Introduction
Background and main questions
The study of polynomial ergodic averages originated from the connection between multiple recurrence and arithmetic configurations. Furstenberg’s ergodic proof of Szemerédi’s theorem [14 ] connected multiple recurrence with arithmetic progressions. The polynomial recurrence theorem of Bergelson and Leibman [2 ] extended this connection to polynomial configurations and commuting transformations. A central problem is to characterize the parts of a dynamical system that determine the limits of these averages.
Let ℓ ≥ 3 \ell\ge3 ℓ ≥ 3 , [ ℓ ] : = { 1 , … , ℓ } [\ell] := \{1,\ldots,\ell\} [ ℓ ] := { 1 , … , ℓ } , and let T 1 , … , T ℓ T_1,\ldots,T_\ell T 1 , … , T ℓ be invertible commuting measure-preserving transformations of a standard probability space ( X , X , μ ) (X,\mathcal{X},\mu) ( X , X , μ ) . For integer polynomials p = ( p 1 , … , p ℓ ) \mathbf{p}=(p_1,\ldots,p_\ell) p = ( p 1 , … , p ℓ ) with p i ( 0 ) = 0 p_i(0)=0 p i ( 0 ) = 0 , write
A N p ( f 1 , … , f ℓ ) = 1 N ∑ n = 1 N ∏ i = 1 ℓ T i p i ( n ) f i , f i ∈ L ∞ ( μ ) , (1) A_N^{\mathbf{p}}(f_1,\ldots,f_\ell)=\frac{1}{N}\sum_{n=1}^{N}\prod_{i=1}^{\ell}T_i^{p_i(n)}f_i,\qquad f_i\in L^\infty(\mu), \tag*{(1)} A N p ( f 1 , … , f ℓ ) = N 1 n = 1 ∑ N i = 1 ∏ ℓ T i p i ( n ) f i , f i ∈ L ∞ ( μ ) , (1)
where T f = f ∘ T Tf=f\circ T T f = f ∘ T . For linear and polynomial iterations, Tao [24 ] and Walsh [25 ] proved norm convergence, respectively. To describe limits more precisely, one seeks a collection of factors ( A i ) i = 1 ℓ (\mathcal{A}_i)_{i=1}^{\ell} ( A i ) i = 1 ℓ such that, upon replacing each f i f_i f i with E ( f i ∣ A i ) \mathbb{E}(f_i\mid\mathcal{A}_i) E ( f i ∣ A i ) , the difference in the average tends to zero in L 2 ( μ ) L^2(\mu) L 2 ( μ ) . Such a collection of factors is called a system of characteristic factors.
For a single transformation, the Host–Kra–Ziegler structure theory [16 , 27 ] , combined with the theory of equidistribution of polynomials on nilmanifolds, provides the characteristic factors for polynomial averages [17 , 22 ] .
For several commuting transformations, people wish to use the factors corresponding to each T i T_i T i separately. Chu, Frantzikinakis, and Host established Host–Kra factor control for various transformations, applicable to polynomials with distinct degrees [5 ] . Following the estimates on the seminorm in [6 , 7 ] , Frantzikinakis and Kuca proved that for any family of pairwise linearly independent polynomials, the finite-order Host–Kra factors of the individual transformations form a system of characteristic factors [13 Theorem 2.8] . This theorem does not require ergodicity, and the necessary order depends only on the number of polynomials and their maximum degree. Section 2.4 of the same paper also discusses a version of this result for prime numbers. For a family of linearly independent polynomials, Frantzikinakis and Kra proved that the rational Kronecker factors are characteristic for averages taken over integers for a single ergodic transformation [12 Theorem 1.1] . Frantzikinakis and Kuca extended this result to commuting transformations, covering both averages over integers and averages over prime numbers [13 Theorems 2.10 and 2.13] .
For a family of polynomials that are pairwise linearly independent but collectively satisfy a linear relation, the above results leave two questions open: what is the required degree for each coordinate, and which parts of the corresponding Host–Kra factor determine the limit of the average? The first of these questions already arises in the following average:
A N ( f 1 , f 2 , f 3 ) = 1 N ∑ n = 1 N T 1 n f 1 T 2 n 2 f 2 T 3 n 2 + n f 3 . (2) A_N(f_1,f_2,f_3)=\frac{1}{N}\sum_{n=1}^{N}T_1^n f_1\,T_2^{n^2} f_2\,T_3^{n^2+n} f_3. \tag*{(2)} A N ( f 1 , f 2 , f 3 ) = N 1 n = 1 ∑ N T 1 n f 1 T 2 n 2 f 2 T 3 n 2 + n f 3 . (2)
The family of polynomials t , t 2 , t 2 + t t,t^2,t^2+t t , t 2 , t 2 + t are pairwise linearly independent, but the third polynomial is the sum of the first two. When T 1 = T 2 = T 3 = T T_1=T_2=T_3=T T 1 = T 2 = T 3 = T is ergodic, the Kronecker factor Z 1 ( T ) \mathcal{Z}_1(T) Z 1 ( T ) coincides with the characteristic factor [10 Theorem B] . For general commuting transformations, the existence of a control by a factor of finite order has been established [13 Theorem 2.8] . In [21 Section 6.3, Problem 36] , Kuca asks about the optimal order, noting that it is not even clear whether Z 1 ( T 1 ) , Z 1 ( T 2 ) , Z 1 ( T 3 ) \mathcal{Z}_1(T_1),\mathcal{Z}_1(T_2),\mathcal{Z}_1(T_3) Z 1 ( T 1 ) , Z 1 ( T 2 ) , Z 1 ( T 3 ) are sufficient. This problem examines whether the low-order description in a single transformation case can be extended to several commuting transformations.
The second question is inspired by existing work on the single-transformation case, which reduces the problem to affine systems and factors generated by quasi-eigenfunctions [10 ] . The study of Weyl systems and polynomial complexity [3 , 20 ] relates characteristic orders to affine models and polynomial identities. Even if the Host–Kra complexity is equal to the Weyl complexity, it cannot be concluded that, in a given commutative system, each Host–Kra factor can be replaced with a subfactor generated by polynomial phases for positive powers of the corresponding transformation. We investigate when this substitution holds and allow for linear relationships within a family of polynomials.
In this paper, we answer the first question for (2) : for all commuting systems, the optimal coordinatewise Host–Kra factor orders are 4, 2, 2. Therefore, the first-order conclusion for a single transformation no longer holds in general commuting systems. For a pair of commuting affine transformations on T 5 \mathbb{T}^5 T 5 , one obtains three matching lower bounds simultaneously, with the possibility of setting T 2 = T 3 T_2=T_3 T 2 = T 3 .
The upper bound comes from a more general reduction result, namely, reduction to a factor generated by polynomial phases for positive powers of the corresponding transformation. We prove that as long as the dimension of the space spanned by each nonempty subfamily of polynomials over the rational number field is greater than half its size, these smaller factors constitute a set of characteristic factors. This criterion covers all triples of pairwise linearly independent polynomials and also applies to longer families of polynomials with linear relationships. This reduction and the order 4, 2, 2 hold true for both integer and prime averages. Furthermore, the orders of the phase factors for this three-term polynomial family are optimal.
Phase factors and a rank criterion
Now we define the phase factor and state the result precisely. Let P \mathbb{P} P be the primes and π ( N ) = ∣ P ∩ [ 1 , N ] ∣ \pi(N)=|\mathbb{P}\cap[1,N]| π ( N ) = ∣ P ∩ [ 1 , N ] ∣ . For N ≥ 2 N\ge2 N ≥ 2 , put
A N , P P ( f 1 , … , f ℓ ) = 1 π ( N ) ∑ p ≤ N p ∈ P ∏ i = 1 ℓ T i p i ( p ) f i . (3) A^{\mathbb{P}}_{N,\mathbb{P}}(f_1,\ldots,f_\ell)=\frac{1}{\pi(N)}\sum_{\substack{p\le N\\p\in\mathbb{P}}}\prod_{i=1}^{\ell}T_i^{p_i(p)}f_i. \tag*{(3)} A N , P P ( f 1 , … , f ℓ ) = π ( N ) 1 p ≤ N p ∈ P ∑ i = 1 ∏ ℓ T i p i ( p ) f i . (3)
Let A N , P A_{N,\mathbb{P}} A N , P denote the prime-number version of (1.2). For averages over primes, the definition of the characteristic factor tuple remains the same. By Walsh’s theorem and the comparison results in [11 ] , these averages over primes converge.
We adopt the following conventions: ∥ f ∥ U d + 1 ( T ) = 0 \|f\|_{U^{d+1}(T)}=0 ∥ f ∥ U d + 1 ( T ) = 0 if and only if E ( f ∣ Z d ( T ) ) = 0 \mathbb{E}(f\mid\mathcal{Z}_d(T))=0 E ( f ∣ Z d ( T )) = 0 ; see [16 , 18 ] . For a function ϕ \phi ϕ of modulus one, let Δ T ϕ = ( T ϕ ) ϕ ‾ \Delta_T\phi=(T\phi)\overline{\phi} Δ T ϕ = ( T ϕ ) ϕ , and define its iterates recursively. For d ≥ 1 d\ge1 d ≥ 1 , let
W d ( T ) = σ μ ( { ϕ ∈ L ∞ ( μ ) : ∣ ϕ ∣ = 1 , Δ T q d + 1 ϕ = 1 for some q ≥ 1 } ) , \mathcal{W}_d(T)=\sigma_\mu\left(\left\{\phi\in L^\infty(\mu):|\phi|=1,\ \Delta_{T^q}^{d+1}\phi=1\text{ for some }q\ge1\right\}\right), W d ( T ) = σ μ ( { ϕ ∈ L ∞ ( μ ) : ∣ ϕ ∣ = 1 , Δ T q d + 1 ϕ = 1 for some q ≥ 1 } ) ,
where σ μ \sigma_\mu σ μ denotes the completed sigma-algebra generated by the indicated functions. We call W d ( T ) \mathcal{W}_d(T) W d ( T ) the phase factor of order d d d .
The functions satisfying Δ T q d + 1 ϕ = 1 \Delta_{T^q}^{d+1}\phi=1 Δ T q d + 1 ϕ = 1 are classical quasi-eigenfunctions for positive integer powers of T T T . Abramov established a structural theory for totally ergodic systems with quasi-discrete spectrum [1 ] ; see also [15 ] . For a fixed q q q , following the convention in [15 Section 2.3] , these generating functions are of order at most d d d with respect to T q T^q T q . Equivalently, adopting the notation from [4 Definition 1.13] , W d ( T ) = ⋁ q ≥ 1 Abr < d + 1 ( T q ) \mathcal{W}_d(T)=\bigvee_{q\ge1}\operatorname{Abr}_{<d+1}(T^q) W d ( T ) = ⋁ q ≥ 1 Abr < d + 1 ( T q ) , where the completion is taken. The equivalence of different definitions of quasi-eigenfunction will be reviewed in § 2 . It is important to consider positive integer powers here because, even if a transformation is ergodic, the space of higher-order quasi-eigenfunctions may change upon taking powers [9 Theorem 1.4(i)] .
Lemma 2.4 establishes that W d ( T ) ⊆ Z d ( T ) \mathcal{W}_d(T)\subseteq\mathcal{Z}_d(T) W d ( T ) ⊆ Z d ( T ) , and this inclusion can be strict, even when considering all positive integer powers. For an ergodic translation on the standard Heisenberg nilmanifold, W 2 ( T ) = Z 1 ( T ) ⊊ Z 2 ( T ) \mathcal{W}_2(T)=\mathcal{Z}_1(T)\subsetneq\mathcal{Z}_2(T) W 2 ( T ) = Z 1 ( T ) ⊊ Z 2 ( T ) ; see [10 Propositions 2.1(i), 2.4 and Theorem 2.2] and § 7.2 . Therefore, even if the corresponding Host–Kra seminorm is not zero, the phase factor theorem may still deduce that the average tends to zero.
For J ⊆ [ ℓ ] J\subseteq[\ell] J ⊆ [ ℓ ] , put r p ( J ) = dim Q span Q { p j : j ∈ J } r_p(J)=\dim_{\mathbb{Q}}\operatorname{span}_{\mathbb{Q}}\{p_j:j\in J\} r p ( J ) = dim Q span Q { p j : j ∈ J } . The following criterion can be verified using the coefficient matrices of the respective families of polynomials.
Theorem A. Let D ≥ 1 D\ge1 D ≥ 1 and ℓ ≥ 3 \ell\ge3 ℓ ≥ 3 . There exists an integer s = s ( D , ℓ ) ≥ 1 s=s(D,\ell)\ge1 s = s ( D , ℓ ) ≥ 1 with the following property. Let p 1 , … , p ℓ ∈ Z [ t ] p_1,\ldots,p_\ell\in\mathbb{Z}[t] p 1 , … , p ℓ ∈ Z [ t ] have degree at most D D D and satisfy p i ( 0 ) = 0 p_i(0)=0 p i ( 0 ) = 0 . Assume that
∣ J ∣ < 2 r p ( J ) ( ∅ ≠ J ⊆ [ ℓ ] ) . (4) |J|<2r_p(J)\qquad(\varnothing\ne J\subseteq[\ell]). \tag*{(4)} ∣ J ∣ < 2 r p ( J ) ( ∅ = J ⊆ [ ℓ ]) . (4) For every system of invertible commuting measure-preserving transformations T 1 , … , T ℓ T_1,\ldots,T_\ell T 1 , … , T ℓ , the factors W s ( T 1 ) , … , W s ( T ℓ ) \mathcal{W}_s(T_1),\ldots,\mathcal{W}_s(T_\ell) W s ( T 1 ) , … , W s ( T ℓ ) are characteristic for both (1.1) and (1.3). More precisely, if f i ∈ L ∞ ( μ ) f_i\in L^\infty(\mu) f i ∈ L ∞ ( μ ) and E ( f j ∣ W s ( T j ) ) = 0 \mathbb{E}(f_j\mid\mathcal{W}_s(T_j))=0 E ( f j ∣ W s ( T j )) = 0 for some j ∈ [ ℓ ] j\in[\ell] j ∈ [ ℓ ] , then
∥ A N P ( f 1 , … , f ℓ ) ∥ 2 ⟶ 0 , ∥ A N , P ( f 1 , … , f ℓ ) ∥ 2 ⟶ 0. \|A_N^{\mathbb{P}}(f_1,\ldots,f_\ell)\|_2\longrightarrow0,\qquad
\|A_{N,\mathbb{P}}(f_1,\ldots,f_\ell)\|_2\longrightarrow0. ∥ A N P ( f 1 , … , f ℓ ) ∥ 2 ⟶ 0 , ∥ A N , P ( f 1 , … , f ℓ ) ∥ 2 ⟶ 0.
The rank condition excludes the zero polynomial and implies pairwise linear independence. For three polynomials, it is exactly equivalent to each being pairwise linearly independent. The order s ( D , ℓ ) s(D,\ell) s ( D , ℓ ) is independent of the coefficients, the system, and the functions involved.
At the boundary where equality holds, there is a specific obstruction to the strict inequality. Proposition 7.3 examines p j ( t ) = t + j t 2 p_j(t)=t+jt^2 p j ( t ) = t + j t 2 , 0 ≤ j ≤ 3 0\le j\le3 0 ≤ j ≤ 3 . Every nonempty subfamily satisfies ∣ J ∣ ≤ 2 r p ( J ) |J|\le2r_p(J) ∣ J ∣ ≤ 2 r p ( J ) , with equality occurring only for the full family of four polynomials. On a totally ergodic Heisenberg nilsystem, the same set of four functions are orthogonal to W s ( T ) \mathcal{W}_s(T) W s ( T ) for every s ≥ 1 s\ge1 s ≥ 1 , while their integer average has a non-zero limit. Here, all four transformations are identical. Therefore, increasing the phase order cannot remove this obstacle. This example excludes the possibility of uniformly replacing < < < by ≤ \le ≤ in (1.4).
By Lemma 3.1, (1.4) is equivalent to the following statement: for every i i i , there exist two linearly independent subfamilies that cover the whole family of polynomials and intersect only at p i p_i p i . This can be derived from the finite vector form of Edmonds’ matroid partition theorem [8 ] , with a convenient sufficient condition.
Corollary 1.1. Let D ≥ 1 D \ge1 D ≥ 1 and ℓ ≥ 3 \ell\ge3 ℓ ≥ 3 , and let s = s ( D , ℓ ) s=s(D,\ell) s = s ( D , ℓ ) be as in Theorem A. If p 1 , … , p ℓ ∈ Z [ t ] p_1,\ldots,p_\ell\in\mathbb{Z}[t] p 1 , … , p ℓ ∈ Z [ t ] have degree at most D D D , zero constant terms, and every ⌊ ℓ / 2 ⌋ + 1 \lfloor\ell/2 \rfloor+1 ⌊ ℓ /2 ⌋ + 1 of them are linearly independent over Q \mathbb{Q} Q , then W s ( T 1 ) , … , W s ( T ℓ ) \mathcal{W}_s(T_1),\ldots,\mathcal{W}_s(T_\ell) W s ( T 1 ) , … , W s ( T ℓ ) are characteristic for both (1.1) and (1.3). In particular, it suffices that every ℓ − 1 \ell-1 ℓ − 1 polynomials are linearly independent.
