Abstract. A famous theorem of Szemerédi asserts that any set of integers of positive upper density will contain arbitrarily long arithmetic progressions. In its full generality, we know of four types of arguments that can prove this theorem: the original combinatorial (and graph-theoretical) approach of Szemerédi, the ergodic theory approach of Furstenberg, the Fourier-analytic approach of Gowers, and the hypergraph approach of Nagle-Rödl-Schacht-Skokan and Gowers. In this lecture series we introduce the first, second and fourth approaches, though we will not delve into the full details of any of them. One of the themes of these lectures is the strong similarity of ideas between these approaches, despite the fact that they initially seem rather different.
IntroductionThese lecture notes will be centred upon the following fundamental theorem of Szemerédi: This theorem is rather striking, because it assumes almost nothing on the given set A -other than that it is large -and concludes that A is necessarily structured in the sense that it contains arithmetic progressions of any given length k. This is a property special to arithmetic progressions (and a few other related patterns). Consider for instance the question asking whether a set A of positive density must contain a triplet of the form {x, y, x + y}. (Compare with the triplet {x, y, x+y 2 }, which is an arithmetic progression of length three.) It is then clear that the odd numbers, which are certainly a set of positive upper density, do not contain such triples (see however Theorem 6.1 below). Or for another example, consider whether a set of positive upper density must contain a pair {x, x + 2}. The multiples of 3 provide an immediate counterexample. (This is basically why the methods from [25] can leverage Szemerédi's theorem to show that the primes contain arbitrarily long arithmetic progressions, but are currently unable to make any progress whatsoever on the twin prime conjecture.) But the arithmetic progressions seem to be substantially more "indestructable" than these other types of patterns, in that they seem to occur in any large set A no matter how one tries to rearrange A to eliminate all the progressions.We have contrasted Szemerédi's theorem with some negative results where the selected pattern need not occur. Now let us give the opposite contrast, in which it becomes very 1991 Mathematics Subject Classification. 11N13, 11B25, 374A5. The author is supported by a grant from the Packard Foundation. Note that if we could just set a = b in these parallelograms then we could find infinitely progressions of length three. Alas, things are not so easy, and while progressions are certainly intimately related to parallelograms (and more generally to higher-dimensional parallelopipeds, for which an analogue of Proposition 1.2 can be easily located), the existence of the latter does not instantly imply the existence of the former without substantial additional effort. For example, one can easily modify Proposition 1.2 to locate, for any k ≥ 1, infinitely ma...