Professional Documents
Culture Documents
NARCISSE RANDRIANANTOANINA
arXiv:2108.06341v1 [math.OA] 13 Aug 2021
Abstract. Let M be a semifinite von Nemann algebra equipped with an increasing filtration
(Mn )n≥1 of (semifinite) von Neumann subalgebras of M. For 0 < p < ∞, let hcp (M) denote the
noncommutative column conditioned martingale Hardy space and bmoc (M) denote the column
“little” martingale BMO space associated with the filtration (Mn )n≥1 .
We prove the following real interpolation identity: if 0 < p < ∞ and 0 < θ < 1, then for
1/r = (1 − θ)/p,
hcp (M), bmoc (M) θ,r = hcr (M),
1. Introduction
Hardy space theory takes on many forms and appears in many aspects of mathematics such as
harmonic analysis, PDE’s, functional analysis, probability theory, and many others. Interpolation
spaces between Hardy spaces in various contexts have a long history. We refer to the articles
[13, 23, 35] for some background on interpolations between classical Hardy spaces from harmonic
analysis and [18, 47] for interpolations between Hardy spaces from martingale theory. On the
other hand, the theory of noncommutative martingales has seen rapid development in many
directions. Indeed, since the establishment of the noncommutative Burkholder-Gundy inequalities
in [37], many classical inequalities are now understood for this context. We refer to the book
[36, Chap. 14] for a summary of some of the main inequalities from noncommutative martingale
theory. Further references relevant to our purpose are [1, 7, 24, 26, 32, 44]. The main focus of
the present article is on Hardy spaces arising from noncommutative martingale theory. More
specifically, column/row Hardy spaces defined from column/row conditioned square functions
initiated by Junge and Xu in [26] in connection with the noncommutative Burkholder/Rosenthal
inequalities. These spaces are generally referred to as conditioned Hardy spaces and its column
(resp. row) version is usually denoted by hcp (resp. hrp ). We would like to emphasize that
this particular class of martingale Hardy spaces is instrumental in classical theory. We refer to
the monograph [47] for more in depth treatment of the classical setting. Likewise, the Hardy
spaces hcp and hrp are proven to be fundamental objects in various aspects of the new progress
made in the noncommutative martingale theory during the last several years. For instance, the
formulation of the noncommutative Burkholder inequality for the case 1 < p < 2 was mainly due
to a reformulation of the case 2 ≤ p < ∞ as equivalence of norms involving the spaces hcp and
hrp . The asymmetric Doob maximal inequalities in [17] were derived from properties of hcp and
hrp where 1 ≤ p < 2. The general theme of the present article is on interpolation spaces for the
quasi-Banach space couple (hcp , hcq ) when 0 < p < q < ∞.
To motivate our consideration, let us review some interpolation results from classical martingale
theory. Let (Ω, F, P) be a probabilityS space and (Fn )n≥1 be an increasing sequence of σ-subfields
of F satisfying the condition F = σ( n≥1 Fn ). For 0 < p ≤ ∞, denote by Hp (Ω) (resp. hp (Ω))
the martingale Hardy space defined by square functions (resp. conditioned square functions) and
BMO(Ω) (resp. bmo(Ω)) the martingale BMO (resp. martingale little BMO) space associated
with the filtration (Fn )n≥1 . We refer to [14, 47] for precise definitions and properties of these
spaces along with discussions on their importance for the classical theory. It is well established
in the literature that the spaces BMO(Ω) and bmo(Ω) play important role in interpolation
theory as they may be used as natural substitutes for L∞ (Ω). Our motivation comes from the
following three classical results that involved martingale BMO spaces as one of the endpoints of
interpolations:
1
(1.1) H1 (Ω), BMO(Ω) θ = Hr (Ω), 0 < θ < 1 and = 1 − θ;
r
1 1−θ
(1.2) Hp (Ω), BMO(Ω) θ ( Hr (Ω), 0 < p < 1, 0 < θ < 1, and = < 1;
r p
1 1−θ
(1.3) hp (Ω), bmo(Ω) θ,r = hr (Ω), 0 < p < ∞, 0 < θ < 1, and = ,
r p
where [·, ·]θ (resp. (·, ·)θ,r ) denotes the complex (resp. real) interpolation method. The identity
(1.1) and the inclusion (1.2) were obtained by Janson and Jones in [18] while (1.3) was established
in its present form by Weisz in [46]. It is a natural question to consider if the three assertions
stated above still hold for the noncommutative setting. The first result in this direction is due
to Musat in [32] who proved a noncommutative analogue of (1.1) where as introduced in [37],
the noncommutative Hardy space H1 is the sum of the column version and the row version and
the BMO space is defined as the intersection of column BMO and the row BMO. Later, Bekjan
et al. established in [1] that the noncommutative analogue of (1.3) holds for the Banach space
range. That is, (1.3) remains valid for the interpolation couple (hc1 , bmoc ). To the best of our
knowledge, [1] and [32] are the only articles available in the literature that contain substantial
advances in the study of interpolation spaces of noncommutative martingale Hardy spaces to date.
The present paper extends the interpolation results from [1] to the case 0 < p < 1 thus providing
a full noncommutative generalization of (1.3) for column/row spaces. We refer to Theorem 3.5
below for detailed formulation. We also obtained interpolation spaces between spaces of adapted
sequences which may be viewed as the right substitute for (1.1) when 0 < p < 1. This is stated in
Theorem 3.13. We should point out that the result on the spaces of adapted sequences appears
to be new even for the classical setting. Moreover, the noncommutative analogue of (1.2) can be
easily deduced from the result on adapted sequences.
Our method of proof is very different from [1, 32]. In fact, the techniques used in both [1] and
[32] appear to work only for Banach couples. Our main objective is to compare K-functionals
for the couple (hcp , hcq ) where 0 < p < q < ∞ to those associated to the well-known couple
INTERPOLATION 3
(Lp , Lq ) associated with appropriate amplified von Neumann algebras. In the classical setting,
this type of reduction is usually achieved through some strategic use of stopped martingales. We
note that a the time of this writing there is no direct analogue of stopping times available for the
noncommutative setting but the so-called Cuculescu projections ([8]) are often used as a substitute
for stopping times. However, Cuculescu projections do not appear to be efficient enough to provide
the desired truncations. This makes our approach very different from the classical setting as found
in [18, 46]. Our method is based on the so-called algebraic atomic decompositions. The notion
of algebraic atoms were introduced by Perrin in [33] for Junge’s noncommutative conditioned
Lp spaces and the Hardy spaces hcp for 1 ≤ p < 2 and found to be instrumental in the study
of Doob’s maximal inequality for martingale in noncommutative Hardy spaces ([17]). Algebraic
atomic decompositions for the Hardy space hcp with 0 < p < 1 was recently studied in [7] by
constructive approach. Another recent development in this direction is that the noncommutative
Orlicz-Hardy space hcΦ admits algebraic atomic decomposition for convex function Φ satisfying
some natural conditions. Using insights from the constructive approach used in [7] and [44],
we established that noncommutative conditioned Orlicz-Hardy spaces admit version of algebraic
atomic decompositions when the Orlicz function is p-convex and q-concave for 0 < p < q < 2 (see
Theorem 2.8 for details). These more general atomic decompositions turn out to be one of the
decisive tools we used for our results on K-functionals for the couple (hcp , hcq ) when the distance
between p and q is sufficiently small. The results on K-functionals coupled with the well-known
Wolff’s interpolation theorem provide the full noncommutative generalization of (1.3). The case
of spaces of adapted sequences is handle with the same techniques. That is, estimating the
K-functionals through some version of algebraic atomic decompositions.
The paper is organized as follows. In the next section, we review the basics of noncommutative
symmetric quasi-Banach spaces following the formulation of [28, 50] and give full detailed accounts
of the construction of noncommutative martingale conditioned Hardy spaces associated with
symmetric quasi-Banach spaces. We also formulate and prove the algebraic decomposition for
noncommutative conditioned spaces associated with Orlicz function spaces and for spaces of
adapted sequences.
Section 3 contains formulations and proofs of our principal results: noncommutative gener-
alization of (1.2), noncommutative generalization of (1.3), and an extension of (1.1) to the full
range when spaces of adapted sequences are used.
In Section 4, we explore some further applications of our results and methods from Section 2
and Section 3 to various inequalities involving noncommutative martingale Hardy spaces in the
context of general symmetric spaces of measurable operators.
on (0, ∞) which are bounded except on a set of finite measure and for f ∈ M, f µ(f ) is precisely
the usual decreasing rearrangement of the function |f |. We refer to [38] for more information on
noncommutative integration.
We denote by L0 , the space of measurable functions on the interval (0, ∞). Recall that a
quasi-Banach function space (E, k · kE ) of measurable functions on the interval (0, ∞) is called
symmetric if for any g ∈ E and any f ∈ L0 with µ(f ) ≤ µ(g), we have f ∈ E and kf kE ≤ kgkE .
Throughout, all function spaces are assumed to be defined on the interval (0, ∞).
Let E be a symmetric quasi-Banach function space. We define the corresponding noncommu-
tative space by setting: n o
E(M, τ ) = x ∈ M f : µ(x) ∈ E .
Equipped with the quasi-norm kxkE(M,τ ) := kµ(x)kE , the linear space E(M, τ ) becomes a com-
plex quasi-Banach space ([28, 50]) and is usually referred to as the noncommutative symmetric
space associated with (M, τ ) corresponding to E. We remark that if 0 < p < ∞ and E = Lp ,
then E(M, τ ) is exactly the usual noncommutative Lp -space Lp (M, τ ) associated with (M, τ ).
In the sequel, E(M, τ ) will be abbreviated to E(M).
Other classes of examples that are relevant for our purpose are the class of Orlicz spaces and
the class of Lorentz spaces. We review these two classes for convenience.
A function Φ : [0, ∞) → [0, ∞) is called an Orlicz function whenever it is strictly increasing,
continuous, Φ(0) = 0, and limu→∞ Φ(u) = ∞. The Orlicz space LΦ is the collection of all f ∈ L0
for which there exists a constant c such that IΦ (|f |/c) < ∞ where the modular functional IΦ (·)
is defined by: Z ∞
IΦ (|g|) = Φ(|g(t)|) dt, g ∈ L0 .
0
The space LΦ is equipped with the Luxemburg quasi-norm:
f
= inf c > 0 : IΦ (|f |/c) ≤ 1 .
L Φ
Under the above quasi-norm, one can easily see that E(M; ℓc2 ) is a quasi-Banach space. The closed
subspace of E(M; ℓc2 ) consisting of adapted sequences will be denoted by E ad (M; ℓc2 ). That is,
n o
E ad (M; ℓc2 ) = (an )n≥1 ∈ E(M; ℓc2 ) : ∀n ≥ 1, an ∈ E(Mn ) .
Note that for 1 < p < ∞, it follows from the noncommutative Stein inequality that Lad c
p (M; ℓ2 )
c
is a complemented subspace of Lp (M; ℓ2 ). One should not expect such complementation if one
merely assume that E is quasi-Banach symmetric function space.
Next, we will discuss conditioned versions of the spaces defined earlier. Consider the linear
space FS consisting of all x ∈ M such that there exists a projection e ∈ M1 , τ (e) < ∞, and
x = exe. We should note that if M is finite, then FS = M. For every n ≥ 1 and 0 < p ≤ ∞, we
define the space Lcp (M, En ) to be the completion of FS with respect to the quasi-norm:
1/2
x
c =
En (x∗ x)
.
Lp (M,En ) p/2
We would like to emphasize here that if x = exe ∈ FS is as described above, then En (x∗ x) =
eEn (x∗ x)e is a well-defined operator in M and since τ (e) < ∞, it follows that En (x∗ x) ∈ Lp/2 (M).
