Hardy inequalities with optimal constants and remainder terms

Filippo Gazzola
Dipartimento di Scienze T.A.
Via Cavour 84
I-15100 Alessandria

Hans-Christoph Grunau
Fachgruppe Mathematik
Universitat
D-95440 Bayreuth

Enzo Mitidieri
Dipartimento di Scienze Matematiche
Via A. Valerio 12/1
Universit`a degli Studi di Trieste
I-34100 Trieste
May 4, 2000

Abstract
We show that in the classical Hardy inequalities with optimal constants in W01,p (), W02,2 (),
W 2,2 W01,2 () and also in further higher order Sobolev spaces remainder terms may be added.
Here is any bounded domain. For the Hardy inequality in W01,p there is a further Lp -norm
provided p 2, while for 1 < p < 2 we obtain a remainder term in Lq -norms with q < p. In
higher order Sobolev spaces besides the L2 -norm further singularly weighted L2 -norms arise.

Introduction

Hardys inequality in dimensions n > 2

u

W01,2 ()

(n 2)2
|u| dx
4

u2
dx.
|x|2

(1)

is one of the really classical Sobolev embedding inequalities, see [H, HLP]. Here and in what follows,
Rn is a bounded domain. Although we do not explicitly assume that 0 , we always have
this particularly interesting case in mind. Closely related and proven analogously is the Lp -version
we have
 Z

Z
np p
|u|p
1,p
p
u W0 () :
|u| dx
dx;
(2)
p
p
|x|

here we assume n > p 1. With Rellichs inequality in dimensions n > 4, see[R]:

u

W02,2 ()

n2 (n 4)2
(u) dx
16

e-mail: gazzola@unipmn.it
Hans-Christoph.Grunau@uni-bayreuth.de

mitidier@univ.trieste.it

u2
dx
|x|4

(3)

R
and its generalizations to estimates for |x| (u)2 dx, see Lemma 2 below, it is not difficult to
obtain Hardy inequalities in higher order Sobolev spaces W0k,2 (), where n > 2k. See e.g. [DH, Mi].
One has, if k = 2m even
!Z
Z
2m
2
Y
(n
+
4m

4`)
u2
(m u)2 dx
dx
(4)
4m
4

|x|
`=1

and if k = 2m + 1 is odd:
Z

|m u|2 dx

2m+1
Y
`=1

(n + 4m + 2 4`)2
4

!Z

u2
dx.
|x|4m+2

(5)

All the constants given here cannot be enlarged. For a more extensive survey and bibliography and
also for historical remarks we refer again to [DH].
The need to have refined versions of Hardy inequalities like (1) became first obvious in the
paper [BV] of Brezis and Vazquez, where among others the Gelfand problem
u = exp(u),

u 0,

u W01,2 ()

(6)

was studied for positive parameters and in dimensions n > 2. There exists a > 0, such that
for 0 , this problem is solvable, while for > it is not. Further, for = 2(n 2) and
= B the unit ball, one has the singular solution
using = 2 log |x| W01,2 (B).
The linearization of (6) around this singular solution leads to the Hardy-type operator
Llin =

2(n 2)
.
|x|2

This operator is studied in [BV] in order to find out among others whether or not the singular
solution is also extremal in the sense that it corresponds to . For this purpose Brezis and
Vazquez prove by means of a reduction of dimension-technique the following Hardy inequality
with optimal constant and a remainder term:
 2/n Z
Z
Z
u2
en
(n 2)2
1,2
2
u W0 () :
|u| dx
dx + 2
u2 dx.
(7)
2
4
||

|x|

Here n 2, and 2 denotes the first eigenvalue of the Laplace operator (the subscript 2 refers to
the p-Laplacian with p = 2, see Definition 1 below) in the two dimensional unit disk. Further en
and || denote the n-dimensional Lebesgue measure of the unit ball B Rn and the domain
respectively. For further applications and variants of (7) we refer to [VZ].
Remainder terms appear also in other Sobolev inequalities, and in the context of nonlinear
eigenvalue problems, see e.g. [BL, GG].
The goal of the present paper is to find remainder terms for all the Hardy inequalities (2), (3),
(4), (5) mentioned above. In a future paper we also want to address applications of these refined
Hardy inequalities.
In Section 2 the W 1,p -case is studied, corresponding to the p-Laplace operator. For p 2,
carefully exploiting the convexity properties of the function 7 p , we show that an extra Lp norm may be added. For p (1, 2) we find it somehow unexpected that we can show the existence
only of Lq -remainder terms with 1 q < p.
2

In Sections 3 and 4 we show that remainder terms

Z
Z
u2
u2 dx, . . . ,
dx
2k2
|x|

may be added in (4) and (5), in Section 3 under Navier boundary conditions 0 = u = u = . . . on
and in Section 4 under Dirichlet conditions 0 = u = u = u = . . . on .
As the class of functions is larger in Section 3, the constants in front of the remainder terms are
smaller than in Section 4. The constants which appear in Section 4 seem to be more natural, when
the order k of the Sobolev space becomes large. On the other hand, the reduction to the radially
symmetric case is more involved, because the symmetrization argument of Section 3 fails and
because iterative procedures will not give largest possible constants. For this reason it may be also
of interest in its own, how the result in W0k,2 (B) for radial functions is extended to nonsymmetric
functions.

