Sobolev Inequalities in Arbitrary Domains

Advances in Mathematics 293 (2016) 644696
Contents lists available at ScienceDirect
Advances in Mathematics
www.elsevier.com/locate/aim
Sobolev inequalities in arbitrary domains

Andrea Cianchi a, , Vladimir Mazya b,c
a
Dipartimento di Matematica e Informatica U.Dini, Universit di Firenze,
Viale Morgagni 67/A, 50134 Firenze, Italy
b
Department of Mathematics, Linkping University, SE-581 83 Linkping, Sweden
c
Department of Mathematical Sciences, M&O Building, University of Liverpool,
Liverpool L69 3BX, UK
a r t i c l e i n f o a b s t r a c t
Article history: A theory of Sobolev inequalities in arbitrary open sets in Rn is

Received 15 September 2015 established. Boundary regularity of domains is replaced with
Accepted 1 February 2016 information on boundary traces of trial functions and of their
Available online 3 March 2016
derivatives up to some explicit minimal order. The relevant
Communicated by Erwin Lutwak
Sobolev inequalities exhibit the same critical exponents as
MSC: in the classical framework. Moreover, they involve constants
46E35 independent of the geometry of the domain, and hence yield
46E30 genuinely new results even in the case when just smooth
domains are considered. Our approach relies upon new
Keywords: representation formulas for Sobolev functions, and on ensuing
Sobolev inequalities pointwise estimates which hold in any open set.
Irregular domains 2016 Elsevier Inc. All rights reserved.
Boundary traces
Optimal norms
Representation formulas

This research was partly supported by the Research Project of Italian Ministry of University and
Research (MIUR) Prin 2012 (grant number 2012TC7588) Elliptic and parabolic partial dierential
equations: geometric aspects, related inequalities, and applications, and by GNAMPA of the Italian INdAM
(National Institute of High Mathematics).
* Corresponding author.
E-mail address: cianchi@uni.it (A. Cianchi).
http://dx.doi.org/10.1016/j.aim.2016.02.012
0001-8708/ 2016 Elsevier Inc. All rights reserved.
A. Cianchi, V. Mazya / Advances in Mathematics 293 (2016) 644696 645
1. Introduction
The aim of this paper is to develop a theory of Sobolev embeddings, of any order
m N, in arbitrary open sets in Rn , with n 2. As usual, by an m-th order Sobolev
embedding we mean an estimate for a norm of the h-th order weak derivatives (0 h
m 1) of any m times weakly dierentiable function in in terms of norms of some of
its derivatives up to the order m.
The classical theory of Sobolev embeddings involves ground domains satisfying
suitable regularity assumptions. For instance, a formulation of the original theorem by
Sobolev reads as follows. Assume that is a bounded domain satisfying the cone prop-
erty, m N, 1 < p < m n
, and F() is any continuous seminorm in W m,p () which does
not vanish on any (non-zero) polynomial of degree not exceeding m 1. Then there exists
a constant C = C() such that

u np C m uLp () + F(u) (1.1)
L nmp ()
for every u W m,p (). Here, W m,p () denotes the usual Sobolev space of those func-
tions in whose weak derivatives up to the order m belong to Lp (), and m u stands
for the vector of all (weak) derivatives of u of order m.
It is well known that standard Sobolev embeddings are spoiled in presence of domains
with bad boundaries. In particular, inequalities of the form (1.1) do not hold, at least
np
with the same critical exponent nmp , in irregular domains. The interplay between the
geometry of the domain and Sobolev inequalities, even in frameworks more general than
the Euclidean one, has over the years been the subject of extensive investigations, along
diverse directions, by a number of authors. Their results are the object of a rich literature,
which includes the papers [3,5,6,16,7,8,12,13,18,19,2527,29,28,31,32,37,39,43,49,47,48,
5052,5456,6062,64,65,68,69] and the monographs [2124,41,62,66]. Various classical
problems in the theory of spaces of dierentiable functions, involving possibly irregular
sets, have attracted new interest in the last two decades, and have been considered in
such contributions as [10,17,20,33,34,45,40,46,63,67,70].
In order to avoid any assumption on , we deal with Sobolev inequalities from an
unconventional perspective. The underling idea of our results is that suitable information
on boundary traces of trial functions can replace boundary regularity of .
The inequalities that will be established have the form

h uY (,) C m uX() + N (u) , (1.2)
where m N, h N0 , X() is a Banach function norm on with respect to Lebesgue

measure Ln , Y (,) is a Banach function norm with respect to a possibly more general
measure , and N () is a (non-standard) seminorm on , depending on the trace of u,

and of its derivatives up to the order m12 . Here, N0 = N {0}, and [] denotes integer
part. Moreover, 0 u stands just for u, and we shall denote 1 u also by u.
646 A. Cianchi, V. Mazya / Advances in Mathematics 293 (2016) 644696
Some distinctive traits of the inequalities to be presented can be itemized as follows:
No regularity on is a priori assumed. In particular, the constants in (1.2) are

independent of the geometry of .
The critical Sobolev exponents, or, more generally, the optimal target norms, are the
same as in the case of regular domains.

The order m1 2 of the derivatives, on which the seminorm N () depends, is
minimal for an inequality of the form (1.2) to hold without any additional assumption
on .
Let us emphasize that, in view of these features, our inequalities provide original
conclusions even when just applied to smooth trial functions u on regular domains .
A rst-order Sobolev inequality on arbitrary domains Rn , of the form (1.2),
where X() = Lp (), Y (, ) = Lq (), and N () = Lr () , with 1 p < n, r 1
and q = min{ n1 rn np
, np } was established in [60] via isoperimetric inequalities. Sobolev
inequalities of this kind, but still involving only rst-order derivatives and Lebesgue
measure, have recently received renewed attention. In particular, the paper [57] makes
use of mass transportation techniques to address the problem of the optimal constants for
p (1, n), the problem when p = 1 having already been solved in [60]. Sharp constants
in inequalities in the borderline case when p = n are exhibited in [58].
In the present paper, we develop a completely dierent approach, which not only
enables us to establish arbitrary-order inequalities, which cannot just be derived via
iteration of rst-order ones, but also augments the rst-order theory, in that more general
measures and norms are allowed.
Our point of departure is a new pointwise estimate for functions, and their derivatives,
on arbitrary possibly unbounded and with innite measure domains . Such estimate
involves a novel class of double-integral operators, where integration is extended over
Sn1 . The relevant operators act on a kind of higher-order dierence quotients of
the traces of functions and of their derivatives on .
In view of applications to norm inequalities, the next step calls for an analysis of
boundedness properties of these operators in function spaces. To this purpose, we prove
their boundedness between optimal endpoint spaces. In combination with interpolation
arguments based on the use of Peetre K-functional, these endpoint results lead to point-
wise bounds, for Sobolev functions, in rearrangement form. As a consequence, Sobolev
inequalities on an arbitrary n-dimensional domain are reduced to considerably simpler
one-dimensional inequalities for Hardy type operators.
With this apparatus at disposal, we are able to establish inequalities involving
Lebesgue norms, with respect to quite general measures, as well as YudovichPohozaev
Trudinger type inequalities in exponential Orlicz spaces for limiting situations. The
compactness of corresponding RellichKondrashov type embeddings, with subcritical
exponents, is also shown. Inequalities for other rearrangement-invariant norms, such as
Lorentz and Orlicz norms, could be derived. However, in order to avoid unnecessary
additional technical complications, this issue is not addressed here.
One major motivation for our research is to provide an eective functional framework
for boundary value problems for partial dierential equations, and variational problems,
in domains lacking any regularity, to which the theory of Sobolev spaces available in the
existing literature does not apply. In this connection, let us mention the following simple
instance, which yet gives the avor of the generality allowed by the use of an inequality
like (1.2). Consider the problem

2 u = div F in ,
(1.3)
u = 0, B u = 0 on ,
where is any open set in Rn , n 3, 2 is the bi-Laplace operator, F : Rn

is a given function, and B is the (second-order) boundary operator generated by the
minimization of the quadratic functional

2 2
| u| + 2F u dx

among functions u vanishing on . Here, the dot stands for scalar product in Rn .
2n
Assume that F L n+2 (), where n+2
2n
is the Hlder conjugate of the critical Sobolev
2n
exponent n2 . A special case of Theorem 6.3, Section 6, tells us that there exists a
constant C = C(n) such that the Sobolev inequality
u 2n C2 uL2 ()

L n2 ()
holds for every function u vanishing on . (Incidentally, notice that, instead, an inequal-
ity of this kind fails if 2 u is replaced just with u, unless is suciently regular.)
A standard argument relying upon Riesz representation theorem in Hilbert spaces then
yields the existence of a unique solution u to (1.3), whatever is.
An analysis of more general problems in arbitrary domains goes beyond the scope of
the present contribution, and is the subject of a work in progress.
The paper is organized as follows. In the next section we oer a brief overview of
some Sobolev type inequalities, in basic cases, which follow from our results, and we
discuss their novelty and optimality. Section 3 contains some preliminary denitions and
results. The statement of our main results starts with Section 4, which is devoted to
our key pointwise inequalities for Sobolev functions on arbitrary open sets. Estimates
in rearrangement form are derived in the subsequent Section 5. In Section 6, Sobolev
type inequalities in arbitrary open sets are shown to follow via such estimates. Examples
which demonstrate the sharpness of our results are exhibited in Section 7. In particular,
Example 7.4 shows that inequalities of the form (1.2) may possibly fail if N (u) only
depends on derivatives of u on up to an order smaller than [ m1 2 ]. Finally, in the
Appendix, some new notions, which are introduced in the denitions of the seminorms
N (), are linked to classical properties of Sobolev functions.
2. A taste of results
In order to give an overall idea of the content of this paper, we enucleate hereafter a
few basic instances of the inequalities that can be derived via our approach.
We begin with two examples which demonstrate that our conclusions lead to new
results also in the case of rst-order inequalities, namely in the case when m = 1 in (1.2).
Let be any open set in Rn , and let be a Borel measure on such that (Br )
Cr for some C > 0, and (n 1, n], and for every ball Br with radius r. Clearly, if
= Ln , then this condition holds with = n.
Assume that 1 < p < n and r > 1, and let s = min{ n1 r p
, np }. Then

uLs (,) C uLp () + uLr () (2.1)
for some constant C and every function u with bounded support, provided that
Ln () < , () < and Hn1 () < . Here, Hn1 denotes the (n 1)-dimensional
Hausdor measure. In particular, if r = p(n1) p
np , and hence s = np , then (2.1) holds even
if the assumption on the niteness of these measures is dropped; in this case, the constant
C depends only on n, p, . Inequality (2.1) follows via a general principle contained in
Theorem 6.1, Section 6. It extends a version of the Sobolev inequality for measures, on
regular domains [62, Theorem 1.4.5]. It also augments, as far as measures are concerned,
the result of [57], whose approach, though yielding sharp constants, is conned to norms
evaluated with respect to the Lebesgue measure. Let us point out that, by contrast,
our method, being based on representation formulas, need not lead to optimal norms in
inequalities with p = 1.
Consider now the borderline case corresponding to p = n. As a consequence of Theo-
rem 6.1 again, one can show that

uexp L n1
n
(,)
C uL n () + u n
exp L n1 ()
, (2.2)
for some constant C and every function u with bounded support, provided that
Ln () < , () < and Hn1 () < . Here, exp L n1 n
(,)
and exp L n1
n
()
denote norms in Orlicz spaces of exponential type on and , respectively. In-
equality (2.2) on the one hand extends the YudovichPohozaevTrudinger inequality
to possibly irregular domains; on the other hand, it enhances, under some respect, a re-
sult of [58], where optimal constants are exhibited in estimates involving the weaker
norm in exp L(), and just for the Lebesgue measure.
Let us now turn to higher-order inequalities. Focusing, for the time being, on second-
order inequalities may help to grasp the quality and sharpness of our conclusions in this
framework. In the remaining part of this section, we thus assume that m = 2 in (1.2);
we also assume, for simplicity, that = Ln .
First, assume that h = 0. Then we can prove (among other possible choices of the
exponents) that, if 1 < p < n2 , then

u pn C 2 uLp () + u p(n1) + u p(n1) , (2.3)
L n2p () V 1,0 L np () L n2p ()
for some constant C = C(p, n), in particular independent of , and every function u
pn
with bounded support. Note that n2p is the same critical Sobolev exponent as in the
case of regular domains. Here, V 1,0 Lr () denotes, for r [1, ], the seminorm given
by
uV 1,0 Lr () = inf gLr () , (2.4)