Proof. For each i i i , divide [ ℓ ] ∖ { i } [\ell]\setminus\{i\} [ ℓ ] ∖ { i } into two sets whose sizes differ by at most one, and add i i i to these two sets. The size of each resulting set is at most ⌊ ℓ / 2 ⌋ + 1 \lfloor\ell/2 \rfloor+1 ⌊ ℓ /2 ⌋ + 1 , so the corresponding subfamilies are linearly independent. Applying Lemma 3.1 and Theorem A yields the conclusion. The last assertion follows from ⌊ ℓ / 2 ⌋ + 1 ≤ ℓ − 1 \lfloor\ell/2 \rfloor+1\le\ell-1 ⌊ ℓ /2 ⌋ + 1 ≤ ℓ − 1 .
Optimal orders for three iterates
For t , t 2 , t 2 + t t,t^2,t^2+t t , t 2 , t 2 + t , by Theorem A one can reduce the average to phase factors of some finite order. Subsequently, recurrence relations and polynomial identities give the exact order.
Theorem B. Let T 1 , T 2 , T 3 T_1,T_2,T_3 T 1 , T 2 , T 3 be invertible commuting measure-preserving transformations of a standard probability space ( X , X , μ ) (X,\mathcal{X},\mu) ( X , X , μ ) . Then the factors W 4 ( T 1 ) , W 2 ( T 2 ) , W 2 ( T 3 ) \mathcal{W}_4(T_1),\mathcal{W}_2(T_2),\mathcal{W}_2(T_3) W 4 ( T 1 ) , W 2 ( T 2 ) , W 2 ( T 3 ) are characteristic for both A N A_N A N and A N , P A_{N,\mathbb{P}} A N , P . More precisely, for any f 1 , f 2 , f 3 ∈ L ∞ ( μ ) f_1,f_2,f_3\in L^\infty(\mu) f 1 , f 2 , f 3 ∈ L ∞ ( μ ) , if
E ( f 1 ∣ W 4 ( T 1 ) ) = 0 , or E ( f 2 ∣ W 2 ( T 2 ) ) = 0 , or E ( f 3 ∣ W 2 ( T 3 ) ) = 0 , \mathbb{E}(f_1\mid\mathcal{W}_4(T_1))=0,\qquad\text{or}\qquad\mathbb{E}(f_2\mid\mathcal{W}_2(T_2))=0,\qquad\text{or}\qquad\mathbb{E}(f_3\mid\mathcal{W}_2(T_3))=0, E ( f 1 ∣ W 4 ( T 1 )) = 0 , or E ( f 2 ∣ W 2 ( T 2 )) = 0 , or E ( f 3 ∣ W 2 ( T 3 )) = 0 , then both averages tend to zero in L 2 ( μ ) L^2(\mu) L 2 ( μ ) . For both averaging schemes if the conclusion is required to hold for all such systems, then none of the three orders can be reduced even if the constraint T 2 = T 3 T_2=T_3 T 2 = T 3 is imposed.
Corollary 1.2. Under the same system assumptions, the factors Z 4 ( T 1 ) , Z 2 ( T 2 ) , Z 2 ( T 3 ) \mathcal{Z}_4(T_1),\mathcal{Z}_2(T_2),\mathcal{Z}_2(T_3) Z 4 ( T 1 ) , Z 2 ( T 2 ) , Z 2 ( T 3 ) are characteristic for both A N A_N A N and A N , P A_{N,\mathbb{P}} A N , P . In particular,
∥ f 1 ∥ U 5 ( T 1 ) = 0 or ∥ f 2 ∥ U 3 ( T 2 ) = 0 or ∥ f 3 ∥ U 3 ( T 3 ) = 0 \lVert f_1\rVert_{U^5(T_1)}=0\qquad\text{or}\qquad\lVert f_2\rVert_{U^3(T_2)}=0\qquad\text{or}\qquad\lVert f_3\rVert_{U^3(T_3)}=0 ∥ f 1 ∥ U 5 ( T 1 ) = 0 or ∥ f 2 ∥ U 3 ( T 2 ) = 0 or ∥ f 3 ∥ U 3 ( T 3 ) = 0 implies that both averages tend to zero in L 2 ( μ ) L^2(\mu) L 2 ( μ ) . For both averaging schemes if the conclusion is required to hold for all such systems, then none of the three orders can be reduced—even if the constraint T 2 = T 3 T_2=T_3 T 2 = T 3 is imposed.
The upper bound of this corollary is given by W d ( T ) ⊆ Z d ( T ) \mathcal{W}_d(T)\subseteq\mathcal{Z}_d(T) W d ( T ) ⊆ Z d ( T ) . For the lower bounds, Proposition 5.1 gives commuting affine transformations on T 5 \mathbb{T}^5 T 5 , with T 2 = T 3 T_2=T_3 T 2 = T 3 along with functions of modulus one that satisfy
∥ f 1 ∥ U 4 ( T 1 ) = 0 , ∥ f 2 ∥ U 2 ( T 2 ) = ∥ f 3 ∥ U 2 ( T 3 ) = 0 , \lVert f_1\rVert_{U^4(T_1)}=0,\qquad\lVert f_2\rVert_{U^2(T_2)}=\lVert f_3\rVert_{U^2(T_3)}=0, ∥ f 1 ∥ U 4 ( T 1 ) = 0 , ∥ f 2 ∥ U 2 ( T 2 ) = ∥ f 3 ∥ U 2 ( T 3 ) = 0 ,
Furthermore, the summand in (1.2) remains the same function of modulus one for every n n n . Consequently, even without imposing restrictions on the factors associated with the other two coordinates, none of the three factor orders can be reduced. Within the hierarchies of Host–Kra factors and phase factors, these orders cannot be lowered if the conclusion is required to hold for all commuting systems. In particular, the minimal common Host–Kra order is four.
Outline of the proofs
For the polynomial average of integer parameters, by [13 Theorem 2.8] , each f i f_i f i can be replaced by E ( f i ∣ Z s ( T i ) ) \mathbb{E}(f_i\mid\mathcal{Z}_s(T_i)) E ( f i ∣ Z s ( T i )) without changing the L 2 L^2 L 2 limit of the average. Applying [19 Theorem 1.2] , we approximate these functions in L 2 L^2 L 2 on s s s -step nilsystem bundles. Since the starting points π 1 ( x ) , … , π ℓ ( x ) \pi_1(x),\ldots,\pi_\ell(x) π 1 ( x ) , … , π ℓ ( x ) all depend on the same x x x , we may study the polynomial orbits originating from the point ( π 1 ( x ) , … , π ℓ ( x ) ) (\pi_1(x),\ldots,\pi_\ell(x)) ( π 1 ( x ) , … , π ℓ ( x )) .
By Leibman’s theorem [23 ] , the limiting orbit measure is a finite average of Haar measures on compact orbits H z 1 , … , H z r Hz_1,\ldots,Hz_r H z 1 , … , H z r , where H ≤ ∏ i G i ∘ H\le\prod_i G_i^\circ H ≤ ∏ i G i ∘ is a connected subgroup. For each i i i , the rank condition guarantees the existence of two linearly independent subfamilies of polynomials, both of which contain p i p_i p i , and their union contains the whole family of polynomials. Using [13 Theorem 2.10] , we show that the projection of H H H onto the two corresponding products of groups are both surjective. By calculating the commutator for each coordinate, we obtain ∏ i [ G i ∘ , G i ∘ ] ⊆ H \prod_i[G_i^\circ,G_i^\circ]\subseteq H ∏ i [ G i ∘ , G i ∘ ] ⊆ H . Therefore, we can take the commutator quotient for each coordinate, and then use the torus characters to obtain the polynomial phase for suitable positive integer powers of each transformation.
For the family of polynomials t , t 2 , t 2 + t t,t^2,t^2+t t , t 2 , t 2 + t , we use Weyl’s theorem, recurrence properties, and interpolation to show that the degrees of the three phase polynomials modulo Q \mathbb{Q} Q are at most 4 4 4 , 2 2 2 , and 2 2 2 , respectively, at almost every point where the average along a fixed residue class has a nonzero limit. By Lemma 4.2 , these degree bounds modulo Q \mathbb{Q} Q give the characteristic factors W 4 ( T 1 ) W_4(T_1) W 4 ( T 1 ) , W 2 ( T 2 ) W_2(T_2) W 2 ( T 2 ) , and W 2 ( T 3 ) W_2(T_3) W 2 ( T 3 ) . Finally, using an example involving commuting affine transformations, it is shown that none of these three orders can be reduced.
For polynomial averages over prime numbers, using [11 Proposition 3.6] , the proof can be reduced to the case of integer averages over arithmetic progressions. Since the restriction to an arithmetic sequence does not change the rank of any family of polynomials, in general we still only need the order s ( D , ℓ ) s(\mathcal{D},\ell) s ( D , ℓ ) . For t , t 2 , t 2 + t t,t^2,t^2+t t , t 2 , t 2 + t , by constructing a finite cyclic extension, we can choose terms in each residue class, thus obtaining the same upper bound of order 4 , 2 , 2 4,2,2 4 , 2 , 2 . Furthermore, since the term being averaged in the aforementioned affine transformation example is independent of n n n , this example also proves the optimality of the three orders in the case of prime averaging.
Organization
In Section 2 2 2 , we recall some definitions and related results which will be used later. In Section 3 3 3 , we prove the general integer theorem. Sections 4 4 4 proves the upper bounds 4 , 2 , 2 4,2,2 4 , 2 , 2 , and Section 5 5 5 proves their optimality. In Section 6 6 6 we treat prime averages and the intersection limits. In Section 7 7 7 we give further bounds and examples, including the obstruction at the rank boundary.
AI-use disclosure
AI-assisted tools were used to help extend the results for polynomial averages over the integers to averages over the primes. The paper was written by the author. The author has verified all mathematical arguments and takes full responsibility for the correctness of all proofs in the paper.
Acknowledgments
The author is grateful to his advisor, Professor Siming Tu, for introducing him to the subject and for providing continued guidance, support, and encouragement.
Preliminaries
We write N = { 1 , 2 , … } \mathbb{N}=\{1,2,\ldots\} N = { 1 , 2 , … } , T = R / Z \mathbb{T}=\mathbb{R}/\mathbb{Z} T = R / Z , and S 1 = { z ∈ C : ∣ z ∣ = 1 } \mathbb{S}^1=\{z\in\mathbb{C}:|z|=1\} S 1 = { z ∈ C : ∣ z ∣ = 1 } , and put e ( t ) = exp ( 2 π i t ) e(t)=\exp(2\pi it) e ( t ) = exp ( 2 π i t ) and μ ∞ = { { e ( t ) : t ∈ Q } } \mu_\infty=\{\{e(t):t\in\mathbb{Q}\}\} μ ∞ = {{ e ( t ) : t ∈ Q }} . For a probability measure ν \nu ν , let L p ( A , ν ) L^p(\mathcal{A},\nu) L p ( A , ν ) denote the space of all A \mathcal{A} A -measurable functions in L p ( ν ) L^p(\nu) L p ( ν ) . When the measure is clear from the context, we use the shorthand notation ∥ f ∥ p \lVert f\rVert_p ∥ f ∥ p and L p ( A ) L^p(\mathcal{A}) L p ( A ) . We denote the conditional expectation by E ( f ∣ A ) \mathbb{E}(f\mid\mathcal{A}) E ( f ∣ A ) ; in L 2 ( ν ) L^2(\nu) L 2 ( ν ) , this corresponds to the orthogonal projection onto L 2 ( A , ν ) L^2(\mathcal{A},\nu) L 2 ( A , ν ) . The inner product is linear in its first argument. Let L ∞ ( ν ) L^\infty(\nu) L ∞ ( ν ) be the space of essentially bounded measurable functions, with the norm ∥ f ∥ ∞ = ess sup ν ∣ f ∣ \lVert f\rVert_\infty=\operatorname{ess\,sup}_\nu|f| ∥ f ∥ ∞ = ess sup ν ∣ f ∣ . For a locally compact space M M M , let B ( M ) \mathcal{B}(M) B ( M ) denote its Borel σ \sigma σ -algebra, C ( M ) C(M) C ( M ) the space of complex-valued continuous functions on M M M , and C c ( M ) C_c(M) C c ( M ) the subspace consisting of functions with compact support. For a Borel probability measure μ \mu μ on M M M , define supp μ = { x ∈ M : μ ( U ) > 0 for every open neighbourhood U of x } \operatorname{supp}\mu=\{x\in M:\mu(U)>0\text{ for every open neighbourhood }U\text{ of }x\} supp μ = { x ∈ M : μ ( U ) > 0 for every open neighbourhood U of x } . We say that μ \mu μ has full support if supp μ = M \operatorname{supp}\mu=M supp μ = M .
Measure-preserving systems and factors
A measure-preserving system is a quadruple ( X , X , μ , T ) (X,\mathcal{X},\mu,T) ( X , X , μ , T ) in which ( X , X , μ ) (X,\mathcal{X},\mu) ( X , X , μ ) is a standard probability space and T : X → X T:X\to X T : X → X is an invertible measure-preserving transformation. We write ( X , μ , T ) (X,\mu,T) ( X , μ , T ) when the sigma-algebra is understood. We use the convention T f = f ∘ T Tf=f\circ T T f = f ∘ T . All sets, functions, and maps are measurable. In the standard Borel model, functions and maps are Borel-measurable. Equalities, inequalities, and invariance are understood modulo null sets unless explicitly stated pointwise.
A factor of ( X , X , μ , T ) (X,\mathcal{X},\mu,T) ( X , X , μ , T ) is a completed T T T -invariant sub-sigma-algebra of X \mathcal{X} X . A factor map π : ( X , μ , T ) → ( Y , ν , S ) \pi:(X,\mu,T)\to(Y,\nu,S) π : ( X , μ , T ) → ( Y , ν , S ) satisfies π ∗ μ = ν \pi_*\mu=\nu π ∗ μ = ν and π ∘ T = S ∘ π \pi\circ T=S\circ\pi π ∘ T = S ∘ π . The corresponding factor sigma-algebra on X X X is the completion of π − 1 Y \pi^{-1}\mathcal{Y} π − 1 Y , where Y \mathcal{Y} Y is the sigma-algebra of Y Y Y . We identify functions on Y Y Y with their pullbacks to X X X . For a family F \mathcal{F} F of measurable sets or functions, σ ν ( F ) \sigma_\nu(\mathcal{F}) σ ν ( F ) denotes the ν \nu ν -completion of the sigma-algebra generated by F \mathcal{F} F . All joins of sigma-algebras are understood to be completed.
Let I ( T ) \mathcal{I}(T) I ( T ) denote the sigma-algebra of sets A ∈ X A\in\mathcal{X} A ∈ X such that μ ( T − 1 A △ A ) = 0 \mu(T^{-1}A\mathbin{\triangle} A)=0 μ ( T − 1 A △ A ) = 0 . The transformation T T T is ergodic if every A ∈ I ( T ) A\in\mathcal{I}(T) A ∈ I ( T ) has measure zero or one. It is totally ergodic if T q T^q T q is ergodic for every integer q ≥ 1 q\ge1 q ≥ 1 . The rational Kronecker factor of T T T is
K r a t ( T ) = ⋁ m ≥ 1 I ( T m ) = ⋁ j ≥ 1 I ( T j ) . \mathcal{K}_{\mathrm{rat}}(T)=\bigvee_{m\ge1}\mathcal{I}(T^m)=\bigvee_{j\ge1}\mathcal{I}(T^j). K rat ( T ) = m ≥ 1 ⋁ I ( T m ) = j ≥ 1 ⋁ I ( T j ) .
Host–Kra seminorms and phase factors
Let ( X , X , μ , T ) (X,\mathcal{X},\mu,T) ( X , X , μ , T ) be a measure-preserving system. For f ∈ L ∞ ( μ ) f\in L^\infty(\mu) f ∈ L ∞ ( μ ) , write Δ T f = ( T f ) f ‾ \Delta_Tf=(Tf)\overline{f} Δ T f = ( T f ) f , Δ T 0 f = f \Delta_T^0f=f Δ T 0 f = f , and Δ T k + 1 f = Δ T ( Δ T k f ) \Delta_T^{k+1}f=\Delta_T(\Delta_T^kf) Δ T k + 1 f = Δ T ( Δ T k f ) . The Host–Kra seminorms satisfy
∥ f ∥ U 1 ( T ) = ∥ E ( f ∣ I ( T ) ) ∥ 2 , ∥ f ∥ U r + 1 ( T ) 2 r + 1 = lim H → ∞ 1 H ∑ h = 1 H ∥ Δ T h f ∥ U r ( T ) 2 r . \|f\|_{U^1(T)}=\|\mathbb{E}(f\mid\mathcal{I}(T))\|_2,\qquad\|f\|_{U^{r+1}(T)}^{2^{r+1}}=\lim_{H\to\infty}\frac{1}{H}\sum_{h=1}^{H}\|\Delta_{T^h}f\|_{U^r(T)}^{2^r}. ∥ f ∥ U 1 ( T ) = ∥ E ( f ∣ I ( T )) ∥ 2 , ∥ f ∥ U r + 1 ( T ) 2 r + 1 = H → ∞ lim H 1 h = 1 ∑ H ∥ Δ T h f ∥ U r ( T ) 2 r .
