Notes PDF

FINITE DIMENSIONAL NORMED SPACES
TOM SANDERS
In this course we shall study the classical theory of Banach spaces with an eye to its
quantitative aspects. The overarching structure follows that of the notes [Gar03] by Garling
entitled ‘Classical Banach Spaces’, but we also borrow heavily from the notes [Nao10] of
Naor entitled ‘Local Theory of Banach Spaces’, and the book [Woj91] of Wojtaszczyk
entitled ‘Banach Spaces for Analysts’.
In terms of prerequisites it will be useful to have taken a basic course on Banach spaces.
In the Oxford undergraduate degree there are three particularly helpful courses:
(a) B4.1 Banach Spaces, maths.ox.ac.uk/courses/course/26298/synopsis;
(b) B4.2 Hilbert Spaces, maths.ox.ac.uk/courses/course/26299/synopsis;
(c) C4.1 Functional Analysis, maths.ox.ac.uk/courses/course/26335/synopsis.
To agree notation we shall recap the relevant material when we come to need it, and while
we shall not dwell on ideas already developed in other courses we shall try to direct the
interested reader to a suitable source. Finally, the book [Bol99] of Bollobás may also serve
as a useful companion.
The course is constructed from the perspective that examples are essential, and there will
be an examples sheet available at people.maths.ox.ac.uk/sanders/ to which problems
will be added.
1. Introduction
We start by recalling some basic definitions and examples. Suppose that F is either R
or C, and X is a vector space over F. A norm on X is a function } ¨ } : X Ñ R that is
(i) (Homogenous) }λx} “ |λ|}x} for all λ P F, x P X;
(ii) (Sub-additive) }x ` y} ď }x} ` }y} for all x, y P X;
(iii) (Non-degenerate) }x} “ 0 implies that x “ 0X .
The pair pX, } ¨ }q is then said to be a normed space, and F is said to be the base field
or field of scalars.
The norm } ¨ } induces a natural metric on X defined via
dpx, yq :“ }x ´ y} for all x, y P X,
and pX, } ¨ }q is said to be a Banach space if X is complete as a metric space with respect
to this norm. If we say X is a Banach space without mentioning the norm then the norm
will be denoted } ¨ }X .
Last updated : 21st December, 2015.
In these notes we shall use bold text when making formal definitions, and italics for emphasis and
informal definitions. Text marked in blue was not lectured.
1
2 TOM SANDERS
1.1. Why restrict the base field to R and C? 1 The definition of normed space as
above is a little unsatisfactory because of the (apparently) artificial way we have restricted
attention to the fields R and C. Some of this is explained by the fact that the homogeneity
property of a norm makes reference to the absolute value defined on C.
It seems, then, that we could consider any sub-field of C, but since we are interested
in complete normed spaces, it follows that the underlying field must be complete and the
only complete sub-fields of C are R and C.
That being said, there is a more general notion of absolute value: given a field F an
absolute value on F is a map | ¨ | : F Ñ R such that
(i) (Multiplicative) |λ||µ| “ |λµ| for all λ, µ P F;
(ii) (Sub-additive) |λ ` µ| ď |λ| ` |µ| for all λ, µ P F;
(iii) (Non-degenerate) |λ| ě 0 with equality if and only if λ “ 0.
For example, if F is a finite field, then there is only one absolute value on F, the trivial one,
taking each non-zero x to 1 (and taking 0F to 0). There are more exotic absolute values
though: given a prime p we define
ˇa ˇ
ˇ nˇ
ˇ p ˇ :“ pń where pa, pq “ 1 “ pb, pq.
b p
This defines an absolute value on Q called the p-adic absolute value, and these absolute
values play an important role in number theory.
Examining the p-adic absolute values defined above more carefully one sees that they not
only satisfy the sub-additivity property, but in fact enjoy a stronger ultrametric property
viz.
|λ ` µ|p ď maxt|λ|p , |µ|p u for all λ, µ P Q.
We call an absolute value with this stronger property non-Archimedean, and otherwise
it is called Archimedean.
Absolute values induced metrics on fields in the same way that norms do and as before,
it is natural to ask that our field be complete with respect to this metric. Somewhat
surprisingly it turns out that any field which is complete with respect to an Archimedean
valuation is equivalent (in an appropriate sense) to R or C. For details see [Neu99, Theorem
4.2].
To summarise the discussion then, we are led to consider the case when our base field is
either R or C (the case we shall consider), or when it supports an absolute value enjoying the
ultra-metric property. Fields enjoying this latter property give rise to ‘non-Archimedean
functional analysis’ and the interested reader may wish to start with the monograph [vR78]
(reviewed in [Tai79]).
There are many examples of Banach spaces; we start with some of the so-called ‘classical’
spaces.
1This is off the main topic of the course, but is nonetheless a worthy question.
FINITE DIMENSIONAL NORMED SPACES 3
Example 1.2 (`p -spaces). Suppose that 1 ď p ă 8. We write `p for the set of F-valued
sequences x “ px1 , x2 , . . . q such that
˜ ¸1{p
8
ÿ
p
}x} :“ |xn | ă 8.
n“1
It is easy to check that `p is a vector space and } ¨ } defines a norm on `p so as to make it

into a Banach space.
By considering the limit as p Ñ 8 (either heuristically from the definition of the norm
or formally as a direct limit of the system in Example 2.2) we are lead to Tchebychev
space, denoted `8 , and defined to be the space of F-valued sequences endowed with the
norm
}x} :“ supt|xn | : n P N u.
1.3. Separability. In a certain sense the space `8 is too big, and we capture this with
the concept of separability. A Banach space X is said to be separable if it contains a
countable dense subset – we think of this set as a way in which we might ‘generate’ X.
Now, `8 is not separable as can be seen by noting that the set of vectors E :“ t1A :
A Ă Nu is 1-separated i.e.
}v ´ w}8 ě 1 for all v, w P E with v ‰ w.
It follows that any dense subset of `8 must contain at least one vector for every vector in
E, and hence be uncountable.
It is often easy to restrict attention to separable Banach spaces. Indeed, if X is a Banach
space and E Ă X is countable then the closure of the vector space generated by E is a
closed and separable2 subspace of X.
Example 1.4 (Convergent sequence spaces: c0 and cc ). In `8 (and, indeed, `p for 1 ď p ă

8) there is a natural countable set E :“ te1 , e2 , . . . u where
n´1 times
hkkikkj
(1.1) en :“ p0, . . . , 0, 1, 0, . . . q for each n P N.
The vector space generated by E is the space of finitely (compactly) supported sequences,
denoted cc (or sometimes c00 ) and its closure in `8 is denoted c0 , the space of sequences
tending to 0. By construction c0 is separable.
Given the above example we might ask what the closure of the vector space generated
by cc is in `p for 1 ď p ă 8. It is easy to check that this is actually the whole space `p ,
and so `p is separable whenever 1 ď p ă 8.
2The reader may wish to check this: all sums are finite, and both R and C have countable dense subsets.
4 TOM SANDERS
2. Operators
We shall be interested in understanding relationships between Banach spaces, and these
relationships are encoded by operators. If X and Y are Banach spaces over the same base
field F then we write LpX, Y q for the space of continuous linear operators X Ñ Y .
This is naturally endowed with a norm called the operator norm and defined by
}T }XÑY :“ supt}T x}Y : }x}X ď 1u.
With this norm LpX, Y q forms a Banach space over the base field F.
Example 2.1. Suppose that Y is a Banach space with base field F. Then there are two
natural maps
ψ : LpF, Y q Ñ Y ; T ÞÑ T 1F and φ : Y Ñ LpF, Y q; y ÞÑ pλ ÞÑ λyq.
It is easy to check that ψ ˝ φ is the identity on Y and φ ˝ ψ is the identity on LpF, Y q.
Moreover
}ψpT q} “ }T } for all T P LpF, Y q and }φpyq} “ }y} for all y P Y.
To all intents and purposes LpF, Y q and Y are ‘the same’.
This example leads us to some definitions. We say that T P LpX, Y q is a short map if
}T } ď 1; it is an isometry if
}T x} “ }x} for all x P X;
and it is an isometric isomorphism if it is a surjective isometry. Equivalently if it is
short and has a short inverse.
Sightly extending this new terminology the conclusion of Example 2.1 above is sim-
ply that LpF, Y q and Y are isometrically isomorphic because there is some isometric
isomorphism between them.
Example 2.2 (Nesting of `p -spaces). Whenever 1 ď q ď p ď 8 we have
}x}`p ď }x}`q for all x P `q .
It follows that the maps
ιqÑp : `q Ñ `p ; x ÞÑ x
are short, but if q ă p then they are not isometries.
In fact more is true and `p and `q are not isometrically isomorphic unless p “ q.
2.3. Linear functionals. Given a Banach space X over a field F, an operator in LpX, Fq
has a special name – it is called a linear functional – and we call this space of linear
functionals the dual space of X and denote it X ˚ .
It may be worth noting that X 1 is sometimes used in place of X ˚ , although more often X 1
is used to mean the algebraic dual of X, that is the set of all (not necessarily continuous)
linear functionals from X to F.
Example 2.4 (The structure of `˚p ). Suppose that 1 ă p ă 8 and write q for the conju-
gate exponent to p, that is 1{p ` 1{q “ 1. It turns out that `˚p is isometrically isomorphic
to `q as we shall now see.
If y P `q then there is a map φy P `˚p defined by
8
ÿ
x ÞÑ φy pxq “ xx, yy :“ xi y i .
i“1
This is easily seen to be linear and well-defined by Hölder’s inequality, which also tells us
that
|φy pxq| ď }x}`p }y}`q for all x P `p ;
thus }φy } ď }y}`q . In fact we have equality: consider x defined so that
xi yi }y}q´1 q
`q “ |yi | p and xi “ 0 when yi “ 0q,
which can easily be checked to lie in the unit ball of `p . On the other hand φy pxq “ }y}`q
as claimed. It follows (checking linearity in y) that the map y ÞÑ φy is a linear isometry
from `q to `˚p .
It turns out that y ÞÑ φy is an isometric isomorphism. To see that this map is surjective
(and hence an isometric isomorphism), suppose that φ P `˚p and let y P `8 be defined so
that yi :“ φpei q. We should like to show that y P `q and φ “ φy . Consider the vector x
defined such that
xi yi “ |yi |q and xi “ 0 if yi “ 0.
Write Pn x for the projection of x into the first n co-ordinates and note that
n
ÿ n
ÿ
φpPn xq “ x i yi “ |yi |q .
i“1 i“1
On the other hand

n
ÿ n
ÿ n
ÿ
}Pn x}p`p “ p
|xi | “ |yi | pq´1qp
“ |yi |q ,
i“1 i“1 i“1
and hence
˜ ¸1{q
n
ÿ
|yi |q ď }φ}.
i“1
Taking limits we conclude that y P `q . It follows that φ and φy restricted to cc agree, but
then cc is dense in `p and so φ “ φy as required.
One can show that `˚1 is isometrically isomorphic to `8 similarly. The first part also
goes through for `8 so that `1 embeds in `˚8 . However, in some models of ZF, for example
those for which the Hahn-Banach theorem holds there are many more functionals in `˚8
than those produced by `1 . On the other hand there are other models of ZF in which `˚8
is isometrically isomorphic to `1 . (See, for example, [Vät98].)
6 TOM SANDERS
2.5. Bilinear forms. Given Banach spaces X and Y over a field F an operator in LpX, Y ˚ q
is called a bilinear form. The reason for this name is that if T P LpX, Y ˚ q then
T pαx ` α1 x1 qpβy ` β 1 y 1 q “ αβT pxqpyq ` αβ 1 T pxqpy 1 q ` α1 βT px1 qpyq ` α1 β 1 T px1 qpy 1 q
for all x, x1 P X, y, y 1 P Y and α, α1 , β, β 1 P F. That is to say T induces a map
X ˆ Y Ñ F; px, yq ÞÑ T pxqpyq
that is bilinear.
Example 2.6. Although barely warranting the status of an example it will be useful to
note that if Y “ F then LpX, Y ˚ q is isometrically isomorphic to X ˚ .
2.7. Topologies of pointwise convergence. Given a set T and a vector space V of
functions T Ñ F the topology of pointwise convergence on V is defined to be the
weakest topology on V such that the evaluation functions
V Ñ F; f ÞÑ f ptq are continuous for all t P T.
This topology is rather useful in practice because of the following result.
Proposition 2.8. Suppose that X and Y are separable Banach spaces. Then the topology
of pointwise convergence on K, the unit ball3 of LpX, Y ˚ q, is metrisable4 and (sequentially)
compact.
Proof. Since X and Y are separable there are sequences pxm qm Ă X and pyn qn Ă Y , dense
in X and Y respectively. We define a metric by putting
8
ÿ
dpS, T q :“ 2´pn`mq mint|Spxm qpyn q ´ T pxm qpyn qq|, 1u for all S, T P LpX, Y ˚ q.
n,m“1
First we shall show that all the maps S ÞÑ dpS, T q (T P K) are continuous in the topology
of pointwise convergence. To see this note that the maps
(2.1) S ÞÑ mint|Spxm qpyn q ´ T pxm qpyn qq|, 1u for m, n P N
are continuous in the topology of pointwise convergence as they are the composition of the
maps
S ÞÑ Spxm qpyn q
which are continuous by definition of the topology; and
F Ñ R; λ ÞÑ mint|λ ´ T pxm qpyn q|, 1u
which are continuous by direct calculation (note that T pxm qpyn q is just a constant element
of F).
On the other hand S ÞÑ dpS, T q is a uniform limit of weighted sums of maps of the form
(2.1) and so is, itself, continuous. It follows that the topology of pointwise convergence is
at least as strong as the topology induced by d.
3In the operator norm.
4Meaning it is homeomorphic to a metric space.
In the other direction we trivially have that

S ÞÑ Spxm qpyn q
is continuous with respect to d for all pm, nq P N2 . Since pxm qm and pyn qn are dense in X
and Y respectively, it follows that for any x P X and y P Y there are sequences xmj Ñ x
and ymj Ñ y. Now, since K is norm-bounded, we see that the functions
p¨qpxmj qpynj q Ñ p¨qpxqpyq uniformly,
and hence the map
S ÞÑ Spxqpyq
is the uniform limit of continuous functions and so is continuous (with respect to d). But
x and y were arbitrary, so we conclude that the topology induced on K by d is at least as
strong as the topology of pointwise convergence.
Combining the two directions we have shown that the topology of pointwise convergence
is the same as that induced by d.
To see that K is (sequentially) compact suppose that pTj qj is a sequence of operators in
K and proceed by diagonalisation. Let n, m : N Ñ N be such that r ÞÑ pnprq, mprqq is a
bijection N Ñ N2 . Let Tj,1 be a subsequence of Tj such that Tj,1 pxmp1q qpynp1q q converges;
Tj,2 be a subsequence of Tj,1 such that Tj,2 pxmp2q qpynp2q q; and so on. This is possible since
the Tj s are bounded and bounded subsets of F are sequentially compact.
For the diagonal subsequence pTj,j qj we then have that Tj,j pxm qpyn q converges as j Ñ 8
for every fixed m and n, and it is a simple exercise to check that since pxm qm is dense in
X, and pyn qn is dense in Y we have that Tj,j pxqpyq converges for all x P X and y P Y .
One might suppose that the above proof also shows that the whole of LpX, Y ˚ q is metris-
able in the topology of pointwise convergence. In fact it shows that this topology on
LpX, Y ˚ q is a refinement of the topology induced on LpX, Y ˚ q by d, but it is not equal
unless the space is finite dimensional.
We also remark now that the proof above makes use of the Axiom of Dependent Choice:
we iteratively extract convergence subsequences. It follows that if one were trying to make
use of this in a finite setting it would be difficult, but then it is already difficult to make
use of sequential compactness in such a setting.
If X is a Banach space and V “ X ˚ then the topology of pointwise convergence is called
the weak-* topology and it is particularly useful because of the following theorem which
is an immediate corollary of Proposition 2.8 with Y “ F.
Theorem 2.9 (Sequential Banach-Alaoglu theorem, [Bat14, Theorems 5.7 and 5.9]). Sup-
pose that X is a separable Banach space. Then the unit ball in X ˚ is metrisable and
(sequentially) compact.
3. Spaces of continuous functions

Spaces of continuous functions are prototypical Banach spaces. They may seem bigger
than the sequence spaces we have considered before but it actually turns out that in many
cases they are not.
8 TOM SANDERS
Suppose that T is a compact metrisable space. We write CpT q for the space of con-
tinuous F-valued functions on T endowed with the norm
}f } :“ supt|f ptq| : t P T u.
It is easy to check that this is a Banach space (the uniform limit of continuous functions is
continuous). It is also separable, as is easy to see in the explicit case when T “ r0, 1s (with
the usual metric). (The details may be found in [Bel14, Example 5.4].) More generally this
is an application of the Stone-Weierstrass theorem but we do not pursue this here. (See
[Kec95, Theorem 4.19].)
As an aside we note that it may seem a little odd to talk about homeomorphisms rather
than isometries of metric spaces. There is a parallel here with Banach spaces where we
have two notions of equivalence: spaces can be (continuously) isomorphic (which we shall
properly define later) or isometrically isomorphic. It turns out that isometry in both cases
is often too restrictive.
3.1. Embedding in CpT q. One of the reasons that spaces of continuous functions are
important is that every separable Banach space with a reasonable dual can be viewed as
a subspace of a space of continuous functions on some compact metrisable space.
The argument will proceed by embedding a space into its double dual, but this can only
work if there are sufficiently many linear functionals i.e. if the dual is reasonably rich. One
way of capturing this is to ask that the map
ΦX : X Ñ X ˚˚ ; x ÞÑ pφ ÞÑ φpxqq
be an isometry. There are a number of reasons to think that this is reasonable. The
obvious one is that the Hahn-Banach theorem (with the attendant assumption that some
fragment5 of AC holds) can be used to prove it. For details see [Bel14, Theorem 7.3] and
[Bel14, Corollary 7].
While some fragment of choice is necessary in general, it is possible to show that ΦX is
isometric for many spaces without any such assumption, and moreover this often yields a
way to compute the isometry. For example, in 2.4 we showed in all but name that Φ`p is
an isometry for any 1 ď p ă 8, and are given a very easy way to index the elements of `˚p
(at least when p ą 1).
Theorem 3.2. Suppose that X is a separable Banach space. Then there is a short map
ψ : X Ñ CpKq for some compact metrisable space K; if ΦX is an isometry then this map
is an isometry.
Proof. Write K for the unit ball of X ˚ endowed with the topology of pointwise convergence.
By Theorem 2.9 it is metrisable and compact. Now consider the map
ψ : X Ñ CpKq; x ÞÑ pk ÞÑ kpxqq.
5For
separable Banach spaces it is possible to prove using a slight weakening of the Axiom of Dependent
Choice, and it turns out it is equivalent to this weakening in a suitable sense. (See [BS86] for details.)
This map is well-defined since K is topologised with the topology of pointwise convergence,
and so the image of x is a bonafide continuous map on K; the map is linear since the
elements of K are linear functionals; finally, the map is isometric because
}ψpxq}CpKq “ supt|kpxq| : k P Ku “ }ΦX pxq}X ˚˚ “ }x}
since ΦX is an isometry.
3.3. Universality of Cpr0, 1sq. It turns out that not only can (many) separable Banach
spaces be embedded into spaces of continuous functions, but in fact they can be embed-
ded into a particular space of continuous functions. Suppose that S and T are compact
metrisable spaces and ρ : T Ñ S is a continuous surjection. Then
CpSq Ñ CpT q; f ÞÑ f ˝ ρ
is an isometric linear map.
Our task now will be to find surjections from some well-known space to an arbitrary
compact metrisable space. We do this in two steps, starting with the Cantor set or count-
ably infinite dyadic compactum. Write D2 for the two point topological space with discrete
topology, and then put
ź8
8
D2 :“ D2
i“1
considered as a space endowed with the product topology. Equivalently, D28 is the set
t0, 1uN endowed with the metric
8
ÿ
(3.1) dpx, yq “ 2í |xi ´ yi |.
i“1
This space is called the countably infinite dyadic compactum. The Cantor set, on
the other hand, is defined to be the set
# +
8
ÿ xi
∆ :“ 2 i
: x P t0, 1uN
i“1
3
endowed with the subspace topology inherited from r0, 1s. It is an uncountable closed
subset of r0, 1s, and there is a natural homeomorphism between the Cantor set and the
countably infinite dyadic compactum, so we shall use the two interchangeably.
Proposition 3.4. Suppose that K is a compact metrisable space. Then there is a contin-
uous surjection f : D28 Ñ K.
Proof. Since K is metrisable we take it to be endowed with a metric d and since it is
compact we may take K, we may rescale the metric so that the closed unit ball about
some point in this metric is the whole of the space. (If x0 P K is some point, then
x ÞÑ dpx, x0 q is continuous and on a compact space and so it is bounded.)
10 TOM SANDERS
We define some sequences of closed neighbourhoods6 and non-negative integers itera-

tively. The integers shall be denoted j1 , j2 , . . . and the neighbourhoods will
pBH q, pBX1 qX1 Pt0,1uj1 , . . . , pBX1 ,...,Xn qX1 Pt0,1uj1 ,...,Xn Pt0,1ujn ,
such that for all 0 ď m ď n:
(i) BX1 ,...,Xm is a closed neighbourhood for all X1 , . . . , Xm ;
(ii) diam BX1 ,...,Xm ď 2´m for all X1 , . . . , Xm ;
(iii) Ť
BX0 ,...,Xm Ă BX0 ,...,Xm´1 for all X1 , . . . , Xm ;
(iv) Xm Pt0,1ujn BX0 ,...,Xm Ą BX0 ,...,Xm´1 for all X0 , . . . , Xm´1 .
Setting BH :“ K establishes the above for n “ 0. At stage n we consider the set BX1 ,...,Xn
for some X1 , . . . , Xn . Since BX1 ,...,Xn is closed and K is compact, the space BX1 ,...,Xn with
induced metric is compact. The cover
(3.2) tBX1 ,...,Xn X tx P K : dpx, yq ă 2´pn`1q u : y P Ku
is an open cover in the space BX1 ,...,Xn with induced metric and hence has a finite sub-cover
of size spX1 , . . . , Xn q. Let jn`1 be minimal such that
2jn`1 ě maxtspX1 , . . . , Xn q : X1 P t0, 1uj1 , . . . , t0, 1ujn u.
By adding in repetitions if necessary, index the sets in the finite sub-cover of (3.2) by
elements of t0, 1ujn`1 and label them
pBX1 ,...,Xn`1 qXn`1 Pt0,1ujn`1 .
Now, for x P t0, 1uN we write π1 pxq for px1 , . . . , xj1 q, π2 pxq for pxj1 `1 , . . . , xj1 `j2 q etc.. The
set
8
č
Bπ1 pxq,...,πn pxq
n“1
contains exactly one element since it is the intersection of a nested sequence of closed
non-empty sets with diameter tending to 0; we define f pxq to be that element.
It remains to check surjectivity and continuity. Suppose that k P K, then k P BH , and
by (iv) there is some X1 P t0, 1uj1 such that k P BX1 ; then there is some X2 P t0, 1uj2 such
that k P BX1 ,X2 ; and so on. Note that we certainly have
č8
kP BX1 ,...,Xn ,
n“1
and so letting x be the member of t0, 1uN generated by letting π1 pxq “ X1 , π2 pxq “ X2 ,
etc. we see that f pxq “ k and we have proved surjectivity.
Finally, for continuity, suppose that xn Ñ x. Then there is some N such that for all
n ą N we have dpxn , xq ă 2´pj1 `¨¨¨`jn q , where the metric here is (3.1). It follows that
πi pxn q “ πi pxq for all 1 ď i ď n, and hence
f pxn q, f pxq P BX1 ,...,Xn .
6Recall
that a closed neighbourhood is a closed set containing an open set e.g. a closed ball of
positive radius.
Since BX1 ,...,Xn has diameter at most 2ń we conclude that the distance between f pxn q and
f pxq in K is at most 2ń . Thus f pxn q Ñ f pxq and we have established continuity.
We are now in a position to establish the so-called universality of Cpr0, 1sq.
Corollary 3.5 (Banach-Mazur Theorem). Suppose that X is a separable Banach space
and ΦX is an isometry. Then there is an isometric embedding of X into Cpr0, 1sq.
Proof. We apply Theorem 3.2 to get an isometry X Ñ CpKq for some compact metrisable
space K. We know from the proof of Theorem 3.2 that K is actually the unit ball of the
dual space and is hence convex. Now, by Proposition 3.4 there is a continuous surjection
f : D28 Ñ K, and hence a continuous surjection g : ∆ Ñ K. Since ∆ is a closed subset of
r0, 1s we can define
h : r0, 1s Ñ K; x ÞÑ λf pyq ` p1 ´ λqf pzq
where y :“ infty 1 P ∆ : x ď y 1 u and z :“ suptz 1 P ∆ : x ě z 1 u, and x “ λy ` p1 ´ λqz. This
map h is a continuous surjection and so by the remarks at the start of §3.3 the result is
proved.
There are a number of other applications of the surjectivity of the Cantor set and the
interested reader may wish to consult [Ben98].
3.6. The dual of CpT q. Throughout this section take T to be a compact metrisable space,
and if specific examples are helpful then consider the case T “ r0, 1s.
To understand the dual of CpT q it will be useful to understand CpT q as a topological
vector space. A topological vector space is a vector space endowed with a topology
making vector addition and scalar multiplication continuous. Any Banach space X is an
example of a topological vector space when the underlying vector space is endowed with the
topology induced by the norm. Any topological vector space V has a dual space, defined
to be the vector space of continuous linear functionals on V , so that if X is a Banach space
then the dual spaces of X as a topological vector space with topology induced by the norm
is the same as the dual space of X considered as a Banach space.
There are two ways in which we consider CpT q as a topological vector space: first, with
CpT q endowed with the topology induced by the norm; secondly, with CpT q endowed the
the topology of bounded pointwise convergence i.e. we say fn Ñ f if
p}fn }qn is bounded and fn ptq Ñ f ptq for all t P T.
The space CpT q endowed with the topology of bounded pointwise convergence then has a
dual space, V , and it turns out that V “ CpT q˚ . It is easy to see that V Ă CpT q˚ since
if pfn qn convergence uniformly then it convergences in the bounded pointwise topology. In
the other direction this is essentially the content of the Bounded Convergence Theorem.
Theorem (Bounded Convergence Theorem). Suppose that φ P CpT q˚ and pfn qn is a se-
quence of bounded continuous functions on T with fn Ñ 0 pointwise. Then φpfn q Ñ 0 i.e.
φ is continuous when CpT q is endowed with the topology of bounded pointwise convergence.
12 TOM SANDERS
For a nice direct proof of this for the Riemann integral on r0, 1s see [Lew86].
The vector space CpT q with the topology of bounded pointwise convergence is not,
in general, a metric space (indeed, despite being sequentially defined it is not even first
countable), but (like all topological vector spaces) it nevertheless has a notion of Cauchy
sequence. In particular, we say pfn qn is a Cauchy sequence if and only if pfn qn is
bounded and pfn ptqqn is Cauchy for every t P T . We can then talk about the sequential
completion (sometimes semi-completion) of CpT q. Forming this abstractly is a little
complicated for reasons that will become clear in a moment, but the space of all bounded
functions on T is sequentially complete (since F is sequentially complete) and so we can
take the sequential closure of CpT q in this space, and we denote this closure LBaire
8 pT q –
the elements are the bounded Baire measurable functions.
The space LBaire
8 pT q is sequentially complete (since it is sequentially closed in a se-
quentially complete space) and every element of CpT q˚ extends to a continuous linear
functional on LBaire
8 pT q endowed with the topology of bounded pointwise convergence i.e.
˚
for all φ P CpT q there is a linear map φ̃ : LBaire
8 pT q Ñ F such that
φ̃pf q “ φpf q for all f P CpT q;

and
lim φ̃pfn q “ φ̃pf q whenever fn Ñ f in LBaire
8 pT q
nÑ8
with the bounded pointwise topology.