According to [24], for every 0 < p ≤ ∞, there exists an isometric right Mn -module map
un,p : Lcp (M, En ) → Lp (Mn ; ℓc2 ) such that
(2.1) un,p (x)∗ un,q (y) = En (x∗ y) ⊗ e1,1 ,
whenever x ∈ Lcp (M; En ), y ∈ Lcq (M; En ), and 1/p + 1/q ≤ 1. An important fact about these
maps is that they are independent of p as the index p in the presentation of [24] was only needed
to accommodate the non-tracial case. Below, we will simply use un for un,p .
For 0 < p ≤ ∞, the range of un is complemented in Lp (Mn , ℓc2 ). In fact, as proved in [24,
Proposition 2.8(iii)], there exists a contractive projection Qn from Lp (Mn ; ℓc2 ) onto the range of
un such that for every ξ ∈ Lp (Mn ; ℓc2 ),
Qn (ξ)∗ Qn (ξ) ≤ ξ ∗ ξ.
This fact will be used in the sequel.
Let F be the collection of all finite sequences a = (an )n≥1 in FS. For 0 < p < ∞, we defined
the conditioned space Lcond
p (M; ℓc2 ) to be the completion of the linear space F with respect to the
quasi-norm:
(2.2)
a
cond c =
σc (a)
Lp (M;ℓ2 ) p
Now, we generalize the notion of conditioned spaces to the setting of symmetric spaces of
operators. This is done in steps.
• Assume first that E is a symmetric quasi-Banach function space satisfying Lp ∩ L∞ ⊆ E ⊆
Lp + L∞ for some 0 < p < ∞ and Lp ∩ L∞ is dense in E. This is the case for instance when E is
a separable fully symmetric quasi-Banach function space. For a given sequence a = (an )n≥1 ∈ F,
we set:
(an )
cond =
σc (a)
E(M) =
U ((an ))
E(M⊗B(ℓ2 (N2 ))) .
E (M;ℓc )
2
This is well-defined and induces a quasi-norm on the linear space F. We define the quasi-Banach
space E cond (M; ℓc2 ) to be the completion of the quasi normed space (F, k · kE cond (M;ℓc2 ) ). The space
E cond (M; ℓc2 ) will be called the column conditioned space associated with E. It is clear that U
extends to an isometry from E cond (M; ℓc2 ) into E(M⊗B(ℓ2 (N2 ))) which we will still denote by
U.
Below, we use the notation E 0,cond (M; ℓc2 ) for the closure of the linear space F0 = {a = (an ) ∈
F : a1 = 0} in E cond (M; ℓc2 ). One can easily see that for a = (an )n≥1 ∈ F, we have
n
o
max
E1 (|a1 |2 )1/2
E(M1 ) ,
(an )n≥2
E cond (M;ℓc ) ≤
(an )n≥1
E cond (M;ℓc ) .
2 2
It is clear that E cond (M; ℓc2 ) as defined is a linear quasi-normed space. We claim that it is
complete. Indeed, if (xν )ν≥1 is a Cauchy sequence in E cond (M; ℓc2 ), then it converges to some
x ∈ (Lp + Lq )cond (M; ℓc2 ). Since (U (xν ))ν≥1 is also a Cauchy sequence in the quasi-Banach
space E(M⊗B(ℓ2 (N2 ))) and E(M⊗B(ℓ2 (N2 ))) ⊆ (Lp + Lq )(M⊗B(ℓ2 (N2 ))), it follows that x ∈
E cond (M; ℓc2 ) and (xν )ν≥1 converges to x in E cond (M; ℓc2 ). We should note here that if Lp ∩ L∞
is dense in E then the above definition coincides with the one described in the previous bullet.
We also define
E 0,cond (M; ℓc2 ) = (Lp + Lq )0,cond (M; ℓc2 ) ∩ E cond (M; ℓc2 ).
The direct sum stated in (2.4) still applies.
Remark 2.2. At the time of this writing, we do not know of any suitable definition for conditioned
space associated with Lp + L∞ when 0 < p < 2. It is also unclear if our definition of E cond (M; ℓc2 )
for non separable space E ⊂ Lp + Lq when 0 < p < 2 is independent of the isometry U .
We now recall the construction of column conditioned martingale Hardy spaces. As in the
conditioned spaces, we describe the noncommutative conditioned Hardy spaces in steps. Let
F(M ) be the collection of all finite martingale (xn )1≤n≤N for which xN ∈ FS. We can easily see
that for every 0 < p ≤ ∞, F(M ) ⊆ hcp (M). As in the case of conditioned spaces, F(M ) is dense
in hcp (M) when 0 < p < ∞.
• First, assume that E ⊆ L2 + L∞ . In this case, column conditioned square functions are well-
defined for bounded martingales in E(M). We define hcE (M) to be the collection of all bounded
martingale x in E(M) for which sc (x) ∈ E(M). We equip hcE (M) with the norm:
x
c =
sc (x)
.
hE E(M)
8 RANDRIANANTOANINA
The column conditioned Hardy space hcE (M) is the completion of (F(M ), k·khcE ). Clearly, the map
x 7→ (dxn ) (from F(M ) into F) extends to an isometry from hcE (M) into E cond (M; ℓc2 ) which we
denote by Dc . In particular, hcE (M) is isometrically isomorphic to a subspace of E(M⊗B(ℓ2 (N2 )))
via the isometry U Dc . We should note here that if L2 ∩ L∞ is dense in E and E ⊆ L2 + L∞ , then
the two definitions provide the same space.
• Assume now that E ⊂ Lq + Lq for 0 < p, q < ∞. As in the case of conditioned spaces, we set
n o
hcE (M) = x ∈ hcLp +Lq (M) : U Dc (x) ∈ E(M⊗B(ℓ2 (N2 )))
Since the operator U is independent of the index, one can easily see that the space hcE (M) is
independent of p and q. Moreover, one can verify as in the case of conditioned spaces that the
quasi-normed space (hcE (M), k · khcE ) is complete. Furthermore, if E is such that L2 ∩ L∞ is dense
in E then hcE (M) coincides with the one defined through completion considered in the second
bullet.
The following fact will be used in the sequel.
Lemma 2.3. Let 0 < p < q < ∞ and assume that E ⊆ Lp + Lq . The Hardy space hcE (M) is 1-
complemented in E cond (M; ℓc2 ). More precisely, there is an onto map Π : E cond (M; ℓc2 ) → hcE (M)
with kΠk = 1 and ΠDc is the identity map in hcE (M).
Proof. Let n ≥ 1 and b ∈ FS. Using the Kadison-Schwarz inequality for conditional expectations,
we have
(2.5) En−1 (|En (b) − En−1 (b)|2 ) ≤ En−1 (|b|2 ).
We define Π : F → hcE (M) by setting:
X
a = (an )n≥1 7→ En (an ) − En−1 (an ).
n≥1
It follows from (2.5) that sc (Π(a)) ≤ σc (a) for every a ∈ F. If F is dense in E cond (M; ℓc2 ), then
Π extends to a bounded linear map from E cond (M; ℓc2 ) onto hcE (M) with kΠ : E cond (M; ℓc2 ) →
hcE (M)k ≤ 1. Clearly, if x ∈ hcE (M), we have ΠDc (x) = x. This verifies the lemma for the case
where E is separable.
For the general case, we observe that from the fact that F is dense in (Lp + Lq )cond (M; ℓc2 ), the
inequality on conditioned square functions above can be restated as:
(2.6) U Dc (Π(y)) ≤ U (y), y ∈ (Lp + Lq )cond (M; ℓc2 ).
If y ∈ E cond (M; ℓc2 ), then by definition y ∈ (Lp +Lq )cond (M; ℓc2 ) and U (y) ∈ E(M⊗B(ℓ2 (N2 ))).
By the separable case, Π(y) ∈ hcLp +Lq (M). Moreover, it follows from (2.6) that U D(Π(y)) ∈
E(M⊗B(ℓ2 (N2 ))). This means, Π(y) ∈ hcE (M) with kΠ(y)khcE ≤ kykE cond (M;ℓc2 ) .
INTERPOLATION 9
We refer to [6, 11, 19, 22, 26, 42, 43] for more information on noncommutative Hardy spaces
associated with symmetric spaces of measurable operators. In the sequel, noncommutative column
Hardy spaces associated with Orlicz space LΦ will be denoted by HΦ c (M) and hc (M) while those
Φ
associated with the Lorentz space Lp,q will be denoted by Hp,q (M) and hcp,q (M).
c
We conclude this subsection with a description of the dual space of the Hardy scpace hc1 (M).
A martingale x belongs to the column little bmo space denoted by bmoc (M) if
n
o
x
c = max
x1
; sup sup
En (|xm − xn |2 )
1/2 .
bmo ∞ ∞
m 1≤n≤m
c
The Banach space (bmo (M), k · kbmoc ) was introduced in [34] for finite case where it was shown
that it coincides with the dual of the Hardy space hc1 (M). The proof given there can be easily
generalized to the semifinite case. We record this for further use:
(2.7) (hc1 (M))∗ = bmoc (M)
with equivalent norms.
We note that the Orlicz function Θ can be taken to be the inverse of the function t 7→
t−1/2 Φ−1 (t)for t > 0. We refer to [31] for more details.
We will also make use of the following function:
Z ∞
(2.10) Ψ(t) = (t2−p s−p + t2−q s−q )−1 dµ(s)
0
where µ is the positive Borel measure from the representation of Φ in (2.8). The reason for the
consideration of Ψ is summarized in the next lemma. The first three items are straightforward
generalizations of [44, Proposition 3.3] while the last item can be deduced as in [44, Lemma 3.4].
Lemma 2.4. (i) Ψ(t) ≈p,q t−2 Φ(t);
(ii) Θ Ψ(t)−1/2 ≈p,q Φ(t);
(iii) t 7→ Ψ(t1/2 ) is operator monotone decreasing;
10 RANDRIANANTOANINA
We now introduce a concept of atoms for conditioned space constructed from the Orlicz function
space LΦ .
Definition 2.5. A sequence x ∈ LΦ (M; ℓc2 ) is called an algebraic Lc,cond
Φ -atom if it admits a
factorization x = α . β where
P
(i) α = j<n αn,j ⊗ en,j is a strictly lower triangular matrix in L2 (M⊗B(ℓ2 (N))) with
X 1/2
α
= kαn,j k22 ≤ 1;
2
j<n
(ii) β ∈ Lad c
≤ 1.
Θ (M; ℓ2 ) with β L c
Θ (M;ℓ2 )
The above definition was motivated by the case of Lcond r (M; ℓc2 ) for 1 ≤ r < 2 introduced for
the first time in [33] and explored further in [17]. We also refer to [7] where the same concept was
considered for the case of conditioned Hardy space hcr (M) for the range 0 < r ≤ 1. More recently,
the case of Orlicz-conditioned Hardy space hcϕ (M) associated with convex Orlicz function ϕ was
formulated in [44] in the context of ϕ-moments.
We should note that since strictly lower triangular matrices were used in the definition of
algebraic atoms, it follows that if x = (xn )n≥1 is an algebraic Lc,cond
Φ -atom then x1 = 0.
The next lemma can be deduced as in the first part of [33, Theorem 3.6.10] using the factor-
ization LΦ = L2 ⊙ LΘ . We include the argument for the sake of completeness.
Lemma 2.6. Every algebraic Lc,cond
Φ -atom belongs to L0,cond
Φ (M; ℓc2 ). More precisely, if x is an
algebraic Lc,cond
Φ -atom then kxkLcond (M;ℓc ) ≤ 1.
