Sergey Astashkin and Konstantin Lykov - Extrapolation Description of Rearrangement Invariant Spaces and Related Problems

Extrapolation description
of rearrangement invariant spaces

and related problems
Sergey Astashkin and Konstantin Lykov
Abstract. Problems of the extrapolation theory of operators
in the scale of L
p
spaces are considered. Mainly we focus on the
problem of an extrapolation description of rearrangement invariant
(r.i.) spaces on [0, 1]. Generalizing JawerthMilmans approach we
introduce and study a new class SE of extrapolation r.i. spaces
with respect to the scale of L
p
spaces. In spite of the fact that the
class SE contains almost all known by now extrapolation spaces
close to L
it has a rather simple description. We completely

characterize Marcinkiewicz and Lorentz spaces belonging to this
class and present some examples of not extrapolation r.i. spaces
X such that X L
p
for every p < . Under mild assumptions it
is proved that Peetres Kfunctional of a couple (X, L
) has an
extrapolation description if and only if X SE. Moreover, it is
investigated the problem when an Orlicz space is representable as
an intersection of L
p
spaces with scalar weights. As an applica-
tion, a sharp Yano type theorem for operators bounded in L
p
is
established.
1 Introduction and preliminaries
The starting point of the extrapolation theory of operators is the fol-
lowing remarkable theorem proved by Japanese mathematician S.Yano
in 1951.
2000 subject classication: Primary 46M35, 46E30
Key words and phrases: Extrapolation spaces, rearrangement invariant spaces,
extrapolation of operators, Yanos theorem, Orlicz spaces, Peetres K-functional
The rst author was partially supported by the RFBR grant 09-01-08236
1
2 S. Astashkin and K.Lykov
Theorem 1.1 ([45]). (a) Suppose that a linear operator T is bounded
in L
p
for every p (1, p
0
), p
0
> 1, and
(1) T
L
p
L
p
C(p 1)
, p (1, p
0
),
for some > 0, with a constant C > 0 independent of p. Then,
T : L(log L)
L
1
,
where the Zygmund space L(log L)
consists of all measurable on [0,1]

functions x(t) such that
x
L(log L)
=
1
0
log
(e/t)x
(t) dt < .
Here and next, x
(t) stands for the non-increasing rearrangement of

|x(t)|.
(b) If T is a linear operator bounded in L
p
for all p > p
0
and
(2) T
L
p
L
p
Cp
1/
, p (p
0
, ),
for some > 0, with a constant C > 0 independent of p, then
T : L
Exp L
.
The Zygmund space Exp L
consists of all measurable functions x(t)

such that
x
Exp L
= sup
0<t1
log
1/
(e/t)x
(t) < .
Extrapolation description of rearrangement invariant spaces 3
In fact, the paper [45] contains only the part (a) of Theorem 1.1
but in the classical Zygmunds book Trigonometric series both parts
were published (see [47, Theorem 12.4.41]). Yanos theorem at once has
found valuable applications to studying classical operators of analysis
such that the maximal Hardy-Littlewood operator and the operator of
trigonometric conjugate function [45, 47]. Their norms in L
p
spaces
satisfy estimates of the form (1) or (2) and therefore, by Theorem 1.1, we
immediately obtain the appropriate estimates in Zygmund spaces. Thus
Yanos theorem can be treated as a converse statement with respect to
interpolation theorems: estimates of operator norms in L
p
-spaces imply
these in appropriate limiting spaces of this scale. So, Yanos result has
been called an extrapolation theorem.
Soon after that some more papers devoted to extrapolation problems
were appeared (see, for instance, [46, 43, 19]). In particular, in [46] and
[43] Yanos result was really rediscovered.
Later, in nineties, B. Jawerth and M. Milman laid down the founda-
tions of the general extrapolation theory, which studies natural limiting
spaces associated with various interpolation scales and provides esti-
mates for norms of appropriate operators [16, 17]. Mainly, they consider
scales generated by the real method of interpolation.
Recall that for any Banach couple (A
0
, A
1
) (i.e., A
0
and A
1
are Ba-
nach spaces linearly and continuously embedded in a common Hausdor
topological vector space) and t > 0, the Peetre Kfunctional is dened
as follows:
K(t, a; A
0
, A
1
) = inf{a
0
A
0
+ta
1
A
1
: a = a
0
+a
1
, a
i
A
i
, i = 0, 1}.
Then, for every 0 < < 1 and 1 q , the space

A
,q
consists of all
a A
0
+A
1
such that
a
A
,q
:=
_
_
0
_
t
K(t, a; A
0
, A
1
)
_
q
dt
t
_
_
1/q
< if q <
and
a
A
,
:= sup
0<t<
t
K(t, a; A
0
, A
1
) < .
If we replace in the last formulas the function t
with an arbitrary
positive function W(t) we get the space

A
W,q
.
As extrapolation methods (functors) B.Jawerth and M.Milman in-
troduce the sum and the intersection of families of Banach spaces. Let
{A
}
0<<1
be a family of Banach spaces such that there exist two Ba-
nach spaces U and V for which the continuous inclusions
U A
V (0 < < 1)
hold with uniformly bounded norms. Then the sum
(A
) =
_
a =
in V : a
for some and
<
_
,
with the norm a
(A
)
= inf
, where inmum is taken over all

admissible representations of a (A
). Analogously, the intersection

(A
) consists of all a
0<<1
A
such that
a
(A
)
= sup
0<<1
a
A
< .
In what follows, the expression f g means that cf g Cf
for some constants c > 0 and C > 0 independent of all or of a part of
arguments of f and g. Finally, if X is a Banach space and b > 0, then
bX consists of the same elements as X and x
bX
= bx
X
.
JawerthMilmans approach to calculation of norms of extrapolation
spaces is based on two key facts. Firstly, and functors commute
with appropriate interpolation functors. For example,
(3)
0<<1
_
M()
A
,
_
=

A
W,
,
where M() is an arbitrary positive function and W(t) = sup
M().
The second fact is valid for any tempered function M() which satises
the conditions
M() M (/2) as 0 and M() M
_
1 +
2
_
as 1.
By this assumption, we have the following stability: for all 1 q, r
0<<1
_
M()
A
,q
_
=
0<<1
_
M()
A
,r
_
,
where
A
,q
:= ((1 )q)
1/q

A
,q
. In particular, for the Banach couple
A = (L
1
, L
) of the spaces on the segment [0, 1] we have
A
,1
L
p
=
A
,p

A
,
for = 1
1
p
uniformly with respect to 1 p . Therefore, assuming that (p)
(2p) as p we obtain
M() :=
_
1
1
_
= (p) (2p) = M
_
1 +
2
_
as 1,
whence
1p<
((p)L
p
) =
0<<1
_
1
1
_

A
,
.
Hence, using (3), we are able to nd an extrapolation description of
Zygmund spaces from Yanos theorem (see, for instance, [34, p. 22-23]).
For every > 0 and p
0
> 1 we have
(4)
1<p<p
0
_
(p 1)
L
p
_
= L(log L)
and
(5)
pp
0
_
p
1/
L
p
_
= Exp L
.
These relations recover the spaces L(log L)
and ExpL
as limiting
spaces of the scale of L
p
-spaces and, in addition, immediately imply both
parts of Yanos theorem. In fact, for example, assuming T
L
p
L
p

Cp
1/
, by (5), we obtain
Tx
ExpL
sup
pp
0
p
1/
Tx
p
C sup
pp
0
x
p
= Cx
.
Next, we will concentrate on studying the extrapolation functor.
By (5),
(6) x
ExpL
sup
pp
0
p
1/
x
p
=
_
_
_x
p
p
1/
_
_
_
L
(p
0
,)
.
Therefore, the Zygmund space Exp L
is the result of applying a weigh-

ted L
-norm to the function f

x
(p) := x
p
. In [18], L
is replaced
with the more general space L
r
(1 r ) which allows to obtain
new extrapolation spaces. In [6] and [9], we focused on the following in
a sense opposite problem: which rearrangement invariant spaces such
that X L
p
for every p < have an extrapolation description with
respect to the scale of L
p
-spaces or, in other words, when the norm of
an r.i. space X can be expressed by L
p
norms?
First of all, recall basic denitions of the theory of r.i. spaces on [0, 1]
(its detailed exposition can be found in [21, 22]). A Banach function
space X is called a Banach lattice if the conditions y = y(t) X and
|x(t)| |y(t)| imply that x = x(t) X and x y. A Banach
lattice X is said to be an rearrangement invariant (r.i.) space if from
y = y(t) X and x
(t) y
(t) it follows that x = x(t) X and

x
X
y
X
.
The simplest and most important example of an r.i. space is given
by L
p
space (1 p ), with the usual norm
x
p
=
_
1
0
|x(t)|
p
dt
_
1/p
(1 p < ) and x
= ess sup
0t1
|x(t)|.
Note that L
is the smallest space among all r.i. spaces on [0, 1] and

L
1
is the largest one [21, Theorem 2.4.1]. Let 1 < p < , 1 q .
Then, the space L
p,q
is dened as the set of all functions for which the
following quasi-norm is nite:
x
p,q
:=
_
_
q
p
1
0
_
t
1/p
x
(t)
_
q
dt
t
_
_
1/q
(1 q < )
and
x
p,
= sup
0<t1
t
1/p
x
(t).
Replacing in the previous formulas x
(t) by x
(t) :=
1
t
t
0
x
(s) ds, we
get in L
p,q
an equivalent r.i. norm for every 1 < p < , 1 q . In
particular, L
p
= L
p,p
.
An another natural generalization of L
p
spaces are Orlicz spaces [20].
Let M(u) be an Orlicz function, that is, an increasing convex function
on [0, ) such that M(0) = 0. The Orlicz space L
M
consists of all
measurable functions x(t) on [0, 1] such that the function M
_
|x(t)|
u
_
L
1
for some u > 0, with the norm
x
L
M
= inf
_
_
_
u > 0 :
1
0
M
_
|x(t)|
u
_
dt 1
_
_
_
.
In the case of the power function M(u) = u
p
(1 p < ) we obtain
L
p
space; if M(u) is equivalent to the function e
u
( > 0) for all u

large enough, L
M
coincides with the Zygmund space Exp L
appeared
in Yanos theorem. Moreover, the exponential Orlicz space ExpL
N
is
generated by the function e
N(t)
1 (N(t) is an Orlicz function).
Other examples of r.i. spaces are Lorentz and Marcinkiewicz spaces.
Let (t) be an increasing concave function on [0, 1], with (0) = 0. The
Marcinkiewicz space M() consists of all measurable functions x(t) such
that
x
M()
= sup
0<s1
(s)
s

s
0
x
(t)dt < .
The Lorentz space
r
() (1 r < ) is dened by the norm
x
r
()
=
_
_
1
0
x
(t)
r
d(t)
_
_
1/r
.
An important characteristic of an r.i. space X is its fundamental
function
X
(t) =
(0,t)
X
, where
(0,t)
is the characteristic function of
the interval (0, t). In particular,
L
p
(t) = t
1/p
,
M()
(t) =
1
()
(t) = (t),
L
M
(t) =
1
M
1
(1/t)
.
Moreover, () :=
1
() is the smallest space and M() is the largest
space among all r.i. spaces with the fundamental function (t) [21,
Theorems 4.5.5 and 4.5.7]. Every fundamental function (t) is quasi-
concave (this means that (0) = 0, (t) > 0, (t) increases, and (t)/t
decreases). Every such function is equivalent to its least concave ma-
jorant (t); more exactly, 1/2 (t) (t) (t) (0 t 1) [21,
Theorem 4.1.1]. An r.i. space X is said to have the Fatou property if the
conditions 0 x
n
x a.e. on [0, 1] and sup
nN
x
n
< imply x X
and x
n
x.
An Orlicz space L
M
close to L
coincides with the Marcinkiewicz

space with the same fundamental function . More precisely [23, 42],
L
M
= M() if and only if 1/ L
M
. In view of the expression for
the fundamental function of L
M
, the last condition is equivalent to the
following: there exists a > 0 such that
(7)
1
0
M
_
1
M
1
_
1
t
__
dt < .
In particular, a Zygmund space ExpL
and an exponential Orlicz space

