Vous êtes sur la page 1sur 13

Curved Elements in the Finite Element Method. I Author(s): Milos Zlamal Source: SIAM Journal on Numerical Analysis, Vol.

10, No. 1 (Mar., 1973), pp. 229-240 Published by: Society for Industrial and Applied Mathematics Stable URL: http://www.jstor.org/stable/2156389 . Accessed: 30/09/2011 15:23
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Society for Industrial and Applied Mathematics is collaborating with JSTOR to digitize, preserve and extend access to SIAM Journal on Numerical Analysis.


SIAM J. NUMER. ANAL. Vol. 10, No. 1, March 1973




Abstract.Curved elements, introducedby the author in [13] and [14], which are suitable for solvingboundaryvalue problemsof the second orderin plane domains withan arbitrary boundary theDirichlet fora W(21)-elliptic are discussed.An approximation theorem is proved, problem equation is consideredas a model problemand errorbounds are derived.

1. Introduction. The finiteelementmethod is the Ritz-Galerkinmethod The main feature usingspecial subspaces of trialfunctions. of thesesubspacesis thatthebasis functions vanishoverall but a fixed number ofelements intowhich thegivendomain is divided.The triangular elements are theelements mostoften used by engineers whensolvingboundaryvalue problemsin two variables.The reason is obviouslythat it is easy to approximatean arbitrary boundaryby a the polygonal domain. If we use piecewiselinear polygon and to triangulate functions as trialfunctions, theapproximation oftheboundaryby a polygonis a naturalone. However,ifwe use functions whichare piecewisepolynomials of a higherdegree,we cannot expect to retainthe same accuracyalong the curved partoftheoriginalboundary as insidethedomainor along thepolygonalpartof theboundary. Numericalevidencewas givenin [14] and at thesame time, curved elements wereproposedwhichweresuitableforuse along thecurvedpartof the boundary. in [14] are closelyassociatedwithisoparaThe curvedelements constructed metricelements. Isoparametric elementswere first introduced by Irons [4] and theyare wellknownin thetechnical literature (see Strangand Fix [10]). The curvedelements are a naturalgeneralization of thetriangular elements. From the computational point of view,theymake it possibleto construct finitewhichare subspaces of the energy dimensional spaces of trialfunctions space of in arbitrary value problems secondorderboundary planedomains.The numerical thatbyusingthemwe can get and suggest results givenin [14] are very promising is a polygon thesame orderofaccuracyas in thecase whentheoriginal boundary thisquestionsee [11] and [3]). elements and we apply thetriangular (concerning The aim ofthisand ofa subsequentpaper (see thelast paragraphofthispaper)is theorem and to derive to provean approximation to justify thecurvedelements, for a W(2)-elliptic whichis theDirichlet a modelproblem error boundsfor problem theorderof thatthehypothesis boundsconfirm equation.These error concerning accuracyis correct. 2. Curvedelements. We considera bounded domain in the x, y-planewith a boundaryF. We assumethatF is sufficiently smooth, i.e.,thatF is ofclass Ck+ 1 ifF belongsto Ck+l 1 piecewiseand Theorems 1, 2, 3 are (in fact,it is sufficient validifwe add theword"piecewise")withk sufficiently largeto fulfill our requirements. In anycase we assumek ? 1.
* Receivedby the editorsDecember23, 1971,and in revisedform March 20, 1972.

of Brno, Obrancuimiru 21, Brno, Czechot ComputingCenter of the Technical University slovakia.This workwas done at the Mathematical Institute oftheCzechoslovakAcademyofScience in Brno. 229



along F by completed The elements intowhichQ willbe dividedare triangles of F of an arc and the P1P3 c segments elements whoseboundaryconsists P1P2, side and 9 P2P3. Here P1, P3 are two pointsofF and P2 Ec Q. Let h be thegreatest ourselves P1P2P3.It is easyto see thatifwe restrict thesmallest angleofthetriangle zero as h -+ 0, i.e., such that, is boundedaway from to elements (1)