In W01,p , 1 < p < n

We are always restricted to 1 < p < , if not stated differently.

Definition 1. Let X := {v C 1 ([0, 1]) : v 0 (0) = v(1) = 0} \ {0} and

R1

p = inf R01
X

rp1 |v 0 (r)|p dr
rp1 |v(r)|p dr

Remark. For all p > 1 we have p > 0 and if p N, then p is the first eigenvalue of p in Rp .
Theorem 1. Let Rn be a bounded domain, p (1, n), and let en = |B1 (0)| and || denote the
n-dimensional Lebesgue measure of the unit ball and of the domain resp.
(a) If p 2, then for every u W01,p () we have
Z

|u| dx

np
p

p Z

|u|p
dx + p
|x|p

en
||

p/n Z

|u|p dx.

(8)

(b) If 1 < p < 2, then for any q [1, p) there exists a constant C = C(n, p, q) > 0 such that for
every u W01,p () there holds
Z

|u|p dx

np
p

p Z

|u|p
dx + C
|x|p

en
||

(p/n)+(p/q)1 Z

p/q
|u|q dx
.

(9)

Remarks.
(i) Let n 3. If we put q := np/(n + 4 2p) in the case p < 2 the constant C in front of
the Lq -norm in (9) may be chosen such that C(n, p, q) 2 as p % 2, cf. (17) below: this
shows the continuity between (8) and (9) for p approaching 2. For p & 1 our constants
C(n, p, q) & 0 for any choice of admissible q.
3

(ii) If p = 1 and = B1 (0) the Hardy constant

n1=

inf

uW01,1 (B1 (0))\{0}

|u| dx
R |u|
|x| dx

is attained on any positive smooth and radially decreasing function u with u||x|=1 = 0.
This means that for p = 1 no remainder term may be added and shows that necessarily
C(n, p, q) & 0 as p & 1.
(iii) For related inequalities with weights being the inverse of the distance from the boundary of
we refer to [BM, MS].
To prove Theorem 1 we need the following elementary inequalities, which give a quantitative estimate from below for the convexity behaviour of power-like functions. Similar inequalities in a more
general context but with smaller constants were obtained in [L].

Lemma 1. Let p 1 and , be real numbers such that 0 and 0. Then

p
if p 2,

|| ,
p
p1
p
( ) + p
1
2

, if 1 p 2.
p(p 1)
2
( + ||)2p
Proof. By means of integration by parts or Taylors formula we find
Z 1
p
p1
p
2
( ) + p = p(p 1)
(1 t) ( t )p2 dt.

(10)

We assume first that p 2. If 0 one has t t || and the remainder term in (10) becomes
Z 1
Z 1
p(p 1) 2
(1 t) ( t )p2 dt p(p 1)||p
(1 t) tp2 dt = ||p .
0

For 0 we estimate t (1 t) || and obtain

Z 1
Z 1
2
p2
p
p(p 1)
(1 t) ( t )
dt p(p 1)||
(1 t)p1 = (p 1) ||p ||p .
0

Let now 1 p < 2, that means that p 2 < 0. Here we estimate ( t )p2 ( + ||)p2 , and
there follows:
Z 1
Z 1
2
2
1
2
p2
p(p 1)
(1 t) ( t )
dt p(p 1)
(1

t)
dt
=
p(p

1)
.
2
( + ||)2p 0
( + ||)2p
0

Proof of Theorem 1. After (Schwarz) symmetrization and rescaling we may assume that is the
unit ball, which we simply denote with B = B1 (0), and that u W01,p () is nonnegative, radially
symmetric and nonincreasing. We recall that symmetrization leaves Lq -norms of functions itself
4

unchanged, increases Lp -norms with the singular weight |x|p (see e.g. [AL, Theorem 2.2]) and
decreases Lp -norms of the gradient (see [AL, Theorem 2.7]). By density we may further assume
that u is as smooth as needed.
d
So, in what follows we may write u(r), u0 (r) = dr
u(r); it holds |u(x)| = |u0 (r)| with r = |x|
and we have to find a lower bound for

 Z
Z 1
n p p 1 np1
n1 0
p
I :=
r
|u (r)| dr
r
u(r)p dr.
p
0
0
Similarly as in [BV] a suitable transformation reduces the dimension and replaces not only the
partial integration, which is usually performed to show Hardy inequalities, but also uncovers a
remainder term. We set
v(r) := r(n/p)1 u(r),
so that
u0 (r) =

u(r) = r1(n/p) v(r),

(11)

n p n/p
r
v(r) + r1(n/p) v 0 (r).
p

Since u is radially non-increasing we have for r (0, 1]:

n p v(r)

v 0 (r),
p
r

(12)

and the expression I to be estimated becomes


p

 Z
Z 1
n p p 1 v(r)p
p1 n p v(r)
0
I=
r

v (r)
dr
dr.
p
r
p
r
0
0
By (12) we may apply Lemma 1 with
=

n p v(r)

p
r

= v 0 (r).

and

For p 2 we conclude by virtue of v(0) = v(1) = 0:

I p
p

np
p

p1 Z

1
p1

v(r)

v (r) dr +

rp1 |v 0 (r)|p dr

rp1 v(r)p dr = p

rn1 u(r)p dr

and (8) follows.

