In search of a counterexample to the Lovász conjecture

It is a celebrated result of John Dixon, (The probability of generating the symmetric group, (subscription) Math. Z. 110 (1969), 199–205.) that if one choose two random permutations in the symmetric group , uniformly (i.e. each with probability ), and independently, the probability that the two permutations generate the whole group tends to as . It is clear that this probability will never be greater than , since there is a probability that the two permutations will both be even, in which case you could only generate, at most, the alternating group. Interestingly enough, Dixon’s paper covers this possibility, and he actually shows that the probability that two permutations generate the alternating group tends to as .

Equivalently, if two random elements of the alternating group are chosen uniformly and randomly, the probability that they generate the group tends to as . This leads me to my question — what is the probability that the -regular Cayley graph with generators is not Hamiltonian, as ?

Showing that this probability is bounded away from would provide a counterexample for a notorious problem about vertex-transitive graphs. So we might expect that this is hard. But is it even possible that it is true, or is there some obvious reason that such graphs will tend to be Hamiltonian?

Another approach in the same spirit would be computational rather than asymptotic. Suppose we look at thousands of random Cayley graphs on the alternating groups and , for example. It is straightforward to check that they are connected. Is it within reach for a cleverly designed algorithm on modern computers to conclusively rule out Hamiltonicity for a -regular graph on or vertices? I would also be happy with a computer-aided proof that the conjecture is false.

Historical note: It is called the Lovász conjecture, even though he just asked the question (and perhaps conjectured the other way). I am under the impression that some prominent people in this field have felt that the answer should be no. In particular Babai does not believe it.

Analyzing card shuffling machines

Diaconis, Fulman, and Holmes have uploaded a preprint titled, “Analysis of Casino Shelf Shuffling Machines.” The paper provides a brief overview of the venerable history of mixing time of card shuffling, all the way back to early results by Markov and Poincaré, and their main point is to analyze a model of shuffle that had not been studied previously. What I found most interesting, though, was their account of successfully convincing people in the business of making card shuffling machines that their machines weren’t adequately mixing up the cards. They gave the manufacturers one mathematical argument, based on total variation distance, that they didn’t accept, and then another argument, based on a card guessing game, that they did.

I’ll describe the card guessing game. I flip through a deck of 52 cards, one card at a time, and before I flip a card you try to guess what it will be. Let’s say you have a perfect memory for every card that’s already been flipped, so you obviously won’t guess those. On the other hand, if the cards are in a truly random order to start out, you obviously don’t have any better strategy than to guess uniformly among the remaining cards. An easy analysis gives that your best possible expected number of correct guesses is . On the other hand, the authors described a strategy (conjectured to be best possible) that allows one to guess an average of cards correctly, on a totally ordered deck run through the shelf shuffling machine only once. This suggests strongly that the cards are not sufficiently random.

This analysis convinced the company to have the shelf shuffling machine make two passes through the deck, rather than one as they had initially hoped. The president of the company told them that “We are not pleased with your conclusions, but we believe them and that’s what we hired you for.”

The fundamental group of random 2-complexes

Eric Babson, Chris Hoffman, and I recently posted major revisions of our preprint, “The fundamental group of random 2-complexes” to the arXiv. This article will appear in Journal of the American Mathematical Society. This note is intended to be a high level summary of the main result, with a few words about the techniques.

The Erdős–Rényi random graph is the probability space on all graphs with vertex set , with edges included with probability , independently. Frequently and , and we say that asymptotically almost surely (a.a.s) has property if as .

A seminal result of Erdős and Rényi is that is a sharp threshold for connectivity. In particular if , then is a.a.s. connected, and if , then is a.a.s. disconnected.

Nathan Linial and Roy Meshulam introduced a two-dimensional analogue of , and proved an analogue of the Erdős-Rényi theorem. Their two-dimensional analogue is as follows: let denote the probability space of all 2-dimensional (abstract) simplicial complexes with vertex set and edge set (i.e. a complete graph for the 1-skeleton), with each of the triangles included independently with probability .

Linial and Meshulam showed that is a sharp threshold for vanishing of first homology . (Here the coefficients are over . This was generalized to for all by Meshulam and Wallach.) In other words, once is much larger than , every (one-dimensional) cycle is the boundary of some two-dimensional subcomplex.