> So,

S0 = const.> 0,

FIG. 1. Curvedelement

linesP2P1and P2P3 withthestraight thenthearc P1P3 has no otherintersections and together withthe segments than P1 and P3, respectively, P1P2 and P2P3 it ofthisJordancurvewillbe denotedby T and forms a Jordancurve.The interior called a curvedelement. trialfunctions Ifonlytriangular elements are applied,thenthecorresponding which are polyi.e., functions used most are piecewise polynomialfunctions, In case of second order boundary nomials of a certaintype on each triangle. are mustbe chosen such thatthe trialfunctions value problemsthe polynomials to insurethat as well as necessary continuouson Q. This conditionis sufficient of a such trialfunctions belong to the Sobolev space W(l)(i). The construction of polynomialin a triangleT (we use the same notationas forcurvedelements) of a suitablepolynomial an arbitrary shape can be reducedto the construction in the unit triangleT1 withvertices a-plane. R1(O, 0), R2(1,0), R3(0, 1) in the X, If (xj, yj),j = 1,2, 3, are the coordinatesof the vertices Pj of T, then the linear mapping
X = X0(A,l)
X1 + X-24 + X3J,

y = Yo(4,')

YI ? Y24 ? Y31l,


xj- x1,

Yj = Yj -Y,

= 2, 3,

T. Choosingsuitably T1one-to-one on theclosed triangle maps theclosed triangle a polynomial p(x,y) in T by setting r(c,i) in T1 we getthedesiredpolynomial (3) p(x, y) = r[4(x,y),n(x,y)],

where4 = 4(x,y), j = n(x,y) is theinverse mappingto (2). T theonlydifference in thepreceding construction In case ofa curved element one of theform is thatthelinearmappingwill be replacedby a nonlinear (4)
x = x(4, C) Y xo(4, C) + (1 - 4 - C)(5), + (1

= y(4,'C) _Yo(4,'C)




The mapping(4) will have the property that forh sufficiently small it maps T1 one-to-one on T, theJacobianJ(4,ri)of thismappingdoes not vanishon T1, the functions @(C), T(q) are smoothand 'D(O)= T(O) = 0. The function p(x,y) in (3) is, ofcourse,in generalnot a polynomial in T; however, it is a polynomial of one variablealong the sides P1P2 and P2P3. This property the factthat followsfrom themapping(4) is linearalong thesides R1R2 and R2R3 whichare mapped onto P1P2 and P2P3, respectively, and it will guaranteethe continuity of the trial functions across a commonside of two adjacent elements. We add thatin actual it is not necessary computations to carryout theinversion 4 = 4(x,y), i1 = q(x,y) 2 ofthenonlinear in [14]). mapping (4) (see Remark at theend ofthesecondsection the functions theseexplanatory remarks we can define After 4D(q)and T(11). The boundaryF can be dividedintoa finite ofarcs each ofwhichhas the number parametric representation x = (p(s), y = /(s), a < s < b, with functionsp(s), /(s) belongingto Ck+1 and such that at least one of the from zero on <a, b>. Let s, and S3 be the values derivatives s(s), /(s) is different oftheparameters We set P1 and P3, respectively. to thevertices corresponding

+ S3) -iXl

S3 =S3

() s.


+ S3t) -


Y ~~~~~1-i;


The point q = 1 is only an apparent singularity because we shall see that it is possibleto extend(D(q) and T(q) forrj = 1 in such a way thattheybelongto Ck forriE <0, 1>. We summarizeall properties of the mapping(4) in the following theorem. THEOREM 1. Let F be of class Ck+ 1, k > 1. If (1) holds and h is sufficiently is on T. The JacobianJ(d,ti) of thismapping small,then(4) maps T1 one-to-one on zero T1, theside R1R3 is mapped on thearc P P3, thesidesR1R2 differentfrom andR2R3are linearly mapped on thesidesP1P2andP2P3,respectively. The mapping are ofclass Ck. In addition, as wellas itsinverse mapping

