A crash course in interpolation theory

Ian Tice
Department of Mathematical Sciences
Carnegie Mellon University
October 28, 2024

Contents
0 Overview 2
0.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
0.2 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
0.3 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1 Classical interpolation theory 5

1.1 Analysis tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.1.1 The Minkowski and Hardy inequalities . . . . . . . . . . . . . . . . . . . . . 5
1.1.2 The distribution function . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.1.3 The layer cake theorem and its consequences . . . . . . . . . . . . . . . . . 11
1.1.4 Decreasing rearrangements . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.1.5 The Hardy-Littlewood rearrangement inequality . . . . . . . . . . . . . . . 18
1.1.6 Lorentz spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.2 Interpolation theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.2.1 The real method of Marcinkiewicz . . . . . . . . . . . . . . . . . . . . . . . 29
1.2.2 The complex method of Riesz-Thorin . . . . . . . . . . . . . . . . . . . . . 34
1.2.3 Comparing and contrasting the two interpolation theorems . . . . . . . . . . 38
1.3 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
1.3.1 The Hardy-Littlewood maximal function . . . . . . . . . . . . . . . . . . . . 38
1.3.2 The Hausdorff-Young inequality and its variants . . . . . . . . . . . . . . . . 39
1.3.3 Integral operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

2 Abstract interpolation theory 49

2.1 Compatible pairs of Banach spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
2.1.1 Reminders about the basics of Banach spaces . . . . . . . . . . . . . . . . . 50
2.1.2 Compatible Banach spaces and their sums and intersections . . . . . . . . . 51
2.1.3 Intermediate spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.1.4 The K and J functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
2.2 Interpolating between compatible Banach spaces . . . . . . . . . . . . . . . . . . . . 61
2.2.1 Heuristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.2.2 The interpolation spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
2.2.3 Some special cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

1
 2.3 Further properties of interpolation spaces . . . . . . . . . . . . . . . . . . . . . . . . 72
2.3.1 Equivalent norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
2.3.2 Reiteration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
2.4 Examples and applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
2.4.1 Interpolation of Lebesgue spaces . . . . . . . . . . . . . . . . . . . . . . . . . 81
2.4.2 Interpolating between Lp and W 1,p . . . . . . . . . . . . . . . . . . . . . . . 84
2.4.3 Interpolating between C 0 and C 1 . . . . . . . . . . . . . . . . . . . . . . . . 89

0 Overview
0.1 Introduction
To say that the Lebesgue spaces, which arise as natural generalizations of the spaces of integrable
and essentially bounded functions on a measure space, are useful in analysis is a ridiculous un-
derstatement, evident to anyone with even minor experience in the subject. A terse explanation
of the ubiquity of these spaces is that they possess a host of functional analytic properties that
provide powerful tools for working with them. Interpolation theory has its genesis in the study of
a particularly useful set these properties related to the mutual embeddings of the Lebesgue spaces
and the implications of these on the study of operators between Lebesgue spaces. Arguably, it is the
latter, which centers on the theorems of Marcinkiewicz and Riesz-Thorin, that is the most useful in
practice, as it allows one to study the mapping properties of operators on a wide range of Lebesgue
spaces (the interpolation spaces) by first understanding the properties in a restricted setting (the
endpoints of the interpolation). Abstract interpolation theory seeks to extend these ideas and tools
into a more general setting.
The purpose of these notes is to provide a brief introduction to this theory and to highlight some
of the powerful tools it provides for use in applications, in particular the idea of fractional regularity.
As a warning, the title should be taken seriously: the notes are by no means a complete study of
this topic, and huge portions of the theory and many important results are completely ignored.
The reasons for this are two-fold. First, these notes originated in the course Function Spaces and
Generalized Regularity, taught jointly by the author and Giovanni Leoni at the Scuola Matematica
Interuniversitaria Summer School in Cortona, Italy from July 5th to 16th of 2021 (though, it should
be noted that, due to the pandemic, the course was actually taught through Zoom from the author’s
basement). This school ran only for two weeks with fifteen hours of lectures on interpolation theory
and fifteen more on fractional Sobolev spaces (taught by Leoni), and the material recorded here
is already a strict superset of what it was possible to cover in the interpolation lectures. Second,
there are many very good and very thorough books on the subject, and a deeper study should begin
by looking at these. In particular, the texts of Bennett and Sharpley [1], Bergh and Löfström [2],
Brudnyĭ and Krugljak [3], Lunardi [5], and Triebel [7] are all recommended.
The notes are organized as follows. In the remainder of this overview we review notation used
throughout the notes and provide a quick motivation for our study by examining some interpola-
tion properties of Lebesgue spaces In Section 1 we develop the classical theory of interpolation in
Lebesgue spaces mentioned above. To properly frame this, we first need a number of tools from
advanced real analysis: the distribution function, rearrangements, and Lorentz spaces. These are
developed in Section 1.1. In Section 1.2 we prove the main theorems on interpolation in Lebesgue
spaces: the classical theorems of Marcinkiewicz and Riesz-Thorin. In Section 1.3 we provide some
examples of applications of these theorems by studying the Hardy-Littlewood maximal function,

2
various estimates of the Fourier transform (the Hausdorff-Young theorem and its variants), and
certain integral operators, including those of convolution and Riesz potential type.
In Section 2 we introduce the abstract interpolation theory that arises as a generalization of the
Marcinkiewicz theorem. This is known as the real method of interpolation, as it relies entirely on
real variable techniques, in contrast with a second known method that relies crucially on complex
variables and holomorphic functions, which is entirely ignored in these notes. In the abstract
framework we first need to build a scale of spaces analogous to the scale of spaces obtained by varying
p in the Lebesgue context. The preliminary functional analysis for this is carried out in Section
2.1. Then in Section 2.2 we construct interpolation spaces from given pairs of compatible Banach
spaces. In Section 2.3 we prove a couple important theorems about abstract interpolation. In
Section 2.4 we then study three concrete examples of the spaces obtained through real interpolation
and demonstrate some applications.

0.2 Notation
We will employ the following notational conventions throughout these notes.

1. We use F to denote either of the fields R or C. All vector spaces are over F, where F is allowed
to be either. The natural numbers, N, include 0.

2. We will often write R+ = (0, ∞) = {x ∈ R | 0 < x}.

3. Given a metric space X, x ∈ X, and r > 0, we define the open balls B(x, r) = {y ∈
X | d(x, y) < r} and the closed balls B[x, r] = {y ∈ X | d(x, y) ≤ r}.

4. Given normed vector spaces X and Y over a common field, we write L(X; Y ) = {T : X →
Y | T is bounded and linear} and endow it with the usual operator norm

kT kL(X;Y ) = sup{kT xkY | kxkX ≤ 1}. (0.2.1)

5. Given a measure space (X, M, µ) and 1 ≤ p ≤ ∞ we will write Lp (X; F) for the space
of F−valued p−integrable functions. When F = R we will typically abbreviate Lp (X) =
Lp (X; F). If we want to emphasize the measure we will write Lpµ (X; F).

6. We will write `p (Z; F) = Lpµ (Z; F) for µ the counting measure.

7. Given a measure space (X, M, µ) we will write S(X; F) for the space of simple functions on
X and Sf in (X; F) for the space of simple functions with support of finite measure.

8. For 1 ≤ n ∈ N we write ωn = Ln (B(0, 1)) for the n−dimensional Lebesgue measure (written
Ln ) of the unit ball in Rn , and we write αn = Hn−1 (∂B(0, 1)) for the (n − 1)−dimensional
Hausdorff measure (written Hn−1 ) of the unit sphere in Rn . These are related via
Z Z 1
αn
ωn = dx = αn rn−1 dr = . (0.2.2)
B(0,1) 0 n

9. Given topological vector spaces X and Y , we write X ,→ Y if X ⊆ Y is a vector subspace

and the inclusion map I : X → Y is continuous.

3
0.3 Motivation
Let (X, M, µ) be a measure space and 1 ≤ p ≤ ∞. We recall two basic facts about Lebesgue spaces
that serve as the initial motivation for developing the theory of interpolation. For the first consider
f ∈ Lp (X; F)\{0} for 1 < p < ∞ and let t ∈ R+ . Then we can write

f = f χ{|f |>t} + f χ{|f |≤t} =: f1 + f2 . (0.3.1)

Then f2 ∈ L∞ (X; F) and kf2 kL∞ ≤ t, while

|f |p−1
Z Z Z
1
kf1 kL1 = |f1 | dµ = |f | dµ ≤ |f | p−1 dµ = p−1 kf kpLp , (0.3.2)
X {|f |>t} {|f |>t} t t

which in particular means that f1 ∈ L1 (X; F). This shows that we can always decompose a generic
f ∈ Lp (X; F) into a sum of elements of L1 (X; F) and L∞ (X; F), which is noteworthy since 1 and
∞ are the extreme points of the set [1, ∞]. In fact, we can take this a bit further by realizing that
the parameter t ∈ R+ can be tuned. Indeed, we can optimize the right side of the bound
1
kf1 kL1 + kf2 kL∞ ≤ kf kpLp + t (0.3.3)
tp−1
over t ∈ R+ to see that the minimal value occurs when

1 − (p − 1)t−p kf kpLp = 0 ⇔ t = (p − 1)1/p kf kLp . (0.3.4)

Using this choice of t in the estimate above shows that

kf kL1 +L∞ := inf{kakL1 + kbkL∞ | f = a + b, a ∈ L1 , b ∈ L∞ }

 
−1/p0 1/p
≤ (p − 1) + (p − 1) kf kLp , (0.3.5)

which actually shows the stronger result that Lp (X; F) ,→ L1 (X; F) + L∞ (X; F), where the latter
space is endowed with the infimum norm written above.
The second fact we wish to recall from Lebesgue theory has to do with multiple inclusions.
Suppose that 1 ≤ p0 < p1 ≤ ∞ and let p0 < p < p1 be given by
1 1−θ θ
= + for θ ∈ (0, 1). (0.3.6)
p p0 p1

In other words, 1/p is the convex interpolation between 1/p0 and 1/p1 . Suppose f ∈ Lp0 (X; F) ∩
Lp1 (X; F). If p1 < ∞, then Hölder’s inequality allows us to estimate
Z Z Z p(1−θ)/p0 Z pθ/p1
p (1−θ)p θp p0 p1
|f | dµ = |f | |f | dµ ≤ |f | dµ |f | dµ , (0.3.7)
X X X X

which implies that f ∈ Lp (X; F) and

kf kLp ≤ kf kL1−θ θ
p0 kf kLp1 . (0.3.8)

The same is true if p1 = ∞, and we leave it as an exercise to verify this. In fact, we can take
this a bit further by recalling that we can endow the space Lp0 (X; F) ∩ Lp1 (X; F) with the norm

4
kf kLp0 ∩Lp1 = max{kf kLp0 , kf kLp1 }. Then the previous estimate shows that kf kLp ≤ kf kLp0 ∩Lp1 ,
and so Lp0 (X; F) ∩ Lp1 (X; F) ,→ Lp (X; F).
Note that if we combine these results, we deduce that for all 1 < p < ∞ we have the embeddings

L1 (X; F) ∩ L∞ (X; F) ,→ Lp (X; F) ,→ L1 (X; F) + L∞ (X; F), (0.3.9)

which we can think of as telling us that the Lp (X; F) spaces somehow interpolate between the
spaces on the left and right. This and the above results show that the Lebesgue spaces have very
interesting interpolation properties. The goal of these notes is to further develop these ideas, first
in the context of Lebesgue spaces and their natural generalization, Lorentz spaces, and second in
the completely abstract setting of “compatible” Banach spaces. This leads to a set of powerful tools
that have many applications in analysis and PDE, for example.

1 Classical interpolation theory

In this section we develop the interpolation theory of Lebesgue and Lorentz spaces. First we develop
the necessary analytic tools. Then we prove the theorems of Marcinkiewicz and Riesz-Thorin. We
conclude the section with some applications of these theorems.

1.1 Analysis tools

Here we collect a number of analytic tools that are useful in interpolation theory.

1.1.1 The Minkowski and Hardy inequalities

We begin by proving a version of Minkowski’s inequality for integrals.
Theorem 1.1.1 (Minkowski’s inequality, integral form). Suppose that (X, MX , µ) and (Y, MY , ν)
are σ−finite measure spaces and let (X × Y, Mx ⊗ MY , µ ⊗ ν) be the associated product space. Let
1 ≤ p < ∞ and f : X × Y → [0, ∞] be a µ ⊗ ν−measurable function. Then
Z Z p 1/p Z Z 1/p
p
f (x, y)dν(y) dµ(x) ≤ (f (x, y)) dµ(x) dν(y). (1.1.1)
X Y Y X

Proof. We begin by making a reduction. We claim that it suffices to prove the result under the
extra assumptions that X and Y are finite and that f is bounded. Indeed, if the result is proved
in this case then the monotone convergence theorem allows us to extend it to the case of X and Y
σ−finite with f bounded by using decompositions of X and Y into countably many finite-measure
subsets. In turn, the monotone convergence theorem again allows us to extend to general f by
passing to the limit with finite truncations of f . This proves the claim, so we henceforth assume
that X and Y are finite and f is bounded.
When p = 1 Tonelli’s theorem tells us that (1.1.1) actually holds as an equality. Suppose then
that 1 < p < ∞ and set p0 = p/(p − 1) ∈ (1, ∞). We may assume without loss of generality that
the left side of (1.1.1) is non-zero, as otherwise the result is trivially true. Note also that the right
side of (1.1.1) is finite due to assumptions of X, Y and f .
Define the measurable function F : X → [0, ∞) via
Z
F (x) = f (x, y)dν(y). (1.1.2)
Y

5
If F = 0 for µ−a.e. x ∈ X, then there’s nothing to prove since the left side of (1.1.1) vanishes in
this case. We may assume, then that this is not the case. Then by Tonelli’s theorem and Hölder’s
inequality we can bound
Z Z p Z Z 
p
kF kLp (X) = f (x, y)dν(y) dµ(x) = f (x, y)dν(y) (F (x))p−1 dµ(x)
X Y X Y
Z Z Z Z 1/p Z 1/p0
p−1 p p
= f (x, y)(F (x)) dµ(x)dν(y) ≤ (f (x, y)) dµ(x) (F (x)) dµ(x) dν(y)
Y X Y X X
Z Z 1/p
= kF kp−1
Lp (X)
p
(f (x, y)) dµ(x) dν(y). (1.1.3)
Y X

By assumption, 0 < kF kLp (X) < ∞, so we can divide both sides by by kF kLp−1
p (X) to deduce (1.1.1).

Remark 1.1.2. Minkowski’s integral inequality can also be proved using duality arguments.
Next we prove another very useful pair of inequalities, due to G.H. Hardy.
Theorem 1.1.3 (Hardy’s inequalities). If f : (0, ∞) → [0, ∞) is Lebesgue measurable, s > 0, and
1 ≤ p < ∞, then
Z ∞ Z x p 1/p Z ∞ 1/p
1 p p−s−1 p
f (t)dt dx ≤ x f (x) dx . (1.1.4)
0 xs+1 0 s 0

and p 1/p 1/p

Z ∞ Z ∞ Z ∞
s−1 p p+s−1 p
x f (t)dt dx ≤ x f (x) dx . (1.1.5)
0 x s 0

Proof. Using the change of variable t = xr shows that

Z x Z 1
1 f (rx)
(s+1)/p
f (t)dt = (s+1)/p−1
dr. (1.1.6)
x 0 0 x

Then p 1/p Z p 1/p

Z ∞ Z x ∞ Z 1
1 f (rx)
f (t)dt dx = dr dx . (1.1.7)
0 xs+1 0 0 0 x(s+1)/p−1
Minkowski’s integral inequality then shows that
∞ 1 p 1/p 1 ∞ 1/p
f (rx)p
Z Z Z Z
f (rx)
dr dx ≤ dx dr, (1.1.8)
0 0 x(s+1)/p−1 0 0 xs+1−p
but we may again change variables to write
Z ∞ Z ∞
f (rx)p s−p f (x)p
dx = r dx. (1.1.9)
0 xs+1−p 0 xs+1−p
Thus
Z 1 Z ∞ 1/p ∞ 1/p Z 1 ∞ 1/p
f (rx)p f (x)p
Z Z
s/p−1 p p−s−1 p
dx dr = dx r dr = x f (x) dx .
0 0 xs+1−p 0 xs+1−p 0 s 0
(1.1.10)

6
Chaining these together gives the first inequality.
The second identity follows similarly:
Z ∞ Z ∞ p 1/p Z ∞ Z ∞ p 1/p
s−1 (s−1)/p+1
x f (t)dt dx = x f (rx)dr dx
0 x 0 1
Z ∞ Z ∞ p Z ∞ Z ∞ p
s−1+p p 1 s−1+p p
≤ x f (rx) dx dr = y f (y) dy dr
1 0 1 r1+s/p 0
Z ∞ p
p s−1+p p
= y f (y) dy . (1.1.11)
s 0

Remark 1.1.4. One particularly interesting case of Hardy’s inequalities is occurs when p > 1 and
s = p − 1, in which case
Z ∞  Z x p 1/p Z ∞ 1/p
1 p p
f (t)dt dx ≤ f (x) dx . (1.1.12)
0 x 0 p−1 0

This tells us that if f ∈ Lp ((0, ∞)) then the average function A : (0, ∞) → [0, ∞) defined by

1 x
Z
A(x) = f (t)dt (1.1.13)
x 0

is such that A ∈ Lp ((0, ∞)).

1.1.2 The distribution function

We now establish an important relationship between the integral of a function (or rather the integral
of its pth power), and the size of the sets where |f | is large.

Theorem 1.1.5 (Chebyshev’s inequality). Let (X, M, µ) be a measure space, 1 ≤ p < ∞, and
f : X → F be measurable. Then for each t ∈ (0, ∞) we have that
Z
1
µ({x ∈ X | |f (x)| > t}) ≤ p |f |p dµ. (1.1.14)
t X

Proof. Let t ∈ (0, ∞). We compute

|f |p
Z Z Z
1
µ({x ∈ X | |f (x)| > t}) = dµ ≤ p
dµ ≤ p |f |p dµ. (1.1.15)
{|f |>t} {|f |>t} t t X

This suggests that we introduce some notation.

Definition 1.1.6. Let (X, M, µ) be a measure space and f : X → F be measurable. We define the
distribution function of f to be df : [0, ∞) → [0, ∞] given by

df (t) = µ({x ∈ X | |f (x)| > t}). (1.1.16)

Let’s consider some examples.

7
Example 1.1.7. Let (X, M, µ) be a measure space. Suppose that E ∈ M and f = χE . Then for
t ≥ 1 we have that {x ∈ X | χE (x) > t} = ∅, while for t ∈ [0, 1) we have that {x ∈ X | χE (x) >
t} = E. Thus (
µ(E) if t ∈ [0, 1)
dχE (t) = (1.1.17)
0 if t ∈ [1, ∞).
When µ(E) < ∞ we can write this as dχE = µ(E)χ[0,1) . 4
Example 1.1.8. Let (X, M, µ) be a measure space. Suppose that f : X → [0, ∞) is a finite simple
function given by
Xn
f= ai χEi (1.1.18)
i=1
where E1 , . . . , En ∈ M are finite measure sets that are pairwise disjoint, and 0 < an < · · · < a1 < ∞.
Set an+1 = 0. For j = 1, . . . , n set
j
X
bj = µ(Ei ). (1.1.19)
i=1

Let t ∈ [0, ∞). If t ≥ a1 then clearly df (t) = µ(∅) = 0. If t ∈ [a2 , a1 ), then |f (x)| > t if and only if
x ∈ E1 , so df (t) = µ(E1 ). More generally, if t ∈ [aj+1 , aj ) then
j
[
|f (x)| > t ⇔ x ∈ Ei (1.1.20)
i=1

and hence df (t) = bj . Assembling this information shows that

n
X
df (t) = bj χ[aj+1 ,aj ) (t). (1.1.21)
j=1

Thus the distribution function of a finite simple function is again a finite simple function. 4
Example 1.1.9. Let (X, M, µ) be a measure space. Suppose that f : X → [0, ∞) is a simple
function that is not finite. We may then write
n
X
f= ai χEi (1.1.22)
i=1

where E1 , . . . , En ∈ M are pairwise disjoint, and 0 < an < · · · < a1 < ∞. Since f is not a finite
simple function, there exists 1 ≤ i ≤ n such that µ(Ei ) = ∞. We then define

m = min{1 ≤ i ≤ n | µ(Ei ) = ∞}. (1.1.23)

If m = 1 then it’s clear that

(
∞ for 0 ≤ t < a1
df (t) = (1.1.24)
0 for a1 ≤ t.

Assume then that 1 < m ≤ n. For j = 1, . . . , m − 1 set

j
X
bj = µ(Ei ). (1.1.25)
i=1

8
If t ∈ [aj+1 , aj ) for 1 ≤ j ≤ m − 1 then
j
[
|f (x)| > t ⇔ x ∈ Ei (1.1.26)
i=1

and hence df (t) = bj . On the other hand, if 0 ≤ t < am then df (t) = ∞. Assembling this
information shows that
(
∞ for 0 ≤ t < am
df (t) = Pm−1 (1.1.27)
j=1 bj χ[aj+1 ,aj ) (t) for am ≤ t.

From this analysis we can write

n
X
df (t) = bj χ[aj+1 ,aj ) (t) (1.1.28)
j=1

with the understanding that bj = ∞ for j ≥ m. Thus the distribution function of a simple function
is again a simple function. 4

Example 1.1.10. Let (X, M, µ) be a measure space and f : X → F be measurable. Then

df (t) = µ({x ∈ X | |f (x)| > t}) = d|f | (t). (1.1.29)

This shows that f and |f | : X → [0, ∞) have the same distribution functions. 4

The next result establishes some basic properties of the distribution function.

Proposition 1.1.11. Let (X, M, µ) be a measure space and f, g : X → F be measurable. Then the
following hold.

1. If |g| ≤ |f | a.e. then dg ≤ df on [0, ∞).

2. If c ∈ F\{0}, then dcf (t) = df (t/ |c|) for all t ∈ [0, ∞).

3. If t, s ∈ [0, ∞), then df +g (t + s) ≤ df (t) + dg (s).

4. If t, s ∈ [0, ∞), then df g (ts) ≤ df (t) + dg (s).

5. df is nonincreasing and right continuous.

Proof. The first two items are trivial. For the third item note that if |f (x) + g(x)| > t + s then
either |f (x)| > t or |g(x)| > s since otherwise

|f (x)| + |g(x)| ≤ t + s < |f (x) + g(x)| ≤ |f (x)| + |g(x)| , (1.1.30)

a contradiction. Thus

{x ∈ X | |f (x) + g(x)| > t + s} ⊆ {x ∈ X | |f (x)| > t} ∪ {x ∈ X | |g(x)| > s}, (1.1.31)

which immediately implies that df +g (t + s) ≤ df (t) + dg (s). This proves the third item. The fourth
item follows from a similar argument, which we leave as an exercise.

9
 We now turn to the proof of the fifth item. It’s obvious that df is nonincreasing. Fix t ∈ [0, ∞)
and suppose that {tn }∞ ∞
n=` ⊂ (t, ∞) is such that tn → t. Extract a decreasing subsequence {tnm }m=` .
Note that ∞
[
{x ∈ X | |f (x)| > tnm } = {x ∈ X | |f (x)| > t} (1.1.32)
m=1

and that
{x ∈ X | |f (x)| > tnm } ⊆ {x ∈ X | |f (x)| > tnm+1 }. (1.1.33)
Hence
∞
!
[
df (t) = µ {x ∈ X | |f (x)| > tnm } = lim µ({x ∈ X | |f (x)| > tnm }) = lim df (tnm ).
m→∞ m→∞
m=1
(1.1.34)
On the other hand, since df is nondecreasing we have that

lim df (tnm ) = lim df (tn ), (1.1.35)

m→∞ n→∞

and so
df (t) = lim df (tn ). (1.1.36)
n→∞

This proves that df is right continuous, completing the proof of the fifth item.

Now we examine how the distribution function behaves with respect to sequences of functions.
The next result should be thought of as analogous to the monotone convergence theorem and Fatou’s
lemma.

Proposition 1.1.12. Let (X, M, µ) be a measure space. Suppose that for each n ≥ ` ∈ Z the
function fn : X → F is measurable. Further suppose that f : X → F is measurable. Then the
following hold.

1. If {|fn |}∞ ∞
n=` is a.e. nondecreasing and |fn | → |f | a.e. as n → ∞, then {dfn }n=` is nondecreas-
ing and dfn → df pointwise as n → ∞.

2. If
|f | ≤ lim inf |fn | a.e. in X, (1.1.37)
n→∞

then
df ≤ lim inf dfn in [0, ∞). (1.1.38)
n→∞

Proof. We first prove the first item. Proposition 1.1.11 shows that {dfn }∞
n=` is nondecreasing, so
it suffices to prove that dfn → df pointwise as n → ∞. Fix t ∈ [0, ∞). Modifying the functions
fn on null sets if necessary, we may assume without loss of generality that {|fn |}∞
n=` is pointwise
nondecreasing. Then

{x ∈ X | |fn (x)| > t} ⊆ {x ∈ X | |fn+1 (x)| > t} (1.1.39)

for each n ≥ `, and

∞
[
{x ∈ X | |fn (x)| > t} = {x ∈ X | |f (x)| > t}. (1.1.40)
n=1

10
Thus
∞
!
[
df (t) = µ {x ∈ X | |fn (x)| > t} = lim µ({x ∈ X | |fn (x)| > t}) = lim dfn (t), (1.1.41)
n→∞ n→∞
n=1

which completes the proof of the first item.

Next we prove the second item. For n ≥ ` define the measurable function gn : X → [0, ∞)
via gn (x) = inf m≥n |fm (x)|. By assumption we then have that {|gn |}∞ n=` is a.e. nondecreasing and
gn = |gn | → g := lim inf |fn | a.e. as n → ∞. Consequently, the first item tells us that
n→∞

dg = lim dgn . (1.1.42)

n→∞

On the other hand, by assumption we have that |f | ≤ |g| = g a.e. in X, and by construction we
have that |gn | ≤ |fn | a.e. in X. Consequently, Proposition 1.1.11 implies that

df ≤ dg = lim dgn ≤ lim inf dfn . (1.1.43)

n→∞ n→∞

This proves the second item.

Next we record a useful result that allows us to compute the distribution function for some map
given that we know the distribution function of the map restricted to essentially disjoint sets.

Proposition 1.1.13. Let (X, M, S µ) be a measure space. Suppose that I 6= ∅ is countable and
{Xi }i∈I ⊆ M is such that X = i∈I Xi and µ(Xi ∩ Xj ) = 0 for i, j ∈ I with i 6= j. Let f : X → F
be measurable, and for each i ∈ I define the measurable function fi = f χXi . Then
X
df (t) = dfi (t) for all t > 0. (1.1.44)
i∈I

Proof. Exercise.

1.1.3 The layer cake theorem and its consequences

The following result establishes a deep connection between the distribution function of a map and
integrals related to the map.

Theorem 1.1.14 (Layer cake theorem). Let (X, M, µ) be a measure space and let ν be a Borel
measure on [0, ∞) such that the map ϕ : [0, ∞) → [0, ∞] given by ϕ(t) = ν([0, t)) is finite for every
t ∈ [0, ∞). Let f : X → [0, ∞) be measurable. Then ϕ ◦ f is measurable and
Z Z ∞
ϕ ◦ f dµ = df dν. (1.1.45)
X 0

Proof. First note that ϕ is nondecreasing and ϕ(0) = ν(∅) = 0, so ϕ ◦ f is µ−measurable.

Since f is measurable we can pick a sequence {fn }∞n=0 ⊂ S(X; [0, ∞)) such that fn ≤ fn+1 on
X and fn → f pointwise as n → ∞. For each n ∈ N write
Nn
X
fn = ai,n χEi,n (1.1.46)
i=1

11
such that 0 < aNn ,n < · · · < a1,n . Define aNn +1,n = 0 as well as
j
X
bj,n = µ(Ei,n ) ∈ [0, ∞] for 1 ≤ j ≤ Nn . (1.1.47)
i=1

According to Examples 1.1.8 and 1.1.9, we have that dfn (t) = 0 for t ≥ a1,n and
dfn (t) = bj,n for t ∈ [aj+1,n , aj,n ). (1.1.48)
Hence, Z ∞ X
dfn dν = bj,n ν([aj+1,n , aj,n )), (1.1.49)
0 j∈Jn

where
Jn = {1 ≤ j ≤ Nn | ν([aj+1,n , aj,n )) > 0}. (1.1.50)
Note that
ϕ(aj,n ) − ϕ(aj+1,n ) = ν([aj+1,n , aj,n )) > 0 for j ∈ Jn . (1.1.51)
On the other hand, Z X
ϕ ◦ fn dµ = ϕ(ai,n )µ(Ei,n ), (1.1.52)
X i∈In

where
In = {1 ≤ i ≤ Nn | ϕ(ai,n ) > 0}. (1.1.53)
According to (1.1.51) we have that Jn ⊆ In . Note also that since ϕ is nondecreasing the set In has
the property that if i ∈ In then {1, . . . , i} ⊆ In , and so
j ∈ Jn ⇒ {1, . . . , j} ⊆ In . (1.1.54)
Now, from (1.1.52)and (1.1.54) we can compute (employing the convention that ∞ · 0 = 0 here)
Z X X Nn
X
ϕ ◦ fn dµ = ϕ(ai,n )µ(Ei,n ) = µ(Ei,n ) ν([aj+1,n , aj,n ))
X i∈In i∈In j=i
XX XX
= µ(Ei,n )ν([aj+1,n , aj,n ))χ{i,...,Nn } (j) = µ(Ei,n )ν([aj+1,n , aj,n ))χ{i,...,Nn } (j)
i∈In j∈Jn j∈Jn i∈In
j
X X X X
= ν([aj+1,n , aj,n )) µ(Ei,n ) = ν([aj+1,n , aj,n )) µ(Ei,n )
j∈Jn i∈In ∩{1,...,j} j∈Jn i=1
X
= ν([aj+1,n , aj,n ))bj , (1.1.55)
j∈Jn

and upon combining this with (1.1.49) we deduce that

Z Z ∞
ϕ ◦ fn dµ = dfn dν. (1.1.56)
X 0

Since ϕ is nondecreasing we know that {ϕ ◦ fn }∞

n=0 is also nondecreasing and converges pointwise to
ϕ ◦ f as n → ∞. Proposition 1.1.12 implies that {dfn }∞n=0 is nondecreasing and dfn → df pointwise
as n → ∞. Combining these facts with the monotone convergence theorem, we may send n → ∞
in (1.1.56) to conclude that (1.1.45) holds.