g
where the inmum is taken among all Borel functions g on such that
|u(x) u(y)| |x y|(g(x) + g(y)) for Hn1 -a.e. x, y , (2.5)
and Lr () denotes a Lebesgue space on with respect to the measure Hn1 . The
function g appearing in (2.5) is an upper gradient, in the sense of [38], for the restriction
of u to , endowed with the metric inherited from the Euclidean metric in Rn, and
with the measure Hn1 . In [38], a denition of this kind, and an associated seminorm
given as in (2.4), were introduced to dene rst-order Sobolev type spaces on arbitrary
metric measure spaces. In the last two decades, various notions of upper gradients and
of Sobolev spaces of functions dened on metric measure spaces, have been the object
of investigations and applications. They constitute the topic of a number of papers and
monographs, including [4,11,35,39,42,53].
Let us stress that, although the new term u p(n1) on the right-hand side
V 1,0 L np ()
of (2.3) can be dropped when is a regular, say Lipschitz, domain, it is indispensable
if is arbitrary. This can be shown by taking into account a domain as in Fig. 1 (see
Example 7.1, Section 7).
As in the case of regular domains, if p > n2 , then the Lebesgue norm on the left-hand
side of (2.3) can be replaced by the norm in L . Indeed, if r > n 1, then

uL () C 2 uLp () + uV 1,0 Lr () + uL () (2.6)
for any open set such that Ln () < and Hn1 () < , for some constant C, and
for any function u with bounded support. In particular, the constant C depends on
only through Ln () and Hn1 ().
In the limiting situation when n 3, p = n2 and r > n 1, a YudovichPohozaev
Trudinger type inequality of the form

uexp L n2
n
()
C 2 uL n2 () + uV 1,0 Lr () + uexp L n2
n
()
(2.7)
Fig. 1. Example 7.1, Section 7.
holds for some constant C independent of the geometry of , and every function u
with bounded support, provided that Ln () < and Hn1 () < . The norms
exp L n2
n
()
and exp L n2
n
()
are the same exponential norms appearing in the
YudovichPohozaevTrudinger inequality on regular domains, and in its boundary trace
counterpart.
Consider next the case when still m = 2 in (1.2), but h = 1. From our estimates one
can infer that, if 1 < p < n and r 1, and is any open set with Ln () < and
Hn1 () < , then

uLq () C 2 uLp () + uV 1,0 Lr () (2.8)
for some constant C independent of the geometry of , and every function u with
bounded support, where

rn np

q = min n1 , np . (2.9)
In particular, if r = p(n1) np
np , and hence q = np , then the constant C in (2.8) depends
only on n and p.
Inequality (2.8) is optimal under various respects. For instance, if is regular, then, as
a consequence of (1.1), the seminorm uV 1,0 Lr () can be replaced just with uLr ()
on the right-hand side. By contrast, this is impossible in a domain as in Fig. 2, for
any q [1, np
np
], whatever r is see Example 7.2, Section 7.
The question of the optimality of the exponent q given by (2.9) can also be raised.
The answer is armative. Actually, domains like that of Fig. 3 show that such exponent
q is the largest possible in (2.8) if no regularity is imposed on (Example 7.3, Section 7).
When p > n, inequality (2.8) can be replaced with

uL () C 2 uLp () + uV 1,0 L () , (2.10)
for some constant C independent of the geometry of , and every function u with
bounded support, provided that Ln () < and Hn1 () < .
Finally, in the borderline case corresponding to p = n, an exponential norm is involved

again. Under the assumption that Ln () < and Hn1 () < , one has that

uexp L n1
n
()
C 2 uLn () + uV 1,0 exp L n1
n
()
(2.11)
for some constant C, depending on only through Ln () and Hn1 (), and for every
function u with bounded support. Here, the seminorm V 1,0 exp L n1
n
()
is dened as
in (2.4), with the norm Lr () replaced with the norm exp L n1
n
()
. Again, the
exponential norms in (2.11) are the same optimal Orlicz target norms for Sobolev and
trace inequalities, respectively, on regular domains.
3. Preliminaries
Let be any open set in Rn , n 2. Given x , dene
x = {y : (1 t)x + ty for every t (0, 1)}, (3.1)
and
()x = {y : (1 t)x + ty for every t (0, 1)}. (3.2)
They are the largest subset of and , respectively, which can be seen from x. It is
easily veried that x is an open set. The following proposition tells us that ()x is a
Borel set.
Proposition 3.1. Assume that is an open set in Rn , n 2. Let x . Then the set
()x , dened by (3.2), is Borel measurable.
Proof. Given any r Q (0, 1), dene
()x (r) = {y : (1 t)x + ty for every t (0, r)}.
If y ()x (r), then there exists > 0 such that B (y) ()x (r). Thus, for
each r Q (0, 1), the set ()x (r) is open in , in the topology induced by Rn . The
conclusion then follows from the fact that ()x = rQ(0,1) ()x (r). 2
Next, we dene the sets
( Sn1 )0 = {(x, ) Sn1 : x + t for some t > 0}, (3.3)
and
( Sn1 ) = ( Sn1 ) \ ( Sn1 )0 . (3.4)

Clearly,
( Sn1 )0 = Sn1 if is bounded. (3.5)
Let
: ( Sn1 )0 Rn (3.6)
be the function dened as
(x, ) = x + t, where t is such that x + t ()x .
In other words, (x, ) is the rst point of intersection of the half-line {x + t : t > 0}
with .
Given a function g : R, with compact support, we adopt the convention that
g((x, )) is dened for every (x, ) Sn1 , on extending it by 0 on ( Sn1 ) ;
namely, we set
g((x, )) = 0 if (x, ) ( Sn1 ) . (3.7)
Let us next introduce the functions
a : Sn1 [, 0) and b : Sn1 (0, ] (3.8)
given by

|(x, ) x| if (x, ) ( Sn1 )0 ,
b(x, ) = (3.9)
otherwise,
and
a(x, ) = b(x, ) if (x, ) Sn1 . (3.10)
Proposition 3.2. The function is Borel measurable. Hence, the functions a and b are
Borel measurable as well.
Proof. Assume rst that is bounded, so that ( Sn1 )0 = Sn1 . Consider a

sequence of nested polyhedra {Qk } invading , and the corresponding sequence of func-
tions {k }, dened as , with replaced with Qk . Such functions are Borel measurable,
by elementary considerations, and hence is also Borel measurable, since k converges
to pointwise.
Next, assume that is unbounded. For each h N, consider the set h = Bh (0),
where Bh (0) is the ball, centered at 0, with radius h. Let h and bh be the functions,
dened as and b, with replaced with h . Since h is bounded, then we already know
that bh is Borel measurable. Moreover, bh converges to b pointwise. Hence, b is Borel

measurable as well, and in particular the set ( Sn1 )0 , which agrees with {b < },
is Borel measurable. Finally, the function h is Borel measurable, inasmuch as h is a
bounded set. Moreover, h converges to pointwise on the Borel set ( Sn1 )0 . Thus,
is Borel measurable. 2
Given m N and p [1, ], we denote by V m,p () the Sobolev type space dened
as

V m,p () = u : u is m-times weakly dierentiable in , and |m u| Lp () .
(3.11)
Let us notice that, in the denition of V m,p (), it is only required that the derivatives
of u of the highest order m belong to Lp (). Replacing Lp () in (3.11) with a more
general Banach function space X() leads to the notion of m-th order Sobolev type
space V m X() built upon X().
For k N0 , we denote as usual by C k () the space of real-valued functions whose
k-th order derivatives in are continuous up to the boundary. We also set
Cbk () = {u C k () : u has bounded support}. (3.12)
Clearly,
Cbk () = C k () if is bounded.
Let = (1 , . . . , n ) be a multi-index with i N0 for i = 1, . . . , n. We adopt the

notations || = 1 + +n , ! = 1 ! n !, and =
1 n for R . Moreover,
1 n n
|| u
we set D u = x1 ...xnn for u : R.

1
We need to extend the notion of upper gradient g for the restriction of u to
appearing in (2.5) to the case of higher-order derivatives. To this purpose, let us denote
by g k,j , where k N0 and j = 0, 1, (k, j) = (0, 0), any Borel function on such that:
(i) If k N, j = 0, and u Cbk1 (),

(2k 2 ||)! (y x)
(1) D u(y) D u(x)
||
(k 1 ||)!! |y x|2k1
||k1
g k,0 (x) + g k,0 (y) (3.13)
for Hn1 -a.e. x, y .

(ii) If k N, j = 1, and u Cbk (),
n

(2k 2 ||)! (y x)
(1) D xi (y) D xi (x)
|| u u
(k 1 ||)!! |y x|2k1
i=1 ||k1
g k,1 (x) + g k,1 (y) (3.14)
for Hn1 -a.e. x, y .

(iii) If k = 0, j = 1, and u Cb0 (),
|u(x)| g 0,1 (x) (3.15)
for Hn1 -a.e. x .
Note that inequality (3.13), with k = 1, agrees with (2.5), and hence g 1,0 has the same
role as g in (2.5). Let us also point out that, as (2.5) extends a classical property of the
gradient of weakly dierentiable functions in Rn , likewise its higher-order versions (3.13)
and (3.14) extend a parallel property of functions in Rn endowed with higher-order weak
derivatives. This is shown in Proposition A.1 of the Appendix.
In analogy with (2.4), we introduce the seminorm given, for r [1, ], by
uV k,j Lr () = inf g k,j Lr () (3.16)

g k,j
where k, j and u are as above, and the inmum is extended over all functions g k,j
fullling the appropriate denition among (3.13), (3.14) and (3.15). More generally, given
a Banach function space Z() on with respect to the Hausdor measure Hn1 , we
dene
uV k,j Z() = inf g k,j Z() . (3.17)

g k,j
Observe that, in particular,
uV 0,1 Z() = uZ() .
4. Pointwise estimates
In the present section we establish our rst main result: a pointwise estimate for
Sobolev functions, and their derivatives, in arbitrary open sets. In what follows we dene,
for k N,

0 if k is odd,
(k) = (4.1)
1 if k is even.
Theorem 4.1 (Pointwise estimate). Let be any open set in Rn , n 2. Assume that
m N and h N0 are such that 0 < mh < n. Then there exists a constant C = C(n, m)
such that
g [ k+h+1
mh1
|m u(y)| 2 ],(k+h)
((y, ))
| u(x)| C
h
dy + dHn1 () dy
|x y|nm+h |x y|nk
k=1 Sn1

h+1
+ g[ 2 ],(h) ((x, )) dHn1 () for a.e. x , (4.2)
Sn1
[ m1 ] k+h+1
for every u V m,1 () Cb 2 (). Here, g [ 2 ],(k+h)
is any function as in
(3.13)(3.15), and convention (3.7) is adopted.
Remark 4.2. In the case when m h = n, and is bounded, an estimate analogous

1
to (4.2) can be proved, with the kernel |xy|nm+h in the rst integral on the right-hand
C
side replaced with log |xy| . The constant C depends on n and the diameter of . If
m h > n, and is bounded, then the kernel is bounded by a constant depending on
n, m and the diameter of .
Remark 4.3. Under the assumption that
m1
u = u = = [ 2 ] u = 0 on , (4.3)
h+1
one can choose g [ 2 ],(h) = 0 for k = 0, . . . , m h 1 in (4.2). Hence,

|m u(y)|
|h u(x)| C dy for a.e. x . (4.4)
|x y|nm+h

A special case of (4.4), corresponding to h = m 1, is the object of [62, Theorem 1.6.2].

Remark 4.4. As already mentioned in Section 1, the order m1 2 of the derivatives
prescribed on , which appears on the right-hand side of (4.2), is minimal for Sobolev
type inequalities to hold in arbitrary domains. This issue is discussed in Example 7.4,
Section 7 below.
A key step in the proof of Theorem 4.1 is contained in the next lemma, which deals
with the case when h = m 1 in Theorem 4.1.
Lemma 4.5. Let be any open set in Rn , n 2.
(i) If u V 21,1 () Cb1 () for some N, then

|22 u(x)|

|21 u(y)|
C dy + g 1,1
((x, )) dH n1
() for a.e. x , (4.5)
|x y|n1
Sn1
for some constant C = C(n, ).