It is worth noting that because CpT q endowed with the topology of bounded pointwise
convergence is not first countable its completion cannot be formed by quotienting the
space of Cauchy sequences. Equivalently the set of limits of Cauchy sequences of functions
in CpT q is not sequentially closed. In fact we call this set of limits the space of Baire
one functions. When T “ r0, 1s this contains, for example, the indicator function of the
rationals with denominator at most n. It does not, however, contain their limit – the
indicator function of the rationals. For any ordinal n the Baire n functions are those
functions obtained as limits of Cauchy sequences of Baire functions of class less than n.
The fact that every element of CpT q˚ can be extended as above is really the key feature
of the dual space from our perspective. That being said, it is possible to describe these
functionals even more explicitly, and we turn to this now.
The Baire sets of T are the sets in the σ-algebra generated by the elements of CpT q
and a Baire measure is a finite measure on the σ-algebra of Baire sets. If µ is such then
the map
ż
CpT q Ñ F; f ÞÑ f dµ
is a continuous linear functional on CpT q.

It turns out that the converse of the above construction is also true. In the case of
T “ r0, 1s this is due to Riesz [Rie10]; more generally the result is due to Kakutani
[Kak41].
Theorem 3.7 (Riesz-Kakutani representation theorem). Suppose that T is a compact

metrisable space and φ P CpT q. Then there is a unique finite Baire7 measure on T such
that ż
φpf q “ f dµ for all f P CpT q.
We shall not include a proof of this result in the course. This is partly because it would
take us rather far afield and partly because the key property we shall need is the exten-
sion property described above and that follows from the rather straightforward bounded
convergence theorem for continuous functions and the completion of topological vector
spaces.
3.8. Dual maps. Dual spaces give rise to dual maps. In particular, given T : X Ñ Y a
continuous linear map between Banach spaces X and Y , we write
T ˚ : Y ˚ Ñ X ˚ ; y ˚ ÞÑ px ÞÑ y ˚ pT xqq.
This is easily seen to be a well-defined linear map and we have
}T ˚ } “ supt|y ˚ pT xq| : }x} ď 1 and }y ˚ } ď 1u ď }T }.
3.9. Isometries between spaces of continuous functions. We saw at the start of §3.3
that if there is a continuous surjection between two compact metrisable spaces S and T
then there is an isometric embedding from CpT q into CpSq. Extending this a little, if there
is a homeomorphism between S and T then there is an isometric isomorphism between
CpT q and CpSq. Interestingly it turns out that the converse is true as we shall now prove
following [Cam66] and [Ami65].
Theorem 3.10 (Robust Banach-Stone Theorem). Suppose that S and T are compact
metrisable spaces and Φ : CpSq Ñ CpT q, and Ψ : CpT q Ñ CpSq are continuous linear
inverses of each other with }Φ}}Ψ} ă 2. Then S and T are homeomorphic.
7It is worth noting that this is usually stated for Borel measures. The Borel σ-algebra on T is the
σ-algebra generated by the topology on T . The Baire σ-algebra is certainly a sub-algebra of the Borel
σ-algebra, but it is not, in general, equal. In our case they are equal because the spaces we consider
are second countable (meaning that the topology has a countable base) and so we shall not be overly
concerned with the distinction. The question of whether a probability measure on the Baire σ-algebra
of a normal topological space X can be extended to a measure on the Borel σ-algebra of X is called the
measure extension problem and a discussion may be found in [KM11]. (Here normal means that every
two disjoint closed sets have disjoint open neighbourhoods.)
One of the reasons that Baire measures are rather nice is that, unlike Borel measures, Baire probability
measures are automatically regular [Fre06, 412D]. A measure µ on a topological space is regular if
µpSq “ suptµpCq : C Ă S and C is closedu.
Mařı́k’s extension theorem [Mař57] shows that if X is countably paracompact then every Baire probability
measure on X extends to a regular Borel probability measure on X. Here countably paracompact
means that for every countable open cover U there is a open cover U 1 consisting of open subsets of the
sets in U, such that every point in X has a neighbourhood intersecting finitely many elements of U 1 . The
general measure extension problem is still open.
There is much more to be said here which we shall not concern ourselves with, but a gentler introduction
to some of the differences between Baire and Borel measures may be found in [Arv96].
14 TOM SANDERS
Proof. It is easiest to work with the duals so that we get maps

Φ˚ : CpT q˚ Ñ CpSq˚ and Ψ˚ : CpSq˚ Ñ CpT q˚
with }Φ˚ }}Ψ˚ } ă 2.
We shall find a homeomorphism between S and T by finding maps between the δ-
measures in CpSq˚ and CpT q˚ . This will happen in stages: write T 1 for the set of t P T for
which there is some s P S such that
|Φ˚ pδt qptsuq| ą }Φ˚ }{2.
Since }Φ˚ pδt q} ď }Φ˚ }}δt } “ }Φ˚ } we see that there can be at most one such s; we write
φ : T 1 Ñ S for the function8 taking t P T 1 to this unique s P S. We will show that this map
is a continuous surjection and similarly for the equivalent map associated with Ψ˚ instead
of Φ˚ . These two functions will turn out to be mutually inverse and we shall be done.
For each t P T 1 we let µt be a measure such that
Φ˚ pδt q “ αt δφptq ` µt where αt “ Φ˚ pδt qptφptquq,
so µt K δφptq (meaning µt ptφptquq “ 0) and |αt | ą }Φ˚ }{2, by definition of T 1 . These last
two facts entail
(3.3) }µt } “ }Φ˚ pδt q} ´ |αt | ă }Φ˚ }{2,
which will be important later.
Now we turn to showing that φ is a continuous surjection.
Claim. φ is surjective.
Proof. Suppose that s P S and let pfs,n qn converge to 1tsu in the bounded pointwise topol-
ogy. First, by the Bounded Convergence Theorem,
ż ż
Φ pδt qptsuq “ 1tsu dΦ pδt q “ lim fn,s dΦ˚ pδt q “ lim Φpfn,s qptq.
˚ ˚
nÑ8 nÑ8
It follows that pΦpfn,s qqn has a limit in the bounded pointwise topology. However, by
another application of the Bounded Convergence Theorem, we have
ż ż
1 “ lim fn,s dδs “ lim Φpfn,s qdΨ˚ pδs q
nÑ8 nÑ8
ż ż
“ lim Φpfn,s qdΨ pδs q “ Φ˚ pδt qptsuqdΨ˚ pδs qptq.
˚
nÑ8
Now, if |Φ˚ pδt qptsuq| ď }Φ˚ }{2 for all t P T then

1 ď p}Φ˚ }{2q ¨ }Ψ˚ pδs q} ď }Φ˚ }}Ψ˚ }{2 ă 1
which is a contradiction. It follows that there is some t P T such that |Φ˚ pδt qptsuq| ą
}Φ˚ }{2, and hence φptq “ s as required.
Claim. φ is continuous.
8At this stage for all we know the domain may be the empty set.
Proof. By the closed graph theorem for metric spaces9 it suffices to show that if tn Ñ t in
T 1 and φptn q Ñ s in S, then φptq “ s. We exploit the definition of T 1 to write
Φ˚ pδtn q “ αtn δφptn q ` µtn
where |αtn | ą }Φ˚ }{2 and (following (3.3)) }µtn } ă }Φ˚ }{2.
By passing to a subsequence if necessary we may assume that pαtn qn converges to some
α. Since tn Ñ t and φptn q Ñ s we have δtn Ñ δt and δφptn q Ñ δs in the weak-* topologies
on CpT q˚ and CpSq˚ respectively. On the other hand Φ is continuous and so Φ˚ is weak-*
to weak-* continuous and hence Φ˚ pδtn q Ñ Φ˚ pδt q in the weak-* topology. Since pαn qn also
converges it follows that µtn Ñ µ in the weak-* topology, and we have
Φ˚ pδt q “ αδs ` µ,
with |α| ě }Φ˚ }{2, }µ} ď }Φ˚ }{2. Since }µ} ď }Φ˚ }{2 we can conclude that
Φ˚ pδt qptxuq ď }Φ˚ }{2 for all x ‰ s.
However, since t P T 1 we also have
|Φ˚ pδt qptφptquq| ą }Φ˚ }{2
Hence s “ φptq as required.
Just as we defined T 1 and the function φ there is a set S 1 Ă S and a continuous surjection
ψ : S 1 Ñ T with
Ψ˚ pδs q “ βs δψpsq ` νs
where νs K δψpsq and βs P F has |βs | ą }Ψ˚ }{2.
It remains to show that ψ and φ are inverses of each other. Suppose that s P S 1 . Then
since φ is surjective there is some t P T 1 such that φptq “ s, and
(3.4) Φ˚ pδt q “ αt δs ` µt
where |αt | ě }Φ˚ }{2, µt K δs , and }µt } ă }Φ˚ }{2. On the other hand
Ψ˚ pδs q “ βs δψpsq ` νs
where |βs | ě }Ψ˚ }{2, νs K δψpsq and }νs } ă }Ψ˚ }{2. Combining these we get that
δt “ Ψ˚ pΦ˚ pδt qq “ αt βs δψpsq ` Ψ˚ pµt q ` αt νs .
Now if ψpsq ‰ t then since νs ptψpsquq “ 0 we must have
Ψ˚ pµt qptψpsquq “ ´αt βs
and hence that
1 “ δt pttuq “ |Ψ˚ pµt qpttuq| ď }Ψ˚ pµt q} ´ |αt ||βs | ` |αt |}νs }
ă }Ψ˚ }}Φ˚ }{2 ` |αt |p}νs } ´ |βs |q ă }Ψ˚ }}Φ˚ }{2 ă 1.
9The closed graph theorem in this case simply says that if X is a topological space and Y is a sequentially
compact metric space, then f : X Ñ Y is sequentially continuous if and only if its graph is sequentially
closed. This can be proved by passing to subsequences in Y .
16 TOM SANDERS
This contradiction ensures that ψpsq “ t, and hence φpψpsqq “ s. Similarly ψpφptqq “ t for
all t P T 1 .
Finally, if t P T then there is some s P S 1 such that ψpsq “ t, but then φptq “ s. It
follows that the image of φ is S 1 , but we know that φ is surjective. We conclude that
S “ S 1 ; similarly T “ T 1 . The result is proved.
For reference the Banach-Stone theorem usually means the above theorem in the case
when Φ (and Ψ) are isometries. We have called the above the Robust Banach-Stone
theorem (although this is not standard) because of the wiggle room in the hypotheses.
The proof presented above essentially follows [Cam66], although there it is established
for non-compact spaces too provided we replace CpSq and CpT q by C0 pSq and C0 pT q, the
continuous functions vanishing at infinity. In this extended setting Cambern [Cam70] gave
an example to show that the constant 2 in Theorem 3.10 is best possible.
Example 3.11. Consider S :“ tn´1 : n P Nu Y t0u Y tn´1 : n P Nu and T :“ tn´1 : n P
Nu Y t0u Y tn : n P Nu where both are endowed with the subspace topology from R. We
then define $
’gp0q
& if s “ 0
Ψpgqpsq :“ gpń q ` gpnq if s “ n´1
´1
’
%gpń´1 q ´ gpnq if s “ ń´1 .
It can be checked that Ψ is invertible and }Ψ´1 }}Ψ} “ 2, although S and T are not
homeomorphic.
4. Isomorphisms and the structure of `p spaces

The Banach-Stone theorem provides us with plenty of examples of spaces that are not
isometrically isomorphic, for example Cpr0, 1s Y r2, 3sq and Cpr0, 2s Y t3uq. We know that
these spaces are not isometrically isomorphic because the underlying topological spaces are
not homeomorphic. On the other hand if we consider
Φ : Cpr0, 1s Y r2, 3sq Ñ Cpr0, 2s Y t3uq
¨ $ ˛
&f pxq
’ if x P r0, 1s
f ÞÑ ˝x ÞÑ f px ` 1q ´ f p2q ` f p1q if x P p1, 2s ‚,
˚ ‹
’
%f p2q ´ f p1q if x “ 3
it is a continuous linear map of norm 3, and it has an inverse map
Ψ : Cpr0, 2s Y t3uq Ñ Cpr0, 1s Y r2, 3sq
˜ # ¸
f pxq if x P r0, 1s
f ÞÑ x ÞÑ ,
f px ´ 1q ` f p3q if x P r2, 3s
which is a continuous linear map of norm 2. We say that Cpr0, 1sYr2, 3sq and Cpr0, 2sYt3uq
are isomorphic.
Formally, given Banach spaces X and Y we say that they are isomorphic and write
X – Y if there are continuous linear maps Φ : X Ñ Y and Ψ : Y Ñ X that are mutually
inverse – such maps are called isomorphisms.
It turns out that the above example is completely typical and while CpT q and CpSq are
only isometrically isomorphic if S and T are homeomorphic, if S and T are any uncountable
compact metrisable spaces then CpT q is isomorphic to CpSq. This result is due to Miljutin
[Mil66]. (See [Woj91, §III.D, Theorem 19].)
Some properties of Banach spaces, e.g. separability, are preserved by isomorphism. This
tells us straight away that `8 is not isomorphic to any `p for 1 ď p ă 8. Furthermore,
if X – Y then X ˚ – Y ˚ . From this and the work of Example 2.4 we see that `1 is
not isomorphic to any `p with 1 ă p ă 8 since its dual is not separable. (It is also not
isomorphic to `8 , but that is for the reason previously mentioned.)
The question remains, what about the other `p spaces? It is already a useful exercise to
prove that `p and `q are not isometrically isomorphic if p ‰ q, but actually more is true
and it is the purpose of this section to prove the following result.
Theorem 4.1. Suppose that 1 ă p ă q ă 8. Then `p is not isomorphic to `q .
Our first attempt at an isomorphism is to take the identity map `p Ñ `q which (as we
saw in Example 2.2) is a short map. The problem is that this map is not surjective: there
are elements of `q that are not in `p as can be seen by considering the vector λ P `q defined
by λj :“ j ´2{pp`qq for all j P N, which has no pre-image in under the above identity map.
The proof we shall give of Theorem 4.1 will revolve around the idea that if T : `p Ñ `q is
continuous then we shall be able to find subspaces X ď `p and Y ď `q such that T pXq “ Y ,
and isomorphisms φ : `p Ñ X and ψ : `q Ñ Y , such that T is (almost) diagonal when
restricted to X i.e. some scalars pτi qi such that
T pφ´1 pei qq “ τi ψpei q for all i P N.
This will lead to a contradiction in the essentially the same way as above.
A key step in the previous paragraph is finding the subspaces X and Y , and to do
this we need there to be a lot of subspaces of `p isomorphic to `p . (In fact we shall see
later in Proposition 4.16 that every infinite dimensional complemented10 subspace of `p is
isomorphic to `p .)
We now need some notation. Define the linear maps
PN : `8 Ñ cc ; x ÞÑ px1 , . . . , xN , 0, . . . q,
for each N P N. We think of these as linear maps rather than operators because we
shall view them as maps from (and to) many different vector subspace of `8 (and vector
superspaces of cc ), with different norms.
We say that pyn q8 n“1 is a block basic sequence if there is a sequence of integers 0 “
j0 ă j1 ă . . . such that
Pjn pyn q “ yn and Pjn´1 pyn q “ 0 for all n P N.
10This will be defined later.
18 TOM SANDERS
The idea is that yn is supported on ejn´1 `1 , . . . , ejn , so that the supports of the yn s are
disjoint. We say that pyn q8 n“1 is a normalised block basic sequence if additionally
}yn } “ 1 for all n P N.
As an example, peni q8i“1 is a normalised block basic sequence in `p for any increasing
sequence of naturals n1 ă n2 ă . . . .
The reason that normalised block basic sequences are useful is that they provide us with
access to a huge number of subspaces of `p that are isometrically isomorphic to `p as the
following lemma captures.
Lemma 4.2. Suppose that pyn qn is a normalised block basic sequence in `q for 1 ď q ă 8.
Then the map
8
ÿ
Φ : `q Ñ `q ; λ ÞÑ λi yi
i“1
is an isometric linear map.
Proof. First, we need to check that the map is well-defined. Writing SN for the partial
sums on the right i.e.
N
ÿ
SN :“ ΦpPN λq “ λn yn for all N P N0 ,
n“1
with the usual convention about the empty sum so that S0 “ 0, we see that for naturals
N ą M we have
N
ÿ N
ÿ N
ÿ
}SN ´ SM }q`q “ } λn yn }q “ |λn |q }yn }q “ |λn |q “ }PN λ ´ PM λ}q ,
n“M `1 n“M `1 n“M `1
because the support of the yi s is disjoint. Hence if λ P `q then pPN λqN is Cauchy in
`q , and so pSN qN is Cauchy in `q , whence limN Ñ8 SN exists in `q and Φ is well-defined.
Furthermore, taking M “ 0 the above tells us that
› ›
›ÿ8 ›
› λi yi › “ lim }SN } “ lim }PN λ} “ }λ}
› ›
›i“1 › N Ñ8 N Ñ8
and so the map is an isometry. Finally, Φ is trivially linear on the vector subspace cc , and
the map is continuous (since it is an isometry), but cc is dense in `q and hence Φ is linear
on `q .
We shall also need a way to extract normalised block basic sequences from other se-
quences, and this lemma captures that.
Lemma 4.3. Suppose that pxn qn is a sequence of unit vectors in `q (1 ď q ă 8) with
PN xn Ñ 0 (in `q ) as n Ñ 8 for all N P N. Then there is a subsequence pxni qi and a
normalised block basic sequence pyi qi such that
}xni ´ yi } ď 2í for all i P N.
Proof. We start by constructing sequences 0 “: n0 ă n1 ă n2 ă . . . and 0 “: j0 ă j1 ă . . .

iteratively such that
(4.1) }xni ´ pPji ´ Pji´1 qxni } ď 2´pi`1q for all i P N.
Suppose that k P N0 and that we have chosen n0 , . . . , nk and j0 , . . . , jk such that (4.1) holds
for 1 ď i ď k. (Note that if k “ 0 this just means we have chosen n0 “ 0 and j0 “ 0.) By
hypothesis there is some nk`1 ą nk such that
}Pjk xnk`1 } ď 2´pk`3q .
On the other hand since xnk`1 P `p (and p ă 8) there is also some jk`1 ą jk such that
}xnk`1 ´ Pjk`1 xnk`1 } ď 2´pk`3q .
It follows by the triangle inequality that
}xnk`1 ´ pPjk`1 ´ Pjk qxnk`1 } ď 2´pk`2q ,
and (4.1) holds for i “ k ` 1. With this construction it is immediate that the auxiliary
sequence pzi qi defined by zi :“ pPji ´ Pji´1 qxni is a block basic sequence. By the triangle
inequality again we see that
|1 ´ }zi }| ď }xni ´ zi } ď 2´pi`1q for all i P N,
and we set yi :“ zi {}zi }. The triangle inequality then once again tells us that
}xni ´ yi } ď }xni ´ zi } ` }zi ´ yi } ď 2´pi`1q ` |}zi } ´ 1| ď 2í
as required.
Proof of Theorem 4.1. Suppose that T : `p Ñ `q is a continuous linear map with a contin-
uous inverse. We write zn :“ T en {}T en }`q . (This is well-defined because T is invertible.)
Claim. For all N P N we have PN zn Ñ 0 (in `q ) as n Ñ 8.
Proof. Write e˚i : `q Ñ F for the continuous linear functional defined by e˚i pxq “ xi for each
x P `p . Then note that PN zn Ñ 0 in `q as n Ñ 8 for all N P N if and only if e˚i pzn q Ñ 0
as n Ñ 8 for all i P N. We shall prove this second statement; suppose that i P N.
Since T is invertible we have
|e˚i pzn q| ď |e˚i pT pen qq|}T en }´1 ď }T ´1 }}en }´1 |e˚i pT pen qq| ď }T ´1 }|e˚i pT pen qq|.
Now x ÞÑ e˚i pT xq is a continuous linear functional on `q and so, by Example 2.4 (since
q ă 8), there is some w P `p1 (where 1{p ` 1{p1 “ 1) such that
8
ÿ
e˚i pT xq “ xj wj for all x P `q .
j“1
In particular, e˚i pT en q “ wn . Since p ą 1 and so p1 ă 8, we have that wn Ñ 8 as n Ñ 8,

and hence |e˚i pzn q| Ñ 0 as n Ñ 8 and the so-called ‘second statement’ above is proved,
establishing the claim.
20 TOM SANDERS
By Lemma 4.3 we conclude that there is some subsequence pzni qi and normalised block
basic sequence pyi qi such that
}T peni q ´ yi }T eni }`q }`q ď 2í }T eni }`q for all i P N.
Now, suppose that λ P `8 and N P N. Then by Lemma 4.2 (applied in `q to the normalised
N
block basic sequence pyi q8
i“1 and vector pλi }T peni q}`q qi“1 ) we have
› ›q
N N ›ÿN ›
` ´1 ´1 ˘q ÿ ÿ ˇ ˇ q
}T } }PN λ}`q “ }T ´1 }´q |λi |q ď ˇλi }T pen q}`q ˇ “ ›› λi }T pen q}`q yi ›› ,
i i
i“1 i“1
›i“1 ›
`q
and so by the triangle inequality we have

› ›
›ÿN › ÿN
´1 ´1
}T } }PN λ}`q ď › λi T peni q› ` |λi |2í }T eni }`q .
› ›
›i“1 › i“1 `q
However, again by Lemma 4.2 (since peni q8

i“1 is a normalised block basic sequence in `p ),
we also have
› › › ˜ ¸› › ›
›ÿN › › N
ÿ › ›ÿN ›
› λi T peni q› “ ›T λi eni › ď }T } › λi eni › “ }T }}PN λ}`p .
› › › › › ›
›i“1 › › i“1
› ›i“1 ›
`q `q `p
Combining these (and using nesting of norms between `8 and `p ) we get

}PN λ}`q ď }T }}T ´1 }p}PN λ}`p ` }PN λ}`8 q ď 2}T }}T ´1 }}PN λ}`p .
Since p ă q we can choose λ here to get a contradiction. Specifically, take λj :“ j ´2{pp`qq
so that
˜ ¸1{q
N
ÿ
ωN Ñ8 p1q “ j ´2q{pp`qq “ }PN λ}`q ď 2}T }}T ´1 }}PN λ}`p
j“1
˜ ¸1{p
N
ÿ
ď 2}T }}T ´1 } j ´2p{pp`qq “ ON Ñ8 p1q.
j“1
This contradiction shows that no such T can exist and hence proves the theorem.
4.4. Products, coproducts and direct sums. It will be useful for us to be able to build
new Banach spaces from old, and decompose existing Banach spaces into simpler pieces.
To this end we shall take a moment to set out some of the basic constructions. Much of
this has been covered in detail elsewhere (e.g. [Bat14, §1.1]) so for the most part we simply
record the essentials.
Recall
š that in a general category a coproduct of two objects X and Y is an object
X Yšfor which there are morphisms iX , iY such that for any object Z and morphisms
f : X Y Ñ Z there are maps i1 and i2 such that the following diagram commutes.
iX š iY
X X Y Y
f
i1 i2
Z
ś
Similarly a product of two objects X and Y is an object X Y for which there are
morphisms πX , πY such that forśany object Z and morphism π1 : Z Ñ X and π2 : Z Ñ Y
there is a morphism f : Z Ñ X Y such that the following diagram commutes.
πX ś πY
X X Y Y
f π2
π1
Z
The category Ban1 has products and coproducts of Banach spaces over the same field and
these can be described as follows.
Lemma 4.5 (Products and coproducts in Ban1 ). Suppose that X and Y are Banach
spaces over a field F. Then
š
(i) (Coproducts) X Y is isometrically isomorphic to the vector space direct sum
X ‘ Y endowed with the norm }px, yq} :“ }x}X ` }y}Y , and the short maps x ÞÑ
px, 0Y q and y ÞÑ
ś p0X , yq;
(ii) (Products) X Y is isometrically isomorphic to the vector space direct sum of
X ‘ Y endowed with the norm }px, yq} :“ maxt}x}X , }y}Y u, and the short maps
px, yq ÞÑ x and px, yq ÞÑ y; ś
(iii) (Isomorphism)
š The map iX πX ìY πY is a norm 2 continuous linear map X Y Ñ
X Y with a norm 2 inverse.
We leave the proof of this as an exercise.
In this section we have been interested in the question of when two Banach spaces
are continuously isomorphic, rather than when they are isometrically isomorphic. This
notion of isomorphism is the categorical notion of isomorphism in TopVect the category
of topological vector spaces with continuous linear maps.
This different perspective will have us looking at normable spaces rather than normed
spaces in much the same way we looked at metrisable spaces rather than metric spaces
in earlier sections. A topological vector space X is normable if the topology on X is
induced by a norm. Many of the notions we have discussed before work well for topological
vector spaces. In particular, X is separable if it has a countable dense subset, and it is
sequentially complete if every Cauchy sequence in X converges. Here a sequence pxn qn
is Cauchy if, given a local base B about 0, then for all V P B there is some N P N such
that xn ´ xm P V for all n, m ą N .
Lemma 4.6 (Topological vector space invariants). Suppose X and Y are isomorphic topo-
logical vector spaces. Then X is normable iff Y is normable; X is separable iff Y is sepa-
rable; X is sequentially complete iff Y is sequentially complete.
22 TOM SANDERS
One sometimes describes a normable, sequentially complete topological space as Ba-

nachable, although this is not a terribly attractive word.
In TopVect every product is isomorphic to a coproduct and vice-versa – we call these
objects biproducts or direct sums. Formally, given topological vector spaces X and Y over
a field F then the vector space X ‘Y endowed with the product topology is both a product
and a coproduct – we call it the topological direct sum or topological biproduct.
The key point for us is the following lemma.
Lemma 4.7. Suppose that X and Y are Banachable topological vector spaces over a field
F. Then there is a norm } ¨ } on the vector space X ‘ Y such that x ÞÑ px, 0Y q, y ÞÑ p0X , yq,
px, yq ÞÑ x and px, yq ÞÑ y are all continuous, and any such norm induces the product
topology on X ‘ Y . In particular, X ‘ Y is Banachable.
Proof. Suppose that } ¨ }X and } ¨ }Y are norms inducing the topologies on X and Y
respectively. Then }px, yq} :“ }x}X ` }y}Y is a norm such that the maps x ÞÑ px, 0Y q,
y ÞÑ p0X , yq, px, yq ÞÑ x and px, yq ÞÑ y are all continuous.
On the other hand, if } ¨ } is a norm on X ‘ Y such that x ÞÑ px, 0Y q, y ÞÑ p0X , yq,
px, yq ÞÑ x and px, yq ÞÑ y are continuous then
}x}X “ Op}px, yq}q and }y}Y “ Op}px, yq}q whence }x}X ` }y}Y “ Op}px, yq}q,
and
}px, yq} ď }px, 0Y q} ` }p0X , yq} “ Op}x}X ` }y}Y q,
and so } ¨ } is equivalent to the norm px, yq ÞÑ }x}X ` }y}Y .
Finally it is easy to check that this really is the product topology on X ‘ Y .
Some of the most useful Banach spaces giving rise to the topological vector space X ‘ Y
are defined as follows. For p P r1, 8s we write X ‘p Y for the space X ‘ Y endowed with
the norm
}px, yq} :“ p}x}pX ` }y}pY q1{p ,
with the natural convention for p “ 8. Note that these are norms on X ‘ Y of the type
described in Lemma 4.7, so they induce the product topology on X ‘ Y . In this language
Lemma 4.5 tells us that X ‘1 Y is a coproduct of X and Y (in Ban1 ), and similarly X ‘8 Y
is a product of X and Y (in Ban1 ). The last part of Lemma 4.5 then tells us that these
two spaces are continuously isomorphic.
Example 4.8. The space `p ‘p `p is isometrically isomorphic to `p . To see this simply note
that the maps
`p ‘p `p Ñ `p ; px, yq ÞÑ px1 , y1 , x2 , y2 , . . . q
and
`p Ñ `p ‘p `p ; x ÞÑ ppx1 , x3 , x5 , . . . q, px2 , x4 , . . . qq
are isometric isomorphisms.
This last construction and the `p -space construction (see Example 1.2) can be fused.
Given a Banach space X we write `p pXq for the set of vectors px1 , x2 , . . . q where xi P X
such that
8
ÿ
}xi }pX ă 8,
i“1
endowed with the norm ˜ ¸1{p
8
ÿ
}x}`p pXq :“ }xi }pX .
i“1
The elements of this space are sometimes called p-summable, and it is easy to check that
they form a Banach space in much the same way one does for `p “ `p pFq. This space has
some useful properties.
Lemma 4.9. Suppose X and Y are Banach spaces and p P r1, 8s. Then
(i) `p pX ‘p Y q is isometrically isomorphic to `p pXq ‘p `p pY q;
(ii) X ‘p `p pXq is isometrically isomorphic to `p pXq;
(iii) if X – Y , i.e. X is continuously isomorphic to Y , then `p pXq – `p pY q.
None of these is difficult; we leave the proof as another exercise.
Example 4.10. Building on Example 4.8 we have that `p p`p q is isometrically isomorphic
to `p . Let φ, ψ : N Ñ N be such that N Ñ N2 ; n ÞÑ pφpnq, ψpnqq is a bijection. Then
pφpnqq
`p p`p q Ñ `p ; pxpiq q8 8
i“1 ÞÑ pxψpnq qn“1
is a well-defined isometric isomorphism.

