General Lagrange and Hermite Interpolation in R" with lpplications to Finite Element Methods

P. G. CIARLET& P. A. RAVIART
Communicated by J.L. LIONS
1. Introduction and Preliminaries

Recently there has been considerable interest in approximation theory in several variables, one of the main reasons for this being that it can be applied to obtain error estimates for various finite element methods of numerical approximation (we have given some references to these methods and their use in the bibliography, but the list is far from exhaustive). Basically, there are two different types of approach. The first one consists in prescribing a set X of points in R" and then interpolating the values of a given function u at the points of 2; by a polynomial of a given degree. This is Lagrange interpolation. If, in addition, the values of some derivatives of the function u are also to be interpolated at some points of ~, then we have an Hermite interpolation problem. Generally, the error bounds are given in term of the sup-norm of some high-order derivatives of the interpolated function; see, for example, the papers of STRANG [29], STRANG & FIX [30], ZLAMAL [35, 36], ~ENi~EK [34], THACHER [32], THACHER & MILNE [33], HALL [20], NICOLAIDES[22, 23, 24], FREDERICKSON[17], SALZER [27], GUENTHER& ROETMAN [18]. The work of COATMI~LEC[14] is also particularly important and relevant in this respect. In most cases, a special role is played by the canonical Cartesian coordinates, and indeed, this limits in some ways the obtainable results as well as the practical examples which can be handled in this way. However, a more powerful coordinatefree approach can be used, such as in OARLET & WAGSCHAL [13], and in some parts of COATMI~LEC[14]. ~ 2 and 3 of the present paper depend upon the same idea, and their results can be considered as a generalization of [13], which was itself a generalization of [35]. Our main result concerns the following situation: given a finite set ~ c R" and given a function u, the interpolation problem consists in finding a polynomial fi of degree < k , in such a way that the values of u and fi and of some corresponding derivatives of u and fi are equal at some specified points of X. Under the sole assumption that this problem has a unique solution, we prove the error estimate (cf. Theorems 2 and 4) (1.1)

sup(llD'~u(x)-Dmfi(x)ll; xeK}<CMk+l ~ pm

hk+ l

for all O<m<_k

12 Arch. Rational Mech. Anal., Vol. 46

178

P.G. CIARLET,~ P. A. RAVLART:

where K = closed convex hull of 27, Mk+ 1= sup {liD~+1 u(x)ll ; xeK}, h = diameter of K, p = supremum of the diameters of the inscribed spheres of K, C = numerical constant independent of 27 (in a sense to be made more precise in the sequel). The inequality (1.1) is obtained from generalized Taylor expansions (el. Theorems 1 and 3), which contain as special cases those given in [13]. The second approach, taken up in w is more abstract, and perhaps more elegant from a mathematical standpoint. The approximation scheme is defined as follows: given a bounded open subset f2 of R n, we associate with any function u in the Sobolev space W k+ 1,p(f2) a unique "approximation" llu. With the sole assumption t h a t / / l e a v e s invariant all polynomials of degree < k (apart from this condition, 1-lu can be any "reasonable" function), we obtain (cf. Theorem 5) (1.2) where
U I~+l,p,s~= flok+l u IILp(~),

Ilu-l"lullw,,.,,r

hk+ 1 P"

forall 0 < m < k + l

h = diameter of f2, p = supremum of the diameters of the inscribed spheres of f2, C = numerical constant independent of t2; this general result is then applied to the Lagrange and Hermite interpolations described in w167 and 3 (el. Theorem 6). This type of approximation is not necessarily defined as a Lagrange or Hermite interpolation problem, but of course it contains such problems as special cases. Our approach is not new, and in some instances, is very similar to that of BRAMBLE [7], BRAMBLE &~HILBERT [8, 9], BRAMBLE t~r ZL.&MAL [10], for example. However our results seem to have a fairly new and more concise form, which generalizes in one way or another these previous results, whether it be that we consider Sobolev spaces Wm'P(O) for p~=2, or that the dimension n is arbitrary, or else that we can handle more practical examples, once the coordinate-free approach developed in ~ 2 and 3 is used. For this second approach we note the following relevant papers: AUmN [1, 2], BABU~KA [3, 4], DI GUGLIELMO [19], FIX & STRANG [16], BIRKHOFF, SCHULTZ & VARGA [6], NITSCHE [25, 26], SCHULTZ [28]. Several examples are examined in the text, each stemming from some particular finite element method. It is in this sense that the "applications to finite element methods" mentioned in the title are to be understood. The associated error bounds obtained in the solution of elliptic problems are not explicitly mentioned, since their derivation is almost immediate. Basically, our examples involve "elements" which are n-simplices of R n, because it is perhaps for those types of elements that our results appear to be the newest. We have not considered the multivariate

Lagrange and Hermite Interpolation

179

Hermite approximations of BmrmOFF, SCHULTZ VARGA [6] and SCHULTZ [28], but our results could also be applied to these cases. The geometrical dependence of the error estimates upon the geometry of the domain is exhibited by the two parameters h and p of (1.1) or (1.2). Thus, the "smallest angle 0 of the triangle" (cf. Remark 3) used by ZL~MAL [35, 36], STRg~qG [29] and others in the case where n = 2 as a measure of the "badness" of the approximation for derivatives, is here replaced by the ratio p/h, which is its natural generalization to the case where n is arbitrary. In practice, one often thinks of a family ( ~ ) ~ of sets, with associated parameters h~ and p~. As in [13, condition (6.18)], it is then natural to define a regular family as a family for which one has (1.3)

hi<~pi

for all i~I,

for some constant ~>0. For such regular families the error bounds of (1.1) can be converted immediately into the following form: (1.4)

sup(llDmu(x)-DmF~(x)ll; ~ K } = O ( h k+l-m) x

for all O < m < k

(and similarly for (1.2)), and this is generally the form which is found in the literature. As in [13], heavy use will be made of differential calculus with Fr6chet derivatives, for which we refer to [11] or [15]. We now briefly explain some of oursubsequent notations. Given two vector spaces over the same scalar field, we denote by Lek(X; Y) the space of k-linear mappings Ak: xk--* Y, where X k is the Cartesian product of X, k times with itself (we recall that a mapping is k-linear if and only if it is linear with respect to any one of the k variables when the other ( k - 1 ) variables are kept fixed). When k = 1, we shall write ~ (X; Y) in lieu of L,e (X; Y), 1 and ~ (X) if Y= X. Given a function u: R" ~ R, its k-th derivative D k u(a) at a point a of R" (whenever it is defined) is an element of .~k(Rn; R) which is symmetric in the sense that Dk u (a) . (~1, ~2..... ~k) =Ok u (a) . (~1, r ..... 40,) for any permutation a: i ~ a ~ of the set {1, 2 . . . . . k}. In case r we shall simply write Dk u(a) 9(4, 4 .... , r u(a) 9(4)k, and likewise, if 4i = r for all 2 < i ~ k, for all 1 < i < k ,

Ok u(a) 9(41, ~..... ~)=Dk u(a) " (~1, (~)k-a).

The norm of D k u(a) is defined as IIDku (a)II = sup {ID~u (a). (~1, 42
.... '

~91; 1 ~ II--<1, 1 < i < k} 1

where I1" II denotes the usual Euclidean norm over R". If R n is equipped with its canonical basis (e~, ez . . . . . e,), then the usual partial derivatives are given by ~u

(a)=Du(a). (e~),
12"

a2u ax~ax--------f. (a)=D2u(a)" (ei, ej),

etc .....