(ii) If u V 2,1 () Cb1 () for some N, then

|2 u(y)|
|21 u(x)| C dy + g ,0 ((x, )) dHn1 () (4.6)
|x y|n1
Sn1
for some constant C = C(n, ). Here, g 1,1 and g ,0 are functions as in
(3.13)(3.15), and convention (3.7) is adopted.
Our proof of Lemma 4.5 in turn requires the following representation formula for the
(2 1)-th order derivative of a one-dimensional function in an interval, in terms of
its 2-th derivative in the relevant interval, and of its derivatives up to the order 1
evaluated at the endpoints.
Lemma 4.6. Let < a < b < . Assume that W 2,1 (a, b) for some N. Then
(21) (t)
t 2 a b b a + b 2
= Q21 ( ) d Q21
(2)
(2) ( ) d
ba ba
a t

1
(2 k 2)! 1
+ (2 1)!(1) (1)k+1 (k) (b) + (k) (a) (4.7)
k!( k 1)! (b a) 2k1
k=0
for t (a, b). Here, (k) denotes the k-th order derivative of , and Q21 is the poly-
nomial of degree 2 1, obeying
Q21 (t) + Q21 (t) = 1 for t R, (4.8)
and
(1) (1)
Q21 (1) = Q21 (1) = = Q21 (1) = 0. (4.9)
Proof. Let us represent as
(t) = (t) + (t) for t (a, b), (4.10)

where and are the solutions to the problems

(2) (t) = (2) (t) in (a, b),
(4.11)
(k) (a) = (k) (b) = 0 for k = 0, 1, . . . , 1,
and

(2) (t) = 0 in (a, b),
(4.12)
(k) (a) = (k) (a), (k) (b) = (k) (b) for k = 0, 1, . . . , 1,
respectively. Let us rst focus on problem (4.11). We claim that
t 2 a b
(21)
(t) = Q21 (2) ( ) d
ba
a
b a + b 2
Q21 (2) ( ) d for t (a, b), (4.13)
ba
t
where Q21 is as in the statement. In order to verify (4.13), let us consider the auxiliary
problem

2 (s) = (s) in (1, 1),
(4.14)
(k) (1) = 0, k = 0, 1, . . . , 1,
where L1 (a, b) is any given function. Let : [1, 1]2 R be the Green function
associated with problem (4.14), so that
1
(s) = (s, r)(r) dr for s [1, 1]. (4.15)
1
The function takes an explicit form ([14]; see also [36, Section 2.6]), given by
1sr

|sr|
(s, r) = C|s r| 21

(t2 1)1 dt for s = r, (4.16)
1
where C = C() is a suitable constant. One can easily see from formula (4.16) that (s, r)
is a polynomial of degree 2 1 in s for xed r, and a polynomial of degree 2 1 in r
for xed s, both in {(s, r) [1, 1]2 : s > r}, and in {(s, r) [1, 1]2 : s < r}. Moreover,
(s, r) = (s, r). In particular, if s > r, one has that
1

1 (1)1j
(s, r) = C (1 sr)2j+1 (s r)22k2
j=0
j 2j + 1
1

1 (1)1j
(s r)21 . (4.17)
j=0
j 2j + 1
Thus, if s > r,
1
1
21 1 (1)+j 2j+1 1 (1)1j
(s, r) = C(2 1)! r ,
s21 j=0
j 2j + 1 j=0
j 2j + 1
(4.18)
a polynomial of degree 2 1 in r, depending only on odd powers of r. Let us denote

this polynomial by Q21 (r). It follows from (4.18) that Q21 (1) = 0. Moreover,
1

Q21 21 1
(r) = (s, r) = C(2 1)! (1)+j r2j
r r s21 j=0
j
= C(2 1)!(r2 1)1 . (4.19)
Thus, Q21 vanishes, together with all its derivatives up to the order 1, at 1,
namely Q21 fullls (4.9). Equation (4.18) also tells us that Q21 (s) Q21 (0) is an
odd function, and hence
Q21 (s) + Q21 (s) = 2Q21 (0) for s R. (4.20)
Since is an even function,
21 21
s21
(s, r) = s21
(s, r) = Q21 (r) if 1 s < r 1.
Thus, (2 1)-times dierentiation of equation (4.15) yields
s 1
(21)
(s) = Q21 (r)(r) dr Q21 (r)(r) dr for s [1, 1]. (4.21)
1 s
Since (2) = , an integration by parts in (4.21), equation (4.20), and the fact that
Q21 (1) = 0, tell us that
s
(21)
(s) = 2Q21 (0) (21)
(s) Q21 (r) (21) (r) dr
1
1
Q21 (r) (21) (r) dr for s [1, 1]. (4.22)
s
Owing to the arbitrariness of , equation (4.22) ensures that 2Q21 (0) = 1. Equa-
tion (4.8) thus follows from (4.20).
The function dened as

(t) = 2tab
ba for t [a, b]
is thus the solution to problem (4.11), and the representation formula (4.13) follows via
a change of variables in (4.21).
Consider next problem (4.12). The function is a polynomial of degree 2 1, and
(21) is a constant which, owing to the two-point Taylor interpolation formula (see
e.g. [30, Chapter 2, Section 2.5, Ex. 3]), is given by
1
d (t) d1 (t)
(21) = (2 1)! +
dt1 (t b) |t=a dt1 (t a) |t=b
1 1

d (t) d (t)
= (2 1)! + . (4.23)
Leibnitz dierentiation rule for products yields

d1 (t) d1 (t)
+

1
(2 k 2)! 1
= (1) (1)k+1 (k) (b) + (k) (a) . (4.24)
k!( k 1)! (b a) 2k1
k=0
Equation (4.7) follows from (4.13), (4.23) and (4.24). 2
Proof of Lemma 4.5. Given x and Sn1 , let a(x, ) and b(x, ) be dened as
in (3.10) and (3.9), respectively.
We begin with the proof of (4.5) for = 1. If u V 1,1 () Cb0 (), then, by a standard
property of Sobolev functions, for a.e. x the function
[0, b(x, )] t u(x + t)
belongs to V 1,1 (0, b(x, )) for Hn1 -a.e. Sn1 , and
d
u(x + t) = u(x + t) for a.e. t [0, b(x, )].
dt
Hence, for any such x and ,

b(x,)
u((x, )) u(x) = u(x + t) dt, (4.25)

0
where convention (3.7) is adopted. Integrating both sides of equation (4.25) over Sn1
yields

b(x,)
nn u(x) = u((x, )) dH n1
() u(x + t) dt dHn1 (), (4.26)
Sn1 Sn1 0
n
where n = 2 /(1 + n2 ), the Lebesgue measure of the unit ball in Rn . One has that

b(x,)
b(x,)
1
u(x + t) dt dH n1
() = u(x + ) tn1 d dHn1 ()
tn1
Sn1 0 Sn1 0

u(y) (y x)
= dy. (4.27)
|x y|n
x
Inequality (4.5), with = 1, follows from (4.26) and (4.27).

Let us next prove (4.6). If u V 2,1 () Cb1 (), then for a.e. x , the function
[a(x, ), b(x, )] t u(x + t)
belongs to V 2,1 (a(x, ), b(x, )) for Hn1 -a.e. Sn1 . Consider any such x and . If
(x, ) ( Sn1 )0 , then, by Lemma 4.6,
t 2 a(x, ) b(x, ) d2

d21
u(x + t) = Q21 u(x + ) d
dt21 b(x, ) a(x, ) d 2
a(x,)

b(x,)
a(x, ) + b(x, ) 2 d2
Q21 u(x + ) d
b(x, ) a(x, ) d 2
t
+ (2 1)!(1)
k
k+1 d u(x+t)
k

1
(2 k 2)! (1) dtk |t=b(x,)
+ d u(x+t)
dtk |t=a(x,)
(4.28)
k!( k 1)! (b(x, ) a(x, ))2k1
k=0
for t (a(x, ), b(x, )). If, instead, (x, ) ( Sn1 ) , then

d2
d21 t d 2 u(x + ) d, if b(x, ) = ,
u(x + t) = t d2 (4.29)
dt21 u(x + ) d, if a(x, ) = ,
d 2
for t (a(x, ), b(x, )) (if both b(x, ) = and a(x, ) = , then either expression
on the right-hand side of (4.29) can be exploited).
We have that
dk k!
u(x + t) = D u(x + t) for a.e. t (a(x, ), b(x, )), (4.30)
dtk !
||=k
for k = 1, . . . , 2. From (4.28)(4.30) we infer that
(2 1)!
D u(x)
!
||=21

(2 1)!
= (Sn1 ) (x, ) D+ u(x + ) d
!
||=21 0 ||=1
0 2 a(x, ) b(x, ) d2

+ (Sn1 )0 (x, ) Q21 u(x + )d
b(x, ) a(x, ) d 2
a(x,)

b(x,)
a(x, ) + b(x, ) 2 d2
Q21 u(x + ) d
b(x, ) a(x, ) d 2
0

1
(2 k 2)!
+ (2 1)!(1)
k!( k 1)!
k=0

k! k+1
D u(x + b(x, )) + D u(x + a(x, ))
(1)
, (4.31)
! (b(x, ) a(x, ))2k1
||=k
where the upper limit of integration in the rst integral on the right-hand side is either
+, or according to whether b(x, ) = or a(x, ) = .
Denote by {P } the system of all homogeneous polynomials of degree 2 1 in the
variables 1 , . . . , n such that

P () dHn1 () = ,
Sn1
P ()
where stands for the Kronecker delta. On multiplying equation (4.31) by (21)! , and
integrating over Sn1 one obtains that
D u(x)
!

+
= (Sn1 ) (x, )P () D+ u(x + ) d dHn1 ()
!
Sn1 ||=21 0
||=1

P ()
+ (Sn1 )0 (x, )
(2 1)!
Sn1
0 2 a(x, ) b(x, ) d2

Q21 u(x + )d
b(x, ) a(x, ) d 2
a(x,)

b(x,)
a(x, ) + b(x, ) 2 d2
Q21 u(x + ) d dHn1 ()
b(x, ) a(x, ) d 2
0

+ (1) P ()
Sn1

(2 || 2)! (1)
||+1
D u(x + b(x, )) + D u(x + a(x, ))
dHn1 ().
( || 1)!! (b(x, ) a(x, ))2||1
||1
(4.32)
There exist constants C = C(n, ) and C = C (n, ) such that

+
(Sn1 ) (x, )P () D +
u(x + ) d dH n1
()
!
Sn1 ||=21 0
||=1

|2 u(x + )| n1
C | u(x + )| d dH
2 n1
() = C d dHn1 ()
n1
Sn1 0 Sn1 0

| u(y)|
2
C dy. (4.33)
|x y|n1

Next, we claim that there exists a constant C = C(, n) such that
0 2 a(x, ) b(x, ) d2

(x, ) P () Q u(x + )d
(S n1 )0 21
b(x, ) a(x, ) d 2
Sn1 a(x,)

b(x,)
a(x, ) + b(x, ) 2 d2

P () Q21 u(x + )d dH n1
()
b(x, ) a(x, ) d 2
0

| u(y)|
2
C dy. (4.34)
|x y|n1

In order to prove (4.34), observe that


(Sn1 )0 (x, )P ()
Sn1
0 2 a(x, ) b(x, ) d2

Q21 u(x + )d dHn1 ()
b(x, ) a(x, ) dr2
a(x,)

= (Sn1 )0 (x, )P ()
Sn1
a(x,)
2r + a(x, ) + b(x, ) d2
Q21 u(x r)dr dHn1 ()
b(x, ) a(x, ) d 2
0

= (Sn1 )0 (x, )P ()
Sn1
a(x,)
2r + a(x, ) + b(x, ) d2
Q21 u(x + r)dr dHn1 ()
b(x, ) a(x, ) dr2
0

= (Sn1 )0 (x, )P ()
Sn1

b(x,)
a(x, ) + b(x, ) 2r d2
Q21 u(x + r)dr dHn1 (),
b(x, ) a(x, ) dr2
0
where we have made use of the fact that P () = P () if || = 2 1, and of (3.10).

Thus,
0 2 a(x, ) b(x, ) d2

(Sn1 )0 (x, ) P () Q21 u(x + )d
b(x, ) a(x, ) d 2
Sn1 a(x,)

b(x,)
a(x, ) + b(x, ) 2 d2

P () Q21 u(x + )d dH n1
()
b(x, ) a(x, ) d 2
0

b(x,)
a(x, ) + b(x, ) 2r d2

= 2 P () Q21 u(x + r) drdH n1
()
b(x, ) a(x, ) dr2
Sn1 0

= 2 P ()
Sn1

b(x,)
a(x, ) + b(x, ) 2r (2)!

Q21 D u(x + r) drdHn1 ()
b(x, ) a(x, ) !
0 ||=2

b(x,)
|2 u(y)|
C | u(x + r)| drdH
2 n1
() C dy
|x y|n1
Sn1 0 x
for some constants C = C(n, ) and C = C (n, ). Hence, inequality (4.34) follows.
Finally, by denition (3.13), there exists a constant C = C(n, ) such that

(Sn1 )0 (x, )(1) P ()

Sn1
(2 || 2)!

( || 1)!!
||1

(1)||+1 D u(x + b(x, )) + D u(x + a(x, ))
dH n1
()
(b(x, ) a(x, )) 2||1

C (Sn1 )0 (x, ) g ,0 (x + a(x, )) + g ,0 (x + b(x, )) dHn1 ()
Sn1

= 2C g ,0 ((x, )) dHn1 (). (4.35)
Sn1
Combining (4.32)(4.35) yields (4.6).