R1
The case 1 < p < 2 requires greater effort. Again by Lemma 1 and observing 0 v p1 v 0 dr = 0,
it follows that
Z 1
rp1 |v 0 (r)|2
1
I p(p 1)
(13)

2p dr.
2
np v(r)
0
0 (r)|

+
|v
p
r
In order to estimate this term from below we introduce a further regularizing factor r , where
> 0 will be chosen below in dependence on n, p, q. Consequently we will miss the Lp -norm for
the remainder term. Application of Holders inequality yields
Z

r+p1 |v 0 (r)|p dr

2/p

Z
0

Z
0

rp(p1)/2 |v 0 (r)|p

np
p

np
p

2
p
(2p)p/2 !
n p v(r)

+ |v 0 (r)|
dr
p
r

+(p1)(2p)/2
(2p)p/2 r

+ |v 0 (r)|

p  2p
Z 1

p
v(r)
rp1 |v 0 (r)|2
n

+ |v 0 (r)|
dr
(14)
r2/(2p) rp1
2p dr
p
r
0
0 (r)|
v(r)
+
|v
r
v(r)
r

For the second integral we find by means of a Hardy inequality in dimension (p + ), which is
proved for radial functions as in any integer dimension:
p

Z 1
0
2/(2p) p1 n p v(r)

+ |v (r)|
dr
r
r
p
r
0

p Z 1

p
Z 1
p1 n p
+p1 v(r)
p1
2
r
dr + 2
r+p1 |v 0 (r)|p dr
p
r
0
0


Z 1
np p
p1
2
+1
r+p1 |v 0 (r)|p dr

0
where we have also used the inequality (a + b)p 2p1 (ap + bp ) (a, b 0) and the fact that
r2/(2p) < r for all r (0, 1). Combining this estimate with (14) and inserting the result into
(13) we arrive at:
1
I p(p 1)2(p1)(p2)/p
2



np

p

+1

(p2)/p Z

r+p1 |v 0 (r)|p dr.

(15)

R1
In order to estimate 0 r+p1 |v 0 (r)|p dr from below we need a slightly different constant than in
Definition 1. For the same set X and for p > 1 we set
R1

,p = inf R01
X

r1 |v 0 (r)|p dr
r1 |v(r)|p dr

With help of this first eigenvalue of the radial p-Laplacian in dimension , we obtain from (15):
1
I p(p 1)2(p1)(p2)/p
2



np

p

(p2)/p
Z
+1
p+,p

r+p1 v(r)p dr.

(16)

For q [1, p) we choose := n(p q)/(2q), apply again Holders inequality and substitute u
according to (11):
1

Z

n1

u(r) dr

p/q

Z

1

(+n1)q/p

u(r)

(n1) pq
pq
p

dr

p/q

 (pq)/q Z 1
 (pq)/q Z 1
2
2
+n1
p

r
u(r) dr =
r+p1 v(r)p dr.
n
n
0
0
We insert this estimate into (16) and obtain finally
Z
B

|u|p dx

np
p

p Z
B

|u|p
dx + C(n, p, q)
|x|p
6

Z
B

|u|q dx

p/q

with
1
C(n, p, q) = p(p1)2(p1)(p2)/p
2

1
2en

(pq)/q 

2q(n p)
n(p q)

p

+1

(p2)/p

p+n(pq)/(2q),p . (17)

With the choice q = pn/(n + 4 2p) this constant tends to 2 as p % 2 as stated in the remark
above, because 2,p depends continuously on p.

Remark. We have to leave open the question, whether in the case 1 < p < 2 in Theorem 1 the
Lq -remainder term may be replaced with a suitable multiple of the Lp -norm. However, we have
the following proposition showing that the result for p 2 cannot hold equally for any smaller p.

Proposition 1. Let = B Rn (n 2) be the unit ball. For any p (1, 2) there is a nonnegative
radially nonincreasing function up W01,p (B) such that

 Z
Z
Z
|up |p
np p
p
|up | dx <
dx + p
|up |p dx.
p
p
|x|
B
B
B
Proof. We fix p (1, 2), n 2. Let v C 1 ([0, 1]) be a nonnegative radially nonincreasing first
eigenfunction for p defined in Definition 1, i.e.
Z

rp1 |v 0 (r)|p dr = p

rp1 v(r)p dr,

v 0 (0) = v(1) = 0.