(7) (8)

c1h2 <

)I ? IJ(4,


c1 = const. > 0, 1 < il < k,'

Dx(E, ,l) = 0(h111), DYy(d, i) = 0(h11)

D1d(x,y)= 0(h'), D?j(x,y) = 0(h-), lil= 1. Proof.We first extendthe definition of (D(i) and T(q) for q = 1 and find estimates forthesefunctions and theirderivatives. Both @D(q) and T(PQ) are of the formG(q) = g(q)/(l - q), whereg(tI)is of class Ck and g(O)= g(1) = 0. For such a function,


= (-1)j+ l(l Gli)(1l)



'1 (T (-


1)(T) dT9 = s.

j = O,*,


Evidently (9) is trueforj we get


0. Suppose thatitis trueforj

+ (-1)?+2(1

Integrating byparts, gs

L-gS+ 1)(r)

s s1


Here i

= (il, i2), lil = il + i2, D'u(x,y) = a1tIu/Ox810y2, D'v(4,')


232 Hence (s + 1)G(s)(,i)+

g(S+ 1)(jj)


( _ 1)s+2(1


{ (z

s + 1lg(s+ 2)() dz.

On theotherhand,by differentiating theequality(1 -q)G(tq) = g(1) we obtain


,)G(s+ 1)(,1) = (s + 1)G(s)(ti) +


and we see that(9) is truefor j = s + 1. Now from thatG(i)(t),j = 0, , k,can be extended (9) itfollows continuously mean value theorem we findthat forri= 1 because usingthefirst
lim G(j)(ij) =


From (9) we also get estimates of D(D()(i) and T(P)(q1), j = 0, then (S1 + S3A) - Xx3, (q)= =S(S1

k. If g(ij)

+ S3) 0S')l(S1


P(S1)] = S3[0(S1

+ S3)

(S1 + S30')]

+ S3f"),



Therefore Ifj > 2, then

g() (s23).

g(i( 1= (

3)jp(j)(Sl +


0( 3

Because eitherkI(s)l > m,m = const.> 0, or II(s)I > m on <a, b>, we easily see thatS3 = 0(h) and that the finalestimates whichfollowfrom(9) and whichwe writeforT(ij), too, are
(D(q) = 0(h2),

T(q) = 0(h2),

= 0(hi+1), T(j)(q1) j = 1, = 0(hi+1), ,k. @~~~(D()(q1) We now prove that the mapping(4) is injective. From the sine theoremit followsthat the smallestside of the trianglePjP2P3 is greater than or equal to > h sin 90and it holds that x + h sin 90.Therefore (11)
? coh or IY21_ coh, 1x21 with c0 =

(4) in thisway:

sin 9o. Let us consider the first case and write the firstequation in
X Xl X35 (1 -

ONt(O) = [X2


As 'D(t) = 0(h2) and Ix21> cohwe can compute4 from thisequation ifh is suffismall.Doing thisand replacing4 in thesecondequation in (4) we come to ciently an equation oftheform Now an easy computation, y = X(11). usingtheestimates (10) and (11), gives (12) where
X'(M)= Jo/X2+ 0(h2),
X2 X3

JO =








P1P2P3)whichmakes The absolutevalue ofJO is equal to 2 x (area ofthetriangle After some elementary calculationswe obtain it possibleto estimate IJo1. (13)
< IJolsin,90h2 2

