Arxiv:2011.14515V2 [Math.CO] 17 Dec 2020

arXiv:2011.14515v2 [math.CO] 17 Dec 2020 oin flreesta a eipsdon imposed be can that largeness of notions 2020 Let .Tplgcldnmc n icws ydtct 29 syndeticity piecewise and dynamics References Topological 8. .Dsodn esadgnrlztoso qaerenmes12 SL numbers in squarefree sets of Discordant generalizations and sets 5. times Discordant visiting from sets 4. Discordant 3. Preliminaries 2. Introduction 1. .Lresbesof subsets Large anti-recurrence and 7. Recurrence 6. ICRATST N ROI ASYTHEORY RAMSEY ERGODIC AND SETS DISCORDANT A ahmtc ujc Classification. Subject Mathematics apeo u results. interestin our historically of these sample o about results unify new and obtain generalize and classical We Ra the Szemerédi’s theorem. of between and questions complexity theorem many in in difference involved are the kind manifest this of Sets groups. tv pe est,wihw call we which density, upper itive Abstract. est;peeiesyndetic. piecewise density; ia asyter,egdcter,nme hoy n top i and mathematics, theory, dynamics. of number theory, bolic corners ergodic various theory, from Ramsey draw sical we way, the Along easbe fteset the of subset a be • • • Keywords: IAYBRESN AEHRN N UHLRAGHAVAN RUSHIL AND HURYN, JAKE BERGELSON, VITALY rdcn ayeape fdsodn esin sets discordant squarefr of the examples of many generalization producing wide-ranging a measure. introduce discord positive We between having analogy sets syste intimate dense dynamical an nowhere exhibit in we recurrence setting to this sets discordant connect We eso ht o n onal bla group abelian countable in any Φ o for densities that, compute show to We method unified sets. a discordant develop We group. berg d Φ ( A = ) G eepoetepoete fnnpeeiesnei eswi sets syndetic non-piecewise of properties the explore We n any and , c asyter;egdcter;tplgcldnmc;uppe dynamics; topological theory; ergodic theory; Ramsey Here . { 0 , 1 d } 2 c Φ ( G ∈ Z eoe est ln Φ. along density denotes ) 1. N (0 , fpstv nees ebgnb enn some defining by begin We integers. positive of ) hr xssadsodn set discordant a exists there 1), Introduction Contents discordant 51,37A44. 05D10, 1 A . ncutbeaeal (semi-) amenable countable in , G Z n ønrsequence Følner any , , es eei small a is Here sets. g Z d a e Waerden’s der van n h Heisen- the and , lgcladsym- and ological syter and theory msey dconstructions ld cuigclas- ncluding A n esand sets ant enumbers, ee s n in and ms, ⊆ G hpos- th these f with r 32 26 10 22 19 5 1 2 VITALY BERGELSON, JAKE HURYN, AND RUSHIL RAGHAVAN

• A is an IP set if it contains a ﬁnite sums set, i.e., a set of the form

FS((an)n∈N)= ai : I ⊆ N is finite and nonempty ( ) Xi∈I for some infinite sequence (an)n∈N of positive integers. • A is thick if it contains arbitrarily long intervals: for each ℓ ∈ N, there exists a ∈ N such that {a,a +1,...,a + ℓ − 1}⊆ A. • A is syndetic if its gaps are bounded: this means that, if we write A = {a1 < a2 < ···}, the successive difference ai+1 − ai is bounded. Equivalently, A r is syndetic if there exist t1,...,tr ∈ N such that N = i=1 A − ti, where A − t = {x ∈ N: x + t ∈ A}. • A is piecewise syndetic if it can be written as the intersectionS of a thick set with a syndetic set. Equivalently, A is piecewise syndetic if there exist r t1,...,tr ∈ N for which i=1 A − ti is thick. • The upper density of A is S |A ∩{1,...,n}| d(A) = lim sup . n→∞ n When the limit in question exists, this quantity is called the density of A, denoted d(A). A set should be thought of as large in terms of (upper) density if this quantity is positive. Although we have called each of the above properties a “notion of largeness”, they are not on equal footing. For all these definitions, N is large, and if A has a large subset, then A is itself large. However, only some of these notions of largeness are partition regular, meaning that if A is large then, for any finite partition A = r i=1 Ci, at least one Ci is large. This third property underlies much of classical Ramsey theory. The partition regularity of IP sets is a famous theorem of Hindman S[28], and the partition regularity of piecewise syndeticity is a classical fact, often rediscovered (see, e.g., [29, Lemma 2.4] and [22, Theorem 1.24]).1 The partition regularity of positive upper density is an easy exercise for the reader, as is checking that thickness and syndeticity fail to be partition regular. One of the earliest partition results of Ramsey theory is van der Waerden’s r theorem, which says that given any finite partition N = i=1 Ci, at least one Ci is AP-rich, i.e., contains arithmetic progressions of any finite cardinality [35]. A useful equivalent form of this theorem states that the familyS of AP-rich sets is partition regular. Van der Waerden’s theorem has a corresponding density version, Szemerédi’s theorem [34], one of whose many equivalent formulations guarantees that any set having positive upper density is AP-rich. In seeking a density version of Hindman’s theorem, Erd˝os asked whether d(A) > 0 is enough to imply that A contains shift of a finite sums set. Of course we cannot ask that A itself be contain a finite sums set; consider, e.g., the set of odd numbers. This question was answered negatively in unpublished work of E. Straus, who showed that, for any ε > 0, one can find A ⊆ N with d(A) ≥ 1 − ε such that A − n is not an IP set for any n ∈ Z. In [19, p. 105], Erd˝os writes the following:

I ﬁrst hoped that Hindman’s theorem on all subsums P εiai can be gen- eralised for sets of positive density, but the following example of E. Straus

1 r A related fact, that if N = =1 Ci then at least one Ci is piecewise syndetic, appears to be Si ﬁrst observed by Brown [15, Lemma 1]. DISCORDANT SETS ANDERGODIC RAMSEYTHEORY 3

seems to give a counterexample to all such attempts: Let p1 < p2 < · · · be a set of primes tending to inﬁnity fast, and let P εi < ∞. Consider the set of integers a1 < a2 < · · · so that ai 6≡ α (mod pi) where |α| ≤ εipi.

The density of this sequence is Qi(1 − εi) > 0, and it is not difficult to see that this sequence furnishes the required counterexample. See [4, Theorem 2.20] for more discussion and a detailed presentation of Straus’s construction. One may also ask if there is a density version of the fact that piecewise syndeticity is partition regular. Along these lines one might hope that any set of positive density is piecewise syndetic. Here Straus’s construction again provides a counterexample, since if no shift of A is IP, then A is necessarily non-piecewise syndetic. To see why, recall that if A is piecewise syndetic, then a union of finitely many shifts of A is thick. It is not too hard to see that any thick set is IP, so by Hindman’s theorem on the partition regularity of IP sets, one of these shifts must also be IP. It is these sets we are interested in—those which are non-piecewise syndetic but have positive (upper) density. We will call such sets discordant. Another classical example of a discordant set is the set of squarefree numbers. However, this set has shifts which are IP [12, Theorem 2.8], and so has a different nature than Straus’s counterexample. More constructions of discordant sets can be found in [13] along with applications to Ramsey theory. A fruitful analogy can be drawn between discordant sets and nowhere dense sets of positive measure, both being small in a “geometric” sense but large in an “analytic” sense. This analogy will be developed and more closely investigated in the following sections. Just as every piecewise syndetic set contains a shift of an IP set, it follows from van der Waerden’s theorem that every piecewise syndetic set is also AP-rich. However, given that every piecewise syndetic set is AP-rich, the partition regularity r of piecewise syndeticity tells us that, for any finite partition N = i=1 Ci, there is some i for which Ci is AP-rich. Thus van der Waerden’s theorem is equivalent, in this informal sense, to the fact that every piecewise syndetic set isS AP-rich. In light of this, if one could prove that every discordant set is AP-rich, then Sze- merédi’s theorem would follow at once upon invoking van der Waerden’s theorem. Thus the existence of discordant sets can be interpreted as a sign that Szemerédi’s infamously difficult theorem is a deeper thing than van der Waerden’s.2 The strik- ing abundance of contexts in which discordant sets arise, some of which will be provided in the ensuing sections, can be taken as evidence for the versatility of Szemerédi’s theorem and strengthens our interest in discordant sets. The distinction between van der Waerden’s theorem and Szemerédi’s theorem can also be profitably analyzed from the point of view of dynamics and ergodic theory. Indeed, as initiated by Furstenberg’s ergodic-theoretic proof of Szemerédi’s theorem [21], one can elucidate the properties of a subset of N by examining the dynamical and ergodic properties of the orbit closure of its indicator function. To N be more explicit, given A ⊆ N, consider its indicator function 1A ∈{0, 1} , where we view {0, 1}N as having the product topology (making {0, 1}N homeomorphic to

2Another equivalent form of Szemerédi’s theorem says that any subset of N having positive upper Banach density is AP-rich. In contrast to density, for which one averages along the intervals {1,...,n}, the upper Banach density of a set A is defined to be the supremum of all densities of A averaged along any sequence of intervals of increasing length. Any piecewise syndetic set has positive upper Banach density, a fact which clarifies the relationship between van der Waerden’s and Szemerédi’s theorems and strengthens the perspective we take here. 4 VITALY BERGELSON, JAKE HURYN, AND RUSHIL RAGHAVAN the Cantor set). Then, letting T : {0, 1}N → {0, 1}N denote the left-shift operator (i.e., T (a1,a2,... ) = (a2,a3,... )), define the orbit closure of 1A as

n O(1A)= {T (1A): n ∈ N}.

The resulting topological dynamical system (O(1A),T ) can be made into a measure- preserving system. In this setting, the difference in difficulty between van der Waerden’s and Szemerédi’s theorems is indicated by the existence or nonexistence of nontrivial minimal subsystems of O(1A). In the presence of a nontrivial minimal subsystem, a purely topological-dynamical recurrence theorem, the topological analogue to van der Waerden’s theorem, can be used to obtain a streamlined proof of the classical van der Waerden’s theorem [23, Theorem 0.3]. On the other hand, when O(1A) does not possess nontrivial minimal subsystems, a considerably more difficult proof via measurable dynamics is necessary to prove Szemerédi’s theorem. In particular, those sets A ⊆ N for which O(1A) has a minimal subsystem are exactly the piecewise syndetic sets (Theorem 44). This dynamical and ergodic approach leads not only to a proof of these classical theorems but also to far-reaching generalizations which still elude purely combinatorial proofs, such as versions of the polynomial Szemerédi theorem (e.g. [10, Theorem 1.1] and [11, Theorem 7.3]). Finally, because the existence of discordant sets is a nontrivial byproduct of the interplay between the notions of density and piecewise syndeticity, both very basic definitions of Ramsey theory, these special sets are particularly nice vehicles for ad- vancing Ramsey theory in different settings and showing the bountiful connections between Ramsey theory and other areas of mathematics. Therefore our paper is not only about discordant sets but also is meant as an accessible (but admittedly very non-exhaustive) demonstration of the diversity of Ramsey theory.

Outline of the paper. This paper concerns itself with the study of subsets of countable semigroups, so some definitions must be introduced to allow us to work in this generality. This is done in Section 2, where we also provide the necessary prerequisites from Ramsey theory, ergodic theory, and topological dynamics. In Section 3 we give an ergodic-theoretic construction of discordant sets that can be done in many semigroups, generalizing a construction based on irrational rotation of the torus R/Z. In particular, we prove that the sets of return times to certain nowhere dense sets of positive measure form discordant sets. This section lays the foundation for the analogy between nowhere dense sets of positive measure and discordant sets, which we revisit and reexamine several times. In Section 4, we prove in two ways that the squarefree numbers are discordant. The first proof is purely algebraic while the second uses the dynamical results of the previous section. We also give several wide-ranging generalizations of this fact. Our view narrows in Section 5, where we take a brief excursion to SL2(Z). Al- though the classical notions of density fail to exist in these semigroups since SL2(Z) is not amenable, we define a natural analogue of density which allows us to construct discordant sets using results of the preceding section. In Section 6, we study a special kind of discordant set which is particularly bad for recurrence in measure-preserving systems. We establish the existence of such discordant sets with density c for each c ∈ (0, 1) for any countable abelian group and also prove the existence of discordant sets with density c for each c ∈ (0, 1) in any countable commutative cancellative semigroup. DISCORDANT SETS ANDERGODIC RAMSEYTHEORY 5

We shift focus in Section 7. First, we generalize the notion of a disjunctive {0, 1}- sequence, one in which every finite {0, 1}-word occurs, to an arbitrary countable semigroup, and show that such sequences are “topologically typical” in the sense that the family of such sequences is comeager. This allows us to state that, in cancellative semigroups, non-piecewise syndetic sets (and hence discordant sets) are topologically atypical. We then consider special types of disjunctive sequences by measuring the limiting frequency with which every finite word occurs. In contrast to how one defines a normal sequence by saying that these frequencies approach 2−n, where n is the length of the word, we define an extremely non-averageable, or ENA, sequence by declaring that these frequencies should fail to converge as badly as possible. We prove that ENA sequences are topologically typical in arbitrary countable amenable cancellative semigroups, strengthening the previous result for such semigroups. Finally, in Section 8 we give a topological-dynamical characterization of non- piecewise syndetic sets in arbitrary countable semigroups by looking at their orbit closures. We conclude the paper by giving combinatorial applications of these dynamical results.

