Siberian Mathematical Journal, Vol. 47, No. 5, pp.

797812, 2006
Original Russian Text Copyright c _2006 Astashkin S. V. and Lykov K. V.
EXTRAPOLATORY DESCRIPTION FOR THE LORENTZ
AND MARCINKIEWICZ SPACES CLOSE TO L

S. V. Astashkin and K. V. Lykov UDC 517.982.27

Abstract: We introduce the notion of a rearrangement invariant extrapolation space on [0, 1]. We
obtain some sucient conditions (also necessary in some cases) for the Marcinkiewicz and Lorentz
spaces to be extrapolation spaces. We generalize and improve the previous results, which enables us
to determine the possible limits of such description of spaces and thereby to establish assertions of the
Yano-type theorem. In particular, we present some examples of the spaces close to L

in a sense but
lacking this description.
Keywords: rearrangement invariant space, Lorentz space, Marcinkiewicz space, extrapolation, inter-
polation
Introduction
It is well known that many important operators of analysis, for example the maximal Hardy
Littlewood operator, the singular Hilbert operator, and the conjugate function operator in harmonic
analysis, act boundedly in the L
p
-spaces for 1 < p < but unboundedly at the endpoints of this scale,
i.e., in L

and L
1
(or at least in one of those spaces). This circumstance was one of the reasons giving
rise to the extrapolation assertions the rst of which is apparently the classical Yano theorem (see [1]
or [2, Chapter 12, Theorem 4.41]). Its part on the operators bounded in the spaces close to L

can be
stated as follows: If T is a bounded linear operator in L
p
= L
p
[0, 1] for all suciently large p < and
the estimate |Tx|
Lp
Cp

|x|
Lp
holds for some > 0, where C > 0 is independent of x = x(t) and ,
then T acts boundedly from L

to the Zygmund space Exp L

1/
with the norm
|x|
ExpL
1/ = sup
0<t1
log

2
_
2
t
_
x

(t).
Here and in the sequel x

(t) is the nonincreasing rearrangement of [x(t)[.

The next important step in the development of extrapolation theory is connected with the names of
Jawerth and Milman; developing the general approaches of the theory, they revealed the source of the
Yano-type theorems; namely, existence of the extrapolation constructions making it possible to describe
the limit spaces of such theorems [35]. In particular, they demonstrated (for example, see [5, pp. 2223])
that
|x|
ExpL
1/ sup
1<p<
(p

|x|
Lp
). (1)
It is easy to see that the Yano theorem is an immediate consequence of (1) and another elementary
extrapolation relation |x|
L
= sup
1<p<
|x|
Lp
.
More general spaces and constructions were considered in [6, 7]. In particular, extrapolatory descrip-
tion was obtained for the exponential Orlicz spaces ExpL

( is an Orlicz function on (0, )) consisting

of all functions x = x(t), t [0, 1], satisfying
|x|
ExpL
= inf
_
u > 0 :
1
_
0
_
exp
_
[x(t)[
u
_
1
_
dt 1
_
< .
Samara. Translated from Sibirski Matematicheski Zhurnal, Vol. 47, No. 5, pp. 974992, SeptemberOctober, 2006.
Original article submitted October 27, 2005.
0037-4466/06/47050797 c _2006 Springer Science+Business Media, Inc. 797
Namely [7, Corollary 1], for an arbitrary Orlicz function we have
|x|
ExpL
sup
1<p<
|x|
(p)
p
. (2)
The exponential Orlicz spaces generalize Zygmund spaces, while constituting a subclass of the rear-
rangement invariant Marcinkiewicz spaces which play an important role in many applications of operator
theory. This brings about the question of the possibility of extrapolatory description for rearrangement
invariant spaces, in particular, Marcinkiewicz and Lorentz spaces. Lately, such description turned out
to be interesting not only in its own right. It proved to be an important tool for studying convergence
of series over orthonormal systems in the spaces close to L

. In this connection, we mention the

articles [810] which are far from lling a complete list of references.
The article is organized as follows: In 1 we give the basic denitions and some preliminaries from
the theory of rearrangement invariant spaces. Moreover, here we introduce the denition of extrapolation
spaces (close to L

) with respect to the L

p
scale (1 < p < ).
The main results, presented in 2, deal with the extrapolatory description for Marcinkiewicz spaces.
Here we study two versions of this description. One of them is universal in the sense that every extrapola-
tion Marcinkiewicz space admits such a description. We obtain some necessary and sucient conditions
for a Marcinkiewicz space to be an extrapolation space in terms of the properties of the generating func-
tion of the space under study. On the one hand, these results are the far-reaching generalizations of (1)
and (2) and, on the other hand, they enable us to nd the possible limits of this description for spaces
and so of the assertions similar to the Yano theorem. In particular, we present some examples of the
spaces close to L

in a sense but lacking this description.