180

P.G. CIARLET~: P. A. RAVIART:

so that as a consequence of the definition of

IIDk u(a)l[,

we have

[ ~-~ (a)[ <__llOu(a)ll<C1(n)max {l~x (a)l ;

]X'-ff'~-~ ~

l <i<n},
....

[<l'D2u(a)ll<C2(n)max{~(a)l;l<i'j<n}'etc

for some constants C1 (n), C2 (n), etc . . . . . which depend only upon n; this last observation shows that error estimates such as those of (1.1), (2.15), and (3.15) can immediately be converted into similar estimates involving the usual partial derivatives. 2. Lagrange Interpolation Given an integer k=> 1, we let -< k defined over R ~, and we let

Pk denote

the space of polynomials of degree

N=N(k)=dimP k.
We shall say that N distinct points ai of R ~ form a

k-unisolvent set

provided that, given any real numbers gi, 1 <-i< N, there exists one and only one polynomial P6Pk such that

p(ai)=gt, polating polynomial if it

l<_i<N.

Given a function u defined on a k-unisolvent set 27, we say that fi is its interis the unique polynomial of degree < k with the property that

?~ = u (al), (ai)

1 < i < N.

Example1. Let K be a nondegenerate n-simplex of R ~ with vertices a l, 1 _<_i< n + 1. We denote by aij = aj~ the mid-point of the edge joining the vertex a~ to the vertex aj, i*-j. Then the set
2 7 1 = f a "ln+l I. iJi= 1

is 1-unisolvent, and the set X =(a,},= 1 u{a,j}l~_e<j~,+l is 2-unisolvent (cf. [13]). If ~ is the interpolating polynomial of a given function u on 27i (resp. 2711), We shall say that ~ is an approximation of type I (resp. of type II) of the function u. Now let 27= {al}~= l be a set of points of R n. We say that a subset K c R n is Y-admissible if and only if whenever it contains a point x, it also contains the closed segments joining the points x and ai, for all 1 _<i<N (in other words, K is star-shaped with respect to any point of 27). Given a function u: K~R, we say
11 n+ 1

Lagrange and Hermite Interpolation that u belongs to the class ~k+ t (K) if and only if the Taylor formula

181

u(ai) = u ( x ) + D u ( x ) . ( a t - x ) + "'" + ~ . Dku(x) 9(a~-x) k

(2.1)
1 -~ (k + 1)~ Dk+ ' u (rh(x)) 9( a i - x) k+1

holds for all points x ~ K and all 1 < i<N, where (2.2)

tli(x)=Oix+(1-Oi)ai

for some 0 < 0 1 < l .

In particular, this will be the case if K is the closed convex hull of F,, ue C k (K), and D k+ 1 u(x) exists for all points x in K. Theorem 1. Let k be a fixed integer > 1, and let there be given a k-unisolvent set Z = {at}~= 1 of points of R". Let K ~ R" be a Z-admissible set, and let u ~ "~+1 (K) be given. Then, for any point x e K and any integer m with O<m<_k, one has (2.3)

Ofi(x)=Omu(x)+

S (Dk+Xu(rh(X)). (ai--x)k+l}ompdx) (k + 1)! , ~ , ' TM

where the pi's are the unique polynomials of degree -<k such that
(2.4)

Pi (a j) = Jij,

1 ~ i, j <= N,
N

and rh(X) is of the form (2.2).

Proof. From the definition of the polynomials Pi it follows that ~ = ~, u(ai)Pt,
N t=1

and thus D m ~ = ~, u (ai) D~ P, for any m __>1. Let x be a point of K. Since U ~ ' k + t (K),
i=l

we can replace u(a3 by (2.5) in the above expressions. In this fashion, we obtain for any integer m with 0_< m _ k,

D'?t(x)= ~=ol ~= { D ' u ( x ) ' ( a , - x ) ' } D'p,(x)

(2.5) 1 N 4 (k + 1)~ ,=~,{ok +' u (~li(X))" ( a s - x) *+1 } O,~pi(x).

We shall now use the fact that ff=u whenever ueP k. This will yieldthe (somehow unexpected) result that in the last expression,

(2.6) 9i=1

{Dtu(x) .(ai-x)l}Dmpi(x)=

li

mU(X) for l=m,

forO l ml
for m + l < l < k .

In [13], this fact was observed in some particular cases, but was left unexplained. We begin by proving that
N

{Dl u (x). (ai-- x)'} D'n pi(x) = 0

i=l

182

P.G. CIARLET~; P. A. RAVIART"

for the values of l belonging to the (possibly empty) set {0, 1, ..., m - 1}. If u is a constant .4 0, then (2.5) reduces to
N

O= ~ {Ao(a~-x)} Dmp~(x),
i=1

for all x~R ~, AveR.

In particular, this is true for any x ~ K a n d for Ao = u(x), so that the term corresponding to l=0 in (2.5) is equal to zero. The proof of (2.6) will now be achieved by induction on /. If (2.6) holds for I=0, 1..... Io for some l o < m - 2 , then (2.5) yields (since Iv + 1 < m - 1)
N

(2.7)

0 = ~ {A,o. (ai-x) l~ Dmp~(x),

i=1

for all x~R ~, Aloe&alo(R~; R),

since to any element A~o~o(Rn;R) we can associate a polynomial pEPto such

that Dto p(x)= `4tofor all x~ R" (for example, choose x~ R" -- p(x)= l-~`4to . (x)'o).
Thus, we obtain
N

lol

0 = ~ {Ol~
i=1

9(ai--x) t~ OZpi(x),

for all x~K,

by letting Ato=Dt~ u(x) in (2.7). Therefore, (2.6) is proved for O < l < m - 1 . If l=m, then by observing that f i = u if uePm, we get in the same fashion,

Am =
so that in particular

Am" (ai-x) m} Dmpi(x),

for

all

xeR n, Am6.~CPm(R R), n',

Dmu(x)='-~ E {Dmu(x) 9(ai-x)m}Dmpi(x), re!i= 1

for all

x~K.

The proof of (2.6) for m + 1 < I< k is identical, and is omitted.

Remark 1. With the same assumptions as in Theorem 1, we can expand the value u(x) for any x e K via the multipoint Taylor formula
N

(2.8)

u (x) = ~, u (as) Pi(x) + ~1(Ok+t u, x)

i=1

where the functions p~ do not depend upon the function u, and the remainder, denoted by ~(Dk x), depends upon the (k+l)-st derivative of u evaluated at a finite number of points of K, but does not depend upon lower order derivatives of u. This is an immediate consequence of (2.3) with m = 0 . Before proving error estimates (Theorem 2), we must examine some geometrical properties of point sets of R". Let 2~={a~}~=1 and ~ ={d~}~= l be two sets of N points of R ". Then we say that the two sets are equivalent if and only if there exists an invertible element B~.~(R ~) and a vector b~R ~ such that (2.9)

a~=Ba~+b,

l<i<N.

Lemma 1. Let ,~={ai}~=l be a k-unisolvent set, and let 27={a~}~=1 be an equivalent set of points. Then ~ is k-unisolvent.