Exp L
N
coincide with the Marcinkiewicz space M() generated by the
function (t) log
1/
(e/t) and (t) = 1/N
1
(log(e/t)), respectively.
Let us recall the denition from [6], where instead of weighted L
r
-
parameters arbitrary Banach lattices are considered. Let F be a Banach
lattice of measurable functions on [1, ) (it will be called an extrapola-
tion parameter). The set L
F
consists of all functions x(t) for which the
function f
x
(p) := x
p
F, with the norm
x
L
F
:= f
x
F
.
It is not hard to check that L
F
is an r.i. space such that L
F
L
p
for
every p < .
Denition 1.2 ([6]). We will say that an r.i. space X on [0, 1] is an
extrapolation space with respect to the scale of L
p
spaces (X E) if there
is a Banach lattice F on [1, ) such that X = L
F
(with equivalence of
norms).
The simplest example of such a space is the space L
. Since
f
L
= lim
p+
f
p
,
then L
= L
L
[1,)
. Other examples are Zygmund spaces Exp L
( >
0) (see (6)) and exponential Orlicz spaces Exp L
N
since, by [4],
(8) x
ExpL
N sup
pp
0
x
p
N
1
(p)
.
Note that in both cases we can take for a parameter F an appropriate
weighted L
space.
Some basic properties of Banach spaces L
F
were studied in [27].
Moreover, in [6] necessary and sucient conditions under which impor-
tant in applications Marcinkiewicz and Lorentz spaces belong to the
class E have been found. Next, in the paper [9] the following interesting
subclass of strong or nice extrapolation spaces was singled out. If X is
an r.i. space on [0, 1], then by

X we denote the Banach lattice of all
measurable on [1, ) functions f such that
f
X
:= f
_
log
2
(2/t)
_
X
< .
Denition 1.3 ([9]). An r.i. space X is called a strong extrapolation
space (X SE) if X = L
X
(with equivalence of norms).
In particular, relations (6) and (8) demonstrate that all Zygmund
spaces Exp L
( > 0) and exponential Orlicz spaces ExpL

N
are strong
extrapolation spaces.
Thus, for a strong extrapolation space X the corresponding extrap-
olation parameter

X is completely and explicitly determined by X. We
observe the following surprised phenomenon: the norm of any function
in every strong extrapolation r.i. space X is equivalent to the norm in
X of its L
p
norm. More precisely, x
X

_
_
x
log
2
(2/t)
_
_
X
.
In [9], a rather simple characterization of strong extrapolation spaces
has been found. Namely, an r.i. space X SE if and only if the
operator Sx(t) = x(t
2
) is bounded in X. Moreover, in a partial case
including Marcinkiewicz and Lorentz spaces a simpler criterion holds. In
Section 2, we review these results. Moreover, we introduce the concept of
a tempered extrapolation parameter generalizing the Jawerth-Milmans
idea of a tempered weight. Then, we prove that an r.i. space X
SE if and only if X = L
F
with a tempered extrapolation parameter
F (see Theorem 2.2). This implies, in particular, that every limiting
space of the L
p
scale near to L
whose norm can be calculated by

Jawerth-Milmans approach is a strong extrapolation space. At the same
time, in this section some examples of extrapolation but not strong
extrapolation Marcinkiewicz spaces are presented. By Theorem 2.2,
every such space is generated by a not tempered parameter and therefore
cannot be obtained by Jawerth-Milmans approach. Section 3 contains
a review of results from [9] about an extrapolation description of the
K-functional K(t, x; X, L
) provided that X is a strong extrapolation

space (Theorem 3.1). In Section 4, it is investigated the problem when
an Orlicz space L
M
is representable in the form:
(9) L
M
=
1p<
((p)L
p
) ,
with some bounded measurable function (p). We nd some sucient
conditions (see Theorems 4.6 and 4.7) and prove extrapolation relations
for corresponding dual spaces which are both Orlicz and Lorentz spaces.
In concluding part of this section we provide some examples of equalities
of the form (9) both with tempered and not tempered functions (p).
At last, Section 5 contains some concluding remarks and applications.
Firstly, we consider a discrete version of the extrapolation functor
and show that in contrast to the case of the continuous scale any ex-
trapolation functor always gives on the scale {L
k
}
k=1
an Orlicz space.
Secondly, we prove a sharp extrapolation Yano type theorem. Finally,
using a description of strong extrapolation spaces derived in Theorem
2.2, we nd sucient conditions for the rearrangement invariance of
Rademacher multiplicator spaces.
The dilation function of a positive function (t) (0 t 1) is dened
as
M
(s) = sup
0<tmin(1,1/s)
(st)
(t)
, 0 < s < .
There are the numbers
= lim
s0+
ln M
(s)
ln s
and
= lim
s+
ln M
(s)
ln s
which are called the lower and upper dilation indices of . For a quasi-
concave 0
1 [21, p. 76]. Finally, given Banach couple

(X
0
, X
1
), a Banach space X is an interpolation space with respect to
(X
0
, X
1
) if X
0
X
1
X X
0
+ X
1
and, for every linear operator T
bounded in X
0
and in X
1
, T is bounded in X.
2 Strong extrapolation spaces
The next denition is a generalization of the Jawerth-Milman idea of a
tempered weight.
Denition 2.1. An extrapolation parameter F will be called tempered
if the operator Df(p) := f(2p) is bounded in F.
Recall that the concepts of extrapolation and strong extrapolation
r.i. spaces (with respect to the scale of L
p
spaces) were been intro-
duced in [6] and [9], respectively (see Denitions 1.2 and 1.3 in the
Introduction). One of the main results of this article is the following
characterization of strong extrapolation spaces.
Theorem 2.2. For any r.i. space X the following conditions are equiv-
alent:
(a) X SE;
(b) the operator Sf(t) = f(t
2
) is bounded in X.
(c) X = L
F
with a tempered extrapolation parameter F;
To prove Theorem 2.2 and some other statements we need the fol-
lowing lemma.
Lemma 2.3. The operator Sx(t) = x(t
2
) is bounded in an r.i. space X
if and only if Sx
X
Cx
X
for some C > 0 and all x X.
Proof. Since necessity is obvious, we should prove only suciency. Let
x E and let > 0. If
A = {t [0, 1] : | x(t) |> } and B = {t [0, 1] : | x(t
2
) |> },
then B = (A) with (u) =

u. Hence (see, for instance, [36, Lem-
ma 9.5.1]),
{t : |Sx(t)| > } = (B) =
A
du
2
{t: x
(t)>}
0
du
2
u
= {t : Sx
(t) > } .
Thereby (Sx)
(t) Sx
(t) and, by assumption, we obtain

Sx X and Sx
X
Sx
X
Cx
X
.
Proof of Theorem 2.2. Firstly, show that the conditions (a) and (b) are
equivalent.
Let x = x(t) be a measurable function on [0, 1]. Consider the func-
tion x(t) := x
log
2
2/t
(0 < t 1). Since
x
p

_
_
x
[0,2
p+1
]
_
_
p

1
2
x
(2
p+1
) (p 1),
then x
X

1
2
x
X
. Therefore, by the denition of a strong extrapola-
tion space, condition (a) of the theorem is equivalent to the inequality
(10) x
X
Cx
X
,
which holds for all x X and some C > 0.
First, we assume S to be bounded in X. Since S
n
x
X
S
n
x
X
(n = 1, 2, . . . ), we infer
x(2
p+1
) = x
p
2
_
_
x

[0,2
p
)
_
_
p
2
n=0
_
_
_x

[2
2
n+1
p
,2
2
n
p
)
_
_
_
p
2
n=0
2
2
n
x
_
2
2
n+1
p
_
,
for all p 1, or equivalently
x(2t) 2
n=0
2
2
n
S
n+1
x
(t) (0 < t 1/2).

Hence,
x(2t)
X
2
n=0
2
2
n
S
n+1
x
X
C
1
x
X
.
Since the dilation operator
2
x(t) = x(t/2) is bounded in every r.i. space
and
2
XX
2 [21, Theorem 2.4.4], the last inequality implies (10)
with C = 2C
1
.
Let now conversely X = L
X
, i.e., x X whenever x X and
inequality (10) holds. If t = 2
p+1
(p 1), then
x
(t
2
) x
(2
2p
) 4
_
_
x

[0,2
2p
]
_
_
p
4x
p
= 4 x(t).
Therefore, Sx
X
4 x
X
4Cx
X
. Applying Lemma 2.3, we con-
clude that S is bounded in X.
Next, prove that conditions (a) and (b) imply (c). In fact, with (a)
at hand we may take F =

X. Then, by condition (b) and again by
the boundedness of the dilation operator
2
x(t) = x(t/2) in X, for an
arbitrary f = f(p) F we have
f(2p)
F
= f(2 log
2
(2/t))
X
= f(log
2
(4/t
2
))
X

Cf(log
2
(4/t))
X
2Cf(log
2
(2/t))
X
= 2Cf
F
.
At last, we check that (c) implies (b). Suppose that X = L
F
, where
F is a tempered extrapolation parameter. Applying the HolderRogers
inequality, we obtain
1
0
|Sx(t)|
p
dt =
1
x(t
2
)
p
dt =
1
0
|x(s)|
p
ds
2
_
_
1
0
|x(s)|
4p
ds
_
_
1
4
_
_
1
0
1
(2
s)
4
3
ds
_
_
3
4
,
whence Sx
p
C
1
x
4p
. Since F is a tempered parameter, then
Sx
X
C
2
_
_
_Sx
p
_
_
_
F
C
1
C
2
_
_
_x
4p
_
_
_
F
C
3
_
_
_x
p
_
_
_
F
C
4
x
X
,
which implies that the operator S is bounded in X.
Corollary 2.4. An r.i. space X SE if and only if there exists a
constant C > 0 such that
_
_
_
n=1
x
[2
2n
,2
2n+1
)
[2
n
,2
n+1
)
_
_
_
X
Cx
X
for all x X.
Proof. First of all, using the boundedness of the dilation operator
2
in
X [21, Theorem 2.4.4] once more, we obtain
1
2
_
_
_
k=1
x
(2
k
)
[2
k
,2
k+1
)
_
_
_
X
x
X
2
_
_
_
k=1
x
(2
k+1
)
[2
k
,2
k+1
)
_
_
_
X
This together with the following elementary inequality
x
(2
k+1
) 2
_
_
_x
[2
k
,2
k+1
)
_
_
_
k
x
(2
k
) (k = 1, 2, . . . )
imply that
_
_
_
k=1
_
_
_x
[2
k
,2
k+1
)
_
_
_
k
[2
k
,2
k+1
)
_
_
_
X
x
X