Φ 2
Using the above notion of atoms, we may naturally consider the following concept of atomic
decompositions:
Definition 2.7. A sequence x ∈ LΦ (M; ℓc2 ) is said to admit an algebraic Lc,cond
Φ -atomic decom-
position if X
x= λk a(k) ,
k
where for each k, a(k) is either an algebraic Lc,cond
Φ -atom Por a(k) belongs to the unit ball of
the conditioned space LcΦ (M; E1 ) and λk ∈ C satisfying p
k |λk | < ∞ for 0 < p < 1 and
P c,cond
k |λk | < ∞ for 1 ≤ p < 2. Since Φ is p-convex and algebraic LΦ -atoms belong to the unit
cond c c,cond
ball of LΦ (M; ℓ2 ), it follows that if x admits an algebraic LΦ -atomic decomposition then it
cond c
belongs to LΦ (M; ℓ2 ).
Following [33], the corresponding algebraic atomic column conditioned space Lcond c
Φ,aa (M; ℓ2 ) is
defined to be the completion of the space of all x that admit algebraic Lc,condΦ -atomic decom-
cond c c,cond
positions in the space LΦ (M; ℓ2 ). If x admits an algebraic LΦ -atomic decomposition, we
set: X 1/p
kxkLcond (M;ℓc ) = inf |λk |p for 0 < p ≤ 1
Φ,aa 2
k
and X
kxkLcond (M;ℓc ) = inf |λk | for 1 < p < 2,
Φ,aa 2
k
where the infimum is taken over all decompositions of x as described above. We refer the reader
to [7] for a more in depth discussion on the need to separate the two cases.
The following result generalizes the atomic decomposition of conditioned Lp -spaces from [33]
in two directions: it is valid for Orlicz spaces and also cover the quasi-Banach space range.
This will play a crucial role in the next section. The approach of [33] was by duality which is not
applicable to the present situation since we are dealing with not necessarily convex functions. Our
constructive proof given below combined ideas from [7] and the case of Hardy spaces associated
with convex Orlicz functions considered in [44]. This may be of independent interest.
Theorem 2.8. Let Φ be p-convex and q-concave for 0 < p ≤ q < 2. Then the two spaces
Lcond c cond c
Φ,aa (M; ℓ2 ) and LΦ (M; ℓ2 ) coincide (with constant of isomorphism depending only on p and
q).
12 RANDRIANANTOANINA
We should point out here that since the function t 7→ Ψ(t1/2 ) is operator monotone decreasing
and λ2 σj2 ≤ λ2 σj+1
2 , we have Ψ(λσ −1 − Ψ(λσ )−1 ≥ 0.
j+1 ) ≤ Ψ(λσj ). Taking inverses, Ψ(λσj+1 ) j
Thus, taking 1/2-power in the expression above is justified.
The column sequence is defined by setting:
(
β1,1 := Ψ(λσ2 )−1/2 ;
(2.12) 1/2
βm,1 := Ψ(λσm+1 )−1 − Ψ(λσm )−1 , for m ≥ 2.
Then α = (αn,j )j<n is a strictly lower triangular matrix and β = (βm,1 )m≥1 is an adapted
sequence. Moreover, for every n ≥ 2, it clearly follows from the definition that
X
(α.β)n,1 = αn,j .βj,1 = xn .
j≥1
That is, we have the factorization x = α . β. We claim that the product α . β satisfies the desired
norm estimate. We begin with the L2 -norm of α.
2 X X
α
=
αn,j
2
2 2
n≥2 1≤j<n
X X
= τ Ψ(λσn )|xn |2 Ψ(λσn ) Ψ(λσ2 )−1 + Ψ(λσj+1 )−1 − Ψ(λσj )−1
n≥2 2≤j<n
X
2
= τ |xn | Ψ(λσn ) .
n≥2
INTERPOLATION 13
We may deduce from Lemma 2.4(iv) that there is a constant Cp,q so that
2
(2.13)
α
≤ Cp,q λ−2 τ Φ(λσc (x)) .
2
′ so that
On the other hand, Lemma 2.4(ii) implies that there exists a constant Cp,q
h X 1/2 i h i
(2.14) τ Θ |βm,1 |2 = τ Θ Ψ(λσc (x))−1/2 ≤ Cp,q′
τ Φ(λσc (x)) .
m≥1
′ } + 1. Since t 7→ t−p Φ(t) is non-decreasing, one can easily verify that
Let Kp,q = max{Cp,q , Cp,q
−1 Φ(K 1/p
Φ(t) ≤ Kp,q p,q t). Using this fact, we get from (2.13) and (2.14) that
2
α
≤ λ−2 τ Φ(Kp,q 1/p
λσc (x))
2
and h X 1/2 i
τ Θ |βm,1 |2 1/p
≤ τ Φ(Kp,q λσc (x)) .
m≥1
1/p 1/p
Choose λ so that τ ≤ 1. This can be achieved with λ−1 = Kp,q
σc (x)
L (M) .
Φ(Kp,q λσc (x))
Φ
1/p
With the above choice of λ, we clearly have
α
2 ≤ Kp,q
x
Lcond (M;ℓc ) and
β
L (M;ℓc ) ≤ 1.
Φ 2 Θ 2
This proves the desired estimate and therefore the second part of the theorem.
To conclude the proof, we apply the case of F0 with the direct sum (2.4). We see that every
y ∈ F admits an algebraic Lc,cond
Φ -atomic decomposition with kykLcond (M;ℓc ) .p,q kykLcond (M;ℓc ) .
Φ,aa 2 Φ 2
Since F is dense in Lcond c cond c cond c
Φ (M; ℓ2 ), it follows that LΦ,aa (M; ℓ2 ) = LΦ (M; ℓ2 ).
Remark 2.9. Using λ = 1 in the proof above, we also obtain a moment version P of the preceding
0
theorem: given a sequence x = (xn )n≥1 ∈ F , the column matrix x = n≥1 xn ⊗ en,1 admits
a factorization x = α . β with α is a strictly lower triangular in L2 (M⊗B(ℓ2 (N))) and β is an
adapted column matrix satisfying
X
α
+ τ Θ ( |βn,1 |2 1/2
) . p,q τ Φ σ c (x) .
2
n≥1
We consider now the version of algebraic atomic decomposition for martingale Hardy spaces.
The next consideration generalizes notions from [7] and [44] for non necessarily convex Orlicz
functions.
Definition 2.10. Let 0 < p < q < 2 and Φ be an Orlicz function that is p-convex and q-concave.
An operator
P x ∈ LΦ (M) is called an algebraic hcΦ -atom, whenever it can be written in the form
x = n≥1 yn bn , with yn and bn satisfying the following conditions:
(i) En (yn ) = 0 and bn ∈ LΘ (Mn ) for all n ≥ 1;
X
2
X 1/2
(ii)
yn
≤ 1 and
|bn |2
≤ 1.
2 LΘ (M)
n≥1 n≥1
14 RANDRIANANTOANINA
Following [7, 44], this concept of atoms naturally leads to the consideration of the corresponding
atomic decomposition for conditioned Orlicz-Hardy spaces: we say that an operator x ∈ LΦ (M)
admits an algebraic hcΦ -atomic decomposition if
X
x= λk ak ,
k
where for each
P k, ak is an algebraic hcΦ -atom an element of the unit ball of LΦ (M1 ), and λk ∈ C
or P
satisfying k |λk |p < ∞ for 0 < p ≤ 1 and k |λk | < ∞ for 1 < p < 2. The corresponding
algebraic atomic column martingale Hardy space hcΦ,aa (M) is defined to be the space of all x
which admit a algebraic hcΦ -atomic decomposition and is equipped with
X 1/p
kxkhcΦ,aa = inf |λk |p for 0 < p ≤ 1
k
and X
kxkhcΦ,aa = inf |λk | for 1 < p < 2,
k
where the infimum are taken over all decompositions of x as described above.
We note that the above concepts were introduced in [44] for the case where Φ is a convex
function. Our main focus here is the case where Φ is p-convex with 0 < p < 1.
The next result is an extension of [7, Theorem 3.10] to the case of Orlicz function spaces. It
follows immediately from Theorem 2.8 and the complementation result stated in Lemma 2.3.
Corollary 2.11. Let 0 < p < q < 2 and Φ is an Orlicz function that is p-convex and q-concave.
Then
hcΦ (M) = hcΦ,aa (M).
More precisely, if x ∈ hcΦ (M), then x admits a unique decomposition x = x1 + y where x1 ∈
LΦ (M1 ) and y is a scalar multiple of an algebraic hcΦ -atom.
Indeed, if Π : Lcond c c
Φ (M; ℓ2 ) → hΦ (M) denotes the norm one projection described in Lemma 2.3
and x = α . β is a algebraic Lc,cond
Φ -atom with α being a strictly lower triangular matrix in
L2 (M⊗B(ℓ2 (N))) and β is an adapted column matrix in LΘ (M; ℓc2 ), then we have
XX
Π(x) = dn (αn,k )βk,1
n n>k
X X
= dn (αn,k ) βk,1
k n>k
X
= ak βk,1 .
k
P 2 2
Clearly, we have for every k ≥ 1, Ek (ak ) = 0. Moreover, k≥1 kak k2 ≤ kαk2 . This shows
in particular that Π(x) is an algebraic hcΦ -atom. The assertions in the corollary follow from
combining this fact with Theorem 2.8 and direct sum.
We conclude this section with a companion of Theorem 2.8 for spaces of adapted sequences.
It may be viewed as the algebraic atomic decompositions for spaces of adapted sequences. This
will be used in the next section.
Proposition 2.12. Let Φ be a p-convex and q-concave Orlicz function for 0 < p < q < 2. If
x = (xn )n≥1 is a sequence in Lad c ad c
Φ (M; ℓ2 ) then there exists a sequence β = (βn )n≥1 in LΘ (M; ℓ2 )
and a lower triangular matrix α with the following properties:
(i) α ∈ L2 (M⊗B(ℓ2 (N)));
INTERPOLATION 15
First, we note that since square functions are well-defined for elements of Lad c
Φ (M; ℓ2 ), reduction
to finite sequences or sequences of finite supports is not necessary.
Sketch of the proof. Let x = (xn )n≥1 ∈ Lad c
Φ (M; ℓ2 ). The construction is an adaptation of the
proof of Theorem 2.8 but using square functions in place of conditioned square functions.
For each n ≥ 1, Sc,n (x) ∈ LΦ (Mn ). That is, (Sc,n (x))n≥1 is an adapted sequence. We simply
write ςn for Sc,n (x). As before, we may assume that the ςn ’s are invertible with bounded inverse.
Similarly, (Ψ(ςn ))n≥1 is an adapted sequence with the Ψ(ςn )’s being invertible.
As in the proof of Theorem 2.8, fix λ > 0 and set
(
αn,1 := xn Ψ(λςn )Ψ(λς1 )−1/2 ;
1/2
αn,j := xn Ψ(λςn ) Ψ(λςj )−1 − Ψ(λςj−1 )−1 , for 2 ≤ j ≤ n;
and (
β1,1 := Ψ(λς1 )−1/2 ;
1/2
βm,1 := Ψ(λςm )−1 − Ψ(λςm−1 )−1 , for m ≥ 2.
A slight difference here is that unlike in the proof of Theorem 2.8, we do not make the indexing
shift. The factorization x = α . β is straightforward. Clearly, (βm,1 )m≥1 is an adapted sequence.
Moreover, as xn ∈ LΦ (Mn ), it follows that αn,j is affiliated with Mn and (αn,j )1≤j≤n is a lower
triangular matrix.
−1 1/p
We may choose λ exactly as in the proof of Theorem
2.8.
That is, λ
=
Kp,q kSc (x)kLΦ (M) .
With this choice, the verification of the fact that
α
2 .
β
L (M;ℓc ) .p,q
x
L (M;ℓc ) follows the
Θ 2 Φ 2
same reasoning as in the proof of Theorem 2.8 and is left to the reader.
On the other hand, if x admits a factorization x = α . β, then from the fact that LΦ = L2 ⊙ LΘ ,
we may conclude as in the proof of Lemma 2.6 that
kxkLΦ (M;ℓc2 ) ≤ kαk2 . kβkLΘ (M;ℓc2 ) .