According to (11) we consider

u(r) := r1(n/p) v(r);


this function, however, will not be in W01,p (B). So, for 0, np
we introduce functions, which
p
are regularized around 0:
v (r) := r v(r);

u (r) := r1(n/p) v (r) = r+1(n/p) v(r);

and we want to estimate the following expression from above:


 Z
Z 1
n p p 1 np1
n1 0
p
r
|u (r)| dr
r
u (r)p dr
p
0
0

p

 Z
Z 1
v
(r)
n p p 1 1
n

p1
0
=
r

v (r)
dr
r v (r)p dr =: I .
p
r
p
0
0
Here we observe that for (n p)/p
v0 (r) = r v 0 (r) + r

v(r)
v (r)
n p v (r)

.
r
r
p
r

v (r)
That means first that u is nonincreasing and further that for all r, := np
and := v0 (r)
p r
are in the region which was considered in Lemma 1. A close look at its proof shows that for 0,
and p < 2:
( )p p p1 + p + ||p ,

and equality holds only if = 0 or = 0. As on any compact interval (0, 1], v and v0 converge
v (r)
uniformly to v and v 0 resp., we find an interval [r0 , r1 ] (0, 1] such that we have = np
and
p r
0
= v (r) in a suitable compact subset of (0, ) (, 0) for small enough, say < . Hence
there is a number > 0 such that we have on [r0 , r1 ] uniformly in < :
p





 


n p p1 v (r) p1 0
n p p v (r) p
n p v (r)

v0 (r) + p
v (r)
|v0 (r)|p .
p
r
p
r
p
r
On the whole interval [0, 1] the l.h.s. of this expression is at least nonpositive. We conclude that
we have a positive constant c0 = c0 (p, r0 , r1 ) such that for all < :


Z r1
Z
Z 1
n p p1 1 0
p1
p1
r
dr p
v (r)v (r)
dr +
rp1 |v0 (r)|p dr
I
p
0
0
r0


Z 1
n p p1
p
p
c0
(v (1) v (0) ) +
rp1 |v 0 (r)|p dr
p
0
Z 1


rp1 |v 0 (r)|p1 + |v0 (r)|p1 |v0 (r) v 0 (r)| dr
+p
0
Z 1
Z 1
c0 + p
rp1 |v (r)|p dr + C
rp1 (1 rp ) dr
0
0
Z 1




+C
rp1 1 + p1 r(1)(p1) (1 r ) + r1 dr
0
Z 1


c0 + p
rn1 |u (r)|p dr + O p1
0
Z 1
< p
rn1 |u (r)|p dr
0

for small enough; up (x) := u (|x|) satisfies the stated reverse inequality.

Remark. Let X := {v C 1 ([0, 1]) : v 0 (0) = v(1) = 0} \ {0} and

R1 0
r|v (r)|p dr

p = inf R01
,
p
X
0 r|v(r)| dr
then by choosing = 2 p in (15) we obtain
 Z


(p2)/p

Z
Z
np p
np p
|u|p 1
p
(p1)(p2)/p

|u|
+ p(p 1)2
+1
p
|x|2p |u|p
p
p
2
2p
B
B |x|
B
for all radially symmetric u W01,p (B). As p 2 this inequality tends to (8) with p = 2 and this
shows further continuity as p approaches 2.

In higher order Sobolev spaces under Navier boundary conditions

Also here a similar reduction of dimension-technique as in [BV] and as in the preceding section
will be employed. In order to find the largest possible constants also for the remainder terms we
need a notation for first eigenvalues of the Laplacian in appropriate space dimensions.

Definition 2. Let X = {v C 1 ([0, 1]) : v 0 (0) = v(1) = 0, v 6 0} and for n N let

R1

rn1 |v 0 (r)|2 dr
(n) = inf R0 1
.
n1 v(r)2 dr
X
0 r
This notation should not be confused with that of Section 2.

Proposition 2. Let B Rn be the unit ball. The first eigenvalue 0 of the following Navier
boundary value problem
 2
u = u in B,
u = u = 0 on B,
satisfies
2

(n) = 0 =

inf

uW 2,2 W01,2 (B)\{0}

2
BR(u) dx
.
2
B u dx

Proof. The first equality is due to the fact that the biharmonic operator u 7 2 u under Navier
boundary conditions u = u = 0 on B is actually the square of the Laplacian under Dirichlet
conditions. For the second equality, see e.g. [V, Lemma B3].