Inequality (4.5), with 2, follows on applying (4.6) with u replaced with its rst-
order derivatives. 2
Proof of Theorem 4.1. For simplicity of notation, we consider the case when h = 0, the
[ m1
2 ]
proof in the general case being analogous. Let u V m,1 () Cb (). By inequal-
ity (4.5) with = 1,

|u(y)|
|u(x)| C dy + g 0,1 ((x, )) dHn1 () for a.e. x . (4.36)
|x y|n1
Sn1
From (4.36) and an application of inequality (4.6) with = 1 one obtains that

|2 u(z)| dz dy
|u(x)| C
|y z|n1 |x y|n1

1,0 dHn1 () dy 0,1 n1
+ g ((y, )) + g ((x, )) dH ()
|x y|n1
Sn1 Sn1

|2 u(z)| dHn1 () dy
C dz + g 1,0 ((y, ))
|x z|n2 |x y|n1
Sn1

+ g 0,1
((x, )) dH n1
() for a.e. x , (4.37)
Sn1
for some constants C = C(n) and C = C (n). Note that in the last inequality we have
made use of a special case of the well known identity

1 f (z) f (z)
dz dy = C dz for a.e. x Rn , (4.38)
|x y|n |y z|n |x z|n
Rn Rn Rn
which holds for some constant C = C(n, , ) and for every compactly supported inte-
grable function f , provided that > 0, > 0 and + < n.
Inequality (4.37) in turn yields, via an application of inequality (4.5) with = 2,

|3 u(z)| dHn1 () dy
|u(x)| C dz + g 1,1 ((y, ))
|x z|n3 |x y|n2
Sn1

dHn1 () dy
+ g 1,0 ((y, )) + g 0,1 ((x, )) dHn1 ()
|x y|n1
Sn1 Sn1
for a.e. x , (4.39)
for some constant C = C(n). A nite induction argument, relying upon an alternate
iterated use of inequalities (4.6) and (4.5) as above, eventually leads to (4.2). 2
5. Estimates in rearrangement form
The pointwise bounds established in the previous section enable us to derive rearrange-
ment estimates for functions, and their derivatives, with respect to any Borel measure
on such that
(Br (x) ) C r for x and r > 0, (5.1)
for some (n 1, n] and some constant C > 0. Here, Br (x) denotes the ball, centered
at x, with radius r.
Recall that, given a measure space R, endowed with a positive measure , the de-
creasing rearrangement : [0, ) [0, ] of a -measurable function : R R is
dened as
(s) = inf{t 0 : (|| > t}) s} for s [0, ).
The operation of decreasing rearrangement is not linear. However, one has that
( + ) (s) (s/2) + (s/2) for s 0, (5.2)
for every measurable functions and on R.

Any function shares its integrability properties with its decreasing rearrangement ,
since
({|| > t}) = L1 ({ > t}) for every t 0.
As a consequence, any norm inequality, involving rearrangement-invariant norms, be-

tween the rearrangements of the derivatives of Sobolev functions and the rearrangements
of its lower-order derivatives, immediately yields a corresponding inequality for the orig-
inal Sobolev functions. Thus, the rearrangement inequalities to be established hereafter
reduce the problem of n-dimensional Sobolev type inequalities in arbitrary open sets to
considerably simpler one-dimensional Hardy type inequalities see Theorem 6.1, Sec-
tion 6 below.
Theorem 5.1 (Rearrangement estimates). Let be any open bounded open set in Rn ,
n 2. Let m N and h N0 be such that 0 < m h < n. Assume that is a Borel
measure in fullling (5.1) for some (n 1, n] and for some C > 0. Then there
exist constants c = c(n, m) and C = C(n, m, , C ) such that
n
s
h nm+h
r n |m u|Ln (r)dr
nm+h
| u| (cs) C s | u|Ln (r)dr +
m
0 n
s
n1

s
mh1
n1k
[ k+h+1 ],(k+h)
+ s g 2 Hn1
(r)dr
k=1 0

r
n1k k+h+1
+ n1 g[ 2 ],(k+h)
Hn1
(r)dr
n1
s
n1
s
n1
[ h+1 ],(h)
+s g 2 Hn1
(r)dr for s > 0, (5.3)
0
[ m1 ] k+h+1
for every u V m,1 () Cb 2 (). Here, () is dened as in (4.1), and g [ 2 ],(k+h)
denotes any Borel function on fullling the appropriate condition from (3.13)(3.15).
Remark 5.2. In inequality (5.3), and in what follows, when considering rearrangements
and norms with respect to a measure , Sobolev functions and their derivatives have
to be interpreted as their traces with respect to . Such traces are well dened, thanks
to standard (local) Sobolev inequalities with measures, owing to the assumption that
(n 1, n] in (5.1). An analogous convention applies to the integral operators to be

considered below.
In preparation for the proof of Theorem 5.1, we introduce a few integral operators,
and prove pointwise estimates for their rearrangements.
Let be any open set in Rn . We dene the operator T as

Tg(x) = |g((x, ))| dHn1 () for x , (5.4)
Sn1
at any Borel function g : R. Here, and in what follows, we adopt convention (3.7).
Note that, owing to Fubinis theorem, Tg is a measurable function with respect to any
Borel measure in .
For (0, n), we denote by I the classical Riesz potential operator given by

f (y)
I f (x) = dy for x , (5.5)
|y x|n

at any f L1 (), and we call N the operator dened as

g(y)
N g(x) = dHn1 (y) for x , (5.6)
|x y|n

at any function g L1 ().

Finally, we dene the operator Q as the composition
Q = I T. (5.7)
Namely,

dHn1 () dy
Q g(x) = |g((y, ))| for x , (5.8)
|x y|n
Sn1
for any Borel function g : R.

Our analysis of these operators requires a few notations and properties from interpo-
lation theory.
Assume that R is a measure space, endowed with a positive measure . Given a pair
X1 (R) and X2 (R) of normed function spaces, a function X1 (R) +X2 (R), and s R,
we denote by K(, s; X1 (R), X2 (R)) the associated Peetres K-functional, dened as

K(s, ; X1 (R), X2 (R)) = inf 1 X1 (R) + s2 X2 (R) for s > 0.
=1 +2
We need an expression for the K-functional (up to equivalence) in the case when X1 (R)
and X2 (R) are certain Lebesgue or Lorentz spaces, and R is one of the measure spaces
mentioned above. Recall that, given > 1, the Lorentz space L,1 (R) is the Banach
function space of those measurable functions on R for which the norm

(s)s1+ ds
1
L,1 (R) =
0
is nite. The Lorentz space L, (R), also called Marcinkiewicz space or weak-L space,
is the Banach function space of those measurable functions on R for which the quantity
L, (R) = sup s (s)

1
s>0
is nite. Note that, in spite of the notation, this is not a norm. However, it is equivalent
to a norm, up to multiplicative constants depending on , obtained on replacing (s)
s
with 1s 0 (r)dr.
It is well known that
s

1
K(, s; L (R), L (R)) = (r) dr for s > 0, (5.9)
0
for every L1 (R) + L (R) [9, Chapter 5, Theorem 1.6]. If > 1, then
K(, s; L, (R), L (R)) r (r)L (0,s )

1
for s > 0, (5.10)
for every L, (R) + L (R) [44, Equation (4.8)]. Moreover,

s
(r) dr r (r) dr
1
1
K(, s; L (R), L ,1
(R)) +s for s > 0, (5.11)
0 s
for every L1 (R) + L,1 (R) [44, Theorem 4.2]. In (5.10) and (5.11), the notation
means that the two sides are bounded by each other up to multiplicative constants
depending on .
Let R and S be positive measure spaces. An operator L dened on a linear space
of measurable functions on R, and taking values into the space of measurable functions
on S, is called sub-linear if, for every 1 and 2 in the domain of L and every 1 , 2 R,
|L(1 1 + 2 2 )| |1 ||L1 | + |2 ||L2 |.
A basic result in the theory of real interpolation tells us what follows. Assume that L
is a sub-linear operator as above, and Xi (R) and Yi (S), i = 1, 2, are normed function
spaces on R such that
L : Xi (R) Yi (S) (5.12)
with norms not exceeding Ni , i = 1, 2. Here, the arrow denotes a bounded operator.
Then,
K(L, s; Y1 (S), Y2 (S)) max{N1 , N2 }K(, s; X1 (R), X2 (R)) for s > 0, (5.13)
for every X1 (R) + X2 (R).
Lemma 5.3. Let be an open set in Rn , n 2, and let (0, n). Assume that is any
Borel measure in fullling (5.1) for some (n , n] and for some C > 0. Then
there exists a constant C = C(n, , , C ) such that
N g , CgL1 () (5.14)
L n (,)
for every g L1 ().
Proof. We make use of an argument related to [1,2]. Given g L1 () and t > 0, dene
Et = {x : |N g(x)| > t}, and denote by t the restriction of the measure to Et .
By Fubinis theorem,

|g(y)|
t(Lt ) = t dt (x) |N g(x)|dt (x) dHn1 (y) dt (x)
|x y|n

dt (x)dHn1 (y)
|g(y)| . (5.15)
|x y|n

Next,

dt (x)
= (n ) n+1 dt (x) d
|x y|n
0 {x:|xy|n >n }

(n ) n+1 t (B (y)) d. (5.16)
0
From (5.15) and (5.16) we deduce that, for each xed r > 0,

t(Lt ) (n ) |g(y)| n+1 t (B (y)) d dHn1 (y)
0
r
n+1
= (n ) |g(y)|t (B (y)) dHn1 (y) d
0

n+1
+ (n ) |g(y)|t (B (y)) dHn1 (y) d. (5.17)
r
We have that
r
n+1 |g(y)|t (B (y)) dHn1 (y) d
0
r
n+1+ d = C rn+ gL1 () .
n+
C gL1 () (5.18)
0
On the other hand,

n+1
|g(y)|t (B (y)) dH n1
(y) d (Lt ) n+1 |g(y)| dHn1 (y) d
r r
n+
= (Lt ) rn gL1 () . (5.19)
1
(Lt )
Combining (5.17)(5.19), and choosing r = C , yield
rn+ rn+
t(Lt ) (n )gL1 () C + (Lt )
n+ n
n n
n+ gL () C (Lt )1

= 1
. (5.20)
Thus,
n n
t(Lt )
n+ C

gL1 () for t > 0. (5.21)
Hence, inequality (5.14) follows. 2
Proposition 5.4. Let be an open set in Rn . Then

|g(y)|
|g((x, ))| dHn1 () 2n dHn1 (y) for x , (5.22)
|x y|n1
Sn1
for every Borel function g : R. Here, convention (3.7) is adopted.
Proof. We split the proof in steps. Fix x .

Step 1. Denote by : Rn \ {x} Sn1 the projection function into Sn1 given by
yx
(y) = for y Rn \ {x}.
|y x|
Then
1
Hn1 ((E)) Hn1 (E) (5.23)
dist(x, E)n1
for every set E .

The function is dierentiable, and |(y)| |y x|1 for y Rn \ {x}. Thus, the
restriction of to E is Lipschitz continuous, and
1
|(y)| for y E. (5.24)
dist(x, E)
Inequality (5.24) implies (5.23), by a standard property of Hausdor measure see

e.g. [59, Theorem 7.5].
Step 2. We have that

dHn1 (y)
Hn1 ((E)) 2n for every Borel set E . (5.25)
|x y|n1
E
The following chain holds:

dHn1 (y)
= Hn1 ({y E : |x y|n+1 > t}) dt
|x y|n1
E 0

= Hn1 ({y E : |x y| < }) d( 1n )
0

k+1
2
= Hn1 ({y E : |x y| < }) d( 1n )

kZ
2k
2k+1

Hn1 ({y E : |x y| < 2k }) d( 1n )
kZ
2k

= (1 2n+1 ) 2k(n1) Hn1 ({y E : |x y| < 2k })
kZ

(1 2n+1 ) 2k(n1) Hn1 ({y E : 2k1 |x y| < 2k }). (5.26)
kZ
Since dist(x, {y E : 2k1 |x y| < 2k }) 2k1 , by (5.23)
Hn1 ({y E : 2k1 |x y| < 2k })

C2(k1)(n1) Hn1 (({y E : 2k1 |x y| < 2k })) (5.27)
for k Z. From (5.26) and (5.27) we deduce that

dHn1 (y)
1
2k(n1) 2(k1)(n1) Hn1 (({y E : 2k1 |x y| < 2k }))
|x y|n1 2
E kZ

= 2n Hn1 (({y E : 2k1 |x y| < 2k }))
kZ
n

2 Hn1 kZ ({y E : 2k1 |x y| < 2k })

= 2n Hn1 (kZ {y E : 2k1 |x y| < 2k })
= 2n Hn1 ((E)), (5.28)
whence inequality (5.25) follows.

Step 3. Conclusion.
We have that
({y ()x : |g(y)| > t}) = { Sn1 : |g((x, ))| > t} for t > 0.
Thus, by (5.25),

dHn1 (y)
2n Hn1 ({ Sn1 : |g((x, ))| > t}) for t > 0.
|x y|n1
{y()x :|g(y)|>t}
Hence,

|g(y)| dHn1 (y)
dHn1 (y) = dt
|x y|n1 |x y|n1
()x 0 {y()x :|g(y)|>t}

n
2 Hn1 ({ Sn1 : |g((x, ))| > t}) dt
0

= 2n |g((x, ))| dHn1 (). (5.29)
Sn1
Inequality (5.22) is thus established. 2
Lemma 5.5. Let be an open set in Rn , n 2. Assume that is any Borel measure
in fullling (5.1) for some (n 1, n] and for some C > 0. Then there exists a
constant C = C(n, , C ) such that
n1
s
(Tg) (s) Cs
n1
gH n1 (r) dr for s > 0, (5.30)
0
for every Borel function g : R.

