Martingales Junge

a
r
X
i
v
:
m
a
t
h
/
0
2
0
6
0
6
2
v
1

[
m
a
t
h
.
O
A
]

6

J
u
n

2
0
0
2
DOOBS INEQUALITY FOR NON-COMMUTATIVE
MARTINGALES
MARIUS JUNGE
Abstract. Let 1 p < and (x

n
)
nN
be a sequence of positive elements in a non-
commutative L
p
space and (E
n
)
nN
be an increasing sequence of conditional expectations,
then
_
_
_
_
_
n
E
n
(x
n
)
_
_
_
_
_
p
c
p
_
_
_
_
_
n
x
n
_
_
_
_
_
p
.
This inequality is due to Burkholder, Davis and Gundy in the commutative case. By
duality, we obtain a version of Doobs maximal inequality for 1 < p .
Introduction:
Inspired by quantum mechanics and probability, non-commutative probability has be-
come an independent eld of mathematical research. We refer to P.A. Meyers exposition
[Me], the successive conferences on quantum probability [AvW], the lecture notes by Jajte
[Ja1, Ja2] on almost sure and uniform convergence and nally the work of Voiculescu,
Dykema, Nica [VDN] and of Biane, Speicher [BS] concerning the recent progress in free
probability and free Brownian motion. Doobs inequality is a classical tool in probability
and analysis. Transferring classical inequalities into the non-commutative setting theory
often requires an additional insight. Pisier, Xu [PX, Ps3] use functional analytic and com-
binatorial methods to establish the non-commutative versions of the Burkholder-Gundy
square function inequality. The absence of stopping time arguments, at least until the
time of this writing, imposes one of the main diculties in this recent branch of martin-
gale theory.
The formulation of Doobs inequality for non-commutative martingales faces the following
problem. For an increasing sequence of conditional expectations (E
n
)
nN
and a positive
operator x in L
p
, there is no reason why sup
n
E
n
(x) or sup
n
[E
n
(x)[ should be an element
in L
p
or represent a (possibly unbounded) operator at all. Using Pisiers non-commutative
vector-valued L
p
-space L
p
(N;
) we can overcome this problem and at least guess the right

formulation of Doobs inequality. However, Pisiers denition is restricted to von Neumann
algebras with a -weakly dense net of nite dimensional subalgebras, so-called hypernite
Junge is partially supported by the National Science Foundation.

1
2 MARIUS JUNGE
von Neumann algebras. But maximal inequalities are also interesting for free stochastic
processes where the underlying von Neumann algebra is genuinely not hypernite. All
these obstacles disappear for the so-called dual version of Doobs inequality: For every
sequence (x
n
)
nN
of positive operators
_
_
_
_
_
n
E
n
(x
n
)
_
_
_
_
_
p
c
p
_
_
_
_
_
n
x
n
_
_
_
_
_
p
.
(DD
p
)
In the commutative case this inequality is due to Burkholder, Davis and Gundy [BDG]
(even in the more general setting of Orlicz norms). Since it is crucial to understand
our approach to Doobs inequality, let us indicate the duality argument relating (DD
p
)
and Doobs inequality in the commutative case. Indeed, (DD
p
) implies that T(x
n
) =
n
E
n
(x
n
) denes a continuous linear map between L
p
(
1
) and L
p
. The norm of |T
|
yields the best constant in Doobs inequality for the conjugate index p
=
p
p1
.
_
_
_
_
sup
n
[E
n
(x)[
_
_
_
_
p
= |T
(x)|
L
p
()
|T| |x|
p
= c
p
|x|
p
.
Personally, I learned this argument after reading Dilworths paper [Di]. But I am sure it
is known to experts in the eld, see Garcias [Ga] for the general theory (and [AMS] for
the explicit equivalence). (DD
p
) admits an entirely elementary proof in the commutative
case (see again [AMS]). This elementary proof still works in the non-commutative case
for p = 2, see Lemma 3.1. It is the starting point of our investigation. We recommend
the reader (not familiar with modular theory) to start in section 3 where interpolation
is used to extend (DD
p
) to 1 p 2 and suitable norms are introduced to make the
above duality argument work in the non-commutative case. In section 4, we establish
the dual version (DD
p
) in the range 2 p < using duality arguments which rely on
Pisier/Xus version of Steins inequality in combination with techniques from Hilbert C
-
modules. By duality, we obtain the non-commutative Doob inequality in the more delicate
range 1 < p 2. The heart of our arguments rely on the (apparently new) connection
between Hilbert C
-modules and non-commutative L

p
spaces presented in section 2. These
duality techniques are necessary because p > 1 and 0 a b implies 0 a
p
b
p
only for
commuting operators. This is very often used in the elementary approach to commutative
martingales inequalities as in Garsias book [Ga].
Let us formulate our main results for nite von Neumann algebras. If : N C is
a normal, tracial state, i.e. (xy) = (yx), then the space L
p
(N, ) is dened by the
completion of N with respect to the norm
|x|
p
= ((x
x)
p
2
)
1
p
.
NON-COMMUTATIVE DOOB INEQUALITY 3
We refer to the rst section for more precise denitions and references. Given a subalgebra
M N, the embedding : L
1
(M, ) L
1
(N, ) is isometric because [x[ =
x M for
all x M. The dual map E =
: N M yields a conditional expectation satisfying

E(axb) = aE(x)b
for all a, b M and x N, see [Tk, Theorem 3.4.]. Since E is trace preserving, E extends
to a contraction E : L
p
(N, ) L
p
(N, ) with range L
p
(M, ). In the following, we con-
sider an increasing sequence (N
n
)
nN
N of von Neumann subalgebras with conditional
expectations (E
n
)
nN
. We recall that an element x is positive if it is of the form x = y
y.
Theorem 0.1. Let 1 p < , then there exists a constant c
p
depending only on p such
that for every sequence of positive elements (x
n
)
nN
L
p
(N, )
_
_
_
_
_
n
E
n
(x
n
)
_
_
_
_
_
p
c
p
_
_
_
_
_
n
x
n
_
_
_
_
_
p
.
(DD
p
)
Note the close relation to the non-commutative Stein inequality, see [PX, Theorem 2.3.],
_
_
_
_
_
n
E
n
(x
n
)
E
n
(x
n
)
_
_
_
_
_
p

2
2p
_
_
_
_
_
n
x
n
x
n
_
_
_
_
_
p
.
Using Kadisons inequality E
n
(x
n
)
E
n
(x
n
) E
n
(x
n
x
n
), it turns out that (DD
p
) is stronger
than Steins inequality. However, Steins inequality combined with the theory of Hilbert
C
-modules yields one of the fundamental inequalities in the proof of Theorem 0.1. Using a
Hahn-Banach separation argument `a la Grothendieck-Pietsch, we deduce Doobs maximal
inequality.
Theorem 0.2. [Doobs maximal inequality] Let 1 < p and x L
p
(N, ), then
there exist a, b L
2p
(N, ) and a sequence of contractions (y
n
) N such that
E
n
(x) = ay
n
b and |a|
2p
|b|
2p
c
p
|x|
p
Here c
p
is the constant in Theorem 0.1 for the conjugate index p
=
p
p1
. In particular,
for every positive x L
p
(N, ), there exists a positive b L
p
(N, ) such that
E
n
(x) b
for all n N.
In the case of hypernite von Neumann algebras this is equivalent to the corresponding
vector-valued inequality, see [Ps2], and therefore justies the name maximal inequality.
4 MARIUS JUNGE
In the semi-commutative case where N
n
= L
(,
n
, ) M and
n
are increasing -
algebras this inequality is stronger than the vector-valued Doob inequality, see Remark
5.5. In particular, this applies for random matrices. The inequality can be extended
to a continuous index set under suitable density assumptions, for examples for Cliord
martingales or free stochastical processes. Since these modications are rather obvious,
we omit the details.
Clearly, maximal inequalities immediately imply almost sure convergence. Therefore it is
not surprising that Theorem 0.2 implies almost uniform convergence convergence of the
martingale truncations (E
n
(x)) for p 2 and bilateral convergence for 1 < p 2 in
case of a tracial state. We refer to [Ja1, Ja2] for the denition of these notions and more
details. In the tracial case the bilateral convergence of the martingale truncations is known
by a result of Cuculescu [Cu] even for martingales in L
1
(N, ). Therefore Theorem 0.2
provides a alternative approach to these results but only for p > 1. However, the maximal
inequality discussed in [Ja1] cannot easily be interpolated to obtain Theorem 0.2 as in the
real case. In a subsequent paper [DJ2], we will apply the maximal inequality of Theorem
0.2 in Haagerup L
p
spaces in order to obtain (bilateral) almost sure convergence for all
states thus underlining the strength of these maximal inequalities.
Preliminary results and notation are contained in section 1. Section 5 contains immedi-
ate applications to submartingales and conditional expectations associated to actions of
groups.
I am indebted to Q. Xu for many discussion and support. I want to thank A. Defant for
the discussion about almost everywhere convergence. The knowledge of similar results for
almost sure convergence of unconditional sequences, see [DJ1], have been very encouraging.
I would like to thank Stanislaw Goldstein for initiating a correction in the proofs of Lemma
2.3 and Lemma 3.2.
1. Notation and preliminary results
As a shortcut, we use (x
n
), (x
nk
) instead of (x
n
)
nN
, (x
nk
)
n,kN
for sequences indexed by
the natural numbers N or its cartesian product N
2
. We use standard notation in operator
algebras, as in [Tk, KR]. In particular, B(H) denotes the algebra of bounded operators
on a Hilbert space H and /(H), / denote the subalgebra of compact operators on H,
2
,
respectively. The letters N, M will be used for von Neumann algebras, i.e. subalgebras
of some B(H) which are closed with respect to the -weak operator topology. We refer
to [Dx, Tk] for the dierent locally convex topologies relevant to operator algebras. For
n N we denote by M
n
(N) the von Neumann algebra of n n matrices with values in
N. We will briey use M
n
for M
n
(C). Given C
-algebras /, B, we denote by /B the

minimal tensor product. For von Neumann algebras N B(H
1
), M B(H
2
), we use
N

M for the closure of N M B(H
1
H
2
) in the -weak operator topology. Let us
recall that a von Neumann algebra is seminite if there exists a normal, seminite faithful
trace. A trace is a positive homogeneous, additive function on N
+
= x
x[ x N, the
cone of positive elements of N, such that for all increasing nets (x
i
)
i
with supremum in N
and for all x N
n) (sup
i
x
i
) = sup
i
(x
i
).
s) For every 0 < x there exists 0 < y < x such that 0 < (y) < .
f) (x) = 0 implies x = 0.
t) For all unitaries u N: (uxu
) = (x).
A positive homogeneous, additive function w : N
+
[0, ] satisfying n), s), f), but not
the last property t), is called a n.s.f. (normal, seminite, faithful) weight.
It will be worthwhile to clarify the dierent notions of non-commutative L
p
-spaces. If
is a trace then
m() =
_
n
i=1
y
i
x
i
n N,
n
i=1
[(y
i
y
i
) + (x
i
x
i
)] <
_
is the denition ideal on which there exists a unique linear extension : m() C
satisfying (xy) = (yx). The L
p
-(quasi)-norm is dened for x m() by
|x|
p
= ((x
x)
p
2
)
1
p
.
Then L
p
(N, ) is the completion of m() with respect to the L
p
-norm. (For p < 1 smaller
ideal is needed in order to guarantee that ([x[
p
) is nite.) We refer to [Ne, Te, KF, Ye] for
more on this and the fact that L
p
(N, ) can be realized as unbounded operators aliated
to N.
The starting point of Kosakis [Ko] denition of an L
p
-space is a normal faithful state
on a von Neumann algebra N. Then N acts on the Hilbert space L
2
(N, ) obtained by
completing N with respect to the norm
|x|
L
2
(N,)
= (x
x)
1
2
.
The modular operator is an (unbounded) operator obtained from the polar decomposi-
tion S = J
1
2
of the antilinear operator S(x) = x
on L
2
(N, ) , see [KR, Section 9.2.]. We
6 MARIUS JUNGE
denote by
t
: N N the modular automorphism group dened by
t
(x) =
it
x
it
.
Let us recall the standard notation
x.(y) = (xy) .
For each t there is a natural map I
t
: N N
(N
the unique predual of N) dened by

I
t
(x) =
t
(x). .
According to [Ko, Theorem 2.5.], there is a unique extension I
z
: N N
such that for

xed x the function f
x
: z [ 1 Im(z) 0 N
, f
x
(z) = I
z
(x) is analytic and
satises
f
x
(t)(y) = (y
t
(x)) and f
x
(i + t)(y) = (
t
(x)y) .
The density of the algebra of analytic elements shows that for 0 1 the map I
i
is
injective. By complex interpolation, the Banach space
L
p
(N, , ) = [I
i
(N), N
] 1
p
is dened by specifying |x|
0
=
_
_
I
1
i
(x)
_
_
N
and |x|
1
= |x|
N
. We refer to [Te, Fi] for
further information.
Haagerups abstract L
p
space [Ha1, Te] is dened for every von Neumann algebra N using
the crossed product N
w R with respect to the modular automorphism group of a n.s.f.