The proof is complete.
Remark 2.13. Using factorizations of operator-valued triangular matrices (or more generally,
elements of Hardy spaces associated with semifinite version of subdiagonal algebras in the sense
of Arveson), the existence of a factorization x = α . β, where α is a lower triangular and β is a
column matrix, is clear. We refer to [37] for this fact. The main point of Proposition 2.12 is that
when x is an adapted sequence, we can choose the sequence β to be adapted and the matrix α
to satisfy the extra property stated in item (ii). These additional facts are very important in the
next section.
Let (A0 , A1 ) be a compatible couple of quasi-Banach spaces in the sense that both A0 and A1
embed continuously into some topological vector space Z. This allows us to define the spaces
A0 ∩ A1 and A0 + A1 . These are quasi-Banach spaces when equipped with quasi-norms:
n
o
x
= max
x
,
x
A0 ∩A1 A0 A1
and
n
o
x
= inf
x0
+
x1
: x = x0 + x1 , x0 ∈ A0 , x1 ∈ A1 ,
A0 +A1 A0 A1
respectively.
Definition 3.1. A quasi-Banach space A is called an interpolation space for the couple (A0 , A1 )
if A0 ∩ A1 ⊆ A ⊆ A0 + A1 and whenever a bounded linear operator T : A0 + A1 → A0 + A1 is
such that T (A0 ) ⊆ A0 and T (A1 ) ⊆ A1 , we have T (A) ⊆ A and
T : A → A
≤ c max
T : A0 → A0
,
T : A1 → A1
for some constant c.
If A is an interpolation space for the couple (A0 , A1 ), we write A ∈ Int(A0 , A1 ). Below, we are
mostly interested in two well-known specific interpolation methods generally referred to as real
method and complex method.
We begin with a short discussion of the real interpolation method. A fundamental notion for
the construction of real interpolation spaces is the K-functional which we now describe. For
x ∈ A0 + A1 , we define the K-functional by setting for t > 0,
n
o
K(x, t) = K x, t; A0 , A1 = inf
x0
A0 + t
x1
A1 : x = x0 + x1 , x0 ∈ A0 , x1 ∈ A1 .
Note that each t > 0, x 7→ K(x, t) gives an equivalent quasi-norm on A0 + A1 . There is also the
dual notion of J-functionals which is defined for y ∈ A0 ∩ A1 and t > 0,
n
o
J(y, t) = J y, t; A0 , A1 = max
y
A0 , t
y
A1 .
If 0 < θ < 1 and 0 < γ < ∞, we recall the real interpolation space Aθ,γ = (A0 , A1 )θ,γ by
x ∈ Aθ,γ if and only if
Z ∞ γ dt 1/γ
x
= t −θ
K x, t; A0 , A1 < ∞.
(A0 ,A1 )θ,γ t
0
If γ = ∞, we define x ∈ Aθ,∞ if and only if
x
= sup t−θ K(x, t; A0 , A1 ) < ∞.
(A0 ,A1 )θ,∞
t>0
For 0 < θ < 1 and 0 < γ ≤ ∞, k · kθ,γ is a quasi-norm and (Aθ,γ , k · kθ,γ ) is a quasi-Banach
space. Moreover, the space Aθ,γ is an interpolation space for the couple (A0 , A1 ) in the sense of
Definition 3.1. There is also an equivalent description of Aθ,γ using the J-functionals for which
we refer to [3, 30] for the exact formulation.
It is worth noting that the real interpolation method is well understood for the couple (Lp0 , Lp1 )
for both the classical case and the noncommutative case. We record here that Lorentz spaces
can be realized as real interpolation spaces for the couple (Lp0 , Lp1 ). More precisely, if N is a
semifinite von Neumann algebra, 0 < p0 < p1 ≤ ∞, 0 < θ < 1, and 0 < q ≤ ∞ then, up to
equivalent quasi-norms (independent of N ),
Lp0 (N ), Lp1 (N ) θ,q = Lp,q (N )
where 1/p = (1 − θ)p0 + θ/p1 . In particular, we have
Lp0 (N ), Lp1 (N ) θ,p = Lp (N )
INTERPOLATION 17
with 1/w3 = (1−ψ2 )/r2 +ψ2 /q. Using Wolff’s interpolation theorem with B1 = Ap,γ1 , B2 = Ar2 ,γ ,
B3 = Aw3 ,γ3 , and B4 = Aq,γ2 , we obtain that
Ar2 ,γ = (Ap,γ1 , Aq,γ2 )θ,γ
θ2 ψ2
where θ = . As before, one can verify that 1/r2 = (1 − θ)/p + θ/q.
1 − θ2 + θ2 ψ2
⋄ Case 3. Assume r3 ∈ [w2 , q).
As in the previous case, let w3 ∈ (w1 , w2 ). Since w3 , r3 , q ∈ I, we have
Ar3 ,γ = (Aw3 ,γ3 , Aq,γ2 )θ3 ,γ
where 1/r3 = (1 − θ3 )/w3 + θ3 /q. Next, we apply Case 1 with w3 in place of r1 in order to get
that
Aw3 ,γ3 = (Ap,γ1 , Aq,γ2 )ψ3 ,γ3
with 1/w3 = (1 − θ3 )/p + θ3 /q. The desired statement can be deduced as in Case 2.
Combining the three cases above, we may state that if p ∈ I \ J, q ∈ J \ I, 0 < γ1 , γ2 , γ ≤ ∞,
0 < θ < 1, and 1/r = (1 − θ)/p + θ/q, then
Ar,γ = (Ap,γ2 , Aq,γ2 )θ,γ ,
which is equivalent to the statement that the family (Ap,γ )p,γ∈(0,∞] forms a real interpolation
scale on the interval I ∪ J.
We now state the primary result of the paper. It is an extension of [1, Theorem 4.8] to the full
range 0 < p < ∞ and a noncommutative generalization of (1.3) (see also [46]).
Theorem 3.5. If 0 < θ < 1, 0 < p < ∞, and 0 < λ, γ ≤ ∞, then for 1/r = (1 − θ)/p,
hcp,λ (M), bmoc (M) θ,γ = hcr,γ (M)
with equivalent quasi-norms.
Before we proceed, we should point out that for 0 < p < ∞ and 0 < λ ≤ ∞, the pair
(hcp,λ (M), bmoc (M)) forms a compatible couple. Indeed, as described in the preliminary section,
hcp,λ (M) embeds isometrically into Lp,λ (M⊗B(ℓ2 (N2 ))) which in turn is continuously embeded
into the topological vector space L0 (M⊗B(ℓ2 (N2 ))). Now we verify that bmoc (M) also embeds
continuously into L0 (M⊗B(ℓ2 (N2 ))). This is immediate if M is finite as bmoc (M) ⊆ hc1 (M) ⊆
L0 (M⊗B(ℓ2 (N2 ))). When M is infinite, this can be achieved as follows: since P M1 is semifinite,
choose a family of mutually disjoint finite projections (ej )j∈J in M1 so that j∈J ej = 1 for the
strong operator topology. Let x = (xn )n≥1 ∈ bmoc (M) and for each j ∈ J, set xej = (xn ej )n≥1 . It
is clear that xej ∈ bmoc (M) and s2c (xej ) = ej s2c (x)ej . Since τ (ej ) < ∞, xej ∈ hc2 (M). Therefore
U Dc (xej ) is well-defined in L0 (M⊗B(ℓ2 (N2 ))). Moreover, the module property of U implies
2
P U Dc (xej ) = U Dc (xej ).(ej ⊗ Id) where Id is the identity of B(ℓ2 (N )). This
that P implies that
j∈J U Dc (xej ) converges in measure. This shows in particular that the map x →
7 j∈J U Dc (xej )
c 2
provides a continuous embedding of bmo (M) into L0 (M⊗B(ℓ2(N ))).
Similarly, if 0 < r, s ≤ ∞, then Lcondr (M; ℓc2 ), Lcond
s (M; ℓc2 ) is a compatible couple as both
spaces embed continuously into (Lr + Ls )cond (M; ℓc2 ).
We need some preparation for the proof. We consider the Orlicz structure of the symmetric
function space Lr + tLs for any given 0 < r < s < ∞ and t > 0. To describe this structure, we
consider the following Orlicz function:
(r,s)
Φt (u) = min{ur , ts us }, u ≥ 0.
INTERPOLATION 19
(r,s)
One can verify that Φt (·) is r-convex and s-concave. According to [15, Lemma 3.2], we have
for every f ∈ Lr + Ls and t > 0,
(3.1) 2−1−1/r
f
Φ(r,s) ≤ K f, t; Lr , Ls ≤ 2
f
Φ(r,s) .
t t
In particular, the Orlicz space LΦ(r,s) coincides with the space Lr +tLs with isomorphism constant
t
depending only on the index r and therefore independent of t.
Let 0 < p, q, r, s < 2 and assume that p < q, 1/p = 1/2 + 1/r, and 1/q = 1/2 + 1/s. One can
easily check that
Lp + tLq = L2 ⊙ (Lr + tLs ).
It follows by identification that the following factorization holds for the corresponding Orlicz
spaces:
LΦ(p,q) = L2 ⊙ LΦ(r,s) ,
t t
(p,q) −1
Proof. For the first item (i), we write u−2 Φt (u) = min{u−2+p , tq u−2+q } = max{u2−p , t−q u2−p }
and note that
2−1 (u2−p + t−q u2−p ) ≤ max{u2−p , t−q u2−p } ≤ u2−p + t−q u2−p .
It follows from taking inverses that
(p,q) (p,q) (p,q)
ψt (u) ≤ u−2 Φt (u) ≤ 2ψt (u).
Next, item (iii) can be seen as follows: if a and b are positive operators with a ≤ b then since
0 < 1 − (p/2) < 1 and 0 < 1 − (q/2) < 1, we have
a1−(p/2) + tq a1−(q/2) ≤ b1−(p/2) + tq b1−(q/2) .
Taking inverse operators, we see that
(p,q) (p,q)
ψt (b1/2 ) ≤ ψt (a1/2 ).
(p,q)
This shows that u 7→ ψt (u1/2 ) is operator monotone decreasing.
20 RANDRIANANTOANINA
With the equivalence (i) on hand, the proof of (iv) is identical to the proof of [44, Lemma 3.4]
so we leave it to the reader. It remains to verify (ii). From the equivalence (i), it suffices to prove
that
(r,s) (p,q) (p,q)
Φt u(Φt (u))−1/2 .p,q Φt (u).
(p,q) (r,s)
Since u(Φt (u))−1/2 = max{u1−p/2 , t−q/2 u1−q/2 }, we have from the definition of Φt (·) that
(3.2) n o
(r,s) (p,q)
Φt u(Φt (u))−1/2 = min max{(u1−p/2 )r , (t−q/2 u1−q/2 )r }, max{ts (u1−p/2 )s , ts (t−q/2 u1−q/2 )s } .
Note that (u1−p/2 )r = up and ts (t−q/2 u1−q/2 )s = tq uq . If u1−(p/2) ≥ t−q/2 u1−q/2 then tq uq ≥ up .
(p,q)
In particular, Φt (u) = up . Then (3.2) implies that
(r,s) (p,q) (p,q)
Φt u(Φt (u))−1/2 = min up , ts (u1−(p/2) )s ≤ Φt (u).
(p,q)
Similarly, if u1−(p/2) ≤ t−q/2 u1−q/2 then tq uq ≤ up and therefore, Φt (u) = tq uq . We can deduce
from (3.2) that
n o
(r,s) (p,q) (p,q)
Φt u(Φt (u))−1/2 = min (t−q/2 u1−q/2 )r , tq uq ≤ Φt (u).
This completes the proof.