4.11. Complemented subspaces. Suppose that X is a Banach space and Y and Z are
closed subspaces of X with Y X Z “ t0u. Then we call Y ` Z an internal direct sum,
because the natural map py, zq ÞÑ y ` z is a continuous isomorphism from Y ‘ Z to Y ` Z.
Given a Banach space X we say that Y ď X is complemented in X if there is a
subspace Z ď X such that Y ` Z is an internal direct sum and X “ Y ` Z. Note that in
this case there may be many different spaces Z such that Y ` Z is direct and X “ Y ` Z.
Internal direct sums are closely related to projections: a projection on a Banach space
X is a continuous linear map π : X Ñ X such that π 2 “ π.
Lemma 4.12. Suppose that X is a Banach space and Y is a closed subspace of X. Then
Y is complemented in X if and only if there is a continuous linear projection π : X Ñ X
with image Y .
Example 4.13 (Continuous linear functionals). Suppose that X is a Banach space, φ P
X ˚ and y P X has φpyq ‰ 0. Then, writing Y for the space generated by y, we have
X “ Y ` ker φ, and the sum is direct. To see this note that the map
φpxq
π : X Ñ X; x ÞÑ x ´ y
φpyq
is a projection of norm at most 1 ` }φ}}y}{|φpyq|, and Y and ker φ are both complemented
in X.
24 TOM SANDERS
Example 4.14. Given a sequence of vectors pxi qi we define their span to be

# +
n
ÿ
Spanppxi qi q :“ λi xi : λ P Fn , n P N0 ,
i“1
and write Spanppxi qi q for the closure of this span. If pxi qi is a normalised block basic
sequence in `p (1 ď p ă 8) then we know from Lemma 4.2 that Spanppxi qi q is isometrically
isomorphic to `p . Furthermore, it is complemented in `p as we shall now see.
By Example 2.4 we know that for each xi there is a short functional φi P `˚p such that
φi pxi q “ 1. We define
8
ÿ
π : `p Ñ `p ; x ÞÑ φi pxqxi
i“1
and it is easy to check that it is a short projection with image Spanppxi qi q.

It is worth noting that not every closed subspace of a Banach space is complemented.
For example, the Phillips-Sobczyk Theorem11 is the assertion that c0 is not complemented
in `8 .
In the proof of Theorem 4.1 we needed an abundance of subspaces of `p that were
isomorphic to `p . Of course any finite dimensional12 subspace of `p will not be isomorphic
to `p , because it cannot be isomorphic as a vector space. Curiously, however, every infinite
dimensional subspace of `p contains a (complemented) copy of `p .
Lemma 4.15. Suppose X is a closed infinite dimensional subspace of `p (1 ď p ă 8).
Then there is a subspace W ď X with W – `p and W complemented in `p .
Proof. We start by constructing a block basic sequence in a way that is not dissimilar
to that in the proof of Lemma 4.3. We shall construct vectors y1 , y2 , . . . , unit vectors
x1 , x2 , . . . and integers 0 “: j0 ă j1 ă . . . such that
Pji´1 yi “ 0, Pji yi “ yi and }xi ´ yi } ď 2´pi`2q .
Suppose that we have constructed y1 , . . . , yk , x1 , . . . , xk , and j0 , . . . , jk for some k P N0 .
Consider the linear map Pjk : X Ñ `p . The image is finite dimensional, but X is infinite
dimensional, so there is some unit vector xk`1 P X with Pjk xk`1 “ 0. Since xk`1 P `p and
p ă 8 there is some jk`1 ą jk such that
}xk`1 ´ Pjk`1 xk`1 } ď 2´pi`2q ;
11Thefact that c is not complemented in `8 is proved by Phillips in [Phi40, 7.5], and Sobcyzk noted
in [Sob41] that this can be used to show that c0 is not complemented in `8 . This latter assertion is the
one often discussed because there is a short proof by Whitley [Whi66]. The details of that proof may be
found in [Bat14].
12To be clear a Banach space X is said to be finite dimensional if the underlying vector space is finite
dimensional and we write dim X for this dimension of the underlying space.
let yk`1 :“ Pjk`1 xk`1 and we are done. Putting zi :“ yi {}yi } the sequence pzi qi is a
normalised block basic sequence by design. We now consider the map
N
ÿ N
ÿ
Ψ : Spanppzi qi q Ñ Spanppxi qi q; λi zi ÞÑ λ i xi .
i“1 i“1
which is well-defined and linear.

Claim. We have
1
}Ψpzq ´ z} ď }z} for all z P Spanppzi qi q.
2
Proof. The key to this is the following consequence of the fact that }zi ´ xi } ď 2´pi`1q for
all i P N. If z P Spanppzi qi q then z “ N
ř
i“1 λi zi for some N P N0 and scalars λ1 , . . . , λN P F.
But then
› ˜ ¸ ˜ ¸› › ›
N N N N
› ÿ ÿ › ÿ 1 1 1 ›ÿ ›
Ψ λ z λ z z λ z
› › › ›
› i i ´ i i › ď |λ i |}x i ´ i } ď }λ}`8 ď }λ}`p “ › i i›
› i“1 i“1
› i“1
2 2 2 ›i“1
›
by Lemma 4.2. The claim follows.
Writing Z :“ Spanppzi qi q and W :“ Spanppxi qi q, and noting that
1 3
}z} ď }Ψpzq} ď }z} for all z P Spanppzi qi q.
2 2
by the claim, we see that Ψ extends to a continuous linear map from Z to W with a
continuous inverse – Z and W are isomorphic. By Lemma 4.2 Z is isomorphic to `p and
the first part of the conclusion is proved since W ď X.
For the second part note by Example 4.14 that Z is complemented in `p so there is a
space V ď `p such that `p “ Z ` V is direct. More than this, the associated projection
π : `p Ñ `p ; z ` v ÞÑ z is short i.e. }π} “ 1.
The map Φ : `p Ñ `p ; u ÞÑ Ψpπpuqq ` pu ´ πpuqq is a continuous linear map since Ψ and
π are such, and
1 1 1
}Φpuq ´ u} “ }Ψpπpuqq ´ πpuq} ď }πpuq} ď }π}}u} “ }u} for all u P `p .
2 2 2
It follows that }Φ ´ I} ď 1{2, whence Φ is invertible, and it remains to note that Φ ˝ π ˝ Φ´1
is a continuous linear projection with image W . The result is proved.
We are now in a position to prove the final result of this section.
Proposition 4.16. Suppose that X is an infinite dimensional complemented subspace of
`p for 1 ď p ă 8. Then X – `p .
Proof. We shall use Pelczyński’s method [Pel60] and the previous lemma.
By Lemma 4.15 there is a subspace W ď X with W complemented in `p and W – `p .
Since W is complemented in `p , there is a projection π : `p Ñ `p with image W . This
projection must be the identity on W and since W ď X, when restricted to X it becomes
26 TOM SANDERS
a projection on X, and so W is complemented in X. It follows that there is a subspace

U ď X such that X – W ‘ U and hence
X – W ‘ U – `p ‘ U – p`p ‘ `p q ‘ U – `p ‘ p`p ‘ U q – `p ‘ X.
On the other hand `p – X ‘ V and so
`p – `p p`p q – `p pX ‘ V q – `p pXq ‘ `p pV q
– pX ‘ `p pXqq ‘ `p pV q
– X ‘ p`p pXq ‘ `p pV qq
– X ‘ `p pX ‘ V q – X ‘ `p
Of course `p ‘ X – X ‘ `p and the result is proved.
Pelczyński’s method above is part of a family of related arguments including the Eilenberg-
Mazur swindle and the Cantor-Schröder-Bernstein Theorem; we shall prove the latter be-
low.
Theorem 4.17 (Cantor-Schröder-Bernstein Theorem). Suppose that there are injections
X Ñ Y and Y Ñ X, then there is a bijection between X and Y .
Proof. Instead of working in the category of `p -spaces we work in the category Set, where
direct sums are replaced by disjoint union and we write A – B to mean there is a bijection
between A and B. Write g : X Ñ Y for the given injection and put A :“ Y zgpXq. Then
there is a bijection #
g ´1 pyq if y P gpXq
Y Ñ X \ A; y ÞÑ
y otherwise.
We write Y – X \ A; similarly the injection Y Ñ X gives rise to a set B such that
X – Y \ B. It follows by associativity and commutativity of disjoint union that
X – X \ Z where Z :“ A \ B.
The key difference between this argument and that in Proposition 4.16 is that we cannot
apply Lemma 4.15; instead of finding a copy of `p in the ambient space (as in Proposition
4.16) we find a copy of a suitable infinite disjoint union of Z with itself.
Write f for the bijection X Ñ X\Z, put Wi :“ tx P X : x, f pxq, . . . , f i´1 pxq PŮ
X, f i pxq P
Zu and W :“ tx P X : x, f pxq, ¨ ¨ ¨ P Xu. It is easy to check that X “ W \ i Wi and
f i pWi q “ Z. Thus we can define the map
#
x if x P W
X Ñ W \ pZ ˆ Nq; x ÞÑ i
pf pxq, iq if x P Wi .
This map is trivially injective on W , and if pf i pxq, iq “ pf j pyq, jq then i “ j and hence x “ y
since f is a injection; we conclude that the map is injective. Moreover, if pz, iq P Z ˆN then
there is some x P Wi such that f i pxq “ z and hence the map is surjective. We conclude
that X \ pZ ˆ Nq – X.
The key point now (and also in the proof of Proposition 4.16) is to exploit the idea
behind Hilbert’s Hotel; by that argument we have for any set C that
C \ pC ˆ Nq – C ˆ N.
This is captured as with the second property of Lemma 4.9; the other properties are
pC ˆ Nq \ pD ˆ Nq – pC \ Dq ˆ N, and if C – D then C ˆ N – D ˆ N,
for any sets C and D. It follows from these (and our earlier definitions of A and B) that
pA \ Bq ˆ N – pA ˆ Nq \ pB ˆ Nq
– pA \ pA ˆ Nqq \ pB ˆ Nq
– A \ ppA ˆ Nq \ pB ˆ Nqq
– A \ ppA \ Bq ˆ Nq – ppA \ Bq ˆ Nq \ A.
Hence we conclude that
X – W \ ppA \ Bq ˆ Nq – W \ pppA \ Bq ˆ Nq \ Aq
– pW \ ppA \ Bq ˆ Nqq \ A – X \ A – Y.
The result is proved.
It is natural to wonder if a result of the above type holds for Banach spaces. In particular,
if X is complemented in Y and Y is complemented in X, then is X – Y . Such a result does
not hold as was shown by Gowers in [Gow96]. (In actual fact he established the stronger
result that there is a Banach space Z such that Z – Z ‘ Z ‘ Z, but Z fl Z ‘ Z.)
As an aside we remark that the proof of Proposition 4.16 effectively decomposes into
two parts: the first uses Lemma 4.15 to show that if X is complemented in `p (and infinite
dimensional) then `p is complemented in X; secondly, that a Schröder-Bernstein result
holds for Banach spaces when one of the spaces is `p .
5. Banach-Mazur distance
Associated to the notion of isomorphism is the Banach-Mazur distance defined be-
tween two spaces X and Y by
dBM pX, Y q :“ inft}Φ}}Φ´1 } : Φ : X Ñ Y is an isomorphism.u.
This, or rather log dBM pX, Y q, is a (pseudo-)metric and in this language we showed in
Theorem 4.1 that
dBM p`p , `q q “ 8 if 1 ă p ă q ă 8.
The Robust Banach-Stone theorem (Theorem 3.10) can also be written in this language
and it says
dBM pCpSq, CpT qq ă 2 ñ S is homeomorphic to T.
28 TOM SANDERS
A little care is needed here: if S and T are homeomorphic then CpSq and CpT q are
isometrically isomorphic and so dBM pCpSq, CpT qq “ 1. However, if dBM pX, Y q “ 1 it need
not be the case that X and Y are isometric13; we call such spaces almost isomorphic.
If X and Y are finite dimensional spaces then dBM pX, Y q is finite if and only if X and
Y have the same dimension. Moreover, we have the following lemma showing that the
infimum in the definition of dBM is achieved.
Lemma 5.1. Suppose that X and Y are finite dimensional Banach spaces and dBM pX, Y q “
K ă 8. Then there are maps T : X Ñ Y and S : Y Ñ X such that T S “ IY and ST “ IX
and }T }}S} “ K – we say that X is K-isomorphic to Y .
Proof. For every n P N there are linear maps Tn ?: X Ñ Y and Sn : Y Ñ X such that
Tn Sn “ IY and Sn Tn “ IX , and }Tn }, }Sn } ď K ` 1{n. Since X and Y are finite
dimensional we can pass to a subsequence such that Tnj Ñ T and Snj Ñ S in operator
norm. The required properties of S and T follow immediately.
Which yields the following as an immediate corollary.
Corollary 5.2. Suppose that X is a finite dimensional Banach space and dBM pX, Y q “ 1.
Then X is isometrically isomorphic to Y .
While two finite dimensional Banach spaces are isomorphic if and only if they have the
same dimension, the Banach-Mazur distance lets us quantify this.
Example 5.3 (`np spaces). We write `np for the vector space Fn endowed with the norm
˜ ¸1{p
ÿn
p
}x} :“ |xi | .
i“1
This is evidently an n-dimensional Banach space, and

dBM p`np , `nq q ď }I}`np Ñ`nq }I}`nq Ñ`np .
Now if 1 ď p ď q ď 8, then by Hölder’s inequality we have
1 1
}x}`np ď }x}`nq n p ´ q and }x}`nq ď }x}`np ,
1 1
and it follows that dBM p`np , `nq q ď n p ´ q . It turns out that when p ě 2 (or q ď 2) this is
tight.
A natural question arises as to what happens with more general spaces. In particular,
given two n-dimensional spaces X and Y , the Banach-Mazur distance between them is
certainly finite, but is there a bound uniform in the dimension? The answer follows from
the next simple proposition.
Proposition 5.4. Suppose that X is an n-dimensional Banach space. Then dBM pX, `n1 q ď
n.
13See, for example, [PB79, 2.1] or the examples sheet.
For finite dimensional spaces the dual is rather richer than we have come to expect in
our Choice-deprived world and, in particular, we have the following useful lemma.
Lemma 5.5 (Auerbach’s lemma, [Woj91, §II.E, Lemma 11]). Suppose that X is a finite
dimensional Banach space. Then there is a unit biorthogonal system in X ˆ X ˚ , meaning
there are unit vectors x1 , . . . , xn P X and φ1 , . . . , φn P X ˚ such that
φi pxj q “ δij for all 1 ď i, j ď n.
Proof. Let e˚1 , . . . , e˚n be a basis of X ˚ . (This is a purely algebraic fact since all linear
functions are continuous in finite dimensions, so X ˚ “ X 1 , the algebraic dual of X.)
We define the map
Ψ : X n Ñ F; pz1 , . . . , zn q ÞÑ detppe˚j pzi qqni,j“1 q
which is trivially continuous. It Ψpz1 , . . . , zn q “ 0 then the rows of the matrix pe˚j pzi qqni,j“1
are linearly dependent and so there are scalars pλi qni“1 such that
e˚j pλ1 z1 ` ¨ ¨ ¨ ` λn zn q “ λ1 e˚j pz1 q ` ¨ ¨ ¨ ` λn e˚j pzn q “ 0 for all 1 ď j ď n.
Since pe˚j qnj“1 is a basis and ΦX is an injection it follows that
λ1 z1 ` ¨ ¨ ¨ ` λn zn “ 0,
and hence z1 , . . . , zn are linearly dependent. Since there are subsets of n vectors in X that
are linearly independent we conclude that Ψ is not identically 0.
We also have that Ψ is multi-linear: if we fix pzi qi‰k then
n
ÿ
Ψpz1 , . . . , zk´1 , z, zk`1 , . . . , zn q “ p´1ql´1 detppe˚j pzi qqi‰k,j‰l qe˚l pzq,
l“1
and so z ÞÑ Ψpz1 , . . . , zk´1 , z, zk`1 , . . . , zn q is sum of linear maps in z and hence linear.
Continuity tells us that at Ψ has a maximum modulus on B n , the n-fold product of the
unit ball, B, of X (which is compact since X is finite dimensional). Multi-linearity tells us
that this is achieved for a vector px1 , . . . , xn q P B n with }xi } “ 1 for all 1 ď i ď n. Since
Ψ is not identitically 0 we have Ψpx1 , . . . , xn q ‰ 0 and we can define
φi pxq :“ Ψpx1 , . . . , xi´1 , x, xi`1 , . . . , xn qΨpx1 , . . . , xn q´1 ,
which is linear since Ψ is multi-linear. Since x1 , . . . , xn have been chosen to maximise the
modulus of Ψ over B n we certainly have that }φi } “ 1. As noted before }xi } “ 1, and
finally φi pxi q “ 1 and φi pxj q “ 0 if i ‰ j. This last fact is because φi pxj q is just the
determinant of a matrix in which the ith and jth rows are both pe˚1 pxj q, . . . , e˚n pxj qq. Thus
the rows are not linearly independent and so the determinant is 0.
Proof of Proposition 5.4. Suppose that x1 , . . . , xn P X is a basis of unit vectors for X, and
consider the map
ÿn
Φ : `n1 Ñ X; pλi qni“1 ÞÑ λi xi .
i“1
30 TOM SANDERS
This certainly has norm at most 1. The problem is that the inverse map might have very
large norm unless the xi are chosen carefully; we choose them using Auerbach’s lemma.
By biorthogonality we have
ˇ ˜ ¸ˇ › › › ›
ˇ ÿn ˇ ›ÿn › ›ÿ n ›
|λj | “ ˇφj λi xi ˇ ď }φj } › λi xi › “ › λi xi › .
ˇ ˇ › › › ›
ˇ i“1
ˇ ›i“1 › ›i“1 ›
It follows that › ˜ ¸› › ›
› n
ÿ › ÿ n ›ÿ n ›
› ´1
›Φ λi xi › “ |λi | ď n › λi xi › ,
› › ›
› i“1
› i“1
› i“1
›
i.e. }Φ´1 } ď n as required.
The triangle inequality for (the log of) the Banach-Mazur distance then tells us that for
any two n-dimensional spaces X and Y we have dBM pX, Y q ď n2 . In fact this bound can
be reduced to n by passing through `n2 rather than `n1 . Indeed, Hilbert space (of which `n2
is a key example) will play an important role in much of the rest of the course.
Since we have mentioned `n2 and `n1 it is rather natural to consider the other extreme:
`8 . Of course p`n1 q˚ is isometrically isomorphic to `n8 , and for finite dimensional spaces we
n
have that X is isometrically isomorphic to X ˚˚ (via ΦX ), and dBM pX, Y q “ dBM pX ˚ , Y ˚ q.

It follows that if X is n-dimensional then
dBM pX, `n8 q “ dBM pX ˚˚ , p`n1 q˚ q “ dBM pX ˚ , `n1 q ď n
by Proposition 5.4. Estimating the worst case for this distance is an open problem due
to Pelczyński [Pel84]. According to [DLAT10] the best known upper bound is due to
Giannopoulos [Gia95] who showed
dBM pX, `n8 q “ Opn5{6 q;
there is also a construction of a space X due to Szarek [Sza90] such that
dBM pX, `n8 q “ Ωpn1{2 log nq.
5.6. Near isometries. As noted above, if X is finite dimensional it is easy to check that
ΦX is an isometric isomorphism of X. This means that we can apply Theorem 3.5 to
see that X is isometrically isomorphic to a subspace of Cpr0, 1sq. The finite dimensional
analogue of Cpr0, 1sq is `n8 , and while (as we shall see shortly) it is not the case the every
finite dimensional space can be isometrically embedded in `n8 for some n, it is nearly the
case.
Proposition 5.7 ([Woj91, §II.E, Proposition 13]). Suppose that X is a n-dimensional
Banach space over R. Then there is an isomorphism Φ : X Ñ `N8 where N ď
Ópnq
with
p1 ´ q}x} ď }Φpxq} ď }x} for all x P X.
Proof. Let S be a maximal -separated subset of K, the unit ball of X ˚ , so that if φ, ψ P S

then
}φ ´ ψ} ą whenever φ ‰ ψ.
By the triangle inequality and the separation of S, none of the balls of radius {2 centred
at the elements of S overlap and so, writing µ for Lebesgue measure14 on Rn , we have
|S|µpKqp{2qn “ µpS ` tx P X : }x} ď {2uq

ď µptx P X : }x} ď p1 ` {2quq “ p1 ` {2qn µpKq.
(Note that K is closed and bounded so has finite measure; it contains a non-empty open
neighbourhood and so the measure is positive.) A bound on S follows on rearranging.
It remains to note that putting N “ |S| and
Φ : X Ñ `N
8 ; x ÞÑ pφs pxqqsPS
we have a φ with the desired properties. In particular, if x P X then there is some φ P X ˚

with }φ} “ 1 and φpxq “ }x} (see Exercise sheet). But then there is some s P S such that
}φs ´φ} ď , and it follows that ||φs pxq|´|φpxq|| ď }x} and we see that }Φpxq} ě p1´q}x}.
On the other hand all the elements φs are in the unit ball of the dual so we certainly have
}Φpxq} ď }x} and the result is proved.
Again, in the language of the Banach-Mazur distance, for every ą 0 and n-dimensional
Banach space X there is some N ď Ópnq and subspace Y ď `N 8 such that
dBM pX, Y q ď 1 ` .
It turns out that this is essentially best possible as we shall see in the next section.
5.8. The Banach-Mazur distance between `np and `nq . In the other direction from the
arguments above we shall show later in the course that for 1 ă p ă q ă 8
dBM p`np , `nq q Ñ 8 as n Ñ 8.
14So far we have only discussed Baire and Borel measures. There are various ways to define Lebesgue
measure, but one is as the completion of the unique translation invariant regular Borel measure on R
assigning mass 1 to r0, 1s. Any translation invariant regular Borel measure on R is called a Haar measure
and it turns out that such measures (exists and) are unique up to scaling. We shall discuss Haar measures
arising from actions of groups on compact spaces in §7 and which this does not cover R acting on itself, it
is a short step to this extension.
A measure is said to be complete if the measure of every subset of a set of measure 0 has measure 0.
The Borel σ-algebra on R has Borel sets having measure 0 in the Haar measure that have subsets that are
not measurable – the Haar measure is incomplete. However, given a measure µ on a measure space there
is a unique minimal completion – passing to this completion is easy and will not concern us further.
32 TOM SANDERS
6. Hilbert space
At the other end of the spectrum from the spaces of continuous functions we saw in §3
is Hilbert space. Recall that H is a Hilbert space if it is a Banach space with a norm
satisfying the parallelogram law i.e.
2}x}2 ` 2}y}2 “ }x ´ y}2 ` }x ` y}2 for all x, y P H.
It follows from this that } ¨ } is induced by an inner product.
Example 6.1. Hilbert spaces give rise to some surprising isometric isomorphisms. In
particular, the space `2 is isometrically isomorphic to L2 pr0, 1sq as can be shown (following
[Rad22]) with the Radamacher system of functions on r0, 1s:
rn pxq :“ sgn sinp2n πxq for all x P r0, 1s.
It may be most helpful to simply draw these. The map
8
ÿ
pλn q8
n“1 ÞÑ λn rn
n“1
is then an isometric isomorphism, and we leave the verification of this as an exercise.