12
 The most important consequence of the Layer Cake theorem is the following corollary.

Corollary 1.1.15. Let (X, M, µ) be a measure space and f : X → F be measurable. Then for
1 ≤ p < ∞ we have that Z Z ∞
p
|f | dµ = ptp−1 df (t)dt. (1.1.57)
X 0

Proof. Let νp be the Borel measure on [0, ∞) given by

Z
νp (E) = ptp−1 dt, (1.1.58)
E

where here dt is Lebesgue measure on [0, ∞). Then νp ([0, t)) = tp and we deduce from the layer
cake representation and Example 1.1.10 that
Z Z ∞ Z ∞
p p−1
|f | dµ = pt d|f | (t)dt = ptp−1 df (t)dt. (1.1.59)
X 0 0

1.1.4 Decreasing rearrangements

We know that the distribution function is nonincreasing and right continuous. This does allow
for the possibility of jump discontinuities. If none existed, then we would be free to invert the
distribution function. Since we can’t do this in general, we do the next best thing. The resulting
function, which we define now, will play a key role in our subsequent analysis.

Definition 1.1.16. Let (X, M, µ) be a measure space and f : X → F be measurable. We define the
decreasing rearrangement of f to be the function f # : [0, ∞) → [0, ∞] given by

f # (t) = inf{s ∈ [0, ∞) | df (s) ≤ t}. (1.1.60)

We will prove momentarily that f # is nonincreasing, so it is a bit of an abuse of notation to call

f # the decreasing rearrangement. Nevertheless, this is the standard notation, so we adopt it here.

Let’s consider some examples.

Example 1.1.17. Let (X, M, µ) be a measure space. Suppose that E ∈ M and f = χE . We know
from Example 1.1.7 that (
µ(E) if t ∈ [0, 1)
dχE (t) = (1.1.61)
0 if t ∈ [1, ∞).
In particular this implies that dχE (s) ≤ µ(E) for all s ∈ [0, ∞), and thus (χE )# (t) = 0 for all
t ≥ µ(E). Also, if t < µ(E) then dχE (s) ≤ t if and only if s ∈ [1, ∞), which implies that
(χE )# (t) = 1. Hence (
1 if t ∈ [0, µ(E))
(χE )# (t) = (1.1.62)
0 if t ∈ [µ(E), ∞).
4

13
Example 1.1.18. Let (X, M, µ) be a measure space. Suppose that f : X → [0, ∞) is a finite
simple function given by
Xn
f= ai χEi (1.1.63)
i=1

where E1 , . . . , En ∈ M are finite measure sets that are pairwise disjoint, and 0 < an < · · · < a1 < ∞.
Set an+1 = 0. We saw in Example 1.1.8 that
n
X
df (t) = bj χ[aj+1 ,aj ) (t), (1.1.64)
j=1

where for 1 ≤ j ≤ n we set

j
X
bj = µ(Ei ) ∈ (0, ∞). (1.1.65)
i=1

Set b0 = 0. Then an argument in the same spirit as Example 1.1.8, which we leave as an exercise,
shows that n
X
#
f (t) = aj χ[bj−1 ,bj ) (t). (1.1.66)
j=1

This proves that the rearrangement of a finite simple function is a finite simple function.
4

Example 1.1.19. Let (X, M, µ) be a measure space. Suppose that f : X → [0, ∞) is a simple
function that is not finite given by
Xn
f= ai χEi (1.1.67)
i=1

where E1 , . . . , En ∈ M are finite measure sets that are pairwise disjoint, and 0 = an+1 < an < · · · <
a1 < ∞. We saw in Example 1.1.9 that
(
∞ for 0 ≤ t < am
df (t) = P (1.1.68)
1≤j≤m−1 bj χ[aj+1 ,aj ) (t) for am ≤ t,

where
m = min{1 ≤ i ≤ n | µ(Ei ) = ∞}, (1.1.69)
j
X
bj = µ(Ei ) ∈ (0, ∞] if 1 ≤ j ≤ m, (1.1.70)
i=1

and we have employed the convention that the sum over an empty set of indices is 0. From this we
find that if we set b0 = 0 then we can write
m
X
#
f (t) = aj χ[bj−1 ,bj ) (t). (1.1.71)
j=1

This proves that the rearrangement of a non-finite simple function is a non-finite simple function.
4

14
Example 1.1.20. Let (X, M, µ) be a measure space, f : X → F be measurable and define |f | :
X → [0, ∞) via |f | (x) = |f (x)|. We know from Example 1.1.10 that

df (t) = µ({x ∈ X | |f (x)| > t}) = d|f | (t). (1.1.72)

Hence f # = |f |# . 4

Our next result establishes some of the basic properties of the decreasing rearrangement.

Proposition 1.1.21. Let (X, M, µ) be a measure space and suppose that f, g : X → F are measur-
able. Then the following hold.

1. f # is nonincreasing.

2. If |g| ≤ |f | a.e. then g # ≤ f # on [0, ∞).

3. If c ∈ F then (cf )# = |c| f # .

4. If t ∈ [0, ∞) and df (t) < ∞, then f # (df (t)) ≤ t.

5. If t ∈ [0, ∞) and f # (t) < ∞, then df (f # (t)) ≤ t.

6. If t, s ∈ [0, ∞), then (f + g)# (t + s) ≤ f # (t) + g # (s).

7. If t, s ∈ [0, ∞), then (f g)# (t + s) ≤ f # (t)g # (s).

Proof. The first item follows directly from the fact that df is nonincreasing.
To prove the second item we note that Proposition 1.1.11 tells us that dg ≤ df , and the inequality
g ≤ f # then follows easily from the definition. To prove the third item we first note that it suffices
#

to prove the result when c 6= 0, as the equality is trivial when c = 0. For c ∈ C\{0} and t ∈ [0, ∞)
the second item of Proposition 1.1.11 implies that

{s ≥ 0 | dcf (s) ≤ t} = {s ≥ 0 | df (s/ |c|) ≤ t} = |c| {s ≥ 0 | df (s) ≤ t}, (1.1.73)

which then implies that (cf )# (t) = |c| f # (t), completing the proof of the third item.
The fourth item follows trivially since if df (t) < ∞ then

f # (df (t)) = inf{s ≥ 0 | df (s) ≤ df (t)} ≤ t. (1.1.74)

For the fifth item we note that if f # (t) < ∞ then by definition for each n ∈ N we can find
rn ∈ [f # (t), f # (t) + 2−n ) such that df (rn ) ≤ t. Clearly rn → f # (t) as n → ∞, so the right
continuity of df implies that
df (f # (t)) = lim df (rn ) ≤ t. (1.1.75)
n→∞

This proves the fifth item.

We now turn to the proof of the sixth item. Fix t, s ∈ [0, ∞). Without loss of generality we may
assume that f # (t)+g # (s) < ∞, in which case we can set a = f # (t) < ∞ and b = g # (s) < ∞. Then
the fifth item tells us that df (a) ≤ t and dg (b) ≤ s. Consequently, the third item of Proposition
1.1.11 implies that
df +g (a + b) ≤ df (a) + dg (b) ≤ t + s (1.1.76)
and so by definition (f + g)# (t + s) ≤ a + b = f # (t) + f # (s). The sixth item is proved.

15
 To prove the seventh item we assume without loss of generality that f # (t) · g # (s) < ∞ and
again set a = f # (t) < ∞ and b = g # (s) < ∞. Then the third item of Proposition 1.1.11 tells us
that
df g (ab) ≤ df (a) + dg (b) ≤ t + s (1.1.77)
and so (f g)# (t + s) ≤ ab = f # (t)g # (s). The proves the seventh item.

Now we record the most important features of decreasing rearrangements.

Theorem 1.1.22. Let (X, M, µ) be a measure space and suppose that f : X → F is measurable.
Let f # : [0, ∞) → [0, ∞] be its decreasing rearrangement. Then the following hold.
1. For s, t ∈ [0, ∞) we have that s < f # (t) if and only if t < df (s).

2. f # is right continuous.

3. If s ∈ [0, ∞), then

df (s) = µ({x ∈ X | |f (x)| > s}) = L1 ({t ∈ [0, ∞) | f # (t) > s}), (1.1.78)

where L1 denotes Lebesgue measure on R. In particular df = df # .

4. For every 1 ≤ p < ∞ we have that

Z Z ∞
p
|f | dµ = (f # (t))p dt (1.1.79)
X 0

and we also have that

esssup |f | = inf{C > 0 | |f | ≤ C a.e. in X} = sup f # (t) = f # (0). (1.1.80)

t>0

Proof. To prove the first item fix s, t ∈ [0, ∞). If s < f # (t) then s ∈ / {r | df (r) ≤ t}, so we must
have that t < df (s). Now suppose that t < df (s). Proposition 1.1.11 guarantees that df is right
continuous and nonincreasing, so we can pick ε > 0 such that t < df (r) for all 0 ≤ r < s + ε. This
means that [0, s + ε] ∩ {r | df (r) ≤ t} = ∅, and hence s < s + ε ≤ f # (t). This completes the proof
of the first item.
To prove the second item we fix T ∈ [0, ∞). If f # (T ) = 0 then f # (r) = 0 for all r ≥ T , and so
f is trivially right continuous at T . Suppose then, that f # (T ) > 0 and pick 0 < s < f # (T ). The
#

first item then implies that T < df (s), which then allows us to choose T < r < df (s). In turn, the
first item implies that s < f # (t) for all t ∈ [T, r]. Then since f # is nonincreasing we deduce that

s ≤ lim+ f # (t) ≤ f # (T ). (1.1.81)

t→T

This inequality holds for arbitrary 0 < s < f # (T ), and hence

f # (T ) = lim+ f # (t), (1.1.82)

t→T

which proves that f # is right continuous at T . This proves the second item.
To prove the third item fix s ∈ [0, ∞). The first item then implies that

L1 ({t ∈ [0, ∞) | f # (t) > s}) = L1 ({t ∈ [0, ∞) | df (s) > t}) = L1 ([0, df (s))) = df (s), (1.1.83)

16
which proves the third item.
Finally, we prove the fourth item. The third item tells us that df = df # , so Corollary 1.1.15
implies the fourth item for 1 ≤ p < ∞. Finally, we have that

esssup |f | = inf{s ≥ 0 | df (s) = 0} = inf{s ≥ 0 | df # (s) = 0} = sup f # (t), (1.1.84)

t>0

which completes the proof of the fourth item.

The following result serves as an analog of the monotone convergence theorem and Fatou’s
lemma for decreasing rearrangements.
Proposition 1.1.23. Let (X, M, µ) be a measure space, and suppose that for each n ≥ ` ∈ Z the
function fn : X → F is measurable. Further suppose that f : X → F is measurable. Then the
following hold.
1. If {|fn |}∞ # ∞
n=` is a.e. nondecreasing and |fn | → |f | a.e. as n → ∞, then {fn }n=` is nondecreas-
ing and fn# → f # pointwise as n → ∞.

2. If
|f | ≤ lim inf |fn | a.e. in X, (1.1.85)
n→∞

then
f # ≤ lim inf fn# in [0, ∞). (1.1.86)
n→∞

Proof. We begin with the proof of the first item. According to Proposition 1.1.21 we have that
#
fn# ≤ fn+1 ≤ f # for all n ≥ `, and so

lim fn# (t) ≤ f # (t) for all t ∈ [0, ∞). (1.1.87)

n→∞

Suppose, by way of contradiction, that there exists t ∈ [0, ∞) such that

lim fn# (t) < f # (t) (1.1.88)

n→∞

and choose s ∈ [0, ∞) such that

lim fn# (t) < s < f # (t). (1.1.89)
n→∞

By the first item of Theorem 1.1.22 we then have that t < df (s). Proposition 1.1.12 tells us that
dfn → df as n → ∞, so we can find N ≥ ` such that n ≥ N implies that t < dfn (s) ≤ df (s), which
particular implies that s < fN# (t), a contradiction since {fn# }∞
n=` is nondecreasing. This proves the
first item.
To prove the second let us suppose, again by way of contradiction, that

lim inf fn# (t) < f # (t) (1.1.90)

n→∞

for some t ∈ [0, ∞). In particular we can pick s ∈ [0, ∞) such that

lim inf fn# (t) < s < f # (t). (1.1.91)

n→∞

Then again by Theorem 1.1.22 and Proposition 1.1.12 we have that

t < df (s) ≤ lim inf dfn (s), (1.1.92)

n→∞

17
which implies that there exists N ≥ ` such that

t < dfn (s) for all n ≥ N, (1.1.93)

and hence that

s < fn# (t) for all n ≥ N. (1.1.94)
Consequently,
s ≤ lim inf fn# (t) < s, (1.1.95)
n→∞

a contradiction. Hence
f # (t) ≤ lim inf fn# (t) for all t ∈ [0, ∞), (1.1.96)
n→∞

which proves the second item.

1.1.5 The Hardy-Littlewood rearrangement inequality

Our goal now is to prove a key inequality related to rearrangements. We begin with a lemma.
Lemma 1.1.24. Let (X, M, µ) be a measure space. Suppose that f : X → F is measurable, and let
E ∈ M. Then for every t ∈ [0, ∞) we have that

df χE (t) ≤ min{df (t), µ(E)} (1.1.97)

and
(f χE )# (t) ≤ f # (t)χ[0,µ(E)) (t). (1.1.98)
Moreover,
Z Z µ(E)
|f | dµ ≤ f # (t)dt. (1.1.99)
E 0

Proof. Fix t ∈ [0, ∞). Since E ⊆ X we may then bound

df χE (t) = µ({x ∈ E | |f (x)| > t}) ≤ min{df (t), µ(E)}. (1.1.100)

With this inequality established we then see that

{s ∈ [0, ∞) | df (s) ≤ t} ⊆ {s ∈ [0, ∞) | df χE (s) ≤ t}, (1.1.101)

which implies that (f χE )# (t) ≤ f # (t). Similarly, if µ(E) ≤ t then df χE (s) ≤ µ(E) ≤ t for all
s ∈ [0, ∞) and hence (f χE )# (t) = 0. Thus

(f χE )# (t) ≤ f # (t)χ[0,µ(E)) (t). (1.1.102)

The inequality (1.1.99) then follows directly from the last bound and Corollary 1.1.15.

We now have the tools needed to prove the aforementioned inequality. It serves as a sort of in-
termediate inequality that nestles between the terms encountered in the standard Hölder inequality.
Theorem 1.1.25 (Hardy-Littlewood rearrangement inequality). Let (X, M, µ) be a measure space,
and suppose that f, g : X → [0, ∞] are measurable. Then
Z Z ∞
f gdµ ≤ f # (t)g # (t)dt. (1.1.103)
X 0

18
Proof. We begin by making two reductions. First, it suffices to prove the result under the assump-
tion that f 6= 0 and g 6= 0, i.e. neither f nor g is trivial, as in this case the inequality is trivially
satisfied. Second, we claim that it suffices to prove the result when f is simple. Indeed, suppose the
result is proved whenever f, g : X → [0, ∞) and f is simple. We then choose a sequence {fn }∞ n=0 of
non-negative simple functions such that fn % f a.e. as n → ∞. Applying the result, we find that
Z Z ∞
fn gdµ ≤ fn# (t)g # (t)dt (1.1.104)
X 0

According to Proposition 1.1.23 we have that fn# % f # on [0, ∞) as n → ∞, so the monotone

convergence theorem allows us to send n → ∞ in (1.1.104) to deduce that (1.1.103) holds. This
proves the claim.
Assume then that f, g : X → [0, ∞) and that f is simple. We may write
n
X
f= ai χEi , (1.1.105)
i=1

where E1 , . . . , En ∈ M are pairwise disjoint, and 0 := an+1 < an < · · · < a1 < ∞. We saw in
Examples 1.1.18 and 1.1.19 that
m
X
#
f (t) = aj χ[bj−1 ,bj ) (t), (1.1.106)
j=1

where (
n if f is finite
m= (1.1.107)
min{1 ≤ i ≤ n | µ(Ei ) = ∞} otherwise ,
b0 = 0, and
j
X
bj = µ(Ei ) ∈ (0, ∞] if 1 ≤ j ≤ m. (1.1.108)
i=1
Sj
Write F0 = ∅ and for 1 ≤ j ≤ n set Fj = i=1 Ei and dj = aj − aj+1 . This allows us to rewrite
n
X
f= dj χFj . (1.1.109)
j=1

Similarly, we have the identity

m−1
X
dj χFj (x) = f (x) − am for all x ∈ Fm−1 . (1.1.110)
j=1

Now we use Theorem 1.1.22, Lemma 1.1.24, (1.1.106) and (1.1.110) to bound
Z Z Z Z Z Z
f gdµ = f gdµ = f gdµ + f gdµ ≤ f gdµ + am gdµ
X Fn Fm−1 Fn \Fm−1 Fm−1 Fn \Fm−1
Z Z m−1
X Z Z
= (f − am )gdµ + am gdµ = dj g χFj dµ + am g χFn dµ
Fm−1 Fn j=1 X X

m−1
X Z ∞ Z ∞ m−1
X Z µ(Fj ) Z µ(Fn )
# # #
= dj (g χFj ) (t)dt + am (g χFn ) (t)dt ≤ dj g (t)dt + am g # (t)dt.
j=1 0 0 j=1 0 0

(1.1.111)

19
We then compute
m−1
X Z µ(Fj ) m−1
X Z µ(Fj )
#
dj g (t)dt = (aj − aj+1 ) g # (t)dt
j=1 0 j=1 0

m−1
X Z µ(Fj ) m−1
X Z µ(Fj )
#
= aj g (t)dt − aj+1 g # (t)dt
j=1 0 j=1 0

m−1
X Z µ(Fj ) m
X Z µ(Fj−1 ) Z µ(Fm−1 ) m−1
X Z µ(Fj )
# # #
= aj g (t)dt− aj g (t)dt = −am g (t)dt+ aj g # (t)dt.
j=1 0 j=1 0 0 j=1 µ(Fj−1 )

(1.1.112)
Chaining the previous two estimates together then reveals that
Z m−1
X Z µ(Fj ) Z µ(Fn )
#
f gdµ ≤ aj g (t)dt + am g # (t)dt
X j=1 µ(Fj−1 ) µ(fm−1 )
Z ∞ m−1
X Z µ(Fn )
#
= aj χ[bj−1 ,bj ) (t)g (t)dt + am g # (t)dt. (1.1.113)
0 j=1 µ(Fm−1 )

We now split to cases. If f is finite, then m − 1 = n − 1 and µ(Fn ) < ∞, so (1.1.106) allows us
to compute
Z ∞ m−1
X Z µ(Fn ) Z ∞ n
X
# #
aj χ[bj−1 ,bj ) (t)g (t)dt + am g (t)dt = aj χ[bj−1 ,bj ) (t)g # (t)dt
0 j=1 µ(Fm−1 ) 0 j=1
Z ∞
= f # (t)g # (t)dt. (1.1.114)
0

On the other hand, if f is not finite, them µ(Fn ) = µ(Fm ) = bm = µ(Em ) = ∞, so (1.1.106) allows
us to compute
Z ∞ m−1
X Z µ(Fn )
#
aj χ[bj−1 ,bj ) (t)g (t)dt + am g # (t)dt
0 j=1 µ(Fm−1 )
Z ∞ m−1
X Z bm Z m
∞X
# #
= aj χ[bj−1 ,bj ) (t)g (t)dt + am g (t)dt = aj χ[bj−1 ,bj ) (t)g # (t)dt
0 j=1 bm−1 0 j=1
Z ∞
= f # (t)g # (t)dt. (1.1.115)
0

Combining (1.1.113), (1.1.114), and (1.1.115) then proves (1.1.103).

We proved the Hardy-Littlewood inequality for non-negative functions, but it also works just as
well for more general maps.
Corollary 1.1.26. Let (X, M, µ) be a measure space and let f, g : X → F be measurable. Then
Z Z Z ∞
f gdµ ≤ |f g| dµ ≤ f # (t)g # (t)dt. (1.1.116)
X X 0

Proof. Exercise.

20
1.1.6 Lorentz spaces
Our aim now is to define a larger class of functions than the Lebesgue spaces. To motivate this we
begin with the following simple result.

Proposition 1.1.27. Let (X, M, µ) be a measure space, 1 ≤ p < ∞, and f : X → F be measurable.

Then
sup t(df (t))1/p = sup t1/p f # (t). (1.1.117)
t>0 t>0

Proof. Write
A = sup t(df (t))1/p and B = sup t1/p f # (t). (1.1.118)
t≥0 t≥0

We will prove that B ≤ A ≤ B.

Suppose, by way of contradiction, that A < B and pick s ∈ (0, ∞) such that A < s < B. By
the definition of B we may then choose t > 0 such that s < t1/p f # (t). By Theorem 1.1.22 we then
see that
s #
 s  s h  s i1/p
< f (t) ⇒ t < d f ⇒ s < df 1/p ≤ A, (1.1.119)
t1/p t1/p t1/p t
which is a contradiction. Thus B ≤ A.
Now suppose, again by way of contradiction, that B < A and pick B < s < A. Choose t > 0
such that s < t(df (t))1/p . Then Theorem 1.1.22 again tells us that

sp
 p
s # sp
 
# s
< df (t) ⇒ t < f ⇒s< f ≤ B, (1.1.120)
tp tp t tp

which is a contradiction. Thus A ≤ B.

Consider a function f ∈ Lp (X; F). We know from Theorem 1.1.22 that
Z ∞ 1/p Z ∞ 1/p
# p 1/p # p dt
kf kLp = (f (t)) dt = (t f (t)) = (·)1/p f # Lpµ ((0,∞)
, (1.1.121)
0 0 t

where we have defined the measure µ = dt/t on (0, ∞). On the other hand, Chebyshev’s inequality
and Proposition 1.1.27 tell us that

sup t1/p f # (t) = sup t(df (t))1/p ≤ kf kLp . (1.1.122)

t>0 t>0

The null sets of µ and L1 (standard Lebesgue measure) agree, and so L∞ ∞

µ ((0, ∞)) = LL1 ((0, ∞)).
Thus
sup t1/p f # (t) = (·)1/p f # (·) L∞ µ((0,∞)) . (1.1.123)
t>0

These computations suggest that there is interesting information encoded in the quantities

(·)1/p f # (·) Lqµ ((0,∞))

for 1 ≤ q ≤ ∞. (1.1.124)

Indeed, there is! We begin our exploration of this by defining some new spaces.

Definition 1.1.28. Let (X, M, µ) be a measure space.

21
 1. Let f : X → F is measurable, 1 ≤ p ≤ ∞, and 1 ≤ q ≤ ∞. For q < ∞ we define
Z ∞ 1/q
1/p # q dt
|||f |||Lp,q = (t f (t)) ∈ [0, ∞], (1.1.125)
0 t
and for q = ∞ we define
|||f |||Lp,∞ = sup{t1/p f # (t) | t > 0} ∈ [0, ∞]. (1.1.126)

2. For 1 ≤ p, q ≤ ∞ we define the Lorentz space Lp,q to be

Lp,q (X; F) = Lp,q (X, µ; F) = {[f ]' | f : X → F is measurable and |||f |||Lp,q < ∞} (1.1.127)
where once more [·]' is the equivalence class generated by a.e. equality. We will write
Lp,q (X) = Lp,q (X; R).
Let’s consider some examples. The first shows that Sf in (X; F) ⊆ Lp,q (X; F) when p, q < ∞.
Example 1.1.29. Let (X, M, µ) be a measure space. Suppose that f : X → [0, ∞) is a finite
simple function given by
Xn
f= ai χEi (1.1.128)
i=1
where E1 , . . . , En ∈ M are finite measure sets that are pairwise disjoint, and 0 < an < · · · < a1 < ∞.
Set an+1 = 0. We saw in Example 1.1.18 that if we set
j
X
bj = µ(Ei ). (1.1.129)
i=1

for 1 ≤ j ≤ n, and b0 = 0, then

n
X
f # (t) = aj χ[bj−1 ,bj ) (t). (1.1.130)
j=1

For 1 ≤ p, q < ∞ we then compute

Z ∞ n Z bj n
q dt p X q  q/p
X 
q 1/p # q q/p−1 q/p
|||f |||Lp,q = (t f (t)) = aj t dt = aj bj − bj−1 . (1.1.131)
0 t j=1 bj−1 q j=1

4
The next example shows just how inclusive the Lorentz space Lp,∞ can be.
Example 1.1.30. Let 1 ≤ p < ∞ and consider X = Rn with Lebesgue measure. For α > 0 define
f : Rn → R via f (0) = 0 and f (x) = |x|−α . Then
1 1
t< α ⇔ |x| < 1/α (1.1.132)
|x| t
and so
df (t) = ωn t−n/α (1.1.133)
n p,∞
where ωn = L (B(0, 1)). Consequently, f ∈ L if and only if p = n/α, in which case
|||f |||Lp,∞ = ωn1/p = ωnα/n . (1.1.134)
The bound 1 ≤ p requires that
0 < α ≤ n. (1.1.135)
In particular, if α = n/p for a given 1 ≤ p, then f ∈ Lp,∞ (Rn ). 4

22
 We used Proposition 1.1.27 to justify our choice of the form of |||·|||Lp,q , and this result showed that
|||f |||Lp,∞ = supt>0 t(df (t))1/p , i.e. we can recharacterize this quantity in terms of the distribution
function. We can do something similar for the integral quantities as well.

Proposition 1.1.31. Let (X, M, µ) be a measure space, V be a finite dimensional normed vector
space, 1 ≤ q < ∞, and 0 < α < ∞. Then
Z ∞
q ∞ q
Z
α # q dt ds
t (f (t)) = s (df (s))α . (1.1.136)
0 t α 0 s
In particular, for 1 ≤ p < ∞ we have that
Z ∞ 1/q
1/q p q/p ds
|||f |||Lp,q = p s (df (s)) . (1.1.137)
0 s

Proof. The second assertion follows from the first by setting α = q/p. To prove the first we compute,
using Tonelli’s theorem and the fundamental theorem of calculus:
Z ∞ Z ∞ Z f # (t) Z ∞ Z ∞
α # q dt α−1 q−1 α−1
t (f (t)) = t qs tdsdt = qsq−1 χ(0,f # (t)) (s)dsdt
0 t 0 0 0 0
Z ∞ Z ∞ Z ∞ Z
=q sq−1 α−1
t χ(0,f # (t)) (s)dtds = q s q−1
tα−1 dtds. (1.1.138)
0 0 0 {t>0 | f # (t)>s}

However, from Theorem 1.1.22 we know that

{t > 0 | s < f # (t)} = {t > 0 | t < df (s)} = (0, df (s)). (1.1.139)

Thus,
Z ∞ Z ∞ Z df (s) Z ∞
α # q dt q−1 α−1 q
t (f (t)) =q s t dtds = sq−1 (df (s))α ds, (1.1.140)
0 t 0 0 α 0

which is the desired equality.

Remark 1.1.32. The identity used here can be significantly generalized. Indeed, if µ, ν are Radon
measures on [0, ∞) define ϕµ , ϕν : [0, ∞) → [0, ∞) via ϕµ (t) = µ([0, t)) and ϕν (t) = ν([0, t)). Then
for Radon measures µ and ν on [0, ∞) we have that
Z ∞ Z ∞
#
ϕµ ◦ f (t)dν(t) = ϕν ◦ df (s)dµ(s). (1.1.141)
0 0

We leave it as an exercise to verify this.

Now we record some very basic properties of Lorentz spaces.

Proposition 1.1.33. Let (X, M, µ) be a measure space and 1 ≤ p, q ≤ ∞. Then the following hold.

1. Lp,p (X; F) = Lp (X; F) and |||·|||Lp,p = k·kLp .

2. If 1 ≤ q < ∞ and µ(X) > 0 then L∞,q (X; F) = {0}.

23
 3. If 1 ≤ p < ∞ and 1 ≤ q ≤ ∞ or p = q = ∞, then

Sf in (X; F) ⊆ Lp,q (X; F), (1.1.142)

where we recall that Sf in (X; F) denotes the set of simple functions of finite support.
Proof. Exercise.
At this point it is not clear why we have used the symbol |||·||| in place of the usual norm symbol
k·k. The reason, unfortunately, is that |||·|||Lp,q is not actually a norm in general. We investigate
this now.
Proposition 1.1.34. Let (X, M, µ) be a measure space and 1 ≤ p, q ≤ ∞. Let f, g : X → F be
measurable and suppose that |||f |||Lp,q , |||g|||Lp,q < ∞. Then the following hold.
1. |||f |||Lp,q ≥ 0 and |||f |||Lp,q = 0 if and only if f = 0 a.e.

2. If α ∈ F then |||αf |||Lp,q = |α| |||f |||Lp,q .

3. We have that |||f + g|||Lp,q ≤ 21/p |||f |||Lp,q + 21/p |||f |||Lp,q .

4. Lp,q (X; F) is a vector space.

Proof. The first and second items are trivial. We know from Proposition 1.1.21 that

(f + g)# (t) ≤ f # (t/2) + g # (t/2). (1.1.143)

If p, q < ∞, then Minkowski’s inequality provides us with the bound

Z ∞ 1/q Z ∞ 1/q
1/p # q dt 1/p # q dt
|||f + g|||Lp,q = (t (f + g) (t)) ≤ (t f (t/2))
0 t 0 t
Z ∞ 1/q Z ∞ 1/q Z ∞ 1/q
1/p # q dt 1/p 1/p # q dt 1/p 1/p # q dt
+ (t g (t/2)) =2 (t f (t)) +2 (t g (t/2))
0 t 0 t 0 t
= 21/p |||f |||Lp,q + 21/p |||g|||Lp,q . (1.1.144)

A similar argument, which we leave as an exercise, shows that (1.1.144) also holds when either
p = ∞ or q = ∞. This completes the proof of the third item. The fourth item then follows from
the second and third.
In general the bound in the third item cannot be improved, as we show in the following example.
Example 1.1.35. Let 1 ≤ p < ∞ and define f, g : (0, 1) → R via f (x) = x−1/p and g(x) =
(1 − x)−1/p . Note that for t > 0
1
f (x) > t ⇔ p > x (1.1.145)
t
and so (
1 if 0 ≤ t < 1
df (t) = p
(1.1.146)
1/t if 1 < t.
which together with Proposition 1.1.27 implies that

|||f |||Lp,∞ = sup t(df (t))1/p = 1. (1.1.147)

t>0

24
Similarly, |||g|||Lp,∞ = 1.
Next write h = f + g and note that
1 1
0 = h0 (x) = − x−1/p−1 + (1 − x)−1/p−1 ⇔ x = 1 − x ⇔ x = 1/2. (1.1.148)
p p

From this and the fact that h diverges at 0 and 1 we deduce that f achieves its minimum at x = 1/2
and
min h(x) = h(1/2) = 21/p + 21/p = 21+1/p . (1.1.149)
0<x<1

Thus,
df +g (t) = dh (t) = 1 for 0 < t < 21+1/p , (1.1.150)
and in turn this and Proposition 1.1.27 imply that

|||f + g|||Lp,∞ ≥ lim t11/p = 21+1/p . (1.1.151)

t→21+1/p

From these calculations we then find that

|||f + g|||Lp,∞
≥ 21/p > 1. (1.1.152)
|||f |||Lp,∞ + |||g|||Lp,∞

When combined with the bound from the previous proposition, we deduce that this bound is actually
an equality.
4

We now give a name to the norm variant we have discovered.