In 3 we consider similar questions for the Lorentz spaces predual to Marcinkiewicz spaces. The
necessary and sucient conditions found here for some quite denite extrapolatory description of these
spaces improve the recent results of S. F. Lukomskii [10] who applied them to studying convergence of
the Fourier series.
1. Denitions and Preliminaries
Henceforth by an embedding of one Banach space into the other we mean a continuous embedding;
i.e., X
1
X
0
means that x X
1
implies x X
0
and |x|
X
0
C|x|
X
1
for some C > 0. Writing X
1
= X
0
for some Banach spaces, we mean that they coincide and their norms are equivalent. The relation F
1
F
2
means that cF
1
F
2
CF
1
for some c > 0 and C > 0; moreover, these constants are independent of all
or a part of the arguments of F
1
and F
2
.
Denote by F the class of increasing continuous concave functions on [0, 1] such that (0) = 0 and
(t) > 0 (0 < t 1). Recall that, for every continuous quasiconcave function (t) on [0, 1] (i.e., (t) is
positive and increasing and (t)/t is decreasing for t > 0) such that (0) = 0, there is F for which
(t) (t) [11, p. 70].
Henceforth we speak of rearrangement invariant (symmetric) function spaces; a detailed exposition
of this theory can be found in the monographs [1113]. Recall that a Banach space X of measurable
functions on [0, 1] is called a rearrangement invariant space if the following are satised:
(a) y = y(t) X and [x(t)[ [y(t)[ imply that x = x(t) X and |x| |y|;
(b) if y = y(t) X and x

(t) = y

(t) then x X and |x| = |y| (x

(t) denotes the rearrangement

of [x(t)[, i.e., the nonincreasing function equimeasurable with [x(t)[).
The simplest and important example of rearrangement invariant spaces is given by the L
p
-spaces
(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)[.
798
The embedding L
p
L
q
holds (with constant 1) for p > q. Moreover, L

is the smallest space among

all rearrangement invariant spaces on [0, 1], as mentioned in the introduction, and the norm of an arbitrary
function x(t) in L

is expressed in terms of its norms in L

p
for p < : |x|

= sup
1<p<
|x|
p
.
Other examples of the rearrangement invariant spaces are the Lorentz and Marcinkiewicz spaces.
Let (t) F. The Marcinkiewicz space M() consists of all measurable functions x(t) on [0, 1] with the
nite norm
|x|
M()
= sup
0<s1
(s)
s
s
_
0
x

(t) dt.
The norm in the Lorentz space
p
() (1 p < ) is dened as follows:
|x|
p()
=
_
1
_
0
(x

(t))
p
d(t)
_1
p
.
The important characteristic of a rearrangement invariant space X is its fundamental function

X
(t) = |
(0,t)
|
X
, where
(0,t)
denotes the characteristic function (indicator) of the interval (0, t). In
particular,
M()
(t) = (t) and
p()
(t) = ((t))
1
p
. Moreover, () :=
1
() is the smallest space and
M() is the largest space among all rearrangement invariant spaces with the fundamental function (t)
[11, pp. 160, 162].
The dilation function of a positive function (t), t (0, 1], is dened by the relation
M

(s) = sup
0<tmin(1,1/s)
(st)
(t)
, 0 < s < .
Furthermore, the numbers

= lim
s0+
log M

(s)
log s
and

= lim
s
log M

(s)
log s
are called the lower and upper dilation indices of (t). If F then 0

1 [11, p. 76].
Henceforth we are interested in the rearrangement invariant spaces X such that X L
p
for all
p < . In particular, we will wonder when the norm in these spaces is expressed in terms of the
L
p
-norm. Moreover, we can construct rearrangement invariant spaces of the above-indicated form by
ourselves. Namely, if F is an ideal Banach space of functions on [1, ) then we can consider the set L
F
of all functions x(t) for which = (p) := |x|
p
F. It is easy to show that this set (with the usual
identication of the functions coinciding almost everywhere) becomes a rearrangement invariant space
under the norm:
|x|
L
F
:= ||
F
.
Instead of a function space as a parameter of an extrapolation space, we can also take an ideal Banach
space of sequences, requiring that the sequences |x|
n

n=1
belong to this space.
Definition 1. Let X be a rearrangement invariant space on [0, 1] such that X L
p
for all p < .
We say that X is an extrapolation space (as p ) if there is an ideal Banach space F for which
X = L
F
. We denote the set of all extrapolation spaces by E .
If F is an ideal Banach space of functions on [1, ) and w = w(p) is a positive function (weight)
then we denote by F(w) the ideal space of functions x = x(t) with the norm |x|
F(w)
= |xw|
F
. There is
a similar denition for sequence spaces.
799
2. Extrapolatory Description for Marcinkiewicz Spaces
In this section we wonder when a Marcinkiewicz space M() coincides with a space of the form L
F
,
i.e., when M() is an extrapolation space.
Let (t) F and (t) = t/(t). In the case of lim
t0+
(t) > 0 we can assume that (t) = t and
then M() = L
1
. Otherwise the function is absolutely continuous [11, p. 71] and therefore has the
derivative

. Suppose that
lim
t0+
(t) = 0 and

L
p
for all p < . (3)
It follows from the denition of the norm of a Marcinkiewicz space that
s
_
0
x

(t) dt |x|
M()
s
_
0

(t) dt, 0 < s 1.