We use their characterization of Z d ( T ) \mathcal{Z}_d(T) Z d ( T ) ; see [16 , 18 ] and [13 Section 3.2.1] for details. The mean ergodic theorem gives
lim H → ∞ 1 H ∑ h = 1 H ∫ Δ T h f d μ = ∫ ∣ E ( f ∣ I ( T ) ) ∣ 2 d μ = ∥ f ∥ U 1 ( T ) 2 . \lim_{H\to\infty}\frac{1}{H}\sum_{h=1}^{H}\int\Delta_{T^h}f\,d\mu=\int|\mathbb{E}(f\mid\mathcal{I}(T))|^2\,d\mu=\|f\|_{U^1(T)}^2. H → ∞ lim H 1 h = 1 ∑ H ∫ Δ T h f d μ = ∫ ∣ E ( f ∣ I ( T )) ∣ 2 d μ = ∥ f ∥ U 1 ( T ) 2 .
By induction,
∥ f ∥ U r ( T ) 2 r = lim H 1 → ∞ 1 H 1 ∑ h 1 = 1 H 1 ⋯ lim H r → ∞ 1 H r ∑ h r = 1 H r ∫ Δ T h r ⋯ Δ T h 1 f d μ . (5) \|f\|_{U^r(T)}^{2^r}=\lim_{H_1\to\infty}\frac{1}{H_1}\sum_{h_1=1}^{H_1}\cdots\lim_{H_r\to\infty}\frac{1}{H_r}\sum_{h_r=1}^{H_r}\int\Delta_{T^{h_r}}\cdots\Delta_{T^{h_1}}f\,d\mu. \tag*{(5)} ∥ f ∥ U r ( T ) 2 r = H 1 → ∞ lim H 1 1 h 1 = 1 ∑ H 1 ⋯ H r → ∞ lim H r 1 h r = 1 ∑ H r ∫ Δ T h r ⋯ Δ T h 1 f d μ . (5)
For q ≥ 1 q\ge1 q ≥ 1 and r ≥ 2 r\ge2 r ≥ 2 ,
∥ f ∥ U r ( T ) ≤ ∥ f ∥ U r ( T q ) ≤ q r / 2 r ∥ f ∥ U r ( T ) ; \|f\|_{U^r(T)}\le\|f\|_{U^r(T^q)}\le q^{r/2^r}\|f\|_{U^r(T)}; ∥ f ∥ U r ( T ) ≤ ∥ f ∥ U r ( T q ) ≤ q r / 2 r ∥ f ∥ U r ( T ) ;
see [13 Lemma 3.1] . Moreover, for any d ≥ 1 d\ge1 d ≥ 1 , Z d ( T q ) = Z d ( T ) \mathcal{Z}_d(T^q)=\mathcal{Z}_d(T) Z d ( T q ) = Z d ( T ) .
Definition 2.1. Let ( X , X , μ , T ) (X,\mathcal{X},\mu,T) ( X , X , μ , T ) be a measure-preserving system and let d ≥ 1 d\ge1 d ≥ 1 be an integer. Define
P d ( T ) = { ϕ ∈ L ∞ ( μ ) : ∣ ϕ ∣ = 1 , Δ T q d + 1 ϕ = 1 for some integer q ≥ 1 } . \mathcal{P}_d(T)=\left\{\phi\in L^\infty(\mu):|\phi|=1,\ \Delta_{T^q}^{d+1}\phi=1\text{ for some integer }q\ge1\right\}. P d ( T ) = { ϕ ∈ L ∞ ( μ ) : ∣ ϕ ∣ = 1 , Δ T q d + 1 ϕ = 1 for some integer q ≥ 1 } . The phase factor for finite powers of order d d d is
W d ( T ) = σ μ ( P d ( T ) ) . \mathcal{W}_d(T)=\sigma_\mu(\mathcal{P}_d(T)). W d ( T ) = σ μ ( P d ( T )) . Denote A d ( T ) = span C P d ( T ) \mathcal{A}_d(T)=\operatorname{span}_{\mathbb{C}}\mathcal{P}_d(T) A d ( T ) = span C P d ( T ) .
We next recall Newton’s formula for functions satisfying Δ S L ϕ = 1 \Delta_S^L\phi=1 Δ S L ϕ = 1 . For a ∈ R a\in\mathbb{R} a ∈ R , write [ a ] = a + Q ∈ R / Q [a]=a+\mathbb{Q}\in\mathbb{R}/\mathbb{Q} [ a ] = a + Q ∈ R / Q . For P ( t ) = ∑ k a k t k ∈ R [ t ] P(t)=\sum_k a_kt^k\in\mathbb{R}[t] P ( t ) = ∑ k a k t k ∈ R [ t ] , put [ P ] ( t ) = ∑ k [ a k ] t k [P](t)=\sum_k[a_k]t^k [ P ] ( t ) = ∑ k [ a k ] t k . Here R / Q \mathbb{R}/\mathbb{Q} R / Q is regarded as a vector space over Q \mathbb{Q} Q , and we set deg 0 = − ∞ \deg0=-\infty deg 0 = − ∞ . This notation will also be used in Section 4.
Lemma 2.2. Let ( X , X , μ , T ) (X,\mathcal{X},\mu,T) ( X , X , μ , T ) be a measure-preserving system, let L ≥ 1 L\ge1 L ≥ 1 be an integer, and let ϕ ∈ L ∞ ( μ ) \phi\in L^\infty(\mu) ϕ ∈ L ∞ ( μ ) satisfy ∣ ϕ ∣ = 1 |\phi|=1 ∣ ϕ ∣ = 1 and Δ T L ϕ = 1 \Delta_T^L\phi=1 Δ T L ϕ = 1 . Then there are a measurable T T T -invariant set X T , ϕ X_{T,\phi} X T , ϕ with μ ( X T , ϕ ) = 1 \mu(X_{T,\phi})=1 μ ( X T , ϕ ) = 1 and real polynomials P T , ϕ , x P_{T,\phi,x} P T , ϕ , x of degree less than L L L , with coefficients measurable in x x x , such that the following hold.
(i) For every x ∈ X T , ϕ x\in X_{T,\phi} x ∈ X T , ϕ and n ∈ Z n\in\mathbb{Z} n ∈ Z ,
ϕ ( T n x ) = ∏ k = 0 L − 1 ( Δ T k ϕ ( x ) ) ( n k ) = e ( P T , ϕ , x ( n ) ) . \phi(T^n x)=\prod_{k=0}^{L-1}\left(\Delta_T^k\phi(x)\right)^{\binom{n}{k}}=e(P_{T,\phi,x}(n)). ϕ ( T n x ) = k = 0 ∏ L − 1 ( Δ T k ϕ ( x ) ) ( k n ) = e ( P T , ϕ , x ( n )) .
(ii) For every x ∈ X T , ϕ x \in X_{T,\phi} x ∈ X T , ϕ , h ∈ Z h \in\mathbb{Z} h ∈ Z , and integer d ≥ 0 d \ge0 d ≥ 0 , we have [ P T , ϕ , T h x ] ( t ) = [ P T , ϕ , x ] ( t + h ) [P_{T,\phi,T^h x}](t) = [P_{T,\phi,x}](t+h) [ P T , ϕ , T h x ] ( t ) = [ P T , ϕ , x ] ( t + h ) and
deg [ P T , ϕ , x ] ≤ d ⟺ Δ T k ϕ ( x ) ∈ μ ∞ for every d < k < L . \deg[P_{T,\phi,x}] \le d \quad\Longleftrightarrow\quad\Delta_T^k\phi(x) \in\mu_\infty\text{ for every } d < k < L. deg [ P T , ϕ , x ] ≤ d ⟺ Δ T k ϕ ( x ) ∈ μ ∞ for every d < k < L .
If d ≥ L − 1 d \ge L-1 d ≥ L − 1 , both sides of the equivalence hold automatically.
Proof. Set ψ k = Δ T k ϕ \psi_k = \Delta_T^k\phi ψ k = Δ T k ϕ for 0 ≤ k ≤ L 0 \le k \le L 0 ≤ k ≤ L , and put
X T , ϕ = ⋂ n ∈ Z T − n { x : ∣ ϕ ( x ) ∣ = 1 , ψ L ( x ) = 1 } . X_{T,\phi} = \bigcap_{n\in\mathbb{Z}} T^{-n}\{x : |\phi(x)| = 1,\ \psi_L(x) = 1\}. X T , ϕ = n ∈ Z ⋂ T − n { x : ∣ ϕ ( x ) ∣ = 1 , ψ L ( x ) = 1 } .
Then X T , ϕ X_{T,\phi} X T , ϕ is T T T -invariant and μ ( X T , ϕ ) = 1 \mu(X_{T,\phi}) = 1 μ ( X T , ϕ ) = 1 . For every x ∈ X T , ϕ x \in X_{T,\phi} x ∈ X T , ϕ and 0 ≤ k < L 0 \le k < L 0 ≤ k < L , we have
∣ ψ k ( x ) ∣ = 1 , ψ k ( T x ) = ψ k ( x ) ψ k + 1 ( x ) , ψ L ( x ) = 1. |\psi_k(x)| = 1,\qquad\psi_k(Tx) = \psi_k(x)\psi_{k+1}(x),\qquad\psi_L(x) = 1. ∣ ψ k ( x ) ∣ = 1 , ψ k ( T x ) = ψ k ( x ) ψ k + 1 ( x ) , ψ L ( x ) = 1.
Choose measurable a k ( x ) ∈ [ 0 , 1 ) a_k(x) \in[0,1) a k ( x ) ∈ [ 0 , 1 ) with e ( a k ( x ) ) = ψ k ( x ) e(a_k(x)) = \psi_k(x) e ( a k ( x )) = ψ k ( x ) on X T , ϕ X_{T,\phi} X T , ϕ , and set a k = 0 a_k = 0 a k = 0 elsewhere. Define
P T , ϕ , x ( t ) = ∑ k = 0 L − 1 a k ( x ) ( t k ) , ( t 0 ) = 1 , ( t k ) = t ( t − 1 ) ⋯ ( t − k + 1 ) k ! ( k ≥ 1 ) . P_{T,\phi,x}(t) = \sum_{k=0}^{L-1} a_k(x)\binom{t}{k},\qquad\binom{t}{0}=1,\qquad\binom{t}{k}=\frac{t(t-1)\cdots(t-k+1)}{k!}\quad(k\ge1). P T , ϕ , x ( t ) = k = 0 ∑ L − 1 a k ( x ) ( k t ) , ( 0 t ) = 1 , ( k t ) = k ! t ( t − 1 ) ⋯ ( t − k + 1 ) ( k ≥ 1 ) .
These polynomials have degree less than L L L and measurable coefficients. The formula below also follows from [15 (2.1)–(2.2)] . Here, for every n ∈ Z n \in\mathbb{Z} n ∈ Z , by Pascal's identity we have
∏ k = 0 L − 1 ψ k ( T x ) ( n k ) = ∏ k = 0 L − 1 ( ψ k ( x ) ψ k + 1 ( x ) ) ( n k ) = ∏ k = 0 L − 1 ψ k ( x ) ( n + 1 k ) . \prod_{k=0}^{L-1}\psi_k(Tx)^{\binom{n}{k}}
=
\prod_{k=0}^{L-1}\bigl(\psi_k(x)\psi_{k+1}(x)\bigr)^{\binom{n}{k}}
=
\prod_{k=0}^{L-1}\psi_k(x)^{\binom{n+1}{k}}. k = 0 ∏ L − 1 ψ k ( T x ) ( k n ) = k = 0 ∏ L − 1 ( ψ k ( x ) ψ k + 1 ( x ) ) ( k n ) = k = 0 ∏ L − 1 ψ k ( x ) ( k n + 1 ) .
The product at n = 0 n=0 n = 0 is ϕ ( x ) \phi(x) ϕ ( x ) . Since T T T is invertible, by induction one can prove (i) for all n ∈ Z n \in\mathbb{Z} n ∈ Z .
For x ∈ X T , ϕ x \in X_{T,\phi} x ∈ X T , ϕ and h , n ∈ Z h,n \in\mathbb{Z} h , n ∈ Z , part (i) gives
e ( P T , ϕ , T h x ( n ) ) = ϕ ( T n + h x ) = e ( P T , ϕ , x ( n + h ) ) . e(P_{T,\phi,T^h x}(n)) = \phi(T^{n+h}x) = e(P_{T,\phi,x}(n+h)). e ( P T , ϕ , T h x ( n )) = ϕ ( T n + h x ) = e ( P T , ϕ , x ( n + h )) .
Since the degree of the polynomial P T , ϕ , T h x ( t ) − P T , ϕ , x ( t + h ) P_{T,\phi,T^h x}(t)-P_{T,\phi,x}(t+h) P T , ϕ , T h x ( t ) − P T , ϕ , x ( t + h ) is less than L L L and the polynomial takes integer values at 0 , 1 , … , L − 1 0,1,\ldots,L-1 0 , 1 , … , L − 1 , Lagrange interpolation shows that all its coefficients are rational. This proves [ P T , ϕ , T h x ] ( t ) = [ P T , ϕ , x ] ( t + h ) [P_{T,\phi,T^h x}](t) = [P_{T,\phi,x}](t+h) [ P T , ϕ , T h x ] ( t ) = [ P T , ϕ , x ] ( t + h ) . For the remaining part, use
[ P T , ϕ , x ] ( t ) = ∑ k = 0 L − 1 [ a k ( x ) ] ( t k ) , [ a k ( x ) ] = 0 ⟺ ψ k ( x ) ∈ μ ∞ . [P_{T,\phi,x}](t) = \sum_{k=0}^{L-1}[a_k(x)]\binom{t}{k},\qquad[a_k(x)] = 0 \quad\Longleftrightarrow\quad\psi_k(x) \in\mu_\infty. [ P T , ϕ , x ] ( t ) = k = 0 ∑ L − 1 [ a k ( x )] ( k t ) , [ a k ( x )] = 0 ⟺ ψ k ( x ) ∈ μ ∞ .
If [ a k ( x ) ] = 0 [a_k(x)] = 0 [ a k ( x )] = 0 for all d < k < L d < k < L d < k < L , then deg [ P T , ϕ , x ] ≤ d \deg[P_{T,\phi,x}] \le d deg [ P T , ϕ , x ] ≤ d .
Conversely, assume that [ a k ( x ) ] ≠ 0 [a_k(x)] \ne0 [ a k ( x )] = 0 for some k > d k>d k > d . Take the largest such k k k . Since ( t k ) \binom{t}{k} ( k t ) has leading coefficient 1 / k ! 1/k! 1/ k ! , the coefficient of t k t^k t k in [ P T , ϕ , x ] [P_{T,\phi,x}] [ P T , ϕ , x ] is [ a k ( x ) ] / k ! ≠ 0 [a_k(x)]/k! \ne0 [ a k ( x )] / k ! = 0 . Hence
deg [ P T , ϕ , x ] ≥ k > d . \deg[P_{T,\phi,x}] \ge k > d. deg [ P T , ϕ , x ] ≥ k > d .
□
Remark 2.3. For a measure-preserving system ( X , X , μ , T ) (X,\mathcal{X},\mu,T) ( X , X , μ , T ) , define the Abramov factor
Abr < d + 1 ( T ) = σ μ { ϕ ∈ L ∞ ( μ ) : ∣ ϕ ∣ = 1 , Δ T d + 1 ϕ = 1 } . \operatorname{Abr}_{<d+1}(T)=\sigma_\mu\{\phi\in L^\infty(\mu):|\phi|=1,\ \Delta_T^{d+1}\phi=1\}. Abr < d + 1 ( T ) = σ μ { ϕ ∈ L ∞ ( μ ) : ∣ ϕ ∣ = 1 , Δ T d + 1 ϕ = 1 } .
This agrees with the mixed-difference definition in [4 Definition 1.13] . Indeed, Lemma 2.2 shows that Δ T d + 1 ϕ = 1 \Delta_T^{d+1}\phi=1 Δ T d + 1 ϕ = 1 implies Δ T n 1 ⋯ Δ T n d + 1 ϕ = 1 \Delta_{T^{n_1}}\cdots\Delta_{T^{n_{d+1}}}\phi=1 Δ T n 1 ⋯ Δ T n d + 1 ϕ = 1 for all n 1 , … , n d + 1 ∈ Z n_1,\ldots,n_{d+1}\in\mathbb{Z} n 1 , … , n d + 1 ∈ Z . The converse follows by taking n 1 = ⋯ = n d + 1 = 1 n_1=\cdots=n_{d+1}=1 n 1 = ⋯ = n d + 1 = 1 .
The phase factor W d ( T ) \mathcal{W}_d(T) W d ( T ) allows the same difference identity to hold for a positive power of T T T , with the power depending on the generating function. Thus
W d ( T ) = ⋁ q ≥ 1 Abr < d + 1 ( T q ) . \mathcal{W}_d(T)=\bigvee_{q\ge1}\operatorname{Abr}_{<d+1}(T^q). W d ( T ) = q ≥ 1 ⋁ Abr < d + 1 ( T q ) .
In particular, Abr < d + 1 ( T ) ⊆ W d ( T ) \operatorname{Abr}_{<d+1}(T)\subseteq\mathcal{W}_d(T) Abr < d + 1 ( T ) ⊆ W d ( T ) .