Definition 1.1.36. Let V be a vector space. A function |||·||| : V ×V → [0, ∞) is called a quasinorm
if

1. |||v||| = 0 if and only if v = 0,

2. |||αv||| = |α| |||v||| for every v ∈ V and α ∈ F,

3. There exists a constant C ≥ 1 such that |||v + w||| ≤ C(|||v||| + |||w|||) for all v, w ∈ V .

We say V is a quasinormed space if it is equipped with a quasinorm |||·|||. Given a quasinorm we

can define convergent and Cauchy sequences in the obvious way. With these notions in hand we
can say that V is quasi-Banach if every Cauchy sequence is convergent.

Example 1.1.37. If (X, M, µ) is a measure space, then Lp,q (X; F) is a quasinormed space. 4

Quasinorms are almost as good as norms, but there are certain technical results that they
don’t provide. For example, the following version of the generalized triangle inequality holds in
quasinormed spaces.

Proposition 1.1.38. Let V be a vector space equipped with a quasinorm |||·||| with constant C ≥ 1.
If v1 , . . . , vm ∈ V for m ≥ 2, then
m
X
|||v1 + · · · + vm ||| ≤ C m−1 |||vi ||| . (1.1.153)
i=1

25
Proof. Exercise. Hint: induct on m ≥ 2.
Next we show an essential inclusion result for the Lorentz spaces: Lp,q (X; F) ⊆ Lp,r (X; F) when
q < r.
Theorem 1.1.39. Let (X, M, µ) be a measure space, 1 ≤ p ≤ ∞, and 1 ≤ q < r ≤ ∞. If
f ∈ Lp,q (X; F) then f ∈ Lp,r (X; F) and
 1/q−1/r
q
|||f |||Lp,r ≤ |||f |||Lp,q . (1.1.154)
p
In particular, we have the subspace inclusions
Lp,1 (X; F) ⊆ Lp,q (X; F) ⊆ Lp,r (X; F) ⊆ Lp,∞ (X; F). (1.1.155)
Proof. We first consider the case r = ∞, in which case q < ∞ by assumption. Fix t > 0 and write
 Z t 1/q
1/p q 1/p q ds
t = (s ) . (1.1.156)
p 0 s
Then since f # is nonincreasing we can bound
 Z t 1/q  Z t 1/q  1/q
1/p # q 1/p # q ds q 1/p # q ds q
t f (t) = (s f (t)) ≤ (s f (s)) ≤ |||f |||Lp,q . (1.1.157)
p 0 s p 0 s p
Since t > 0 was arbitrary, we deduce that f ∈ Lp,∞ (X; F) and
 1/q
1/p # q
|||f |||Lp,∞ = sup t f (t) ≤ |||f |||Lp,q . (1.1.158)
t>0 p
Now suppose that 1 ≤ q < r < ∞. We compute
Z ∞ Z ∞ Z ∞
1/p # r dt 1/p # r−q 1/p # q dt r−q dt
(t f (t)) = (t f (t)) (t f (t)) ≤ |||f |||Lp,∞ (t1/p f # (t))q
0 t 0 t 0 t
= |||f |||Lp,∞ |||f |||qLp,q , (1.1.159)
r−q

which then implies that

 1/q−1/r
1−q/r q/r q
|||f |||Lp,r ≤ |||f |||Lp,∞ |||f |||Lp,q ≤ |||f |||Lp,q . (1.1.160)
p

Next we establish a version of Hölder’s inequality. We begin with a version for functions taking
values in a field F.
Theorem 1.1.40 (Hölder’s inequality for Lorentz spaces). Let (X, M, µ) be a measure space and
0 0
1 ≤ p, q ≤ ∞. Suppose that f ∈ Lp,q (X; F) and g ∈ Lp ,q (X; F), where
1 1 1 1
+ 0 = 1 and + 0 = 1. (1.1.161)
p p q q
Then f g ∈ L1 (X; F) and
Z Z
f gdµ ≤ |f | |g| dµ ≤ |||f |||Lp,q |||g|||Lp0 ,q0 . (1.1.162)
X X

26
Proof. It clearly suffices to prove only the second bound in (1.1.162). To prove it we use Theorem
1.1.25 and the standard Hölder inequality to bound
Z ∞ Z ∞ 1/p 0
t1/p #
Z
# # t #
|f | |g| dµ ≤ f (t)g (t)dt = f (t) 1/q0 g (t)dt ≤ |||f |||Lp,q |||g|||Lp0 ,q0 . (1.1.163)
X 0 0 t1/q t
This is the desired estimate.

A typical exercise in Lebesgue theory shows that the Lebesgue spaces satisfy a nice interpolation
property: inclusion in Lp0 and Lp1 for p0 < p1 implies inclusion in Lp for p0 < p < p1 . The same
is true in Lorentz spaces, though the inclusion is actually a bit stronger than we might expect. We
conclude our initial discussion of Lorentz spaces by exploring this now.

Theorem 1.1.41. Let (X, M, µ) be a measure space, 1 ≤ p0 < p1 ≤ ∞, and 1 ≤ q0 , q1 ≤ ∞. If

f ∈ Lp0 ,q0 (X; F) ∩ Lp1 ,q1 (X; F), then f ∈ Lp,q (X; F) for every p0 < p < p1 and 1 ≤ q ≤ ∞, and
 1/q  θ/q0  (1−θ)/q1  1/q
p q0 q1 p0 p1
|||f |||Lp,q ≤ + |||f |||θLp0 ,q0 |||f |||1−θ
Lp1 ,q1 (1.1.164)
q p0 p1 p − p0 p1 − p

where
1/p − 1/p1
θ= ∈ (0, 1). (1.1.165)
1/p0 − 1/p1
Proof. We will prove the result when q < ∞ and leave the case q = ∞ as an exercise. The result
is trivial if f = 0, so we may assume that f 6= 0.
Fix T ∈ (0, ∞). Theorem 1.1.39 guarantees that f ∈ Lp0 ,∞ (X; F) ∩ Lp1 ,∞ (X; F), so we have the
estimate
f # (t) ≤ min{t−1/p0 |||f |||Lp0 ,∞ , t−1/p1 |||f |||Lp1 ,∞ }. (1.1.166)
We may then bound
Z ∞ Z T Z ∞
q dt q dt dt
|||f |||qLp,q = 1/p #
(t f (t)) = (t 1/p #
f (t)) + (t1/p f # (t))q
0 t 0 t T t
Z T Z ∞
dt dt
≤ |||f |||qLp1 ,∞ (t1/p−1/p1 )q + |||f |||qLp0 ,∞ (t1/p−1/p0 )q
0 t T t
q
|||f |||Lp1 ,∞ α |||f |||qLp0 ,∞ −β
= T + T (1.1.167)
α β
for
q q q q
α= − > 0 and β = − > 0. (1.1.168)
p p1 p0 p
If a, b > 0, then the map T 7→ aα−1 T α + bβ −1 T −β achieves its global minimum at
 1/(α+β)
b
Tmin = , (1.1.169)
a

and the minimal value is  

1 1
+ aβ/(α+β) bα/(α+β) . (1.1.170)
α β

27
 Since f 6= 0 we may then minimize the right side of (1.1.41) as a function of T . Doing so, we
arrive at the bound
 1/q  1/q
p p0 p1
|||f |||Lp,q ≤ + |||f |||θLp0 ,∞ |||f |||1−θ
Lp1 ,∞ . (1.1.171)
q p − p0 p1 − p
Then (1.1.164) follows by chaining this bound together with the bound of Theorem 1.1.39.
Finally, we turn to the issue of completeness of Lorentz spaces, establishing that they are indeed
quasi-Banach.
Theorem 1.1.42. Let (X, M, µ) be a measure space and 1 ≤ p, q ≤ ∞. Then Lp,q (X; F) is a
quasi-Banach space.
Proof. We will prove the result when F = R and leave the details of how to extend this to F = C
as an exercise.
If p = ∞ and q < ∞ then Lp,q (X) = L∞,q (X) = {0} and is thus trivially complete. If p = q = ∞
then Lp,q (X) = L∞ (X) and is thus complete due to the completeness of the Lebesgue spaces. It
remains to handle the case 1 ≤ p < ∞ and 1 ≤ q ≤ ∞. We will prove the result under the extra
assumption that q < ∞ and leave the case q = ∞ as an exercise.
Suppose that {fn }∞ p,q
n=` ⊆ L (X) is Cauchy. According to Proposition 1.1.27 and Theorem 1.1.39
we have that  1/q
1/p 1/p # q
sup t(dg (t)) = sup t g (t) ≤ |||g|||Lp,q (1.1.172)
t>0 t>0 p
for every g ∈ Lp,q (X). Applying this to g = fn − fm shows that {fn }∞n=` is Cauchy in measure, and
so we can find a measurable function f : X → R and a subsequence {fnk }∞ k=` such that fnk → f
a.e. in X as k → ∞.
For ` ≤ j ∈ Z we have that
f − fnj = lim fnk − fnj a.e. in X, (1.1.173)
k→∞

and so Proposition 1.1.23 implies that

(f − fnj )# ≤ lim inf (fnk − fnj )# in [0, ∞). (1.1.174)
k→∞

This and Fatou’s lemma then allow us to bound

Z ∞ Z ∞
q 1/p # q dt dt
f − fnj Lp,q = (t (f − fnj ) (t)) ≤ lim inf (t1/p (fnk − fnj )# (t))q
0 t 0 k→∞ t
Z ∞
dt q
≤ lim inf (t1/p (fnk − fnj )# (t))q = lim inf fnk − fnj Lp,q
. (1.1.175)
k→∞ 0 t k→∞

Upon sending j → ∞ and utilizing the fact that {fn }∞ p,q

n=` is Cauchy in L (X), we deduce that
fnk − f → 0 in Lp,q (X) as k → ∞. In turn, since quasi-Cauchy sequences are bounded (the proof
for Cauchy sequences works), this and Proposition 1.1.23 imply that
Z T Z T Z ∞
1/p # q dt 1/p # q dt dt
(t f (t)) ≤ lim inf (t fnj (t)) ≤ lim inf (t1/p fn#j (t))q
0 t j→∞ 0 t j→∞ 0 t
q
= lim inf fnj Lp,q < ∞ (1.1.176)
j→∞
p,q
for all T > 0, and hence that f ∈ L (X). We leave it as an exercise to check that Cauchy sequences
in quasinormed spaces with convergent subsequences are actually convergent. Thus fn → f in
Lp,q (X) as n → ∞, which proves that Lp,q (X) is complete.

28
1.2 Interpolation theorems
Our goal in this section is to exploit the properties of the Lebesgue and Lorentz spaces in order to
prove some interpolation results for operators between them. The point of this is that it is often
easy to check that an operator (say, linear) is bounded between certain pairs of spaces. We will look
for results that allow us to extend from these special pairs to other pairs that interpolate between.
We will prove two main results: the Marcinkiewicz and Riesz-Thorin interpolation theorems.

1.2.1 The real method of Marcinkiewicz

We begin with some definitions, starting with the notion of truncation.

Definition 1.2.1. Let (X, M, µ) be a measure space.

1. We write
L0 (X; F) = {[f ]' | f : X → F is measurable}. (1.2.1)
We know from measure theory that L0 (X; F) is a vector space over F.

2. Given f ∈ L0 (X; F) and t ∈ [0, ∞) we define the truncation of f to be ft : X → F given by

(
f (x) if |f (x)| ≤ t
ft (x) = (1.2.2)
0 if |f (x)| > t.

Clearly ft ∈ L0 (X; F).

3. Suppose that E ⊆ L0 (X; F). We say that E is closed under truncation if f ∈ E implies that
ft ∈ E for every t ∈ [0, ∞).

Example 1.2.2. Simple functions are closed under truncation, as are finite simple functions. 4

Example 1.2.3. Let (X, M, µ) be a measure space. Then Lp,q (X; F) is closed under truncation for
every 1 ≤ p, q ≤ ∞. The same is true of Lp0 ,q0 (X; F) ∩ Lp1 ,q1 (X; F) and Lp0 ,q0 (X; F) + Lp1 ,q1 (X; F).
To see this note that if fr is the truncation of the measurable function f : X → F, then
(
df (t) − df (r) for 0 ≤ t < r
dfr (t) = (1.2.3)
0 for t ≥ r

and so

fr# (t) = inf{s | dfr (s) ≤ t} = inf{0 ≤ s < r | df (s) − df (r) ≤ t} = f # (df (r) + t). (1.2.4)

In particular, this means that

fr# (t) ≤ min{f # (t), r} (1.2.5)
and so fr ∈ Lpi ,qi (X; F) whenever f ∈ Lpi ,qi (X; F).
4

We now prove one of our two main interpolation results.

29
Theorem 1.2.4 (Marcinkiewicz). Let (X, M, µ) and (Y, N, ν) be measure spaces and let 1 ≤
pi , ri , si ≤ ∞ for i = 0, 1 be such that p0 < p1 , r0 6= r1 , s0 = 1, and
(
1 if p1 < ∞
s1 = (1.2.6)
∞ if p1 = ∞.

Let U ⊆ L0 (X; F) be a subspace closed under truncation, and suppose that T : U → L0 (Y ; F) is

such that there exist A, C0 , C1 > 0 such that
|T (f + g)| ≤ A (|T (f )| + |T (g)|) ν − a.e. in Y (1.2.7)
for every f, g ∈ U , and
|||T (f )|||Lri ,∞ (Y ;F) ≤ Ci |||f |||Lpi ,si (X;F) (1.2.8)
for every f ∈ U ∩ Lpi ,si (X; F) and i = 0, 1. Then for every θ ∈ (0, 1) there exists a constant
C = C(θ, pi , ri , si , A, Ci ) > 0 such that
|||T (f )|||Lr,q (Y ;F) ≤ C |||f |||Lp,q (X;F) (1.2.9)
for every f ∈ U ∩ Lp,q (X; F), where 1 ≤ q ≤ ∞ and
1 1−θ θ 1 1−θ θ
= + and = + . (1.2.10)
p p0 p1 r r0 r1
Proof. Note that p0 < p < p1 . We will prove the result under the assumption that 1 ≤ p1 , q < ∞
and leave the remaining cases as an exercise. Throughout the proof we will employ the convention
of writing C > 0 for a generic constant depending on θ, p0 , p1 , r0 , r1 , q, A, C0 , and C1 that may
change from line to line. Let f ∈ U ∩ Lp,q (X; F). We divide the rest of the proof into several steps.
Step 1 – Set up
The relations (1.2.10) and some simple algebra show that
1/r0 − 1/r 1/p0 − 1/p 1/r − 1/r1 1/p − 1/p1
=θ= and =1−θ = (1.2.11)
1/r0 − 1/r1 1/p0 − 1/p1 1/r0 − 1/r1 1/p0 − 1/p1
and hence we may define
1/r0 − 1/r 1/r0 − 1/r1 1/r − 1/r1
γ := = = > 0. (1.2.12)
1/p0 − 1/p 1/p0 − 1/p1 1/p − 1/p1
For τ > 0 define fτ , gτ : X → F via
(
f (x) if |f (x)| ≤ f # (τ γ )
fτ (x) = (1.2.13)
0 otherwise
and (
f (x) if |f (x)| > f # (τ γ )
gτ (x) = (1.2.14)
0 otherwise.
Since U is closed under truncation we have that fτ ∈ U , and since gτ = f − fτ and U is a vector
space we have that gτ ∈ U . By construction we have that |fτ | , |gτ | ≤ |f |, and so fτ# , gτ# ≤ f # .
Additionally, a simple computation, which we leave as an exercise, reveals that
(
f # (τ γ ) if 0 < t < τ γ
fτ# (t) ≤ (1.2.15)
f # (t) if τ γ ≤ t

30
and (
f # (t) if 0 < t < τ γ
gτ# (t) ≤ (1.2.16)
0 if τ γ ≤ t.
Step 2 – Estimates for fτ and gτ
Hölder’s inequality and the bounds 1/p1 < 1/p < 1/p0 imply that

Z ∞ Z τγ Z τγ 1/q0 Z τ γ 1/q
dt dt dt
q 0 /p −q 0 /p q dt
t1/p0 gτ# (t) ≤ t1/p0 f # (t) ≤ t 0 1/p #
(t f (t))
0 t 0 t 0 t 0 t
Z τ γ 1/q
dt
≤ Cτ γ(1/p0 −1/p) (t1/p f # (t))q < ∞ (1.2.17)
0 t

and
Z ∞ Z τγ Z ∞
1/p1 dt 1/p1 dt dt
t fτ# (t) ≤ f (τ )# γ
t + t1/p1 f # (t)
0 t 0 t τγ t
Z ∞ 1/q0 Z ∞ 1/q
# γ γ/p1 q 0 /p1 −q 0 /p dt 1/p # q dt
≤ p1 f (τ )τ + t (t f (t))
τγ t τγ t
Z ∞ 1/q
# γ γ/p1 γ(1/p1 −1/p) 1/p # q dt
≤ Cf (τ )τ + Cτ (t f (t)) < ∞ (1.2.18)
τγ t

from which we deduce that

gτ ∈ Lp0 ,1 (X; F) and fτ ∈ Lp1 ,1 (X; F) (1.2.19)

for every τ > 0.

We know from (1.2.8) that

|||T (h)|||Lri ,∞ (Y ;F) ≤ C |||h|||Lpi ,1 (X;F) for i = 0, 1 (1.2.20)

whenever h ∈ U ∩ Lpi ,1 (X; F). Applying this to fτ and gτ shows that

|||T (gτ )|||Lr0 ,∞ (Y ;F) ≤ C |||gτ |||Lp0 ,1 (X;F) and |||T (fτ )|||Lr1 ,∞ (Y ;F) ≤ C |||fτ |||Lp1 ,1 (X;F) , (1.2.21)

which, when combined with the estimates (1.2.17) and (1.2.18), implies that for each τ > 0 we have
the bounds Z ∞ Z τγ
# −1/r0 1/p0 # dt −1/r0 dt
[T (gτ )] (τ ) ≤ Cτ t gτ (t) ≤ Cτ t1/p0 f # (t) (1.2.22)
0 t 0 t
and
Z ∞ Z ∞
# −1/r1 1/p1 dt γ/p1 −1/r1 −1/r1 dt
[T (fτ )] (τ ) ≤ Cτ t fτ# (t) ≤ Cf (τ )τ # γ
+ Cτ t1/p1 f # (t) . (1.2.23)
0 t τγ t

Step 3 – Estimates from (1.2.7)

From the bound (1.2.7) we deduce that

|T (f )| = |T (fτ + gτ )| ≤ A |T (fτ )| + A |T (gτ )| ν − a.e. in Y, (1.2.24)

31
and so Proposition 1.1.21 and Example 1.1.20 tell us that

[T (f )]# (t) ≤ A[|T (fτ )| + |T (gτ )|]# (t) ≤ A[|T (fτ )|]# (t/2) + A[|T (gτ )|]# (t/2)
= A[T (fτ )]# (t/2) + A[T (gτ )]# (t/2) (1.2.25)
for all t > 0, and in particular for t = 2τ . Consequently, a change of variable and Minkowski’s
inequality imply that
Z ∞ 1/q
1 1/r # q dτ
|||T (f )|||Lr,q (Y ;F) = (τ [T (f )] (2τ ))
21/r 0 τ
Z ∞ 1/q
1/r # # q dτ
≤C (τ [[T (fτ )] (τ ) + [T (gτ )] (τ )])
0 τ
Z ∞ 1/q Z ∞ 1/q
1/r # q dτ 1/r # q dτ
≤C (τ [T (fτ )] (τ )) +C (τ [T (gτ )] (τ )) . (1.2.26)
0 τ 0 τ
Step 4 – Synthesis and conclusion
Multiplying (1.2.22) and (1.2.23) by τ 1/r and employing the algebraic relations in (1.2.12) shows
that
Z τγ Z τγ
1/r # 1/r−1/r0 1/p0 # dt γ(1/p−1/p0 ) dt
τ (T (gτ )) (τ ) ≤ Cτ t f (t) = Cτ t1/p0 f # (t) (1.2.27)
0 t 0 t
and
Z ∞
1/r # # γ γ/p1 +1/r−1/r1 1/r−1/r1 dt
τ [T (fτ )] (τ ) ≤ Cf (τ )τ + Cτ t1/p1 f # (t)
τγ t
Z ∞
# γ γ/p γ(1/p−1/p1 ) dt
≤ Cf (τ )τ + Cτ t1/p1 f # (t) . (1.2.28)
τγ t
From (1.2.27), the change of variable z = τ γ (which implies dτ /τ = γ −1 dz/z), and Hardy’s inequality
we may then bound
Z ∞ 1/q Z ∞ Z τ γ q 1/q
1/r # q dτ γq(1/p−1/p0 ) 1/p0 # dt dτ
(τ [T (gτ )] (τ )) ≤C τ t f (t)
0 τ 0 0 t τ
Z ∞ Z z q 1/q Z ∞ 1/q
q(1/p−1/p0 ) 1/p0 # dt dz q−q(1/p0 −1/p)−1 1/p0 # q dt
=C z t f (t) ≤C t [t f (t)] q
0 0 t z 0 t
Z ∞ 1/q
dt
=C [t1/p f # (t)]q = C |||f |||Lp,q (X;V ) . (1.2.29)
0 t
Similarly, (1.2.28) allows us to bound
Z ∞ 1/q Z ∞ q
1/r # q dτ γq/p # γ q dτ
(τ (T (fτ )) (τ )) ≤C τ [f (τ )]
0 τ 0 τ
Z ∞ Z ∞ q q
γq(1/p−1/p1 ) 1/p1 # dt dτ
+C τ t f (t)
0 τγ t τ
Z ∞ q Z ∞ Z ∞ q q
1/p # q dz q(1/p−1/p1 ) 1/p1 # dt dz
=C [z f (z)] +C z t f (t)
0 z 0 z t z
Z ∞ q
dt
≤C [t1/p f # (t)]q = C |||f |||Lp,q (X;F) . (1.2.30)
0 t

32
 We now chain together the bounds (1.2.26), (1.2.29), and (1.2.30) to deduce that

|||T (f )|||Lr,q (Y ;F) ≤ C |||f |||Lp,q (X;F) . (1.2.31)

This is (1.2.9).
Some remarks are in order.
Remark 1.2.5. With a bit of elementary but tedious work, the exact form of the constant C =
C(θ, pi , ri , si , A, Ci ) > 0 that appears in the final estimate can be tracked in the proof of the
Marcinkiewicz theorem. If we were to do so we would find that the constant blows up as θ → 0
and θ → 1. This is, of course, not surprising: we should not expect to be able to improve our
assumed estimates for free as a byproduct of the theorem.
Remark 1.2.6. In the Marcinkiewicz theorem we do not assume that the map T is linear, and in
particular we make no assumptions about how T behaves relative to scalar multiplication. There is
then no good reason for us to require the functions on X and on Y take values in the same field.
In fact, an examination of the proof shows that the exact same results hold if we replace L0 (X; F)
and L0 (Y ; F) with L0 (X; F1 ) and L0 (Y ; F2 ). In other words, the fields we use as the target spaces
for our functions can be different. This is often useful in practical applications of the theorem.
Remark 1.2.7. Probably the most common use of Theorem 1.2.4 occurs in showing that the map T
is bounded between Lebesgue spaces. In order to get a Lebesgue estimate from the theorem we need
p ≤ r and q = r:

kT (f )kLr (Y ;F) = |||T (f )|||Lr,r (Y ;F) ≤ C |||f |||Lp,r (X;F) ≤ C 0 |||f |||Lp,p (X;F) = C 0 kf kLp (X;F) (1.2.32)

for every f ∈ U ∩ Lp (X; V ), where C and C 0 are some constants depending on the parameters. In
order to guarantee that p ≤ r we typically assume that pi ≤ ri for i = 0, 1, which then implies that
p ≤ r. This places a fairly serious restriction on the range of validity of the Marcinkiewicz theorem,
though there are many uses of the result anyway.
In the event that Lp,q (X; F) ⊆ U in the Marcinkiewicz theorem, we immediately find that we
can view T : Lp,q (X; F) → Lr,q (Y ; F) as a bounded map in the sense that |||T (f )|||Lr,q ≤ C kf kLp,q .
However, in practice we often want to utilize a space U that is much smaller than Lp,q (X; F) but
possibly dense. In this case we can actually use Marcinkiewicz to extend T to Lp,q (X; F), provided
that T satisfies a slightly more restrictive condition than stated in the theorem. We explore this in
the next example. We will return to more practical applications of the theorem soon.
Example 1.2.8. Assume the hypotheses of Theorem 1.2.4. Suppose that T : U → L0 (Y ; F) is
real-valued, i.e. T (f ) is real a.e. in Y for each f ∈ U . Further suppose that T is a sublinear
operator, i.e. for f, g ∈ U we have that

T (f + g) ≤ T (f ) + T (g) and |T (−f )| = |T (f )| a.e. in Y. (1.2.33)

Consequently, for f, g ∈ U we can estimate

T (f ) ≤ T (f − g) + T (g) ⇒ T (f ) − T (g) ≤ T (f − g) ≤ |T (f − g)| a.e. in Y, (1.2.34)

and so upon reversing the roles of f and g we find that

T (g) − T (f ) ≤ |T (g − f )| ⇒ − |T (f − g)| = − |T (g − f )| ≤ T (f ) − T (g) a.e. in Y. (1.2.35)

33
Combining these, we deduce that

|T (f ) − T (g)| ≤ |T (f − g)| a.e. in Y, (1.2.36)

which implies that

|||T (f ) − T (g)|||Lr,s (Y ) ≤ |||T (f − g)|||Lr,s (Y ) . (1.2.37)
Combining this estimate with the one provided by the Marcinkiewicz interpolation theorem then
shows that
|||T (f ) − T (g)|||Lr,q (Y ) ≤ |||T (f − g)|||Lr,q (Y ) ≤ C |||f − g|||Lp,q (X) . (1.2.38)
Now, if we know that U is dense in Lp,q (X; V ), then we may use the fact that Lr,s (Y ) is quasi-
Banach (proved in Theorem 1.1.42) to uniquely extend T to a sublinear operator T : Lp,q (X; F) →
Lr,q (Y ; F) satisfying
|||T (f )|||Lr,q ≤ C |||f |||Lp,q . (1.2.39)
4

1.2.2 The complex method of Riesz-Thorin

We now aim for the proof of the second interpolation result. It will be tied to the structure of
the complex numbers. We begin by recalling the maximum modulus principle for holomorphic
functions.

Theorem 1.2.9 (Maximum modulus principle). Suppose that ∅ 6= U ⊆ C is open and connected
and that f : U → C is holomorphic in U . If there exists z ∈ U such that |f (z)| = maxU |f |, then f
is constant in U . In particular, if Ū is compact and f extends continuously to Ū then

max |f | = max |f | . (1.2.40)

Ū ∂U

Now we prove a very interesting estimate known as Hadamard’s three lines lemma. It will be
the workhorse for our interpolation result.

Lemma 1.2.10 (Hadamard’s three lines lemma). Let R = {z ∈ C | 0 ≤ Re(z) ≤ 1}, and suppose
that f ∈ Cb0 (R; C) is holomorphic in R◦ . Further suppose that M0 , M1 > 0 are such that

|f (0 + iy)| ≤ M0 and |f (1 + iy)| ≤ M1 for all y ∈ R. (1.2.41)

Then for every x ∈ [0, 1] we have that

|f (x + iy)| ≤ M01−x M1x . (1.2.42)

Proof. Note first that for z = x + iy ∈ R we have that

M01−z M1z = M01−x M1x ≥ min{M0 , M1 } > 0. (1.2.43)

This allows us to define the functions g, gn ∈ Cb0 (R; C) (here 1 ≤ n ∈ N) via

f (z) 2
g(z) = 1−z z
and gn (z) = g(z)e(z −1)/n . (1.2.44)
M0 M1

34
The boundedness of g follows from the boundedness of f and (1.2.43), while the boundedness of gn
follows since for z = x + iy we have that

|f (z)| (x2 −y2 −1)/n kf kC 0 (R)

−y 2 /n
|gn (z)| = 1−x e ≤ b
e . (1.2.45)
M0 M1x min{M0 , M1 }

This estimate also tells us that

|gn (z)| ≤ 1 for z ∈ ∂R (1.2.46)
and that for each n ≥ 1 there exists rn > 0 such that

|Im(z)| ≥ rn ⇒ |gn (z)| ≤ 1. (1.2.47)

Clearly g and gn for n ≥ 1 are holomorphic in R◦ . Fix n ≥ 1. For each m ≥ 1 define the
rectangle Rm = {z ∈ R◦ | |Im(z)| < rn + m}. According to the estimates (1.2.46) and (1.2.47) we
have that for n ∈ N
max |gn | ≤ 1, (1.2.48)
∂Rm

and so the modulus principle guarantees that

max |gn | ≤ 1. (1.2.49)

Rm

Sending m → ∞ then shows that

sup |gn | ≤ 1. (1.2.50)
R

Finally, since gn (z) → g(z) as n → ∞ for each z ∈ R, we deduce that |g(z)| ≤ 1 on R, and hence
that (1.2.42) holds.

We now have all the tools needed to prove the second interpolation result, the Riesz-Thorin
interpolation theorem.

Theorem 1.2.11 (Riesz-Thorin). Let (X, M, µ) and (Y, N, ν) be measure spaces and suppose that
1 ≤ p0 , p1 , q0 , q1 ≤ ∞. If q0 = q1 = ∞ then also suppose that Y is σ−finite. Write U = S(X; C) ∩
Lp0 (X; C) ∩ Lp1 (X; C) and suppose that T : U → L0 (Y ; C) is a linear map and that there exist
M0 , M1 > 0 such that

kT f kLq0 (Y ;C) ≤ M0 kf kLp0 (X;C) and kT f kLq1 (Y ;C) ≤ M1 kf kLp1 (X;C) (1.2.51)

for all f ∈ U . Let θ ∈ [0, 1] and 1 ≤ p, q ≤ ∞ be given by

1 1−θ θ 1 1−θ θ
= + and = + . (1.2.52)
p p0 p1 q q0 q1
Then
kT f kLq (Y ;C) ≤ M01−θ M1θ kf kLp (X;C) for all f ∈ U. (1.2.53)
In particular, the map T extends to a bounded linear map T : Lp (X; C) → Lq (Y ; C), satisfying
(1.2.53).

35
Proof. First note that once (1.2.53) is established, the extension of T to a bounded linear operator
T : Lp (X; C) → Lq (Y ; C) follows immediately from the density of U in Lp (X; C). Thus it suffices
to prove (1.2.53). Throughout the proof we adopt the convention that if c = d = ∞, then c/d = 1.
Let R = {z ∈ C | Re(z) ∈ [0, 1]} ⊂ C. Define the functions P, Q : R → C via

P (z) = (1 − z)p/p0 + zp/p1 (1.2.54)

Q(z) = (1 − z)q 0 /q00 + zq 0 /q10 . (1.2.55)

By construction we have that

0 ≤ min{p/p0 , p/p1 } ≤ Re(P (z)) ≤ max{p/p0 , p/p1 } < ∞

(1.2.56)
0 ≤ min{q 0 /q00 , q 0 /q10 } ≤ Re(Q(z)) ≤ max{q 0 /q00 , q 0 /q10 } < ∞.

