Numer. Math.

25, 215--229 (1976)

9 by Springer-Verlag t976

Sard Kernel Theorems on Triangular Domains

with Application to Finite Element Error Bounds*
Robert E. Barnhill and John A. Gregory

Received September 12, 1972

Summary. Error bounds for interpolation remainders on triangles are derived by

means of extensions of the Sard Kernel Theorems. These bounds are applied to the
Galerkin method for elliptic boundary value problems. Certain kernels are shown to
be identically zero under hypotheses which are, for example, fulfilled by tensor product
interpolants on rectangles. This removes certain restrictions on how the sides of the
triangles and/or rectangles tend to zero. Explicit error bounds are computed for piece-
wise linear interpolation over a triangulation and applied to a model problem.

1. I n t r o d u c t i o n

In this paper, a Kernel Theorem of Sard [8] is extended to construct error

bounds for interpolation remainders defined on triangles. The Kernel Theorem
provides an exact representation of linear functionals which are admissible on
spaces of functions with a prescribed smoothness. The theory has application to
the finite element analysis of elliptic boundary value problems, since the inter-
polation remainder is an upper bound on the finite element remainder in the
energy norm (see Section 2).
The theory of Sard is well suited to the calculation of interpolation remainders
defined on rectangles. Birkhoff, Schultz, and Varga [6] use the theory to derive
bounds for tensor product Hermite interpolation. We show that the theory c a n
be extended to treat triangular and other domains (see Section 3). This theory
provides a constructive method of computing the constants in the error bounds,
which the Bramble-Hilbert Lemma approach does not yield.
In Section 4, we prove a Zero Kernel Theorem which states sufficient conditions
for certain of the Sard kernels to be identically zero. This theorem implies that
finite element remainder functionals do not involve all possible derivatives of a
certain order, and this permits avoidance of mesh restrictions. In particular,
tensor product interpolants satisfy the hypotheses of the Zero Kernel Theorem
(see Section 5) and thus the mesh restrictions in Birkhoff, Schultz, and Varga are
not necessary. Lagrange interpolants for triangles also are covered by the Zero
Kernel Theorem, in Section 6.
* The research of R. E. Barnhill was supported by The National Science Foundation
with Grant GP 20293 to the University of Utah, the Science Research Council with
Grants B/SR/9652 at Brunel University and B/RG/61876 at Dundee University, a
N.A.T.O. Senior Fellowship in Science, and The University of Utah Research Com-
mittee.
t7 Numer. Math., Bd. 25
216 1R. E. Barnhill and J. A. Gregory

We conclude with a computed example of the constants in the error bounds

for piecewise linear interpolation in Section 6 and their application to a model
elliptic problem in Section 7.

2. Interpolation Remainders and the Galerkin Method

In Sections 5 and 6 we consider bounds for interpolation remainders of a
bivariate function F (x, y) defined on the Sobolev space W,n (f2), where D is a
rectangle or a triangle. More generally, let s be a simply connected bounded
region that satisfies a restricted cone condition in the xy-plane. The Sobolev
space W,n (~2), I ~ r ~ ~ , integer n ~ 0, is the space of functions such that all
generalized derivatives of order ~ n are in L, (f2). A norm for W,n (f2) is

01~1
where 0c=(~i, ~), I~l ---~1 q-~2, and D ~ - 0x~,aya, .
Let P be an interpolation projector defined on FE W,'(~2). The remainders of
interest in the Sobolev space are
cqh+k
Rh, k[F(x'y)] -- OshOtk R[F(x,y)] O<h+k<n, (2.2)
where
R iF(x, y)] ~F(s, t) --P [F(s, t)]. (2.3)
For fixed (s, t) (2.2) and (2.3) define linear functionals on F (x, y)E W,* (12).
If the interpolation function is piecewise defined over a subdivision of a poly-
gonal region ~2 into a union of disjoint triangular elements f2e, then each element
can be considered separately since
"
IIFII , .,j , 1 <_,<_ (2.4)
Interpolation remainder theory has application to finite element remainder
theory. Following Varga [9], we consider linear elliptic operators in divergence
form:
Lu(x,y) = Z (--t)I~IDe[p~(x,Y) D~u(x,Y)] (2.5)
I~l<~
where the p~ are in Loo(12). The nonhomogeneous boundary value problem
corresponding to L is to find uE W2n(12) such that:
Lu(x, y) =g(x, y), (2.6)
Dgu(x,y) =l~(x,y), (x,y)eO0 for O=<[fll ~ n - - l . (2.7)
The homogeneous problem is that all the /g are identically zero, the relevant
O O
Sobolev space then being called W2* (s A norm in W," (~) is