Lagrange and Hermite Interpolation

183

ProoL Let/~i, l < i < N , be the unique polynomials of degree =<k such that ~ (a j) = 6 ~j, 1 < i, j < N. The polynomials P t deftned by (2.10) x~R"~p~(x)=~i(B-l(x-b)), l<i<N,

are such that pi(aj)= 6iy, 1 < i, j < N, which proves that 27 is k-unisolvent. Example 2. Let K (resp. K) be a nondegenerate n-simplex of R" with vertices ai (resp. ai) and with mid-points a~i (resp. aij) on the edges (cf. Example 1). Then the 1-unisolvent sets ~=x and , - - t ~ = a are equivalent, and likewise, the 2-unisolvent sets
271I

..~ai~i= 1 k.l{aU}l<_i<j~_n+ 1

~n + 1

and

~ l I ----S a ( n J il= l3 f a ~. ijSl<=i<j~_n+l "~ ~ i+ k

are equivalent. Given a set 27={ai}N=1 of N points of R", we henceforth denote by K=K(27) the closed convex hull of 27, and we associate with K the two following geometrical parameters. (2.11) (2.1 2) h = h ( Z ) = d i a m e t e r of K, p = p (Z)= sup {diameter of the spheres contained in K}

(in many applications, K is an n-simplex, so that p is the diameter of the inscribed sphere). If 27 is another set of points of R", we shall likewise denote by/~ its closed convex hull, and by ~ and t3 the two parameters defined as in (2.11)-(2.12). If 27 is any k-unisolvent set ( k > 1), the interior/~ of K is nonempty, so that p is >0. The following simple result will be of crucial importance in proving the error estimates. Lemma 2. Let 27={ai}~= 1 and ~={ai}~'= x be two equivalent k-unisolvent sets, with ai=Bai+b, l <-i<-N, for some Be.~q'(R") and some b e R ~. I f II" II denotes the operator norm induced by the usual Euclidean vector norm of R ~, we have (2.13) IlBll<h^ p and

IIn-Xll<~ P
with

Proof. Since ~ > 0 , we can write

Ilnllas flln~ll. Ilnll--sup~, ~R"

II~ll=~}
with ~, ~e/~,

Each vector ~ e R ~ which satisfies by definition of ~. The vector

I1~11 can be written as ~ = ~ - s =~

B ~ = ( B y + b)-(B'~ + b)

184

P . G . CIARLET & P. A. RAVIART:

is thus of the form B ~ = y - z with y, zeK; thus [[B~[[ __<hby ddinition of h, from which the first inequality of (2.14) follows at once. The proof of the second inequality of (2.14) is identical. Theorem 2. Let Z = {ai}~=, be a k-unisolvent set of points of R", and let h and p be defined as in (2.11)-(2.12). Let ue:~-k+~ (K) be given with (2.14)
Mk+ , = s u p { [[Ok+ ' U(X)[[ ; x e K } < + oo.

I f u is the unique interpolating polynomial of degree < k of u, we have for any integer m with 0 <_m <_k, hk+ 1 (2.15) sup { liD mu ( x ) - O m~ (x) ll ; x E K } <=C M k +, pm for some constants

(2.16)

C = C ( n , k, m, F.)

which are the same for all equivalent k-unisolvent sets and which can be computed once and for all in a k-unisolvent set Z equivalent to Z.

ProoL Since the closed convex hull K = K(Z) is always Z-admissible, the formula (2.3) holds for any point x~K. Thus, for any such point,
N

(2.17)

1 Z[{Dk+Xu(rh(X)).(a,_x)k+,}lllDmp,(x)ll [IOm u(x)-Omfi(x)l[ < (k + l)-------~,=,

By definition of h and Mk+ 1, we first have (2.18) [{O k+' u (q,(x)) . (a,-- x) k+ '} I <= +, h k+1. Mk

Next, let us choose once and for all a k-unisolvent set ,~ equivalent to Z. From the proof of Lemma 1 (of. (2.10)), the associated polynomials p~ a n d / ~ satisfy
p,(x)=~,(a-l(x-b))

for all xeR",

so that, for any vectors ~, e R", 1 _<# < m,

Dmp,(x) "(~1, 42..... ~m)= D ' ~ , ( B - ' ( x - b ) ) . ( B - ' ~.,, B - ' ~2..... B - ' ~,,),

whence sup {[[Omp,(x)l[; x e g } < sup { Ilom/3,(B -a (x - b))[1; x e K } IIB- 'll m. Since the image of K under the mapping ~--* x = B J + b is precisely K, we can write

sup { IID" (n-' (x -

b))ll; x e K } = sup

{ll Dm/~i (x)I1 ; ~ g}.

F r o m Lemma 2, ][B-* l[m< ~/pm. Combining this with (2.18) in (2.17), we obtain (2.15) with

C(n, k, m, s
which completes the proof.

h" S'N ~eg}, ( k + 1)! ,~1 sup {llDm/3,(~)ll ;

Remark 2. The quantities sup {I[Dm/~R~)I] ; ; e R } for m > 1 can all be bounded above in terms of sup {[~(~)1; 5 ~ g } , for one can prove that if P~Pk and K is

Lagrange and Hermite Interpolation

185

any compact convex subset of R ~, then the following generalization of MARKOFF'S inequality in one variable holds (cf. [14, w sup

{llDp(x)l[;xeK}<C-b-~ sup {Ip(x) l; xeK}

k2n

where C is a numerical constant and a(K) is a geometrical parameter associated with K (note that tr(K) reduces to the diameter of the inscribed sphere when K is an n-simplex). Remark 3. It is clear that the estimates of (2.15) for m > 1 are better when the ratio hip is small. The intuitive significance of this is that one should not consider k-unisolvent sets Z whose closed convex hull is " t o o flat", i.e., which is " a l m o s t " contained in an ( n - 1)-dimensional linear manifold of R n. For example, if K is a 2-simplex (i.e., a triangle) in R 2, one has the estimate

1 <h< 2 2 t a n O = p = sin0 2
where 0 is the smallest angle of K, which shows that the smaller 0 is, the poorer the estimate is. Example3. Consider the approximations of types I and II described in Example 1. In both cases, p is equal to the diameter of the inscribed sphere of the n-simplex K. Let us denote by 21 (x), 22 (x), ..., 2n+ 1(x) the barycentric coordinates of a point x of R ~ with respect to the vertices ai, 1 < i<n + 1, of K. Then approximation of type I is given by [13, Theorem 1]
n+l

fi(x)= ~
t=l

u(a~)2~(x),
^ n+l

so that (for k = 1,

N=N(k)=n+ 1)
^ N N

C(n, 1,0,27)=-~-.v,~lsup{lA,(x)l; ~c~R}= 2 ' C(n, 1, 1, 23--~-.v,~lsup(llO2,(~)ll"' 3c~}__< 2

n+l

/3'

the last inequality being a consequence of [13, Lemma 1]. To obtain the lowest possible ratio 7i/~, one should pick for ~ an equilateral n-simplex, for which [12, Theorem 3]

