To Edinburgh yesterday for the admission of new Fellows of the Royal Society of Edinburgh.

The grey city was looking at its best in the spring sun: people were sitting on the grass in Princes Street Gardens, or crowding round the Scott monument.

Unlike its similarly-named counterpart in London, the RSE is not just a scientific society; among the new Fellows was a writer I had heard of (Ian Rankin) and a businesswoman I hadn’t, as well as two mathematicians from my corridor in St Andrews (Rosemary Bailey and Ineke De Moortel); two other new Fellows from St Andrews, theologian (and colleague of mine in the far past at Merton College) Tom Wright, and poet Don Paterson, were unable to be present. One of the three corresponding fellows admitted was a Chinese landscape ecologist. But certainly most of the new fellows worked in medicine, biology or chemistry.

At the reception afterwards, I happened to mention that I had spent four and a half years as an undergraduate and teaching fellow in Brisbane, whereupon I was taken to see the portrait of Sir Thomas Brisbane (from Brisbane House in Largs), president of the RSE for 28 years, and previously governor of New South Wales. According to the RSE’s description, he “explored the north of Australia” and discovered the Brisbane River, after whom the city was named. But I suspect that he actually discovered it in the way that governors usually discover things, by sending somebody out to look (in this case John Oxley).

]]>

Actress Elaine Cassidy likes knitting. She says,

I love maths, and knitting is just maths, really.

Now mathematicians have thought about weaving, but not (as far as I know) about knitting. Are we missing a trick there?

]]>

This morning, I nearly found my own head on a pole on this bridge.

With Celia Glass and Robert Schumacher, I am looking at acyclic orientations of graphs. There is a simple formula for the number of acyclic orientations of the graph obtained from a complete graph on *n* vertices by deleting *r* pairwise disjoint edges:

.

On first inspection, this is a straightforward inclusion-exclusion formula, and should have a straightforward proof. Maybe it does, but I was unable to find one.

Any acyclic orientation of a graph *G* arises from a total order of the vertices, or in other words (if the vertex set is {1,…,*n*}) a permutation of this set. Call two permutations *equivalent* if they give the same acyclic orientation of *G*. Our job is to count equivalence classes of this relation.

Now inclusion-exclusion works fine for subsets but not so well for equivalence relations. I spent a bit of time trying to get Möbius inversion on the partition lattice to work, without success. Eventually I found a proof. First, you figure out that the size of the equivalence class containing a permutation is 2^{k}, where *k* is the number of “missing edges” mapped to consecutive positions by the permutation. Then we proceed as follows. Write down a formula for the number of permutations in which some given “missing edges” are mapped to consecutive positions. Then use inclusion-exclusion to count the permutations for which these and no more are mapped to consecutive positions. Divide by 2^{k} to get the number of equivalence classes into which these permutations fall. Sum over all subsets of the missing edges to get the total number of equivalence classes. Then a chain of formal manipulations:

- reverse the order of summation;
- gather up terms where the set of missing edges has a given size;
- notice that the Binomial Theorem can be applied;
- now gather up terms where the set in the outer summation has a given size.

The final line gives the required formula.

This description may not be clear enough for you to reconstruct. But am I missing something simple?

]]>

**26th British Combinatorial Conference
University of Strathclyde
3–7 July 2017
**

Make a note of the dates! Further details will follow. In the meantime, bookmark the website:

http://combinatorics.cis.strath.ac.uk/bcc2017/

]]>

The 25th **British Combinatorial Conference** takes place from 6 to 10 July 2015, at the University of Warwick.

**There is still time to register!** Don’t forget, do it now. Here is the link to the registration page, and here is the conference website.