However, this is an inequality between K -functionals on the functions x(t) and

(t) in the pair of

spaces (L
1
, L

) (for example, see [11, p. 108]). Since L

p
is an interpolation space between (L
1
, L

) with
constant 1, applying [11, p. 130], we conclude that x L
p
for all 1 p < and
|x|
p
|x|
M()
|

|
p
, 1 p < ,
or
sup
p1
|x|
p
|

|
p
|x|
M()
.
The last inequality means that
M() L
F
, (4)
where F

= L

(1/|

|
p
); moreover, the norm of (4) is equal to 1.
Theorem 1. Let F. The following are equivalent:
(1) M() E ;
(2) M() = L
F
;
(3) there is C > 0 such that
(t) C sup
p1
t
1
p
|

|
p
, 0 < t 1. (5)
Proof. First of all note that the relations of (3) are consequences of each of the conditions (1)(3)
(in particular, in the case of (1) and (2) they follow from the denition of extrapolation space). There-
fore, in all three cases we can assume (4) valid. Now, it is easy to see that (5) is an inequality between
the fundamental functions of M() and L
F
. Since M() is the largest space among all rearrangement
invariant spaces with a given fundamental function [11, p. 162], inequality (5) is equivalent to the em-
bedding L
F
M(). Consequently, by the previous remark, (5) is valid if and only if M() = L
F
.
Therefore, (2) and (3) are equivalent.
Since the implication (2) (1) is obvious, we are left with proving the converse. Assume there
is an ideal Banach space F
1
such that M() = L
F
1
. If a = a(p) F

then [a(p)[ |a|

F
|

|
p
,
p < . Since

M(), the function b = b(p) = |

|
p
belongs to F
1
and consequently (since F
1
is an ideal space) from the inequality [a(p)[ |a|
F
b(p) we obtain a F
1
and |a|
F
1
|a|
F
|b|
F
1
,
whence F

F
1
. Thereby L
F
L
F
1
= M(). Since the reverse embedding to (4) is always valid, the
theorem is proved.
Now, suppose that the upper dilation index of is nontrivial, i.e.,

< 1. Then, by [11, p. 156],

|x|
M()
sup
0<u1
(u)x

(u),
800
whence, from monotonicity of x

(t) and quasiconcavity of (t), we obtain

|x|
M()
sup
p1
(2
p
)x

(2
p
). (6)
In this case x
0
(t) =
1
(t)
M(). Therefore, it follows from (4) and the inequality

that
|1/|
p
|

|
p
with constants independent of p [1, ). Consequently, we can replace (5) of Theorem 1
for

< 1 with the condition

(t) C sup
p1
t
1/p
|1/|
p
, 0 < t 1. (7)
Corollary 1. Let F be such that

< 1. Then M() E if and only if

sup
t(0;1]
_
inf
p1
_
1/t
_
0
_
(t)
(st)
_
p
ds
_1
p
_
< .
Proof. Since, under the conditions of the corollary, (5) and (7) are equivalent, by Theorem 1,
M() E if and only if
sup
t(0;1]
_
inf
p1
_
(t)
_
1
t
1
_
0
1
(u)
p
du
_1
p
_
_
< .
Using the change of variable u = st, we obtain the assertion.
Embedding (4), which is always valid, can be considered as some upper approximation of the
Marcinkiewicz space by an extrapolation space. Find a similar lower approximation in the case when

< 1. Since
|x|
p

_
2
p
_
0
(x

(t))
p
dt
_
1/p

1
2
x

(2
p
),
from (6) we derive
|x|
M()
C sup
p1
(2
p
)|x|
p
(8)
with some constant C > 0 independent of x. It follows from (8) and (4) that
L
F
M() L
F
, (9)
where F

= L

((2
p
)).
In particular, M() is an extrapolation space, if L
F
= L
F
. At the same time, as follows from
Theorem 1, for the last equality to hold, it is sucient that M() and L
F
coincide. We will nd out
when this happens.
Lemma 1. If

< 1 then M() = L

F
if and only if the condition
1
_
0
dt
((t))
p
C
p
1
((2
p
))
p
, p 1, (10)
holds with some constant C > 0 independent of p.
Proof. Since

, it follows from (10) that (2

p
) C
1
1

p
. Consequently, F

, whence
L
F
L
F
. Combining this with (9), we conclude that L
F
= M() = L
F
.
801
Now, let M() = L
F
. Since x
0
(t) =
1
(t)
M() by (6), from the embedding M() L
F
we
obtain
(2
p
)|1/|
p
sup
p1
(2
p
)|x
0
|
p
C
1
|x
0
|
M()
= C.
Since the inequalities are equivalent to (10), the lemma is proved.
Show that (10) is equivalent to a more explicit property of .
Definition 2. We say that a function (t) F satises the
2
-condition (
2
) if there is > 0
such that
(t) (t
2
) (11)
for all t [0, 1].
Lemma 2. Let (t) F. For (t) to satisfy the
2
-condition, it is necessary and sucient that (10)
be valid.
Proof. If (10) is valid then, recalling that (t) is increasing, we obtain
2
2p
(2
2p
)
p

