When does embed into ?
The question goes back to Banach. This will be an excuse to review some modern tools. Banach space theory has evolved into metric geometry. The leitmotiv will be: how does one prove non-embeddability?
1. Linear embeddings
1.1. Banach’s question
Notation. measurable functions on . -summable sequences.
in its norm.
Definition 1 A linear operator between normed spaces is a -isomorphic embedding, where , if there exists such that ,
Example 1 embeds in isometrically.
Remark. is not isomorphic to when . This is a nontrivial fact, probably going back to Banach.
Question (Banach). For which does embed isomorphically into ?
Theorem 2 (Banach 1932, Paley 1936) does not embed into unless or .
Theorem 3 (Kadec 1958) embeds isometrically into if or .
Theorem 2 is much harder than Theorem 3 (although older).
1.2. Proof of Kadec’ theorem
Let be a sequence of iid standard gaussian random variables on a -finite probability space.
Gaussians have the following property: if standard gaussian r.v. and are independent, the random variable has the same distribution as . Indeed, both have the same characteristic function.
Define the embedding as follows,
Then has the same distribution as . Therefore
Question. Why doesn’t this work for other values of ? In other words, do there exist iid symmetric random variables such that has the same distribution as ?
The answer arises in Paul Lévy’s book.
Theorem 4 (Levy 1951) Such random variables exist iff . They are called standard -stable random variables, and for , they satisfy
Therefore,
is finite iff .
This shows that embeds isometrically into for .
Remark 1 To prove that embeds into in this range, one needs two more nontrivial results:
- is finitely representable into .
- If is finitely representable in and is separable, then embeds into .
This ends our excursion in the construction of embeddings. From now on, we switch to nonembeddability results.
1.3. Linear distorsion
Definition 5 The linear distorsion of into is the smallest such that there exists a -isomorphic embedding of into . It is denoted by
and if , by
Thus we have stated and partly proved that
Goal. Understand the asymptotics of when are in the second range.
Theorem 6 Let . Then
Easy. The upper bounds are easy, the embedding is identity or identity to followed with Kadec’s isometric embedding.
Our goal is to prove lower bounds, by designing invariants.
2. Smoothness and convexity in spaces
2.1. Smoothness and convexity
Definition 7 (Ball – Carlen – Lieb 1993) Fix . Say a Banach space is -uniformly smooth with constant if
Fix . Say a Banach space is -uniformly convex with constant if
The best constants are denoted by and .
Exercise. (Lindenstrauss’ duality formula à la Ball-Carlen-Lieb). For any normed space and ,
2.2. The case of spaces
Theorem 8 (Clarkson’s inequality) For , .
For , .
Proof. Since the inequality involves only , is suffices to prove that for all ,
By monotonicity of norms and the parallelogram identity,
Theorem 9 For , .
For , .
2.3. Two lemmata
The proof of Theorem 10 requires two lemmata, the first one is Bonami’s hypercontractivity inequality on the -cube.
Lemma 10 (Bonami’s two-point inequality) Let . For ,
Proof. One can assume that and . Using Taylor’s expansion, one checks that
The second lemma illustrates a general principle: an inequality for with constant one can be nothing but a relaxation of the parallelogram identity in Euclidean space.
Lemma 11 (Hanner’s inequality) Let . For ,
One first checks that for , the numbers
satisfy
(again, one can assume that and ; then it amounts to the monotonicity of a function).
Hanner’s inequality follows: assuming that , set , , to get
Integrating with respect to yields
2.4. Proof of Theorem 10
Let , . We show that
Indeed, first apply Bonami’s inequality.
Conjecture. Does Hanner’s inequality hold in Schatten class, i.e. for matrices where the norm of singular values is used?
Known for and (Ball-Carlen-Lieb). Heinavaara 2022 proves this for .
Conjecture. For ,
Theorem 12 (D. Schechtman 1995) Yes for .
3. Martingales in Banach spaces
Definition 13 Let be a Banach space. A (Paley-Walsh) martingale with values in is a sequence of functions such that for all ,
The basic example is
for given vectors .
Remark. If , the set is orthogonal, so
Definition 14 (Pisier 1975) A Banach space has martingale type , , il for every -valued martingale ,
has martingale cotype , , il for every -valued martingale ,
These properties can be thought of as relaxations of the identity that holds in Hilbert space.
3.1. Smoothness/convexity versus type/cotype
These properties follow from the convexity properties introduced earlier.
