Lp -SPACES FOR 0 < p < 1

KEITH CONRAD

1. Introduction
In a first course in functional analysis, a great deal of time is spent with Banach spaces,
especially the interaction between such spaces and their dual spaces. Banach spaces are a
special type of topological vector space, and there are important topological vector spaces
that do not lie in the Banach category, such as the Schwartz spaces.
The most fundamental theorem about Banach spaces is the Hahn-Banach theorem, which
links the original Banach space with its dual space. What we want to illustrate here is a
wide collection of topological vector spaces where the Hahn-Banach theorem has no obvious
extension because the dual space is zero. The model for a topological vector space with
zero dual space will be Lp [0, 1] when 0 < p < 1. After proving the dual of this space is {0},
well see how to make the proof work for other Lp -spaces, with 0 < p < 1. The argument
eventually culminates in a pretty theorem from measure theory (Theorem 4.2) that can be
understood at the level of a first course on measures and integration.

2. Banach Spaces and Beyond

In this section, to provide some context, we recall some basic classes of vector spaces that
are important in analysis.
Throughout, our vector spaces are real vector spaces. All that we say would go through
with minimal change to complex vector spaces.
Definition 2.1. A norm on a vector space V is a function || || : V R satisfying
||v|| 0, with equality if and only if v = 0,
||v + w|| ||v|| + ||w|| for all v and w in V .
||cv|| = |c|||v|| for all scalars c and v V .
Given a norm on a vector space, we get a metric by d(v, w) = ||v w||.
Example 2.2. On Rn , we have the sup-norm |x|sup = max1in |xi | and the L2 -norm
|x|2 = x x = ( ni=1 |xi |2 )1/2 . The metric coming from the L2 -norm is the usual notion of
P
distance on Rn . The metric on Rn coming from the sup-norm has balls that are actually
cubes. These two norms give rise to the same topology on Rn . Actually, all norms on Rn
yield the same topology.
Example 2.3. The space C[0, 1] of continuous real-valued functions on [0, 1] has the sup-
R1
norm |f |sup = supx[0,1] |f (x)| and the L2 -norm |f |2 = ( 0 |f (x)|2 dx)1/2 . While functions
that are close in the sup-norm are close in the L2 -norm, the converse is false: a function
whose graph is close to the x-axis except for a tall thin spike is near 0 in the L2 -norm but
not in the sup-norm.
Definition 2.4. A Banach space is a vector space V equipped with a norm || || such that,
with respect to the metric defined by d(v, w) = ||v w||, V is complete.
1
2 KEITH CONRAD

Example 2.5. Under either norm in Example 2.2, Rn is a Banach space.

Example 2.6. In the sup-norm, C[0, 1] is a Banach space (convergence in the sup-norm is
exactly the concept of uniform convergence). But in the L2 -norm, C[0, 1] is not a Banach
space. That is, C[0, 1] is not complete for the L2 -norm.
Definition 2.7. For a Banach space V , its dual space is the space of continuous linear
functionals V R, and is denoted V .
Continuity is important. We do not care about arbitrary linear functionals (as in linear
algebra), but only those that are continuous. One of the important features of a Banach
space is that we can use continuous linear functionals to separate points.
Theorem 2.8. Let V be a Banach space. For each non-zero v V , there is a V such
that (v) 6= 0. Thus, given distinct v and w in V , there is a V such that (v) 6= (w).
Theorem 2.8 is a special case of the Hahn-Banach theorem, and can be found in any text
on functional analysis. Even this special case cant be proved in a constructive way (when
V is infinite-dimensional). Its general proof depends on the axiom of choice.
Example 2.9. On C[0, 1], the evaluation maps ea : f 7 f (a), for a R, are linear func-
tionals. Since |f (a)| |f |sup , each ea is continuous for the sup-norm. If f 6= 0 in C[0, 1],
there is some a such that f (a) 6= 0, and then ea (f ) = f (a) 6= 0.
Using the sup-norm topology on C[0, 1], the dual space C[0, 1] is the space of bounded
Borel measures (or Riemann-Stieltjes integrals) on [0, 1], with the evaluation maps ea cor-
responding to point masses.
Definition 2.10. A topological vector space is a (real) vector space V equipped with a
Hausdorff topology in which addition V V V and scalar multiplication R V V
are continuous.
Note the Hausdorff condition is included in the definition. We wont be meeting any
non-Hausdorff spaces.
Example 2.11. The usual topology on Rn makes it a topological vector space. So does
the discrete topology on Rn . The usual topology on Rn is the only nondiscrete Hausdorff
topology that makes it a topological vector space. (This is not trivial. Try the case of
dimension 1.)
Example 2.12. Any Banach space is a topological vector space.
Definition 2.13. A subset of a vector space is called convex if, for any v and w in the
subset, the line segment tv + (1 t)w, for 0 t 1, is in the subset.
More
Pmgenerally, if a subset is convex
Pm and v1 , . . . , vm are in the subset, then any weighted
sum c v
i=1 i i with ci 0 and i=1 ci = 1 is in the subset. In particular, the subset
contains the average (1/m) m
P
v
i=1 i .
Definition 2.14. A topological vector space is called locally convex if the convex open sets
are a base for the topology: given any open set U around a point, there is a convex open
set C containing that point such that C U .
Example 2.15. Any Banach space is locally convex, since all open balls are convex. This
follows from the definition of a norm.
 Lp -SPACES FOR 0 < p < 1 3