In fact it turns out that any separable infinite dimensional Hilbert space is isometrically
isomorphic to `2 .
Hilbert spaces, like finite dimensional Banach spaces, have a rich dual structure. For
every x P H, the map y ÞÑ xx, yy is a continuous linear map on H and it turns out (this
is the Riesz representation theorem) that these are all such maps. A key ingredient in one
proof of this is the following lemma.
Lemma 6.2. Suppose that H is a Hilbert space, x, y P H?and φ P H ˚ have }x}, }y} ď 1,
}φ} ď 1 and |φpxq ´ 1|, |φpyq ´ 1| ă . Then }x ´ y} “ Op q.
Proof. We shall work with real Hilbert space so that φpzq ď }φ}}z} “ }z} for all z P H. (The
complex case is not substantially more difficult.) By linearity and the triangle inequality
we have
|φpx ` yq ´ 2| ă |φpxq ´ 1| ` |φpyq ´ 1| ă 2,
and it follows that
4 ě 2}x}2 ` 2}y}2 “ }x ` y}2 ` }x ´ y}2
ě φpx ` yq2 ` }x ´ y}2 ą p2 ´ 2q2 ` }x ´ y}2 .
Rearranging gives the result.
As an immediate corollary we get the Riesz representation theorem.
Corollary 6.3 (Riesz representation theorem). Suppose that H is a Hilbert space and
φ P H ˚ . Then there is some x P H with }x} “ }φ} such that φpzq “ xx, zy for all z P H.
Proof. We may take }φ} “ 1 and hence there is a sequence pxn qn of unit elements in H
such that φpxn q Ñ 1. By Lemma 6.2 the sequence pxn qn is Cauchy and so converges to
some x P H. It follows that }x} “ 1 and φpxq “ 1 by continuity of φ.
Now, suppose that y P ker φ is a unit vector and δ ą 0. Then
}x}2 ˘ 2xx, δyy ` δ 2 }y}2 “ }x ˘ δy}2 ě φpx ˘ δyq2 “ 1,
from which it follows that |xx, yy| ď δ{2. Letting δ Ñ 0 tells us that xx, yy “ 0. Hence
ker φ Ă txuK , and so φpzq “ xx, zy for all z P H.
6.4. Near isometries revisited. Returning to Proposition 5.7 we are now in a position
to show that it is best possible
Proposition 6.5 ([Nao10, Lemma 30]). Suppose that F “ R and there is a linear map
Φ : `n2 Ñ `N
8 such that p1 ´ q}x}`n
2
ď }Φpxq}`N8 ď }x}`n2 for all x P `n2 . Then N “ ´Ωpnq .
Proof. We write φi : `n2 Ñ R for the continuous linear functional taking x P `n2 to the ith
coordinate of Φpxq – there are N of them – and then we define the caps
Ki :“ tz P `n2 : }z}`n2 “ 1 and φi pzq ě 1 ´ u for 1 ď i ď N.
By the first inequality in the hypothesis we have

N
ď
Ki “ tz P `n2 : }z}`n2 “ 1u.
i“1
Since }Φ} ď 1 we see that }φi } ď 1 and so |φi pzq ´ 1| “ Opq for all z P Ki , and hence by
Lemma 6.2 there is some absolute constant C ą 0 such that
?
}x ´ y} ď C for all x, y P Ki ;
Let zi P Ki (if Ki is non-empty; if it is empty ignore it) so that, writing Br for the ball in
`n2 of radius r, we have Ki Ă zi ` BC ? . It follows that
˜ ¸
N
ď ďN
n
` ˘
B1` zB1´ Ă tz P `2 : }z} “ 1u ` B “
? ? ? Ki ` B Ă
? zi ` Bp1`Cq? .
i“1 i“1
We conclude that
? ?
pp1 ` qn ´ p1 ´ qn qµpB1 q “ µpB1`? zB1´? q
?
ď N µpBp1`Cq? q “ N pp1 ` Cq qn µpB1 q.
Since B1 is closed it is measurable, and since it contains a neighbourhood of the origin we

have µpB1 q ‰ 0. Dividing we then get the result on rearranging.
34 TOM SANDERS
7. Haar measure
Measure was crucial to the arguments in Propositions 5.7 and 6.5, and is available in
the finite dimensional setting because the unit ball is compact. In those propositions the
measure was Lebesgue measure on Rn , the key property of which was that it is invariant
under translation. It turns out, however, that any group action on a compact space admits
an invariant measure. (Of course Rn is not compact so this is not a generalisation, but
these statements are closely related.)
Suppose we have a compact metric space T with metric d, and a group G acting isomet-
rically on T so
dpgx, gyq “ dpx, yq for all x, y P T and g P G.
We call a measure µ on the Baire sets of T a G-Haar measure if
ż ż
f pgxqdµpxq “ f pxqdµpxq for all g P G,
i.e. the measure is invariant under the group action15. µ will be called a G-Haar proba-
bility measure if it is a G-Haar measure and a probability measure.
Theorem 7.1 (Haar measure, [Nao10, Theorem 3]). Suppose that G is a group acting
isometrically on a compact metric space T with metric d. Then there is a G-Haar probability
measure on T .
We shall give a proof of this result due to [Maa35]; the historical context comes from
[Jac84], and our treatment is from [Nao10]. One of Maak’s insights was that one could
make use of Hall’s marriage theorem to prove this, although he proved his own variant
with Hall’s theorem appearing a little later.
Theorem 7.2 (Hall’s marriage theorem). Suppose that G is a finite bipartite graph with
vertex sets V and W such that16 for any S Ă V we have ΓpSq :“ tw P W : v „ wu at least
as large as S. Then there is an injective choice function ψ : V Ñ W such that v „ ψpvq.
Proof. This appears in the course C8.3 Combinatorics as [Sco15, Theorem 3] along with a
far more extensive discussion. We shall include a brief proof here for completeness.
We shall proceed by induction on the number of edges in the graph. The result is trivial
for the empty graph, and we split each step of the induction into two cases:
Case (A). There is some H ‰ V 1 Ĺ V and |ΓpV 1 q| “ |V 1 |
15Notethat really this the property of being Haar is a function of the action not the group. The same
group might act in completely different ways, in which case the Haar measures may be different. Consider,
for example, the space T “ F22 endowed with the metric dpx, yq “ 1 if and only if x ‰ y, and the actions
of F2 on T defined by λ ÞÑ px ÞÑ x ` p0, λqq and λ ÞÑ px ÞÑ x ` pλ, 0qq. These actions are isometric and
the measure µ on T defined by assigning mass 1{2 to the points p0, 0q and p0, 1q and mass 0 everywhere
else is a Haar probability measure with respect to the first action, but not the second.
16This property is called the Hall property.
We decompose into two bipartite graphs: let G1 have vertex sets V1 :“ V 1 and W1 :“
ΓpV 1 q, and G2 have vertex sets V2 :“ V zV 1 and W2 :“ W zΓpV 1 q. Both of these graphs
have the Hall property: for the first one, if S Ă V1 then ΓG pSq Ă ΓG pV1 q Ă W1 , and hence
|ΓG1 pSq| “ |ΓG pSq| ě |S|.
Now, if S Ă V2 then
|ΓG pS Y V q| ě |S| ` |V | “ |S| ´ |ΓpV q|,
whence
|ΓG2 pSq| ě |ΓG pS Y V q| ´ |W1 | “ |ΓG pS Y V q| ´ |V | ě |S|,
as required. Since H ‰ V 1 ‰ V we see that both G1 and G2 contain an edge and so both
contain strictly few edges than G. The inductive hypothesis applies and we get a function
ψ1 : V1 Ñ W1 and ψ2 : V2 Ñ W2 ; ψ is just the combination of these functions.
Case (B). For all H ‰ V 1 Ĺ V we have |ΓpV 1 q| ą |V 1 |.
In this case G contains an edge from an element v P V ; remove it to get G 1 . It follows
that
|ΓG 1 pV 1 q| ě |ΓG pV 1 q| ´ 1 ě |V 1 |
whenever V 1 ‰ V . If the set V 1 :“ V ztvu is non-empty then
|ΓG 1 pV q| ě |ΓG 1 pV 1 q| “ |ΓG pV 1 q| ě |V 1 | ` 1 “ |V |,
and so G 1 has the Hall property and we can apply the inductive hypothesis. The final
possibility is that V “ tvu in which case the result is trivial.
The result now follows by complete induction since the cases are exhaustive.
Hall’s marriage theorem [Hal35] (which it turns out is a special case of a result of König
from [Kön16]) can be proved by induction or by duality.
Proof of Theorem 7.1. The most obvious idea is to construct a functional in the same way
one constructs the Riemann integral on r0, 1s. (Although r0, 1s is not a group, it is nearly,
and is certainly illustrative.) This is not quite possible because we have no analogue of
open interval – a sort of open set for which we can write down a ‘length’ – however, if we
only initially want to integrate continuous functions there is another way to define it.
Suppose that x1 , . . . , xn in r0, 1s and consider the functionals
n
1ÿ
φn : Cpr0, 1sq Ñ R; f Ñ
Þ f pxi q.
n i“1
Since r0, 1s is compact every function f P Cpr0, 1sq is uniformly continuous and so once
there are sufficiently many suitably spread out xi s this will be a good approximation to the
integral of f . We might hope to use compactness (Theorem 2.9) to take a limit of these
and then extract a measure by the Riesz-Kakutani representation theorem (Theorem 3.7).
Picking the xi s is slightly delicate. We would naturally pick them independently and
uniformly at random from r0, 1s, but that gets us back to where we started. There is a
36 TOM SANDERS
metric notion of independence here – δ-separation which has a dual concept of δ-covering.
We say that S Ă T is δ-covering if
ď
T Ă Bps, δq where Bps, δq :“ tt P T : dps, tq ď δu.
sPS
Instead of picking a sequence x1 , . . . , xn in r0, 1s (or T ) uniformly at random, we shall pick

it to be δ-covering of minimum 17 size.
We turn to the proof proper, and start with a claim which captures what we need about
coverings of minimum size.
Claim. Suppose that δ ą 0, and S and T are δ-coverings of T , and S has minimum size
amongst all δ-coverings. Then there is a function ψ : S Ñ T such that dpψpsq, sq ă 2δ.
Proof. Form a bipartite graph with (disjoint) vertex sets S and T , and connect s P S to
t P T if and only if dps, tq ă 2δ. This graph has the Hall property: if S Ă S has |ΓpSq| ă |S|
(where ΓpSq “ tt P T : dps, tq ă 2δu), then consider the set U :“ pSzSq Y ΓpSq. This has
17It
is important here that we take the minimum size rather than a minimal δ-covering set. This can
be seen by considering the case T “ r0, 1s again and considering δ :“ 1{4n. If x0 , . . . , x2n are the points
ˆ ˙
1
0, 2δ, 4δ, . . . , p2n ´ 2qδ, 2nδ “ , p2n ` 2qδ, . . . , p4n ´ 2qδ, 4nδ “ 1,
2
then they form a set that is a minimal δ-covering of r0, 1s, in the sense that no element can be removed.
The functionals
2n
1 ÿ
ψn : Cpr0, 1sq Ñ R; f ÞÑ f pxi q
2n ` 1 i“0
converge weakly to the usual integral on r0, 1s, exactly as we should like. However, if we perturb the first
half slightly by some (progressively larger multiple of a) very small amount η to get
ˆ ˙
1
2η, 2δ ` 4η, 4δ ` 6η, . . . , p2n ´ 2qδ ` 2nη, 2nδ “ , p2n ` 2qδ, . . . , p4n ´ 2qδ, 4nδ “ 1,
2
then this sequence is no longer δ-covering since there are gaps
pδ ` 2η, δ ` 4ηq, p3δ ` 4η, 3δ ` 6ηq, . . . , pp2n ´ 3qδ ` p2n ´ 2qη, p2n ´ 3qδ ` 2nηq,
provided p2n ´ 3qδ ` 2nη ă p2n ´ 1qδ (i.e. η ă δ{n). If we cover these gaps by adding in the n ´ 1 elements
δ ` 3η, 3δ ` 5η, . . . , p2n ´ 3qδ ` p2n ´ 1qη,
the resulting sequence which, we relabel x0 , . . . , x3n´1 , has 3n elements and is minimal δ-covering in the
sense that we cannot remove any element and still have a δ-covering (again, provided η ă δ{n). Arranging
η “ δ{2n, say, we then have that the functionals
3n´1
1 ÿ
ψn : Cpr0, 1sq Ñ R; f ÞÑ f pxi q
3n i“0
approach
ż1 ż 1{2
Cpr0, 1sq Ñ R; f ÞÑ f pxqdx ` f pxqdx.
0 0
To summarise, we have found a sequence of functionals each of which is formed by averaging over a
minimal covering set which does not converge to the usual integral – it turns out that if we use functionals
of minimum size they will.
size strictly smaller that S. Since T is δ-covering we have some T 1 Ă T such that
ď ď
Bps, δq Ă Bpt, δq.
sPS tPT 1
Ť
We may certainly take T 1 Ă ΓpSq, since for any t R ΓpSq we have Bpt, δqX sPS Bps, δq “ H.
It follows that ď ď
Bps, δq Ă Bpt, δq,
sPS tPΓpSq
and hence U is δ-covering. Since S is of minimum size we conclude the aforementioned

graph has the Hall property. It follows that there is a map ψ : S Ñ T such that dpψpsq, sq ă
2δ.
Since T is compact, the cover
tts P T : dps, tq ă 1{nu : t P T u
has a finite sub-cover Cn , and hence tt : ts P T : dps, tq ă 1{nu P Cn u is a finite 1{n-covering
subset of T . It follows that there is a 1{n-covering subset of T of minimum size; let Tn be
such a set and let φn be the linear functional
1 ÿ
CpT q Ñ C; f ÞÑ f ptq.
|Tn | tPT
n
The sequence pφn qn is in the unit ball of CpT q˚ and so by the sequential Banach-Alaoglu
theorem (Theorem 2.9) we see that there is a subsequence φnj Ñ φ in the topology of
pointwise convergence. Considering the constant function 1, we see that |φnj p1q| ě 1, and
so |φp1q| ě 1. On the other hand φ remains in the unit ball of CpT q˚ and so }φ} “ 1.
Similarly, if f ě 0 then φnj pf q ě 0 and so φpf q ě 0.
We now turn to showing that φ is G-invariant. For each g P G we write τg pf q for the
function t ÞÑ f pg ´1 ptqq. Fix f P CpT q and g P G; we shall show that φpτg pf qq “ φpf q.
Suppose ą 0. Since T is compact, f is uniformly continuous and there is some δ ą 0
such that |f psq ´ f ptq| ă whenever dps, tq ă 2δ. Let n be such that
|φn pτg pf qq ´ φpτg pf qq| ă and |φn pf q ´ φpf q| ă .
Now
1 ÿ 1 ÿ
φn pτg pf qq “ f pg ´1 ptqq “ f puq.
|Tn | tPT |Tn | uPg´1 T
n n
´1 ´1
Since G acts isometrically, the set g Tn “ tg ptq : t P Tn u is 1{n-covering. It also has
minimum size since it is the same size at Tn , and so by the earlier Claim there is an injective
map ψ : Tn Ñ g ´1 Tn such that dpψptq, tq ă 2δ. Since Tn and g ´1 Tn have the same size and
ψ is injective it follows that ψ is a bijection and hence
1 ÿ
φn pτg pf qq “ f pψptqq.
|Tn | tPT
n
38 TOM SANDERS
By the triangle inequality and choice of and δ earlier we have

1 ÿ
|φn pτg pf qq ´ φn pf q| ď |f pψptqq ´ f ptq| ă .
|Tn | tPT
n
By the triangle again we therefore have

|φpτg pf qq ´ φpf q| ď |φpτg pf qq ´ φn pτg pf qq| ` |φn pτg pf qq ´ φn pf q| ` |φn pf q ´ φpf q| ă 3.
However, was arbitrary and hence φpτg pf qq “ φpf q as required.
Finally, to extract a measure we apply the Riesz-Kakutani representation theorem (The-
orem 3.7) to φ. This gives us a Baire measure µ on T such that
ż
φpf q “ f dµ for all f P CpT q.
The measure is a probability measure since µpAq “ φp1A q ě 0 for all measurable A, and
µpT q “ φp1q “ 1. The measure is G-invariant since µpgAq “ φpτg p1A qq “ φp1A q “ µpAq
for all measurable A.
There are many examples of groups acting on compact metric spaces.
Example 7.3 (Isometries of Banach spaces). Suppose that X is an n-dimensional Banach
space. Then the group of isometries of X induces isometries of K the closed unit ball of
X. This space is compact since X is finite dimensional and so it has a Haar probability
measure.
The isometry group of every real Banach space includes ˘I, but it can be the case that
these are the only isometries. Indeed, Jarosz [Jar88] showed that any real Banach space
can be equipped with an equivalent norm so that the only isometries are ˘I. More than
this he showed that for any countable group G there is an equivalent norm on CpT q (with
F “ R) such that the group of isometries is (isomorphic to) G ˆ t´1, 1u.
At the other end of the spectrum if X “ `n2 then the group of isometries of X is On , and
this gives rise to a rather rich group of isometries of the unit ball in Euclidean space Rn .
Example 7.4 (The group of isometries of a metric space). Given a compact metric space
T we put
IsompT q :“ tg : T Ñ T s.t. g is an isometry of T u,
which is a group. It also easy to check that it itself becomes a metric space via
dIsom pg, hq :“ suptdpgptq, hptqq : t P T u.
Usefully we also have the following.
Claim. pIsompT q, dIsom q is a compact metric space.
Proof. We proceed as in the proof of the sequential Banach-Alaoglu theorem (Theorem

2.9) by diagonalisation. The slight difference is that we are not given that T is separable
(in the metric sense meaning that it has a countable dense subset), however it follows that
T is separable since it is a compact metric space (assuming countable choice, see [KT01],
though we shall need dependent choice for the rest of the proof so this is not unreasonable).
Let ptj q8
j“1 be a countable dense subset of T .
Suppose that pgn qn is a sequence in G. We need to find a subsequence that converges
in dIsom . We define a pointwise convergent subsequence as follows: let gn,0 :“ gn for all
n P N, and for each j P N let pgn,j q8 8
n“1 be a subsequence of pgn,j´1 qn“1 such that gn,j´1 ptj q
converges (possible since T is a compact metric space). We now consider the sequence
pgn,n q8
n“1 . This converges pointwise for every tj , but since the tj s are dense and the gn,n s
are isometries it follows that it converges pointwise for all t P T ; write
gptq :“ lim gn,n ptq.
nÑ8
Note that for all n P N we have

|dpgptq, gpsqq ´ dpt, sq| ď dpgptq, gn,n ptqq
`|dpgn,n ptq, gn,n psqq ´ dpt, sq| ` dpgn,n psq, gpsqq
“ dpgptq, gn,n ptqq ` dpgn,n psq, gpsqq
since gn, n is an isometry. The right hand side now tends to 0 as n Ñ 8, and it follows
that g is an isometry. Furthermore, gn,n Ñ g in dIsom . To see this, suppose ą 0. Then
by compactness of T and density of ptj qj there is some J P N such that tBptj , q : j ď Ju
is an open cover of T . Let N P N be such that dpgn,n ptj q, gptj qq ă for all 1 ď j ď J and
n ě N . Hence, for all t P T there is some j ď J with t P Bptj , q and hence
dpgn,n ptq, gptqq ď dpgn,n ptq, gn,n ptj qq ` dpgn,n ptj q, gptj qq ` dpgptj q, gptqq ă 3.
We conclude that gn,n Ñ g in dIsom as required.
If any group induces a transitive action of isometries on a space then the group of
isometries of that space is evidently transitive. Although this is a rare property it gives
rise to a rather useful uniqueness result for Haar measure.
Theorem 7.5 (Uniqueness of Haar measure). Suppose that T is a compact metric space,
G acts transitively and isometrically on T , and µ and ν are G-Haar probability measures
on T . Then µ “ ν.
Proof. By quotienting we may suppose that the kernel of G is trivial, and hence we view
G as a subgroup of IsompT q. We write G for the closure18 of G in IsompT q. Since IsompT q
18It is not completely trivial to think of an example of a transitive faithful action of a group G on a
compact metric space T where G is not a compact subspace of IsompT q. One way to arrive at such an
example from the classical groups is to consider actions on the complex sphere CSn´1 :“ tx P Cn : }x}`n2 “
1u.
The group Upnq acts isometrically on CSn´1 via pU, xq ÞÑ U x. (It is not the whole group of isometries,
because CSn´1 can be embedded isometrically into R2n . It is then` isometric to the sphere S 2n´1 , whose
2n
˘ 2
group of isometries is (isometrically isomorphic to) O2n . This is a 2 -dimensional sub-manifold of R4n ,
2
whereas Upnq, considered as a sub-manifold of R4n , is only n2 -dimensional.)
The group Upnq is also a closed subgroup of the group of all isometries, but if it were not we would
be done because it turns out it acts transitively. In fact it has a subgroup SUpnq, the kernel of the group
40 TOM SANDERS
is compact, G is also compact and the group G acts on the compact metric space G by
G ˆ G Ñ G; pg, φq ÞÑ gpφq :“ φ ˝ g ´1 .
This action is isometric since
dIsom pgpφq, gpψqq “ suptdpφpg ´1 ptqq, ψpg ´1 ptqqq : t P T u

“ suptdpφptq, ψptqq : t P T u “ dIsom pφ, ψq,
and it follows from Theorem 7.1 that there is a probability measure κ on G such that
ż ż
f pgpφqqdκpφq “ f pφqdκpφq for all f P CpGq and g P G.
Suppose that µ is a G-Haar probability measure on T .

Claim. µ is also a G-Haar probability measure on T i.e.
ż ż
f pgptqqdµptq “ f ptqdµptq for all f P CpT q and g P G.
Proof. Suppose that f P CpT q. Then f is uniformly continuous and so for all ą 0 there
is some δ ą 0 such that |f pxq ´ f pyq| ă whenever dpx, yq ă δ. Given g P G there is some
n´1
Upnq under the determinant map, which acts transitively on CS for n ą 1. It will be enough to have
the following claim (although the extension to all n ą 1 is not much harder).
Claim. SUp2q acts transitively on CS1 .
Proof. To see this, suppose that x P CS1 , and write x “ px1 , x2 q P C2 where |x1 |2 ` |x2 |2 “ 1. Then the
matrix
ˆ ˙
x1 ´x2
x2 x1
takes p1, 0q to x and is an element of SUp2q. Since SUp2q is a group it follows that we can take any x P CS1
to any y P CS´1 via p1, 0q. The claim is proved.
Putting ∆ :“ texpp2πiqqI : q P Qu, where I is the identity matrix, we see that
∆ “ texpp2πiθqI : θ P Ru,
and both ∆ and ∆ are subgroups commuting with SUpnq. Hence, letting Hn be the subgroup generated
by ∆ and SUpnq, we have that
Hn “ ∆ SUpnq and Hn “ ∆ SUpnq.
Since ∆ X SUpnq “ tIu we conclude that Hn ‰ Hn and hence Hn is not compact. However, H2 acts
transitively on CS1 by the Claim since it contains SUp2q and this shows H2 is a construction of the desired
type.
g 1 P G such that dIsom pg, g 1 q ă δ. Hence

ˇż ż ˇ ˇż ż ˇ
ˇ ˇ ˇ 1
ˇ
ˇ f pgptqqdµptq ´ f ptqdµptqˇ ă ˇ f pgptqqdµptq ´ f pg ptqqdµptqˇ
ˇ ˇ ˇ ˇ
ˇż ż ˇ
ˇ ˇ
` ˇˇ f pg 1 ptqqdµptq ´ f ptqdµptqˇˇ
ˇż ˇ
ˇ 1
ˇ
“ ˇ pf pgptqq ´ f pg ptqqqdµptqˇˇ ă .
ˇ
Since ą 0 was arbitrary the claim follows.

By the G-invariance of T we have
ż ż ż ż
f dµ “ κpGq f dµ “ f ptqdµptqdκpgq
T T G T
ż ż
“ f pgptqqdµptqdκpgq.
G T
The function G ˆ T Ñ F; pg, tq ÞÑ f pgptqq is continuous since f is continuous and

dpgn ptn q, gptqq ď dpgn ptn q, gptn qq ` dpgptn q, gptqq
ď dIsom pgn , gq ` dpgptn q, gptqq,
so the right hand side tends to 0 as n Ñ 8.
Since G is closed, it is compact and hence so is G ˆ T and so pg, tq ÞÑ f pgptqq is a
continuous function on a compact space. It follows from Fubini’s theorem19 that
ż ż ż ż
f pgptqqdµptqdκpgq “ f pgptqqdκpgqdµptq.
G T T G
Let s P T be a fixed element. By transitivity of G we see that for every t P T there is some
ht P G such that ht psq “ t. It follows that
ż ż ż
f pgptqqdκpgq “ f pht pgqptqqdκpgq “ f pgpsqqdκpgq
G G G
19We have in mind here the Fubini theorem of Bourbaki [Bou52], an easier reference for which may be
the paper [LW12] the purpose of which is a wide generalisation of Fubini’s theorem. In our language the
theorem is as follows.
Theorem ([LW12, 1.1]). Suppose that S and T are compact Hausdorff spaces and µ is a Baire measure
on S and ν is a Baire measure on T . Then there is a Baire measure κ on S ˆ T such that
ż ż ż
f ps, tqdµpsqdνptq “ f ps, tqdκps, tq “ f ps, tqdνptqdµpsq for all f P CpS ˆ T q.
There is a slight subtlety here that we do not see because the spaces we are considering are compact. In
general the product of two Baire σ-algebras is not the Baire σ-algebra of the product, and so the product
measure κ above is not in general just the product µ ˆ ν. Of course it does correspond to a tensor product
of the appropriate spaces of functionals and this is what the Theorem is capturing.
42 TOM SANDERS
by G-invariance of κ. Hence
ż ż ż ż
f dµ “ f pgpsqqdκpgqdµptq “ f pgpsqqdκpgq.
T T G G
The right hand side is independent of µ, and hence µ “ ν
Example 7.6 (Surface measure on the sphere). The pn ´ 1q-sphere is defined to be

S n´1 :“ tx P `n2 : }x}`n2 “ 1u,
so it is an pn´ 1q-dimensional surface in n-dimensional space. It naturally inherits a metric
from the norm on `n2 , and with this metric becomes a compact metric space. (The map
x ÞÑ }x}2`n2 is continuous and so S n´1 is closed. It is also bounded in a finite dimensional
space and so is compact.)
The group Autp`n2 q of automorphisms of `n2 (that is linear isometries of `n2 ) has a natural
action on S n´1 via
Autp`n2 q ˆ S n´1 Ñ S n´1 ; pφ, xq ÞÑ φpxq.
This is an isometric action by definition of the metric on S n´1 and it is transitive. Tran-
sitivity follows since if e1 , f1 P S n´1 then we can extend (by the Gramm-Schmidt process)
e1 to an orthonormal basis e1 , . . . , en of `n2 and f1 to an orthonormal basis f1 , . . . , fn of `n2 .
Then there is a well-defined linear isometry
n
ÿ n
ÿ
φ: `n2 Ñ `n2 ; λi ei ÞÑ λi fi ,
i“1 i“1
which has φpe1 q “ f1 .

