JAN MALY
Contents
1. Hardys inequality
2. Rearrangement
3. Lorentz spaces
4. Rearrangement and Lorentz spaces in Rn
5. Embedding to continuous functions
6. Embedding into the Lorentz space
7. The critical case
8. Sobolev inequalities
9. Compact embedding
10. Notes
References
1
1
3
6
8
10
13
14
16
18
18
1. Hardys inequality
1.1. Theorem (Hardys inequality). Let 1 p < and < p 1. Suppose that f L1 (0, T ) and
Z t
F (t) =
f (s) ds.
0
2. Rearrangement
2.1. Distribution and rearrangement of a function. Let (X, S, ) be a -finite measure space and
f : X R be a measurable function. We define the distribution function of f as
f (s) = ({|f | > s}),
s > 0,
t > 0.
(c) f = (f ) .
(d) fR , f are nonincreassing,
right
R
R continuous (and thus lower semicontinuous).
f (s) ds
Proof. By the Fubini theorem, both integrals are equal to the product measure of the set
{(x, s) X (0, ) : s < |f (x)|}.
2.4. Theorem (Hardy-Littlewood-P
olya inequality).
Z
Z
|f g| d
f (t)g (t) dt.
X
0<r<f (x)
0<s<g(x)
dr ds d(x)
{f > s} {g > r} dr ds
(0,)2
n
o
min {f > s} , {g > r} dr ds
(0,)2
{f > s} {g > r} dr ds
=
(0,)2
Z
Z
=
0<r<f (t)
0<s<g (t)
Z
=
dr ds dt
2.5. Double-star operator and test functions. We define
Z
Z 1
1 t
f (s) ds =
f (rt) dr.
f (t) =
t 0
0
Notice that f has nothing to do with (f ) . (Of course, (f ) = f always holds.)
It is easily seen for f L1 (X) that f is continuous, nonincreasing and f f .
We say that g is a test function for f (t) if g : X R is measurable, 0 g 1/t and
Z
g d 1.
X
Proof. Suppose that g Gt . By Observations 2.2(c), the integral of g is 1. Using the monotonicity of
f and g we obtain
Z
Z
Z t
f (s) ds = f (t).
|f g| d
f (s)g (s) ds
t 0
X
0
To prove converse inequality, we denote
t0 = ({f > f (t)}),
and set
(
g=
1
t
tt
t(t00 t0 )
t00 = {f f (t)}
on {f > f (t),
on{f = f (t)}.
|f |g d = f (t).
2.7. Proposition (Subadditivity). Let : (0, ) R be right continuous, nonnegative and nonincreasing. Then the operator
Z
(t)f (t) dt
f 7
0
f (t) ((t, )) dt =
f (t) dt d(s) =
sf (s) d(s),
0
3. Lorentz spaces
3.1. Lorentz norm. Let 1 m , 1 q . We define
1
kf kLm,q = kt 7 t m f (t)kLq ( dt ) .
t
We observe
kf kqLm,q =
(t1/m f (t))q
dt
,
t
q < ,
We define
n
o
Lm,q = f : f : X R measurable, kf kLm,q < .
It is easily seen that L,q = {0} for q < and k kL, = k k . From Observation 2.2(e) we obtain
that k kLp,p = k kp for all p (notice that |f p | = (f )p ). In what follows we will show that the triangle
inequality holds for k km,q if 1 q m and that the Lorentz norm is equivalent to a genuine norm
1
kf kL(m,q) = kt 7 t m f (t)kLq ( dt ) .
t
if m > 1. Within the range 1 = m < q the Lorentz spaces are not normable.
3.2. Proposition. The Lorentz space Lm,q is linear.
3
(3)
If f (t/2) + g (t/2) s/2, then f (t/2) s/2 and thus f (s/2) t/2, similarly g (s/2) t/2. Using
(3) we obtain
(f + g) (s) t/2 + t/2 = t.
It follows that
(f + g) (t) 2 f (t/2) + g (t/2) .
From this we easily deduce that all the Lorentz spaces are linear.