The next proposition constitutes the decisive step toward our proof of Theorem 3.5. We
formulate it here for the more general conditioned spaces.
Proposition 3.7. Let ν be a positive integer with ν ≥ 2. Assume that 2/(ν + 1) < p ≤ 2/ν and
p < q < 2/(ν − 1). If x ∈ F, then for every t > 0,
K x, t; Lcond
p (M; ℓc2 ), Lcond
q (M; ℓc2 ) ≈p,q K σc (x), t; Lp (M), Lq (M) .
Proof. • We observe first that one inequality in the equivalence follows easily from the isometric
embeddings described in the preliminary section. Indeed, let x ∈ F and set N = M⊗B(ℓ2 (N2 )).
Recall that for 0 < r ≤ ∞, the map U : Lcond r (M; ℓc2 ) → Lr (N ) is an isometry satisfying the
identity
|U (x)| = σc (x) ⊗ e1,1 ⊗ e1,1 .
It follows that for every t > 0,
K σc (x), t; Lp (M), Lq (M) = K |U (x)|, t; Lp (N ), Lq (N )
= K U (x), t; Lp (N ), Lq (N )
≤ K x, t; Lcond
p (M; ℓc2 ), Lcond
q (M; ℓc2 )
which verifies one inequality with constant 1.
• The reverse inequality is more involved. We consider the following two finite sequences of
indices: set p0 = p and q0 = q and for 1 ≤ m ≤ ν − 1,
1/pm−1 = 1/2 + 1/pm and 1/qm−1 = 1/2 + 1/qm .
From earlier discussions, the following factorization holds for the respective Orlicz spaces:
L (pm−1 ,qm−1 ) = L2 ⊙ LΦ(pm ,qm ) .
Φt t
We note that by Lemma 3.6, atomic decompositions stated in Theorem 2.8 and Proposition 2.12
apply to each of these factorizations.
The main idea in the argument below is to repeatedly apply the above factorization until one
gets indices that are strictly larger than 1. At that point, splittings into adapted sequences are
possible via the noncommutative Stein inequality. The restriction imposed on the values of p and
INTERPOLATION 21
q is needed in the argument since at every step (except the last one) we need both indices pm
and qm to remain in the open interval (0 , 2) so that algebraic atomic decompositions from the
previous section can be applied.
We now present the details of the proof. Assume first that x ∈ F0 (the general case will be
dealt later).
Fix t > 0. We apply Theorem 2.8 to the Orlicz space LΦ(p,q) . We have the following factoriza-
t
tion:
(3.3) x = α(1) . β (1)
where α(1) is a strictly lower triangular matrix with entries in L2 (M) and β (1) is an adapted
column matrix in LΦ(p1 ,q1 ) (M; ℓc2 ) satisfying:
t
(1)
(1)
(3.4)
α
.
β
.p,q
σc (x)
L .
2 L (M;ℓc )
(p ,q ) 2 (M) (p,q)
Φt 1 1 Φt
Next, we inductively construct finite sequences {α(m) : 2 ≤ m ≤ ν −1} and {β (m) : 1 ≤ m ≤ ν −1}
(m)
where the α(m) ’s are lower triangular matrices taking values in L2 (M) satisfying αn,j ∈ L2 (Mn )
for 1 ≤ j ≤ n, and β (m) ’s are adapted column matrices. Both sequences satisfy for 2 ≤ m ≤ ν − 1:
This is done by applying Proposition 2.12 to β (m−1) ∈ Lad(pm−1 ,qm−1 ) (M; ℓc2 ). The constant de-
Φt
pends on pm−1 and qm−1 but since they depend on p and q respectively, we may state that for
each step, the constant depends on p and q.
Clearly, the above construction induces a factorization:
x = α(1) . . . α(ν−1) . β (ν−1) .
Now, we consider the adapted sequence β (ν−1) ∈ L (pν−1 ,qν−1 ) (M; ℓc2 ). By identification, we have
Φt
β (ν−1) ∈ (Lpν−1 + tLqν−1 )(M; ℓc2 ) = Lpν−1 (M; ℓc2 ) + tLqν−1 (M; ℓc2 ).
It follows from the definition of sum of two Banach spaces that β (ν−1) admits a decomposition
β (ν−1) = ξ (1) +ξ (2) with ξ (1) ∈ Lpν−1 (M; ℓc2 ) and ξ (2) ∈ Lqν−1 (M; ℓc2 ) satisfying the norm estimate:
(1)
(2)
(ν−1)
(3.7)
ξ
c + t ξ
c ≤ 2 β
c .
Lpν−1 (M;ℓ2 ) Lqν−1 (M;ℓ2 ) (Lpν−1 +tLqν−1 )(M;ℓ2 )
The important fact here is that 1 < pν−1 < qν−1 < ∞. By applying the noncommutative Stein
(1)
inequality ([37]), we may replace ξ (1) and ξ (2) by adapted sequences ζ (1) = {En (ξn )}n≥1 and
(2)
ζ (2) = {En (ξn )}n≥1 satisfying:
(1)
(2)
(ν−1)
(3.8)
ζ
c + t
ζ
c ≤ C p,q
β
.
Lp
ν−1(M;ℓ )
2 Lq (M;ℓ )
ν−1 2 (Lp +tLq )(M;ℓc ) ν−1 ν−1 2
The constant Cp,q can be taken to be equal to 2 max{γpν−1 , γqν−1 } where γr is the constant from
the noncommutative Stein inequality for 1 < r < ∞. This justifies that Cp,q depends only on p
and q since pν−1 and qν−1 depend on p and q respectively.
Next, we consider two sequences of operators by setting:
x(1) = α(1) . . . α(ν−1) . ζ (1) and x(2) = α(1) . . . α(ν−1) . ζ (2) .
22 RANDRIANANTOANINA
Since β (ν−1) = ζ (1) + ζ (2) , we clearly have x = x(1) + x(2) . We claim that x(1) ∈ Lcond p (M; ℓc2 )
and x(2) ∈ Lcond
q (M; ℓc2 ). Indeed, let y (1) = α(2) . . . α(ν−1) . ζ (1) and y (2) = α(2) . . . α(ν−1) . ζ (2) .
(ν−1)
The fact that ζ (1) is adapted and the property that αn,j ∈ L2 (Mn ) for 1 ≤ j ≤ n implies
that βb (ν−2) := α(ν−1) .ζ (1) is an adapted sequence. Moreover, by Hölder’s inequality, we have
βb(ν−2) ∈ Lpν−2 (M;ℓc2 ) with
(ν−2)
βb
≤
α(ν−1)
2 .
ζ (1)
Lp (M;ℓc ) .
Lp (M;ℓc )
ν−2 2 ν−1 2
One can work backward and set βb(m−1) = α(m) . βb(m) for 2 ≤ m ≤ ν − 2 to see that y (1) is an
adapted column sequence. Moreover, a repeated use of Hölder’s inequality yields:
(1)
ν−1
Y
y
≤
α(m)
ζ (1)
< ∞.
Lp1 (M;ℓc2 ) 2 Lp c
ν−1 (M;ℓ2 )
m=2
Since α(1) is strictly lower triangular and y (1) is adapted, it follows that x(1) = α(1) . y (1) ∈
Lcond
p (M; ℓc2 ). Similar argument can be applied to deduce that x(2) ∈ Lcond q (M; ℓc2 ).
(1) (2)
We now verify that the decomposition x = x + x provides the desired inequality between
the two K-functionals. Indeed, we have the following norm estimates:
K x, t; Lcond
p (M; ℓc2 ), Lcond
q (M; ℓc2 ) ≤
x(1)
Lcond (M;ℓc ) + t
x(2)
Lcond (M;ℓc )
p 2 q 2
ν−1
Y
≤
α(m)
.
ζ (1)
c + t
ζ (2)
Lq c
2 Lp ν−1 (M;ℓ2 ) ν−1 (M;ℓ2 )
m=1
ν−1
Y
(m)
(ν−1)
.p,q
α
.
β
c
2 (Lp ν−1 +tLqν−1 )(M;ℓ2 )
m=1
ν−1
Y
(m)
(ν−1)
.p,q
α
.
β
c .
2 L (p ,q ) (M;ℓ2 )
m=1 Φt ν−1 ν−1
Using the fact that for every 0 < r ≤ ∞, hcr (M) embeds isometrically into a 1-complemented
subspace of Lcond
r (M; ℓc2 ), the next result follows immediately from Proposition 3.7.
Corollary 3.8. Let ν be a positive integer with ν ≥ 2. Assume that 2/(ν + 1) < p ≤ 2/ν and
p < q < 2/(ν − 1). If y is a finite martingale in F(M ) then for every t > 0,
K y, t; hcp (M), hcq (M) ≈p,q K sc (y), t; Lp (M), Lq (M) .
The next lemma is the infinite version of the Banach space case. It will be deduced from
Lemma 3.10 and Wolff’s interpolation theorem. Since our approach differs from [1], we include
the details.
24 RANDRIANANTOANINA
for interpolation ([3, Theorem 3.7.1]) that if 1 < p < r < ∞, then
hcr (M) = hcp (M), bmoc (M) θ,r
which is the desired conclusion.
Proposition 3.12. If 0 < θ < 1, 0 < p, q < ∞, and 0 < γ ≤ ∞, then for 1/r = (1 − θ)/p + θ/q,
Lcond
p (M; ℓc2 ), Lcond
q (M; ℓc2 ) θ,γ = Lcond c
r,γ (M; ℓ2 ),
Next, we deduce the Banach space range using complementation: if 1 < p < r < q < ∞ and
1/r = (1 − ψ)/p + ψ/q, then
(3.11) Lad c ad c ad c
r (M; ℓ2 ) = Lp (M; ℓ2 ), Lq (M : ℓ2 ) ψ,r .
Using (3.10) and (3.11), we can repeat the inductive argument used in the proof of Theorem 3.5
with the intervals I1 = (1, ∞) and Iν = (2/(ν + 1), 2/(ν − 1)) for ν ≥ 2, to conclude that the
family {Lad c
p (M; ℓ2 )}p∈(0,∞);γ∈(0,∞] forms a real interpolation scale.
Assume that 1 ≤ p < ∞. We recall that the map D(x) = (dxn )n≥1 is an isometric embedding
of Hpc (M) into Lad c
p (M; ℓ2 ). Using the noncommutative Stein inequality when 1 < p < ∞ and the
noncommutative Lépingle-Yor inequality when p = 1 ([39]), the linear map
Π (an )n≥1 = (an − En−1 (an ))n≥1
is simultaneously bounded from Lad c c
p (M; ℓ2 ) onto D Hp (M) for all 1 ≤ p < ∞. As a result, one
can immediately deduce from Theorem 3.13 that the interpolation
(H1c (M), Hqc (M))θ,r = Hrc (M)
for 0 < θ < 1, 1 < q < ∞, and 1/r = (1 − θ) + θ/q holds. Due to this fact, we may view
Theorem 3.13 as an extension of the real interpolation version of [32] to the quasi-Banach space
26 RANDRIANANTOANINA
range. However, when 0 < p < 1, the spaces of adapted sequences cannot be replaced by column
martingale Hardy spaces. In fact, only one inclusion holds if column martingale Hardy spaces are
used. The next result may be viewed as a noncommutative generalization of [18, Theorem 2]. We
also direct the reader to its companion Corollary 3.19 below and a discussion on the non validity
of the reverse inclusions. We refer to [32, 37] for the definition of BMOc (M).
Corollary 3.15. Assume that 0 < p < 1 and 0 < θ < ∞ are such that 1/r = (1 − θ)/p < 1.
Then
Hpc (M), BMO c (M) θ,r ⊂ Hrc (M).
