Cai Heng Li, Gabriel Verret and myself are currently advertising for a two year postdoc to work on our ARC grant entitled `Symmetries of finite digraphs’. The deadline is the 19th of February. More details are available at
Over the last few years, Minion (I’m giving the link because if you try to search for it, you’ll be swamped by Despicable Me merchandise) has become one of my “go-to” tools, mostly because for certain specific types of searches it seems to perform just as well as, or even better than, a bespoke program. And of course, writing, debugging and optimising a bespoke program is enormously time-consuming, while writing a Minion model is usually very quick.
Most recently, I had an email from Brendan (McKay), who has found three planar cubic graphs of girth 5 that are hypohamiltonian – a graph is hypohamiltonian if it is not hamiltonian, but for every vertex , the graph is hamiltonian. Brendan’s graphs had 76 vertices, so not huge, but big enough.
This is the story of the shameful conjecture, which started with Dominic Welsh, was almost solved by Paul Seymour, and eventually resolved by my friend Fengming Dong from Nanyang in Singapore.
The conjecture concerns a graph parameter called the mean colour number defined as follows: If is an -vertex graph, then is the average number of colours used over all proper colourings of using a palette of n colours.
As this is a slightly strange parameter, an example is probably in order, and it seems only appropriate to use the Welsh graph (that is, the graph obtained by deleting one edge from ) for this purpose.
Using a palette of colours, there are colourings that use all four of the colours. There are colourings that use a fixed set of colours (the two vertices of degree two must be assigned the same colour), and there are ways of choosing colours of out , making a total of colourings using three colours. So the mean colour number of the Welsh graph is as the colourings are evenly divided between – and -colourings (there are no colourings with fewer than colours).
Using some straightforward identities, it can be shown that
where is the chromatic polynomial of .
The chromatic polynomial of the Welsh graph is so for this graph the expression has the value as expected.
It is obvious that the complete graph has the largest mean colour number, because every colouring requires colours, so .
At the other extreme is the empty graph on vertices, which has chromatic polynomial and thus
and as gets large the factor famously converges to from below.
But hang on, why did I say “the other extreme”? How do I know that the empty graph has the lowest mean colour number? Of course, it seems intuitively obvious that adding an edge to a graph can only increase the average number of colours required to colour it. Unfortunately, while it is intuitively obvious, it is not actually true that adding edges increases the mean colour number. Mike Mosca (yes, the quantum crypto guy from Waterloo) found an infinite family of counterexamples, of which the smallest has only 6 vertices.
So is it even true that is the extreme graph? Bartels and Welsh conjectured that this is indeed the case – namely that for any -vertex graph
For a graph theorist, it is a little embarrassing to define a parameter applicable to all graphs, with the complete graph trivially at one extreme, and then be unable to show that the empty graph is the other extreme. In fact, more than just embarrassing, perhaps even positively shameful.
Thus the shameful conjecture was christened.
Now Paul Seymour enters the ring. Normally if he turns his attention to some topic, even briefly, he quickly vacuums up and solves most of the known unsolved conjectures, generates some much harder and deeper conjectures, and then moves on. In this case though, he didn’t quite knock off the shameful conjecture. Instead, he proved that for any n-vertex graph ,
As , this value is just a little bit too low to prove the shameful conjecture.
The shameful conjecture hung on for another few years, but ultimately Fengming Dong (who has resolved quite a number of difficult chromatic and flow polynomial questions), used one of his typically intricate and ingenious proofs to show that the shameful conjecture is indeed true.
I don’t know if the shameful conjecture has stimulated any further research, or if the mean colour number has found any other uses, but I’ve always liked the story…
Written with StackEdit.
All the universities in Australia are having their “Open Day” events, which usually consists of a mix of a children’s carnival plus some information on courses and degrees for secondary school students. Just when you thought the marketing fraternity could not get lower, the geniuses at La Trobe University have introduced an instagrants campaign. The idea was that if you took a selfie in front of a designated research poster, then $150 would be given by the university to that research area. I wonder how long it will take the other universities to come up with this idea too!
We’ve just heard that John has been promoted, so let the celebrations begin!
Well done, John!
Last week I was at the 8th Slovenian Conference on Graph Theory. This was the latest in what is commonly known as the ‘Bled conference’ but this year was in Kranjska Gora. This meant that the conference excursion was to Lake Bled. It was a very enjoyable conference with lots of interesting talks and it was good to catch up with lots of people. I was one of the plenary speakers and my talk was entitled ‘Bounding the number of automorphisms of a graph’. This surveyed the recent work on the Weiss conjecture and its generalisation the PSV conjecture. It also discussed my recent work with Luke Morgan on the PSV conjecture for semiprimitive groups with a nilpotent regular normal subgroup. More details can be found on my slides.
Automorphisms of a 76 million vertex graph
I’ve been meaning to write a post about practical graph automorphism finding for a while, and the last few days experience has finally prompted me to do so now.