Proof. Let f, g be nonnegative measurable functions on X. Then t m 1 [(f + g) ]q1 (t) is nonincreasing
and thus by Proposition 2.7
Z
q
t m 1 [(f + g) (t)]q1 h (t) dt
h 7
0
m
m
[(f + g) ] (t) dt
t
[(f + g) ] (t)f (t) dt +
t m 1 [(f + g) ]q1 (t)g (t) dt
t
0
0
0
Z q
11/q hZ q
1/q
t m 1 [(f + g) ]q (t) dt
t m 1 [f ]q (t) dt
0
0
1/q i
Z q
1
q
+
t m [g ] (t) dt
0
and a cancellation yields the result. The cancellation is legitimate by Proposition 3.2.
3.4. Theorem (Equivalent norm). If 1 < m < , then k kL(m,q) is a genuine norm equivalent to
k kLm,q .
Proof. The fact that this is a genuine norm follows from the subadditivity of the double-star operator,
Proposition 2.7. Since f f , the inequality
kf kLm,q kf kL(m,q)
is trivial. The converse inequality follows from the Hardy inequality (Theorem 1.1) if q < . For q =
we use the estimate
Z t
Z t
1
1
1
1
1
t m f (t) = t m 1
f (s) ds t m 1
s m s m f (s) ds
0
0
Z t
1
1
1
t m 1 sup{s m f (s)}
s m ds
s
m
m1
0
1
sup{s m f (s)},
t > 0.
3.5. Remark. We will not prove here that k kLm,q is not norm for m < q and k kL1,q is not equivalent
to any norm for q > 1. Instead of this, we will demonstrate this phenomenon on a particular case q = .
Example of failure of the triangle inequality in Lm, ((0, 2)) is the following: f (x) = min{1, x1/m },
g(x) = f (2 x). Then f = g = f , (f + g) (2) = 1 + 21/m . We have
kf kLm, = kgkLm, = 1,
fn (x) =
1 X
1
.
2n i=1 |x nk |
Then fn belongs to the convex hull of the unit ball, but fn C log n.
4
kf kLm,q = m1/q ks 7 s f
(s)kLq ( ds ) .
s
(4)
the equality follows from Observation 2.2a. For q < we use the Fubini theorem to obtain
Z
Z
q
q
1
m
t
[f (t)] dt =
rqm [f (rm )]q mrm1 dr
0
Z f (r p )
Z
=m
0
Z
rq1 qsq1 ds dr = m
0
Z [f (s)]1/m
Z
=m
Z0
=m
q(rs)q1 dr ds
m (f )
sq1 qrq1 dr ds
0
q1
[f (s)] m ds.
t>0
or as
sup s[f (s)]1/m .
s>0
3.7. Theorem (H
older type inequality).
Z
|f g| d kf kLm,q kgkLm0 ,q0 .
X
1
p
1
p0
= 1).
q
m [f (t)]
Z
0
s m 1 ds
q
m
s m 1 [f (s)]q ds
q
m
s m 1 [f (s)]q ds,
which proves (a) for Q = . In the general case we refer to the previous estimate and obtain
Z
Z
Q dt
Qq 1 q dt
1
1
m
t f (t)
=
t m f (t)
t m f (t)
t
t
0
0
Z
Qq
q dt
1
1
sup s m f (s)
t m f (t)
t
s>0
0
Z
Z q
Qq
q
q
q
m
s m 1 [f (s)]q ds
s m 1 [f (s)]q ds
0
0
Z q
Q/q
Q
1
q q
= m
s m 1 [f (s)]q ds
.
0
s>0
m1 M1
(X)
1
m
1
M
1
sup s M f (s) .
s>0
3.9. Remark. The inclusion LM,Q Lm,q fails if the measure of the space is infinite.
3.10. Lemma. Suppose that Ej are pairwise disjoint measurable subsets of X and f Lm,q (X), 1
q m. Then
X
m
kf E km
Lm,q kf kLm,q .
j
Let S = inf{s > 0 : f (s) = 0}. (S = if f is strictly positive everywhere.) Holders inequality yields
Z S
mq 1
Z S
m/q Z S
q
q
1
q/m
sq1 fm (s) ds
j
Lm,q
j
Z
mq 1 XZ
q
sq1 fm (s) ds
Z
0
j
S
(s) ds
m/q
q/m
sq1 f (s) ds
= mm/q kf km
Lm,q .