2
2p
_
0
dt
((t))
p

1
_
0
dt
((t))
p
C
p
1
((2
p
))
p
,
whence (2
p
) 4C(2
2p
) for all p 1.
After the change 2
p
= t the last inequality becomes (11) for t
1
2
. Hence, by quasiconcavity of (t),
(11) holds for all t [0, 1] with the constant 8C.
Now, take
2
. Then the inequality (t)
n
(t
2
n
) holds for all n N; in particular, for t = 2
p
we have
(2
p
)
n
(2
2
n
p
), p 0.
Now, choose n
0
N so that the inequality 2
n
(n +1) log
2
n holds for all n n
0
. If p 1 then
1
_
0
dt
((t))
p
2
2
n
0
p
2
2
n
0p
_
0
dt
((t))
p
= 2
2
n
0
p
+

n=n
0
2
2
n
p
_
2
2
n+1
p
dt
((t))
p
2
2
n
0
p
+

n=n
0
2
2
n
p
((2
2
n+1
p
))
p
2
2
n
0
p
+

n=n
0

(n+1)p
2
2
n
p
1
((2
p
))
p
2
2
n
0
p
1
((2
p
))
p
+

n=n
0
2
np
= 2
2
n
0
p
2
n
0
p
1 2
p
1
((2
p
))
p
C
p
1
((2
p
))
p
,
where C = 2
2
n
0
; thus, (10) is satised.
Using Lemmas 1 and 2, we can easily obtain
Theorem 2. Let (t) F and

< 1. The following are equivalent:

(1)
2
;
(2) (2
p
) 1/|

|
p
(p 1);
(3) M() = L
F
;
(4) M() = L
f
, where f

= l

((2
n
)).
Proof. Equivalence of (1)(3) follows from Lemmas 1 and 2, since the reverse inequality to (10)
is always valid. From monotonicity of f
x
(p) = |x|
p
(p 1) and quasiconcavity of (t) we derive
L
F
= L
f
; hence, each of these conditions is equivalent to (4).
802
It follows from Theorem 2 that, for
2
, both embeddings in (9) become equalities. If (t) does
not satisfy the
2
-condition then, at least for

< 1, the embedding L

F
M() becomes strict. Is
it possible that the equality M() = L
F
remains valid in this case? We give an example showing that
the answer is armative.
Example 1. Consider a function (t) F such that (t) (t) = e
1log

e/t
for some (0, 1).
Such a function in the class F exists indeed, since (t) > 0,

(t) > 0, and

(t) < 0 for t (0, t

0
)
(where t
0
is suciently small) and lim
t0+
(t) = 0.
Show that M() E . It is easy to see that

= 0 < 1. Therefore, it suces to prove (7) with

(t) = (t).
Estimate the L
p
-norm of
1

as follows:
_
_
_
_
1

_
_
_
_
p
p
=
1
_
0
e
p log

e/tp
dt = e
p+1
+
_
1
e
ps

s
ds.
Represent the integrand in the form e
ps

s
= e
ps

s
e
(1)s
, where (0, 1) is a parameter, and nd
the maximum of f

(s) = ps

s. Since f

(s) = ps
1
, s
0
= (p/)
1
1
is a maximum point of f

(s).
Consequently, for s [1, ), we have f

(s) f

(s
0
) = (1 )p
1
1
(/)

1
and
_
_
_
_
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
.
Thus,
_
_
_
_
1

_
_
_
_
p
exp
_
(1 )
_
p

_
1
__
1
1
_1
p
.
Put =
p

1+p

, where =

1
. Then
(1/(1 ))
1
p
= (1 + p

)
1
p
2p

p
< C <
for 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
p

(1 )

1
(1 )

1
(p

+ C

)
and consequently
|1/|
p
C
1
exp((1 )

1
p

1
). (12)
To prove (7), verify that the inequality
(t)|1/|
p
C
2
t
1/p
holds for p =
1

log
1
e/t with the constant C
2
= C
1
e. Indeed, on the one hand, by (12),
(t)|1/|
p
C
1
exp((1 ) log

e/t) exp(1 log

e/t) = C
2
exp( log

e/t).
On the other hand,
t
1/p
= exp(log t/p) = exp( log
1
e/t(1 log e/t)) exp( log

e/t).
Thus, M() E . At the same time, we see easily that ,
2
.
For the function (t) of Example 1 we have

= 0. The same is true for all functions satisfying the

2
-condition. In this connection, the question arises: Is there a Marcinkiewicz space M() such that
simultaneously M() E and

> 0? A partial answer to this question is contained in

803
Proposition 1. Let M() E . If

< 1 then

= 0.
Proof. Take

(0, 1). Then M

(s) > s

for some > 0 if s > 1. Consequently, for every s > 1,

there is t
0
= t
0
(s) (0, 1/s) such that
(st
0
)
(t
0
)
> s

. (13)
Since

< 1, the condition M() E is equivalent to (7) or the inequality

(t) C
1
sup
p2/
t
1/p
|1/|
p
, 0 < t 1. (14)
Choose s > 1 so that s
/2
> C
1
. Since p 1 p/2 for p 2/, by (13), we have
_
_
_
_
1

_
_
_
_
p
p
=
1
_
0
du
(u)
p

t
0
_
0
du
(u)
p

t
0
(t
0
)
p
>
s
p
t
0
(st
0
)
p
= s
p1
st
0
(st
0
)
p
(s
/2
)
p
st
0
(st
0
)
p
for such ps. Hence, we obtain
(t
1
) > C
1
t
1/p
1
|1/|
p
for t
1
= st
0
and all p 2/, which contradicts (14).
Corollary 2. For every (0, 1), there is a function F such that