Since the action is transitive it follows from Theorems 7.1 and 7.5 that there is a unique
Autp`n2 q-Haar probability measure on S n´1 ; we denote this σn´1 .
Example 7.7 (Random automorphisms of `n2 ). The group of automorphisms from Example
7.6 is a compact metric space in its own right. Indeed, we can view it as a subset ofLp`n2 , `n2 q,
and so endow it with a metric via
dpφ, ψq :“ }φ ´ ψ},
where the norm is the operator norm. Since `n2 is finite dimensional, Lp`n2 , `n2 q is finite
dimensional, and hence the unit ball is compact. On the other hand Autp`n2 q is a closed
subset20 of the unit ball. Indeed, if φn Ñ φ then }φn pxq} Ñ }φpxq} for all x P `n2 , but
}φn pxq} “ }x} and hence }φpxq} “ }x}. It follows that pAutp`n2 q, dq is a compact metric
space.
Now Autp`n2 q acts on itself isometrically via
Autp`n2 q ˆ pAutp`n2 q, dq Ñ pAutp`n2 q, dq; pφ, ψq ÞÑ φ ˝ ψ,
20Notethat it is not equal to the unit sphere in Lp`n2 , `n2 q equipped with the operator norm, since there
are certainly norm 1 maps `n2 Ñ `n2 that are not invertible.
since
}ψpφ ´ φ1 q} “ supt}ψpφ ´ φ1 qpxq} : }x} ď 1u
“ supt}pφ ´ φ1 qpxq} : }x} ď 1u “ }φ ´ φ1 }
for all φ, φ1 P Lp`n2 , `n2 q. More over every action of a group on itself by multiplication
is transitive and so by Theorems 7.1 and 7.5 there is a unique Autp`n2 q-Haar probability
measure on Autp`n2 q; we denote it µn .
When a group G is a compact metric group then it can be seen as acting isometrically
on itself by multiplication and the G-Haar measure on G is just called the Haar measure
on G. In fact more generally we can consider any group endowed with a locally compact
Hausdorff topology that it compatible with its group structure. Such a group acts on itself
and supports an invariant measure, called a Haar measure (see e.g. [Alf63]), although there
there is not (in general) a natural normalisation.
Example 7.8. The multiplicative group GLn pRq of invertible n ˆ n real matrices can be
2
considered as a topological subspace of Rn . This topology is locally compact and Hausdorff
and it is easy to check that the group operations are continuous. It follows that with this
topology GLn pRq becomes a locally compact Hausdorff group and it turns out that Haar
measure exists on GLn pRq via
ż
1 2
f ÞÑ f pAq n
dλn pAq
GLn pRq | detpAq|
2 2
where λn is Lebesgue measure on Rn . To see that this is left invariant note that if B “ CA
then the Jacobian J “ C b I where I is the identity in GLn pRq, indeed
n
ÿ BBij
Bij “ Cik Akj , and hence “ Cik δjl .
k“1
BAkl
Thus the Jacobian determinant is det J “ pdet Cqn , and by the usual rule of substitution
in integration (see, e.g. [Rud87, Theorem 7.26]) that
ż ż
1 n2 1 2
f pCAq n
dλ pAq “ f pBq n
| detpCq|n dλn pAq
GLn pRq | detpAq| GL pRq | detpBq|
ż n
1 2
“ f pBq n
dλn pBq,
GLn pRq | detpBq|
as claimed.
8. Concentration inequalities
Concentration of measure and concentration inequalities are incredibly powerful tools in
mathematics. There are a lot of references, but we shall start with some ideas of Kahane
[Kah60]; see [BK00, §1.1] or [Ver12, §5.2.3] for a modern presentation.
Given a probability space pΩ, Pq, we shall seek to understand it through its spaces of
real-valued random variables – that is measurable functions Ω Ñ R where R is thought of
44 TOM SANDERS
as endowed with the Lebesgue σ-algebra. One particularly nice type of random variables
present in some spaces are Gaussians. We say that X is Gaussian with mean µ and
variance σ 2 and write X „ N pµ, σ 2 q if
ż8
1
PpX ě tq “ ? expp´px ´ µq2 {2σ 2 qdx.
2π t
Gaussians are enormously important in many different ways, but for us now we shall be
interested in their tails which are very sparse. In particular, if X „ N p0, 1q then we have
the following related facts:
(i) (Tail estimates) we have the estimate21,
Pp|X| ě tq ď expp´t2 {2q for all t ě 0;
(ii) (Bounded moment growth) the moments of Gaussian’s can be computed explicitly
(see [GR00]) where22 we have
ˆ ˙1{p
? Γppp ` 1q{2q
}X}Lp pPq “ 2 whenever p ě 1.
Γp1{2q
These can be estimated using the fact that Γpxq “ px ´ 1qΓpx ´ 1q giving
?
}X}Lp pPq “ Op pq whenever p ě 1;
21To check this just note that for y, t ě 0 we have py ` tq2 ě y 2 ` t2 and so
ż8
2
Pp|X| ě tq “ ? expp´x2 {2qdx
2π t
ż8
2
“ ? expp´py ` tq2 {2qdy
2π 0
ż8
2
ď ? expp´py 2 ` t2 q{2qdy
2π 0
ż8
2 2
“ expp´t {2q ¨ ? exppý 2 {2qdy “ expp´t2 {2q
2π 0
whenever t ě 0.
22Here Γ denotes Euler’s gamma function defined by
ż8
Γptq “ xλ´1 expp´xqdx for all λ P R.
0
(iii) (Moment generating function) the moment generating function of X can also be
explicitly computed23
EexppλXq “ exppλ2 {2q for all λ P R.
The first of these properties is very useful – it tells us that the tail of X is rather sparse –
and it turns out that all of these properties are, in a certain sense, equivalent. That being
said many spaces do not support Gaussian random variables and we should like a rough
cousin that nevertheless reflects these features. To this end we shall say that a random
variable X having EX “ 0 is sub-Gaussian if there is some c ą 0 such that
EexppλXq ď exppc2 λ2 {2q for all λ P R,
and write24 SubpΩq for the set of random variables on Ω that are sub-Gaussian. Allied to
this we define the following quantity on sub-Gaussian random variables
}X}SubpΩq :“ inftc ą 0 : EexppλXq ď exppc2 λ2 {2q for all λ P Ru,
so that
SubpΩq “ tX : Ω Ñ R s.t. X is measurable and }X}SubpΩq ă 8u.
Example 8.1. Suppose that X is a random variable on the probability space Ω such that
X „ N p0, σq. Then X P SubpΩq and }X}SubpΩq “ σ. To see this note that if X „ N p0, σ 2 q
then ż
EexppλXq “ exppλxq expp´x2 {2σ 2 qdx
ż
“ exppλ σ {2q expp´px ´ λσ 2 q2 {2σ 2 dx “ exppλ2 σ 2 {2q.
2 2
It follows that }X}SubpΩq “ σ.

Given the notation it should not be a surprise that } ¨ }SubpΩq is a norm.
Lemma 8.2 (Sub-Gaussian norm). The set SubpΩq is a vector space over R and } ¨ }SubpΩq
defines a (semi-)norm25 on the space.
Proof. First, if }X}SubpΩq “ 0 then for each η ą 0 we have
EexppλXq ď exppη 2 λ2 {2q for all λ P R.
For each λ P R we can take the limit as η Ñ 0 and conclude that
EexppλXq ď 1 for all λ P R.
23We have
ż8
1
EexppλXq “ ? exppλxq expp´x2 {2qdx
2π ´8
ż8
1 2
“ ? exppλ {2q expp´px ´ λq2 {2qdx “ exppλ2 {2q
2π ´8
for all λ P R.
24This is not standard notation.
25What we mean here is that }X}
SubpΩq “ 0 implies that X “ 0 almost everywhere.
46 TOM SANDERS
Now, suppose that ą 0. Then

PpX ą q ď EexppλpX ´ qq “ expp´λqEexppλXq ď expp´λq.
Since ą 0 the right hand side tends to 0 as λ Ñ 0. We conclude that PpX ą q “ 0 and
similarly that PpX ă q “ 0. Hence
˜ ¸
8
ď
PpX ‰ 0q “ P tω : |Xpωq| ą 1{nu “ 0,
n“1
by continuity of probability measures. It follows that X “ 0 almost everywhere as claimed.

(Conversely if X “ 0 almost everywhere then clearly }X}SubpΩq “ 0.)
The rest of the lemma will follow if we can show that whenever X, Y P SubpΩq we have
X ` Y P SubpΩq and }X ` Y }SubpΩq ď }X}SubpΩq ` }Y }SubpΩq , and whenever X P SubpΩq
and α P R we have αX P SubpΩq and }αX}SubpΩq “ |α|}X}SubpΩq .
First, suppose that X, Y P SubpΩq. Then for any σX ą }X}SubpΩq and σY ą }Y }SubpΩq
we have
2 2
EexppλXq ď exppσX λ {2q and EexppλY q ď exppσY2 λ2 {2q
for all λ P R.
By Hölder’s inequality applied with conjugate exponents p and q we then have
EexppλpX ` Y qq ď pEexpppλXqq1{p pEexppqλY qq1{q
` 2 2 2
˘1{p ` ˘1{q
ď exppσX p λ {2q exppσY2 q 2 λ2 {2q
2
“ exppppσX ` qσY2 qλ2 {2q.
Taking p “ pσX ` σY q{σX and q “ pσX ` σY q{σY we get that
EexppλpX ` Y qq ď expppσX ` σY q2 λ2 {2q,
and it follows that X ` Y P SubpΩq and }X ` Y }SubpΩq ď σX ` σY . Taking infima over
admissible σX , σY s gives the triangle inequality.
Finally, homogeneity is pretty straight forward since if α P R and σX ą }X}SubpΩq then
2 2
EexppλXq ď exppσX λ {2q for all λ P R,
and so
2 2 2
EexppλpαXqq “ EexppλαXq ď exppσX α λ q,
and hence }αX}SubpΩq ď |α|σX , and αX P SubpΩq. Taking infima over admissible σX s it
follows that }αX}SubpΩq ď |α|}X}SubpΩq . Since X and α were arbitrary and we now know
αX P SubpΩq we also have }X}SubpΩq ď |α|´1 }αX}SubpΩq and so }αX}SubpΩq “ |α|}X}SubpΩq .
The Lemma is proved.
In fact it turns out that SubpΩq is a Banach space as we shall see shortly. First, however,
we look at another example.
Example 8.3. Suppose that X P L8 pPq and EX “ 0. Then }X}SubpΩq ď }X}L8 pPq .
To see this note first that since } ¨ }SubpΩq is homogenous it suffices to consider the case
when }X}L8 pPq “ 1 (the case X “ 0 almost everywhere being trivial). Now, exppλyq ď
cosh λ ` y sinh λ for all λ P R and ´1 ď y ď 1. Hence
EexppλXq ď Epcoshpλq ` X sinhpλqq “ cosh λ ď exppλ2 {2q
for all λ P R, and so }X}SubpΩq ď 1. The claim is proved.
What makes the sub-Gaussian norm so powerful is the way that it interacts with inde-
pendence, and to that end we have the following lemma which is sometimes described as
rotation invariance.
Lemma 8.4. Suppose that X, Y P SubpΩq are independent. Then
b
}X ` Y }SubpΩq ď }X}2SubpΩq ` }Y }2SubpΩq .
Proof. Suppose that σX ą 0 and σY ą 0 are such that for all λ P R we have
2 2
EexppλXq ď exppσX λ {2q and EexppλY q ď exppσY2 λ2 {2q.
Then by independence we have
2
EexppλpX ` Y qq “ EexppλXqEexppλY q ď expppσX ` σY2 qλ2 {2q
a
2
for all λ P R. It follows that }X ` Y }SubpΩq ď σX ` σY2 and taking infima over admissible
σX , σY s gives the result.
This lemma is hugely powerful. Consider a two point space Ω “ t´1, 1u with measure
assigning equal mass to each point (and σ-algebra the power-set of Ω). The space Ωn does
not support a Gaussian because it is finite, but it does support a lot of bounded random
variables. In particular, consider the coordinate functions Xi : Ωn Ñ R; x ÞÑ xi . These are
random variables with
EXi “ 0 and }Xi }L8 pPq “ 1 for all 1 ď i ď n.
It follows from the triangle inequality that
n
ÿ n
ÿ
} Xi }SubpΩq ď }Xi }SubpΩq ď n,
i“1 i“1
but because all of these random variables are independent Lemma 8.4 tells us that
˜ ¸1{2
n
ÿ ÿn
} Xi }SubpΩq ď }Xi }2SubpΩq ď n1{2 .
i“1 i“1
Of course, this leaves the question of what we do with this. The norm } ¨ }SubpΩq was defined
to copy the moment generating function of random variables having a normal distribution.
As we saw at the start of the section those random variables also have good tail estimates
and it is this consequence of small sub-Gaussian norm that we are most interested in. The
next lemma collects this idea.
48 TOM SANDERS
Lemma 8.5. Suppose that X P SubpΩq. Then the following are equivalent in the sense
that each implies the other with Ki “ OpKj q.
(i) (Bounded L2k -norm growth)
?
}X}L2k pPq ď K1 k for all k P N;
(ii) (Bounded Lp -norm growth)
?
}X}Lp pPq ď K2 p for all p ě 1;
(iii) (Bounded exponential mean)
EexppX 2 {2K32 q ď 2;
(iv) (Chernoff tail estimate)
Pp|X| ą tK4 q ď 2 expp´t2 {2q for all t ą 0;
(v) (Moment generating function)
}X}SubpΩq ď K5 .
Proof. Of course (i) implies (ii) with K2 ď K1 , and conversely (ii) implies (i) with K1 ď
?
2K2 .
We shall now show that (i) implies (iii) implies (iv) implies (i), and (i) implies (v) implies
(iv) implies (i) and we shall be done.
Claim. (i) implies (iii) for some K3 ď OpK1 q.
?
Proof. We choose K3 “ OpK1 q (e.g. K3 “ 2 expp1qK1 ) such that
8 ˆ ˙k
ÿ expp1qK1
? ď 2.
k“0
2K3
Since k! ě k k expp´kq for all k P N we have
8 8 8 ˆ ˙2k
ÿ 1 1 2k
ÿ 1 K12k k k ÿ expp1qK1
EX ď ď ? .
k“0
k! 2k K32k k“0
k! 2k K32k k“0
2K3
Taking polynomial approximations to exppX 2 {2K32 q we can apply the Monotone Conver-
gence Theorem [Rud87, 1.26] to get
˜ ˙2k ¸ ÿ
8 ˆ 8
ÿ 1 X 1 1
EexppX 2 {2K32 q “ E ? “ k K 2k
EX 2k ď 2;
k“0
k! 2K3 k“0
k! 2 3
it follows that (iii) holds for some K3 “ OpK1 q as required.

Claim. (iii) implies (iv) for some K4 ď K3 .
Proof. For any t ą 0 we have
Pp|X| ą tK3 q exppt2 {2q ď exppt2 {2qEexppp|X|2 ´ t2 K32 q{2K32 q
“ EexppX 2 {2K32 q ď 2,
from which (iv) follows for some K4 ď K3 .
Claim. (iv) implies (i) for some K1 “ OpK4 q.

Proof. We have
ż8
}X}2k
L2k pPq “ 2ks2k´1 Pp|X| ą sqds
0
ż8 ż8
2k 2k´1 2k
“ K4 2ks Pp|X| ą sK4 qds ď 2K4 2ks2k´1 expp´s2 {2qdt.
0 0
The integral now just corresponds to moments of a normal distribution with mean 0 and
variance 1, and because they are integral moments they are easy to compute by parts. For
r ą 1 we have
ż8 ż8
r 2
“ r´1 2
‰8
s expp´s {2qds “ ´s expp´s {2q 0 ` pr ´ 1qsr´2 expp´s2 {2qds
0 0
ż8
“ pr ´ 1q sr´2 expp´s2 {2qds;
0
hence by induction we have }X}2k
L2k pPq “ OpK42 kqk . Thus, we have that (i) holds for some
K1 “ OpK4 q.
Claim. (i) implies (v) for some K5 “ OpK1 q.
Proof. For any λ P R we have
8 ˆ ˙k
2
ÿ 1 expp2qλ2 K12
exppOpλ K12 qq “
k“0
k! 4
8 ˆ ˙k
ÿ expp2qλ2 K12
ě
k“0
4k
8 8
ÿ expp2kqλ2k E|X|2k ÿ λ2k EX 2k
ě ě ,
k“0
p2kq2k k“0
p2kq!
and hence, by the Monotone Convergence Theorem again [Rud87, 1.26], we have
8
ÿ pλXq2k
EcoshpλXq “ E ď exppOpλ2 K12 qq.
k“0
p2kq!
Of course 2 cosh y ` y ´ 1 ě exppyq for all y P R and so, since EλX “ 0, we have
EexppλXq ď 2EcoshpλXq ´ 1 ď 2 exppOpλ2 K12 qq ´ 1 “ exppOpλ2 K12 qq.
It follows that (v) holds for some K5 “ OpK1 q.
Claim. (v) implies (iv) for some K4 ď K5 .
Proof. For every σX ą K5 we have (for λ ą 0) that
PpX ą tσX q ď EexppλpX ´ tσX qq
“ expp´tλσX qEexppλXq
2 2
ď expp´tλσX ` σX λ {2q “ expp´t2 {2q expp´pλσX ´ tq2 {2q.
50 TOM SANDERS
Optimising by taking λ “ t{σX (which is positive) gives

PpX ą tσX q ď expp´t2 {2q.
A similar argument tells us that
PpX ă ´tσX q ď expp´t2 {2q,
and hence
Pp|X| ą tσX q ď 2 expp´t2 {2q.
Letting En :“ tω : |X| ą tpK5 ` 1{nqu we see that 8
Ť
n“1 En “ tω : |X| ą tK5 u and so by
continuity of probability we have
n
ď
Pp|X| ą tK5 q “ lim Pp En q “ lim PpEn q ď 2 expp´t2 {2q,
nÑ8 nÑ8
m“1
and we have (iv) for some K4 ď K5 .

In light of the above it is natural to define another norm on SubpΩq by
}X}SG1 :“ suptp´1{2 }X}Lp pPq : p ě 1u,
and it turns out this is an equivalent norm. One of the advantages of this definition is that
it can be more easily extended to random variables without mean 0. This is not essentially
more general but can be more convenient.
Lemma 8.6. The space SubpΩq equipped with the norm } ¨ }SubpΩq is complete – it is a
Banach space.
Proof. Suppose that pXn qn is a Cauchy sequence in } ¨ }SubpΩq . By the triangle inequality
it follows that p}Xn }SubpΩq qn is Cauchy in the reals, and hence converges and is bounded
above by some constant S i.e. S is such that }Xn }SubpΩq ď S for all n P N.
Lemma 8.5 part (ii) tells us that } ¨ }L2 pPq is dominated by } ¨ }SubpΩq . It follows that pXn qn
converges to some X in L2 , and hence it converges almost everywhere and |Xpωq| ă 8
almost everywhere.
We now fix λ P R. Since x ÞÑ exppλxq is continuous we conclude that exppλXn q converges
to exppλXq almost everywhere, and exppλXq is finite almost everywhere.
The collection pexppλXn qqn is uniformly integrable. To see this write Yn :“ exppλXn q,
which is a non-negative random variable, and note that
EYn 1t|Yn |ěKu ď K ´1 E|Yn |2 “ K ´1 Eexpp2λXn q
ď K ´1 expp}Xn }2SubpΩq λ2 {2q ď K ´1 exppS 2 λ2 {2q.
It follows that for K sufficiently large as a function of S and λ, the left hand side is less
than any given . Thus pYn qn “ pexppλXn qqn is uniformly integrable as claimed.
It follows by the Uniform Integrability Theorem (see e.g. [Wil91, Theorem 13.7] and,
if necessary, [Wil91, Lemma 13.5] to pass from almost sure convergence to convergence in
probability) that exppλXn q Ñ exppλXq in L1 pPq and hence
EexppλXq “ lim EexppλXn q ď lim exppλ2 }Xn }2SubpΩq {2q
nÑ8 nÑ8
ˆ ´ ¯2 ˙
2
“ exp λ lim }Xn }SubpΩq {2
nÑ8
by continuity of x ÞÑ exppλ2 x2 {2q. Since limnÑ8 }Xn }SubpΩq is independent of λ, and λ was
arbitrary we conclude that X P SubpΩq and }X}SubpΩq ď limnÑ8 }Xn }SubpΩq .
One application of the above is the following so-called Chernoff-type result.
Proposition 8.7 (Chernoff-Hoeffding bound). Suppose that X1 , . . . , Xn are independent
random variables with mean µ and }Xi }L8 pPq ď 1. Then
n
ÿ
Pp| Xi ´ µn| ě µq ď 2 expp´2 {8q
i“1
Proof. By nesting of norms we have |µ| ď 1 and so }Xi ´ µ}L8 pPq ď 2, and hence }Xi ´
µ}SubpΩq ď 2 by Example 8.3. Since pXi qni“1 is an independent sequence of random variables,
we conclude that pXi ´ µqi is an independent sequence of random variables and hence by
Lemma 8.4 we have
› › › › d
n n n
›ÿ › ›ÿ › ÿ ?
› Xi ´ µn› “ › pXi ´ µq› }Xi ´ µ}2SubpΩq ď 2 n.
› › › ›
ď
›i“1 › ›i“1 › i“1
SubpΩq SubpΩq
By Lemma 8.5 (in fact the proof of the claim (v) implies (iv)) we have for all t ą 0 that
n
ÿ ?
Pp| Xi ´ µn| ě t2 nq ď expp´t2 {2q.
i“1
?
Taking t “ n{2 the result follows.
Moment generating functions were a key tool in the above arguments, and we can only
really hope to extend them to functions for which these exist, at least somewhere. Much
as with power series, when mgfs exist for some values, it follows that they exist for many
values.
Lemma 8.8. Suppose that X is a random variable with mean 0 and variance σ 2 , and there
are some λ1 ă 0 ă λ2 such that Eexppλi Xq ă 8. Then EexppλXq is a real analytic power
series with radius of convergence at least mint|λ1 |, |λ2 |u and
EexppλXq “ exppσ 2 λ2 {2 ` Oλ1 ,λ2 ;λÑ0 pλ3 qq,
in that region.
52 TOM SANDERS
Proof. Note that for λ in the given range we have

exppλXq ď exppλ1 Xq ` exppλ2 Xq
almost everywhere, and so by dominated convergence (see e.g. [Wil91, Theorem 5.9]) we
conclude that EexppλXq exists.
Now for the estimate at the origin. For |λ| ď λ0 we have
ˇ ˇ
N k kˇ N
λ X ˇ ÿ λk |X|k
ˇÿ
ď exppλ1 Xq ` exppλ2 Xq,
ˇ
ˇ ˇ ď
ˇk“0 k! ˇ k“0 k!
and hence by dominated convergence
8
k ´k
ÿ λk EX k
E|X| “ Oλ1 ,λ2 p|λ0 | k!q and EexppλXq “ .
k“0
k!
Given the bound on the growth of the moments it follows that the convergence on the right
is locally uniform whenever |λ| ă λ0 , and hence the right hand function is real analytic in
|λ| ă λ0 . Evaluating the first few terms of the power series we get
λ2
EexppλXq “ 1 ` λEX ` EX 2 ` Oλ1 ,λ2 pλ3 q
2
“ exppσ 2 λ2 {2 ` Oλ1 ,λ2 ;λÑ0 pλ3 qq,
since 1 ` x “ exppx ` OxÑ0 px2 qq. The second conclusion follows.
In light of this lemma we make the following definition. We say that a random variable
X with EX “ 0 is pσ 2 , bq-sub-exponential if
EexppλXq ď exppλ2 σ 2 {2q whenever |λ| ď 1{b
with the obvious convention for b “ 0 (which corresponds to the case of sub-Gaussian
random variables). Lemma 8.8 tells us that if X has an mgf at a positive and negative
value then it is pσ 2 , bq-sub-exponential for some parameters σ 2 and b, although we should
be clear that this σ 2 need not be the variance of X.
As a side remark it may be worth explaining that the name comes from the fact that if
the tail of a distribution decays exponentially, meaning there are constants c, C ą 0 such
that
Pp|X| ą tq ď exppćtq for all t ą C,
then it can be shown that X has a moment generating function for |λ| ă c. Indeed,
it is possible to prove an analogue of Lemma 8.5 showing that a distribution being sub-
exponential is essentially equivalent to it having an exponentially decaying tail (for t large),
and also equivalent to having the moment condition
}X}Lp pPq “ Oppq for all p ě 1.
In this regime moments determine the distribution of a random variable (see [Bil95, The-
orem 30.1]), whereas more generally they do not (see [Bil95, Example 30.2]).
There are two key analogues of our work on sub-Gaussian random variables which will
be useful. The first is a version of Lemma 8.4.
Lemma 8.9. Suppose that X1 , X2 are independent sub-exponential random variables with
parameters pσ12 , b1 q and pσ22 , b2 q respectively. Then X1 ` X2 is sub-exponential with param-
eters pσ12 ` σ22 , maxtb1 , b2 uq.
Proof. Simply note that
EexppλpX1 ` X2 qq “ EexppλX1 qEexppλX2 q ď exppλ2 σ12 {2q exppλ2 σ22 {2q
whenever |λ| ď 1{b1 and |λ| ď 1{b2 . The result follows.
We then have a crucial concentration result which takes into account the heavier tail
admissible in sub-exponential distributions.
Proposition 8.10. Suppose that X is a pσ 2 , bq-sub-exponential random variable. Then
#
expp´t2 {2q whenever 0 ď t ď σ{b
PpX ą tσq ď .
expp´tσ{2bq whenever t ą σ{b
Proof. The proof is just the proof of the claim (v) implies (iv) in Lemma 8.5. Specifically,
for 0 ď λ ď 1{b we have
PpX ą tσq ď EexppλpX ´ tσqq “ expp´λtσqEexppλXq
ď expp´λtσq exppλ2 σ 2 {2q
“ expp´t2 {2 ` pλσ ´ tq2 {2q.
If t ď σ{b then we can take λ “ t{σ and we get the first case. Otherwise, take λ “ 1{b and
we have
PpX ą tσq ď exppσ{bpσ{2b ´ tqq ď expp´tσ{2bq
since t ě σ{b and so pσ{2b ´ tq ă ´t{2. The result is proved.
Corollary 8.11. Suppose that X is a pσ 2 , bq-sub-exponential random variable. Then
Pp|X| ą tσq ď 2 maxtexpp´t2 {2q, expp´tσ{2bqu.
Proof. This is immediate from the triangle inequality and Proposition 8.10 applied to X
and ´X (the latter is easily seen to be pσ 2 , bq-sub-exponential).
It will now be useful for us to record some rather important examples of sub-exponential
random variables that are not Gaussian.
Example 8.12 (χ2 -distributions). Suppose X is a random variable with X „ N p0, 1q.
Then Y :“ X 2 ´ 1 has mean 0 and is p4, 4q-sub-exponential. To see this simply note that
EexppλY q “ expp´λqEexppλX 2 q
ż8
1
“ expp´λq ¨ ? exppλx2 q expp´x2 {2qdx
2π ´8
ż8
2 1
“ expp´λσ q ¨ ? expp´x2 p1 ´ 2λq{2qdx
2π ´8
ż8
1 expp´λq
“ expp´λq ¨ a exppú2 {2qdu “ ?
2πp1 ´ 2λq ´8 1 ´ 2λ
54 TOM SANDERS
?
provided λ ă 1{2. But expp´λq{ 1 ´ 2λ ď expp2λ2 q whenever |λ| ă 1{4 and we have the
claim.
This example is important because it will let us establish a concentration result for the
standard Gaussian. We define the standard centred Gaussian on `n2 to be the measure
γ n determined by
ż ż ˆ ˙
n 1 1 2
f pxqdγ pxq :“ ? n f pxq exp ´ }x}`n2 dx for all f P LBaire 8 p`n2 q,
2π 2
and dx is the usual measure on `n2 i.e. Lebesgue measure on Rn restricted to Baire sets.
It may be worth noting here that we have only defined Baire sets for compact metric
spaces. They can also be defined for locally compact Hausdorff spaces: the Baire σ-algebra
is minimal σ-algebra such that all the continuous functions having compact support are
measurable. It is a sub-algebra of the Lebesgue σ-algebra; we choose to restrict to it to make
it compatible with our work so far. As with compact spaces, by definition the continuous
functions of compact support are then dense in LBaire
8 in the topology of bounded pointwise
n
convergence. Of course technically The notation of γ should be suggestive of the idea that
it arises as a product. ? n´1
n
The vast majority of the measure? γ is concentrated on the sphere nS , i.e. the
sphere in n-dimensions of radius n.
Proposition 8.13 (Concentration of Gaussian measure). For P p0, 1s we have
γ n ptx P `n2 : |}x}2`n2 ´ n| ą nuq ď 2 expp´2 n{8q.
Proof. The space `n2 as a probability space when endowed with the measure γ n , and the co-
ordinate projection maps Xi : `n2 Ñ R; x ÞÑ xi are mutually independent random variables
with Xi „ N p0, 1q. It follows from Example 8.12 that Xi2 ´ 1 is p4, 4q-sub-exponential, and
so by Lemma 8.9 that x ÞÑ }x}2`n2 ´ n is p4n, 4q-sub-exponential. Corollary 8.11 then gives
us that
? ?
Pp|}x}2`n2 ´ n| ą t2 nq ď 2 maxtexpp´t2 {2q, expp´t n{4qu
?
for any t ą 0 and the result follows on setting t “ n{2.
In fact the constant in the exponent can be improved (see [Bar05, Corollary 2.3]) but
we shall not pursue this here.
We shall use the above to help us push results for Gaussians onto results for spheres.
The key example of this will be with projections. As before we write
Pk : `n2 Ñ `n2 ; px1 , . . . , xn q ÞÑ px1 , . . . , xk , 0, . . . , 0q.
A question which we shall be interested in is what happens to the norm?of elements x P S n´1
under projection by Pk . Since the measure of γ n is concentrated on nS n´1 we expect x
picked with γ n to have }x}2 « n. On the other hand
ż ż
f pPk xqdγ pxq‘ “ ’ f pyqdγ k pyq
n
n n
for all suitable f , where the inverted commas reflect the fact P
? k maps `2 Ñ `2 rather than
`n2 Ñ `k2 . Thus we might expect Pk x to be concentrated on ?kS k´1 so that }Pk x}2 « k.
To summarise, then, we expect that if x P S n´1 then }Pk x} « kn.
To capture the inverted comma part of this argument we have the following corollary.
Corollary 8.14. For P p0, 1s we have
γ n ptx P `n2 : |}Pk x}2`n2 ´ k| ą kuq ď 2 expp´2 k{8q.
Proof. We have
γ n ptx P `n2 : |}Pk x}2`n2 ´ k| ą kuq “ γ k ptx P `k2 : |}x}2`n2 ´ k| ą kuq ď 2 expp´2 k{8q
by Proposition 8.13
To capture the push forward from Gaussian concentration to the sphere involves our
work on Haar measure.
Proposition 8.15. For all P p0, 1s we have the estimate
´! ˇn ˇ )¯
σn´1 x P S n´1 : ˇ }Pk x}2`n2 ´ 1ˇ ą ď 2pexpp´2 k{72q ` expp´2 n{72qq.
ˇ ˇ
k
Proof. We shall define a measure τ on S n´1 as a pushforward of the standard centred
Gaussian. We then establish this concentration result for this measure τ , before showing
that it is invariant under the action of Autp`n2 q. By uniqueness of Haar measure this will
force τ to be σn´1 .
Claim. There is a Baire probability measure τ on S n´1 such that
ż ż
(8.1) f dτ “ f px{}x}qdγ n pxq for all f P LBaire
8 pS n´1 q.
`n
2 zt0u
Proof. First we check that the right hand side of the above is well-defined. For each m P N
and f P LBaire
8 pS n´1 q define
ˆ ˙
n x
km pf q : `2 Ñ R; x ÞÑ p1 ´ gm pxqqf φm pxq,
}x}`n2 ` 1{m
where pgm qm is a sequence of continuous non-negative functions mapping into r0, 1s and
tending pointwise to 1`n2 zt0u , and pφm qm is a sequence of continuous non-negative functions
of compact support mapping into r0, 1s and tending pointwise to 1`n2 . If g P CpS n´1 q has
}g} ď 1 then km pgq P Cc p`n2 q and }km pgq} ď 1, and hence if f P LBaire 8 pS n´1 q has }f } ď 1
then we have km pf q P LBaire
8 p`n2 q and }km pf q} ď 1, since LBaire 8 pS n´1 q is the closure of
n´1
CpS q in the topology of bounded pointwise convergence, and similarly for LBaire 8 p`n2 q.
On the other hand the integrand in (8.1) is the point-wise limit of functions km pf q, and
hence is itself an element of LBaire
8 p`n2 q. γ n is a finite Baire measure and hence the integral
is well defined. It follows that
ż
Φpf q :“ f px{}x}`n2 qdγ n pxq for all f P LBaire
8 pS n´1 q
`n
2 zt0u
56 TOM SANDERS
is a well-defined linear functional on LBaire