4
_
_
_
k=1
_
_
_x
[2
k
,2
k+1
)
_
_
_
k
[2
k
,2
k+1
)
_
_
_
X
.
Applying Theorem 2.2, we complete the proof.
Theorem 2.2 implies that every r.i. space being an interpolation
space with respect to a couple of strong extrapolation spaces is a strong
extrapolation space as well. In particular, we obtain
Corollary 2.5. Let N
1
and N
2
be two Orlicz functions and X be an
interpolation r.i. space with respect to the couple (Exp L
N
1
, Exp L
N
2
).
Then X is a strong extrapolation space.
Remark 2.6. As it is easily seen, instead of the norm in

X we can use
in Theorem 2.2 the seminorm
f
X,p
0
= f
[p
0
,)
X
,
with some xed p
0
> 1.
Remark 2.7. Recall that using Jawerth-Milmans theory [18] we may
identify an extrapolation space (with respect to the L
p
-scale) of the form
(r)
((p)L
p
) with the norm
x
(r)
_
_
p
0
_
(p)x
p
_
r
dp
_
_
1
r
if r <
and
x
() sup
pp
0
_
(p)x
p
_
whenever
(11) (p) (2p) as p .
Without loss of generality, we may assume that (p/2) C(p) if p
2p
0
. Thus,
_
_
p
0
((p)f(2p))
r
dp
_
_
1/r
=
_
_
2p
0
((p/2)f(p))
r
dp
2
_
_
1/r
C
_
_
p
0
((p)f(p))
r
dp
_
_
1/r
,
which implies that the extrapolation parameter F = L
r
() is tempered.
Therefore, all r.i. spaces of the form
(r)
((p)L
p
), with (p) satisfying
condition (11), possess property (c) from Theorem 2.2 and, according to
this theorem, are strong extrapolation spaces.
Remark 2.8. The construction resulting in strong extrapolation spaces
as well as the
(r)
-functor with a tempered weight is in a sense sta-
ble. Consider instead of the scale of L
p
spaces the scale of L
p,q
-spaces
(1 < p < ), where a q [1, ] is xed, and a given Banach lattice
F on [1, ) dene the space L
F,q
consisting of all functions x = x(t)
measurable on [0, 1] such that f
x,q
(p) := x
p,q
F. Arguing in the
same way as in the proof of Theorem 2.2, it can be showed that either
of its conditions (a), (b), and (c) is equivalent to either of the following
ones:
(d) for every q [1, ] X = L
X,q
, i.e., x
X

_
_
x
log
2
(2/t),q
_
_
X
(x X);
(e) there is a q [1, ] such that X = L
X,q
.
In some important cases condition (b) under which an r.i. space
belongs to the class SE is equivalent to a much simpler condition [6].
Denition 2.9. We say that a nonnegative function (t) on [0, 1] sat-
ises the
2
condition (
2
) whenever there exists a C > 0 such
that
(12) (t) C(t
2
), 0 t 1.
Proposition 2.10. Given an r.i. space X on [0, 1], consider the follow-
ing conditions:
(i) the operator Sx(t) = x(t
2
) is bounded in X;
(ii) the fundamental function
X
(t) of X satises the
2
condition.
Then (i) implies (ii). Conversely, if X is an interpolation space with
respect to a couple ((), M()) of Lorentz and Marcinkiewicz spaces
with the same fundamental function (t) (in particular, if X coincides
with () or with M()), then (ii) implies (i).
Proof. Since the implication (i) (ii) is obvious it suces to prove the
opposite one.
Let X = (). It is well known that the boundedness in a Lorentz
space of a linear operator that is continuous in the space of all nite
a.e. measurable functions with respect to convergence in measure is
equivalent to its boundedness on the set of characteristic functions [21,
Lemma 2.5.2]. Since for the operator S the last is equivalent to (12),
then in this case we done.
Examine now the case of a Marcinkiewicz space. Firstly, if
2
,
then for every x M()
(t)Sx
(t) C(t
2
)x
(t
2
) C sup
0<s1
(s)x
(s) Cx
M()
.
It is immediate that the condition
2
implies
= 0. Hence, using
the last inequality and [21, Theorem 2.5.3], we infer
Sx
M()
C
sup
0<t1
Sx
(t)(t) C
Cx
M()
.
Thereby it remains to refer to Lemma 2.3.
At last, let X be an interpolation space with respect to a couple
((), M()). Then, obviously,
X
and therefore
2
. Since by
the previous arguments S is bounded in () and in M(), then it is
bounded also in X.
Theorem 2.2 and the last proposition imply the following result (a
weaker version was proved earlier in [26]).
Theorem 2.11. Let be an increasing concave function on [0, 1] and let
X be an interpolation space with respect to the couple ((), M()) (in
particular, X may be () or M()). Then X is a strong extrapolation
space if and only if
2
.
Corollary 2.12. Suppose
2
and 1 r < . Then
(13) x
M()
sup
1p<
(2
p
)x
p
and
(14) x
r
()

_
_
1
x
r
p
2
p
(2
p
)dp
_
_
1
r
.
Relations (13) and (14) were established for the rst time in [6].
Earlier, in [25], Lukomskii has proved the second of them by a stronger
assumption:
(t) Ct
(t
2
) (0 < t 1). In the same paper, he has
applied (14) by consideration of the problem of convergence of trigono-
metric series in r.i. spaces.
Relations (13) and (14) can be obtained also from Jawerth-Milman
theory using the functors
(r)
. Let us show that for Marcinkiewicz
spaces. Suppose
2
. Then, by [34, formula (4.6)],
(15) x
((2
p
)L
p
)
sup
s>0
K(s, x; L
1
, L
)
(s)
,
where
(t) =
t
sup
p1
(2
p
)t
1/p
.
Firstly, sup
p1
(2
p
)t
1
p
(t)/2, for every t <
1
2
. Conversely, let t =
2
p
1
and C be a constant from inequality (12). In the case, when p p
1
,
we have (2
p
)t
1/p
(t). If p p
1
, then p
1
[2
n1
p, 2
n
p], for some
positive integer n. Thereby
(2
p
)t
1/p
C
n
(2
2
n
p
)2
p
1
/p
C
n
2
2
n1
(t) C
1
(t),
where C
1
= sup
nN
C
n
2
2
n1
< . Thus,
sup
p1
(2
p
)t
1/p
(t),
which implies that (t) t/(t). Therefore, by (15),
sup
1p<
(2
p
)x
p
= x
((2
p
)L
p
)
sup
s>0
(s)
s
s
0
x
(t) dt = x
M()
,
and we come to (13).
In general, the condition
2
is not sucient for an r.i. space X
with the fundamental function to be strong extrapolation.
Theorem 2.13. There is an r.i. space X which is a closed subspace of
a Marcinkiewicz space M(), with
2
, such that X SE.
Proof. Let = (t) be an increasing concave function on [0, 1] such that
(+0) = (0) = 0 and
2
. Dene a sequence {t
n
}
n=0
[0, 1] by
induction as follows: t
0
= 1 and t
n
are chosen arbitrarily so that
(t
2
n1
)
(t
n
)
> n.
Clearly, t
n
0. We put
x(t) =
n=1
1
(t
2
n1
)

(t
2
n
,t
2
n1
]
(t)
and
X
0
:= {y M() : y
(t) B x(t) for some B, > 0 and all t (0, 1]} .

It is easy to see that X
0
is a linear subset of the Marcinkiewicz space
M(). Dene X as the closure of X
0
in M(). Then X is an r.i.
space endowed with the norm of M(). In particular, the fundamental
function of this space
X
(t) coincides with (t) and therefore satises
the
2
condition.
Show now that the function x
1
(t) := x(t
2
) X. It suces to check
that
(16) inf
yX
0
x
1
y
M()
> 0.
Let y X
0
be arbitrary. Applying the well-known inequality [21, The-
orem 2.3.1]
a b
M()
a
M()
,
we obtain
(17) x
1
y
M()
x
1
y
M()
(x
1
y
)
(t
2
n
,t
n
]
M()
.
Choose a suciently large n N so that
1)
(t
2
n1
)
(t
n
)
> 2B and
_
1 +
1
_
t
2
n
t
n
,
where B > 0 and (0, 1) are such that y
(t) B x(t). In this case

(x
1
y
)
(t
2
n
,t
n
]
M()
(x
1
(t
n
) B x(t))
(t
2
n
/,t
n
]
M()
=
=
_
x
1
(t
n
) B x(t
2
n1
)
_
(t
n
t
2
n
/)
_
x(t
2
n
) B x(t
2
n1
)
_
(t
2
n
) =
=
_
1
(t
2
n
)

B
(t
2
n1
)
_
(t
2
n
)
_
1
(t
n
)

B
(t
2
n1
)
_
(t
2
n
) =
=
1
(t
n
)
_
1
B(t
n
)
(t
2
n1
)
_
(t
2
n
) >
(t
2
n
)
(t
n
)

1
2

1
2C
.
Hence, (17) implies that for every y X
0
x
1
y
M()

1
2C
,
and thus (16) is proved. Finally, Theorem 2.2 guarantees that the r.i.
space X is not a strong extrapolation space.
By [6, Theorem 1], a Marcinkiewicz space M() is an extrapolation
space if and only if
(18) (t) C sup
p1
t
1/p
1/
p
, 0 < t 1.
Moreover, the last condition implies that
(19) x
M()
sup
p1
x
p
1/
p
.
Now, assume that
2
. Then, by [6, Lemma 1],
1/
p

C
(2
p
)
(p 1),
and therefore we have (18). Moreover, it is obvious that in this case (19)
implies (13). In this connection, a natural question arises: Is there an ex-
trapolation Marcinkiewicz space M() generated by a function
2
? In other words, is there an extrapolation space which cannot be ob-
tained by Jawerth-Milmans theory ? We give an example showing that
the answer is armative (other examples of this sort will be presented
in Section 4).
Example 2.14. Consider an increasing concave function (t) such that
(t) (t) = e
1ln
e/t
for some (0, 1). Such a function exists since
(t) > 0,
(t) > 0,
(t) < 0 for t (0, t

0
) (where t
0
is suciently
small) and lim
t0+
(t) = 0.
Show that M() E. It suces to prove (18) with instead of .
Firstly,
1/
p
p
=
1
0
e
p ln
(e/t)p
dt = e
p+1
1
e
ps
s
ds.
Let us represent the integrand in the form: e
ps
s
= e
ps
s
e
(1)s
,
where (0, 1) is a parameter, and consider the function f
(s) =
ps
s (s 1). It is not hard to check that max

s1
f
(s) = (1 )
p
1
1
(/)
1
. Consequently,
1/
p
p
e
p+1
exp
_
(1 )p
1
1
_
1
_
0
e
(1)s
ds
= e
p+1
exp
_
(1 )p
1
1
_
1
_
1
1
,
whence
1/
p
exp
_
(1 )
_
p
_
1
_
_
1
1
_
1/p
.
Put = p
/(1 +p
), where = /(1 ). Then

(1 )
1/p
= (1 +p
)
1/p
2p
/p
< C
for some C > 0 and all p 1. Moreover, since (1 + x)
1 + C
x for
all x [0, 1], we have
(1 )
_
p
_
1
= p
_
1 +
1
p
(1 )
1
(1 )
1
(p
+C
),
and thus
(20) 1/
p
C
1
exp((1 )
1
p
1
).
To prove (18), verify that the inequality
(21) (t) 1/
p
C
2
t
1/p
holds for p =
1
ln
1
e/t with the constant C
2
= C
1
e. In fact, on the
one hand, by (20),
(t) 1/
p
C
1
exp((1ln
e/t)+(1) ln
e/t) = C
2
exp( ln
e/t).
On the other hand,
t
1/p
= exp((ln t)/p) = exp((ln
1
e/t)(1 ln e/t)) exp( ln
e/t).
Thus, (21) is proved and M() E. At the same time, it is not hard
to check that
2
.
3 Extrapolation description of Peetres Kfunc-
tional
The extrapolation theory developed in [16, 17, 18, 34] allows us to reverse
in a sense the interpolation process. In particular, in the case of scales
generated by the real method of interpolation it is possible to recover the
Kfunctional corresponding to the initial Banach couple. More precisely,
for any Banach couple