Acknowledgements. We would like to thank Dona Strauss for helpful conver- sations. Her input was instrumental in guiding us toward the proof of Theorem 37. The latter two authors were partially supported by the Ohio State Arts and Sciences Undergraduate Research Scholarship.

2. Preliminaries Throughout this paper, G will stand for a countable semigroup. When we say that G is a left cancellative semigroup, we mean that for any g,h1,h2 ∈ G, if gh1 = gh2 then h1 = h2; G is right cancellative if instead h1g = h2g implies h1 = h2; and G is cancellative if it is both left and right cancellative. Given g ∈ G, define gA = {ga: a ∈ A} and g−1A = {x ∈ G: gx ∈ A}; the sets Ag and Ag−1 are defined analogously. Also, given B ⊆ G, define

AB = {ab: a ∈ A, b ∈ B}, A−1B = a−1B, and AB−1 = Ab−1. a[∈A b[∈B For any set X, we will write P(X) to denote the set of subsets of X. We write Y c for the complement of Y , when the superset is understood. In particular, for A ⊆ G, Ac always denotes the complement relative to G.

2.1. Thickness, syndeticity, and piecewise syndeticity. Let A ⊆ G. • A is right thick if for any finite F ⊆ G there exists g ∈ G such that Fg ⊆ A. • A is right syndetic if there exists some finite H ⊆ G such that G = H−1A. • A is right piecewise syndetic if there is a finite H ⊆ G for which H−1A is right thick. Since G is not assumed to be commutative, there are also “left” versions of these definitions obtained by replacing Fg with gF and H−1A with AH−1. We warn the reader that the left and right versions do not necessarily coincide in noncommutative 6 VITALY BERGELSON, JAKE HURYN, AND RUSHIL RAGHAVAN

(semi)groups.3 Henceforth we will deal only with the right-handed definitions and omit the word “right”. Note that, if (Fn) is an increasing sequence of finite subsets of G satisfying G = Fn, then A ⊆ G is thick if and only if Fngn ⊆ A for some sequence (gn) in G. WeS now outline the relationship between thick,S syndetic, and piecewise syndetic sets. First, the family of syndetic sets and the family of thick sets are dual, in the following sense. Lemma 1. Let A ⊆ G. Then A is syndetic if and only if Ac is not thick. Proof. Suppose A is syndetic, so G = H−1A for some finite H ⊆ G. Let g ∈ G. Since g ∈ H−1A, there exists h ∈ H with hg ∈ A. Hence Hg ∩ A 6= ∅, i.e., Hg * Ac. This proves that Ac is not thick. Inversely, suppose A is not syndetic, and let F ⊆ G be finite. Since F −1A 6= G, −1 c c find g ∈ G such that g∈ / Fn A. Then Fg ⊆ A , so A is thick. From this duality we obtain a useful equivalent description of piecewise syndetic sets, which we described in introduction for subsets of (N, +). Lemma 2. Let A ⊆ G. Then A is piecewise syndetic if and only if there exist a syndetic set S ⊆ G and a thick set T ⊆ G such that A = S ∩ T . Proof. First suppose A is piecewise syndetic, and find a finite H ⊆ G such that H−1A is thick. Let T = A ∪ H−1A, and let S = A ∪ T c, so that A = S ∩ T . Pick any x ∈ G; we will show that G = (Hx ∪{x})−1S. Indeed, let g ∈ G. If g∈ / x−1S, then xg ∈ Sc = Ac ∩ T ⊆ H−1A, so g ∈ (Hx)−1S. Now suppose S ⊆ G is syndetic and T ⊆ G is thick. Fix a finite H ⊆ G satisfying G = H−1S. Let F ⊆ G be finite and pick g ∈ G satisfying HFg ⊆ T . Then for any f ∈ F we can find h ∈ H such that fg ∈ h−1S. We also know that hfg ∈ T , so fg ∈ h−1(S ∩ T ). This proves that Fg ⊆ H−1(S ∩ T ), and hence that H−1(S ∩ T ) is thick.

2.2. Densities in semigroups. In this subsection, G is assumed to be a countable amenable cancellative semigroup. Given these assumptions, amenability is equivalent to the existence of a Følner sequence.4 A (left) Følner sequence in G is a

3Here are some examples to show that the left- and right-handed notions really are different. In F2, the free group on a and b, it is easily seen that F2a is right thick but not left thick, and left syndetic but not right syndetic. For a less noncommutative example, consider the following subset of the Heisenberg group H3(Z): 1 x y A = 0 1 z ∈ H3(Z): |y|≤ z . n 0 0 1 o It can be shown that A is left thick but not right piecewise syndetic (and thus not right thick). The group H3(Z) is “almost abelian”, having nilpotence class 2, so we see from this example that not much noncommutativity is needed to separate the right- and left-handed notions. See [8, Section 2] for many similar examples. 4 Amenability is commonly defined as follows. For each g ∈ G, define λg : G → G by λg (h)= gh. Then G is called (left) amenable if there exists a positive normalized linear functional L on the space B(G) of bounded functions G → R which satisfies L(f ◦ λg ) = L(f) for each f ∈ B(G); L is called an invariant mean for G. For countable left cancellative semigroups, the existence of Følner sequences and amenability are equivalent [1]. See [32] for a comprehensive overview of the notion of amenability. DISCORDANT SETS ANDERGODIC RAMSEYTHEORY 7 sequence Φ = (Φn) of nonempty finite subsets of G satisfying |Φ ∩ gΦ | |Φ △ gΦ | lim n n =1, or equivalently lim n n =0, n→∞ |Φn| n→∞ |Φn| for any g ∈ G. The upper density of A ⊆ G with respect to Φ is defined as

|A ∩ Φn| dΦ(A) = lim sup . n→∞ |Φn| When the corresponding limit exists this quantity is the density of A with respect to Φ, denoted dΦ(A), and we also define the lower density dΦ(A) to be the correspond- c ing lim inf. Note that dΦ(A)=1−dΦ(A ). Observe also that it follows quickly from the pigeonhole principle and the cancellativity of G that Følner sequences satisfy |Φn| → ∞, guaranteeing a degree of non-degeneracy for these definitions. A subset A ⊆ G will be called discordant if it is not piecewise syndetic and there is some Følner sequence Φ in G such that dΦ(G) > 0. We would like to also guarantee, when possible, the stronger condition dΦ(A) > 0, or better, dΦ(A) > 0. In the introduction, we defined density in (N, +) with respect to the Følner sequence ({1,...,n})n∈N. We will reserve d, leaving the Følner sequence unspecified, to refer to this density (and similarly with d and d). A set A ⊆ N for which d(A) exists is sometimes said to have natural density. Closely related are densities defined with respect to the Følner sequences ({−n,...,n}) in Z and ({−n,...,n}d) in Zd; we will also use d to refer to these without danger of ambiguity. Upper density has some useful properties, which we briefly describe here. The most primitive is that it is invariant under shifts. In other words, for any g ∈ G −1 and A ⊆ G, one has dΦ(gA)= dΦ(g A)= dΦ(A). We record the following similar result as we will use it later on. (The proof of shift-invariance can be seen in the proof of Lemma 3.)

Lemma 3. Suppose g ∈ G and A ⊆ G satisfy A ∩ gA = ∅. Then dΦ(A ∪ gA) = 2 dΦ(A).

Proof. It is clear that dΦ is subadditive. For the opposite inequality, ﬁnd n ∈ N such that |A ∩ Φn|/|Φn| > d(A) − ε/3 and |gΦn \ Φn|/|Φn| < ε/3. Then |(A ∪ gA) ∩ Φ | |A ∩ Φ | |gA ∩ gΦ | |gΦ \ Φ | n ≥ n + n − n n . |Φn| |Φn| |Φn| |Φn|

Now since |gA ∩ gΦn| = |g(A ∩ Φn)| = |A ∩ Φn| by cancellativity, we get

|(A ∪ gA) ∩ Φn| ε ε ε > dΦ(A) − + dΦ(A) − − > 2 dΦ(A) − ε. |Φn| 3 3 3 More discussion of thickness, syndeticity, piecewise syndeticity, the various notions of density, and other deﬁnitions of Ramsey theory can be found in, among other places, [8, Sections 1 and 2]. 2.3. Partition regularity and Straus’s construction. To begin this subsection we will prove two lemmas about piecewise syndeticity and partition regularity. The following lemma states that the family piecewise syndetic subsets of any countable semigroup is partition regular. A diﬀerent proof of this fact appears in [27, Lemma 7.2]. Since, as we will see, non-piecewise syndetic sets can be seen as discrete analogues of nowhere dense sets, Theorem 4 is a sort of discrete analogue to the Baire category theorem on unions of nowhere dense sets. We will not use Theorem 8 VITALY BERGELSON, JAKE HURYN, AND RUSHIL RAGHAVAN

4 later on, but include it for completeness and to support the discussion given in the introduction and the analogy we are hoping to emphasize. A different proof, using topological dynamics, will be presented in the final section (Theorem 47). Theorem 4. Suppose A ⊆ G is piecewise syndetic. Then for any finite partition r A = i=1 Ci, there exists i for which Ci is piecewise syndetic. Proof.S We will prove this for r = 2, and the claim will follow by induction. By Lemma 2, we may write A = S ∩ T for a syndetic set S ⊆ G and thick T ⊆ G. Let ′ c ′ ′ c ′ S = C2 ∩ S, so we get C1 = S ∩ T and C2 = S ∩ (S ) . Since either S is syndetic ′ c 5 or (S ) is thick by Lemma 1, one of C1 and C2 must be piecewise syndetic. The following easy but important lemma is another connection between piecewise syndeticity and partition regularity. We record it as the general phenomenon behind why Straus’s construction produces discordant sets. We will apply this lemma by starting with a family B ⊆P(G) and extending it to a G-invariant family A = {gB : B ∈B and g ∈ G}. Lemma 5. Suppose A⊆P(G) satisfies the following: (a) A contains all thick subsets of G; (b) A is partition regular; (c) if B ⊆ A ⊆ G and B ∈ A, then A ∈ A; and (d) if A ∈ A and g ∈ G then gA ∈ A. Then A contains all piecewise syndetic subsets of G. Proof. Let A ⊆ G be piecewise syndetic and find a finite set H ⊆ G such that −1 −1 −1 −1 H A is thick. Then H A = h∈H h A, so by (a) and (b), h A ∈ A for some h ∈ H. Thus h(h−1A) ∈ A, so finally A ∈ A since h(h−1A) ⊆ A. S With this lemmas, we give a simplified version Straus’s construction to obtain a set with positive lower density, no shift of which is IP.

Theorem 6. Let (an) be a sequence in N satisfying 1/an < 1. Then the following set is discordant: P A = N \ (anN + n − 1). N n[∈ Proof. We leave it as an exercise to verify that every thick subset of N is IP; thus, by Hindman’s theorem and Lemma 5, it suﬃces to show that A is not a shift of an IP set (that is, a set of the form B + t for some IP set B ⊆ N and t ∈ N). This follows immediately from the construction of A and the observation that if B ⊆ N is IP, then B ∩ kN 6= ∅ for all k ∈ N. To see that d(A) > 0, we compute that |A ∩{1,...,k}| k − (k/a + k/a + ··· + k/a ) 1 ≥ 1 2 n > 1 − k k N an nX∈ for each k

5Note that this proof does not use the deﬁnition of piecewise syndeticity, but only that the family of piecewise syndetic sets is the family of intersections of two dual families (“dual” in the sense of Lemma 1). In other words, we have shown that if A, B ⊆ P(G) are any two dual families of subsets of G, then the family {A ∩ B : A ∈ A and B ∈ B} is guaranteed to be partition regular. DISCORDANT SETS ANDERGODIC RAMSEYTHEORY 9

For conciseness, we excuse ourselves from proving that d(A) exists. The reader is encouraged to look toward the sources we have mentioned or, alternatively, Theorem 22 to see how this could be done. One may also alter the set described in Theorem 6 to obtain a discordant set without invoking Hindman’s theorem as follows. Pick (an) satisfying n/an < 1 and set A = N \ (anN + {0,...,n − 1}). One shows similarly that d(A) > 0, and a totally self-contained proof that A is not piecewise syndetic is notP hard. S Note that Theorem 6 gives us discordant sets with lower densities arbitrarily close to 1. (A discordant set cannot have upper density equal to 1, since any set of upper density 1 is thick.) By generalizing the notion of an IP set, a similar construction can be done in many diﬀerent groups. Discussion of this construction and a complete characterization of such groups is given in [18, Chapter 4] and [7]. 2.4. Preliminaries from ergodic theory and dynamics. In this section we review some basic notions and theorems of ergodic theory and dynamics. This section is utilitarian, and hence we do not venture past what we will need. For this reason we omit proofs or provide a rudimentary outline and refer the reader elsewhere. Let (X, B,µ) be a probability space. A map T : X → X is measure preserving if µ(T −1E) = µ(E) for each E ∈ B. If G is a countable semigroup, a measure- preserving action of G on (X, B,µ) is an action of G on X such that x 7→ gx is measure-preserving for each g ∈ G. In this case, we say that µ is G-invariant. We refer to (X, B,µ,T ) and (X, B,µ,G) as measure-preserving systems. Given a measure-preserving system, one has the following version of the von Neumann ergodic theorem. Theorem 7. Suppose G is a countable amenable cancellative semigroup with a Følner sequence (Φn) and let (X, B,µ,G) be a measure-preserving system. Let f ∈ L2(X), and let f ∗ be the orthogonal projection of f onto the subspace of L2(X) consisting of G-invariant functions. Then 1 lim f(gx)= f ∗(x) in L2(X). n→∞ |Φn| gX∈Φn Proof. See [36, Theorem 1.14.1] for a proof in the case G = Z. The proof for more general G is analogous (cf. [3, Theorem 4.15]). With additional constraints on the action of G, more can be said. The system (X, B,µ,G) is called ergodic if there are no nontrivial G-invariant subsets, that is, if E ∈ B satisﬁes µ(E △ g−1E) = 0 for all g ∈ G, then µ(E) ∈ {0, 1}. In this case, the orthogonal projection of f onto the subspace of G-invariant functions is

