific ctisfmnt q,ue l e tl:.iorGroe 1,l. svapplique e t l q o n a donc i'exis-
tcnce c t l * u n i c i t 6 dc l a solution u &c ((6.5).
Corn dans lVexempleprgcttent (6.5) discrbtiske se met sous l a forne

S i f h converge discrdtenegt v f r s f sn peut appliquer l e t h 6 o r h e 5.1, qui
a s s u r e l P e x i s t c n c e c t i 9 m Z c i t 6 de l a sclutior. % de (6.6) c t de plus nous
avons :
;
ph B 2- u sans ?;P( 17) f o r t .
VXI - H6oolution du pro5l&e d i s c r 6 t i s 6 par une mgthode

itGrat ive.
On s e propose dans cc paragraphc de dancer w e m6thode i t g r a t i v e per-
mettant cle r6soudre dms Vh (de dimei~sionf i n i e ) lsEquation
All ul1 fll ; f h donn6 dans V'

h
fi
Plus ~69Eralemcnt s o i t un espace de Hilbert s u r 53 muni du produit sca-
laire ( ,) e t de l a nolme I/ . On se propose de rksoudre dans % i q 6 q u -
t i o n Au = f 4 pour f dom6 dans I!.!, oG A e s t .un op6rateur non-lingnire de
1;dms .X3. On sc donne ?me qq.-4ication S 6 ) verifiant :

2 (&:,g
(7.2) 11 & s t e w e c o n ~ t u i t cc > 0 t e l l e que (Su,u) 2 G l l ~ 2l l

On posc [ul2 = (Su,u). [u] d e f i n i t SLU. $ >mi.nome kquivdente 2 11 (1.
On s u ~ p o s eque Isop&rateurA v g r i f i e :
('7.41 Qttelque s o i t N > 0 ,3 un? constante C(M) t e l l e qut
U,V a ?I>, avec [?1] _t , [v] -< N alors
(?.u-:Lv,W )(c(:!) [u-V] [w] d W E %

Nous avons avons l e
Th60rsme :,1
Avec I e s hypotheses (7.1), (7'21, (7.3), (7.4) llQquation 1Zu = f adnet
une solution unique u e '2 e t l a s u i t e un dgfinie par l f i t 6 r a t i o n

S U ~ =~ Sun
+ ~ - p(aun - f )
converge fortemcnt vers u pour p > 0 c t u c % c~nvenablementchoisis.
0
-gmonstration
-*.-."--.- : L9existence e t i 9 u r r i c i t 6 r6sultent du t h e o r h e 1.1.
Dfautre part, l f i t 6 r a t i o n a un sens puisque S e s t bijective.
En retranchant menbre 5 membre :
.- Su r Sun .-. Su - p(Aun - Au)

Posons an = un-u il vient :
Soit No t e l que [u] 2 N0/2 ; on f a i t llhypoth?se de recurrence
[un-u] = [rn] 2 No/2. Ceci entraine que pJ No. Or
( A U ~ - A U , S ~ ~ ( ~) ~=- A
[sW'(hn-~u
U) )] 2
e t d7aprZs (7.4) e t 19hypothZse. de r6currence on a pour C(N ) =. C
0
ce qui entraine
On c h o i s i t p t e l que 9 < 1 e t l'hypoth~se 2c rrcurrence e s t vgrifi6e. ?ar
consdquent c --+O. Ce qui achsve l a d6monstration.

n
En p a r t i c u l l e r l a s u i t e u convcrae vcrs u pour uo = O ,

n
2
N = , C=C(No) e t 0 =k/c 2 c e q u i d o n n e 0 = 1 - - b < 1.
0 kG opt c2
VIII - ."pplicztions 2 l a rEsolution numerique de certaines

Bqwtions aux dEriv6es p s r t i e l l e s .
Dans ce qui s u i t nous nous proposons d'appliqwr & quelques exemples
l e s d t h o d e s prdcedentes .
avec
-n
Au =-id 0. ( I D ~ U ~Diu)
~ ' t su s > o
1=1 I
pour f aonn6 dans ~ ~ ( 0 )

~onsidgronsl c probldme discrEtis6 associ6 5 (8.1) : On cherche ~1 E Vh
t e l que
(8.2) nh '-'ti =
avec
n
r, P-2
Ah%=-{giVi(I~i\l Vi\)+s\ 0 0
T?16or2me 8.1
i ) l l l q u a t i o n ((HI) adnet une solution unique u r t y p ( n )
i i ) Le probldme (8.2) admet une solution unique % E Vh. De plus nous
avons
%% --+. u dans I~$$*(Q) f o r t

0
i i i ) L'iiquation (8.2) peut s e r6sou&:! par 1 7 i t k r a t i o n s u i v m t e :
$' = .;: *-
n
.(,\i,%-.fn)
avec p convc~iablemntchoisi e t ~10 = 0.

D6~ons~,ratior,
--"-"-"..-----.- : Lrs points i ) c t i i ) ont &j?i6t6 d6montr6s aux para-
graphes 4 e t 6. Pour dcmcntrer i i i ) , upnliquons l c s r d s u l t a t s du th6orEmc
(%,vh)h = (~~,q!)$
Posons
2
Come : ( \ u ~ ; - ~ $ v ~ , u ~ -2. ~s ~lly;vhll
)~
L
on a avec l e s notations du thkorsme 7.1. k = s
Calcul de C(N)
S u ~ ~ s o nque
s [%I 5 9 rvhl h]l lo
on a :
[%12 = hl
T
h2...hn lUII(~)12
ce qui entraine
[\j2 -> hl...hn sup lu,,(H)I 2
De msme nous avons v x
2
(8.3) 1 %(XI I 2 h.I 4 lo"'n 6 [%I 2 hi TT;
h n 1
I !J!
Par a i l l e u r s
n
( % % - q h ,wh)h = & ( I viYr I vivh lP2vivh. viwh +
' s(%'vh5wt,)h
ce qui donne
n n
1
I (%\-$'hs~hU 2 (&(I vi%lp2viuh-! 'ivhlPZpivh[
' !whll
orpour lul2ti9 , (61(ii1 , act 6ri.elsona:
I lulPs2a - 161 ~ ~ 5f ( lp ll) 1, - 81