weight w. (In our applications, we can assume w = for a n.f. state.) If N acts faithfully
on a Hilbert space H, then the crossed product N
w R is dened as the von Neumann

algebra dened on L
2
(R, H) and generated by
(x)((t)) =
w
t
((t)) and (s)(t) = (t s) .
Then N
w R, see [PTA], is seminite and admits a unique trace such that the dual
action
s
(x) = W(s)xW(s)
satises (
s
(x)) = e
s
(x). Here W(s) is dened by the phase shift
W(s)(t) = e
ist
(t) .
The dual action satises
s
((x)) = (x) and moreover
(N) = x N
w R [
s
(x) = x , for all s R (1.1)
Let us agree to identify N with (N) in the following. L
p
(N) is dened to be the space
of unbounded, -measurable operators aliated to N
w R such that for all s R
s
(x) = e
s
p
x .
Note that the intersection L
p
(N) L
q
(N) is 0 for dierent values p ,= q. There is a
natural isomorphism between N
and L
1
(N) such that for every normal functional N
there is a unique a
L
1
(N) associated satisfying
(a
x) = (
_
R
s
(x))
for all positive x N
w
t
R. The key point in this construction is the denition of the
trace function tr : L
1
(N) C (corresponding to the integral in the commutative case)
given by
tr(a
) = (1) .
Let
1
p
+
1
p
= 1 and x L
p
(N), y L
p
(N). Then we have the trace property
tr(xy) = tr(yx) .
The polar decomposition x = u[x[ of x L
p
(N) satises u N and
|x|
p
= tr([x[
p
)
1
p
.
N acts as a left and right module on L
p
(N) and more generally Holders inequality
|xy|
r
|x|
p
|y|
q
(1.2)
holds whenever
1
p
+
1
q
=
1
r
. As for seminite von Neumann algebras, there is a positive
cone L
p
(N)
+
in L
p
(N) consisting of elements in L
p
(N) which are positive as unbounded
operators aliated to N
w R. Following [Te, Proposition 33, Theorem 32], we deduce

for 0 x y L
p
(N) and
1
p
+
1
p
= 1
|x|
p
= sup
zL
p
(N)
+
,z
p
1
tr(zx) sup
zL
p
(N)
+
,z
p
1
tr(zy) = |y|
p
. (1.3)
In the sequel, we will often use the following simple observation.
Lemma 1.1. Let 0 < p and x, y L
p
(N) such that x
x y
y. If p is the left
support projection of y, then xy
1
p is a well-dened element in N of norm less than one.
Proof: We note that a = xy
1
p is aliated with N
w R and
p(y
1
)
xy
1
p p(y
1
)
yy
1
p p
8 MARIUS JUNGE
shows that a is a contraction and in particular -measurable. Moreover, we have
p =
s
(p) =
s
(yy
1
p) = e
s
p
y
s
(y
1
p)
and hence
s
(y
1
p) = e
s
p
y
1
p . (1.4)
Therefore, the equality
s
(a) =
s
(x)
s
(y
1
p) = e
s
p
xe
s
p
y
1
p = a
shows with (1.1) that a is a contraction in N.
In the -nite case, Kosakis L
p
-space is isomorphic to Haagerups L
p
space, see [Ko,
section 8]. Indeed, given a n. f. state with corresponding density D in L
1
(N)

= N
,
then
t
(x) = D
it
xD
it
and for all x N
|I
i
(x)|
[I
i
(N),N]
1
p
=
_
_
_D
p
xD
1
p
_
_
_
Lp(N)
.
We recall that an element N is analytic, if t
t
(x) extends to an analytic function with
values in N. The

-closed subalgebra of analytic elements will be denoted by /. The
following Lemma is probably well-known, see [Ko], [JX, Lemma 1.1.]. We add a short
proof for the convenience of the reader.
Lemma 1.2. Let 0 < p < , then D
1
2p
/
+
D
1
2p
is dense in L
p
(N)
+
and ND
1
p
is dense in
L
p
(N) and for 1 p the map J
p
: L
p
(N) L
1
(N), J
p
(x) = xD
1
1
p
is injective.
Proof: D is a -measurable operator with support projection 1. Therefore, xD
1
1
p
= 0
implies that x = xD
1
1
p
D
1
p
1
is a well-dened -measurable operator and equals 0. Hence,
J
p
is injective for 1 p . Let
1
p
+
1
p
= 1. We show the density of D

1
2p
/
+
D
1
2p
in L
p
(N)
+
for 1 p < . If this is not the case, the Hahn-Banach theorem implies the existence
of x L
p
(N)
+
and y L
p
(N)
sa
such that tr(yD
1
2p
aD
1
2p
) = tr(D
1
2p
yD
1
2p
a) 0 for all
a /
+
and tr(xy) > 0. Using the -strong density of /
+
in N
+
, see [PTA], we deduce that
y = D
1
2p
D
1
2p
yD
1
2p
D
1
2p
is negative and hence tr(xy) 0, a contradiction. Let
1
2
p 1
and y L
p
(N)
+
, then we can approximate y
1
2
by an element x = D
1
4p
aD
1
4p
, a /
+
and
hence Holders inequality implies that x
2
y = x(x y
1
2
) + (x y
1
2
)y
1
2
approximates y.
Since a is analytic, we observe that x
2
= D
1
4p
aD
1
4p
D
1
4p
aD
1
4p
= D
1
2p
i
4p
(a)
i
4p
(a)D
1
2p
is in D
1
2p
/
+
D
1
2p
. By induction, we deduce the density of D
1
2p
/
+
D
1
2p
in L
p
(N)
+
for all
0 < p < . Since /D
1
p
= D
1
2p
/D
1
2p
and every element is a linear combination of 4
positive elements, we obtain the assertion.
2. Hilbert C
-modules and L
p
-spaces
In this section, we will analyze some Banach spaces related to the dual version of Doobs in-
equality and combine Kasparovs stabilisation theorem for Hilbert C
-modules with Steins

inequality for non-commutative martingales proved by Pisier and Xu [PX]. Let N M
be a von Neumann subalgebra and E : M M be a normal conditional expectation
onto N. E is normal if the dual map E
satises E
(M
) M
and hence has a predual

map E
: M
such that E = (E
. In order to simplify the exposition, we will

assume in addition that : M C is a normal faithful state satisfying = [
N
E.
Using Kosakis interpolation spaces [Ko, Proposition 4.1], it is very easy to check that E
extends to certain L
p
-spaces. However, in our context it is more convenient to work with
the Haagerup L
p
-spaces.
Lemma 2.1. Under the previous assumptions let D be the density of in L
1
(M) and
t
be the modular group, then E
(D) = D, and for all x M, y M
E(
t
(x)) =
t
(E(x)) and tr(E
(y)) = tr(y) .
Proof: The rst assertion follows from
tr(Dx) = (x) = (E(x)) = tr(DE(x)) = tr(E
(D)x)
which is valid for all x M. For E(
t
(x)) =
t
(E(x)) see [C, Lemme 1.4.3]. In order to
prove the third assertion let M
and a
be the corresponding density then

tr(E
(a
)) = E(1) = (E(1)) = (1) = tr(a
) .
We will need several approximation results.
Lemma 2.2. Let / be a

-closed (not necessarily norm closed) but -strongly dense sub-
algebra of M. For 0 < p < the space /D
1
p
is norm dense in L
p
(N).
10 MARIUS JUNGE
Proof: According to Lemma 1.2, MD
1
p
is norm dense in L
p
(M). According to Kaplanskys
density theorem [Tk, Theorem II.4.8], every element in the unit ball of M is the strong
-
limit of elements in the unit ball of /. An application of the following Lemma 2.3 yields
the assertion.
Lemma 2.3. Let 0 < p < , M be a von Neumann algebra, a L
1
(M)
+
,(x
N be
a bounded net and x N. If x
a
1
p
converges to xa
1
p
in norm for some 1 p < , then
this is true for all 0 < p < . In particular, if (x
N converges to x strongly, then

for all b L
p
(N), x
b converges to xb in norm.
Proof: We consider the set I =
_
0 < p <
lim
a
1
p
= xa
1
p
_
. Here, we refer to
convergence in norm. Let us rst observe that 0 < q < p I implies q I. Indeed,
according to Holders inequality (1.2), the linear map M
r
: L
p
(N) L
q
(N) dened
by M
l
(x) = xa
1
q
1
p
is continuous and therefore, the convergence of x
a
1
p
implies the
convergence of x
a
1
q
= M
r
(x
a
1
p
). Let us now show that 1 p I implies 2p I, i.e.
_
_
_(x
x)a
1
2p
_
_
_
2p
= 0.
We note that
_
_
_(x
x)a
1
2p
_
_
_
2
2p
=
_
_
_a
1
2p
(x
)(x
x)a
1
2p
_
_
_
p
and dene the analytic function f : 0 Im(z) 1 L
p
(N) given by
f
(z) = exp((z
1
2
)
2
) a
z
p
(x
)(x
x)a
1z
p
.
Since a
it
xa
it
L
p
(N) for every x L
p
(N) this is well dened and analytic in the interior.
For z = 1 +it, we get
|f
(1 +it)|
p
exp(
1
4
)
_
_
_a
1+it
p
(x
)(x
x)a
it
p
_
_
_
p
exp(
1
4
)
_
_
_a
1
p
(x
)
_
_
_
p
|x
x|
exp(
1
4
)
_
_
_(x
x)a
1
p
_
_
_
p
sup
|x
x|
.
Similarly,
|f
(it)|
p
exp(
1
4
)
_
_
_a
it
p
(x
)(x
x)a
1it
p
_
_
_
p
exp(
1
4
) sup
|x
_
_
_(x
x)a
1
p
_
_
_
p
.
Therefore, p I implies with the three line Lemma
lim
_
_
_a
1
2p
(x
)(x
x)a
1
2p
_
_
_
p
= lim
_
_
_
_
f
(
1
2
)
_
_
_
_
p
lim
sup
t
max|f
(it)|
p
, |f
(1 +it)|
p
exp(
1
4
) sup
|x
x|
lim
_
_
_(x
x)a
1
p
_
_
_
p
= 0 .
Hence, if I [1, ) is non empty, then I = (0, ) and the rst assertion is proved. Let
b L
p
(N). We use the polar decomposition of b
to nd a partial isometry u such that

b = [b
[u. Then a = [b
[
p
is in L
1
(N)
+
. If x
converges to x strongly then x
a
1
2
converges
in norm to xa
1
2
. This means 2 I and hence I = (0, ). In particular, p I and hence
x
[b
[ = x
a
1
p
converges to x[b
[. This immediately implies that x
b = x
[b
[u converges
to xb = x[b
[u.
In this paper, we will frequently use the space L
p
(M;
C
2
) L
p
(B(
2
)
M) of column
matrices with values in L
p
(M).
Corollary 2.4. Let 0 < p < and / -strongly dense

-closed subalgebra of M, then
the space of nite sequences (a
n
D
1
p
)
n
such that a
n
/ is dense in L
p
(M;
C
2
).
Proof: Let p
n
be the orthorgonal projection onto the rst n unit vectors in
2
, then p
n
1
converges to 1 1 in the strong
topology. Hence, the space of nite sequences is dense

in L
p
(M;
C
2
). We can apply Lemma 2.2 for each coordinate to obtain the assertion.
In the next step we refer to [JX, Proposition 2.3] to understand how a state preserving
conditional expectation extends to a contraction E
p
on L
p
(M) and in fact E
= E
1
. We
will later often use these facts without further reference and drop the index p or

because
it is easily determined by the context.
Proposition 2.5. Let 1 p , then there is a contraction E
p
: L
p
(M) L
p
(M) such
that for all x M and 0 1
E
p
(D
1
p
xD
p
) = D
1
p
E(x)D
p
E
p
satises E
p
(x
) = E
p
(x) and E
p
(x) is positive for all positive x. Moreover, if
1
s
=
1
p
+
1
q
+
1
r
1, a L
p
(N), b L
q
(N) and x L
r
(M), then
E
s
(axb) = aE
r
(x)b .
12 MARIUS JUNGE
Conditional expectations (or, more generally, positive operator-valued weights) are closely
connected to Hilbert C
-modules. An excellent reference for the few facts and the notation
we need in this paper is Lances book [La]. We recall that a Hilbert C
-module /
over a C
-algebra / is a right /-module / together with an /-valued sesquilinear form

, ) : // / such that for all a / and x, y /
x, ya) = x, y)a ,
x, y)
= y, x) .
The norm in / is given by |x|
M
= |x, x)|
1
2
A
. We should note that the sesquilinear form
is assumed to be linear in its second and antilinear in its rst component. The standard
example of a Hilbert C
-module over a C
-algebra / is the space H

A
, the closure of nite
sequences (x
n
) / with respect to the norm
|(x
n
)| =
_
_
_
_
_
n
x
n
x
n
_
_
_
_
_
1
2
A
.
We can identify H
A
with the space of column matrices H
A
/ /. In the context of
von Neumann algebras / = N, we derive a similar example using the space of column
matrices C(N) B(
2
)
N which consists of sequences (x

n
) such that
n
x
n
x
n
converges
in the -weak operator topology. Then the sesquilinear form
(x
n
), (y
n
)) = (x
1
, x
2
, )(y
1
, y
n
, )
n
x
n
y
n
converges in the -weak operator topology and satises the axioms of a Hilbert C
-module.
Let us indicate how this concept applies to non-commutative L
p
-spaces. Indeed, Ses :
L
p
(N;
C
2
) L
q
(N;
C
2
) L
r
(N) dened by
Ses((x
n
), (y
n
)) =
n
x
n
y
n
, where
1
r
=
1
p
+
1
q
is still a sesquilinear form. We will omit Ses and simply write (x
n
)
(y
n
) to remind the
natural matrix multiplication. Holders inequality immediately implies
_
_
_
_
_
n
x
n
y
n
_
_
_
_
_
r
_
_
_
_
_
(
n
x
n
x
n
)
1
2
_
_
_
_
_
p
_
_
_
_
_
(
n
y
n
y
n
)
1
2
_
_
_
_
_
q
(2.1)
whenever
1
r
=
1
p
+
1
q
. Our aim is to extend this type of inequalities to the sesquilinear form
x, y) = E(x
y)
given by a conditional expectation E : M N. We denote by L
(M, E) the completion

of M with respect to the norm
|x|
L(M,E)
= |E(x
x)|
1
2
.
In analogy with the case p = , we introduce the following notation:
Denition 2.6. Let 0 < p and x = aD
1
p
MD
1
p
, then
|x|
Lp(M,E)
=
_
_
_D
1
p
E(a
a)D
1
p
_
_
_
1
2
p
2
.
The completion with respect to this norm is denoted by L
p
(M, E).
Remark 2.7. Let us note that for p 2, the precaution x MD
1
p
is unnecessary because
for all x L
p
(N) the conditional expectation E(x
x) is well-dened and
|x|
Lp(N,E)
= |E(x
x)|
1
2
p
2
|x
x|
1
2
p
2
= |x|
p
.
By norm density of MD
1
p
in L
p
(M), we obtain the same closure. However, for p < 2 we
no longer dispose of the continuity of E on L
p
2
(N) (even in the commutative case). This
justies our slightly articial denition.
Our next step will be to prove the triangle inequality for p 1 using Kasparovs stabili-
sation result.
Proposition 2.8. Let N M, E and as above and assume in addition that M
is sep-
arable. For all 0 < p , there exists an isometric right N-module map u
p
: L
p
(M, E)
L
p
(N;
C
2
). Moreover,
i) For all x L
p
(M, E) and y L
q
(M, E)
u
p
(x)
u
q
(y) = E(x
y) .
ii) Let
1
r
=
1
q
+
1
p
. For all x L
p
(M, E) and y L
q
(M, E)
|E(x
y)|
r
|x|
Lp(M,E)
|y|
Lq(M,E)
.
iii) If 0 < p < there exists a contractive projection Q
p
onto the image of u
p
such that
for all z L
p
(N;
C
2
)
Q
p
(z)
Q
p
(z) z
z .
14 MARIUS JUNGE
iv) If 1 < p , p
< and
1
p
+
1
p
= 1, then
Q
p
= Q
p
.
In particular, L
p
(M, E) is a normed space for 1 p < and p-normed for 0 < p 1.
Proof: Let X M be a -strongly dense separable C
-algebra and let F be the Hilbert