X f dµ, so Theorem 7 states that 1 R lim f(gx)= f dµ in L2(X). n→∞ |Φn| X gX∈Φn Z If X is a compact topological space on which G acts by continuous maps, the pair (X, G) is a (topological) dynamical system. If E ⊆ X and x ∈ X, we deﬁne RE(x) = {g ∈ G: gx ∈ E}; this is analogous to a set of visiting times of x to E. The system (X, G) is minimal if there is no proper nonempty closed subset of X invariant under the action of G. Equivalently, the action of G on X is minimal if, for each x ∈ G, its orbit {gx: g ∈ G} is dense. 10 VITALY BERGELSON, JAKE HURYN, AND RUSHIL RAGHAVAN

When X is a compact metric space with a Borel σ-algebra B, we say that (X, G) is uniquely ergodic if there is exactly one G-invariant Borel probability measure on X, i.e., there exists exactly one normalized measure µ on B which makes (X, B,µ,G) into a measure-preserving system. Under the condition of unique ergodicity, one has the following convenient pointwise ergodic theorem. Theorem 8. Let (X, G) be a topological dynamical system where X is a compact metric space and G is a countable amenable cancellative group. Assume µ is the unique G-invariant Borel probability measure and let (Φn) be a Følner sequence in G. Then, for each f : X → R having discontinuities on a set of measure zero and each x ∈ X, 1 lim f(gx)= f dµ. n→∞ |Φn| gX∈Φn Z Proof. For continuous f, this can be proved in complete analogy with the well- known proof of the case G = Z; see [36, Theorem 6.19]. For more general f, the statement follows by a standard approximation argument. A fact we will find useful is that a minimal action by isometries on a compact metric space is automatically uniquely ergodic. 2.5. Cantor spaces. The following construction will be very useful. Given any sequence (Xn)n∈N of finite sets having the discrete topology, their topological product X = Xn is a compact space homeomorphic to the classical Cantor set. We will use this construction frequently, and we will call such a space a Cantor space. To makeQ the notation less cluttered and easier to read, we will use boldface letters for elements of a Cantor space, e.g., x = (xn) ∈ X. (This is also consistent with our use of 1 for the indicator function.) It will be convenient to also give Cantor spaces a metric space structure. Giving each Xn the discrete metric δn, the space X can be made into a compact metric space via the product metric δ given by −n (1) δ(x, y)= 2 δn(xn,yn). N nX∈ Assuming x 6= y, let n ∈ N be the smallest integer satisfying xn = yn and xn+1 6= yn+1 (take n = 0 if x1 6= y1). Then δ is equivalent to the more common and conceptually simpler metric δ′ defined by δ′(x, y)=1/(n + 1). Observe that, given a collection of bijections fn : Xn → Xn for each n ∈ N, the induced map f : X → X is an isometry of X. There is one more construction that we describe here. Analogously to the natural topological and metric structures on X, we can endow each Xn with the normalized counting measure. The resulting (uniform) product measure µ is a probability measure on the Borel σ-algebra B of X satisfying µ(V (y1,...,yn))=1/(|X1| · · · |Xn|). Clearly, µ is invariant under the isometries f described in the previous paragraph.

3. Discordant sets from visiting times As foreshadowed by the introduction, discordant sets naturally arise from dynamical systems, one of the most familiar being irrational rotation of the one- dimensional torus T = R/Z. Our search for discordant sets is motivated by the fact that if α ∈ T is irrational and U ⊆ T has nonempty interior, then {n ∈ Z: nα ∈ U} DISCORDANT SETS AND ERGODIC RAMSEY THEORY 11 is syndetic. The punchline of this direction of reasoning is that if E ⊆ T is nowhere dense and has positive measure, then for almost every x ∈ T, the set of visiting times RE(x) = {n ∈ Z: x + nα ∈ E} is discordant. Here RE(x) is non-piecewise syndetic for any x ∈ T by Lemma 9 below, whereas the positive upper density of RE(x) is guaranteed only for almost every x by an application of the von Neumann ergodic theorem.6 Theorem 10 is a generalization of this construction which yields many discordant sets in many settings, and is the core of our analogy between discordant sets and nowhere dense sets of positive measure. Let us note that the restriction “almost every x” is a signiﬁcant and unavoidable constraint. One can construct a closed nowhere dense set E ⊆ T of positive measure for which {n ∈ Z: x + nα ∈ E} is empty for inﬁnitely many x ∈ T. The importance of the irrationality of α is that (T, x 7→ x + α) is minimal if and only if α is irrational. In our more general setting we replace this system with a minimal action of a semigroup G on a compact space X. In this setting, we wish to connect properties of RE (x) = {g ∈ G: gx ∈ E} as a subset of G with the properties of E as a subset of X. Lemma 9 connects the piecewise syndeticity of RE (x) with the topological properties of E and Theorem 10 connects the density of RE(x) with the measure of E, allowing us to construct discordant sets. −1 −1 In analogy with our notation for subsets of G, we write A E = a∈A a E for A ⊆ G and E ⊆ X. S Lemma 9. Let (X, G) be a minimal topological dynamical system, where G is a countable cancellative semigroup. Then, for any nowhere dense E ⊆ X and x ∈ X, the set RE(x) is not piecewise syndetic.

Proof. It is a classical fact that (X, G) is minimal if and only if RV (x) is syndetic for each open V ⊆ X and x ∈ X (e.g., [22, Theorem 1.15]). We reproduce the argument here for the convenience of the reader. The “if” direction is evident, so suppose V ⊆ X is open and x ∈ X. Let Y = G−1V ; we claim that Y = X. Indeed, if y ∈ X \ Y , then gy∈ / V for all g, so y cannot have dense orbit, contradicting the minimality of (X, G). By compactness, there is a finite set H ⊆ G such that X = H−1V . Now fix g ∈ G and find h ∈ H such that gx ∈ h−1V . Then −1 hg ∈ RV (x), so G = H RV (x), proving that RV (x) is syndetic. −1 Now let E ⊆ X be nowhere dense and let F ⊆ G be finite. Then F RE(x) = −1 −1 RF −1E(x), and F E is nowhere dense. Let V be an open set lying in X \ F (E). −1 Then F RE(x) is disjoint from the syndetic set RV (x), and so cannot be thick. Since F was arbitrary, RE(x) cannot be piecewise syndetic.

Suppose (X, G) is a topological dynamical system. If G is amenable, there always exists a G-invariant measure µ on X by a version of the Bogolyubov-Krylov theorem (see [36, Theorem 6.9] for a proof of the case G = Z; the theorem for general G is proved analogously). Moreover, µ can taken to be ergodic by invoking the so-called ergodic decomposition. Theorem 10. Let (X, G) be a minimal topological dynamical system, where X is compact and G is a countable amenable cancellative semigroup with a Følner

6 In fact, one can easily see that RE (x) has positive natural density for almost every x by using the pointwise ergodic theorem. 12 VITALY BERGELSON, JAKE HURYN, AND RUSHIL RAGHAVAN sequence Φ = (Φn). Let µ be a G-invariant probability measure on the Borel σ- algebra B of X. If E ⊆ X is nowhere dense with positive measure, then there exists z ∈ X such that RE(z) is discordant. If, in addition, (X, B,µ,G) is ergodic, dΦ(RE (z)) ≥ µ(E) for almost every z ∈ X. It is worth noting that if X is an uncountable compact metric space, then X possesses a nowhere dense set of measure c (with respect to the G-invariant measure µ) for any c ∈ (0, 1).

Proof. Let f = 1E. By Theorem 7, 1 lim f(gx)= f ∗(x), n→∞ |Φn| gX∈Φn where f ∗ is the orthogonal projection of f onto the space of G-invariant functions 2 2 ′ in L (X,µ), and the convergence is in L . Extract a subsequence Φ = (Φnk )k∈N −1 ∗ such that limk→∞ |Φnk | f(gx)= f (x) almost everywhere. For each k, g∈Φnk 1 P 1 1 f(gx) dµ = f(gx) dµ = µ(E)= µ(E). |Φn | |Φn | |Φn | k g∈Φ k g∈Φ k g∈Φ Z Xnk Xnk Z Xnk Hence, by the dominated convergence theorem, f ∗ dµ = µ(E). Thus, there is a set of positive measure on which f ∗ > 0. Choose z such that f ∗(z) > 0 and R 1 ∗ dΦ′ (RE (z)) = lim f(gz)= f (z). k→∞ |Φn | k g∈Φ Xnk ∗ Then dΦ(RE (z)) ≥ dΦ′ (RE(z)) = f (z) > 0, and RE(z) is not piecewise syndetic by Lemma 9. In the presence of ergodicity, f ∗ is the constant function f dµ = µ(E), so we −1 ∗ may take z to be any element for which limk→∞ |Φnk | g∈Φ f(gx) = f (x). R nk As we noted, such z form a subset of X of full measure. P 4. Discordant sets and generalizations of squarefree numbers It is known that the set of squarefree numbers in N is discordant, and in this section we describe a wide-ranging generalization of this fact and examine in detail some special cases. 4.1. A characterization and a construction. We will prove non-piecewise syndeticity in two ways; the first uses a characterization (Lemma 11) of non-piecewise syndetic sets, and the second uses Lemma 9. Ramsey theory aficionados will notice that Lemma 11 can be rephrased as a characterization of sets satisfying the dual property, in the sense of Lemma 1, to piecewise syndeticity, often denoted PS∗. Lemma 11. Let G be a countable semigroup and let A ⊆ G. Then A is not piecewise syndetic if and only if, for any finite H ⊆ G, there exists a syndetic S ⊆ G such that A ∩ HS = ∅.

Proof. Let H ⊆ G be ﬁnite and consider the set BH = {g ∈ G: Hg ∩ A = ∅}. c Observe that g ∈ BH if and only if hg ∈ A for some h ∈ H, that is, if and only −1 −1 if g ∈ H A. Thus, by Lemma 1, BH is not syndetic if and only if H A is not thick. In particular, BH fails to be syndetic for all ﬁnite H ⊆ G if and only if A is not piecewise syndetic. DISCORDANT SETS AND ERGODIC RAMSEY THEORY 13

Before constructing non-piecewise syndetic sets using Lemma 11, we ﬁrst recall a general group-theoretic form of the Chinese remainder theorem.7 Lemma 12 (Chinese remainder theorem). Let G be a group with normal subgroups N1,...,Nn which satisfy

(2) (N1 ∩···∩ Ni−1 ∩ Ni+1 ∩···∩ Nn)Ni = G for each i ∈{1,...,n}, n n and write N = i=1 Ni. Then the map ϕ: G/N → i=1 G/Ni defined by ϕ(gN)= (gN1,...,gNn) is an isomorphism. T Q Proof. We give a proof since it is quick. It is immediate from this first isomorphism theorem that ϕ is a well-defined injective homomorphism. To prove surjectivity, for each g ∈ G and i ∈ {1,...,n}, we will find h ∈ G satisfying ϕ(hN) = (N1,...,Ni−1,gNi,Ni+1,...,Nn). By hypothesis, we may find x ∈ N1 ∩···∩ Ni−1 ∩ Ni+1 ∩···∩ Nn and y ∈ Ni such that xy = g. Setting h = x gives the desired outcome. We now combine the Chinese remainder theorem with Lemma 11 and Lemma 9 to obtain an abundance of non-piecewise syndetic sets. Although in Lemma 11 we must remove “thicker and thicker” syndetic sets to guarantee non-piecewise syndeticity, Theorem 13 shows that in special situations (when the syndetic subsets are suitably “arithmetically/algebraically independent”, as the Chinese remainder theorem will guarantee) this is not necessary.

Theorem 13. Let G be a countable group and (Nn) be a sequence of normal finite- index subgroups of G such that for any n ≥ 2, the subgroups N1,...,Nn satisfy (2). Then, for any sequence (gn) in G, the set A = G \ gnNn is not piecewise syndetic. S In principle, the set A may be empty. Below we will see many examples where this is not the case. Notice that we can equivalently state the conclusion of Theorem 13 as “every piecewise syndetic subset of G contains an element (in fact, infinitely many elements) of gnNn”. We will give two proofs of this fact. n Proof 1. Fix F = S{f1,...,fn} ⊆ G. Let N = i=1 Ni, and let ϕ be the isomorphism guaranteed by the Chinese remainder theorem. Then, setting x = −1 −1 −1 T ϕ (f1 g1N1,...,fn gnNn), we have fix ∈ giNi for each i ∈ {1,...,n}, so F (xN) ∩ A = ∅. Since the subgroups Ni have finite index, N is syndetic, so Lemma 11 completes the proof. Our second proof of Theorem 13, though longer, uses Lemma 9 and strengthens our analogy between discordant sets and nowhere dense sets of positive measure. In the remainder of this subsection we will use the definitions made in Subsections 2.4 and 2.5.