En posant a = vi\(x) , 6 = vivh(x) , il vient compte tenu de (8.3)
n n
Donc :
.
.n 411:'~~'
1 (L\\-:'hvh,wh)h 1 5 (p-1) !i.L+
1 -e
h.
l l ~ ~2112
l ~ ~ a 5
!i=l 4 IIwhII 2 1 1 2
1 hi
Donc on a avec l e s notations &u tth6or&me 7.2.
ce qui donne
Donc '1
0
=2 11~+11 ,,._,e s t
I
me constant^^: inde72ndante de h. En posnnt :
1.t
s 1 i t g r a t ion n
P =-
2 {+'
=% ,- f(3
?A%-
11
fh ) e s t convergente d 1 ap&s l e
C th6or;me 7.1.
Remarque 8.1.
2
3 i f E L (II)
on pourrait arendre aussi No = -
IlfhlL2
ir
Remarque 8.2.
0
posone 0 = 1 .- 2- uh = 0, nous avons a l o r s l a formule de majora-
c2
tion de l s e r r e u r suivante :
Le nombre d.'it6rtions n6cessaires pour avoir [<-%I < E p u t &re estim6
n o
2 Log lull ..
2 Log E
Log 0
bemple 8.2.
On cherche u r W;*P(Q) v g r i f i a n t 116puation
(8.6) Au = f pour f donn6e dans L (Q)

P
avec
Consid6rons l e probldme discr6tis6 s?ssoci6 : on cterche \ c Vh v 6 r i f i a n t :
avec
Th6orSme 8.2.
i ) ~ l & c l u a t i o n(8.6) admet une solution unique u r \tvp(n)

i i ) Le problzme (R.7) admet une solution unique ~1 r Vh. De plus on a
u aans u:'P( Q) f o r t
PhUh -hG3
i i i ) L1&quationd i s c r e t i s 6 (fi.7) ?cut s e r6eouare & l ' a i d e de l a for-
mule i t b r a t i v e suivante :
avec
P 2
, p convenablement choisi et, yl0 = 0
Shy, = - f . Viy,
iamonstration : Les points
--."-----..--- i ) e t i i ) ont 6t6 d6montrC dans l e s p r a -
graphes 4 e t 6 . Pour i i i ) , nous appliquons l e s r e s u l t a t s du theorzrne 7.1.
sn posant : "ab = Vh (.% ) 6u produit s c l a i r e (uh,vh), = (I+,,V~).~
n
(% ' h,Yh 9 y,.'vh)h .f ' ElI / v i ( v v h ) l 2g
L
il cn r6sultc quc k :
: s avvec l e s riotations 2u t h e o r h e 7.1.
Calcul
-- de B
On a
o r L c L, w e c injectisn continue :)our y 2 2. Donc

F L
Donc Yo = cte ind6pendente rle i?. En yceznt i. = --

S
1' i d r a t ion
c"(w)
3 n+1 3 n n
-
hUF, " h"h :)(:k~i'fk*)
e s t converceate. D' oh l e thbor?rne.
2
1) s i f e L (Q) on pourrait prendre aussi
ncus avons alors la formule de majoration de llcrreur suivante :
o; C (Q)est une constante nc &pendant que de ilouvert Q.

1
3) Si n w alors on ;
Log 0
Remrque 8.4.
On suppse que f t L2 ( a ) . Si l'ouvert Q = (0,l) * (0,l) et hl = !i2 =h

alors on a :
Cas particulier
Si n - nous swns ~ l o r s
(~1~1);
Pour ceux qui elinteressent aux resultats numdriques cf llarticle
de Brezis-Sibony B paraftre dans Arch.Rat. Math. and Mech,

BIBLIOORAPIIIE
--
[I] AUBIIt ~ p ~ r o x i m a t i o&es

n eespaccs de Cistributions e t des
op6rateurs d i f f g r e ~ t i c l s ( T h h e 1966)
[2] H. BREZSS 0p6rateurs rnonctonor. C.R. Ac. Sc. de Faris ( d v r i l 1967)
[3] F. RR3WI)ER Existence anu uniqxcn?cr sheorems f o r solutions 02-
non l i n e a r boundarj valuc problems.

Proceding of symposie i n q l l i e d mathematics , vol 17, .
p. 24-49, h e r . h t n . Soc. 1965.
[k] J. LEKAY e t J.L. LIONS Quel,;ues r 6 s u l t a t s de Visik sur l e s pro-

blemes e l l i p t i q u e s non l i n g a i r e s par l e s m6thodes de
E!i~ity-Browder .
Bull. Soc. Math. France 93, 1965, p. 97-107.
t6] R.S. VARGA Matrix i t e r a t i v e analysis, Prentice Hall.
[5J PETRYSHYN 1 ) On the e x t e n s i o n and t h e s o l u t i o n of non l i n e a r

equations .- I l l i q o i s J o u r n a l of Mathematics
vol. ION0 2 , June 1966
2 ) Cours C.I.M.E. 1967
(C. I. M. E. )
V. THOMEE
"SQME TOPICS IN STABILITY THEORY FOR PARTIAL DIFFERENCE

OPERATORS

'SOME TOPICS IN STABILITY THEORY FOR PARTIAL DIFFERENCE OPE-
RATORSW .
by
Vidar Thomee
(University of G0teborg)
1 . -----------
Stability in L
P
d
Let W denote L (R ) for 1 -
P P
 . . J d ) . ,
n
m
F o r l a t e r use we shall also introduce the Sobolev space W of distribu-
...ax,
dd P P
t i o n s s u c h that D ~ ~ = ~ 1 - ~ ~ ~t wPs f o rM/ d l =I ~ tJ-~( . <m~ ~ ~
j
(d= ( a .. .
,q d ) ) This i s also a Banach space with norm
4
~ e t =W
: , W; = myo Wm ,d * = W*m . By Sobolev's imbedding
w
theorem, W consists of all infinitely differentiable functions with
P
Ddu C Wp for a11 0( . For 1 -< p - < co , W* i s dense in W When
P P
.
u is a N-vector u = (ul, ..., , we use again the above definitions
where in ( l . l ) \ u / = (El~jl 2
) '12
U ~ )
.
j
(I.2]
Consider the initial-value problem
- P(x, D)u 2
3t- /dl5
- Pd (x) D~ u, t-
> 0,
(1.3) u(x, 0) = V(X) ,