Babson, Hoffman, and I showed that the threshold for vanishing of is much larger: up to some log terms, the threshold is . In other words, you must add a lot more random two-dimensional faces before every cycle is the boundary of not any just any subcomplex, but the boundary of the continuous image of a topological disk. A precise statement is as follows.

Main result Let be arbitrary but constant. If then , and if then , asymptotically almost surely.

It is relatively straightforward to show that when is much larger than , a.a.s. . Almost all of the work in the paper is showing that when is much smaller than a.a.s. . Our methods depend heavily on geometric group theory, and on the way to showing that is non-vanishing, we must show first that it is hyperbolic in the sense of Gromov.

Proving this involves some intermediate results which do not involve randomness at all, and which may be of independent interest in topological combinatorics. In particular, we must characterize the topology of sufficiently sparse two-dimensional simplicial complexes. The precise statement is as follows:

Theorem. If is a finite simplicial complex such that for every subcomplex , then is homotopy equivalent to a wedge of circle, spheres, and projective planes.

(Here denotes the number of -dimensional faces.)

Corollary. With hypothesis as above, the fundamental group is isomorphic to a free product , for some number of ‘s and ‘s.

It is relatively easy to check that if then with high probability subcomplexes of on a bounded number of vertices satisfy the hypothesis of this theorem. (Of course itself does not, since it has and roughly as approaches .)

But the corollary gives us that the fundamental group of small subcomplexes is hyperbolic, and then Gromov’s local-to-global principle allows us to patch these together to get that is hyperbolic as well.
This gives a linear isoperimetric inequality on which we can “lift” to a linear isoperimetric inequality on .

But if is simply connected and satisfies a linear isoperimetric inequality, then that would imply that every -cycle is contractible using a bounded number of triangles, but this is easy to rule out with a first-moment argument.

There are a number of technical details that I am omitting here, but hopefully this at least gives the flavor of the argument.

An attractive open problem in this area is to identify the threshold for vanishing of . It is tempting to think that , since this is the threshold for vanishing of for every integer . This argument would work for any fixed simplicial complex but the argument doesn’t apply in the limit; Meshulam and Wallach’s result holds for fixed as , so in particular it does not rule out torsion in integer homology that grows with .

As far as we know at the moment, no one has written down any improvements to the trivial bounds on , that . Any progress on this problem will require new tools to handle torsion in random homology, and will no doubt be of interest in both geometric group theory and stochastic topology.

A conjecture concerning random cubical complexes

Nati Linial and Roy Meshulam defined a certain kind of random two-dimensional simplicial complex, and found the threshold for vanishing of homology. Their theorem is in some sense a perfect homological analogue of the classical Erdős–Rényi characterization of the threshold for connectivity of the random graph.

Linial and Meshulam’s definition was as follows. is a complete graph on vertices, with each of the triangular faces inserted independently with probability , which may depend on . We say that almost always surely (a.a.s) has property if the probability that tends to one as .

Nati Linial and Roy Meshulam showed that if is any function that tends to infinity with and if then a.a.s , and if then a.a.s .

(This result was later extended to arbitrary finite field coefficients and arbitrary dimension by Meshulam and Wallach. It may also be worth noting for the topologically inclined reader that their argument is actually a cohomological one, but in this setting universal coefficients gives us that homology and cohomology are isomorphic vector spaces.)

Eric Babson, Chris Hoffman, and I found the threshold for vanishing of the fundamental group to be quite different. In particular, we showed that if is any constant and then a.a.s. and if then a.a.s. . The harder direction is to show that on the left side of the threshold that the fundamental group is nontrivial, and this uses Gromov’s ideas of negative curvature. In particular to show that the is nontrivial we have to show first that it is a hyperbolic group.

[I want to advertise one of my favorite open problems in this area: as far as I know, nothing is known about the threshold for , other than what is implied by the above results.]

I was thinking recently about a cubical analogue of the Linial-Meshulam set up. Define to be the one-skeleton of the -dimensional cube with each square two-dimensional face inserted independently with probability . This should be the cubical analogue of the Linial-Mesulam model? So what are the thresholds for the vanishing of and ?