A = (A
0
, A
1
) we have
a
0<<1(t
A
,q
)
K(t, a; A
0
, A
1
),
with a constant independent of a A
0
+ A
1
and t > 0 [35, formu-
la (B.26)]. Here, as above,
A
,q
:= ((1 )q)
1
q
A
,q
.
Let us consider the problem when a Kfunctional K(t, f; X, L
),
where X is an r.i. space on [0, 1], may be described by the extrapola-
tion procedure from the previous section. We show that under certain
assumptions it is possible if and only if X is a strong extrapolation space.
Theorem 3.1. Let X be an r.i. space on [0, 1].
(i) If X SE, then
(22) K(t, x; X, L
) K(t, x
log
2
(2/u)
; X, L
),
with constants independent of x X and t > 0;
(ii) If there is a Banach lattice F on [1, ), F L
, such that
(23) K(t, x; X, L
) K(t, x
p
; F, L
[1, )),
with constants independent of x X and t > 0, then the fundamental
function of X satises the
2
condition.
To prove this theorem, we need several auxiliary assertions. The rst
of them is proved in [28, Example 3]. Since the proof of the second one
is similar, it is omitted as well.
Let X be an r.i. space on [0, 1], X = L
. Then its fundamental

function (t) satises the condition lim
t0+
(t) = 0. Without loss of
generality, we may assume that (t) is strictly increasing on [0, 1] and
(1) = 1.
Lemma 3.2. For every x X and 0 < t 1 we have
x
[0,
1
(t)]
X
K(t, x; X, L
) 2 x
[0,
1
(t)]
X
.
Let now X E, i.e. X = L
F
, with a Banach lattice F on [1, ). If
X = L
, then there exists x X such that lim

p+
x
p
= . Hence,
the function
(24) (u) :=
[u,)
F
(u 1)
satises the condition lim
u+
(u) = 0. Without loss of generality, it
may be assumed that is strictly decreasing on [1, ) and (1) = 1.
Lemma 3.3. Let a F be an increasing nonnegative function on [1, )
and 0 < t 1. Then
a
[
1
(t),)
F
K(t, a; F, L
[1, )) 2 a
[
1
(t),)
F
.
In what follows, a connection between the fundamental function
of an extrapolation space X = L
F
and the function dened by (24)
plays an important role. In particular, if X SE, then
(u) =
[u,)
X
=
[u,)
(log
2
(2/s))
X
= (2
1u
) (u 1),
whence
(25) (t) = (log
2
(2/t)) (0 < t 1).
In the general case, we have the following: for every h (0, 1) there is
A = A(h) > 0 such that
(26)
_
hlog
2
2
t
_
A(t) if 0 < t 2
11/h
.
In fact, if p hlog
2
(2/t), then we have t
1/p
t
1
hlog
2
2/t
2
1/h
, and the
denitions of and imply that (t) C
1
2
1/h
_
hlog
2
2
t
_
, where C
is the constant of equivalence of norms in X and L
F
. Thereby we obtain
(26), with A = 2
1/h
C.
In the proof of the second part of Theorem 3.1 we will use the next
statement.
Lemma 3.4. Let X = L
F
, where F is a Banach lattice on [1, ), and
let be the fundamental function of X. Then the following conditions
are equivalent:
(a) there exists C
1
> 0 such that
(27) (t) C
1
_
_
_t
1/p
[
1
((t)),)
_
_
_
F
for all 0 < t 1;
(b) there exist C
2
> 0 and 0 < h < 1 such that
(28) (t) C
2
_
hlog
2
2
t
_
for all 0 < t 2
11/h
.
Proof. First, we assume that (b) holds. In the case when 0 < t 2
11/h
the denition of yields
(t) C
2
_
_
[hlog
2
2/t,)
_
_
F
2
1/h
C
2
_
_
_t
1
hlog
2
2/t
[hlog
2
2/t,)
_
_
_
F
2
1/h
C
2
_
_
_t
1/p
[hlog
2
2/t,)
_
_
_
F
.
Moreover, since the function
1
decreases, then (26) implies that
1
(A(t)) hlog
2
(2/t) if 0 < t 2
11/h
.
Combining this with the previous inequality, we infer
(t) 2
1/h
C
2
_
_
_t
1/p
[
1
(A(t)),)
_
_
_
F
.
At the same time, Lemma 3.3 and the concavity of the Kfunctional
yield
_
_
_t
1/p
[
1
(A(t)),)
_
_
_
F
K(A(t), t
1/p
; F, L
[1, ))
AK((t), t
1/p
; F, L
[1, )) 2A
_
_
_t
1/p
[
1
((t)),)
_
_
_
F
.
Thus, (27) is proved for 0 < t 2
11/h
and hence for all 0 < t 1.
Prove now the opposite implication (a) (b). Let h be for now
an arbitrary number from the interval (0, 1/2). If 0 < t 2
11/h
then
t < 1/2. Thereby
log
2
t
1log
2
t

1
2
, which implies t
h
1
log
1
2
2/t
2
1/(2h)
.
Therefore, assuming that hlog
2
(2/t) >
1
((t)), we obtain that
_
_
_t
1/p
[
1
((t)),hlog
2
(2/t))
_
_
_
F
t
h
1
log
1
2
2/t
_
_
[
1
((t)),)
_
_
F
2
1/(2h)
(t).
Thus, choosing h (0, 1/2) so that 2C
1
2
1/(2h)
< 1, by (27), we have
(t) C
1
_
_
_t
1/p
[
1
((t)),)
_
_
_
F
C
1
_
_
_t
1/p
[
1
((t)),hlog
2
(2/t))
_
_
_
F
+C
1
_
_
_t
1/p
[hlog
2
(2/t),)
_
_
_
F
C
1
2
1/(2h)
(t) +C
1
(hlog
2
(2/t)) (1/2)(t) +C
1
(hlog
2
(2/t)) ,
whence
(t) 2C
1
(hlog
2
(2/t)) (0 < t 2
11/h
).
Since in the case when hlog
2
2/t
1
((t)) the last inequality is a
straightforward consequence of (27), the proof is completed.
Proof of Theorem 3.1. Since for X = L
both statements of the the-

orem are obvious, we conne exposition to the case when X = L
.
Therefore, we may assume that the fundamental function of X sat-
ises the conditions mentioned before Lemma 3.2. Since (1) = 1, we
have that x
X
x
(x L
), whence
(29) K(t, x; X, L
) = x
X
(t 1).
Moreover, a
X
a
L
[1,)
(a L
[1, )) and, therefore, if a

X
then
K(t, a;

X, L
[1, )) = a
X
(t 1).
The last equality and (29) imply that under the condition X = L
X
we
have
K(t, x; X, L
) K(t, x
p
;

X, L
[1, )) (t 1).
Thus, taking into account Lemmas 3.2 and 3.3 and equality (25), we see
that (i) will be proved as soon as we will show that
(30)
_
_
_
_

1
(t)
0
x
(s)
p
ds
_
1/p
_
_
_
X
x
p
[V (t),)
X
(0 < t 1),
where V (t) := log
2
_
2/
1
(t)
_
.
First of all, in the case when p V (t)
_

1
(t)
0
x
(s)
p
ds
_
1/p

1
(t)
1/p
x
p

1
(t)
1/V (t)
x
p
.
Since
log
2
z
1log
2
z
1 (0 < z 1), the denition of V (t) and the previous
inequality yield
_

1
(t)
0
x
(s)
p
ds
_
1/p
1
2
x
p
(p V (t)),
whence
(31)
_
_
_
_

1
(t)
0
x
(s)
p
ds
_
1/p
_
_
_
1
2
x
p
[V (t),)
X
(0 < t 1).
To prove the opposite inequality, we may assume that
(32) {s [0, 1] : x
(s) > 0} [0,

1
(t)].
If 2
1p

1
(t) (which is equivalent to the inequality p V (t)), then
x
(2
1p
) 2
_
2
1p
0
x
(s)
p
ds
_
1/p
2 x
p
[V (t),)
(p).
From this and (32) it follows
x
X
= x
(2
1p
)
X
2 x
p
[V (t),)
X
.
On the other hand, since X is a strong extrapolation space, then
_
_
_
_

1
(t)
0
x
(s)
p
ds
_
1/p
_
_
_
X
Cx
X
.
The last two relations yield
_
_
_
_

1
(t)
0
x
(s)
p
ds
_
1/p
_
_
_
X
2Cx
p
[V (t),)
X
,
which combined with (31) implies (30). Thus, the proof of (i) is com-
pleted.
Proceed with the proof of (ii). Firstly, since F L
[1, ), we have
K(t, a; F, L
[1, )) a
F
(t 1)
for all a F. Combining this with (29) and (23) we obtain that X = L
F
.
Hence, applying Lemmas 3.2 and 3.3 (we may assume that the function
dened by (24) satises the conditions stated before the last of them)
and hypothesis (23), we infer
_
_
_
_

1
(t)
0
x
(s)
p
ds
_
1/p
_
_
_
F
x
p
[
1
(t),)
F
(0 < t 1).
Choosing x =
[0,u)
and t = (u), where 0 < u 1, we conclude that
(u) u
1/p
F
C
_
_
_u
1/p
[
1
((u)),)
_
_
_
F
.
Thus, Lemma 3.4 yields (28). This fact and (26) imply that
(u) C
2
(hlog
2
(2/u)) = C
2
_
(h/2) log
2
(2/u)
2
_
C
2
_
(h/2) log
2
(2/u
2
)
_
C
2
A(h/2)(u
2
)
for some h (0, 1) and all 0 < u 2
1/21/h
, i.e.,
2
. The proof is
completed.
As it was showed there are Marcinkiewicz spaces M() E such
that
2
(see Example 2.14). At the same time, from Theorems 2.11
and 3.1 it follows the following result.
Corollary 3.5. If X is an interpolation space with respect to a Banach
couple ((), M()) (in particular, if X = () or X = M()), then
the following conditions are equivalent:
(i) there exists a Banach lattice F on [1, ) such that F L
and
equivalence (23) is valid with constants independent of x X and t > 0;
(ii) X is a strong extrapolation space.
In both cases relation (23) is fullled for F =

X.
Example 3.6. Consider an exponential Orlicz space ExpL
generated
by the function e
(u)
1, where is an increasimg convex function
on the half-line (0, ) and (0) = 0. As it is well known [23, 42],
we have ExpL
= M() (with equivalence of norms), where (u) =