Write f ∈ U and g ∈ Sf in (Y ; C) as
J
X K
X
f= rj vj χEj and g = sk wk χFk (1.2.57)
j=1 k=1

for {Ej }Jj=1 ⊆ M and {Fk }Kk=1 ⊆ N pairwise disjoint and rj , sk ∈ (0, ∞), and |vj | = |wk | = 1 for
1 ≤ j ≤ J and 1 ≤ k ≤ K. Since g is a finite simple function we know that ν(Fk ) < ∞ for
1 ≤ k ≤ K.
Note that for each j = 1, . . . , J we have that vj χEj ∈ U , so

T (vj χEj ) ∈ Lq0 (Y ; C) ∩ Lq1 (Y ; C) ⊆ Lq (Y ; C) (1.2.58)

and hence that T f ∈ Lq (Y ; C). Moreover, since ν(Fk ) < ∞, Hölder’s inequality implies that
T (vj χEj ) ∈ L1 (Fk ; C), and hence
Z
T (vj χEj )dν ∈ C for 1 ≤ j ≤ J, 1 ≤ k ≤ K. (1.2.59)
Fk

For z ∈ R define
J
X K
X
P (z) Q(z)
fz = rj vj χEj ∈ S(X; C) and gz = sk wk χFk ∈ Sf in (Y ; C). (1.2.60)
j=1 k=1

We then define the function F ∈ Cb0 (R; C) via

Z J X
X K Z 
P (z) Q(z)
F (z) = gz T fz dν = rj sk wk T (vj χEj )dν ∈ C. (1.2.61)
Y j=1 k=1 Fk

Note that the boundedness of F on R follows from the fact that Re(P (z)) and Re(Q(z)) satisfy
(1.2.56). Furthermore, for each j, k the maps
P (z) Q(z)
R◦ 3 z 7→ rj ∈ C and R◦ 3 z 7→ sk ∈C (1.2.62)

are holomorphic, and thus F is holomorphic on R◦ .

36
 Suppose that z ∈ R is such that Re(z) = 0. In this case, if p0 , q00 < ∞ then
p P (z) q 0 Q(z) 0
rj 0 = rjp and sk0 = sqk , (1.2.63)

which implies that

p/p
0 q 0 /q 0
kfz kLp0 (X;C) = kf kLp (X;C) and kgz kLq00 (Y ;C) = kgkLq0 (Y
0
;C)
. (1.2.64)

We leave it as an exercise to verify that (1.2.64) also holds if either p0 = ∞ or q00 = ∞.

Now suppose that z ∈ R is such that Re(z) = 1. In this case, if p1 , q10 < ∞ then
p P (z) q 0 Q(z) 0
rj 1 = rjp and sk1 = sqk , (1.2.65)

and so
p/p
1 q 0 /q 0
kfz kLp1 (X;C) = kf kLp (X;C) and kgz kLq10 (Y ;C) = kgkLq0 (Y
1
;C)
. (1.2.66)
We again leave it as an exercise to verify that (1.2.66) also holds if either p1 = ∞ or q10 = ∞.
Now, if z ∈ R is such that Re(z) = 0 then (1.2.51), (1.2.64), and Hölder’s inequality allow us to
estimate

|F (z)| ≤ kT fz kLq0 (Y ;C) kgz kLq00 (Y ;C) ≤ M0 kfz kLp0 (X;V ) kgz kLq00 (Y ;C)
0 p/p q 0 /q 0
= M0 kf kLp (X;C) kgkLq0 (Y
0
;C)
. (1.2.67)

Similarly, if z ∈ R is such that Re(z) = 1 then (1.2.51), (1.2.66) and Hölder’s inequality imply that

|F (z)| ≤ kT fz kLq1 (Y ;C) kgz kLq10 (Y ;C) ≤ M1 kfz kLp1 (X;C) kgz kLq10 (Y ;C)
1p/p q 0 /q 0
= M1 kf kLp (X;V ) kgkLq0 (Y ;C) . (1.2.68)
1

Lemma 1.2.10 and (1.2.52) then imply that

1−θ  θ
q 0 /q00 q 0 /q10

p/p0 p/p1
|F (θ)| ≤ M0 kf kLp (X;C) kgkLq0 (Y ;C)
M1 kf kLp (X;C) kgk 0
Lq (Y ;C)

= M01−θ M1θ kf kLp (X;C) kgkLq0 (Y ;C) . (1.2.69)

Since P (θ) = Q(θ) = 1 by construction, we then have that fθ = f , gθ = g, and

Z
gT f dν = |F (θ)| ≤ M01−θ M1θ kf kLp (X;C) kgkLq0 (Y ;C) . (1.2.70)
Y

Since we know from above that T f ∈ Lq (Y ; C), we may then employ a standard result in Lebesgue
theory to bound
Z
kT f kLq (Y ;C) = sup{ gT f dν | g ∈ Sf in (Y ; C) and kgkLq0 (Y ;C) ≤ 1}
Y
≤ M01−θ M1θ kf kLp (X;C) , (1.2.71)

which is (1.2.53).

37
1.2.3 Comparing and contrasting the two interpolation theorems
It is worth pointing out the differences between the Marcinkiewicz and Riesz-Thorin theorems.
There is a trade-off between the two results. Marcinkiewicz let’s us have a more general type
of operator (not necessarily linear, but satisfying (1.2.7)) and works in more general spaces (the
Lorentz scale) with weak estimates (bounds in Lri ,∞ ), but we pay a constant that depends on
the p, q, r, s parameters and blows up as we approach the endpoints. For Lebesgue spaces it is
also subject to the restrictions discussed in Remark 1.2.7. Riesz-Thorin requires a more restrictive
operator (linear), and works only for a stricter class of spaces (stronger estimates required) over the
complex field, but it gives better interpolated bounds (the constants don’t blow up), and it is not
subject to the restrictions of Remark 1.2.7. Note that neither theorem is stronger than the other:
they are simply different.

1.3 Applications
Our goal now is to demonstrate various uses of the Marcinkiewicz and Riesz-Thorin theorems.

1.3.1 The Hardy-Littlewood maximal function

Let f : Rn → F be locally integrable. Then the Hardy-Littlewood maximal function M(f ) : Rn →
[0, ∞] is defined via Z
1
M(f )(x) = sup n
|f | . (1.3.1)
r>0 ωn r B(x,r)

Note that for a fixed x ∈ Rn the continuity of the map R+ 3 r 7→ ωn1rn B(x,r) |f | ∈ R shows that
R

the supremum over r > 0 can be replaced by the supremum over r ∈ Q ∩ R+ . The latter set is
countable, so M(f ) is measurable. It is then a trivial matter to verify that the Hardy-Littlewood
maximal function, viewed as a map M : L1loc (Rn ; F) → L0 (Rn ; F), is sublinear (see Example 1.2.8).
Let f 6= 0 be locally integrable. We may trivially bound

kM(f )kL∞ ≤ kf kL∞ . (1.3.2)

On the other hand, since f 6= 0 we can pick δ > 0 and 0 < R < ∞ such that
Z
|f | ≥ δ. (1.3.3)
B(0,R)

Then for x ∈ Rn with |x| > R have that B(0, R) ⊆ B(x, R + |x|) and 1/(2 |x|) ≤ 1/(|x| + R), so we
can bound
Z Z
1 1 δ
M(f )(x) ≥ |f | ≥ |f | ≥ n . (1.3.4)
n
ωn (|x| + R) B(x,|x|+R) n
ωn (|x| + R) B(0,R) 2 ωn |x|n

Hence, Z Z
δ dx
M(f )(x)dx ≥ n = ∞, (1.3.5)
B(0,R)c 2 ωn B(0,R)c |x|n
and we deduce from this that M(f ) is never integrable when f is nontrivial and locally integrable.
In spite of this failure, we can get a weaker estimate that will be sufficient for using Marcinkiewicz.

38
Theorem 1.3.1. Let 1 < p < ∞ and 1 ≤ q ≤ ∞. The Hardy-Littlewood maximal operator extends
uniquely to a bounded sublinear operator M : Lp,q (Rn ; F) → Lp,q (Rn ; F), i.e. there exists a constant
C = C(p, q) > 0 such that

|||M(f )|||Lp,q ≤ C |||f |||Lp,q for all f ∈ Lp,q (Rn ; F). (1.3.6)

In particular, by taking q = p we have that M : Lp (Rn ; F) → Lp (Rn ; F) is a bounded sublinear

operator.
Proof. Let f ∈ L1 (Rn ; F). If x ∈ Rn , t > 0, and M(f )(x) > t then there exists rx > 0 such that
Z
|f | > tωn rxn , (1.3.7)
B(x,rx )

and hence [
Et := {x ∈ Rn | M(f )(x) > t} ⊆ B(x, rx ). (1.3.8)
x∈Et

Vitali’s lemma allows us to extract a countable subcollection {B(xi , ri )}i that is pairwise disjoint
and satisfies the bound X
Ln (B(xi , ri )) ≥ CLn (Et ) (1.3.9)
i

for a constant C = C(n) depending on the dimension n. Then

Z Z
n 1 X n 1 X 1 1
dM(f ) (t) = L (Et ) ≤ ωn ri ≤ |f | ≤ |f | = kf kL1 . (1.3.10)
C i Ct i B(xi ,ri ) Ct Rn Ct

Consequently,
|||M(f )|||L1,∞ ≤ C1 kf kL1 (1.3.11)
for some constant C1 > 0.
We now know that M maps L∞ (Rn ; F) to L∞ (Rn ; F) and L1 (Rn ; F) to L1,∞ (Rn ; F) in a bounded
way. The Marcinkiewicz interpolation theorem and Example 1.2.8 then tell us that for each 1 <
p < ∞ and 1 ≤ q ≤ ∞ there exists C > 0 such that

|||M(f )|||Lp,q ≤ C |||f |||Lp,q (1.3.12)

for all f ∈ Lp,q (Rn ; F). In particular, choosing q = p shows that

kM(f )kLp ≤ C kf kLp for all 1 < p < ∞. (1.3.13)

1.3.2 The Hausdorff-Young inequality and its variants

The Fourier transform is the map ˆ· : L1 (Rn ; C) → L∞ (Rn ; C) defined via
Z
ˆ
f (ξ) = f (x)e−2πix·ξ dx. (1.3.14)
Rn

It’s a simple matter to verify that ˆ· is a bounded linear map,

fˆ ≤ kf kL1 , (1.3.15)
L∞

39
and that fˆ is uniformly continuous. Actually, the Riemann-Lebesgue lemma (a proof of which can
be found in any book on Fourier analysis) shows that

lim fˆ(ξ) = 0 (1.3.16)

|ξ|→∞

so ˆ· actually maps into the space of uniformly continuous functions that decay to zero at infinity.
Remarkably, for f, g ∈ L1 (Rn ; C) ∩ L2 (Rn ; C) we have that
Z Z
¯
f ḡ = fˆĝ¯ and kf kL2 = fˆ , (1.3.17)
Rn Rn L2

which allows us to uniquely extend ˆ· to a unitary bijection ˆ· : L2 (Rn ; C) → L2 (Rn ; C) satisfying the
above for all f, g ∈ L2 (Rn ; C). Again, a proof can be found in any book on Fourier analysis.
We will take these facts as given and use them as inputs in interpolation theory. The classic
Hausdorff-Young inequality applies Riesz-Thorin to these estimates to deduce further boundedness
properties of the Fourier transform.

Theorem 1.3.2 (Hausdorff-Young). Let 1 ≤ p ≤ 2. Then the Fourier transform extends to a

0
bounded linear map ˆ· : Lp (Rn ; C) → Lp (Rn ; C). In particular,

fˆ ≤ kf kLp for all f ∈ Lp (Rn ; C). (1.3.18)

Lp0

Proof. The above bounds tell us that

fˆ ≤ kf kL1 for all f ∈ L1 (Rn ; C) (1.3.19)

L∞

and
fˆ = kf kL2 for all f ∈ L2 (Rn ; C). (1.3.20)
L2
The result then follows from these bounds and the Riesz-Thorin interpolation theorem after we note
that if
1 1−θ θ 1 1−θ θ
= + and = + (1.3.21)
p 2 1 q 2 ∞
then
1 1
+ =1 (1.3.22)
q p
and hence q = p0 .
Note that we also could have used Marcinkiewicz to deduce the boundedness of ˆ· from Lp (Rn ; C)
0
to Lp (Rn ; C) since p ≤ p0 for 1 ≤ p ≤ 2. However, this would result in a worse estimate for the norm
than that provided by Riesz-Thorin. We can still use Marcinkiewicz to learn something, though.

Theorem 1.3.3 (Hausdorff-Young, Lorentz variant). Let 1 < p < 2 and 1 ≤ q ≤ ∞. Then the
0
Fourier transform extends to a bounded linear operator ˆ· : Lp,q (Rn ; C) → Lp ,q (Rn ; C).

Proof. Exercise.
Remarkably, this isn’t the end of the story! We can derive another variant that employs certain
weights on the Fourier side.

40
Theorem 1.3.4 (Hausdorff-Young, weighted variant). Let 1 < p < 2 and p ≤ q ≤ p0 . Then there
exists a constant C > 0 such that
Z !1/q
q dξ
fˆ(ξ) 0 ≤ C kf kLp (1.3.23)
Rn |ξ|n(1−q/p )
for every f ∈ Lp (Rn ; C).
Proof. Define the measure µ on Rn via µ = dξ/ |ξ|2n . We will write Lpµ (Rn ; C) and Lp,q n
µ (R ; C) for
the Lebesgue and Lorentz spaces relative to the measure µ, and when we omit the µ we mean the
usual spaces relative to Lebesgue measure.
Define the map T : L1 (Rn ; C) + L2 (Rn ; C) → L0µ (Rn ; C) via

T f (ξ) = |ξ|n fˆ(ξ). (1.3.24)

Using (1.3.17), we may bound
Z 1/2
2 dξ
kT f kL2µ = 2n
|ξ| fˆ(ξ) = fˆ = kf kL2 . (1.3.25)
Rn |ξ|2n L2

Next, we claim that

|||T f |||L1,∞
µ
≤ ωn kf kL1 , (1.3.26)
where ωn is the volume of the unit ball in Rn (see the Section 0.2). To see this first note that if
ξ ∈ Rn is such that t < |T f (ξ)|, then (1.3.15) implies that
 1/n
n ˆ n t
t < |ξ| f (ξ) ≤ |ξ| kf kL1 ⇒ < |ξ| , (1.3.27)
kf kL1
and hence
1/n
{ξ ∈ Rn | |T f (ξ)| > t} ⊂ B(0, t1/n / kf kL1 )c . (1.3.28)
Thus, for t ∈ R+ we may bound
∞
rn−1 dr αn kf kL1
Z Z
n dξ
µ({ξ ∈ R | |T f (ξ)| > t}) ≤ = αn = , (1.3.29)
B(0,t1/n /kf k
1/n c
L1
) |ξ|2n t1/n /kf k
1/n
L1
r2n n t

which implies (1.3.26) since ωn = αn /n.

With (1.3.25) and (1.3.26) in hand, we may apply the Marcinkiewicz interpolation theorem to
deduce that for each 1 < p ≤ 2 there exists C > 0 such that
Z !1/p
p dξ
fˆ(ξ) = kT f kLpµ ≤ C kf kLp (1.3.30)
Rn |ξ|n(2−p)
for all f ∈ Lp (Rn ; C). When q = p this is (1.3.23) since
p
1− = 2 − p, (1.3.31)
p0
as is readily verified. In the general case of p ≤ q ≤ p0 we define θ ∈ [0, 1] via q = (1 − θ)p + θp0 , or
equivalently
q−p p0 − q 1 − q/p0 1 − q/p0
θ= 0 and 1 − θ = 0 = = . (1.3.32)
p −p p −p 1 − p/p0 2−p

41
Then from (1.3.30), the basic Hausdorff-Young inequality from Theorem 1.3.2, and Hölder’s in-
equality we may bound
 p 1−θ
Z q dξ
Z fˆ(ξ) 
p0
θ
fˆ(ξ) 0 =   fˆ(ξ) dξ
Rn |ξ|n(1−q/p ) Rn |ξ|n(2−p)
!1−θ Z θ
p0
Z p dξ (1−θ)p 0
≤ fˆ(ξ) fˆ(ξ) dξ ≤ C kf kLp kf kθp q
Lp = C kf kLp , (1.3.33)
Rn |ξ|n(2−p) Rn

and we deduce that !1/q

Z q dξ
fˆ(ξ) 0 ≤ C kf kLp . (1.3.34)
Rn |ξ|n(1−q/p )
This is (1.3.23).

1.3.3 Integral operators

Let (X, M, µ) and (Y, N , ν) be measure spaces and U ⊆ L0 (Y ; F) be a subspace. We now turn our
attention to maps T : U → L0 (X; F) of the form
Z
T f (x) = K(x, y)f (y)dν(y), (1.3.35)
Y

where K : X × Y → F is a given measurable “kernel” function. Of course, we need to worry about

hypotheses on K and f that will guarantee that this integral makes sense, and we will identify
sufficient conditions for this. When the operator is well-defined it is obviously linear, and such
linear maps are referred to as integral operators.
Our first result on integral operators assumes that the kernel satisfies some uniform integrability
conditions.
Theorem 1.3.5. Let (X, M, µ) and (Y, N, ν) be σ−finite measure spaces, and let 1 ≤ p ≤ ∞. Let
K : X × Y → F be measurable with respect to M ⊗ N and suppose there exists a constant A ≥ 0
such that
Z
|K(x, y)| dµ(x) ≤ A for ν − a.e. y ∈ Y
X
Z (1.3.36)
|K(x, y)| dν(y) ≤ A for µ − a.e. x ∈ X.
Y

Then the following hold.

R
1. If f ∈ Lp (Y ; F) then for µ−a.e. x ∈ X the integral Y K(x, y)f (y)dν(y) is well-defined, and
the function T f : X → F defined a.e. via
Z
T f (x) = K(x, y)f (y)dν(y) (1.3.37)
Y

is such that T f ∈ Lp (X; F) and

kT f kLp (X;F) ≤ A kf kLp (Y ;F) . (1.3.38)

42
 2. The induced map T : Lp (Y ; F) → Lp (X; F) is linear and bounded.
Proof. First note that it suffices to prove the first item, as the second follows immediately from the
first. Second, note that we may reduce to proving the result with F = C since R ⊆ C and T f is
clearly real-valued when K and f are. If p = 1 or ∞, then the first item follows from Fubini-Tonelli
or Hölder, respectively. Then the Riesz-Thorin interpolation theorem implies that the first item
continues to hold for general 1 < p < ∞.
Remark 1.3.6. We also could have applied Marcinkiewicz to prove the theorem, but the constant
would have been worse.
We can parlay the result of Theorem 1.3.5 into a proof of Young’s inequality.
Theorem 1.3.7 (Young’s inequality). Let 1 ≤ p, q, r ≤ ∞ satisfy
1 1 1
+ =1+ . (1.3.39)
p q r
0
Then for each f ∈ Lp (Rn ; F) the linear map Γf : L1 (Rn ; F) ∩ Lp (Rn ; F) ∩ S(Rn ; F) → L0 (Rn ; F)
given by Z
Γf g(x) = f (y)g(x − y)dy (1.3.40)
Rn
is well-defined and extends uniquely to a bounded linear map Γf : Lq (Rn ; F) → Lr (Rn ; F), and
kΓf gkLr ≤ kf kLp kgkLq for all g ∈ Lq (Rn ; F). (1.3.41)
Proof. First note that the conditions on p, q, r imply that
1 1 1 1
= 0 + ≥ ⇒ p ≤ r ≤ ∞. (1.3.42)
p q r r
If we define K : Rn × Rn → F via K(x, y) = g(x − y), then we may use Theorem 1.3.5 with
A = kgkL1 to see that
kΓf gkLp ≤ kgkL1 kf kLp , (1.3.43)
so Γf g is well-defined on L1 (Rn ; F). If r = p, then q = 1 and the result follows from this and a
standard density argument.
On the other hand, Hölder’s inequality shows that
0
kΓf gkL∞ ≤ kgkLp0 kf kLp for all g ∈ Lp (Rn ; F). (1.3.44)
If r = ∞, then q = p0 and again the result follows from this and a standard density argument.
Suppose, then, that p < r < ∞ in which case we can define θ ∈ (0, 1) via
1 1−θ θ 1−θ
= + = . (1.3.45)
r p ∞ p
This implies that
1−θ θ θ θ 1 1 1
+ 0 =1−θ+θ− =1− =1+ − = . (1.3.46)
1 p p p r p q
We may then apply Riesz-Thorin to deduce that the map Γf extends to a bounded linear operator
Γf : Lq (Rn ; F) → Lr (Rn ; F) with
kΓf gkLr ≤ kf kLp kgkLq , (1.3.47)
and the result is proved.

43
 This suggests a definition.

Definition 1.3.8. Let 1 ≤ p, q, r ≤ ∞ satisfy

1 1 1
+ =1+ . (1.3.48)
p q r

We define the bilinear map ∗ : Lp (Rn ; F) × Lq (Rn ; F) → Lr (Rn ; F) via

Z Z
f ∗ g(x) = f (y)g(x − y)dy = f (x − y)g(y)dy. (1.3.49)
Rn Rn

By Young’s inequality this is well-defined and bounded:

kf ∗ gkLr ≤ kf kLp kgkLq . (1.3.50)

Remark 1.3.9. If we replace Rn with Zn and Lebesgue measure by counting measure in Young’s
inequality, then the argument pushes through without modification and shows that if we define
X
f ∗ g(m) = f (k)g(m − k), (1.3.51)
k∈Z

then f ∗ g ∈ `r (Zn ; F) when f ∈ `p (Zn ; F) and g ∈ `q (Zn ; F) with (1.3.39) satisfied. Moreover,
kf ∗ gk`r ≤ kf k`p kgk`r .

In the above we have exploited Riesz-Thorin in order to get bounds on integral operators.
It turns out that we can also use the Marcinkiewicz theorem, and this allows us to put weaker
assumptions on the kernel. To prove this we fist need a lemma.

Lemma 1.3.10. Let (X, M, µ) be a measure space. Let f : X → F be measurable, τ ∈ R+ , and set

E(τ ) = {x ∈ X | |f (x)| > τ }. (1.3.52)

Define hτ , gτ : X → V via
f f
hτ = f χE(τ )c + τ χE(τ ) and gτ = f − hτ = (|f | − τ )χE(τ ) . (1.3.53)
|f | |f |

Then (
df (t) if t < τ
dgτ (t) = df (t + τ ) and dhτ (t) = (1.3.54)
0 if τ ≤ t.

Proof. First note that |hτ (x)| ≤ τ for all x, so dhτ (t) = 0 for t ≥ τ . For t < τ we then compute

dhτ (t) = µ({x ∈ E(τ )c | |f (x)| > t}) + µ(E(τ )) = µ({x ∈ X | |f (x)| > t}) = df (t), (1.3.55)

which proves the equality for hτ . On the other hand,

f (x)
dgτ (t) = µ({x ∈ E(τ ) | (|f (x)| − τ ) > t}) = µ({x ∈ X | |f (x)| − τ > t}) = df (t + τ ).
|f (x)|
(1.3.56)

44
 With the lemma in hand, we can now prove a variant of Theorem 1.3.5.
Theorem 1.3.11. Let (X, M, µ) and (Y, N, ν) be σ−finite measure spaces, and let 1 < p, q, r < ∞
be such that
1 1 1
+ =1+ . (1.3.57)
p q r
Let K : X × Y → F be measurable with respect to M ⊗ N, and suppose there exists a constant A ≥ 0
such that
|||K(·, y)|||Lq,∞ (X;F) ≤ A for ν − a.e. y ∈ Y
(1.3.58)
|||K(x, ·)|||Lq,∞ (Y ;F) ≤ A for µ − a.e. x ∈ X.

Then the following hold.

R
1. If f ∈ Lp (Y ; F) then for µ−a.e. x ∈ X the integral Y K(x, y)f (y)dν(y) is well-defined, and
the function T f : X → F defined a.e. via
Z
T f (x) = K(x, y)f (y)dν(y) (1.3.59)
Y
p
is such that T f ∈ L (X; F).
2. The induced map T : Lp (Y ; F) → Lr (X; F) is linear and bounded, and there exists a constant
C > 0 depending only on p, q, r, such that

kT f kLr (X;F) ≤ CA kf kLp (Y ;F) for all f ∈ Lp (Y ; F) (1.3.60)

Proof. First note that

1 1 1 1 1 1 1
= + − 1 = − 0 = − 0, (1.3.61)
r p q p q q p
and this tells us that p < q 0 and q < p0 . Next note that the result is trivial if A = 0, so we can
assume that A > 0.
Consider 0 6= f ∈ Lp (Y ; F). Assume for now that kf kLp = 1 and A ≤ 1. Let τ > 0 and define

E(τ ) = {(x, y) ∈ X × Y | |K(x, y)| > τ }, (1.3.62)

and
K
K1 = (|K| − τ )χE(τ ) and K2 = K − K1 . (1.3.63)
|K|
Define the operators T1 and T2 with K replaced by K1 and K2 , respectively.
By the Lemma 1.3.10 we may compute
Z ∞ Z ∞
τ 1−q
Z
|K1 (x, y)| dν(y) = dK(x,·) (t + τ )dt ≤ t−q dt = , (1.3.64)
Y 0 τ q−1
and similarly
τ 1−q
Z
|K1 (x, y)| dµ(x) ≤ . (1.3.65)
X q−1
Thus Theorem 1.3.5 tells us that T1 f is defined a.e. and
τ 1−q τ 1−q
kT1 f kLp ≤ kf kLp = . (1.3.66)
q−1 q−1

45
 On the other hand, since q < p0 we may compute
Z τ τ 0
p0 τ p −q
Z Z
p0 0 p0 −1 0 0
|K2 (x, y)| dν(y) = p t dK(x,·) (t)dt ≤ p tp −1−q dt = . (1.3.67)
Y 0 0 p0 − q
This bound and Hölder’s inequality imply that T2 f is defined for a.e. x ∈ X and that
0 1/p0  1/p0
p0 τ p −q

r
kT2 f kL∞ ≤ kf kLp = τ q/r . (1.3.68)
p0 − q q

These two arguments combine to show that T f = T1 f + T2 f is well-defined a.e. in X.

Due to the decomposition T f = T1 f + T2 f we can bound, for any t ∈ R+ ,

dT f (t) ≤ dT1 f (t/2) + dT2 f (t/2), (1.3.69)

and so if we choose  r/q  

t q r/(qp0 )
τ= (1.3.70)
2 r
then (1.3.68) shows that kT2 f kL∞ ≤ t/2 and hence dT2 f (t/2) = 0. With this choice made we then
use Chebyshev to bound
p p
2τ 1−q
 
2 kT1 f kLp
dT f (t) ≤ dT1 f (t/2) ≤ ≤
t (q − 1)t
2p−[(1−q)pr]/q  q [(1−q)(pr)]/(qp0 ) −p+[(1−q)pr]/q Cp
= t = r (1.3.71)
(q − 1)p r t

for some constant C = C(p, q, r) > 0, due to the algebraic calculation

 
(1 − q)pr r r
− p = p − 0 − 1 = −p = −r. (1.3.72)
q q p

The bound
1
dT f (t) ≤ C (1.3.73)
tr
was proved under the assumption that kf kLp = 1, but for general f 6= 0 we can apply this estimate
to f / kf kLp to deduce that  r
kf kLp
dT f (t) ≤ C . (1.3.74)
t
This yields the bound
|||T f |||Lr,∞ ≤ C kf kLp (1.3.75)
for all p, r satisfying the relation to q in the hypotheses.
Since 1 < p < ∞ we may choose 1 < p0 < p < p1 < ∞ and define 1 < r0 < r < r1 < ∞ via
1 1 1 1 1 1
+ =1+ and + =1+ . (1.3.76)
q p0 r0 q p1 r1
Define θ ∈ (0, 1) via
1 1−θ θ
= + (1.3.77)
p p0 p1

46
and note that
 
1−θ θ 1 1−θ θ 1 1 1
+ = (1 − θ + θ) −1 + + = + −1= . (1.3.78)
r0 r1 q p0 p1 q p r
Further note that pi ≤ ri for i = 1, 2 by construction, so p ≤ r.
The estimates above tell us that

|||T f |||Lri ,∞ (X;F) ≤ Ci kf kLpi (Y ;F) for i = 1, 2 and f ∈ Lpi (Y ; F) (1.3.79)

for constants Ci > 0. Then according to this and the Marcinkiewicz interpolation theorem there
exists a constant C > 0 such that

kT f kLr (X;F) ≤ C kf kLp (X;F) (1.3.80)

for all f ∈ Lp (Y ; F). Finally, this estimate was derived under the auxiliary assumption that A ≤ 1,
but the general case follows from this special case applied to K/A.
As an immediate consequence we get a very useful variant of Young’s inequality.

Theorem 1.3.12 (Young’s inequality, weak-strong form). Let 1 < p, q, r < ∞ satisfy
1 1 1
+ =1+ . (1.3.81)
p q r
If f ∈ Lp (Rn ; F) and g ∈ Lq,∞ (Rn ; F), then the function f ∗ g : Rn → F given by
Z
f ∗ g(x) = f (y)g(x − y)dy (1.3.82)
Rn

is well-defined a.e., measurable, and belongs to Lr (Rn ; F). Moreover, there exists a constant C =
C(p, q, r) > 0 such that
kf ∗ gkLq ≤ C kf kLp |||g|||Lq,∞ (1.3.83)
for every f ∈ Lp (Rn ; F) and g ∈ Lq,∞ (Rn ; F).

Proof. This follows from Theorem 1.3.11 by setting K(x, y) = g(x − y) and A = |||g|||Lq,∞ .

Remark 1.3.13. This theorem can again be extended to functions defined on the integer lattice Zn .
We will not need this, though, so we don’t record it in a precise form.

As a brief glimpse of the power of this generalized version of Young’s inequality we present the
following result on the Riesz potentials.