The following lemma gives the properties of W d ( T ) \mathcal{W}_d(T) W d ( T ) needed in the proofs below.
Lemma 2.4. Let ( X , X , μ , T ) (X,\mathcal{X},\mu,T) ( X , X , μ , T ) be a measure-preserving system and let d ≥ 1 d \ge1 d ≥ 1 be an integer. Then the following hold.
(i) If S S S is an invertible measure-preserving transformation of ( X , X , μ ) (X,\mathcal{X},\mu) ( X , X , μ ) commuting with T T T , then S − 1 W d ( T ) = W d ( T ) S^{-1}\mathcal{W}_d(T)=\mathcal{W}_d(T) S − 1 W d ( T ) = W d ( T ) . In particular, the sigma-algebra W d ( T ) \mathcal{W}_d(T) W d ( T ) is T T T -invariant.
(ii) For every q ≥ 1 q \ge1 q ≥ 1 ,
W d ( T q ) = W d ( T ) , I ( T q ) ⊆ W d ( T ) ⊆ Z d ( T ) . \mathcal{W}_d(T^q)=\mathcal{W}_d(T), \qquad\mathcal{I}(T^q)\subseteq\mathcal{W}_d(T)\subseteq\mathcal{Z}_d(T). W d ( T q ) = W d ( T ) , I ( T q ) ⊆ W d ( T ) ⊆ Z d ( T ) . (iii) For every f ∈ L 2 ( W d ( T ) ) f\in L^2(\mathcal{W}_d(T)) f ∈ L 2 ( W d ( T )) and ε > 0 \varepsilon>0 ε > 0 , there exists U ∈ A d ( T ) U\in\mathcal{A}_d(T) U ∈ A d ( T ) such that
∥ f − U ∥ 2 < ε . \lVert f-U\rVert_2<\varepsilon. ∥ f − U ∥ 2 < ε .
Proof. Fix ϕ ∈ P d ( T ) \phi\in\mathcal{P}_d(T) ϕ ∈ P d ( T ) . Choose m ≥ 1 m\ge1 m ≥ 1 such that Δ T m d + 1 ϕ = 1 \Delta_{T^m}^{d+1}\phi=1 Δ T m d + 1 ϕ = 1 , and put R = T m R=T^m R = T m . By Lemma 2.2 (i), for almost every x x x and every n ∈ Z n\in\mathbb{Z} n ∈ Z , we have
ϕ ( R n x ) = e ( P R , ϕ , x ( n ) ) , deg P R , ϕ , x ≤ d . \phi(R^n x)=e(P_{R,\phi,x}(n)),\qquad\deg P_{R,\phi,x}\le d. ϕ ( R n x ) = e ( P R , ϕ , x ( n )) , deg P R , ϕ , x ≤ d .
For every h ≥ 1 h\ge1 h ≥ 1 , Δ R h ϕ ( R n x ) = e ( P R , ϕ , x ( n + h ) − P R , ϕ , x ( n ) ) \Delta_R^h\phi(R^n x)=e(P_{R,\phi,x}(n+h)-P_{R,\phi,x}(n)) Δ R h ϕ ( R n x ) = e ( P R , ϕ , x ( n + h ) − P R , ϕ , x ( n )) . Since deg P R , ϕ , x ≤ d \deg P_{R,\phi,x}\le d deg P R , ϕ , x ≤ d , taking d + 1 d+1 d + 1 differences gives Δ R h d + 1 ⋯ Δ R h 1 ϕ = 1 \Delta_R^{h_{d+1}}\cdots\Delta_R^{h_1}\phi=1 Δ R h d + 1 ⋯ Δ R h 1 ϕ = 1 for all h 1 , … , h d + 1 ≥ 1 h_1,\ldots,h_{d+1}\ge1 h 1 , … , h d + 1 ≥ 1 . In particular, Δ R q d + 1 ϕ = 1 \Delta_{R^q}^{d+1}\phi=1 Δ R q d + 1 ϕ = 1 for every q ≥ 1 q\ge1 q ≥ 1 .
For (i), suppose that S S S commutes with T T T . For ϕ ∈ P d ( T ) \phi\in\mathcal{P}_d(T) ϕ ∈ P d ( T ) , choose m ≥ 1 m\ge1 m ≥ 1 such that Δ T m d + 1 ϕ = 1 \Delta_{T^m}^{d+1}\phi=1 Δ T m d + 1 ϕ = 1 . Then Δ T m d + 1 ( ϕ ∘ S ± 1 ) = ( Δ T m d + 1 ϕ ) ∘ S ± 1 = 1 \Delta_{T^m}^{d+1}(\phi\circ S^{\pm1})=(\Delta_{T^m}^{d+1}\phi)\circ S^{\pm1}=1 Δ T m d + 1 ( ϕ ∘ S ± 1 ) = ( Δ T m d + 1 ϕ ) ∘ S ± 1 = 1 . Thus ϕ ∘ S \phi\circ S ϕ ∘ S and ϕ ∘ S − 1 \phi\circ S^{-1} ϕ ∘ S − 1 belong to P d ( T ) \mathcal{P}_d(T) P d ( T ) . Since W d ( T ) = σ ( P d ( T ) ) \mathcal{W}_d(T)=\sigma(\mathcal{P}_d(T)) W d ( T ) = σ ( P d ( T )) , S − 1 W d ( T ) = W d ( T ) S^{-1}\mathcal{W}_d(T)=\mathcal{W}_d(T) S − 1 W d ( T ) = W d ( T ) . Taking S = T S=T S = T shows that W d ( T ) \mathcal{W}_d(T) W d ( T ) is T T T -invariant.
For (ii), by definition, we have P d ( T q ) ⊆ P d ( T ) \mathcal{P}_d(T^q)\subseteq\mathcal{P}_d(T) P d ( T q ) ⊆ P d ( T ) . Conversely, if Δ T m d + 1 ϕ = 1 \Delta_{T^m}^{d+1}\phi=1 Δ T m d + 1 ϕ = 1 , the argument above gives Δ T m q d + 1 ϕ = 1 \Delta_{T^{mq}}^{d+1}\phi=1 Δ T m q d + 1 ϕ = 1 . Therefore P d ( T q ) = P d ( T ) \mathcal{P}_d(T^q)=\mathcal{P}_d(T) P d ( T q ) = P d ( T ) and W d ( T q ) = W d ( T ) \mathcal{W}_d(T^q)=\mathcal{W}_d(T) W d ( T q ) = W d ( T ) . If B ∈ I ( T q ) B\in\mathcal{I}(T^q) B ∈ I ( T q ) , put ϕ = 1 − 21 B \phi=1-2\mathbf{1}_B ϕ = 1 − 2 1 B . Then ∣ ϕ ∣ = 1 |\phi|=1 ∣ ϕ ∣ = 1 and Δ T q ϕ = 1 \Delta_{T^q}\phi=1 Δ T q ϕ = 1 . Thus ϕ ∈ P d ( T ) \phi\in\mathcal{P}_d(T) ϕ ∈ P d ( T ) , and 1 B = 1 − ϕ 2 ∈ L ∞ ( W d ( T ) ) \mathbf{1}_B=\frac{1-\phi}{2}\in L^\infty(\mathcal{W}_d(T)) 1 B = 2 1 − ϕ ∈ L ∞ ( W d ( T )) . This proves I ( T q ) ⊆ W d ( T ) \mathcal{I}(T^q)\subseteq\mathcal{W}_d(T) I ( T q ) ⊆ W d ( T ) .
It remains to prove W d ( T ) ⊆ Z d ( T ) \mathcal{W}_d(T)\subseteq\mathcal{Z}_d(T) W d ( T ) ⊆ Z d ( T ) . Fix ϕ ∈ P d ( T ) \phi\in\mathcal{P}_d(T) ϕ ∈ P d ( T ) and choose m ≥ 1 m\ge1 m ≥ 1 with Δ T m d + 1 ϕ = 1 \Delta_{T^m}^{d+1}\phi=1 Δ T m d + 1 ϕ = 1 . Put R = T m R=T^m R = T m and g = E ( ϕ ∣ Z d ( R ) ) g=\mathbb{E}(\phi\mid\mathcal{Z}_d(R)) g = E ( ϕ ∣ Z d ( R )) . Since Δ R d + 1 ϕ = 1 \Delta_R^{d+1}\phi=1 Δ R d + 1 ϕ = 1 , the preceding argument gives Δ R h d + 1 ⋯ Δ R h 1 ϕ = 1 \Delta_R^{h_{d+1}}\cdots\Delta_R^{h_1}\phi=1 Δ R h d + 1 ⋯ Δ R h 1 ϕ = 1 for all h 1 , … , h d + 1 ≥ 1 h_1,\ldots,h_{d+1}\ge1 h 1 , … , h d + 1 ≥ 1 . By (5) , ∥ ϕ ∥ U d + 1 ( R ) = 1 \lVert\phi\rVert_{U^{d+1}(R)}=1 ∥ ϕ ∥ U d + 1 ( R ) = 1 . Since g = E ( ϕ ∣ Z d ( R ) ) g=\mathbb{E}(\phi\mid\mathcal{Z}_d(R)) g = E ( ϕ ∣ Z d ( R )) , we have ∥ ϕ − g ∥ U d + 1 ( R ) = 0 \lVert\phi-g\rVert_{U^{d+1}(R)}=0 ∥ ϕ − g ∥ U d + 1 ( R ) = 0 . Thus
1 = ∥ ϕ ∥ U d + 1 ( R ) ≤ ∥ g ∥ U d + 1 ( R ) + ∥ ϕ − g ∥ U d + 1 ( R ) = ∥ g ∥ U d + 1 ( R ) ≤ ∥ ϕ ∥ U d + 1 ( R ) = 1. 1=\lVert\phi\rVert_{U^{d+1}(R)}\le\lVert g\rVert_{U^{d+1}(R)}+\lVert\phi-g\rVert_{U^{d+1}(R)}=\lVert g\rVert_{U^{d+1}(R)}\le\lVert\phi\rVert_{U^{d+1}(R)}=1. 1 = ∥ ϕ ∥ U d + 1 ( R ) ≤ ∥ g ∥ U d + 1 ( R ) + ∥ ϕ − g ∥ U d + 1 ( R ) = ∥ g ∥ U d + 1 ( R ) ≤ ∥ ϕ ∥ U d + 1 ( R ) = 1.
Hence ∥ g ∥ U d + 1 ( R ) = 1 \lVert g\rVert_{U^{d+1}(R)}=1 ∥ g ∥ U d + 1 ( R ) = 1 .
Since ∣ ϕ ∣ = 1 |\phi|=1 ∣ ϕ ∣ = 1 , conditional expectation gives ∣ g ∣ ≤ 1 |g|\le1 ∣ g ∣ ≤ 1 . For all h 1 , … , h d + 1 ≥ 1 h_1,\ldots,h_{d+1}\ge1 h 1 , … , h d + 1 ≥ 1 ,
∣ Δ R h d + 1 ⋯ Δ R h 1 g ( x ) ∣ = ∏ ω ∈ { 0 , 1 } d + 1 ∣ g ( R ω 1 h 1 + ⋯ + ω d + 1 h d + 1 x ) ∣ ≤ ∣ g ( x ) ∣ . \left|\Delta_R^{h_{d+1}}\cdots\Delta_R^{h_1}g(x)\right|=\prod_{\omega\in\{0,1\}^{d+1}}\left|g\left(R^{\omega_1h_1+\cdots+\omega_{d+1}h_{d+1}}x\right)\right|\le|g(x)|. Δ R h d + 1 ⋯ Δ R h 1 g ( x ) = ω ∈ { 0 , 1 } d + 1 ∏ g ( R ω 1 h 1 + ⋯ + ω d + 1 h d + 1 x ) ≤ ∣ g ( x ) ∣.
Therefore, for every H 1 , … , H d + 1 ≥ 1 H_1,\ldots,H_{d+1}\ge1 H 1 , … , H d + 1 ≥ 1 ,
∣ 1 H 1 ⋯ H d + 1 ∑ h 1 = 1 H 1 ⋯ ∑ h d + 1 = 1 H d + 1 ∫ X Δ R h d + 1 ⋯ Δ R h 1 g d μ ∣ ≤ 1 H 1 ⋯ H d + 1 ∑ h 1 = 1 H 1 ⋯ ∑ h d + 1 = 1 H d + 1 ∫ X ∣ Δ R h d + 1 ⋯ Δ R h 1 g ∣ d μ ≤ ∫ X ∣ g ∣ d μ . \begin{aligned}
\left|\frac{1}{H_1\cdots H_{d+1}}\sum_{h_1=1}^{H_1}\cdots\sum_{h_{d+1}=1}^{H_{d+1}}\int_X\Delta_R^{h_{d+1}}\cdots\Delta_R^{h_1}g\,d\mu\right| \\
&\le\frac{1}{H_1\cdots H_{d+1}}\sum_{h_1=1}^{H_1}\cdots\sum_{h_{d+1}=1}^{H_{d+1}}\int_X\left|\Delta_R^{h_{d+1}}\cdots\Delta_R^{h_1}g\right|\,d\mu\\
&\le\int_X|g|\,d\mu.
\end{aligned} H 1 ⋯ H d + 1 1 h 1 = 1 ∑ H 1 ⋯ h d + 1 = 1 ∑ H d + 1 ∫ X Δ R h d + 1 ⋯ Δ R h 1 g d μ ≤ H 1 ⋯ H d + 1 1 h 1 = 1 ∑ H 1 ⋯ h d + 1 = 1 ∑ H d + 1 ∫ X Δ R h d + 1 ⋯ Δ R h 1 g d μ ≤ ∫ X ∣ g ∣ d μ .
Taking the successive limits in (5) , we obtain 1 = ∥ g ∥ U d + 1 ( R ) 2 d + 1 ≤ ∫ X ∣ g ∣ d μ ≤ 1 1=\lVert g\rVert_{U^{d+1}(R)}^{2^{d+1}}\le\int_X|g|\,d\mu\le1 1 = ∥ g ∥ U d + 1 ( R ) 2 d + 1 ≤ ∫ X ∣ g ∣ d μ ≤ 1 . Thus ∣ g ∣ = 1 |g|=1 ∣ g ∣ = 1 .
Since g g g is the orthogonal projection of ϕ \phi ϕ onto L 2 ( Z d ( R ) ) L^2(\mathcal{Z}_d(R)) L 2 ( Z d ( R )) , ∫ X ∣ ϕ − g ∣ 2 d μ = ∫ X ∣ ϕ ∣ 2 d μ − ∫ X ∣ g ∣ 2 d μ = 0 \int_X|\phi-g|^2\,d\mu=\int_X|\phi|^2\,d\mu-\int_X|g|^2\,d\mu=0 ∫ X ∣ ϕ − g ∣ 2 d μ = ∫ X ∣ ϕ ∣ 2 d μ − ∫ X ∣ g ∣ 2 d μ = 0 . Thus ϕ ∈ L 2 ( Z d ( R ) ) \phi\in L^2(\mathcal{Z}_d(R)) ϕ ∈ L 2 ( Z d ( R )) . Since
∥ f ∥ U d + 1 ( T ) ≤ ∥ f ∥ U d + 1 ( T m ) ≤ m ( d + 1 ) / 2 d + 1 ∥ f ∥ U d + 1 ( T ) \lVert f\rVert_{U^{d+1}(T)}\le\lVert f\rVert_{U^{d+1}(T^m)}\le m^{(d+1)/2^{d+1}}\lVert f\rVert_{U^{d+1}(T)} ∥ f ∥ U d + 1 ( T ) ≤ ∥ f ∥ U d + 1 ( T m ) ≤ m ( d + 1 ) / 2 d + 1 ∥ f ∥ U d + 1 ( T )
then Z d ( T m ) = Z d ( T ) \mathcal{Z}_d(T^m)=\mathcal{Z}_d(T) Z d ( T m ) = Z d ( T ) . Every ϕ ∈ P d ( T ) \phi\in\mathcal{P}_d(T) ϕ ∈ P d ( T ) is therefore Z d ( T ) \mathcal{Z}_d(T) Z d ( T ) -measurable, proving W d ( T ) ⊆ Z d ( T ) \mathcal{W}_d(T)\subseteq\mathcal{Z}_d(T) W d ( T ) ⊆ Z d ( T ) .