= , M() L
p
for all
p < , and M() , E .
Proof. We start with constructing a function c(t)
1
(t)
and searching the latter in the form
c(t) =
+

i=0
c
i

(2
i1
;2
i
]
(t), where 1 = c
0
c
1
c
n
. . . .
For the function (t) =
1
c(t)
to be equivalent to some function (t) F, it is sucient that the
conditions
c
i
c
i+1
2c
i
, i = 0, 1, . . . , and c
i
as i (15)
are satised.
Indeed, if (15) is valid then (t) increases and for k i we obtain c
k
2
ki
c
i
, whence
2
k
(2
k
) 2
i
(2
i
).
Now, if t
1
< t
2
then t
1
[2
k1
, 2
k
) and t
2
[2
i1
, 2
i
), k i. Moreover,
(t
2
)
t
2
2
i+1
(2
i
) 2
k+1
(2
k
) 2
k+2
(2
k1
) 4
(t
1
)
t
1
.
Thereby the function (t) is equivalent to some concave function [11, p. 69]. By (15), lim
t0+
(t) = 0.
Eventually, (t) (t) F.
For M() to be embedded into L
p
for all p < , it is necessary that
|1/|
p
< , 1 p < ,
or, equivalently,
|c|
p
=
_
+

i=0
c
p
i
2
i1
_1
p
< , 1 p < . (16)
804
Now, choose c
i

+
i=0
so that (15) and (16) hold and

= (0, 1). First of all, assume that

(0) n
0
= 1;
(1) n
2j+1
= 2n
2j
, j = 0, 1, . . . ;
(2) n
2j+2
= n
2j+1
+ j, j = 0, 1, . . . .
Now, dene a sequence c
i

+
i=0
as follows:
(a) c
0
= c
1
= 1;
(b) c
i
= c
n
2j
if n
2j
i n
2j+1
, j = 0, 1, . . . ;
(c) c
i
= c
n
2j+1
2
(in
2j+1
)
if n
2j+1
i n
2j+2
, j = 0, 1, . . . .
Relation (15) follows from the denition of c
i

+
i=0
. To check (16), we carry out some transformations:
+

i=2
c
p
i
2
i
=
+

k=1
n
2k+1
1

i=n
2k
2
p
k

j=1
(n
2j
n
2j1
)
2
i
+
+

k=1
n
2k+2
1

i=n
2k+1
2
p
k

j=1
(n
2j
n
2j1
)
2
p(in
2k+1
)
2
i
=
+

k=1
n
2k+1
1

i=n
2k
2
p
k(k1)
2
i
+
+

k=1
n
2k+2
1

i=n
2k+1
2
p(
k(k1)
2
+in
2k+1
)i
.
By (0)(2), n
2k+1
n
2k
n
2k1
2
k
for k > 0. Therefore,
I
1
:=
+

k=1
n
2k+1
1

i=n
2k
2
p
k(k1)
2
i
2
+

k=1
2
p
k(k1)
2
n
2k
2
+

k=1
2
p
k(k1)
2
2
k
<
and similarly
I
2
:=
+

k=1
n
2k+2
1

i=n
2k+1
2
p(
k(k1)
2
+in
2k+1
)i
2
+

k=1
2
p(
k(k1)
2
+k)2
k
< .
Eventually,
+

i=2
c
p
i
2
i
= I
1
+ I
2
< ;
therefore, (16) is valid. Straightforward calculation demonstrates that

= . Consequently, by Propo-
sition 1, M() , E .
The following implication played a key role in the proof of the embedding M() L
F
(see (4))
(where x = x(t) is an arbitrary measurable function on [0, 1]):
t
_
0
x

(s) ds
t
_
0

(s) ds for all t [0, 1] |x|

p
|

|
p
for all p 1.
As is easily seen, the reverse embedding L
F
M() holds if and only if the converse implication holds:
|x|
p
|

|
p
for all p 1
t
_
0
x

(s) ds C
t
_
0

(s) ds for all t [0, 1].

By Corollary 2, there is F such that M() L
p
for all p < , but M() , E ; i.e., the last is false
for such function . We will give an assertion improving this remark and simultaneously demonstrating
the diculties we face in interconnecting the interpolation and extrapolation constructions.
805
Proposition 2. There is no universal constant C > 0 such that the inequality
|x|
p
|y|
p
, p 1,
valid for two functions x, y L

, would imply that

t
_
0
x

(s) ds C
t
_
0
y

(s) ds, 0 < t 1. (17)

Proof. Take n N, n 3, =
1
2(n1)
n
, and = (2)
1
n
=
1
n1
. Consider the functions
x(t) =
_
1
1
n
t
_

[0,n)
(t) and y(t) =
[0,)
(t) +
[,1)
(t).
Then |y|
p
p
= +
p
(1 ), |x|
p
p
=
n
p+1
, and
|x|
p
p
|y|
p
p
(18)
for all p 1. Indeed, we can rewrite the last inequality as follows:
n
p + 1
+
p
(1 ), p 1.
If n p + 1 then the inequality is obvious. Assume that 1 p n 1. It suces to show that
n
p + 1