Since topological vector spaces are homogeneous (we can use addition to translate neigh-
borhoods around one point to neighborhoods around other points), the locally convex con-
dition can be checked by focusing at the origin: the open sets around 0 need to contain a
basis of convex open sets.
Example 2.16. The space C(R) of continuous real-valued functions on R does not have
the sup-norm over all of R as a norm: a continuous function on R could be unbounded.
But C(R) can be made into a locally convex topological vector space as follows. For each
positive integer n, define a semi-norm | |n by
|f |n = sup |f (x)|.
|x|n

This is just like a norm, except it might assign value 0 to a non-zero function. That is,
a function could vanish on [n, n] without vanishing everywhere. Of course, if we take n
large enough, a non-zero continuous function will have non-zero n-th semi-norm, so the
total collection of semi-norms | |n as n varies, rather than one particular semi-norm, lets
us distinguish different functions from each other. Using these semi-norms, define a basic
open set around g C(R) to be all functions close to g in a finite number of semi-norms:
{f C(R) : |f g|n1 , . . . , |f g|nr }
for a choice of finitely many semi-norms and an . These sets are a basis for a topology
that makes C(R) a locally convex topological vector space.
Definition 2.17. When V is a topological vector space, its dual space V is the space of
continuous linear functionals V R.
Theorem 2.8 generalizes to all locally convex spaces, as follows.
Theorem 2.18. Let V be a locally convex topological vector space. For any distinct v and
w in V , there is a V such that (v) 6= (w).
Proof. See the chapters on locally convex spaces in [3] or [9]. 

Lets meet some topological vector spaces that are not locally convex.
R1
Example 2.19. Let L1/2 [0, 1] be the measurable f : [0, 1] R such that 0 |f (x)|1/2 dx
is finite, with functions equal almost everywhere identified. (We need to make such an
identification, since integration does not distinguish between functions that are changed on
a set of measure 0.)
R1
The function d(f, g) = 0 |f (x) g(x)|1/2 dx is a metric on L1/2 [0, 1]. The topology
L1/2 [0, 1] obtains from this metric is not locally convex. To see why, consider any open ball
around 0:  Z 1 
1/2 1/2
f L [0, 1] : |f (x)| dx < R .
0
We will show, for any > 0, that the -ball around 0 contains functions whose average lies
outside the ball of radius R. That violates the meaning of local convexity.
Pick > 0 and n 1. Select n disjoint intervals in [0, 1] (they need not cover all of [0, 1]).
Call them A1 , A2 , . . . , An . Set fk = (/(Ak ))2 Ak , where is Lebesgue measure (so (Ak )
R1
is simply the length of Ak ). Then 0 |fk (x)|1/2 dx = , so every fk is at distance from
4 KEITH CONRAD

Pn
0. However, since the fk s are supported on disjoint sets, their average gn = (1/n) k=1 fk
satisfies Z 1 n Z
1/2 1 X 1
|gn (x)| dx = 1/2 |fk (x)|1/2 dx = n1/2 .
0 n k=1 0