3.11. Lemma (Absolute ontinuity of the norm). Let {Ek } be a sequence of measurable sets,
\
E1 E2 , . . . ,
Ek = 0.
k
4. Rearrangement and Lorentz spaces in Rn
4.1. Lemma. Suppose that f L1loc (Rn ). Then
Z
s(M f ) (Cs)
|f (x)| dx,
{|f |>s}
s > 0,
Proof. Without loss of generality we may assume f 0. We set C = 5n + 1. Then for each z {M f >
Cs} there exists a ball Bz = B(z, rz ) such that
Z
Z
Cs|Bz |
f (x) dx
f (x) dx + s|Bz |,
Bz {f >s}
Bz
so that
Z
(C 1)s|Bz |
f (x) dx.
Bz {f >s}
f (x) dx.
{f >s}
4.2. Lemma. Suppose that f L1loc (Rn ). Then
Z
|f (x)| dx Cs|{M f > s}|
s > 0,
{M f >s}
f (x) dx
B(z,5rz )
B(y,10rz )
B(zj ,5rj )
4.3. Theorem (HerzRiesz inequality). Suppose that f L1loc (Rn ). Then
C 0 f (t) (M f ) (t) Cf (t),
t > 0,
f (t) (M f ) Cf (t)
f (x) dx
f (x) dx
{f >f (t)}
t
{f >f (t)}
f ( ) d = t f (t).
Thus
Cf (t) {s 0 : (M f ) (s) t},
which yields
(M f ) (t) Cf (t).
To prove the converse inequality, we use Lemma 4.2 with s = (M f ) (t) and denote G = {M f > s},
F = Rn \ G = {M f s}. Then obviously |G| = |{M f > (M f ) (t)}| t and f F M f F s a.e. on
Rn . (Thus also (f F ) s and (f F ) s.) We have
Z t
Z
(5)
t(f G ) (t) =
(f G ) ( ) d =
f (x) dx Cs|G| Cst
0
and
t(f F ) (t) ts.
(6)
Since the double star operator is subadditive (Proposition 2.7), adding together (5) and (6) we obtain
tf (t) Cts = Ct(M f ) (t),
which concludes the proof.
4.4. Theorem (Boundedness of the maximal operator). If f Lm,q (Rn ), 1 < m < , then
kM f kLm,q Ckf kLm,q .
Proof. By Theorem 4.3 and Theorem 3.4
kM f kLm,q Ckf kL(m,q) Ckf kLm,q .
4.5. Proposition (Regularization). Let 1 < m < , 1 q < . Suppose that u Lm,q (Rn ) and is
a standard mollifier. Then u u in Lm,q (Rn ) as 0.
Proof. Using Luzins theorem we decompose u to a part with small norm and to a continuous function
with a compact support. For the part v with a small norm we use the estimate
v(x) CM v(x),
which implies
k vkLm,q CkM vkLm,q CkvkLm,q .
by Theorem 4.4.
5. Embedding to continuous functions
Let B = B(0, R) be such that |B| = |E|. Then also |B \ E| = |E \ B|, hence
Z
Z
Z
Z
|x|1n dx
R1n =
R1n
|x|1n dx.
E\B
Thus, adding
R
EB
E\B
B\E
B\E
5.2. Theorem. Suppose that u Ln,1 () and B = B(y, r) be a ball containing z. Then u has
a continuous representative, which satisfies the estimate
Z
(8)
u(z) u CkuB kLn,1 .
B
Proof. We may assume that u has a compact support, = Rn and z = 0. Suppose first that u is smooth.
Denote f = |u|B . By the potential estimate
Z
Z
C
|x|1n f (x) dx.
u(z)
B
This may be rewriten using the distribution function and Lemma 5.1 as
Z
Z Z
Z
1/n
|x|1n dx ds C
f (s) ds,
u(z) u C
B
{f >s}
which is the desired estimate. Now, in the non-smooth case, we aproximate u by standard mollifications
and observe that the estimate shows that the j u is a uniform Cauchy sequence. It follows that there
is a continuous representative for which the estimate (8) holds.
5.3. Corollary. Suppose that u Ln,1 () and B = B(y, r) be a ball. Then for the continuous
representative of u
oscB u CkuB kLn,1 ,
where we denote
oscE u = diam(u(E)).
Proof. For z, z 0 B we infer from Theorem 5.2
Z
Z
0
0
|u(z ) u(z)| u(z) u + u(z ) u CkuB kLn,1 .