1/
1
(log
2
(2/u)) . Since
1
is a concave function, then
1
_
log
2
(2/u
2
)
_

1
(2 log
2
(2/u)) 2
1
(log
2
(2/u)) .
Therefore, (u) 2(u
2
), which implies that
2
. Thus, ExpL
SE and
K(t, f; ExpL
, L
)
_
_
f
p
[
1
(t),)
_
_
F
(0 < t 1).
In this case,
1
(t) = (1/t) (0 < t 1) and, by (13), F = L
((2
p
)).
Thereby
K(t, f; ExpL
, L
) sup
p(1/t)
f
p
1
(p)
(0 < t 1),
whence
K(t, f; L
, ExpL
) tK(1/t, f; ExpL
, L
) t sup
p(t)
f
p
1
(p)
(t 1).
The last relation coincides with the statement of Proposition 1 in [4]
(see also Theorem 2 in [1]) to within change of variables.
In the power case, i.e., (t) = t
, we arrive at the classical Zygmund

space ExpL
and the equivalence

K(t, f; L
, ExpL
) t sup
pt
f
p
p
1/
= t sup
pt
f
p
p
is valid for every > 0 [2, Theorem 2].
4 Extrapolation description of Orlicz spaces.
In this section we look for conditions under which an Orlicz space is
a limiting space of the L
p
scale or, in other words, when it is possi-
ble to describe the norm of an Orlicz space by L
p
norms. Analogous
problems and various applications to dierential equations, imbedding
theorems, probability theory, orthogonal series have been considered in
[1, 2, 4, 6, 12-14, 26, 29-33, 37-40]. In particular, Mamontov [29] [33]
has developed the techniques of special integral transformations and
used them to prove some imbedding theorems for Orlicz spaces and in-
tersections of L
p
spaces.
Note that our exposition is close to that by Mamontov and Ostrovsky
[38] [40]. The main idea is to look at L
p
spaces as Orlicz spaces
and a key observation is the following: an intersection of Orlicz spaces
generated by functions M
( I) often coincides with the Orlicz space,

which is generated by the envelope of the family {M
}
I
, i.e., by the
function M(u) = sup
(u).
Lemma 4.1. Let M
(u) ( I) be arbitrary Orlicz functions and let

L
:=
I
L
M
, where L
M
are Orlicz spaces generated by the functions

M
. Then
(33) L
M
L
and x
L
x
L
M
,
where M(u) := sup
(u) (u > 0).

Moreover,
L
M
(t) =
L
(t) (0 t 1).
Proof. Since (33) is obvious, we should prove only the equality of funda-
mental functions. By the denition of the space L
, we have
L
M
(t)
(t). Therefore,
1
0
M
[0,t]
(s)
(t)
_
ds 1
for all t (0, 1] and I. On the other hand, for any > 0 and
t (0, 1] there is = (, t) such that
M
_
1
(t)
_
(1 +)M
_
1
(t)
_
.
Thus,
1
0
M
_
[0,t]
(s)
(t)
_
ds (1 +)
1
0
M
[0,t]
(s)
(t)
_
ds 1 +,
which implies
L
M
(t)
L
(t). The opposite inequality is a consequence

of (33).
Let (L
p
) :=
1p<
((p)L
p
) , where (p) is a nonnegative boun-
ded measurable function on [1, ). It is not hard to check that (L
p
) is
an r.i. space on [0, 1] and (L
p
) =
p
0
p<
((p)L
p
) with equivalence
of norms whenever (p
0
) > 0 (p
0
> 1).
Setting M
(u) = M
p
(u) = (p)
p
u
p
, we see that the space (L
p
)
coincides with the space L
from Lemma 4.1 with equality of norms. We

will examine under which conditions L
= L
M
, where M is the envelope
of the family {M
p
}
1p<
, i.e., M(u) = sup
1p<
(p)
p
u
p
(u > 0). Note
that it can be assumed that (p) 0 as p since, otherwise,
M(u) = for u > 1, or, equivalently, (L
p
) = L
. We begin with
the following result.
Theorem 4.2. Let M
p
(u) = (p)
p
u
p
and let
M(u) := sup
p1
M
p
(u)
1
M
p
(u)R(p) dp, for u > u
0
,
where a nonnegative function R(p) satises the condition
(34)
1
R(p)
p
dp < ,
with some > 0. Then (L
p
) = L
M
.
Proof. Suppose that x (L
p
) and x
(L
p
)
< 1. By the denition
of the functions M
p
, we have
1
0
M
p
_
|x(t)|
_
dt
1
p
,
with from hypothesis (34). Therefore,
1
0
M
_
|x(t)|
_
dt M(u
0
) +
1
1
M
p
_
|x(t)|
_
R(p) dp dt =
= M(u
0
)+
1
_
_
1
0
M
p
_
|x(t)|
_
dt
_
_
R(p) dp M(u
0
)+
1
R(p)
p
dp < ,
which implies that x L
M
. The opposite imbedding follows from
Lemma 4.1.
Next, we want to clarify when an Orlicz function may be represented
in the form
(35) M(u) = sup
p1
(p)u
p
,
with some function (p).
Lemma 4.3. A function M(u) may be represented in the form (35) if
and only if
(36) sup
p1
_
inf
v>0
M(v)
v
p
_
u
p
= M(u).
Proof. Firstly, if equality (36) holds then, setting (p) := inf
v>0
M(v)
v
p
,
we obtain (35).
Conversely, assume that (35) is fullled with some (p). Then
(p)
M(u)
u
p
, for all u > 0 and p 1. Therefore, (p) inf
v>0
M(v)
v
p
and, nally,
M(u) = sup
p1
(p)u
p
sup
p1
_
inf
v>0
M(v)
v
p
_
u
p
sup
p1
_
M(u)
u
p
_
u
p
= M(u).
Suppose that a function M(u) may be represented in the form (35).
Putting

(p) := sup
qp
(q), we see that
M(u) = sup
p1
(p)u
p
for all u 1. Since Orlicz spaces generated by equivalent functions have
equivalent norms we may assume that (p) decreases and (1) = 1. In
this case M(u) = u if 0 u 1.
After change of variables u = e
t
, v = e
s
and taking the logarithm of
both sides of equality (36) we obtain that
log M(e
t
) = sup
p1
_
pt sup
sR
{sp log M(e
s
)}
_
, t R.
Since log M(e
t
) = t if t < 0, we may put in the right hand side of
the last equality the supremum over all p R. By the well-known
FenchelMoreau theorem [44, Remark 1.63, p. 26], the last means that
the function N(t) := log M(e
t
) is convex. Recall that the Legendre
conjugate function N
for a function N is dened by

N
(p) = sup
tR
{pt N(t)}.
In our case N
(p) increases for p > 1, N
(1) = 0, and N
(p) = + if
p < 1. Thus, Lemma 4.3 implies the following statement.
Theorem 4.4. An Orlicz function M(u) may be represented in the form
(35) if and only if
(37) M(u) = e
N(log u)
,
with some convex function N such that N(t) = t if t < 0. Moreover,
(p) = e
N
(p)
.
Remark 4.5. In the case when N(t) = qt + C (t > 0, q 1) we have
M(u) = C
1
u
q
(u > 1) and (p) = 0 for p > q.
The following results follow from Theorems 4.2, 4.4 and Lemma 4.1.
Theorem 4.6. Let a function M be of the form (37) with some convex
function N(t) such that lim
t+
N(t)/t = . If
(38) e
N(t)
1
e
ptN
(p)+Cp
dp, for t t
0
,
with some C > 0, then L
M
= (L
p
), where
(39) (p) = e
(p)
p
.
Proof. According to Theorem 4.4
M(u) = sup
p1
M
p
(u), where M
p
(u) = (p)
p
u
p
= e
p log uN
(p)
.
After change of variables t = log u in (38) we see that conditions of
Theorem 4.2 hold for R(p) = e
Cp
, = e
C+1
, and u
0
= e
t
0
.
Theorem 4.7. Let a function M be of the form (37) with some convex
function N(t) and let the Orlicz space L
M
coincide with a Marcinkiewicz
space. Then L
M
= (L
p
), where is dened by (39).
Proof. By Theorem 4.4 and Lemma 4.1, L
M
(L
p
) and the fun-
damental functions of these spaces are equal. Since L
M
is a Marcin-
kiewicz space being the largest space among all r.i. spaces with the
same fundamental function [21, Theorem 2.5.7], we deduce the opposite
imbedding.
Remark 4.8. In [38], a somewhat similar result is proved. Moreover,
in the same article its applications can be found as well.
Remark 4.9. Note that the function N
(p) for p large enough is dened

by values of N(p) for large p. Therefore, by the standard extending
argument, it is not hard to check that Theorems 4.6 and 4.7 will be still
valid if equality (37) is assumed to be only for u u
0
(u
0
> 0), where
N(t) is a function convex for t t
0
.
Remark 4.10. Recall that an Orlicz space L
M
coincides with a Mar-
cinkiewicz space if and only if the function M satises condition (7).
Suppose M can be represented in the form (37). Then after logarithmic
change of variables we see that (7) is equivalent to the following:
(40)
s
0
e
N(N
1
(s)C)s
ds < ,
with some constant C > 0. Moreover, the change N
1
(s) = x allows to
rewrite condition (40) as follows
(41)
x
0
e
N(xC)N(x)
N
(x) dx < .
Next, using the convexity of the function N together with (40) and (41),
we obtain the following conditions guaranteeing the coincidence of the
corresponding Orlicz and Marcinkiewicz spaces:
(42)
s
0
exp
_
Cs
N
1
(s)
_
ds <
and
(43)
x
0
exp
_
CN(x)
x
_
N
(x) dx < .
Analogously, by the inequality
CN
(x) < N(x C) N(x),

we conclude that the Orlicz and Marcinkiewicz spaces do not coincide
whenever for arbitrary C > 0 we have
(44)
x
0
e
CN
(x)
N
(x) dx = .
Let us state a result dual with respect to Theorem 4.7.
Theorem 4.11. Let be an increasing concave function on [0, 1] such
that (t) te
L(log(1/t))
, with a concave positive function L(u) on [0, +)
satisfying the condition lim
u+
L(u)/u = 0. If the Lorentz space ()
coincides with an Orlicz space, then
() =
1<p<p
0
_
e
N
(p
)
p
L
p
_
,
where N
(p) is a Legendre conjugate function to the Orlicz function

N(t) = L
1
(t) and p
= p/(p 1).
Proof. Firstly, the Marcinkiewicz space M( ) , where (t) = t/(t),
coincides with the Orlicz space L
M
, M(u) = e
N(log u)
. According to
Theorem 4.7
(45) y
M( )
sup
p
0
<p
<
exp
_
(p
)
p
_
y
p
.
By [21, Theorem 2.5.4],
x
()
sup
_
_
_
1
0
x(t)y(t) dt : y L
, y
M( )
= 1
_
_
_
.
Hence, for any representation x =
i=1
x
i
(convergence in L
1
) we have
0
x(t)y(t) dt
i=1
0
x
i
(t)y(t) dt
i=1
x
i
p
y
p
i=1
e
N
(p
)
p
x
i
p
e
(p
)
p
y
p
sup
p
_
e
(p
)
p
y
p
i=1
e
N
(p
)
p
x
i
p
_
.
Combining the last inequality with the previous one and with (45), we
conclude that
()
1<p<p
0
_
e
N
(p
)
p
L
p
_
.
To prove the opposite inequality, we estimate the fundamental func-
tion of the space :=
1<p<p
0
_
e
N
(p
)
p
L
p
_
. Since by (45),
t
(t)
=
(0,t)
M( )
sup
p
0
<p
<
e
(p
)
p