Taking n large enough (depending on ), we see the distance between gn and 0 in L1/2 [0, 1]
can be made as large as desired.
Example 2.20. For 0 < p < 1, let Lp [0, 1] be the set of measurable functions f : [0, 1]
R1
R such that 0 |f (x)|p dx < , with functions equal almost everywhere identified. The
R1
function d(f, g) = 0 |f (x) g(x)|p dx is a metric on Lp [0, 1]. With the inherited metric
topology, Lp [0, 1] is not locally convex, by exactly the same argument as in the previous
example (where p = 1/2). Indeed, in the previous construction, replace the exponent 2
R1
in the definition of fk with 1/p, and at the end youll get 0 |gn (x)|p dx = n1p . Since
1 p > 0, the distance between gn and 0 can be made arbitrarily large with suitable choice
of n. In fact, what this means is that the only convex open set in Lp [0, 1] is the whole space.
This method of constructing topological vector spaces that are not locally convex is due
to Tychonoff [10, pp. 768769], whose actual example was the sequence space `1/2 = {(xi ) :
P 1/2 [0, 1].
i1 xi < } rather than the function space L
One cant push a result like Theorem 2.18 to all topological vector spaces, as the next
result [4] vividly illustrates.
Theorem 2.21. For 0 < p < 1, Lp [0, 1] = {0}. That is, the only continuous linear map
Lp [0, 1] R is 0.
Proof. We argue by contradiction. Assume there is Lp [0, 1] with 6= 0. Then has
image R (a nonzero linear map to a one-dimensional space is surjective), so there is some
f Lp [0, 1] such that |(f )| 1.
Using this choice of f , map [0, 1] to R by
Z s
s 7 |f (x)|p dx.
0
This is continuous, so there is some s between 0 and 1 such that
Z s
1 1
Z
p
(2.1) |f (x)| dx = |f (x)|p dx > 0.
0 2 0
Let g1 = f [0,s] and g2 = f (s,1] , so f = g1 + g2 and |f |p = |g1 |p + |g2 |p . So
Z 1 Z s
1 1
Z
p p
|g1 (x)| dx = |f (x)| dx = |f (x)|p dx,
0 0 2 0
R1 p 1 1
R p
hence 0 |g2 (x)| dx = 2 0 |f (x)| dx. Since |(f )| 1, |(gi )| 1/2 for some i. Let
R1 R1 R1
f1 = 2gi , so |(f1 )| 1 and 0 |f1 (x)|p dx = 2p 0 |gi (x)|p dx = 2p1 0 |f (x)|p dx. Note
2p1 < 1.
Iterate this to get a sequence {fn } in Lp [0, 1] such that |(fn )| 1 and
Z 1 Z 1
p p1 n
d(fn , 0) = |fn (x)| dx = (2 ) |f (x)|p dx 0,
0 0
a contradiction of continuity of . 
 Lp -SPACES FOR 0 < p < 1 5

3. More Spaces with Dual Space 0

We recall the definition of Lp -spaces for any measure space and then extend Theorem
2.21 quite generally.
Definition 3.1. Let (X, M, ) be a measure space. For any p > 0, set
Z
def
Lp () = {f : X R : f measurable and |f |p d < },
X
with functions that are equal almost everywhere being identified with one another.
The metric used on Lp () is
( R
1/p
X |f g|p d , if p 1,
d(f, g) = R p
X |f g| d, if 0 < p < 1.
The reason for these choices, and a discussion of common properties of Lp () for all p > 0,
is discussed in the appendix.
The spaces Lp () for p 1 have their metric coming from a norm, so they are locally
convex. We saw in Examples 2.19 and 2.20 that, for 0 < p < 1, Lp [0, 1] is not locally convex.
For a measure space (X, M, ) and 0 < p < 1, is Lp () ever locally convex?
Theorem 3.2. For 0 < p < 1, Lp () is locally convex if and only if the measure assumes
finitely many values.
Proof. If takes finitely many values, then X is the disjoint union of finitely many atoms,
say B1 , . . . , Bm . A measurable function is constant almost everywhere on each atom, so
Lp () is topologically just Euclidean space (of dimension equal to the number of atoms of
finite measure), which is locally convex.
Now assume takes infinitely many values. We will extend the idea from Example 2.19
to show Lp () is not locally convex.
Since has infinitely many values, there is a sequence of subsets Yi X such that
0 < (Y1 ) < (Y2 ) < . . . .
From the sets Yi , we can construct recursively a sequences of disjoint sets Ai such that
(Ai ) > 0.
Fix > 0. Let fk = (/(Ak ))1/p Ak , so X |fk |p d = . If Lp () is locally convex, then
R
any open set around 0 contains a convex open set around 0, which in turn contains some
-ball (and thus every fk ). We will show an average of enough fk s is arbitrarily far from 0
in the metric onP Lp (), and that will contradict local convexity.
Let gn = n nk=1 fk . Since the fk s are supported on disjoint sets, X |gn |p d =
1
R
1 Pn 1p . Since p < 1, n1p becomes arbitrarily large as n . Thus,
np k=1 = n
p
L () is not locally convex. 
In Theorem 2.21, we saw Lp [0, 1] = {0}. Is the dual of Lp (), for general measure spaces
and 0 < p < 1, also {0} (when Lp () is not locally convex)?
First we treat the boring case.
Theorem 3.3. If the measure contains an atom with finite measure, then Lp () 6= {0}.
Proof. Let B be an atom with finite measure. Any measurable function f : X R is
constant almost everywhere on B. Call the almost-everywhere common value (f ). The
reader can check is a non-zero continuous linear functional on Lp (). 
6 KEITH CONRAD