The speakers and titles are:

- Manuel Bodirsky (Dresden): “Ramsey classes: examples and constructions”
- David Conlon (Oxford): “Recent developments in graph Ramsey theory”
- Stefanie Gerke (Royal Holloway): “Controllability and matchings in random bipartite graphs”
- Gil Kalai (Jerusalem): “Some old and new problems in combinatorial geometry I: Around Borsuk’s problem”
- Tomasz Łuczak (Poznan): “Randomly generated groups”
- Gary McGuire (Dublin): “Curves over finite fields and linear recurring sequences”
- Sergey Norin (Montréal): “New tools and results in graph minor structure theory”
- Nik Ruškuc (St Andrews): “Well quasi-order in combinatorics: embeddings and homomorphisms”
- Chaoping Xing (Nanyang): “Constructions of block codes from algebraic curves over finite fields”

I hope to see many of you there!

]]>

As always, this year’s colloquia provided very good instruction and entertainment. And, as always in my descriptions, I will not cover the talks uniformly!

At the QM day, the first talk was by Stéphan Thomassé, who was very complimentary about the state of combinatorics in the UK, and indeed quoted recent results by Keevash, Kuhn and Osthus in his talk. It was a lovely talk. He started with the context, which is nowhere-zero flows in graphs. A *k*-flow, for short, is a flow on an orientation of the graph which takes values from the set {1,…,*k*−1} such that the flow into any vertex is equal to the flow out. (A result of Tutte, which has always smacked of magic to me, is that we can take the values of the flow to be non-zero elements of any abelian group of order *k*, without changing the class of graphs having such a flow.) Tutte made two conjectures, both still open:

- a 2-edge-connected graph has a 5-flow;
- a 4-edge-connected graph has a 3-flow.

Stéphan remarked that a sensible way to weaken the conjectures is to replace 5 (in the first) and 4 (in the second) by larger numbers. The first was achieved by Kirkpatrick (in his thesis, which was ignored) and five years later by Jaeger, giving the value 8. The proof is so nice that I will give it in outline below. Seymour reduced 8 to 6, and there matters stand.

For the 8-flow theorem, we use the abelian group formulation, and show first that we can assume the graph is 3-edge-connected (if there is an edge cut of size 2, the two edges carry equal flow in opposite directions, so we may cut them and join the ends to get smaller graphs). Now double the edges of the graph to obtain a 6-edge-connected graph. A theorem of Nash-Williams shows that this graph has three edge-disjoint spanning trees; their complements can be enlarged to Eulerian subgraphs (see below). Now define a flow in (Z_{2})^{3} by putting a flow (*a,b,c*) on an edge *e*, where *a* is 1 if the edge lies in the first Eulerian subgraph and 0 otherwise, and similarly for the other two entries. It is easily checked that this is a flow.

[Why can the complement of a spanning tree be enlarged to an Eulerian subgraph? This depends on the fact that 2*m* vertices of a tree can be “matched” by *m* disjoint paths – an easy exercise, try it! Take the set of vertices of odd degree in the complement of the tree, and add the joining paths to this complement.]

For the other conjecture, Carsten Thomassen proved in 2012 that an 8-edge-connected graph has a 3-flow. According to Stéphan, Thomassen had been working on this problem for 40 years! Along the way, he came up with another nice conjecture. Given a tree *T*, there is a number *c* (depending on *T*) such that a *c*-edge-connected graph, whose number of edges is divisible by the number in *T*, can be decomposed into copies of *T*. He proved this for stars, but paths seemed much harder.

Stéphan and his co-authors have shown that 24-connected is enough, provided we add a minimum degree condition: given a path *P*, there is a number *d* (depending on *P*) such that a 24-edge-connected graph with minimal degree at least *d*, whose number of edges is divisible by the number in *P*, can be decomposed into copies of *P*.

The next talk was by Alex Scott; a very nice talk, but very similar to his Glasgow talk which I already described.

After lunch we had a nice survey by Olof Sisask (formerly at Queen Mary) of some new results in additive combinatorics, depending on a “continuity” condition for convolutions; then a talk by Anita Liebenau on the minimal degree of *minimal Ramsey graphs* (graphs in which, say, any *r* edge-colouring contains a monochromatic copy of the complete graph of order *k*, but when a vertex or edge is deleted this is no longer true).