I just did some “back of the envelope” calculations which surprised me. It looks like must be much larger (in particular bounded away from zero) before either homology or homotopy is killed. Here is what I think probably happens. For the sake of simplicity assume here that is constant, although in realty there are terms that I am suppressing.

(1) If then a.a.s , and if then a.a.s .

(2) If then a.a.s. , and if then a.a.s. .

Perhaps in a future post I can explain where the numbers and come from. Or in the meantime, I would be grateful for any corroborating computations or counterexamples.

Topological Turán theory

I just came across the following interesting question of Nati Linial.

If a two-dimensional simplicial complex has vertices and faces, does it necessarily contain an embedded torus?

I want to advertise this question to a wider audience, so I’ll explain first why I think it is interesting.

First of all this question makes sense in the context of Turán theory, a branch of extremal combinatorics. The classical Turán theorem gives that if a graph on vertices has more than edges then it necessarily contains a complete subgraph on vertices. This is tight for every and .

One could ask instead how many edges one must have before there is forced to be a cycle subgraph, where it doesn’t matter what the length of the cycle is. This is actually an easier question, and it is easy to see out that if one has edges there must be a cycle.
It also seems more natural, in that it can be phrased topologically: how many edges must be added to vertices before we are forced to contain an embedded image of the circle?

What is the right two-dimensional analogue of this statement? In particular, is there a constant such that a two-dimensional simplicial complex with vertices and at least two-dimensional faces must contain an embedded sphere ? If so, then this is essentially best possible. By taking a cone over a complete graph on vertices, one constructs a two-complex on vertices with faces and no embedded spheres. Without having thought about it at all, I am not sure how to do better.

In any case, the corresponding question for torus seems more interesting, but for different reasons. In a paper with Eric Babson and Chris Hoffman we looked at the fundamental group of random two-complexes, as defined by Linial and Meshulam, and found the rough threshold for vanishing of the fundamental group. To show that the fundamental group was nontrivial when the number of faces was small required a lot of work — in particular, in order to apply Gromov’s local-to-global method for hyperbolicity, we needed to prove that the space was locally negatively curved, and this meant classifying the homotopy type of subcomplexes up to a large but constant size.

It turned out that the small subcomplexes were all homotopy equivalent to wedges of circles, spheres, and projective planes. In particular, we show that there are not any torus subcomplexes, at least not of bounded size. (Linial may have recently shown that there are not embedded tori, even of size tending to infinity with .) On the other hand, just on the other side of the threshold embedded tori abound in great quantity. It is interesting that something similar happens in the density random groups of Gromov — that the threshold for vanishing of the density random group corresponds to the presence of tori subcomplex in the naturally associated two-complex. It is not clear to me if this is a general phenomenon, coming geometrically from the fact that a torus admits a flat metric.

Some of the great successes of the probabilistic method in combinatorics have been in existence proofs when constructions are hard or impossible to come by. It would be nice to have interesting or extremal topological examples produced this way. Nati’s question suggests an interesting family of extremal problems in topological combinatorics, and it might make sense that in certain cases, random simplicial complexes have nearly maximally many faces for avoiding a particular embedded subspace.

Update: Nati pointed me to the paper Sós, V. T.; Erdos, P.; Brown, W. G., On the existence of triangulated spheres in $3$-graphs, and related problems. Period. Math. Hungar. 3 (1973), no. 3-4, 221–228.
Here it is shown that is the right answer for the sphere. Their lower bound is constructive, based on projective planes over finite fields. Nati said that being initially unaware of this paper, he found a probabilistic proof that works just as well as a lower bound for every fixed 2-manifold. So it seems that the main problem here is to find a matching upper bound for the torus.

The Rado complex

A simplicial complex is said to be flag if it is the clique complex of its underlying graph. In other words, one starts with the graph and add all simplices of all dimensions that are compatible with this -skeleton. A subcomplex of a flag complex is said to be induced if it is flag, and if whenever vertices and is an edge of , we also have that is an edge of .

Does there exist a flag simplicial complex with countably many vertices, such that the following extension property holds?