For (iii), let ϕ , χ ∈ P d ( T ) \phi,\chi\in\mathcal{P}_d(T) ϕ , χ ∈ P d ( T ) . Choose m 1 , m 2 ≥ 1 m_1,m_2\geq1 m 1 , m 2 ≥ 1 such that Δ T m 1 d + 1 ϕ = 1 \Delta_{T^{m_1}}^{d+1}\phi=1 Δ T m 1 d + 1 ϕ = 1 and Δ T m 2 d + 1 χ = 1 \Delta_{T^{m_2}}^{d+1}\chi=1 Δ T m 2 d + 1 χ = 1 . Taking q = m 1 m 2 q=m_1m_2 q = m 1 m 2 , we have Δ T q d + 1 ϕ = Δ T q d + 1 χ = 1 \Delta_{T^q}^{d+1}\phi=\Delta_{T^q}^{d+1}\chi=1 Δ T q d + 1 ϕ = Δ T q d + 1 χ = 1 . Hence Δ T q d + 1 ( ϕ χ ) = 1 \Delta_{T^q}^{d+1}(\phi\chi)=1 Δ T q d + 1 ( ϕ χ ) = 1 and Δ T q d + 1 ϕ ‾ = 1 \Delta_{T^q}^{d+1}\overline{\phi}=1 Δ T q d + 1 ϕ = 1 . Thus ϕ χ , ϕ ‾ ∈ P d ( T ) \phi\chi,\overline{\phi}\in\mathcal{P}_d(T) ϕ χ , ϕ ∈ P d ( T ) . Since 1 ∈ P d ( T ) 1\in\mathcal{P}_d(T) 1 ∈ P d ( T ) , A d ( T ) \mathcal{A}_d(T) A d ( T ) is an algebra containing the constants and closed under complex conjugation. Moreover, σ ( A d ( T ) ) = W d ( T ) \sigma(\mathcal{A}_d(T))=\mathcal{W}_d(T) σ ( A d ( T )) = W d ( T ) . By the monotone class theorem, every bounded W d ( T ) \mathcal{W}_d(T) W d ( T ) -measurable function can be approximated in L 2 L^2 L 2 by functions in A d ( T ) \mathcal{A}_d(T) A d ( T ) . Since bounded functions are dense in L 2 ( W d ( T ) ) L^2(\mathcal{W}_d(T)) L 2 ( W d ( T )) , the proof is complete. □
We need the following L 2 L^2 L 2 approximation lemma.
Lemma 2.5. Let ( X , X , μ ) (X,\mathcal{X},\mu) ( X , X , μ ) be a probability space, let ℓ ≥ 1 \ell\geq1 ℓ ≥ 1 , and let B N : ( L ∞ ( μ ) ) ℓ → L 2 ( μ ) B_N:(L^\infty(\mu))^\ell\to L^2(\mu) B N : ( L ∞ ( μ ) ) ℓ → L 2 ( μ ) , N ≥ 1 N\geq1 N ≥ 1 , be complex multilinear maps. Suppose that a constant C ≥ 0 C\geq0 C ≥ 0 satisfies
∥ B N ( f 1 , … , f ℓ ) ∥ 2 ≤ C ∥ f i ∥ 2 ∏ j ≠ i ∥ f j ∥ ∞ \lVert B_N(f_1,\ldots,f_\ell)\rVert_2\leq C\lVert f_i\rVert_2\prod_{j\neq i}\lVert f_j\rVert_\infty ∥ B N ( f 1 , … , f ℓ ) ∥ 2 ≤ C ∥ f i ∥ 2 j = i ∏ ∥ f j ∥ ∞ for all bounded f 1 , … , f ℓ f_1,\ldots,f_\ell f 1 , … , f ℓ , i ∈ [ ℓ ] i\in[\ell] i ∈ [ ℓ ] , and N ≥ 1 N\geq1 N ≥ 1 . For each i i i , let V i ⊆ H i ⊆ L ∞ ( μ ) V_i\subseteq H_i\subseteq L^\infty(\mu) V i ⊆ H i ⊆ L ∞ ( μ ) be complex linear spaces, with V i V_i V i dense in H i H_i H i in L 2 L^2 L 2 . If ∥ B N ( v 1 , … , v ℓ ) ∥ 2 → 0 \lVert B_N(v_1,\ldots,v_\ell)\rVert_2\to0 ∥ B N ( v 1 , … , v ℓ ) ∥ 2 → 0 whenever v i ∈ V i v_i\in V_i v i ∈ V i , then ∥ B N ( f 1 , … , f ℓ ) ∥ 2 → 0 \lVert B_N(f_1,\ldots,f_\ell)\rVert_2\to0 ∥ B N ( f 1 , … , f ℓ ) ∥ 2 → 0 whenever f i ∈ H i f_i\in H_i f i ∈ H i .
Proof. We extend the assertion one coordinate at a time. Suppose it holds when the first r − 1 r-1 r − 1 coordinates belong to H i H_i H i and the remaining coordinates belong to V i V_i V i . Fix f i ∈ H i f_i\in H_i f i ∈ H i for i ≤ r i\leq r i ≤ r and v i ∈ V i v_i\in V_i v i ∈ V i for i > r i>r i > r . For every v r ∈ V r v_r\in V_r v r ∈ V r , multilinearity gives
lim sup N → ∞ ∥ B N ( f 1 , … , f r , v r + 1 , … , v ℓ ) ∥ 2 ≤ C ∥ f r − v r ∥ 2 ∏ i < r ∥ f i ∥ ∞ ∏ i > r ∥ v i ∥ ∞ . \limsup_{N\to\infty}\lVert B_N(f_1,\ldots,f_r,v_{r+1},\ldots,v_\ell)\rVert_2\leq C\lVert f_r-v_r\rVert_2\prod_{i<r}\lVert f_i\rVert_\infty\prod_{i>r}\lVert v_i\rVert_\infty. N → ∞ lim sup ∥ B N ( f 1 , … , f r , v r + 1 , … , v ℓ ) ∥ 2 ≤ C ∥ f r − v r ∥ 2 i < r ∏ ∥ f i ∥ ∞ i > r ∏ ∥ v i ∥ ∞ . Approximate f r f_r f r in L 2 L^2 L 2 by elements of V r V_r V r while keeping the other functions fixed. The right-hand side tends to zero. Starting at r = 1 r=1 r = 1 and continuing to r = ℓ r=\ell r = ℓ proves the result. □
Nilmanifolds and nilsystems
For a group G G G , write 1 G 1_G 1 G for its identity and [ a , b ] = a b a − 1 b − 1 [a,b]=aba^{-1}b^{-1} [ a , b ] = ab a − 1 b − 1 for the commutator of a , b ∈ G a,b\in G a , b ∈ G . For subgroups A , B ≤ G A,B\leq G A , B ≤ G , let [ A , B ] [A,B] [ A , B ] be the subgroup generated by their commutators. The lower central series is defined by γ 1 ( G ) = G \gamma_1(G)=G γ 1 ( G ) = G and γ r + 1 ( G ) = [ G , γ r ( G ) ] \gamma_{r+1}(G)=[G,\gamma_r(G)] γ r + 1 ( G ) = [ G , γ r ( G )] for r ≥ 1 r\geq1 r ≥ 1 . For an integer s ≥ 1 s\geq1 s ≥ 1 , we call G G G s s s -step nilpotent if γ s + 1 ( G ) = { 1 G } \gamma_{s+1}(G)=\{1_G\} γ s + 1 ( G ) = { 1 G } . For a Lie group G G G , G ∘ G^\circ G ∘ denotes the connected component of 1 G 1_G 1 G .
Let G G G be an s s s -step nilpotent Lie group and let Γ ≤ G \Gamma\leq G Γ ≤ G be a discrete cocompact subgroup. Then M = G / Γ M=G/\Gamma M = G /Γ is called an s s s -step nilmanifold. We write g x gx g x for the left action g ( h Γ ) = g h Γ g(h\Gamma)=gh\Gamma g ( h Γ ) = g h Γ and m M m_M m M for the unique G G G -invariant Borel probability measure on M M M . For τ ∈ G \tau\in G τ ∈ G , define R τ : x ↦ τ x R_\tau:x\mapsto\tau x R τ : x ↦ τ x (x ∈ M x\in M x ∈ M ) and the system ( M , m M , R τ ) (M,m_M,R_\tau) ( M , m M , R τ ) is called an s s s -step nilsystem.
For a nilmanifold M M M , Comp ( M ) \operatorname{Comp}(M) Comp ( M ) denotes the finite set of connected components of M M M . Any R ∈ Comp ( M ) R\in\operatorname{Comp}(M) R ∈ Comp ( M ) is clopen and can be written as R = G ∘ g R Γ / Γ R=G^\circ g_R\Gamma/\Gamma R = G ∘ g R Γ/Γ for some g R ∈ G g_R\in G g R ∈ G . Let m R = ( m M ∣ R ) / m M ( R ) m_R=(m_M|_R)/m_M(R) m R = ( m M ∣ R ) / m M ( R ) be its normalized Haar measure. For a closed subgroup H ≤ G H\leq G H ≤ G with a compact orbit H x Hx H x , we write m H x m_{Hx} m H x for the H H H -invariant probability measure on that orbit. A continuous character of a torus K K K is a continuous homomorphism χ : K → S 1 \chi:K\to\mathbb{S}^1 χ : K → S 1 .
Let C = [ G ∘ , G ∘ ] C=[G^\circ,G^\circ] C = [ G ∘ , G ∘ ] . The commutator quotient map is ρ : M → Z \rho:M\to Z ρ : M → Z , where Z = { C x : x ∈ M } Z=\{Cx:x\in M\} Z = { C x : x ∈ M } with the quotient topology, C x = { c x : c ∈ C } Cx=\{cx:c\in C\} C x = { c x : c ∈ C } , and ρ ( x ) = C x \rho(x)=Cx ρ ( x ) = C x . For any connected component R = G ∘ g R Γ / Γ R=G^\circ g_R\Gamma/\Gamma R = G ∘ g R Γ/Γ of M M M , let Λ R = G ∘ ∩ g R Γ g R − 1 \Lambda_R=G^\circ\cap g_R\Gamma g_R^{-1} Λ R = G ∘ ∩ g R Γ g R − 1 . The quotient Z R = ρ ( R ) Z_R=\rho(R) Z R = ρ ( R ) is identified with a torus by
Z R ≅ G ∘ / ( C Λ R ) ≅ T d R , ρ ( h g R Γ ) = h ( C Λ R ) ; Z_R\cong G^\circ/(C\Lambda_R)\cong\mathbb{T}^{d_R},\qquad\rho(hg_R\Gamma)=h(C\Lambda_R); Z R ≅ G ∘ / ( C Λ R ) ≅ T d R , ρ ( h g R Γ ) = h ( C Λ R ) ;
see for example [23 ] , Section 1.10. Hence Z = ⨆ R ∈ Comp ( M ) Z R Z=\bigsqcup_{R\in\operatorname{Comp}(M)}Z_R Z = ⨆ R ∈ Comp ( M ) Z R and ρ − 1 ( ρ ( x ) ) = C x \rho^{-1}(\rho(x))=Cx ρ − 1 ( ρ ( x )) = C x .
Reduction to phase factors
In this section, we prove the integer assertion of Theorem A. We first express the rank condition in terms of two linearly independent subfamilies, then use nilsystem models of the individual Host–Kra factors.
The rank condition and independent subfamilies
For a family of polynomials p = ( p 1 , … , p ℓ ) \mathbf{p}=(p_1,\ldots,p_\ell) p = ( p 1 , … , p ℓ ) with a finite index set, let r p ( J ) = dim Q span Q { p j : j ∈ J } r_{\mathbf{p}}(J)=\dim_{\mathbb{Q}}\operatorname{span}_{\mathbb{Q}}\{p_j:j\in J\} r p ( J ) = dim Q span Q { p j : j ∈ J } . The following lemma provides an equivalent form of the rank condition.
Lemma 3.1. Let ℓ ≥ 1 \ell\ge1 ℓ ≥ 1 and p 1 , … , p ℓ ∈ Q [ t ] p_1,\ldots,p_\ell\in\mathbb{Q}[t] p 1 , … , p ℓ ∈ Q [ t ] . The following are equivalent.
(i) For every nonempty J ⊆ [ ℓ ] J\subseteq[\ell] J ⊆ [ ℓ ] ,
∣ J ∣ < 2 r p ( J ) . |J|<2r_{\mathbf{p}}(J). ∣ J ∣ < 2 r p ( J ) . (ii) For every i ∈ [ ℓ ] i\in[\ell] i ∈ [ ℓ ] , there exist J i , K i ⊆ [ ℓ ] J_i,K_i\subseteq[\ell] J i , K i ⊆ [ ℓ ] such that
i ∈ J i ∩ K i , J i ∪ K i = [ ℓ ] , i\in J_i\cap K_i,\qquad J_i\cup K_i=[\ell], i ∈ J i ∩ K i , J i ∪ K i = [ ℓ ] , and both ( p j ) j ∈ J i (p_j)_{j\in J_i} ( p j ) j ∈ J i and ( p j ) j ∈ K i (p_j)_{j\in K_i} ( p j ) j ∈ K i are linearly independent over Q \mathbb{Q} Q .
Proof. Assume (ii) holds. Fix a nonempty J ⊆ [ ℓ ] J\subseteq[\ell] J ⊆ [ ℓ ] and choose i ∈ J i\in J i ∈ J . Since J i ∪ K i = [ ℓ ] J_i\cup K_i=[\ell] J i ∪ K i = [ ℓ ] , i ∈ J i ∩ K i i\in J_i\cap K_i i ∈ J i ∩ K i , and both ( p j ) j ∈ J i (p_j)_{j\in J_i} ( p j ) j ∈ J i and ( p j ) j ∈ K i (p_j)_{j\in K_i} ( p j ) j ∈ K i are linearly independent over Q \mathbb{Q} Q we have
∣ J ∣ ≤ ∣ J ∩ J i ∣ + ∣ J ∩ K i ∣ − 1 ≤ 2 r p ( J ) − 1 < 2 r p ( J ) . |J|\le|J\cap J_i|+|J\cap K_i|-1
\le2r_{\mathbf{p}}(J)-1<2r_{\mathbf{p}}(J). ∣ J ∣ ≤ ∣ J ∩ J i ∣ + ∣ J ∩ K i ∣ − 1 ≤ 2 r p ( J ) − 1 < 2 r p ( J ) . This proves (i).
Conversely, suppose (i) holds and fix i ∈ [ ℓ ] i\in[\ell] i ∈ [ ℓ ] . Take J = { i } J=\{i\} J = { i } with p i ≠ 0 p_i\ne0 p i = 0 . Consider the quotient vector space
V = span Q { p 1 , … , p ℓ } / Q p i . V=\operatorname{span}_{\mathbb{Q}}\{p_1,\ldots,p_\ell\}/\mathbb{Q}p_i. V = span Q { p 1 , … , p ℓ } / Q p i . and let p ‾ j = p j + Q p i ∈ V \overline{p}_j=p_j+\mathbb{Q}p_i\in V p j = p j + Q p i ∈ V for j ≠ i j\ne i j = i . Then for every A ⊆ [ ℓ ] ∖ { i } A\subseteq[\ell]\setminus\{i\} A ⊆ [ ℓ ] ∖ { i } ,
dim Q span Q { p ‾ j : j ∈ A } = r p ( A ∪ { i } ) − 1. \dim_{\mathbb{Q}}\operatorname{span}_{\mathbb{Q}}\{\overline{p}_j:j\in A\}=r_{\mathbf{p}}(A\cup\{i\})-1. dim Q span Q { p j : j ∈ A } = r p ( A ∪ { i }) − 1. Applying (i) to A ∪ { i } A\cup\{i\} A ∪ { i } gives ∣ A ∣ + 1 < 2 r p ( A ∪ { i } ) |A|+1<2r_{\mathbf{p}}(A\cup\{i\}) ∣ A ∣ + 1 < 2 r p ( A ∪ { i }) . Since both sides are integers,
∣ A ∣ ≤ 2 ( r p ( A ∪ { i } ) − 1 ) = 2 dim Q span Q { p ‾ j : j ∈ A } . |A|\le2\bigl(r_{\mathbf{p}}(A\cup\{i\})-1\bigr)=2\dim_{\mathbb{Q}}\operatorname{span}_{\mathbb{Q}}\{\overline{p}_j:j\in A\}. ∣ A ∣ ≤ 2 ( r p ( A ∪ { i }) − 1 ) = 2 dim Q span Q { p j : j ∈ A } . By Edmonds’s partition theorem [8 Theorem 1] , the vectors ( p ‾ j ) j ≠ i (\overline{p}_j)_{j\ne i} ( p j ) j = i can therefore be partitioned into two linearly independent subfamilies. Thus
[ ℓ ] ∖ { i } = A i ∪ ˙ B i , [\ell]\setminus\{i\}=A_i\mathbin{\dot\cup}B_i, [ ℓ ] ∖ { i } = A i ∪ ˙ B i , where ( p ‾ j ) j ∈ A i (\overline{p}_j)_{j\in A_i} ( p j ) j ∈ A i and ( p ‾ j ) j ∈ B i (\overline{p}_j)_{j\in B_i} ( p j ) j ∈ B i are both linearly independent.