C
-module generated by N and X. Then F is a countably generated Hilbert N-module

and according to [La, Theorem 6.2.] there exists a unitary w : F H
N
H
N
such that
w(f, x)
w(g, y) = f, g) + x
y .
Let

Q : F H
N
F be the projection onto the rst coordinate. This projection carries
over to Q = w

Qw
1
: H
N
w(F). We dene u : F H
N
to be the restriction of w to
the rst component. Clearly, u preserves the N-valued sesquilinear form. For 0 < p ,
we dene the map u
p
: FD
1
p
L
p
(N;
C
2
) by
u
p
(aD
1
p
) = u(a)D
1
p
.
Then for all a, b F
D
1
p
E(a
b)D
1
q
= D
1
p
u(a)
u(b)D
1
q
= u
p
(aD
1
p
)
u
q
(bD
1
q
) .
This justies i) for the subsets FD
1
p
, FD
1
q
, respectively. In particular, we obtain
_
_
_aD
1
p
_
_
_
2
Lp(M,E)
=
_
_
_D
1
p
E(a
a)D
1
p
_
_
_
p
2
=
_
_
_u
p
(aD
1
p
)
u
p
(aD
1
p
)
_
_
_
p
2
=
_
_
_u
p
(xD
1
p
)
_
_
_
2
p
which implies that u
p
is isometric when restricted to FD
1
p
. According to the following
Lemma 2.10 and Proposition 2.9, FD
1
p
is dense and hence i) follows. Assertion ii) is an
immediate consequence of (2.1). Indeed,
_
_
_D
1
p
E(a
b)D
1
q
_
_
_
r
=
_
_
_u
p
(aD
1
p
)
u
q
(bD
1
q
)
_
_
_
r
_
_
_u
p
(aD
1
p
)
_
_
_
p
_
_
_u
q
(bD
1
q
)
_
_
_
q
=
_
_
_aD
1
p
_
_
_
Lp(M,E)
_
_
_bD
1
p
_
_
_
Lq(M,E)
.
Since L
p
(N;
C
2
) is a Banach space, respectively a complete p-normed space, and u
p
is
isometric, we deduce the last assertion. The projection Q
p
: L
p
(N;
C
2
) L
p
(N;
C
2
) is
densely dened for a sequence (z
n
D
1
p
) such that z
n
= 0 for n n
0
by
Q
p
[(z
n
D
1
p
)] = Q[(z
n
)]D
1
p
.
(Here and in the following we will use the brackets [ ] to indicate that Q is applied to the
sequence (z
n
).) Since w is a unitary in the sense of Hilbert N-modules, we deduce for
w
1
[(z
n
)] = f + (a
n
)
Q[(z
n
)]
Q[(z
n
)] =
Q(f + (a
n
)),

Q(f + (a
n
))) = f, f) f, f) +
n
a
n
a
n
= f + (a
n
), f + (a
n
)) = w(f + (a
n
)), w(f + (a
n
))) =
n
z
n
z
n
.
In particular,
_
_
_Q
p
[(z
n
D
1
p
)]
_
_
_
2
p
=
_
_
_D
1
p
Q[(z
n
)]
Q[(z
n
)]D
1
p
_
_
_
p
2
_
_
_
_
_
D
1
p
n
z
n
z
n
D
1
p
_
_
_
_
_
p
2
=
_
_
_(z
n
D
1
p
)
_
_
_
2
Lp(N;
C
2
)
.
Hence, Q
p
extends by density, see Corollary 2.4, to a contraction on L
p
(N;
C
2
). This proves
iii). In order to prove iv) it suces again by density to consider nite sequences z =
(z
n
D
1
p
) and z = ( z
n
D
1
p
) with z
n
, z
n
F. Using again w
1
[(z
n
)] = f + (a
n
), w
1
[( z
n
)] =
f + ( a
n
), we deduce from the fact that w is a unitary module map
tr(Q
p
(z)
z) = tr(D
1
p
Q[(z
n
)]
( z
n
)D
1
p
)
= (Q[(z
n
)], ( z
n
)))
= (
Q(f + (y
n
)),

f + ( a
n
)))
= (E(f

f))
= (f + (a
n
),

Q(

f + ( a
n
)) ))
= tr(z
Q
p
( z)) .
Let us briey clarify how the unitary module map w : F H
N
H
N
extends to an
isometry onto C(N). We recall that M acts naturally on L
2
(M) and that p
E
= E
2
:
L
2
(M) L
2
(M) is the orthogonal projection onto L
2
(N). Hence, we obtain an isometric
isomorphism
L
(M, E) = Mp
E
B(L
2
(N), L
2
(M)) B(L
2
(M)) .
Note that in general p
E
is not in M. Via this inclusion, we dispose of all the relevant
locally convex operator topologies on L
(M, E) and on
L
(M, E) C(N) B(L

2
(N), L
2
(M)
2
(L
2
(N))) .
We denote by L
st
(M, E) the closure with respect to the -strong topology. The following
example due to O. Ramcke shows that this might be dierent from L
(M, E). Let M =

16 MARIUS JUNGE
L
([0, 1]
2
) and E(f)(t, s) =
_
1
0
f(t, r)dr. Then, the norm on L
(M, E) is the norm in

L
(L
2
). Given disjoint sets I
k
of measure 2
k
the function
f =
k
1
I
k
2
k
2
1
I
k
yields an element in -strong closure not belonging to L
(M, E). The next proposition

shows that L
st
(M, E) is isomorphic to a complemented module in C(N). In particular,

we will nd a sequence (x
j
) L
st
(M, E) and completely contractive module maps u

j
:
M N such that for all x M
x =
j
x
j
u
j
(x)
For inclusions with nite index, a nite sum of this form suces and more can be said
about the coecients E(x
i
x
i
), see [PP]. In any case
1 = E(1) =
j
u
j
(1)
E(x
j
x
j
)u
j
(1)
shows that we are dealing with a partition of unity.
Proposition 2.9. Under the above assumptions, w extends to an isometric isomorphism
between the space L
st
(M, E)C(N) and C(N). Moreover, for every element x in the unit
ball of L
(M, E) there exists a net x
in the unit ball of L
(M, E) F converging to x
in the strong
topology and Q extends to a projection onto the image of L

st
(M, E).
Proof: We note that for h L
2
(N) and x F C(N)
|w(x)(h)|
2
= (w(x)
w(x)(h), h) = (x
x(h), h) = |x(h)|
2
2
.
Hence, w preserves the strong and -strong topology. Therefore, we obtain a natural
extension, also denoted by w, to the -strong closure L
st
(M, E) C(N) of F H
N
with
values in C(N). To show that w, w
1
remain contractions, it suces to show that every
element in the unit ball of L
st
(M, E) C(N), C(N), can be approximated with respect

to the strong topology by elements in the unit ball of F H
N
, H
N
, respectively. This
follows immediately from Kaplanskys density theorem [Tk, Theorem II.4.8] for C(N). To
obtain the assertion for the -strong closure of Mp
E
, we note that F is -strong
dense in
M, hence Fp
E
is -strong
dense in L
st
(M, E). Then, we follow the proof of Kaplanskys

density theorem, see [Tk, Theorem II.4.8], and note that the function f : F B(L
2
(N))
dened there by
f(xp
E
) = 2xp
E
(1 + (p
E
x
xp
E
))
1
= 2xp
E
(1 + (E(x
x))
1
FN F
satises f(F) F. Hence, the proof of Kaplanskys density theorem applies and therefore
every element in the unit ball of
L
st
(M, E) = Mp
E
strong
= Mp
E
strong
([Tk, Theorem 2.6]) can be approximated by a net f(x
) in the unit ball of F. Since

Q = w
1

Qw corresponds to the projection onto the rst component in L
st
(M, E) C(N),
the last assertion follows easily.
Lemma 2.10. Under the previous assumptions FD
1
p
is dense in L
p
(M, E). The Cauchy-
Schwarz inequality 2.8 ii) and i) also holds for p = or q = .
Proof: Let us assume 2 p < rst. Using Kaplanskys density theorem [Tk, Theorem
II.4.8], the unit ball of F is strong
-dense in the unit ball of M. Using Lemma 2.3, we

deduce that FD
1
p
is dense in L
p
(M). As observed in Remark 2.7, the inclusion L
p
(M)
L
p
(M, E) is contractive and dense, hence the assertion follows. For p < 2, we observe
that the inclusion i : L
2
(M, E) L
p
(M, E), i(aD
1
2
) = aD
1
p
is contractive using Holders
inequality
_
_
_aD
1
p
_
_
_
p
=
_
_
_D
1
p
E(a
a)D
1
p
_
_
_
1
2
p
2
_
_
_D
1
2
E(a
a)D
1
2
_
_
_
1
2
1
=
_
_
_aD
1
2
_
_
_
L
2
(M,E)
.
By denition of L
p
(M, E), the image of i is dense. Hence, FD
1
p
= i(FD
1
2
) is dense in
L
p
(M, E). To prove the Cauchy-Schwarz inequality if p = r < and q = , we x
x = bD
1
p
and y L
st
(M, E) of norm less than one. Then there exists a net a
p
E
in the
unit ball of F such that a
converges to y with respect to the strong
topology. Then the

strong
convergence of b
p
E
implies the strong
convergence of E(b
) = p
E
b
p
E
.
Therefore the norm convergence of D
1
p
E(b
) to D
1
p
E(b
y) follows from Lemma 2.3. In

particular according to Proposition 2.15 ii),
_
_
_D
1
p
E(b
y)
_
_
_
p
= lim
_
_
_D
1
p
E(b
)
_
_
_
p
lim
_
_
_D
1
p
E(b
b)D
1
p
_
_
_
1
2
p
2
|E(a
)|
1
2
_
_
_bD
1
p
_
_
_
Lp(M,E)
.
18 MARIUS JUNGE
Thus, by density, every norm one element in L
st
(M, E) induces a contractive multiplier

on L
p
(M, E). The case p = and q < is similar. The case p = and q = is
classical and follows from the fact that E : M B(L
2
(M)) is completely positive and
therefore admits a dilation E(x) = v(x)v
for a contraction v and a

-representation .
Hence, we obtain
|E(x
y)|
B(H)
= |v(x
y)v
| |v(x
)(x)v
|
1
2
|v(y
)(y)v
|
1
2
= |E(x
x)|
1
2
|E(y
y)|
1
2
.
Now we turn our attention to the duality between L
p
(N, E) and L
p
(N, E). Let us point
out that we use the antilinear duality bracket
(x, y) = tr(x
y)
between L
p
(N;
C
2
) and L
p
(N;
C
2
). The following Lemma is a standard application of the
Hahn-Banach theorem.
Lemma 2.11. Let 1 < p < and
1
p
+
1
p
, X
p
L
p
(M;
C
2
) and X
p
L
p
(M;
C
2
) be
subspaces and Q
p
: L
p
(M;
C
2
) L
p
(M;
C
2
) be a projection onto X
p
such that Q
p
is a
projection onto X
p
, then
X
p
= X
p
.
and for every dense subset X X
p
and x X
p
|x| |Q| sup
_
tr(y
x)
y X, |y|
X
p
< 1
_
.
Proof: Since (L
p
(M;
C
2
))
= L
p
(M;
C
2
), the antilinear map : X
p
X
p
(y)(x) = tr(y
x)
is obviously contractive. Let f : X C be a norm one functional, then we can apply the
Hahn-Banach extension theorem and (L
p
(M;
C
2
))
= L
p
(M;
C
2
) to obtain z L
p
(M;
C
2
)
such that
f(x) = tr(z
x) = (z, x) .
Clearly,
(z, x) = (z, Q
p
(x)) = (Q
p
(z), x) = tr(Q
p
(z)
x) .
(Dened in this way Q
p
is linear.) Since, Q
p
(z) X
p
we deduce that is surjective. The
last formula follows from the Hahn-Banach theorem, i.e.
|x| = sup
f1
[f(x)[ = sup
z1
[tr(z
x)[ = sup
z1
[tr(Q
p
(z)
x)[
|Q
p
| sup
y1,yX
p
[tr(y
x)[
The supremum is unchanged if restricted to a dense subset.
Corollary 2.12. Let 1 < p, p
< ,
1
p
+
1
p
= 1 and M
be separable, then L
p
(M, E)
=
L
p
(M, E) holds isometrically.
In the last part of this section, we investigate the space L
p
(M, (E
n
);
C
2
), a generalization
of L
p
(M, E), which is important for the dual version of Doobs inequality. We feel that
the space L
p
(M, E) is easier to understand and more directly connected to Hilbert C
-
modules. Let (E
n
) be a sequence of conditional expectations E
n
: M M onto von
Neumann subalgebras N
n
such that E
n
= for all n N. Pisier, Xu proved a non-
commutative version of Steins inequality. To formulate the version we need here, we
consider the subspace L
cond
p
(M;
C
2
) L
p
(M;
C
2
(N
2
)) of double indexed sequences (x
nk
)
such that x
nk
L
p
(N
n
) for all k N. We refer to [PX, JX] for the proof of the following
theorem.
Theorem 2.13 (Steins inequality). Let 1 < p , p
< such that

1
p
+
1
p
= 1 and (E
n
)
be a sequence of conditional expectations such that E
n
E
m
= E
min{n,m}
. Then the linear
map ST
p
: L
p
(M;
C
2
(N
2
)) L
cond
p
(M;
C
2
) dened by ST
p
[(x
nk
)] = (E
n
(x
nk
)) is a bounded
projection onto L
cond
p
(M;
C
2
) and satises ST
p
= ST
p
. In particular,
L
cond
p
(M;
C
2
)
= L
cond
p
(M;
C
2
)
with equivalent norms depending only on p.
Denition 2.14. Let 0 < p , the space L
p
(M, (E
n
);
C
2
) is the completion of the space
of sequences (a
n
D
1
p
), a
n
M with respect to the norm
_
_
_(a
n
D
1
p
)
_
_
_
Lp(M,(En);
C
2
)
=
_
_
_
_
_
D
1
p
n
E
n
(a
n
a
n
)D
1
p
_
_
_
_
_
1
2
p
2
.
20 MARIUS JUNGE
Proposition 2.15. Let 0 < p and in addition M
separable, then there is an isomet-

ric embedding u
p
: L
p
(M, (E
n
);
C
2
) L
cond
p
(M;
C
2
) and for p < a norm one projection
R
p
onto the image of u
p
. Moreover,
i) if
1
p
+
1
q
=
1
r
, then for all (x
n
) L
p
(M, (E
n
);
C
2
) and (y
n
) L
q
(M, (E
n
);
C
2
)
u
p
(x
n
)
u
q
(y
n
) =
n
E
n
(x
n
y
n
)
and
_
_
_
_
_
n
E
n
(y
n
x
n
)
_
_
_
_
_
r
|(y
n
)|
Lq(M,(En);
C
2
)
|(x
n
)|
Lp(M,(En);
C
2
)
.
ii) if 1 < p, p
< such that