Then we had the highlight of the day, for me at least: another proof of the existence of *t*-designs, about which Peter Keevash talked last year. This one was by Ron Peled, and the approach was completely different. To say that a certain kind of design exists is equivalent to saying that, if you choose the blocks independently at random with some fixed probability, then there is a non-zero probability (maybe quite small!) of obtaining the design. Now suppose we could estimate this probability, with techniques from probability theory, and show that it was non-zero: wouldn’t that be nice? Well, that is what Ron and his co-authors have done.

It is interesting to contrast their result with Keevash’s. He was able to show that *t*-(*v,k*,λ) designs exist for all sufficiently large *v* satisfying the usual necessary conditions (for fixed *t,k*,λ). In particular, Steiner systems (λ = 1) exist whenever the appropriate conditions hold and *v* is large enough. He wouldn’t be drawn on how large is large enough, though he claimed it is not astronomically large.

By contrast, Peled and his co-authors require λ to be large, so no Steiner systems (though they give a formula for how large it has to be, containing an undetermined constant). But they don’t require *t* and *k* to be fixed; for example, *k* might be as large as the square root of *v*. Moreover, their estimate for the probability translates into an estimate of how many such designs there are: it is a function of the parameters times (1+o(1)), where the o(1) is explicit.

The method also constructs other things such as orthogonal arrays and uniformly *t*-transitive sets of permutations.

I will briefly outline how it works. Let *M* be an integer matrix with rows and columns indexed by sets *B* and *A* respectively (and think of *B* as typically much larger than *A*). Let *x* be the vector obtained by averaging the rows of *M*. We look for a relatively small set of rows so that the average of the rows in this set is also *x*. The theorem guarantees that this can be done, under suitable conditions:

- the entries of
*M*are bounded by a constant; - the space orthogonal to the row space has a basis of integer vectors with small
*l*^{1}-norm (bounded by another constant: this is related to the LDPC, or low-density parity check, condition of coding theory, in ways which are not yet understood); - the matrix has a group of symmetries acting transitively on the rows.

This is proved by analysing a certain random walk on the lattice spanned by the rows of the matrix, showing that the three conditions imply a “local central limit theorem” for the limiting distribution of the random walk.

For the design application, we take *M* to be the incidence matrix of *k*-sets against *t*-sets in a set of size *v*. The row sum average vector for the whole set is a constant vector; and a subset of rows is a *t*-design if and only if the average over this subset if also a constant vector (necessarily the same).

The final talk, by Greg Sorkin, was on “VCG auctions”; I confess that I found it more interesting than I had expected.

The second day was as usual at LSE. The weather had changed for the worse but enthusiasm was undampened. But I am getting tired, so my write-ups may be shorter.

Apart from the British Combinatorial Committee, who had a meeting before the start ot the talks, we kicked off with two talks on the connection between combinatorics and statistical mechanics.

Will Perkins started by giving a “birthday inequality”, an inequality for the probability that no two of *n* people share a birthday. The bound came from noting that the probability that the next person has a different birthday from all those considered so far is minimised when all these have different birthdays: he called this the “repulsion inequality”. Then he considered the hard spheres problem. The goal is to find the probability that *n* spheres of fixed radius *r*/2 placed in the *d*-dimensional torus don’t overlap. An upper bound from this would follow from a similar “repulsion inequality”, which would say that the proportion of the torus covered by the spheres of radius *r* with the same centres is maximised if the smaller spheres are pairwise disjoint. A reasonable-looking statement, but much harder: he can prove it in certain ranges, but it is not true in general! Indeed, it fails in 24 dimensions, and the reason is the Leech lattice, which gives a very dense sphere-packing.