# 0 for From (13) and (12) it follows(takinginto account Ix21 > coh) that X'(tq) 0 < t1 < 1 ifh is sufficiently smalland thisprovesthatthemapping(4) is injective. of1D(t) From the construction As it is continuouson T1 it is a homeomorphism. and T(t1)one immediately sees thatthe side R1R3 is mapped onto the arc P1P3, and thesides R1R2 and R2R3 are linearly mapped onto the sides P1P2 and P2P3, of a Jordancurve onto the A homeomorphism maps the interior respectively. of T1 onto T. (4) is a homeomorphism interior ofitsimage.Therefore theJacobianJ(4, ri)we easilycomputethat Concerning J(d,1) = JO+ 0(h3). get (6). Further, the (13) we immediately Taking again into account the estimate thattheinverse mapmapping(4) is ofclass Ck and as J(d,ri)# 0 on T1it follows (7) and (8) followfrom(10), (11) ping is also of thisclass. Finally,the estimates and (6). approximating subspaces.We considerall "triangula3. Finite-dimensional tions" of the givendomain Q into triangles completedalong the boundaryF by curvedelements such thatthe union of theirclosuresis Q and any two elements are eitherdisjointor have a common vertexor a common side. To every"trisideand 9 thesmallest h,9.his thelargest angulation"we associatetwoparameters: withthe angle of all triangles (in case of a curvedelementwe mean the triangle We restrict same vertices). ourselvesto "triangulations" such that9 is bounded zeroas h -- 0,i.e.,thecondition "triangulation" (1) is satisfied. On every awayfrom ofthiskindwe construct belonging a finite-dimensional subspaceoftrialfunctions to W"')(Q) and ifneeded satisfying the Dirichletboundaryconditionon a given part of F. On each elementthe trial functions are definedin the form(3), where is the inverse = (x, y),t1 = tq(x, mappingto themapping(4). This is truefor y) a triangular element, too, because in thiscase we can choose 'D(t) = T(1) =0 and For our ri)we choose a suitableclass of polynomials. (4) is equal to (2). For r(d, of the (2n - 1)th purposethe mostsuitableseemsto be the class of polynomials degree,n = 1,2, .., introducedby Koukal [5]. He proved that a polynomial by the values than 2n - 1 is uniquelydetermined m(x, y) of degreenot greater D'1(Pj), Iil ?n1, j=1,2, 3, DV1(Po), il < n - 2. ofthis ofgravity of a triangleand P0 is thecenter Here P1, P2, P3 are vertices Now a trialfunction by thevalues v(x,y) is uniquelydetermined triangle. (14) Iil < n- 1; DVv(Pr), D'V(POq), Iil < n- 2, are centers ofgravity and POq ofthegiven"triangulation" wherePrare all vertices we mean thepoint of a curvedelement ofall elements (by the "centerofgravity" of the triangle T1 in the mapping(4)). whichis theimage ofthe centerof gravity

- ~~~2




For each element T with verticesP1, P2, P3 we choose the values D1r(Rj), p(x,y) il ? n - 1,j = 1,2, 3, and DYr(Ro),il ? n - 2, such thatforthe function i(X, Y)], Y), r[N(x, < n - 1, j = 1, 2, 3, Jil D'p(Pj) = D'v(Pj),