1
p
+
1
p
= 1 and ST
p
, ST
p
denote the projection onto
L
cond
p
(M;
C
2
), L
cond
p
(M;
C
2
), respectively, then
(R
p
ST
p
)
= R
p
ST
p
.
Proof: The proof of the Cauchy-Schwarz inequality for p = q = and the isometric
embedding is similar to the last part of the proof in Lemma 2.10. Since we will not need
it, we omit the details. For each n N, we x the isometric isomorphism u
n
p
: L
p
(M, E
n
)
L
p
(N
n
;
C
2
) and dene u
p
: L
p
(M, (E
n
);
C
2
) L
cond
p
(M;
C
2
) by
u
p
[(x
n
)] = (u
n
p
(x
n
)
kn
) ,
i.e. we apply u
n
p
to x
n
and obtain a double indexed sequence. For nite sequences, we
apply Proposition 2.8 i) and obtain
u
p
[(x
n
)]
u
q
[(y
n
)] =
n
u
n
p
(x
n
)
u
n
q
(x
n
) =
n
E
n
(x
n
y
n
) . (2.2)
In particular, u
p
is isometric when restricted to nite sequences. For p < an easy Cauchy
sequence argument implies that u
p
extends isometrically to L
p
(M, (E
n
);
C
2
). Similarly as
in the proof of Proposition 2.8, we deduce from (2.2) the Cauchy-Schwarz inequality. If
Q
n
p
: L
p
(N
n
;
C
2
) L
p
(N
n
;
C
2
) denotes the projection onto the image of u
n
p
, then
R
p
[(x
nk
)] = (Q
n
p
[(x
nk
)
k
]
kn
)
is certainly well-dened for nite sequences. However, we deduce from Proposition 2.8 iii)
R
p
[(x
nk
)]
R
p
[(x
nk
)] =
k
Q
n
p
[(x
nk
)
k
]
k
Q
n
p
[(x
nk
)
k
]
k

k
x
nk
x
nk
and therefore R
p
extends to a contraction for all p < . Finally, we observe that Propo-
sition 2.8 iii) implies assertion ii):
tr(Q
p
[(E
n
(x
nk
))]
(y
nk
)) =
k
tr(Q
n
p
[(E
n
(x
nk
))
k
]
k
y
nk
)
=
k
tr(Q
n
p
[(E
n
(x
nk
))
k
]
k
E
n
(y
nk
))
=
k
tr(E
n
(x
nk
)
Q
n
p
[(E
n
(y
nk
))
k
]
k
)
=
nk
tr(x
nk
Q
n
p
[(E
n
(y
nk
))
k
]
k
)
= tr((x
nk
)
Q
p
[(E
n
(y
nk
))]) .
We want to remove the additional assumption that M has separable predual.
Lemma 2.16. Let M be a von Neumann algebra, a normal faithful state, E
n
a sequence
of normal conditional expectations onto von Neumann subalgebras N
n
. For every separable
subalgebra A M there exists a subalgebra A B M with separable dual, for all n N
E
n
(B) B, and a normal conditional expectation

E : M B such that

E = .
Proof: The proof is a modication of [Ki, Corollary 3.5]. Let A
1
be separable,

-closed
(but not necessarily norm closed) algebra such that for all x A,
sup
yN,y1
[(yx)[ = sup
yA
1
,y1
[(yx)[
and moreover E
n
(A) A
1
for all n N. Repeating this process, we obtain a separable
-closed subalgebra A
k
A
k
such that the embedding i : L
1
(A
, ) L
1
(N, ),
i(x.) = x. is isometric. The dual map

E = i
: M M is a normal conditional
expectation onto the -weak operator closure B of A
. By construction, we have

E =
and E
n
(A
) A
N
n
. Since, E
n
is -strongly continuous, E
n
(B) is a von Neumann
subalgebra of N
n
B for all n N.
22 MARIUS JUNGE
Theorem 2.17. Let 1 p , then L
p
(M, (E
n
);
C
2
) is a Banach space and the Cauchy-
Schwarz inequality 2.15 i) holds. Let 1 < p , p
< with
1
p
+
1
p
= 1 and
p
the constant
from 2.13. If the sequence (N
n
) is either increasing or decreasing and (x
n
) a sequence in
L
p
(M, (E
n
);
C
2
), then
|(x
n
)|
Lp(M,(En);
C
2
)

p
sup
_
n
tr(D
1
p
n
x
n
)
_
_
_(b
n
D
1
p
)
_
_
_
L
p
(M,(En);
C
2
)
1
_
.
Proof: Since the triangle inequality and the Cauchy-Schwarz inequality are checked for
two sequences, it suces to consider a countably generated subalgebra B of M. Indeed,
according to Lemma 2.16 we can even assume that E
n
(B) B and there exists a -
preserving conditional expectation

E : M B. Then L
p
2
(B), is a subspace of L
p
2
(M).
Hence, L
p
(B, (E
n
);
C
2
) is a subspace of L
p
(M, (E
n
);
C
2
) and hence there is no loss of
generality to assume that M
is separable. Then rst assertions follow from Proposition

2.15. For the last inequality, we apply Lemma 2.11 to the image u
p
(L
p
(M, (E
n
);
C
2
)),
where u
p
is the isometric embedding from Proposition 2.15. Indeed, the projection R
p
ST
p
satises the assumptions of Lemma 2.11 and the norm is bounded by the universal constant
p
from Theorem 2.13. Finally, we note that u
p
preserves the duality bracket
tr(u
p
[(b
n
D
1
p
)]
u
p
[(a
n
D
1
p
)]) =
n
tr(D
1
p
E
n
(b
n
a
n
)D
1
p
) =
n
(E
n
(b
n
a
n
))
=
n
(b
n
a
n
) =
n
tr(D
1
p
n
a
n
D
1
p
) .
Therefore the duality formula from Lemma 2.11 is also valid for L
p
(M, (E
n
);
C
2
).
3. The dual version of Doobs inequality for 1 p 2
In this section, we start with an elementary proof of the dual version of Doobs inequality
for p = 2 and show how the complex interpolation method can be used to extend the
inequality to the interval 1 p 2. Then we provide the duality argument which justies
the name dual version of Doobs inequality. In the following, we consider a normal
faithful state , a von Neumann algebra N and a sequence of von Neumann subalgebras
N
n
with conditional expectations E
n
: N N such that
E
n
=
for all n N. (Note that in case of a tracial state such conditional expectations always
exist [Tk].) Let us stress that in addition to the last section we also always assume that
the sequence N
n
is increasing.
Lemma 3.1. Let (x
n
) be a sequence of positive elements in L
2
(N), then
_
_
_
_
_
n
E
n
(x
n
)
_
_
_
_
_
2
2
_
_
_
_
_
n
x
n
_
_
_
_
_
2
.
Proof: By monotonicity, (1.3) and Corollary 2.4, it suces to prove this inequality for
nite sequences. Using Proposition 2.5, Lemma 2.1 and positivity as in (1.3), see [Te,
Proposition 33], we deduce from Holders inequality
_
_
_
_
_
n
E
n
(x
n
)
_
_
_
_
_
2
2
=
nk
tr(E
n
(x
n
)E
k
(x
k
))
=
nk
tr(E
n
(x
n
)E
k
(x
k
)) +
n>k
tr(E
n
(x
n
)E
k
(x
k
))
=
nk
tr(E
k
(E
n
(x
n
)x
k
)) +
n>k
tr(E
n
(x
n
E
k
(x
k
)))
=
nk
tr(E
n
(x
n
)x
k
) +
n>k
tr(x
n
E
k
(x
k
))
=
k
tr(
_
nk
E
n
(x
n
)
_
x
k
) +
n
tr(x
n
_
k<n
E
k
(x
k
)
_
)
k
tr(
_
n
E
n
(x
n
)
_
x
k
) +
n
tr(x
n
_
k
E
k
(x
k
)
_
)
= 2tr(
_
n
x
n
__
k
E
k
(x
k
)
_
) 2
_
_
_
_
_
n
x
n
_
_
_
_
_
2
_
_
_
_
_
n
E
n
(x
n
)
_
_
_
_
_
2
.
Hence, we get
_
_
_
_
_
n
E
n
(x
n
)
_
_
_
_
_
2
2
_
_
_
_
_
n
x
n
_
_
_
_
_
2
.
24 MARIUS JUNGE
Lemma 3.2. If (DD
2
) holds with constant c
2
and 1 p 2, then for all sequences (x
n
)
and (y
n
) in L
2p
(N)
_
_
_
_
_
n
E
n
(x
n
y
n
)
_
_
_
_
_
p
c
2(p1)
p
2
_
_
_
_
_
n
x
n
x
n
_
_
_
_
_
1
2
p
_
_
_
_
_
n
y
n
y
n
_
_
_
_
_
1
2
p
.
Proof: Let us rst prove the assertion for nite sequences and p = 2 or p = 1. We start
with (DD
1
). Using Lemma 2.1, we get
_
_
_
_
_
n
E
n
(x
n
x
n
)
_
_
_
_
_
1
= tr(
n
E
n
(x
n
x
n
)) =
n
tr(E
n
(x
n
x
n
))
=
n
tr(x
n
x
n
) =
_
_
_
_
_
n
x
n
x
n
_
_
_
_
_
1
.
By the density of elements of the form x
n
= a
n
D
1
p
, the Cauchy-Schwarz inequality, see
Theorem 2.17, implies with Lemma 3.1 for p 1, 2
_
_
_
_
_
n
E
n
(x
n
y
n
)
_
_
_
_
_
p
_
_
_
_
_
n
E
n
(x
n
x
n
)
_
_
_
_
_
1
2
p
_
_
_
_
_
n
E
n
(y
n
y
n
)
_
_
_
_
_
1
2
p
c
p
_
_
_
_
_
n
x
n
x
n
_
_
_
_
_
1
2
p
_
_
_
_
_
n
y
n
y
n
_
_
_
_
_
1
2
p
,
(3.1)
where c
2
is the constant given in the assumption and c
1
= 1. To use interpolation, we
consider nite sequences (x
n
) and (y
n
) such that
_
_
_
_
_
n
x
n
x
n
_
_
_
_
_
p
= 1 =
_
_
_
_
_
n
y
n
y
n
_
_
_
_
_
p
.
We dene X =
n
x
n
x
n
, Y =
n
y
n
y
n
. Their support projections are denoted by q
X
and q
Y
and are in N, see [Te, Proposition 4. 2) c), Proposition 12]. Since X
1
2
q
X
, q
Y
Y

1
2
are well-dened unbounded operators, we can dene
v
n
= x
n
X
1
2
q
X
, w
n
= y
n
Y
1
2
q
Y
Note that x
n
x
n
X and y
n
y
n
Y implies x
n
= x
n
q
X
, y
n
= y
n
q
Y
, respectively and
according to Lemma 1.1 we have v
n
N, w
n
N. Then, we observe
n
v
n
v
n
= q
X
X
1
2
XX
1
2
q
X
q
X
1 ,
n
w
n
w
n
= q
Y
Y
1
2
Y Y
1
2
q
Y
q
Y
1 .
Let be determined by
1
p
=
1
1
+

2
. According to Kosakis interpolation [Ko]
[L
2,L
(N, ), L
4,L
(N, )]
= L
2p,L
(N, ) and [L
2,R
(N, ), L
4,R
(N, )]
= L
2p,R
(N, )
with respect to the state (x) = tr(Dx). By approximation, we may assume that there are
continuous functions X(z), Y (z) on the strip 0 Re(z) 1 with values in N, analytic
in the interior, such that X
1
2
= D
1
2p
X(), Y
1
2
= Y ()D
1
2p
and
sup
t
max|D
1
2
X(it)|
2
, |D
1
4
X(1 +it)|
4
1,
sup
t
max|Y (it)D
1
2
|
2
, |Y (1 +it)D
1
4
|
4
1.
Now, we note Kosakis symmetric interpolation result
[L
1,sym
(N, ), L
2,sym
(N, )]
= L
p,sym
(N, ) .
Hence, by (3.1) and Holders inequality we get
_
_
_
_
_
n
E
n
(X
1
2
v
n
w
n
Y
1
2
)
_
_
_
_
_
p
=
_
_
_
_
_
n
D
1
2p
E
n
(X()v
n
w
n
Y ())D
1
2p
_
_
_
_
_
p
sup
t
_
_
_
_
_
n
D
1
2
E
n
(X(it)v
n
w
n
Y (it))D
1
2
_
_
_
_
_
1
1
sup
t
_
_
_
_
_
n
D
1
4
E
n
(X(1 +it)v
n
w
n
Y (1 +it))D
1
4
_
_
_
_
_
2
sup
t
_
_
_
_
_
n
D
1
2
X(it)v
n
v
n
X(it)
D
1
2
_
_
_
_
_
1
2
1
_
_
_
_
_
n
D
1
2
Y (it)
n
w
n
Y (it)
D
1
2
_
_
_
_
_
1
2
1
c
2
sup
t
_
_
_
_
_
n
D
1
4
X(1 +it)v
n
v
n
X(1 +it)
D
1
4
_
_
_
_
_
2
1
sup
t
_
_
_
_
_
n
D
1
4
Y (1 +it)
n
w
n
Y (1 +it)D
1
4
_
_
_
_
_
2
2
sup
t
_
_
_D
1
2
X(it)X(it)
D
1
2
_
_
_
1
2
1
sup
t
_
_
_D
1
2
Y (it)
Y (it)D
1
2
_
_
_
1
2
1
sup
t
_
_
_D
1
4
X(1 +it)X(1 +it)
D
1
4
_
_
_
1
2
2
c
2
sup
t
_
_
_D
1
4
Y (1 +it)
Y (1 +it)D
1
4
_
_
_
1
2
2
c
2
= c
2(p1)
p
2
.
The assertion is proved.
26 MARIUS JUNGE
Proof of Theorem 0.1 in the case 1 p 2: For 1 p 2 and a sequence of
positive elements (z
n
) L
p
(N), we can apply Lemma 3.2 to x
n
= y
n
= z
1
2
n
and deduce
the assertion for the sequence (z
n
).
The duality argument relies on the following norm for sequences (x
n
) L
p
(N)
|(x
n
)|
Lp(N;
1
)
= inf
_
_
_
_
_
nj
v
nj
v
nj
_
_
_
_
_
1
2
p
_
_
_
_
_
nj
w
jn
w
jn
_
_
_
_
_
1
2
p
.
Here the inmum is taken over all (double indexed) sequences (v
nj
) and (w
nj
) such that
for all n
x
n
=
j
v
nj
w
jn
.
We require norm convergence for p < and convergence in the -weak operator topology
for p = . In fact, we think of x
n
being obtained by matrix multiplication of a row with a
column vector. We denote by L
p
(N;
1
) the set of all sequences admitting a decomposition
as above.
Remark 3.3. This norm is motivated by the following characterization of a normal, de-
composable map T :
N, see [Pa]. Indeed, a normal map is decomposable if and only