Theorem 2. Let Rn , n 4, be a bounded domain and denote as above by || its n-dimensional

Lebesgue measure and by en = |B|, where B = B1 (0) is the unit ball. Then for any u W 2,2
W01,2 () one has:
Z

(u) dx

n2 (n 4)2
16

u2
dx
4
|x|
 2/n Z
 4/n Z
n(n 4)
en
u2
en
2
+
(2)
dx + (4)
u2 dx.
2
2
||
|x|
||

(18)

Proof. Step 1. By scaling it suffices to consider || = en . Further we want to show first that
we may restrict ourselves to = B and radial functions u(r). For this purpose we assume that
(18) has already been shown in the radial setting, and let now be arbitrary with || = en and
u W 2,2 W01,2 (). We define f = u, let w W 2,2 W01,2 (B) be a (radial) strong solution of
(
w = f in B,
w=0

on B,

where f 0 denotes the Schwarz symmetrization of f . By [T, Theorem 1] we know that w

u 0, so that
Z
Z
Z
Z
2
2
2
(u) dx =
f dx =
(f ) dx =
(w)2 dx

Z
B

w2
|x|2k

dx

Z
B

(u )2
|x|2k

dx

u2
|x|2k

dx,

k = 0, 1, 2;

for the last inequality we again refer to [AL, Theorem 2.2]. Assuming (18) in the radial setting,
the same follows for any domain and any u W 2,2 W01,2 ().
Step 2. We assume now that = B and that u is radially symmetric: u = u(r), r = |x|. Similarly
as in (11) we introduce the transformation
v(r) := r(n/2)2 u(r),
so that
u(r) = r

2(n/2)

u(r) = r2(n/2) v(r),

3
v + v0
r
00

1
n(n 4)rn/2 v,
4

and the 4-dimensional Laplacian of v arises. In the following integration by parts only the
boundary datum u(1) = v(1) = 0 is needed:
Z 1
Z
n2 (n 4)2 1 n5 2
2
n1
r
(u) dr
r
u dr
16
0
0
2

Z 1 
Z 1 
Z 1
3
1
1
=
r3 v 00 + v 0
dr + n(n 4)
r v 00 v 0 v dr n(n 4)
v 0 v dr
r
2
r
0
0
0
Z 1
Z 1
1
r3 (4 v)2 dr + n(n 4)
=
r (2 v) v dr,
2
0
0
where 4 v and 2 v denote the radial Laplace operator in 4 and 2 dimensions respectively. According to Proposition 2 we proceed with
Z 1
Z
n2 (n 4)2 1 n5 2
2
n1
r
(u) dr
r
u dr
16
0
0
Z 1
Z 1
1
2
3 2
(4)
r v dr + n(n 4) (2)
r v 2 dr
2
0
0
Z 1
Z 1
1
u2
= (4)2
rn1 u2 dr + n(n 4) (2)
rn1 2 dr,
2
r
0
0
thereby completing the proof of Theorem 2.

Combining this result with the original extended Hardy inequality of Brezis-Vazquez (7) already
mentioned in the introduction, we have for third order Sobolev spaces:

Theorem 3. Let Rn , n 6, be a bounded domain. Then for any u W 3,2 W01,2 () with
u = 0 on , i.e. u W01,2 (), one has:
Z
Z
u2
(n + 2)2 (n 2)2 (n 6)2
|u|2 dx
dx
6
64
|x|

 2/n Z


en
1
u2
4
2
+
3(n 2) + 8(n 2) + 16 (2)
dx
4
16
||
|x|

  4/n Z
(n 2)2
n(n 4)
en
u2
2
2
+
(4) +
(2)
dx
2
4
2
||
|x|
 6/n Z
en
2
+ (2) (4)
u2 dx.
||

10

Proof. First we may reduce to = B and the radial setting by combining the argument from the
proof of Theorem 2 and [AL, Theorem 2.7].
The application of the Brezis-Vazquez inequality (7) reduces the order of terms by one:
1

rn1 |u|2 dr

(n 2)2
4

rn3 (u)2 dr + (2)

rn1 (u)2 dr.

The second term will be estimated directly with help of Theorem 2, while for the singular term
Z

rn3 (u)2 dr

we have to extend that proof by means of the transformation

v(r) := r(n/2)3 u(r),

u(r) = r3(n/2) v(r),

so that
1

=
=

Z
(n + 2)2 (n 6)2 1 n7 2
(u) dr
r
u dr
16
0



Z 1 
Z
2 0 2
(n + 2)(n 6) 1
5 0
3
00
r 4 v + v
dr
r v + v v dr
r
2
r
0
0
Z 1
Z 1
r3 (4 v)2 dr + 4
r2 (4 v) v 0 dr
0
0
Z 1
Z


(n + 2)(n 6) 1
0 2
+4
r v
dr +
r (2 v) v dr
2
0
0
Z 1
Z

1 1

1
r3 (4 v)2 dr + 2 r2 v 0 (r)2 0 +
(n 2)2 + 8
r (2 v) v dr
2
0
0
Z 1
Z 1


1
2
3 2
2
r v dr +
(4)
(n 2) + 8 (2)
rv 2 dr
2
0
0
Z 1
Z 1


1
2
2
n3 2
(n 2) + 8 (2)
rn5 u2 dr.
(4)
r
u dr +
2
0
0
2

n3

Note that only u(1) = v(1) = 0 was needed. Collecting terms yields the stated inequality.

