In which spaces is every curve Lebesgue-Pettis integrable?

0.06 MB·English

by  Heinrich von Weizsäcker

In which spaces every curve is Lebesgue-Pettis integrable?
Heinrich v. Weizsa¨cker
Kaiserslautern, July 26, 2012
Dedicated to the memory of Erik G. F. Thomas
2
1
0 1. Introduction. Around 1980 P. Dierolf und J. Voigt raised the question which locally convex
2
real vector spaces E are sufficiently ’complete’ to have the property in the title. The following text

l
u is a revised translation of an unpublished manuscript of that time.
J
Our primary result is that the property in the title (for continuous curves) is equivalent to the ’metric

5
convex compactness property’ in the sense of Voigt (1992, [2]): the closed convex hull of a compact
2
metrizable subset is again compact. There ([2],Theorem 0.1) an observation of H. Pfister (1981) is

]
A reproduced to the effect that the metric compactness property is equivalent to the Pettis integrability
F of all E-valued continuous functions on compact metric spaces with a Borel measure µ. So our main
. point is that it suffices to restrict the attention to the unit interval and the Lebesgue measure. In
h
t [2] the metric convex compactness property is compared to other completeness properties like the
a
m so-called Mackey-completeness; moreover in that paper it is shown that various spaces of linear
[ operators, e.g. the compact operators in a Banach space, do have this property with respect to the