Let
a(u, v) = ~. f f p,(x,y) D"u(x,y)D'v(x,y) dxdy. (2.8)
 Sard Kernel Theorems and Finite Element Analysis 217

Then the weak problem corresponding to (2.6) and (2.7) is to find u satisfying (2.7)
and such t h a t
o

a (u, v) = (g, v) for all v in W2* (Y2). (2.9)

We consider interpolants ~ to u, where the interpolation conditions are the
following :
L~(~) =L~(u), i = t . . . . . I,
Mj (~/) = Mj (u), i = t ..... J,
and the L i and M i are interpolation functionals such t h a t the L i (u) are unknown
and the Mi(u ) are known a priori from (2.7).
Let V h be an (I + J ) - d i m e n s i o n a l subspace of VV2*(f2) such t h a t the L i and M i
are linearly independent over V h. Then V h has a basis of functions {B~ (x, Y)}~=I
and {Ci(x, Y)}~=1 t h a t are b i o r t h o n o r m a l with respect to the L~ and M i E2]. L e t
S h be the subset of W~n (f2) which consists of functions U of the form
/ y
U (x, y) -~ Y~ a,B~(x, y) + Y, Uj(u) Ci(x, y)
i=i i=i

where the a i are numbers. L e t sho be the m-dimensional subspace generated b y

the B~. The Galerkin method is to find U in S * such t h a t
a (U, v) = (g, v) for all v in Soh. (2.t0)
Under the a s s u m p t i o n that
o
S h0 "~- W2/o~
2 ~l,
the following l e m m a applies:
L e m m a 2.1. The Galerkin a p p r o x i m a t i o n U is the best approximation from S h
to u in the energy n o r m induced b y the inner product a (u, v). T h a t is,
a(u--U,u--U)~a(u--a,u--a) forallainS h. (2.11)
In fact,
a(u-- U, u-- U) +a(~-- U, ~-- U) =a(u--~, u--~). (2.12)

3. Sard Kernel Theorems and Interpolation Remainder Theory

The Kernel Theorems of Sard [8] involve Taylor series expansions of a function
F(x, y) about a point (a, b). These expansions have a rectangular domain of
influence a n d Sard restricts the theory to function spaces defined on rectangles.
However, the theory can be extended to bounded regions /2 which satisfy the
following p r o p e r t y :
Property 3.1. There is a rectangular coordinate system and a point (a, b)E~
such t h a t for all (x, y ) E ~ the rectangle with opposite corners at (a, b) a n d (x, y) is
contained in ~, where the sides of the rectangle are parallel to the axes x = 0 and
y = 0 of the rectangular coordinate system.
I f / 2 is a rectangle, then (a, b) can be an a r b i t r a r y point in the rectangle. If
is a triangle, then (a, b) can be the point on the longest side which is at the foot
of the perpendicular to this side from the opposite vertex. Alternatively, for a
right-angled triangle the v e r t e x at the right angle can be t a k e n as (a, b).
17"
2t 8 R.E. Barnhill and J. A. Gregory

Sard Taylor expansion. Sard defines a function space "boldface Bp,q (g2)" of
functions for which a certain type of Taylor expansion exists, where p and q are
positive integers and p + q ----n. The Taylor series expansion involves the triangular
scheme of all derivatives of order less than or equal to n executed in a certain
manner, see Fig. 3.t. The Taylor expansion is presented in the following theorem
for FE Cn ((2) and the space boldface Bp,q is then developed in Corollary 3.t.
(0, n)
N N N ~i ) (ni-, i)

(O,q) I ~ ~(p,q)
\
~ )

(o,o) ! ~ ~ : _ (n, 0)
Fig. 3.1. Index triangle of partial derivatives

(x, y)

(x, ~) (a, ~)
(~, ~)

(a, y')

(x, b) (a, b)
(~, b)
Fig. 3.2. Function arguments in Taylor expansion

Theorem 3.1. Let T'EC"(g2) where g2 is a region which satisfies Property 3.t.
Then F has the following Taylor expansion at (x, y) about (a, b):
F(x, y) = ~, (x--a)CO(y--b)r b)
i+j<n
x