8 pS n´1 q. By the Riesz-Kakutani Theorem 3.7
applied to Φ restricted to CpS n´1 q there is a finite Baire measure τ on S n´1 such that
ż
Φpf q “ f dτ for all f P CpS n´1 q.
Since the closure of CpS n´1 q in the topology of bounded pointwise convergence is LBaire
8 pS n´1 q,
the bounded domination theorem gives the equality of the claim. Finally, it is immediate
that τ is a non-negative measure because the functional is non-negative. Similarly if f is
identically 1 then Φpf q “ 1 whence it is a probability measure.
We write F for the indicator function of the set we are interested in i.e.
# ˇ ˇ
ˇn 2
1 if ˇ k }Pk x}`n2 ´ 1ˇ ą
ˇ
n´1
F :S Ñ R; x ÞÑ
0 otherwise.
which is Baire-measurable since x ÞÑ }Pk x}2`n2 is continuous. By the claim defining τ we
have
ż ˜# ˇ › ›2 ˇ +¸
ˇn › x › ˇ
F dτ “ γ n x P `n2 zt0u : ˇ ››Pk › ´ 1ˇ ą
ˇ ˇ
ˇk }x} ›`n ˇ
2
´! ˇn ˇ )¯
“ γ n x P `n2 zt0u : ˇ }Pk x}2`n2 ´ }x}2`n2 ˇ ą }x}2`n2
ˇ ˇ
´! ˇn k ˇ )¯
n n ˇ 2 2 ˇ 2
“ γ x P `2 : ˇ }Pk x}`n2 ´ }x}`n2 ˇ ą }x}`n2 ,
k
since γ n pt0uq “ 0. Now, by the triangle inequality
ˇn ˇ ˇn ˇ ˇ ˇ
ˇ }Pk x}2`n2 ´ }x}2`n2 ˇ ď ˇ }Pk x}2`n2 ´ nˇ ` ˇ}x}2`n2 ´ nˇ ,
ˇ ˇ ˇ ˇ ˇ ˇ
k k
2
so if the left hand side is at least }x}`n2 then either
ˇ ˇ ˇn ˇ
ˇ 2
ˇ}x}`n2 ´ nˇ ą }x}2`n2 or ˇ }Pk x}2`n2 ´ nˇ ą }x}2`n2 .
ˇ ˇ ˇ
2 k 2
Since the first possibility implies
ˇ
ˇ 2
ˇ
ˇ}x}`n2 ´ nˇ ą n ě n,
ˇ
2` 3
we conclude that either this holds, or else
ˇn ˇ p1 ´ 3 q
ˇ }Pk x}2`n2 ´ nˇ ą }x}2`n2 ě n ě n.
ˇ ˇ
k 2 2 3
It follows that
ż
n
´!
n ˇ
ˇ 2 ˇ )¯ n
´!
n ˇ
ˇ
2
ˇ )¯
F dτ ď γ x P `2 : }x} ´ n ą n ` γ x P `2 : ˇ}Pk x}`n2 ´ k ˇ ą k .
ˇ ˇ
3 3
The required estimate now follows from Proposition 8.13 and Corollary 8.14.
To complete the proposition we now establish invariance of τ
Claim. The measure τ is invariant under the action of Autp`n2 q i.e.

ż ż
f pφpxqqdτ pxq “ f pxqdτ pxq for all f P LBaire
8 pS n´1 q, φ P Autp`n2 q.
Proof. First it is intuitively obvious (and not too difficult to prove) that γ n is invariant
under the action of Autp`n2 q.
Sub-claim. For all φ P Autp`n2 q we have
ż ż
f pφpxqqdγ pxq “ f pxqdγ n pxq for all f P LBaire
n
8 p`n2 q.
Proof. Consider the change of variables y “ φpxq. We have

n
ÿ Byi
yi “ xei , φpek qyxk , and hence “ xei , φpek qy,
k“1
Bxk
thus the Jacobian determinant is detpxei , φpek qyqi,j “ detpφq. Of course | det φ|2 “ detpφ˚ φq “
det ι “ 1. Moreover, since φ P Autp`n2 q we have }φ´1 y}`n2 “ }y}`n2 , and by the usual rule of
substitution in integration [Rud87, Theorem 7.26] (which applies since φ is one-to-one and
differentiable) we have
ż ż ˆ ˙
n 1 2
f pφpxqqdγ pxq “ f pφpxqq exp ´ }x}`n2 dx
2
ż ˆ ˙
1 ´1 2
“ f pyq exp ´ }φ y}`n2 | det φ´1 |dy
2
ż ˆ ˙ ż
1 2
“ f pyq exp ´ }y}`n2 dy “ f pyqdγ n pyq,
2
and the sub-claim is proved.
Now, if φ P Autp`n2 q then the claim defining τ and the previous sub-claim tells us that
ż ż
f pφpzqqdτ pzq “ f pφpxq{}x}qdγ n pxq
n
` zt0u
ż2
“ 1`n2 zt0u pφpxqqf pφpxq{}φpxq}qdγ n pxq
ż
“ 1`n2 zt0u pxqf px{}x}qdγ n pxq
ż
“ f pzqdτ pzq.
We conclude that τ is invariant under the action of Autp`n2 q on S n´1 completing the proof
of the claim.
By Theorem 7.5 and the definition of σn´1 in Example 7.6 we have that τ “ σn´1 and
the result is proved.
58 TOM SANDERS
With these results in hand we now turn to the Johnson-Lindenstrauss theorem [JL84].
Theorem 8.16 (Johnson-Lindenstrauss Theorem). Suppose that x1 , . . . , xn are elements
of a Hilbert space H. Then for all P p0, 1s there is an orthogonal projection π : H Ñ H
such dim Im π “ Op´2 log nq and
c
n
p1 ´ q}xi ´ xj } ď }πpxi q ´ πpxj q} ď p1 ` q}xi ´ xj } for all i, j.
k
Proof. There are many approaches to this result but they all revolve around the same set
of ideas; we shall follow [Bar05]. The plan is to take a random projection of H onto an
Op´2 log nq-dimensional subspace and show that with high probability the norms ofa any
2 n
set of n elements on the unit sphere are just scaled by a factor of almost exactly k
.
These elements will be the pairs xi ´ xj and this will give us the result.
There are two key observations: first, we can pick a random element of the sphere by
picking a random automorphism of `n2 and applying it to a fixed element of the sphere (we
shall prove this rigorously in the claim below); secondly, we can pick a random projection,
by picking a random automorphism of `n2 and composing it with a fixed projection.
By restricting to the space generated by x1 , . . . , xn we may certainly assume that dim H ď
n. On the other hand any m-dimensional Hilbert space is isometric to `m 2 (just take an
orthonormal basis of the space and map it to the canonical basis of `2 ), and `m
m
2 embeds
isometrically into `n2 .
Recall that µn is the Autp`n2 q-Haar probability measure on Autp`n2 q defined in Example
7.7.
Claim. For all x P S n´1 we have
ż ż
f pφpxqqdµn pφq “ f pyqdσn´1 pyq for all f P LBaire
8 pS n´1 q.
Proof. We consider the functional

ż
Φ: LBaire
8 pS n´1 q Ñ R; f ÞÑ f pφpxqqdµn pφq.
First, to see that it is well-defined we note that φ ÞÑ φpxq is continuous and so if f :

S n´1 Ñ R is continuous the φ ÞÑ f pφpxqq is continuous. Hence Φ is defined on CpS n´1 q
and by the definition of Baire sets and the bounded domination theorem we see that Φ is
a well-defined linear functional on LBaire
8 pS n´1 q. It is continuous and of norm 1 and so by
the Riesz-Kakutani theorem (Theorem 3.7) we see that there is a finite Baire measure τ
on S n´1 such that ż
Φpf q “ f dτ for all f P CpS n´1 q.
It is a probability measure since f ě 0 implies Φpf q ě 0 and Φp1q “ 1. More than this by
the definition of Baire sets and the bounded domination theorem we see that
ż ż
f pyqdτ pyq “ f pφpxqqdµn pφq for all LBaire
8 pS n´1 q.
Finally for all ψ P Autp`n2 q and f P LBaire

8 pS n´1 q we have
ż ż
f pψpyqqdτ pyq “ f pψpφpxqqqdµn pφq
ż ż ż
“ f pψpφqpxqqdµn pφq “ f pφpxqqdµn pφq “ f pyqdτ pyq.
It follows that τ is a Autp`n2 q-Haar probability measure on S n´1 . By Theorem 7.5 and
Example 7.6 it then follows that τ “ σn´1 as claimed.
For each pair pi, jq P rns2 let
" c *
n n
Ei,j :“ φ P Autp`2 q : p1 ´ q}xi ´ xj } ď }Pk φpxi ´ xj q} ď p1 ` q}xi ´ xj } ,
k
so that writing yi,j :“ pxi ´ xj q{}xi ´ xj } we have
ˆ" c *˙
n n
µn pEi,j q “ µn φ P Autp`2 q : p1 ´ q}yi,j } ď }Pk φpyi,j q} ď p1 ` q}yi,j }
k
´! ˇn ˇ )¯
n 2 2ˇ 2
ě 1 ´ µn φ P Autp`2 q : ˇ }Pk φpyi,j q} ´ }yi,j } ˇ ą }yi,j }
ˇ
´! ˇnk ˇ )¯
“ 1 ´ µn φ P Autp`2 q : ˇ }Pk φpyi,j q}2 ´ }φpyi,j q}2 ˇ ą }φpyi,j q}2
n ˇ ˇ
´! ˇ nk ˇ )¯
n´1 ˇ 2 2ˇ 2
“ 1 ´ σn´1 y P S : ˇ }Pk y} ´ }y} ˇ ą }y}
k
by the claim. By Proposition 8.15 we get
µn pEi,j q ě 1 ´ 2pexpp´2 k{72q ` expp´2 n{72qq,
and hence ˜ ¸
č
µn Ei,j ě 1 ´ n2 pexpp´2 k{72q ` expp´2 n{72qq
1ďiăjďn
by the union bound. It follows that there is some k “ Op´2 log nq such that with positive
probability all of the events Ei,j occur, and hence, taking π :“ Pk φ, a suitable projection
exists.
The bound here was shown to be tight up to a logarithmic factor by Alon [Alo03].
9. Khintchine’s inequality
As immediate corollary of Example 8.3, Lemma 8.4 and Lemma 8.5 from §8 we have
Khitnchine’s inequality.
Proposition 9.1 (Khintchine’s inequality). Suppose that p P r2, 8q and X1 , . . . , Xn are
random variables with PpXi “ ai q “ PpXi “ ái q “ 1{2. Then
¨ ˜ ¸1{2 ˛
ÿ ? ÿ
} Xi }Lp pPq “ O ˝ p }Xi }2L2 pPq ‚.
i i
60 TOM SANDERS
This can be bootstrapped to the following.
Theorem 9.2 (Marcinkiewicz-Zygmund inequality). Suppose ř that p P r2, 8q and we are

given independent random variables X1 , . . . , Xn P Lp pPq with E i Xi “ 0. Then
˜ ¸
ÿ ? ÿ 1{2
} Xi }Lp pPq “O p} |Xi |2 }Lp{2 pPq .
i i
Proof. For complex random variables the result follows from the real case by taking real
and imaginary parts and applying the triangle inequality.
We now proceed in two parts. First we prove the inequality with the Xi s assumed
symmetric (that is when Xi „ ´Xi ). We partition Ω according to the multi-index k P
pt´8u Y Zqn so that
Ωk “ tω P Ω : 2ki ď |Xi | ă 2ki `1 for all 1 ď i ď nu,
with the convention that Xi “ 0 if ki “ ´8. The sets Ωk are measurable since the Xi s
are measurable and if PpΩk q ‰ 0 we write Pk for the probability measure induced on Ωk
by P i.e. Pk pAq “ PpAq{PpΩk q for every measurable A Ă Ωk , and write Xi,k :“ Xi |Ωk .
Since Xi P Lp pPq we get Xi,k P Lp pPk q and hence, by nesting of norms, that EXi,k exists.
Symmetry and the definition of Ωk then tell us that
ż
1
EXi,k “ Xi dP
PpΩk q Ωk
ˆż ż ˙
1
“ Xi dP ` Xi dP
PpΩk q ´2ki `1 ăXi ď´2ki 2ki ďXi ă2ki `1
ˆ ż ż ˙
1
“ ´ ´Xi dP ` Xi dP “ 0,
PpΩk q 2ki ď´Xi ă2ki `1 2ki ďXi ă2ki `1
since Xi „ ´Xi .
By Example 8.3 we have }Xi,k }SubpΩk q ď 2ki `1 , and so by Lemma 8.4 we have
ÿ cÿ
} Xi,k }SubpΩk q ď 22pki `1q
i i
cÿ
ď 2 inft |Xi,k pωq|2 : ω P Ωk u.
i
It follows from Lemma 8.5 that

$ ˜ ¸p{2 , ˜ ¸p{2
ÿ & ÿ . ż ÿ
} Xi,k }pLp pPk q “ inf O p |Xi,k pωq|2 : ω P Ωk ď O p |Xi,k |2 dPk .
i
% i
- i
ř
Since Lp pPq is a normed space we have i Xi P Lp pPq and by the Dominated Convergence
theorem we have
ÿ ÿ ÿ
} Xi }pLp pPq “ } Xi |Ωk }pLp pPq
i kPpt´8uYZqn i
ÿ ÿ
“ } Xi,k }pLp pPk q PpΩk q
kPpt´8uYZqn :PpΩk q‰0 i
ż ˜ÿ ¸p{2
ÿ
“ Oppqp{2 PpΩk q |Xi,k |2 dPk
kPpt´8uYZqn :PpΩk q‰0 i
ÿ p{2
“ Oppqp{2 } Xi }Lp{2 pPq .
i
The claimed bound follows for the case of symmetric random variables.
Now we suppose that the variables X1 , . . . , Xn are given as in the hypotheses of the
proposition. We let Y1 , . . . , Yn be such that Xi „ Yi and X1 , . . . , Xn , Y1 , . . . , Yn are in-
dependent i.e. we consider the probability space pP2 , Ω2 q. We now apply the symmetric
result to the variables Xi ´ Yi to get that
˜ ¸
ÿ ? ÿ 1{2
} pXi ´ Yi q}Lp pPˆPq “ O p} |Xi ´ Yi |2 }Lp{2 pPˆPq
i i
˜ ¸
? ÿ 1{2
“ O p} |Xi |2 }Lp{2 pPˆPq .
i
ř
But then it follows from nesting of norms and the fact that E i Yi “ 0 that
ÿ ÿ ÿ
} Xi }Lp pPq “ } Xi ´ E Yi }Lp pPq
i i i
˜ ˇ ˇp ¸1{p
ˇÿ ÿ ˇ
1 ˇ
Eω ˇ Xi pωq ´ Eω1 Yi pω qˇ
ˇ
“
ˇ i i
ˇ
˜ ˇ ˜ ¸ˇp ¸1{p
ˇ ÿ ÿ ˇ
1
Xi pωq ´ Yi pω q ˇ
ˇ ˇ
“ Eω ˇEω1
ˇ i i
ˇ
˜ ˇ ˇp ¸1{p
ˇÿ ÿ ˇ
1 ˇ
Eω Eω ˇ Xi pωq ´ Yi pω qˇ
ˇ
ď 1
ˇ i i
ˇ
ÿ
“ } pXi ´ Yi q}Lp pPˆPq ,
i
and the result is proved.

For random variables satisfying the hypotheses of Khintchine’s inequality the Lp{2 -
norm on the right is an L1 -norm, and there is something close to this true for variables
62 TOM SANDERS
in the generality considered above called ř Rosenthal’s inequality. Indeed, suppose that
X1 , . . . , Xn P Lp pPq are independent and E i Xi “ 0. Then
¨ $˜ ¸1{p ,˛
ÿ p & ÿ ÿ .
(9.1) } Xi }Lp pPq “ O ˝ max }Xi }pLp pPq , } Xi }L2 pPq ‚.
i
log p % i i
-
For p large the second term in the max takes over and we recover a strengthening of
Khintchine’s inequality. Of course, precisely when this takes over depends on the specific
variables Xi and how large their Lp mass is compared to their L2 mass – that is how often
they take very large values.
The p dependence in (9.1) is best possible (up to the precise constant; see [JSZ85, Ute85,
FHJ` 97] for details), and it is weaker than that for the Marcinkiewicz-Zygmund inequality.
This fits with the fact that the critical distributions for Rosenthal’s inequality are Poisson
whereas for the Marcinkiewicz-Zygmund inequality they are Gaussians.
There are so called vector valued or Banach space valued variants of the above inequal-
ities. These start with Kahane’s proof [Kah64] of a vector-valued version of Khintchine’s
inequality. A vector-valued version of Rosenthal’s inequality was proved by Talagrand in
[Tal89], with a subsequent neater proof in [KS91]. We shall follow some of the ideas in this
latter paper to give a proof of Kahane’s result.
Extending Example 4.8, given a Banach space Z we write Lp pP; Zq for the set of Z-valued
measurable functions such that
}f }Lp pP;Zq :“ pExPΩ }f pxq}pZ q1{p ă 8;
the function } ¨ }Lp pP;Zq is a (semi-)norm. The integral here is called the Bochner integral
and enjoys many of the properties one might expect. It is worth noting that there is some
choice here regarding what we regard as the σ-algebra on Z. We could simply take the
Borel σ-algebra on Z induced by the norm-topology on Z. More natural for us is to take
the Baire σ-algebra, which is, of course, the minimal σ-algebra such that every continuous
f : Z Ñ R is measurable.
Suppose that L is a category in which each object is an Lp pPq-space for some probability
space pΩ, Pq, though not necessarily all Lp -spaces are in the category, and the morphisms
are short maps. Suppose further that b : L ˆ L Ñ L is a tensor product (with unit
Lp pδq, where δ is a δ-measure – this an Lp -space with a one-point probability space) on L
making it monoidal. (See the classic [Kel64] for the various coherence conditions required,
or [ML98] for a book covering the topic.) This can just be thought of as the usual tensor
product here, so
Lp pP1 q b Lp pP2 q :“ Lp pP1 ˆ P2 q,
which satisfies all the necessary coherence axioms. This could be slightly subtler for reasons
we have mentioned before: suppose that L contained spaces all of which are of the form
Lp pµq where µ is some Baire, not necessarily probability measure. There are examples of
Baire measures µ and ν where µ ˆ ν is not Baire, and so the tensor product could not be
defined as above. That being said there is a tensor product on this category and hopefully
that goes some way to explain why it is the notion of tensor product which we are focusing
on.
We now have a map
˚ : L ˆ Ban1 Ñ Ban1 ; pLp pPq, Zq ÞÑ Lp pP; Zq
which is an action of the monoidal category L on Ban because we have a natural isomor-
phism with components the isometric isomorphisms
αLp pP1 q,Lp pP2 q,Z : pLp pP1 q b Lp pP2 qq ˚ Z Ñ Lp pP1 q ˚ pLp pP2 q ˚ Zq
ppx, yq ÞÑ f px, yqq Ñ px ÞÑ py ÞÑ f px, yqqq,
and another natural isomorphism with components the isometric isomorphisms
λZ : Lp pδq ˚ Z Ñ Z; f ÞÑ f pωq,
where ω is ‘the’ point in the probability space on which δ is defined. The coherence
requirements for this can be found in [JK02, 1.1–1.3].
We have the following extension of Proposition 9.1.
Theorem 9.3 (Khintchine-Kahane inequality). Suppose that p P r2, 8q and X1 , . . . , Xn are
random variables taking values in a Banach space Z with PpXi “ ai q “ PpXi “ ái q “ 1{2.
Then ˜ ¸
ÿ ? ÿ
} Xi }Lp pP;Zq “ O p} Xi }L2 pP;Zq .
i i
The key point here is that the dependence in the big-O does not depend on the dimension
of the space generated by the ai s; if we allow that dependence then the above follows from
Khintchine’s inequality 9.1.
ř 1{2
It may be slightly surprising that the sum on the right is not p i }Xi }2Z q , however
it cannot be. Consider the example when Z :“ `8 pt0, 1un q and let ai : t´1, 1un Ñ R be
the vector with ai pxq “ 1 if xi “ 1 and ai pxq “ ´1 if xi “ ´1. Then }ai }Z “ 1 and if
x P t´1, 1un then
ÿ n ÿn
n ě } σi ai }Z ě xi ai pxq “ n.
i“1 i“1
It follows that if the Xi s are are in Theorem 9.3 then
˜ ¸1{p
ÿ ÿ p
} Xi }Lp pP;Zq “ E} Xi }Z “ n,
i i
ř 1{2 ?
while p i }Xi }2Z q “ n. It follows that if we wanted this quantity on the right we would
not be able to have a constant independent of n.
For the proof we shall work from the notes [Ver10, Chapter 2] and [Gar03]. The key
tool there (and in [KS91]) is an inequality due to Bonami [Bon70, Chaptaire III, Lemme
3, p378] (our Lemma 9.4). This was also proved by Beckner in [Bec75, Lemma 1], who
seems to have been unaware of the work of Bonami. Both Bonami and Beckner then note
that the inequality tensorises well in [Bon70, Chaptaire III, Lemme 1, p375] and [Bec75,
64 TOM SANDERS
Lemma 2] respectively; Beckner describes his Lemma 2 as a generalisation of an important

lemma of Nelson and Segal.
We shall start by proving, in the language of Beckner [Bec75], the two-point inequality
below. While the proof is rather careful, it is straightforward to prove something of this
type.
Lemma 9.4 (The ‘two-point’ inequality). Suppose that 2 ď p ă 8. Then for all 0 ď y ă x
we have
ˆ ˙1{p ˆ ˙1{2
px ` θp yqp ` px ´ θp yqp px ` yq2 ` px ´ yq2
ď
2 2
where θp :“ pp ´ 1q´1{2 .
Proof. Since x ą 0 we can divide out by x, and the inequality will follow if we can prove
it for x “ 1.
Before we go about the calculation proper it is worth noting that there is certainly some
choice of θ “ ΩpÑ8 pp´1{2 q such that
ˆ ˙1{p ˆ ˙1{2
p1 ` θyqp ` p1 ´ θyqp p1 ` yq2 ` p1 ´ yq2
ď
2 2
?
whenever 0 ď y ă 1. Indeed, put θ “ c{ p for some small (but absolute) c ą 0 then
ˆ ˙1{p ˆ ˙1{p
p1 ` θyqp ` p1 ´ θyqp exppθypq ` expp´θypq
ď
2 2
“ pcoshpθypqq1{p ď exppθ2 y 2 p{2qq “ 1 ` OpÑ8 pc2 y 2 q.
A sufficiently small choice of c will give the claim.
The proof of the actual inequality is not much harder than the above – it is just more
careful. First note that the inequality is equivalent to
ˆ ˙1{p
p1 ` zqp ` p1 ´ zqp a
ď 1 ` pp ´ 1qz 2 whenever 0 ď z ă θp ,
2
which will be easier to handle using calculus because we shall require fewer applications of
the Chain Rule. We put
ˆ ˙
1 p1 ` zqp ` p1 ´ zqp 1 ` ˘
φpzq :“ log ´ log 1 ` pp ´ 1qz 2 ,
p 2 2
and shall show that φpzq ď 0 for 0 ď z ă θp . First
pp1 ` zqp´1 ´ p1 ´ zqp´1 q pp ´ 1qz
φ1 pzq “ p p
´
p1 ` zq ` p1 ´ zq 1 ` pp ´ 1qz 2
p1 ` zqp´1 p1 ´ pp ´ 1qzq ´ p1 ´ zqp´1 p1 ` pp ´ 1qzq
“ .
pp1 ` zqp ` p1 ´ zqp qp1 ` pp ´ 1qz 2 q
Write ψpzq for the numerator of the last expression. Then
ψ 1 pzq “ pp ´ 1qp1 ` zqp´2 p1 ´ pp ´ 1qzq ´ pp ´ 1qp1 ` zqp´1