Next, Lex Schrijver (who was introduced with his name pronounced correctly by Jan van den Heuvel) talked about some variations of a result he proved a few years ago with Freedman and Lovász. Given any real symmetric *n*×*n* matrix *A*, one can define a partition function on finite graphs. He gave some conditions that such a partition function must satisfy; the theorem asserts that these conditions characterise partition functions. (The conditions are slightly different over the real and the complex numbers, a feature which recurred in his talk.) He went on to analogous situations: first the analogue for edge colourings (thinking of the classical case as vertex colourings); then link invariants. Connections with graph limits and with Vassiliev knot invariants.

After lunch, we heard from Frank Vallentin about packing superspheres in 3-space. He began with a survey of packing spheres. A supersphere (a notion invented and popularised by Piet Hein) is a ball in the *l ^{p}* metric, for

Then Christina Goldschmidt told us about scaling limits of Galton–Watson trees. The simplest model gives “random compact **R**-trees”, which are almost surely 3-branching everywhere, but more complicated models allow higher-order branching. The simple model is related to Brownian motion in a way she explained, and the more complicated models to Lévy processes. But I am out of my depth here!

Alexey Pokrovskiy talked about two very nice results about tournaments:

- for some constant
*c*, a*ck*^{2}-connected tournament has*k*edge-disjoint Hamiltonian cycles; - for some constant
*c*, a*ck*-connected tournament is*k*-linked. (This condition means that, given two disjoint*k*-tuples of distinct vertices, there are*k*edge-disjoint paths such that the*i*th path goes from the*i*th point of the first tuple to the*i*th point of the second.)

My keenest followers may recall that I mentioned in a recent post the fact that (strongly) connected tournaments have Hamiltonian cycles (in connection with a problem about idempotent generators for the transformation semigroup). I feel almost certain that some of Alexey’s results and ideas will be useful in work on this problem. As soon as I can devise a nice test question, I have promised to dangle it in front of him!

The day ended with the Norman Biggs lecture, given this year by Tim Gowers. He talked about a result he and Emanuele Viola have proved. Let *G* be the group SL(2,*q*). Let *A* and *B* be “large” subsets of *G*×*G*. If (*a*_{1},*a*_{2}) and (*b*_{1},*b*_{2}) are random elements of *A* and *B* respectively, then the “interleaved product” *a*_{1}*b*_{1}*a*_{2}*b*_{2} is approximately uniformly distributed over *G*. He took us through a series of reductions of the problem, which came in the end to counting the solutions of a single polynomial equation in five variables over the field of order *q*. The polynomial turns out, after some work, to be almost always absolutely irreducible, so the Lang–Weil theorem gives an estimate for the number of solutions, which does the job.

The theorem has an application to communication complexity in computer science, which he explained. He speculated that a similar result will hold for all finite groups of Lie type and bounded rank. The key property of SL(2,*q*) needed for the proof is that all its non-trivial conjugacy classes have approximately the same size.

So in most respects, it was business as usual; the organisers are very good at choosing speakers who have something to say and know how to say it. However, there was a new feature of this year’s colloquia: the inside front cover of the program book carried the message “When tweeting about the colloquia, please use the hashtag #CC2015″. Apparently this feature was well used (but how would I know?). If you did use this feature, I welcome your feedback!