+ Z (Y--b) r f (x--s162163 b) ds
~<qy
a x
(3.t)
+ f (y_ y)(q-l~ f (x_ ~)r ~,q(~, y) d~ dy
b a
Y
+ Z (x--a) r f ( y _ y)t.-i-x)F~,._i(a, :9)dy,
i<p b
 Sard Kernel Theorems and Finite Element Analysis 2t 9

where p and q are positive integers such that p + q = n and (x--a) Ii)=- (x--a)i/il
etc.
Proo/. For FEC~(~2) the following single variable expansions can be obtained
by integration by parts (cf. Figs. 3A and 3.2)
q--I y
F(x, y) = X (Y--b)I'~ Fo,i( x, b) + f (y--y)(q-~) Fo,q(x, y) d5, (3.2)
p--1
Fo,~(~, ~) = Z (~-a)"~ F,,da, Y) + f (x- ~)~P-~Fp,~(~, Y) d~, (3.3)
i~O a

n--j-1
~,j(~, b) = E (x--a) {i) Fi, i(a, b) + f (x-- s F~_i.i(s b) ds (3.4)
i=o
n--i--1
~ , d ~, Y)= jE= q (33-- b){i-q} Fi,i(a, b) + [ (y_y,)r ~,,,_~(a, y') dy'. (3.S)
b

The Taylor expansion (3.t) is derived by combining Eqs. (3.2)-(3.5), where, by

integration by parts,

f ( y _ y)(q-1) / ( y _ y , ) { p - , - s ) F~,,,_, (a, y') dy'd y

b b
Y
= f (y--y)(~-~-tlFi,,_i(a, y) dy. Q.E.D.
b

Corollary 3.1. Let F satisfy the following properties on ~2:

(i) F(x, y) is q--I times continuously differentiable with respect to y and
Fo, q-1 (x, y) is absolutely continuous with respect to y.
(ii) Fo, q(x, #) is p--1 times continuously differentiable with respect to x and
F~_x,q(x, ~) is absolutely continuous with respect to x, almost everywhere 3Z
(iii) Yo, i(x, b) is n - - j - - i times continuously differentiable witk respect to
x and F,,_~_l,i(x, b) is absolutely continuous with respect to x, 0----<7"~ q - - t
(iv) F~, q(a, y) is n - - i - - t times continuously differentiable with respect to y
and Fi,,_i_l(a, y) is absolutely continuous with respect to Y.
Then the Taylor expansion (3.t) exists.
Proo[. The properties (i)-(iv) are required for the existence of expansions
(3.2)-(3.5). The property (ii) need exist only a.e. y since it is required under the
Lebesgue integral in (3.2). Q.E.D.
Remark. In Corollary 3.1 the differentiations are performed according to a
particular ordering rule consistent with the expansion (3.2)-(3.5), namely
F i i __ /-)1-1a /-p D ~ o F

where D~-~ 8/8 x, Dy ~ ~]ay and io ~ rain (i, q).

Properties (i)-(iv) are an equivalent definition of the function space boldface
Bp, q (Y2) to that in Sard E8, p. 172], when ~ is a rectangle.
220 R.E. Barnhill and J. A. Gregory

Corollary 3.2. Let FEW,n(I2), I ~ r ~ oo. Then the Taylor expansion (3A)
exists a.e. (a, b) where the derivatives are now generalized derivatives.
The Taylor expansion (3.t) can he amended to an expansion involving definite
integrals by the following device: Let

max_ (x) -----~, r a i n (x) =_~

(x, b)r (x, b)E.O
(3.6)
m a x (y) =/~, m i n (y) -----
(a, y)e~ (a, y)e~

and let the function ~o be defined by

if a ~ s
(a, ~7, x)= - if x<=s (3.7)
otherwise.

Then the Taylor expansion (3.1) can be written as

F(x, y)= E (x--a)(O(Y--b)(i) ~,i(a,b)

i+i<n

+ ~, f (x--~)("-i-~)~(a, ~2,x)(y--b)~F,,_j,i(~, b) ds
i<q ~_ (3.8)
+ yf (x-- s s x)(y--y)Cq-s)~o(b, y, y)F~,q(s y) ds dy

+ ~. f (x--a){O(y--y)C"-i-1)~o(b, y,y)Fc,n_i(a, 37) dy.

i<p

Sard Kernel Theorem. The Sard Kernel Theorem applies to a class of ad-
missible functionals defined on the space boldface Bp, r as follows:

Definition 3.1. The admissible functionals R on boldface Bp,q(f2), where f2

is a region satisfying Property 3.t, are of the following form:

R[F(x, y)]= E f f Fi,i(x, Y) d#~'i(x, Y)

i<p t2
i<q

+ ~. fFi,i(x, b) dtd'i(x) (3.9)

i+i<n ~_
i>p

+ ~. /Fi,j(a, y) dt*~'i(y),
i+i<n fl
i>q

where the/,~' i are functions of bounded variation and _~, ~, ~, and • are defined by
(3.61.
Example. Let R IF] = ~ , i (c, d), (c, d) E12.
Then R IF] = f f ~ , i (x, y) d/~ (x, y)
o
where
/ , ( x , y ) = { t 0 c<x, d<y
otherwise.
 Sard Kernel Theorems and Finite Element Analysis 221

Theorem 3.2 (Kernel Theorem). Let fl satisfy Property 3A and let R be an

admissible functional on boldface Bp, q(/2). If F (x, y) ~ boldface Bp, q(~2), then
R [F] : E c~'i F~,i (a, b)
i+j<n

+ F, f Kn-i'i(2) F~_/,i(s b) ds
i < q ~_
+ f f KP'q(2, y) Fp,q (2, y) d2 d y (3.1o)
D

+ E ]g"n-'(Y) F~,~_~(a, y) dy,

i<p _~
where
ci'i=R(x,y)[(x--a)(i)(y--b)(i)], i +i <n, (3.11)
K,,-i, i (2) = R(.,y) [(x-- 2) ('-i-x) ~p(a, 2, x) (y-- b)~i)], J < q, 2r x, (3.12)
Ki' "-i (9) = R(.,y) [(x-- a) (0 (y-- y)(.-i-x) y2(b, y, y)],
i<p, yCJy, (3.t3)
KP'q(s y)=R(,,,y)[(x--i)(P-1)~p(a, 2, x)(y--y)(q-1)~p(b, Y, Y)], 2r yCjy.
(3.t4)
The notation R(,,y) means that R is applied to functions in the variables x and y.
J x is the jump set consisting of the points of discontinuity of the functions of
total variation [#~-i-1, i] (x), j < q, and J y is dual. f x is the jump set consisting
of the points of discontinuity of [#~-Li' I (x, y)evaluated at y (x) ----max {y},
i' < q, and J y is dual. (~,y)~a
Remark. The jump sets are the points at which the kernels (3.t2)-(3A4) are
undefined, but these sets are of measure zero with respect to the Lebesgue-
Stieltjes integrals in (3 A 0).
Proo/ o/ Theorem 3.2. The application of the functional R defined by (3.9) to
the Taylor expansions (3.8), and the use of Fubini's Theorem to change the order
of integration, give the desired result. Q.E.D.
Remark. If the functional is not admissible, then it can be applied to the
Taylor expansion (3A) directly.
Corollary 3.3. Let FE W,n(/2), t --<r ~ oo. Then the Kernel Theorem 3.2 is true
a.e. (a, b) where the derivatives are now generalised derivatives.
Remark. For the interpolation remainder functionals (2.2) the corresponding
kernels (3A2)-(3A4) depend on s and t and will be written:
Kn-i'i(s, t; 2), Ki'n-i(s, t; y), KP'q(s, t; 2, y). (3A5)
Also, if ~ = p + q is chosen such that ~P~-I -~{xiYi/0 <=i +j <=n--t} is contained
in the precision set of the interpolation operator P, then the ci, i in (3.t t) are zero
for the remainder functionals Rh, k.

4. Zero Kernels
Theorem 4.1 (Zero Kernel Theorem). Let R be a linear operator which maps
F (x, y)Eboldface Bp, q(/2) to functions of (s, t) such that R is an admissible linear
222 R.E. Barnhill and J. A. Gregory

functional for fixed s and t. Also let

~+ h
Rh,k[F(x, Y)] -- ~sh~tk R[F(x, y)], h, k > 0 , (4.1)
be an admissible functional for fixed s and t. Let Pi (x) be a polynomial in x of
degree _<i and qi (Y) be a polynomial in y of degree =<i. Then the Sard kernels for
Rh,~ have the property that
K~'n-~(s,t; 3~)= 0, O<=i<min{h,p}, (4.2)
if
R [Pi(x)g(y)] =r(s, t) (4.3)
is such that r (s, t) considered as a function of s alone is a polynomial of degree
< h. Dually
K"-i'i(s, t; ~) = 0 , 0 _--<i< rain{k, q}, (4.4)
if
R El(x) qi(Y)] =-r(s, t) (4.5)
is such that r (s, t) considered as a function of t alone is a polynomial of degree < k.
Proo I. It is sufficient to prove (4.2) as (4.4) is a dual result. Let i be an integer
such that 0 ~ i < r a i n (h, p). Then the kernel K~"-i(s, t; ~) corresponding to R is
K~ '*-~(s, t; y) = R(~,y)[(z--a) (~(y--y)('-i-*)y~ (b, y, y)], yr ]y.
which, considered as a function of s, is a polynomial of degree < h, by the hypo-
thesis (4.3). Thus the kernel Ki"~-i(s, t; y) corresponding to Rh, ~ is
~h+k .
K~'S-~(s't;Y) - ~sh~tk K~"-~(s,t;Y) =0, O<=i<min{h,p}. Q.E.D.
Schematically, the possibilities for the domain of influence of a functional
Rh, ,, which satisfies the hypotheses of the Zero Kernel Theorem in the Sard
space boldface Bp,q, are shown in Fig. 4A.