[Extension property] For every finite or countably infinite flag simplicial complex and vertex , and for every embedding of as an induced subcomplex , can be extended to an embedding of as an induced subcomplex.

It turns out that such a does exist, and it is unique up to isomorphism (both combinatorially and topologically). Other interesting properties of immediately follow.

– contains homoemorphic copies of every finite and countably infinite simplicial complex as induced subcomplexes.

– The link of every face is homeomorphic to itself.

– The automorphism group of acts transitively on -dimensional faces for every .

– Deleting any finite number of vertices or edges of and the accompanying faces does not change its homeomorphism type.

Here is an easy way to describe . Take countably many vertices, say labeled by the positive integers. Choose a probability such that , and for each pair of integers , connect to by an edge with probability . Do this independently for every edge.

This is sometimes called the Rado graph, and because it is unique up to isomorphism (and in particular because it does not depend on ) it is sometimes also called the random graph. It is also possible to construct the Rado graph purely combinatorially, without resorting to probability. The I have in mind is of course just the clique complex of the Rado graph.

We can filter the complex by setting to be the induced subcomplex on all vertices with labels , and this allows us to ask more refined questions. (Now the choice of affects the asymptotics, so we assume .) From the perspective of homotopy theory, is not a particularly interesting complex; it is contractible. (This is an exercise, one should check this if it is not obvious!) However, has interesting topology.

As , the probability that is contractible is going to . It was recently shown that has asymptotically almost surely (a.a.s.) at least nontrivial homology groups, concentrated around dimension . For comparison, the dimension of is .

I think one can probably show using the techniques from this paper that there is a.a.s. no nontrivial homology above dimension or below dimension . It is still not clear (at least to me) what happens between dimensions and . It seems that a naive Morse theory argument can give that the expected dimension of homology is small in this range, but to show that it is zero would take a more refined Morse function. Perhaps a good topic for another post would be “Morse theory in probability.”

Another question: given a non-contractible induced subcomplex (say an embedded -dimensional sphere) on a set of vertices, how many vertices should one expect to add to before becomes contractible in the larger induced subcomplex? For example, it seems that once you have added about vertices, it is reasonably likely that one of these vertices induces a cone over , but is it possible that the subcomplex becomes contractible with far fewer vertices added?

Threshold behavior for non-monotone graph properties

One of my research interests is what you might call non-monotone graph properties.

A seminal result of Erdős and Rényi is that if , then the random graph is a.a.s. disconnected, while if then is a.a.s. connected. This can be made more precise; for example, if with , then as . Connectivity is an example of a monotone graph property, meaning a set of graphs either closed under edge deletion or edge contraction. Other examples would be triangle-free, k-colorable, and less than five components. The fact that every monotone graph property has a sharp threshold is a celebrated theorem of Friedgut and Kalai.

More generally, a graph property is a way of assigning numbers to finite graphs that is invariant under graph isomorphism, and increases or decreases with edge deletion or addition. Examples would be the number of triangle subgraphs, chromatic number, and number of connected components.

Much of random graph theory is concerned with monotone properties of random graphs. But it is not hard to think of examples of non-monotone properties. For example, let the number of induced four-cycle subgraphs. Another example would be the number of complete -subgraphs that are not contained in any -subgraph. When is very large,  , then a.a.s. every three vertices share some common neighbor, so every -subgraph is contained in a -subgraph and . Similarly, when is very small, , then a.a.s. there are no subgraphs, and . For intermediate values of , , the expected value of is .

So the expected value of the graph property is unimodal in edge density, and we still have threshold behavior. Can Friedgut and Kalai’s result be extended to a more general setting that includes these cases?

This is a simple and perhaps natural combinatorial example, but my original motivation was topological. I wrote a paper about topological properties of random clique complexes, available on the arXiv, which turn out to be non-monotone generalizations of the original Erdős-Rényi theorem. I will continue to write more about this and other related examples sometime soon, and as I have time, but in the meantime if anyone knows any nice examples of non-monotone graph properties, please leave a comment.

One more thought for now. It seems in many of the most natural examples we have of non-monotone properties , the expected value of over the probability space is basically a unimodal function in the underlying parameter . Can you give any natural examples of bimodal or multimodal graph properties? (Pathological examples?)