if there are sequences (x
n
) N, (y
n
) N such that T(e
n
) = y
n
x
n
and
_
_
_
_
_
n
y
n
y
n
_
_
_
_
_
N
_
_
_
_
_
n
x
n
x
n
_
_
_
_
_
N
< .
Lemma 3.4. If (DD
p
) holds, then
_
_
_
_
_
n
E
n
(x
n
y
n
)
_
_
_
_
_
p
c
p
_
_
_
_
_
n
x
n
x
n
_
_
_
_
_
1
2
p
_
_
_
_
_
n
y
n
y
n
_
_
_
_
_
1
2
p
,
The linear map T : L
p
(N;
1
) L
p
(N;
1
), T((x
n
)) = (E
n
(x
n
)) satises
|T| c
p
.
Proof: We have seen in (3.1) that the rst inequality is an immediate consequence of
the Cauchy-Schwarz inequality, see Theorem 2.17. As for the second assertion, we can
assume that N
is separable and use the Kasparov maps u

n
p
: L
p
(N, E
n
) L
p
(N
n
;
C
2
)
from Proposition 2.8. Let x
n
=
j
v
nj
w
nj
, then
|(E
n
(x
n
))|
Lp(N;
1
)
=
_
_
_
_
_
(
nj
u
n
p
(v
nj
)
u
n
p
(w
nj
))
_
_
_
_
_
Lp(N;
1
)
_
_
_
_
_
nj
u
n
p
(v
nj
)
u
n
p
(v
nj
)
_
_
_
_
_
1
2
p
_
_
_
_
_
nj
u
n
p
(w
nj
)
u
n
p
(w
nj
)
_
_
_
_
_
1
2
p
=
_
_
_
_
_
nj
E
n
(v
nj
v
nj
)
_
_
_
_
_
1
2
p
_
_
_
_
_
nj
E
n
(w
nj
w
nj
)
_
_
_
_
_
1
2
p
c
p
_
_
_
_
_
nj
v
nj
v
nj
_
_
_
_
_
1
2
p
_
_
_
_
_
nj
w
nj
w
nj
_
_
_
_
_
1
2
p
.
Taking the inmum over all these decompositions, we obtain the assertion.
Let us state some elementary properties of the space L
p
(N;
1
).
Lemma 3.5. For 1 p the set L
p
(N;
1
) is a Banach space. For 1 p < , the
set L
0
p
of elements
x
n
=
j
v
nj
w
jn
such that card(j, n)[v
nj
,= 0 or w
jn
,= 0 < is dense in L
p
(N;
1
). If (x
n
) is a sequence
then
_
_
_
_
_
n
x
n
_
_
_
_
_
p
|(x
n
)|
Lp(N;
1
)
and equality holds if all the x
n
s are positive.
Proof: The proof of the triangle inequality is completely elementary, see also [Ps2], but
essential. Indeed, let > 0 and
x
n
=
j
1
v
nj
1
w
j
1
n
and y
n
=
j
2
v
nj
2
w
j
2
n
,
28 MARIUS JUNGE
such that
_
_
_
_
_
nj
1
v
nj
1
v
nj
1
_
_
_
_
_
p
=
_
_
_
_
_
nj
1
w
j
1
n
w
j
1
n
_
_
_
_
_
p
(1 +) |(x
n
)|
Lp(N;
1
)
,
_
_
_
_
_
nj
2
v
nj
2
v
nj
2
_
_
_
_
_
p
=
_
_
_
_
_
nj
2
w
j
2
n
w
j
2
n
_
_
_
_
_
p
(1 +) |(y
n
)|
Lp(N;
1
)
.
We have
x
n
+ y
n
=
j
1
v
nj
1
w
j
1
n
+
j
2
v
nj
2
w
j
2
n
and the triangle inequality in L
p
(N) implies
_
_
_
_
_
nj
1
v
nj
1
v
nj
1
+
nj
2
v
nj
2
v
nj
2
_
_
_
_
_
p
_
_
_
_
_
nj
1
v
nj
1
v
nj
1
_
_
_
_
_
p
+
_
_
_
_
_
nj
2
v
nj
2
v
nj
2
_
_
_
_
_
p
(1 +)(|(x
n
)|
Lp(N;
1
)
+ |(y
n
)|
Lp(N;
1
)
) .
Similarly,
_
_
_
_
_
nj
1
w
j
1
n
v
j
1
n
+
nj
2
w
j
2
n
w
j
2
n
_
_
_
_
_
p
(1 +)(|(x
n
)|
Lp(N;
1
)
+|(y
n
)|
Lp(N;
1
)
)
and the assertion follows with 0. We consider the spaces of column matrices, row
matrices, L
p
(N;
C
2
(N
2
)), L
p
(N;
R
2
(N
2
)) L
p
(B(
2
(N
2
))
N), respectively. Using,

((x
nk
) (y
nk
)) = (
k
x
nk
y
nk
)
nN
,
we deduce that L
p
(N;
1
) is isomorphic to a quotient space of the projective tensor product
L
p
(N;
R
2
(N
2
))
L
p
(N;
C
2
(N
2
)), see [DF] for a denition and basic properties of the
projective tensor product. Hence L
p
(N;
1
) is complete. The image under of pairs of
nite sequences generates L
0
p
. According to Corollary 2.4 nite sequences are dense in the
column and row spaces and hence L
0
p
is dense in L
p
(N;
1
). Finally, let (x
n
) be a sequence
of positive elements. Clearly, x
n
= x
1
2
n
x
1
2
n
and hence
|(x
n
)|
Lp(N;
1
)

_
_
_
_
_
n
x
n
_
_
_
_
_
p
.
On the other hand if x
n
=
j
v
nj
w
nj
, we deduce from Holders inequality
_
_
_
_
_
n
x
n
_
_
_
_
_
p
=
_
_
_
_
_
nj
(e
1,nj
v
nj
)(e
nj,1
w
nj
)
_
_
_
_
_
p
_
_
_
_
_
nj
e
1,nj
v
nj
_
_
_
_
_
2p
_
_
_
_
_
nj
e
nj,1
w
nj
_
_
_
_
_
2p
=
_
_
_
_
_
nj
v
nj
v
nj
_
_
_
_
_
1
2
p
_
_
_
_
_
nj
w
nj
w
nj
_
_
_
_
_
1
2
p
.
Taking the inmum yields the assertion.
Inspired by Pisiers vector-valued L
p
(N, ;
) space, we dene for 0 < p

_
_
_
_
sup
n
[x
n
[
_
_
_
_
p
= |(x
n
)|
Lp(N;)
= inf
xn=aynb
|a|
2p
|b|
2p
sup
n
|y
n
|
N
,
where the inmum is taken over all a, b L
2p
(N) and all bounded sequences (y
k
). If N
is a hypernite, nite von Neumann algebra, this space coincides with L
p
(N, ;
) in the
sense of Pisier [Ps2]. The rst (formal) notation is suggestive and facilitates understand-
ing our inequalities in view of the commutative theory. For positive elements, we will
drop the absolute value. Let us note that Haagerups work [Ha2] shows that the equal-
ity L
1
(N;
) = L
1
(N)
(operator space projective tensor product) only holds for

injective von Neumann algebras. However, this does not aect the following factorization
result which is, nowadays, a standard application of the Grothendieck-Pietsch version of
the Hahn-Banach theorem, see [Ps1, Ps2].
Proposition 3.6. Let 1 p < . If p = 1, then L
1
(N;
1
) =
1
(L
1
(N)) holds with equal
norms. If 1 < p , p
< satisfy
1
p
+
1
p
= 1, then
L
p
(N;
1
)
= L
p
(N;
)
holds isometrically.
30 MARIUS JUNGE
Proof: If z
n
= ay
n
b and (y
n
) is a bounded sequence, we deduce from Holders inequality
for all (v
nj
) L
2p
(N), (w
jn
) L
2p
(N)
n
tr(z
n
j
v
nj
w
jn
)
nj
tr(ay
n
bv
nj
w
jn
)
nj
tr(y
n
bv
nj
w
jn
a)
sup
n
|y
n
|
nj
|bv
nj
w
jn
a|
1
sup
n
|y
n
|
nj
|bv
nj
|
2
2
_1
2
_
nj
|w
jn
a|
2
2
_1
2
= sup
n
|y
n
|
_
_
_
_
_
b
nj
v
nj
v
jn
b
_
_
_
_
_
1
2
1
_
_
_
_
_
a
nj
w
nj
w
jn
a
_
_
_
_
_
1
2
1
sup
n
|y
n
|
|b|
2p
_
_
_
_
_
nj
v
nj
v
nj
_
_
_
_
_
1
2
p
|a|
2p
_
_
_
_
_
nj
w
jn
w
jn
_
_
_
_
_
1
2
p
.
Hence, L
p
(N;
) L
p
(N;
1
)
. Using
1
(L
1
(N)) L
1
(N;
1
), we deduce the equality
1
(L
1
(N)) = L
1
(N;
1
). Now we show that for 1 < p < all the functionals are in
L
p
(N;
). Let : L
p
(N;
1
) C be a norm one functional. Using
1
(L
p
(N))
L
p
(N;
1
), we can assume that there exists a sequence (z
n
) L
p
(N) such that
[(x
n
)] =
(zn)
[(x
n
)] =
n
tr(z
n
x
n
) .
Let us denote by B = B
+
L
p
(N)
the positive part of the unit ball in L
p
(N). B is compact
when equipped with the (L
p
(N), L
p
(N))-topology. The denition of L
p
(N;
1
) implies
with the geometric/arithmetric mean inequality
nj
tr(z
n
v
nj
w
jn
)
[(
j
v
nj
w
jn
)
n
]
_
_
_
_
_
nj
v
nj
v
nj
_
_
_
_
_
1
2
p
_
_
_
_
_
nj
w
jn
w
jn
_
_
_
_
_
1
2
p
1
2
sup
c,dB
[
nj
tr(v
nj
v
nj
c) +
nj
tr(w
jn
w
jn
d)] .
Since the right hand side remains unchanged under multiplication with signs
nj
, we deduce
nj
[tr(z
n
v
nj
w
jn
)[
1
2
sup
c,dB
[
nj
tr(v
nj
v
nj
c) +
nj
tr(w
jn
w
jn
d)] .
Following the Grothendieck-Pietsch separation argument as in [Ps1], we observe that the
C given by the functions
f
v,w
(c, d) =
n
[tr(v
n
v
n
c) + tr(w
n
w
n
d) 2[tr(z
n
v
n
w
n
)[]
is disjoint from the cone C
= g[ supg < 0. Here v = (v

n
) and w = (w
n
) are nite
sequences and hence f
v,w
is continuous with respect to the product topology on B B.
Since f
v,w
+f
v, w
can be obtained by taking the ( v
n
, w
n
)s to the right of the nite sequence
(v
n
, w
n
), we deduce that C is a cone. Hence, there exist a measure on B B and a
scalar t such that for all g C
and f C
_
BB
g d < t
_
BB
f d .
Since we are dealing with cones, it turns out that t = 0 and is positive. Therefore, we
can and will assume that is a probability measure. We dene the positive elements c
and d by their projections
a =
_
BB
c d(c, d) , b =
_
BB
d d(c, d) .
By convexity of B, we deduce a, b B. Hence, we obtain
n
2 [tr(z
n
v
n
w
n
)[
_
BB
n
[tr(v
n
v
n
c) + tr(w
n
v
n
d)] d(c, d)
=
n
_
BB
tr(v
n
v
n
c) d(c, d) +
_
BB
tr(w
n
w
n
d) d(c, d)
=
n
[tr(v
n
v
n
a) + tr(w
n
w
n
b)] .
Using once more 2st = inf
r>0
(rs)
2
+ (r
1
t)
2
, we get
n
[tr(z
n
v
n
w
n
)[
_
n
tr(v
n
v
n
a)
_1
2
_
n
tr(w
n
w
n
b)
_1
2
=
_
n
_
_
_a
1
2
v
n
_
_
_
2
2
_1
2
_
n
_
_
_b
1
2
w
n
_
_
_
2
2
_1
2
.
(3.2)
Let q
a
, q
b
N be the support projections of a, b, respectively. Consider, d
a
= a
p
+ (1
q
a
)D(1 q
a
), D the density of . Then
da
(x) = tr(d
a
x) is a normal, faithful state on N
and according to Lemma 1.2 d
1
2
a
N is dense in L
2
(N). Hence,
q
a
d
1
2
a
N = a
p
2
N = a
1
2
a
p
2p
N a
1
2
L
2p
(N)
32 MARIUS JUNGE
shows that a
1
2
L
2p
(N) is dense in q
a
L
2
(N). Similarly, b
1
2
L
2p
(N) is dense in q
b
L
2
(N) and
therefore (3.2) implies that for every n N there is a contraction T
n
: q
a
L
2
(N) q
b
L
2
(N)
such that for all v, w L
2p
(N)
tr(wz
n
v) = (b
1
2
w
, T
n
(a
1
2
v)) = tr(wb
1
2
T
n
(a
1
2
v)) .
This means T
n
is a bounded extension of the densely dened hermitian form
(b
1
2
q
b
(h
), z
n
a
1
2
q
a
(h)) = (b
1
2
q
b
h
, z
n
a
1
2
q
a
h)
Using the density of a
1
2
L
2p
(N) and b
1
2
L
2p
(N) it is easily checked that q
b
T
n
q
a
is aliated
with N. Since T
n
is bounded, we deduce q
b
T
n
q
a
N. On the other hand, we have for
v L
2p
(N) and w L
2p
(N)
[tr(z
n
(1 q
a
)vw)[ tr(a(1 q
a
)vv
)
1
2
tr(w
wb)
1
2
= 0
and
[tr(z
n
vw(1 q
b
)[ tr(avv
)
1
2
tr(w
w(1 q
b
)b)
1
2
= 0 .
This shows z
n
= q
b
z
n
q
a
and therefore
z
n
= q
b
z
n
q
a
= q
b
b
1
2
q
b
b
1
2
z
n
a
1
2
q
a
a
1
2
q
a
= b
1
2
y
n
a
1
2
.
The assertion is proved because L
0
p
is dense in L
p
(N;
1
) and hence the functional is
uniquely determined by the sequence (z
n
).
Remark 3.7. Let 1 p < and (z
n
) L
p
(N) a sequence of positive elements, then
_
_
_
_
sup
n
z
n
_
_
_
_
p
= sup
_
_
_
n
tr(z
n
x
n
)
x
n
0 ,
_
_
_
_
_
n
x
n
_
_
_
_
_
p
1
_
_
_
.
Moreover, there exists a positive element a L
2p
(N) and a sequence of positive elements
y
n
such that
z
n
= ay
n
a and |a|
2
2p
sup
n
|y
n
|
=
_
_
_
_
sup
n
z
n
_
_
_
_
p
.
For positive elements (x
n
) L
p
(N)
+
, we also have
_
_
_
_
_
n
x
n
_
_
_
_
_
p
= sup
_
n
tr(x
n
z
n
)
z
n
0 , |(z
n
)|
L
p
(N;)
1
_
and therefore the cones of positive sequences in L
p
(N;
1
) respectively L
p
(N;
) are in
duality.
Proof: For positive elements (z
n
) satisfying
n
tr(z
n
x
n
)
_
_
_
_
_
n
x
n
_
_
_
_
_
p
,
for all sequences of positive elements (x
n
) L
p
(N)
+
, we deduce
j,n
[tr(z
n
v
nj
v
nj
)[
n
tr(z
n
(
j
v
nj
v
nj
))
_
_
_
_
_
nj
v
nj
v
nj
_
_
_
_
_
p
= sup
cB
j,n
tr(v
nj
v
nj
c) .
Using the Hahn-Banach separation argument in the space of continuous functions on B,
we obtain a positive element a in the unit ball of L
p
(N) such that
n
[tr(z
n
v
n
v
n
)[
n
tr(v
n
v
n
a) .
Since a
1
2
z
n
a
1
2
is positive, this inequality still ensures that all the y
n
= a
1
2
z
n
a
1
2
are
positive contractions in N. The last equality follows immediately from Lemma 3.5 and
the duality between the positive parts of L
p
(N) and L
p
(N).
Remark 3.8. Proposition 3.6 and Remark 3.7 can easily be modied for uncountable (or-
dered) index sets by requiring the inequality for all countable (ordered) subsets or for an
essential supremum. This is helpful in the context of continuous ltrations.
The required duality argument is now very simple.
Lemma 3.9. Let 1 < p and
1
p
+
1
p
= 1 assume that (DD