Proof. By Proposition 5.4, there exists a constant C = C(n) such that
Tg(x) CN1 |g|(x) for x , (5.31)
for every Borel function g : R. Hence, by Lemma 5.3, with = 1, there exists a
constant C = C(n, , C ) such that
T gL n1
,
(,)
CgL1 () (5.32)

On the other hand,

0 T g(x) gL () dHn1 () = nn gL () for x , (5.33)
Sn1
and hence
T gL (,) nn gL () (5.34)
for every Borel function g : R. We thus deduce from (5.9), (5.10), (5.13), (5.32)
and (5.34) that
s(Tg) (s n1 ) r (Tg) (r)L (0,s n1 K(Tg, s; L n1 , (, ), L (, ))

n1

)
s

CK(g, s; L (), L ()) = C
1
gH n1 (r) dr for s > 0, (5.35)
0
for some constant C = C(n, , C ), and for every Borel function g : R. Hence,
inequality (5.30) follows. 2
Lemma 5.6. Let be an open set in Rn , n 2, and let (0, n). Assume that is any
Borel measure in fullling (5.1) for some (n , n] and for some C > 0. Then,
n
s
(I f ) (s) n
fL n (r) dr r
n
C s + n fL n (r) dr for s > 0, (5.36)
0 n
s
for every f L1 ().
Proof. A standard weak-type inequality for Riesz potentials tells us that there exists a
constant C1 = C1 (n, , , C ) such that
I f , C1 f L1 () (5.37)
L n (,)
for every f L1 () (a proof of inequality (5.37) follows, in fact, along the same lines as
that of (5.14)). Furthermore, there exists a constant C2 = C2 (n, , , C ) such that
I f L (,) C2 f n ,1 (5.38)
L ()
n
for every f L ,1 (). Inequality (5.38) can be derived from (5.37), applied with = Ln
and = n, via a duality argument. Indeed,

|f (y)|
I f L () I |f |L () = sup |h(x)| dy dx
hL1 () 1 |y x|n

|h(x)|
= sup |f (y)| dy dx
hL1 () 1 |y x|n

sup Cf n ,1 I h n ,
L () L n ()
hL1 () 1
sup C f n ,1 hL1 () C f n ,1 (5.39)

L () L ()
hL1 () 1
n
for some constants C = C(n, ) and C = C (n, , , C ), and for every f L ,1 ().
Note that the rst inequality holds owing to a Hlder type inequality in Lorentz spaces.
As shown by a standard convolution argument, the space of continuous functions is
n
dense in L ,1 (). Inequality (5.39) then implies that the function I f is continuous if
n
f L ,1 (). Thus, I f L (,) I f L () , and (5.38) follows from (5.39).
By (5.37) and (5.38), via (5.10), (5.11) and (5.13), we deduce that there exists a
constant C = C(n, , , C ) such that
s(I f ) (s n ) r
n
(I f ) (r) K(I f, s; L n , (, ), L (, ))

L (0,s n )
n
CK(f, s; L1 (), L ,1 ())
n
s
n

n
fL n (r) dr +s r n fL n (r) dr for s > 0, (5.40)
0 n
s n
where the equivalence is up to multiplicative constants depending on n, , , C . Hence,

(5.36) follows. 2
Lemma 5.7. Let be an open set in Rn , n 2, and let (0, n 1). Assume that
is any Borel measure in fullling (5.1) for some (n , n] and for some C > 0.
Then there exists a constant C = C(n, , , C ) such that
n1
s
n1
(Q g) (s) C s n1

gH n1 (r) dr + r n1
gH n1 (r) dr for s > 0,
0 n1
s
(5.41)
Proof. By inequality (5.32), with = Ln , there exists a constant C = C(n) such that
T gL n1
n ,
()
CgL1 () (5.42)
for every Borel function g : R. Moreover, there exists a constant C = C(n, , , C )

such that
I f , Cf L n1
n , (5.43)
L n1 (,) ()
n
for every f L n1 , (). Indeed, by (5.36), for any such f ,
I f ,
L n1 (,)
n1
= sup s (I f ) (s)
s>0
n
s
n1
n
fL n (r) dr fL n (r) dr
1
C sup s +s r n
s>0
0 n
s
n
s
n1 n
n1
r n n
1 n1
Cf L n1
n ,
()
sup s r n dr + s dr
s>0
0 n
s
= C f L n1
n ,
()
, (5.44)
where C is the constant appearing in (5.36), and C = C (n, , , C ). If follows

from (5.42) and (5.43) that
Q g , CgL1 () , (5.45)
L n1 (,)
for some constant C = C(n, , , C ), and for every Borel function g : R.

On the other hand, by (5.30), applied with = Ln , there exists a constant C = C(n, )
such that
n1
s n

1+ n
1+ n n1
Tg n ,1 = (Tg)Ln (s)s ds C s n gH n1 (r) dr ds
L ()
0 0 0

1+ n n1 n
1+ n1
= gH n1 (r) s n ds dr = n1 gH n1 (r)r dr
0 n 0
r n1
n1 gL ,1 ()
n
= n1 (5.46)
for every Borel function g : R. Coupling inequalities (5.46) and (5.38) tells us that
Q gL (,) Cg n1 ,1

(5.47)
L ()

Now, by (5.10), (5.11), (5.13), (5.45) and (5.47), there exists a constant C =
C(n, , , C ) such that
s(Q g) (s n1 ) r
n1
(Q g) (r)

L (0,s n1 )
K(Q g, s; L n1 , (, ), L (, ))

n1
CK(g, s; L1 (), L ,1 ())
n1

s n1
n1

gH n1 (r) dr +s r n1
gH n1 (r) dr for s > 0,
0 n1
s n1
for every Borel function g : R, where equivalence holds up to multiplicative

constants depending on n, , , C . Inequality (5.41) follows. 2
Proof of Theorem 4.1. Inequality (4.2) can be written as

mh1
k+h+1 h+1
|h u(x)| C Imh (|m u|)(x) + Qk g [ 2 ],(k+h) (x) + T g [ 2 ],(h) (x)
k=1
for a.e. x .
Hence, (5.3) follows via Lemmas 5.55.7, owing to property (5.2) of rearrangements. 2
6. Sobolev inequalities
We present here a sample of Sobolev type inequalities that can be established via the
universal pointwise and rearrangement estimates of Sections 4 and 5, respectively. We
limit ourselves to inequalities for standard norms, such as Lebesgue norms and Orlicz
norms of exponential or logarithmic type, which naturally come into play when borderline
exponents in the Sobolev norms are taken into account. Measures satisfying (5.1) will
be included in our results. Let us emphasize, however, that inequalities for more general
norms can be derived from the relevant pointwise bounds. Virtually, any Sobolev type
inequality for rearrangement-invariant norms, which holds in regular domains, has a
counterpart in arbitrary domains, provided that appropriate boundary seminorms are
employed.
A key tool in our approach is the reduction principle to one-dimensional inequalities
stated in Theorem 6.1 below for Sobolev inequalities involving arbitrary rearrangement-
invariant norms. Recall that a rearrangement-invariant space X(R) on a measure
space R, endowed with a positive measure , is a Banach function space (in the sense of
Luxemburg) endowed with a norm X(R) such that
X(R) = X(R) whenever = . (6.1)
Every rearrangement-invariant space X(R) admits a representation space X(0, ),

namely another rearrangement-invariant space on (0, ) such that
X(R) = X(0,) for every X(R). (6.2)
In customary situations, an expression for the norm X(0,) immediately follows

from that of X(R) . The Lebesgue spaces and the Lorentz spaces, whose denition
has been recalled above, are standard instances of rearrangement-invariant spaces. The
exponential spaces, which have already been mentioned in Section 2, can be regarded as
special examples of Orlicz spaces. The Orlicz space LA (R) built upon a Young function
A : [0, ) [0, ], namely a left-continuous convex function which is neither identically
equal to 0 nor to , is a rearrangement-invariant space equipped with the Luxemburg
norm given by

|(x)|
LA (R) = inf >0 : dx 1 . (6.3)

R
The class of Orlicz spaces includes the Lebesgue spaces, since LA (R) = Lp (R) if A(t) = tp
for p [1, [, and LA (R) = L (R) if A(t) = (1,) (t). Given > 0, we denote by

exp L (R) the Orlicz space built upon a Young function equivalent to A(t) = et 1 near
innity. If either p > 1 and R, or p = 1 and 0, we denote by Lp (log L) (R) the
Orlicz space built upon the Young function A(t) = tp log (c + t), where c is a suciently
large positive number.
We refer to [9] for a comprehensive account of rearrangement-invariant spaces.
Theorem 6.1 (Reduction principle for Sobolev inequalities). Let be any open set in Rn ,
n 2. Assume that is a measure in fullling (5.1) for some (n 1, n], and for
some constant C . Let m N, and h N0 be such that 0 < m h < n. Assume that
X(), Y (, ) and Xk (), k = 0, , m h 1, are rearrangement-invariant spaces
such that

n

nm+h s

s (r)dr CX(0,) , (6.4)

0
Y (0,)

r n (r)dr
nm+h
CX(0,) , (6.5)

n
s Y (0,)

n1

nk1 s

s (r)dr CX k (0,) , k = 1, . . . , m h 1, (6.6)

0
Y (0,)

nk1
r n1 (r)dr CX k (0,) , k = 1, . . . , m h 1, (6.7)

n1
s Y (0,)

n1

n1 s

s (r) dr CX 0 (0,) , (6.8)

0
Y (0,)
for some constant C, and for every non-increasing function : [0, ) [0, ). Then
there exists a constant C = C (n, C) such that

mh1
h uY (,) C m uX() + u k+h+1 (6.9)
V [ 2 ],(k+h) Xk ()
k=0
[ m1
2 ]
for every u V m X() Cb ().
Remark 6.2. The statement of Theorem 6.1 can be somewhat generalized, in the sense
that assumptions (6.4)(6.8) can be weakened if either () < , or Ln () < ,
or Hn1 () < . Specically: if () < , it suces to assume that there exists
L (0, ) such that inequalities (6.4)(6.8) hold with the integral operators multiplied
by (0,L) on the left-hand sides; if Ln () < , it suces to assume that inequalities
(6.4)(6.5) hold with replaced by (0,M ) for some M (0, ); if Hn1 () < ,
it suces to assume that inequalities (6.6)(6.8) hold with replaced by (0,N ) for
some N (0, ). Then inequality (6.9) holds, but with C depending also either on
L and (), or on M and Ln (), or on N and Hn1 () < , according to whether
() < , or Ln () < , or Hn1 () < .
Our rst application of Theorem 6.1 yields the following Sobolev type inequality, in
arbitrary domains, with usual exponents.
Theorem 6.3 (Sobolev inequality with measure). Let be any open set in Rn , n 2.
Assume that is a measure in fullling (5.1) for some (n 1, n], and for some
constant C . Let m N, and h N0 be such that 0 < m h < n. If 1 < p < mh
n
, then
there exists a constant C = C(n, m, p, , C ) such that

mh1
h u p C m uLp () + u k+h+1 p(n1)
L n(mh)p (,) V [ 2 ],(k+h) L n(mhk)p ()
k=0
(6.10)
[ m1
2 ]
for every u V m,p () Cb ().
The next result tells us that, as in the classical Rellich theorem, the Sobolev embedding
p
corresponding to inequality (6.10) is pre-compact if the exponent n(mh)p is replaced
with any smaller one, and () < .
Theorem 6.4 (Compact Sobolev embedding with measure). Let , , n, m and h be as

in Theorem 6.3. Assume, in addition, that () < . If 1 q < n(mh)p
p
, and {ui }
[ m1 ]
is a bounded sequence in V m,p () Cb 2 () endowed with the norm appearing on the
right-hand side of (6.10), then {h ui } is a Cauchy sequence in Lq (, ).
n
The limiting case when p = mh , which is excluded from Theorem 6.3, is considered
in the next statement, which provides us with a YudovichPohozaevTrudinger type
inequality in arbitrary domains.
Theorem 6.5 (Limiting Sobolev inequality with measure). Let and be as in The-
orem 6.3. Assume, in addition, that Ln () < , () < and Hn1 () < .
Let m N and h N0 be such that 0 < m h < n. Then there exists a constant
C = C(n, m, , C , Ln (), (), Hn1 ()) such that
h u n
exp L n(mh) (,)

mh1
C m uL mh
n + u k+h+1 ],(k+h) n1 (mh)(nk1)
() V[ 2 L k (log L) nk ()
k=1

+ u h+1 ],(h) n (6.11)
V[ 2 exp L n(mh) ()
n [ m1
2 ]
for every u V m, mh () Cb ().
n
The super-limiting regime, where p > mh , is the object of the following theorem.
Theorem 6.6 (Super-limiting Sobolev inequality). Let be a open set in Rn , n 2, such

that Ln () < and Hn1 () < . Assume that m N, h N0 , and 0 < m h < n.
n
If p > mh and pk > n1 k for k = 1, . . . , m h 1, then there exists a constant
C = C(n, m, p, p1 , . . . , pmh1 , Ln (), Hn1 ()) such that
h uL ()

mh1
C uLp () +
m
u k+h+1 + u h+1 (6.12)
V [ 2 ],(k+h) Lpk () V [ 2 ],(h) L ()
k=1
[ m1
2 ]
for every u V m,p () Cb ().
Proof of Theorem 6.1. Let u be any function as in the statement. Inequalities (5.3) and
(6.4)(6.8) tell us that
h
| u| (cs)
Y (0,)
n
s
nm+h
C
s | m
u|L n (r)dr
+
r nm+h
n | m
u|L n (r)dr

Y (0,) n Y (0,)
0 s
n1

s
mh1
n1k [ k+h+1 ],(k+h)
+ s g (r)dr

2
Hn1
k=1 Y (0,)
0

[ k+h+1 ],(k+h)
+
r n1k
n1 g 2
Hn1
(r)dr

Y (0,)
n1
s
s
n1

n1 [ h+1 ],(h)
+
s g 2
Hn1
(r)dr

Y (0,)
0

[ k+h+1
mh1

C |m u|Ln (s) + g 2 ],(k+h) Hn1 (s) (6.13)
X(0,) X k (0,)
k=0
for some constants C depending on n, and C depending on the constant C appearing in

(6.4)(6.8) and on n. A property of rearrangement-invariant spaces [9, Proposition 5.11,
Chapter 3] ensures that the norm on left-most side of (6.13) is bounded from below by
min{1, 1c }|h u| (s)Y (0,) . Hence, from (6.13) and (6.2) we deduce that

mh1
[ k+h+1

h uY (,) C m uX() + g 2 ],(k+h) (6.14)
Xk ()
k=0
for some constant C depending on the constant C appearing in (6.4)(6.8) and on n.