Theorem 1.3.14 (Hardy-Littlewood-Sobolev inequality). Let α ∈ R satisfy 0 < α < n. For

f ∈ Sf in (Rn ; F) define the Riesz potential of f to be the measurable function Iα f : Rn → F given by
Z
f (y)
Iα f (x) = α dy. (1.3.84)
Rn |x − y|

Then Iα extends to a bounded linear operator Iα : Lp (Rn ; F) → Lr (Rn ; F) for 1 < p < n/(n − α)
and
1 1 n−α
= − . (1.3.85)
r p n

47
Proof. Note that
1 1
t< α ⇔ |x| < 1/α (1.3.86)
|x| t
and so if we write g(x) = 1/ |x|α then
tn/α dg (t) = ωn (1.3.87)
and we deduce that g ∈ Ln/α,∞ . We can then apply Theorem 1.3.12 to deduce that Iα ∈
L(Lp (Rn ; F); Lr (Rn ; F)) whenever
1 1 n−α np
= − ⇔r= . (1.3.88)
r p n n − p(n − α)

These bounds for the Riesz potential cannot be improved. Indeed, we have the following example.
Example 1.3.15. Suppose that there exists C > 0 such that

kIα f kLr ≤ C kf kLp for all f ∈ Lp . (1.3.89)

Pick f ∈ Lp (Rn )\{0} such that f ≥ 0 a.e., which implies that Iα f 6= 0. For λ > 0 consider fλ ∈ Lp
given by fλ (x) = f (λx). We then compute

kfλ kLp = λ−n/p kf kLp . (1.3.90)

Also Z Z
f (λy) f (λy)
Iα fλ (x) = dy = λα dy = λα−n Iα f (λx), (1.3.91)
Rn |x − y|α Rn |λx − λy|α
and so
kIα fλ kLr = λα−n−n/r kIα f kLr . (1.3.92)
Thus
λα−n−n/r kIα f kLr ≤ λ−n/p C kf kLp . (1.3.93)
Since both of the norms are nonzero and we are free to send λ → 0 and λ → ∞ in this inequality,
we see that we can avoid a contradiction if and only if
n n 1 1 n−α np
− =− +α−n⇔ = − ⇔r= . (1.3.94)
p r r p n n − p(n − α)
This scaling argument shows that the relationship between p and r from the theorem is a necessary
condition. Note that we need 1 ≤ p ≤ n/(n − α) for this to make sense at all. The scaling argument
does not address what happens at the endpoint cases, p = 1 and p = n/(n − α), though. We will
show that the Riesz potentials are unbounded in these cases.
Let p = 1, which means that r = n/α. Consider f ∈ L1 (Rn ; F)\{0} real valued such that f ≥ 0
a.e. Pick R > 0 and δ > 0 such that Z
|f | ≥ δ. (1.3.95)
B(0,R)

Then for |x| ≥ R and |y| ≤ R we have that |x − y| ≤ |x| + |y| ≤ |x| + R ≤ 2 |x|. Thus, for |x| ≥ R
we can bound Z Z
f (y) f (y) δ
Iα f (x) ≥ α dy ≥ α dy ≥ α , (1.3.96)
B(0,R) |x − y|
α
B(0,R) 2 |x| 2 |x|α

48
and so Z  n/α Z
n/α δ dx
|Iα f (x)| dx ≥ = ∞, (1.3.97)
B(0,R)c 2α B(0,R)c |x|n
which in particular means that Iα f ∈/ Lr (Rn ; F).
Let p = n/(n − α), which means that r = ∞. Fix 1 < β < p and define f : Rn → R via
(
|x|−n/p |log |x||−β/p for 0 < |x| < 1/e
f (x) = (1.3.98)
0 otherwise.

Then since 1 < β,

Z 1/e Z ∞
dr ds
kf kpLp = αn = < ∞, (1.3.99)
0 r(log(1/r))β e s(log s)β
so f ∈ Lp (Rn ; F). However, if 0 < δ < 1/e, δ < |y| < 1/e, and |x| < δ, then |x − y| ≤ |x| + |y| <
δ + |y| < 2 |y|, and so
Z Z
f (y) 1 f (y)
Iα f (x) ≥ α dy ≥ α α dy
{δ<|y|<1/e} |x − y| 2 {δ<|y|<1/e} |y|

αn 1/e rn−1 dr αn 1/δ

Z Z Z
1 dy ds
= α α+n/p
= α n β/p
= α .
2 {δ<|y|<1/e} |y| [log(1/ |y|)]β/p 2 δ r [log(1/r)] 2 e s[log s]β/p
(1.3.100)

Since β/p < 1, Z ∞

ds
= ∞, (1.3.101)
e s[log s]β/p
/ L∞ (Rn ; F).
and we conclude that Iα f ∈
4

2 Abstract interpolation theory

We have now seen that the Lebesgue and Lorentz spaces admit a rich interpolation theory that is
extremely useful in the sense that it allows us to gain information about the mapping properties of
certain types of operators by first understanding how the operator acts on two fixed pairs of such
spaces. Our goal now is to generalize these techniques to the context of more abstract spaces, and
to thereby distill the essential ideas at the core of what we’ve done so far. In turn, we can hope
that this will provide a clearer perspective on the Lebesgue and Lorentz interpolation as well as
equip us with tools analogous to the Marcinkiewicz and Riesz-Thorin theorems that we can use in
a more general context.
A serious problem with this program immediately presents itself: in the context of Lebesgue and
Lorentz spaces we naturally have a continuum of parameters with which can interpolate, namely the
values 1 ≤ p, q ≤ ∞, but in a more general context it’s not obvious where such parameters will come
from. This means that our starting point will not be the development of the interpolation theory for
operators but rather the construction of the interpolation spaces themselves. Once we have these in
hand, we can then turn to the question of building the corresponding operator interpolation theory.
Fortunately, these spaces can be constructed, and they do admit a beautiful and useful theory
of operator interpolation. Once we have developed this in this section, we will return to the

49
context of the previous section and demonstrate how what we saw there fits into this framework.
Unfortunately, due to time constraints, we will not be able to fully develop this theory in these
notes. Because of this, we have chosen to focus solely on the so-called real method of interpolation,
which is, roughly-speaking, the abstract generalization of the Marcinkiewicz interpolation theorem.
There is a corresponding abstract generalization of the Riesz-Thorin theorem known as the complex
method that we will completely ignore.

2.1 Compatible pairs of Banach spaces

We begin by laying the groundwork for the construction of the interpolation spaces. The crucial
observation was described in the overview: if (X, M, µ) is a measure space and 1 < p < ∞, then

L1 (X; F) ∩ L∞ (X; F) ,→ Lp (X; F) ,→ L1 (X; F) + L∞ (X; F). (2.1.1)

In fact, it’s possible to prove the same result with Lp (X; F) replaced by Lp,q (X; F) for any 1 ≤ q ≤ ∞.
What this tells us is that the spaces we use for interpolation always nest between the extreme spaces
L1 (X; F) ∩ L∞ (X; F) and L1 (X; F) + L∞ (X; F). This suggests that if we want to extend our theory
we should start by looking at pairs of Banach spaces X0 and X1 for which we can make sense of
X0 ∩ X1 and X0 + X1 . We turn our attention to this now.

2.1.1 Reminders about the basics of Banach spaces

We first record a couple quick reminders about the structure of general Banach spaces. The first
gives a characterization of completeness in terms of infinite sums.

Theorem 2.1.1. Let X be a normed vector space. The following are equivalent.

1. X is Banach.
P∞
2. If {xn }∞ kxn k is convergent in R, then ∞
P
n=` ⊆ X and n=` n=` xn is convergent in X.
P∞
Proof. Suppose first that X is Banach and that n=` kxn k is convergent in R. Let ε > 0. Since R
is complete we may choose N ≥ ` such that m ≥ n ≥ N implies that
m
X m
X
xk ≤ kxk k < ε (2.1.2)
k=n k=n

partial sums { N ∞
P
and hence theP n=` xn }N =` are Cauchy and hence convergent due to the completeness
∞
of X. Hence n=` xn converges in X.
Now suppose that (2) holds and let {xn }∞ n=` ⊆ X be Cauchy. We may extract a subsequence
∞ −k
{xnk }k=` such that xnk+1 − xnk < 2 for all k ≥ `. Set yk = xnk+1 − xnk for k ≥ `. Then
∞
X ∞
X
kyk k ≤ 2−k < ∞, (2.1.3)
k=` k=`
P∞
and so by (2) we have that k=` yk converges in X. However,
K
X ∞
X
yk = xnK+1 − xn` ⇒ lim xnk = xn` + yk ∈ X, (2.1.4)
k→∞
k=` k=`

50
i.e. the subsequence {xnk }∞
k=` converges in X. Basic real analysis tells us that Cauchy sequences
with convergent subsequences must be convergent, so we find that {xn }∞ n=` is convergent. Hence X
is complete.

Our second result records two useful ways to check that we have an embedding X ,→ Y when
X and Y are Banach spaces.
Theorem 2.1.2. Let X and Y be Banach spaces and suppose that X ⊆ Y as a vector subspace.
Then the following are equivalent.
1. X ,→ Y , i.e. the inclusion map I : X → Y is continuous.

2. There exists a constant C > 0 such that kxkY ≤ C kxkX for all x ∈ X.

3. If {xn }∞
n=` ⊆ X, xn → x in X, and xn → y in Y , then x = y.

Proof. The equivalence of the first and second follow from the equivalence of boundedness and
continuity for linear maps, and the equivalence and the second and third follow from the closed
graph theorem.

2.1.2 Compatible Banach spaces and their sums and intersections

If we want to form the intersection X0 ∩ X1 and the sum X0 + X1 for general pairs of spaces X0
and X1 , then we clearly need that the vector space structure of these spaces are compatible in some
way. In fact, it will be crucial to have a sort of topological compatibility condition as well. We
define this notion now.
Definition 2.1.3. Suppose that X0 , X1 are Banach spaces over the field F.
1. We say X0 and X1 are compatible if there exists a Hausdorff topological vector space Z over
F such that Xi ,→ Z is a subspace for i = 0, 1 and the inclusions are continuous.

2. Suppose that X0 , X1 are compatible. We define

X0 ∩ X1 = {z ∈ Z | z ∈ X0 and z ∈ X1 } (2.1.5)

and
X0 + X1 = {z ∈ Z | z = x0 + x1 for x0 ∈ X0 and x1 ∈ X1 }. (2.1.6)
Define k·kX0 ∩X1 : X0 ∩ X1 → R via

kxkX0 ∩X1 = max{kxkX0 , kxkX1 } (2.1.7)

and define k·kX0 +X1 : X0 + X1 → R via

kxkX0 +X1 = inf{kx0 kX0 + kx1 kX1 | x = x0 + x1 for x0 ∈ X0 and x1 ∈ X1 }. (2.1.8)

Note: we will sometimes write k·k0+1 = k·kX0 +X1 and k·k0∩1 = k·kX0 ∩X1 as shorthand.
We can now mimic the usual proof that Lp0 ∩Lp1 and Lp0 +Lp1 are complete in this more general
context.
Theorem 2.1.4. Suppose that X0 and X1 are compatible Banach spaces. Then the following hold.

51
 1. k·kX0 ∩X1 is a norm, and X0 ∩ X1 is a Banach space when equipped with this norm.
2. k·kX0 +X1 is a norm, and X0 + X1 is a Banach space when equipped with this norm.
3. We have the continuous embeddings

X0 ∩ X1 ,→ Xi ,→ X0 + X1 for i = 0, 1. (2.1.9)

Proof. We begin with the proof of the first item. The fact that k·kX0 ∩X1 is a norm follows directly
from fact that k·kXi is a norm for i = 0, 1; we leave it as an exercise to check the details. Consider
a Cauchy sequence {xn }∞ n=` ⊂ X0 ∩ X1 . Then the definition of the norm shows that the sequence
is Cauchy in both X0 and in X1 , and since these are both assumed to be complete, there exist
x ∈ X0 and y ∈ X1 such that xn → x in X0 and xn → y in X1 as n → ∞. Due to the continuous
embeddings Xi ,→ Z for i = 0, 1 we have that xn → x and xn → y in Z, but limits in Z are unique
since Z is Hausdorff, and hence x = y ∈ X0 ∩ X1 . From this we readily deduce that

lim kx − xn kX0 ∩X1 = 0 (2.1.10)

n→∞

and so X0 ∩ X1 is a Banach space.

We now turn to the proof of the second item, which is a bit more involved than the proof of
the first. First we show that k·kX0 +X1 is a norm. Clearly kxkX0 +X1 ≥ 0 and k0kX0 +X1 = 0. On
the other hand, if kxkX0 +X1 = 0 then for every n ∈ N we can find yn ∈ X0 and zn ∈ X1 such that
x = yn + zn and
kyn kX0 + kzn kX1 < 2−n . (2.1.11)
Then yn → 0 in X0 and zn → 0 in X1 and since the inclusions Xi ,→ Z are continuous we have that
x = yn + zn → 0 as n → ∞ in Z, and so x = 0. Homogeneity and the triangle inequality are easy
and left as an exercise to verify. Thus, k·kX0 +X1 is a norm.
∞
P∞We now prove completeness for X0 + X1 . Consider a sequence {xn }n=` ⊆ X0 + X1 such that
n=` kxn kX0 +X1 < ∞. For each n ≥ ` we can pick yn ∈ X0 and zn ∈ X1 such that

kyn kX0 + kzn kX1 < kxn kX0 +X1 + 2−n , (2.1.12)

which in turn means that ∞ ∞

X X
kyn kX0 + kzn kX1 < ∞. (2.1.13)
n=` n=`

Since X0 and X1 are Banach, we can use Theorem 2.1.1 to see that we have the convergence
∞
X ∞
X
yn = y in X0 and zn = z in X1 . (2.1.14)
n=` n=`

Set x = y + z ∈ X0 + X1 and note that for N ≥ ` we have

N N
! N
!
X X X
x− xn = y − yn + z − zn , (2.1.15)
n=` n=` n=`

which implies that

N
X N
X N
X
x− xn ≤ y− yn + z− zn → 0 and N → ∞. (2.1.16)
n=` X0 +X1 n=` X0 n=` X1

52
 P∞
Thus, n=` xn = x ∈ X0 + X1 , where the sum convergences in the X0 + X1 norm, and so by
Theorem 2.1.1 we conclude that X0 + X1 is Banach. This proves the second item.
Finally, we prove the third item. The subspace inclusions X0 ∩ X1 ⊆ Xi ⊆ X0 + X1 are obvious,
so we only need to prove continuity. We have the bound
kxkXi ≤ kxkX0 ∩X1 for every x ∈ X0 ∩ X1 , (2.1.17)
which proves that X0 ∩ X1 ,→ Xi . On the other hand, if x ∈ Xi then x = x + 0 and so
kxkX0 +X1 ≤ kxkXi , (2.1.18)
which shows that Xi ,→ X0 + X1 .

Remark 2.1.5. Notice that in the proof we don’t really exploit the assumption that Z is a topological
vector space in the sense that we never use the continuity of scalar multiplication. This shows that
we could in principle weaken the assumptions on Z in the notion of compatibility and require that Z
is merely an Abelian topological group in which X0 and X1 are subgroups with respect to the additive
group structure of these vector spaces.

2.1.3 Intermediate spaces

We now have the ability to form the Banach spaces X0 ∩ X1 and X0 + X1 when X0 and X1 are
compatible. As the next step we introduce the notion of a space intermediate to these, which
generalizes the relations we saw in (2.1.1).
Definition 2.1.6. Suppose that X0 and X1 are compatible Banach spaces. A Banach space X is
said to be intermediate to X0 and X1 if
X0 ∩ X1 ,→ X ,→ X0 + X1 . (2.1.19)
Let’s consider some examples.
Example 2.1.7. X0 ∩ X1 and X0 + X1 are both trivially intermediate to X0 and X1 . 4
Example 2.1.8. Theorem 2.1.4 shows that both X0 and X1 are intermediate to X0 and X1 . 4
Example 2.1.9. From (2.1.1) we know that Lp (X; F) is intermediate to L1 (X; F) and L∞ (X; F)
when 1 < p < ∞. The same is true of Lp,q (X; F) for 1 < p < ∞ and 1 ≤ q ≤ ∞. 4
Given a pair of intermediate spaces Y0 and Y1 between compatible spaces X0 and X1 we clearly
have that Y0 and Y1 are compatible, and so we can repeat our previous constructions. The next
elementary result shows how the resulting spaces sit inside the existing ones.
Proposition 2.1.10. Let X0 and X1 be compatible Banach spaces, and suppose that Y0 and Y1 are
intermediate to X0 and X1 . Then Y0 ∩ Y1 and Y0 + Y1 are intermediate to X0 and X1 and we have
the following embeddings:

Y0

X0 ∩ X1 Y0 ∩ Y1 Y0 + Y1 X0 + X1 . (2.1.20)

Y1

53
Proof. Exercise.
Given three topological vector spaces that nest via X ,→ Y ,→ Z we have two natural operations
we perform to construct new spaces: we can take the closure of X in Y , and we can take the closure
of Y in Z. We now give this idea a special name in the context of intermediate spaces.
Definition 2.1.11. Suppose that X0 and X1 are compatible Banach spaces and that X is interme-
diate to X0 and X1 . We define the space X as the closure of X0 ∩ X1 in X, and the space X̄ as the
closure of X is X0 + X1 . The space X is called the lower closure and X̄ is called the upper closure.
Clearly, X and X̄ are Banach, and we have the continuous inclusions

X0 ∩ X1 ,→ X ,→ X ,→ X̄ ,→ X0 + X1 , (2.1.21)

which in particular means that X and X̄ are intermediate to X0 and X1 .

At this point the only spaces that we know are intermediate to a compatible pair X0 and X1
are X0 , X1 , X0 ∩ X1 , and X0 + X1 . It turns out that we can characterize the upper and lower
closures of these in terms of sum and intersection operations in a very nice way. Indeed, we have
the following fundamental result.
Theorem 2.1.12. Let X0 and X1 be compatible Banach spaces. Then the following hold.
1. X0 = X0 ∩ X̄1 and X1 = X̄0 ∩ X1 .

2. X̄0 = X0 + X1 and X̄1 = X0 + X1 .

3. X0 ∩ X1 = X0 ∩ X1 = X0 ∩ X1 = X0 ∩ X1 , X0 ∩ X̄1 = X0 , and X̄0 ∩ X1 = X1 .

4. X̄0 + X̄1 = X̄0 + X1 = X0 + X̄1 = X0 + X1 , X0 + X̄1 = X̄1 , and X̄0 + X1 = X̄0 .

5. X0 + X1 = X0 ∩ X1 = X̄0 ∩ X̄1 = X0 + X1 .
Proof. We begin with a bit of notation. If A and B are normed vector spaces and we have the
subspace inclusion A ⊆ B, then we write cl(A, B) ⊆ B for the closure of A in the topology of B.
We leave it as an exercise to verify that if A ,→ B ,→ C for normed vector spaces A, B, and C,
then
cl(A, B) ,→ cl(A, C) and cl(A, C) ,→ cl(B, C). (2.1.22)
From (2.1.22) and the fact that each Xi is intermediate to X0 and X1 we see that

Xi = cl(X0 ∩ X1 , Xi ) ,→ cl(Xi , Xi ) = Xi (2.1.23)

and

Xi = cl(X0 ∩ X1 , Xi ) ,→ cl(X0 ∩ X1 , X0 + X1 ) ,→ cl(X1−i , X0 + X1 ) = X̄1−i (2.1.24)

so that
Xi ,→ Xi ∩ X̄1−i . (2.1.25)
To complete the proof of the first item it then suffices to show that this embedding is surjective.
Fix i ∈ {0, 1} and let x ∈ Xi ∩ X̄1−i . Then for every ε > 0 there exists xε ∈ X1−i such that
kx − xε k0+1 < ε, which in turn allows us to choose yε ∈ Xi and zε ∈ X1−i such that x − xε = yε + zε
and
kyε ki + kzε k1−i < ε. (2.1.26)

54
Upon rearranging, we find that

Xi 3 x − yε = xε + zε ∈ X1−i , and hence xε + zε ∈ X0 ∩ X1 . (2.1.27)

Then
kx − (xε + zε )ki = kyε ki < ε for all ε > 0 (2.1.28)
and we deduce that x ∈ cl(X0 ∩ X1 , Xi ) = Xi . Thus the embedding (2.1.25) is surjective, and the
first item is proved.
For the second item we first note that by (2.1.24) we have

X1−i + Xi ,→ X1−i + X̄1−i = X̄1−i . (2.1.29)

We claim that this embedding is a surjection. Indeed, for a fixed i ∈ {0, 1} and x ∈ X̄1−i ⊆ X0 + X1
we can write x = x0 + x1 for x0 ∈ X0 and x1 ∈ X1 . Then xi = x − x1−i ∈ X̄1−i ∩ Xi , but by the
first item X̄1−i ∩ Xi = Xi , so
x = x1−i + xi ∈ X1−i + Xi . (2.1.30)
This proves the claim, which then completes the proof of the second item.
The third and fourth items follows directly from the first and second, so it remains only to prove
the fifth. From the first item we have

X0 + X1 = X0 ∩ X̄1 + X̄0 ∩ X1 ,→ X̄0 ∩ X̄1 + X̄0 ∩ X̄1 = X̄0 ∩ X̄1 . (2.1.31)

Let x ∈ X̄0 ∩ X̄1 . By the second item we can then find a ∈ X0 , b ∈ X1 , c ∈ X0 , and d ∈ X1 such
that
x = a + b = c + d. (2.1.32)
Upon rearranging, this implies that

a − c = d − b ∈ X0 ∩ X1 , (2.1.33)

and hence
a = c + (a − c) ∈ X0 + X0 ∩ X1 ⊆ X0 , (2.1.34)
so
x = a + b ∈ X0 + X 1 (2.1.35)
and thus the embedding (2.1.31) is surjective, so X0 + X1 = X̄0 ∩ X̄1 .
Next we note that (2.1.22) shows

X0 + X1 = X0 ∩ X1 = cl(X0 ∩ X1 , X0 + X1 ) ,→ cl(Xi , X0 + X1 ) = X̄i for i = 0, 1 (2.1.36)

so
X0 ∩ X1 ,→ X̄0 ∩ X̄1 . (2.1.37)
To conclude we will show that this embedding is also surjective. Let x ∈ X̄0 ∩ X̄1 . Since x ∈ X̄0 ,
for each ε > 0 we can find xε ∈ X0 , aε ∈ X0 , and bε ∈ X1 such that x − xε = aε + bε and
kaε k0 + kbε k1 < ε. Similarly, since x ∈ X̄1 , for each ε > 0 we can find yε ∈ X1 , cε ∈ X0 , and dε ∈ X1
such that x − yε = cε + dε and kcε k0 + kdε k1 < ε. Then

xε + aε + bε = x = yε + cε + dε , (2.1.38)

55
and upon rearranging we see that
xε + aε − cε = −bε + yε + dε ∈ X0 ∩ X1 . (2.1.39)
Then x − (xε + aε − cε ) = bε + cε and
kx − (xε + aε − cε )k0+1 ≤ kcε k0 + kbε k1 < 2ε, (2.1.40)
from which we deduce that x ∈ X0 ∩ X1 , and (2.1.37) is surjective. Hence, X0 ∩ X1 = X̄0 ∩ X̄1 ,
and the fifth item is proved.
The final item of this theorem shows that there is a special role played by the single space
X0 + X1 = X0 ∩ X1 = X̄0 ∩ X̄1 = X0 + X1 . (2.1.41)
The following diagram summarizes the relation of this space to the others we have discussed.

X0 ∩ X1 X0 X0 X̄0

X0 ∩ X1 X0 + X1 = X0 ∩ X1 = X0 + X1 = X̄0 ∩ X̄1 X̄0 + X̄1

X1 X1 X̄1 X0 + X1
(2.1.42)
We might hope at this point that we could iterate the constructions we’ve done so far to get
even more spaces. For instance, we could take the upper closure of Xi or the lower closure of X̄i .
However, the above diagram indicates another crucial fact: iterating these constructions doesn’t
produce anything else. Indeed, it immediately shows the following.
Corollary 2.1.13. Let X0 and X1 be compatible Banach spaces. Then for i ∈ {0, 1}, the upper
closure of Xi , the lower closure of X̄i , and the space X0 + X1 = X0 ∩ X1 = X̄0 ∩ X̄1 = X0 + X1
coincide.
The upshot of this analysis is that while the upper and lower closure operations do allow us to
produce five new spaces intermediate to X0 and X1 , we are stuck with these five and can proceed
no further with the closure, sum, and intersection operations alone. If we want to construct more
intermediate spaces, we need a new idea. We will turn to the development of this idea momentarily,
but first we record another simple consequences of the above diagram.
Corollary 2.1.14. Let X0 and X1 be compatible Banach spaces. The following hold.
1. If Xi ,→ X1−i , then
Xi ,→ X̄i = X1−i = X0 ∩ X1 = X0 + X1 ,→ X1−i . (2.1.43)

2. If X0 ∩ X1 is a closed subspace of X0 + X1 , then

X0 ∩ X1 = X0 = X1 = X0 + X1 = X0 ∩ X1 = X̄0 ∩ X̄1 = X0 + X1 . (2.1.44)

3. If Xi ⊆ X1−i is a closed subspace, then

Xi = X0 = X1 = X0 + X1 = X0 ∩ X1 = X̄0 ∩ X̄1 = X0 + X1 ⊆ X1−i . (2.1.45)

Proof. Exercise.

56
2.1.4 The K and J functions
Our starting point for constructing more intermediate spaces to a given compatible pair of Banach
spaces X0 and X1 is the simple realization that we can introduce a positive real parameter’s worth
of equivalent norms on the spaces X0 ∩ X1 and X0 + X1 . We give these a special name now.
Definition 2.1.15. Let X0 and X1 be compatible Banach spaces. We define the maps K : (X0 +
X1 ) × R+ → R and J : (X0 ∩ X1 ) × R+ → R via

K(x, t) = inf{kx0 k0 + t kx1 k1 | x = x0 + x1 for xi ∈ Xi , i = 0, 1} (2.1.46)

and
J(x, t) = max{kxk0 , t kxk1 }. (2.1.47)
We begin our study of these new functions with the following proposition, which records the
basic properties of the K function.
Proposition 2.1.16. Let X0 and X1 be compatible Banach spaces. Then the following hold.
1. For each t ∈ R+ the map X0 + X1 3 x 7→ K(x, t) ∈ R is a norm that’s equivalent to the usual
norm.

2. For each x ∈ X0 + X1 the map K(x, ·) is continuous, nondecreasing, and concave.

3. For each x ∈ X0 + X1 , the map R+ 3 t 7→ K(x, t)/t ∈ R is nonincreasing, and we have the
inequalities

min{1, t/s}K(x, s) ≤ K(x, t) ≤ max{1, t/s}K(x, s) for all t, s ∈ R+ . (2.1.48)

4. Let K̃ : (X0 + X1 ) × R+ → R be the K map obtained by flipping the indices on X0 and X1 ,

i.e.
K̃(x, t) = inf{kx1 k1 + t kx0 k0 | x = x0 + x1 for xi ∈ Xi , i = 0, 1}. (2.1.49)
Then
K(x, t)
= K̃(x, 1/t) for all x ∈ X0 + X1 and t ∈ R+ . (2.1.50)
t
Proof. The fact that K(·, t) defines a norm is left as an exercise. Then the first item follows from
the simple inequality

min{1, t} kxkX0 +X1 ≤ K(x, t) ≤ max{1, t} kxkX0 +X1 . (2.1.51)

Now let x = x0 + x1 for xi ∈ Xi and suppose that 0 < t < s < ∞. Then

K(x, t) ≤ kx0 k0 + t kx1 k1 ≤ kx0 k0 + s kx1 k1 , (2.1.52)

but since this holds for all such decompositions, we deduce that

K(x, t) ≤ K(x, s). (2.1.53)

Similarly, for θ ∈ [0, 1] we have that

θK(x, t)+(1−θ)K(x, s) ≤ θ[kx0 k0 +t kx1 k1 ]+(1−θ)[kx0 k0 +s kx1 k1 ] = kx0 k0 +[θt+(1−θ)s] kx1 k1 ,

(2.1.54)

57
and since this holds for all such decompositions, we deduce that

θK(x, t) + (1 − θ)K(x, s) ≤ K(x, θt + (1 − θ)s). (2.1.55)

This proves all of the second item except continuity, which we delay momentarily.
For the third item again let x = x0 + x1 with xi ∈ Xi and let s, t ∈ (0, ∞). If t ≤ s then the
first item shows that K(x, t) ≤ K(x, s) = max{1, t/s}K(x, s). On the other hand, if s < t, then
s s s
K(x, t) ≤ [kx0 k0 + t kx1 k1 ] = kx0 k0 + s kx1 k1 ≤ kx0 k0 + s kx1 k1 , (2.1.56)
t t t
and since this holds for all such decompositions we find that st K(x, t) ≤ K(x, s), which implies
that K(x, t) ≤ max{1, t/s}K(x, s). This proves the second stated estimate, but the first follows by
reversing the roles of s and t. This proves the third item.
Finally, we complete the proof of the second item by proving that K(x, ·) is continuous. Fix
s ∈ (0, ∞). Then

K(x, s) = lim min{1, t/s}K(x, s) = lim max{1, t/s}K(x, s), (2.1.57)

t→s t→s

so the inequalities of the third item show that K(x, s) = limt→s K(x, t), and we conclude that
K(x, ·) is continuous. We leave the proof of the fourth item as an exercise.

Next we record a corresponding result on the basic properties of the J function.

Proposition 2.1.17. Let X0 and X1 be compatible Banach spaces. Then the following hold.

1. For each t ∈ R+ the map X0 ∩ X1 3 x 7→ J(x, t) ∈ R is a norm that’s equivalent to the usual
norm on X0 ∩ X1 .

2. For each x ∈ X0 ∩ X1 the map J(x, ·) is continuous, nondecreasing, and convex.

3. For each x ∈ X0 ∩ X1 , the map R+ 3 t 7→ J(x, t)/t ∈ R is nonincreasing, and we have the
inequalities

min{1, t/s}J(x, s) ≤ J(x, t) ≤ max{1, t/s}J(x, s) for all t, s ∈ R+ . (2.1.58)

4. Let J˜ : (X0 ∩ X1 ) × R+ → R be the J map obtained by flipping the indices on X0 and X1 , i.e.
˜ t) = max{kxk , t kxk }.
J(x, (2.1.59)
1 0

Then
J(x, t) ˜ 1/t) for all x ∈ X0 ∩ X1 and t ∈ R+ .
= J(x, (2.1.60)
t
Proof. We will only prove the second and third items and leave the first and fourth as exercises.
If s ≤ t then s kxk1 ≤ t kxk1 ≤ J(x, t) and kxk0 ≤ J(x, t), so J(x, s) ≤ J(x, t), which shows that
J(x, ·) is nondecreasing.
If s, t ∈ R+ and θ ∈ [0, 1], then [θs + (1 − θ)t] kxk1 ≤ θJ(x, s) + (1 − θ)J(x, t), while kxk0 =
θ kxk0 + (1 − θ) kxk0 ≤ θJ(x, s) + (1 − θ)J(x, t), so J(x, θs + (1 − θ)t) ≤ θJ(x, s) + (1 − θ)J(x, t).
Thus J(x, ·) is convex.

58
 If t ≤ s then J(x, t) ≤ J(x, s) by the above. On the other hand if s < t, then
s
J(x, t) = max{(s/t) kxk0 , s kxk1 } ≤ max{kxk0 , s kxk1 } = J(x, s). (2.1.61)
t
Thus J(x, t) ≤ max{1, t/s}J(x, s). The estimates of the third item follow from this and the inequal-
ity obtained by reversing t and s. This also shows that R+ 3 t 7→ J(x, t)/t ∈ R is nondecreasing.
Continuity follows directly from the bounds of the third item, and the proof of the second and third
items is complete.
The K and J functions are related in a useful way, as the following result shows.

Proposition 2.1.18. Let X0 and X1 be compatible Banach spaces. The for each x ∈ X0 ∩ X1 and
s, t ∈ R+ we have that K(x, s) ≤ min{1, s/t}J(x, t).