Similarly [13, Theorem 3], approximation of type II is given by

n+! i=1
l <=i<j<=n+ l

with

p~(x)= 2 (2i(x)) 2 - 2i(x), pij(x)=42i(x)2j(x),

1 < i < n + l,

l<i<j<n+l,

186

P.G. CIARLET,r p. A. RAVIART;

so that we can again compute upper bounds for the constants C(n, 2, m, ~) for m = 0, 1 and 2. Thus the error estimates of Corollaries 2 and 4 of [13] are direct consequences of Theorem 2, with the new estimate corresponding to m = 2 in the case of approximations of type II. 3. Hermite

Interpolation

Let k be a given integer ~ 1. Then the natural extension of Hermite interpolation in one variable should be as follows: we are given an integer s with 1 < s < k, and (s+ 1) sets of points (3.1) Z~
"( i S i = l '

~176 z t --~) ~ it~N'' " , a Ji=l, -

Z s ~ ' ( raS~N" iJi=l"

9 ; r In a given set ,,r __ f ~ r ' l . N r t, we assume that air ~a~ if i~j, but we do not exclude -~,,~i= the possibility that a~.= a~; for some 1 < i < N, and some 1 < i' < iV,, if r ~=r'. Indeed, in practical applications one has

Z~ cZ~-1 ~ . . . ~ Z t ~ z o , although this assumption is not necessary for our theory. We say that (3.2) Z=Z~ 1 u ... u Z ~

is a k-unisolvent set provided that, given any A~.~,(R"; R), 1 <i<N,, O<r<s, there exists one and only one polynomial PePk such that

D'p(a~)=A~,

l < i < N , , O<r<s,

with the convention that Lao (R"; R)= R.

Example 4. Let K be a nondegenerate n-simplex of R" with vertices a~, 1 _<i< n + l . We denote by a~jk=aj~k . . . . . akj~ the barycentre of each triangle with vertices a , aj and a k, with i=~j, j=i=k, k#-i. Then the set
tzo_j- a zm=z~ 1 with /El
t =/
).n+l --~. i / i = 1 / J i = 1,

u~'a
L

ijkJi,

~n+l

j,k=l

;a " t n + l

i~=J'J#k'k*i

is 3-unisolvent [13, Theorem 5], and the associated interpolating function (as defined below) will be called an approximation of type IlL The above formulation is, however, not general enough: in many cases, all the conditions needed to define the derivative D' p (a~.) as an element of A~ (R"; R) cannot be specified at all the points a[, the reason being that we need to match the number of given parameters with the number of coefficients of a polynomial of degree k. Instead, here it is only required that some " p a r t i a l " derivatives D'p(a'i). (~l, ~z, ..., ~,) are to be interpolated for ( ~ , ~2,-.., ~,) belonging to some subset X: of (R")'. For example, if R ~ is equipped with the canonical basis (ek)~= 1, it can be required that -(32P (a2), 1 < k < n, must be interpolated at some ~k point a~, but the cross-derivatives ~2 p (a~), k 4:1, are not to be considered. In this

Lagrange and Hermite Interpolation

187

case, X ~ = ~){(e k, ek)}, whereas the full knowledge of the second derivative
k=l

D2p (a?) would require that, for instance, X~ = 6 {(ek, e,)}. z

k,l= l

Accordingly, the Hermite interpolation problem is defined as follows: We are given a set Z of the form (3.2), each Z' being as in (3.1); in addition, to each point a[,for 1 <r<s, we associate a subset Xr of (Rn)~ (X.', is of course assumed to be nonempty, and (r ~2. . . . , ~,)EXr implies that all ~i's are ~e0). Then we say that the set 27 is k-unisolvent if and only if, given any A'~.2~,(R"; R), l<i<N,, O<r<s, there exists one and only one polynomial p~Pe such that, for all 1 < i<N, and all 0 < r < s, (3.3) D'p(a~). (~1, ~ .... , e r ) = A ~ . (r

~2 . . . . . ~r)

for all

(r

~2, ..., r

with the convention that (3.3) reduces to p(a~ ~ when r = 0 . Likewise, given a function u such that D" u(a'~) is defined for 1 <i<N, and O<r<s, we say that is its interpolating polynomial if it is the unique polynomial of degree < k such that (3.4)
D ' ~ ( a ~ ) . (~1, ~2 . . . . . ~ , ) = D " u ( a g " (~1,

~2 ..... ~r)

for all (~1, ~2,-.., ~,)~X[, l<i<N,, O<r<s, with the convention that this reduces to fi (a ~ = u (a ~ for r = 0.

Example 5. Let K be a triangle in R 2 with vertices as, 1 ___i < 3, and let atj =aj~ denote the mid-points of the edges joining ai to aj, for i #j. Then, given a function u, the interpolation problem consists in finding a polynomial fieP5 such that
fi(ai)=u(ai), Dfi(ai)=Du(ai), O2fi(ai)=D2u(ai),
1_<i<3, 1<i<3, 1_<i_<3, l<i<j<3

(3.5)

Off

Ov (aij)=-O-v-(aiy)'

du

where O/Ov denotes the (interior) normal derivative along the boundary of K. This type of approximation was considered by ZLAMAL[35], who showed that the problem has a unique solution. In fact, the above set of conditions implies that Dfi(ai~ ) is known, since fi is a polynomial of degree 5 in one variable along the edge joining as to aj, with known values for itself and its first and second derivatives at the points ai and aj. As will become apparent later on, the key observation is that we can replace conditions (3.5) by

(3.5')

D?t(aij).(ak--aiy)=Du(aij).(ak--aij),

l<i<j<3

where k # i and k#j. The associated interpolation function will be called an

approximation of type V.

188 The set

P.G. CIARLET& P. A. RAVIART:

ZV=~O UZ1 U Z 2

where Z~ = {a,}~=,, Z'={a,}~3=,u{a,j}31 Z2={ai}Lt with with

Xil=R 2,
XI2=(R2) 2,

Xi~=ak--aij,

is clearly 5-unisolvent. Let us notice that for X~ (resp. X~), we could as well take the sets )(.1,={aj-a~, ak--a,} (resp. X~ ={(aj-a,, aj-a,), ( a j - a i, ak--at), (ak--a~, ak--ai)}). We shall now prove the analog of Theorem 1 for Hermite interpolation, but for the sake of simplicity we shall restrict ourselves to the case s = 1. The extension to the case s___2 offers no difficulties, but it is long and cumbersome and for this reason will be omitted. Theorem 3. Let there be given a k-unisolvent set

Z=Z~

1, with Z ~- - [ ~176 iJi=l

and

Z l --~ a i11s'l , t Ji=

such that with each point a~, I < i< N1, is associated a subset Xt of R", in each of which some vectors ~h, ~2 ..... ~a, form a basis. Let K = R ~ be a Z-admissible set, and let u e ~ "k+1(K) be given. Then, for any point x e K and any integer m with O< m < k, one has 1 No D'fi(x)=Dmu(x)-~ ( k + l ) t ~=1{Ok+ Xu(~i~(x)) "(a~ x) k +l } Omp~ (x)
(3.6) 1 ~ d, +-~-v i~t k=t {0

U(qlk(X))'(r

--x)k)} O'~ plk(x)

where the p~ ' s and p~k'S are the unique polynomials of degree < k such that
0 0 Pi (aj)=6ij,

1-<i, j<<_No, l<-i<No, l < j < N x , l<_k<_dj, l < i < N 1, l_-<j_-<No, l<k<-d i, l<=i, j<=N1, l~_k, l<=di,

(3.7)