(h, k) (h, k) x~ ~ ( p q)

Fig. 4.1. Non-zero kernels in the Zero Kernel Theorem

Remark. Although not stated in the Theorem, the c~'j, (3.1t), are zero for
0 ~ i < min (h, p), 0 ~/" < min (k, q).

5. Error Bounds for Tensor Product Interpolation

We next show that the kernels (4.2) and (4.4) are always zero for tensor
product schemes with sufficient polynomial precision.
 Sard Kernel Theorems and Finite Element Analysis 223

Theorem 5.1. Tensor product interpolants of polynomial precision of at least

h - - t in the variable x and at least k - - t in the variable y satisfy the hypotheses
(4.3) and (4.5) of the Zero Kernel Theorem.
Proo/. P a tensor product interpolation operator implies that P is of the form
e = P, P~ =P~P, (5.1)

where P, is an operator in the variable x and Py is dual in y. Therefore, if F (x, y) =

pi(x) g(y) where Pi(x) is a polynomial in x of degree i < h, then
P [p~(x) g(y)] =PyPx [p~ (x) g(y)] =Py [pi(x) g(y)]
= p (x)P Eg(y)3 y).
r(x, y) satisfies (4.3), so that (4.2) follows. The argument for (4.5) is dual.
Q.E.D.
Birkhoff, Schultz, and Varga derive error bounds for tensor product Hermite
interpolation using the Sard Kernel Theorem. Here, P, is defined b y
N N
P, EI(x)J = Y + Y. (5.2)
i=O 4=0

where the ~bi(x) and ~ i ( x ) = ( - - l ) i ~ i ( l - - x ) are the cardinal basis functions for
Hermite two point Taylor interpolation on [0, 1]. Pv is dual and the resulting
tensor product interpolant is on S = [0, t ] • [0, t ]. An important observation is
that the point (a, b) of the Taylor expansion can be chosen as the point (s, t) of the
remainder functionals Rh, k (cf. Property 3.1) where Rh, ~ is defined by (2.2). The
remainder functional Rh,k is then admissible in boldface B~, m(S) where m = N + 1
and 0 < h + k < 2 N + t. (Rh, k is precise for the set ~ N + I . ) An analysis similar
to that in Birkhoff, Schultz, and Varga using the Sard Kernel Theorem, but with
the application also of the Zero Kernel Theorem then gives that
2m--rain(h, m)
IIR,, fjll ,r y, Cr O<=h+k<2m, p ~ r (5.3)
/'=rain(k, m)

where the C i are constants (see Gregory [7] for further details). The choice of
(a, b) = (s, t) enables the derivation of the above bound in W,s'~ (S) which is not
possible in application of the Sard Kernel Theorem on a triangle (cf. (6.2)).
The summation over the range min(k, m)<~i~2m--mirt(h, m) in ($.3) is a
consequence of the Zero Kernel Theorem. A change of variable leads to the bound
2 m--min(h, m)
C H s~-i-h+Up-Ur K i-k+llp-xlr F, ~
i=mm(k, ~) (5.4)
O<=h+k<2m, p ~ r .

for the Hermite interpolation remainder Rh,k for the function F defined on S =
[0, HI • [0, K]. Now, when h, k <=m(h + k # 2 m ) the summation in (5.4) is over
k <~ ~ 2 m - - h and the exponents of H and K are then greater than or equal to
zero and not simultaneously zero. Thus negative exponents of H and K are not
possible in this case, which removes the need of a "regular" mesh restriction.
224 R.E. Barnhill and J. A. Gregory

6. Error Bounds for Interpolation on Triangles

Let P be an interpolation operator on the standard triangle T with vertices at
(t, 0), (0, t), and (0, 0), which is precise for the set ~ - 1 but not all of ~ . Consider
the remainder functional

R CF (x, Y)] - - F (s, t) -- P EF (s, t)].