B
5.4. Theorem. If u Ln,1 (), then the continuous representative of u is a.e. differentiable.
Proof. We denote
lipf (x) = lim sup
yx
|f (y) f (x)|
.
|x y|
(9)
oscB(x,rx ) f
C
kuB(x,rx ) kLn,1
r
r
By the Vitali type covering theorem, we find a disjointed sequence B(xj , rj ) of balls from {B(x, rx ) : x
E } such that
E
B(xj , 5rj ).
j=1
X
j=1
rjn C
j=1
C
kuknLn,1
n
It follows that
|{x : lipf (x) = +}| = 0.
By the Rademacher-Stepanov theorem, the function u is differentiable a.e.
5.5. Theorem. If n > 1 and u Ln,1 (Rn ), then there exists c R such that uc C0 (Rn ). Moreover,
ku ck CkukLn,1 .
Proof. If z, z 0
/ B(0, 2R), then there is a chain of 3 balls in Rn \ B(0, R) which connects z and z 0 , and
thus
oscRn \B(0,2R) u CkukLn,1 (Rn \B(0,R))
and the expression on the right tends to zero as R by Lemma 3.11.
9
6.3. Proposition (Weak type estimate). Suppose that u W 1,1 (B) and 0 is a median of u in B. Let
s > 0. Then
Z
1
|u(x)| dx,
(11)
s|{u > s}|1 n C
B
C|G|
n1
G |x y|
B
B
Thus
Z
1
1 n
s |G|
C
|u(y)| dy.
B
1,1
6.4. Lemma. Suppose that u W (B) and 0 is a median of u in B. Let 0 a b < . Suppose that
1 q. Then
Z
q1
q0
n
q
C
|u|q dx |{u > a}|
.
(12)
(b a) |{u b}|
{a<u<b}
u a on a < u < b,
v = b a on u b,
0
on u a.
Then v has zero median in B and v = 0 a.e. outside {a < u < b}. Applying Proposition 6.3 to v and
s = (b a), we obtain immediately (12) for q = 1. If q > 1, using the Holder inequality at the right hand
side of (11) with s = b a we obtain
10
Z
q1
10
10
n
n
q
= s |{v s}|
C
|v(x)|q dx
|{0 < v < s}|
(b a) |{u b}|
B
Z
1/q
10
q
=C
|u|q
|{a < u < b}|
,
{a<u<b}
6.5. Theorem (Peetre). Let B be a ball. Suppose that u W 1,1 (B) and 0 is a median of u in B.
1
Suppose that 1 < p < n and p is the Sobolev exponent, i.e. p
= p1 n1 . Let 1 q . Then
kukLp ,q CkukLp,q
with C depending only on p, q and n.
10
bk = u (tk+1 ),
k = 0, 1, . . . .
Write
Ek = {ak < u < bk }.
Since
tk 2tk+1 2|{u bk }|,
(13)
|{u > ak }| tk ,
we obtain
q q q
u (tk+1 ) u (tk ) tk n Ctq1
k
(14)
Let
g=
tkp
|u|q dx.
Ek
E .
k
If q < p, then
q
g (tkp
and thus
g (tk ) tkp
}| |{u > ak }| tk
k = 0, 1, 2,
which implies
q
g (t) Ct p 1 ,
(15)
t > 0.
If q = p, then g 1 and thus (15) holds as well. Applying (15) and the Hardy-Littlewood-Polya
inequality (Theorem 2.4) to f = |u|q and g we obtain
Z
Z
Z
q
X
p 1
q
tk
|u| dx =
f (x)g(x) dx C
f (t)g (t) dt
Ek
k=0
(16)
Z
C
q
p 1
[|u| (t)]q dt
= kukqLp,q .
Now, we have
t p n 1 [u (t)]q dt
0
X Z tk q q
C
t p n 1 [u (t)]q dt
kukqp ,q =
(17)
tk+1
q
q
pn
tk
[u (tk )]q .
(18)
q
(b a)q .
1
If b < a, then the left hand side of (18) is negative, otherwise (1 )b (b a) and thus (18) holds as
well. Since u (t0 ) = 0 and 2tk+1 = tk , we can use (18) to estimate
q q
X qq
X qq
q q X pq nq
4 pn
1
tkp n [u (tk )]q = 34 p n
tk+1 [u (tk+1 )]q
tkp n [u (tk )]q
3
k
(19)
q
q
pn
tk
q q
2 pn
3
tkp
q
n
Getting together (17), (19), (14) and (16) we obtain the assertion.