(0,t)
= sup
p
0
<p
<
e
(p
)
p
t
1
p
,
then
(t) =
(0,t)
inf
1<p<p
0
e
N
(p
)
p

(0,t)
p
=
t
sup
p
0
<p
<
e
(p
)
p
t
1
p
(t).
Recall that a Lorentz space is the smallest one among all r.i. spaces with
the same fundamental function [21, Theorem 2.5.5]. Therefore, the last
inequality implies that
() (
)
1<p<p
0
_
e
N
(p
)
p
L
p
_
,
which completes the proof.
It is instructive to compare Theorem 4.11 with the following result
which may be proved using Jawerth-Milmans theory.
Consider the Banach couple

A = (L
1
, L
) of the spaces on [0, 1] and

the corresponding spaces

A
,1,J
of the real Jmethod of interpolation
(for its denition see, for instance, [10, 3.2]). It is well known that in
this case the spaces

A
,1,J
coincide (up to equivalence of norms) with
the Lorentz spaces L
p,1
, for = 1 1/p [10, Theorem 5.2.1]. Therefore,
by [34, Ch. 2, formula (2.16)],
0<<
0
_
M()
A
,1,J
_
= (),
where
(t) = t (1/t) , (t) := inf
0<<
0
M()t
.
Moreover, if M is a tempered function as 0, i.e. M() M(2),
then from [34, Ch. 2, Theorem 5] it follows that
0<<
0
_
M()
A
,1,J
_
=
0<<
0
_
M()
A
,r
_
(see Introduction as well). Let (q) be a nonnegative function on (1, )
such that (q) (2q) as q . Setting M() := 1/(q), where
= 1/q, we see that M() M(2) as 0. Since
A
,1
L
p
=
A
,p

A
,
, for = 1
1
p
=
1
p
,
with uniformly bounded constants, then the previous relations imply
that
1<p<p
0
_
1
(p
)
L
p
_
= (), where (t) =
t
sup
q
0
<q<
(q)t
1/q
,
provided if (q) (2q) as q .
We see that this result is completely consistent with Theorem 4.11,
but the last one contains essentially weaker assumptions on a weight
(q). We may say the same concerning to Theorem 4.7 as well.
In conclusion, we consider some examples of extrapolation Orlicz
spaces. Emphasize that only the rst of them (which appeared in the
second part of Yanos theorem) may be derived using JawerthMilmans
theory. In this connection, note that an Orlicz space L
M
is a strong
extrapolation space if and only if the function M satises the following
condition: there is C > 0 such that
(46) M
1
(u) CM
1
(
u) (u 1).
In fact, let be the fundamental function of L
M
. Since (t) =
1
M
1
(1/t)
,
then (46) means that
2
. Therefore, by Theorem 2.2 and Proposi-
tion 2.10, the condition L
M
SE implies (46). On the other hand, from
(46) it follows that
1
0
M
_
1
C
M
1
_
1
t
__
dt
1
0
M
_
M
1
_
1
t
__
dt = 2.
Hence we obtain (7) and L
M
coincides with the Marcinkiewicz space
M(). Then, by Theorem 2.11, L
M
is a strong extrapolation space.
In the following examples we will repeatedly make use of the property
indicated in Remark 4.9.
Example 4.12. Let > 0 and M(u) = e
u
if u u
0
, where u
0
> 0
is large enough. This function is representable in the form (37), with
N(t) = e
t
. It is not hard to verify that this function satises condition
(40) and hence the corresponding Orlicz space ExpL
coincides with the

Marcinkiewicz space M(), where (t) = log
1/
(e/t). Therefore, by
Theorem 4.7, the corresponding Orlicz space ExpL
is an extrapolation
one. Moreover, elementary calculations show that
N
(p) =
p
log
_
p
,
and we infer
x
ExpL
sup
p1
x
p
p
1/
.
Example 4.13. Let > 1 and > 0. Consider the Orlicz space L
M
,
generated by the function M
,
(u) = e
log
(u)
if u u
,
, where u
,
>
0 is large enough. This function is of the form (37), with N
,
(t) = t
.
Moreover, since condition (42) holds, L
M
,
is a Marcinkiewicz space.
Taking into account that
N
,
(p) =
p
()
1
if p p
0
, where
1
+
1
= 1,
and using Theorem 4.7, we infer
L
M
,
=
_
exp
_
p
1
()
1
_
L
p
_
,
which means that
x
L
M
,
sup
p1
x
p
exp
_
p
1
()
1
_
.
In the case when = 1/ these expressions are simplied and for the
function M
(u) = e
1
log
(u)
(u u
) we have the following extrapola-

tion formula
x
L
M
sup
p1
x
p
exp
_
p
1
_
.
Example 4.14. Let 1. Consider the Orlicz space generated by the
function of the form (37), with N(t) = N
(t) = t log
t if t t
, where
t
is large enough. Since N
(t) grows at innity essentially more slowly

than the function N
,
(t) from the previous example, then the Orlicz
space L
M
, where M
(u) = e
N
(log u)
, is wider than the Orlicz space
L
M
,
from Example 4.13. However, L
M
is an extrapolation space as
well. In fact, since
x
0
e
2 log
x
log
xdx < if 1,
then (43) holds. Therefore, L
M
coincides with the corresponding Mar-

cinkiewicz space.
It is not hard to nd the Legendre conjugate function to an Orlicz
function which is equivalent to N
(see, for example, [20]). But it is

insucient because the corresponding Orlicz spaces may be dierent.
At the same time, to nd the Legendre conjugate function N
in the
general case is not an easy thing. However, if = 1 and = 2 elementary
calculations show that
N
1
(p) = e
p1
and N
2
(p) = pe
p+11
/
p + 1,
respectively. Therefore, by Theorem 4.7, we have
x
L
M
1
sup
p1
x
p
exp
_
e
p1
p
_
and
x
L
M
2
sup
p1
x
p
exp
_
p+11
p + 1
_
,
where
M
1
(u) = exp (log ulog log u) and M
2
(u) = exp
_
log ulog
2
log u
_
,
respectively.
Note that in the case = 1 we may make use of Theorem 4.6 as
well, without exploiting the fact of the coincidence of the Orlicz and
Marcinkiewicz spaces. We must check only condition (38), which is a
consequence of the following inequality
e
t log t
0
e
pte
p1
+p
dp if t t
0
.
In fact, the integrand attains its maximum at the point p
0
= log(t+1)+1.
Therefore,
0
e
pte
p1
+p
dp
p
0
+
1
t
p
0
e
pte
p1
+p
dp
1
t
exp
__
p
0
+
1
t
_
t e
p
0
+
1
t
1
+p
0
+
1
t
_
=
= exp
_
(t + 1) log(t + 1) +t + 2 +
1
t
(t + 1)e
1
t
log t
_
e
t log t
if t is large enough, because of
lim
t+
_
(t + 1) log(t + 1) +t + 2 +
1
t
(t + 1)e
1
t
log t t log t
_
= 1.
The last example shows that there are Orlicz spaces generated by
functions of the form (37) that do not satisfy the condition of Theorem
4.7.
Example 4.15. Let N
be of the same form as in the previous example

but now (0, 1). Since for arbitrary C > 0
x
0
e
C log
x
log
xdx = ,
then (44) holds, and the corresponding Orlicz space doesnt coincide
with any Marcinkiewicz space. Analogously, Theorem 4.7 cannot be
used for the function N(t) = t log log t if t t
0
.
5 Concluding remarks and some applications.
5.1 Discrete extrapolation relations.
Here we conne to consideration of L
p
spaces with integer p. Firstly,
show that in contrast to the case of the continuous scale any extrapola-
tion functor of the type always gives on the scale {L
k
}
k=1
an Orlicz
space. In other words,
x
d
(L
k
)
:= sup
kN
(k)x
k
is an Orlicz norm, for an arbitrary nonnegative sequence {(k)}
k=1
.
Taking into account Lemma 4.1, it is sucient to check that
L
:=
k=1
L
M
k
L
M
,
where
M
k
(u) := (k)
k
u
k
(k N), M(u) = sup
k=1,2,...
(k)
k
u
k
.
Consider the function G(u) :=
k=1
M
k
(u). If x L
and x
L
< 1,
then
1
0
M
k
(|x(t)|) dt 1 for all k N. Therefore,
1
0
G
_
|x(t)|
2
_
dt =
k=1
1
0
M
k
_
|x(t)|
2
_
dt
k=1
1
2
k
= 1,
whence x
L
M
x
L
G
2, i.e., x L
M
.
Remark 5.1. Analogous arguments show that x = sup
k>2
(k)x
log k
is an Orlicz norm, for an arbitrary nonnegative sequence {(k)}
k=1
.
Now, compare a discrete extrapolation method with its correspond-
ing continuous version. The following example shows that in the case
when (p) tends to zero (as p ) not too fast these methods are
equivalent, i.e., give the same extrapolation spaces.
Example 5.2. Let be an arbitrary quasi-concave function on [0, 1].
Then, it follows easily that
sup
kN
(2
k
)x
k
sup
p1
(2
p
)x
p
.
In particular, this implies that for the Orlicz spaces L
M
from Example
4.13 if 2 we obtain
L
M
=
d
(exp(k
1
/)L
k
), where
1
+
1
= 1,
i.e.,
x
L
M
sup
kN
x
k
e
k
1
/
.
Remark 5.3. It is not hard to prove that the analogous result holds
for every extrapolation method generated by a tempered Banach lat-
tice F (see Section 2). However, Example 5.2 shows that discrete and
continuous methods may coincide by essentially weaker assumptions.
In general, a discrete extrapolation method and its corresponding
continuous version give, of course, dierent spaces.
Example 5.4. Let x(t) = e
log
3/4
1/t
(0 < t 1). Using asymptotic
Laplaces method, it can be proved that
(47) x
p
e
p
3
, where =
27
256
.
In fact, after change of variables log 1/t = s, we have
x
p
p
=
1
e
ps
3/4
s
ds.
The integrand increases on the interval [1, s
p
) and decreases on (s
p
, ),
where s
p
:= (3p/4)
4
. Therefore,
x
p
p

s
p
s
p
1
e
ps
3/4
s
ds exp
_
p
_
(3p/4)
4
1
_
3/4
(3p/4)
4
+ 1
_
,
whence x
p
ce
p
3
. On the other hand, if p is large enough, then
x
p
p
e
p+p
4
+
e
p
e
s
2
ds 2e
p+p
4
.
Combining this with the previous inequality, we obtain (47). Thus,
(48) lim
p
x
p+1/2
x
p
= .
Setting
(p) :=
k=1
1
x
k
[k,k+1)
(p)
and taking into account (48), we see that
(p) 0 as p , x
d
(L
k
), but x (L
p
).
5.2 A sharp extrapolation Yano type theorem.
As usual, let x
(t) =
1
t
t
0
x
(s) ds. Given > 0 and an r.i. space

X dene the space X(log
) as follows: it consists of all functions

x(t) such that x
(t) log
2
(2/t) X and it is endowed with the norm
x
X(log
)
= x
(t) log
2
(2/t)
X
. Such spaces arise in the theory of
interpolation of weak type operators [21, 2.6] as well as in studying
geometry of r.i. spaces (for instance, see [5, Theorem 2.17]).
First, we show that the replacement of two stars with one star in
the norm of X(log
) leads to an equivalent functional provided that X

is a strong extrapolation space.
Proposition 5.5. Let X SE. Then X(log
) consists of all functions

x(t) such that x
(t) log
2
(2/t) X and
(49) x
X(log
)
x
(t) log
2
(2/t)
X
.
Proof. Since x
(t) decreases, we obtain the inequality

x
(t) log
2
(2/t)
[1/2,1]
X
2x
(t)
[1/2,3/4]
X
2 3
(t) log
2
(2/t)
[1/4,1/2]
X
.
Hence,
x
X(log
)
x
(t) log
2
(2/t)
[0,1/2]
X
+ x
(t) log
2
(2/t)
[1/2,1]
X
(1 + 2 3
)x
(t) log
2
(2/t)
[0,1/2]
X
.
Further, assume that t [0, 1/2]. Then,
x
(t) =
1
t
n=0
t
2
n
t
2
n+1
x
(s) ds
n=0
x
_
t
2
n+1
_
t
2
n
1
.
Therefore,
x
(t) log
2
(2/t)
n=0
x
_
t
2
n+1
_
log
2
_
2/t
2
n+1
_
log
2
_
2/t
2
n+1
_
log
2
(2/t)
t
2
n
1
n=0
S
n+1
_
x
(t) log
2
(2/t)
_
2
(n+1)
_
1
2
_
2
n
1
,
where, as above, Sx(t) = x(t
2
). By Theorem 2.2, the operator S is
bounded in X and hence
x
X(log
)
= x
(t) log
2
(2/t)
X
(1 + 2 3
)
_

n=0
S
n+1
2
(n+1)+12
n
_
_
_
x
(t) log
2
(2/t)
_
_
X
= C
_
_
x
(t) log
2
(2/t)
_
_
X
.
Since x
(t) x
(t), the opposite inequality is obvious.