1
2

p
, 1 p n 1. (19)
Consider the function (p) = (p+1)
p
. Since

(p) =
p
(1+(p+1) log ), (p) is decreasing if 1/

e.
Therefore, it suces to verify (19) for p = n 1 which is done by simple substitution: = (2)
1
n
.
Thereby (18) is valid. Simultaneously, if
K (t, x; L
1
, L

) =
t
_
0
x

(s) ds (0 t 1)
then
K (n, x; L
1
, L

)
K (n, y; L
1
, L

)
=
n/2
+ (n 1)
=
n
2 + 2(n 1)
=
n
4
+ as n +,
and so (17) is false.
3. Extrapolatory Description for Lorentz Spaces
S. F. Lukomskii [10] proved the following: If there is > 0 such that the inequality

(t) t

(t
2
) (20)
holds for almost all t [0, 1] then the Lorentz space
p
() is an extrapolation space; more exactly,
|x|
p()

_
+

n=1
|x|
p
n
2
n

(2
n
)
_1
p
. (21)
In this section we nd some necessary and sucient conditions for (21) to be valid and also give
an example of F for which (20) is false whereas (21) holds.
Denote by g
p
the space l
p
(2
n/p

(2
n
)
1/p
) in which the norm coincides with the right-hand side
of (21). Its function-space analog is the space G
p
= L
p
(2
s/p

(2
s
)
1/p
) of functions on [1, ).
806
Theorem 3. The following are equivalent for each p [1, ):
(1)
2
;
(2)
p
() = L
Gp
;
(3)
p
() = L
gp
.
To prove the theorem, we need Lemma 3 below in which the role of the extrapolation parameter is
played by the space

G
p
of functions on [p, ) with the norm
|y|

Gp
=
_
+
_
p
[y(s)[
p
2
s

(2
s
) ds
_1
p
.
Lemma 3. For every p [1, ), the rearrangement invariant space L

Gp
is p-convex.
Proof. We need to prove that
_
_
_
_
_
_
n

k=1
[x
k
[
p
_1
p
_
_
_
_
_
L
Gp
C
_
n

k=1
|x
k
|
p
L
Gp
_1
p
(22)
for arbitrary n N and x
1
, x
2
, . . . , x
n
from L

Gp
. Since, for s p, the space L
s
[0; 1] is p-convex with
constant 1 [12, Proposition 1.d.5], we have
_
_
_
_
_
_
n

k=1
[x
k
[
p
_1
p
_
_
_
_
_
s

_
n

k=1
|x
k
|
p
s
_1
p
.
Thereby
_
_
_
_
_
_
n

k=1
[x
k
[
p
_1
p
_
_
_
_
_
L
Gp

_
+
_
p
_
n

k=1
|x
k
|
p
s
_
2
s

(2
s
) ds
_
1
p
=
_
n

k=1
|x
k
|
p
L
Gp
_1
p
,
and (22) is proved.
Proof of Theorem 3. First of all, note that L

Gp
in Lemma 3 coincides with L
Gp
. Indeed, the
embedding L
Gp
L

Gp
is obvious. On the other hand,
_
+
_
1
|x|
p
s
2
s

(2
s
) ds
_1
p

_
p
_
1
|x|
p
s
2
s

(2
s
) ds
_1
p
+
_
+
_
p
|x|
p
s
2
s

(2
s
) ds
_1
p
2
_
2p
_
p+1
|x|
p
s
2
s

(2
s
) ds
_1
p
+
_
+
_
p
|x|
p
s
2
s

(2
s
) ds
_1
p
3 |x|
L
G
,
whence we nd that L
Gp
L

Gp
. Therefore, in the proof below, instead of L
Gp
we can consider L

Gp
.
It is easy to show that
L

Gp

p
() (23)
807
for every F. Indeed, since x

(2
s
) 2|x|
s
(s 1), we have
|x|
p
p()
2
p
1/2
p
_
0
x

(t)
p
d(t) = 2
p
log 2
+
_
p
x

(2
s
)
p
2
s

(2
s
) ds 2
p+1
log 2|x|
p
L
Gp
.
As is well known [12, p. 54], if X is a p-convex Banach function space then we can consider the
Banach space X
(p)
of all functions x = x(t) such that [x(t)[
1
p
X with the norm |x|
X
(p)
= |[x[
1
p
|
p
X
. By
Lemma 3, this procedure applies to L

Gp
; moreover, (
p
())
(p)
= (). Thereby, by (23),
(L

Gp
)
(p)
(). (24)
The Lorentz space () is the smallest space among all rearrangement invariant spaces with the same
fundamental function [11, p. 160]. Therefore, the reverse embedding to (24), i.e., () (L

Gp
)
(p)
or,
equivalently,
p
() L

Gp
, is valid if and only if

p
(t) C(t), 0 < t 1, (25)
where
p
(t) is the fundamental function of (L

Gp
)
(p)
. Using the denition of (L

Gp
)
(p)
, we infer the
inequality
+
_
p
t
p
s
2
s

(2
s
) ds C(t), 0 < t 1.
Since both sides represent quasiconcave functions, it suces to validate this relation for t = 2
j
, j 2p.
After substitution we obtain
p
_