d
where x C R , u = u(x, t), and v = v(x) a r e complex N-vectors and
Pd (x) a r e NxN matrices with elements which we assume for simplici-
ty to be in 4
The initial-value problem (1.2), (1.3) i s said to be correctly posed
in W if P = P(x, D) (considered a s a densely defined closed operator
P
in W ) i s the infinitesimal generator of a C semi-group of operators
P 0
E(t) on W for t -
> 0 , that is the family of bounded operators
P .
E(t), t > 0 , satisfies
E(0) = I = the identity operator ,
The operator E(t) i s r e f e r r e d to a s the solution operator connected

with the initial-value problem.
The definition of correctness thus depends on p ; the c a s e most
discussed in the literature i s the case of correctness in LIn the ca-
2
s e of constant coefficients, using Fourier t r a n s f o r m s one easily finds
that the initial-value problem i s correctly posed in L 2 if and only if
for any T-
> 0 ,
F o r a NxN matrix A with eigenvalues i J. ,j = 1,. .., N ,

we define
A (A) = max Re A. .
j J
We also set
ReA = -21 (A t A*) .
Clearly Re A is a hermitian matrix. F o r hermitian NxN matrices,
A -< B means (Au, u) -
< (Bu, u) f o r a l l N-vectors u, w h e r e
N
A n e c e s s a r y condition f o r (1.4) to hold i s that (1.2) ,(I.

.- 3) is c o r r e c -
tly posed in Petrowsky's s e n s e , namely
(1.5) sup [ A ( P (I) ) ; ~ E R <~ m) .

N e c e s s a r y and sufficient conditions have been given by K r e i s s [I47 . It i s
not o u r a i m h e r e to give a complete ac'count of his work, but we s t a t e
one s u c h r e s u l t which we shall need in Section 2 :
The condition (1.4) holds if t h e r e a r e positive constants
T h e o r e m 1.1.
d
7
C1, C2 and for e a c h / g R a h e r m i t i a n m a t r i x H({) such that
and
Re(H( 5 ) P ( 5 1 C2 I a
On t h e o t h e r hand, if (1.4) folds we have (1.5) and t h e r e a r e positive

constants C , C and f o r each 16 R~ a positive definite h e r m i t i a n
1' C2 3
matrix H( ) satisfhying (1.6) and
We shall consider the c a s e of a hermitian (hyperbolic) s y s t e m
In the c o n s t a ~ l tcoefficient c a s e we obtain the c o r r e c t n e s s in at once L

2
f r o m the fact that e x p ( t P ( 5 ) ) = exp (ti A . j .) is then a unitary n a -
J J
trix. The c o r r e c t n e s s in L 2 in t h e c a s e of variable coefficients w a s pro-
ved by F r i e d r i c h s l 7 ] . As f o r the c a s e of Wp, p / 2, we have the
following t h e o r e m by R r e n n e r [4 :
Theorem 1.2. Let 1 -
<p -< m, p # 2 , Assume that the system (1.7) has
constant coefficients. Then the corresponding initial-value problem i s cor-
rectly posed in W if and only if
P
This condition is the necessary and sufficient condition for the simul-
taneous diagonalizability of the matrices A j = 1,
j'
.. .
,d, and s o the
result means that apart from the case when the system can be brought
into the forin
with real diagonal matrices D . , the initial -value problem for (1.7) i s
J
not correctly posed in
P'
W p# 2 .
For the approximate solutioli of the initial-value problem (1.2) ,
(1.3) we' consider operators of the form
where h is a small positive number, and A, a r e explicit diffe-

1, h
rence operators
A. v(x) =
1,h
Z
p
a
1p
(x. h)v(x+/h) ,
where the summation is over a finite set of p = ( p ...,,jd),,$ J

integer,
and a.p(x, h) a r e NSN matrices which a r e polynomials in h with coefficients
1
in'& . If A1, = I, the operator i s explicit, in other c a s e s implicit.
Assume that (1.2) , (1.3) i s correctly posed in W

P
. Let k be a
small positive number tied to h by the relation k/hM = 1= constant.
The idea is then to choose the operator Eh SO a s to be able to appro-
ximate E(nk)v by E ~ V: Eh i s said to be consistent with E(k) if
there is a dense s e t of initial-values v in W such that
P
u(x, t) = E(t) v wm for t -> 0 and
P
(1.9) u(x, t t k ) = Ehu(x, t) + o(k) in W when
P
h+O .
The operator Eh is said to approximate E(k) with order of accu-
racy /L if o(k) in (1.9) can be replaced by kO(hF ) . In applications,
condition (1.9) turns out to be a purely algebraic condition on the coef-
ficients.
The operator E i s said to be stable in W if for any T > 0,
h P
We have the well-known Lax1 equivalence theorem [16] :
Theorem 1.3. Assume that the initial-value problem (1.2), (1.3) is correc-
tly posed in W and let Eh be consistent with E(k) Then the sta-
P
.
bility condition (1.10) is equivalent to convergence ; for any v g Wp ,
t-
> 0, and nay pair of sequences {hj);' . inj): with hJ . 4 , n.k.+t
J J
when j -t oo, one has
It should be noticed that stability is necessary for convergence only

if one demands (1.11) for all ; for individual
v 6W
P EWP One
can have convergence even without stability. Generally this depends on the
regularity of v; for an interesting example with analytic initial data and
highly unstable difference operator, see Dahlquist 151 . Although in
principle one may thus have convergence without stability, in practices
round-off e r r o r s then cause problems. We will return to these questions
in Section 3 .
Again the case most discussed in the literature is the case of stabi-
lity in L2 . F o r constant coefficients, using Fourier transforms one
finds that Eh i s stable in L if and only if for any T > 0 ,
2
where E ( I ) is the symbol (or amplification matrix) of E
h h'
A necessary condition for stability i s the von Neumann condition, (?€(A)

i s the spectral radius of A) ,
Setting
'1 2
lulH = (~u,u) . IAJ~ = s u p J ~ u l / u~ l H
/ .
one has the following analogue of Theorem 1.1 for difference operators:
Theorem 1.4. The condition (1.12) holds if there a r e positive constants
d
ho, C1, C2, and f o r each S C R and h < h
- 0 a hermitian matrix Hh ( )
such that
and
On the other hand, if '(1.12) holds wd have (1.14) and if 0-