Dpi(aj).(~k)=O, 1 o Pik(aj)=O, Dp~(aJ).(~)D=6,j6kt,

and, for some 0 < 0~ O~k< 1,

(3.8)

rl~176

x +(1--O~ ~
No N1

q~k(x)=O~kX+(1--O~k)a ~.
d~

Proof, It is readily seen that the interpolating polynomial fi is given by fi= Z u(a~
i=1

~-, Z {Du(al). (r
/=1 k = l

so that, for any 0 < m _ k,

NO NI di

Draft= E u(a~
i=X

E {Ou(a~). (r

/=1 k=l

Lagrange and Hermite Interpolation Let x 6 K be given. Since u ~ 3 "k+ 1 (K), we may write u (a ~ = u (x) + D u (x). (a ~ - x) +... + ~
1

189

O k u (x). (a ~ - x) k

I -F (k + l)~ Dk+1 uOlO(x)). (aO_x)~+1

for I < i < N o, and O u ( a l ) . (r 9( r for I < i < N D l<k<di, 9(r D O 2 u ( x ) . (ilk, al - x ) + ... 4 ( k - 1 1)-------~ k u (x)

k+lu

(~l[k(X))'(r

k)

where r/~ and r/]k are of the form (3.8). Thus, we have

D"~(x)= ~o~,~ ,co' u(x) ( : - ~)') D"po(~)

+~, (~_ ~)~,~, ~{D' u(~). (r (a:-~)'-'11 D'p;,O,)
1 No 4 (k + 1)~i~, {Dk+1 u(tl~ (x)). (a ~ - x) k + 11D mpO (x) k
I N~

k 1 No

d,

I N, d, { Dk+ +-~. ,~=, ~_, ' U(tll'(x))'(r

Om p l ( x)"

Using the fact that ~ = u whenever u e P k, it can now be shown, exactly as in the proof of Theorem I, that

,~o~,~ ~D'u(x). (ao-~)') D"po(x)

k 1 N, + ~ (l-- 1)' ~
1=1 "i=l

k 1 No

a, ~ {D~ U(X)" (~k, (a~ -- x)k)} D ~ p~ (x)-- D" u (x),

k=l

which completes the proof. As in w we must introduce the notion of equivalent k-unisolvent sets. Let ~ _ raO~No ,, f,,~N, u f,a s[N"
--~ iJi=l~tji=l k-)''" t Oi=l

and

.~_~O~No

--~. i J i = l k " l l .

S~IIN, ,

i JI=I~J'"LJ(

/~N.

iji=l

be two sets of points of R ", a subset X[ (resp. X~.) of (R")' being associated with each point a~ (resp. ~7) for 1 < i < N , , 1 < r < s . Then we say that the two sets are equivalent if and only if there exists an invertible element Be'2' (R ~) and a vector b~R" such that (3.9) and
(3.10) X/r--'~{(r 42 . . . . . Cr)~(Rn)r'~ r

a'~=B~[+b,

l<i<N,,

O<r<s,

for all (~1, ~2) "") ~ , ) E - ~ } ,

1 <i<_N,, 1 <r<_s.

190

P.G. CIARLETrye. A. RAVIART: P.

Example6. It is clear that condition (3.10) is automatically satisfied at a point a~ if the derivative D r fi (a~) is specified there, in which case we may choose X r = X f = ( R " ) r. The same holds if the elements (~1, ~2. . . . . ~r) of X~ are such that each vector ~, 1 < i<r, is expressed as a linear combination of points of Z. This is the case for approximation of type III (cf. Example 4) and for approximation of type V (cf. Example 5), once the interpolating conditions (3.5) are replaced by conditions (3.5'). Lemma 3. I f two sets Z and ~ are equivalent in the above sense, and if Z is k-unisolvent, then Z is k-unisolvent.
Proof. We give the proof in the case s = 1. Let fro and ff~ be the polynomials associated with the k-unisolvent set 2 and which satisfy (3.7). It is readily seen that the polynomials pO and P~k defined by
x~R"~p~176 l <_i<_No, l <i<N~, l <_k<d,,

(3.11)

x~R"~p]k(x)=~,~(B-~(x-b)),

satisfy (3.7), which shows that the set Z is k-unisolvent. We can now prove our main result.

Theorem 4. Let there be given a k-unisolvent set of points of R ~

Z=Z~ s, with Zr=Ya r~N" I. / J i = 1, O<_r<_s,

such that each point a~, I < i < Nr, 1 <r<=s, is associated with asubset XF of(R") r. Let K c R" be a Z-admissible set, and let u ~ 3 "k+ l (K) be given, with

(3.12)

Mk+ l=sup{llDk+ l u(x)ll ; x~K} < + c~.

I f fi is the interpolating polyomial of u, i.e., uEPk and

(3.13) (3.14)

fi(a~176

l <-i<-No,

Dr u(a~) " (~x, ~2..... ~r)=Or u(a~) " (~1, ~2..... ~r) l < r < s , we have for any integer m with

for all (~1,r ..... ~,)eXr, l < i < N , , O<m<k,

(3.15)

hk+ l sup { IIDm u (x) - D m~t(x) ll ; x eK} < C M k +x - pm ' -

for some constants

(3.16)

C = C ( n , k, m, ~)

which are the same for all equivalent k-unisolvent sets and which can be computed once and for all in a k-unisolvent set ~ equivalent to Z. The parameters h and p are defined as in (2.11)-(2.12).

Lagrange and Hermite Interpolation

191

Proof. For simplicity, we shall give the proof in the case s = 1. From (3.6), we obtain for any point xeK,

IIDrn U (X) -- D" u (x)II ~ - - 1

( k + 1)! =
1 N1

~=~{D~+~u(,tO(x)).(a?_x)~+~}l llD'p~
di

No

+-~. ,~=~k~= l(Dk+' u(rl:k(X)) "(r (a~- x)k)} I liD" P~k(X)ll.

By definition of h and

Mk+ 1, we have
(a ~ x) k+'}l_- MR+ 1 h k+ t, < (al -- x)k)} I < Mk +, I[~h[I hk.

]{Dk+l u(q ~ I{Dk+~ u (r/~k(X))" (r

Now let Z be a k-unisolvent set equivalent to Z. From the definition of equivalence, it follows that the vectors ~k are related to the corresponding vectors $~k by (cf. (3.10))

G=B'G.

Thus

IIGII ~ h II~hll
by Lemma 2, so that

r (al_x)k)}l<Mk+~hk+X II~:kl]
Next, as in the proof of Theorem 2 we have

sup{llDmp~

xeK}-< sup {llOm/3,(~)11 ; ~ e K : } pm -

~m
'

sup (liD" ph(x)ll ; x EK} ~ sup { HD" ff,Xk ; ~E~} (~)ll

using (3.11). Hence, we obtain (3.15) with
~m No

prn '

C(n, k, m, ~ ) = -(- + 1)! ,~xsup ~llD, ff,(~)[i ; k x-~ r o

k! /3 i=1 k=l

II~hll sup{llDmph(~)ll ; ~ K } .

Example 7. If we consider approximations of type III, as defined in Example 4, we obtain h4 sup{HD'u(x)-D'fi(x)ll;xeK}<CM4--~for m = 0,1, 2 and 3. P These bounds were announced in [13]. Similarly, we obtain for approximations of type V (of. Example 5),

sup{llD"u(x)-D'fi(x)ll; xEK}<=CM6--dP

h6

for 0 < m < 5 .

