Flat Ultrafilters (Michigan Logic Seminar Sept 21, 2011)

25 Sep 2011

Remember how our about page says that Booles' Rings is about best practices for an academic homepage? Ok, let's try one: making notes to talks available.

Some introductory remarks

Skip this section if you only want mathematics.

Wednesday, I gave a short talk about flat ultrafilters at our Logic Seminar here at the University of Michigan (as announced on Set Theory Talks, the talk was recorded and the video ~~will be online eventually~~ is now online).

Flat Ultrafilters (2011/09/21 University of Michigan Logic Seminar) from Peter Krautzberger on Vimeo.

When I visited Toronto in June, Ilijas Farah introduced me to this somewhat strange new type of utrafilter on $\omega$ . My talk this week was mostly about the results from a 2009 paper by Ilijas Farah, N. Christopher Philips and Juris Steprans, doi, but some of what I am going to write are insights I gained while discussing this notion with François Dorais and Andreas Blass. So, actually, this post is attempting another best practice: notes from reading a paper.

I would like to explain something about the result that P-points are not flat. When I started looking at their paper with Francois Dorais, we first re-proved that selective ultrafilters are not flat -- instead of the (possibly stronger) functional analytic result from the paper we used the combinatorial definition of flat ultrafilters. Then we worked on improving that proof and, with Andreas's help, got it to work for P-points.

Only after all of this happened did we notice that at the very end, the paper had already announced that they have a proof that P-points are not flat. Last week, Ilijas kindly send me some slides with further results on flat ultrafilters; even though the proof for P-points isn't in there I would guess from the formulation on the slides that our proofs are essentially the same.

Long story short, the proof "P-points are not flat" below is "our" proof even though the result (and most likely the proof) should be credited to at least Steprans (according to the slides).

I will focus on my own interpretation of these notions, i.e., my rephrasing of the definition of flat ultrafilters; they might look different from what you'll find in their paper even though it is formally the same notion.

Alright, nuff said. My talk was designed to first sketch the main functional analysis result of their paper. So the first two sections will be void of much explanations or proofs. But they motivate why the notion of flat ultrafilter is as strange as it is -- it simply came out of those considerations.

Some functional analysis

Let $H$ be a (the) separable, infinite dimensional Hilbert space, i.e.,

H\cong l_ 2(\omega) = \{ s: \omega \rightarrow \mathbb{C}, \sum_ {n\in \omega} \left\vert s_ n\right\vert ^2 \lt \infty \}.

Then $B(H)$ is the space of bounded, linear operators, i.e.,

B(H) = \{ T: H\rightarrow H: T \text{ linear }, (\exists M)(\forall v\in H) \left\Vert Tv\right\Vert \leq M\left\Vert v\right\Vert \},

and $K(H)$ the space of compact operators

K(H) = cl( \{ T \in B(H): T[H] \text{ finite dimensional } \}).

Finally, the Calkin Algebra is the quotient $C(H) = B(H) / K(H)$ .

Farah, Philips and Steprans were interested in the relative commutant of subalgebras $A\subseteq B$ :

F_ A(B) = \{ a\in A: (\forall b\in B) ab=ba \}.

The relative commutant in the ultrapower

If $p \in \beta \omega$ is an ultrafilter, we can consider the norm-ultrapowers of $B(H)$ and $C(H)$ which I'll denote by $B(H)^p, C(H)^p$ . That is

B(H)^p = \ell_ \infty(B(H) / \{c \in \ell_ \infty(B(H)): \lim_ p \left\Vert c\right\Vert = 0\}.

(Similarly for $C(H)^p$ .)

Kirchberg had shown that $F_ {C(H)^p}(C(H)) = \mathbb{C} \cdot 1$ regardless of $p$ and asked it the analogous statement might be true of $B(H)$ .

As a response, Farah, Philips and Steprans showed that $F_ {B(H)^p}(B(H))$ depends on the choice of $p$ (under CH); the results from their paper relevant to this talk are as follows.

Theorem (Farah, Philips, Steprans)

If $p$ is a flat ultrafilter (see below), then $F_ {B(H)^p}(B(H)) \neq \mathbb{C} \cdot 1$ .

If $p$ is selective, then $F_ {B(H)^p}(B(H)) = \mathbb{C} \cdot 1$ ; in particular, selective ultrafilters are not flat.

Flat ultrafilters exist under ZFC.

It was announced that P-points are not flat.

Curious fact: under CH all ultrapowers of $B(H)$ are isomorphic (via model theoretic arguments). Yet, by the above, if we take a flat and a selective ultrafilter, no such isomoprhism can map $B(H)$ to $B(H)$ (or else the relative commutants would be isomorphic). This is an odd situation.

If you look at the recording you will, as usual, see lots of confusion on my part. In particular, I remembered that the first result in the theorem was an equivalence. It turned out that the paper does not say so and this is an open question. In my defence, the slides did give it as an equivalence.

I'll only give the proofs of the last two statements as well as some further observations on flat ultrafilters.

Flat ultrafilters

What are flat ultrafilters? Well, let's first introduce the assisting structure of a flatness scale, a set of sequences in $[0,1]$ .