= D1v(Po), il ? n - 2. D1p(PO) This is easy to do by differentiating theequalityr(,,ii) = p[x(,, ii), y(,,ii)]. in thisway belongsto W"2')(Q) The trialfunction as it assumes v(x,y) defined This follows from the thesame valueson a commonside oftwoadjacenttriangles. factthatit is a polynomialof one variablealong a side such as we noted before. Its degreeis not greater than 2n - 1 and, as the values Dv(Pr), )il _ n - 1, are assume the same function given, the polynomials to adjacent elements belonging in thedirection oftheside up to the(n - 1)thorderincluvalues and derivatives are identically remarks: siveat bothendsoftheside; hencethey equal. We add three (i) v(x,y) has continuousderivatives up to the (n - 1)th order inclusiveat all thecomputavertices ofthepolynomials Pr;(ii) theconstruction r(,, ij) is easyfrom can be done once forall polytional pointof view and almost all computations on a fixed nomialsbecause thesepolynomials are defined triangle, namelyon T1; the DirichletboundaryconditionvIr,= 0 on F1 c F, (iii) ifthereis prescribed to the then this condition means that the polynomialsr(E,qri) corresponding withone side lyingon F1 have to satisfy elements r(O, ri)0_ whichis again easy to fulfill. Let u(x,y) be a function We denoteby V' theclass ofall such trialfunctions. from W(')(Q), k ? n + 1. Accordingto the Sobolev theorem(see, e.g., [1]) it thereexiststhe interpolate belongsto C`- 1 and therefore Un(x, y) of u(x,y), i.e., thefunction fromVnsuch that
D1Uh(Pr)= D1u(Pr), DUh(Poq) = D'u(Poq),

? n - 1, Jil

Iil < n - 2.

norm of L2(Q) and W"2')(Q). As usual the norm in W(')(Q), s

Ifulr, = 0 and ifthe "triangulation" is chosen in such a way thatthearcs of the curved elementsbelong eitherto F1 or to F - Fl, then the interpolate also assumes zero values on F1. We want to estimatethe difference u - Un in the


0, 1,

,is defined


IiI?s Q


y)l2 dx dy.

We shall also use the seminorm


dx dy. JID'u(x,y)12

THEOREM2. Let thegiven"triangulations" satisfy (1), u(x,y) belongto W(k)(Q) and F be ofclass Ck+ 1 withk > n +- 1. Then


Unllj Q < ChsilulIls'Q,

j = 0, 1,

s = min (2n, k),

whereC is a constant which does notdependon u and h.2


the same in any two places. In the sequel C will denotea genericconstantnot necessarily






Proof.We restrict ourselvesto the case j = 1. If (15) is truefork ? 2n,it is certainlytrue for k > 2n, also. Hence we have to prove that IIu -UhII Q U1(,Q < Ch fork < 2n. Obviously it is sufficient to prove that IIU - U111T < Chk-IIUIIk,T for any element T We consider the functions u )

u[x(4, ii), y(', i)], Uh(, ii)


U'[x(4, ii), y(4, ii)] and using (6) and (8) we get


U'hlll,T -< CIIU

To estimate jl -Ujll we apply a lemma by Brambleand Hilbert[2] which forour purposewe formulate in thisway: Let Q be a boundeddomainwitha boundary ofclass C?. IfF(u) is a bounded
linear functional on W(k)(Q), IF(u)I
-< C1IIluIllk,Q'

and ifF(p) = 0 foreverypolynomialp of degreeless than k, thenthereexistsa constantK1 depending on Q onlysuch that
IF(u)I < KlCllUlk,Q

the norm of the quotient space W(k)(Q)/Pkl_, where Pk-

In thisformulation the lemma followsfromthe equivalenceof lUlkQ and of


all polynomials of degreenot greater than k - 1 (see, e.g.,[6, p. 112]). We choose forF thefunctional F(u) = (u - Uh, where w is a function )h1,Tl from ) Qdenotes W(21)(T1) which willbe determined laterand ( thescalarproduct in W(2l)(Q). Now UCis evidently theinterpolate ofa(d,ij) on thetriangle T1. Hence Uj(4, ij) is a linear combination of the values D1i(Rj), Iil? n - 1,j = 1,2, 3, and D1i(RO), Iij < n - 2. These values are bounded by CIP11k,T, according to the Sobolev theoremas we assume k ? n + 1. Therefore, T( C1akT1 and _ 1U1K we have + C lUlflk,Tj} ? CwK0)1,Tj1aIk,TlIF(a)I< 110W 11,TJ{11U1T1, ifu is a polynomialofdegreeless thank and k < 2n,theinterpolate Further of u on T1 is equal to u whichfollowsfrom the unicity of the interpolate on any Therefore triangle. F(r) = 0 forall polynomialsof degreeless than k and using stantK 1is an absoluteconstant becauseitdependson T1only.Setting o = uawe have
