A note on polynomial approximation in Sobolev spaces

R. Verf urth
Fakult at f ur Mathematik, Ruhr-Universit at Bochum, D-44780 Bochum, Germany
E-mail address: rv@silly.num1.ruhr-uni-bochum.de
Date: October 1997
Summary: For domains which are star-shaped w.r.t. at least one point, we give new bounds on
the constants in Jackson-inequalities in Sobolev spaces. For convex domains, these bounds do not
depend on the eccentricity. For non-convex domains with a re-entrant corner, the bounds are uniform
w.r.t. the exterior angle. The main tool is a new projection operator onto the space of polynomials.
Key words: Jackson inequalities; polynomial approximation; Sobolev spaces
Resume: Pour des domaines etoiles on donne des nouvelles bornes sur les constants dans les
inegalites de Jackson pour les espaces de Sobolev. Pour des domaines convexes, les bornes ne
dependent pas de lexcentricite. Pour des domaines non-convexes ayant un point rentrant, les bornes
sont uniformes par rapport ` a langle exterieur. Loutil central est un nouvel operateur de projection
sur lespace des polynomes.
Mots clefs: Inegalites de Jackson; approximation polynomiale; espaces de Sobolev
AMS Subject classication: 41A17, 41A10, 65N15, 65N30
1. Introduction and main results
In this note we derive new upper bounds on the constant c
m,j
in the Jackson-type
inequality
sup
uH
m+1
()
inf
pIP
m
|u p|
H
j
()
|u|
H
m+1
()
c
m,j
d
m+1j
0 j m.
Here, is a bounded open domain in IR
n
, n 2, with diameter d. H
k
(), k IN, and
L
2
() := H
0
() denote the usual Sobolev- and Lebesgue-spaces equipped with the
standard norms .
H
k
()
and semi-norms |.|
H
k
()
. IP
m
is the space of all polynomials
in n variables of degree at most m.
The interest in sharp explicit bounds on the constants c
m,j
originates from the
a posteriori error analysis of nite element discretizations of partial dierential equa-
tions. The correct calibration of many popular a posteriori error estimators depends,
among others, on sharp explicit error estimates for suitable Clement-type interpo-
lation operators. These estimates in turn employ the constants c
m,j
(cf., e.g., [1,
4]).
Specically, we consider two types of domains: (1) convex domains and (2) non-
convex domains which are star-shaped w.r.t. at least one interior point. If is convex,
we obtain the bound
c
m,j

jm1

n +j 1
j

1/2
((m+ 1 j)!)
1/2

m+1j
n

n/2
. (1.1)
1
(As usual, x denotes the largest integer less or equal x.) If is not convex, but
star-shaped w.r.t. the point x , we obtain the bound
c
m,j
max

2
K
1
() +K
2
(), K
3
()

(m+1j)/2

n +j 1
j

1/2
((m+ 1 j)!)
1/2

m+1j
n

n/2
.
(1.2)
Here, the number is given by
:= max
y
|y x|
2
/ min
y
|y x|
2
(|.|
2
denotes the Euclidean norm in IR
n
) and the functions K
1
, K
2
, and K
3
are dened
by
K
1
(z) :=4z
n2
3z
2
K
2
(z) :=
4
n(n + 2)
[z
n2
z
2
],
K
3
(z) :=

lnz
1
2
+
1
2
z
2
, if n = 2,
2
n(n2)
z
n2

1
n2
+
1
n
z
2
, if n 3.
The main tool in establishing these bounds is a new projection operator of
H
m
() onto IP
m
which is dened in Section 2. In Sections 3 and 4 we will prove
estimates (1.1) and (1.2), respectively. In Section 5 we will shortly comment on the
generalization to L
q
-norms.
The best estimates of c
m,j
, which are known to us, are due to Dur an [2]. When
comparing his results with ours we observe two major dierences.
First, if is convex, the bound of [2] is proportional to d
n/2
||
1/2
, where || is the
n-dimensional Lebesgue-measure of . Thus, it is not uniformly bounded w.r.t. to
the eccentricity d
n
||
1
, whereas our bound does not depend on this parameter.
Second, if has a re-entrant corner, the bound of [2] tends to innity when the
exterior angle at the corner approaches zero. Our bounds, on the contrary, are uniform
w.r.t. the exterior angle. To see this, we may assume that is the intersection of the
unit ball with the complement of a cone with base at the origin and angle
(0, ) having the positive x
1
-axis as its axis of symmetry. An elementary geometrical
argument than shows that is star-shaped w.r.t. (
1
2
, 0, ..., 0) and that =
1
2
{5 +
4 cos