Inequality (6.9) follows on taking the inmum in (6.14), for each xed k, among all
k+h+1
functions g [ 2 ],(k+h) . 2
Proof of Theorem 6.3. Since, for any measure space R, a representation space of the
Lebesgue space Lp (R) is just Lp (0, ), the conclusion can be easily deduced from The-
orem 6.1, via standard one-dimensional Hardy type inequalities for Lebesgue norms (see
e.g. [62, Section 1.3.2]). 2
Proof of Theorem 6.4. Fix any > 0. Then, there exists a compact set K such that
( \ K) < . Let C0 () be such that 0 1, = 1 in K. Thus, K supp(),
the support of , and hence
(supp(1 )) ( \ K) < . (6.15)
Let be an open set, with a smooth boundary, such that supp() . Let {ui }
[ m1 ]
be a bounded sequence in V m,p () Cb 2 (). Then, by Theorem 6.3 (applied with
= Ln ), it is also bounded in the standard Sobolev space W m,p ( ). By a weighted
version of Rellichs compactness theorem [62, Theorem 1.4.6/1], {h ui } is a Cauchy
sequence in Lq ( , ), and hence there exists i0 N such that
h ui h uj Lq ( ,) < (6.16)
if i, j > i0 . On the other hand, by Hlders inequality,
(1 )(h ui h uj )Lq (,)

p(nmp)q
h ui h uj p (supp(1 )) pq
L nmp (,)
p(nmp)q
C ui m1 ] + uj m1 ] pq
V m,p ()C [ 2 () V m,p ()C [ 2 ()
p(nmp)q
C pq (6.17)
for some constants C and C independent of i and j. From (6.16) and (6.17) we infer
that
h ui h uj Lq (,) h ui h uj Lq ( ,) + (1 )(h ui h uj )Lq (,)

p(nmp)q
+ C pq (6.18)
if i, j > i0 . Owing to the arbitrariness of , inequality (6.18) tells us that {h ui } is a

Cauchy sequence in Lq (, ). 2
Proof of Theorem 6.5. If R is a nite measure space, equipped with a measure , then
the norm of a function in the Orlicz space exp L (R), with > 0, is equivalent, up to
multiplicative constants depending on and (R), to the functional
1
1 + log (R) (s) .
s L (0,(R))
Moreover, the norm in the Orlicz space Lp log L(R) is equivalent, up to multiplicative
constants depending on p, and (R) to the functional

1 + log (R) p (s) .
s Lp (0,(R))
Thus, owing to Theorem 6.1 and Remark 6.2, inequality (6.11) will follow if we show
that

n

n(mh) s

()
n(mh)
s 1 + log s n
(r)dr CL mh n ,
(0,Ln ())
0 L (0,())
(6.19)

n(mh) n(mh)
()
1 + log s n
r n (r)dr CL mh
n ,
(0,Ln ())
n
s L (0,())
(6.20)
for every non-increasing function : [0, ) [0, ) with support in [0, Ln ()], and

n1
s
nk1
()
n(mh)
s 1 + log s n
(r)dr

0
L (0,())
(mh)(nk1)
C 1 + log H s()
n1
n(n1)
(s) n1 , k = 1, . . . , m h 1,
L k (0,Hn1 ())
(6.21)

()
n(mh)
nk1
1 + log n
r n1 (r)dr
s
n1
s L (0,())
(mh)(nk1)
C 1 + log H s()
n1
n(n1)
(s) n1 , k = 1, . . . , m h 1,
L k (0,Hn1 ())
(6.22)

n1
s
n1
()
n(mh)

s 1 + log s n
(r) dr

0
L (0,())
n(mh)
C 1 + log H s()
n1
n
(s)L (0,Hn1 ()) , (6.23)
for some constant C and every non-increasing function : [0, ) [0, ) with support
in [0, Hn1 ()].
Inequalities (6.19)(6.23) are consequences of classical weighted Hardy type inequali-

ties [62, Section 1.3.2]. 2
Proof of Theorem 6.6. Inequality (6.12) follows from Theorem 6.1 and Remark 6.2, via
weighted Hardy type inequalities [62, Section 1.3.2]. 2
7. Sharpness of results
In this section we work out in detail some examples, announced in Sections 1 and 2,
in connection with certain sharpness features of the inequalities presented above.
Example 7.1. We observed in Section 2 that the term u p(n1) can be dropped
V 1,0 L np ()
on the right-hand side of (2.3) if is a regular domain. Here, we show that, by contrast,
the term in question is indispensable for an arbitrary domain. To this purpose, we exhibit
a domain Rn for which the inequality

u pn C 2 uLp () + u p(n1) (7.1)
L n2p () L n2p ()
fails for 1 < p < n2 , for every constant C independent of u. The relevant domain is the
union of a sequence of axially symmetric cup-shaped subdomains k about the xn -axis,
which are connected by thin cylinders Hk joining k with k1 (Fig. 1, Section 2). Each
subdomain k is a set of revolution about the xn -axis of the form

k = x : |x | < (xn xkn + k ) , xn (xkn , xkn + hk )
for some xkn > 0 and 0 < k < hk , with hk 0+ as k , and (0, 1). The cylinder
Hk has a basis of radius k . Dene the sequence {uk } by
xn xk
uk (x) = 1 hk
n
for x k ,
uk = 0 in j for j = k and in Hj for j = k, k + 1, and is continued to Hk and Hk+1 in

such a way that u C 2 ().
One can verify that
[(n1)+1](n2p)
uk pn hk np
, (7.2)
L n2p ()
[(n2)+1](n2p)
uk p(n1) hk (n1)p

, (7.3)
L n2p ()
as k , and
2 uk Lp () = 2 uk Lp (Hk Hk+1 ) (7.4)

for k N. If k decays to 0 suciently fast as k , the norm 2 uk Lp (Hk Hk+1 )

decays arbitrarily fast to 0. Thus, inequality (7.1) fails when tested on the sequence uk ,
whatever C is.
Example 7.2. Our purpose here is to demonstrate that, whereas the seminorm
uV 1,0 Lr () can be replaced with uLr () in (2.8) when is a regular domain,
this is impossible, in general, if no regularity on is retained. Precisely, we construct
an open set in R2 for which the inequality

uLq () C 2 uLp () + uL () (7.5)
for u V 2,p () C() fails for 1 < p < 2 and for every q 1. The relevant set is
represented in Fig. 2, Section 2.
Let u : R be a function such that u C 2 () C 0 (), u(x, y) = 1 + b1k (y 1 ck )
if (x, y) Rk , u(x, y) = 0 if (x, y) R, and u(x, y) depends only on y in Nk . One has
that
uL () = 2, (7.6)
and

ak bk
uL1 () uL1 (Rk ) = = ak = . (7.7)
bk
k=1 k=1 k=1
On the other hand,

p1
2 uLp () = 2 upLp (Nk ) .
k=1
Thus,
2 uLp () < , (7.8)
provided that the sequence dk decays suciently fast to 0. Equations (7.6)(7.8) tell us
that inequality (7.5) cannot hold in .
Example 7.3. We are concerned here with the sharpness of the exponent q given by (2.9)
in inequality (2.8). An open set Rn is produced where inequality (2.8) fails if q
exceeds the right-hand side of (2.9). Consider the domain Rn , with n 3, depicted
for n = 3 in Fig. 3, Section 2. By the standard Sobolev inequality, one necessarily has
q np
np
. Thus, it suces to show that
rn
q . (7.9)
n1
Let {uk } be a sequence of functions uk : R enjoying the following properties:

uk C 2 () C 0 (); uk (x , xn ) = 1 + b1k (xn 1 ck ) if (x , xn ) Rk ; uk (x , xn ) depends
only on xn on Nk ; uk (x , xn ) = 0 if (x , xn ) / Rk Nk . One has that, for k N,
nq
uk Lq () uk Lq (Rk ) = bk q ,
n1r
uk V 1,0 Lr () Cbk r
,
and
2 uk Lp () = 2 uk Lp (Nk ) ,
for some constant C. Thus, inequality (2.8) entails that

nq n1r
bk q C 2 uk Lp (Nk ) + bk r (7.10)
for some constant C, and for every k N. The norm on the right-hand side of (7.10)
decays to 0 arbitrarily fast, provided that dk tends to 0 fast enough. Hence, if (2.8) holds,
then q must necessarily satisfy (7.9).

Example 7.4. We conclude by showing that the number m1 2 of derivatives to be
prescribed on , appearing in our inequalities, is minimal, in general, for an m-th order
Sobolev inequality to hold in an arbitrary domain . This will be demonstrated by two
examples.
First, given p > 1 and h, i, n N such that p(m h) < n and 0 h i < m 2 , we
produce a counterexample to the inequality
h u pn Cm uLp () (7.11)

L np(mh) ()
for all u V m,p () C i1 () such that u = u = = i1 u = 0 on . Note that

m1
the condition i < m 2 is equivalent to i 1 < 2 .
Second, in the case when p(m h) > n > p max{m i, 2i h} and 0 h < i < m 2
we produce a counterexample to the inequality
h uL () Cm uLp () (7.12)
for all u V m,p () C i1 () such that u = u = = i1 u = 0 on .

To this purpose, consider a domain similar to the one constructed in Example 7.1,
save that the sequence of cup-shaped subdomains k is replaced with a sequence of balls
Bk (xk ), with radius k to be chosen later, again connected by thin cylinders (Fig. 4).
Let vk : Bk (xk ) [0, 1] be the function dened as
|xxk |2 i
vk (x) = 1 2
k
for x Bk (xk ). (7.13)
We have that vk = 0 on Bk (xk ) for 0 i 1, and hence, given i and

k (0, 2k ), there exists a positive constant c such that
| vk | ck i i
k (7.14)
in an neighborhood of Bk (xk ). Moreover, if 2i, then there exists a positive

constant c such that
| vk | ck (7.15)
in a subset of Bk (xk ) of Lebesgue measure k n , whereas, if > 2i, then
vk = 0. (7.16)
Next, denote by yk and zk the north and the south poles of Bk (xk ), respectively, and let
: [0, ) [0, 1] be a smooth function, which vanishes in [0, 12 ] and equals 1 in [1, ).
Let us dene the function wk : [0, ) as
wk (x) = vk (x)(|x yk |/k )(|x zk |/k ) for x Bk (xk ), (7.17)
and wk = 0 elsewhere, where k will be chosen later.

If j 2i,
|j wk | ck j (7.18)
in a subset of Bk (xk ) of Lebesgue measure k n , and, if j > 2i,
j wk = 0 in Bk (xk ) \ (Bk (yk ) Bk (zk )). (7.19)
Thus, there exists a constant c such that

min{j,2i}
|j wk | c j
k | vk |

in Bk (yk ) Bk (zk ). (7.20)
=0
Consequently, if j m, then
i
|j wk | cij
k k in Bk (yk ) Bk (zk ), (7.21)
for some constant c. Hence, if j > 2i, then
ck pi k
p(ij)+n
j wk pLp (B (xk )) (7.22)
k
for some constant c. On the other hand, if j 2i, then

p(ij)+n
j wk pLp (B (xk )) c k npj + k pi k (7.23)
k
for some constant c.