Proof. Since x ∈ X0 ∩ X1 we have that K(x, s) ≤ kxk0 ≤ J(x, t) and K(x, s) ≤ s kxk1 =
(s/t)t kxk1 ≤ (s/t)J(x, t), so

K(x, s) ≤ min{J(x, t), (s/t)J(x, t)} = min{1, s/t}J(x, t). (2.1.62)

It turns out that the K function’s asymptotic behavior at 0 and ∞ encodes some useful informa-
tion related to inclusion in spaces that we have seen before. In turn, this information is equivalent
to an extremely useful decomposition that will be essential in our subsequent work with the K and
J functions. We record this result now.

Theorem 2.1.19. Let X0 and X1 be compatible Banach spaces. Then the following hold for x ∈
X0 + X 1 .

1. lim K(x, t) = 0 if and only if x ∈ X̄1 .

t→0

2. lim K(x, t)/t = 0 if and only if x ∈ X̄0 .

t→∞

3. The following are equivalent.

(a) lim K(x, t) = lim K(x, t)/t = 0.

t→0 t→∞

(b) x ∈ X0 ∩ X1 = X0 + X1 .
P
(c) There exists a sequence {xn }n∈Z ⊆ X0 ∩ X1 such that x = n∈Z xn , where the series
converges in X0 + X1 .

Moreover, if any (and hence all) of these holds, then for any 1 < r < ∞ the sequence {xn }n∈Z
in (c) can be chosen such that J(xn , rn ) ≤ 2(1 + r)K(x, rn ) for each n ∈ Z.

Proof. We begin by proving the first item. Suppose that lim K(x, t) = 0. Then for every ε > 0
t→0
there exists δ > 0 such that 0 < t < δ implies K(x, t) < ε. In particular, for every ε > 0 we can
pick yε ∈ X0 and zε ∈ X1 such that kyε k0 + (δ/2) kzε k1 < ε. In turn, this implies that

kx − zε k0+1 ≤ kyε k0 < ε for every ε > 0, (2.1.63)

and since zε ∈ X1 we deduce from this that x ∈ X̄1 .

59
 Conversely, assume that x ∈ X̄1 . Then for every ε > 0 we can pick xε ∈ X1 such that
kx − xε k0+1 < ε. In turn, we may choose yε ∈ X0 and zε ∈ X1 such that x − xε = yε + zε and
kyε k0 + kzε k1 < ε. Then x = yε + (xε + zε ) and for any t ∈ (0, ∞) we may estimate

K(x, t) ≤ kyε k0 + t kxε + zε k1 < ε + t kxε + zε k1 . (2.1.64)

Hence,
lim K(x, t) ≤ ε for every ε > 0, so lim K(x, t) = 0. (2.1.65)
t→0 t→0

This completes the proof of the first item.

The second item follows from the first and the fourth item of Proposition 2.1.16. The equivalence
of (a) and (b) in the third item follows from the first two items and Theorem 2.1.12. We will show
that (a) ⇔ (c) to conclude. These are clearly equivalent if x = 0, so we may assume without loss
of generality that x 6= 0.
Suppose that (a) holds and let 1 < r < ∞. For each n ∈ Z we can choose yn ∈ X0 and zn ∈ X1
such that x = yn + zn and
kyn k0 + rn kzn k1 < 2K(x, rn ). (2.1.66)
Then (a) implies that
lim kyn k0 = lim kzn k1 = 0. (2.1.67)
n→−∞ n→∞

For n ∈ Z define
xn = yn − yn+1 = zn+1 − zn ∈ X0 ∩ X1 . (2.1.68)
By construction, for each n ∈ Z we have the estimate

J(xn , rn ) ≤ max{kyn k0 + kyn+1 k0 , rn kzn k1 + rn kzn+1 k1 }

2
≤ max{2K(x, rn ) + 2K(x, rn+1 ), 2K(x, rn ) + K(x, rn+1 )}
r
2
≤ max{2K(x, rn ) + 2 max{1, r}K(x, rn ), 2K(x, rn ) + max{1, r}K(x, rn )}
r
= 2(1 + r)K(x, rn ). (2.1.69)

Moreover, for N ∈ N we may compute

X
x− xn = x − yN + y−N +1 = zN + y−N +1 , (2.1.70)
|n|≤N

which then implies that

X
x− xn ≤ ky−N +1 k0 + kzN k1 → 0 as N → ∞. (2.1.71)
|n|≤N
0+1

Thus, (c) holds, and we have proved that (a) ⇒ (c).

P
Now suppose that (c) holds. Let ε > 0, and choose N ∈ N such that x − |n|≤N xn < ε.
0+1
Then we may estimate

X X X X
K(x, t) ≤ K(x − xn , t) + K( xn , t) ≤ max{1, t} x − xn +t xn (2.1.72)
|n|≤N |n|≤N |n|≤N |n|≤N
0+1 1

60
in order to see that lim K(x, t) ≤ ε for every ε > 0 and hence lim K(x, t) = 0. Similarly, we may
t→0 t→0
bound
P P
K(x, t) K(x − |n|≤N xn , t) K( |n|≤N xn , t)
≤ +
t t t
max{1, t} X 1 X
≤ x− xn + xn (2.1.73)
t t
|n|≤N |n|≤N
0+1 0

in order to see that lim K(x, t)/t ≤ ε for every ε > 0 and hence lim K(x, t)/t = 0. Thus, (a) holds,
t→∞ t→∞
and we have proved that (c) ⇒ (a), which completes the proof of the third item.

2.2 Interpolating between compatible Banach spaces

We now have all of the tools needed to begin constructing new intermediate spaces. In fact, with
what we now know about the K and J functions, there are many options available to us. We begin
our discussion with some heuristic considerations in order to justify the definition we will ultimately
work with.

2.2.1 Heuristics
We have now seen that for each t ∈ R+ the function K(·, t) defines an equivalent norm on X0 + X1
when X0 and X1 are compatible Banach spaces. Moreover, since K(x, ·) is continuous for each
x ∈ X0 + X1 , it’s measurable for any choice of a Radon measure on R+ . These facts suggest that we
might use some sort of Lebesgue norms relative to a Radon measure µ on R+ to build new spaces
intermediate to X0 and X1 . This leads us to the following definition.

Definition 2.2.1. Let X0 and X1 be compatible Banach spaces. Suppose that µ is a Radon measure
on R+ and consider a continuous weight function w : R+ → R+ . Let 1 ≤ p ≤ ∞. We define
k·kX(µ,w,p) : X0 + X1 → [0, ∞] via

kxkX(µ,w,p) = kwK(x, ·)kLpµ < ∞, (2.2.1)

where we recall that Lpµ (R+ ) denotes Lp on R+ with respect to the measure µ. We define the space

X(µ, w, p) = {x ∈ X0 + X1 | kxkX(µ,w,p) < ∞}. (2.2.2)

Note that the weight w can be absorbed into the measure when 1 ≤ p < ∞, so the main utility of
the weight is seen when p = ∞.

Now, based on our prior discussion of intermediate spaces, the natural question that arises is
when we can guarantee that X(µ, w, p) is intermediate. To address this question, first note that if
x ∈ X0 ∩ X1 then
K(x, t) ≤ min{kxk0 , t kxk1 } ≤ min{1, t} kxkX0 ∩X1 , (2.2.3)
and consequently
kxkX(µ,w,p) ≤ kxkX0 ∩X1 kw min{1, ·}kLpµ . (2.2.4)

61
On the other hand, for x ∈ X0 + X1 we have K(x, t) ≥ min{1, t} kxkX0 +X1 , so

kxkX(µ,w,p) ≥ kxkX0 +X1 kw min{1, ·}kLpµ . (2.2.5)

From these calculations we learn that if the space X(µ, w, p) is nontrivial, then we have the inclusion
w min{1, ·} ∈ Lpµ (R+ ), and conversely this inclusion implies the embeddings

X0 ∩ X1 ,→ X(µ, w, p) ,→ X0 + X1 . (2.2.6)

To prove that X(µ, w, p) is intermediate, we also need to verify that it is complete. For this note
that the assumption that µ is Radon guarantees that w max{1, ·} is locally in Lpµ (R+ ).
Theorem 2.2.2. Let X0 and X1 be compatible Banach spaces and 1 ≤ p ≤ ∞. Suppose that µ is a
Radon measure on R+ , w : R+ → R+ is continuous, and w min{1, ·} ∈ Lpµ (R+ ). Then the following
hold.
1. The space X(µ, w, p) is a Banach space intermediate to X0 and X1 .

2. If w max{1, ·} ∈ Lpµ (R+ ), then X(µ, w, p) = X0 + X1 with equivalent norms.

R R
3. Suppose that p < ∞. If R+ (w(t))p dµ(t) = ∞, then X(µ, w, p) ,→ X̄1 . If R+ (tw(t))p dµ(t) =
R R
∞, then X(µ, w, p) ,→ X̄0 . If R+ (w(t))p dµ(t) = R+ (tw(t))p dµ(t) = ∞, then X(µ, w, p) ,→
X0 ∩ X1 = X0 + X 1 .
Proof. From the above analysis we know that

kxkX0 +X1 kw min{1, ·}kLpµ ≤ kxkX(µ,w,p) ≤ kxkX0 ∩X1 kw min{1, ·}kLpµ (2.2.7)

and so X0 ∩ X1 ,→ X(µ, w, p) ,→ X0 + X1 . To conclude we only need to prove that X(µ, w, p) is

complete.
Let {xn }∞
n=` ⊆ X(µ, w, p) be Cauchy. According to (2.2.7), the sequence is also Cauchy in the
Banach space X0 + X1 , and so converges in X0 + X1 to some x. For t ∈ R+ we may estimate

K(x − xn , t) ≤ K(xn − xm , t) + K(x − xm , t) ≤ K(xn − xm , t) + max{1, t} kx − xm kX0 +X1 (2.2.8)

and
K(x, t) ≤ K(xn , t) + K(x − xn , t). (2.2.9)
Then for 1 < λ < ∞ we may bound

kwK(x − xn , ·)kLpµ ([λ−1 ,λ]) ≤ kxn − xm kX(µ,w,p) + kx − xm kX0 +X1 kw max{1, ·}kLpµ ([λ−1 ,λ]) (2.2.10)

and

kwK(x, ·)kLpµ ([λ−1 ,λ]) ≤ kxn kX(µ,w,p) + kx − xn kX0 +X1 kw max{1, ·}kLpµ ([λ−1 ,λ]) . (2.2.11)

We take the limsup in (2.2.11) to see that

kwK(x, ·)kLpµ ([λ−1 ,λ]) ≤ lim sup kxn kX(µ,w,p) < ∞. (2.2.12)
n→∞

Sending λ → ∞ and employing the monotone convergence theorem if p < ∞, we deduce from this
that x ∈ X(µ, w, p).

62
 Now let ε > 0 and choose N ≥ ` such that n, m ≥ N implies kxn − xm kX(µ,w,p) < ε. Then for
n ≥ N we deduce from (2.2.10) that

kwK(x − xn , ·)kLpµ ([λ−1 ,λ])

 
≤ lim sup kxn − xm kX(µ,w,p) + kx − xm kX0 +X1 kw max{1, ·}kLpµ ([λ−1 ,λ]) ≤ ε (2.2.13)
m→∞

for all 1 < λ < ∞. Again sending λ → ∞ and using the monotone convergence theorem if p < ∞,
we deduce that
n ≥ N ⇒ kx − xn kX(µ,w,p) = kwK(x − xn , ·)kLpµ (R+ ) ≤ ε, (2.2.14)
which means that xn → x in X(µ, w, p) as n → ∞. Hence, X(µ, w, p) is complete, and the first
item is proved.
Suppose now that w max{1, ·} ∈ Lpµ (R+ ). Then from the estimate K(x, t) ≤ max{1, t} kxkX0 +X1
we find that
kxkX(µ,w,p) ≤ kw max{1, ·}kLpµ kxkX0 +X1 (2.2.15)
for all x ∈ X0 + X1 . This and (2.2.7) imply that k·kX(µ,w,p) and k·kX0 +X1 are equivalent norms and
X(µ, w, p) = X0 + X1 . This proves the second item.
We now prove the third item, assuming that p < ∞. Let 0 6= x ∈ X(µ, w, p). If x ∈ / X̄1 , then
Theorem 2.1.19 implies that there exists ε > 0 such that K(x, t) > ε for all t > 0, and hence
Z 1/p
p
∞ > kxkX(µ,w,p) ≥ ε (w(t)) dµ(t) . (2.2.16)
R+
R
Hence, if R+ (w(t))p dµ(t) = ∞, then x ∈ X̄1 . Similarly, if x ∈ / X̄0 , then Theorem 2.1.19 implies
that there exists ε > 0 such that K(x, t)/t > ε for all t > 0, and hence
Z  p 1/p Z 1/p
p K(x, t) p
∞ > kxkX(µ,w,p) = (tw(t)) dµ(t) ≥ε (tw(t)) dµ(t) . (2.2.17)
R+ t R+
R
In turn, this means that if R+ (tw(t))p dµ(t) = ∞, then x ∈ X̄0 . This proves the first two assertions
of the third item, and the third assertion follows from these and Theorem 2.1.12, which completes
the proof of the third item.

Remark 2.2.3. One consequence of this result can be stated nicely if we know a priori that X0 ∩ X1
is nontrivial. Indeed, in this case we know that X(µ, w, p) is a nontrivial Banach space intermediate
to X0 and X1 if and only if w min{1, ·} ∈ Lpµ (R+ ).

The simplest candidate to consider is w(t) = t−α for some α ∈ R and µ standard Lebesgue
measure on R+ . Then it’s easy to check that w min{1, ·} ∈ L∞ (R+ ) if and only if 0 ≤ α ≤ 1, while
for 1 ≤ p < ∞ we have that w min{1, ·} ∈ Lp (R+ ) if and only if
Z 1 Z ∞
dt dt
(α−1)p
< ∞ and < ∞, (2.2.18)
0 t 1 tαp
which in turn is equivalent to
1 1
<α<1+ . (2.2.19)
p p

63
To ensure that we pick a useful weight for p < ∞, we then consider θ ∈ (0, 1) and set
 
1 1 1
α = (1 − θ) + θ 1 + = + θ, (2.2.20)
p p p

which means that Z  p 1/p

K(x, t) dt
kxkX(µ,w,p) = . (2.2.21)
R+ tθ t
The upshot of this analysis is that the most natural choice is not Lebesgue measure, but rather the
measure µ = dt/t. This is actually a very nice choice on R+ , as it is a Haar measure for the locally
compact Abelian group R+ . Thus, it seems like a good candidate to study is the space X(µ, ()−θ , p).
Note that if p < ∞, then
tp dt
Z Z
dt
1+pθ
= 1+pθ
= ∞, (2.2.22)
R+ t R+ t

and so the third item of Theorem 2.2.2 shows that X(µ, ()−θ , p) ,→ X0 ∩ X1 . On the other hand,

K(x, t) K(x, t)
sup θ
< ∞ ⇒ lim K(x, t) = lim = 0 ⇒ x ∈ X0 ∩ X1 , (2.2.23)
t>0 t t→0 t→∞ t

and again we find that X(µ, ()−θ , p) ,→ X0 ∩ X1 . Finally, note that w max{1, ·} ∈
/ Lpµ (R+ ) for any
1 ≤ p ≤ ∞, so the second item of the theorem does not apply and there is hope that X(µ, ()−θ , p)
is strictly smaller than X0 + X1 .

2.2.2 The interpolation spaces

The above heuristics lead us to change our notation slightly and make the following definitions.

Definition 2.2.4. Let X0 and X1 be compatible Banach spaces.

1. For 1 ≤ p ≤ ∞ and θ ∈ (0, 1) we define the space

(X0 , X1 )θ,p = {x ∈ X0 + X1 | kxkθ,p < ∞} (2.2.24)

where k·kθ,p : X0 + X1 → [0, ∞] is defined for p < ∞ by

Z  p 1/p
K(x, t) dt
kxkθ,p = (2.2.25)
R+ tθ t

and for p = ∞ by
K(x, t)
kxkθ,∞ = sup . (2.2.26)
t>0 tθ
Clearly, we have that (X0 , X1 )θ,p = X(µ, (·)−θ , p) for µ = dt/t, and so Theorem 2.2.2 im-
plies that (X0 , X1 )θ,p is a Banach space intermediate to X0 and X1 and that (X0 , X1 )θ,p ,→
X0 ∩ X1 = X0 + X 1 .

2. For p = ∞ and θ ∈ {0, 1} we define the space

(X0 , X1 )θ,∞ = {x ∈ X0 + X1 | kxkθ,∞ < ∞} (2.2.27)

64
 where k·kθ,∞ : X0 + X1 → [0, ∞] is defined by

K(x, t)
kxkθ,∞ = sup . (2.2.28)
t>0 tθ

Clearly, (X0 , X1 )θ,∞ = X(µ, (·)−θ , ∞) for µ = dt/t, and so Theorem 2.2.2 implies that
(X0 , X1 )θ,p is a Banach space intermediate to X0 and X1 . However, in this case the available
embeddings are Xθ ,→ (X0 , X1 )θ,∞ ,→ Xθ for θ ∈ {0, 1} (the former is trivial and the latter
follows from Theorem 2.1.19).

Three remarks are in order.

Remark 2.2.5. When 1 ≤ p < ∞, we don’t include the endpoint cases θ ∈ {0, 1} since for
0 6= x ∈ X0 + X1 , Z ∞ Z ∞
p dt p dt
(K(x, t)) ≥ (K(x, 1)) =∞ (2.2.29)
0 t 1 t
and p p Z
Z ∞   1
K(x, t) dt K(x, 1) dt
≥ = ∞. (2.2.30)
0 t t 1 0 t
However, when p = ∞, there is useful information encoded in the quantities

K(x, t) K(x, t)
kxk0,∞ = sup K(x, t) = lim K(x, t) and kxk1,∞ = sup = lim . (2.2.31)
t>0 t→∞ t>0 t t→0 t

Remark 2.2.6. If we define K̃(x, t) = K(x, t)/t, then R+ 3 t 7→ K̃(x, t) ∈ [0, ∞] is nonincreasing,
and
kxkθ,p = (·)1−θ K̃(x, ·) p . (2.2.32)
Lµ

Since K̃(x, ·) and its rearrangement coincide (exercise: verify this claim), we deduce that

kxkθ,p = K̃(x, ·) , (2.2.33)

L1/(1−θ),p (R+ )

and we naturally arrive at a nice connection between Lorentz spaces and our new spaces.

Remark 2.2.7. For every choice of θ ∈ (0, 1) and 1 ≤ p ≤ ∞ we know that (X0 , X1 )θ,p ,→
X0 ∩ X1 = X0 + X1 . This highlights the special role played by the latter space in interpolation
theory: it serves as a container space for everything we construct using this method. Note, though,
that in general we do not know that (X0 , X1 )∞,θ is contained in this space when θ ∈ {0, 1}. It is
easy, though, to use Theorem 2.1.19 to see that

Xθ ,→ (X0 , X1 )θ,∞ ,→ X̄θ . (2.2.34)

Our next result establishes the fundamental embedding properties of our new spaces as we vary
the parameter 1 ≤ p ≤ ∞. The result should be contrasted with Theorem 1.1.39.

Theorem 2.2.8. Let X0 and X1 be compatible Banach spaces, 1 ≤ p ≤ ∞, and θ ∈ (0, 1). Then
the following hold.

1. We have that (X0 , X1 )θ,p = (X1 , X0 )1−θ,p with identical norms.

65
 2. If x ∈ (X0 , X1 )θ,p , then
K(x, t) ≤ (θp)1/p tθ kxkθ,p for all t ∈ R+ (2.2.35)

with the usual understanding that (θp)1/p = 1 for p = ∞.

3. If p < q ≤ ∞, then (X0 , X1 )θ,p ,→ (X0 , X1 )θ,q and
kxkθ,q ≤ (θp)1/p−1/q kxkθ,p for all x ∈ (X0 , X1 )θ,p . (2.2.36)

Proof. We’ll prove the first two results for p < ∞ and leave the case p = ∞ as an exercise. We use
the fourth item of Proposition 2.1.16 and a change of variables s = 1/t to compute
Z  p Z
p K(x, t) dt
p ds
kxk(X0 ,X1 )θ,p = θ
= sθ K(x, 1/s)
R+ t t R+ s
Z !p
K̃(x, s) ds
= 1−θ
= kxkp(X1 ,X0 )1−θ,p , (2.2.37)
R+ s s

where K̃ is determined by switching the roles of X0 and X1 . This proves the first result.
Next we use the fact that K(x, ·) is nondecreasing from Proposition 2.1.16 to bound
Z ∞ p Z ∞
p K(x, s) ds p ds (K(x, t))p
kxkθ,p ≥ ≥ (K(x, t)) = (2.2.38)
t sθ s t s1+θp θptθp
for every t ∈ R+ . Upon rearranging, we then complete the proof of the second item.
Now suppose that p < q ≤ ∞. If q = ∞, then the second item implies that kxkθ,∞ ≤
(θp)1/p kxkθ,p , so it remains to consider the case q < ∞. In this case we estimate
Z  q−p  p !1/q
K(x, t) K(x, t) dt 1−p/q p/q
kxkθ,q = θ θ
≤ kxkθ,∞ kxkθ,p ≤ (θp)1/p−1/q kxkθ,p , (2.2.39)
R+ t t t
which completes the proof of the third item.
We can organize the results of Remark 2.2.7 and Theorem 2.2.8 as the following diagram, which
highlights the fact that for each θ ∈ (0, 1) we have a continuum of nested intermediate spaces
indexed by 1 ≤ p ≤ ∞, while for θ ∈ {0, 1} we have outlier spaces. We typically arrange the spaces
with θ increasing as we move down, so here 0 < θ0 < θ < θ1 < 1.
X0 X0 (X0 , X1 )0,∞ X̄0

(X0 , X1 )θ0 ,1 (X0 , X1 )θ0 ,p (X0 , X1 )θ0 ,∞

X0 ∩ X1 (X0 , X1 )θ,1 (X0 , X1 )θ,p (X0 , X1 )θ,∞ X0 ∩ X1 X0 + X1

(X0 , X1 )θ1 ,1 (X0 , X1 )θ1 ,p (X0 , X1 )θ1 ,∞

X1 X1 (X0 , X1 )1,∞ X̄1

(2.2.40)

66
 Next we study how our spaces depend on the parameter θ by examining the intersection of two
of them.
Proposition 2.2.9. Let X0 and X1 be compatible Banach spaces, 1 ≤ p ≤ ∞, and 0 < θ0 < θ <
θ1 < 1. Then (X0 , X1 )θ0 ,p ∩ (X0 , X1 )θ1 ,p ,→ (X0 , X1 )θ,p , and
θ1 −θ θ−θ0

kxkθ,p ≤ kxkθθ01,p−θ0 kxkθθ11,p−θ0 (2.2.41)

for every x ∈ (X0 , X1 )θ0 ,p ∩ (X0 , X1 )θ1 ,p .
Proof. Write θ = (1 − α)θ0 + αθ1 for α ∈ (0, 1). If p < ∞ then we use Hölder’s inequality to bound
Z  p Z  (1−α)p  αp
p K(x, t) dt K(x, t) K(x, t) dt (1−α)p
kxkθ,p = θ
= θ0 θ1
≤ kxkθ0 ,p kxkαp
θ1 ,p , (2.2.42)
R+ t t R+ t t t
and the stated bound follows by solving for α and taking pth roots. A similar argument works for
p = ∞.
One particularly interesting consequence of this result is the interpolation estimate recorded in
the following.
Corollary 2.2.10. Let X0 and X1 be compatible Banach spaces, 1 ≤ p ≤ ∞, and θ ∈ [0, 1]. There
exists a constant C ∈ R+ such that if x ∈ X0 ∩ X1 , then
kxkθ,p ≤ C kxk1−θ
0 kxkθ1 for all x ∈ X0 ∩ X1 . (2.2.43)
Proof. Exercise.
Proposition 2.2.9 has the benefit of giving a nice constant on the right side but only works with
the same value of p in all the spaces. Next we prove that the spaces can be improved at the expense
of a worse constant on the right. This result should be contrasted with that of Theorem 1.1.41.
Theorem 2.2.11. Let X0 and X1 be compatible Banach spaces, and 0 ≤ θ0 < θ < θ1 ≤ 1. Then
(X0 , X1 )θ0 ,∞ ∩ (X0 , X1 )θ1 ,∞ ,→ (X0 , X1 )θ,1 , and
θ1 −θ θ−θ0
 
1 1 −θ0 −θ0
kxkθ,1 ≤ + kxkθθ01,∞ kxkθθ11,∞ (2.2.44)
θ1 − θ θ − θ0
for every x ∈ (X0 , X1 )θ0 ,∞ ∩ (X0 , X1 )θ1 ,∞ .
Proof. It suffices to prove the stated estimate, and for this we may also assume that x 6= 0. Due to
the assumed inclusions, we have the bound
K(x, t) ≤ min{tθ0 kxkθ0 ,∞ , tθ1 kxkθ1 ,∞ }. (2.2.45)
Set !1/(θ1 −θ0 )
kxkθ0 ,∞
T = ∈ R+ . (2.2.46)
kxkθ1 ,∞
Then
T tθ1 kxkθ1 ,∞ dt ∞ tθ0
kxkθ0 ,∞ dt
Z Z Z
K(x, t) dt
= +
R+ tθ t 0 tθ t T tθ t
θ1 −θ θ1 −θ θ−θ0
 
T 1 1 1 θ1 −θ0 θ1 −θ0
= kxkθ1 ,∞ + kxkθ1 ,∞ = + kxk θ0 ,∞ kxkθ1 ,∞ (2.2.47)
θ1 − θ (θ − θ0 )T θ−θ0 θ1 − θ θ − θ0

67
 Now that we have developed some of the essential properties of our interpolation spaces, we re-
turn to the question of whether these spaces admit a corresponding theory of operator interpolation.
It turns out that they do, and the theory is relatively simple in comparison to the Marcinkiewicz
and Riesz-Thorin theorems.

Theorem 2.2.12. Let X0 and X1 be compatible Banach spaces and Y0 and Y1 be compatible Banach
spaces over the same field. Suppose that T : X0 +X1 → Y0 +Y1 is a linear map such that T (Xi ) ⊆ Yi
and T ∈ L(Xi ; Yi ) for i ∈ {0, 1}. Then for every 1 ≤ p ≤ ∞ and θ ∈ (0, 1) we have that
T ∈ L((X0 , X1 )θ,p , (Y0 , Y1 )θ,p ) and

kT kL((X0 ,X1 )θ,p ,(Y0 ,Y1 )θ,p ) ≤ kT k1−θ θ

L(X0 ;Y0 ) kT kL(X1 ;Y1 ) . (2.2.48)

Proof. We write KX for the K function associated to X0 and X1 and KY for the K function
associated to Y0 and Y1 . The result is trivial if T = 0, so we may assume that T 6= 0. Let
x = x0 + x1 ∈ (X0 , X1 )θ,p . Then T x = T x0 + T x1 , T x0 ∈ Y0 , and T x1 ∈ Y1 , so for t ∈ R+

KY (T x, t) ≤ kT x0 kY0 + t kT x1 kY1 ≤ kT kL(X0 ;Y0 ) kx0 kX0 + t kT kL(X1 ;Y1 ) kx1 kX1
!
t kT kL(X1 ;Y1 )
= kT kL(X0 ;Y0 ) kx0 kX0 + kx1 kX1 . (2.2.49)
kT kL(X0 ;Y0 )

This holds for all such decompositions of x, and thus

KY (T x, t) ≤ kT kL(X0 ;Y0 ) KX (x, t kT kL(X1 ;Y1 ) / kT kL(X0 ;Y0 ) ) (2.2.50)

for t ∈ R+ .
Let µ denote the measure dt/t on R+ and f : R+ → R be measurable. For λ ∈ R+ write
fλ (t) = f (λt). Then we may compute

(·)−θ fλ (·) Lpµ

= λθ (·)−θ f (·) Lpµ
. (2.2.51)

Combining (2.2.50) and (2.2.51), we then find that

!θ
kT kL(X1 ;Y1 ) 1−θ
kT xk(Y0 ,Y1 )θ,p ≤ kT kL(X0 ;Y0 ) kxk(X0 ,X1 )θ,p = kT kL(X 0 ;Y0 )
kT kθL(X1 ;Y1 ) kxk(X0 ,X1 )θ,p ,
kT kL(X0 ;Y0 )
(2.2.52)
and this estimate completes the proof.
In practice, we often verify the hypotheses of Theorem 2.2.12 with the help of the following
lemma.

Lemma 2.2.13. Let X0 and X1 be compatible Banach spaces and Y0 and Y1 be compatible Banach
spaces over the same field. Suppose that for i ∈ {0, 1} we have Ti ∈ L(Xi ; Yi ) such that T0 = T1 on
X0 ∩ X1 . Then there exists a unique T ∈ L(X0 + X1 ; Y0 + Y1 ) such that T |Xi = Ti for i ∈ {0, 1}.
Moreover, we have the bound

kT kL(X0 +X1 ;Y0 +Y1 ) ≤ max{kT kL(X0 ;Y0 ) , kT kL(X1 ;Y1 ) }. (2.2.53)

68
Proof. Suppose that x = x0 + x1 = w0 + w1 for x0 , w0 ∈ X0 and x1 , w1 ∈ X1 . Then x0 − w0 =
w1 − x1 ∈ X0 ∩ X1 and so T0 (x0 − w0 ) = T1 (w1 − x1 ) by hypothesis. Upon rearranging, this implies
that T0 (x0 )+T1 (x1 ) = T0 (w0 )+T1 (w1 ). From this we deduce that the mapping T : X0 +X1 → Y0 +Y1
defined by T (x0 + x1 ) = T0 (x0 ) + T1 (x1 ) is well-defined and linear. This map is bounded since if
x = x0 + x1 , then


kT xkY0 +Y1 ≤ kT0 x0 kY0 + kT1 x1 kY1 ≤ max{kT kL(X0 ;Y0 ) , kT kL(X1 ;Y1 ) } kx0 kX0 + kx1 kX1 (2.2.54)

and since this holds for all such decompositions we have the bound

kT xkY0 +Y1 ≤ max{kT kL(X0 ;Y0 ) , kT kL(X1 ;Y1 ) } kxkX0 +X1 . (2.2.55)

This proves the existence of the desired T . Uniqueness follows since if S is any other such operator,
the condition S|Xi = Ti implies that Sx = Sx0 + Sx1 = T0 x0 + T1 x1 = T x for all x = x0 + x1 ∈
X0 + X1 .
It’s also possible to prove some interpolation results for nonlinear operators. We will demonstrate
the basic principle in a particularly simple case in which the nonlinear map satisfies a Lipschitz-type
condition. For more sophisticated versions see the paper of Tartar [6].