In order to iterate further we quote from [DH, Theorem 12], [Mi, Theorem 3.3]:
Lemma 2. Let Rn be a sufficiently smooth bounded domain and < n 4. Then for every
with u| = 0 we have
u C 2 ()
Z

(u)2
(n 4 )2 (n + )2
dx

|x|
16

u2
dx.
|x|+4

We emphasize that it is sufficient to have u| = 0.

With help of this result it is not difficult to see that there are remainder terms in Hardy inequalities
of arbitrary order.

11

Corollary 1. Let Rn be a sufficiently smooth domain and let k N, 2k n. Then there

exist constants c1 , . . . , ck , depending only on k, n and ||, such that for every u W 2,k () with
j u| = 0 for j N0 and 2j < k there holds:
(a) if k even, k = 2m, m N:
Z

1
( u) dx 2m
4

2m
Y

(n + 4(m j))

j=1

2m

X
u2
dx
+
c`
|x|4m
`=1

u2
|x|4m2`

dx;

(b) if k odd, k = 2m + 1, m N0 :
Z

|m u|2 dx

1
42m+1
+

2m+1
X
`=1

2m+1
Y

c`

j=1

(n + 4m + 2 4j)2
u2

|x|4m+22`

u2
dx
|x|4m+2

dx.

The existence of some constants c1 , . . . , ck as above is easily shown, of particular interest would be
their respective largest possible value. But already the proof of Theorem 3 indicates that iterative
methods under Navier boundary conditions 0 = u = u = . . . on yield constants which cannot
be easily given in closed form and which presumably will not be optimal.
One may hope that things will improve under the more restrictive Dirichlet conditions, i.e.
u W0k,2 (). This will be discussed in the following section.

In higher order Sobolev spaces under Dirichlet boundary conditions

Our first goal will be to adapt Theorem 2 to Dirichlet conditions, i.e. to the space W02,2 () instead
of W 2,2 W01,2 (). At least when is a ball or close to a ball the estimation constants will be
considerably larger. On the other hand the symmetrization argument in the proof of Theorem 2
fails: the function w used in that proof is not in the class W02,2 (B) to be considered here. We
overcome this difficulty by a relatively involved argument, which bases on positivity properties of
Green functions and works only if is replaced with the circumscribed ball.
As the technical difficulties to obtain best possible constants increase rapidly with the order
k of the space W0k,2 (), we give a complete discussion only for the improved Hardy inequality in
W02,2 (). For the general case we only give a conjecture to show the expected behaviour of the
remainder terms, when the order of spaces become arbitrarily large.
In this section also eigenvalues of biharmonic and more general polyharmonic operators will
play an important role.

12

Definition 3. Let B Rn denote again the unit ball and let

(m u)2 dx

BR

inf
,
k,2
W0 (B)\{0}
u2 dx


B
()k , n =
R

m u|2 dx

B |

R
inf
,

k,2
2
W0 (B)\{0}
B u dx

k = 2m, m N;

k = 2m + 1, m N0 .

Remark. The notation (n) of Section 3 is a special case by (n) = (() , n).


Taking into account Proposition 2 we immediately see ()2 , n (, n)2 . With help
of qualitative properties of first eigenfunctions in particular under Navier boundary conditions,
also strict inequality can be shown. Moreover
the

 ratio of these eigenvalues is rather large, by
elementary calculations one finds e.g. ()2 , 1 = 31.285243 . . . while (() , 1)2 = 4 /16 =
6.088068 . . .. For n = 
4, the case which is needed below, a very rough estimate according to [PS, p.
2
57] gives () , 4 j12 j22 = 387.23 . . . while (() , 4)2 = j14 = 215.56 . . ..

Theorem 4. Let Rn , n 4, be a bounded domain, BR (0). Then for every u W02,2 ()

one has:
Z
Z
n2 (n 4)2
u2
2
(u) dx
dx
4
16

|x|
Z
Z

4
n(n 4)
u2
2
+
(, 2) R2
dx
+

()
,
4
R
u2 dx. (19)
2
2
|x|

Proof. By trivial extension we have W02,2 () , W02,2 (BR (0)), after suitable scaling we may assume
that = B.
If u is additionally radially symmetric, we proceed as in Step 2 of the proof of Theorem 2. In its
last conclusion we may instead exploit that
 u satisfies
 homogeneous Dirichlet boundary conditions
2
2
2
and replace (4) = (, 4) with () , 4 , thereby proving (19) for radial u.
It remains to extend (19) to arbitrary functions u W02,2 (B). The idea is that the formal EulerLagrange equation
a minimum problem associated with (19) and with the eigenvalue parameter
R for
2
in front of the B u dx-term is linear and has only radial coefficients. Hence radialization of any
solution would give again a solution. That the latter solution is not the trivial one has to be avoided
e.g. with starting with a positive solution. Hence a minimizer could be chosen radial. However,
the mentioned eigenvalue problem is critical in so far, as the infimum may not be attained.
For that reason we consider for ` N a sequence of relaxed minimum problems:
` :=

inf

W02,2 (B)\{0}

F` (v)
,
2
B v dx

where

Z
n2 (n 4)2
1
v2
F` (v) =
(v) dx
1
dx
4
16
`
B
B |x|

Z
1
n(n 4)
v2

(, 2) 1
dx.
2
2
`
B |x|
Z

13

R
As we may already use Theorem 2 we have F` (v) 1` B (v)2 dx for every v W02,2 (B). That
R 2
2,2
means that for every minimizing
sequence
v