11
Ek
1
1 p
Ctk
u (t0 ).
t>0
and thus
1
a 2 s p b,
Then we have
1
1 n
bt0
1
1 n
C(b a)t0
1
1 p
C
{a<u<b}
Thus
1/p
kukLp , Ct0
kukLp,
b CkukLp, .
6.6. Theorem. Let B be a ball. Suppose that u BV (B) and 0 is a median of u in B. Then
kukL1 ,1 Ckuk1
with C depending only on n.
Proof. For u W 1,1 (B) the preceeding proof works. For u BV (B) we use a molifying argument.
s > 0.
t > 0,
Then
where C deponds only on n.
Proof. Let t > 0. Let F (0, ) be a measurable set with |F | < t. We find pairwise disjoint intervals
(i , i ] such that
i 2i , i = 1, 2, . . . ,
[
X
F (i , i ] and
(i i ) < t.
i
Denote
We have
|Ei | i i
and thus
(20)
|E| < t.
12
i = 1, 2, . . . ,
|v(s)| ds =
i
(u )0 (s) ds
i
0
(2i )1/n [u (i ) u (i )]
Z
C
|u(x)| dx.
Ei
Summing over i and applying the HardyLittlewoodPolya inequality and (20) we obtain
Z
|v(s)| ds
F
XZ
|v(s)| ds
Z
C
|u(x)| dx
ZE
(E ) (s)|u| (s) ds
C
0
Z t
|u| (s) ds = t|u| (t).
C
0
Passing to supremum with respect to all measurable sets F (0, ) with |F | < t we obtain
tv (t) Ct|u| (t).
(21)
with C depending only on n.
Proof. We may assume that u is bounded. We set tk = 2k+1 , use Lemma 6.4 with
ak = u (tk ),
bk = u (tk+1 ),
k = 1, 2 . . . .
Since
tk 2tk+1 2|{u bk }|,
|{u > ak }| tk ,
we obtain
n
u (tk+1 ) u (tk ) C
(22)
|u|n dx.
{ak <u<bk }
m1
X
ank+1
anm
ank
=
mn1
(k + 1)n1
k n1
k=1
m1
X
k=1
m1
X 1
ank
1
n
.
k
(k + 1)n1
k n1
(k + 1)n1
ank+1
k=1
13
1
X
X
X
ank+1 ank
1
ank
n
C
a
k
kn
k n1
(k + 1)n1
(k + 1)n1
k=1
k=1
(23)
k=1
k=1
an1
k+1 (ak+1 ak )
(k + 1)n1
X
ak+1 ak
n n1 X
k=1
k=1
ank+1 1 n1
.
(k + 1)n
X
u (tk )n
k=1
kn
X
X
n
ank
=
C
ak+1 ak C
|u|n dx,
kn
B
k=1
k=1
7.2. Theorem (Trudinger inequality). Let B Rn be a ball of measure 1. Suppose that u W 1,1 (B)
and 0 is a median of u in B. Then there exists > 0 depending only on n such that
Z
n0
|u|
(24)
e ( kukn ) dx 2.
B
Z
0
ds
(n 1)
s logn (e/s)
Z
0
(u (s))n ds
s logn (e/s)
e ( kukn ) dx =
dt 2.
e kukn
dt
t
B
0
0
8. Sobolev inequalities
1,1
8.1. Theorem (Global inequality on the entire space). Suppose that u Wloc
(Rn ) and u Lp,q , where
1 p < n and 1 q . Then there exists a constant c R such that u c Lp ,q (Rn ). Moreover
ku ckLp ,q CkukLp,q
with C depending only on p, q and n.
Proof. Let ck be medians of u on Bk := B(0, 2k ). Set
(
u ck in Bk ,
uk =
0
outside Bk
Then by Theorems 6.5 and 3.8
kuk kqLp ,q CkukqLp,q (Bk ) CkukqLp,q (Rn ) .
(25)
For the constants we have
Z
Z
1/p
|ck1 ck | C |u ck | dx C |u ck |p dx
C2kn/p kuk kLp ,q
Bk
kn/p
C2
Bk
kukLp,q (Rn ) .
Hence the sequence ck has a limit c. We may assume that c = 0. Then uk u a.e. and thus
u lim inf (uk ) .
k
14
(26)
then
Z
1/q
Z
1/p
|u u
B |q
Cr |u|p
,
(27)
where u
B is the average of u on B and C depends only on n, p and q.
Proof. We may assume that u has zero median in B. Suppose first that there is an equality in (26).