strong operator topology.
1
v We also show that the property in the title is equivalent to another property which, in view of earlier
4
work of E. G. F. Thomas (1975, e.g. [1]), appears to be the right condition to allow for a satisfactory
3
0 extension of the Bochner integral beyond the domain of Banach space valued functions.
6
.
7 2. The main result.
0
2
1 Theorem 1 A real locally convex Hausdorff space E has the metric compactness property if and
v: only if for every continuous map f : [0,1] → E the Pettis integral 1f(t) dt exists in E.
0
i R
X
Proof: We first remark that a Borel map from a toplogical space X into E is ν-Pettis integrable for
r
a a Borel measure ν on X if and only if the image measure µ = ν ◦f−1 has a barycenter (resultant)
r(µ) in E. In fact
he′,r(µ)i = he′,xi dµ = he′,f(t)i dµ(t)
Z Z
X
for all e′ ∈ E′ and thus the Pettis integral of f if it exists is indeed just this barycenter.
1. Assume the metric compactness property. If f : [0,1] → E is continuous then the image set
K = f([0,1]) is compact und metrizable. Its closed convex hull is compact by assumption. Hence
the image of the Lebesgue measure λ on [0,1] has a barycenter and thus f is λ-Pettis integrable by
the preceding remark.
2. Conversely assume that every continuous curve f : [0,1] → E is λ-Pettis integrable. Combining
Pfister’s remark (Theorem 0.1 in [2], see the introduction) and the above it suffices to show that
every probability measure µ on E with compact metrizable support K has a barycenter in E.
a. First we assume that µ = σ◦g−1 for a continuous map g : [0,1] → E and a non-atomic probability
measure σ on [0,1] with supp σ = [0,1]. Let F(t) := σ((0,t]). Then F is strictly increasing and σ is
the image of λ under the continuous map F−1. Therefore µ = λ◦(g ◦F−1)−1. Thus the barycenter
of µ is just the Pettis integral 1g(x) dσ(x) = 1g(F−1(t)) dt which exists by assumption.
0 0
R R
b. Next we drop the support condition on σ in the previous argument: Suppose that µ = ρ ◦ g−1
for a non-atomic probability measure ρ on [0,1]. Consider σ = 1(ρ+λ). Then supp σ = [0,1] and
2
ρ = 2σ −λ. By the previous step
1 1 1
r(µ) = g(x) dρ = 2 g(x) dσ − g(x) dx ∈ E.
Z Z Z
0 0 0
c. Now let µ be a non-atomic probability measure with compact metrizable support K ⊂ E. Then
there is a continuous map from the Cantor set D onto K. By linear interpolation in the intervals
which constitute the complement of the Cantor set this map can be extended to a continuous map
g : [0,1] → K. The measure ρ on [0,1] with distribution function F(t) = µ(g([0,t]) is non-atomic
with image measure µ. Thus according to the previous step r(µ) ∈ E.
d. Finally let µ be an arbitrary probability measure with compact metrizable support K ⊂ E. The
discrete part of µ can be written in the form µ = ∞ a δ as an infinite combination of point
d i=1 i xi
masses δ . Without loss of generality we may assuPme that x = 0 for no i. We replace each of
xi i
these point masses by the uniform distribution on the vector interval [1x , 3x ]. Then the resulting
2 i 2 i
probability measure is non-atomic, and it is concentrated on the compact metrizable set [1, 3]K.
2 2
Thus it has a barycenter and this barycenter clearly is also the barycenter of the original µ. This
concludes the proof. q.e.d.
3. An Extension
We call a positive Borel-measure µ on a topological space X metrically regular, if
µ(B) = sup{µ(K) : K ⊂ B, K compact and metrizable}
holds for all Borel-sets B. A function g is called µ-Lusin-measurable, if for every Borel-set D ⊂ X
we have µ(D) = sup{µ(K) : K ⊂ D compact, g continuous}. Clearly the image measure of a finite
|K
metrically regular µ under a Lusin-measurable function is again metrically regular.
Theorem 2 A real locally convex Hausdorff space E has the metric compactness property if and
only if the following holds:
Let X be a topological space and let ν be a metrically regular Borel measure on X. Let B ⊂ E be
a bounded closed absolutely convex set with the associated gauge functional ϕ : E → R . Let the
B +
function g : X → E be ν-Lusin-measurable with ϕ ◦ g dν < ∞. Under these conditions g is
X B
ν-Pettis-integrable. R
Proof: Let us remark first that the gauge functional of a bounded closed absolutely convex set B
due to Hahn-Banach has the representation
ϕ (e) = sup{he′,ei : e′ ∈ E′,suphe′,xi ≤ 1},
B
x∈B
hence it is lower semi-continuous and in particular Borel measurable.
Assume now that the condition of the theorem holds. Then indeed for every Borel measure ν on
a compact metric space X every continuous E-valued funktion is Pettis-integrable. As expalained
above this implies the metric compactness property of E.
Conversely assume that the metric compactness property is satisfied and let X, ν, B and g be given
as indicated. The real valued function defined by h(x) := ϕ (g(x)) is ν-Lusin-measurable and even
B
ν-integrable. Replacing, if necessary, dν by hdν and g by g · 1 , we may and will assume that
h {h>0}
ν(X) < ∞ and that g(X) ⊂ B. The image measure µ = ν ◦ g−1 is finite, metrically regular and
concentrated on B.
In order to prove the integrability of g it thus suffices to show that each finite, metrically regular
previous discussion. In the general case, let K ⊂ B,n ∈ N be disjoint compact metrizable sets such
n
that µ(K ) = µ(E) < ∞. There is a sequence (b ) of positive numbers with b → 0 sufficiently
n n n n
slowlyPsuch that ∞ µ(Kn) < ∞. Consider the measure µ˜ defined by
n=1 bn
P
∞ µ(b−1A∩K )
µ˜(A) = n n .
b
X n
n=1
This finite metrically regular measure is concentrated on the set K = b K ∪ {0}. This set is
n n n
compact and metrizable; in fact K is bounded and b → 0 and henceSthe sets U = b K ∪
n n n m n≥m n n
{0} form a base of neighbourhoSods of the point 0 in the relative topology of K. TShus µ˜ has a
barycenter and this barycenter is also the barycenter of µ because for each e′ ∈ E′ we have
∞ ∞
1
e′ dµ˜ = he′,b xi dµ = he′,xi dµ = e′ dµ.
Z b Z n Z Z
Xn=1 n Kn Xn=1 Kn
q.e.d.
Remark The function g in the theorem is in an obvious way equivalent to a Bochner integrable
function with values in the canonical Banach space associated to the seminorm ϕ . This justifies
B
the last remark in our introduction.
Acknowledgement. The author is indebted to J. Voigt for helpful comments and for quoting the
original manuscript in his 1992 paper [2]; moreover to H. Gloeckner, whose curiosity recently was
raised by that citation. This renewed interest triggered the present version of this text.
References
[1] E.G.F. Thomas. Integration of Functions With Values in Locally Convex Suslin Spaces. Trans.
Am. Math. Soc., 212:61-81, 1975.
[2] J. Voigt. On the convex compactness property for the strong operator topology. Note Mat.,
12:259-269, 1992.