Proof. Let 0 < η < 1 such that 1 = (1 − η)/p + η/r. Let x ∈ Hpc (M), Hrc (M) η,1 . Since
Huc (M) ⊂ Lad c
u (M; ℓ2 ) for u ∈ {p, r}, we have
≈ kdxk(Lad c ad c
p (M;ℓ2 ),Lr (M;ℓ2 ))η,1
3.3. The complex method. We now turn our attention to the case of complex interpolation
method which we now briefly review.
Let S (respectively, S) denote the open strip {z : 0 < Re z < 1} (respectively, the closed
strip {z : 0 ≤ Re z ≤ 1}) in the complex plane C. Let A(S) be the collection of C-valued
functions that are analytic on S and continuous and bounded on S. For a compatible couple of
complex quasi-Banach
Pn spaces (A0 , A1 ), we denote by F0 (A0 , A1 ) the family of functions of the
form f (z) = k=1 fk (z)xk with fk ∈ A(S) and xk ∈ A0 ∩ A1 . We equip F0 (A0 , A1 ) with the
quasi-norm: n
o
f
= max sup
f (it)
, sup
f (1 + it)
.
F0 (A0 ,A1 ) A0 A1
t∈R t∈R
Then F0 (A0 , A1 ) becomes a quasi-Banach space. For 0 < θ < 1, the complex interpolation norm
on A0 ∩ A1 is defined by:
n
o
x
= inf
f
: f (θ) = x, f ∈ F0 (A0 , A1 ) .
[A0 ,A1 ]
θ F0 (A0 ,A1 )
The complex interpolation space (of exponent θ) [A0 , A1 ]θ is defined as the completion of the
quasi-normed space (A0 ∩ A1 , k · k[A0 ,A1 ]θ ).
As in the real method, complex interpolations of the couple (Lp , Lq ) (for 0 < p < q ≤ ∞) are
well-known. Indeed, if N is a semifinite von Neumann algebra and 0 < θ < 1, then
Lp (N ), Lq (N ) θ = Lr (N )
isometrically for 1/r = (1 − θ)/p + θ/q. This fact comes from [49, Theorem 4.1] (see also [38]).
The next result is the version of Theorem 3.5 for the complex method.
INTERPOLATION 27
1 1−θ θ
Theorem 3.16. If 0 < p, q < ∞, 0 < θ < 1, and = + , then
r p q
c
c c
hr (M) = hp (M), hq (M) θ
with equivalent quasi-norms.
For the proof, we will use the next result which provides a connection between complex inter-
polation method and real interpolation method that is valid for quasi-Banach spaces. We should
note that for Banach spaces, the inequality in the next theorem is actually an equivalence but for
quasi-Banach spaces only one inequality is valid in its full generality.
Theorem 3.17 ([9, Theorem 3]). Let (A0 , A1 ) be a compatible couple of quasi-Banach spaces.
Let 0 < θj < 1, 0 < θ < 1, and 0 < γj ≤ ∞. Denote Ej = (A0 , A1 )θj ,γj for j = 0, 1. If
1/γ = (1 − θ)/γ0 + θ/γ1 and λ = (1 − θ)θ0 + θθ1 then for every a ∈ A0 ∩ A1 ⊂ E0 ∩ E1 ,
a
≤ C
a
.
[E0 ,E1 ] θ (A0 ,A1 ) λ,γ
Proof of Theorem 3.16. Assume that 0 < p < q < ∞ and fix 0 < p0 < p. Then according
to Theorem 3.5, hcp (M) = (hcp0 (M), bmoc (M))θ0 ,p and hcq (M) = (hcp0 (M), bmoc (M))θ1 ,q where
1/p = (1 − θ0 )/p0 and 1/q = (1 − θ1 )/p0 . We may state from Theorem 3.17 that if 1/γ =
(1 − θ)/p + θ/q and λ = (1 − θ)θ0 + θθ1 then for every a ∈ hcp0 (M) ∩ bmoc (M),
a
c ≤ C
a
c .
p
c
[h (M),h (M)]
q θ (h (M),bmoc (M))
p0 λ,γ
One can easily verify that (1− λ)/p0 = 1/γ = 1/r and therefore we may deduce from Theorem 3.5
that
(3.12)
a
c c ≤ C ′
a
c .
[hp (M),hq (M)]θ hr
Proposition 3.18. If 0 < θ < 1 and 0 < p, q < ∞ then for 1/r = (1 − θ)/p + θ/q, the following
hold:
cond
Lp (M; ℓc2 ), Lcond
q (M; ℓc2 ) θ = Lcond
r (M; ℓc2 )
and
Lad c ad c ad c
p (M; ℓ2 ), Lq (M; ℓ2 ) θ = Lr (M; ℓ2 )
with equivalent quasi-norms.
As in the case of real interpolation, we may view the second assertion in the proposition as
an extension of Musat’s result to the quasi-Banach space range. Moreover, using [48, Lemma 1]
(which is valid for quasi-Banach spaces), one can adapt the argument used in the proof of Corol-
lary 3.15 to show that the noncommutative analogue of (1.2) holds:
Corollary 3.19. Assume that 0 < p < 1 and 0 < θ < ∞ are such that 1/r = (1 − θ)/p < 1.
Then
c
Hp (M), BMO c (M) θ ⊂ Hrc (M).
Our method of proof only applies under the assumption that r > 1. We should point out that
the reverse inclusion does not hold even in the classical setting. An example exhibited in [18,
S a probability space (Ω, F, P) and an increasing filtration of σ-fields
p. 66] shows that there exists
(Fn )n≥1 of F with F = σ( n≥1 Fn ) and such that if 0 < p < 1 < r < ∞, then for 1/r = (1−θ)/p,
Hp (Ω), BMO(Ω) θ 6= Hr (Ω).
All results stated in this section have row counterparts. However, at the time of this writing, it
is unclear if for 0 < p, q < 1, the interpolation results for column/row conditioned Hardy spaces
have counterparts to the couple of diagonal Hardy spaces (hdp (M), hdq (M)).
equipped with the norm khkE (1/p) = k |h|1/p kpE . According to [5, Corollary 4.6], E (1/p) ∈
Int(L1 , Lq/p ). Let F be a Banach function space so that E (1/p) = (L1 , Lq/p )F ;K . We make
INTERPOLATION 29
the observation from Homlsted’s formula ([16, Theorem 4.1]) that if a is a positive function in
Lp + Lq then:
h i1/p
K a, t; Lp , Lq ≈p,q K ap , tp ; L1 , Lq/p .
Let x ∈ E. We have
p
x
=
|x|p
(1/p)
E E
≈E
t 7→ K(|x|p , t; L1 , Lq/p )
F
p
≈E
t 7→ K(|x|, t1/p ; Lp , Lq )
F
p
≈E
t 7→ K(|x|, t1/p ; Lp , Lq )
(p) .F
Let Z = {f ∈ L0 : t 7→ f (t1/p ) ∈ F (p) } with the quasi-norm kf kZ = kt 7→ f (t1/p )kF (p) . We see
now that E = (Lp , Lq )Z;K .
Below, we will also use the J-method version of the interpolation associated with function
space which we now briefly describe: for x ∈ E0 + E1 , we recall that by a representation of x with
respect to the couple (E0 , E1 ), we mean a measurable function u : (0, ∞) → E0 ∩ E1 satisfying
Z ∞
dt
x= u(t)
0 t
where the convergence is taken in E0 + E1 . Recall that for y ∈ E0 ∩ E1 and t > 0,
J(y, t; E0 , E1 ) = max{kykE0 ; tkykE1 }.
For a given function space F, we define the quasi-norm
kxk(E0 ,E1 )F ;J := inf kt 7→ J(u(t), t; E0 , E1 )kF
where the infimum is taken over all representation u(·) of x with respect to the couple (E0 , E1 ).
The interpolation space (E0 , E1 )F ;J is defined as the collection of all x ∈ E0 + E1 for which
kxk(E0 ,E1 )F ;J < ∞. We refer to [30] for more on this interpolation method.
From the fact that every interpolation space of the couple (Lp , Lq ) is given by a K-method,
the following can be easily deduced from our results on K-functionals from the previous section:
Proposition 4.3. Let ν be a positive integer with ν ≥ 2. Assume that 2/(ν + 1) < p ≤ 2/ν and
p < q ≤ 2/(ν − 1). If E ∈ Int(Lp , Lq ) with E = (Lp , Lq )F ;K for a quasi Banach function space F
then the following hold:
hcE (M) = (hcp (M), hcq (M) F ;K ,
E cond (M; ℓc2 ) = Lcond
p (M; ℓc2 ), Lcond
q (M; ℓc2 ) F ;K ,
and
E ad (M; ℓc2 ) = Lad c ad c
p (M; ℓ2 ), Lq (M; ℓ2 ) F ;K .
The first assertion in the preceding observation motivates the following more general question:
assume that E0 and E1 are symmetric quasi-Banach
function spaces and E ∈ Int(E0 , E1 ), does
c c c
it follow that hE (M) ∈ Int hE0 (M), hE1 (M) ?
Note that if for j ∈ {0, 1}, Ej ∈ Int(Lpj , Lqj ) for some 1 < pj , qj < ∞, then one can de-
duce from Junge’s representation and interpolation that hcEj (M) embeds complementably into
Ej (M⊗B(ℓ2 (N2 ))). Thus, in this special case, the answer to the above question is clearly positive.
Even for the particular case where E0 = Lp and E1 = Lq , we do not know if the assumptions
on p and q in Proposition 4.3 can be removed when 0 < p < q ≤ 1. This of course is closely
related to asking whether the statement about K-functionals in Proposition 3.7 is valid for any
0 < p < q ≤ 1.
30 RANDRIANANTOANINA
Next, we make the observation that a well-known result for convex functions extends to the
general setting of p-convex functions for 0 < p < 1.
Proposition 4.4. Let 0 < p < q < ∞, and E ∈ Int(Lp , Lq ). There exists a constant cE such
that the following holds: if f , g are functions such that kgkLΦ ≤ kf kLΦ for every function Φ that
is p-convex and q-concave, and if f ∈ E, then g ∈ E with
kgkE ≤ cE kf kE .
Proof. Let f and g as in the statement of the proposition. Using the Orlicz space description of
Lp + tLq in the previous section, we have from the assumption that
K(g, t; Lp , Lq ) .p,q K(f, t; Lp , Lq ), t > 0.
According to Proposition 4.2, the interpolation space E is given by a K-method. Fix a function
space F so that for every h ∈ E,
khkE ≈E kt 7→ K(h, t; Lp , Lq )kF .
We may now deduce from the inequality on K-functionals that
kgkE ≈E kt 7→ K(g, t; Lp , Lq )kF
.E kt 7→ K(f, t; Lp , Lq )kF
≈E kf kE .
This verifies the desired conclusion.
Our first result in this section is a comparison between conditioned column space and column
space for a class of symmetric spaces.
Theorem 4.5. Let 0 < p < q < 2 and F ∈ Int(Lp , Lq ). There exists a constant CF such that for
any x ∈ F cond (M; ℓc2 ), the following holds:
x
cond
c ≤ CF x c .
F (M;ℓ2 ) F (M;ℓ2 )
Similarly, if Φ is an Orlicz function that is p-convex and q-concave for 0 < p < q < 2 then there
exists a constant Cp,q so that for any sequence x = (xk )k≥1 with σc (x) ∈ LΦ (M),
τ Φ Sc (x) ≤ Cp,q τ Φ σc (x)
Proof. Let Φ be an Orlicz function that isPp-convex and q-concave. According to Theorem 2.8, if
x = (xk )k≥1 in F, the column vector x = k≥1 xk ⊗ ek,1 admits a factorization x = α . β where α
is a strictly lower triangular matrix taking values in L2 (M) and β ∈ Lad c
Θ (M; ℓ2 ) where Θ is the
Orlicz function satisfying LΦ = L2 ⊙ LΘ (as described in Theorem 2.8) and
α
.
β
cond
c ≤ Cp,q x c .