Proof 2. Consider the Cantor space X = G/Nn. Given g1,...,gn ∈ G, deﬁne (3) V (g1,...,gn)= {x ∈ X : xiNi =QgiNi for each i ∈{1,...,n}} Sets having the form given in (3) constitute an open base for the topology on X. Let G act by componentwise multiplication on X, i.e., gx = (gxnNn)n∈N. The Chinese remainder theorem guarantees that this action is minimal. To see this, let

7The authors found this useful formulation of the Chinese remainder theorem in the StackEx- change discussion SE2954401. 14 VITALY BERGELSON, JAKE HURYN, AND RUSHIL RAGHAVAN x ∈ X and g1,...,gn ∈ G be given. By the Chinese remainder theorem, ﬁnd h ∈ G −1 −1 so that (hN1,...,hNn) = (g1x1 N1,...,gnxn Nn). Then hx ∈ V (g1,...,gn) since −1 hxiNi = hNixi = gixi Nixi = giNi. This proves that x has dense orbit. To apply Lemma 9, we simply observe that A = RE(1), where

(4) E = {x ∈ X : xn ∈/ Nn for each n ∈ N} and 1 ∈ X denotes the trivial sequence 1 = (Nn)n∈N. The set E is nowhere dense, since it is closed and is not a superset of any set given in (3). As one might expect from this second proof, we can obtain a density result for Theorem 13 by applying Theorem 10.

Theorem 14. Let G, X, and (Nn) be as in Proof 2 of Theorem 13, and set cn = |G/Nn| for each n ∈ N. Suppose G is amenable, possessing a Følner sequence Φ. Let µ be the probability measure on the Borel σ-algebra B of X satisfying µ(V (h1,...,hn)) = 1/(c1 ··· cn). Finally, given x ∈ X, set Ax = G \ xnNn. Then, for almost every x ∈ X, S 1 (5) dΦ(Ax)= 1 − . N cn nY∈ Proof. Since G acts minimally on X by isometries, the system (X, B,µ,G) is uniquely ergodic, and in particular ergodic. Let E ⊆ X be as in (4). Then by Theorem 10, we have, for almost every x ∈ X, 1 dΦ(Ax)= µ(E)= 1 − . N cn nY∈ (That µ(E)= (1 − 1/cn) can be easily checked since E is closed.) In practice,Q Theorem 14 will not be useful for constructing explicit discordant sets as we cannot easily control for which x equation (5) holds. 4.2. Some number-theoretic examples of discordant sets. In this subsection we apply Theorem 13 to construct some discordant sets. In the following subsection we provide proofs of positive density for all but one of the constructions. The most primitive example is the following. Example 15. The set Q of squarefree integers is not piecewise syndetic: Q = Z \ p2Z. p prime[ It is classical that d(Q)=6/π2, so that Q is discordant. Example 15 can be generalized a great deal. We provide ﬁve generalizations below.

Example 16. Let B = (bn) be a sequence of pairwise coprime positive integers and let FB denote the set of B-free integers

FB = Z \ bnZ, N n[∈ consisting of integers not divisible by any element of B. Then FB is not piecewise syndetic, and if 1/bn < ∞, then d(FB)= (1 − 1/bn) [25, Theorem 9, Section V.5]. We provide a proof of this fact below (Theorem 22). A well-known but P Q DISCORDANT SETS AND ERGODIC RAMSEY THEORY 15 notable special case is that the set of k-free integers, those divisible by no perfect kth power, has density 1/ζ(k).

Example 17. Let K be a ﬁnite extension of Q, let k ≥ 2, and let QK,k denote the k-free integers K: k QK,k = OK \ p ,

p⊆OK prime[ ideal where OK denotes the ring of integers in K. Since Dedekind domains such as OK have unique factorization of ideals into prime ideals, we may apply Theorem 13 (or an appropriate ring-theoretic version of the Chinese remainder theorem) to show d that QK,k is not piecewise syndetic. Moreover, as K is a finite extension, OK =∼ Z as additive groups, where d is the degree of K/Q. Define Φ to be a sequence of d 1 balls (pulled back from Z to Ok under this isomorphism), with respect to the L norm, centered at 0 and having increasing radius. Then dΦ(QK,k)=1/ζK (k) as shown in [17, Corollary 4.6], where ζK denotes the Dedekind zeta function of K. k Equivalently, dΦ(QK,k)= p(1 − 1/N(p) ), where N(p) denotes the absolute norm of p, i.e., the cardinality of the residue field O /p. Hence Q is discordant. Q K K,k Example 18. Given k ∈ N and a prime p, write ep(k) to denote the exponent on p in the prime factorization of k. In [4, Theorem 3.9], it is shown in a somewhat cumbersome way that, for any infinite set P of primes and any corresponding collection of nonempty subsets (Yp)p∈P of N, the set

A = {k ∈ N: ep(k) ∈/ Yp for each p ∈ P } is not piecewise syndetic. This strengthens the non-piecewise syndeticity result of Example 15 and, in certain cases, Example 16, since it is clear that A is a superset min(Yp) of the (p )p∈P -free numbers. Below, we state this result in better generality and deduce it easily from Theorem 13. Given any b ∈ N and k ∈ Z, deﬁne eb(k) to be the largest power of b dividing k (set eb(0) = ∞). Now let B = (bn) be a sequence of pairwise coprime positive integers and let u = (un) be any sequence of positive integers. Deﬁne u FB = {k ∈ Z: ebn (k) 6= un for each n ∈ N}.

u un+1 un u Then FB ⊆ Z \ (bn N + bn ), so FB is not piecewise syndetic by Theorem 13. In Theorem 23 below, we prove that, if b−un < ∞, then S n u P bn − 1 d(FB)= 1 − . un+1 N bn nY∈ Example 19. The set C consisting of pairs (a,b) of coprime integers can be ex- pressed as C = (Z × Z) \ (pZ × pZ), p prime[ making clear that C is not piecewise syndetic by Theorem 13. It is well known that d(C)=6/π2, which will follow from the more general Theorem 25 (in other words, the probability that two “randomly chosen” integers are coprime is 6/π2; see, e.g., [26, Theorem 332]). Many possible generalizations in this direction can be imagined. 16 VITALY BERGELSON, JAKE HURYN, AND RUSHIL RAGHAVAN

Example 20. Our final example will take place in a noncommutative setting. Consider the Heisenberg group H3(Z), the matrix group 1 a c 0 1 b : a,b,c ∈ Z .     0 0 1    3 Associating these matrices with triples (a,b,c), we view H3(Z) as Z with the group operation (a1,b1,c1) · (a2,b2,c2) = (a1 + a2,b1 + b2,a1b2 + c1 + c2). For each n ∈ N, 3 the kernel nZ of the entrywise reduction map H3(Z) → H3(Z/nZ) is a normal 3 subgroup of H3(Z) of index n . Hence, for any sequence (bn) of pairwise coprime 3 positive integers, the set H3(Z) \ bnZ is not piecewise syndetic in H3(Z) by Theorem 13. We will see in Theorem 26 that this set has density (1 − 1/b3 ) with S n respect to an appropriate Følner sequence in H3(Z). Q 4.3. Computing density. Before computing the density of some of the non- piecewise syndetic sets we constructed in Subsection 4.2, we make some definitions and prove a general lemma. Suppose we have a countable amenable cancellative semigroup G with a Følner sequence Φ. Given a function f : G → Z, we define 1 AΦ(f) = lim sup f(g), n→∞ |Φn| gX∈Φn and likewise AΦ(f) and AΦ(f) in the obvious way. We make the abbreviations A(f), A(f), and A(f) analogously with the notation for density. We will be most interested in the case when AΦ(f) exists. Assuming this limit exists, it has some obvious but useful properties. First, given E ⊆ G, AΦ(1E) = dΦ(E). Second, if f1,f2 : G → Z satisfy f1 ≤ f2, then AΦ(f1) ≤ AΦ(f2). Finally, f 7→ AΦ(f) is finitely additive: AΦ(f1 + f2)= AΦ(f1)+ AΦ(f2). In order to prove that a set of the form G \ En has density, we will impose a number of nice properties on the family (En). We adopt the notation EI = Ei S i∈I for I ⊆ N. In addition, we use Pk(N) to refer to the set of k-element subsets of N. T Specifically, we say that (En) is Φ-inclusion-exclusion good, or Φ-I.E. good if

(a) dΦ(En) exists for each n ∈ N, and n∈N dΦ(En) < ∞; (b) EI = ∅ for every inﬁnite I ⊆ N; P (c) for every ﬁnite I ⊆ N, dΦ(EI )= i∈I dΦ(Ei); and (d) for each k ∈ N, AΦ( N 1E )= N dΦ(EI ). I∈Pk( ) IQ I∈Pk( ) For Φ-I.E. good families, theP following lemmaP states that an asymptotic version of the inclusion-exclusion principle holds.

Lemma 21. Suppose G and Φ are as above and (En) is a Φ-I.E. good family of subsets of G. Write F = G \ En. Then dΦ(F)= (1 − dΦ(En)). n Proof. Clearly dΦ(F) ≤ (1S− dΦ(En)), since F ⊆QG \ i=1 Ei for each n ∈ N. Now let n ∈ N. We will demonstrate that Q S dΦ(F) ≥ 1 − dΦ(Ei)+ dΦ(EI ) −···− dΦ(EI ); N N N Xi∈ I∈PX2( ) I∈PX2n−1( ) properties (a) and (c) of Φ-I.E. goodness guarantee that each sum is finite. To do this, it suffices, by the finite additivity of f 7→ AΦ(f) and property (d) of Φ-I.E. DISCORDANT SETS AND ERGODIC RAMSEY THEORY 17 goodness, to show

(6) 1F ≥ 1 − 1Ei + 1EI −···− 1EI . N N Xi∈ I∈PX2( ) I∈PX2n−1 Let f denote the function on the right side of (6) and let r ∈ N. If r ∈ F, then clearly f(r) = 1. Otherwise, due to property (b) of Φ-I.E. goodness, we may suppose r is contained in exactly k ≥ 1 elements of (En). Then 2n−1 k k − 1 f(r)= (−1)i = − ≤ 0, i 2n − 1 i=0 X so indeed 1F ≥ f. This veriﬁes that dΦ(F)= dΦ(F)= (1 − dΦ(En)). Using this lemma, we compute the density of sets constructedQ in Examples 16 and 18.

Theorem 22. Let B = (bn) be a sequence of pairwise coprime positive integers satisfying 1/bn < ∞. Then d(FB)= (1 − 1/bn).

Proof. WeP prove this statement for FBQ= N \ bnN for convenience, but it is of no consequence, since FB is symmetric about 0. We must show that (bnN) is I.E. good. Only property (d) is not immediate. S

Write f = 1bnN. Since 1b1N ≤ 1b1N + 1b2N ≤···≤ f, clearly A(f) ≥ 1/bn. To prove A(f) ≤ 1/bn, the key point is that, for each k,n ∈ N, P P k P 1 1 1 N(i) ≤ , k bn b i=1 n X which is just a statement of the fact that there are no more than k/bn multiples of bn in {1,...,k}. It follows that, for each k ∈ N, k 1 1 f(i) ≤ , k b i=1 N n X nX∈ whence A(f)= A(f)= 1/bn. Thus (bnN) is I.E. good.

Theorem 23. Let B =P (bn) be a sequence of pairwise coprime positive integers −un and let u = (un) be a sequence of positive integers for which bn < ∞. Then

u bn − 1 P d(FB)= 1 − . un+1 N bn nY∈ Proof. As in Theorem 22 we confine ourselves to N. For each n ∈ N define Bn = u {k ∈ N: ebn (k) = un}, so that FB ∩ N = N \ Bn. We aim to show that (Bn) is an I.E. good family. (The density d(Bn) is easy to compute but is also derived below.) Clearly property (b) of I.E. goodness isS satisfied. ui Given I ⊆ N, define also BI = i∈I Bi, bI = i∈I bi, and rI = i∈I bi . Observe that BI ⊆ rI N. Define CI = BI /rI to be the set of positive integers not T Q Q divisible by bi for any i ∈ I. Observe that CI is bI -periodic, i.e., for each k ∈ N, one has k ∈ CI if and only if bI + k ∈ CI . In particular, d(CI )= |CI ∩{1,...,bI }|/bI. ∼ Now consider the isomorphism ϕ: Z/bI Z → i∈I Z/bi afforded by the classical Chinese remainder theorem. An element x ∈ Z/bI Z is in CI if and only if ϕ(x) Q 18 VITALY BERGELSON, JAKE HURYN, AND RUSHIL RAGHAVAN is nonzero in every coordinate, from which it follows that |CI ∩ {1,...,bI }| = i∈I (bi − 1). From this we get Q 1 b − 1 d(B )= d(r C )= (b − 1) = i = d(B ), I I I i ui+1 i bI rI bi Yi∈I Yi∈I Yi∈I proving that property (c) above holds for (Bn). In addition, the formula we have obtained for d(Bi) and the hypotheses on B and u verify property (a). Now fix k ∈ and set f = 1 . As in Theorem 22 we have A(f) ≥ N I∈Pk(N) BI d(B ). To prove that A(f) ≤ d(B ), fix ε> 0. Select N ∈ such that I P I N P P 1 ε < . bI 2 I∈Pk(N) max(XI)>N

Now let I ∈ Pk(N) and note that, since BI = rI CI where CI is bI -periodic, we have, for each m ≥ abI rI ,

|BI ∩{1,...,m}| (a + 1) · |B ∩{1,...,bI rI }| a +1 ≤ = · d(BI ). m abI rI a Find a ∈ for which d(B )/a < ε/2 and set R = max b r . N I∈Pk(N) I I∈Pk({1,...,N}) I I Then, for each m ≥ aR, P m 1 |B ∩{1,...,m}| |B ∩{1,...,m}| f(k)= I + I m m m k=1 I∈Pk({1,...,N}) I∈Pk(N) X X max(XI)>N a +1 1 ≤ · d(BI )+ a rI I∈Pk({1,...,N}) I∈Pk(N) X max(XI)>N

<ε + d(BI ) N I∈PXk( )

(we have |BI ∩{1,...,m}|/m ≤ 1/rI since BI ⊆ rI N). This veriﬁes that (Bn) is indeed I.E. good and completes the proof.