< Y <1
there a r e positive constants ho, C1. C2 , and f o r each if R~ and
h h a hermitian matrix
-< 0
Hh({ ) satisfying (1.15) and
Proof. F o r y = 0 , see Kreiss [13] . The variant with Y > 0 i s due

to Widlund [26] .
Consider the case of a hermitian system (1.7) . F o r operators con-
sistent with such systems, Kreiss [15] defined the difference operator
E to be dissipative of o r d e r u ( v even) if (with the natural generaliza-
h
tion of (1.13) to variable coefficients) ,
V. Thomee
d
(1.16) p(Eh(x,{)) < l - C l l h l 1v +C2k. 5 R,h lj L n .
and has been able to prove :
Theorem 1.5. Let v be an even natural number. Assume that Eh is

consistent with the hermitian system (1.7) , has order of accuracy I, -1,
and is dissipative of order v . Then it i s stable in L
2 '
Certain results have been proved also in the case of accuracy of
order v -2 .
Consider now a hermitian system (1.7) with constant coefficients.
It follows from the proof of Theorem 1.2. that if p { 2, except for
the case that the system can be brought into the form (1.8) , there do
not exist difference operators consistent with (1.7) which a r e stable in
W
P
.
We therefore do not essentially restrict the generality by conside-
ring the scalar equation
a u =p-a
- p real .
a t a x ,
We then want to discuss the stability in W p f 2, of consistent ope-
P'
rators
with constant .
a . In the case of an implicit operator, the sum may be
J
infinite. Introducing the function
we have stability in L if and only i f a({) (1 for 5 6 Rd , For

p # 2 we need :
2 I I-
Theorem 1.6. Let 1 -< p -< co , p # 2. The operator Eh in (1.18) i s
stable in W if and only if one of the following Mo conditions i s
P
satisfied, natnely
ij t
a) a ( { ) = ce 1 Ic ( = 1;
b) la( { )I < 1 except for at most a finite number of points
9
.
q=l.. .., Q , i n 1 { 1 ~ n where a ( { ) = 1. F o r q = l . . . ..Q,
there a r e constants oq, pq. 5 where (Y is real, Re p > 0 , and
q 9
i s an even natural number such that
q
-
Proof, See [22] [23] . Here we only want to sketch the proof in the
case p = m and
(1.19) la({) 0, v even .

(Notice that (1.19) and (1.21) a r e exactly the assumptions of Theorem
1.5 in this particular case. ) By (1.20) and (1.21) we get by simple
estimates, the second of which after two integrations by parts ,
and hence we obtain easily

To prove the necessity, assume that a({) does not satisfy
(1.21) . Because of (1.19) we then have
(1.22) a ( < ) = e x p ( i a j + i ~ P q ( 5 ) - v J V ( 1 ~+ ( l ) ) ) , 5-0,

cu real, q({ ) polynomial , q(0) f 0, 1 O .
It i s then easy t o prove
and using a lemma by van d e r Corput ,
max 1 anjl 5 C2n- 1/P
j
and s o
1 1
anj
2
C1
-1 - -1
: .1 lL = fJanjJ
J
max
j
a
I njl
>F
2
n
P JJ
which contradicts the stibility since p <u .

As an application, consider for the solution of the equation (1.17)
the Lax-Wendroff [17] operator,
Ehv(x) = z1 ( P2 I 2 + p l )v(x t h) t (1 - P 2 .I2) v (a) t

ti ( P 2 .I2 - p I ) v ( x - h)
with
a ( { ) = p 2 I 2 cos I + iplsin j t l - P2 A 2 .
This can be described a s the most accurate explicity operator for (1.17)
based on the thlcee poilits x+h and'x. Here
and s o E i s stable in L < 1. On the other

if and only i f J p l .I -
h 2
hand, if 0 < l p l . A < l v e have a(/) C l f o r 0 < 1 1 1-
< n and
1 4
a ( ( ) = e x p ( ~ l 'i6 { - - ~ ~ ( l - p ~ t ~o (~[ ) )i ) l, ~/ + o ,
It follows from Theorem 1.6 that E
h
is unstable in W
P
for p # 2.
By the above proof we have
1112
Serdjukova [19] , PO] and Hedstrom 183 , [9] have, by using

more refined techniques of estimating the a . above, been able to gi-
n.l
ve more precise estimates of the growth of JIE& for the case when
Eh is stable in
2
L
but unstable.in the other W
P
.
In the particular
case of the Lax-Wendroff scheme, the exact result is
more generally, when a ( 5 ) has the form (1.22) one has
The instability present here i s of course quite weak; we shall return to

its influence on convergence in Section 3.
The proof of the sufficiency part of Theorem 1.6. for p = u3 is
due to John [ll] and Strang [21] . The proof of the necessity part
I
in
r 2 3 J for p < oa uses results about Fourier multipliers on L ; the
P
sufficiency part i s a trivial consequence of the result for p = oa .
Theorem 1.6 has also been generalized to variable coefficients in 1243
,. P a r a b o l i c d i f f e r e n c e o p e r a t o r s .
................................
Consider a s in Section 1 an initial-value problem
(2.2) u(x, 0) * v(x) ,
with coefficients in 4 . With the notation of Section 1, we say that

the sjrstem (2.1) is parabolic in Petrowskyts sense if
In this case we have correctness in $ ; one can actually even estimate
the derivatives of the solution :

Theorem 2.1. Let (2.1) be parabolic in Petrowskyts sense. Then the
initial value problem (2. l ) , (2.2) is correctly posed, and for any differen-
tial operator Q of order q and any T >0 there i s a positive
constant C such that
(2.4) -
O < t < T.
Proof. See e.g. Friedman [6] .