Atoms with infinite measure are invisible as far as integration is concerned (integrable
functions must vanish on them), so we may assume our measure space has no atoms of
infinite measure.
What remains is the case of a nonatomic measure, and Theorem 2.21 generalizes to this
case as follows.
Theorem 3.4. If (X, M, ) is a nonatomic measure space and 0 < p < 1, then Lp () = 0.
Lets try to prove Theorem 3.4 by understanding more conceptually the proof that
Lp [0, 1] = 0. To make that proof work in the setting of Theorem 3.4, suppose there is
a non-zero Lp () . Then, by scaling, there is an f Lp () such that |(f )| 1. Wed
it like to find some A M such that
Z Z
p 1
(3.1) |f | d = |f |p d.
A 2 X
Then set g1 = f A and g2 = f XA , and repeating the ideas of the proof of Theorem 2.21
will eventually yield a contradiction, so Lp () = 0.
But how do we construct A (depending on f ) that makes (3.1) correct? When X = [0, 1],
as in Theorem 2.21, we considered the collection of integrals
Z s
|f |p dx
0

as s runs from 0 to 1. These are integrals over the sets [0, s], which can be thought of as a
path of intervals in the space of measurable subsets of [0, 1], starting with {0} and ending
with the whole space [0, 1]. The Intermediate Value Theorem helped us find an s such that
Rs p R1 p
0 |f | dx = (1/2) 0 |f | dx.
To extend this idea to any measure space (X, M, ) in place of [0, 1], note M can topol-
ogized by the semi-metric d (A, B) = (AB), where is the symmetric difference oper-
ation:
(3.2) AB = A B A B
(For instance, d (A, ) = (A).) With this metric, integration over measurable subsets of
X defines a function M R by
Z
(3.3) f (A) = |f |p d,
A
that is continuous.
For each f Lp (), suppose we can find a path (continuous function) h : [0, 1] M,
depending on f , such that
Z Z Z
p p
(3.4) |f | d = 0, |f | d = |f |p d.
h(0) h(1) X

Then composing h with f gives the function

Z
s 7 |f |p d,
h(s)

which is continuous from [0, 1] to R. We can apply theRIntermediate Value RTheorem to this
continuous function of s and thus find an s such that h(s) |f |p d = (1/2) X |f |p d.
 Lp -SPACES FOR 0 < p < 1 7

However, we havent explained how to construct a path in M satisfying (3.4). It is not

clear (to me) how to define such a path in M in general. We will use another approach, but
will return to this path idea at the end of the paper.
The key step to finish this proof of Theorem 3.4 is to reformulate the problem more
directly in terms of pure measure theory.

4. Values of a Measure
We start with a nonatomic measure space (X, M, ), and p (0, 1). For each f Lp (),
the map f : M R given by (3.3) defines a finite measure on X (even if (X) = ). We
want to find an A M such that f (A) = (1/2)f (X), i.e., A cuts X in half as far as f is
concerned.
Notice that f depends not so much on f , but on |f |p , and this is a function in L1 (, R).
Lemma 4.1. If (X, M, ) is a nonatomic measure space, then for any nonnegative F
L1 (, R), F d is a finite nonatomic measure.
R
Proof. Without loss of generality, X F d > 0. Since F is measurable,
Z Z
F d = sup gd,
X 0gF X
where the sup is over simple g. For 0 g F , g is a step map, say
X
g= i Ai
with i 0 and 0 < (Ai ) < . R R
Assume F d has an atom A, Rso = A F d > 0 and for B A, B F d = 0 or . Our
goal is to get the contradiction A F d = 0. Since A is an atom for F A , without loss of
generality F = 0 off of A, so Ai A for all i. The continuity of integration as a function of
the set implies Z
lim F d = 0.
(B)0 B

R i ) < , Ai contains a subset B with arbitrarily

Since is nonatomic and 0 < (A R small
positive measure. For such B, B F d = 0 since B Ai A. Therefore B gd = 0, so
g = 0 Ralmost everywhere
R on B, so i = 0 since (B) > 0. This is true for all i, so g = 0.
Thus X F d = 0, so A F d = 0, a contradiction. 
Thus, for f Lp (), the measure f in (3.3) is a finite nonatomic measure on X.
Theorem 4.2. If (X, M, ) is any finite nonatomic measure space, then for all t in [0, (X)]
there is some A M such that (A) = t.
Before proving Theorem 4.2, we note that Lemma 4.1 and Theorem 4.2 help us fill in the
hole in our proof of Theorem 3.4. Indeed, when (X, M, ) is nonatomic and f Lp (, R),
Lemma 4.1 tells us |f |p d is a finite
R nonatomic Rmeasure on X, and Theorem 4.2 then tells
us there is an A M such that A |f |p d = 12 X |f |p d. Using this result and iteration,
we can show Lp () = 0 for 0 < p < 1 by the ideas in the proof of Theorem 2.21.
Remark 4.3. There are theorems that say a finite nonatomic measure space (X, M, )
often looks like the Borel measurable subsets of [0, (X)] with Lebesgue measure. Then
the element of M corresponding to [0, t] solves Theorem 4.2. However, such theorems need
X or M to be suitable separable metric spaces. For our purposes, we do not need such
8 KEITH CONRAD