We next compute the density of the sets constructed in Examples 19 and 20. To do so, we will need a technical lemma.

Lemma 24. Let G and H be countable amenable cancellative semigroups with Følner sequences Φ and Ψ, respectively. Let (Dn) be a Φ-I.E. good family of subsets of G and (En) be a Ψ-I.E. good family of subsets of H. Then (Dn×En) is a (Φ×Ψ)- I.E. good family of G × H, where Φ × Ψ=(Φn × Ψn)n∈N. Proof. The fact that Φ × Ψ is a Følner sequence for G × H is easy and left to the reader. We outline the proof that (Dn × En) satisﬁes property (d) of (Φ × Ψ)-I.E. goodness; the others are straightforward. We have, for each k ∈ N,

But since

|DI ∩ Φn| AΦ 1D = lim = dΦ(DI ), I n→∞  N  N |Φn| N I∈PXk( ) I∈PXk( ) I∈PXk( )   and likewise with EI , proving property (d) reduces to verifying the following claim: given doubly-indexed sequences (ak,n) and (bk,n) of nonnegative quantities satisfying

• limn→∞ ak,n = ak and limn→∞ bk,n = bk for each k, and • limn→∞ k∈N ak,n = ak < ∞ and limn→∞ k∈N bk,n = bk < ∞, one has lim a b = a b . n→∞P k∈N k,n k,nP k k P P Theorem 25. LetP d ≥ 2 and let (Pbn,1),..., (bn,d) be sequences such that, for each i, (bn,i) consists of pairwise coprime positive integers. Then

d 1 d Z \ (bn,1Z ×···× bn,dZ) = 1 − . N ! N bn,1 ··· bn,d n[∈ nY∈ Proof. This follows from Theorem 22 via repeated applications of Lemma 24.

3 To apply Lemma 24 to the Heisenberg group H3(Z), viewed as Z with the group operation described in Example 20, a laborious but straightforward computation shows that 2 2 (7) Φ = ({−n,...,n}×{−n ,...,n }×{−n,...,n})n∈N is a Følner sequence in H3(Z). Hence Theorem 22 and Lemma 24 together give the following.

Theorem 26. Let (bn) be a sequence of pairwise coprime positive integers. Then, letting Φ be as in (7), we have

3 1 dΦ H3(Z) \ bnZ = 1 − 3 . N ! N bn n[∈ nY∈

5. Discordant sets in SL2(Z) Let G denote the free semigroup generated by a and b with identity 1 ∈ G. It is well known that G is not amenable. To see why, suppose Φ is a Følner sequence for G, and form the partition G = {1} ∪ aG ∪ bG. One of these parts must have positive upper density with respect to Φ, and since {1} cannot, suppose without loss 2 of generality that c = dΦ(aG) > 0. Then aG,baG,b aG, . . . are disjoint subsets of G each with upper density c> 0. But Lemma 3 guarantees that this is impossible, which is our desired contradiction. This argument can be modified without too much difficulty to show that nonabelian free groups are not amenable, and moreover any group possessing a nonabelian free subgroup is not amenable. We now turn our attention to the group SL2(Z). Since SL2(Z) contains nonabelian free subgroups,8 it is not amenable. However, we will define a natural density-like notion for subsets of SL2(Z) and construct sets which are large with respect to this notion but not piecewise syndetic. Although this density does not

8For example, it is well-known that the matrices 1 2 and 1 0 generate a free subgroup of 0 1 2 1 SL2(Z) (see [33]). 20 VITALY BERGELSON, JAKE HURYN, AND RUSHIL RAGHAVAN enjoy shift-invariance (which in the amenable case is assured by Følner sequences), we consider this result an interesting aside which has independent value. In particular, it makes more robust the analogy with nowhere dense sets of positive measure, showing that it extends beyond the amenable case. We ﬁrst leverage Theorem 13 to produce an ample supply of non-piecewise syndetic subsets of SL2(Z). Recall the congruence subgroups of SL2(Z), which are subgroups of the form a b a b 1 0 Γ(n)= ∈ SL (Z): ≡ (mod n) . c d 2 c d 0 1 The subgroup Γ(n) is normal in SL2(Z) for each n ∈ N since it is the kernel of the homomorphism ϕn : SL2(Z) → SL2(Z/nZ) deﬁned by reducing each matrix entrywise modulo n.

Theorem 27. Let (bn) be a sequence of pairwise coprime positive integers, and define A = SL2(Z) \ Γ(bn). Then A is not piecewise syndetic. Proof. Fix any coprimeS m,n ∈ N. Since Γ(m) ∩ Γ(n)=Γ(mn), to apply Theorem 13 it suffices to show that Γ(m)Γ(n)=SL2(Z). We claim that we may always find, for any T ∈ SL2(Z/mZ), some S ∈ SL2(Z) such that ϕm(S) = T and ϕn(S) = I. Granted this, fix S ∈ SL2(Z), and find S1,S2 ∈ SL2(Z) such that ϕm(S1)= ϕm(S), −1 −1 ϕn(S1) = I, ϕm(S2) = I, and ϕn(S2) = ϕn(S). Then SS1 S2 ∈ ker(ϕm) ∩ ker(ϕn) =Γ(mn), so there exists T ∈ Γ(mn) such that S = (TS2)S1 ∈ Γ(m)Γ(n). Hence Γ(m)Γ(n)=SL2(Z), as desired. a b To prove the claim, let T = c d , treating the entries as elements of Z. Find x, y ∈ Z such that mx + ny = 1 and set a′ b′ mx + any bny S′ = = , c′ d′ cny mx + dny so S′ ≡ T (mod m) and S′ ≡ I (mod n). Now ad − bc ≡ 1 (mod m), so gcd(a,b) is invertible mod m and hence coprime with m, and from this one checks that mn is coprime with gcd(a′,b′). Letting q denote the product of primes dividing a′ but not b′, we see that a′ and b′ + mnq are coprime. Find s ∈ Z such that a′ b′ + mnq det = mns +1 c′ d′ and u, v ∈ Z such that a′u + (b′ + mnq)v = 1. Then a′ b′ + mnq S = c′ + mnsv d′ − mnsu is an element of SL2(Z) satisfying ϕm(S)= T and ϕn(S)= I.

We will now introduce density in SL2(Z). Define a b F = ∈ SL (Z): |a|, |b|, |c|, |d|≤ n . n c d 2 We then define the upper density of A ⊆ SL2(Z) to be |A ∩ F | d(A) = lim sup n , n→∞ |Fn| and define d and d analogously. We first make some crude estimates on the growth of the sets Fn and Γ(k) ∩ Fn. DISCORDANT SETS AND ERGODIC RAMSEY THEORY 21

Lemma 28. There is a positive constant C and a positive integer N such that 2 |Fn|≥ Cn for all n ≥ N.

2 a b Proof. We will produce, for suﬃciently large n, at least Cn matrices c d ∈ SL2(Z) such that max{|a|, |b|, |c|, |d|} ∈ {|a|, |b|} and |a|, |b|, |c|, |d| ≤ n. As we saw in Example 19, |{a,b ∈ Z: |a|, |b|≤ n and gcd(a,b)=1}| π2 lim = . n→∞ (2n + 1)2 6 Thus there is a constant C and positive integer N such that for all n ≥ N, there are at least Cn2 pairs (a,b) such that max{|a|, |b|} ≤ n and gcd(a,b) = 1. Given such a pair (a,b), using the Euclidean algorithm, we can ﬁnd c, d ∈ Z with |c| ≤ |a| a b and |d| ≤ |b| such that ad − bc = 1. Then c d ∈ Fn, as desired.

Lemma 29. For each k ≥ 2, there is a positive constant Ck such that for all n ∈ N, 2 |Γ(k) ∩ Fn|≤ Ckn . Moreover, limk→∞ Ck =0. Proof. Deﬁne a b Λ(k)= ∈ Γ(k): max{|a|, |b|, |c|, |d|} ∈ {|a|, |b|} . c d By swapping the rows, one sees that |Γ(k) ∩ Fn|≤ 2 · |Λ(k) ∩ Fn|. There are at most ((2n + 1)/k)2 pairs (a,b) ∈ Z2 with |a|, |b| ≤ n such that a ≡ 1 (mod k) and b ≡ 0 (mod k). In particular, there are at most ((2n + 1)/k)2 pairs (a,b) with gcd(a,b) = 1. Given such a pair, one can ﬁnd, using the Euclidean algorithm, c, d ∈ Z such that ad − bc = 1, |d| ≤ |b|, |c| ≤ |a|. Then {(x, y) ∈ Z2 : ax − by =1} = {(c + ma,d + mb): m ∈ Z}, so there are at most three pairs (x, y) such that ax − by = 1 and max{|x|, |y|} ≤ a b max{|a|, |b|}, and in particular there are at most three (x, y) such that x y ∈ Λ(k) ∩ Fn. 2 Combining all of this, we have |Γ(k) ∩ Fn|≤ 6((2n + 1)/k) . With these estimates, we can show that under certain conditions, the sets in Theorem 27 have positive lower density.

Theorem 30. Let Ck and C be as above and let (bn) be a sequence of pairwise coprime integers such that Cbn /C < 1. Then A = SL2(Z) \ Γ(bn) has positive lower density. P S Proof. For 2 ≤ n 0, as desired. P 22 VITALY BERGELSON, JAKE HURYN, AND RUSHIL RAGHAVAN

6. Recurrence and anti-recurrence Let G be a countable group. A set A ⊆ G is a set of recurrence if for all measure preserving systems (X, B,µ,G) and for all Y ⊆ X with µ(Y ) > 0, there is a g ∈ A not equal to the identity such that µ(Y ∩ gY ) > 0. Recall that in a measure-preserving system, we assume µ(X) = 1, so such sets do exist. Call A ⊆ G an anti-recurrence set, or an AR set, if for all g ∈ G, gA is not a set of recurrence. We will begin by showing that any piecewise syndetic sets is a shift of a set of recurrence, so any AR set with positive upper density is therefore discordant. Lemma 31. Let T ⊆ G be thick. Then T is a set of recurrence. Proof. Let T ⊆ G be thick. First, we will show that there is an inﬁnite sequence −1 (gn) of elements of G such that for all pairs i, j ∈ N with i < j, gj gi ∈ T . Indeed, such a set can be constructed recursively: given g1,...,gn, we can ﬁnd gn+1 such −1 that gn+1{g1,g2,...,gn}⊆ T . Let (X, B,µ,G) be a measure preserving system, and let Y ⊆ X have µ(Y ) > 0. −1 There exist gi,gj with i < j such that µ(gj Y ∩giY ) > 0. Then, µ(Y ∩gj giY ) > 0, as desired. r Lemma 32. Let A be a set of recurrence partitioned as A = i=1 Ai. Then at least one Ai is a set of recurrence. S Proof. Assume that no Ai is a set of recurrence. Then, for each i, there is a measure preserving system (Xi, Bi,µi, G) and Yi ⊆ Xi with µi(Yi) > 0 such that for all g ∈ Ai, µi(Yi ∩ gYi) = 0. The actions of G on each Ai induce an action on X = X1 ×···× Xn, which is measure preserving with respect to the product measure µ of the measures µi. Let Y = Y1 ×···×Yn. We have µ(Y1 ×···×Yn) > 0, but for all g ∈ A, since g ∈ Ai for some i, we have µi(Yi ∩ gYi) = 0, and thus µ(Y ∩ gY ) = 0, meaning that A is not a set of recurrence. This establishes the contrapositive. Theorem 33. If A is AR, then A is not piecewise syndetic. Proof. Apply Lemma 5, noting the previous two lemmas. In [5], it was noted that Straus’s example of a discordant set was also an AR set. We will follow their terminology and refer to AR sets with positive upper density as Straus sets. In what follows, we investigate the question of whether a given discordant set is a Straus set. We begin by showing that Example 16 provides large class of non-examples.