For difference operators E of the form discussed in Section 1,
h
Widlund [26] has defined a concept of parabolicity which can be conside-
red a s a discrete counterpart of (2.3) , namely ,
(Compare the definition (1.16) of a dissipative operator Eh). It can then

be proved :
Theorem 2.2. Assume that Eh is consistent with (2.1) and satisfies (2.5),
and that Q i s a difference o p e r a t o r consistent with a differential ope-
h
r a t o r of o r d e r q. Then f o r any T > 0 t h e r e is a positive constant C
s u c h that
Proof. See [27] . The proof goes back to John [li] and Aronson [2] ,
[3] , and depends on e s t i m a t e s of a d i s c r e t e fundamental solution.Actual-
l y the r e s u l t is .valid in much m o r e general situations, and p e r m i t s

multi-step s c h e m e s , variations in the coefficients a l s o in t, and low r e -
gularity.
It might b e considered natural to make inequalities like ( 2 . 4 ) and
( 2 . 6 ) definitions of parabolicity. In o r d e r to investigate where such defi-
nitions lead we shall r e s t r i c t o u r s e l v e s in the r e s t of this section to the
c a s e of constant coefficients in P and Eh , and we shall consider only
s o that F o u r i e r - t r a n s f o r m s can b e conveniently applied. We then s a y
L2
that the s y s t e m ( 2 . 1 ) with constant P, (x) = Pa is parabolic if the initial
value problem ( 2 . 1 ) , ( 2 . 2 ) is c o r r e c t l y posed in and if f o r ally
L2
differential o p e r a t o r Q and any positive T , T ,
Similarly, we s a y that the o p e r a t o r Eh with coefficients independent of

x i s parabolic in L2 if it i s stable in L 2 , and if for any difference
operator Qh consistent with a differential o p e r a t o r Q ,
We have the following analogue of T h e o r e m 1 . 3 :
Theorem 2 . 3 . Assume that Eh i s consistent with the parabolic equation

( 2 . 1 ) (with constant coefficients) . Then Eh
i s parabolic if and only if for
any difference o p e r a t o r Qh consistent with a differential o p e r a t o r Q
of order q, any v c L, , any t -> 0 which i s > 0 if q > 0 , and any
p a i r of sequences { h.)'J 1 ,{nj) f , with h --r 0 , n.k,+ t when j + CO,
j J J
we have
n.
11% Eh
v - Q ~ ( t ) 1v --r 0, when j +m .
j j
Proof . See E5] .
We now want to characterize algebraically systems (2.1) and operators
Eh which a r e parabolic in the present sense. By Parsevalts relation
conditions (2.7) and (2.8) a r e equivalent to
d
(2.9) sup{l Q ( l ) e x p (t P ( I ) ) ~ ; 0 < T -
< t-
< T,lcR ] < w 3
and
(2.10) s u p ( I Q ~ ( I ) E ~ (;~ O) <~ T-< n k -
< T , I € R ~ } <m,
respectively. Here E ( 5 ) is the symbol of the difference operator E h and

h
similarly for Qh( 5 ) .
Consider first the case of a system. We then have :
Theorem 2.4. Assume that the initial-value problem (2.1) , (2.2) i s correc-
tly posed in L 2 . Then it i s parabolic in L
2
if and only if there a r e po-
sitive constants C such that
1' C2'
-
Proof. F o r details see 25 . We give a short sketch . The suSf'~c~~c~~ir.).
of
the conditions follows easily from the inequality
N- 1
exp (t A) -< exp(t A (A)) (2t A )' ,
j=O
which holds for any NxN matrix A. To prove the necessity of conditiori
(2.11) , we notice that by the parabolicity condition (2.9) , we have
and s o with
-
A(r) = max A (P(t) ) ,
t=r
(2.13)
-
A(r) -< log C - log (1 t r) -+ -03 when r -. 03
-
Using the Seidenberg-Tarski elimination theorem one can prove that A(r)
is algebraic in r for large r and thus by (2.13), there is a (rational)
positive number p and a positive C such that
- -
( r )= 2
1
(1 t 0 1 ) when r - m .
This implies (2.11) .
In general it i s necessary in the formulation of Theorem 2.4 to ex-
plicitly assume the correctness of the initial-value problem. There are,
however, some cases when the correctness follows from (2.11) . One such
case is when P ( [ ) is a normal matrix, in particular if P ( ( ) is scalar,,
An example of this is offered by the equation
where
2
Re P ( F ) = R e ( i t ) + ( i t )
3
= -5
2
.
One other case when correctness i s automatic is when (2.8) holds with
= M ; this i s the case of parabolicity in Petrowsky's sense. An
example where (2.11) is satisfied without correctness is given by (N=2)
where
Systems which satisfy (2.11) a r e called parabolic in Silovts sense;

the present parabolicity concept is thus more restrictive than Silovts.
V, Thomke
Using Theorem 1.1 we can give a characterization of parabolic

equations which contain at the same time the correctness and the condi-
tion (2.11):
Theorem 2.5. The initial-value problem (2.1) , (2.2) is parabolic if and
only if there a r e positive constants C , C 2, C 3, p and for each r e a l
1
a hermitian matrix H( 1) such that
and
(2.14) Re (H(I)p({) ) ( (-c21~I"+ C 3 ) I -
Our aim is now to similarly characterize parabolic difference opera-
tors. We have :
Theorem 2.6. Assume that the operator Eh is consistent with the equa-
tion (2. l ) and stable in
2
.
Then E is parabolic in L if and only
L
h 2
if there a r e positive constants C1, C2, ho , v such that
Proof. See [25]

p
. The proof is similar to that of Theorem 2.4 but mo-
r e complicated . The inequality
( P = P (A)) , which is analogous to (2.12), is used to prove the sufficien-

cy of (2.15) for (2.10) . To prove the necessity one uses again the
Seidenberg-Tarski theorem, but this time trigonometric polynomials take
the place of ordinary polynomials.
Also in this case the stability has to be explicitly assumed; it is
easy to give examples where (2.15) is satisfied by a n unstable operator
E
h
.
Again, the case of a normal matrix Eh({) , in particular a scalar
Eh({) , and the case v = M in (2.15) a r e exceptions ;-in this latter case E,
is parabolic in the s e n s e of Widlund.
Using T h e o r e m 1.4 one c a n prove the following counterpart of
T h e o r e m 2.5 :
Theorem 2 . 7 . The o p e r a t o r E is parabolic in L if and only if t h e r e
h 2
a r e positive constants C1, C2, C3, ho, D and f o r each h < h and
2 - 0
5 E R~ a positive definite m a t r i x H
h (5 ) with
and
The l a r g e s t possible p and u in (2.11) (or (2.14)) and (2.15)

