Partition regularity of some quadratic equations Joint work with N. Frantzikinakis Ergodic Theory with Connections to Arithmetic Heraklion, June 3, 2013 – I – A Theorem of partition regularity Definition. A family of finite subsets of N is partition regular if, for every finite partition N = C1 ∪ · · · ∪ Cr of N, at least one element Cj of this partition contains a set in this family. The notion of partition regularity of an equation f (x1, . . . , xk ) = 0 with integer unknowns is defined in a similar way. We restrict to non trivial solutions, that is with distinct values of x1, . . . , xk . 1 Examples. – Schur Theorem (1916): The equation x + y = z is partition regular. – van der Waerden Theorem (1927): The family of arithmetic progressions of given length is partition regular. – Rado’s Theorem (1933) characterizes the systems of linear equations that are partition regular. – Less is known for non linear equations. Classical problem: are the equations x2 + y 2 = z 2 and x2 + y 2 = 2z 2 partition regular? – The polynomial van der Waerden Theorem of Bergelson and Leibman (1996) provides examples of non linear configurations. These configurations are invariant under translations. 2 Definition (Partition regularity of an equation with a free variable). The equation p(x, y, λ) = 0 is partition regular if, for every finite partition N = C1 ∪ · · · ∪ Cr of N there exist x 6= y in the same subset Cj and an arbitrary λ ∈ N with p(x, y, λ) = 0. Examples. – The equation x − y = λ2 is partition regular (Sárközy 1978, Furstenberg). This is a particular case of the polynomial van der Waerden Theorem. – The equation x + y = λ2 is partition regular (Khalfalah & Szemerédi 2006). In this talk we consider equations of the form p(x, y, λ) = 0 where p is a quadratic homogeneous polynomial. These equations are not translation invariant and not linear in x and y. 3 Open questions: Are the equations x2 + y 2 = λ2 and x2 + y 2 = 2λ2 partition regular? Theorem The equations: 16x2 + 9y 2 = λ2 and x2 + y 2 − xy = λ2 (and many others) are partition regular. In the sequel we restrict to the equation 16x2 + 9y 2 = λ2. 4 – II – Reduction to a theorem about multiplicative functions Parametrization We obtain a parametrization of the solutions of the equation 16x2 + 9y 2 = λ2 by letting x = km(m + 3n) ; y = k(m + n)(m − 3n) and λ = k(5m2 + 9n2 + 6mn) for acceptable k, m, n ∈ N, meaning such that x and y are positive and distinct. This parametrization satisfies: • it is invariant under dilation; • x and y are products of two linear forms in the variables m and n. • m has the same coefficient in the four linear forms. It is possible to explicitly characterize the homogeneous quadratic equations p(x, y, λ) = 0 admitting a similar parametrization of the family of solutions. The results of this talk are valid for all these equations. 5 Partition regularity and density Translation invariant equations that are partition regular satisfy often the stronger property of density regularity, meaning that the equation admits a solution in every set of integers of positive density. Furstenberg’s correspondence principle allows to deduce density regularity from a recurrence result in ergodic theory. The equation 6x2 + 9y 2 = λ2 is not translation invariant. Open question. Is it true that every set of integers of positive density contains a non trivial solution (x, y)? 6 Multiplicative density The equation 6x2 + 9y 2 = λ2 is invariant under dilations, and this leads us to use the multiplicative density. Let {p1 < p2 < p3 < . . . } be the set of primes. For every N , let ΦN = m m m p1 1 p2 2 . . . pN N : 0 ≤ m1, m2, . . . , mN < N n o . (ΦN : N ≥ 1) is an example of a multiplicative Følner sequence: for every r ∈ Q+, ΦN \ rΦN |ΦN | → 0 when N → +∞ where rΦN = {rx : x ∈ ΦN } ∩ N. Definition. The multiplicative density of the subset E of N is |E ∩ ΦN | dmult(E) = lim sup . |ΦN | N →+∞ 7 Since the multiplicative density is subadditive, partition regularity follows from: Theorem (Multiplicative density regularity). Every subset E of N with positive multiplicative density contains a non trivial solution (x, y) of the equation 16x2 + 9y 2 = λ2. 