Let be the free group on two generators, and let be the endomorphism defined on generators by and . We define *Sapir’s group* to be the ascending HNN extension

This group was studied by Crisp-Sageev-Sapir in the context of their work on right-angled Artin groups, and independently by Feighn (according to Mark Sapir); both sought (unsuccessfully) to determine whether contains a subgroup isomorphic to the fundamental group of a closed, oriented surface of genus at least 2. Sapir has conjectured in personal communication that does not contain a surface subgroup, and explicitly posed this question as Problem 8.1 in his problem list.

After three years of thinking about this question on and off, Alden Walker and I have recently succeeded in finding a surface subgroup of , and it is the purpose of this blog post to describe this surface, how it was found, and some related observations. By pushing the technique further, Alden and I managed to prove that for a fixed free group of finite rank, and for a* random endomorphism* of length (i.e. one taking the generators to random words of length ), the associated HNN extension contains a closed surface subgroup with probability going to 1 as . This result is part of a larger project which we expect to post to the arXiv soon.

The context of this problem is Gromov’s notorious question:

**Question(Gromov):** Does every 1-ended hyperbolic group contain a surface subgroup?

Actually, it is not at all clear if Gromov really asked this question, or what sort of answer he expected. There is a discussion of this in the introduction to a recent paper by Henry Wilton. A positive answer to this question is known in only a few special cases, including

- Coxeter groups (Gordon-Long-Reid)
- Graphs of free groups with cyclic edge groups and (Calegari)
- Fundamental groups of hyperbolic 3-manifolds (Kahn-Markovic)
- Certain doubles and graphs of free groups with cyclic edge groups (Kim-Wilton, Kim-Oum, Kim, Wilton)

(this list is not exhaustive). One strategy to find a surface subgroup is to define a class of groups with the property that every one-ended hyperbolic group contains a subgroup in the class , and then to show that every group in this class further contains a surface subgroup. A reasonable candidate for the class is the class of *one-ended graphs of free groups*. The logic behind this choice is that it is very easy to produce many free subgroups of a one-ended hyperbolic group (in fact, this is more or less the only kind of subgroup one knows how to produce) by Klein’s pingpong argument, and one could perhaps argue that because there are so many such subgroups, that intersect in quite rich and interesting ways, a sufficiently rich collection is one-ended while at the same time has the structure of a graph of groups. On the other hand, the structure of a graph of free groups is similar in some ways to the structure of a Haken 3-manifold, and one knows enough about the components of the graph (i.e. the free factors) that one can try to build a surface subgroup by amalgamating surface-with-boundary subgroups along cyclic subgroups of the edge groups.

Anyway, this is more philosophy than mathematics, but it does partly explain why the class has been widely studied by geometric group theorists interested in Gromov’s question. One important class of graphs of groups are the HNN extensions, whose underlying graphs consist of a single vertex and a single edge joining this vertex to itself. An (injective) endomorphism of a free group thus gives rise to an HNN extension in the class .

Now, suppose is a map from a surface subgroup to . There is a homomorphism sending to 0 and the conjugating element to . The kernel intersected with the image of will determine an infinite cyclic cover of , and one would like to determine whether this map is injective. We can think of as an infinite union of subsurfaces with boundary, where each is attached to and , and contained in a conjugate of the subgroup . If we identify each with for , then we can think of . Let denote the union of the with . Evidently it is sufficient to show that the inclusion of to is injective, since any loop in the kernel of is conjugate into . This is convenient, since we can discuss surface-with-boundary subgroups of a fixed free group, and essentially ignore the endomorphism .

The first thing to check is that each separate inclusion is injective. Each may be represented by a certain kind of diagram, called a *fatgraph*. Basically, a fatgraph is a graph in the usual sense, together with a choice of cyclic ordering of the edges incident to each vertex. A fatgraph embeds canonically as the spine of some surface which itself deformation retracts back to , in such a way that the cyclic order on edges inherited from the embedding agrees with the fatgraph structure. The oriented edges of are labeled with reduced words in in such a way that the labels on opposite sides of an edge of are inverse in . In this way, a fatgraph “represents” a surface-with-boundary mapping to . Here is an example of a (disconnected) fatgraph, whose underlying surface is homeomorphic to the union of two 4-punctured spheres:

Now, the fundamental group of every (component of every) fatgraph is free, but the map to is not necessarily injective. Stallings gave a celebrated criterion for a simplicial map from a graph to a rose (i.e. a standard graph with fundamental group ) to be injective, namely that the map should be *folded*, or equivalently, that the map should be an immersion on the link of every vertex. In terms of fatgraphs, this means that there should be at most one incoming edge at each vertex with each label. The graph pictured above is folded in this sense. Notice if every boundary word is reduced, a 2- or 3-valent vertex is necessarily (locally) folded.

OK, this is a criterion that will certify that an individual might be injective, when represented as a fatgraph. What about the dynamics of ? Notice that the endomorphism has a particularly nice property: if we think of it as representing a self-map of the standard rose to itself, then the map is an *immersion*, in the sense of Stallings. This means that if each of the surfaces is represented by a folded fatgraph , then each will be folded if is. This suggests the following definition:

**Definition.** A fatgraph with associated surface is *-folded* if there is a decomposition of its boundary into and in such a way that (with the opposite orientation), and satisfying the following properties:

- The graph is Stallings folded
- Every -vertex in (i.e. the images under of the vertices of ) is associated to a 2-valent vertex of
- No vertex of is associated to more than one -vertex in
- No vertex of is associated to more than one vertex in

When we talk about a vertex of being “associated” to a vertex of we mean that the vertex of maps to the given vertex of under the deformation retraction of to (this deformation retraction is simplicial when restricted to ).