192

P.G. CIARLET P. A. RAVIART:

These bounds were proved by ZLAMAL[35, Theorem 3] for 0<m_<4. Again, we emphasize that central in our approach is the fact that the interpolating condition (3.5) must be written in a form which is the same for all equivalent 5-unisolvent sets (triangles in the present case). This requirement is not satisfied by (3.5) but is satisfied by (3.5'). Finally, we want to mention a development of the preceeding approach which is a generalization of the Hermite interpolation problem in the following sense. We are still given a set ,~ of the form (3.2), and, given a function u, it is again required to find an interpolate ~ such that the interpolating conditions (3.4) are satisfied. However the condition that ~ be a polynomial of degree < k will now be replaced by the condition that ~ belong to some finite-dimensional vector space P such that (a) the set ~ is P-unisolvent, i.e., there exists one and only one ~ e P which satisfies conditions (3.4) for all 1 <_i<_N,, and all O<_r<s, (3.17) (3.18)

(b) P k c P for some fixed k, and thus

(r ~ = u whenever UePk.

In all the cases previously considered we had P=Pk, but the necessity for considering such more general problems will be made clearer in Example 8 below. Let us just mention that the exact analog of Theorem 4 holds without modification. It suffices to observe that Lemma 3 is now replaced by

I.emma 4. I f two sets Z and ~ are equivalent in the sense of (3.9)-(3.10) and if Z is P-unisolvent, then ~ is/3-unisolvent, if we let
(3.19) /3 = { i f i : / ~ R ; / 3 ( ~ ) = p(B~ + b), Y x e K , Y peP}.

Remark 4. As a space of functions, /3 might in general be different from P (except if P = Pk for some k), but this is irrelevant. Example 8. The set Z ul described in Example 4 is 3-unisolvent. We shall define an approximation of type III' by specifying that the values of the interpolating polynomial ~ at the points aij k will no longer be considered as free parameters (and which could thus be equated to the values u(aok ) in approximation of type III), but instead that these values u(a~jk) will be replaced by the following linear combination of the values of the approximated function u, together with its first derivative, at the vertices a f, aj and ak:

(3.20)

(atjk) = 89 (a,) + u (a j) + u (ak) } {u -- ~ {D u (a~) .(ai- ai j k) + D u (ay). (aj - a~j k) + D u (ak) . (ak -- ai jk)}.

This is a generalization of the approximation considered in [36] (for an engineering source, see [5]). Although the order of accuracy obtained is one less than the corresponding order of accuracy for approximations of type III (as will be shown below), this method has the practical advantage that the computing time for the associated finite-element method is smaller than for approximations of type III, or even for approximations of type II (of. [36]). To study this type of approximation, we prove

Lagrange and Hermite Interpolation

193

Lemma 5. Let T be a triangle with vertices at, 1~_i<3, in R" (n> 2),and let a be its barycentre. Then, if ueP2, one has
3 3
i=l

(3.21)

u(a)=89~ u(a~)-~ ~ Du(at). (at-a).

i=1

Proof. Denoting by Ke Aa2(R~, R) the constant second order derivative of u, one has u(a~)=u(a)+Du(a). (a~-a)+89 (a~- a) 2, 1 < i < 3 ;
3

hence, since ~ ( a i - a ) = 0 by definition of a, we obtain

i=1 3 3

(3.22)

~ u (as) = 3 u (a) + 89 K . (a t - a) 2. ~
i=l i=l

On the other hand, we have

Du(at)=Du(a)+K.(ai-a ),
3

1<i<3,

whence, again using the relation ~ ( a t - a ) = 0 , there results

i=l 3 3
i=1 3

(3.23)

~ Du(a~). ( a t - a ) = ~ K . (a~-a) 2
i=l i=1

and (3.21) follows by eliminating ~ K. (a~-a) 2 between (3.22) and (3.23). From the preceeding lemma it follows that u = ~ whenever u~P2 (instead of P3 for approximations of type III). We can then apply the theory developed in ~ 2 and 3: the Taylor series will then be written with a remainder involving the third derivative of u instead of the fourth; hence we obtain for approximations of type III' h3 sup{llDmu(x)-Dmfi(x)ll; x~K}-<CM 3 -~- for m=0, 1, which is the generalization of the bounds obtained in [36] for the case n = 2. In approximations of type III', the space P consists of the subspace of P3 for which the homogeneous relation (3.20) holds for all i+j, j~ek, k +i.

4. Interpolation in Sobolev Spaces For any integer m_~ 1 and any 1 <p___ oo, we denote by w m ' P ( ~ ' ~ ) the Sobolev space of (classes of) real-valued functions which, together with all their partial distributional derivatives of order ~rn, belong to LP(~2). It wiU be convenient here to use the norms and semi-norms respectively given by*

Ilull"P'n= l~o IIDullp, o] ~

lulm.~,o=llD'~ullp,

* Here, for example, liDSullp,~ means (SII~ u(x)llP dx)X/p, etc.

13 Arch. Rational Mech. Anal., Vol. 46

194 for 1 =<p< oo, and by

[lUJ]. . . .

P.G. CrARL~T& P. A. RAVIART:

o=max{J[Dlul]~, ~; O<l<_m},

lUlm, 00,0= [[DmUH ao,s~

for p = oo. Those norms and semi-norms are readily seen to be equivalent to the usual ones. W e denote by (wm'P(O)) ' the strong dual space of Wm'F(Q), by (f, u) the pairing between an elementfe(Wm'P(f2)) ' and an element ue wm'P(~2), and by

'lfll*m'g'~=sup[[V[]m,v,~l(f'v)[; veWm, p(~), v#O } {

the dual norm. We begin by proving a preliminary result, which is essentially Theorem 2 of [8] (see also [7]). We give the proof for the sake of completeness. Lemma 6. Let 0 be a bounded open subset of R n with a continuous boundary (in the sense of [21 ]), let p be given with 1 <p < oo, let k >= be a fixed integer, and 0 let f e ( W k +1, p(~)), be such that (4.1) (f, u ) = 0

for all ueP k.

Then there exists a constant

(4.2)
such that
(4.3)

c = c (n, k, p, t2)

[(f,u)l<=C[lfl[*+l,p,s~Ju[k+l,v,~

for all ueWk+l'v(O).

Proof. From a result given in [21, Theorem 7.2, page 112] for the case I < p < oo and which can be easily extended to the case p = oo, it follows that the semi-norm I" Ik+ 1,n, ~ is a norm on the quotient space W k +1, n (O)/pk equivalent to the quotient norm [[[u] [[wk+,. p(s~)/ek= inf{ I[u + V[[k+1, p, D ; vePk},

where [u] denotes the equivalence class of any of its elements u. Thus there exists a constant C= C(n, k, p, ~) such that

[[[U][[Wk+,,p(~)/ek<=C]U[k+l,p.~
Given an f e ( W k +1,. (s SO that

for all uewk+I"P(s

which satisfies (4.1), we have

(f,u)=(f,u+v)

forallueWk+1'P(O)

andall

VePk,

[(f, u)[ <=[[f []*+l,p,~inf { l[u + v[[k+l.p.~;

<ell * fllk+l,p, ~ lu Ik+,,~,~,

V~Pk}

which completes the proof. We need another technical result: Lemma 7. Let ~2 be a bounded open subset of R n with a continuous boundary, let p be given with 1 <=p< oo, let k >O be a fixed integer, and let m be an integer

Lagrange and Hermite Interpolation

195

with O<m<k + 1. Let HeLe(wk+I,P(O); wm'p(f2)) be such that

(4.4)

IIu=u

for all U~Pk.