hypotheses, since all we are asking about is the set of values of the measure , not the
structure of X or M.
The proof of Theorem 4.2 we now give is taken from [8, Exercise 18-28, p. 247].
Proof. Let S1 = {A M : (A) < t}. Choose A1 S1 such that (A1 ) > sup{(A) : A
S1 } 1.
Let S2 = {A M : (A) < t, A1 A}. Choose A2 S2 such that (A2 ) > sup{(A) :
A S2 } 1/2.
Assuming Sn is defined, choose An Sn such that
(An ) > sup{(A) : A Sn } 1/n.
Then let Sn+1 = {A M : (A) < t, An A}.
Now we have A1 A2 A3 X. Define
A = An M,
so
(4.1) (A ) = lim (An ) t.
The rest of the proof is devoted to showing (A ) = t.
If A A and (A) < t, then A Sn+1 . Thus (An+1 ) > (A) 1/(n + 1), so
(A ) (A), so (A) = (A ). We have shown (A ) t and any measurable set
containing A has measure (A ) if its measure is less than t.
Having constructed A by a recursion from below, we now approach A by a recursion
from above. Let T1 = {B M : (B) > t, A B}. (For instance, X T1 .) Choose
B1 T1 such that
(B1 ) < inf{(B) : B T1 } + 1.
Let T2 = {B M : (B) > t, A B B1 }. If Tn is defined, choose Bn Tn such that
(Bn ) < inf{(B) : B Tn } + 1/n.
Let Tn+1 = {B M : (B) > t, A B Bn }. Repeat this for all n, so
A B3 B2 B1 .
Let B = Bn M, so
(4.2) (B ) = lim (Bn ) t.
If A B B and (B) > t, then B Tn+1 , so
(Bn+1 ) < (B) + 1/(n + 1),
so (B ) (B), so (B) = (B ).
Putting together what we have found for sets between A and B ,
A A, (A) < t = (A) = (A ),
A B B , (B) > t = (B) = (B ).
We use this to show (B A ) = 0, which implies (A ) = t and will finish the proof.
Notice we have yet to use the fact that is nonatomic.
Let Y B A , so
A Y A B , Y A = .
 Lp -SPACES FOR 0 < p < 1 9

If (Y A ) < t then (Y A ) = (A ), so (Y ) = 0. If (Y A ) > t then (Y A ) =

(B ), so (Y ) = (B A ). Thus, every measurable subset of B A has measure
0, (B A ), or t (A ). Since is nonatomic, we conclude that (B A ) = 0. 

For an alternate proof of Theorem 4.2, see [5, Lemma 2]. For an extension of Theorem
4.2 to vector-valued measures, see [6].
This completes a proof that, for 0 < p < 1, Lp -spaces of a nonatomic measure space
have zero dual space. The machinery of functional analysis is built on the Hahn-Banach
Theorem, so how can we study these kinds of Lp -spaces, or more generally a topological
vector space that is not locally convex? See [1].

5. Theorem 4.2 Using Convexity

We now give a second proof of Theorem 4.2 (which will give a second proof that Lp () =
{0} when is nonatomic) based on an abstract convexity property for the measurable
subsets of any finite nonatomic measure space. The idea we use is briefly indicated in [7,
Exercise 41.2, p. 174].
Start with a finite measure space (M, ). (For instance, (M, ) is any measure space and
= F d for some nonnegative F L1 ().) A metric can be introduced on (M, ) by
d (A, B) = (AB),
where is the symmetric difference (see (3.2)). Here it is crucial that is a finite measure,
or this metric is not always defined. Actually, d is only a semimetric, so we look at M, the
metric space coming from M after elements at distance zero from one another are identified.
Step 1: The space M is complete.
Let {An } be d -Cauchy sequence. Then (in fact, equivalently) {An } is an L1 -Cauchy
sequence in L1 (). Let f be an L1 -limit of this sequence of characteristic functions. A
subsequence converges pointwise almost everywhere to f , so the set {x : f (x) = 1} is a
limit of {An } in M.
Step 2: Call a metric space (M, ) convex if for any x 6= y there is some z 6= x, y in M such
that (x, y) = (x, z) + (z, y). Note such a z need not be uniquely determined by the two
distances (x, z) and (z, y). For example, take M = S 2 with its surface metric, x the north
pole, y the south pole, and z any point along a chosen latitude of M . (When M Rn ,
this definition of convex might not match the usual notion: an open star-shaped region in
Rn is not convex in the usual sense but is convex in the abstract sense above. However, for
closed subsets of Rn using the induced metric from Rn , the above notion of convex does
match the usual meaning of the term.)
We show that (M, d ) is convex if (and only if!) is nonatomic.
If M is convex and (A) > 0, so d (A, ) > 0, then there exists a C such that d (A, C) > 0,
d (C, ) > 0, and d (A, ) = d (A, C) + d (C, ). That is
(AC) > 0, (C) > 0, (A) = (AC) + (C).
Therefore (AC) = (A) (C) > 0, so (C A) = 0, so (C) = (A C). This
implies 0 < (A C) < (A). Since A C A, this shows is nonatomic.
Conversely, if d (A, B) > 0, we want to find C such that d (A, C) and d (C, B) are both
positive and
d (A, B) = d (A, C) + d (B, C),
10 KEITH CONRAD