W
(B)
with
k
0
B vk dx = 1 and limk F` (vk ) = ` ,


we have boundedness of

kuk kW 2,2 (B)

0

uk * u

W02,2 (B)

kN

. After selecting a subsequence we may assume that

and uk u strongly in L2 (B). The bilinear form


Z
Z
n2 (n 4)2
1
vw
F` (v, w) :=
v w dx
1
dx
4
16
`
B
B |x|

Z
1
vw
n(n 4)
(, 2) 1
dx

2
2
`
B |x|

associated with F` (v) defines a scalar product on W02,2 (B), which by

Z
Z
1
2
(v) dx F` (v, v)
(v)2 dx = kvkW 2,2 (B)
0
` B
B
gives an equivalent norm. If we consider just for the following argument W02,2 (B) with F` ( . , . ) as
scalar product, then lower semicontinuity of the corresponding norm in the weak topology gives
` = lim inf F` (vk ) F` (v).
k

We have further
1 = lim

k B

vk2 dx

v 2 dx.

Hence F` (v) = ` , and v W02,2 (B) is an optimal (nontrivial) function for ` . Consequently v is a
weak solution of the Euler-Lagrange equation





1 n(n 4)
v
1 n2 (n 4)2 v
2
v = 1
+ 1
(, 2) 2 + ` v in B,
4
(20)
`
16
|x|
`
2
|x|

v = v = 0
on B.
Next we show that v is of fixed sign. Assume by contradiction that there are subsets B + , B B
both with positive measure such that v > 0 on B + and v < 0 on B . We may now apply a
decomposition method explained in detail in [GG, Section 3], cf. also [Mo]. Let
n
o
K = u W02,2 (B) : u 0
be the closed convex cone of nonnegative functions and
n
o
K = u W02,2 (B) : (u , u) 0 for all u K
R
the dual cone. Here (u , u) = B (u ) (u) dx is the standard scalar product in W02,2 (B). K is
the cone of all weak subsolutions of the biharmonic equation in B under Dirichlet conditions. By
a comparison result of Boggio [Bo, p. 126] the elements of K are nonpositive. We decompose
v = v1 + v2 ,

v1 K,

v2 K ,

v1 v2 .

As 0 6 v1 0, 0 6 v2 0, replacing v = v1 + v2 with v1 v2 0 yields

Z
Z
2
F` (v1 + v2 ) > F` (v1 v2 ),
(v1 + v2 ) dx <
(v1 v2 )2 dx,
B

14

R
contradicting the minimality of F` (v)/ B v 2 dx.
Hence we may assume 0 6 v 0. Considering polar coordinates x = r, r [0, 1], || = 1 and
integrating the Euler-Lagrange equation (20) over {|| = 1} shows that
Z
1
v(r ) d() W02,2 (B)
w(r) :=
n en ||=1
is a radial weak solution of (20). By virtue of 0 6 v 0, we also have 0 6 w 0. As we may
already use the Hardy inequality (19) for the radial function w we may conclude:
R
R w2
n2 (n4)2 R w2
n(n4)
2
(, 2) B |x|
2 dx
F` (w)
16
2
B (w) dx
B |x|4 dx
R
` = R
>
2
2
B w dx
B w dx


2
() , 4 .
To sum up, we have shown that for any ` N one has for every u W02,2 (B):


Z
Z
1 n2 (n 4)2
u2
2
(u) dx
1
dx
4
`
16
B
B |x|


Z
Z


1 n(n 4)
u2
2
+ 1
(, 2)
dx + () , 4
u2 dx.
2
`
2
|x|
B
B
Letting ` , the stated Hardy inequality (19) follows.

We conclude with some remarks concerning general higher order Sobolev spaces. We recall that
W0k,2 () is equipped with the norm
Z

(m u)2 dx, if k = 2m, m N;

kuk2W k,2 () =
Z

|m u|2 dx, if k = 2m + 1, m N0 .