Theorem 2.2.14. Let Y0 and Y1 be compatible Banach spaces, and let X0 and X1 be Banach spaces
such that X1 ,→ X0 . Suppose that f : X0 → Y0 is such that f (X1 ) ⊆ Y1 and there exist constants
A0 , A1 ∈ R+ such that

kf (x) − f (w)kY0 ≤ A0 kx − wkX0 for all x, w ∈ X0 (2.2.56)

and
kf (x)kY1 ≤ A1 kxkX1 for all x ∈ X1 . (2.2.57)
Then for θ ∈ (0, 1) and 1 ≤ p ≤ ∞ we have that f ((X0 , X1 )θ,p ) ⊆ (Y0 , Y1 )θ,p , and

kf (x)k(Y0 ,Y1 )θ,p ≤ A1−θ θ

0 A1 kxk(X0 ,X1 )θ,p for all x ∈ (X0 , X1 )θ,p . (2.2.58)

Proof. Write KY for the K function associated to Y0 and Y1 and KX for the K function associated
to X0 and X1 . Let x ∈ (X0 , X1 )θ,p and write x = x0 + x1 for x0 ∈ X0 and x1 ∈ X1 . Then f (x1 ) ∈ Y1
and f (x) − f (x1 ) ∈ Y0 by hypothesis, so f (x) = (f (x) − f (x1 )) + f (x1 ) ∈ Y0 + Y1 . Consequently,
for t ∈ R+ we may estimate

KY (f (x), t) ≤ kf (x) − f (x1 )kY0 + t kf (x1 )kY1 ≤ A0 kx − x1 kX0 + tA1 kx1 kX1
 
tA1
= A0 kx0 kX0 + kx1 kX1 . (2.2.59)
A0

This holds for all such decompositions of x, so we deduce that

KY (f (x), t) ≤ A0 KX (x, tA1 /A0 ). (2.2.60)

Arguing as in the proof of Theorem 2.2.12, we conclude from this that

kf (x)k(Y0 ,Y1 )θ,p ≤ A01−θ Aθ1 kxk(X0 ,X1 )θ,p . (2.2.61)

69
2.2.3 Some special cases
We now turn our attention to some special cases in which we know more about the relation between
X0 and X1 . We begin with a very simple result that shows that if X0 ∩ X1 is trivial, then the
interpolation spaces are also trivial.

Proposition 2.2.15. Let X0 and X1 be compatible Banach spaces, 1 ≤ p ≤ ∞, and θ ∈ (0, 1). If
X0 ∩ X1 = {0}, then (X0 , X1 )θ,p = {0}.

Proof. We know from Theorem 2.2.2 and Corollary 2.1.14 that {0} = X0 ∩ X1 ,→ (X0 , X1 )θ,p ,→
{0} = X0 ∩ X1 .

It often occurs in practice that Xi ,→ X1−i , in which case X0 ∩ X1 = Xi and X0 + X1 = X1−i . It

is then possible to find equivalent norms on (X0 , X1 )θ,p that are useful in the sense that they only
involve strict subsets of R+ . We record this result now.

Theorem 2.2.16. Let X0 and X1 be compatible Banach spaces and 1 ≤ p ≤ ∞. Suppose that
θ ∈ (0, 1) if p < ∞ and θ ∈ [0, 1] if p = ∞. Then the following hold.

1. If X0 ,→ X1 , then X0 ,→ (X0 , X1 )θ,p ,→ X1 , and for any T ∈ R+ an equivalent norm on

(X0 , X1 )θ,p is given by the map

kxk + ∞ (t−θ K(x, t))p dt 1/p ∈ [0, ∞] if p < ∞
R

1 T t
X1 3 x 7→ (2.2.62)
kxk1 + sup t−θ K(x, t) ∈ [0, ∞] if p = ∞.
t>T

2. If X1 ,→ X0 , then X1 ,→ (X0 , X1 )θ,p ,→ X0 , and for any T ∈ R+ an equivalent norm on

(X0 , X1 )θ,p is given by the map
 R 1/p
kxk + T (t−θ K(x, t))p dt

∈ [0, ∞] if p < ∞
0 0 t
X0 3 x 7→ −θ
(2.2.63)
kxk0 + sup t K(x, t) ∈ [0, ∞]
 if p = ∞.
0<t<T

Proof. If X0 ,→ X1 , then X0 ∩ X1 = X0 and X0 + X1 = X1 . Similarly, if X1 ,→ X0 , then

X0 ∩ X1 = X1 and X0 + X1 = X0 . As such, we only need to prove the equivalence of the norms.
For the first item we note that since X0 + X1 = X1 ,

min{1, t} kxk1  min{1, t} kxkX0 +X1 ≤ K(x, t) ≤ t kxk1 , (2.2.64)

and so for 0 < t < T we have that

K(x, t)  t kxk1 . (2.2.65)
Consequently,
Z T 1/p
−θ p dt
(t K(x, t))  kxk1  sup t−θ K(x, t) (2.2.66)
0 t 0<t<T

and the equivalence for the first item follows.

For the second item, we instead bound

min{1, t} kxk0  min{1, t} kxkX0 +X1 ≤ K(x, t) ≤ kxk0 (2.2.67)

70
to see that for T ≤ t < ∞,
K(x, t)  kxk0 . (2.2.68)
Consequently,
Z ∞ 1/p
−θ p dt
(t K(x, t))  kxk0  sup t−θ K(x, t) (2.2.69)
T t T <t

and the equivalence for the second item follows.

As an immediate consequence of this new equivalence, we find a number of interesting embed-
dings.

Corollary 2.2.17. Let X0 and X1 be Banach spaces and 0 < θ0 < θ1 < 1. Then the following hold.

1. If X0 ,→ X1 , then we have the embeddings (X0 , X1 )θ0 ,∞ ,→ (X0 , X1 )θ1 ,1 , (X0 , X1 )0,∞ ,→
(X0 , X1 )θ0 ,1 , and (X0 , X1 )θ1 ,∞ ,→ X1 = X̄0 ,→ (X0 , X1 )1,∞ = X1 .

2. If X1 ,→ X0 , then we have the embeddings (X0 , X1 )θ1 ,∞ ,→ (X0 , X1 )θ0 ,1 , (X0 , X1 )1,∞ ,→
(X0 , X1 )θ1 ,1 , and (X0 , X1 )θ0 ,∞ ,→ X0 = X̄1 ,→ (X0 , X1 )0,∞ = X0 .

Proof. In light of Corollary 2.1.14 and Theorem 2.2.16, applied with T = 1, it suffices to observe
that Z ∞ θ0 −θ1
t dt 1 1 1
< ∞, θ1 ≤ 1, and ≤ θ1 for t > 1 (2.2.70)
1 t t t t
for the first item, and
Z 1 θ1 −θ0
t dt 1 1 1
< ∞, θ
≤ , and 1 ≤ θ0 for 0 < t < 1 (2.2.71)
0 t t 1 t t
for the second item.
It’s convenient to again organize what we know in a diagram. If X0 ,→ X1 , then we have the
following zig-zag embedding diagram with 0 < θ0 < θ1 < 1.

X0 = X0 (X0 , X1 )0,∞

(X0 , X1 )θ0 ,1 (X0 , X1 )θ0 ,p (X0 , X1 )θ0 ,∞

(2.2.72)
(X0 , X1 )θ1 ,1 (X0 , X1 )θ1 ,p (X0 , X1 )θ0 ,∞

X1 = X̄0 X1 = (X0 , X1 )1,∞

On the other hand, if X1 ,→ X0 , then we have the following zig-zag embedding diagram with

71
0 < θ0 < θ1 < 1.

X0 = X̄1 X0 = (X0 , X1 )0,∞

(X0 , X1 )θ0 ,1 (X0 , X1 )θ0 ,p (X0 , X1 )θ0 ,∞

(2.2.73)

(X0 , X1 )θ1 ,1 (X0 , X1 )θ1 ,p (X0 , X1 )θ0 ,∞

X1 = X1 (X0 , X1 )1,∞

These diagrams highlight the interesting fact that if Xi ,→ X1−i , then (X0 , X1 )i,∞ is the smallest of
the interpolation spaces we have constructed, and (X0 , X1 )1−i,∞ is the largest.

2.3 Further properties of interpolation spaces

We now aim to derive some more properties of our abstract interpolation spaces.

2.3.1 Equivalent norms

So far we have not really employed the J function in our analysis of the interpolation spaces. It turns
out to play a role through the following theorem, which establishes the existence of an equivalent
norm on (X0 , X1 )θ,p that utilizes the J function. In the theorem we will also establish equivalence
with a discretized version of our previous norm. Roughly speaking, the idea is that we can return
to the context of Theorem 2.2.2 and exploit the monotonicity properties
−θ 0
P of K(x, ·) to switch from
the measure µ = dt/t and weight w(t) = t to the measure µ = n∈Z δrn for some r > 1, but
with the same weight.
Theorem 2.3.1. Let X0 and X1 be compatible Banach spaces, 1 ≤ p ≤ ∞, and θ ∈ (0, 1). Fix
1 < r < ∞. Then there exist constants C0 , C1 , C2 , C3 ∈ R+ such that

C0 kxkθ,p ≤ C1 {r−θn K(x, rn )}n∈Z `p

X
−θn n
≤ inf{ {r J(xn , r )}n∈Z `p
|x= xn for {xn }n∈Z ⊆ X0 ∩ X1 }
n∈Z

≤ C2 {r−θn K(x, rn )}n∈Z `p

≤ C3 kxkθ,p (2.3.1)

for all x ∈ X0 + X1 . Consequently, all three of these quantities define equivalent norms on
(X0 , X1 )θ,p .
Proof. We will first prove the existence of constants A0 , A1 ∈ R+ such that

A0 kxkθ,p ≤ {r−θn K(x, rn )}n∈Z `p

≤ A1 kxkθ,p (2.3.2)

for all x ∈ X0 + X1 . Suppose initially that p < ∞. Since r > 1 we can write
Z XZ rn+1
p dt dt
(K(x, t)) = (K(x, t))p . (2.3.3)
R+ t1+θp n∈Z rn t1+θp

72
For each n ∈ Z we can also compute
Z rn+1
1 rθp − 1 rθp − 1
 
dt 1 1 1 1
= − = = . (2.3.4)
rn t1+θp θp rnθp r(n+1)θp rnθp θprθp r(n+1)θp θp
From these and the fact that K(x, ·) is nondecreasing, we deduce that
p p
rθp − 1 X K(x, rn ) rθp − 1 X K(x, rn+1 )
 Z 
p dt
≤ (K(x, t)) ≤
θprθp n∈Z rnθ R+ t1+θp θp n∈Z r(n+1)θ
p
rθp − 1 X K(x, rn )

= . (2.3.5)
θp n∈Z rnθ

This proves (2.3.2) when p < ∞. In the case p = ∞ we note that

K(x, t) K(x, t) K(x, rn ) K(x, t) K(x, rn+1 )
sup θ
= sup sup θ
and (n+1)θ ≤ θ
≤ nθ
for rn < t ≤ rn+1
t>0 t n∈Z rn <t≤rn+1 t r t r
(2.3.6)
and argue similarly. Thus, (2.3.2) is proved.
Next we will prove the existence of constants B0 , B1 ∈ R+ such that

B0 {r−θn K(x, rn )}n∈Z `p

X
−θn
≤ inf{ {r J(xn , rn )}n∈Z `p
|x= xn for {xn }n∈Z ⊆ X0 ∩ X1 }
n∈Z

≤ B1 {r−θn K(x, rn )}n∈Z `p

(2.3.7)

for all x ∈ X0 + X1 , and once this is established the proof is complete.

Suppose initially that the infimum
P in the middle of (2.3.7) is finite, and pick any sequence
{xn }n∈Z ⊆ X0 ∩ X1 such that x = n∈Z xn . For n ∈ Z we can use Propositions 2.1.16 and 2.1.18
to estimate
X X
r−nθ K(x, rn ) ≤ r−nθ K(xm , rn ) ≤ r−nθ min{1, rn−m }J(xm , rm )
m∈Z m∈Z
X
= min{1, rn−m }r−(n−m)θ r−mθ J(xm , rm ). (2.3.8)
m∈Z

Define the sequences k = {kn }n∈Z , d = {dn }n∈Z , j = {jn }n∈Z ⊂ R+ via

kn = r−nθ K(x, rn ), dn = min{1, rn }r−nθ , and jn = r−nθ J(xn , rn ), (2.3.9)

and note that X X X

kdk`1 = dn = r−nθ + rn(1−θ) < ∞. (2.3.10)
n∈Z 0≤n n<0

Then the (2.3.8) implies that

kn ≤ (d ∗ j)n for all n ∈ Z, (2.3.11)
and so Young’s inequality (see Remark 1.3.9) implies that

{r−θn K(x, rn )}n∈Z `p

= kkk`p ≤ kd ∗ jk`p ≤ kdk`1 kjk`p = kdk`1 {r−θn J(xn , rn )}n∈Z `p
.
(2.3.12)

73
This holds for all such decompositions of x, and hence
X
{r−θn K(x, rn )}n∈Z `p
≤ kdk`1 inf{ {r−θn J(xn , rn )}n∈Z `p
|x= xn for {xn }n∈Z ⊆ X0 ∩ X1 }.
n∈Z
(2.3.13)
−θn n
Now suppose that {r K(x, r )}n∈Z `p < ∞. According to (2.3.2), we then know that x ∈
(X0 , X1 )θ,p ,→ X0 ∩ X1 , and so thePthird item of Theorem 2.1.19 allows us to choose a sequence
{wn }n∈Z ⊆ X0 ∩ X1 such that x = n∈Z wn and J(wn , rn ) ≤ 2(1 + r)K(x, rn ) for n ∈ Z. Then
X
inf{ {r−θn J(xn , rn )}n∈Z `p
|x= xn for {xn }n∈Z ⊆ X0 ∩ X1 }
n∈Z
−θn
≤ {r J(wn , rn )}n∈Z `p
≤ 2(1 + r) {r−θn K(x, rn )}n∈Z `p
, (2.3.14)

and together with (2.3.13) this proves (2.3.7).

Of course, the reason we care about the new equivalent norms is that they allow us to prove
things that were out of reach with the original norm. To demonstrate what we can do with the J
formulation we now record an amusing embedding result that complements that of Theorem 2.2.11.

Theorem 2.3.2. Let X0 and X1 be compatible Banach spaces and 0 < θ0 < θ < θ1 < 1. Then

(X0 , X1 )θ,∞ ,→ (X0 , X1 )θ0 ,1 + (X0 , X1 )θ1 ,1 . (2.3.15)

P P
P x ∈ (X0 , X1 )θ,∞
Proof. Let and write x = n∈Z xn for {xn }n∈Z ⊆ X0 ∩ X1 . Define y = m<0 xm
and z = m≥0 xm . Then
X X
2−nθ0 K(y, 2n ) ≤ 2−nθ0 K(xm , 2n ) ≤ 2−nθ0 min{1, 2n−m }J(xm , 2m )
m<0 m<0
J(xm , 2m ) X −nθ0
≤ sup mθ
2 min{1, 2n−m }2mθ , (2.3.16)
m∈Z 2 m<0

and X X
2−nθ0 min{1, 2n−m }2mθ ≤ 2−(n−m)θ0 min{1, 2n−m }2m(θ−θ0 ) . (2.3.17)
m<0 m<0

Hence,
2−nθ0 K(y, 2n ) ≤ (d ∗ e)n for n ∈ Z, (2.3.18)
where
dn = 2−nθ0 min{1, 2n } and en = 2n(θ−θ0 ) χ(−∞,0) (n). (2.3.19)
Since d, e ∈ `1 (Z) we may use Young’s theorem to bound

J(xm , 2m )
{2−nθ0 K(y, 2n )}n∈Z `1
≤ kdk`1 kek`1 sup , (2.3.20)
m∈Z 2mθ

and so Theorem 2.3.1 allows us to pick a constant C0 ∈ R+ such that

kykθ0 ,1 ≤ C0 kxkθ,∞ . (2.3.21)

74
 On the other hand,
X X
2−nθ1 K(z, 2n ) ≤ 2−nθ1 K(xm , 2n ) ≤ 2−nθ1 min{1, 2n−m }J(xm , 2m )
m≥0 m≥0
J(xm , 2m ) X −nθ0
≤ sup mθ
2 min{1, 2n−m }2mθ , (2.3.22)
m∈Z 2 m≥0

and X X
2−nθ1 min{1, 2n−m }2mθ = 2−(n−m)θ1 min{1, 2n−m }2m(θ−θ1 ) . (2.3.23)
m≥0 m≥0

Hence,
2−nθ1 K(y, 2n ) ≤ (f ∗ g)n for n ∈ Z, (2.3.24)
where
fn = 2−nθ1 min{1, 2n } and gn = 2n(θ−θ1 ) χ[0,∞) (n). (2.3.25)
Then f, g ∈ `1 (Z), so again we can use Young’s inequality and Theorem 2.3.1 to pick a constant
C1 ∈ R+ such that
kzkθ1 ,1 ≤ C1 kxkθ,∞ . (2.3.26)
We now know that x = y + z with y ∈ (X0 , X1 )θ0 ,1 and z ∈ (X0 , X1 )θ1 ,1 . Moreover, (2.3.21) and
(2.3.26) show that
kxk(X0 ,X1 )θ +(X0 ,X1 )θ1 ,1 ≤ kykθ0 ,1 + kzkθ1 ,1 ≤ max{C0 , C1 } kxkθ,∞ . (2.3.27)
0 ,1

This proves the asserted embedding.

We can summarize the results in Theorems 2.2.11 and 2.3.2 in the following diagram, in which
0 < θ0 < θ < θ1 < 1.

(X0 , X1 )θ0 ,1 (X0 , X1 )θ0 ,∞

(X0 , X1 )θ0 ,∞ ∩ (X0 , X1 )θ1 ,∞ (X0 , X1 )θ,1 (X0 , X1 )θ,∞ (X0 , X1 )θ0 ,1 + (X0 , X1 )θ1 ,1

(X0 , X1 )θ1 ,1 (X0 , X1 )θ1 ,∞

(2.3.28)
As another use of our new norm, we prove that X0 ∩ X1 is dense in (X0 , X1 )θ,p for 1 ≤ p < ∞
and θ ∈ (0, 1).
Theorem 2.3.3. Let X0 and X1 be compatible Banach spaces, 1 ≤ p < ∞, and θ ∈ (0, 1). Then
X0 ∩ X1 is dense in (X0 , X1 )θ,p .
Proof. Let x ∈ (X0 , X1 )θ,p and pick
P 1 < r < ∞. According to Theorem 2.3.1, we can choose
{xn }n∈Z ⊆ X0 ∩ X1 such that x = n∈Z xn (convergence in X0 + X1 ) and
!1/p
X
(r−θn J(xn , rn ))p < ∞. (2.3.29)
n∈Z

75
Let ε > 0 and choose N ∈ N such that
 1/p
X
 (r−θn J(xn , rn ))p  < C0 ε, (2.3.30)
|n|≥N

P P
where C0 ∈ R+ is the constant from Theorem 2.3.1. Then |n|<N xn ∈ X0 ∩X1 and x− |n|<N xn =
P
|n|≥N xn , with this series again converging in X0 + X1 , and so Theorem 2.3.1 implies that

 1/p
X X
C0 x − xn ≤ (r−θn J(xn , rn ))p  < C0 ε. (2.3.31)
|n|<N |n|≥N
θ,p

Since ε > 0 was arbitrary, we conclude that X0 ∩ X1 is dense in (X0 , X1 )θ,p .

2.3.2 Reiteration
For a pair of compatible Banach spaces X0 and X1 , we now know that (X0 , X1 )θ0 ,p0 and (X0 , X1 )θ1 ,p1
are also compatible spaces. A natural question then arises: what happens if we interpolate between
these new spaces? Do we get something new, or do we end up with another interpolation space? In
order to answer this question we first need to introduce some machinery, starting with the following
simple result.

Proposition 2.3.4. Let X0 and X1 be compatible Banach spaces, 1 ≤ p ≤ ∞, and θ ∈ (0, 1). There
exists a constant C ∈ R+ such that

kxkθ,p ≤ C inf t−θ J(x, t) for all x ∈ X0 ∩ X1 . (2.3.32)

t>0
P
Proof. Suppose x ∈ X0 ∩ X1 . For any m ∈ Z we have that x = n∈Z xn , where xn = x if n = m
and xn = 0 otherwise, so Theorem 2.3.1 with r = 2 implies that

J(x, 2m )
C0 kxkθ,p ≤ {2−θn J(xn , 2n )}n∈Z `p
= . (2.3.33)
2θm
Consequently,
J(x, 2m )
C0 kxkθ,p ≤ inf . (2.3.34)
m∈Z 2θm
However, for 2n ≤ t ≤ 2n+1 we have that

1 J(x, 2m ) J(x, 2n ) J(x, t)

θ
inf θm
≤ (n+1)θ
≤ , (2.3.35)
2 m∈Z 2 2 tθ
from which we deduce that
2θ J(x, t)
kxkθ,p ≤ inf . (2.3.36)
C0 t>0 tθ

In light of the second item of Theorem 2.2.8 and the estimate of Proposition 2.3.4, we are led
to introduce the following idea.

76
Definition 2.3.5. Let X0 and X1 be compatible Banach spaces and X be intermediate to X0 and
X1 . Let θ ∈ [0, 1].

1. We say that X is of K−type θ if there exists a constant C > 0 such that

sup t−θ K(x, t) ≤ C kxkX for all x ∈ X. (2.3.37)

t>0

2. We say that X is of J−type θ if there exists a constant C > 0 such that

kxkX ≤ C inf t−θ J(x, t) for all x ∈ X0 ∩ X1 . (2.3.38)

t>0

3. We say that X is of type θ if X is of K−type θ and of J−type θ.

Let’s consider a simple example.

Example 2.3.6. Let X0 and X1 be compatible Banach spaces. If x ∈ X0 +X1 , then K(x, t) ≤ kxk0
for all t ∈ R+ , and so X0 and X0 are of K−type 0. Similarly, if x ∈ X0 ∩ X1 , then kxk0 ≤ J(x, t)
for all t ∈ R+ , so X0 and X0 are of J−type 0. Consequently, X0 and X0 are of type 0.
On the other hand, K(x, t) ≤ t kxk1 for all x ∈ X0 + X1 and t ∈ R+ , so X1 and X̄1 are of
K−type 1. If x ∈ X0 ∩ X1 then t kxk1 ≤ J(x, t) for all t ∈ R+ , and so X1 and X1 are of J−type 1
as well. Thus, X1 is of type 1.
Let θ ∈ (0, 1) and 1 ≤ p ≤ ∞. We know from Theorem 2.2.8 and Proposition 2.3.4 that
(X0 , X1 )θ,p is of type θ. 4

It turns out that we can exactly characterize which spaces are of J−type and K−type θ in terms
of embeddings with familiar spaces. We record this now.

Proposition 2.3.7. Let X0 and X1 be compatible Banach spaces and X be intermediate to X0 and
X1 . Then the following hold for every θ ∈ [0, 1].

1. X is of K−type θ if and only if X ,→ (X0 , X1 )θ,∞ .

2. X is of J−type θ if and only if

(
(X0 , X1 )θ,1 ,→ X if θ ∈ (0, 1)
(2.3.39)
Xθ ,→ X if θ ∈ {0, 1}.

3. The following are equivalent for θ ∈ (0, 1).

(a) X is of type θ.
(b) (X0 , X1 )θ,1 ,→ X ,→ (X0 , X1 )θ,∞ .
(c) X is intermediate to (X0 , X1 )θ,1 and (X0 , X1 )θ,∞ .

4. The following are equivalent for θ ∈ {0, 1}.

(a) X is of type θ.
(b) Xθ ,→ X ,→ (X0 , X1 )θ,∞ .
(c) X is intermediate to Xθ and (X0 , X1 )θ,∞ .

77
Proof. The first item follows directly from the definition of K−type, and the third and fourth items
follow from the first and second together with Theorem 2.2.8, so we only need to prove the second.
Suppose initially that θ ∈ (0, 1). If (X0 , X1 )θ,1 ,→ X, then there is a constant C > 0 such that
kxkX ≤ C kxkθ,p for all x ∈ X0 + X1 , and so X is of J−type θ byPProposition 2.3.4. Conversely,
suppose that X is of J−type θ. Let x ∈ (X0 , X1 )θ,1 , and write x = n∈Z xn for {xn }n∈Z ⊆ X0 ∩ X1 .
Then
X X J(xn , 2n )
kxn kX ≤ C = C kxkθ,1 , (2.3.40)
n∈Z n∈Z
2nθ
P
and so Theorem 2.1.1 shows that x = n∈Z xn with the series converging in X; moreover,
X
kxkX ≤ kxn kX ≤ C kxkθ,1 . (2.3.41)
n∈Z

Hence, (X0 , X1 )θ,1 ,→ X, and the second item is proved in the case θ ∈ (0, 1).
Now assume that θ ∈ {0, 1}. Suppose Xθ ,→ X. Then there exists a constant C > 0 such that
kxkX ≤ C kxkθ since the norm on the space Xθ is precisely k·k0 . Thus, for x ∈ X0 ∩ X1 we have that
kxkX ≤ C kxkθ ≤ Ct−θ J(x, t) for all t ∈ R+ , and we deduce that X is of J−type θ. Conversely,
suppose that X is of J−type θ, so that there exists a constant C > 0 such that kxkX ≤ Ct−θ J(x, t)
for all t ∈ R+ and x ∈ X0 ∩ X1 . In particular, for a fixed x ∈ X0 ∩ X1 we an send t → 0 if
θ = 0 and t → ∞ if θ = 1 to see deduce from this that kxkX ≤ C kxkθ . Now let x ∈ Xθ and
pick a sequence {xn }∞ n=` ⊆ X0 ∩ X1 such that xn → x in Xθ . For m, n ≥ ` we then have that
kxn − xm kX ≤ C kxn − xm kθ , which implies that {xn }∞n=` is Cauchy in X, and hence convergent in
X to x (thanks to the compatibility of X0 and X1 ). Hence,

kxkX = lim kxn kX ≤ C lim kxn kθ = C kxkθ for all x ∈ Xθ , (2.3.42)

n→∞ n→∞

and we deduce that Xθ ,→ X. This completes the proof of the second item when θ ∈ {0, 1}.
We can make a variant of the diagram (2.2.40) to indicate how to think about the location of
the spaces of type θ ∈ [0, 1] within the collection of interpolation spaces. In the following diagram
we again write 0 < θ0 < θ < θ1 < 1 and indicate a generic space X(ψ) as a space of type ψ, which
must lie in the indicated position along the horizontal lines.

X0 X(0) (X0 , X1 )0,∞ X̄0

(X0 , X1 )θ0 ,1 X(θ0 ) (X0 , X1 )θ0 ,∞

(X0 , X1 )θ,1 X(θ) (X0 , X1 )θ,∞ X0 ∩ X1 X 0 + X1 (2.3.43)

(X0 , X1 )θ1 ,1 X(θ1 ) (X0 , X1 )θ1 ,∞

X1 X(1) (X0 , X1 )1,∞ X̄1

The notion of type is exactly what we need to answer the question raised above. We now state
the answer as the important “reiteration theorem.”

78
Theorem 2.3.8 (Reiteration theorem). Let X0 and X1 be compatible Banach spaces and Y0 and
Y1 be intermediate spaces to X0 and X1 . Suppose that 0 ≤ θ0 < θ1 ≤ 1, 1 ≤ p ≤ ∞, σ ∈ (0, 1), and
θ = (1 − σ)θ0 + σθ1 ∈ (0, 1). Then the following hold.

1. If Yi is of K−type θi for i ∈ {0, 1}, then

(Y0 , Y1 )σ,p ,→ (X0 , X1 )θ,p . (2.3.44)

2. If Yi is of J−type θi for i ∈ {0, 1}, then

(X0 , X1 )θ,p ,→ (Y0 , Y1 )σ,p . (2.3.45)

3. If Yi is of type θi for i ∈ {0, 1}, then

(X0 , X1 )θ,p = (Y0 , Y1 )σ,p (2.3.46)

with equivalence of norms on these spaces.

Proof. First note that the first and second items imply the third, so we must only prove these two.
In order to keep the association between the K and J functions and the pair of spaces clear, we
will write

KX (x, t) = inf{kxkX0 + t kxkX1 | x = x0 + x1 }, JX (x, t) = max{kxkX0 , t kxkX1 } (2.3.47)

and

KY (x, t) = inf{kxkY0 + t kxkY1 | x = x0 + x1 }, JY (x, t) = max{kxkY0 , t kxkY1 }. (2.3.48)

Note also that σ satisfies

θ − θ0
σ= . (2.3.49)
θ1 − θ0
We begin with the proof of the first item. Suppose that x ∈ (Y0 , Y1 )σ,p ,→ Y0 + Y1 and let
x = y0 + y1 for yi ∈ Yi . By hypothesis, there is a constant C > 0 such that

KX (y, t) ≤ Ctθi kykYi for all y ∈ X0 + X1 and t ∈ R+ . (2.3.50)

Thus,

KX (x, t) ≤ KX (y0 , t) + KX (y1 , t) ≤ Ctθ0 ky0 kY0 + Ctθ1 ky1 kY1 = Ctθ0 ky0 kY0 + tθ1 −θ0 ky1 kY1 ,

(2.3.51)
and since this holds for all such decompositions, we find that

KX (x, t) ≤ Ctθ0 KY (x, tθ1 −θ0 ). (2.3.52)

If p = ∞, then this and (2.3.49) imply that

KX (x, t) KY (x, t)
sup θ
≤ C sup tθ0 −θ KY (x, tθ1 −θ0 ) = C sup , (2.3.53)
t>0 t t>0 t>0 tσ

79
while if p < ∞ this, (2.3.49), and a change of variables show that
Z 1/p Z 1/p
−θ dt θ0 −θ θ1 −θ0 dt
(t KX (x, t)) ≤C (t KY (x, t ))
R+ dt R+ dt
Z 1/p
C −σ dt
= (t KY (x, t)) . (2.3.54)
(θ1 − θ0 )1/p R+ dt

We deduce from these that (Y0 , Y1 )σ,p ,→ (X0 , X1 )θ,p , which proves the first item.
P We now turn to the proof of the second item. Suppose that x ∈ (X0 , X1 )θ,p and write x =
n∈Z xn for {xn }n∈Z ⊆ X0 ∩ X1 and the series converging in X0 + X1 . Then by Proposition 2.1.18,
X X
KY (x, 2(θ1 −θ0 )n ) ≤ KY (xm , 2(θ1 −θ0 )n ) ≤ min{1, 2(θ1 −θ0 )(n−m) }JY (xm , 2(θ1 −θ0 )m ). (2.3.55)
m∈Z m∈Z

For each m ∈ Z we can use the fact that Yi is of J−type θi to bound

JY (xm , 2(θ1 −θ0 )m ) = max{kxm kY0 , 2(θ1 −θ0 )m kxm kY1 } ≤ C2−mθ0 JX (xm , 2m ), (2.3.56)

and we can use (2.3.49) to compute

2−(θ1 −θ0 )σn min{1, 2(θ1 −θ0 )(n−m) }2−mθ0 = 2−(θ−θ0 )n min{1, 2(θ1 −θ0 )(n−m) }2−mθ0
= 2−θ(n−m) min{2θ0 (n−m) , 2θ1 (n−m) }2−mθ . (2.3.57)

Thus, we have the estimate

X
2−(θ1 −θ0 )σn KY (x, 2(θ1 −θ0 )n ) ≤ C dn−m 2−mθ JX (xm , 2m ) (2.3.58)
m∈Z

for
dn = 2−θn min{2θ0 n , 2θ1 n }. (2.3.59)
Since X X
kdk`1 = 2(θ1 −θ)n + 2−n(θ−θ0 ) < ∞ (2.3.60)
n≤0 0<n

we can employ Young’s inequality and Theorem 2.3.1 to bound

kxk(Y0 ,Y1 )σ,p ≤ C kdk`1 {2−mθ JX (xm , 2m )}m∈Z `p

. (2.3.61)

This holds for all such decompositions of x, and so again Theorem 2.3.1 shows that

kxk(Y0 ,Y1 )σ,p ≤ C kdk`1 kxk(X0 ,X1 )θ,p . (2.3.62)

We deduce from this that (X0 , X1 )θ,p ,→ (Y0 , Y1 )σ,p , which completes the proof of the second item.