2
jp
s
d(2
s
) C(2
j
).
Since
j
_

2
jp
s
d(2
s
) (2
j
),
the previous inequality is equivalent to the inequality
p
_
j
2
jp
s
d(2
s
) C(2
j
). (26)
Prove that (26) holds if and only if
2
.
Assume (26) valid. Then, on the one hand,
j/2
_
j
2
jp
s
d(2
s
) C(2
j
)
for j 2p and, on the other hand,
j/2
_
j
2
jp
s
d(2
s
) 2
2p
((2
j/2
) (2
j
)).
808
Hence,
(2
j/2
) (2
2p
C + 1)(2
j
), j 2p;
now, by the standard arguments involving the quasiconcavity of , we arrive at
2
.
Suppose, vice versa, that
2
. First of all, note that in this case it suces to prove (25) only
for the points t = 2
j
, where j = 2
m
, m N. Indeed, assume that (25) is valid for such t and take
j [2
m
, 2
m+1
]. Then

p
(2
j
)
p
(2
2
m
) C(2
2
m
) C(2
2
m+1
) C(2
j
),
where is the constant from the
2
-condition. Thus, in the proof of (26) we may assume that j = 2
m
,
m N, and j p. In this case we obtain
p
_
j
2
jp
s
d(2
s
)
1
_
j
2
jp
s
d(2
s
)
1
_
j
2
j
s
d(2
s
) =
m

k=1
2
k1
_
2
k
2
j
s
d(2
s
)

k=1
2
2
mk
(2
2
k1
)
m

k=1
2
2
mk

mk+1
(2
2
m
)
= (2
j
)
m1

l=0
2
2
l
+l log
2

< (2
j
)
+

l=0
2
2
l
+l log
2

= C(2
j
),
where C =

+
l=0
2
2
l
+l log
2

< , and (26) is valid. Consequently,
p
() L

Gp
.
Thus, (1) and (2) are equivalent. To complete proving, show that L
Gp
= L
gp
. Indeed, for s [n, n+1]
we nd
1
2
|x|
p
n
2
n

(2
n
) |x|
p
s
2
s

(2
s
) 2|x|
p
n+1
2
n1

(2
n1
)
and so
|x|
p
Lgp
=
+

n=1
|x|
p
n
2
n

(2
n
) 2
+

n=1
n+1
_
n
|x|
p
s
2
s

(2
s
) ds
= 2
+
_
1
|x|
p
s
2
s

(2
s
) ds = 2|x|
p
L
Gp
4
+

n=1
|x|
p
n+1
2
n1

(2
n1
) 4|x|
p
Lgp
.
Therefore, the norms in L
Gp
and L
gp
are equivalent.
Note that if F satises (20) then
2
(it suces to integrate (20)). We will give an example
demonstrating that the converse is false; i.e., there is a function F belonging to
2
but violating (20).
Example 2. We denote by F
1
the class of piecewise linear functions in F of the form
(2
k
) =

mk

m
and (t) is linear on [2
k1
, 2
k
], k Z
+
.
For a such function to belong to F, it is necessary and sucient that
(1)
k
> 0 ((t) is increasing);
(2)

k0

k
< ((t) is nite);
(3)

k

k+1
2 ((t) is concave).
809
By the properties of the functions in F, it suces to verify the
2
-condition only at the points of
the form t
k
= 2
k
. For F
1
, this condition is equivalent to the requirement

kn

k
C

k2n

k
, n Z
+
. (27)
Condition (20) for F
1
takes the form:

k

1
min(
2k+1
,
2k
), k Z
+
.
Therefore, does not satisfy (20) if
lim
k+

2k
= +. (28)
Thus, it suces to construct an example of a function F
1
satisfying (27) and (28).
Assume that n
0
4 is a full square natural number satisfying
(n +

n)
2
2

n
n
3
2
(29)
for n n
0
. Introduce the two sequences of natural numbers:
n
i
= n
2
i1
and m
i
= min
_
k n
i
: k
2
2
kn
i
n
3
2
i
_
, i N. (30)
By (29) and (30), m
i
n
i
+

n
i
and the intervals [n
i
, m
i
] (i = 0, 1, . . . ) are pairwise disjoint. Moreover,
none of the intervals of the form [n, 2n] (n N) can intersect more than one interval of the form [n
i
, m
i
].
Indeed, since n
i
4, we have n
2
i
> 2(n
i
+

n
i
). Consequently, if i is the least number such that
[n
i
, m
i
] [n, 2n] ,= then m
i
n and
n
i+1
m
i
n
2
i
n
i

n
i
> n
i
+

n
i
m
i
n,
whence n
i+1
> m
i
+ n 2n. Therefore, [n
i+1
, m
i+1
] [n, 2n] = .
Consider (t) such that

k
=

1
10
if k = 0, 1, 2;
1
k
2
if k 3 and k , [n
i
, m
i
);
2
n
i
k 1
n
3/2
i
if k [n
i
, m
i
).
We rst verify the conditions (1)(3) for (t) to belong to F
1
. Condition (1) is obvious. Since