Then there exists a constant

(4.5)

C=C(n, k, p, f2)

(the same as in Lemma 5) such that

(4.6)

Ilu--Flullm, p,o~C[lI--Flll~cw~+..<m;w~,.cm) lUlk

Given g~(Wm'P(f2)) ', the linear form defined by f: u --, (f, u) =(g, u -Flu)

for all u~ W k+ l'P(f2).

Proof.

is continuous over wk+I'P(~) (which is contained in wm'p(O)) since ']fll~'+l, ~ = s u p { ,(g,u-Flu),

P' Ilull~+x,p,~ ;

u~Wk

=<]lgllm*,,,osup ~ llu--Flu[lm,',~ ,. u~Wk+ l,'(O), u ~:O} ~ , ~

--< IIgll*, p, ~ IlI-Flllsecw~+ 1.,r

wm.'<m)"

Since (f, u) = 0 for any u EPk, the conclusion follows by applying Lemma 5 combined with the fact that [31, Theorem 4.3-B, page 186]
Ilu-Flullm p , o = s u p { ] ( g , u -, F l u ) ] ; ' Ilgllm,~,~

g~(Fvm,P(f2)),, g~0}.

This completes the proof. The proof of the fundamental error estimates (4.13) of Theorem 5 will depend upon the notion of equivalent domains, as in the preceeding sections. Let 12 (resp. ~) be a bounded open subset of R n. Then we say that 12 and ~ are equivalent if and only if there exists an invertible element B~Ae (R n) and a vector b s R ~ such that (4.7) O={xER*; x=B~c+b for each Sr

With each function u defined over I2, we associate a function fi defined over by letting (4.8)

ft(~c)=u(B~c+b)

for each ~c~12.

As will become clear from the proof of Theorem 5, the correspondence u--, is an isomorphism between Wm' p (I2) and Wm' p (~) for each m and p. Likewise, if 17 is an element o f ~ ( W k + l ' p (f2); W m ' p (f2)), we associate with H an element/7 e ~ ( W k+1,1,(~); wm,p (~)) by letting (4.9)
13"

Hft =Hu

for each u~W k+1, P(O).

196

P.G. CIARLET& P. A. RAVIART:

It can easily be seen that Pk is left invariant by H if and only if it is left invariant by/I. Finally, with each bounded open subset 12 of R", we associate as in w the two geometrical parameters: (4.10) (4.11) h = diameter of 12, p = s u p {diameter of the spheres contained in ~};

we leave it to the reader to verify that Lemma 2 holds without modification. Theorem 5. Let 12 be a bounded open subset of R" with a continuous boundary, let p be given with l <__p<oo, let k >=O be a fixed integer, and let m be an integer with O<m<_k + 1. Let H ~ . ~ ( w R + l'~ WIn'P(12)) be such that (4.12)

H u =u

for all u e P k.

Then for any uewR+l'P(12) (and for h small e n o u g h / f p < oo; cf. (4.18)),
hk+l

(4.13)

Ilu--l-lUllm, p,O<Cdlulk+x,p,~
A A

pro,

for some constants

(4.14) ~r k, p, O, n )

(where 17 is defined as in (4.9)) which are the same for all equivalent domains 12 and which can be computed once and for all in a domain ~ equivalent to 12.
Proof. If D:u is any distributional derivative of u, then D l ~ is defined for each ~ e ~ by

Dt fi(~r 9(~1, ~2..... ~l)=Dt u ( B x + b) 9(B~I, B~2 ..... B~l),

so that for 1 = p < oo

IlD' ~t(Fr p dfx~- I[BII~PSHDt u(B?x-b b)l[Pd~

D

= IIell~pIdet(B) 1-1 S liD' u (x) IIp dx, that is, (4.15) ]u t~,p.~< ]IBIt~ldet(B)l-~/~ l~.p, ~,

for all 1 __<p_~oo, if we understand that l / p = 0 if p = oo. Similarly, we obtain

(4.16)

[ulz.p,~<llB-~ll~ldet(B)l~/~l~ll, p,~.

It follows from (4.16) that (4.17) IlullP ~,~__<ldet(B)l~ IIB-all~Pl~ It, p, ~_-<ldet(B)lllB-ali m" Ilullm,~,~ p ^ ~
1=0

under the assumption that 1 < (4.18)

liB-x I[; this in turn is certainly satisfied for

h__</~,

Lagrange and Hermite Interpolation

197

i.e., for h small enough (once the domain 1] is chosen) since we then have from Lemma 2 _~_ 1 1< < Ilnll <llB-Xl[" Thus for h satisfying (4.18), we obtain from (4.17)

(4.19)

Ilu -rlult..p,9<-_ldet(B)

ll/Ptln-XlI" Ilfi- B fi I1.,,~, ~.

From Lemma 6, we have (4.20) [lfi-nfilr,,,v,~<f(n,

k, p, fi)[lI-~qll~tw~+..,<~); ~-.,<~))Ifilk+l.p.h.

Therefore, by using the inequality (4.15) for l = k + 1 and the estimates (2.13) of Lemma 2, we finally obtain the inequality (4.13) with
^m

(4.21)

c~= Cg(n, k, p, ~ , / I ) = ~

C(n, k, p, fi)I[l-/Ill.~,tw~

w,-.,tfi))

where C(n, k, p, ~) is the constant (4.2) of Lemma 5. This completes the proof. We shall now apply the above theorem to the generalized Hermite interpolation defined at the end of w In so doing, we generalize similar results obtained in [10] for the case p = 2, n = 2, and obtained in [34] for a restricted class of polynomial approximations. Likewise, we generalize some of the results of [6], [8] and [9] which were obtained for spline and Hermite interpolation over "rectangles" of R ~.
T h e o r e m 6. Let there be given a P-unisolvent set T. of the form (3.2). I f u e C ' ( K ) (we recall that K is the closed convex hull of X), we let ~ denote (as before) its Hermite interpolation, in the sense that fi is the unique element of P such that

(4.22)

O" u(ar) 9(~l, r .... , ~,)=Oru(ar) " (~1, ~2..... ~,)

for all (~1, ~2, ..., ~r)~Xr, l < i < N , , O<r<s. We also assume that

(4.23)
so that

Pk c P C W k+ l' P(K)

for some fixed integer k >=O

(4.24)

fi=u

whenever

U~Pk.