the lemma by Bramble and Hilbert we get IF(a)l ? K1CIIwI 1,TjIaIk,T1. The con-

denotes the subspace of



C?UIk,Tl. U~hl1,Tj <


Going back to the x, y-planewe easily findby means of (6) and (7) that whichtogether with(16) and (17) gives < Ch'k-1 11U1k,T
IIU ?1


Remark.It is easy to prove a bound analogous to (15) withthe maximum norm. 4. Application to a Dirichlet We applythesubspaces V" forsolving problem. the Dirichletproblemto the equation



Y -




+ g(X1,X2)U = f(Xl, X2)



element method.Concerning thecoefficients of the operatorLu and by thefinite the right-hand sidef(x1, x2) we assume that (19)

a12(x1,x2) = a21(x1,x2), a

i,j =1

aij(X1,x2)4ii ?


oc= const. > 0, g(x1, x2) > 0,

2(Q), f(x1,x2)e W?k-2)(Q).





If F is of class Ck, it followsfrom(19) and (20) that the solution u belongs to and W(2k)(Q)

(21) (see,e.g.,[1]).

llUllk,Q <

Cllf Ilk-2,Q

bilinearfunctional Let us denoteby a(v,w) the symmetric

a(v,w) = (22)



= 1

x2)8 ax. + X(X ,X2)VWdx1dx2

Ox xj

aw1 a~v

functional and It is a bounded W?")-elliptic


2 V11 11

< a(v,v) <M

V112n 11

all vEW0(1)(Q). for

fromC'(Q) Here W(2MP(Q) is the completionin the norm - 111 of all functions withcompact support. The solutionu satisfies (23) a(u, v) = (f,v)o, forall v E (1)0). The finiteelementsolution un is that functionfrom Vn (constructed so that v]r= 0 forall v E Vn) whichsatisfies (24)
a(un, v) =

(f,v)0Q forall vE Vn.

Our aim is to derivebounds in the L2-normand in the norm ,1 forthe discretization errore = u - U. THEOREM 3. Let thegiven "triangulations" satisfy (1), letthecoefficients ofthe Lu and theright-hand sidef satisfy operator (19) and (20) withk ? 2 and let F be ofclass Cr + 1, r = max(n + 1,k). Then


< Chs-il3f llelJIj, lls-2,n,

s = min (2n, k).

j = 0, 1,

Proof.Firstwe prove(25) forj = 1. In the usual way we derive


e 11 U I-IV 1,Q < C 11

for all vE Vn.

It followsfrom (23) and (24) that (27) a(e, v) = 0 forall v Vn. forall v E Vn, Hence unis the projection of u onto Vnin the norm lIv Therefore, II = [a(v,V)]112. a(e, e) < a(u-v, u-v) and using(22) we get(26).

to assume thatg(x1,x2) is measurableand bounded. 3 In case k = 2 it is sufficient






Ifk ? n + 1,we can set v = U' (accordingto a previousremarkU' assumes 1 < Chsl fls- 2Q. zerovalueson F) and (15), (26) and (21) giveimmediately j1ell We smoothu in the of smoothing. technique For k < n + 1 we use the familiar by the formula uf is given function way proposedby Strang[9]. The smoothed u(x) = h

2 u(y)p(x h

dy= h

u(x - Y)P(i) dy,

where x = (x1,x2), y = (Y1,Y2), u(y)C W(k)(R2) is the extensionof u and the in [8, Theorem2]: introduced function p has the properties smoothing (i) any polynomialof degreeless than k is preserved; ? 1. 1x12 (ii) p has compact support:p(x) = 0 whenever Strangprovedthatforu E W(k)(0), (28) ju - uljn _

j ?


namely, We add anotherestimate, (29)

iuijQ < Ch

j ? k.

u(x) and get To prove(29) we differentiate




y)p(YI) dy = h2

x Dfu(r)p(

) dr,

11 = k.