]]>

- Rosemary A. Bailey and Peter J. Cameron,
*The eigenvalues of averaged symmetric matrices*. [Our first joint paper, rejected and never published. The paper sets the problem of finding the eigenspaces of the*n*×*n*matrix obtained by averaging the images of a real symmetric matrix under a transitive group acting on {1,…,*n*}, and why statisticians want to do this.] - Peter J. Cameron,
*Yet another generalisation of Fisher’s inequality*. [Proved in response to a question from one of my first PhD students at Queen Mary, Mark Whelan. An*m*-part*t*-design on a set*X*of points is a collection of functions from*X*to {1,…,*m*} such that the inverse images of each point in all functions have the same size, and the number of functions taking prescribed values on*t*distinct points depend only on the prescribed values and not on the chosen points. For*m*= 2, this is equivalent to a complementary pair of*t*-designs. I show that an*m*-part 2-design has at least (*m*−1)*v*functions, where*v*is the number of points.] - Peter J. Cameron,
*Locally projective spaces and the transitivity of parallelism*. Submitted to a journal, possibly*Math. Z.*, in 1976; rejected, then abandoned. [A study of geometries with the property that, if three hyperplanes have the property that two of the three intersections are codimension-2 subspaces, then the third is empty or codimension-2.] - Peter J. Cameron,
*Paired suborbits*. [Sims showed that paired subconstituents in a finite transitive permutation group must have a common no-trivial quotient; in my thesis I used his method to prove other results of this sort, and this note extends the ideas further.] - Peter J. Cameron,
*Automorphism groups of parallelisms*. [These are parallelisms of the design of all*t*-subsets of an*n*-set, where*t*divides*n*. Some of this went into my book on parallelisms.]

There are also some handwritten papers:

- Three things I wrote as an undergraduate: one takes the set of unordered
*n*-tuples over some mathematical structure (ordered set, field, metric space) and explores giving it a related structure; one considers the set of 2×2 real matrices commuting with a given matrix and attempts to mimic the structure of the complex numbers; the third takes the set {1,2,…,*mn*}, partitions it into*m*sets of size*n*, takes the product of the numbers in each set and adds these products, and asks: what is the minimum value? - Three different constructions starting from a Hadamard matrix of order
*n*and producing symmetric Hadamard matrices of order*n*^{2}with constant diagonal and constant row and column sums. - A generalisation of the outer automorphism of M
_{12}: take a pair of 3-designs, in each of which the complement of a block is a block, and there are correspondences between pairs of points in one design and complementary pairs of blocks in the other. - Subsets of projective planes having a constant size intersection with every secant.
- Yet more generalisations of Fisher’s inequality, some of which got into my BCC papers in 1973 and 1977. One document considers designs in vector spaces; the other, designs in association schemes.
- Some characterisations of 2-arc transitive graphs (if they have enough quadrangles, they are known: this grew out of my thesis, but unlike there, it did not assume vertex-primitivity.
- Relations between paired subconstituents of transitive permutation groups. (Sims showed that they must have a common no-trivial quotient; in my thesis I used his method to prove other results of this sort, and this note extends the ideas further.)
- A paper with Peter Neumann characterising the permutation groups which are 3/2-transitive on the
*k*-subsets of the domain, for some*k*: the groups S_{7}and A_{7}are the only examples.

The most interesting to me are the undergraduate pieces, since it might seem as if a lot of the mathematics I did grew from considering the structure of the set of *k*-subsets of an *n*-set.

]]>

First, the background.

Let *A* be an *n*×*n* matrix. The *permanent* of *A* is calculated in a similar way to the determinant (as a sum over permutations), but without the alternating signs. In other words, for each matching from rows to columns, take the *n* elements of the matrix picked out by the matching (those whose row and column numbers are matched), and multiply them; then sum the results over all *n*! matchings.

Unlike the determinant, which can be computed with *O*(*n*^{3}) arithmetic operations, the permanent (despite its simpler definition) is hard to compute, indeed it is #P-hard. However, it is important, since it counts various things of interest. In the special case where the entries of *A* are all 0 or 1, the term corresponding to a matching is 1 if all the corresponding matrix entries are 1, and is 0 otherwise. So, regarding such a matrix as describing a bipartite graph (whose vertices are the rows and columns, with an edge wherever there is a 1 in the matrix), the permanent of *A* is the number of matchings (sets of *n* disjoint edges) in the graph.

A square matrix is said to be *doubly stochastic* if its entries are all non-negative and its row and column sums are all 1. The reason for the name is that a non-negative matrix with row sums 1 can be regarded as the transition matrix of a Markov chain with *n* states; the column sum condition essentially means that we can run the Markov chain backwards.