(or (2.16)) , respectively, a r e r e f e r r e d to a s the o r d e r s of parabolicity.
It can be proved that
i) if (2.1) is parabolic of o r d e r p , then t h e r e e x i s t s a n o p e r a t o r
Eh
consistent with (2.1) and parabolic of o r d e r p ;
ii) if Eh is consistent with (2.1) , and (2.11) and (2.16) hold, then
u i p ;
iii) if (2.1) is parabolic of o r d e r p , then t h e r e exist o p e r a t o r s
Eh
consistent with (2.1) which a r e unstable, o t h e r s which a r e stable but
not parabolic, and s t i l l o t h e r s which a r e parabolic, but of o r d e r u <p .
F o r details, s e e [25] .
3. T h e r a t e o f c o n v e r g e n c e .
..........................
Consider again an initial-value problem
In the sequel we shall demand not only that the initial-value problem
be correctly posed in W but that it satisfies the stronger require-
P'
ment of the following definition. We say that the initial-value problem is
strongly correctly posed in W if for any m > O . , v E wm implies
P P
E(t) v € wm and there is a constant C such that for all v c w m ,
P m, T P
In particular, this definition implies that E(t). W C wW

P ' P
w
.
It can be proved that if P = P(x, D) has constant coefficients, o r if
it i s of first order, then strong correctness in
W i s an automatic
P
consequence of correctness. Further , systems which a r e parabolic in
Petrowsky's sense a r e strongly correctly posed in W for any p with
P
-
l l P < W .
We shall also introduce a more general parabolicity concept than in
Section 2. We shall say that the system (3.1) i s strongly parabolic of o r -
der b in W if the initial-value problem (3.1) ;(3.2) i s correctly
P
posed in W if v € W implies Da E(t)v 6 W for all a when
P' . P P
t > 0 , and if
One can prove that the o r d e r b of strong parabolicity i s at most

equal to the o r d e r M of the system. Systems which a r e parabolic in
Petrowsky's sense a r e strongly parabolic of o r d e r M in W for any
P
p with -
1< p-
< oo , and systems which a r e parabolic of o r d e r b in
L2 in the sense of Section 2 a r e also strongly parabolic in L2 of
order b .
Consider difference operators of the s a m e form a s previously, na-
mely
F o r the sake of simplicity we shall assume here that i s explicit, s o

Eh
that the summation i s over a finite number of t e r m s only. We have pre-
viously defined consistency of E with E(k) to mean that for any
h
sufficiently smooth solution u(x, t) of (3.1) ,
more precisely, Eh i s said to approximate E(k) with ortler of accu-

racy y if for any such u,
(3.4) u(x, t + k) = E h u(x, t) + kO(hP ), h +0 .

When (3. I), (3.2) i s strongly correctly posed in W , it i s sufficient
P
to assume this condition locally to obtain the following global estimate :
Theorem 3.1. Assume that the initial-value problem (3. I), (3.2) i s stron-
gly correctly posed in W and that Eh approximates E(k) with
P
order of accuracy . Then there exists a constant C such that for
M+P
any v€ W ,
P
Proof. See
- 1181 . The proof consists in expanding E(k)v = u(x, k) and
F v in Taylor s e r i e s around the point (x, 0) , using (3.4), and estimating
v
the remainder t e r m s in integral form. In doing so, it is sufficient to
consider v in the dense subset WW= of
P
1:
.
We now easily obtain the following estimate for the r a t e of conver-
gence :
Theorem 3.2. : Assume that the initial-value problem (3.1) , (3.2) i s
strongly correctly posed inW and that E in stable in W and ap-
P h P
proximates E(k) with order of accuracy y .
Then there i s a constant
M+
C = CT such that for v W , nk -
<T ,
P
Proof. We
- have
n- 1
- E(nk)) v = $-I-' (Eh - E(k)) E (jk) v ,
j=O
and s o by the stability of Eh , Theorem 3.1, and the strong correctness,
which proves the theorem. In special cases, this theorem appears in many
places.
Thus, the situation i s that for initial-values in
W we have (by
P
Lax' equivalence theorem) convergence without any added information on
its rate, and if the initial-values a r e known to be in wMtPWe can
P
conclude that the rete of convergence i s O(hy ) when .
h+O It i s natural
to ask what one can say if the initial data belong to a space "intermedia-
ten to W and wMtq To answer this question we shall introduce some
P P
spaces of functions which a r e interpolation spaces between W- and
P
wm in the sense of the theory of interpolation of Banach spaces (cf
P
.
1;8] and references) .
Let s be a positive r e a l number and write s = S + a , S integer,
0 < 0 -< 1. Set T, u(x) = u(x+ T ) . We then denote by B~ the space
P
of u E W such that the following norm i s finite, namely
P
Thus BS i s defined by Lipschitz type condition for the derivatives of
w
o r d e r S ; these spaces a r e sometimes called Lipschitz spaces. F o r the
Heavyside function
o , x < 0 (d = 11,
(3.6) 1, X L 0,
we have for 1-
 0 arbitrary ,
The main property of these spaces that we will need is then the following
interpolation property; assume that 1-
<p -
< ca , m is a natural number,
and s is a r e a l number with 0<s<m . Then there i s a constant
C such that any bounded 1i n e a r operator A in W with
P
rr < P '
whe have
Theorem 3.2 and (3.8) with A=E
n
h
- E(nk) proves immediately
the following result :
Theorem 3.3. Assume that the initial-value problem (3.1) , (3.2) i s stron-
gly correctly posed in and that E i s stable in W and approxima-
W
P h P
tes E(k) with order of accuracy y .
Then for 0 < s < M + P there i s
S
a constant C=C such that for any v E B , nk -< T ,
s , "I' P
Notice that Y = P ( M t i u )
- 1
grows with y and lim V = '1 .
This means that the estimate (3.9) becomes increasingly bette? for fixed
s when y grows . In other words, if for a given strongly correctly po-
sed initial-value problem one can construct stable difference schemes of a r -
bitrarily high order of accuracy, then given any s > 0 one can obtain
rates of convergence arbitrarily close to O(hS) when h -,0 for all
initial-values in BS
P
.
As an application, consider L -stable operator E with order
an
2 h
of accuracy ,U for the hyperbolic equation
m
and let v= (P X where cp F CO and X is the Heavyside function
(3.6) . By above we have in this case
F o r dissipative operators Eh , stronger results have been obtained i n