Theorem 34. Let B = (bn) be a sequence of pairwise coprime positive integers such that 1/bn < ∞. Define FB = Z \ bnZ. Then FB is a discordant set which is not AR. P S Proof. As noted in Example 16, FB is discordant. Now let b ∈ FB. By [12, Theorem 2.8], FB − b contains an IP set. Since IP sets are sets of recurrence, B − b is also a set of recurrence, and hence B cannot be an AR set. Not all countable amenable groups admit Straus sets. For example, in the group of finitely supported even permutations of N, every set with positive upper Banach density is a set of recurrence (in fact, such a set must be IP). In [5, Theorem 1.23], DISCORDANT SETS AND ERGODIC RAMSEY THEORY 23 it is shown that Straus sets exist in a locally compact, second countable, amenable group if and only if the von Neumann kernel of that group is not cocompact. In such groups, for all ε > 0, there is a Straus set E with d(E) > 1 − ε. Thus, it is natural to ask whether such a group has a Straus set E with d(E) = c for each c ∈ (0, 1). We prove that this is the case for all countable abelian groups in Theorem 37. First, we need the following lemma, a special case of a lemma appearing in [5]. Lemma 35 (cf. [5, Lemma 5.1]). Let G be a countable amenable group and let Φ be a Følner sequence in G. Let (An) be a sequence of subsets of G such that dΦ(An) exists for every n, d(An) < ∞, and dΦ(A1 ∪···∪An) exists for every n. Then there are ′ ′ cofinite subsets A ⊆ An such that for C = A , dΦ(C) = lim dΦ(A1 ∪···∪ An). P n n Before proving Theorem 37, we give a proofS of the case G = Z which does not rely on tools from ergodic theory. This also serves as a model of the proof of Theorem 37 for arbitrary discrete countable abelian groups. For the proof of Theorem 36, the notion of well-distribution for a sequence in T will be used. For a more detailed exposition of this theory, see [31]. A sequence (xn) in T is called well-distributed if for any measurable A ⊆ T with µ(∂A) = 0,

N 1 lim 1A(xn)= µ(A). N−M→∞ N − M n=XM+1 Equivalently, (xn) is well-distributed if for any Følner sequence Φ in Z and any measurable A ⊆ T with µ(∂A) = 0, 1 lim 1A(xk)= µ(A). n→∞ |Φn| kX∈Φn Theorem 36. Let c ∈ (0, 1) and let Φ be a Følner sequence in Z. Then, there is a Straus set A ⊆ Z such that dΦ(A)= c.

Proof. Let t ∈ (0, 1), and let α ∈ T be irrational. For a set A ⊆ T, let RA denote n+1 {n ∈ Z: nα ∈ A}. For n ≥ 0, let rn = t/((2n + 1) · 2 ), let n Bn = (−rn + kα, rn + kα), k=[−n n and let Cn = k=0 Bn. Since each of Bn and Cn have boundaries with measure zero, and since (nα) is well-distributed in T, we have dΦ(RB )= µ(Bn) and dΦ(RC )= S n n µ(Cn) for all n. Since µ(Bn) ≤ t, by Lemma 35, there are coﬁnite subsets ′ R ⊆ RB such that Bn n P ∞ ∞ ′ dΦ RB = lim µ(Cn)= µ Bn . n n→∞ n=0 ! n=0 ! [ [ Let fn(t)= µ(Cn), and let f(t) = limn→∞ fn(t). For all t, ∞ 1 |f(t) − f (t)|≤ µ(B ) ≤ , n n 2n k=Xn+1 so the convergence is uniform. Since the boundary of each interval has measure zero, the functions fn are continuous, so f is continuous. Moreover, limt→1 f(t) = 1 and 24 VITALY BERGELSON, JAKE HURYN, AND RUSHIL RAGHAVAN

′ ′ limt→0 f(t) = 0, so we can choose t so that dΦ RBn =1−c. Let A = Z\ RBn . Then, dΦ(A)= c. It remains to show that A is AR. Let k ∈ Z. Then S S

−rk rk −rk rk (A − k) ∩ n ∈ Z: + nα, + nα ∩ , 6= ∅ 2 2 2 2 is ﬁnite, so A − k cannot be a set of recurrence.

In fact, a version of Theorem 36 holds true for all countable abelian groups.

Theorem 37. Let G be a (discrete) countable abelian group. Let c ∈ (0, 1), and let Φ be a Følner sequence in G. Then, there is a Straus set A ⊆ G such that dΦ(A)= c.

Proof. We will construct a compact group X and a homomorphism ϕ: G → X with dense image. Depending on G, this construction can occur three different ways. First, assume that G has an element of infinite order, i.e., G is not a torsion group. We define a homomorphism from G to X = T as follows. By [20, Theorems 23.1 and 24.1], G can be embedded in T ⊕ N Q, where T is a torsion group. Let ι be the embedding from G into this direct sum. Let g be an element of G with L infinite order. Then, there is a projection ρ from T ⊕ N Q onto Q such that ρ(ι(g)) is nonzero. Choose an irrational α ∈ T. Then the map h 7→ ρ(ι(h))α is a homomorphism from G to T with dense image, so in thisL case, we will let X = T and let ϕ: G → T be such that ϕ(h)= ρ(ι(h))α. Next, assume that G is a torsion group. For p a prime and n ∈ N, let Hp,n be ∞ ∼ the Prüfer p-group Z[p ]. (By definition, for n1,n2 ∈ N, Hp,n1 = Hp,n2 .) By [20, Theorems 23.1 and 24.1], G can be embedded into p prime n∈N Hp,n. Let i be the corresponding inclusion map, and let πp,n be the projection from the direct sum L L onto Hp,n. For convenience, we abbreviate ρp,n = πp,n ◦ i. There are two cases. First, assume there is some pair (p,n) such that |ρp,n(G)| is infinite. Note that k Hp,n is isomorphic to the subgroup {a/p : a ∈ Z, k ∈ N} of T; call this isomorphism ψ. Let ϕ = ψ ◦ ρp,n. Then ϕ: G → T is a homomorphism with dense image, so we set X = T. On the other hand, assume that for all pairs (p,n), |ρp,n(G)| is finite. Then G is embedded in p n ρp,n(G). Since G is a subgroup of a direct sum of finite cyclic groups, by [20, Theorem 18.1], there is an infinite sequence (nj )j∈N of integers L L ∼ greater than 1 such that G = j Znj . Then there is a dense embedding ϕ from G =∼ Z to the X = Z . If each Z is endowed with the discrete metric δ , j nj j njL nj j we can give X the product metric δ(x, y)= 2−jδ (x ,y ). The Haar measure L Q j∈N j j j on X is the product measure of the normalized counting measures of each Znj . As a result, for any a ∈ [0, 1] and x ∈ X, the setP {y ∈ X : δ(x, y) = a} has measure zero. Here, the casework ends, and for G, the following is true. We have a dense homomorphism ϕ: G → X, where X is a compact group. Moreover, X is endowed with an invariant metric δ such that for x ∈ X and a ∈ (0, 1), we have µ({y ∈ X : δ(x, y)= a}) = 0, where µ is the normalized Haar measure on X. We define a natural G-action on X by isometries via gx = x + ϕ(g). Since im(ϕ) is dense, the action is uniquely ergodic, with the normalized Haar measure µ as the unique invariant measure. Thus, by Theorem 8, for any f : X → R for which the DISCORDANT SETS AND ERGODIC RAMSEY THEORY 25 set of discontinuities has measure zero and x ∈ X, 1 (8) lim f(gx)= f dµ. n→∞ |Φn| X gX∈Φn Z Suppose Y ⊆ X is a Borel set with µ(∂Y ) = 0, and write RY = {g ∈ G: ϕ(g) ∈ Y }. Taking x = 0 and f = 1Y , equation (8) becomes 1 dΦ(RY ) = lim 1Y (ϕ(g)) = 1Y dµ = µ(Y ). n→∞ |Φn| X gX∈Φn Z The remainder of the argument proceeds along the same lines as in Theorem 36. n+1 Let t ∈ (0, 1), and for n ≥ 0, let rn = t/(|Φn| · 2 ). Let

Bn = ({x ∈ X : δ(0, x) < rn} + ϕ(g)), g∈[Φn n ′ and let Cn = k=0 Bn. Using Lemma 35, for each n, find a cofinite RBn ⊂ RBn such that S ∞ ∞ ′ dΦ R = lim µ(Cn)= µ Bn . Bn n→∞ n=0 ! n=0 ! [ [ Let fn(t) = µ(Cn). Note that since the construction of the sets Cn depends con- tinuously on t, the functions fn are continuous. Let f = limn→∞ fn. The convergence of (fn) to f is uniform, so f is continuous. Note that limt→1 f(t)=1 ′ and limt→0 f(t) = 0, so we can choose t so that dΦ RBn = 1 − c. Then, if ′ A = G \ R , we have dΦ(A) = c. For any g ∈ G, choose k such that g ∈ Φk. Bn S Then (A − g) ∩{h ∈ G: δ(ϕ(h), 0) < rk,t} is finite, so A − g cannot be a set of recurrence.S Theorem 37 can be extended to countable commutative cancellative semigroups, allowing us to conclude that discordant subsets with any density in (0, 1) exist in, e.g., (N, +) and (N, ×) as well. Corollary 38. Let G be a countable commutative cancellative semigroup, let Φ be a Følner sequence in G, and let c ∈ (0, 1). Then there is a discordant set A ⊆ G with dΦ(A)= c. Proof. We can embed G in a countable abelian group G by a process similar to that of embedding an integral domain in a field. The elements of G can be thought formally as “fractions” of elements of G, so for any g ∈ Ge, there is an h ∈ G such that gh ∈ G. We will show that Φ is a Følner sequence in G (cf. [6,e Theorem 2.12]). Let g ∈ G, and let h ∈ G be such that hg ∈ G. For each en ∈ N, e |Φn △ gΦn| |Φn △ hgΦn| |gΦn △ hgΦn| |Φn △ hgΦn| |Φn △ hΦn| e ≤ + = + , |Φn| |Φn| |Φn| |Φn| |Φn| so |Φ △ gΦ | lim n n =0, n→∞ |Φn| and thus Φ is a Følner sequence in G. By Theorem 37, there is a discordant set A ⊆ G with dΦ(A)= c. Set A = A ∩ G. We claim that A is not piecewise syndetic in G. Suppose, toward the contrapositive,e that there is a finite H ⊆ G such that He −1Ae is thick.e Let F ⊆ G be finitee and select g ∈ G such that Fg ⊆ G. Then there is an h ∈ G such that Fgh ⊆ H−1A. But then H−1A ⊆ H−1A is piecewise e e 26 VITALY BERGELSON, JAKE HURYN, AND RUSHIL RAGHAVAN syndetic in G, which establishes the claim. Thus A is a non-piecewise syndetic subset of G with dΦ(A)= c. e 7. Large subsets of {0, 1}G Let G be a countable semigroup. In this section we will identify a subset A ⊆ G G with its indicator function 1A ∈{0, 1} . We will say that 1A is thick, syndetic, etc. whenever A possesses these properties. Thus, families of subsets of G are identified with subsets of {0, 1}G. To study the topological properties of subsets of {0, 1}G, we give {0, 1}G the product topology, making it a Cantor space. In this space, a base for the topology is given by the cylinder sets, which have the form G V (L1,L2)= {α ∈{0, 1} : α(L1)= {1} and α(L2)= {0}} for finite, disjoint L1,L2 ⊆ G. Let us write

L = {(L1,L2) ∈P(G) ×P(G): L1 and L2 are ﬁnite and disjoint}.

7.1. Background and motivation. Recall that if X is a topological space homeomorphic to the Cantor set, it is a Baire space, meaning that it cannot be written as a countable union of nowhere dense sets. A subset of X which is a countable union of nowhere dense sets is called meager, and a subset of X whose complement is meager is comeager. The fact that X is a Baire space means that no subset can be both meager and comeager, so one can think of the elements of a comeager set as being “topologically generic”. We wish to stress in this section the analogy between topology and measure theory in which “meager” corresponds to “measure 0” and “comeager” corresponds to “measure 1”. We will describe comeager subsets of {0, 1}G, whose elements can thus be thought of as “topologically generic” subsets of G. Let us first motivate our results by N examining them in a familiar setting. Consider the case G = N, where {0, 1} is the space of infinite binary sequences. An infinite binary sequence is called disjunctive if every finite binary word occurs within it. (Formally, α ∈{0, 1}N is disjunctive if, for each k ∈ N and ω ∈{0, 1}k, there exists n ∈ N such that α(n + i − 1) = ω(i) for each 1 ≤ i ≤ k.) The subset of {0, 1}N consisting of disjunctive sequences is large both topologically and measure-theoretically: it is comeager and of measure 1. However, the topological and measure-theoretical worlds diverge once we consider the limiting frequencies with which finite subwords occur. Given a sequence α ∈ {0, 1}N and a finite word ω ∈ {0, 1}k, one says that ω occurs in α with a limiting frequency |{1 ≤ i ≤ n − k + 1: α(i + j − 1) = ω(j) for each 1 ≤ j ≤ k}| (9) lim . n→∞ n It is a classical result of Emile´ Borel [14] that the set of normal sequences, those for which any word of length k occurs as a subword with limiting frequency 2−k, has measure 1. Moreover, as we will see, it is a meager set. On the other hand, one can consider sequences for which limits analogous to (9) (but modified slightly to avoid counting overlapping sequences) fail to exist as badly as possible. For these sequences, the lim sup of the frequency of any finite word is 1 and the corresponding lim inf is 0. We will call such sequences extremely non-averageable (or ENA). It turns out that the set of ENA sequences, in contrast to the set of normal sequences, is comeager [16, Theorem 1]. DISCORDANT SETS AND ERGODIC RAMSEY THEORY 27

We will make the following generalizations of the facts we have just stated: assuming only cancellativity, we prove an analogue in a countable semigroup G to the statement that the set of disjunctive sequences is comeager (Theorem 39); assuming also amenability, we prove an analogue to the statement that the set of ENA sequences is comeager (Theorem 41). A generalization of measure-theoretical normality to countable amenable semigroups and the corresponding results can be found in [6]. Since the set of normal sequences and the set of ENA sequences are disjoint, we see also that the set normal sequences are meager, and that the set of ENA sequences has measure 0 (see [6, Proposition 4.7] regarding the former statement in arbitrary amenable cancellative semigroups). This allows us to give another motivation for the results of this section. One reason we gave for studying discordant sets is that they illustrate how geometric and analytic notions of largeness for subsets of G diverge; in this section, we construct subsets of {0, 1}G on which basic topological and analytic notions of largeness disagree. Finally, we remark that, although we prove the results of this section in the space {0, 1}G, their generalizations to {0,...,n − 1}G (which can be thought of as the space of n-part partitions, or n-colorings, of G) are apparent and can be achieved with the same arguments.