A point (a, b) which satisfies Property 3.t on the standard triangle T is (a, b) =
(0, 0). Consider the Sard space boldface Bn_m,r~(T ), where m = En]2]. Suppose R
is an admissible functional on boldface Bn . . . . (T). Then application of the Sard
Kernel Theorem and the precision of the interpolation operator P give that
l
R f l = X f K"-J,;(s, t; ~)F~_;,j(~, 0) g~
i<m 0

+ff K"-","(s, t; ~, y)F~:,~,,,(~, y)a~ a:~ (6.1)

T
1
+ Y~ fK~'"-~(s,t;y) F~,,,_~(O,y)dy.
i<n--m 0

From (6.t) it follows by the triangle inequality and H61der's inequality that

[IR EC lk,<,,,,-~ ,X J[IlK"-i,+(s, t; ~)k:+ll~<,,,~ll~-;,;(~, o) IL,<~,

j<ra

+ll IlK"-=,"(s, t; ~, ~)[L:(w,w,li~,(,,,)ll~-m,=(~, Y)[L.(~, ~) (6.2)

+ Y, II[IK', "-' (s, t; y)IL:,~,IL,~,,)IIF,,,,_,(o, y)~,,~,,
i<n--m

where Lt~(s, t) denotes the Lp norm over the triangle T with respect to (s, t),
L,I (~) denotes the L,I norm over [0, t ] with respect to ~, etc., and t / r 1 + t/r'l = t,
t / r z + l / r ' ~ = t . The norms of the kernels in (6.2) are constants which can be
estimated. However, (6.2) is not a bound in the Sobolev space WrY(T) because of
the presence of the univariate norm terms. The device of taking (a, b) = (s, t)to
obtain Sobolev space results is not possible here, as it is for the rectangle, since
the rectangular domain of influence of the Sard Taylor expansion would then go
outside the triangle.
Remark. For x, ~E T, (a, b) = (0, 0), (3.7) becomes

{t0 ~<x
~o(0, ~, x) = ' otherwise.

Thus the functions which occur in the kernels can be expressed in terms of the +
function as follows:
--~)C0 if ~ < x ,
I(x
(x--~)10~(0, s x) = ( x - - , ~ ) ~ = [0 otherwise, (6.3)
etc.
The treatment of the functional
~h+h
Rh, k[F] -- OshOt* R [ F ]
 Sard Kernel Theorems and Finite Element Analysis 225

is best considered in relation to the particular interpolation operator P. Firstly,

one must consider whether the functional is admissible, and secondly whether or
not the Zero Kernel Theorem, is applicable. In the particular example of linear
interpolation considered below, the Ri,o and R0,1 functionals are not admissible;
however, results are derived by direct application of the functional to the Sard
Taylor expansion. One example of an interpolant for which the hypotheses of the
Zero Kernel Theorem holds is given in the following theorem:
Theorem 6.1. Let P be the interpolation operator defined by the Lagrange
polynomial of degree N which interpolates F(x, y) at the nodes (x i, Y i ) = ( N '
~),O~i+j~N,i.e.,
P IF(x, y)] -- Y. p~,i(x, y) F(x~, Yi), (6.4)
O~i+i~_N
where
t i--i i j--i I ~--i
P,,i(x, y ) = i! H
~=0
[I
,=0
II
v=0
(6.5)

and i +l' + k = N . Then P satisfies the hypotheses (4.2) and (4.5) of the Zero
Kernel Theorem, where the i and I" of that theorem satisfy i < h ~ N; ~"< k ~ N.
Pro@ We show that P [xig(y)l is a polynomial in x of degree = i for the
nontrivial case 0 ~ i < N, from which the conclusion follows. Now,
N N--j'
P~x~g(Y)] = E E Pi',i'(x,Y) @g(Yj')
i'=0 i'=0
N (6.6)
---- Z ai" (x, y) g (y,.,),
i'=O
where
N--l"
a,,(x, y) = Y, pi, i,(x, y) xi,.i (6.7)
i' =0

The cardinal function pv,i,(x, y) is a polynomial of degree N - - i' in x. Thus for

j' > N - - i , ai,(x, y) is a polynomial in x of degree ~ i . Now for O ~ j ~ N - - i - - 1 ,
N--i
(6.6) with g(y) ----gi(Y) ~ M (Y--Y,) gives

N
P [xigi(Y)] = Y' gi(Yi') ai'(x, Y)
i'=O
N
=gi(Yi) ai(x, Y) + E gi(Yi') aj,(x, y)
i'=N--i+I

= xi gi(Y), by the precision of P.