2
}
1/2
/
1
2
3 for all (0, ).
2
2. A projection operator
For any integer m and any measurable subset B of with positive Lebesgue-measure
|B| we want to dene a projection operator P
m,B
of H
m
() onto IP
m
which has the
following properties
D

(P
m,B
u) = P
mj,B
(D

u), (2.1)

B
D

(u P
m,B
u) = 0 (2.2)
for all u H
m
(), all 0 j m, and all IN
n
with || :=
1
+. . . +
n
= j.
To this end we denote by

B
:=
1
|B|

the mean value of any integrable function on B. For any u H

m
() we recursively
dene polynomials p
m,B
(u), ..., p
0,B
(u) in IP
m
by
p
m,B
(u) :=

IN
n
||=m
1
!
x

B
(D

u) (2.3)
and, for k = m, m 1, ..., 1,
p
k1,B
(u) = p
k,B
(u) +

IN
n
||=k1
1
!
x

B
[D

(u p
k,B
(u))]. (2.4)
We set
P
m,B
u := p
0,B
(u).
Obviously, P
m,B
is a linear operator of H
m
() onto IP
m
. In order to see that
it satises relations (2.1) and (2.2) we rst consider the case j = 0. Relation (2.1)
then is a trivial identity and relation (2.2) immediately follows from equation (2.4)
for k = 1. Next, assume that m 1 and x an arbitrary j with 1 j m and an
arbitrary IN with || = j. Dierentiating equation (2.3) we conclude that
D

p
m,B
(u) =

IN
n
||=mj
1
!
x

B
(D

u)
=p
mj,B
(D

u).
Dierentiating equation (2.4) we recursively obtain for k = m, m 1, ..., j + 1
D

p
k1,B
(u) =D

p
k,B
(u) +

IN
n
||=k1j
1
!
x

B
[D

(u p
k,B
(u))]
=p
kj,B
(D

u) +

IN
n
||=k1j
1
!
x

B
[D

(D

u p
kj,B
(D

u))]
=p
kj1,B
(D

u).
3
This proves that
D

(P
m,B
u) =D

(p
j,B
(u))
=p
0,B
(D

u) = P
mj,B
(D

u)
and thus establishes relation (2.1). Relation (2.2) immediately follows from relation
(2.1) and equation (2.4) for k = 1 with u replaced by D

u.
3. Convex domains
Assume that is convex. Payne and Weinberger [3] proved that

L
2
()

d

||
H
1
()
(3.1)
holds for all H
1
() having zero mean value.
Now, consider an arbitrary u H
m+1
(). Applying inequality (3.1) to :=
u P
m,
u and using relation (2.2) with = 0 yields that
u P
m,
u
2
L
2
()

2
|u P
m,
u|
2
H
1
()
=

2
n

i=1
D
i
(u P
m,
u)
2
L
2
()
.
Applying this estimate recursively to the derivatives of uP
m,
u and using relations
(2.1) and (2.2) we arrive at
u P
m,
u
2
L
2
()

2(m+1)
1i
1
,...,i
m+1
n
D
i
1
...D
i
m+1
u
2
L
2
()
=

2(m+1)
IN
n
||=m+1
(m+ 1)!
!
D

u
2
L
2
()

2(m+1)
max

(m+ 1)!
!
: IN
n
, || = m+ 1

|u|
2
H
m+1
()
.
A simple symmetry argument shows that
max

(m+ 1)!
!
: IN
n
, || = m+ 1

(m+ 1)!

m+1
n

n
and thus completes the proof of estimate (1.1) for the case j = 0.
The case 1 j m follows from the case j = 0, relation (2.1), and the observation
that

IN
n
: || = j

n +j 1
j

.
4
4. Non-convex, star-shaped domains
Assume that is star-shaped w.r.t. the point x and not necessarily convex. Since
the Lebesgue-integral is translation invariant we may from now on assume that x is
the origin. Set
:= min
y
|y|
2
, R := max
y
|y|
2
, :=
R

and denote by B the ball with radius and centre at the origin. We claim that

2
L
2
()
K
1
()
2

2
L
2
(B)
+K
2
()R
2
||
2
H
1
(B)
+K
3
()R
2
||
2
H
1
(\B)
(4.1)
holds for all H
1
(). Estimate (4.1) follows from the proof of Lemma 4.1 in [4].
For completeness we will sketch its proof at the end of this section.
Inserting := u P
m,B
u in estimate (4.1) and invoking relation (2.2) for = 0
and estimate (3.1) with replaced by B, we conclude that
u P
m,B
u
2
L
2
()
K
1
()
2

2
4

2
|u P
m,B
u|
2
H
1
(B)
+K
2
()R
2
|u P
m,B
u|
2
H
1
(B)
+K
3
()R
2
|u P
m,B
u|
2
H
1
(\B)
max