8 A translation to ergodic theory Definition. A measure preserving action of the multiplicative group Q+ on a probability space (X, µ) is a family (Tr : r ∈ Q+), of measurable, invertible, measure preserving transformations of X with Tr Tr = Trs for all r, s ∈ Q+ . Multiplicative version of Furstenberg’s correspondence principle. Let E ⊂ N be a set of positive multiplicative density. There exist a measure preserving action (Tr : r ∈ Q+) of Q+ on a probability space (X, µ) and a subset A of X with µ(A) = dmult(E) and, for every k ∈ N and all r1, . . . , rk ∈ Q+, dmult(r1E ∩ · · · ∩ rk E) ≥ µ(Tr1 A ∩ · · · ∩ Trk A) . 9 We are reduced to show: Theorem (Ergodic formulation). Let (Tr : r ∈ Q+) be a measure preserving action of Q+ on a probability space (X, µ) and let A ⊂ X be a set of positive measure. Then there exist x 6= y ∈ N and λ ∈ N with 16x2 + 9y 2 = λ2 such that µ(Tx−1A ∩ Ty−1A) > 0 . Indeed, if (X, µ), (Tr : r ∈ Q+) and A are given by the correspondence principle, for x, y ∈ N we have dmult({k ∈ N : kx ∈ E and ky ∈ E}) = dmult(x−1E ∩ y −1E) ≥ µ(Tx−1A ∩ Ty−1A). 10 Let (X, µ), Tr and A be as in the last Theorem. We want to show that there exists a non trivial solution (x, y) of the equation such that −1 −1 µ Tx A ∩ Ty A > 0. Using the parametrization and the invariance of µ under the transformations Tr , we are reduced to showing that there exist acceptable m, n ∈ N with Z X Tm(m+3n)(m+n)−1(m−3n)−1 f · f dµ > 0 where f = 1A. We recognize an integral arising in the Spectral Theorem. 11 Multiplicative functions Definition. A multiplicative function is a function χ : N → C, of modulus 1, such that χ(xy) = χ(x)χ(y) for all x, y ∈ N . We write M for the family of multiplicative functions. These functions are often called “completely multiplicative functions”. A multiplicative function is characterized by its value on the primes. Endowed with the pointwise multiplication and with the topology of the pointwise convergence, the family M is a compact abelian group. Its unit is the constant function 1. 12 Every multiplicative function can be extended to a function on Q+ by χ a/b) = χ(a) χ(b) for all a, b ∈ N. The group M of multiplicative functions is the dual group of the multiplicative group Q+, the duality being given by the last formula. Spectral Theorem of actions of Q+. Let (Tr : r ∈ Q+) be a measure preserving action of Q+ on a probability space (X, µ) and let f ∈ L2(µ). Then there exists a finite positive measure ν on the compact abelian group M, called the spectral measure of f , such that Z X Tr f · f dµ = Z M χ(r) dν(χ) for every r ∈ Q+. 13 Let A ⊂ X with µ(A) > 0. Let ν be the spectral measure of f = 1A. Recall that we want to show that there exist acceptable m, n ∈ N such that Z X Tm(m+3n)(m+n)−1(m−3n)−1 f · f dµ > 0. By the Spectral Theorem and the multiplicativity of the functions χ, we are reduced to show: Theorem (Spectral formulation). Let (Tr : r ∈ Q+) be a measure preserving action of Q+ on a probability space (X, µ), A ⊂ X a set of positive measure and ν the spectral measure of 1A. Then there exist m, n ∈ N with m > 3n, m(m + 3n) 6= (m + n)(m − 3n) and Z M χ(m)χ(m + 3n)χ(m + n)χ(m − 3n) dν(χ) > 0. (This integral is allways ≥ 0.) 14 Averaging In fact we show that there are many values of m and n such that this integral is positive: we show that some average of this integral is positive: Since we are proving a result about sets of positive multiplicative density, multiplicative averages may seem more natural, but we use ordinary (additive) averages. Some notation If φ is a function defined on a finite set A, 1 X φ(x). Ex∈A φ(x) = |A| x∈A Same notation for a function of several variables. For every N ∈ N, [N ] = {1, 2, . . . , N } ; ZN = Z/N Z. 15 We find it more convenient to deal with functions defined on a cyclic group than on an interval of N. Notation. f for