A *permutation matrix* (a (0,1)-matrix with a single 1 in each row or column) is doubly stochastic (though the Markov chain is not stochastic at all, but completely deterministic!); its permanent is 1. The matrix with every entry 1/*n* is also doubly stochastic (and represents a Markov chain which forgets where it came from after 1 step), and has permanent *n*!/*n ^{n}*.

In 1926, van der Waerden conjectured that the permanent of a doubly stochastic *n*×*n* matrix is at least *n*!/*n ^{n}*, with equality holding only for the constant matrix. This conjecture was proved in 1979 or 1980 independently by Egorychev and Falikman. However, it is less remembered that, in 1979, Bang and Friedland independently gave the slightly weaker lower bound 1/e

I had been thinking about Latin squares, Steiner triple systems, and what I then called “edge-parallelisms” (or 1-factorisations of complete graphs). These three have many things in common, including a Markov chain (proved for Latin squares but still conjectural in the other cases) for choosing one uniformly at random. Subsequently I invented the notion of a “generalised *t*-design” and remarked that these objects are the generalised 2-designs with block size 3 and λ = 1.

There were at the time upper and lower bounds for the numbers of such objects on *n* points. It was known by many people at the time that a proof of the van der Waerden conjecture would improve the lower bounds so that they would be much closer to the upper bounds. Indeed, the logarithms of the upper and lower bounds would be asymptotically equal, and would be *cn*^{2} log *n*, where *c* is 1 for Latin squares, 1/6 for Steiner triple systems, and 1/2 for edge-parallelisms.

Now it was clear to me in 1979 that the result of Bang and Friedland was good enough to have the same effect. Moreover, I saw that the upper bound for the number of objects having some non-trivial symmetry would be asymptotically smaller than the lower bound for the total number of objects. Hence:

**Theorem** Almost all Latin squares, Steiner triple systems, and edge-parallelisms have trivial automorphism group (in the sense that the proportion of such objects of admissible order having a non-trivial automorphism tends to 0 as *n* tends to infinity).

I thought this was an interesting result, and typed it up. But before I could submit it, or even properly check it, Laci Babai published the result for Steiner triple systems, and so I abandoned my manuscript.

Laci knows of its existence, and over the years has urged me to make it public. I was previously unable to do so since I couldn’t put my hands on it. But now I have found it, so I am happy to put it on the web (for the sake of history, not with any intention of claiming credit retrospectively).

Of course I don’t remember exactly when I wrote this. But the fact that I cite Bang and Friedland but not Egorychev and Falikman indicates that it was in the very small time gap between the appearance of these two pairs of papers, so in 1979 or 1980.

I have typed the manuscript “as is”. Apart from correcting one confusing misprint, and adopting LaTeX conventions such as numbering theorems and lemmas, it is an exact copy of the original (modulo the inevitable typos). In particular, I have not checked the mathematics; while typing it out I thought that the argument in the case of edge-parallelisms is not quite right. I think it can probably be put right.

]]>

I discussed this design before here.

Donald was so pleased with his design that he stuck playing cards to a thick card and mounted them (with commentary) in a frame, so they could be displayed in the foyer of a mathematics department.

This never happened. But the mathematics building at Queen Mary is being refurbished, and as part of clearing Donald’s papers from the emeritus staff offices, I found the original (which, you will note, is slightly different from the version he published).

Below is a photograph of the design. I have been unable to make the commentary legible, so I have put it in the text around the picture.

A “BALANCED DESIGN”

- Each value A, 2, 3, …, 10, J, Q, K appears once in each row.
- Each suit appears in each column.
- Each pair of the values A, 2, 3, …, 10, J, Q, K occurs together in just one column.
- Each row has 3 cards of each of 3 suits, and 4 cards of the other suit.