Now, suppose is -folded. We can glue to by gluing to . Condition 4 implies that the resulting surface is where . In a similar way we can define

and . Now, conditions 2 and 3 imply that every vertex of is obtained by gluing some vertex of to a sequence of 2-valent vertices in various with . In particular, since every vertex of is locally folded, the same is true of every vertex of , and therefore also of . Hence is folded, and thus injective. Since as above, it follows that the suspension of an -folded surface is injective in .

The definition of -folded can be modified for an endomorphism which is not an immersion of . One of the main theorems Alden and I prove is that a “random” endomorphism admits many -folded surfaces in this sense, and therefore the associated HNN extension has (many) surface subgroups. For a random endomorphism of length , the genus of these surfaces will typically be of order at least , but the number will grow at least like for genus .

Now, Sapir’s group is certainly not random in any sense; nevertheless, it is possible to search for an -folded surface. A priori finding an -folded surface with given boundary seems to require trying exponentially many gluings, and is apparently impractical. However, Alden and I are able to show that the search for such a surface can be reduced to a linear programming problem, and thus becomes eminently practical. Sure enough, a computer search rapidly found the following example of an -folded surface in Sapir’s group:

In a bit of detail: the picture above is a fatgraph whose boundary decomposes into three components labeled and one component labeled . There is a 3-fold cover whose boundary decomposes into consisting of three copies of and consisting of three components labeled . The components of are the ones indicated by the blue circles, and one can see that they are embedded, satisfying condition 4. The red dots are the -vertices, and one can check that they are distinct and on 2-valent vertices of the fatgraph. Finally, one can check that the surface is folded in the usual sense of Stallings. It follows that the suspension is an injective surface in Sapir’s group, of genus 31.

(added Thursday, February 21, 2013): Jack Button has just posted a paper to the arXiv making the observation that a random HNN extension of a free group (in the sense of Alden and I, as above) will satisfy the small cancellation condition for any as , with probability , and therefore will be the fundamental group of a special cube complex, by a result of Wise. This is good to know, and underlines the extent to which such HNN extensions resemble 3-manifold groups.

Well done! This is a really great result.

I notice that you don’t mention scl in this post. Does it appear explicitly in the details of the proof, or is its influence here just implicit? (For instance, the ideas here seem related to your joint proof that a random endomorphism is an scl-isometry.)

Hi Henry – thanks for the compliment, which I will happily accept on behalf of Alden and myself!

scl does make an appearance elsewhere in the paper, but proving that has a surface subgroup for random does not directly use scl (but you’re right that there is some overlap with Alden’s and my “Random rigidity” paper). However, a key step of the proof is computer assisted, and it turns out that scallop can be very easily adapted to look for f-folded surfaces just as effectively as for extremal surfaces.

By the way, I did think about mentioning your new result (which is obviously germane to the theme of this post), but I figured you might want to announce it yourself when you are ready. :)

Thanks, Danny. I believe the proof of the new result is now correct and complete, but as it’s in somewhat unfamiliar territory for me I’d like to double-check it. So I would rather leave the multitudes waiting on tiptoes for the moment…

Are these surface subgroups quasiconvex? And my toes are getting tired.

Hi Ian – yes, I believe these surface subgroups should be quasiconvex (although we have not actually written down a proof). In any case, their injectivity is “certified locally”, and there should be some general principle that subgroups of hyperbolic groups whose injectivity is certified locally should be quasiconvex (although it is probably not easy to make that intuition precise).

If your toes send Henry an email, I’m sure he’ll be happy to fill them in.

It would be interesting if you could construct codimension-one quasiconvex subgroups of (hyperbolic) ascending HNN extensions of free groups. One might then be able to cubulate them this way. Guillaume Dufour did this in his thesis for mapping tori of surfaces (using a method of Cooper-Long-Reid and Masters to get quasiconvex surface subgroups).

This is true (it would be a great theorem!), but there’s no particular reason to think that the codimension-one subgroups should be surface groups, is there?

I don’t see a reason right now why these surface subgroups should be codimension one. But of course this is worth thinking about.

The naive analogue of the Cooper–Long–Reid–Masters–Dufour construction (if it works!) would produce *free* subgroups of codimension one.

What is the C-L-R-M-D construction? What would be the analog in this context? What is codimension one about it?

The idea is, in a fibred hyperbolic 3-manifold, to ‘cut and cross-join’ the fibre surface to produce immersed geometrically finite surfaces. The original construction was in C–L–R’s paper ‘Bundles and finite foliations’, I think. Masters showed that there are many such surfaces, in some sense, and Dufour showed that there really are enough to cubulate.

I get it – we have a surface fiber S, and cut it along a loop g, then pull one end of g once around the circle direction. Then one just needs to find an immersed subsurface of S bounding g-phi(g) where phi is the monodromy. Providing g-phi(g) is homologically trivial (eg if g is homologically trivial) such an immersed surface always exists, and can be chosen to represent all but finitely many classes in relative H_2. The surface obtained by gluing up is essential; actually, it’s even Thurston-norm minimizing, and in the boundary of the fibered face, since it’s transverse to the suspension flow (by max flow – min cut).

This is very interesting, since Alden and I actually do look at such surfaces in our paper – we call them “flat surfaces”, and they are pretty common, but not completely omnipresent, because the surface fiber S will have boundary that can’t be ignored so easily. We conjecture they can always be found virtually. But if one just lets them have boundary – so they are free groups rather than closed surface subgroups – they are indeed easy to find. Maybe there are enough to cubulate?