L2 (M⊗B(ℓ2 (N))) LΘ (M;ℓ2 ) LΦ (M;ℓ2 )
where the generalized singular value on the left hand side is taken with respect to M⊗B(ℓ2 (N))
and the one the right hand side is taken with respect to M⊗B(ℓ2 (N2 )). It follows from Proposi-
tion 4.4 that there exists a constant cF such that
µ(ξ)
≤ CF
µ(U (ξ))
.
F F
This is equivalent to
ξ
≤ CF
ξ
F cond (M;ℓc )
F (M;ℓc )
2 2
which is the desired conclusion.
For the Φ-moment version, it suffices to repeat the above argument but using Remark 2.9 in
place of Theorem 2.8.
As an immediate consequence of Theorem 4.5, we have the following extension of the reverse
dual Doob inequality proved in [26, Theorem 7.1] for noncommutative Lp spaces (0 < p < 1) to
the general case of noncommutative symmetric quasi-Banach spaces.
Corollary 4.6. Let E be a symmetric quasi-Banach function space with E ∈ Int(Lp , Lq ) for
0 < p < q < 1. There exists a constant CE so that for any sequence of positive operators (ak ) in
F, the following holds:
X
X
ak
≤ CE
Ek (ak )
.
E(M) E(M)
k≥1 k≥1
Similarly, if Φ is an Orlicz space that is p-convex and q-concave for 0 < p < q < 1, then there
exists a constant Cp,q so that for any sequence of positive operators (ak ) in F, the following holds:
X X
τ Φ ak ≤ Cp,q τ Φ Ek (ak )
k≥1 k≥1
Proof. Let E ∈ Int(Lp , Lq ) with 0 < p < q < 1. Let (ak )k≥1 be a sequence of positive operators
in F. If E (2) is the 2-convexification of E, then E (2) ∈ Int(L2p , L2q ). The conclusion follows
1/2
immediately from the first inequality in Theorem 4.5 using F = E (2) and (xk ) = (ak ). Similar
argument applies to the Φ-moment case using the second inequality in Theorem 4.5 to the Orlicz
function t 7→ Φ(t2 ).
The next result extends [20, Theorem 4.11] to the case of noncommutative martingale Hardy
spaces associated with symmetric function spaces and moment inequalities.
Theorem 4.7. Let 0 < p < q < 2. If F ∈ Int(Lp , Lq ) then there exists a constant CF such that
for every x ∈ hcF (M), the following two inequalities hold:
x
c ≤ CF
x
c
HF (M) hF (M)
and
x
≤ CF
x
hc (M) .
F (M) F
Similarly, if Φ is p-convex and q-concave for 0 < p < q < 2 then there exists a constant Cp,q so
that for every x ∈ hcΦ (M), we have
n o
max τ Φ Sc (x) ; τ Φ |x| ≤ Cp,q τ Φ sc (x) .
Proof. • For the first inequality, let y be a martingale in F(M ). Repeating the argument in the
proof of Theorem 4.5 with the martingale difference sequence (dyk ), we have for every p-convex
and q-concave Orlicz function Φ:
y
c ≤ Cp,q
y
c .
HΦ (M) hΦ (M)
32 RANDRIANANTOANINA
By density, the above inequality shows that there exists a bounded linear map I : hcΦ (M) →
HΦc (M) with I(y) = y for every y ∈ F(M ). For simplicity, for a given x ∈ hc (M), we will denote
Φ
I(x) by x. That is, for x ∈ hcΦ (M) we may state:
µ(Sc (x))
≤ Cp,q
µ(|U Dc (x)|)
LΦ LΦ
where the generalized singular numbers are computed in the appropriate von Neumann alge-
bras. By Proposition 4.4, there exists a constant CF so that whenever x ∈ hcF (M), we have
µ(|U Dc (x)|) ∈ F and
µ(Sc (x))
≤ CF
µ(|U Dc (x)|)
.
F F
This is equivalent to the first inequality:
x
c ≤ CF
x
hc (M) .
HF (M) F
• In view of the argument above, it suffices to verify the second inequality for the case of Orlicz
function spaces. For this special case, our proof below is modeled after the argument used in
[7, Corollary 3.14] for the case of Lp -spaces. Let Φ be a p-convex and q-concave Orlicz function
and x be a martingale in F(M ). By approximation, we assume that for every n ≥ 1, sc,n (x) is
invertible with bounded inverse.
P As above, we denote sc,n (x) by sn and we take s0 = 0.
Let λ > 0. We write x = l≥1 al bl where for every l ≥ 1, we set:
X
al = dxn Ψ(λsn )(Ψ(λsl )−1 − Ψ(λsl−1 )−1 )1/2 and bl = (Ψ(λsl )−1 − Ψ(λsl−1 )−1 )1/2 .
n≥l
Following the same reasoning as in the proof of Theorem 2.8, there exists a constant Cp,q and a
corresponding choice of λ so that
X
1/2
al
2 ≤ C p,q
sc (x)
2 L (M) Φ
l≥1
and
Ψ(λs)−1/2
≤ 1.
L Θ (M)
This yields that the inequality in the statement is verified for Orlicz function spaces. More
precisely,
(4.1)
x
≤ C p,q
sc (x)
.
L (M) Φ L (M) Φ
The case of moment inequalities follows directly from the moment part of Theorem 4.5 and from
using λ = 1 in the proof above.
We now proceed with further application of Proposition 2.12. Below, we show that noncommu-
tative Davis decompositions can be easily deduced from factorizations of adapted sequences. We
refer to [25], [34], and [43] for various forms of noncommutative Davis decompositions. We refer
to [44] for formal definition of the noncommutative vector-valued space LΦ (M; ℓc1 ) used below.
INTERPOLATION 33
Theorem 4.8. Assume that 0 < p < q < 2 and Φ is an Orlicz function that is p-convex and q-
concave. Given an adapted sequence ξ = (ξn )n≥1 in LΦ (M; ℓc2 ), there exist two adapted sequences
y = (yn )n≥1 and z = (zn )n≥1 such that:
(i) ξ = y + z;
(ii) y ∈ LΦ (M; ℓc1 ) with
y
L (M;ℓc ) .p,q
ξkLΦ (M;ℓc2 ) ;
Φ 1
(iii) z = λa where a is an algebraic Lc,cond -atom and λ .p,q
ξkL (M;ℓc ) .
Φ Φ 2
In particular,
y
+
z
Lc,cond (M;ℓc ) .p,q
ξkLΦ (M;ℓc2 ) .
LΦ (M;ℓc1 ) Φ 2
Proof. The key ingredients for the proof are the decomposition from [43, Theorem 3.1] and the
two interpolations results from Proposition 3.12 and Theorem 3.13.
Fix ξ ∈ Lad c ad c
2/3 (M; ℓ2 ) ∩ L2 (M; ℓ2 ). Choose a discrete representation of ξ with respect to the
couple (Lad c ad c
2/3 (M; ℓ2 ), L2 (M; ℓ2 )),
X
ξ= ξ (ν)
ν∈Z
with ξ (ν)
∈ Lad ∩c
2/3 (M; ℓ2 ) Lad c
2 (M; ℓ2 ) for every ν ∈ Z, the series is convergent in Lad c
2/3 (M; ℓ2 ) +
ad c
L2 (M; ℓ2 ), and so that
J ξ (ν) , 2ν , Lad c ad c ν ad c ad c
2/3 (M; ℓ2 ); L2 (M; ℓ2 ) ≤ 4K ξ, 2 , L2/3 (M; ℓ2 ); L2 (M; ℓ2 ) .
We refer to [3, Lemma 3.3.2] for the existence of a representation satisfying the properties discribed
above.
For each ν ∈ Z, we apply [43, Theorem 3.1] to the adapted sequence ξ (ν) : there exists two
adapted sequences y (ν) and z (ν) such that ξ (ν) = y (ν) + z (ν) ,
J y (ν) , 2ν , L2/3 (M⊗ℓ∞ ; L2 (M⊗ℓ∞ )) . J ξ (ν) , 2ν , Lad c ad c
2/3 (M; ℓ2 ); L2 (M; ℓ2 ) ,
and
J z (ν) , 2ν , Lcond c cond
2/3 (M; ℓ2 ); L2 (M; ℓc2 ) . J ξ (ν) , 2ν , Lad c ad c
2/3 (M; ℓ2 ); L2 (M; ℓ2 ) .
34 RANDRIANANTOANINA
The desired inequality follows from combining the last two inequalities.
We isolate the following important example. Motivated by the noncommutative Khintchine
inequality for weak-L1 spaces ([4]), we state below a weak-L1 -version of the Davis-decomposition
for adapted sequences. This appears to be new even for the classical setting.
Example 4.10. There exists a constant C > 0 so that for every adapted sequence ξ ∈ L1,∞ (M; ℓc2 ),
there exist two adapted sequences y and z such that:
(i)
ξ =
x + z;
(ii)
y
L1,∞ (M⊗ℓ∞ ) +
z
Lcond (M;ℓc ) ≤ C
ξ
L1,∞ (M;ℓc ) .
1,∞ 2 2
It is worth pointing out that the preceding example allows us to deduce the noncommutative
weak-type (1, 1) version of the Burkholder/Rosenthal ([41, Theorem 3.1]) from the simpler weak-
type inequality involving square functions given in [40, Theorem 2.1].
The idea used in the proof of Proposition 4.9 can be extended for the case of general symmetric
spaces for the Banach space range. More precisely, we have the following result:
Proposition 4.11. Let E be a Banach function space. If E ∈ Int(L1 , Lq ) for 1 < q < 2, then
E ad (M; ℓc2 ) = E(⊕∞
n=1 Mn ) + E
cond,ad
(M; ℓc2 )
where E cond,ad (M; ℓc2 ) denotes is the subspace of E cond (M; ℓc2 ) consisting of adapted sequences.
Sketch of the proof. It is easy to see that under the assumption, E(⊕∞ ad c
n=1 Mn ) ⊆ E (M; ℓ2 ). On
the other hand, we have from Theorem 4.5 that E cond,ad c ad c
(M; ℓ2 ) ⊆ E (M; ℓ2 ). Thus, we only
need to verify one inequality.
The proof rests upon few facts. The interpolation space E is given by a K-method, the
simultaneous nature of Proposition 4.9 above, and Proposition 4.3.
Since E is given by a K-method, we may fix a Banach function space F such that for any given
semifinite von Neumann algebra N ,
C −1
a
E ≤
a
F ;K
≤
a
E(N )
, a ∈ E(N ).
F ;K
Under the assumption 1 < q < 2 and Proposition 4.3, we may also state that for every sequence
ξ ∈ E ad (M; ℓc2 ),
ξ
≈E
ξ
E ad (M;ℓc ) .
ad c ad c
L1 (M;ℓ2 );Lq (M;ℓ2 ) 2
F ;K
INTERPOLATION 35
and
J z (ν) , 2ν , Lcond
1 (M; ℓc2 ); Lcond
q (M; ℓc2 ) .q K ξ, 2ν , Lad c ad c
1 (M; ℓ2 ); Lq (M; ℓ2 ) .
Consider the following functions defined on the semi-axis (0, ∞):
f (t) = J(y (ν) , 2ν ) for t ∈ [2ν , 2ν+1 ),
From the equivalence of the J-methods and K-methods relative the function space F (see for
instance, [30, Theorem 2.9]) and the equivalence of norms stated at the beginning of the proof,
we may conclude that:
kykE(M⊗ℓ∞ ) + kzkE cond (M;ℓc2 ) .E kξkE ad (M;ℓc2 )
which is the desired inequality.