Set k = k , with to be chosen later. Then, by (7.22) and (7.23),

wk pV m,p (B (xk )) c k pi+[p(im)+n] + k n2pi + k pi+(pi+n)
k

c k pi+[p(im)+n] + k n2pi , (7.24)
where the last inequality holds since p(i m) + n < pi + n, owing to the assumption
that i < m2.
Let us consider (7.11). Note that
n > pi. (7.25)
Indeed, since we are now assuming that p(m h) < n and 0 h i < m
2, we have that
n > p(m h) p(m i) > p(2i i) = pi, namely (7.25).
We may choose such that
pi + [p(i m) + n] < n 2pi. (7.26)
Actually, if p(i m) + n > 0, then inequality (7.26) holds provided that
npi
1<< p(im)+n . (7.27)
Note that the two inequalities in (7.27) are compatible since i < m2 . If, instead, p(i
m) + n 0, then any choice of > 1 is admissible, since pi + [p(i m) + n] pi,
whence (7.26) follows, owing to (7.25).
By (7.24) and (7.26),
wk pV m,p (B (xk )) ck pi+[p(im)+n] , (7.28)

k
for some constant c. Given a sequence {k }, dene u as

u(x) = k wk (x) for x . (7.29)
k=1
Note that u = u = = i1 u = 0 on .
h n
pn
Set q = np(mh) , and choose k = k q for k N. By (7.18), there exists a positive
constant c such that

h
uqLq () = qk | wk | dx c
h q
qk knhq =c 1 = . (7.30)
k=1 Bk (xk ) k=1 k=1
On the other hand, by (7.28) there exists a constant c such that

pi+[p(im)+n]
upV m,p () = pk wk pV m,p (B (xk )) c pk k
k
k=1 k=1

(1)[p(im)+n]
=c k . (7.31)
k=1
Our assumptions ensure that p(i m) + n p(h m) + n > 0. Thus, ( 1)[p(i

m) + n] > 0, and hence the last series in (7.31) converges, provided that k decays to 0
suciently fast. Clearly, equations (7.30) and (7.31) contradict (7.11).
Let us next focus on (7.12). Consider again the function u given by (7.29). Fix
(0, h), and choose k = kh . By (7.18),
h uL () c lim k kh = c lim k = . (7.32)

k k
Moreover, by (7.24),

p(h) pi+[p(im)+n]
upV m,p () = pk wk pV m,p (B (xk )) c k k + kn2pi
k
k=1 k=1

hpppi+[p(im)+n]
=c k + khpp+n2pi . (7.33)
k=1
The assumption n > p max{m i, 2i h} ensures that
hp + n 2pi > 0, and hp pi + [p(i m) + n] > 0,
provided that is suciently large. Since can be chosen arbitrarily small, we may
assume that both exponents of k in the last series of (7.33) are positive, and hence that
uV m,p () < , (7.34)
provided that k decays to 0 fast enough. Equations (7.32) and (7.34) contradict (7.12).
Appendix A
A result in the theory of Sobolev functions tells us that, if u is any weakly dierentiable
function in Rn , then
|u(x) u(y)| C(M (|u|)(x) + M (|u|)(y)) for a.e. x, y Rn , (A.1)
for some constant C. Here, M denotes the maximal function operator dened, for f
L1loc (Rn ), as

1
M f (x) = sup |f (y)| dy for x Rn ,
Bx Ln (B)
B
and B denotes any ball in Rn . Thus, M (|u|) is an upper gradient for u in the sense of
metric measure spaces, as dened in [38].
The following proposition provides us with a higher-order counterpart of (A.1), and
gives grounds for denitions (3.13) and (3.14).
Proposition A.1. Let n, N. Then there exists a constant C = C(, n) such that, if
21,1
u Wloc (Rn ), then

(2 2 ||)! (y x)
(1)||
D u(y) D
u(x)

||1 ( 1 ||)!! |y x|
21

C M (|21 u|)(x) + M (|21 u|)(y) for a.e. x, y Rn . (A.2)
Proof. By [15, Proposition 5.1], if 0 || 2 2, then there exist a measurable function

R21, (u) : Rn Rn R and a constant C = C(, n) such that
(y x) +
D u(y) = D u(x) + R21, (u)(x, y) for a.e. x, y Rn , (A.3)
!
||22||
and
|R21, (u)(x, y)|

C|x y|21|| M (|21 u|)(x) + M (|21 u|)(y) for a.e. x, y Rn . (A.4)
We claim that there exist constants C(, ), for || 1, such that
(2 2 ||)! (y x) [D u(x) + (1)||+1 D u(y)]

( 1 ||)!! |y x|21
||1
C(, )(y x)
= R21, (u)(x, y) for a.e. x, y Rn . (A.5)
|y x|21
||1
Inequality (A.2) will then follow from (A.5) and (A.4).

Let us establish (A.5). By (A.3), after exchanging the order of summation and relabel-
ing the indices, one obtains that there exist constants A(, ) and B(, ), for || 1,
such that
(2 2 ||)! [D u(x) + (1)||+1 D u(y)]

(y x)
( 1 ||)!! |y x|21
||1
(2 2 ||)! (y x)
=
( 1 ||)!! |y x|21
||1

||+1 (y x) +
D u(x) + (1)

D u(x) + R21, (u)(x, y)
!
||22||
(y x) (y x)
= A(, )D (u)(x) + B(, )R21, (u)(x, y)
|y x|21 |y x|21
||22 ||1
(A.6)
for a.e. x, y Rn .
Now, let us choose u = P in (A.6), where P is a polynomial of the form

P(y) = b (y x) for y Rn , (A.7)
||22
for some coecients b R. Clearly,
D P(x) = !b if || 2 2, D P(x) = 0 if || > 2 2. (A.8)

Hence, R21, (P)(x, y) = 0, and from (A.6) we obtain that
(2 2 ||)! [D P(x) + (1)||+1 D P(y)]

(y x)
( 1 ||)!! |y x|21
||1
(y x)
= A(, )!b (A.9)
|y x|21
||22
for a.e. x, y Rn . We next express the leftmost side of (A.6) in an alternative form.
Dene : R R as
(t) = u(x + t) for t R,
where
yx
= .
|y x|
Given j {1, . . . , 2 1}, we have that
j!
(j) (t) = D u(x + t) for a.e. t R. (A.10)
!
||=j
Thus, by the Taylor formula centered at t = 0, if k {1, . . . , 2 1}, then

22
(j) (0)
(k) (|y x|) = |y x|jk + Q21,k ()(|y x|)
(j k)!
j=k

22
1 j!
= |y x|jk D u(x) + Q21,k ()(|y x|) (A.11)
(j k)! !
j=k ||=j
for a.e. x, y Rn , where Q21,k () denotes the remainder in the (2 2 k)-th order
Taylor formula for (k) , centered at t = 0. By (A.10) and (A.11), there exist constants
A (, ) and B (, ) such that
(2 2 ||)! [D u(x) + (1)||+1 D u(y)]

(y x)
( 1 ||)!! |y x|21
||1

1
(2 k 2)! (k) (0) + (1)k+1 (k) (|y x|)

= (1)
k!( k 1)! |y x|2k1
k=0

1
(2 k 2)! 1
= (1)
k!( k 1)! |y x|2k1
k=0

k!
D u(x)
!
||=k
22
j!
1
+ (1)k+1 |y x|jk D u(x) + Q21,k ()(|y x|)
(j k)! !
j=k ||=j
(y x)
= A (, )D u(x)
|y x|21
||22
1
+ B (, )Q21,|| ()(|x y|) (A.12)
|y x|2||1
||1
for a.e. x, y Rn . If P is again a polynomial in t of the form (A.7), then is also a

polynomial of degree not exceeding 2 2, and from (A.8) and (A.12), applied with
u = P, we obtain that
(2 2 ||)! [D P(x) + (1)||+1 D P(y)]

(y x)
( 1 ||)!! |y x|21
||1
(y x)
= A (, )!b (A.13)
|y x|21
||22
for a.e. x, y Rn . Owing to the arbitrariness of the coecients b , we infer from (A.9)
and (A.13) that
A(, ) = A (, ) (A.14)
for every multi-index such that || 2 2. On the other hand, by (4.23) and (4.24),
applied with = = , a = 0 and b = |y x|, and by (A.12) and (A.13),

(21) d1 (t) d1 (t)
0= (t) = +
(t |y x|) |t=0 dt1
dt1 t |t=|yx|

1
(2 k 2)! (k) (0) + (1)k+1 (k) (|y x|)

= (1)
k!( k 1)! |y x|2k1
k=0
(2 2 ||)! (y x) [(1)||+1 D u(y) + D u(x)]
=
( 1 ||)!! |y x|21
||1
(y x)
= A (, )!b (A.15)
|y x|21
||22
for a.e. x, y Rn . By the arbitrariness of the coecients b again, A (, ) = 0 for every

such that || 2 2. Hence, owing to (A.14),
A(, ) = 0 (A.16)
for every such that || 2 2. Equations (A.6) and (A.16) tell us that
(2 2 ||)! D u(x) + (1)||+1 D u(y)