7.2. Disjunctive elements of {0, 1}G. In this subsection we will assume that G is G cancellative. Call α ∈{0, 1} disjunctive if for any (L1,L2) ∈ L there is some g ∈ G 9 for which α ∈ V (L1g,L2g). Cancellativity here guarantees that L1g ∩ L2g = ∅ for all g ∈ G. Note that, for G = N, this deﬁnition specializes to the deﬁnition of disjunctive sequences we gave above. Theorem 39. The set of disjunctive elements of {0, 1}G is comeager.

Proof. For any (L1,L2) ∈ L let AL1,L2 = g∈G V (L1g,L2g), so that the set of disjunctive elements of {0, 1}G is S

A = AL1,L2 . (L1,L\2)∈L Now let (L1,L2), (M1,M2) ∈ L. Since G is cancellative, we can ﬁnd some g ∈ G for which (L1g ∪L2g)∩(M1 ∪M2)= ∅. Then we have 1L1g∪M1 ∈ AL1,L2 ∩V (M1,M2), so that AL1,L2 is dense, since it has nontrivial intersection with each element of the G open base for {0, 1} . Since AL1,L2 is a union of open sets, we have written A as a countable intersection of open dense sets, which proves that A is comeager.

We deduce as a relevant consequence of this theorem that the set of non-piecewise syndetic elements of {0, 1}G is meager, since any disjunctive element of {0, 1}G is thick. In non-cancellative semigroups, the set of piecewise syndetic elements of {0, 1}G may fail to be comeager if G is non-cancellative. For example, if G has an element z ∈ G satisfying gz = z for each g ∈ G, then α ∈ {0, 1}G is piecewise syndetic if and only if α(z) = 1, so the family of piecewise syndetic sets cannot be dense.

9In the notation of Section 8, we can equivalently say that α is disjunctive if and only if, for every (L1,L2) ∈ L, there exists g ∈ G for which gα ∈ V (L1,L2). This deﬁnition has the beneﬁt of making sense for non-cancellative semigroups. 28 VITALY BERGELSON, JAKE HURYN, AND RUSHIL RAGHAVAN

7.3. Extremely non-averageable elements of {0, 1}G. In this subsection we will assume that G is both cancellative and amenable. Fix a Følner sequence Φ in G. We will begin by defining Φ-normality of elements of {0, 1}G, as defined in [6], for comparison with the definition we will give for ENA. We will also utilize the notion of Φ-normality in Theorem 42. Given α ∈ {0, 1}G, one says that α is Φ-normal if, for each finite K ⊆ G and ω ∈{0, 1}K,

|{g ∈ G: Kg ⊆ Φn and α(hg)= ω(h) for each h ∈ K}| 1 lim = |K| . n→∞ |Φn| 2 The following classical lemma regarding Følner sequences, which states that they have a property analogous to thickness, assures us that the notion of Φ-normality is meaningful. Lemma 40. Let Φ be a Følner sequence for a countable cancellative semigroup G. Let K ⊆ G be finite. Then there exists N ∈ N such that, for each n ≥ N, Kgn ⊆ Φn for some gn ∈ G. Proof. Find N large enough that, for each n ≥ N and h ∈ K, 1 |Φ ∩ hΦ | |h(h−1Φ ∩ Φ )| |h−1Φ ∩ Φ | 1 − < n n = n n = n n . |K| |Φn| |Φn| |Φn| −1 Then h∈K h Φn 6= ∅, so let gn be an element of this set. Then Kgn ⊆ Φn. NowT we define extreme non-averageability in {0, 1}G. The following definitions are admittedly unwieldy because we wish to measure the frequency of occurrences without overlap of our finite sequence. We would be pleased if a simpler definition of ENA in an arbitrary countable amenable cancellative semigroup could be found. Fix a finite K ⊆ G. Given a finite set X ⊆ G, define

YK,X = {Y ⊆ X : KY ⊆ X, Ky ∩ Kz = ∅ for all y,z ∈ Y , and |Y | is maximal}. Now let ω ∈{0, 1}K, and α ∈{0, 1}G, and deﬁne |{y ∈ Y : α(hy)= ω(h) for all h ∈ K}| dΦ(K,ω,α) = lim sup max . n→∞ Y ∈YK,Φn |Y |

We observe that dΦ(K,ω,α) is well-defined since YK,Φn = {∅} for only finitely many n by Lemma 40. Define dΦ(K,ω,α) to be the corresponding lim inf. Hence, dΦ(K,ω,α) and dΦ(K,ω,α) give the (upper and lower) limiting frequencies, mea- sured along Φ, with which the finite binary word ω occurs without overlaps in α. For convenience, we write K = {(K,ω): K ⊆ G is finite and ω ∈{0, 1}K}.

We say that α is Φ-ENA if dΦ(K,ω,α) = 1 and dΦ(K,ω,α) = 0foreach(K,ω) ∈ K. With these deﬁnitions in hand, we proceed to the main result of this subsection. Theorem 41. The set of Φ-ENA elements of {0, 1}G is comeager. Proof. For each (K,ω) ∈ K, let G AK,ω = {α ∈{0, 1} : dΦ(K,ω,α) = 1 and dΦ(K,ω,α)=0}, G so that the set of Φ-ENA elements of {0, 1} is (K,ω)∈K AK,ω. Fix (K,ω) ∈ K; it suﬃces to show that A is comeager. K,ω T DISCORDANT SETS AND ERGODIC RAMSEY THEORY 29

′ Find a subsequence Φ = (Φnk )k∈N of Φ which satisﬁes

|Φn | + ··· + |Φn − | lim 1 k 1 =0. k→∞ |Φnk | k Define Ψ1 =Φ1 and, for k ∈ N,Ψk+1 =Φnk+1 \ m=1 Φnm ; let Ψ denote the Følner sequence (Ψn)n∈N. Let m ∈ K. Now for each n ∈ N choose Yn ∈ YK,Ψ and define S n + G Bn = {β ∈{0, 1} : for all y ∈ Yn and h ∈ K, β(hy)= ω(h)} and − G Bn = {β ∈{0, 1} : for all y ∈ Ψn and h ∈ K, β(hy) 6= ω(m)}. ± The definition of YK,Ψn ensures that the sets Bn are nonempty for all large enough + n, and they are clearly open since they are cylinder sets. In particular, Bn = −1 −1 − − V (ω ({1})Yn,ω ({0})Yn), and either Bn = V (∅,KΨn) or Bn = V (KΨn, ∅), − depending on whether ω(m)is1or0. Theset Bn is chosen this way to ensure that for all Yn ∈ YK,Ψn , there is no y ∈ Yn such that β(hy) = ω(h) for all h ∈ K. In ± ∞ ± particular, β(my) 6= ω(m) for all y. Finally, define Cn = m=n Bm. Let (L1,L2) ∈ L. Since the sets Yn are disjoint and L1 and L2 are finite, cancellativity gives that for all large enough n ∈ N there mustS be m ≥ n for which ± ± ± Bm ∩ V (L1,L2) 6= ∅ and hence Cn ∩ V (L1,L2) 6= ∅. It follows that Cn is open + − G and dense. Define C = (Cn ∩ Cn ), so that C is a comeager subset of {0, 1} . Moreover, if γ ∈ C, then T (10) dΨ(K,ω,γ) = 1 and dΨ(K,ω,γ)=0 ′ by construction. Since limk→∞ |Φnk |/|Ψk| = 1, (10) must hold with respect to Φ instead of Ψ, and thus also with respect to Φ. Hence C is a comeager subset of AK,ω, as we had hoped. It is important to note that the property of Φ-ENA, as well as Φ-normality, is dependent on Φ. The following theorem shows that this is true in a strong sense. Theorem 42. Let α ∈ {0, 1}G be disjunctive. Then there exist Følner sequences Ψ1 and Ψ2 of G such that α is simultaneously Ψ1-normal and Ψ2-ENA. G Proof. Let β1,β2 ∈{0, 1} be Φ-normal and Φ-ENA, respectively. For each n ∈ N i −1 i −1 i and i ∈{1, 2}, define Xn =Φn ∩βi ({1}) and Yn =Φn ∩βi ({0}) and find gn ∈ G i i i i i i such that α ∈ V (Xngn, Yngn). Then define the Følner sequences Ψ = (Ψn)n∈N by i i Ψn =Φngn. These Følner sequences have the desired properties.

8. Topological dynamics and piecewise syndeticity Let G denote any countable (not necessarily cancellative) semigroup. In this section we explore piecewise syndeticity of subsets of G in terms of dynamical properties of elements of {0, 1}G and their orbit closures. Some of the results presented below are well-known for subsets of N. Using these results we revisit and examine in greater detail the comparison between van der Waerden’s theorem and Szemerédi’s theorem made in the introduction. We adopt the notation of the previous section. For each g ∈ G, let ρg : G → G be the map h 7→ hg. We will consider the dynamical system ({0, 1}G, G), where G acts G on {0, 1} by gα = α ◦ ρg. In the familiar settings G = N and G = Z, this is just the left-shift map on infinite one- or two-sided sequences. In light of the natural correspondence between {0, 1}G and P(G), we may view this dynamical system 30 VITALY BERGELSON, JAKE HURYN, AND RUSHIL RAGHAVAN instead as (P(G), G). In this case, g acts on P(G) by A 7→ Ag−1. This action is −1 continuous, since for any (L1,L2) ∈ L, we have g V (L1,L2)= V (L1g,L2g). Given α ∈{0, 1}G, we define the orbit closure of α to be O(α)= {gα: g ∈ G}.

Now fix an increasing sequence (Fn) of finite subsets of G such that Fn = G. Let us note that for any α ∈ {0, 1}G, we have β ∈ O(α) if and only if, for all n ∈ N, S there exists g ∈ G such that β(x) = (gα)(x) for each x ∈ Fn. This simple but crucial property of orbit closures in ({0, 1}G, G) system follows immediately from the definition of the product topology on {0, 1}G. We illustrate what this means in the familiar case G = N. In this setting, β ∈ O(α) if and only if, for any n ∈ N, α can be shifted to agree with β on {1,...,n}. Put another way, β ∈ O(α) means that one can find arbitrarily long initial segments of β as subwords of α. Before stating the results, we give convenient reformulations of the notions of syndeticity, piecewise syndeticity, and thickness in terms of indicator functions. An element α ∈{0, 1}G is syndetic if and only if there exists a finite set H ⊆ G such that, for each x ∈ G, α(hx) = 1for some h ∈ H. Similarly, α is piecewise syndetic if and only if there exists an infinite sequence (gi) in G such that, for each n ∈ N and x ∈ Fn, α(hxgn) = 1 for some h ∈ H. Finally, α is thick if and only if 1G ∈ O(α), so by Lemma 1, α is syndetic if and only if 1∅ ∈/ O(α). Lemma 43. An element α ∈{0, 1}G is piecewise syndetic if and only if the orbit closure O(α) contains a syndetic element. Proof. Suppose α ∈ {0, 1}G is piecewise syndetic and find a finite set H ⊆ G and an infinite sequence (gn) in G which witnesses α’s piecewise syndeticity, in the sense of the previous paragraph. We claim that there is a decreasing sequence (Nm) of infinite subsets of N with the property that α(xgn1 )= α(xgn2 ) for each x ∈ Fm and n1,n2 ∈ Nm. Indeed, for any m, there are finitely many functions Fm →{0, 1}, so such a sequence can easily be constructed inductively by applying the pigeonhole G principle. Hence we can define β ∈ {0, 1} by β(x) = α(xgn) = (gnα)(x) for any x ∈ Fm and n ∈ Nm. By construction, β ∈ O(α). To show that β is syndetic, let x ∈ G, find m such that x ∈ Fm, and find n ∈ Nm with n ≥ m. Since x ∈ Fn, we can find h ∈ H such that α(hxgn) = 1. Then β(hx) = α(hxgn) = 1, so β is syndetic. Now suppose α ∈ {0, 1}G and β ∈ O(α) is syndetic. Let H ⊆ G be a finite set which witnesses β’s syndeticity. Fix n ∈ N and find m ∈ N such that HFn ⊆ Fm. Find gm ∈ G such that β(y) = (gmα)(y) for each y ∈ Fm. Let x ∈ Fn and find h ∈ H for which β(hx) = 1. Then hx ∈ Fm, so α(hxgm)= β(hx) = 1. This proves that α is piecewise syndetic.