Corollary 5.6. Under the conditions of the previous proposition, X(log
)
is a strong extrapolation space. Moreover, we have that X(log
) =
L
X(p
)
, where
f
X(p
)
:= f(p) p
X
.
Proof. Formula (49) indicates that the boundedness of S in X implies
its boundedness in X(log
). Therefore, by Theorem 2.2, X(log
) is a
strong extrapolation space as well, i.e., X(log
) = L

X(log
)
. The last
assertion follows now from the isometric equality L

X(log
)
= L
X(p
)
,
which can be checked directly.
Theorem 5.7. Let T be a bounded linear operator in L
p
for all p
p
0
1 and for some > 0 and C > 0
(50) T
L
p
L
p
Cp
(p p
0
).
Then, for every strong extrapolation r.i. space X the operator T is
bounded from X in X(log
).
Moreover, there exists a linear operator T
0
satisfying condition (50)
with the following property: if T
0
is bounded from X SE into an r.i.
space Y, then
(51) X(log
) Y.
Proof. Estimate (50), Corollary 5.6, and Remark 2.6 yield
Tx
X(log
)
C
1
_
_
Tx
p
p
_
_
X,p
0
C
2
x
p
X,p
0
C
3
x
X
and the rst assertion of the theorem is proved.
Given > 0 we consider the operator
T
0
x(t) =
1
t
log
1
(s/t)
s
x(s) ds.
It is not hard to check that the associated operator to T
0
is dened as
follows:
T
0
x(t) =
1
t
t
0
log
1
(t/s)x(s) ds.
Let us estimate
T
0
x
q

1
0
log
1
(1/s)x(st)
L
q
(t)
ds
1
0
s
1/q
log
1
(1/s) dsx
q
.
Therefore, by the equality
1
0
s
1/q
log
1
(1/s) ds =

0
e
t(1/q1)
t
1
dt
=
_
q
q 1
_

0
e
v
v
1
dv =
_
q
q 1
_
(),
where (x) is the Euler gamma-function, we obtain that
T
0
L
p
L
p
= T
L
p
L
p
() p
, where
1
p
+
1
p
= 1,
for every 1 < p < . Thereby T
0
satises condition (50). Suppose that
X SE and that T
0
is bounded from X into an r.i. space Y.
Let x X(log
). Then, for some y X we have

x
(t) y(t) log
2
(2/t) (0 < t 1)
and, by [21, p. 93], we obtain that
(52) x
(t) (y(t) log
2
(2/t))
(t/2) log
2
(4/t) z
(t) log
2
(2/t),
with some z X. Put z
1
(t) = z
(t
2
). By Theorem 2.2 and by the
assumption X SE, we have that z
1
X and
T
0
z
1
(t)

t
t
log
1
(s/t)
s
z
1
(s) ds z
1
(
t)

t
t
log
1
(s/t)
s
ds
=
1
z
(t) log
(1/
t) =
1
2
(t) log
(1/t)

1
2
2
log
2 z
(t) log
2
(2/t)
[0,1/2]
(t).
If T
0
is bounded from X into Y then T
0
z
1
Y and from the last in-
equality and (52) it follows x Y , i.e., we have imbedding (51).
Remark 5.8. If X = L
then the space X(log
) coincides with the

Zygmund space Exp L
1/
. In this case, from Theorem 5.7 it follows the
assertion of the second part of classical Yanos theorem, i.e., if a linear
operator T satises (50) then T is bounded from L
into ExpL
1/
.
Using Theorem 5.7, we obtain the following sucient conditions for
the convergence of orthogonal series in r.i. spaces close to L
.
Corollary 5.9. Let X be a strong extrapolation space and let {
n
}
n=1
be a complete in X orthonormal system of functions such that for every
x X and for all positive integers N we have
_
_
_
N
n=1
c
n
(x)
n
_
_
_
p
Cp
x
p
if p p
0
,
where c
n
(x) are Fourier coecients of x with respect to {
n
}
n=1
. Then,
for every x X the series
n=1
c
n
(x)
n
converges to x in the space
X(log
).
In particular, by [47, Theorems 4.3.16 and 7.6.4] and [15, Theo-
rem 5.3.2], using the interpolation Marcinkiewicz theorem in the same
way as in [26], we deduce that L
p
norms of trigonometric Fourier and
WalshFourier partial sums satisfy inequality (50) if = 1. Therefore,
we obtain
Corollary 5.10. Suppose X is a separable strong extrapolation r.i.
space. Then the Fourier trigonometric and the FourierWalsh series
of a function x X converge to x in the space X(log
1
).
In the case of Lorentz spaces the last corollary was proved by Lukom-
skii [24, 25]. Moreover, he showed that this result is sharp. For Marcin-
kiewicz spaces the analogous result was proved in [26].
5.3 Rearrangement invariance of Rademacher multipli-
cator spaces.
The Rademacher functions are r
n
(t) := sign sin(2
n
t), t [0, 1], n N.
Let R denote the set of all functions of the form
n=1
a
n
r
n
, where the
series converges a.e. For an r.i. space X on [0,1], let R(X) be the closed
linear subspace of X given by R X. The Rademacher multiplicator
space of X is the space (R, X) of all measurable functions x: [0, 1] R
such that x
n=1
a
n
r
n
X, for every
n=1
a
n
r
n
R(X). It is a
Banach function lattice on [0,1] when endowed with the norm
x
(R,X)
= sup
__
_
_x
n=1
a
n
r
n
_
_
_
X
:
n=1
a
n
r
n
X,
_
_
_
n=1
a
n
r
n
_
_
_
X
1
_
.
The space (R, X) can be viewed as the space of operators from R(X)
into the whole space X given by multiplication by a measurable function.
The Rademacher multiplicator space (R, X) was rstly considered
in [11] where it was shown that for a broad class of classical r.i. spaces X
the space (R, X) is not r.i. This result was extended in [5] to include
all r.i. spaces such that the lower dilation index
X
of their fundamental
function
X
satises
X
> 0. This result motivated the study of the
symmetric kernel Sym(R, X) of the space (R, X), that is, the largest
r.i. space embedded into (R, X). The space Sym(R, X) was studied
in [5], where it was shown that, if X is an r.i. space satisfying the Fatou
property and X ExpL
2
, then Sym(R, X) is the r.i. space with the
norm x := x
(t) log
1/2
(2/t)
X
. It was also shown that any space
X that has the Fatou property and that is an interpolation space for
the couple (Llog
1/2
L, L
) can be realized as the symmetric kernel of

a certain r.i. space. The opposite situation is when the Rademacher
multiplicator space (R, X) is r.i. The simplest case of this situation
is when (R, X) = L
. In [3], it was shown that (R, X) = L
holds for all r.i. spaces X that are interpolation spaces for the couple
(L
, ExpL
2
). More exactly, in [7], it was proved that (R, X) = L
if and only if the function log

1/2
(2/t) does not belong to the closure
of L
in X, which will be denoted by X

0
.. In the paper [8], the case
when (R, X) is an r.i. space dierent from L
was investigated. In
particular, the following sucient condition for (R, X) to be r.i. has
been obtained [8, Theorem 3.4].
Theorem 5.11. If X is an r.i. space on [0, 1] such that the operator
Sx(s) = x(s
2
) is bounded in X, then (R, X) is an r.i. space.
The proof in [8] is based on some distribution function estimates. An
extrapolation proof presented here is essentially shorter showing thereby
usefulness of the concept of a strong extrapolation space.
Proof. Note that if log
1/2
(2/t) X
0
then, as it was proved in [7],
(R, X) = Sym(R, X) = L
. Thus, we may assume that log

1/2
(2/t)
X
0
. Let
k
n
:= [
k1
2
n
,
k
2
n
), where n N and k = 1, , 2
n
. In view of
[8, Proposition 3.3], it is sucient to prove the following: there exists
A > 0 such that for all n N and every c
1
c
2
c
2
n 0 we have
(53)
_
_
_
2
n
k=1
c
k
k
n
log
1/2
_
2
t
__
_
_
X
A
_
_
_
2
n
k=1
c
k
k
n
log
1/2
_
2
2
n
t + 1 k
__
_
_
X
.
Firstly, let x be a measurable function such that x L
p
for all p < .
For every 0 < t 1, setting p := log(2/t), we have
x
p
= x
log(2/t)

_
t
0
(x
(s))
log(2/t)
ds
_
1/ log(2/t)
x
(t)t
1/ log(2/t)
1
e
x
(t),
since t
1/ log(2/t)
1/e for 0 < t 1. Therefore,
(54)
_
_
_x
(t) log
1/2
_
2
t
__
_
_
X
e
_
_
_ x
log(2/t)
log
1/2
_
2
t
__
_
_
X
.
If 1 p < , then, by [11, Lemma 1], there exists a Rademacher
sum h
p
=
m
k=n+1
b
k
r
k
such that (b
k
)
l
2
= 1 and
(55) x
p
3p
1/2
x h
p
p
.
By [41] and the relation
x
L
N
sup
0<t1
x
(t) log
1/2
(2/t),
for every k = 1, . . . , 2
n
, we have
_
k
n
h
p
_
(s) =
_
mn
k=1
b
k
r
k
_
(2
n
s)
C
1
log
1/2
_
2
2
n
s
_
[0,2
n
)
(s)(b
k
)
l
2
C
1
log
1/2
_
2
2
n
s
_
[0,2
n
)
(s).
Let x =
2
n
k=1
c
k
k
n
with nonnegative and decreasing coecients c
k
.
The last inequality implies that
(xh
p
)
(s) =
_
2
n
k=1
c
k
k
n
h
p
_
(s)
C
1
_
2
n
k=1
c
k
k
n
log
1/2
_
2
2
n
s + 1 k
__
(s).
Hence, by (55)
x
p
3C
1
p
1/2
_
_
_
2
n
k=1
c
k
k
n
log
1/2
_
2
2
n
s + 1 k
__
_
_
p
.
From this, setting again p = log(2/t), we obtain that
_
_
_ x
log(2/t)
log
1/2
_
2
t
__
_
_
X
3C
1
_
_
_
_
_
_
2
n
k=1
c
k
k
n
log
1/2
_
2
2
n
s + 1 k
__
_
_
log(2/t)
_
_
_
X
.
By Theorem 2.2 and the hypotheses, the previous inequality implies that
_
_
_ x
log(2/t)
log
1/2
_
2
t
__
_
_
X
C
3
_
_
_
2
n
k=1
c
k
k
n
log
1/2
_
2
2
n
s + 1 k
__
_
_
X
.
This inequality and (54) yield (53) and the proof is completed.
References
[1] S.V. Astashkin, Some new extrapolation estimates for the scale of
L
p
spaces// Funct. Anal. Appl. 2003. V. 37, No. 3, 221-224.
[2] S.V. Astashkin, Extrapolation properties of the scale of L
p
spaces//
Sbornik: Math. 2003. V. 194, No. 6, 813832.
[3] S.V. Astashkin, On the multiplier space generated by the Radema-
cher system// 2004. Math. Notes. V. 75, No. 2, 158165.
[4] S.V. Astashkin, Extrapolation functors on a family of scales gen-
erated by the real interpolation method// Siberian Math. J. 2005.
V. 46, No. 2, 205-225.
[5] S.V. Astashkin and G.P. Curbera, Symmetric kernel of Rademacher
multiplicator spaces// J. Funct. Anal. 2005. V. 226, 173192.
[6] S.V. Astashkin and K.V. Lykov, Extrapolatory description for the
Lorentz and Marcinkiewicz spaces close to L
// Siberian Math.
J. 2006. V. 47, No. 5, 797812.
[7] S.V. Astashkin and G.P. Curbera, Rademacher multiplicator spaces
equal to L
// Proc. Amer. Math. Soc. 2008. V. 136, No. 10, 3493