Put J i = A i ∪ { i } J_i=A_i\cup\{i\} J i = A i ∪ { i } and K i = B i ∪ { i } K_i=B_i\cup\{i\} K i = B i ∪ { i } . Then J i ∪ K i = [ ℓ ] J_i\cup K_i=[\ell] J i ∪ K i = [ ℓ ] , J i ∩ K i = { i } J_i\cap K_i=\{i\} J i ∩ K i = { i } . To prove that ( p j ) j ∈ J i (p_j)_{j\in J_i} ( p j ) j ∈ J i is linearly independent, suppose
c i p i + ∑ j ∈ A i c j p j = 0 , c i , c j ∈ Q . c_ip_i+\sum_{j\in A_i}c_jp_j=0,\qquad c_i,c_j\in\mathbb{Q}. c i p i + j ∈ A i ∑ c j p j = 0 , c i , c j ∈ Q . we have
∑ j ∈ A i c j p ‾ j = ∑ j ∈ A i c j p j + Q p i = − c i p i + Q p i = 0 in V . \sum_{j\in A_i}c_j\overline{p}_j=\sum_{j\in A_i}c_jp_j+\mathbb{Q}p_i=-c_ip_i+\mathbb{Q}p_i=0\quad\text{in }V. j ∈ A i ∑ c j p j = j ∈ A i ∑ c j p j + Q p i = − c i p i + Q p i = 0 in V . As ( p ‾ j ) j ∈ A i (\overline{p}_j)_{j\in A_i} ( p j ) j ∈ A i is linearly independent, c j = 0 c_j=0 c j = 0 for all j ∈ A i j\in A_i j ∈ A i . Since p i ≠ 0 p_i\ne0 p i = 0 , we also have c i = 0 c_i=0 c i = 0 . Thus ( p j ) j ∈ J i (p_j)_{j\in J_i} ( p j ) j ∈ J i is independent. The same argument applies to K i K_i K i , proving (ii). □ \square □
Nilsystem bundles and commutator quotients
Definition 3.2 (Nilsystem bundles). Following [19 Definition 2.2] , A nilsystem bundle is a measure-preserving system ( Y , ν , S ) (Y,\nu,S) ( Y , ν , S ) of the form
Y = ∐ a ∈ I Ω a × M a , M a = G a / Γ a , I ⊆ N , Y=\coprod_{a\in I}\Omega_a\times M_a,\qquad M_a=G_a/\Gamma_a,\qquad I\subseteq\mathbb{N}, Y = a ∈ I ∐ Ω a × M a , M a = G a / Γ a , I ⊆ N , ν = ∑ a ∈ I β a ⊗ m a , S ( ω , y ) = ( ω , τ a ( ω ) y ) on Ω a × M a . \nu=\sum_{a\in I}\beta_a\otimes m_a,\qquad S(\omega,y)=(\omega,\tau_a(\omega)y)\quad\text{on }\Omega_a\times M_a. ν = a ∈ I ∑ β a ⊗ m a , S ( ω , y ) = ( ω , τ a ( ω ) y ) on Ω a × M a . Every Ω a \Omega_a Ω a is a standard Borel space with a finite measure β a \beta_a β a , and ∑ a ∈ I β a ( Ω a ) = 1 \sum_{a\in I}\beta_a(\Omega_a)=1 ∑ a ∈ I β a ( Ω a ) = 1 . Every M a M_a M a is a nilmanifold with Haar measure m a = m M a m_a=m_{M_a} m a = m M a . The maps τ a : Ω a → G a \tau_a:\Omega_a\to G_a τ a : Ω a → G a are measurable, and
y ↦ τ a ( ω ) y y \mapsto\tau_a(\omega)y y ↦ τ a ( ω ) y is ergodic on ( M a , m a ) (M_a,m_a) ( M a , m a ) for β a \beta_a β a -almost every ω \omega ω . The sigma-algebra Y \mathcal{Y} Y on Y Y Y is the ν \nu ν -completion of the disjoint-union product sigma-algebra.
For an integer s ≥ 1 s \ge1 s ≥ 1 , the bundle is called s s s -step if every G a G_a G a has nilpotency class at most s s s . Without this qualification, no common bound on these classes is assumed.
For several nilsystem bundles ( Y i , ν i , S i ) (Y_i,\nu_i,S_i) ( Y i , ν i , S i ) , we use an additional index i i i , writing Ω i , a \Omega_{i,a} Ω i , a , M i , a M_{i,a} M i , a , β i , a \beta_{i,a} β i , a , and τ i , a \tau_{i,a} τ i , a .
Lemma 3.3. Let M = G / Γ M=G/\Gamma M = G /Γ be a nilmanifold with Haar measure m M m_M m M and put C = [ G ∘ , G ∘ ] C=[G^\circ,G^\circ] C = [ G ∘ , G ∘ ] . Let Z = { C x : x ∈ M } Z=\{Cx:x\in M\} Z = { C x : x ∈ M } with the quotient topology, and define ρ : M → Z \rho:M\to Z ρ : M → Z by ρ ( x ) = C x = { c x : c ∈ C } \rho(x)=Cx=\{cx:c\in C\} ρ ( x ) = C x = { c x : c ∈ C } . Let m Z = ρ ∗ m M m_Z=\rho_*m_M m Z = ρ ∗ m M . If g ∈ L 2 ( m M ) g\in L^2(m_M) g ∈ L 2 ( m M ) satisfies ∥ g ( c ⋅ ) − g ∥ 2 = 0 \|g(c\cdot)-g\|_2=0 ∥ g ( c ⋅ ) − g ∥ 2 = 0 for every c ∈ C c\in C c ∈ C , then there exists g ~ ∈ L 2 ( m Z ) \widetilde{g}\in L^2(m_Z) g ∈ L 2 ( m Z ) such that g = g ~ ∘ ρ g=\widetilde{g}\circ\rho g = g ∘ ρ m M m_M m M -almost everywhere.
Proof. First assume that ∣ g ∣ ≤ K |g|\le K ∣ g ∣ ≤ K almost everywhere and fix a left Haar measure d h dh d h on G ∘ G^\circ G ∘ . Take a decreasing neighbourhood basis ( V n ) (V_n) ( V n ) at the identity of G ∘ G^\circ G ∘ , and η n ∈ C c ( G ∘ ) \eta_n\in C_c(G^\circ) η n ∈ C c ( G ∘ ) with
η n ≥ 0 , ∫ G ∘ η n ( h ) d h = 1 , supp η n ⊆ V n , g n ( x ) = ∫ G ∘ η n ( h ) g ( h − 1 x ) d h . \eta_n\ge0,\qquad\int_{G^\circ}\eta_n(h)\,dh=1,\qquad\operatorname{supp}\eta_n\subseteq V_n,\qquad g_n(x)=\int_{G^\circ}\eta_n(h)g(h^{-1}x)\,dh. η n ≥ 0 , ∫ G ∘ η n ( h ) d h = 1 , supp η n ⊆ V n , g n ( x ) = ∫ G ∘ η n ( h ) g ( h − 1 x ) d h .
For every F ∈ C ( M ) F\in C(M) F ∈ C ( M ) , by the G ∘ G^\circ G ∘ -invariance of m M m_M m M
∥ g ( h − 1 ⋅ ) − g ∥ 2 ≤ 2 ∥ g − F ∥ 2 + sup x ∈ M ∣ F ( h − 1 x ) − F ( x ) ∣ . \|g(h^{-1}\cdot)-g\|_2\le2\|g-F\|_2+\sup_{x\in M}|F(h^{-1}x)-F(x)|. ∥ g ( h − 1 ⋅ ) − g ∥ 2 ≤ 2∥ g − F ∥ 2 + x ∈ M sup ∣ F ( h − 1 x ) − F ( x ) ∣.
For fixed F F F , compactness gives lim sup h → 1 G ∥ g ( h − 1 ⋅ ) − g ∥ 2 ≤ 2 ∥ g − F ∥ 2 \limsup_{h\to1_G}\|g(h^{-1}\cdot)-g\|_2\le2\|g-F\|_2 lim sup h → 1 G ∥ g ( h − 1 ⋅ ) − g ∥ 2 ≤ 2∥ g − F ∥ 2 . Approximating g g g by continuous F F F therefore gives ∥ g ( h − 1 ⋅ ) − g ∥ 2 → 0 \|g(h^{-1}\cdot)-g\|_2\to0 ∥ g ( h − 1 ⋅ ) − g ∥ 2 → 0 as h → 1 G h\to1_G h → 1 G . Hence
∥ g n − g ∥ 2 ≤ sup h ∈ supp η n ∥ g ( h − 1 ⋅ ) − g ∥ 2 ⟶ 0. \|g_n-g\|_2\le\sup_{h\in\operatorname{supp}\eta_n}\|g(h^{-1}\cdot)-g\|_2\longrightarrow0. ∥ g n − g ∥ 2 ≤ h ∈ supp η n sup ∥ g ( h − 1 ⋅ ) − g ∥ 2 ⟶ 0.
For a ∈ G ∘ a\in G^\circ a ∈ G ∘ , changing variables gives
∣ g n ( a x ) − g n ( x ) ∣ ≤ K ∫ G ∘ ∣ η n ( a h ) − η n ( h ) ∣ d h ⟶ 0 when a → 1 G , |g_n(ax)-g_n(x)|\le K\int_{G^\circ}|\eta_n(ah)-\eta_n(h)|\,dh\longrightarrow0\quad\text{when }a\to1_G, ∣ g n ( a x ) − g n ( x ) ∣ ≤ K ∫ G ∘ ∣ η n ( ah ) − η n ( h ) ∣ d h ⟶ 0 when a → 1 G ,
uniformly in x ∈ M x\in M x ∈ M . Since V x Vx V x is open in M M M for every open neighborhood V V V of 1 G 1_G 1 G in G ∘ G^\circ G ∘ , we have g n ∈ C ( M ) g_n\in C(M) g n ∈ C ( M ) . For c ∈ C = [ G ∘ , G ∘ ] c\in C=[G^\circ,G^\circ] c ∈ C = [ G ∘ , G ∘ ] , we have
∥ g n ( c ⋅ ) − g n ∥ 2 ≤ ∫ G ∘ η n ( h ) ∥ g ( ( h − 1 c h ) ⋅ ) − g ∥ 2 d h = 0. \|g_n(c\cdot)-g_n\|_2\le\int_{G^\circ}\eta_n(h)\|g((h^{-1}ch)\cdot)-g\|_2\,dh=0. ∥ g n ( c ⋅ ) − g n ∥ 2 ≤ ∫ G ∘ η n ( h ) ∥ g (( h − 1 c h ) ⋅ ) − g ∥ 2 d h = 0.
Since g n g_n g n is continuous and m M m_M m M has full support, the equality ∥ g n ( c ⋅ ) − g n ∥ 2 = 0 \|g_n(c\cdot)-g_n\|_2=0 ∥ g n ( c ⋅ ) − g n ∥ 2 = 0 implies g n ( c x ) = g n ( x ) g_n(cx)=g_n(x) g n ( c x ) = g n ( x ) for every c ∈ C c\in C c ∈ C , x ∈ M x\in M x ∈ M . Thus the formula g ~ n ( C x ) = g n ( x ) \widetilde{g}_n(Cx)=g_n(x) g n ( C x ) = g n ( x ) defines a continuous function on Z Z Z by the definition of the quotient topology. Hence g n = g ~ n ∘ ρ g_n=\widetilde{g}_n\circ\rho g n = g n ∘ ρ . By m Z = ρ ∗ m M m_Z=\rho_*m_M m Z = ρ ∗ m M ,
∥ g ~ n − g ~ m ∥ L 2 ( m Z ) = ∥ g n − g m ∥ 2 . \|\widetilde{g}_n-\widetilde{g}_m\|_{L^2(m_Z)}=\|g_n-g_m\|_2. ∥ g n − g m ∥ L 2 ( m Z ) = ∥ g n − g m ∥ 2 .
Hence g ~ n → g ~ \widetilde{g}_n\to\widetilde{g} g n → g in L 2 ( m Z ) L^2(m_Z) L 2 ( m Z ) for some g ~ \widetilde{g} g , and
∥ g − g ~ ∘ ρ ∥ 2 ≤ ∥ g − g n ∥ 2 + ∥ g ~ n − g ~ ∥ L 2 ( m Z ) ⟶ 0. \|g-\widetilde{g}\circ\rho\|_2\le\|g-g_n\|_2+\|\widetilde{g}_n-\widetilde{g}\|_{L^2(m_Z)}\longrightarrow0. ∥ g − g ∘ ρ ∥ 2 ≤ ∥ g − g n ∥ 2 + ∥ g n − g ∥ L 2 ( m Z ) ⟶ 0.
For general g ∈ L 2 ( m M ) g\in L^2(m_M) g ∈ L 2 ( m M ) , put g N = g 1 { ∣ g ∣ ≤ N } g_N=g1_{\{|g|\le N\}} g N = g 1 { ∣ g ∣ ≤ N } . Then g N g_N g N is bounded, is C C C -invariant in L 2 L^2 L 2 , and
∥ g − g N ∥ 2 2 = ∫ { ∣ g ∣ > N } ∣ g ∣ 2 d m M ⟶ 0 , when N → ∞ . \|g-g_N\|_2^2=\int_{\{|g|>N\}}|g|^2\,dm_M\longrightarrow0,\quad\text{when }N\to\infty. ∥ g − g N ∥ 2 2 = ∫ { ∣ g ∣ > N } ∣ g ∣ 2 d m M ⟶ 0 , when N → ∞.
By the bounded case, there exists g ~ N ∈ L 2 ( m Z ) \widetilde{g}_N\in L^2(m_Z) g N ∈ L 2 ( m Z ) such that g N = g ~ N ∘ ρ g_N=\widetilde{g}_N\circ\rho g N = g N ∘ ρ . Then as N , L → ∞ N,L\to\infty N , L → ∞ ,
∥ g ~ N − g ~ L ∥ L 2 ( m Z ) = ∥ g N − g L ∥ 2 ⟶ 0. \|\widetilde{g}_N-\widetilde{g}_L\|_{L^2(m_Z)}=\|g_N-g_L\|_2\longrightarrow0. ∥ g N − g L ∥ L 2 ( m Z ) = ∥ g N − g L ∥ 2 ⟶ 0.
Thus g ~ N \widetilde{g}_N g N converges in L 2 ( m Z ) L^2(m_Z) L 2 ( m Z ) to some g ~ \widetilde{g} g , and
∥ g − g ~ ∘ ρ ∥ 2 ≤ ∥ g − g N ∥ 2 + ∥ g ~ N − g ~ ∥ L 2 ( m Z ) ⟶ 0. \|g-\widetilde{g}\circ\rho\|_2\le\|g-g_N\|_2+\|\widetilde{g}_N-\widetilde{g}\|_{L^2(m_Z)}\longrightarrow0. ∥ g − g ∘ ρ ∥ 2 ≤ ∥ g − g N ∥ 2 + ∥ g N − g ∥ L 2 ( m Z ) ⟶ 0.
:::
Lemma 3.4. Let s ≥ 1 s\ge1 s ≥ 1 , and let ( Y , ν , S ) (Y,\nu,S) ( Y , ν , S ) be an s s s -step nilsystem bundle, where Y = ⨆ a ∈ I Ω a × M a Y=\bigsqcup_{a\in I}\Omega_a\times M_a Y = ⨆ a ∈ I Ω a × M a , M a = G a / Γ a M_a=G_a/\Gamma_a M a = G a / Γ a , I ⊆ N I\subseteq\mathbb{N} I ⊆ N , ν = ∑ a ∈ I β a ⊗ m a \nu=\sum_{a\in I}\beta_a\otimes m_a ν = ∑ a ∈ I β a ⊗ m a , and S ( ω , y ) = ( ω , τ a ( ω ) y ) S(\omega,y)=(\omega,\tau_a(\omega)y) S ( ω , y ) = ( ω , τ a ( ω ) y ) on each Ω a × M a \Omega_a\times M_a Ω a × M a . For each a ∈ I a\in I a ∈ I , put C a = [ G a ∘ , G a ∘ ] C_a=[G_a^\circ,G_a^\circ] C a = [ G a ∘ , G a ∘ ] and Z a = { C a y : y ∈ M a } Z_a=\{C_ay:y\in M_a\} Z a = { C a y : y ∈ M a } , and define ρ a : M a → Z a \rho_a:M_a\to Z_a ρ a : M a → Z a by ρ a ( y ) = C a y \rho_a(y)=C_ay ρ a ( y ) = C a y . Give Z a Z_a Z a the quotient topology. Let
F = σ ν ( { B × ρ a − 1 ( D ) : a ∈ I , B ⊆ Ω a measurable , D ∈ B ( Z a ) } ) . \mathcal{F}=\sigma_\nu\left(\{B\times\rho_a^{-1}(D):a\in I,\ B\subseteq\Omega_a\text{ measurable},\ D\in\mathcal{B}(Z_a)\}\right). F = σ ν ( { B × ρ a − 1 ( D ) : a ∈ I , B ⊆ Ω a measurable , D ∈ B ( Z a )} ) .
Define
V = span C { b ( ω ) v ( y ) : a ∈ I , b ∈ L ∞ ( β a ) , v ∈ C ( M a ) } , \mathcal{V}=\operatorname{span}_{\mathbb{C}}\{b(\omega)v(y):a\in I,\ b\in L^\infty(\beta_a),\ v\in C(M_a)\}, V = span C { b ( ω ) v ( y ) : a ∈ I , b ∈ L ∞ ( β a ) , v ∈ C ( M a )} ,
D = span C { b ( ω ) ( v ( c y ) − v ( y ) ) : a ∈ I , b ∈ L ∞ ( β a ) , v ∈ C ( M a ) , c ∈ C a } , \mathcal{D}=\operatorname{span}_{\mathbb{C}}\{b(\omega)(v(cy)-v(y)):a\in I,\ b\in L^\infty(\beta_a),\ v\in C(M_a),\ c\in C_a\}, D = span C { b ( ω ) ( v ( cy ) − v ( y )) : a ∈ I , b ∈ L ∞ ( β a ) , v ∈ C ( M a ) , c ∈ C a } ,
where each function is defined on Ω a × M a \Omega_a\times M_a Ω a × M a and is zero elsewhere.