The problem of finding the automorphism group of a graph — by this, I mean using software to actually determine the group as a permutation group — is pretty fundamental to a whole range of construction and counting problems. It is obviously very closely related to the graph isomorphism problem, which is to determine whether two graphs are isomorphic. This problem, known as GI, is a problem whose complexity is well known to be unknown (if you see what I mean!). Graph isomorphism is in NP, but it is not known to be either NP-complete or in P, so is a prime candidate for the problem of intermediate complexity that would show that P $\not=$ NP. Finding the automorphism group of a graph is at least as hard as graph isomorphism, because you could use an automorphism-finder to resolve whether G and H are isomorphic by computing the automorphism group of the disjoint union of G and H, and seeing whether any automorphisms swapped the components.
In practice, graph automorphism finding is relatively easy and the automorphism group can be calculated for large graphs, indeed enormous graphs! For more than three decades, the “go-to” software for finding automorphisms of graphs has been Brendan McKay’s program “nauty” which has been used by hundreds, even thousands, of researchers at various points. My thesis, nearly 30 years ago, used one of the early versions of nauty. There are other programs available of course, some of which outperform nauty on particular families of graphs, but none of which match the combination that nauty offers in terms of availability, long-term development and maintenance and overall performance.
Around 25 years ago, I noticed that certain graphs were proving particularly difficult for nauty to process, namely the bipartite point-line incidence graphs of non-desarguesian projective planes. At the time I was working with projective planes of order 9, which have 91 points and 91 lines, and so these were only 182 vertex graphs. I can’t remember the exact numbers, but it took about 0.5 seconds cpu time to process PG(2,9), about 10 seconds to process the Hall plane of order 9 and about 200 hours to find the automorphism group of the Hughes plane of order 9. Of course Brendan immediately put some effort into finding new invariants and other tweaks to make computing the automorphism groups of these projective planes graphs tractable, but they are still much more difficult than most graphs of the same size. (Other geometric graphs such as the incidence graphs of GQs are difficult, but not that difficult.)
A few years ago, Brendan and I were at a meeting in Bristol when we were approached by one of the other delegates who was keen to show us his results on computing automorphism groups of projective planes. He had devised a method of computing the search tree using a hybrid mixture of breadth-first and depth-first search — essentially a single run to the maximum depth of the tree yields an automorphism that can then later be effectively used to prune entire branches of the tree at a very early stage. The results were nothing short of spectacular, with huge projective planes now yielding in seconds rather than days. The researcher in question is called Adolfo Piperno, his technique is called Traces and Adolfo and Brendan have combined forces to refine both techniques and implement them robustly, with the final result being the imaginatively-named “nauty/Traces” package.
So how good is this in practice? For what size of graph can you realistically hope to determine the automorphism group? Pretty big is the answer to that question — in the last few days I’ve managed to use nauty/Traces to determine the automorphism group of a graph with just over 76 million vertices! This graph is a 10-regular graph obtained by taking a primitive representation of the group PSp(6,3) of degree 76422528, which has a non-self paired orbital of valency 5, which thus determines a digraph of out-valency 5. Dropping the directions on the arcs gives us a graph of valency 10. For various reasons, we wanted to know whether the full automorphism group of this graph is just the original PSp(6,3) or bigger.
Of course, it is impossible to store a 76 million x 76 million adjacency matrix, even if it is bit-packed, but fortunately Brendan has recently (in the last few years) added the ability to work with adjacency lists, thus basically using space proportional to the number of edges, rather than the square of the number of vertices. Just forming the file that defined the graph was a tedious process because GAP and Magma can’t do anything fast in groups of this degree, and there are many things that they simply cannot do at all (e.g. find the order of the group!). So I had to “hand-roll” lots of little bits of code that did basic things but in roundabout ways that didn’t use up too much memory and didn’t trigger anything that would never finish.
Once I had the file, the rest was simple. I actually just used the program dreadnaut which is a command-line user interface for nauty/Traces. The input file looked like this:
n=76422528 $=1 g 20385294, 22694895, 50999204, 58103886, 74686745 ; 14658604, 27594574, 45288953, 46264591, 69387673 ; (76422525 lines omitted) 3976598, 18300815, 27911643, 64737249, 74506343 . x
The first two lines set the number of vertices, then the base of the numbering (my vertices are numbered 1 to 76422528, rather than 0 to 76422527). The “g” puts it into graph-input mode, and the next 76422528 lines are the neighbours of each vertex — they are only specified in one direction but unless specified otherwise, nauty/Traces will automatically treat the graph as undirected. Finally, the “x” tells the program to compute the automorphism group.
This whole file — of size 3.6Gb — is then piped into the dreadnaut program with suitable flags to tell it to use Traces and the sparse representation. A little while later, after lots of output of the four permutations, I get the result
1 orbit; grpsize=9170703360; 4 gens; 9 nodes (6 peak);maxlev=2 cpu time = 1550.93 seconds
Unfortunately it wasn’t the result I wanted, because the group of the graph has doubled in size from the guaranteed PSp(6,3) and I was really hoping for this not to happen, because it would have yielded a primitive half-arc transitive graph of the lowest possible valency, whose existence is currently unresolved.
However, I am still hugely impressed that it is possible to find the automorphism group of a 76 million vertex non-trivial graph – in fact I’d really have to say that it’s spectacular!
By way of contrast, consider the NP-complete problem of finding the chromatic number of a graph. This problem is theoretically difficult, but it also seems to be practically difficult – already at 100-200 vertices, there are graphs whose chromatic number I simply cannot determine.
Written with StackEdit.