j again the second integral Differentiating DX(x) = hk- h-2i

that k timesone finds

r) dr,




from which(29) followsby squaringand integrating. ? lu- i/'l, + to (26). As ju - vl Now we return have


we also


IIell, < CIIu- l1Q + CIIVI v

u and forv itsinterpolate. approximating For / we wantto seta smoothfunction its interand consequently Howeverthe choice X = u is not possible because an, saves the remedy polate,do not assumein generalzero values on F. The following situation: / is thesolutionof (31) -AO = -Au, 1r = 0. Denotingao(v,w) As u E C'(Q) and F is of class C'+2, / belongsto W2(n+21(Q). dx1dx2we have SQ [V.1wXI + vX2wX2] f
ao(4, v) = (-Au , v)O0

ao(ui,v) forall v E W2 (Q).

Hence ao(u - t, v) = ao(u


v) forall v E W2 (Q),

and setting v = u - i we getju- (32) ju -

1 ?

1 a.

Using(28) we obtain

Ql 1?Q < ChklUlk,Q.



Further forthe interpolate TnP it holds accordingto (15) that ||/ith|T

< Chn ? Iill1


From (31) and (29) it followsthat

n cl 111n+1JQ -< CIIAU1Q Ulln+l,

? Ch


Hence (33)

I C1,Q?Ch

and from (30), (32), (33) and (21) we obtain Ile|1Q ? Chklllfll

This completesthe proof. The estimate intheL2-norm follows from theinequality (34)
< Ch11 e11 e11 11 O, ? 1,Q

whichwe proveusinga devicebyJ.Nitsche[7] (see also [8]). Let L? = e, P Then (35) Ilel 1Q


(e,L(p)o0Q = a(e, p) = a(e, 9

v) forall v E Vn

(thelast equalityis theconsequenceof(27)). We set forv the Ritz approximation fromVn whichminimizes 9h of q (i.e., the function a(9 - v,q - v) over Vn or, whichis the same, the approximatesolutionof the problema(u,v) = a(9, v) for all vE W1(Q)). From whatwe havejust provedabove it followsthat 1K< ChI11 2,Q. Using (35) and theestimateI PI12,n < ClIe ll0 we get
llellO ? _ Mllell1 Ql ?-h

p 1 1Q?-< Ch 1elK 1 1Q|I2,Q?<


and (34) followsby dividing by lel I0Q. Remark1. The best rate of convergence occursifk = 2n in whichcase


lle0ll o <

Ch2n 11 11 2n -2,Q

< e II,1 ? II


- 2,Q. 2n

For practicalcomputations thespace V2 is especiallysuitable(see [12], [14]) and yetsimple,and we have a sufficiently highaccuracy: (37)

< ChIlIf


U -

|| Chhf12Q

1> x Remark2. If, in addition to the assumptionsof Theorem 3, h*h0 const.> whereh* is thelength ofthesmallestside ofthetriangulation, then the following bound in the maximum normis true:


' maxIe(x,y)l? ChSQ

f lIIf-2,Q,

min (2n,k).

This estimateis, however, certainly not the optimal one. We sketchthe proof: Firstconsiderthe class Pmof polynomials of the givendegreem on a triangle To






withthe greatest side ho. As before, let f(Q,'1) = p[x(4,ii), y(4,ii)]. Then = maxlIP < CO1IIPIO,T1 < Ch-1IIPIIOTO maxipI To T, (COdependson monly).Now letk ? n + 1 so thatthesolutionhas theinterpolate in a pointofthetriangle To, we have Uh. Ifle(x,y)ltakes themaximum
Iel < maxIu maxIel = mrax To To

Ufl + maxIU To

- ul.

side can be easilyestimated termon theright-hand The first (in a way similarto thatwe used to proveTheorem1) and thebound is Chs-1II'U s,To. Concerning the othertermwe use the above assertionand get
Uh UhI ?



? U,IT ChT[o


To +

h Oo

< Ch

u 1


maxIel? Chs-1IlullsQ

< Chs- 1 lf Ils-2,Q.