3501.
[8] S.V. Astashkin and G.P. Curbera, Rearrangement invariance of
Rademacher multiplicator spaces// J. Funct. Anal. 2009. V. 256,
40714094.
[9] S.V. Astashkin and K.V. Lykov, Strong extrapolation spaces and
interpolation// Siberian Math. J. 2009. V. 50, No. 2, 199213.
[10] J. Bergh and J. Lofstrom, Interpolation Spaces, an Introduction.
Springer-Verlag, Berlin, Heidelberg, New York, 1976.
[11] G.P. Curbera, A note on function spaces generated by Rademacher
series// Proc. Edinburgh. Math. Soc. 1997. V. 40, 119126.
[12] D.E. Edmunds and M. Krbec, On decomposition in exponential Or-
licz spaces// Math. Nachr. 2000. V. 213, 7788.
[13] D.E. Edmunds and M. Krbec, Decomposition and Mosers lemma//
Revista Matematica Complutense. 2002. V. 15, No. 1, 5774.
[14] A. Fiorenza and M. Krbec, On an optimal decomposition in Zyg-
mund spaces// Georgian Math. J. 2002. V. 9, No. 2, 271286.
[15] B. Golubov, A. Emov, and V. Skvortsov, Walsh Series And Trans-
forms: Theory And Applications. Kluwer Acad. Publishers, 1991.
[16] B. Jawerth and M. Milman, Extrapolation spaces with applica-
tions// Mem. Amer. Math. Soc. 1991. V. 89, No. 440. 82 pp.
[17] B. Jawerth and M. Milman, New results and applications of
extrapolation theory// Interpolation Spaces and Related Topics,
(M.Cwikel, M.Milman, and R.Rochberg, editors), Israel Math. Con-
fer. Proc., 5. 1992, 81105.
[18] G. Karadzhov and M. Milman, Extrapolation theory: New results
and applications// J. Approx. Theory. 2005. V. 133, No. 1, 3899.
[19] R.A. Kerman, An integral extrapolation theorem with applications//
Stud. Math. 1983. V. 76, 183195.
[20] M.A. Krasnoselskii and Ya.B. Rutickii, Convex Functions and Or-
licz Spaces. Noordho Ltd., Groningen, 1961.
[21] S.G. Krein, Ju.I. Petunin, and E.M. Semenov, Interpolation of Lin-
ear Operators. Amer. Math. Soc., Providence R. I., 1982.
[22] J. Lindenstrauss and L. Tzafriri, Classical Banach spaces 2, Func-
tion spaces. Springer-Verlag, Berlin, Heidelberg, New York, 1979.
[23] G.G. Lorentz, Relations between function spaces// Proc. Amer.
Math. Soc. 1961. V. 12, 127132.
[24] S.F. Lukomskii, Convergence of Walsh series in near-L
spaces//
Math. Notes. 2001. V. 70, No. 6, 804811.
[25] S.F. Lukomskii, Convergence of Fourier series in Lorentz spaces//
East J. Approx. 2003. V. 9, No. 2, 229-238.
[26] K.V. Lykov, Extrapolation in the scale of L
p
spaces and convergence
of orthogonal series in Marcinkiewicz spaces// Vestnik Samarsk.
Univ. 2006. No. 2(42), 2843 (in Russian).
[27] K.V. Lykov, A criterion of separability of an extrapolation space//
Vestnik Samarsk. Univ. 2006. No. 4 (44), 512 (in Russian).
[28] L. Maligranda, The K-functional for symmetric spaces, Lect. Notes
in Math. 1984. V. 1070, 169182. Proc. Conf. Interpolation spaces
and Allied Topics in Analysis, Lund, Aug. 29 Sept. 1, 1983.
[29] A.E. Mamontov, Extrapolation of linear operators from L
p
into
the Orlicz spaces generated by Nfunctions of slow or fast growth,
in: Current Problems of Modern Mathematics, Novosibirsk. Univ.,
Novosibirsk, 1996. V. 2, 95-103 (in Russian).
[30] A.E. Mamontov, Scales of L
p
spaces and their relation to Orlicz
spaces// Vestnik Novosibirsk. Univ. Ser. Mathematics, mechanics,
computer science. 2006. V. 6, No. 2, 33-56 (in Russian).
[31] A.E. Mamontov, Integral representations and transforms of N
functions. I// Siberian Math. J. 2006. V. 47, No. 1, 97-116.
[32] A.E. Mamontov, Integral representations and transforms of N
functions. II// Siberian Math. J. 2006. V. 47, No. 4, 669686.
[33] A.E. Mamontov, Extrapolation from L
p
into Orlicz spaces via in-
tegral transforms of Young functions// J. Anal. Appl. 2006. V. 4,
No 42, 77-118.
[34] M. Milman, Extrapolation and Optimal Decompositions with Appli-
cations to Analysis. Springer-Verlag, Berlin, Heidelberg, New York,
1994. Lect. Notes in Math. V. 1580. 162 pp.
[35] M. Milman, A note on extrapolation theory// J. Math. Anal. Appl.
2003. V. 282, 2647.
[36] I.P. Natanson, The Theory of Functions of Real Variable. Nauka,
Moscow, 1974 (in Russian).
[37] J. Neves, On decompositions in generalized LorentzZygmund spa-
ces// Boll. Unione Mat. Ital. Sez. B Artic. Ric. Mat. (8). 2001. V. 4,
No. 1, 239267.
[38] E. Ostrovsky, Exponential Orlicz Spaces: new Norms and Applica-
tions // Electronic Publ., arXiv/FA/0406534, v.1, 25.06.2004.
[39] E. Ostrovsky, A remark on the inequalities of Bernstein-Markov
type in exponential Orlicz and Lorentz spaces// Electronic Publ.,
arXiv/FA/0411617, v.1, 27.11.2004.
[40] E. Ostrovsky and L. Sirota, Some new moment rearrangement
invariant spaces; theory and applications// Electronic Publ., ar-
Xiv/FA/0605732, v.1, 29.05.2006.
[41] V.A. Rodin and E.M. Semenov, Rademacher series in symmetric
spaces// Anal. Math. 1975. V. 1, 207222.
[42] Ya.B. Rutickii, On some classes of measurable functions// Uspekhi
Matematicheskikh Nauk. 1965. V. 20, No. 4, 205208 (in Russian).
[43] I.B. Simonenko, Interpolation and extrapolation of linear operators
in Orlicz spaces// Matematicheskii Sbornik. 1964. V. 63, No 4, 536-
553 (in Russian).
[44] S. Simons, Minimax and Monotonicity. Lect. Notes in Math. -
Berlin: Springer-Verlag, 1998.
[45] S. Yano, An extrapolation theorem// J. Math. Soc. Japan. 1951.
V. 3, No. 2, 296-305.
[46] V.I. Yudovicz, On some estimates connected with integral operators
and solutions of elliptic equations// Doklady Akademii Nauk SSSR.
1961. V. 138, No. 4, 805808 (in Russian).
[47] A. Zygmund, Trigonometric Series. V.2. Cambridge Univ. Press,
1959.
Sergey Astashkin
Department of Mathematics and Mechanics, Samara State University,
Akad.Pavlova 1, 443011 Samara, Russia
E-mail : astashkn@ssu.samara.ru
Konstantin Lykov
Department of Mathematics and Mechanics, Samara State University,
Akad.Pavlova 1, 443011 Samara, Russia
E-mail : alkv@list.ru

Sergey Astashkin and Konstantin Lykov - Extrapolation Description of Rearrangement Invariant Spaces and Related Problems

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Sergey Astashkin and Konstantin Lykov - Extrapolation Description of Rearrangement Invariant Spaces and Related Problems

Uploaded by

Copyright:

Available Formats

Extrapolation description

of rearrangement invariant spaces

it has a rather simple description. We completely

consists of all measurable on [0,1]

(t) stands for the non-increasing rearrangement of

consists of all measurable functions x(t)

for some and

, where inmum is taken over all

). Analogously, the intersection

) of the spaces on the segment [0, 1] we have

is the result of applying a weigh-

-norm to the function f

(t) it follows that x = x(t) X and

is the smallest space among all r.i. spaces on [0, 1] and

( > 0) for all u

coincides with the Marcinkiewicz

and an exponential Orlicz space

( > 0) and exponential Orlicz spaces ExpL

whose norm can be calculated by

) provided that X is a strong extrapolation

1 [21, p. 76]. Finally, given Banach couple

(t) and, by assumption, we obtain

(t) (0 < t 1/2).

(t) B x(t) for some B, > 0 and all t (0, 1]} .

(t) B x(t). In this case

(t) < 0 for t (0, t

s (s 1). It is not hard to check that max

), where = /(1 ). Then

. Then its fundamental

, then there exists x X such that lim

both statements of the the-

[1, )) and, therefore, if a

(s) > 0} [0,

= M() (with equivalence of norms), where (u) =

, we arrive at the classical Zygmund

and the equivalence

( I) often coincides with the Orlicz space,

(u) ( I) be arbitrary Orlicz functions and let

are Orlicz spaces generated by the functions

(u) (u > 0).

(t). The opposite inequality is a consequence

from Lemma 4.1 with equality of norms. We

for a function N is dened by

(p) increases for p > 1, N

(p) for p large enough is dened

(x) < N(x C) N(x),

(p) is a Legendre conjugate function to the Orlicz function

) of the spaces on [0, 1] and

coincides with the

) we have the following extrapola-

is large enough. Since N

(t) grows at innity essentially more slowly

coincides with the corresponding Mar-

(see, for example, [20]). But it is

be of the same form as in the previous example

(s) ds. Given > 0 and an r.i. space

) as follows: it consists of all functions

) leads to an equivalent functional provided that X

) consists of all functions

(t) decreases, we obtain the inequality

(t), the opposite inequality is obvious.

). Therefore, by Theorem 2.2, X(log