p
) holds with constant c
p
,
then for every y L
p
(N)
_
_
_
_
sup
n
[E
n
(y)[
_
_
_
_
p
c
p
|y|
p
.
Moreover, for every sequence of positive elements (y
n
)
_
_
_
_
_
sup
n
[
jn
E
n
(y
j
)[
_
_
_
_
_
p
c
p
_
_
_
_
_
n
y
n
_
_
_
_
_
p
.
Proof: Indeed, as observed in Lemma 3.4 and using Lemma 3.5, we deduce that (DD
p
)
implies that the linear map T : L
p
(N;
1
) L
p
(N), T((x
n
)) =
n
E
n
(x
n
) satises
34 MARIUS JUNGE
|T| c
p
. By duality and Proposition 3.6, we deduce for all y L
p
(N)
_
_
_
_
sup
n
[E
n
(y)[
_
_
_
_
p
= |T
(y)|
L
p
(N;
1
)
|T
| |y|
p
c
p
|y|
p
.
Given a sequence of positive elements (y
n
), we consider y =
j
y
j
and obtain
_
_
_
_
sup
n
[E
n
(y)[
_
_
_
_
p
c
p
|y|
p
= c
p
_
_
_
_
_
j
y
j
_
_
_
_
_
p
.
However, for positive elements (x
n
) L
p
(N), we deduce by positivity
n
tr(E
n
(
jn
y
j
)x
n
)
j
tr(E
n
(y
j
)x
n
) |(E
n
(y))
nN
|
Lp(N;)
_
_
_
_
_
n
x
n
_
_
_
_
_
p
.
Hence, Remark 3.7 implies
_
_
_
_
_
sup
n
jn
E
n
(y
j
)
_
_
_
_
_
p
_
_
_
_
sup
n
E
n
(y))
nN
_
_
_
_
p
c
p
|y|
p
= c
p
_
_
_
_
_
j
y
j
_
_
_
_
_
p
.
Theorem 3.10. For 1 < p there exists a constant c
p
such that for every sequence
(N
n
) of von Neumann subalgebras with sequence of -invariant conditional expectations
(E
n
) satisfying E
n
E
m
= E
min(n,m)
and for every x L
p
(N) there exist a, b L
2p
(N) and
a bounded sequence (y
n
) N such that
E
n
(x) = ay
n
b and |a|
2p
|b|
2p
sup
n
|y
n
|
c
p
|x|
p
.
If x is positive, one can in addition assume that b = a
and all the y

n
s are positive.
Proof for 2 p : This follows immediately from Lemma 3.9, Lemma 3.2 and Lemma
3.1. Using that E
n
(x) is positive for positive x, the addition follows from Remark 3.7.
4. The dual version of Doobs inequality for 2 p <
In our approach to (DD
p
) in the range 2 p < , our aim is to obtain the same kind
of inequalities for the maximal function as in Garsias book [Ga]. As mentioned in the
introduction, we are forced to use more duality arguments because 0 a b implies
a
only for 0 1 and therefore most of the elementary proofs in Garsias book
are no longer valid in the non-commutative case. We will make the same assumptions
about N, (N
n
), (E
n
) and , D as in the previous section. In particular, the sequence N
n
is supposed to be increasing.
Lemma 4.1. Let 0 x z L
1
(N) such that the support projection of z is 1. Let
1 2, then
tr(z
1
2
(z
)z
1
2
) 2tr(z x) .
Proof: We dene = 1 (0, 1) and observe that 0 x z implies
x
1
z
1
.
We apply Lemma 1.1 to x
2
and z
2
and deduce from x
that v = x
2
z
2
is a
contraction in N. In particular v
= z
2
x
2
N and
a = z
2
x
1+
2
= v
x
1
2
L
2
(N) .
For all elements a, b L
2
(N), we note that (ab)
(ab) 0 implies with tracial property

of the trace
tr(a
b) + tr(b
a) tr(a
a) + tr(b
b) = tr(a
a) + tr(bb
) .
We dene b = z
2
x
1
2
2
which is in L
2
(N) by Holders inequality. Then, we observe
tr(a
b) = tr(x
1+
2
z
2
z
2
x
1
2
2
) = tr(x) ,
tr(b
a) = tr(x
1
2
2
z
2
z
2
x
1+
2
) = tr(x) .
Hence, we deduce
2tr(x) = tr(a
b) + tr(b
a) tr(a
a) + tr(bb
)
= tr(x
1+
2
z
2
z
2
x
1+
2
) + tr(z
2
x
1
2
2
x
1
2
2
z
2
)
= tr(x
1+
2
z
x
1+
2
) + tr(z
2
x
1
z
2
)
tr(x
1+
2
z
x
1+
2
) + tr(z
2
z
1
z
2
)
= tr(x
2
z
1
x
2
) + tr(z) .
This implies
tr(x
2
z
1
x
2
) + tr(z) 2tr(x) + tr(z) + tr(z) = 2tr(z x) .
The assertion follows from z
1
2
x
z
1
2
= aa
L
1
(N) and
tr(x
2
z
1
x
2
) = tr(a
a) = tr(aa
) = tr(z
1
2
x
z
1
2
) .
36 MARIUS JUNGE
The following proposition is a modication of the Theorem in the appendix of Pisier, Xus
paper [PX] and enables us to apply duality.
Proposition 4.2. Let 1 r
< 2 < r , and

1
r
+
1
r
= 1, then for all (x
j
) L
r
(N)
and (y
n
) L
r
(N, (E
n
);
C
2
)
n
tr(y
n
x
n
)
2 |(y
n
)|
L
r
(N,(En);
C
2
)
_
_
_
_
_
j
x
j
x
j
_
_
_
_
_
1
2
r
2
.
Moreover,
n
tr(y
n
x
n
)
2 |(y
n
)|
L
r
(N,(En);
C
2
)
_
_
_
_
_
sup
n
jn
E
n
(x
j
x
j
)
_
_
_
_
_
1
2
r
2
.
Proof: Let us assume that both sequences are nite, i.e. x
j
= 0 = y
j
for j m. By
density, we can moreover assume that y
j
= a
j
D
1
r
with a
j
N and
_
_
_
_
_
m
j=1
D
1
r
E
j
(a
j
a
j
)D
1
r
_
_
_
_
_
r
2
< 1 .
By continuity, we can assume that there is an > 0 such that
_
_
_
_
_
D
2
r
+
m
j=1
D
1
r
E
j
(a
j
a
j
)D
1
r
_
_
_
_
_
r
2
1 .
Let 1 q such that
1
q
+
2
r
= 1. For n N, we dene
S
n
=
_
D
2
r
+
n
j=1
D
1
r
E
j
(a
j
a
j
)D
1
r
_r
2q
L
q
(N
n
) L
q
(N) .
The support projection of S
n
is 1 and D
1
q
S
n
. According to Lemma 1.1, we deduce
that w
n
:= D
1
2q
S
1
2
n
N
n
. Hence
y
n
S
1
2
n
= a
n
D
1
r
1
2
n
= a
n
D
1
2
D
1
2q
S
1
2
n
= a
n
D
1
2
w
n
L
2
(N) .
In particular, S
1
2
n
y
n
L
2
(N) and
S
1
2
n
y
n
y
n
S
1
2
n
= w
n
D
1
2
a
n
a
n
D
1
2
w
n
L
1
(N) .
Moreover,
E
n
(S
1
2
n
y
n
y
n
S
1
2
n
) = E
n
(w
n
D
1
2
a
n
a
n
D
1
2
w
n
) = w
n
D
1
2
E
n
(a
n
a
n
)D
1
2
w
n
= S
1
2
n
D
1
r
E
n
(a
n
a
n
)D
1
r
1
2
n
.
This implies with the Cauchy-Schwarz inequality
n
tr(y
n
x
n
)
n
tr(x
n
y
n
)
n
tr((x
n
S
1
2
n
)(S
1
2
n
y
n
))
n
tr(S
1
2
n
y
n
x
n
S
1
2
n
)
n
tr(S
1
2
n
y
n
y
n
S
1
2
n
)
_1
2
_
n
tr(x
n
x
n
S
n
)
_1
2
=
_
n
tr(E
n
(S
1
2
n
y
n
y
n
S
1
2
n
))
_1
2
_
n
tr(x
n
x
n
S
n
)
_1
2
=
_
n
tr(S
1
2
n
D
1
r
E
n
(a
n
a
n
)D
1
r
1
2
n
)
_1
2
_
n
tr(E
n
(x
n
x
n
)S
n
)
_1
2
.
To estimate the rst term, we dene =
2
r
[1, 2] and notice that

1 = 1
2
r
= 1 2 +
2
r
=
1
q
.
For xed n, we dene x = S
q
n1
and z = S
q
n
. Since
r
2
1, we have
x =
_
D
2
r
+
n1
j=1
D
1
r
E
j
(a
j
a
j
)D
1
r
_r
_
D
2
r
+
n
j=1
D
1
r
E
j
(a
j
a
j
)D
1
r
_r
2
= z .
Then, we note that z
1
2
= z
1
2q
= S
1
2
n
. Hence Lemma 4.1 implies
tr(S
1
2
n
D
1
r
E
n
(a
n
a
n
)D
1
r
1
2
n
) = tr(z
1
2
(z
)z
1
2
) 2tr(z x)
= 2tr(S
q
n
S
q
n1
) .
Therefore, we obtain
n
tr(S
1
2
n
D
1
r
E
n
(a
n
a
n
)D
1
r
1
2
n
) =
n
2tr(S
q
n
S
q
n1
) = 2tr(S
q
m
)
= 2
_
_
_
_
_
D
2
r
+
m
j=1
D
2
r
E
j
(a
j
a
j
)D
2
r
_
_
_
_
_
r
2
r
2
2 .
Now, we want to estimate the second term. Let us dene
j
= S
j
S
j1
.
38 MARIUS JUNGE
and note that
j
L
q
(N
j
). As usual, we set S
1
= 0. Moreover, r
2q implies that
j
is
positive. Then, we deduce with E
j
E
n
= E
min(j,n)
n
tr(E
n
(x
n
x
n
)S
n
) =
jn
tr(E
n
(x
n
x
n
)
j
)
=
j
tr(
nj
E
n
(x
n
x
n
)
j
)
=
j
tr(E
j
_
nj
E
n
(x
n
x
n
)
_
j
)
=
j
tr(
nj
E
j
(x
n
x
n
)
j
)
Now, we can continue in two dierent ways
j
tr(
nj
E
j
(x
n
x
n
)
j
)
_
_
_
_
_
sup
j
E
j
(
nj
x
n
x
n
)
_
_
_
_
_
r
2
_
_
_
_
_
j
_
_
_
_
_
q
=
_
_
_
_
_
sup
j
E
j
(
nj
x
n
x
n
)
_
_
_
_
_
r
2
|S
m
|
q
_
_
_
_
_
sup
j
E
j
(
nj
x
n
x
n
)
_
_
_
_
_
r
2
.
By homogeneity, we obtain the second assertion
n
tr(y
n
x
n
)
2
_
_
_
_
_
sup
j
E
j
(
nj
x
n
x
n
)
_
_
_
_
_
1
2
r
2
|(y
n
)|
L
r
(N,(En);
C
2
)
.
Using for all j N that E
j
(
j
) =
j
, we also get by positivity
j
tr(
nj
E
j
(x
n
x
n
)
j
) =
j
tr(
nj
x
n
x
n
E
j
(
j
))
=
j
tr(
nj
x
n
x
n
j
)
tr(
n
x
n
x
n
j
)
_
_
_
_
_
n
x
n
x
n
_
_
_
_
_
r
2
_
_
_
_
_
j
_
_
_
_
_
q
_
_
_
_
_
n
x
n
x
n
_
_
_
_
_
r
2
|S
m
|
q