Definition
A flatness scale H is a countable subset of
$\{ h:\omega \rightarrow [0,1]: h(0) =1, \lim_ {i \rightarrow \infty} h(i)=0 \}.$

Addendum Nothing spectacular so far -- a flatness scale is just a bunch of sequences converging to $0$ . As such they can have some wild behavior along the way -- but flat ultrafilters tame them.

The original definition of flat ultrafilters is then phrased as follows.

Definition [Farah, Philips, Steprans]
An ultrafilter $p$ is flat if there is a flatness scale $H = (h_ n: n\in \omega)$ such that for every increasing $f: \omega \rightarrow \omega$ with $f(0) \gt 0$
$\lim_ {n\to p} \left\Vert (h_ n - h_ n \circ f)\right\Vert _ \infty = 0.$
In other words, for every $f$ as above and every $\varepsilon\gt 0$ , we have
$\{n\in \omega: \sup_ i \left\vert h_ n(i) - h_ n(f(i))\right\vert \lt \varepsilon \} \in p .$
We then say that $H$ is a flatness scale for p.

One of the goals was to find out how I can best think about the notion of flat ultrafilters.

Addendum What to make of this? The first observation is that flat ultrafilters have a flatness scale such that, given $\varepsilon$ , almost all of the sequences (with respect to the ultrafilter) drop by at most $\varepsilon$ . That's slow, but it's even slower! The functions $f$ in the definition should be thought of as vastly increasing (think: dominating family) and they code huge intervals of the form $[n,f(n)]$ . But even given those vast intervals, still almost all sequences of the flatness scale drop only by $\varepsilon$ on such long intervals. In other words, they never really drop by much, yet they converge to $0$ .

One of the things I found irritating was to think of them as ultrafilters on $\omega$ . Thankfully, the enumeration of $H$ does not matter much. In other words, I don't have to think of $p$ as an ultrafilter on $\omega$ that just happens to come with a map to a flatness scale $H$ -- I can really think of $p$ to be an ultrafilter on $H$ . So allow me to re-phrase.

Definition [reformulation] Given a flatness scale $H$ and an ultrafilter $p$ on $H$ , we say that $p$ is flat , if for every increasing $f:\omega\to\omega$ with $f(0) \gt 0$ and every $\varepsilon\gt 0$ ,
$\{ h \in H: \sup_ i \left\vert h(i) - h (f(i))\right\vert \lt \varepsilon \} \in p .$
In other words,
$\lim_ {p} \left\lVert h - h \circ f)\right\rVert_ \infty = 0.$

Here $\lim_ p$ is the usual limit along the ultrafilter $p$ .

You might wonder if we're not loosing too much information -- after all, the original definition did not forbid repetitions. We'll see later that this is not a problem. For now, I will simply stick to the reformulation.

First there's an obvious question about how complicated a flatness scale can be. So let me show you their construction of flat ultrafilters.

Constructing Flat Ultrafilters

A simple observation

For the construction of flat ultrafilters, the key observation is as follows: all we have to do is find a flatness scale $H$ such that the sets

X_ {f,\varepsilon} = \{h \in H : \sup_ i\left\vert h(i) - h (f(i)\right\vert \lt \varepsilon \}

are infinite for every increasing $f:\omega\to\omega$ with $f(0) \gt 0$ and every $\varepsilon \gt 0$ .

Once we have this, we get the finite intersection property for free since

X_ {\max(f_ 1,f_ 2),\min(\varepsilon_ 1,\varepsilon_ 2)} \subseteq X_ {f_ 1,\varepsilon_ 1} \cap X_ {f_ 2,\varepsilon_ 2}.

In other words, any ultrafilter $p$ containing all the sets $X_ {f,\varepsilon}$ is flat -- witnessed by the flatness scale $H$ .

Luckily, It turns out that there is an extremely simple flatness scale with this property.

Proposition [Farah, Philips, Steprans] There is a flatness scale $H$ such that the set
$X_ {f,\varepsilon} = \{h \in H : \sup_ i \vert h(i) - h(f(i)\vert \lt \varepsilon \}$
is infinite for every increasing $f:\omega\to\omega$ with $f(0) \gt 0$ and every $\varepsilon \gt 0$ .

Proof

Consider the finite subsets of $\omega$ , $[\omega]^{\lt \omega}$ .
We can think of $s\in [\omega]^{\lt \omega}$ as encoding a simple step function, dropping by $1/\vert s\vert$ at each of the elements of $s$ ; in other words, we define $h_ s : \omega \rightarrow [0,1]$ by

h_ s(i) = 1- \frac{\vert s\cap i\vert }{\vert s\vert }.

Clearly, $H_ S := \{ h_ s: s \in [\omega]^{\lt \omega} \}$ is a flatness scale (starting at $1$ and converging to zero, in fact being eventually zero).
Now imagine we're given some increasing $f:\omega\to\omega$ with $f(0) \gt 0$ as well as some $\varepsilon \gt 0$ .
Then we can easily find infinitely many with the following properties:
- $1/\vert s\vert \lt \varepsilon$
- With the natural enumeration of $(s_ i)_ {i\in \vert s\vert }$ of $s$ ,
for each .
But this means that for all $i$ ,

h_ s(f(i)) \geq h_ s(i+1) = \frac{1-\vert s\cap (i+1)\vert }{\vert s\vert }.

It follows that