We expect the following result:

Conjecture 1. Let Rn be a bounded domain, BR (0). Let k N , n 2k. Then for all
u W0k,2 () there holds:
)
  (Y
j
k
X
1 k
2
kukW k,2 ()
((n + 2k 4`)(n 2k 4 + 4`))
4j j
0
j=0
`=1
Z


u2
()kj , 2k 2j R2j2k
dx
(21)
2j
|x|
with the convention that (()0 , 0) = 1.
In the radial setting the terms


()

kj

, 2k 2j

Z

u2
dx
|x|2j

originate from
nen

Z
0

2k2j1

15

2
(kj)/2
2k2j v

dr

if k j is even, and from

Z

nen

r2k2j1



(kj1)/2

2k2j

 0 2

dr

if k j is odd. Here 2k2j is the radial Laplacian in dimension 2k 2j. The radial functions u
and v are related by means of the transformation
v(r) = r(n/2)k u(r),

u(r) = rk(n/2) v(r).

In Conjecture 1 as well as in the proof of Theorem 4 we simply used for these terms eigenvalue
estimates. One may also wish to use Hardy-like inequalities in order to have largest possible
constants in front of the most singular remainder terms. The following Lemma may be applied,
because we assume every function to satisfy homogeneous Dirichlet boundary conditions.
Lemma 3.
(a) Let k 2, j 1, v C 2 ([0, 1]) with v 0 (1) = 0. Then:
Z
0

(k + 1 2j)2
r (j v) dr
4

rk2 v 0

2

dr.

(b) Let k 2, v C 1 ([0, 1]) with v(1) = 0. Then:

Z
0

0 2

(k 1)2
dr
4

rk2 v 2 dr.

To illustrate how the application of this elementary and well-known Lemma shifts less singular
remainder terms to more singular ones, we modify the proof of Theorem 4 and obtain:

Corollary 2. Let Rn , n 4, be a bounded domain, BR (0). Then for every u W02,2 ()

one has:
Z
Z
n2 (n 4)2
u2
2
(u) dx
dx
4
16

|x|
Z
u2
1
+ (n(n 4) + 8) (, 2) R2
dx.
2
2
|x|

References
[AL]

F. Almgren, E. Lieb, Symmetric decreasing rearrangement is sometimes continuous, J.

Amer. Math. Soc. 2 (1989), 683-773

[Bo]

T. Boggio, Sulle funzioni di Green dordine m, Rend. Circ. Mat. Palermo 20 (1905), 97-135.

[BL]

H. Brezis, E. Lieb, Sobolev inequalities with remainder terms, J. Funct. Anal. 62 (1985),
73-86.
16

[BM] H. Brezis, M. Marcus, Hardys inequalities revisited, Ann. Sc. Norm. Sup. Pisa 25 (1997),
217-237
[BV]

H. Brezis, J. L. Vazquez, Blow-up solutions of some nonlinear elliptic problems, Rev. Mat.
Univ. Complutense Madrid 10 (1997), 443-469.

[DH]

E. B. Davies, A. M. Hinz, Explicit constants for Rellich inequalities in Lp (), Math. Z. 227
(1998), 522-523.

[GG] F. Gazzola, H.-Ch. Grunau, Critical dimensions and higher order Sobolev inequalities with
remainder terms, Nonlin. Diff. Eq. Appl. NoDEA, to appear.
[H]

G. H. Hardy, Notes on some points in the integral calculus, Messenger Math. 48 (1919),
107-112.

[HLP] G. H. Hardy, J. E. Littlewood, G. P

olya, Inequalities, Cambridge: University Press, 1934.
[L]

P. Lindqvist, On the equation div(|u|p2 u) + |u|p2 u = 0, Proc. Am. Math. Soc. 109
(1990), 157-164.

[MS]

M. Marcus, I. Shafrir, An eigenvalue problem related to Hardys Lp inequality, preprint

[Mi]

E. Mitidieri, A simple approach to Hardys inequalities, Mat. Zametki (2000), to appear,

downloadable from: http://mathsun1.univ.trieste.it/mitidier/pubblicazioni.html.

[Mo]

J. J. Moreau, Decomposition orthogonale dun espace hilbertien selon deux cones mutuellement polaires, C. R. Acad. Sci. Paris 255 (1962), 238-240.

[PS]

P. Pucci, J. Serrin, Critical exponents and critical dimensions for polyharmonic operators,
J. Math. Pures Appl. 69 (1990), 55-83.

[R]

F. Rellich, Halbbeschr
ankte Differentialoperatoren h
oherer Ordnung, in: J. C. H. Gerretsen
et al. (eds.), Proceedings of the International Congress of Mathematicians Amsterdam 1954,
Vol. III, pp. 243-250, Groningen: Nordhoff, 1956.

[T]

G. Talenti, Elliptic equations and rearrangements, Ann. Scuola Norm. Sup. Pisa, Cl. Sci.,
Ser. IV, 3 (1976), 697-718.

[VZ]

J. L. Vazquez, E. Zuazua, The Hardy constant and the asymptotic behaviour of the heat
equation with an inverse-square potential, J. Funct. Anal. 173 (2000), 103-153.

[V]

R. van der Vorst, Best constant for the embedding of the space H 2 H01 () into
L2N/(N 4) (), Diff. Int. Eq. 6 (1993), 259-276.

17