that is
(AB) = (AC) + (BC).
If (A) or (B) is less than or equal to (A B) (so A B or B A up to measure
zero), say (A) (A B). Then (A) = (A B) and (B) = (A B), so (AB) =
(B) (A) > 0.
Choose Y B A such that 0 < (Y ) < (B) (A) and let C = Y (A B). Then
(AC) = (A) (A B) + (Y ) = (Y ) > 0,

(BC) = (B) (A B) (Y ) = (B) (A) (Y ) > 0,

so
(AC) + (BC) = (B) (A) = (AB).
If (A), (B) > (A B), let C1 A A B and C2 B A B such that
0 < (C1 ) < (A) (A B), 0 < (C2 ) < (B) (A B).
Set C = C1 C2 (A B). Then
AC = (A C1 A B) (B C2 ), BC = (B C2 A B) (A C1 ),
so
(AC) = (A) (C1 ) (A B) + (C2 ) > (C2 ) > 0,
(BC) = (B) (C2 ) (A B) + (C1 ) > (C1 ) > 0,
so
(AC) + (BC) = (A B) (A B) = (AB).
Step 3: The final step will actually be a generalization of Theorem 4.2 to metric spaces.
The application to (M, d ) for nonatomic will come by setting x = and y = X in the
next claim.
Claim: For a complete convex metric space (M, ) and distinct points x and y in M with
0 t (x, y), there is a z such that (x, z) = t (and (x, y) = (x, z) + (z, y)).
The proof of this claim will be a tortuous Zornification.
First we define some notation. For a, b M , let
def
[a, b] = {c M : (a, b) = (a, c) + (c, b)}.
For instance, this set contains a and b. Intuitively, this is the set of points lying on geodesics
from a to b. It is helpful when reading the following discussion to draw many pictures of
line segments with points marked on them. Given t between 0 and (x, y), we will find a
z [x, y] with (x, z) = t.
Some simple properties of these intervals are:
(1) [a, b] = [b, a].
(2) If c [a, b] and b [a, c], then (b, c) = (b, c), so b = c.
Less simple properties are
(3) If b [a, c] then [a, b], [b, c] [a, c].
(4) If b [a, d] and c [b, d] then [a, c], [b, d] [a, d] and [b, c] = [a, c] [b, d] [a, d].
(We will only need that [b, c] lies in the intersection, not equality.)
 Lp -SPACES FOR 0 < p < 1 11

Proof of (3): Without loss of generality, we show [a, b] [a, c]. For x in [a, b],
(a, c) (a, x) + (x, c)
(a, x) + (x, b) + (b, c)
= (a, b) + (b, c)
= (a, c).
Therefore x [a, c].
Proof of (4): By (3), [b, d] [a, d] and [b, c] [b, d]. Therefore c [a, d], so [a, c] [a, d],
so
(a, d) = (a, c) + (c, d)
(a, b) + (b, c) + (c, d)
= (a, b) + (b, d)
= (a, d).
Therefore the inequality is an equality, so b [a, c], so [b, c] [a, c]. Thus [b, c] [a, c][b, d].
For the reverse inclusion, let x [a, c] [b, d]. Then
(a, d) = (a, b) + (b, d)
= (a, b) + (b, c) + (c, d)
(a, b) + (b, x) + (x, c) + (c, d)
= (a, b) + (b, d) (x, d) + (a, c) (a, x) + (c, d)
= 2(a, d) (x, d) (a, x).
Rearranging terms, (a, x) + (x, d) (a, d), so there is equality throughout, so (b, c) =
(b, x) + (x, c). Thus x [b, c].
Now we are ready to investigate geodesics on M . For our fixed x M introduced in
the statement of Step 3, define a partial ordering on M that might be called closer to x
on geodesics by
z1 z2 if and only if z1 [x, z2 ].
In particular, z1 z2 implies (x, z1 ) (x, z2 ).
Lets check this is a partial ordering.
If z1 z2 and z2 z1 , then z1 [x, z2 ] and z2 [x, z1 ], so z1 = z2 by (2).
If z1 z2 and z2 z3 then z1 [x, z2 ] and z2 [x, z3 ]. By (1), z2 [z3 , x] and z1 [z2 , x].
Therefore by (4),
z1 [z1 , z2 ] [x, z3 ],
hence z1 z3 .
Define
A = {z [x, y] : (x, z) t}.
This set is nonempty, since it contains x. We want to apply Zorns Lemma to A with its
induced partial ordering and show a maximal element of A has distance t from x.
Let {zi }iI be a totally ordered subset of A. We want an upper bound. Let
s = sup (x, zi ) t.
iI
for any > 0, there is some i0 such that
s (x, zi0 ) s,
12 KEITH CONRAD