Apelkrans [I] , where also the spreading of discoundinuities i s
discussed.
It is natural to ask if for a parabolic system the smoothing property
of the solution operator can be used to reduce the regularity demends on
the initial data in Theorems 3.2 and 3.3. This is indeed the case; the
result on the r a t e of convergence i s then the following.
Theorem 3.4. Assume that the system (3.1) i s strongly parabolic of
order b W and that E
in is stable in W and approximates E(k)
P h P
with order of accuracy p .
Then for any s > 0 , T > 0 , there is
a constant C=C
ST
such that for -
v ~ ~ ; , n <kT ,
Proof. For
- details, see [18] . Here we will only sketch a proof for the
case v 6 B' where M + p b < s < M t P
P
- .
When b = M the other
cases can be treated similarly ; if b < M the proof in [18 uses slightly 1
more sofisticated interpolation theory.
We shall use (3.5) . For j = 0 we have by the stability and Theorem
and s o by (3.7) and (3.8) , since' s >p ,
For j >0 we have by Theorem 3; 1 and the strong parabolicity ,

and hence by (3.8) ,
where
Adding over j we get
which proves (3.10) in the case considered.

For an earlier particular result in the same direction, see Junco-
s a and Young [12 ] .
We shall complete this section with a simple case of a recent
result by Hedstrom [lo] . Consider a s in Section 1 the initial-value
problem for the equation
a u ,,&
- P real
at ax '
and a consistent difference operator of the form
Ehv(x) = a .v(x i-
J
jh) .
j
Set again
and assume that we have the case that Eh i s stable in L but unstable
2
in W
P'
p # 2; in particular assume that
ct. real, q ( J ) real polynomial, q(0) / 0, 1 <,.U + 1< v ,
V even, Re y > 0 .
(In Section 1 we used ,u instead of p t 1 ; here ,U i s a s above the
order of accuracy.) We then have :
Theorem 3.5. Under the above assumptions, for any t > 0 and s with
S
0 <s -< p +1 there i s a constant C such that for v 6 Bco , nk = t ,
S-
P
-
ltP P t l
, 2
<s< p t 1 ,
-
-P2
h log-l. S =- P t 1
2 '
B
00
2s( v -I)-( v - p -11
In particular this means that for v 6 B:

where P + l < . S < P + l ,
2
- -
the estimate for the r a t e of convergence i s the same a s the one we would
have obtained by Theorem 3 . 3 , in the stable case F o r 0 < s - < P 2f l ,
. -
however, the rate of convergence becomes slower. In the extreme case
s = 0 , we recognize the growth in (1.23) .
REFERENCES
------------
1. Apelkrans , On difference schemes for hyperbolic equations
with discountinuous initial values. To appear.
2, D. G. Aronson, The stability of finite difference approximations to
second order linear parabolic differential equations. Duke Math. .J.
30 (1963), 117-128.
3. D. G. Aronsons, On the stability of certain finite difference approxima-
tions to parabolic systems of differential equations. Numer. Math.
5(1963), 118-137.
4. Ph. Brenner, The Cauchy problem for symmetric hyperbolic systems
in L Math. Scand. 19 (1966), 27-37.
P'
5. G. Dahlquist, Convergence and stability for a hyperbolic difference e-
quation with analytic initial-values. Math. Scand. 2(1954), 91-102.
6. A. Friedman, Generalized functions and partial differential equations.
Prentice-Hall, Englewood Cliffs, New Jersey, 1963.
7. K.O. Friedrichs, Symmetric hyperbolic linear differential equations.
Comm. Pure Appl. Math. 7(1954), 345-392.
8. G. W. Hedstrom, The near-stability of the Lax-Wendroff method.
Numer. Math. 7(1965), 73-77 .
9. G. W. Hedstrom, Norms of powers of absolutely convergent Fourier
series. Michigan Math. J. 13(1966), 393-416.
10. G. W. Hedstrom, The rate of convergence of some difference schemes.
To appear.
11. F. John, On integration of parabolic equations by difference methods.
Comm. P u r e Appl. Math. 5(1962), 155-211.
12. M. L. Juncosa and D. M. Young, On the order of convergence of so-
lutions of a difference equation to a solution of the diffusinn equation.
J. Soc. Indust. Appl. Math. 1(1953), 111-135.
(1
13. H. 0. Kreiss, Uber die Stabilitatsdefinition fiir D i f f e r e n ~ e g l e i c h u n ~ e n