`pp ´ 1qp1 ´ zqp´2 p1 ` pp ´ 1qzq ´ pp ´ 1qp1 ´ zqp´1
“ ´ppp ´ 1qzpp1 ` zqp´2 ´ p1 ´ zqp´2 q.
It follows that ψ 1 pzq ď 0 for 0 ď z ă θp . On the other hand ψp0q “ 0, and so by the
Fundamental Theorem of Calculus we have ψpzq ď 0 for 0 ď z ă θp . It follows that
φ1 pzq ď 0 for 0 ď z ă θp , and since φp0q “ 0 we conclude that φpzq ď 0 for 0 ď z ă θp , and
the result is proved on exponentiating.
The key fact we use is that Lemma 9.4 can be tensored with any Banach space Z.
Lemma 9.5. Suppose 2 ď p ă 8 and X is a random variable with EX “ 0 and taking

values in t´1, 1u. Then for all z, w P Z we have
}z ` θp Xw}Lp pP;Zq ď }z ` Xw}L2 pP;Zq ,
where θp :“ pp ´ 1q´1{2 .
Proof. First, unpacking the notation we have

ˆ ˙1{p
}z ` θp w}p ` }z ´ θp w}p
}z ` θp Xw}Lp pP;Zq “
2
and
ˆ ˙1{2
}z ` w}2 ` }z ´ θp w}2
}z ` Xw}L2 pP;Zq “ .
2
By replacing w by ´w if necessary we may assume }z ` w} ě }z ´ w}. Moreover, since
both sides are continuous in w, it suffices to prove the inequality for w ‰ z. Thus we may
choose x ą y ě 0 such that
x ` y “ }z ` w} and x ´ y “ }z ´ w}.
Then
› ›
› 1 ` θp 1 ´ θ p › 1 ` θp 1 ´ θp
}z ` θp w} “ ›› pz ` wq ` pz ´ wq›› ď px ` yq ` px ´ yq “ x ` θp y.
2 2 2 2
and
› ›
› 1 ` θp 1 ´ θp › 1 ` θp 1 ´ θp
}z ´ θp w} “ ›
› pz ´ wq ` pz ` wq›› ď px ´ yq ` px ` yq “ x ´ θp y.
2 2 2 2
66 TOM SANDERS
These inequalities and Lemma 9.4 then give

ˆ ˙1{p ˆ ˙1{p
}z ` θp w}p ` }z ´ θp w}p px ` θp yqp ` px ´ θp yqp
ď
2 2
ˆ ˙1{p
px ` yqp ` px ´ yqp
ď
2
ˆ ˙1{p
}z ` w}p ` }z ´ w}p
“ .
2
Suppose that X1 , . . . , Xn is a finite collection of independent random variables with
expectation 0 and taking values in t1, ´1u. We write
ź
XS :“ Xi for all S Ă rns
iPS
with the usual convention for the empty product – it is the constant function equal to 1.
Note that the variables pXS qSĂrns are then orthonormal:
#
1 if S “ T ;
xXS , XT yL2 pPq “
0 otherwise.
These random variables generate a (Hilbert) subspace of L2 pPq:
$ ,
& ÿ .
SpanppXS qSĂrns q “ zS XS : zS P F for all S Ă rns .
% -
SĂrns
We shall be interested in the subspace of L2 pP; Zq generated by letting the zS s range

over element of some Banach space. Note that in this neither L2 pP; Zq nor the span are
necessarily Hilbert spaces.
Proposition 9.6. Suppose that X1 , . . . , Xn are independent random variables with EXi “
0 and taking values in t´1, 1u, pzS qSĂrns is a vector of elements of a Banach space Z, and
2 ď p ă 8. Then
› › › ›
› › › ›
› ÿ |S| › › ÿ ›
›
› θp X S zS
›
› ď ›
› X S zS
›
›
›SĂrns › ›SĂrns ›
Lp pP1 ˆ¨¨¨ˆPn ;Zq L2 pP1 ˆ¨¨¨ˆPn ;Zq
´1{2
where θp :“ pp ´ 1q .
Proof. We proceed by induction on n; for n “ 0 the result is vacuous. Suppose we have
proved the result for some n and any Banach space W . Suppose that X1 , . . . , Xn`1 are
independent random variables with EXi “ 0 and taking values in t´1, 1u and pzS qSĂrn`1s
is a vector of elements of Z, and put
wS :“ zS ` θp Xn`1 zSYtn`1u for all S Ă rns.
By the inductive hypothesis we have

› › › ›
› › › ›
› ÿ |S|
› › ÿ
|S|
›
›
› zS θ p X S
›
› “ ›
› w S θp X S
›
›
›SĂrn`1s › ›SĂrns ›
Lp pP1 ˆ¨¨¨ˆPn ;Zq Lp pP1 ˆ¨¨¨ˆPn ;Zq
› ›
› ›
› ÿ ›
ď ›› wS XS ››
›SĂrns ›
L2 pP1 ˆ¨¨¨ˆPn ;Zq
› ›
› ›
› ÿ ÿ ›
“ ›
› zS XS ` θXn`1 zSYtn`1u XS ›› .
›SĂrns SĂrns ›
By Lemma 9.5 applied with the random variable Xn`1 , the Banach space L2 pP1 ˆ¨ ¨ ¨ˆPn ; Zq
and vectors ÿ ÿ
z :“ zS XS and w :“ zSYtn`1u XS ,
SĂrns SĂrns
we get
¨ › ›p ˛1{p
› ›
› ÿ ÿ ›
zS XS ` θXn`1 zSYtn`1u XS ››
˚ ‹
˝EΩn ›
› ‚
¨ › ›2 ˛1{2
› ÿ ÿ ›
› ›
zS XS ` Xn`1 zSYtn`1u XS ››
˚ ‹
ď ˝EΩn ›› ‚
¨ › ›2 ˛1{2 › ›
› › › ›
› ÿ › › ÿ ›
zS XS ›› zS XS ›› .
˚ ‹
“ ˝EΩn ›
› ‚ “›
›
›SĂrn`1s › ›SĂrn`1s ›
L2 pP1 ˆ¨¨¨ˆPn ;Zq L2 pP1 ˆ¨¨¨ˆPn ˆPn`1 ;Zq
Combining this with what we showed earlier the result is proved since
¨ › ›p ˛1{p › ›
› › › ›
› ÿ › › ÿ ›
|S| |S|
zS θp XS › zS θp XS ››
˚ ‹
˝EΩn ›
› › ‚ “› ›
›SĂrn`1s › ›SĂrn`1s ›
Lp pP1 ˆ¨¨¨ˆPn ;Zq Lp pP1 ˆ¨¨¨ˆPn`1 ;Zq

Inequalities of the type in Proposition 9.6 are sometimes called Hypercontractive inequal-
ities following Nelson [Nel73] who proved an analogue for the Banach space Z “ R in one
dimension with X „ N p0, 1q with an optimal constant. The names Bonami’s inequality
and Beckner’s inequality are also used to describe results like Proposition 9.6 in various
degrees of generality.
The Khintchine-Kahane inequality is now a trivial corollary.
68 TOM SANDERS
Theorem (Khintchine-Kahane inequality, Theorem 9.3). Suppose that p P r2, 8q and

Y1 , . . . , Yn are random variables taking values in a Banach space Z with PpYi “ ai q “
PpYi “ ái q “ 1{2. Then
ÿ a ÿ
} Yi }Lp pP;Zq ď p ´ 1} Yi }L2 pP;Zq .
i i
Proof. Apply Proposition 9.6 with ztiu “ ai and zS “ 0 for all S Ă rns with |S| ‰ 1. This
tells us that › › › ›
›ÿ › ›ÿ ›
› θp Xi ai › ď › X i ai › .
› › › ›
› i › › i ›
Lp pP;Zq L2 pP;Zq
The result follows on dividing through by θp´1 .
The constants here are good, although they are not the best. The best constants for
Khintchine’s inequality when Z “ R has received considerable attention and [Haa81] con-
tains the state of the art. The question of the optimal constants in the vector valued
case has received less attention although there is a rather nice paper [LO94] of Latala and
Oleszkiewicz which addresses an important case.
10. Tensor products and Grothendieck’s inequality

In this last section of the course we are going to deal with tensor products and Grothendieck’s
inequality. This is a very important result in the are, and the interested readers are di-
rected to the survey of Pisier [Pis12], the book [Rya02] of Ryan, or the series of survey
articles [DFS02d, DFS02a, DFS02b, DFS02c].
To understand tensor products on Ban1 we shall view it as a multicategory (also known
as a coloured operad or pseudo tensor category). The reader may wish to consult [Lei04,
Chapter 2] for some discussion of classical multicategories. Tensor products are primarily
familiar from the (multi-)category of vector spaces (over a fixed field). Given two vector
spaces V and W over a field K their algebraic tensor product is defined to be the
# +
ÿn
V b W :“ vi b wi : vi P V, wi P W ,
i“1
with the natural vector space operations. This makes VectK into a monoidal category.
The tensor product itself has a universal property: for every vector space Z and bilinear
map ψ : V ˆ W Ñ Z there is a linear map ψr : V b W Ñ Z such that the diagram
ι
V ˆW V bW
ψr
ψ
Z
commutes, where ι : V ˆ W Ñ V b W ; pv, wq ÞÑ v b w, where b between vectors is the
exterior product.
To extend this notion to Ban1 we have to decide what our ‘bilinear maps’ are – the
multimorphisms of Ban1 when we extend it to be a multicategory. Different choices for
the multimorphisms will give rise to different tensor products. The obvious choice is for
them to be short bilinear maps meaning ψ : X ˆ Y Ñ Z is bilinear and
(10.1) }ψpx, yq}Z ď }x}X }y}Y for all x P X, y P Y.
Given Banach spaces X and Y we define their projective tensor product, denoted
X bY
p , to be the completion of X b Y with respect to the norm (and one should check that
this is a bonafide norm)
# +
n
ÿ n
ÿ
}u}^ :“ inf }xi }X }yi }Y : u “ xi b yi for all u P X b Y.
i“1 i“1
With this definition this projective tensor product has the required universal property:
indeed, suppose that ψ : X ˆ Y Ñ Z is a short bilinear map. Since ψ is, in particular,
bilinear there is a linear map φ : X b Y Ñ Z such that φpx b yq “ ψpx, yq for all x P X,
ř
y P Y . For each u P X ˆ Y define ψpuq
p :“ φpuq so that if u “ i xi b yi then
› ›
›ÿ › ÿ ÿ ÿ
}φpuq}Z “ › xi b yi › ď }φpxi b yi q}Z “ }ψpxi , yi q}Z ď }xi }X }yi }Y .
› ›
› i › i i i
Z
From the definition of the projective norm it follows that φ is a short map from X b Y
endowed with } ¨ }^ to Z. On the other hand X b Y is dense in X bYp in this norm and so
φ extends to a short linear map ψr : X bY
p Ñ Z as required.
If we work with Ban where the morphisms are all continuous linear maps then the
hom-sets – the set LpX, Y q – are themselves Banach spaces, and there is an internal hom
functor
Banop ˆ Ban Ñ Ban; pX, Y q ÞÑ LpX, Y q;
see the functor in [Man12, Definition 2.1] for more details.
The map
LpX bY,
p Zq Ñ LpY, LpX, Zqq; T ÞÑ py ÞÑ px ÞÑ T px b yqqq
is an isometric isomorphism and is natural in Y and Z. The fact that it is a bijection
means that Y ÞÑ X bYp is a left adjoint for Z ÞÑ LpX, Zq, and makes X bY
p into the tensor
product so defined.
We have already seen an example of a projective tensor product, we just gave it a
different name in that case.
Proposition 10.1 ([DFS02d, Theorem 1.10]; originally [Gro53, Theorem 3]). Suppose that
X is a Banach space. Then we have an isometric isomorphism from L1 pPqbX
p to L1 pP; Xq.
Proof. We consider the bilinear map
ψ : L1 pPq ˆ X Ñ L1 pP; Xq; pf, xq ÞÑ pω ÞÑ f pωqxq
which has
ż
}ψpf, xq}L1 pP;Xq “ |f pωq|}x}X dω “ }f }L1 pPq }x}X for all f P L1 pPq and x P X,
70 TOM SANDERS
so that (10.1) holds. It follows that there is a short map

ψp : L1 pPqbX
p Ñ L1 pP; Xq
such that
ψpf
p b xqpωq “ f pωqx for all ω P Ω.
We need to show that the map ψp is an isometry. We do this by showing that }ψpuq}
p L1 pP;Xq ě
}u}^ for a suitable set of us. The obvious vectors to choose are
ÿn
u“ 1Ei b xi ,
i“1
where the Ei s are disjoint measurable sets. By the Monotone Convergence Theorem the
span of the indicator functions 1Ei are dense in L1 pPq. On the other hand
› ›
›ÿn › ż ÿn ÿn
}ψpuq}L1 pP;Xq “ › 1Ei xi › }xi }X 1Ei dP “ }1Ei b xi }^ ě }u}^ .
› ›
p “
›i“1 › i“1 i“1
L1 pP;Xq

In §2.5 we recorded the definition of bilinear forms. We can now understand these as
the projective tensor product.
Lemma 10.2 ([DFS02d, Corollary 1.9]; originally [Gro53]). Suppose that X and Y are
Banach spaces. Then there is an isometric isomorphism between LpX, Y ˚ q˚ and X bY
p .
Proof. Consider the bilinear map ψ : X ˆ Y Ñ LpX, Y ˚ q˚ ; px, yq ÞÑ pT ÞÑ T pxqpyqq. The
argument is now an exercise using the universal properties of the tensor product.
On the face of it we might draw a line here: we have made a natural choice of short
bilinear map which we can add into Ban1 as multi-morphisms to make it into a multi-
category. In this multi-category we have a notion of tensor product, and this tensor product
extends to Ban where it coincides with the tensor product defined as a result of the hom-
sets being internal.
Despite how natural the above is there are other ways of endowing Ban1 with a tensor
product. Indeed, in the discussion before Theorem 9.3 we discussed natural tensor products
on Lp -spaces – sub-categories of Ban1 . Those tensor products were discussed in the context
of monoidal categories and it is worth clarifying that endowing a category with a product
in such a way as to make it monoidal is equivalent to endowing it with multimorphisms
as above to make it into a multicategory. This equivalence is discussed in some detail in
[Lei04, §3.3].
The Banch-Mazur theorem (or really Theorem 3.2) told us that a large class of Banach
spaces arise as closed subspaces of Cpr0, 1sq, and in fact in the presence of the Axiom of
Choice, for every Banach space X there is a compact Hausdorff space T such that X is a
closed subspace of CpT q. The sub-Category of Ban1 in which all the objects are of the
form CpT q can be made into a monoidal category in a rather natural way:
CpSqbCpT
r q :“ CpS ˆ T q.
A natural question now presents itself: is this tensor product the same as CpSqbCpT
p q, and
it turns out, in general, it is not. Roughly speaking b r is an example of a different tensor
product corresponding to a reduced collection of ‘allowable’ bilinear maps. Concretely it
is an example of something called the injective tensor product, a general construction we
shall now turn to.
The injective tensor product of Banach spaces X and Y is denoted X bY q and defined
to be the completion of X b Y under the injective norm
}u}_ :“ supt|x˚ b y ˚ puq| : }x˚ }X ˚ ď 1, }y ˚ }Y ˚ ď 1u
for all u P X b Y . Notice here that
ÿ ÿ
x˚ b y ˚ puq “ x˚ pxi qy ˚ pyi q whenever u “ xi b y i ,
i i
and this is well-defined by the universal property of the algebraic tensor product X b Y .
(The map px, yq ÞÑ x˚ pxqy ˚ pyq is bilinear and so gives rise to a linear map X b Y Ñ F
by the universal property.) We should also note that the injective tensor product is not a
norm unless X and Y support sufficient linear functionals – for us it will be sufficient to
regard them as closed subspaces of some spaces of continuous functions.
As with the projective tensor product the injective tensor product has a universal prop-
erty but for a smaller class of bilinear maps, specifically those ψ for which
› ›
(10.2) ›ψ̃puq› ď }u}_ for all u P X b Y
› ›
where ψ̃ is the algebraic extension of ψ to X b Y . Notice that since we are assuming ΦX

and ΦY to be isometric, any such ψ has
› ›
}ψpx, yq} “ ›ψ̃px b yq› ď supt|x˚ b y ˚ px b yq| : }x˚ }X ˚ ď 1, }y ˚ }Y ˚ ď 1u “ }x}X }y}Y ,
› ›
and so satisfies (10.1), but the converse does not hold. The universal property for the
injective tensor product says that given ψ : X ˆ Y Ñ Z such that (10.2) holds there is
some ψq : X bY
q Ñ Z such that ψ factors through ψ. q
We turn to an example: given a Banach space X and compact metrisable space T we
write
CpT ; Xq :“ tf : T Ñ X s.t. f is continuous.u
endowed with the norm
}f } :“ supt}f ptq}X : t P T u.
If X “ CpSq where S is compact and metrisable then the map
CpT ; CpSqq Ñ CpT ˆ Sq; f ÞÑ ppt, sq ÞÑ f ptqpsqq
is an isometric isomorphism.
Proposition 10.3 ([DFS02d, Theorem 1.10]; originally [Gro53, Theorem 3]). Suppose that
X is a Banach space. Then we have an isometric isomorphism from CpT qbX
q to CpT ; Xq.
72 TOM SANDERS
Proof. Consider the bilinear map

ψ : CpT q ˆ X Ñ CpT ; Xq; pf, xq ÞÑ pt ÞÑ f ptqxq.
ř
We check that this map satisfies (10.2): if u “ i fi xi then
#› › +
›ÿ ›
}ψpuq}CpT ;Xq “ sup › fi ptqxi › : t P T
› ›
› i ›
X
#ˇ ˜ ¸ˇ +
ˇ ÿ ˇ
“ sup ˇx˚ fi ptqxi ˇ : t P T, }x˚ } ď 1
ˇ ˇ
ˇ i
ˇ
#ˇ ˇ +
ˇÿ ˇ
“ sup ˇ φpfi qx˚ pxi qˇ : }φ}CpT q˚ ď 1, }x˚ } ď 1
ˇ ˇ
ˇ i ˇ
(
“ sup |φ b x˚ puq| : }φ}CpT q˚ ď 1, }x˚ } ď 1 “ }u}_ ,
and so by the universal property of the injective tensor product there is a short map
ψ̃ : CpT qbX
q Ñ CpT ; Xq.
It remains for us to show that the image is dense. Of course, this is straightforward:
suppose that f : T Ñ X is continuous then by compactness of T , for all ą 0 there is
a cover of open sets pUi qi such that f varies by at most on each Ui . Let f1 , . . . , fn be a
continuous ÿ
fi ptq “ 1 for all t P T ; for all i we have fi : T Ñ r0, 1s,
i
and
}f pxq ´ f pyq} ă whenever x, y P Ui for some i.
Let ti be some element of Ui for each i. Suppose that x P Ui then }f pti q ´ f pxq} ď , and
it follows that
› › › ›
› ÿ › › ÿ ›
›f pxq ´ fi pxqf pti q› ď ` ›f pxq ´ fi pxqf pxq› “ .
› › › ›
› i
› › i
›
ř
Thus there is some u “ i fi b f pti q P CpT q b X such that
› ›
ψ̃puq f ď ,
› ›
› ´ ›
CpT ;Xq
and hence the image of ψ̃ is dense in CpT ; Xq.

The projective and injective tensor products are dual in the sense that if u P X b Y and
t P X ˚ b Y ˚ then t is naturally bilinear on X ˆ Y and hence extends to t̃ on X b Y . We
have
|t̃puq| ď }u}^ }t}_ .
More generally, given a norm } ¨ } on X b Y we define a norm } ¨ }˚ on X ˚ b Y ˚ by
}t}˚ :“ supt|t̃puq| : }u} ď 1u.
This duality was first discussed by Schatten in [Sch50] and used by Grothendieck to great
effect.
We have already seen a couple of possible tensor product norms which, as it happens,
represent the most ‘extreme’ examples. We say a norm } ¨ } on X b Y is a reasonable
cross-norm if
}x b y} ď }x}X }y}Y and }x˚ b y ˚ }˚ ď }x˚ }X ˚ }y ˚ }Y ˚
for all x P X, y P Y, x˚ P X ˚ , y ˚ P Y ˚ . It may be worth noting that it is an easy consequence
of what follows (but see [Rya02, Proposition 6.1, (b)] if necessary) that for any reasonable
cross-norm (on closed subspaces of CpT q) we have equality above.
Proposition 10.4 ([Rya02, Proposition 6.1, (a)]). Suppose that X and Y are Banach
spaces and } ¨ } is a norm on X b Y . Then } ¨ } is a reasonable cross-norm on X b Y if
and only if }u}_ ď }u} ď }u}^ for all u P X b Y .
Proof. In one direct note that if u P X b Y then u “ ni“1 xi b yi and
ř
› ›
›ÿn › ÿ n ÿn
}u} “ › xi b yi › ď }xi b yi } ď }xi }X }yi }Y .
› ›
›i“1 › i“1 i“1
It follows that # +
n
ÿ n
ÿ
}u} ď inf }xi }X }yi }Y : u “ xi b yi .
i“1 i“1
On the other hand
}u}_ “ sup t|x˚ b y ˚ puq| : }x˚ }X ˚ ď 1 and }y ˚ }Y ˚ ď 1u ď }u}.
In the other direction if }u} ď }u}^ then }x b y} ď }x b y}^ ď }x}X }y}Y , and
|x˚ b y ˚ puq| ď }u}_ }x˚ }X ˚ }y ˚ }Y ˚ ď }u}}x˚ }X ˚ }y ˚ }Y ˚ .
It follows that
}x˚ b y ˚ }˚ ď }x˚ }X ˚ }y ˚ }Y ˚ .

One other tensor product norm we shall consider is the following
$ $˜ ¸1{2 ˜ ¸1{2 , ,
& & ÿ ÿ . ÿ .
˚ 2 ˚ 2 ˚ ˚
}u}H :“ inf sup |x pxi q| |y pyi q| : }x }, }y } ď 1 : u “ xi b y i ;
% % i i
- i
-
we leave it to the reader to check that this is a norm, and remark that it is trivial that
}u}H ď }u}^ .
Theorem 10.5 (Grothendieck’s Little Inequality). Suppose that S and T are compact
metrisable spaces. Then if ψ : CpSq ˆ CpT q Ñ F is of the form pf, gq ÞÑ xU f, V gy for short
maps U : CpSq Ñ H and V : CpT q Ñ H, then
|ψ̃puq| ď kG }u}H for all u P CpSq b CpT q
where kG “ π{2 if F “ R, and kG “ 4{π if F “ C.
74 TOM SANDERS
Proof. Notice that in this setting we have

$› ›1{2 › ›1{2 ,
&›ÿ › ›ÿ › ÿ .
2› 2›
}u}H “ inf › |fi | › :u“ f i gi .
› ›
› |gi | ›
%› i › › i › i
-
CpSq CpT q
Since ÿ ÿ
ψ̃puq “ xU fi , V gi y where u “ xi b y i ,
i i
we have by the Cauchy-Schwarz inequality that the result is equivalent to the special case
U “ V i.e. it suffices to show that
˜ ¸1{2 ›˜ ¸1{2 ››
ÿ ›
a › ÿ ›
}U fi }2 ď kG ›› |fi |2 ›
›
i › i ›
CpSq
for all functions f1 , . . . , fn P CpSq.