The last two steps of the above equation cart be considered as an equation in
a i (x, y), from which it follows that a i (x, y) is a polynomial in x of degree _ i,
0 < j-< N - - i - - t . The proof for _P [/(x) y ] is dual. Q.E.D.
Example o[ linear interpolation on the triangle T. The following results are
generalizations of those given in 13arnhill and Whiteman [5]. Further details
including calculations for quadratic interpolation on T can be found in Barnhill
226 R.E. Barnhill and J. A. Gregory

and Gregory [2]. Consider the linear interpolation remainder functionals:

R IF] =F(s, t)-- IF(t, 0) s + F ( 0 , t)t + F ( 0 , 0) (1 --s--t)]
Rl,o[F ] =F~,o(s, t) + F ( 0 , 0 ) - - F ( I , 0) (6.8)
Ro,~[F ] =Fo,x(s, t) + F ( 0 , 0)--F(0, t)
defined on the Sard space boldface BI,I(T ). The functional R is admissible on
boldface B1,1 (T) and thus from (6.2) it follows that
[[R[F][]L,,(,,,)<=CII[F,~o(e, O)[L,(.;)+C;,][F~,I(e, y)[IL,,(,;.,;) (6.9)
+ G IlFo,~(o, y) L,(;),
where
C 1 = [[ IlK 2'0 (s, t; e ) I L l (w)L<~,*,
: [I[IK~ s, t; ~)ILr (6.1o)
c~ = [l IIK~' ~(~, t; ~, y)IL: (~,~)IL (,,,,,
t/r~ + l/rl = t, t/r~ + 1/r'~= t, and
K%~ t; e) :K~ s; e) = R [ ( x - e ) + ] : ( s - e ) + - s ( l -e),
(6.11)
K 1' 1 (s, t; e, y) = R [(x-- e)~ (y-- y)~_] ----(s-- e)~ (t-- 9)~-.
Careful evaluation of (5.10) yields
/ ( ~, 1X ~ ) '1/" { B (p +1 , p+2)}'~P, r~<o% p<o%
c1 /1 / I y,, , (6.12)
(~-kr~T~] ' r1<_-o% p = o %

{S (P14 + 1, Pit'2+ 2)}'/P, fi, p < oo,

{• ,
C2 = \4] ' r$< oo, p= oo, (6.13)

where B is the Beta function.

The functional R1, o and its dual Roa are not admissible on boldface B1,1(T).
However direct application of RI, o to the Sard Taylor expansion in boldface
B1,1 (T) gives

RI,o[F]:RI,o
(/ (x-- e) F~.,o(e, o) d e +Rl,o
[// Fl,x(e, 9)dedy
1
+RI,oI/(Y--Y) Fo,z(O,y)dY ]
x 1 (6.14)
= j K',~ t; e)V~,o(e, o) ,~e + J KO,'(s, t; ~)~0,,(o, Y) d~
0 0

+~ o/i ~,~(e, y) dy de
0 0
 Sard Kernel Theorems and Finite Element Analysis 227

where the first and last terms can be evaluated in Sard kernel form with

K*'~ t; ~ ) = R l , o E ( X - - s ] ----(s-- ~)~_-- t + ~ , ~=s,

(6.t5)
K~ t; y) = R l , o [(y--y)+ ] ----0,
Also

aT F I , I ( X ' •) d y a x = F l , l ( S ' y) d y . (6.t6)

o o o

Remark. For the first kernel in (6.t 5) s ----s is a jump set. The second kernel is
an example of the Zero Kernel Theorem to which Theorem 6.t is applicable.
The Lp norm of (6.t4) and the triangle inequality give

[[R~'~176 = ~ [1%(~)+ F~'x (s' Y) {,,,I (6.17)

where
c~ = It IIK~'~ t; s {6.18)
cf. (6.2).
Now, for p < 0% H61der's inequality gives t h a t

< of 0f Ill]~,;{;}llF,,,(s,~)llL,(~>dsdt
II/ F,,~(s,y)ay I: (s, f) "
1
_-< f Iit*t'~L',,II II~., (s. :~)IL; (~>IIL,{.>at (6.19)
0
I
< 0f (t - t)'/,' t~#: at IIF,,, (~, ~)ILL,(,, ~),
where L,, (p) is over [0, t], Lq (s) is over [0, t -- t], q =- r2/P and hence p __<r 2 _< 0%
q' ~-r2/(rs--p). Thus

IlRl,o f] ]k,{,,,)< C~l)&,o(~, o)tL,{~)+ C, IlFx,l(S, Y) IL,{,,~) (6.20)

where
C 4 = { B (p/r'a + t, 2 -- p -- p/rD} lip, (6.21)
and C 3 is defined by (6.t8) and (6.15). Calculation of the constant C 3 for the case
p = 2 gives
I V Y /2V % , r 1 = 1,
' r 1 = OO,
'
C 8 =/t/2 V3 , r 1' = r 1=2, (6.22)
/ , rl -----0% rl = 1 .