so
s (x, zi ) s
for all i i0 . For i0 i j, s (x, zi ) s and
s (x, zj ) = (x, zi ) + (zi , zj ) s
so (zi , zj ) . Thus {zi } is a Cauchy net, so has a limit ` by completeness of M . We
show this limit is an upper bound in A.
Taking limits,
(x, y) = (x, zi ) + (zi , y) (x, y) = (x, `) + (`, y)
(x, zi ) t (x, `) t.
Thus ` A.
For i j,
(x, zj ) = (x, zi ) + (zi , zj ).
Taking limits over j,
(x, `) = (x, zi ) + (zi , `),
so zi [x, `], so zi ` for all i.
We have justified an application of Zorns Lemma to A. Let m be a maximal element.
That is, m A, and if z A with m [x, z] then z = m.
Let B = {z [y, m] : (y, z) (x, y) t}. Since y B, B is nonempty. Our goal is to
show m B, which is not obvious. Note that the definition of B depends on the existence
of a maximal element of A.
In B, introduce a partial ordering by z1 z2 when z1 [y, z2 ].
As above, B has a maximal element, m0 . Since m [x, y] and m0 [m, y], we get by (4)
that
[m, m0 ] [x, m0 ] [m, y] [x, y].
For z [m, m0 ],
z [x, y] (x, y) = (x, z) + (y, z)
(x, z) t or (y, z) (x, y) t
z A or z B.
Also,
(x, y) = (x, z) + (z, y)
(x, m) + (m, z) + (z, y)
= (x, m) + (m, y) since z [m, y]
= (x, y) since m [x, y].
Therefore (x, z) = (x, m) + (m, z), so m [x, z].
We now have
(x, y) = (x, z) + (z, y) since z [x, y]
(x, z) + (z, m0 ) + (m0 , y)
= (x, m0 ) + (m0 , y) since z [x, m0 ]
= (x, y) since m0 [x, y].
Therefore (y, z) = (z, m0 ) + (m0 , y), so m0 [y, z].
 Lp -SPACES FOR 0 < p < 1 13

Thus if z A then m [x, z] z = m. If z B, then m0 [y, z] z = m0 . Therefore

[m, m0 ] = {m, m0 }, so by convexity of M , m = m0 , hence m A B. Therefore (x, m) t
and (y, m) (x, y) t, so
(x, y) = (x, m) + (m, y) (x, y),
so (x, m) = t. This concludes our second proof of Theorem 4.2.

6. Paths Linking Measurable Sets

Returning to the original idea of showing Lp () has zero dual space (for nonatomic )
by finding a path in the space of measurable sets, the following possibility suggests itself.
Let (M, ) be a complete convex metric space. (Our second proof of Theorem 4.2 gives us
the motivating example.) Is M path connected? To try to show this, recall that weve seen
that for distinct x, y in M and 0 t (x, y), there is zt [x, y] such that (x, zt ) = t.
Perhaps if (x, y) is small enough, each zt is unique, and this would let us construct paths
in M . For a proof that any complete convex metric space is path connected, see [2, Thm.
14.1, p. 41].

Appendix A. Analogies between p 1 and 0 < p < 1

In this appendix, we summarize common features of Lp () for all p > 0.
Any Lp () is a vector space. Closure under scaling is trivial. For closure under addition,
first note that
|f (x) + g(x)| |f (x)| + |g(x)| = |f (x) + g(x)|p (|f (x)| + |g(x)|)p
since the p-th power is increasing on [0, ) for any p > 0. Now we argue separately for
p 1 and 0 < p < 1. When p 1, non-negative numbers a and b satisfy
(a + b)p 2p ap + 2p bp
since a + b 2 max(a, b). Therefore
(A.1) |f (x) + g(x)|p (|f (x)| + |g(x)|)p 2p |f (x)|p + 2p |g(x)|p ,
so X |f + g|p d 2p X |f |p d + 2p X |g|p d < . When 0 < p < 1 and a and b are
R R R
non-negative numbers,
(a + b)p ap + bp ,
so
|f (x) + g(x)|p (|f (x)| + |g(x)|)p |f (x)|p + |g(x)|p ,
so
Z Z Z
p p
(A.2) |f + g| d |f | d + |g|p d < .
X X X
For any p > 0, set
Z 1/p
def p
(A.3) |f |p = |f | d .
X
When f = g almost everywhere, |f |p = |g|p , so | |p is well-defined on Lp ().
When p 1, some work shows | |p satisfies the triangle inequality on Lp ():
|f + g|p |f |p + |g|p .
Therefore | |p is a norm on Lp (), and is called the Lp -norm.
14 KEITH CONRAD