2
K
1
() +K
2
(), K
3
()

R
2
n

i=1
D
i
(u P
m,B
u)
2
L
2
()
.
Applying this estimate recursively to the derivatives of u P
m,B
u and recalling that
R d we arrive at
u P
m,B
u
2
L
2
()
max

2
K
1
() +K
2
(), K
3
()

m+1
d
2(m+1)

1i
1
,...,i
m+1
n
D
i
1
...i
m+1
u
2
L
2
()
max

2
K
1
() +K
2
(), K
3
()

m+1
d
2(m+1)
(m+ 1)!

m+1
n

n
|u|
2
H
m+1
()
.
This proves estimate (1.2) in the case j = 0. The case 1 j m again follows from
the case j = 0 and relation (2.2).
Now we shortly sketch the proof of estimate (4.1). Fix a function H
1
()
and a number r
0
with 0 < r
0
< . Denote by B
0
the ball with radius r
0
and centre
at the origin. Obviously we have

2
L
2
()
=
2
L
2
(\B
0
)
+
2
L
2
(B
0
)
.
5
Introducing polar co-ordinates, the rst term on the right-hand side of this equation
may be estimated as follows

2
L
2
(\B
0
)
=

S
n1

r()
r
0
s
n1
|(s)|
2
dsdS()
2

S
n1

r()
r
0
s
n1
|(s) (r
0
)|
2
dsdS()
+ 2

S
n1

r()
r
0
s
n1
|(r
0
)|
2
dsdS()
=:S
1
+S
2
.
Here, dS denotes the surface element of the unit sphere S
n1
and r : S
n1
IR

+
is
a continuous function such that = {r() : S
n1
}. (Such a function exists
since is assumed to be star-shaped w.r.t. the origin.) Note that R = max
S
n1
r().
Writing
(s) (r
0
) =

s
r
0

t
(t)dt S
n1
and using the Cauchy-Schwarz inequality, the term S
1
can be estimated as follows
S
1
2

S
n1

r()
r
0
s
n1

s
r
0
t
n+1
dt

s
r
0
t
n1
|

t
(t)|
2
dt

dsdS()
2

R
r
0
s
n1

s
r
0
t
n+1
dt

ds||
2
H
1
(\B
0
)
K
3

R
r
0

R
2
||
2
H
1
(\B
0
)
.
Writing
(r
0
) =

r
0
0

t
r
0

n
(t)

dt S
n1
and using the Cauchy-Schwarz inequality, the term S
2
can be etimated as follows
S
2

2r
0
n

R
r
0

n
1

S
n1
r
n1
0

r
0
0

t
r
0

n
(t)

dt

2
dS()

4r
0
n

R
r
0

n
1

S
n1

r
0
0
n
2
r
n+1
0
t
n1
dt

r
0
0
t
n1
|(t)|
2
dt

r
0
0
1
r
n+1
0
t
n+1
dt

r
0
0
t
n1
|

t
(t)|
2
dt

dS()
4

R
r
0

n
1

2
L
2
(B
0
)
+K
2
()R
2
||
2
H
1
(B
0
)
.
Collecting all estimates and observing that 0 < r
0
< was arbitrary, we nally arrive
at the desired result.
6
5. Jackson-inequalities in W
1,q
With obvious modications of the functions K
1
, K
2
, and K
3
, the arguments of Sec-
tions 3 and 4 immediately carry over to general L
q
-spaces with 1 < q < . The
crucial ingredient is an explicit knowledge of a Poincare-constant c
q
such that

L
q
()
c
q
diam()||
W
1,q
()
holds for all open convex subsets of IR
n
and all functions W
1,q
() having zero
mean value on .
From inequality (3.1) we know that c
2

1

. Integration along lines, on the

other hand, yields c

n
1/2
. (Note, that the diameter is always computed using the
Euclidean norm.) Interpolating between L
2
and L

we conclude that
c
q

2
q
n
1
2

1
q
2 q .
For the range 1 < q < 2 a similar estimate is not known to us.
References
[1] C. Carstensen and St. A. Funken: Constants in Clement-interpolation error and
residual based a posteriori error estimates in nite element methods. Report,
Universitat Kiel, 1997 (in preparation).
[2] R.G. Dur an: On polynomial approximation in Sobolev spaces. SIAM J. Numer.
Anal. 20, 985 988 (1983).
[3] L.E. Payne and H.F. Weinberger: An optimal Poincare-inequality for convex
domains. Archive Rat. Mech. Anal. 5, 286 292 (1960).
[4] R. Verf urth: Error estimates for some quasi-interpolation operators. Report, Uni-
versitat Bochum, 1997.
7