2.4 Examples and applications

We now turn our attention to characterizing the interpolation spaces that arise from various natural
choices of X0 and X1 .

80
2.4.1 Interpolation of Lebesgue spaces
We motivated our construction of the abstract interpolation spaces by examining L1 (X; F) and
L∞ (X; F) for a given measure space (X, M, µ), and so it is natural to begin by studying what
results when we use these in our abstract framework. The key idea is contained in the following
beautiful theorem, which relates the decreasing rearrangement to the K function.

Theorem 2.4.1. Let (X, M, µ) be a measure space and f ∈ L0 (X; F). Then the following are
equivalent.

1. f ∈ L1 (X; F) + L∞ (X; F).

Rt
2. There exists t ∈ R+ such that 0
f # (s)ds < ∞.
Rt
3. For each t ∈ R+ we have that 0
f # (s)ds < ∞.

Moreover, if any of these (and hence all) holds, we have that

Z t
K(f, t) = f # (s)ds for all t ∈ R+ . (2.4.1)
0

Proof. The third item trivially implies the second. If the second holds for some t, then by the
#
monotonicity
R r # of f Rwe must have that f # (t) < ∞ or else the integral would be infinite. If 0 < r ≤ t,
t #
then 0 f (s)ds ≤ 0 f (s)ds < ∞. On the other hand, if t < r < ∞, then
Z r Z t Z r Z t
# # #
f (s)ds = f (s)ds + f (s)ds ≤ f # (s)ds + (r − t)f # (t) < ∞. (2.4.2)
0 0 t 0

Thus, the second item implies the third.

Suppose now that the first item holds. Then K(f, t) < ∞ for every t ∈ R+ by Proposition
2.1.16. Let f = g + h for g ∈ L1 (X; F) and h ∈ L∞ (X; F). Then for s ∈ R+ and 0 < ε < 1 we may
may use Proposition 1.1.21 to bound f # (s) ≤ g # ((1 − ε)s) + h# (εs). Thus, for t ∈ R+ we have the
estimate
Z t Z t Z t Z ∞
# # # 1
f (s)ds ≤ g ((1 − ε)s)ds + h (εs)ds ≤ g # (s)ds + th# (0)
0 0 0 1−ε 0
1
= kgkL1 + t khkL∞ . (2.4.3)
1−ε
This holds for all 0 < ε < 1 and all such decompositions, so we deduce that
Z t
f # (s)ds ≤ K(f, t). (2.4.4)
0

Thus, the first item implies the third.

Now suppose the third item holds and let t ∈ R+ . From the above analysis we know that
f # (t) < ∞. Define g, h ∈ L0 (X; F) via
(
f − f # (t) |ff | if |f | > f # (t)
g= (2.4.5)
0 if |f | ≤ f # (t)

81
and (
f # (t) |ff | if |f | > f # (t)
h=f −g = (2.4.6)
f if |f | ≤ f # (t).
Define the set E = {x ∈ X | g(x) 6= 0} = {x ∈ X | |f (x)| > f # (t)} and note that if x ∈ E, then
|f (x)| > 0 and |g(x)| = |f (x)| − f # (t). Proposition 1.1.21 then provides the estimate

µ(E) = df (f # (t)) ≤ t. (2.4.7)

If there exists s such that df (f # (t)) < s < t, then we can use Proposition 1.1.21 again to see that

f # (t) ≤ f # (s) ≤ f # (df (f # (t))) ≤ f # (t). (2.4.8)

Hence, f # (s) = f # (t) for all s ∈ [µ(E), t]. We now have enough information to estimate kgkL1 and
khkL∞ . Indeed,
Z Z Z Z ∞
# #
kgkL1 = |g| dµ = (|f |−f (t))dµ = |f | χE dµ−µ(E)f (t) = (f χE )# (s)ds−µ(E)f # (t)
E E X 0
(2.4.9)
but Lemma 1.1.24 implies that (f χE )# (s) ≤ f # (s)χ(0,µ(E)) (s) for s ∈ R+ , and so
Z µ(E)
kgkL1 ≤ f # (s)ds − µ(E)f # (t). (2.4.10)
0

On the other hand,

Z t
# # # #
t khkL∞ ≤ tf (t) = µ(E)f (t) + (t − µ(E))f (t) = µ(E)f (t) + f # (t)ds
µ(E)
Z t
#
= µ(E)f (t) + f # (s)ds. (2.4.11)
µ(E)

Thus, f ∈ L1 (X; F) + L∞ (X; F) and

Z t
K(f, t) ≤ kgkL1 + t khkL∞ ≤ f # (s)ds. (2.4.12)
0

This completes the proofR that the third item implies the first and that if any of the three items hold
t
we have that K(f, t) = 0 f # (s)ds for all t ∈ R+ .
With the previous theorem in hand, we can now characterize the Lorentz spaces as the abstract
interpolation spaces generated by interpolating between L1 (X; F) and L∞ (X; F).
Theorem 2.4.2. Let (X, M, µ) be a measure space, 1 < p < ∞, and 1 ≤ q ≤ ∞. Then there exists
a constant C > 0 such that
|||f |||Lp,q ≤ kf k1/p0 ,q ≤ C |||f |||Lp,q (2.4.13)
for every f ∈ L0 (X; F). Consequently, we have the algebraic and topological identity

(L1 (X; F), L∞ (X; F))1/p0 ,q = Lp,q (X; F), (2.4.14)

which in particular means that the Lorentz space Lp,q (X; F) admits a norm that generates the same
topology as the quasinorm and makes the space Banach.

82
Proof. Suppose that f ∈ Lp,q (X; F). Then f ∈ Lp,∞ (X; F) and so f # (s) ≤ s−1/p |||f |||Lp,∞ , which
Rt
means that 0 f # (s)ds < ∞ for every t ∈ R+ . Then the previous theorem implies that
Z t
1/p−1 1/p−1
t K(f, t) = t f # (s)ds < ∞ for all t ∈ R+ , (2.4.15)
0

and we can use this and Hardy’s inequality when q < ∞ to deduce that
kf k1/p0 ,q ≤ C |||f |||Lp,q (2.4.16)

for a constant C = C(p, q) > 0. Thus, Lp,q (X; F) ,→ (L1 (X; F), L∞ (X; F))1/p0 ,q .
Conversely, suppose f ∈ (L1 (X; F), L∞ (X; F))1/p0 ,q . Then f ∈ L1 (X; F) + L∞ (X; F), and so the
previous theorem and the fact that f # is nondecreasing imply that
Z t
1/p # 1/p−1
t f (t) ≤ t f # (s)ds = t1/p−1 K(f, t) < ∞ for all t ∈ R+ . (2.4.17)
0

Thus,
|||f |||Lp,q ≤ kf k1/p0 ,q , (2.4.18)
and we deduce that (L1 (X; F), L∞ (X; F))1/p0 ,q ,→ Lp,q (X; F).
Remark 2.4.3. Note that the theorem does apply to the spaces L1,q (X; F) for 1 ≤ q ≤ ∞. In fact, it
can be shows that when q > 1 these spaces are not normable, and we are stuck with the quasi-norm.
As a consequence of this theorem and the reiteration theorem, we can interpolate between
Lorentz spaces as well.
Corollary 2.4.4. Let (X, M, µ) be a measure space, 1 < p0 , p1 < ∞, and 1 ≤ q0 , q1 ≤ ∞. Let
1 ≤ q ≤ ∞, θ ∈ (0, 1), and define 1 < pθ < ∞ via
1 1−θ θ
= + . (2.4.19)
pθ p0 p1
Then
(Lp0 ,q0 (X; F), Lp1 ,q1 (X; F))θ,q = Lpθ ,q (X; F) (2.4.20)
and
(Lp0 (X; F), Lp1 (X; F))θ,pθ = Lpθ (X; F). (2.4.21)
Proof. First note that
1 1−θ θ
0
= 0
+ 0. (2.4.22)
pθ p0 p1
We know from Theorem 2.4.2 that for i ∈ {0, 1}
Lpi ,qi (X; F) = (L1 (X; F), L∞ (X; F))1/p0i ,q (2.4.23)
and is thus of type 1/p0i . Then Theorems 2.3.8 and 2.4.2 combine to show that
(Lp0 ,q0 (X; F), Lp1 ,q1 (X; F))θ,q = (L1 (X; F), L∞ (X; F))1/p0θ ,q = Lpθ ,q (X; F). (2.4.24)
In particular, if we set qi = pi and q = pθ , then we find the Lebesgue interpolation result
(Lp0 (X; F), Lp1 (X; F))θ,pθ = Lpθ (X; F). (2.4.25)

83
2.4.2 Interpolating between Lp and W 1,p
We now aim to interpolate between Lp (Rn ; F) and and the Sobolev space W 1,p (Rn ; F). To do this,
we first need to recall the one definition of the Besov spaces (there are many available with different
degrees of usefulness, depending on the area of intended use - Chapter 17 of Leoni’s book [4] does
a nice job of clarifying the relations among these definitions).

Definition 2.4.5. We define the following.

1. Given a function f : Rn → F and y ∈ Rn define the difference quotient ∆y f : Rn → F via

∆y f (x) = f (x + y) − f (x). (2.4.26)

2. For 1 ≤ p ≤ ∞, 1 ≤ q ≤ ∞, and s ∈ (0, 1) the Besov space of regularity s and integrability

parameters p, q is the space

Bqs,p (Rn ; F) = {[f ] | f : Rn → F is measurable and kf kBqs,p < ∞}, (2.4.27)

where
kf kBqs,p = kf kLp + [f ]Bqs,p , (2.4.28)
and 1/q
k∆y f kqLp
Z
[f ]Bqs,p = dy <∞ (2.4.29)
Rn |y|n+sq
when 1 ≤ q < ∞ and
k∆y f kLp
s,p =
[f ]B∞ sup (2.4.30)
06=y∈Rn |y|s
when q = ∞.

The space Bqs,p (Rn ; F) is a Banach space when endowed with this norm, a fact that we leave as
an exercise to verify. Note that when p = q < ∞, we may use a change of variables to see that
1/p
|f (x) − f (y)|p
Z
[f ]Bps,p = dxdy , (2.4.31)
Rn ×Rn |x − y|n+sp

which shows that Bps,p (Rn ; F) = W s,p (Rn ; F), where the latter is the fractional Sobolev space of
regularity s and integrability p. We now prove that the Besov spaces are the interpolation spaces
between Lp and W 1,p .

Theorem 2.4.6. Now let 1 ≤ p < ∞, 1 ≤ q ≤ ∞, and θ ∈ (0, 1). Then

(Lp (Rn ; F), W 1,p (Rn ; F))θ,q = Bqθ,p (Rn ; F) (2.4.32)

with equivalence of norms.

Proof. Suppose θ,p n

Cc∞ (Rn ) be such that supp(η) ⊆ B(0, 1), η ≥ 0
R first that f ∈ Bq (R ; F). Let η ∈ −n
is radial, and Rn η = 1. For t ∈ R+ write ηt (x) = t η(x/t). Then we write

f = (f − f ∗ ηt ) + f ∗ ηt =: gt + ht . (2.4.33)

84
 Since η is radial, we can compute
Z Z
−n
gt (x) = (f (x) − f (x − y))t η(y/t)dy = (f (x) − f (x + y))t−n η(y/t)dy
Rn Rn
Z
= −∆y f (x)t−n η(y/t)dy. (2.4.34)
Rn

From this and Minkowski’s integral inequality, we have that

Z
kgt kLp ≤ k∆y f kLp t−n η(y/t)dy. (2.4.35)
Rn

If q = ∞, then

k∆y f kLp
Z Z
−θ
t kgt kLp ≤ k∆y f kLp t −n−θ
η(y/t)dy ≤ sup θ
|y/t|θ t−n η(y/t)dy
Rn y6=0 |y| Rn
Z
≤ kf kB∞
θ,p |y|θ η(y)dy = C(η) kf kB∞
θ,p . (2.4.36)
Rn
R
On the other hand, if q < ∞, then we may use the normalization η = 1 together with Hölder’s
inequality (or Jensen’s inequality) to bound
Z
q
kgt kLp ≤ k∆y f kqLp t−n η(y/t)dy, (2.4.37)
Rn

which implies, thanks to Fubini-Tonelli, that

kgt kqLp
Z ∞ Z Z ∞
q
1+θq
dt ≤ k∆y f kLp t−n−1−θq η(y/t)dtdy. (2.4.38)
0 t R n 0

Since supp(η) ⊆ B(0, 1) we then see from this that

∞
kgt kqLp ∞
k∆y f kqLp
Z Z Z Z
dt ≤ kηkL∞ k∆y f kqLp t −n−1−θq
dtdy = C(η, q, θ, n) dy
0 t1+θq Rn |y| Rn |y|n+θq
≤ C kf kqB θ,p . (2.4.39)
q

For the term ht we begin by bounding

kht kLp ≤ kf kLp (2.4.40)

R
with Young’s inequality. We know that Rn ∇η(z)dz = 0 and ∇η(z) = −∇η(−z), so we can compute
Z
∇ht (x) = −(f (x + y) − f (x))t−n−1 ∇η(y/t)dy (2.4.41)
Rn

in order to bound, again with the help of Minkowski,

Z
k∇ht kLp ≤ k∆y f kLp t−n−1 |∇η(y/t)| dy. (2.4.42)
Rn

If q = ∞, then Z
t 1−θ
k∇ht kLp ≤ kf kB∞
θ,p |y|θ |∇η(y)| dy = C(η) kf kB∞
θ,p . (2.4.43)
Rn

85
On the other hand, q < ∞, then we us Hölder’s inequality on (2.4.42) to bound
Z
q −q/q 0
k∇ht kLp ≤ Ct k∆y f kqLp t−n−1 |∇η(y/t)| dy, (2.4.44)
Rn

where we have used the computation

Z 1/q0
−n−1 0 1/q 0 0
t |∇η(y/t)| dy = t−1/q k∇ηkL1 = Ct−1/q when 1 < q. (2.4.45)
Rn

This and Fubini-Tonelli then show us that

Z ∞ Z ∞ Z
q dt −q/q 0 +q−1−θq
(t k∇ht kLp ) 1+θq ≤ C t k∆y f kqLp t−n−1 |∇η(y/t)| dydt
0 t Rn
Z Z ∞0 Z Z ∞
q −n−1−θq q
≤C k∆y f kLp t |∇η(y/t)| dtdy ≤ C k∆y f kLp t−n−1−θq dtdy
Rn 0 Rn |y|
q
k∆y f kLp
Z
≤C n+θq
dy ≤ C kf kqB θ,p . (2.4.46)
Rn |y| q

Now, if q = ∞, then we may combine (2.4.36), (2.4.40), and (2.4.43) with Theorem 2.2.16 to
bound

kf kθ,∞  kf kLp + sup t−θ K(f, t)  kf kLp + sup t−θ kgt kLp + t1−θ kht kLp + t1−θ k∇ht kLp


0<t<1 0<t<1
≤ C kf kB∞
θ,p . (2.4.47)

On the other hand, if q < ∞, then we instead use (2.4.39), (2.4.40), and (2.4.43) with Theorem
2.2.16 to bound
Z 1 Z 1
q q q dt q dt
kf kθ,q  kf kLp + (K(f, t)) 1+θq ≤ kf kLp + (kgt kqLp + tq kht kqLp + tq k∇ht kqLp ) 1+θq
0 t 0 t
q
≤ C kf kB θ,p . (2.4.48)
q

Thus, in either case we deduce that

Bqθ,p (Rn ; F) ,→ (Lp (Rn ; F); W 1,p (Rn ; F))θ,q . (2.4.49)

To conclude, it remains to prove the reverse inclusion.

Suppose that f ∈ (Lp (Rn ; F); W 1,p (Rn ; F))θ,q and write f = g + h for g ∈ Lp (Rn ; F) and h ∈
W 1,p (Rn ; F). From the triangle inequality we have that

kf kLp ≤ kgkLp + khkLp ≤ kgkLp + khkW 1,p ≤ K(f, 1) ≤ C kf kθ,q , (2.4.50)

where in the last inequality we have used Theorem 2.2.8. On the other hand, for y ∈ Rn we may
write ∆y f = ∆y g + ∆y h and use the translation invariance of Lebesgue measure to estimate

k∆y f kLp ≤ k∆y gkLp + k∆y hkLp ≤ 2 kgkLp + k∆y hkLp . (2.4.51)

Next note that if ψ ∈ C 1 (Rn ; F) ∩ W 1,p (Rn ; F), then by the fundamental theorem of calculus,
Z 1
∆y ψ(x) = y · ∇ψ(x + ty)dt, (2.4.52)
0

86
and so Minkowski’s inequality shows that
Z 1
k∆y ψkLp ≤ |y| k∇ψkLp dt = |y| k∇ψkLp . (2.4.53)
0

However, basic Sobolev theory shows that smooth functions are dense in W 1,p (Rn ; F), so this esti-
mate continues to hold for general ψ ∈ W 1,p (Rn ; F) by an approximation argument.
Using (2.4.53) on h, we find that

k∆y hkLp ≤ |y| khkW 1,p , (2.4.54)

and so (2.4.51) shows that

k∆y f kLp ≤ 2 (kgkLp + |y| khkW 1,p ) (2.4.55)
for all decompositions f = g + h with g ∈ Lp (Rn ; F) and h ∈ W 1,p (Rn ; F). Thus, for y 6= 0 we have
that
k∆y f kLp ≤ 2K(f, |y|). (2.4.56)
If q = ∞, this implies that

k∆y f kLp K(f, t)

[f ]B∞
θ,p = sup
θ
≤ 2 sup = 2 kf kθ,∞ . (2.4.57)
y6=0 |y| t>0 tθ

On the other hand, if q < ∞, it instead implies that

k∆y f kqLp
Z ∞
(K(f, |y|))q (K(f, r))q
Z Z
q
[f ]B θ,p = n+θq
dy ≤ 2 n+θq
dy = 2αn 1+θq
dr ≤ C kf kqθ,q . (2.4.58)
q
Rn |y| Rn |y| 0 r

Then (2.4.57) and (2.4.58) combine with (2.4.50) to show that

kf kBqθ,p ≤ C kf kθ,q , (2.4.59)

and we deduce that

(Lp (Rn ; F); W 1,p (Rn ; F))θ,q ,→ Bqθ,p (Rn ; F), (2.4.60)
which completes the proof.

Theorem 2.4.6 allows us to deduce some simple properties of Besov spaces with minimal effort.
We consider two examples of this now.

Example 2.4.7. Using Corollary 2.2.10 and Theorem 2.4.6 in conjunction, we derive the interpo-
lation estimate
kf kBqs,p ≤ C kf k1−s s
Lp kf kW 1,p for all f ∈ W
1,p
(Rn ; F), (2.4.61)
where C ∈ R+ is a constant depending on the parameters. 4

Example 2.4.8. Suppose that 1 < p < n. Then by the Gagliardo-Nirenberg-Sobolev inequality,
∗
we know that W 1,p (Rn ; F) ,→ Lp (Rn ; F), where
1 1 1
∗
= − . (2.4.62)
p p n

87
 ∗
This tells us that the identity map I : Lp (Rn ; F) + W 1,p (Rn ; F) → Lp (Rn ; F) + Lp (Rn ; F) satisfies
I ∈ L(Lp ; Lp ) and L(W 1,p ; Lp ). Then Theorem 2.2.12 shows that I is a bounded linear map from
∗
(Lp (Rn ; F), W 1,p (Rn ; F))s,q = Bqs,p (Rn ; F) to (Lp (Rn ; F), Lp (Rn ; F))θ,q = Lr,q (Rn ; F) for
1 1−s s 1 s
= + ∗ = − . (2.4.63)
r p p p n
This and Theorems 2.4.2 and 2.4.6 then provide a constant C > 0 such that
kf kLr,q ≤ C kf kBqs,p for all f ∈ Bqs,p (Rn ; F). (2.4.64)
In other words, we have the subcritical embedding of the Besov spaces into the Lorentz spaces:
Bqs,p (Rn ; F) ,→ Lr,q (Rn ; F). Since we trivially have the embedding Bqs,p (Rn ; F) ,→ Lp (Rn ; F), we
have that Bqs,p (Rn ; F) ,→ Lp (Rn ; F) ∩ Lr,q (Rn ; F), and we can can then use Theorem 1.1.41 to
further deduce that Bqs,p (Rn ; F) ,→ Lt,1 (Rn ; F) for all p < t < r.
In particular, if we take q = p and note that p ≤ r, then we get the fractional Sobolev embedding
W s,p (Rn ; F) ,→ Lr,p (Rn ; F) ∩ Lp (Rn ; F) ,→ Lr (Rn ; F) ∩ Lp (Rn ; F).
4
Our final example computes the interpolation spaces between Besov and fractional Sobolev
spaces.
Example 2.4.9. Let 1 ≤ p < ∞, 1 ≤ q0 , q1 ≤ ∞, and 0 < s0 , s1 < 1. We know from Theorem
2.4.6 that for i ∈ {0, 1},
Bqsii ,p (Rn ; F) = (Lp (Rn ; F), W 1,p (Rn ; F))si ,q (2.4.65)
and is thus of type si . Let 1 ≤ q ≤ ∞, θ ∈ (0, 1), and define 1 < sθ < 1 via
sθ = (1 − θ)s0 + θs1 . (2.4.66)
Then Theorems 2.3.8 and 2.4.2 combine to show that
(Bqs00 ,p (Rn ; F), Bqs11 ,p (Rn ; F))θ,q = (Lp (Rn ; F), W 1,p (Rn ; F))sθ ,q = Bqsθ ,p (Rn ; F). (2.4.67)
Thus, when we interpolate between Besov spaces with the same first integrability index we get
another Besov space with the same first integrability index. Note that q0 and q1 play no role in
determining the type, so if we set q0 = q1 = p, then we find the fractional Sobolev space interpolation
result
(W s0 ,p (Rn ; F), W s1 ,p (Rn ; F))θ,q = Bqsθ ,p (Rn ; F), (2.4.68)
which in particular means that
(W s0 ,p (Rn ; F), W s1 ,p (Rn ; F))θ,p = W sθ ,p (Rn ; F). (2.4.69)
The above results extend also to the endpoints since Lp (Rn ; F) is of type 0 and W 1,p (Rn ; F) is
of type 1. Indeed, we have that
(Lp (Rn ; F), Bqs11 ,p (Rn ; F))θ,q = Bqθs1 ,p (Rn ; F) and (Lp (Rn ; F), W s1 ,p (Rn ; F))θ,p = W θs1 ,p (Rn ; F)
(2.4.70)
as well as
(Bqs00 ,p (Rn ; F), W 1,p (Rn ; F))θ,q = Bq(1−θ)s0 +θ,p (Rn ; F) (2.4.71)
and
(W s0 ,p (Rn ; F), W 1,p (Rn ; F))θ,p = W (1−θ)s0 +θ,p (Rn ; F). (2.4.72)
4

88
2.4.3 Interpolating between C 0 and C 1
It turns out that the technique we used above works equally well with Lp (Rn ; F) replaced by
Cb0 (Rn ; F) and W 1,p (Rn ; F) replaced by Cb1 (Rn ; F). Here we recall that

Cb0 (Rn ; F) = {f : Rn → F | f is continuous and bounded}, and

Cbk (Rn ; F) = {f : Rn → F | f is k-times differentiable and ∂ α f ∈ Cb0 (Rn ; F) for all |α| ≤ k}
(2.4.73)
are Banach spaces with the norms

kf kC 0 = sup |f (x)| and kf kC k = max k∂ α f kC 0 . (2.4.74)

b
x∈Rn b |α|≤k b

For 0 < α ≤ 1 and k ∈ N we also define

Cbk,α (Rn ; F) = {f ∈ Cbk (Rn ; F) | [f ]C k,α < ∞}, (2.4.75)

where
|∂ α f (x) − ∂ α f (y)|
[f ]C k,α = sup sup , (2.4.76)
|α|≤k x6=y |x − y|α
and we define kf kC k,α = kf kC k + [f ]C k,α . The spaces Cbk,α (Rn ; F) are Banach when endowed with
b b
these norms.
Theorem 2.4.10. For θ ∈ (0, 1) we have that

(Cb0 (Rn ; F), Cb1 (Rn ; F))θ,∞ = (Cb0 (Rn ; F), Cb0,1 (Rn ; F))θ,∞ = Cb0,θ (Rn ; F) (2.4.77)

with equivalence of norms.

Proof. We will only prove that (Cb0 (Rn ; F), Cb1 (Rn ; F))θ,∞ = Cb0,θ (Rn ; F). The second identity follows
from essentially the same argument; we leave it as an exercise to check the details.
First note that Cb1 (Rn ; F) ,→ Cb0 (Rn ; F), so according to Theorem 2.2.16, we can use the equiv-
alent norm
kf kC 0 + sup t−θ K(f, t) (2.4.78)
b
0<t<1

in place of the usual one. We will do so, and by abuse of notation, continue to refer to this quantity
as kf kθ,∞ .
Let f ∈ Cb0,θ (Rn ; C) and let η ∈ Cc∞ (Rn ) be a standard mollifier. For 0 < t < 1 write

f = (f − f ∗ ηt ) + f ∗ ηt =: g + h. (2.4.79)

Then Z
g(x) = (f (x) − f (x − y))t−n η(y/t)dy (2.4.80)
Rn
and we can estimate this via
Z
kgkC 0 ≤ [f ]C 0,θ |y|θ t−n η(y/t) ≤ C(η, θ)tθ kf kC 0,θ (2.4.81)
b b
Rn

for Z
C(η, θ) = |x|θ η(x)dx < ∞. (2.4.82)
Rn

89
On the other hand, since 0 < t < 1,

khkC 0 ≤ kf kC 0 ≤ tθ−1 kf kC 0,θ . (2.4.83)

b b b

Since Z
∂i η(x)dx = 0 for all 1 ≤ i ≤ n, (2.4.84)
Rn
we have that
Z Z
−n−1
∂i h(x) = f (y)t ∂i η((x − y)/t)dy = (f (y) − f (x))t−n−1 ∂i η((x − y)/t)dy, (2.4.85)
Rn Rn

and so
Z
max k∂i hkC 0 ≤ [f ]C 0,θ max |x − y|θ t−n−1 |∂i η((x − y)/t)| dy ≤ C 0 (η, θ)tθ−1 kf kC 0,θ (2.4.86)
1≤i≤n b 1≤i≤n Rn
b

for Z
0
C (η, θ) = sup |x|θ |pi η(x)| dx < ∞. (2.4.87)
1≤i≤n Rn

Hence, for 0 < t < 1 we have that

K(f, t) ≤ kgkC 0 + t khkC 1 ≤ Ctθ kf kC 0,θ (2.4.88)

b b b

for some constant C = C(η, θ) > 0. From this we deduce that

kf kθ,∞ ≤ kf kC 0 + C kf kC 0,θ ≤ C kf kC 0,θ . (2.4.89)

b b b

On the other hand, suppose now that f ∈ (Cb0 (Rn ; F), Cb1 (Rn ; F))θ,∞ and write f = g + h for
g ∈ Cb0 (Rn ; F) and h ∈ Cb1 (Rn ; F). Then

|f (x) − f (y)| ≤ |g(x) − g(y)| + |h(x) − h(y)| ≤ 2 kgkC 0 + khkC 1 |x − y| , (2.4.90)

b b

and since this holds for all such decompositions we deduce that

|f (x) − f (y)| ≤ 2K(f, |x − y|). (2.4.91)

Thus, for 0 < |x − y| < 1 we have that

|f (x) − f (y)|
θ
≤ 2 sup t−θ K(f, t), (2.4.92)
|x − y| 0<t<1

while for 1 ≤ |x − y| < ∞ we have that

|f (x) − f (y)|
≤ |f (x) − f (y)| ≤ 2 kf kC 0 . (2.4.93)
|x − y|θ b

Hence,  
−θ
kf kC 0 + [f ]C 0,θ ≤ 3 kf kC 0 + sup t K(f, t) ≤ 3 kf kθ,∞ . (2.4.94)
b b
0<t<1

90
 Let’s consider an example based on this result and reiteration.

Example 2.4.11. Let 0 < θ0 , θ1 , σ < 1 and write θσ = (1 − σ)θ0 + σθ1 . From Theorems 2.4.10 and
2.3.8 we have the identities

(C 0,θ0 (Rn ; F), C 0,θ1 (Rn ; F))σ,∞ = C 0,θσ (Rn ; F), (2.4.95)

(C 0 (Rn ; F), C 0,θ1 (Rn ; F))σ,∞ = C 0,σθ1 (Rn ; F), (2.4.96)

and
(C 0,θ0 (Rn ; F), C 1 (Rn ; F))σ,∞ = C 0,(1−σ)θ0 +σ (Rn ; F). (2.4.97)
4

References
[1] C. Bennett, R. Sharpley. Interpolation of operators. Pure and Applied Mathematics, 129. Aca-
demic Press, Inc., Boston, MA, 1988.

[2] J. Bergh, J. Löfström. Interpolation spaces. An introduction. Grundlehren der Mathematischen

Wissenschaften, No. 223. Springer-Verlag, Berlin-New York, 1976.

[3] Y. Brudnyĭ, N. Krugljak. Interpolation functors and interpolation spaces. Vol. I. Translated from
the Russian by Natalie Wadhwa. With a preface by Jaak Peetre. North-Holland Mathematical
Library, 47. North-Holland Publishing Co., Amsterdam, 1991.

[4] G. Leoni. A first course in Sobolev spaces. Second edition. Graduate Studies in Mathematics,
181. American Mathematical Society, Providence, RI, 2017.

[5] A. Lunardi. Interpolation theory. Third edition. Appunti. Scuola Normale Superiore di Pisa
(Nuova Serie) [Lecture Notes. Scuola Normale Superiore di Pisa (New Series)], 16. Edizioni della
Normale, Pisa, 2018.

[6] L. Tartar. Interpolation non linéaire et régularité. J. Functional Analysis 9 (1972), 469–489.

[7] H. Triebel. Interpolation theory, function spaces, differential operators. Second edition. Johann
Ambrosius Barth, Heidelberg, 1995.

91