(i) F ⊆ W s ( S ) \mathcal{F}\subseteq\mathcal{W}_s(S) F ⊆ W s ( S ) .
(ii) V \mathcal{V} V is dense in L 2 ( ν ) L^2(\nu) L 2 ( ν ) , and D ⊥ = L 2 ( F , ν ) \mathcal{D}^{\perp}=L^2(\mathcal{F},\nu) D ⊥ = L 2 ( F , ν ) .
(iii) If F ∈ L 2 ( ν ) F\in L^2(\nu) F ∈ L 2 ( ν ) satisfies E ( F ∣ W s ( S ) ) = 0 \mathbb{E}(F\mid\mathcal{W}_s(S))=0 E ( F ∣ W s ( S )) = 0 , then there exist ψ k ∈ D \psi_k\in\mathcal{D} ψ k ∈ D such that ∥ F − ψ k ∥ L 2 ( ν ) → 0 \lVert F-\psi_k\rVert_{L^2(\nu)}\to0 ∥ F − ψ k ∥ L 2 ( ν ) → 0 .
Proof. (i). For each a a a , put q a = ∣ Comp ( M a ) ∣ ! q_a=\lvert\operatorname{Comp}(M_a)\rvert! q a = ∣ Comp ( M a )∣! . Then τ a ( ω ) q a R = R \tau_a(\omega)^{q_a}R=R τ a ( ω ) q a R = R for every ω ∈ Ω a \omega\in\Omega_a ω ∈ Ω a and R ∈ Comp ( M a ) R\in\operatorname{Comp}(M_a) R ∈ Comp ( M a ) .
Fix a a a and a connected component R = G a ∘ g R Γ a / Γ a R=G_a^\circ g_R\Gamma_a/\Gamma_a R = G a ∘ g R Γ a / Γ a of M a M_a M a . Put
Λ R = G a ∘ ∩ g R Γ a g R − 1 , Z a , R = G a ∘ / ( C a Λ R ) . \Lambda_R=G_a^\circ\cap g_R\Gamma_ag_R^{-1},\qquad Z_{a,R}=G_a^\circ/(C_a\Lambda_R). Λ R = G a ∘ ∩ g R Γ a g R − 1 , Z a , R = G a ∘ / ( C a Λ R ) .
Let χ \chi χ be a continuous character of Z a , R Z_{a,R} Z a , R . Put f 0 = χ ∘ ρ a f_0=\chi\circ\rho_a f 0 = χ ∘ ρ a on R R R and χ 0 ( h ) = χ ( h ( C a Λ R ) ) \chi_0(h)=\chi(h(C_a\Lambda_R)) χ 0 ( h ) = χ ( h ( C a Λ R )) for h ∈ G a ∘ h\in G_a^\circ h ∈ G a ∘ . Since ρ a ( h g R Γ a ) = h ( C a Λ R ) \rho_a(hg_R\Gamma_a)=h(C_a\Lambda_R) ρ a ( h g R Γ a ) = h ( C a Λ R ) , f 0 ( h y ) = χ 0 ( h ) f 0 ( y ) f_0(hy)=\chi_0(h)f_0(y) f 0 ( h y ) = χ 0 ( h ) f 0 ( y ) for all h ∈ G a ∘ h\in G_a^\circ h ∈ G a ∘ , y ∈ R y\in R y ∈ R .
Fix ω ∈ Ω a \omega\in\Omega_a ω ∈ Ω a and put u = τ a ( ω ) q a u=\tau_a(\omega)^{q_a} u = τ a ( ω ) q a . Define f r + 1 ( y ) = f r ( u y ) f r ( y ) f_{r+1}(y)=f_r(uy)f_r(y) f r + 1 ( y ) = f r ( u y ) f r ( y ) . For h ∈ G a ∘ h\in G_a^\circ h ∈ G a ∘ , define h 0 ( h ) = h h_0(h)=h h 0 ( h ) = h and h r + 1 ( h ) = [ u , h r ( h ) ] h_{r+1}(h)=[u,h_r(h)] h r + 1 ( h ) = [ u , h r ( h )] . Then h r ( h ) ∈ G a ∘ ∩ γ r + 1 ( G a ) h_r(h)\in G_a^\circ\cap\gamma^{r+1}(G_a) h r ( h ) ∈ G a ∘ ∩ γ r + 1 ( G a ) and h r ( u h u − 1 ) = u h r ( h ) u − 1 h_r(uhu^{-1})=uh_r(h)u^{-1} h r ( u h u − 1 ) = u h r ( h ) u − 1 .
We prove by induction that f r ( h y ) = χ 0 ( h r ( h ) ) f r ( y ) f_r(hy)=\chi_0(h_r(h))f_r(y) f r ( h y ) = χ 0 ( h r ( h )) f r ( y ) . The case r = 0 r=0 r = 0 follows from the definition of f 0 f_0 f 0 . For the induction step,
f r + 1 ( h y ) = f r ( ( u h u − 1 ) u y ) f r ( h y ) = χ 0 ( u h r ( h ) u − 1 ) χ 0 ( h r ( h ) ) f r + 1 ( y ) = χ 0 ( h r + 1 ( h ) ) f r + 1 ( y ) . \begin{aligned}
f_{r+1}(hy)&=f_r((uhu^{-1})uy)f_r(hy)\\
&=\chi_0(uh_r(h)u^{-1})\chi_0(h_r(h))f_{r+1}(y)\\
&=\chi_0(h_{r+1}(h))f_{r+1}(y).
\end{aligned} f r + 1 ( h y ) = f r (( u h u − 1 ) u y ) f r ( h y ) = χ 0 ( u h r ( h ) u − 1 ) χ 0 ( h r ( h )) f r + 1 ( y ) = χ 0 ( h r + 1 ( h )) f r + 1 ( y ) .
Since G a G_a G a has step at most s s s , we have h s ( h ) = 1 G a h_s(h)=1_{G_a} h s ( h ) = 1 G a . Hence f s ( h y ) = f s ( y ) f_s(hy)=f_s(y) f s ( h y ) = f s ( y ) for every h ∈ G a ∘ h\in G_a^\circ h ∈ G a ∘ and y ∈ R y\in R y ∈ R . Since R = G a ∘ g R Γ a / Γ a R=G_a^\circ g_R\Gamma_a/\Gamma_a R = G a ∘ g R Γ a / Γ a , there is ζ ∈ S 1 \zeta\in S^1 ζ ∈ S 1 such that f s ( y ) = ζ f_s(y)=\zeta f s ( y ) = ζ for all y ∈ R y\in R y ∈ R . As u R = R uR=R u R = R ,
f s + 1 ( y ) = f s ( u y ) f s ( y ) = ζ ζ ‾ = 1 f_{s+1}(y)=f_s(uy)f_s(y)=\zeta\overline{\zeta}=1 f s + 1 ( y ) = f s ( u y ) f s ( y ) = ζ ζ = 1
for all $y\in R.
Define Φ \Phi Φ on Y Y Y by
Φ ( ω , y ) = { χ ( ρ a ( y ) ) , ( ω , y ) ∈ Ω a × R , 1 , otherwise . \Phi(\omega,y)=
\begin{cases}
\chi(\rho_a(y)), &(\omega,y)\in\Omega_a\times R,\\
1, &\text{otherwise}.
\end{cases} Φ ( ω , y ) = { χ ( ρ a ( y )) , 1 , ( ω , y ) ∈ Ω a × R , otherwise .
For ( ω , y ) ∈ Ω a × R (\omega,y)\in\Omega_a\times R ( ω , y ) ∈ Ω a × R , we have S q a ( ω , y ) = ( ω , u y ) S^{q_a}(\omega,y)=(\omega,uy) S q a ( ω , y ) = ( ω , u y ) and u R = R uR=R u R = R . Thus Δ S q a r Φ ( ω , y ) = f r ( y ) \Delta_{S^{q_a}}^r\Phi(\omega,y)=f_r(y) Δ S q a r Φ ( ω , y ) = f r ( y ) for every r ≥ 0 r\ge0 r ≥ 0 , and in particular Δ S q a s + 1 Φ = 1 \Delta_{S^{q_a}}^{s+1}\Phi=1 Δ S q a s + 1 Φ = 1 on Ω a × R \Omega_a\times R Ω a × R . The same holds on Y ∖ ( Ω a × R ) Y\setminus(\Omega_a\times R) Y ∖ ( Ω a × R ) , since this set is S q a S^{q_a} S q a -invariant and Φ = 1 \Phi=1 Φ = 1 there. Hence Φ ∈ L ∞ ( W s ( S ) ) \Phi\in L^\infty(\mathcal{W}_s(S)) Φ ∈ L ∞ ( W s ( S )) .
For any measurable B ⊆ Ω a B\subseteq\Omega_a B ⊆ Ω a , we have S − q a ( B × R ) = B × R S^{-q_a}(B\times R)=B\times R S − q a ( B × R ) = B × R . By Lemma 2.4 , B × R ∈ I ( S q a ) ⊆ W s ( S ) B\times R\in\mathcal{I}(S^{q_a})\subseteq\mathcal{W}_s(S) B × R ∈ I ( S q a ) ⊆ W s ( S ) . Since Φ ∈ L ∞ ( W s ( S ) ) \Phi\in L^\infty(\mathcal{W}_s(S)) Φ ∈ L ∞ ( W s ( S )) , we also have 1 B × R Φ ∈ L ∞ ( W s ( S ) ) 1_{B\times R}\Phi\in L^\infty(\mathcal{W}_s(S)) 1 B × R Φ ∈ L ∞ ( W s ( S )) . The continuous characters of the torus Z a , R Z_{a,R} Z a , R generate its Borel sigma-algebra. Consequently, B × ρ a − 1 ( D R ) ∈ W s ( S ) B\times\rho_a^{-1}(D_R)\in\mathcal{W}_s(S) B × ρ a − 1 ( D R ) ∈ W s ( S ) for all measurable B ⊆ Ω a B\subseteq\Omega_a B ⊆ Ω a and D R ∈ B ( Z a , R ) D_R\in\mathcal{B}(Z_{a,R}) D R ∈ B ( Z a , R ) .
For every D ∈ B ( Z a ) D\in\mathcal{B}(Z_a) D ∈ B ( Z a ) , we have
B × ρ a − 1 ( D ) = ⋃ R ∈ Comp ( M a ) B × ρ a − 1 ( D ∩ Z a , R ) ∈ W s ( S ) . B\times\rho_a^{-1}(D)=\bigcup_{R\in\operatorname{Comp}(M_a)}B\times\rho_a^{-1}(D\cap Z_{a,R})\in\mathcal{W}_s(S). B × ρ a − 1 ( D ) = R ∈ Comp ( M a ) ⋃ B × ρ a − 1 ( D ∩ Z a , R ) ∈ W s ( S ) .
Since these sets generate F \mathcal{F} F after completion, we obtain F ⊆ W s ( S ) \mathcal{F}\subseteq\mathcal{W}_s(S) F ⊆ W s ( S ) .
(ii). For measurable B ⊆ Ω a B\subseteq\Omega_a B ⊆ Ω a and E ⊆ M a E\subseteq M_a E ⊆ M a , take v n ∈ C ( M a ) v_n\in C(M_a) v n ∈ C ( M a ) with ∥ v n − 1 E ∥ L 2 ( m a ) → 0 \lVert v_n-1_E\rVert_{L^2(m_a)}\to0 ∥ v n − 1 E ∥ L 2 ( m a ) → 0 . Then
∥ 1 B ( ω ) v n ( y ) − 1 B × E ∥ L 2 ( ν ) 2 = β a ( B ) ∥ v n − 1 E ∥ L 2 ( m a ) 2 ⟶ 0. \left\lVert1_B(\omega)v_n(y)-1_{B\times E}\right\rVert_{L^2(\nu)}^2
=\beta_a(B)\left\lVert v_n-1_E\right\rVert_{L^2(m_a)}^2\longrightarrow0. ∥ 1 B ( ω ) v n ( y ) − 1 B × E ∥ L 2 ( ν ) 2 = β a ( B ) ∥ v n − 1 E ∥ L 2 ( m a ) 2 ⟶ 0.
Finite linear combinations of rectangle indicators on finitely many pieces are dense in L 2 ( ν ) L^2(\nu) L 2 ( ν ) . Hence the functions 1 B ( ω ) v ( y ) 1_B(\omega)v(y) 1 B ( ω ) v ( y ) , with v ∈ C ( M a ) v\in C(M_a) v ∈ C ( M a ) , have dense linear span, and V \mathcal{V} V is dense.
Let H ∈ L 2 ( F , ν ) H \in L^2(\mathcal{F},\nu) H ∈ L 2 ( F , ν ) . Since ρ a ( c y ) = ρ a ( y ) \rho_a(cy)=\rho_a(y) ρ a ( cy ) = ρ a ( y ) , H ( ω , c y ) = H ( ω , y ) H(\omega,cy)=H(\omega,y) H ( ω , cy ) = H ( ω , y ) almost everywhere for each c ∈ C a c\in C_a c ∈ C a . Haar invariance gives
⟨ H , b ( ω ) ( v ( c y ) − v ( y ) ) ⟩ = ∫ Ω a × M a ( H ( ω , c − 1 y ) − H ( ω , y ) ) b ( ω ) ‾ v ( y ) d ( β a ⊗ m a ) = 0. \langle H,b(\omega)(v(cy)-v(y))\rangle
=\int_{\Omega_a\times M_a}\bigl(H(\omega,c^{-1}y)-H(\omega,y)\bigr)\overline{b(\omega)}v(y)\,d(\beta_a\otimes m_a)=0. ⟨ H , b ( ω ) ( v ( cy ) − v ( y ))⟩ = ∫ Ω a × M a ( H ( ω , c − 1 y ) − H ( ω , y ) ) b ( ω ) v ( y ) d ( β a ⊗ m a ) = 0.
Thus L 2 ( F , ν ) ⊆ D ⊥ L^2(\mathcal{F},\nu)\subseteq D^\perp L 2 ( F , ν ) ⊆ D ⊥ .
Since L 2 ( F , ν ) L^2(\mathcal{F},\nu) L 2 ( F , ν ) is closed, it remains to prove that H ⊥ D H\perp D H ⊥ D and H ⊥ L 2 ( F , ν ) H\perp L^2(\mathcal{F},\nu) H ⊥ L 2 ( F , ν ) imply H = 0 H=0 H = 0 . Fix a ∈ I a\in I a ∈ I and a measurable B ⊆ Ω a B\subseteq\Omega_a B ⊆ Ω a . Put h B ( y ) = ∫ B H ( ω , y ) d β a ( ω ) h_B(y)=\int_B H(\omega,y)\,d\beta_a(\omega) h B ( y ) = ∫ B H ( ω , y ) d β a ( ω ) . By Cauchy–Schwarz,
∥ h B ∥ L 2 ( m a ) 2 ≤ β a ( B ) ∫ B × M a ∣ H ∣ 2 d ( β a ⊗ m a ) < ∞ . \|h_B\|_{L^2(m_a)}^2\leq\beta_a(B)\int_{B\times M_a}|H|^2\,d(\beta_a\otimes m_a)<\infty. ∥ h B ∥ L 2 ( m a ) 2 ≤ β a ( B ) ∫ B × M a ∣ H ∣ 2 d ( β a ⊗ m a ) < ∞.
For c ∈ C a c\in C_a c ∈ C a and v ∈ C ( M a ) v\in C(M_a) v ∈ C ( M a ) ,
⟨ h B ( c − 1 ⋅ ) − h B , v ⟩ = ⟨ h B , v ( c ⋅ ) − v ⟩ = ⟨ H , 1 B ( ω ) ( v ( c y ) − v ( y ) ) ⟩ = 0. \langle h_B(c^{-1}\cdot)-h_B,v\rangle=\langle h_B,v(c\cdot)-v\rangle=\langle H,1_B(\omega)(v(cy)-v(y))\rangle=0. ⟨ h B ( c − 1 ⋅ ) − h B , v ⟩ = ⟨ h B , v ( c ⋅ ) − v ⟩ = ⟨ H , 1 B ( ω ) ( v ( cy ) − v ( y ))⟩ = 0.
Density of C ( M a ) C(M_a) C ( M a ) in L 2 ( m a ) L^2(m_a) L 2 ( m a ) gives h B ( c − 1 ⋅ ) = h B h_B(c^{-1}\cdot)=h_B h B ( c − 1 ⋅ ) = h B in L 2 ( m a ) L^2(m_a) L 2 ( m a ) for every c ∈ C a c\in C_a c ∈ C a . By Lemma 3.3 , 1 B ( ω ) h B ( y ) ∈ L 2 ( F , ν ) 1_B(\omega)h_B(y)\in L^2(\mathcal{F},\nu) 1 B ( ω ) h B ( y ) ∈ L 2 ( F , ν ) . Therefore, Fubini’s theorem gives
0 = ⟨ H , 1 B ( ω ) h B ( y ) ⟩ = ∫ M a ( ∫ B H ( ω , y ) d β a ( ω ) ) h B ( y ) ‾ d m a ( y ) = ∥ h B ∥ L 2 ( m a ) 2 . 0=\langle H,1_B(\omega)h_B(y)\rangle=\int_{M_a}\left(\int_B H(\omega,y)\,d\beta_a(\omega)\right)\overline{h_B(y)}\,dm_a(y)
=\|h_B\|_{L^2(m_a)}^2. 0 = ⟨ H , 1 B ( ω ) h B ( y )⟩ = ∫ M a ( ∫ B H ( ω , y ) d β a ( ω ) ) h B ( y ) d m a ( y ) = ∥ h B ∥ L 2 ( m a ) 2 .