Proposition 15 (Pisier) -smoothness (resp. -convexity) implies martingale type (resp. cotype ) with the same constant.
Pisier’s renorming theorem states that the converse is true, up to changing for an equivalent norm.
The advantage of the type/cotype formulation is that the error is multiplicative, hence these properties are isomorphism invariant.
3.2. Proof of Pisier’s “smoothness implies type”
Let be an -valued martingale with . Then
thus has martingale type with constant .
Corollary 16 Let .
If , thenIf , then
Beware that has Rademacher cotype with constant but no nontrivial martingale cotype.
3.3. Proof of Theorem 7, type and cotype range
Let be an embedding of distorsion . Apply type of to . Then
On the other hand,
This yields .
The proofs of the three other lower bounds on are similar. The results are sharp.
Note that the argument used linearity very strongly.
4. Nonlinear embeddings
Given metric spaces and , one can speak of the biLipschitz distorsion of into , denoted by , as the least such that there exists and satisfying
Now we discretize spaces. Let , viewed as an approximation of . What can one say about ?
Theorem 17 Let . Then is of the order of
So we see that a phase transition occurs when .
In certain cases, the upper bounds are smart, but I will not focus on them.
In the unknown range , the best we know now is
The left-hand side would be sharp if the following was true: for every and every , the metric space has a biLipschitz embedding into .
Theorem 18 (Bretagnolle – Dacunha-Castelle – Krivine 1965) This is true for .
In my thesis, we have the following result:
Theorem 19 (Eskenazis – Naor 2016) For , , , the space
does not embed in .
5. Metric type
5.1. Enflo type
Definition 20 (Enflo 1969, modern terminology) A metric space has Enflo type with constant if for all , for all ,
Remark. If is a normed space and , then the lefthand side is
and the righthand side is
so this is really a metric analogue of Rademacher type: for linear spaces,
Theorem 21 (Khot – Naor 2006) For normed linear spaces,
Proof. Let which has martingale type . Define a martingale
From martingale type, we know that . But
where has one sign changed. So
where has length and one sign flipped at position . So summing over yields the expected righthand side. Now add and subtract the expectation,
which completes the proof.
The converse is a recent breakthrough, still in the linear case:
Theorem 22 (Ivanisvili – von Handel – Volberg 2020) For linear spaces, Rademacher type implies Enflo type.
5.2. Proof of distorsion lower bounds for
The upper bound is easy (use global embedding of ).
Here is another easy lower bound:
Assume that (the other case is similar). Let have distorsion with rescaling . Since has Enflo type with constant ,
(one sign change in ), so the righthand side is bounded above by , and the lefthand side is bounded below by , so .
6. Metric cotype
Formally, Rademacher cotype is the reverse of Rademacher type, but a naive delinearization does not work. The cube does not suffice, one needs an extra scaling factor .
Definition 23 (Mendel – Naor 2008) A metric space has metric cotype with constant if for any , there exists such that any function satisfies
Remark. As soon as has at least 2 points, .
Theorem 24 (Mendel – Naor 2008, Giladi – Mendel – Naor 2011) For normed spaces, Rademacher cotype is equivalent to metric cotype, and one can always take .
The major open question is wether one can take . Indeed, this would have many geometric applications.
Theorem 25 (Eskenazis – Mendel – Naor 2019) For normed spaces, martingale cotype implies metric cotype, with .
6.1. Proof of the remaining distorsion lower bounds in Theorem 18
I will cheat and identify with . This is not a serious matter, since
We must show that
Since decreasing decreases the -distorsion, one can assume that . Let have distorsion . Then the lefthand side of metric cotype assumption is , whereas the right hand side is , so .
We see that we really need .
6.2. Proof of Theorem 26
The strategy is inspired from hypercontractivity: smoothing the space using an averaging operator puts functions closer to linear functions, i.e. leads us closer to the linear setting.
Given , we define a martingale
where has length . We view as a smoothing operator.
If and , let . Then
where (1) is an approximation term,
and (2) is a smoothing term,
We first estimate (1) with Hölder,
Jensen’s inequality gives
In other words, contracts , so
which is ok since , this what we want to see on the righthand side.
Next we bound (2). We can replace with mod . For fixed , write the summand in (2) as a telescopic sum,
But
By martingale cotype,
Adding gives the expected bound .
6.3. Final comment
What does for you is that it provides invariance under coarse embeddings, and not merely biLipschitz embeddings. So the story is far from being finished.