(y x)
( 1 ||)!! |y x|21
||1
(y x)
= B(, )R21, (u)(x, y) for a.e. x, y Rn , (A.17)
|y x|21
||1
whence (A.5) follows with C(, ) = B(, ). 2
References
[1] D.R. Adams, Traces of potentials arising from translation invariant operators, Ann. Sc. Norm.
Super. Pisa 25 (1971) 203217.
[2] D.R. Adams, A trace inequality for generalized potentials, Studia Math. 48 (1973) 99105.
[3] A. Alvino, V. Ferone, G. Trombetti, Moser-type inequalities in Lorentz spaces, Potential Anal. 5
(1996) 273299.
[4] L. Ambrosio, P. Tilli, Topics on Analysis in Metric Spaces, Oxford University Press, Oxford, 2004.
[5] T. Aubin, Problmes isoprimetriques et espaces de Sobolev, J. Dierential Geom. 11 (1976)
573598.
[6] F. Barthe, P. Cattiaux, C. Roberto, Interpolated inequalities between exponential and Gaussian,
Orlicz hypercontractivity and isoperimetry, Rev. Mat. Iberoam. 22 (2006) 9931067.
[7] T. Bartsch, T. Weth, M. Willem, A Sobolev inequality with remainder term and critical equations
on domains with topology for the polyharmonic operator, Calc. Var. Partial Dierential Equations
18 (2003) 253268.
[8] M. Belloni, B. Kawohl, A symmetry problem related to Wirtingers and Poincars inequality, J. Dif-
ferential Equations 156 (1999) 211218.
[9] C. Bennett, R. Sharpley, Interpolation of Operators, Academic Press, Boston, 1988.
[10] E. Bierstone, P.D. Milman, W. Pawlucki, Dierentiable functions dened on closed sets. A problem
of Whitney, Invent. Math. 151 (2003) 329352.
[11] A. Bjrn, J. Bjrn, Nonlinear Potential Theory on Metric Spaces, European Mathematical Society
(EMS), Zrich, 2011.
[12] S.G. Bobkov, C. Houdr, Some connections between isoperimetric and Sobolev-type inequalities,
Mem. Amer. Math. Soc. 25 (1997), viii+111.
[13] S.G. Bobkov, M. Ledoux, From BrunnMinkowski to sharp Sobolev inequalities, Ann. Mat. Pura
Appl. 187 (2008) 369384.
[14] T. Boggio, Sulle funzioni di Green dordine m, Rend. Circ. Mat. Palermo 20 (1905) 97135 (in Ital-
ian).
[15] B. Bojarski, Pointwise characterization of Sobolev classes, Tr. Mat. Inst. Steklova 255 (2006) 7187;
English translation in Proc. Steklov Inst. Math. 255 (2006) 6581.
[16] H. Brzis, E. Lieb, Sobolev inequalities with remainder terms, J. Funct. Anal. 62 (1985) 7386.
[17] Y. Brudnyi, P. Shvartsman, Witneys extension problem for multivariate C 1, -functions, Trans.
Amer. Math. Soc. 353 (2001) 24872512.
[18] S. Buckley, P. Koskela, SobolevPoincar implies John, Math. Res. Lett. 2 (1995) 577593.
[19] S. Buckley, P. Koskela, Criteria for embeddings of SobolevPoincar type, Int. Math. Res. Not. 18
(1996) 881902.
[20] Yu.D. Burago, N.N. Kosovskiy, The trace of BV-functions on irregular subsets, Algebra i Analiz 22
(2010) 105126 (in Russian); English translation in St. Petersburg Math. J. 22 (2011) 251266.
[21] Yu.D. Burago, V.A. Zalgaller, Geometric Inequalities, Springer, Berlin, 1988.
[22] V.I. Burenkov, Sobolev Spaces on Domains, B.G. Teubner Verlagsgesellschaft mbH, Stuttgart, 1998.
[23] L. Capogna, D. Danielli, S.D. Pauls, J.T. Tyson, An Introduction to the Heisenberg Group and the
Sub-Riemannian Isoperimetric Problem, Birkhuser, Basel, 2007.
[24] I. Chavel, Isoperimetric Inequalities: Dierential Geometric Aspects and Analytic Perspectives,
Cambridge University Press, Cambridge, 2001.
[25] J. Cheeger, A lower bound for the smallest eigenvalue of the Laplacian, in: Problems in Analysis,
Princeton Univ. Press, Princeton, 1970, pp. 195199.
[26] A. Cianchi, A sharp embedding theorem for OrliczSobolev spaces, Indiana Univ. Math. J. 45 (1996)
3965.
[27] A. Cianchi, Symmetrization and second-order Sobolev inequalities, Ann. Mat. Pura Appl. 183 (2004)
4577.
[28] A. Cianchi, L. Pick, Optimal Gaussian Sobolev embeddings, J. Funct. Anal. 256 (2009) 35883642.
[29] A. Cianchi, N. Fusco, F. Maggi, A. Pratelli, The sharp Sobolev inequality in quantitative form,
J. Eur. Math. Soc. 11 (2009) 11051139.
[30] P.J. Davis, Interpolation and Approximation, Blaisdell Publishing Company, Waltham, MA, 1963.
[31] D.E. Edmunds, R. Kerman, L. Pick, Optimal Sobolev imbeddings involving rearrangement-invariant
quasinorms, J. Funct. Anal. 170 (2000) 307355.
[32] L. Esposito, V. Ferone, B. Kawohl, C. Nitsch, C. Trombetti, The longest shortest fence and sharp
PoincarSobolev inequalities, Arch. Ration. Mech. Anal. 206 (2012) 821851.
[33] C.L. Feerman, A sharp form of Whitneys extension theorem, Ann. of Math. 161 (2005) 509577.
[34] C.L. Feerman, C m extension by linear operators, Ann. of Math. 166 (2007) 779835.
[35] B. Franchi, P. Hajasz, P. Koskela, Denitions of Sobolev classes on metric spaces, Ann. Inst. Fourier
49 (1999) 19031924.
[36] F. Gazzola, H.-C. Grunau, G. Sweers, Polyharmonic Boundary Value Problems. Positivity Preserv-
ing and Nonlinear Higher Order Elliptic Equations in Bounded Domains, Springer-Verlag, Berlin,
2010.
[37] A. Grigoryan, Isoperimetric inequalities and capacities on Riemannian manifolds, in: The Mazya
Anniversary Collection, vol. 1, Rostock, 1998, in: Oper. Theory Adv. Appl., vol. 109, Birkhuser,
Basel, 1999, pp. 139153.
[38] P. Hajasz, Sobolev spaces on an arbitrary metric space, Potential Anal. 5 (1996) 403415.
[39] P. Hajasz, P. Koskela, Isoperimetric inequalities and imbedding theorems in irregular domains,
J. Lond. Math. Soc. 58 (1998) 425450.
[40] P. Hajasz, O. Martio, Traces of Sobolev functions on fractal type sets and characterization of
extension domains, J. Funct. Anal. 143 (1997) 221246.
[41] E. Hebey, Analysis on Manifolds: Sobolev Spaces and Inequalities, Courant Lect. Notes Math.,
vol. 5, AMS, Providence, 1999.
[42] J. Heinonen, Lectures on Analysis on Metric Spaces, Springer-Verlag, New York, 2001.
[43] D. Homan, J. Spruck, Sobolev and isoperimetric inequalities for Riemannian submanifolds, Comm.
Pure Appl. Math. 27 (1974) 715727; A correction to: Sobolev and isoperimetric inequalities for
Riemannian submanifolds, Comm. Pure Appl. Math. 27 (1974) 715725, Comm. Pure Appl. Math.
28 (1975) 765766.
[44] T. Holmstedt, Interpolation of quasi-normed spaces, Math. Scand. 26 (1970) 177199.
[45] A. Israel, A bounded linear extension operator for L2,p (R2 ), Ann. of Math. 178 (2013) 148.
[46] A. Jonsson, H. Wallin, Function Spaces on Subsets of Rn , Harwood Academic Publishers, Chur,
1984.
[47] B. Kawohl, Symmetry results for functions yielding best constants in Sobolev-type inequalities,
Discrete Contin. Dyn. Syst. 6 (2000) 683690.
[48] B. Kawohl, M. Lucia, Best constants in some exponential Sobolev inequalities, Indiana Univ.
Math. J. 57 (2008) 19071927.
[49] R. Kerman, L. Pick, Optimal Sobolev imbeddings, Forum Math. 18 (2006) 535570.
[50] T. Kilpelinen, J. Mal, Sobolev inequalities on sets with irregular boundaries, Z. Anal. Anwend.
19 (2000) 369380.
[51] V.S. Klimov, Imbedding theorems and geometric inequalities, Izv. Akad. Nauk SSSR 40 (1976)
645671 (in Russian); English translation in Math. USSR, Izv. 10 (1976) 615638.
[52] V.I. Kolyada, Estimates on rearrangements and embedding theorems, Mat. Sb. 136 (1988) 323
(in Russian); English translation in Math. USSR Sb. 64 (1989) 121.
[53] P. Koskela, Metric Sobolev spaces, in: Nonlinear Analysis, Function Spaces and Applications, vol. 7,
Czech. Acad. Sci, Prague, 2003, pp. 132147.
[54] P.-L. Lions, F. Pacella, M. Tricarico, Best constants in Sobolev inequalities for functions vanishing
on some part of the boundary and related questions, Indiana Univ. Math. J. 37 (1988) 301324.
[55] E. Lutwak, D. Yang, G. Zhang, Sharp ane Lp Sobolev inequalities, J. Dierential Geom. 62 (2002)
1738.
[56] E. Lutwak, D. Yang, G. Zhang, Optimal Sobolev norms and the Lp Minkowski problem, Int. Math.
Res. Not. 1 (2006) 121.
[57] F. Maggi, C. Villani, Balls have the worst best Sobolev inequalities, J. Geom. Anal. 15 (2005)
83121.
[58] F. Maggi, C. Villani, Balls have the worst best Sobolev inequalities. II. Variants and extensions,
Calc. Var. Partial Dierential Equations 31 (2008) 4774.
[59] P. Mattila, Geometry of Sets and Measures in Euclidean Spaces, Cambridge University Press, Cam-
bridge, 1995.
[60] V.G. Mazya, Classes of regions and imbedding theorems for function spaces, Dokl. Akad. Nauk
SSSR 133 (1960) 527530 (in Russian); English translation in Soviet Math. Dokl. 1 (1960) 882885.
[61] V.G. Mazya, On p-conductivity and theorems on embedding certain functional spaces into a C-
space, Dokl. Akad. Nauk SSSR 140 (1961) 299302 (in Russian).
[62] V.G. Mazya, Sobolev Spaces with Applications to Elliptic Partial Dierential Equations, Springer,
Heidelberg, 2011.
[63] V.G. Mazya, S.V. Poborchi, Dierentiable Functions on Bad Domains, World Scientic, Singapore,
1997.
[64] E. Milman, On the role of convexity in functional and isoperimetric inequalities, Proc. Lond. Math.
Soc. 99 (2009) 3266.
[65] J. Moser, A sharp form of an inequality by Trudinger, Indiana Univ. Math. J. 20 (1971) 10771092.
[66] L. Salo-Coste, Aspects of Sobolev-Type Inequalities, Cambridge University Press, Cambridge,
2002.
[67] P. Shvartsman, Sobolev Wp1 -spaces on closed subsets of Rn , Adv. Math. 220 (2009) 18421922.
[68] G. Talenti, Best constant in Sobolev inequality, Ann. Mat. Pura Appl. 110 (1976) 353372.
[69] G. Zhang, The ane Sobolev inequality, J. Dierential Geom. 53 (1999) 183202.
[70] N. Zobin, Whitneys problem on extendability of functions and an intrinsic metric, Adv. Math. 133
(1998) 96132.

Sobolev Inequalities in Arbitrary Domains

Загружено:

Сведения о документе

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Sobolev Inequalities in Arbitrary Domains

Загружено:

Авторское право:

Доступные форматы

Advances in Mathematics 293 (2016) 644696

Contents lists available at ScienceDirect

Sobolev inequalities in arbitrary domains

Article history: A theory of Sobolev inequalities in arbitrary open sets in Rn is

where m N, h N0 ,  X() is a Banach function norm on with respect to Lebesgue

Some distinctive traits of the inequalities to be presented can be itemized as follows:

No regularity on is a priori assumed. In particular, the constants in (1.2) are

where is any open set in Rn , n 3, 2 is the bi-Laplace operator, F : Rn

u 2n C2 uL2 ()

uV 1,0 Lr () = inf gLr () , (2.4)

|u(x) u(y)| |x y|(g(x) + g(y)) for Hn1 -a.e. x, y , (2.5)

Fig. 1. Example 7.1, Section 7.

Fig. 2. Example 7.2, Section 7.

Fig. 3. Example 7.3, Section 7.

Finally, in the borderline case corresponding to p = n, an exponential norm is involved

Let be any open set in Rn , n 2. Given x , dene

x = {y : (1 t)x + ty for every t (0, 1)}, (3.1)

()x = {y : (1 t)x + ty for every t (0, 1)}. (3.2)

Proof. Given any r Q (0, 1), dene

()x (r) = {y : (1 t)x + ty for every t (0, r)}.

Next, we dene the sets

( Sn1 )0 = {(x, ) Sn1 : x + t for some t > 0}, (3.3)

( Sn1 ) = ( Sn1 ) \ ( Sn1 )0 . (3.4)

( Sn1 )0 = Sn1 if is bounded. (3.5)

be the function dened as

(x, ) = x + t, where t is such that x + t ()x .

g((x, )) = 0 if (x, ) ( Sn1 ) . (3.7)

Let us next introduce the functions

a : Sn1 [, 0) and b : Sn1 (0, ] (3.8)

a(x, ) = b(x, ) if (x, ) Sn1 . (3.10)

Proof. Assume rst that is bounded, so that ( Sn1 )0 = Sn1 . Consider a

that bh is Borel measurable. Moreover, bh converges to b pointwise. Hence, b is Borel

Cbk () = {u C k () : u has bounded support}. (3.12)

Let = (1 , . . . , n ) be a multi-index with i N0 for i = 1, . . . , n. We adopt the

(i) If k N, j = 0, and u Cbk1 (),

g k,0 (x) + g k,0 (y) (3.13)

for Hn1 -a.e. x, y .

(ii) If k N, j = 1, and u Cbk (),

g k,1 (x) + g k,1 (y) (3.14)

for Hn1 -a.e. x, y .

|u(x)| g 0,1 (x) (3.15)

for Hn1 -a.e. x .

uV k,j Lr () = inf g k,j Lr () (3.16)

uV k,j Z() = inf g k,j Z() . (3.17)

Observe that, in particular,

uV 0,1 Z() = uZ() .

Remark 4.2. In the case when m h = n, and is bounded, an estimate analogous

Remark 4.3. Under the assumption that

A special case of (4.4), corresponding to h = m 1, is the object of [62, Theorem 1.6.2].

Lemma 4.5. Let be any open set in Rn , n 2.

(i) If u V 21,1 () Cb1 () for some  N, then

for some constant C = C(n, ).

Q21 (t) + Q21 (t) = 1 for t R, (4.8)

Proof. Let us represent as

(t) = (t) + (t) for t (a, b), (4.10)

where and are the solutions to the problems

respectively. Let us rst focus on problem (4.11). We claim that

(s, r) = C|s r| 21

a polynomial of degree 2 1 in r, depending only on odd powers of r. Let us denote

= C(2 1)!(r2 1)1 . (4.19)

Q21 (s) + Q21 (s) = 2Q21 (0) for s R. (4.20)

Since is an even function,

Thus, (2 1)-times dierentiation of equation (4.15) yields

Leibnitz dierentiation rule for products yields

where m N, h N0 , X() is a Banach function norm on with respect to Lebesgue

u 2n C2 uL2 ()

uV 1,0 Lr () = inf gLr () , (2.4)

uV k,j Lr () = inf g k,j Lr () (3.16)

uV k,j Z() = inf g k,j Z() . (3.17)

uV 0,1 Z() = uZ() .

(i) If u V 21,1 () Cb1 () for some N, then

for some constant C = C(n, ).

Q21 (t) + Q21 (t) = 1 for t R, (4.8)

(s, r) = C|s r| 21

a polynomial of degree 2 1 in r, depending only on odd powers of r. Let us denote

= C(2 1)!(r2 1)1 . (4.19)

Q21 (s) + Q21 (s) = 2Q21 (0) for s R. (4.20)

Thus, (2 1)-times dierentiation of equation (4.15) yields

[0, b(x, )] t u(x + t)

Inequality (4.5), with = 1, follows from (4.26) and (4.27).

[a(x, ), b(x, )] t u(x + t)

t 2 a(x, ) b(x, ) d2

for k = 1, . . . , 2. From (4.28)(4.30) we infer that

0 2 a(x, ) b(x, ) d2

0 2 a(x, ) b(x, ) d2

There exist constants C = C(n, ) and C = C (n, ) such that

Next, we claim that there exists a constant C = C(, n) such that

0 2 a(x, ) b(x, ) d2

0 2 a(x, ) b(x, ) d2

where we have made use of the fact that P () = P () if || = 2 1, and of (3.10).

0 2 a(x, ) b(x, ) d2

L, (R) = sup s (s)

K(, s; L, (R), L (R)) r (r)L (0,s )

T gL (,) nn gL () (5.34)

s(Tg) (s n1 ) r (Tg) (r)L (0,s n1 K(Tg, s; L n1 , (, ), L (, ))

sup C f n ,1 hL1 () C f n ,1 (5.39)

where C is the constant appearing in (5.36), and C = C (n, , , C ). If follows

Q gL (,) Cg n1 ,1

X(R) = X(R) whenever = . (6.1)