Design obtained by D. A. PREECE (Rothamsted Statistics Department) and published in 1982.

This arrangement of cards is a possible design for an experiment on fruit trees.

Each card corresponds to a “plot” of trees, the plots being in 4 rows and 13 columns.

The values A, 2, 3, …, 10, J, Q, K represent 13 treatments from a first set, and the suits represent 4 treatments from a second set.

]]>

So it isn’t so surprising that a book by Ron Aharoni, newly translated into English, doesn’t come to a definite conclusion one way or the other. The best we can do in a book entitled *Mathematics, Poetry and Beauty* is to give many examples of beautiful mathematics and beautiful poetry and discuss what the similarities and differences are.

Ron Aharoni is a mathematician whose field is combinatorics. He has collaboration distance 2 from me (we are both co-authors of Paul Erdős). I hadn’t heard from him for a while. In the book he explains that he made a deliberate move from university to elementary school.

Many, though not all, of his examples of poetry are taken from Israeli poets. I don’t know whether Hebrew is a particularly good language for poetry, but some of these poets pack many layers of meaning into a few words. But other poets appear, including Johann Wolfgang von Goethe, John Donne, Emily Dickinson, Federico Garcia Lorca, Constantine Cafavy, William Carlos Williams, and Matsuo Basho.

Here is an example of the argument. Displacement is a mechanism which, according to Freud, occurs in almost every area of human thought. By focussing on a subsidiary idea, the main message slips through almost unnoticed, although it may be too painful to face directly. Aharoni suggests that this is a technique used by poets for diving inside themselves, and for mathematicians stuck on a problem who look at a seemingly irrelevant detail in the hope of a breakthrough. He gives several examples, both poetic and mathematical.

One of his telling comparisons, expanded over four chapters, is that both a poem and a mathematical proof constitute a game of ping-pong between the abstract and the concrete. A poem can have several such switches in a few lines, as in this example “Written in pencil in the sealed railway-car” by Dan Pagis (translated by Stephen Mitchell):

Here in this carload

I am Eve

with my son Abel

if you see my older boy

Cain son of Adam

tell him that I …

In mathematics, both finding a proof and (if you are kind to your readers) presenting it involve frequent shifts of focus between logical argument and examples. But Aharoni’s conclusion is

… the heart of the poem is given to the concrete, and it is in this direction that the poem goes. This is the diametric opposite of the ping-pong of mathematics, in which the last shot is always towards the abstract.

It is also true that, as he remarks, in many published proofs (most notoriously, those of Gauss, the “fox who effaces his tracks in the sand with his tail”, according to Abel), all traces of the concrete are covered and only the shots to the abstract remain.

This points to another important difference. A finished poem can convey its beauty to any open-minded reader; knowledge of the poet’s biography often gets in the way. But the beauty in mathematics lies in the experience of the discovery of the proof; this can be reproduced to some extent in a reader who follows the argument carefully, but does not reside in the published proof, and still less in the statement of the theorem (in most cases).

And on the same theme, Aharoni invites us to watch the mathematician and the poet at work. The striking secret he reveals is that the mathematician spends most of his time staring into space.

Both poets (as many thinkers have observed) and mathematicians are concerned, not with ever more florid invention, but with the truth. It seems like a different kind of truth. Poets remind us of things we already know. But almost every mathematician, no matter which side they take on the “discovered or invented” question, are in their ordinary work Platonists, and act as if that their mental constructions are “out there”. However, the means they have for diving inside themselves, described in detail by Hadamard in his book, and divided into four stages (preparation, incubation, illumination, and verification) go well beyond the subjective ways that poets operate.

There is much more thought-provoking material in the book, many more dimensions on which mathematics and poetry can be compared. The chapter titles give some indication, including “The miracle of order”, “The power of the oblique”, “Reality or imagination”, “Unexpected combinations”, “Symmetry”, “Content and husk”, and “Change”.

]]>