k0

k

3
10
+

k=1
1
k
2
+

i=0
m
i
1

k=n
i
2
n
i
k
1
n
3/2
i

3
10
+

2
6
+ 2

i=0
1
n
3/2
i
3 + 2

k=1
1
n
k
0
< ,
we see that (2) holds as well.
We are left with verifying (3). If k = 0, 1, 2 then the condition is obvious. If k, k + 1 , [n
i
, m
i
) then

k+1
=
(k + 1)
2
k
2
2
for k 3. If k, k + 1 [n
i
, m
i
) then

k

k+1
= 2 by the denition of
k
. In the case k = n
i
1 for some
i = 0, 1, . . . we have

k+1
=
n
3/2
i
(n
i
1)
2

n
2
i
(n
i
1)
2
2,
810
since n
i
4. We are left with considering the last situation in which k = m
i
1. Using the denition
of m
i
in (30), we obtain

k+1
=
2
n
i
(m
i
1)
m
2
i
n
3/2
i

m
2
i
(m
i
1)
2
2,
since m
i
4.
Thus, F
1
. To check (28), it suces to assert that
lim
i+

n
i

2n
i
= .
Since n
i+1
2n
i
= n
2
i
2n
i
> 0 and m
i
n
i
+

n
i
< 2n
i
, we have 2n
i
(m
i
, n
i+1
) and therefore

2n
i
=
1
(2n
i
)
2
. Consequently,

n
i

2n
i
=
(2n
i
)
2
n
3/2
i
= 4

n
i
+,
and thereby (28) is valid.
Checking (27), we can certainly assume that n 4. By (30),
1
k
2
<
2
n
i
k
n
3/2
i
=
k
for k [n
i
, m
i
). Therefore, we can estimate the sum on the right-hand side of (27) from below as follows:

k2n

k2n
1
k
2

1
2n
, n = 1, 2, . . . . (31)
Represent the sum on the left-hand side of (27) as

kn

k
=
2n1

k=n

k
+

k2n

k
. (32)
Assume that there is i satisfying [n
i
, m
i
] [n, 2n 1] ,= (as observed above, this index i is unique).
Then the rst sum on the right-hand side in (32) is estimated as follows:
2n1

k=n

k

2n1

k=n
1
k
2
+
m
i
1

k=n
i

k

2
n
+
2
n
i
.
Since n < m
i
n
i
+

n
i
< 2n
i
, we have n
i
>
n
2
; therefore, by (31),
2n1

k=n

k

6
n
12

k2n

k
.
Eventually, using (32), we nd that

kn

k
13

k2n

k
;
i.e., (27) is valid.
Remark 1. The above example demonstrates that there exists belonging to
2
but violating (20).
The results of [10] do not apply to such functions but Theorem 3 of the present article does. At the
same time, we should observe that the constructed function (t) is equivalent to the function log
1
(10/t)
which, as is easily veried, satises (20). The point is that (20) (unlike the condition
2
) is not
invariant under passage to equivalent functions.
Remark 2. In Theorem 3 we gave necessary and sucient conditions for
p
() to coincide with the
extrapolation space L
Gp
, where the parameter G
p
has a special form. The authors are unaware whether
there are nevertheless F, ,
2
, such that
p
() = L
H
for some ideal Banach space H.
811
References
1. Yano S., An extrapolation theorem, J. Math. Soc. Japan, 3, No. 1, 296305 (1951).
2. Zygmund A., Trigonometric Series. Vol. 2 [Russian translation], Mir, Moscow (1965).
3. Jawerth B. and Milman M., Extrapolation theory with applications, Mem. Amer. Math. Soc., 89, No. 440, 382
(1991).
4. Jawerth B. and Milman M., New results and applications of extrapolation theory, in: Interpolation Spaces and Related
Topics (Haifa, 1990), Israel Math. Conference Proc., 1992, 5, pp. 81105.
5. Milman M., Extrapolation and Optimal Decompositions: with Applications to Analysis, Springer-Verlag, Berlin (1994)
(Lecture Notes in Math.; V. 1580).
6. Astashkin S. V., On extrapolation properties of the scale of Lp-spaces, Mat. Sb., 194, No. 6, 2342 (2003).
7. Astashkin S. V., Extrapolation functors on a family of scales generated by the real interpolation method, Siberian
Math. J., 26, No. 2, 264289 (2005).
8. Milman M., Extrapolation spaces and a.e. convergence of Fourier series, J. Approx. Theory, 80, No. 1, 1024 (1995).
9. Lukomskii S. F., On the convergence of Walsh series in spaces close to L

, Math. Notes, 70, No. 6, 804811 (2001).

10. Lukomskii S. F., Convergence of Fourier series in Lorentz spaces, East J. Approx., 9, No. 2, 229238 (2003).
11. Kren S. G., Petunin Yu. I., and Semenov E. M., Interpolation of Linear Operators [in Russian], Nauka, Moscow (1978).
12. Lindenstrauss J. and Tzafriri L., Classical Banach Spaces. Vol. 2: Function Spaces, Springer-Verlag, Berlin (1979).
13. Bennett C. and Sharpley R., Interpolation of Operators, Academic Press, Boston (1988).
S. V. Astashkin; K. V. Lykov
Samara State University, Samara, Russia
E-mail address: astashkn@ssu.samara.ru; lykich@list.ru
812