It is possible to show that the result follows by an approximation argument from the
case when S is finite. We shall not do this here, but from now on we take S to have size
N so that CpSq “ `N 8.
Writing hi “ U fi we have
ÿ ÿ ÿ
(10.3) }U fi }2 “ xU fi , hi y “ U ˚ hi pfi q.
i i i
˚
The map U takes H (since it is self-dual) to M pSq. Writing µ for counting measure on
S, since S is finite all the measures in M pSq are absolutely continuous with respect to
counting measure meaning that for any ν P M pSq there is a unique function f P L1 pµq
such that ż ż
gdν “ gf dµ for all g P L1 pνq.
dν
We write dµ for f – it is called the Radon-Nikodym derivative. With this in mind we
define a map
dU ˚ h
W : H Ñ L1 pµq; h ÞÑ ,
dµ
which has }W } “ }U ˚ } “ }U } “ 1. Given the definition of W and (10.3) we have (by
Fubini’s theorem and the Cauchy-Schwarz inequality) that
ÿ ż ÿ
2
}U fi } “ fi psqW hi psqdµpsq
i i
ż ˜ÿ ¸1{2 ˜ ¸1{2
ÿ
2 2
ď |fi psq| |W hi psq| dµpsq
i i
›˜ ¸1{2 ›› ›˜ ¸1{2 ››
› ›
› ÿ 2 › › ÿ 2
›
ď ›› |fi | ›
›
›
› |W hi | ›
› .
› i › › i ›
CpSq L1 pµq
Now, suppose that X1 , . . . , Xn are independent N p0, 1q distributed random variables, and
write X for any instance of Xi . Then
›˜ ¸1{2 ›› › ›
› ›ÿ ›
› ÿ ›
|W hi |2 W h X
› ›
}X}L1 pPq ›› ›
› “ › i i›
› i › i
›
L1 pPˆµq
›
L1 pµq
›
ż › ˜ ¸›
ÿ ›
›W hi Xi pωq › dPpωq
› ›
“
› i
›
L pµq
ż ››ÿ › 1
›
ď }W } › hi Xi pωq› dPpωq
› ›
› i ›
¨ › ›2 ˛1{2
ż ›ÿ ›
ď }W } ˝ › hi Xi pωq› dPpωq‚
› ›
› i ›
˜ ¸1{2
ÿ
“ }W } }hi }2 .
i
As noted earlier we have }W } “ 1 and combining everything we have shown so far we have
›˜ ¸1{2 ›› ˜ ¸1{2
›
ÿ › ÿ › ÿ
}U fi }2 ď ›› |fi |2 ›
› }X}´1
L1 pPq }hi }2 .
i › i › i
CpSq
The definition of the hi s gives the result on computing }X}L1 pPq (which is different depend-
ing on whether F “ R or F “ C).
It is not too difficult to show that the constants above are best possible and the interested
reader may with to consult [Pis12, Theorem 5.1].
In the proof above we focused on the finite dimensional case essentially looking at the
norm } ¨ }H on `N N N N
8 b `8 . Algebraically the space `8 b `8 can be identified with MN pFq –
the space of N ˆ N matrices over F. In this language we have that
! )
˚
}M }H “ inf }U }`N8 Ñ`N2 }V }`N8 Ñ`N2 : M “ U V .
Writing } ¨ }H 1 for the dual norm of } ¨ }H i.e.
}M }H 1 :“ sup ttr M B ˚ : }B}H ď 1u ,
we have
# +
ÿ
}M }H 1 “ sup Mij xfi , gj y : fi , gj P L have }fi }, }gj } ď 1; L is Hilbert. ,
i,j
and what Grothendieck’s Little Inequality says in this language is that

}M }H 1 ď kG }M }H .
76 TOM SANDERS
It is elementary that }M }H ď }M }H 1 .
It turns out that there is a stronger inequality here, relating not just the H-norm, but in
fact the injective tensor norm – in this language the operator norm of M . We concentrate
on the case of F “ R. Suppose that M is an n ˆ n matrix such that
#ˇ ˇ +
ˇÿ ˇ
(10.4) }M }_ “ }M }`n8 Ñ`n1 ď 1 i.e. sup ˇ Mij fi gi ˇ : |fi |, |gj | ď 1 ď 1,
ˇ ˇ
ˇ ij ˇ
then we shall look at }M }H 1 . Writing L2 pXq for L2 pµX q where µX is counting measure on
X and X is a finite set of size d, it will be enough to consider expressions of the form
ÿ ż ÿ
(10.5) Mij xfi , gj yL2 pXq “ Mij fi pxqgj pxqdµX pxq.
ij ij
The Cauchy-Schwarz inequality tells us that

?
(10.6) }f }L8 pXq ď d}f }L2 pXq for all f P L2 pXq,
so we get that
ÿ
| Mij xfi , gj yL2 pXq | ď d sup }fi }L2 pXq }gj }L2 pXq
ij ij
from the hypothesis (10.4). It follows that we can write Kd for the smallest constant such
that for all n ˆ n matrices M satisfying (10.4) and all d-dimensional real Hilbert spaces H
we have
ÿ
| Mij xfi , gj yL2 pXq | ď Kd sup }fi }L2 pXq }gj }L2 pXq .
ij ij
Note that if we restrict to the (Hilbert) subspace generated by pfi qi , pgj qj , then none of the
quantities of concern change and so we have Kd ď K2n .
In this notation our previous argument showed that Kd ď d, whence Kd ď mintd, 2nu,
and Grothendieck’s inequality tells us that Kd is bounded by an absolute constant. We
give a proof following Blei [Ble87].
Theorem 10.6 (Grothendieck’s Inequality). We have that Kd “ Op1q.
Proof. We pick a unit vector v P H uniformly at random so for f, g P H with }f } “ }g} “ 1

we have
xf, gy “ dEv xf, vyxv, gy
¨ ˛ ˜ ¸
ÿ ÿ
E˝ |xf, ek y|2 ‚ ď dC ´2 E |xf, ek y|4
?
k:|xf,ek y|ěC{ d k
and so
ˇ ˇ
ˇÿ ˇ
Kd “ ˇ Mij xfi , gj y|ˇ
ˇ ˇ
ˇ ij ˇ
ˇ ˇ
ˇÿ d ˇ
ˇ ÿ
Mij xfi , ek yxek , gj y|ˇ
ˇ
“ ˇ
ˇk“1 ij ˇ
ˇ ˇ
ÿ d ˇÿ ˇ
ˇ Mij xfi , ek yxek , gj y|ˇ
ˇ ˇ
ď
ˇ
k“1 ij
ˇ
d
ÿ
ď sup |xfi , ek y sup |xek , gj y|
k“1 i j
We continue to assume, as we may, that H “ L2 pXq. In general we cannot do better

than (10.6). However, if the large values of the vectors fi and gj have small L2 -mass then
we can. Let fi and gi be such that
ÿ
Kd “ | Mij xfi , gj yL2 pXq | and }fi }L2 pXq , }gj }L2 pXq ď 1.
ij
Decompose the fi s and gj s into their large and small parts: fi “ fiL ` fiS and gj “ gjL ` gjS
where
# #
L fi pxq if |fi pxq| ě K L gj pxq if |gj pxq| ě K
fi pxq :“ and gj pxq :“
0 otherwise. 0 otherwise.
Then
ÿ ÿ
| Mij xfi , gj yL2 pXq | ď | Mij xfiS , gjS yL2 pXq |
ij ij
ÿ ÿ
`| Mij xfiL , gj yL2 pXq | ` | Mij xfiS , gjL yL2 pXq |
ij ij
2
ď K ` Kd max }fiL }L2 pXq ` Kd max }gjL }L2 pXq .
i j
Since the left hand side is just Kd , we are done if we can show that the two maxima on the
right are small for some K “ Op1q. Of course this is not true, but we can use Khintchine’s
inequality to give us an isometric embedding to a space where it is.
Specifically, let Ω “ t0, 1ud endowed with uniform probability measure P, and pZx qxPX
be a set of d independent ˘1-valued random variables on Ω each having mean 0, and put
1 ÿ 1 ÿ
fri :“ ? fi pxqZx and grj :“ ? gj pxqZx .
d xPX d xPX
It is easy to check that
xfri , grj yL2 pPq “ xfi , gj yL2 pXq .
78 TOM SANDERS
L S
Now, writing fri “ fri ` fri and grj “ grj L ` grj S in the same way as before, we see that
ÿ L
(10.7) | Mij xfri , grj yL2 pPq | ď K 2 ` K2d max }fri }L2 pPq ` K2d max }grj L }L2 pPq .
i j
ij
On the other hand, by Khintchine’s inequality for p “ 4 we have that

ż
2 rL 2 L
K }fi }L2 pPq “ |fri |2 K 2 dP ď }fri }4L4 pPq “ Op}fi }4L2 pXq q “ Op1q,
and similarly for grj L . It follows that there is a choice of K “ Op1q such that the maxima
in (10.7) are each at most 1{4, and hence
ÿ 1
Kd “ | Mij xfri , grj yL2 pPq | ď Op1q ` K2d .
ij
2
Finally K2d ď 2n for all d whence

1 1 1 1
Kd ď Op1q ` Op1q ` ¨ ¨ ¨ ` l Op1q ` l`1 n ď Op1q ` l n
2 2 2 2
for all l. Letting l tend to infinity completes the proof.
References
[Alf63] E. M. Alfsen. A simplified constructive proof of the existence and uniqueness of Haar measure.
Math. Scand., 12:106–116, 1963.
[Alo03] N. Alon. Problems and results in extremal combinatorics. I. Discrete Math., 273(1-3):31–53,
2003. EuroComb’01 (Barcelona).
[Ami65] D. Amir. On isomorphisms of continuous function spaces. Israel J. Math., 3:205–210, 1965.
[Arv96] W. Arveson. Notes on measure and integration in locally compact spaces. 1996, https://math.
berkeley.edu/~arveson/Dvi/rieszMarkov.pdf.
[Bar05] A. Barvinok. Measure concentration. 2005, http://www.math.lsa.umich.edu/~barvinok/
courses.html.
[Bat14] C. Batty. Functional Analysis (C4.1). 2014.
[Bec75] W. Beckner. Inequalities in Fourier analysis. Ann. of Math. (2), 102(1):159–182, 1975.
[Bel14] D. Belyaev. Banach Spaces (B4.1). 2014.
[Ben98] Y. Benyamini. Applications of the universal surjectivity of the Cantor set. Amer. Math. Monthly,
105(9):832–839, 1998.
[Bil95] P. Billingsley. Probability and measure. Wiley Series in Probability and Mathematical Statistics.
John Wiley & Sons, Inc., New York, third edition, 1995. A Wiley-Interscience Publication.
[BK00] V. V. Buldygin and Yu. V. Kozachenko. Metric characterization of random variables and ran-
dom processes, volume 188 of Translations of Mathematical Monographs. American Mathemat-
ical Society, Providence, RI, 2000. Translated from the 1998 Russian original by V. Zaiats.
[Ble87] R. C. Blei. An elementary proof of the Grothendieck inequality. Proc. Amer. Math. Soc.,
100(1):58–60, 1987.
[Bol99] B. Bollobás. Linear analysis. Cambridge University Press, Cambridge, second edition, 1999. An
introductory course.
[Bon70] A. Bonami. Étude des coefficients de Fourier des fonctions de Lp pGq. Ann. Inst. Fourier (Greno-
ble), 20(fasc. 2):335–402 (1971), 1970.
[Bou52] N. Bourbaki. Eléments de mathématique. XIII. Première partie: Les structures fondamentales
de l’analyse. Livre VI: Intégration. Chapitre I: Inégalités de convexité. Chapitre II: Espaces de
Riesz. Chapitre III: Mesures sur les espaces localement compacts. Chapitre IV: Prolongement
d’une mesure; espaces Lp . Actualités Sci. Ind., no. 1175. Hermann et Cie, Paris, 1952.
[BS86] D. K. Brown and S. G. Simpson. Which set existence axioms are needed to prove the separable
Hahn-Banach theorem? Ann. Pure Appl. Logic, 31(2-3):123–144, 1986. Special issue: second
Southeast Asian logic conference (Bangkok, 1984).
[Cam66] M. Cambern. A generalized Banach-Stone theorem. Proc. Amer. Math. Soc., 17:396–400, 1966.
[Cam70] M. Cambern. Isomorphisms of C0 pY q onto CpXq. Pacific J. Math., 35:307–312, 1970.
[DFS02a] J. Diestel, J. Fourie, and J. Swart. The metric theory of tensor products (Grothendieck’s résumé
revisited). II. Bilinear forms and linear operators of type α. Quaest. Math., 25(1):73–94, 2002.
[DFS02b] J. Diestel, J. Fourie, and J. Swart. The metric theory of tensor products (Grothendieck’s résumé
revisited). III. Vector sequence spaces. Quaest. Math., 25(1):95–118, 2002.
[DFS02c] J. Diestel, J. Fourie, and J. Swart. The metric theory of tensor products (Grothendieck’s résumé
revisited). IV. the role of CpKq-spaces and L1 -spaces. Quaest. Math., 25(3):341–361, 2002.
[DFS02d] Joe Diestel, Jan Fourie, and Johan Swart. The metric theory of tensor products (Grothendieck’s
résumé revisited). I. Tensor norms. Quaest. Math., 25(1):37–72, 2002.
[DLAT10] P. Dodos, J. Lopez-Abad, and S. Todorcevic. Banach spaces and Ramsey theory: some open
problems. Rev. R. Acad. Cienc. Exactas Fı́s. Nat. Ser. A Math. RACSAM, 104(2):435–450,
2010.
[FHJ` 97] T. Figiel, P. Hitczenko, W. B. Johnson, G. Schechtman, and J. Zinn. Extremal properties of
Rademacher functions with applications to the Khintchine and Rosenthal inequalities. Trans.
Amer. Math. Soc., 349(3):997–1027, 1997.
[Fre06] D. H. Fremlin. Measure theory. Vol. 4. Torres Fremlin, Colchester, 2006. Topological measure
spaces. Part I, II, Corrected second printing of the 2003 original.
[Gar03] B. Garling. Classical Banach spaces. Lecture notes, Cambridge Part III, 2003.
[Gia95] A. A. Giannopoulos. A note on the Banach-Mazur distance to the cube. In Geometric aspects
of functional analysis (Israel, 1992–1994), volume 77 of Oper. Theory Adv. Appl., pages 67–73.
Birkhäuser, Basel, 1995.
[Gow96] W. T. Gowers. A solution to the Schroeder-Bernstein problem for Banach
spaces. Bulletin of the London Mathematical Society, 28(3):297–304, 1996,
http://blms.oxfordjournals.org/content/28/3/297.full.pdf+html.
[GR00] I. S. Gradshteyn and I. M. Ryzhik. Table of integrals, series, and products. Academic Press,
Inc., San Diego, CA, sixth edition, 2000. Translated from the Russian, Translation edited and
with a preface by Alan Jeffrey and Daniel Zwillinger.
[Gro53] A. Grothendieck. Résumé de la théorie métrique des produits tensoriels topologiques. Bol. Soc.
Mat. São Paulo, 8:1–79, 1953.
[Haa81] U. Haagerup. The best constants in the Khintchine inequality. Studia Math., 70(3):231–283
(1982), 1981.
[Hal35] P. Hall. On representatives of subsets. Journal of the London Mathematical Society, s1-10(1):26–
30, 1935, http://jlms.oxfordjournals.org/content/s1-10/1/26.full.pdf.
[Jac84] K. Jacobs. Ergodic theory and combinatorics. In R. Beals, A. Beck, A. Bellow, and A. Hajian,
editors, Conference in modern analysis and probability, volume 26 of Contemporary Mathe-
matics, pages 171–187. American Mathematical Society, Providence, RI, 1984. Held at Yale
University, New Haven, Conn., June 8–11, 1982, Held in honor of Professor Shizuo Kakutani.
[Jar88] K. Jarosz. Any Banach space has an equivalent norm with trivial isometries. Israel J. Math.,
64(1):49–56, 1988.
[JK02] G. Janelidze and G. M. Kelly. A note on actions of a monoidal category. Theory Appl. Categ.,
9:61–91, 2001/02. CT2000 Conference (Como).
80 TOM SANDERS
[JL84] W. B. Johnson and J. Lindenstrauss. Extensions of Lipschitz mappings into a Hilbert space.
In Conference in modern analysis and probability (New Haven, Conn., 1982), volume 26 of
Contemp. Math., pages 189–206. Amer. Math. Soc., Providence, RI, 1984.
[JSZ85] W. B. Johnson, G. Schechtman, and J. Zinn. Best constants in moment inequalities for linear
combinations of independent and exchangeable random variables. Ann. Probab., 13(1):234–253,
1985.
[Kah60] J.-P. Kahane. Propriétés locales des fonctions à séries de Fourier aléatoires. Studia Math., 19:1–
25, 1960. ř
[Kah64] J.-P. Kahane. Sur les sommes vectorielles ˘un . C. R. Acad. Sci. Paris, 259:2577–2580, 1964.
[Kak41] S. Kakutani. Concrete representation of abstract pM q-spaces. (A characterization of the space
of continuous functions.). Ann. of Math. (2), 42:994–1024, 1941.
[Kec95] A. S. Kechris. Classical descriptive set theory, volume 156 of Graduate Texts in Mathematics.
Springer-Verlag, New York, 1995.
[Kel64] G. M. Kelly. On MacLane’s conditions for coherence of natural associativities, commutativities,
etc. J. Algebra, 1:397–402, 1964.
[KM11] M. Kojman and H. Michalewski. Borel extensions of Baire measures in ZFC. Fund. Math.,
211(3):197–223, 2011.
[Kön16] D. König. Über Graphen und ihre Anwendung auf Determinantentheorie und Mengenlehre.
Math. Ann., 77(4):453–465, 1916.
[KS91] S. Kwapień and J. Szulga. Hypercontraction methods in moment inequalities for series of inde-
pendent random variables in normed spaces. Ann. Probab., 19(1):369–379, 1991.
[KT01] K. Keremedis and E. Tachtsis. Compact metric spaces and weak forms of the axiom of choice.
MLQ Math. Log. Q., 47(1):117–128, 2001.
[Lei04] T. Leinster. Higher operads, higher categories, volume 298 of London Mathematical Society
Lecture Note Series. Cambridge University Press, Cambridge, 2004.
[Lew86] J. W. Lewin. The Teaching of Mathematics: A Truly Elementary Approach to the Bounded
Convergence Theorem. Amer. Math. Monthly, 93(5):395–397, 1986.
[LO94] R. Latala and K. Oleszkiewicz. On the best constant in the Khinchin-Kahane inequality. Studia
Math., 109(1):101–104, 1994.
[LW12] R. B. B. Lucyshyn-Wright. A general Fubini theorem for the Riesz paradigm. 2012,
arXiv:1210.4542.
[Maa35] W. Maak. Eine neue Definition der fastperiodischen Funktionen. Abh. Math. Sem. Univ. Ham-
burg, 11(1):240–244, 1935.
[Man12] O. Manzyuk. Closed categories vs. closed multicategories. Theory Appl. Categ., 26:No. 5, 132–
175, 2012.
[Mař57] J. Mařı́k. The Baire and Borel measure. Czechoslovak Math. J., 7(82):248–253, 1957.
[Mil66] A. A. Miljutin. Isomorphism of the spaces of continuous functions over compact sets of the
cardinality of the continuum. Teor. Funkciı̆ Funkcional. Anal. i Priložen. Vyp., 2:150–156. (1
foldout), 1966.
[ML98] S. Mac Lane. Categories for the working mathematician, volume 5 of Graduate Texts in Math-
ematics. Springer-Verlag, New York, second edition, 1998.
[Nao10] A. Naor. Local theory of Banach spaces. 2010, https://web.math.princeton.edu/~naor/
homepage%20files/localban.pdf.
[Nel73] E. Nelson. The free Markoff field. J. Functional Analysis, 12:211–227, 1973.
[Neu99] J. Neukirch. Algebraic number theory, volume 322 of Grundlehren der Mathematischen Wis-
senschaften [Fundamental Principles of Mathematical Sciences]. Springer-Verlag, Berlin, 1999.
Translated from the 1992 German original and with a note by Norbert Schappacher, With a
foreword by G. Harder.
[PB79] A. Pelczyński and C. Bessaga. Some aspects of the present theory of Banach spaces. In et al.
C. Bessaga, editor, Œuvres. Vol. II, pages 221–302. PWN—Éditions Scientifiques de Pologne,
Warsaw, 1979.
[Pel60] A. Pelczyński. Projections in certain Banach spaces. Studia Math., 19:209–228, 1960.
[Pel84] A. Pelczyński. Structural theory of Banach spaces and its interplay with analysis and probabil-
ity. In Proceedings of the International Congress of Mathematicians, Vol. 1, 2 (Warsaw, 1983),
pages 237–269. PWN, Warsaw, 1984.
[Phi40] R. S. Phillips. On linear transformations. Trans. Amer. Math. Soc., 48:516–541, 1940.
[Pis12] G. Pisier. Grothendieck’s theorem, past and present. Bull. Amer. Math. Soc. (N.S.), 49(2):237–
323, 2012.
[Rad22] H. Rademacher. Einige Sätze über Reihen von allgemeinen Orthogonalfunktionen. Math. Ann.,
87(1-2):112–138, 1922.
[Rie10] F. Riesz. Sur les opérations fonctionnelles linéaires. C. R. Acad. Sci., Paris, 149:974–977, 1910.
[Rud87] W. Rudin. Real and complex analysis. McGraw-Hill Book Co., New York, third edition, 1987.
[Rya02] R. A. Ryan. Introduction to tensor products of Banach spaces. Springer Monographs in Math-
ematics. Springer-Verlag London, Ltd., London, 2002.
[Sch50] R. Schatten. A Theory of Cross-Spaces. Annals of Mathematics Studies, no. 26. Princeton
University Press, Princeton, N. J., 1950.
[Sco15] A. Scott. Combinatorics (C8.3). 2015.
[Sob41] A. Sobczyk. Projection of the space pmq on its subspace pc0 q. Bull. Amer. Math. Soc., 47:938–
947, 1941.
n
[Sza90] S. J. Szarek. Spaces with large distance to l8 and random matrices. Amer. J. Math., 112(6):899–
942, 1990.
[Tai79] M. H. Taibleson. Book Review: Non-Archimedean functional analysis. Bull. Amer. Math. Soc.
(N.S.), 1(5):802–804, 1979.
[Tal89] M. Talagrand. Isoperimetry and integrability of the sum of independent Banach-space valued
random variables. Ann. Probab., 17(4):1546–1570, 1989.
[Ute85] S. A. Utev. Extremal problems in moment inequalities. In Limit theorems of probability theory,
volume 5 of Trudy Inst. Mat., pages 56–75, 175. “Nauka” Sibirsk. Otdel., Novosibirsk, 1985.
[Vät98] M. Väth. The dual space of L8 is L1 . Indag. Math. (N.S.), 9(4):619–625, 1998.
[Ver10] M. Veraar. Probability in Banach spaces, 2010. Available at http://fa.its.tudelft.nl/
seminar/seminar2009_2010/seminar2009_2010.htm.
[Ver12] R. Vershynin. Introduction to the non-asymptotic analysis of random matrices. In Compressed
sensing, pages 210–268. Cambridge Univ. Press, Cambridge, 2012.
[vR78] A. C. M. van Rooij. Non-Archimedean functional analysis, volume 51 of Monographs and Text-
books in Pure and Applied Math. Marcel Dekker, Inc., New York, 1978.
[Whi66] R. Whitley. Mathematical Notes: Projecting m onto c0 . Amer. Math. Monthly, 73(3):285–286,
1966.
[Wil91] D. Williams. Probability with Martingales. Cambridge Mathematical Textbooks. Cambridge
University Press, Cambridge, 1991.
[Woj91] P. Wojtaszczyk. Banach spaces for analysts, volume 25 of Cambridge Studies in Advanced
Mathematics. Cambridge University Press, Cambridge, 1991.
Mathematical Institute, University of Oxford, Radcliffe Observatory Quarter, Wood-

stock Road, Oxford OX2 6GG, United Kingdom
E-mail address: tom.sanders@maths.ox.ac.uk

Notes PDF

Uploaded by

Copyright:

Available Formats

Notes PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Notes PDF

Uploaded by

Copyright:

Available Formats

FINITE DIMENSIONAL NORMED SPACES

It is easy to check that `p is a vector space and } ¨ } defines a norm on `p so as to make it

}v ´ w}8 ě 1 for all v, w P E with v ‰ w.

Example 1.4 (Convergent sequence spaces: c0 and cc ). In `8 (and, indeed, `p for 1 ď p ă

On the other hand

In the other direction we trivially have that

3. Spaces of continuous functions

We define some sequences of closed neighbourhoods6 and non-negative integers itera-

φ̃pf q “ φpf q for all f P CpT q;

with the bounded pointwise topology.

is a continuous linear functional on CpT q.

Theorem 3.7 (Riesz-Kakutani representation theorem). Suppose that T is a compact

Proof. It is easiest to work with the duals so that we get maps

Now, if |Φ˚ pδt qptsuq| ď }Φ˚ }{2 for all t P T then

4. Isomorphisms and the structure of `p spaces

Proof. We start by constructing sequences 0 “: n0 ă n1 ă n2 ă . . . and 0 “: j0 ă j1 ă . . .

In particular, e˚i pT en q “ wn . Since p ą 1 and so p1 ă 8, we have that wn Ñ 8 as n Ñ 8,

and so by the triangle inequality we have

However, again by Lemma 4.2 (since peni q8

Combining these (and using nesting of norms between `8 and `p ) we get

One sometimes describes a normable, sequentially complete topological space as Ba-

is a well-defined isometric isomorphism.

Example 4.14. Given a sequence of vectors pxi qi we define their span to be

and it is easy to check that it is a short projection with image Spanppxi qi q.

which is well-defined and linear.

a projection on X, and so W is complemented in X. It follows that there is a subspace

This is evidently an n-dimensional Banach space, and

have that X is isometrically isomorphic to X ˚˚ (via ΦX ), and dBM pX, Y q “ dBM pX ˚ , Y ˚ q.

Proof. Let S be a maximal -separated subset of K, the unit ball of X ˚ , so that if φ, ψ P S

|S|µpKqp{2qn “ µpS ` tx P X : }x} ď {2uq

we have a φ with the desired properties. In particular, if x P X then there is some φ P X ˚

dBM p`np , `nq q Ñ 8 as n Ñ 8.

is then an isometric isomorphism, and we leave the verification of this as an exercise.

}x}2 ˘ 2xx, δyy ` δ 2 }y}2 “ }x ˘ δy}2 ě φpx ˘ δyq2 “ 1,

Ki :“ tz P `n2 : }z}`n2 “ 1 and φi pzq ě 1 ´ u for 1 ď i ď N.

By the first inequality in the hypothesis we have

Since B1 is closed it is measurable, and since it contains a neighbourhood of the origin we

Instead of picking a sequence x1 , . . . , xn in r0, 1s (or T ) uniformly at random, we shall pick

and hence U is δ-covering. Since S is of minimum size we conclude the aforementioned

By the triangle inequality and choice of  and δ earlier we have

By the triangle again we therefore have

Proof. We proceed as in the proof of the sequential Banach-Alaoglu theorem (Theorem

Note that for all n P N we have

This action is isometric since

dIsom pgpφq, gpψqq “ suptdpφpg ´1 ptqq, ψpg ´1 ptqqq : t P T u

Suppose that µ is a G-Haar probability measure on T .

Claim. SUp2q acts transitively on CS1 .

Putting ∆ :“ texpp2πiqqI : q P Qu, where I is the identity matrix, we see that

Hn “ ∆ SUpnq and Hn “ ∆ SUpnq.

g 1 P G such that dIsom pg, g 1 q ă δ. Hence

Since  ą 0 was arbitrary the claim follows.

The function G ˆ T Ñ F; pg, tq ÞÑ f pgptqq is continuous since f is continuous and

The right hand side is independent of µ, and hence µ “ ν

Example 7.6 (Surface measure on the sphere). The pn ´ 1q-sphere is defined to be

which has φpe1 q “ f1 .

Pp|X| ě tq ď expp´t2 {2q for all t ě 0;

It follows that }X}SubpΩq “ σ.

Now, suppose that  ą 0. Then

by continuity of probability measures. It follows that X “ 0 almost everywhere as claimed.

Proof. Let S be a maximal -separated subset of K, the unit ball of X ˚ , so that if φ, ψ P S

|S|µpKqp{2qn “ µpS ` tx P X : }x} ď {2uq

Ki :“ tz P `n2 : }z}`n2 “ 1 and φi pzq ě 1 ´ u for 1 ď i ď N.

By the triangle inequality and choice of and δ earlier we have

Since ą 0 was arbitrary the claim follows.

Now, suppose that ą 0. Then