the smallest prime ≥ 10N . For every N ∈ N, we write N For χ ∈ M and x ∈ ZN e, χN (x) = χ(x) 0 if x ∈ [N ] ; otherwise. f = N and This is only a technical point, you can forget it and consider that N that χN = χ. 16 Taking averages in the preceding formulation, we get that the theorem of partition regularity follows from: Theorem (Final form). Let (Tr : r ∈ Q+) be a measure preserving action of Q+ on a probability space (X, µ), A ⊂ X a set of positive measure and ν the spectral measure of 1A. Then the lim sup when N → +∞ of Z M Em,n∈Z 1[N ](n) χN (m) χN (m + 3n) χN (m + n) χN (m − 3n) dν(χ) e N is positive. 17 – III – Fourier analysis of multiplicative functions The first tool for the analysis of the expression in the Theorem is the discrete Fourier transform. Definition The discrete Fourier coefficients of a function φ on ZN are fb(ξ) = Ex∈ZN φ(x) e(−xξ/N ) for ξ ∈ ZN where e(x) = exp(2πix). Contrasting with the Möbius function, multiplicative functions can have large Fourier coefficients and even large averages. However, their Fourier coefficients are small on “minor arcs”. 18 A Lemma of Kátai (1986) gives: Corollary (Kátai). For every θ > 0 there exist q ∈ N and c > 0, depending only on θ, such that, for every χ ∈ M, p c if sup En∈[N ] χ(n) e(nα) > θ then there exists p with α − < . q N χ∈M This could be the starting point of the circle method. For each χ ∈ M, χN can be decomposed as a sum of • a structured term, meaning approximatively periodic with small period, • plus a Fourier uniform term, meaning with small Fourier coefficients. 19 Fourier uniformity is measured by the Gowers U 2-norm kφkU 2 of φ. Definition The U 2-norm of a function φ on ZN is kφkU 2 = Ex,t1,t2∈ZN φ(x) φ(x + t1) φ(x + t2) φ(x + t1 + t2) = X 1/4 1/4 4 b |φ(ξ)| . ξ∈ZN If |f | ≤ 1, b(ξ)| ≤ kf k . kf k2 ≤ max | f 2 U2 U 20 – IV – “Higher order Fourier analysis” of multiplicative functions This Fourier analysis of multiplicative functions is not sufficient. Indeed, we want to study Em,n∈Z 1[N ](n) χN (m) χN (m + 3n) χN (m + n) χN (m − 3n). e N But expressions of this form are not controlled by the U 2-norms of the functions in the average. We need the Gowers U 3-norm, that we do not define here. Lemma. If φ1, . . . , φ4 are functions on ZN of modulus ≤ 1 then 1/2 1[N ](n)φ1(m) φ2(m + 3n) φ3(m + n) φ4(m − 3n) ≤ C min kφikU 3 . e N i Em,n∈Z 21 A theorem of decomposition Our main result is a theorem of decomposition of multiplicative functions for the U 3-norm. Weak decomposition theorem for the U 3-norm (simplified statement) For every > 0 there exist q ∈ N and c > 0, independent of χ and of N , such that, for every χ ∈ M, there is a decomposition of χN as a sum χN (x) = χs(x) + χu(x) for x ∈ ZN e where c for every x ∈ ZN χs(x + q) − χs(x) ≤ (Structured part) kχukU 3 ≤ . (U 3-uniform part) N and 22 Some ingredients in the proof of the decomposition theorem: • The Inverse Theorem for the U 3-norm of Green & Tao (2005) leads to study the correlation of multiplicative functions with some nilsequences. • The number theoretic input is again Kátai’s Lemma, now applied to nilsequences. • At this point, we need another “dynamical” ingredient, the Theorem of quantitative equidistribution of orbits in nilmanifolds of Green & Tao (2007). 23 – V – Putting the pieces together We want to show that the lim sup of Z Em,n∈Z 1[N ](n)χN (m)χN (m + 3n)χN (m + n)χN (m − 3n) dν(χ) e N M is positive. Strategy. • We need a more elaborate form of the decomposition theorem, with a stronger control of the uniformity norm and an “explicit” form of the structured part. • The price to pay is the introduction of a third “error” term. • The uniform term has a negligible contribution. • We need also to use the fact that the measure ν on M is not an arbitrary one: it is the spectral measure of some nonnegative function on X. • This allows us to change the set on which the averages are taken, and to eliminate the error term. 24 Thank you for your attention Regular partition of a bottle of ouzo