_
_
_
_
_
n
x
n
x
n
_
_
_
_
_
r
2
.
Again by homogeneity, we deduce
n
tr(y
n
x
n
)
2
_
_
_
_
_
n
x
n
x
n
_
_
_
_
_
1
2
r
2
|(y
n
)|
L
r
(N,(En);
C
2
)
.
Remark 4.3. For r
= r = 2 the assertion is trivially true because L

2
(N, (E
n
);
C
2
) =
2
(L
2
(N)) and
n
|x
n
|
2
2
=
_
_
_
_
_
n
x
n
x
n
_
_
_
_
_
1
= tr(
n
x
n
x
n
) = tr(E
1
(
n
x
n
x
n
))
=
_
_
_
_
_
E
1
(
n
x
n
x
n
)
_
_
_
_
_
1
_
_
_
_
_
sup
j
E
j
(
nj
x
n
x
n
)
_
_
_
_
_
1
.
Proposition 4.2 also shows that the BMO
C
and H
C
1
duality from [PX] is still valid in the
non tracial case.
Corollary 4.4. Let 1 q < and
2q
the constant in Steins inequality, then for all
(x
n
) L
2q
(N)
_
_
_
_
_
n
E
n
(x
n
x
n
)
_
_
_
_
_
q
2
2
2q
_
_
_
_
_
sup
n
E
n
(
jn
x
j
x
j
)
_
_
_
_
_
q
.
Proof: We dene r = 2q and note
_
_
_
_
_
n
E
n
(x
n
x
n
)
_
_
_
_
_
1
2
r
= |(x
n
)|
Lp(N,(En);
C
2
)
.
40 MARIUS JUNGE
Therefore the assertion follows from Proposition 4.2 and Theorem 2.17.
Proof of (DD
p
) for 1 < p < : We dene r = 2p > 2 Let (z
n
) L
p
(N) be a sequence
of positive elements and dene x
n
= z
1
2
n
. Hence by Theorem 2.17 and Proposition 4.2, we
deduce
_
_
_
_
_
n
E
n
(z
n
)
_
_
_
_
_
1
2
p
=
_
_
_
_
_
n
E
n
(x
n
x
n
)
_
_
_
_
_
1
2
p

r
sup
(yn)
L
r
(N,(En);
C
2
)
1
n
tr(y
n
x
n
)
2
_
_
_
_
_
n
x
n
x
n
_
_
_
_
_
1
2
p
.
The assertion follows with c
p
2
2
2p
.
Proof of Theorem 0.2 and 3.10 for 1 < p 2: This follows from (DD
p
) via Lemma
3.9 and Remark 3.7.
Remark 4.5. Let N
be separable and be a functional on L

p
(N, (E
n
);
C
2
), then there
exists a sequence (x
n
) such that
((y
n
)) =
n
tr(x
n
y
n
) and
_
_
_
_
_
sup
n
E
n
(
jn
x
j
x
j
)
_
_
_
_
_
1
2
p
2
d
p
2
||
L
p
(N,(En);
C
2
)
.
Here d
p
2
is the constant in Doobs inequality from Theorem 0.2. The assertion yields an
extension of the BMO
C
-H
C
1
duality for 2 < p < and fails for p = 2. The proof uses
the Kasparov isomorphism from Proposition 2.15. We leave it to the interested reader.
Answering a question by G. Pisier, we can even produce an asymmetric version of Doobs
inequality. Indeed, let 1 < p and consider 1 r, s < such that
2
p
=
1
s
+
1
t
.
Given x L
p
(N) and sequences (v
nj
) L
r
(N), (w
jn
) L
s
(N), we deduce from the
Cauchy-Schwartz inequality 2.13, (DD
s
) and (DD
t
) that
n,j
tr(E
n
(x)v
nj
w
jn
)
n,j
tr(xE
n
(v
nj
w
jn
))
|x|
p
_
_
_
_
_
nj
E
n
(v
nj
w
jn
)
_
_
_
_
_
p
|x|
p
_
_
_
_
_
nj
E
n
(v
nj
v
nj
)
_
_
_
_
_
1
2
s
_
_
_
_
_
nj
E
n
(w
jn
w
jn
)
_
_
_
_
_
1
2
t
c
1
2
s
c
1
2
t
|x|
p
_
_
_
_
_
nj
v
nj
v
nj
_
_
_
_
_
1
2
s
_
_
_
_
_
nj
w
jn
w
jn
_
_
_
_
_
1
2
t
.
Similar as in Proposition 3.6, we deduce the existence of bounded a sequence (z
n
) and
elements a L
2s
(N), b L
2t
(N) such that E
n
(x) = bz
n
a and
|a|
2s
|b|
2t
sup
n
|z
n
|
c
1
2
s
c
1
2
t
|x|
p
.
Note that
1
2t
+
1
2s
=
1
p
and therefore we have proved the following asymmetric version of
Theorem 0.2. (The assumption 2 < q, r is indeed necessary [DJ1].)
Corollary 4.6. Let 1 < p and 2 < q, r such that
1
q
+
1
r
=
1
p
. Then for
every x L
p
(N) there exists a sequence (z
n
) N and a L
q
(N), b L
r
(N) such that
E
n
(x) = az
n
b and
|a|
q
|b|
r
sup
n
|z
n
|
c(p, q, r) |x|
p
.
5. Applications
In this section, we present rst applications of Doobs inequality in terms of submartin-
gales, Doob decomposition. We make the same assumptions about N, (N
n
), (E
n
), and
D as in the previous section and start with almost immediate consequences of the dual
version of Doobs inequality.
Corollary 5.1. Let 1 < p and (z
n
) be an adapted sequence of positive elements, i.e.
z
n
L
p
(N
n
)
+
. If for all n N
z
n
E
n
(z
n+1
) and sup
m
|z
m
|
p
< ,
42 MARIUS JUNGE
then there exist a positive element a L
2p
(N)
+
and a sequence of positive contractions
(y
n
) N such that
z
n
= ay
n
a and |a|
2
2p
c
p
sup
m
|z
m
|
p
.
Proof: It suces to consider p < . We note that for n m
z
n
E
n
(z
n+1
) E
n
(E
n+1
(z
n+2
)) = E
n
(z
n+2
) E
n
(z
m
) .
Let
1
p
+
1
p
= 1. In order to estimate the norm in L

p
(N;
), we refer to Remark 3.7. Given

a nite sequence (x
n
) of positive elements such that x
n
= 0 for n m, we deduce from
(DD
p
)
n
tr(z
n
x
n
)
n
tr(E
n
(z
n
x
n
)) =
n
tr(E
n
(z
n
)E
n
(x
n
))
n
tr(E
n
(z
m
)E
n
(x
n
)) =
n
tr(E
n
(z
m
E
n
(x
n
)))
=
n
tr(z
m
E
n
(x
n
)) tr(z
m
n
E
n
(x
n
))
|z
m
|
p
_
_
_
_
_
n
E
n
(x
n
)
_
_
_
_
_
p
c
p
|z
m
|
p
_
_
_
_
_
n
x
n
_
_
_
_
_
p
Hence, the assertion follows from Remark 3.7.

Corollary 5.2. Let 2 < p and (z
n
) be an adapted seqeunce, i.e. z
n
L
p
(N
n
). If for
all n N
z
n
z
n
E
n
(z
n+1
z
n+1
) ,
then there exist a positive element a L
p
(N) and a sequence (y
n
) of contractions such
that
z
n
= y
n
a and |a|
p
c
1
2
p
p2
sup
m
|z
m
|
p
.
Proof: We apply Corollary 5.1 to z
n
= z
n
z
n
L
p
2
(N
n
) and obtain positive contractions
(v
n
) N and a L
p
(N) such that
z
n
z
n
= av
n
a and |a|
2
p
c
p
p2
sup
m
|z
m
z
m
|p
2
.
If q
a
is the support projection of a, we see that y
n
= v
1
2
n
a
1
q
a
satises the assertion.
Let us mention an immediate application of (DD
p
) in terms of the Doob decomposition of
the square function. Given a martingale sequence x =
k
d
k
(x), where d
k
(x) = E
k
(x)
E
k1
(x) (and E
0
(x) = 0), we recall that one of the square functions is given by
s
c
(x) =
k
d
k
(x)
d
k
(x)
The square function is the discrete analogue of the quadratic variation term see [BS, Ko]
for more details. The Doob decomposition of s
c
(x) is given by the martingale part
V
n
(x) =
n
k=1
d
k
(x)
d
k
(x) E
k1
(d
k
(x)
d
k
(x))
and the predictable part
W
n
(x) =
n
k=2
E
k1
(d
k
(x)
d
k
(x)) L
p
(/
k1
) .
Note that
V
n
(x) + W
n
(x) =
n
k=1
d
k
(x)
d
k
(x) .
Corollary 5.3. Let 2 < p < and x L
p
(N), then
sup
n
max|W
n
(x)|
1
2
p
2
, |V
n
(x)|
1
2
p
2

p
(1 +c
p
2
)
1
2
|x|
p
.
Here
p
is an absolute constant. In particular, there exists an element V
(x) in L
p
2
(N)
such that V
n
(x) = E
n
(V
(x)).
Proof: Using (DD
p
2
) for the sequence (E
k1
) and the non-commutative Burkholder-
Gundy inequality [PX, JX], we deduce
|W
n
(x)|p
2
c
p
2
_
_
_
_
_
n
k=2
d
k
(x)
d
k
(x)
_
_
_
_
_
p
2
c
p
2
2
p
|x|
2
p
.
Hence the triangle implies
|V
n
(x)|p
2
|W
n
(x)|p
2
+
_
_
_
_
_
n
k=1
d
k
(x)
d
k
(x)
_
_
_
_
_
p
2
(c
p
2
2
p
+ 1) |x|
2
p
.
By uniform convexity of L
p
2
(N), we obtain the limit value V
(x) = lim
n
V
n
(x) with the
desired properties.
44 MARIUS JUNGE
The next application yields norm estimates for
n
p
n
E
n
(x)q
n
with respect to a sequence
(p
n
), (q
n
) of disjoint projections. This corresponds to a double sided non-adapted stopping
time.
Corollary 5.4. Let 1 < p , (v
n
), (w
n
) be sequences of bounded elements, then for all
x L
p
(N)
_
_
_
_
_
n
v
n
E
n
(x)w
n
_
_
_
_
_
p
c
p
|x|
p
max
_
_
_
_
_
_
_
_
n
v
n
v
n
_
_
_
_
_
1
2
,
_
_
_
_
_
n
v
n
v
n
_
_
_
_
_
1
2
_
_
_
max
_
_
_
_
_
_
_
_
n
w
n
w
n
_
_
_
_
_
1
2
,
_
_
_
_
_
n
w
n
w
n
_
_
_
_
_
1
2
_
_
_
.
Proof: The case p = is obvious. Hence, we assume 1 < p < . Let y L
p
(N) and
choose y
1
L
2p
(N), y
2
L
2p
(N) such that y = y
1
y
2
and
|y|
p
= |y
1
|
2
2p
= |y
2
|
2
2p
.
Then, according to Lemma 3.4, we deduce from (DD
p
):
tr(
n
v
n
E
n
(x)w
n
y)
n
tr(xE
n
(w
n
yv
n
))
|x|
p
_
_
_
_
_
n
E
n
(w
n
y
1
y
2
v
n
)
_
_
_
_
_
p
|x|
p
c
p
_
_
_
_
_
n
w
n
y
1
y
1
w
n
_
_
_
_
_
1
2
p
_
_
_
_
_
n
v
n
y
2
y
2
w
n
_
_
_
_
_
1
2
p
.
To conclude, we use [PX, Lemma 1.1], see also [JX],
_
_
_
_
_
n
w
n
y
1
y
1
w
n
_
_
_
_
_
p
|y
1
y
1
|
p
max
__
_
_
_
_
n
w
n
w
n
_
_
_
_
_
,
_
_
_
_
_
n
w
n
w
n
_
_
_
_
_
_
= |y|
p
max
__
_
_
_
_
n
w
n
w
n
_
_
_
_
_
,
_
_
_
_
_
n
w
n
w
n
_
_
_
_
_
_
.
A similar argument applies for the other term and therefore taking the supremum over all
y of norm 1 implies the assertion.
Remark 5.5. Let 1 p < . The non-commutative Doob inequality from Theorem 0.2
implies the vector-valued Doob inequality in L
p
(, , ; L
p
(N)).
Proof: It suces to treat the discrete case. Let ^ = L
(, , )
N and (
n
)
nN
be
an increasing sequence of -subalgebras with conditional expectations (E
n
). Let ^
n
=
L
(, , )
N. Then the conditional expectation c

n
onto ^
n
is given by c
n
= E
n

id. Let f L
p
(, , ; L
p
(N)) = L
p
(^). According to Theorem 0.2 there exist a
L
2p
(^), b L
2p
(^) and contractions (z
n
) ^ such that
E
n
id
Lp(N)
(f) = az
n
b and |a|
2p
|b|
2p
c
p
|f|
p
.
Hence, for every and n N
|E
n
(f)()|
Lp(N)
= |a()z
n
()b()|
Lp(N)
|a()|
L
2p
(N)
|z()|
N
|b()|
L
2p
(N)
|a()|
L
2p
(N)
|b()|
L
2p
(N)
.
Holders inequality implies the assertion
_
_
_
sup
n
|E
n
(f)()|
p
Lp(N)
d()
_
_
1
p
_
_
_
|a()|
p
L
2p
(N)
|b()|
p
L
2p
(N)
d()
_
_
1
p
|a|
L
2p
(,,;L
2p
(N)
|b|
L
2p
(,,;L
2p
(N)
= |a|
2p
|b|
2p
c
p
|f|
p
.
In the next application we want to relate group actions with (DD
p
). To illustrate this, we
consider a nite von Neumann algebra N and an increasing sequence (A
n
) N of nite
dimensional subalgebras with 1
N
A
n
. Let N
n
= A
n
be the relative commutant of A
n
in
N. If G
n
denotes the unitary group of A
n
, we have a natural action : G
n
B(L
p
(N))
n
(u)(x) = uxu
such that the conditional expectation on the commutant N