The bound IlRo,1[F] ]}L,{s,,}is the dual of (6.20).

Bounds for linear interpolation over an arbitrary triangle can be obtained b y
affine transformation. For example, the transformation of (6.20) to the triangle ~"
say with sides [0, HI and [0, K] gives

+ c, H"'-". K'+"'-'/"IIP,., L. a->. (6.23)

228 R . E . Barnhill and J. A. Gregory

where L,,(s) is on 0 ~ ]--< H. Such a result (which allows the degenerate case of
one angle of the triangle not being bounded a w a y from zero) is also shown b y
B a b u s k a [t].
Remark. The derivation of error bounds for more general interpolants on a
general triangle b y consideration of an affine transformation from the s t a n d a r d
triangle is only applicable when the interpolant is invariant under such a trans-
formation.
Application of the Sard Kernel T h e o r e m to calculate the constants in the error
bounds for smooth interpolation on triangles is considered in Barnhill and Mans-
field E4].

7. Application to Differential Equations

We consider the model problem
Au----O on O----[0, r e ] • re], (7.1)

u]a ~ = 0 except t h a t u (x, 0) = sin x, 0 =< x =< re,

discussed in Barnhill, Gregory, and W h i t e m a n [3]- F o r linear interpolation on a

triangle Th, we obtain the following bounds:
h8
Hu--Pa ]tL,Crh)~-- ] / ~ , (7.2)

h8
I{u-p~lJ~(~-<_ ~-, (7.3)
IIv (u- pl)IL ~r,, ----<V~ h~, (7.4)
IIv (,, - p,)II'~, --II RI,o u II~-,+ IIRo,1 u IlL
-- Z
(h)
[IIRI,o ~ll~.(~'h,+ llRo,1 ull~,(~h,], (7.5)
where the s u m m a t i o n is over all Th and Th*, Th* being the upper right triangle
analogous to the lower left right triangle T~. Inequalities (7.2), (7.4) and (7.5)
lead to the following bounds:
ll*,-pl IL (,,~ < . 70 hL (7.6)
2~
IIv (u- pl)IlL, (~, -<-- 7~- h. (7.7)
The actual errors at the midpoints of the sides of one subtriangle are 0.0043,
r,
--0.0t07, and 0.0t27, which compare with (7.3) which is 0.078 if h = ~-.

References
t. Babuska, I. : Private communication, Salt Lake City, August, t974
2. Barnhill, R. E., Gregory, J. A. : Sard kernel theorems on triangular and rect-
angular domains with extensions and applications to finite element error bounds.
Technical Report 1 ~, Department of Mathematics, Brnnel University, Uxbridge,
Middlesex, England, July, t972
 Sard Kernel Theorems and Finite Element Analysis 229

3. Barnhill, R. E., Gregory, J. A., Whiteman, J. R. : The extension and application

of Sard kernel theorems to compute finite element error bounds. Mathematical
Foundations of the Finite Element Method with Applications to Partial Differential
Equations, A. K. Aziz (ed.), p. 749-755, New York: Academic Press t972
4. Barnhill, R. E., Mansfield, L. : Error bounds for smooth interpolation in triangles.
J. Approx. Theory 11, 306-318 (t974)
5. Barnhill, R. E., Whiteman, J. R. : Error analysis of finite element methods with
triangles for elliptic boundary value problems. The Mathematics of Finite Elements
and Applications, J. R. Whiteman (ed.), p. 83-112, London: Academic Press 1972
6. Birkhoff, G., Schultz, M. H., Varga, R. S. : Piecewise Hermite interpolation in one
and two variables with applications to partial differential equations. Numer. Math.
11, 232-256 (1968)
7. Gregory, J . A . : Piecewise interpolation theory for functions of two variables.
P h . D . Thesis, Brunel Univ. 1975
8. Sard, A. : Linear Approximation. Mathematical Survey 9, American Mathematical
Society, Providence, Rhode Island, t963
9. Varga, R. S. : The role of interpolation and approximation theory in variational
and projectional methods for solving partial differential equations. I F I P Congress
71, p. t 4 - t 9 , Amsterdam: North Holland 1971

Robert E. Barnhill John A. Gregory

Mathematics Department Mathematics Department
University of Utah Brunel University
Salt Lake City Uxbridge,
Utah 84112 Middlesex
USA England