die partielle Differentialgleichungen approximieren. Nordisk Tidske.
Information-Behandling 2(1962), 153- 181.
(1
14, H.O. Kreiss, Uber sachgem'a'sse Cauchyprobleme. Math. Scand. 13
(1963), 109-128.
15, H.O. Kreiss, On difference approximations of dissipative type for
hyperbolic differential equations. Comm. Pure Appl. Math. 17(1964),
335-353.
16. P.D. Lax and R.D. Richtmyer, Survey of the stability of linear finite
difference equations. Comm. Pure Appl. Math. 9(1956) , 267-293.
P. D. Lax and B. Wendroff, Systems of conservation laws. Comm. Pu-
r e Appl. Math. 13(1960), 217-237.
J.Peetre and V. Thom&e, On the rate of convergence for discrete
initial-value problems. To appear.
S.I. Serdjukova, A study of stability of explicit schemes with constant
real coefficients. Z.Vycis1. Mat. i Mat. Fiz. 3(1963), 365-370.
S.I. Serdjukova, On the stability in C of linear difference schemes
with constant real coefficients. Z. Vycisl. Mat. i Mat. Fiz. 6 (1966);
477-486.
W. G. Strang, Polynomial approximation of Bernstein type. Trans Amer.
Math. Soc. 105(1962), 525-535.
V. ThomCe, Stability if difference schemes in the maximum-norm.
J , Differential Equations 1(1965), 273-292.
V. ThomCe, Stability of difference schemes in L'. XIV Congr. Math.
Scand. 1964. To appear.
V. ThomCe, On maximum-norm stable difference operators, Numeri-
cal Solution of partial differential Equations (Proc. Sympos. Univ.
Maryland, 1965), pp. 125-151. Academic P r e s s . New York.
V. ThomCe, Parabolic difference operators. Math. Scand. 19(1966),
77-107.
0. B. Widlund, On the stability of parabolic difference schemes, Math.
Compt 19(1965), 1-B .
0. B. Widlund, Stability of parabolic difference schemes in the maxi-
mum-norm. Numer. Math. 8 (1966), 186-202.

Numerical Analysis of Partial Differential Equations (PDFDrive)

Caricato da

Informazioni sul documento

Titolo originale

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

Numerical Analysis of Partial Differential Equations (PDFDrive)

Caricato da

Copyright:

Formati disponibili

Jacques Louis Lions ( E d.

Numerical Analysis of Partial Differential Equations

Lectures given at a Summer School of the

ISBN 978-3-642-11056-6 e-ISBN: 978-3-642-11057-3

©Springer-Verlag Berlin Heidelberg 2010

Printed on acid-free paper

I' ALCUNI METODI D I CALCOLO NELLA TEORIA DELLA D I F F U S I O H C D E I

Corso tenuto ad Ispra d a l 3-11 Luglio 1967

valide per x c y P s , essendo d l a d e r i v a t a conormale a as2 e

Osservazione I - Dal punto d i v i s t a f i s i c o l e (1) forniscono l a

1) Positivita dei coefficienti D ~ A, ~ ,gi, Ci (supposti misura-

2) Ellitticit; uniforme e cio8: 4 {A ; p=1,2.. .n, A 6 Rl,x c T )

per ogni i=1,2...g. Inoltre a sari abbastanza regolare da

§ 2 - Soluzione del problema A

siano A , B, C i corrispondenti operatori (matriciali) penerati

Introdotti A, 3 , C il Problema A si riconduce alla.risoluzione

I1 procedimento esistenziale k allora il seguente:

(6) Aou=(A-C)u=f 5 costituita (si osservi la struttura delle

Assumendo fe(f1-I)g (o pin senplicemente ( L Z ) ~ )per le ipotesi

nerale si ha: ui= C io ( C iGo

Ne consegue llesistenza (u=o) di un operatore matriciale di

Essendo Go un operatore limitato da ( L ~ ) ~ +(H:)~, Bij moltipli-

5 3 - Soluzione del Problema I3

a) In tal caso si ha:

(11) A'G'(X,~) = 6 (6x,y misara di Dirac in (xiy))

b) - Determinazione iterativa di A,.

Richiamiamo il Teorema di I. Marek, (6)

e l e (131, come a 1 s o l i t o disaccoppiandosi, permettono d i t r o v a r e

A ciascuna i t e r a z i o n e s i devono p e r c i s r i s o l v . r e problemi d i

essendo R,S,Q m a t r i c i NXN d i elementi n o t i .

L'inversione d e l l e m a t r i c i R 2 s t a t a f a t t a con l'algoritmo di

b) Confronti t r a il metodo iterative variazionale d e s c r i t t o e

F sono l e autosoluzioni d e l l ' o p e r a t o r e d i

Caso tridimensionale: E - 2 , r=3; D , A , B , C c o s t a n t i a p e z z i , simmetria

Osservazione - I1 netodo i t e r a t i v o variazionale f o r n i s c e l'autovalore

4) la condizione iniziale di Cauchy per tutte le incognite:

I1 sistema stazionario associato ad (11 8 ellittioo second0

da cui si vede subito che sulla diagonale principale l'ordine degli

essendo ( = vettore di R 2 H + N l ove !J 5 il numero dei punti del re-

b) Metodo Implicito: 4 ( t ) = ( I - ~ t~ ) " + ( t - A t ) ,

Ad ogni passo temperale c ' 6 da risolvere un sistema del tipo:

Se A t w o c l a l l o r a , i n base a l l a p r o p r i e d 5 del 5 2 , i metodi

In ogni caso per6 c t & il problema d i determinare w o che non

c i s i r i d u c e , come equazione a g l i a u t o v a l o r i per w (essendo w o

Lo schema alternato per il passaggio da tn+ tntlltr[tn,tntl[

fornente u(tn+,) = untl (tntl). La discretizzazione della ( 8 ) ci

(1) S.Albertoni: "Metodi v a r i a z i o n a l i per c e r t i s i s t e m i d i equa-

(2) S .Albertoni - M.Lunel1i -

(3) A.Daneri - A.Daneri - 1 , G a l l i ~ a n i : "A numerical approach t o

(4) 1,Galligani: "Numerical s o l u t i o n s of t h e time dependent d i f -

(6) 1,Marek: " I t e r a t i o n s of l i n e a r unbounded o p e r a t o r i n non

(7) G.Birkoff - R,S.Varna: "Reactor C r i t i c a l i t y and non negative

(8) G. I. Habetler - M.A. Martino: "The multigroup d i f f u s i o n

(9) G.Stampacchia: "Su un problema r e l a t i v o a l l e equazioni d i ti-

"PROBLEMS OF OPTIMIZATION AND NUMERICAL STABILITY'

Corso tenuto ad Ispra dal 3-11 Luglio 1967

Computer Science is a new scientific discipline. An important part

In this paper some results obtained recently in Prague will be given

In this section I shall show some aspects of ideas, which I men-

1. The function f(x) i s a continuous periodic function with

This formula i s the well known trapezoid formula.

(2.4) w (n, B) = inf sup (2 a j f (yj) - J(f) 1

w(n, B) is the minimal possible e r r o r under the assumption that we know