Uk-2 If k < n + 1,we considertheinterpolate and repeatthepreceding proof. we must If we want to use the curvedelementsin practicalcomputations, 9(s) and qi(s) and theirdeprepare subroutinesfor computingthe functions In [14] anotherapproach was introduced which avoids such inputs. rivatives. fora new boundaryF* whichapproxin constructing curvedelements It consists F. The idea is simple:We replacethefuncimatespiecewise theoriginal boundary tions + S3j), (S1 + S3r) byLagrange or Hermite interpolation polynomials 9p(s1 functions ofdegree2n - 1. The corresponding D*(,i),T*(,i) in (4) are (P*(J7), V*(q1) results derivedin [14] by means of of degree2n - 2. The numerical polynomials curved elementswere actually computedin this way. If we denote by V* the finite finite-dimensional element corresponding space and byu* thecorresponding to thisquestionas wellas we shouldfind bounds foru' - u*. We return solution, in a subseof elementstiffness matrices to the problemof numerical integration quentpaper.

The authoris indebtedforhelpful comments to Professors Acknowledgment. J. Descloux and G. Strang. schemes of a highorderof accuracy Addedinproof. Triangular finite element are proposedforcurveddomains and errorbounds are introduced in the paper schemesof a high by V. G. Korneev, The construction of variational difference orderof accuracy,VestnikLeningrad.Univ., 25 (1970), no. 19, pp. 28-40. (In studiedby P. G. Ciarlet and R. A. Russian.) Curved elementsare extensively Raviart in the paper, The combined of curvedboundaries and numerical effect methods (presented at the Symposium integration in isoparametric finite element on the MathematicalFoundationsof the Finite ElementMethod, held in June of Marylandin Baltimore). 26-30, 1972,at the University



Princeton, N.J., 1965. on Sobolevspaces with applicafunctionals oflinear [2] J.H. BRAMBLE AND S. R. HILBERT, Estimation 7 (1970), pp. 113-124. and splineinterpolation, thisJournal, tionto Fouriertransforms in thefinite element Math. Comp., method, elements [3] J. H. BRAMBLE AND M. ZLAMAL, Triangular 24 (1970), pp. 809-820. AIAA J., in stiffness methods, integration applicationof numerical [4] B. M. IRONS, Engineering 11 (1970), pp. 2035-2037. and theirapplicationsto partial differential [5] S. KOUKAL, Piecewise polynomialinterpolations SbornikVAAZ, Brno, 1970,pp. 29-38. (In Czech.). equations, Academia, Prague, 1967. en theorie des equations elliptiques, directes [6] J. NECAS, Lis methodes Numer. Math., des RitzschenVerfahrens, fur die Quasi-Optimalitdt [7] J. NITSCHE, Ein Kriterium 11 (1968), pp. 346-348. rate of variationalof the convergence [8] L. A. OGANESJAN AND P. A. RUCHOVEC, Investigation domain witha schemes for ellipticsecond orderequationsin a two-dimensional difference smooth Z. Vychisl.Mat. i Mat. Fiz., 9 (1969), pp. 1102-1120.(In Russian). boundary, in thefinite element Numer.Math., 19 (1972), pp. 81-98. method, [9] G. STRANG, Approximation Englewood [10] G. STRANG AND G. Fix, An Analysisof theFiniteElementMethod,Prentice-Hall, N.J.,to be published. Cliffs, Numer. Math., 12 (1968), pp. 394-409. element method, finite [11] M. ZLAMAL, On the value problems, procedures for solvingsecond orderboundary [12] , On somefiniteelement Ibid., 14 (1969), pp. 42-48. A finite element of thesecondorderof accuracy,Ibid., 14 (1970), pp. 394-402. procedure [13] Internat. J. Numer. Met. , Thefinite in domainswithcurved method boundaries, element [14] Engrg.,to appear.

on EllipticBoundaryValueProblems, Math. Studies,No. 2, Van Nostrand, [1] S. AGMON, Lectures