However, when 0 < p < 1 the function | |p in (A.3) has a flipped triangle inequality:
|f + g|p |f |p + |g|p . (Try X = [0, 1] with Lebesgue measure, p = 1/2, f = [0,1/2) , and
g = [1/2,1] . Then |f + g|p = 1 while |f |p + |g|p = 211/p < 1.) Taking p-th roots in the
definition of | |p is the source of the problem. However, we do have
|f + g|p 21/p1 (|f |p + |g|p )
for 0 < p < 1, which for all intents and purposes is a fine substitute for the triangle
inequality. (To have the triangle inequality up to a universal scaling factor is harmless.)
Alternatively, we could use X |f |p d as a measure of distance from 0, without the p-th
R
root of the integral, since (A.2) shows the triangle inequality is valid for this. This is the
choice that was made in the main part of the text.
Just to fix ideas, we set the metric on Lp (), for 0 < p < 1, to be
Z
d(f, g) = |f g|p d.
X

However, whether or not we take the p-th root of X |f g|p d will not change the notion
R
of convergence on Lp ().
Some properties of the spaces Lp () for p 1 are
Any sequence of measurable functions that is Lp -convergent has a subsequence that
converges pointwise almost everywhere.
Lp () is complete, and thus is a Banach space for the Lp -norm.
(Holders inequality) If p > 1, and q > 1 is chosen so that 1/p + 1/q = 1, then for
f Lp () and g Lq (), the product f g lies in L1 () and |f g|1 |f |p |g|q .
If fn f in Lp (), then |fn |p |f |p in L1 ().
When X is a locally compact Hausdorff space, and is a regular Borel measure on
X, the space Cc (X) (continuous real-valued functions on X with compact support)
is dense in Lp ().
Since a 7 ap is an increasing function on [0, ) for p > 0, not just for p 1, some results
about Lp -spaces when p 1 work for 0 < p < 1. Here are some examples.
For 0 < p < 1, any sequence of measurable functions that is Lp -convergent has a
subsequence that converges pointwise almost everywhere.
For 0 < p < 1, Lp () is complete.
If 0 < p < 1 and fn f in Lp (), then |fn |p |f |p in L1 ().
When X is a locally compact Hausdorff space, and is a regular Borel measure on
X, the space Cc (X) (continuous real-valued functions on X with compact support)
is dense in Lp () for 0 < p < 1.
Actually, the third item needs a different proof in the case 0 < p < 1. (The usual proof
when p > 1 uses Holders inequality, which breaks down for 0 < p < 1.) Here is the proof.
For 0 < p < 1,
|f |p |f fn |p + |fn |p , |fn |p |f fn |p + |f |p ,
so
||f |p |fn |p | |f fn |p .
Thus Z Z
p p
||f | |fn | | d |f fn |p d 0.
X X
 Lp -SPACES FOR 0 < p < 1 15

Since, for 0 < p < 1, Lp () is a topological vector space complete with respect to
a metric, we have notions of continuity, completeness, and boundedness here. Therefore
several consequences of the Baire category theorem as applied to Banach spaces carry over
without change to Lp () for 0 < p < 1: the Open Mapping Theorem, the Closed Graph
Theorem, and the Principle of Uniform Boundedness.

References
[1] N. Adasch, B. Ernst, and D. Keim, Topological Vector Spaces. The Theory Without Convexity Conditions,
Lecture Notes in Mathematics 639, Springer-Verlag, New York, 1978.
[2] L. Blumenthal, Theory and Applications of Distance Geometry, Clarendon Press, Oxford, 1953.
[3] J. B. Conway, A Course in Functional Analysis, 2nd ed., Springer Verlag, New York, 1990.
[4] M. M. Day, The spaces Lp with 0 < p < 1, Bull. Amer. Math. Soc. 46 (1940), 816823.
[5] P. R. Halmos, On the Set of Values of a Finite Measure, Bull. Amer. Math. Soc. 53 (1947), 138-141.
[6] P. R. Halmos, The Range of a Vector Measure, Bull. Amer. Math. Soc. 54 (1948), 416-421.
[7] P. R. Halmos, Measure Theory, Van Nostrand, New York, 1950.
[8] W. Pfeffer, Integrals and Measures, Marcel Dekker, New York, 1977.
[9] M. Reed and B. Simon, Functional Analysis I, Academic Press, San Diego, 1980.
[10] A. Tychonoff, Ein Fixpunktsatz, Math. Ann. 111 (1935), 767776.