Then by Theorem 6.5 and Theorem 3.8
Z
1/p
Z
Z
Z
1/q
1/q
1/p
+n
n
q
n/q
q
n/q
p
q
p
|u|p dx
|u| dx
=r
|u| dx
Cr
|u| dx
= Cr
B
B
B
B
Z
1/p
= Cr |u|p dx
B
If there is not equality in (26), we find an exponent s [1, n) (the case n = 1 is left to the reader) such
that s < p and s > q. Then using H
older inequality we obtain
Z
1/s
Z
1/q Z
1/s
|u|s dx
Cr |u|s dx
|u|q dx
B
B
B
Z
1/p
p
Cr |u| dx
B
We have
Z
Z
1/q
|
uB | = u dx |u|q dx
B
and thus
Z
1/q
Z
1/q
2 |u|q dx
|u u
B |q dx
B
Rn
Rn
15
Proof. We may assume that || = 1, otherwise we make the rescaling v(y) = r u(y/r), where r is such a
number that rn = ||. Further we may asume that
Z
1/p Z
1/p
|u|p dx
+
|u|p dx
1.
Rn
Rn
We may assume that q > p, otherwise the inequality follows directly from the Holder inequality on
without using the term with the gradient. We find an exponent s [1, n) (the case n = 1 is left to the
reader) such that s p and s q. We may assume that u 0. Denote
w = u v.
v = min{u, 1},
Then w 0 and
Z
|{w > 0}| = |{u > 1}|
up dx 1.
Rn
Then by Theorem 8.1 there exist constants C > 0 and c R; C depending only on n and s, c depending
also on u, such that
Z
1/s Z
1/s Z
1/s Z
1/p
1
|w c|s dx
|w|s dx
=
|w|s dx
|w|p dx
C Rn
Rn
{w>0}
{w>0}
Z
1/p
|u|p dx
1.
Rn
and since w c = c everywhere except on the set {w > 0} of measure 1, the convergence of the integral
on the left implies that c = 0. It follows
Z
1/q Z
1/s
q
(28)
w dx
ws dx
C.
v p dx
Rn
up dx 1.
Rn
[
jk
16
Ej ,
T
and observe that S1 S2 . . . and that k Sk is contained in the null set where {uk } diverges. Hence
|Ek | |Sk | 0 and
Z
Z
Z
aq
q
q
q
(31)
|wk | dx =
|uk | dx
|uk |a dx |Ek |1 a 0.
Ek
9.2. Lemma. Suppose that {uk } is a bounded sequence in W 1,p (Rn ). Then there exists a subsequence of
uk which converges a.e.
Proof. Since the matter is local, we may assume that that {uk } is bounded in W01,1 (B) for a fixed ball
B (otherwise we multiply with a cut-off function). Let 1 be a standard mollification kernel and > 0.
Recall the notation
(x) = n 1 (x/).
Then for each k N
Z
Z
|( uk )(x) uk (x)| dx
Z
(x y)|uk (y) uk (x)| dy dx
n
Rn
ZR Z
C n
|uk (y) uk (x)| dy dx
Rn
Rn
Z
C
(32)
B(x,)
Z
Rn
B(x,)
Z
=C
Rn
B(y,)
|uk (y)|
dy
dx
|y x|n1
|uk (y)|
dx dy
n1
|y x|
Z
C
There exists q (1, ) so that {uk } is locally bounded in Lq (B). This implies that there exists a
subsequence (labelled again as uk ) such that uk converges to some function u weakly in Lq , in particular
uk u
pointwise.
| uk uj | dx + 2C sup
|ui | dx
i
Rn
9.3. Examples. 1. Let be a nontrivial nonnegative smooth function with support in B(0, 1) and
rk 0. Let p, q [1, ) be exponents. Set
1 n
p
uk (x) = rk
Then
(x/rk ).
(|uk | + |uk | ) dx =
Rn
Rn
(rkp ||p + |p ) dy
||q dy,
Rn
Rn
9.4. Theorem (Rellich, Kondrashev). Suppose that || < and that 1 p, q < ,
the embedding W 1,p (Rn ) , Lq () is compact.
1
q
>
1
p
n1 . Then
Proof. Let uk be a sequence bounded in W 1,p (Rn ). We find an exponent s [1, n) (the case n = 1 is
left to the reader) such that s p and s > q. Then, by the embedding theorem (Theorem 8.4, the