Then if u~ wk+ I'P(K), with

(4.25) s < k + 1 --n

i f p<oo or s<-k i f p=oo,

we have (for h small enough if p < ~ )

h~+ I

][u -- fillrn, p, ~ < cg']u [k+ 1, p, ~ om ' for some constants (4.27) c6= Cr

(4.26)

for any O_<m_<k+l,

k, m, 2~,/3)

198

P.G. CIARLET& P. A. RAVlART:

which are the same for all equivalent sets and which can be computed once and for all in a P-unisolvent set ~ equivalent to E. The parameters h and p are defined as
n (2.11)-(2.12). P r o o f . T h e c o n d i t i o n (4.25) guarantees t h a t wk+I'p(K)cCS(K), so t h a t the H e r m i t e i n t e r p o l a t i o n ff is well-defined. F o r a n y ue wk+I'P(K), let IIu=ff. This defines a n element w h i c h is in Wm'P(K) f o r all 0 < m < k + l , and moreover He.~(Wk+I'P(K); Ws'p(K)) since the injection wk+I'P(K)--,CS(K) is continu o u s (cf. [21, w a n d P is finite-dimensional. T h u s the inequality (4.26) is a n i m m e d i a t e consequence of (4.13), the c o n s t a n t c~ of (4.27) b e i n g given b y (4.21) w i t h / I defined b y / 1 ~ = ft. W e leave it to r e a d e r to a p p l y the a b o v e t h e o r e m to the v a r i o u s types of a p p r o x i m a t i o n s (I, II, III, I I I ' a n d V) c o n s i d e r e d in E x a m p l e s 1 to 8.

Acknowledgement. The authors sincerely wish to thank Professor GILBERTSTRANGof M.I.T., with whom they had many helpful discussions while he was visiting Paris. In particular, they discovered many common features in their respective recent works. In this regard, they refer the reader to a forthcoming paper of G. STRANO in Numer. Math., entitled "Approximation in the finite element method".
References
1. AttaiN, J.P., Behavior of the approximate solution of boundary value problems for linear elliptic operators by Galerkin's and finite difference methods. Ann. Scuola Norm. Sup. Pisa 21, 559-639 (1967). 2. AumN, J.P., Approximation des espaces de distributions et des ol~rateurs diff6rentiels. Bull. Soc. Math. France (1967), M6moire No 12. 3. BAEU~KA,IVO, The finite dement method for elliptic differential equations (Syrup. on the Numerical Solution of Partial Differential Equations, Univ. of Maryland, 1970). Numerical Solution of Partial Differential Equations-II (B. HUBBARD,ed.), p. 69-106. New York: Academic Press 1971. 4. BABU~KA,IVO, Error bounds for finite element method. Numer. Math. 16, 322-333 (1971). 5. BF.LL,K., A refined triangular plate bending finite element. Int. J. Numer. Math. Engineering 1, 101-121 (1969). 6. BIRKHOFF,G., M. SCHULTZ, R.S. VAROA,Piecewise Hermite interpolation in one and two variables with applications to partial differential equations. Numer. Math. 11, 232-256 (1968). 7. BRAMBL~,JAMESH., Variational Methods for the Numerical Solution of Elliptic Problems (Lecture notes). Chalmers Institute of Technology and the University of GiSteborg, Sweden, 1970. 8. BRAMBLE,J.H., & S.R. HILBERT, Estimation of linear functionals on Sobolev spaces with applications to Fourier transforms and spline interpolation. SIAM J. Numer. Anal. 7, 112-124 (1970). 9. BRAMBLE,J.H., & S.R. HILBERT, Bounds for a class of linear functionals with applications to Hermite interpolation. Numer. Math. 16, 362-369 (1971). 10. BRAMBLE,JAMESH., & MILO~ZL~,MAL, Triangular elements in the finite element method. Math. Comp. 24, 809-820 (1970). 11. CARTAN,HENRI, Calcul Differentiel. Paris: Hermann 1967. 12. CIARLET,P.G., Some results in the theory of nonnegative matrices. Linear Algebra and Appl. 1, 139-152 (1968). 13. CIARLET,P.G., & C. WAGSCrlAL,Multipoint Taylor formulas and applications to the finite element method. Numer. Math. 17, 84-100 (1971). 14. COATI~LEC,CHRISTIAN, Approximation et interpolation des fonctions diff6rentiables de plusieurs variables. Ann. Sci. Ecole Norm. Sup. (3) 83, 271-341 (1966).

Lagrange and Hermite Interpolation

199

15. DIEUDOr~, J., Foundations of Modem Analysis. New York: Academic Press Inc. 1960. 16. Fix, G., & G. STRANG,Fourier analysis of the finite element method in Ritz-Galerkin theory. Studies in Appl. Math. 48, 265-273 (1969). 17. FREDEgtCI~SON,PAULO., Triangular spline interpolation. Math. Report No. 6, Lakehead Univ. 1970. 18. GUENTHER,R.B., & E.L. ROETMAN,Some observations on interpolation in higher dimensions. Math. Cutup. 24, 517-522 (1970). 19. GUGLIELMO, DI, M6thode des 616ments finis: Une famille d'approximation des espaces F. de Sobolev par les translat6es de p functions. Calcolo 7, 185-234 (1970). 20. HALL,C.A., Bicubic interpolation over triangles. J. Math. Mech. 19, 1-11 (1969). 21. NE~AS,JINDi~ICrt, Les M6thodes Directes en Th6orie des Equations Elliptiques. Paris: Masson 1967. 22. N~COLAmES,R.A., On Lagrange interpolation in n variables. Institute of Computer Sciences (London), Tech. Note ICSI 274, 1970. 23. NICOLAIDES,R.A., On a class of finite elements generated by Lagrange interpolation II. Institute of Computer Science (London), Techn. Note ICSI 329, 1971. 24. NICOLAm~,R.A., On a class of finite elements generated by Lagrange interpolation. To appear. 25. NrrseaE, J., Interpolation in Sobolevschen Funktionenr~iumen. Numer. Math. 13, 334-343 (1969). 26. NrrscnE, J., Lineare Spline-Funktionenund die Methode von Ritz for elliptische Randwertprobleme. Arch. Rational Mech. Anal. 36, 348-355 (1970). 27. SALZER,HERnERTE., Formulas for bivariate hyperosculatory interpolation. Math. Comp. 25, 119-133 (1971). 28. SCHULTZ,M.H., L2-multivariate approximation theory. SIAM J. Numer. Anal. 6, 184-209 (1969). 29. STRANG,GILBERT, The finite element method and approximation theory (Symp. on the Numerical Solution of Partial Differential Equations, Univ. of Maryland, 1970). Numerical Solution of Partial Differential Equations-II 03. HtraBAn.O, ed.), p. 547-583. New York: Academic Press 1971. 30. STRANG,GILBERT,& GEORGEFIx, Chapter III of: An Analysis of the Finite Element Method. To appear. 31. TAYLOR,ANGUSE., Introduction to Functional Analysis. New-York: John Wiley 1958. 32. THACrmR,Jr.,HENRVC., Derivation of interpolation formulas in several independent variables. Ann. New York Acad. Sci. 86, 758-775 (1960). 33. THACt-rER,Jr., HENRYC., & W.E. MmNE, Interpolation in several variables. J. Soc. Indust. Appl. Math. 8, 33--42 (1960). 34. ~ENi~EK,ALEXANDER,Interpolation polynomials on the triangle. Numer. Math. 15, 283-296 (1970). 35. ZLAMAL,MtLO~, On the finite element method. Numer. Math. 12, 394--409 (1968). 36. ZLAMAL,MU.O~,A finite element procedure of the second order of accuracy, l'q-umer.Math. 14, 394-402 (1970). Service de Mathrmatiques Laboratoire Central des Punts et Chaussres Paris and Drpartement de Mathrmatiques Universit6 de Paris VI Paris

(Received September 15, 1971)