Recall that subsystem of a dynamical system (X, G) is a nonempty closed G- invariant subset of X. A minimal subsystem is a subsystem which is itself a minimal dynamical system. By Zorn’s lemma, any dynamical system has a minimal subsystem (see, e.g., [24, Theorem 2.22]). We will call the subsystem of ({0, 1}G, G) consisting of the ﬁxed point {1∅} (the zero function) the trivial subsystem. The following theorem is a restatement of Lemma 43 in terms of the minimal subsystems of orbit closures. DISCORDANT SETS AND ERGODIC RAMSEY THEORY 31

Theorem 44. An element α ∈ {0, 1}G is piecewise syndetic if and only if O(α) possesses a nontrivial minimal subsystem. In particular, if α is discordant, O(α) does not have a nontrivial minimal subsystem. Moreover, sets of positive upper density having this property are exactly the discordant sets. Proof. Suppose α ∈ {0, 1}G is piecewise syndetic. By Lemma 43, there exists β ∈ O(α) which is syndetic. Then O(β) ⊆ O(α) and 1∅ ∈/ O(β), so O(β) has a nontrivial minimal subsystem. Conversely, suppose O(α) possesses a nontrivial minimal subsystem X. Then 1∅ ∈/ X, so every element of X is syndetic. Thus α is piecewise syndetic. The existence or nonexistence of nontrivial minimal subsystems in O(α) has significant combinatorial consequences, which we presently examine. Let us specialize to the case G = N. As stated in the introduction, there is an equivalent formulation of van der Waerden’s theorem as a topological recurrence theorem: Theorem 45 (Topological van der Waerden’s theorem). Let (X,T ) be a mini- mal10 topological dynamical system where X is a compact metric space. Then for any ℓ ∈ N and ε > 0 there exists m, d ∈ N and x ∈ X such that the points x, T dx,...,T (ℓ−1)dx are within ε of each other. Proof. This is a special case of [23, Theorem 1.4]. See also [22, Theorem 2.6] or [2, Corollary 3.9]. From here the classical van der Waerden’s theorem can be derived along the following lines. Suppose α ∈ {0, 1}N is piecewise syndetic. Then O(α) possesses a nontrivial minimal subsystem Y by Theorem 44. Since 1∅ ∈/ Y , every element of Y is syndetic. Giving {0, 1}N the product metric described in Subsection 2.5 and applying Theorem 45 inside (Y,T ) shows that, for any ℓ ∈ N, the system Y possesses an element with a length-ℓ arithmetic progression. Since these elements all lie in O(α), it follows that α is AP-rich. (As we saw in the introduction, this is equivalent, via the partition regularity of piecewise syndetic sets, to the formulation of van der Waerden’s theorem in terms of partitions of N.) This discussion supports our position taken in the introduction on the distinction between van der Waerden’s and Szemerédi’s theorems. Specifically, if the orbit closure of any element of {0, 1}N having positive upper density were to possess a nontrivial minimal subsystem, then Szemerédi’s theorem would follow immediately from the topological version of van der Waerden’s theorem using the argument we gave in the preceding paragraph. As our next combinatorial application of Theorem 44, we will prove the partition regularity of piecewise syndetic sets from a dynamical perspective (cf. Theorem 4). We first address the following question: For which α ∈{0, 1}G is O(α) itself minimal? The answer is that every finite subword occuring in α must occur syndetically

10In the sources, one will not see the condition of minimality explicitly imposed in the formulations of Theorem 45. We of course do not lose generality by making this assumption as we may always pass to a minimal subsystem. Moreover, the topological proofs of this theorem of which the authors are aware, with the notable exception of [9, Proposition L], use minimality in a crucial way. 32 VITALY BERGELSON, JAKE HURYN, AND RUSHIL RAGHAVAN in α, as the following lemma states. The result is well known in the case when G is Z or N; see, e.g., [22, Proposition 1.22]. Lemma 46. Let α ∈ {0, 1}G. Then O(α) is minimal if and only if, for all (L1,L2) ∈ L, the set {g ∈ G: gα ∈ V (L1,L2)} is either empty or syndetic. Proof. This is a consequence of the following general phenomenon: Let (X, G) be a topological dynamical system where X is a metric space, and let x ∈ X. Then O(x)= {gx: g ∈ G} is minimal if and only if, for each open V ⊆ X, RV (x) is either empty or syndetic. Clearly, if O(x) is minimal, then this condition holds (see the proof of Lemma 9). Conversely, suppose y ∈ O(x). We will show that x ∈ O(y), which will prove that O(x) is minimal. To this end, fix n ∈ N and find a finite set H ⊆ G such −1 that H RB(1/(2n),x)(x) = G. Find a sufficiently small ε > 0 that hB(ε,y) ⊆ B(1/(2n),hy) for each h ∈ H. Next find g ∈ G for which gx ∈ B(ε,y) and find h ∈ H satisfying hg ∈ S. Then hgx ∈ B(1/(2n),hy) ∩ B(1/(2n), x), i.e., d(hy,x) < 1/n. This proves that x ∈ O(y), as claimed. Before continuing, we remark that, as in the previous section, the preceding can be easily generalized to statements about {0, 1,...,r}G. We leave the proofs of the generalizations to the interested reader. Using this, we give the promised result on the partition regularity of piecewise syndetic sets. The following proof in the case G = Z can be found in [22, Theorem 1.24]). Theorem 47. Let A ⊆ G be piecewise syndetic and suppose it is partitioned as r A = i=1 Ci. Then Ci is piecewise syndetic for some i. G Proof.S Define α ∈ {0, 1,...,r} by α(g) = 0 if g∈ / A and α(g) = i if g ∈ Ci. By Lemma 44, O(α) has a nontrivial minimal subsystem X. Let β ∈ X. By Lemma 46 and the nontriviality of X, there exists i ∈{1,...,r} such that {g ∈ G: β(g)= i} is syndetic. Define β′ ∈ {0, 1}G by β′(g) = 1 if β(g) = i and β′(g) = 0 otherwise. ′ Then β is a syndetic element in O(1Ci ), so Ci must be piecewise syndetic. It is worth mentioning that Theorem 47 is easily proved with the help of topological algebra in βG, the Stone-Cechˇ compactification of G, that is, the space of ultrafilters on G. It is well known that A ∈ G is piecewise syndetic if and only if some shift g−1A lies in a minimal idempotent of βG [30, Theorem 4.43]. From here partition regularity follows from definition of an ultrafilter. This approach to Ram- sey theory—learning about subsets of G by studying βG—has proved very fruitful, leading to, for example, a slick proof of Hindman’s theorem (see [30, Corollary 5.9] or [3, Theorem 1.2]). We do not pursue this further but direct the curious reader towards the references we have just mentioned. References

[1] L. Argabright and C. Wilde. “Semigroups satisfying a strong Følner condition”. In: Pro- ceedings of the American Mathematical Society 18.4 (1967), pp. 587–591. [2] V. Bergelson. “Ergodic theory and diophantine problems”. In: Topics in Symbolic Dynamics and Applications. London Mathematical Society Lecture Note Series. Cambridge University Press, 2000, pp. 167–206. [3] V. Bergelson. “Minimal idempotents and ergodic Ramsey theory”. In: Topics in Dynamics and Ergodic Theory. London Mathematical Society Lecture Note Series. Cambridge Uni- versity Press, 2003, pp. 8–39. REFERENCES 33

[4] V. Bergelson, M. Beiglböck, N. Hindman, and D. Strauss. “Multiplicative structures in additively large sets”. In: Journal of Combinatorial Theory, Series A 113.7 (2006), pp. 1219– 1242. [5] V. Bergelson, J. C. Christopherson, D. Robertson, and P. Zorin-Kranich. “Finite product sets and minimally almost periodic groups”. In: Journal of Functional Analysis 270.6 (2016), pp. 2126–2167. [6] V. Bergelson, T. Downarowicz, and M. Misiurewicz. “A fresh look at the notion of normality”. In: Annali della Scuola Normale Superiore di Pisa, Classe di Scienze (To appear). [7] V. Bergelson and H. Furstenberg. “WM groups and Ramsey theory”. In: Topology and its Applications 156.16 (2009), pp. 2572–2580. [8] V. Bergelson, N. Hindman, and R. McCutcheon. “Notions of size and combinatorial properties of quotient sets in semigroups”. In: Topology Proceedings 23 (1998), pp. 23–60. [9] V. Bergelson and A. Leibman. “Set-polynomials and polynomial extension of the Hales- Jewett theorem”. In: Annals of Mathematics 150 (1999), pp. 33–75. [10] V. Bergelson, A. Leibman, and E. Lesigne. “Intersective polynomials and polynomial Sze- merédi theorem”. In: Advances in Mathematics 219 (2008), pp. 369–388. [11] V. Bergelson and R. McCutcheon. “An ergodic IP polynomial Szemerédi theorem”. In: Memoirs of the AMS 146.695 (2000), pp. 1–106. [12] V. Bergelson and I. Rusza. “Squarefree numbers, IP sets, and ergodic theory”. In: Paul Erd˝os and his Mathematics. Vol. 4. 5. Janos Bolyai Mathematical Society, 2002. Chap. 8, pp. 147–160. [13] A. Bernardino, R. Pacheco, and M. Silva. “The gap structure of a family of integer subsets”. In: The Electronic Journal of Combinatorics 21.1 (2014). #P1.47. [14] E.´ Borel. “Les probabilités dénombrables et leurs applications arithmétiques”. In: Rendiconti del Circolo Matematico di Palermo 27 (1909), pp. 247–271. [15] T. C. Brown. “An interesting combinatorial method in the theory of locally finite semigroups”. In: Pacific Journal of Mathematics 36 (1971), pp. 285–289. [16] C. Calude and T. Zamfirescu. “Most numbers obey no probability laws”. In: Publicationes Mathematicae Debrecen 54 (Supplement) (1999), pp. 619–623. [17] F. Cellarosi and I. Vinogradov. “Ergodic properties of k-free integers in number fields”. In: Journal of Modern Dynamics 7.3 (2013), pp. 461–488. [18] J. C. Christopherson. “Closed ideals in the Stone-Cechˇ compactification of a countable semigroup and some applications to ergodic theory and topological dynamics”. PhD thesis. The Ohio State University, 2014. [19] P. Erd˝os. “A survey of problems in combinatorial number theory”. In: Annals of Discrete Mathematics 6 (1980), pp. 89–115. [20] L. Fuchs. Infinite Abelian Groups. Elsevier Science, 1970. [21] H. Furstenberg. “Ergodic behavior of diagonal measures and a theorem of Szemerédi on arithmetic progressions”. In: Journal d’Analyse Mathématique 31 (1977), pp. 204–256. [22] H. Furstenberg. Recurrence in Ergodic Theory and Combinatorial Number Theory. Prince- ton University Press, 1981. [23] H. Furstenberg and B. Weiss. “Topological dynamics and combinatorial number theory”. In: Journal d’Analyse Mathématique 34 (1978), pp. 61–85. [24] W. H. Gottschalk and G. A. Hedlund. Topological Dynamics. American Mathematical So- ciety, 1955. [25] H. Halberstam and K. Roth. Sequences. Oxford University Press, 1966. [26] G. H. Hardy and E. M. Wright. An Introduction to the Theory of Numbers. Oxford Uni- versity Press, 2008. [27] N. Hindman. “Algebra in βS and its applications to Ramsey theory”. In: Mathematica Japonica 44 (1996), pp. 581–625. [28] N. Hindman. “Finite sums from sequences within cells of a partition of N”. In: Journal of Combinatorial Theory, Series A 17.1 (1974), pp. 1–11. [29] N. Hindman. “Preimages of points under the natural map from β(N × N) to βN × βN”. In: Proceedings of the American Mathematical Society 37.2 (1973), pp. 603–608. [30] N. Hindman and D. Strauss. Algebra in the Stone-Cechˇ Compactificaton. De Gruyter, 2012. [31] L. Kuipers and H. Neiderreiter. Uniform Distribution of Sequences. Wiley, 1974. [32] A. L. T. Paterson. Amenability. American Mathematical Society, 1988. 34 REFERENCES

[33] I. L. Sanov. “A property of a representation of a free group”. In: Doklady Akademii Nauk SSSR 57 (1947), pp. 657–659. [34] E. Szemer´edi. “On sets of integers containing no k elements in arithmetic progression”. In: Acta Arithmetica 27 (1975), pp. 199–243. [35] B. L. van der Waerden. “Beweis einer Baudetschen Vermutung”. In: Nieuw. Arch. Wisk. 15 (1927), pp. 212–216. [36] P. Walters. An Introduction to Ergodic Theory. Springer-Verlag, 1982.

Department of Mathematics, Ohio State University, Columbus, OH 43210, USA Email address: [email protected]