Remark 4.12. The argument in the proof of Theorem 4.11 can be carried out for the larger class
E ∈ Int(Lp , Lq ) with 2/3 < p < q < 2 to get that every ξ ∈ Lad c ad c
p (M; ℓ2 ) ∩ Lq (M; ℓ2 ) admits a
decomposition into two adapted sequences y and z satisfying:
kyk(Lp (M⊗ℓ∞ ),Lq (M⊗ℓ∞ )F ;J + kzk(Lcond
p (M;ℓc2 );Lcond
q (M;ℓc2 ))F ;J .E kξkE ad (M;ℓc2 ) .
However, we do not know if the equivalence of the J-method and the K-method relative to the
function space F is valid for the case of quasi-Banach couples.
As an immediate application of Proposition 4.11, we deduce the next result which partially
answers a problem from [43, Remark 3.11].
Corollary 4.13. Let E be a Banach function space. If E ∈ Int(L1 , Lq ) for 1 < q < 2, then
c
HE (M) = hdE (M) + hcE (M).
36 RANDRIANANTOANINA
Proof. We only need to verify the inclusion HE c (M) ⊆ hd (M) + hd (M). That is to show the
E E
c (M), the following holds:
existence of a constant αE such that for every x ∈ HE
n
o
inf
xd
d +
xc
c ≤ αE
x
c
hE hE HE
where the infimum is taken over all xd ∈ hdE (M) and xc ∈ hcE (M) such that x = xd + xc .
Let ξ = (dxn )n≥1 ∈ E ad (M; ℓc2 ). It is enough to take martingales xc and xd with for every
n ≥ 1,
dxcn = zn − En−1 (zn ) and dxdn = yn − En−1 (yn ),
where y and z are the adapted sequences from Theorem 4.11. Clearly, x = xd + xc . Since the
map (an )n≥1 7→ (En−1 (an ))n≥1 is a contraction in Lp (⊕∞n=1 Mn ) for every 1 ≤ p < ∞, it follows
by interpolation that it is also bounded in E(⊕∞n=1 M n ). In particular,
kxd khd .E kykE(⊕∞
n=1 Mn )
.
E
On the other hand, since for every n ≥ 1, En−1 |dxcn |2 ≤ En−1 |zn |2 , we immediately get that
kxc khcE ≤ kzkE cond (M;ℓc2 ) .
Combining these two inequalities, we arrive at
kxd khd + kxc khcE .E kξkE ad (M;ℓc2 ) = kxkHcE .
E
Acknowledgments. I am grateful to the anonymous referee for a careful reading of the paper
and for useful comments which improved the presentation of the paper.
References
[1] T. Bekjan, Z. Chen, M. Perrin, and Z. Yin, Atomic decomposition and interpolation for Hardy spaces of
noncommutative martingales, J. Funct. Anal. 258 (2010), no. 7, 2483–2505. MR 2584751 (2011d:46131)
[2] C. Bennett and R. Sharpley, Interpolation of operators, Academic Press Inc., Boston, MA, 1988. MR 89e:46001
[3] J. Bergh and J. Löfström, Interpolation spaces. An introduction, Springer-Verlag, Berlin, 1976, Grundlehren
der Mathematischen Wissenschaften, No. 223. MR MR0482275 (58 #2349)
[4] L. Cadilhac, Noncommutative Khintchine inequalities in interpolation spaces of Lp -spaces, Adv. Math. 352
(2019), 265–296. MR 3961739
INTERPOLATION 37
[5] , Majorization, interpolation and noncommutative Khinchin inequalities, Studia Math. 258 (2021),
no. 1, 1–26. MR 4214351
[6] L. Cadilhac and E. Ricard, Sums of free variables in fully symmetric spaces, arXiv: 1911.06180v3 [math.OA].
[7] Z. Chen, N. Randrianantoanina, and Q. Xu, Atomic decompositions for noncommutative martingales,
arXiv:2001.08775v1 [math.OA].
[8] I. Cuculescu, Martingales on von Neumann algebras, J. Multivariate Anal. 1 (1971), 17–27. MR 45 #4464
[9] M. Cwikel, M. Milman, and Y. Sagher, Complex interpolation of some quasi-Banach spaces, J. Funct. Anal.
65 (1986), no. 3, 339–347. MR 826431
[10] S. Dirksen, Noncommutative and vector-valued Rosenthal inequalities, Ph.D. Thesis dissertation (2011).
[11] , Noncommutative Boyd interpolation theorems, Trans. Amer. Math. Soc. 367 (2015), no. 6, 4079–4110.
MR 3324921
[12] T. Fack and H. Kosaki, Generalized s-numbers of τ -measurable operators, Pacific J. Math. 123 (1986), 269–300.
MR 87h:46122
[13] C. Fefferman and E. M. Stein, H p spaces of several variables, Acta Math. 129 (1972), no. 3-4, 137–193.
MR 447953
[14] A. M. Garsia, Martingale inequalities: Seminar notes on recent progress, W. A. Benjamin, Inc., Reading,
Mass.-London-Amsterdam, 1973, Mathematics Lecture Notes Series. MR 56 #6844
[15] P. Hitczenko and S. Montgomery-Smith, Tangent sequences in Orlicz and rearrangement invariant spaces,
Math. Proc. Cambridge Philos. Soc. 119 (1996), no. 1, 91–101. MR 1356161
[16] T. Holmstedt, Interpolation of quasi-normed spaces, Math. Scand. 26 (1970), 177–199. MR 54 #3440
[17] G. Hong, M. Junge, and J. Parcet, Algebraic Davis decomposition and asymmetric Doob inequalities, Comm.
Math. Phys. 346 (2016), no. 3, 995–1019. MR 3537343
[18] S. Janson and P. W. Jones, Interpolation between H p spaces: the complex method, J. Funct. Anal. 48 (1982),
no. 1, 58–80. MR 671315
[19] Y. Jiao, Martingale inequalities in noncommutative symmetric spaces, Arch. Math. (Basel) 98 (2012), no. 1,
87–97. MR 2885535
[20] Y. Jiao, N. Randrianantoanina, L. Wu, and D. Zhou, Square functions for noncommutative differentially
subordinate martingales, Comm. Math. Phys. 374 (2020), no. 2, 975–1019. MR 4072235
[21] Y. Jiao, F. Sukochev, and D. Zanin, Johnson-Schechtman and Khintchine inequalities in noncommutative
probability theory, J. Lond. Math. Soc. (2) 94 (2016), no. 1, 113–140. MR 3532166
[22] Y. Jiao, F. Sukochev, D. Zanin, and D. Zhou, Johnson–Schechtman inequalities for noncommutative martin-
gales, J. Funct. Anal. 272 (2017), no. 3, 976–1016. MR 3579131
[23] P. W. Jones, On interpolation between H 1 and H ∞ , Interpolation spaces and allied topics in analysis (Lund,
1983), Lecture Notes in Math., vol. 1070, Springer, Berlin, 1984, pp. 143–151. MR 760480
[24] M. Junge, Doob’s inequality for non-commutative martingales, J. Reine Angew. Math. 549 (2002), 149–190.
MR 2003k:46097
[25] M. Junge and M. Perrin, Theory of Hp -spaces for continuous filtrations in von Neumann algebras, Astérisque
(2014), no. 362, vi+134. MR 3241706
[26] M. Junge and Q. Xu, Noncommutative Burkholder/Rosenthal inequalities, Ann. Probab. 31 (2003), no. 2,
948–995. MR 2004f:46078
[27] N. Kalton and S. Montgomery-Smith, Interpolation of Banach spaces, Handbook of the geometry of Banach
spaces, Vol. 2, North-Holland, Amsterdam, 2003, pp. 1131–1175. MR 1 999 193
[28] N. J. Kalton and F. A. Sukochev, Symmetric norms and spaces of operators, J. Reine Angew. Math. 621
(2008), 81–121. MR 2431251 (2009i:46118)
[29] A. Kamińska, L. Maligranda, and L. E. Persson, Convexity, concavity, type and cotype of Lorentz spaces,
Indag. Math. (N.S.) 9 (1998), no. 3, 367–382. MR 1692169
[30] S. G. Kreı̆n, Yu. Ī. Petunı̄n, and E. M. Semënov, Interpolation of linear operators, Translations of Mathematical
Monographs, vol. 54, American Mathematical Society, Providence, R.I., 1982, Translated from the Russian by
J. SzHucs. MR 649411
[31] L. Maligranda, Orlicz spaces and interpolation, Seminários de Matemática [Seminars in Mathematics],
vol. 5, Universidade Estadual de Campinas, Departamento de Matemática, Campinas, 1989. MR 2264389
(2007e:46025)
[32] M. Musat, Interpolation between non-commutative BMO and non-commutative Lp -spaces, J. Funct. Anal. 202
(2003), no. 1, 195–225. MR 1994770
[33] M. Perrin, Inégalités de martingales non commutatives et applications, Ph.D. Thesis dissertation (2011).
[34] , A noncommutative Davis’ decomposition for martingales, J. Lond. Math. Soc. (2) 80 (2009), no. 3,
627–648. MR 2559120 (2011e:46104)
38 RANDRIANANTOANINA
[35] G. Pisier, Interpolation between H p spaces and noncommutative generalizations. I, Pacific J. Math. 155 (1992),
no. 2, 341–368. MR 1178030
[36] , Martingales in Banach spaces, Cambridge Studies in Advanced Mathematics, vol. 155, Cambridge
University Press, Cambridge, 2016. MR 3617459
[37] G. Pisier and Q. Xu, Non-commutative martingale inequalities, Comm. Math. Phys. 189 (1997), 667–698.
MR 98m:46079
[38] , Non-commutative Lp -spaces, Handbook of the geometry of Banach spaces, Vol. 2, North-Holland,
Amsterdam, 2003, pp. 1459–1517. MR 2004i:46095
[39] Y. Qiu, A non-commutative version of Lépingle-Yor martingale inequality, Statist. Probab. Lett. 91 (2014),
52–54. MR 3208115
[40] N. Randrianantoanina, Square function inequalities for non-commutative martingales, Israel J. Math. 140
(2004), 333–365. MR 2005c:46091
[41] , Conditioned square functions for noncommutative martingales, Ann. Probab. 35 (2007), no. 3, 1039–
1070. MR 2319715 (2009d:46112)
[42] N. Randrianantoanina and L. Wu, Martingale inequalities in noncommutative symmetric spaces, J. Funct.
Anal. 269 (2015), no. 7, 2222–2253. MR 3378874
[43] N. Randrianantoanina, L. Wu, and Q. Xu, Noncommutative Davis type decompositions and applications, J.
Lond. Math. Soc. (2) 99 (2019), no. 1, 97–126.
[44] N. Randrianantoanina, L. Wu, and D. Zhou, Atomic decompositions and asymmetric Doob inequalities in
noncommutative symmetric spaces, J. Funct. Anal. 280 (2021), no. 1, 64 pp. MR 4157678
[45] P. Turpin, Convexités dans les espaces vectoriels topologiques généraux, Dissertationes Math. (Rozprawy Mat.)
131 (1976), 221. MR 423044
[46] F. Weisz, Interpolation between martingale Hardy and BMO spaces, the real method, Bull. Sci. Math. 116
(1992), no. 2, 145–158. MR 1168308
[47] , Martingale Hardy spaces and their applications in Fourier analysis, Lecture Notes in Mathematics,
vol. 1568, Springer-Verlag, Berlin, 1994. MR MR1320508 (96m:60108)
[48] T. H. Wolff, A note on interpolation spaces, Harmonic analysis (Minneapolis, Minn., 1981), Lecture Notes in
Math., vol. 908, Springer, Berlin-New York, 1982, pp. 199–204. MR 654187
[49] Q. Xu, Applications du théorème de factorisation pour des fonctions à valeurs opérateurs, Studia Math. 95
(1990), no. 3, 273–292. MR 1060730
[50] , Analytic functions with values in lattices and symmetric spaces of measurable operators, Math. Proc.
Cambridge Philos. Soc. 109 (1991), 541–563. MR 92g:46036