n
is given by
E
n
(x) = E
Nn
(x) =
_
Gn
uxu
d
n
(u) .
Let G =
n
G
n
and the product measure, then
_
_
_
_
_
n
E
n
(x
n
)
_
_
_
_
_
p
_
_
_
G
_
_
_
_
_
n
(u
n
)(x
n
)
_
_
_
_
_
p
p
d(u
1
, u
2
, ..)
_
_
1
p
.
We will show that for a sequence (x
n
) of positive elements even the right hand side can be
estimated by |
n
x
n
|
p
. For simplicity let us use the random variables
n
(x) : G L
p
(N),
given for = (u
1
, u
2
...) by
n
(x)() = u
n
xu
n
.
46 MARIUS JUNGE
For the special case of tensor products of nite dimensional von Neumann algebras the
following theorem implies (DD
p
).
Theorem 5.6. Let 1 < p < and N, (A
n
), (N
n
) be as above and (x
n
) be a sequence of
positive elements, then
_
_
_
G
_
_
_
_
_
n
(x
n
)
_
_
_
_
_
p
p
d
_
_
1
p

p
_
_
_
_
_
n
x
n
_
_
_
_
_
p
.
Here
p
is a constant which only depends on p.
Proof: The assertion is obvious for p = 1 and by interpolation as in Lemma 3.2 it suces
to prove the assertion for p 4. Let (x
n
) be a nite sequence of positive elements. Then,
we observe
_
_
_
G
_
_
_
_
_
n
(x
n
)
_
_
_
_
_
p
p
d
_
_
1
p
_
_
_
_
_
n
E
n
(x
n
)
_
_
_
_
_
p
+
_
_
_
G
_
_
_
_
_
n
(x
n
) E
n
(x
n
)
_
_
_
_
_
p
p
d
_
_
1
p
c
p
_
_
_
_
_
n
x
n
_
_
_
_
_
p
+
_
_
_
G
_
_
_
_
_
n
(x
n
) E
n
(x
n
)
_
_
_
_
_
p
p
d
_
_
1
p
.
Let
n
be the -algebra generated by the rst n coordinates in G =
k
G
k
and ^
n
=
L
(G,
n
, )
N L
(G, , )
N with the corresponding conditional expectation c

n
.
Then, we note
c
n1
(
n
(x
n
)) =
_
kn
G
k
u
n
x
n
u
n
d
n
(u
n
)d
n+1
(u
n+1
) = E
n
(x
n
) .
Hence the n-th martingale dierence of
n
(x
n
) satises
d
n
= c
n
(
k
(x
k
)) c
n1
(
k
(x
k
)) =
n
(x
n
) E
n
(x
n
) .
We apply the non-commutative Rosenthal inequality, see [JX], in this case and obtain
1
r
p
_
_
IE
_
_
_
_
_
n
(x
n
) E
n
(x
n
)
_
_
_
_
_
p
p
_
_
1
p
n
|
n
(x
n
) E
n
(x
n
)|
p
p
_1
p
+
_
_
_
_
_
n
c
n1
(d
n
d
n
+ d
n
d
n
)
_
_
_
_
_
1
2
p
2
2
_
n
|x
n
|
p
p
_1
p
+ 2
1
2
_
_
_
_
_
n
E
n
(x
n
x
n
) + E
n
(x
n
x
n
)
_
_
_
_
_
1
2
p
2
2
_
n
|x
n
|
p
p
_1
p
+ 2
1
2
c
1
2
p
2
_
_
_
_
_
n
x
n
x
n
+ x
n
x
n
_
_
_
_
_
1
2
p
2
Let (
n
) be a sequence of independent Rademacher variables. Using the triangle inequality
and the orthogonality of the (
n
)s, we deduce as in [LP]
max
_
_
_
_
_
_
_
_
n
x
n
x
n
_
_
_
_
_
1
2
p
2
,
_
_
_
_
_
n
x
n
x
n
_
_
_
_
_
1
2
p
2
_
_
_
_
_
IE
_
_
_
_
_
n
x
n
_
_
_
_
_
2
p
_
_
1
2
.
By interpolation, we have
_
n
|x
n
|
p
p
_1
p
_
_
_
_
_
n
x
n
x
n
_
_
_
_
_
1
2
p
2
.
However, since the x
n
are positive, we deduce for any choice of signs
n
with positivity
_
_
_
_
_
n
x
n
_
_
_
_
_
p
_
_
_
_
_
n=1
x
n
_
_
_
_
_
p
+
_
_
_
_
_
n=1
x
n
_
_
_
_
_
p
2
_
_
_
_
_
n
x
n
_
_
_
_
_
p
.
Hence, we obtain
_
_
IE
_
_
_
_
_
n
(x
n
) E
n
(x
n
)
_
_
_
_
_
p
p
_
_
1
p
r
p
(2 + 4c
1
2
p
2
)
_
_
_
_
_
n
x
n
_
_
_
_
_
p
.
Remark 5.7. These methods can also be used to show that for every f L
p
(G; L
p
(N))
there exist a, b L
2p
(; L
2p
(N)) and a sequence of contractions (y
n
) L
()
N such
48 MARIUS JUNGE
that
F
n
(f) = ay
n
b and |a|
2p
|b|
p
c
2
p
|f|
p
where
F
n
(f)(g
1
, g
2
, ...) =
n
(g
n
)f(g
1
, ..., g
n
, ..) .
Note that the crossed product N
(n)
G
n
acts on L
p
(; L
p
(N)) and F
n
somehow removes
the action of G
n
on f. Similar results hold for a von Neumann algebra with a faithful
normal state and -invariant, strongly continuous group actions
n
: G
n
Aut(N) of
compact groups such that the centralizer algebras are increasing or decreasing.
References
[AvW] L. Accardi and M. von Waldenfels: Quantum probability and applications; (Proc. 1988)
Springer Lect. Notes 1442, Berlin-Heidelberg-New-York: Springer-Verlag, 1990.
[AMS] N. Asmar and S. Montgomery-Smith: Littlewood-Payley Theory on solenoids; Colloq. Math.
65, (1993), 69-82.
[BL] J. Bergh and J. L ofstr om: Interpolation spaces. An introduction; Grundlehren der Mathe-
matischen Wissenschaften, No. 223. Springer-Verlag, Berlin-New York, 1976.
[BS] P. Biane and R. Speicher: Stochastic calculus with respect to free Brownian motion and
analysis on Wigner space; Probab. Theory Related Fields 112 (1998), no. 3, 373-409.
[BDG] D.L. Burkholder, B. J. Davis and R.F. Gundy : Integral inequalities for convex functions
of operators on martingales; Proceedings of the Sixth Berkeley Symposium on Mathematical
Statistics and Probability (Univ. California, Berkeley, Calif., 1970/1971), Vol. II: Probability
theory, pp. 223-240. Univ. California Press, Berkeley, Calif., 1972.
[C] A. Connes: Une classication des facteurs de type III; Ann. Sci. Ec. Norm. Super., IV. Ser. 6,
133-252 (1973).
[Cu] I. Cuculescu: Martingales on von Neumann algebras; J. Mult. Anal. 1 (1971), 17-27.
[Di] S.J. Dilworth; Some probabilistic inequalities with applications to functional analysis; (Banach
spaces Merida, 1992), Contemp. Math., 144,53-67, Amer. Math. Soc., Providence, RI, 1993
[DJ1] A. Defant and M. Junge: Almost everywhere convergence of series in non-commuative L
p
-
spaces; preprint
[DJ2] A. Defant and M. Junge: Almost sure convergence and Haagerup L
p
spaces; in preparation
[DF] A. Defant and A. Floret: Tensor norms and operator ideals; North-Holland Mathematics
Studies, 176. North-Holland Publishing Co., Amsterdam, 1993.
[Dx] J. Dixmier: Von Neumann Algebras; North-Holland Amsterdam-New-York-Oxford 1981
[Fi] F. Fidaleo: Canonical operator space structures on non-commutative L
p
spaces; J. Funct. Anal.
169 (1999), no. 1, 226-250.
[Fa] T. Fack: Type and cotype inequalities for non commutative L
p
-spaces; J. Oper. Theory 17,
255-279 (1987).
[Ga] A. Garsia:Martingale Inequalities: Seminar Notes on recent progress; AMS Math. lect. notes
1973
[Ha1] U. Haagerup:L
p
-spaces associated with an arbitrary von Neumann algebra; Algebres
doperateurs et leurs applications en physique mathematique (Proc. Colloq., Marseille, 1977),
pp. 175-184, Colloq. Internat. CNRS, 274, CNRS, Paris, 1979.
[Ha2] U. Haagerup:Injectivity and decomposition of completely bounded maps; in Operator alge-
bras and their connection with Topology and Ergodic Theory, Springer Lect. Notes in Math.
1132(1985), 170-222.
[Ja1] R. Jajte: Strong limit theorems in noncommutative probability Lect. Notes Math. Vol.1110,
Springer-Verlag, Berlin, 1985.
[Ja2] R. Jajte: Strong limit theorems in noncommutative L
2
-spaces; Lect. Notes Math. Vol.1477,
Springer-Verlag, Berlin, 1991.
[JX] M. Junge and Q. Xu: Non-commutative Rosenthal/Burkholder inequalities; preprint
[Ki] E. Kirchberg: On nonsemisplit extensions, tensor products and exactness of group C
-algebras;
Invent. Math. 112 (1993), no. 3,
[KR] R. Kadison and J. Ringrose: Fundamentals of the theory of operator algebras I and II; Grad-
uate Studies in Mathematics 15 & 16, AMS, Providence, RI, 1997
[KF] T. Fack and H. Kosaki: Generalized s-numbers of -measurable operators, Pacic-J.-Math.
123 (1986), no. 2, 269-300.
[Ko] H. Kosaki: Applications of the Complex Interpolation Method to a von Neumann Algebra: Non-
commutative L
p
-spaces; J. of Funct. Anal. 56, 29-78(1984)
[Ko] Claus K oster: Quanten-Marko-Prozesse und Quanten-Brownsche Bewegungen; PhD-thesis
Stuttgart 2000.
[La] E. C. Lance: Hilbert C
-modules. A toolkit for operator algebraists; London Mathematical So-

ciety Lecture Note Series, 210. Cambridge University Press, Cambridge, 1995.
[LP] F. Lust-Piquard: Inegalites de Khintchine dans C
p
(1 < p < ); C. R. Acad. Sci., Paris, Ser. I
303, 289-292 (1986).
[Me] P.A. Meyer: Quantum probablity for probabilists: Lect. Notes in Math. Berlin-Heidelberg-New-
York: Springer-Verlag, 1995.
[Ne] E. Nelson: Notes on non-commutative integration; J. Funct. Anal. 15 (1974), 103-116.
[Pa] V. I. Paulsen: Completely bounded maps and dilations; Pitman Research Notes in Mathematics,
1986.
[PP] M. Pimsner and S. Popa : Entropy and index for subfactors; Ann. Scient Ec. Norm. Sup. 19
(1986), 57-106.
[PTA] G. Pedersen and M. Takesaki: The Radon-Nikodym theorem for von Neumann algebras; Acta
Math. 130 (1973), 53-87.
[Ps1] G. Pisier: Factorization of linear operators and Geometry of Banach spaces; CBMS (Regional
conferences of the A.M.S.) no. 60 (1986), Reprinted with correction 1987.
[Ps2] G. Pisier: Non-commutative vector valued L
p
-spaces and completely p-summing maps; Asterisque
No. 247, (1998).
[Ps3] G. Pisier: An inequality for p-orthogonal sums in non-commutative Lp; preprint
[PX] G. Pisier and Q. Xu : Non-commutative martingale inequalities; Comm. Math. Phys. 189
(1997), no. 3, 667-698.
[Te] M. Terp: L
p
-spaces associated with von Neumann algebras I and II; Copenhagen Univ. 1981.
[Tk] M. Takesaki: Theory of operator algebras I; Springer Verlag New York, 1979.
[VDN] D. Voiculescu, K. Dykema and A. Nica: Free random variables. A noncommutative probabil-
ity approach to free products with applications to random matrices, operator algebras and harmonic
analysis on free groups; CRM Monograph Series, 1. American Mathematical Society, Providence,
RI, 1992.
[Ye] F. Yeadon: Non-commutative L
p
-spaces; Math.-Proc.-Cambridge-Philos.-Soc. 77 (1975), 91-102.
2000 Mathematics Subject Classication: 46L53, 46L52, 47L25.
Key-words: Non-commutative L
p
-spaces, Doobs inequality
Department of Mathematics, University of Illinois at Urbana-Champaign, Urbana, IL
61801
E-mail address: jungemath.uiuc.edu

Martingales Junge

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Martingales Junge

Uploaded by

Copyright:

Available Formats

a

Abstract. Let 1 p < and (x

) we can overcome this problem and at least guess the right

Junge is partially supported by the National Science Foundation.

-modules and non-commutative L

: N M yields a conditional expectation satisfying

-algebras /, B, we denote by /B the

the unique predual of N) dened by

such that for

w R with respect to the modular automorphism group of a n.s.f.

w R is dened as the von Neumann

w R such that for all s R

w R. Following [Te, Proposition 33, Theorem 32], we deduce

= 1. We show the density of D

-modules with Steins

and hence has a predual

. In order to simplify the exposition, we will

(D) = D, and for all x M, y M

be the corresponding density then

)) = E(1) = (E(1)) = (1) = tr(a

N converges to x strongly, then

to nd a partial isometry u such that

converges to x strongly then x

[. This immediately implies that x

topology. Hence, the space of nite sequences is dense

-algebra / is a right /-module / together with an /-valued sesquilinear form

-algebra / is the space H

N which consists of sequences (x

(M, E) the completion

-algebra and let F be the Hilbert

-module generated by N and X. Then F is a countably generated Hilbert N-module

(M, E) C(N) B(L

(M, E). Let M =

(M, E) is the norm in

(M, E). The next proposition

(M, E) is isomorphic to a complemented module in C(N). In particular,

(M, E) and completely contractive module maps u

(M, E) there exists a net x

in the unit ball of L

topology and Q extends to a projection onto the image of L

(M, E) C(N), C(N), can be approximated with respect

(M, E). Then, we follow the proof of Kaplanskys

([Tk, Theorem 2.6]) can be approximated by a net f(x

) in the unit ball of F. Since

-dense in the unit ball of M. Using Lemma 2.3, we

(M, E) of norm less than one. Then there exists a net a

converges to y with respect to the strong

topology. Then the

y) follows from Lemma 2.3. In

(M, E) induces a contractive multiplier

for a contraction v and a

< such that

separable, then there is an isomet-

< such that

is separable. Then rst assertions follow from Proposition

N, see [Pa]. Indeed, a normal map is decomposable if and only

is separable and use the Kasparov maps u

N), respectively. Using,

) space, we dene for 0 < p

(operator space projective tensor product) only holds for

= g[ supg < 0. Here v = (v

= 1 assume that (DD

and all the y