Hence h B = 0 h_B=0 h B = 0 , and ⟨ H , 1 B ( ω ) v ( y ) ⟩ = ⟨ h B , v ⟩ = 0 \langle H,1_B(\omega)v(y)\rangle=\langle h_B,v\rangle=0 ⟨ H , 1 B ( ω ) v ( y )⟩ = ⟨ h B , v ⟩ = 0 for every a a a , B B B , and v ∈ C ( M a ) v\in C(M_a) v ∈ C ( M a ) . The density proved above gives H = 0 H=0 H = 0 . Thus D ⊥ = L 2 ( F , ν ) D^\perp=L^2(\mathcal{F},\nu) D ⊥ = L 2 ( F , ν ) .
( i i i ) (iii) ( iii ) . Part (i) gives L 2 ( F , ν ) ⊆ L 2 ( W s ( S ) , ν ) L^2(\mathcal{F},\nu)\subseteq L^2(\mathcal{W}_s(S),\nu) L 2 ( F , ν ) ⊆ L 2 ( W s ( S ) , ν ) . If E ( F ∣ W s ( S ) ) = 0 \mathbb{E}(F\mid\mathcal{W}_s(S))=0 E ( F ∣ W s ( S )) = 0 , then F ∈ ( D ⊥ ) ⊥ F\in(D^\perp)^\perp F ∈ ( D ⊥ ) ⊥ . Since ( D ⊥ ) ⊥ (D^\perp)^\perp ( D ⊥ ) ⊥ is the L 2 ( ν ) L^2(\nu) L 2 ( ν ) -closure of D D D , there exist ψ k ∈ D \psi_k\in D ψ k ∈ D with ∥ F − ψ k ∥ L 2 ( ν ) → 0 \|F-\psi_k\|_{L^2(\nu)}\to0 ∥ F − ψ k ∥ L 2 ( ν ) → 0 . □
Lemma 3.5. Let ℓ ≥ 3 \ell\geq3 ℓ ≥ 3 , and let ( X , μ , T 1 , … , T ℓ ) (X,\mu,T_1,\ldots,T_\ell) ( X , μ , T 1 , … , T ℓ ) be a system of invertible commuting measure-preserving transformations. Let p 1 , … , p ℓ ∈ Z [ t ] p_1,\ldots,p_\ell\in\mathbb{Z}[t] p 1 , … , p ℓ ∈ Z [ t ] satisfy p i ( 0 ) = 0 p_i(0)=0 p i ( 0 ) = 0 . Suppose that for each i ∈ [ ℓ ] i\in[\ell] i ∈ [ ℓ ] there are J i , K i ⊆ [ ℓ ] J_i,K_i\subseteq[\ell] J i , K i ⊆ [ ℓ ] with i ∈ J i ∩ K i i\in J_i\cap K_i i ∈ J i ∩ K i and J i ∪ K i = [ ℓ ] J_i\cup K_i=[\ell] J i ∪ K i = [ ℓ ] , and that both ( p j ) j ∈ J i (p_j)_{j\in J_i} ( p j ) j ∈ J i and ( p j ) j ∈ K i (p_j)_{j\in K_i} ( p j ) j ∈ K i are linearly independent over Q \mathbb{Q} Q . For each i i i , let π i : ( X , μ , T i ) → ( Y i , ν i , S i ) \pi_i:(X,\mu,T_i)\to(Y_i,\nu_i,S_i) π i : ( X , μ , T i ) → ( Y i , ν i , S i ) be a factor map onto a nilsystem bundle, where
Y i = ∐ a ∈ I i Ω i , a × M i , a , M i , a = G i , a / Γ i , a , Y_i=\coprod_{a\in I_i}\Omega_{i,a}\times M_{i,a},\qquad M_{i,a}=G_{i,a}/\Gamma_{i,a}, Y i = a ∈ I i ∐ Ω i , a × M i , a , M i , a = G i , a / Γ i , a , and S i ( ω , y ) = ( ω , τ i , a ( ω ) y ) S_i(\omega,y)=(\omega,\tau_{i,a}(\omega)y) S i ( ω , y ) = ( ω , τ i , a ( ω ) y ) on each piece. For x ∈ X x\in X x ∈ X , write π i ( x ) = ( ω i , y i ) ∈ Ω i , a i × M i , a i \pi_i(x)=(\omega_i,y_i)\in\Omega_{i,a_i}\times M_{i,a_i} π i ( x ) = ( ω i , y i ) ∈ Ω i , a i × M i , a i , and put G i = G i , a i G_i=G_{i,a_i} G i = G i , a i , Γ i = Γ i , a i \Gamma_i=\Gamma_{i,a_i} Γ i = Γ i , a i , M i = M i , a i M_i=M_{i,a_i} M i = M i , a i , τ i = τ i , a i ( ω i ) \tau_i=\tau_{i,a_i}(\omega_i) τ i = τ i , a i ( ω i ) , and C i = [ G i ∘ , G i ∘ ] C_i=[G_i^\circ,G_i^\circ] C i = [ G i ∘ , G i ∘ ] . Then there exists a measurable set X 0 ⊆ X X_0\subseteq X X 0 ⊆ X with μ ( X 0 ) = 1 \mu(X_0)=1 μ ( X 0 ) = 1 such that the following hold for every x ∈ X 0 x\in X_0 x ∈ X 0 .
(i) The translation y ↦ τ i y y\mapsto\tau_i y y ↦ τ i y is ergodic on ( M i , m M i ) (M_i,m_{M_i}) ( M i , m M i ) , and π i ( T i n x ) = ( ω i , τ i n y i ) \pi_i(T_i^n x)=(\omega_i,\tau_i^n y_i) π i ( T i n x ) = ( ω i , τ i n y i ) for every i ∈ [ ℓ ] i\in[\ell] i ∈ [ ℓ ] and n ∈ Z n\in\mathbb{Z} n ∈ Z .
(ii) The probability measures
λ N , x = 1 N ∑ n = 1 N δ ( τ 1 p 1 ( n ) y 1 , … , τ ℓ p ℓ ( n ) y ℓ ) \lambda_{N,x}=\frac{1}{N}\sum_{n=1}^{N}\delta_{(\tau_1^{p_1(n)}y_1,\ldots,\tau_\ell^{p_\ell(n)}y_\ell)} λ N , x = N 1 n = 1 ∑ N δ ( τ 1 p 1 ( n ) y 1 , … , τ ℓ p ℓ ( n ) y ℓ ) converge weakly on ∏ i = 1 ℓ M i \prod_{i=1}^{\ell}M_i ∏ i = 1 ℓ M i to a probability measure λ x \lambda_x λ x , where δ z \delta_z δ z denotes the Dirac probability measure at z z z .
(iii) For every c i ∈ C i c_i\in C_i c i ∈ C i , i ∈ [ ℓ ] i\in[\ell] i ∈ [ ℓ ] , and Φ ∈ C ( ∏ i = 1 ℓ M i ) \Phi\in C\left(\prod_{i=1}^{\ell}M_i\right) Φ ∈ C ( ∏ i = 1 ℓ M i ) ,
∫ Φ ( c 1 u 1 , … , c ℓ u ℓ ) d λ x ( u ) = ∫ Φ ( u 1 , … , u ℓ ) d λ x ( u ) . \int\Phi(c_1u_1,\ldots,c_\ell u_\ell)\,d\lambda_x(u)=\int\Phi(u_1,\ldots,u_\ell)\,d\lambda_x(u). ∫ Φ ( c 1 u 1 , … , c ℓ u ℓ ) d λ x ( u ) = ∫ Φ ( u 1 , … , u ℓ ) d λ x ( u ) .
Proof. By the bundle definition and the countability of the pairs ( i , n ) (i,n) ( i , n ) , there is a measurable set X ∗ ⊆ X X_\ast\subseteq X X ∗ ⊆ X with μ ( X ∗ ) = 1 \mu(X_\ast)=1 μ ( X ∗ ) = 1 such that, for every x ∈ X ∗ x\in X_\ast x ∈ X ∗ and i ∈ [ ℓ ] i\in[\ell] i ∈ [ ℓ ] , the translation y ↦ τ i y y\mapsto\tau_i y y ↦ τ i y is ergodic on ( M i , m M i ) (M_i,m_{M_i}) ( M i , m M i ) and π i ( T i n x ) = S i n π i ( x ) \pi_i(T_i^n x)=S_i^n\pi_i(x) π i ( T i n x ) = S i n π i ( x ) for all n ∈ Z n\in\mathbb{Z} n ∈ Z . This proves (i) on X ∗ X_\ast X ∗ .
For each x ∈ X ∗ x\in X_\ast x ∈ X ∗ , put R i = G y i ∘ R_i=G_{y_i}^{\circ} R i = G y i ∘ . By [23 Theorem B and Section 1.9] , there are an integer r ≥ 1 r\geq1 r ≥ 1 , a connected closed subgroup H ≤ ∏ i = 1 ℓ G i ∘ H\leq\prod_{i=1}^{\ell}G_i^{\circ} H ≤ ∏ i = 1 ℓ G i ∘ , and points z b = ( z b , 1 , … , z b , ℓ ) ∈ ∏ i = 1 ℓ M i z_b=(z_{b,1},\ldots,z_{b,\ell})\in\prod_{i=1}^{\ell}M_i z b = ( z b , 1 , … , z b , ℓ ) ∈ ∏ i = 1 ℓ M i with compact orbits H z b Hz_b H z b , 1 ≤ b ≤ r 1\leq b\leq r 1 ≤ b ≤ r , such that
λ N , x ⟶ λ x = 1 r ∑ b = 1 r m H z b weakly . \lambda_{N,x}\longrightarrow\lambda_x=\frac{1}{r}\sum_{b=1}^{r}m_{Hz_b}\quad\text{weakly}. λ N , x ⟶ λ x = r 1 b = 1 ∑ r m H z b weakly .
This proves (ii) on X ∗ X_\ast X ∗ . We shall prove (iii) on a measurable subset of X ∗ X_\ast X ∗ of full measure.
For every connected component R R R of M i , a M_{i,a} M i , a , let m R m_R m R be its normalized Haar measure. For fixed i i i , a a a and v ∈ C ( M i , a ) v\in C(M_{i,a}) v ∈ C ( M i , a ) , define two functions on X X X by
v ^ ( x ) = v ( y ) , v 0 ( x ) = ∫ R v d m R if π i ( x ) = ( ω , y ) ∈ Ω i , a × R . \widehat v(x)=v(y),\qquad v_0(x)=\int_R v\,dm_R\quad\text{if }\pi_i(x)=(\omega,y)\in\Omega_{i,a}\times R. v ( x ) = v ( y ) , v 0 ( x ) = ∫ R v d m R if π i ( x ) = ( ω , y ) ∈ Ω i , a × R .
Both functions are zero when π i ( x ) ∉ Ω i , a × M i , a \pi_i(x)\notin\Omega_{i,a}\times M_{i,a} π i ( x ) ∈ / Ω i , a × M i , a . The pair i , a i,a i , a is fixed whenever this notation is used.
Claim 1. For every such v v v ,
E ( v ^ ∣ K r a t ( T i ) ) = v 0 . \mathbb{E}(\widehat v\mid\mathcal{K}_{\mathrm{rat}}(T_i))=v_0. E ( v ∣ K rat ( T i )) = v 0 .
Let τ \tau τ be an ergodic translation on M = G / Γ M=G/\Gamma M = G /Γ , and let k k k be the number of its connected components. Since τ \tau τ is ergodic, no proper nonempty union of connected components can be τ \tau τ -invariant. Hence, for any connected component R R R , M = ⋃ h = 0 k − 1 τ h R M=\bigcup_{h=0}^{k-1}\tau^hR M = ⋃ h = 0 k − 1 τ h R . Thus τ k R = R \tau^kR=R τ k R = R and m M ( R ) = 1 / k m_M(R)=1/k m M ( R ) = 1/ k for every connected component R R R . If D ⊆ R D\subseteq R D ⊆ R is τ k \tau^k τ k -invariant, then
D ′ = ⋃ h = 0 k − 1 τ h D , m M ( τ D ′ △ D ′ ) = 0 , m R ( D ) = m M ( D ′ ) ∈ { 0 , 1 } . D'=\bigcup_{h=0}^{k-1}\tau^hD,\qquad m_M(\tau D'\mathbin{\triangle}D')=0,\qquad m_R(D)=m_M(D')\in\{0,1\}. D ′ = h = 0 ⋃ k − 1 τ h D , m M ( τ D ′ △ D ′ ) = 0 , m R ( D ) = m M ( D ′ ) ∈ { 0 , 1 } .
Hence τ k ∣ R \tau^k|_R τ k ∣ R is ergodic. For R = G ∘ g Γ / Γ R=G^\circ g\Gamma/\Gamma R = G ∘ g Γ/Γ , the open subgroup G R = ⟨ G ∘ , τ k ⟩ G_R=\langle G^\circ,\tau^k\rangle G R = ⟨ G ∘ , τ k ⟩ gives the identification
G R / ( G R ∩ g Γ g − 1 ) ⟶ R , h ( G R ∩ g Γ g − 1 ) ⟼ h g Γ . G_R/(G_R\cap g\Gamma g^{-1})\longrightarrow R,\qquad h(G_R\cap g\Gamma g^{-1})\longmapsto hg\Gamma. G R / ( G R ∩ g Γ g − 1 ) ⟶ R , h ( G R ∩ g Γ g − 1 ) ⟼ h g Γ.
Thus τ k ∣ R \tau^k|_R τ k ∣ R is a nilrotation on a connected space and is totally ergodic by [10 Proposition 2.1(i)] .
Let k k k be the number of components of M i , a M_{i,a} M i , a , and fix m ≥ k m\geq k m ≥ k . Since k ∣ m ! k\mid m! k ∣ m ! , the preceding argument shows that τ i , a ( ω ) m ! ∣ R \tau_{i,a}(\omega)^{m!}|_R τ i , a ( ω ) m ! ∣ R is ergodic for almost every ω \omega ω . For almost every ω ∈ Ω i , a \omega\in\Omega_{i,a} ω ∈ Ω i , a and each connected component R R R , the mean ergodic theorem gives
∫ R ∣ 1 N ∑ n = 1 N v ( τ i , a ( ω ) m ! n y ) − ∫ R v d m R ∣ 2 d m R ( y ) ⟶ 0. \int_R\left|\frac{1}{N}\sum_{n=1}^{N}v\bigl(\tau_{i,a}(\omega)^{m!n}y\bigr)-\int_Rv\,dm_R\right|^2dm_R(y)\longrightarrow0. ∫ R N 1 n = 1 ∑ N v ( τ i , a ( ω ) m ! n y ) − ∫ R v d m R 2 d m R ( y ) ⟶ 0.
This integral is bounded by 4 ∥ v ∥ ∞ 2 4\lVert v\rVert_\infty^2 4 ∥ v ∥ ∞ 2 . Using ( π i ) ∗ μ = ν i (\pi_i)_*\mu=\nu_i ( π i ) ∗ μ = ν i , m i , a ∣ R = m i , a ( R ) m R m_{i,a}|_R=m_{i,a}(R)m_R m i , a ∣ R = m i , a ( R ) m R and dominated convergence, we obtain
∥ 1 N ∑ n = 1 N T i m ! n v ^ − v 0 ∥ L 2 ( μ ) 2 = ∑ R ∫ π i − 1 ( Ω i , a × R ) ∣ 1 N ∑ n = 1 N T i m ! n v ^ ( x ) − ∫ R v d m R ∣ 2 d μ ( x ) = ∑ R ∫ Ω i , a ∫ R ∣ 1 N ∑ n = 1 N v ( τ i , a ( ω ) m ! n y ) − ∫ R v d m R ∣ 2 d m i , a ( y ) d β i , a ( ω ) = ∑ R m i , a ( R ) ∫ Ω i , a ∫ R ∣ 1 N ∑ n = 1 N v ( τ i , a ( ω ) m ! n y ) − ∫ R v d m R ∣ 2 d m R ( y ) d β i , a ( ω ) ⟶ 0. \begin{aligned}
\left\|\frac{1}{N}\sum_{n=1}^{N}T_i^{m!n}\widehat v-v_0\right\|_{L^2(\mu)}^2
&=\sum_R\int_{\pi_i^{-1}(\Omega_{i,a}\times R)}\left|\frac{1}{N}\sum_{n=1}^{N}T_i^{m!n}\widehat v(x)-\int_Rv\,\mathrm{d}m_R\right|^2\mathrm{d}\mu(x)\\
&=\sum_R\int_{\Omega_{i,a}}\int_R\left|\frac{1}{N}\sum_{n=1}^{N}v(\tau_{i,a}(\omega)^{m!n}y)-\int_Rv\,\mathrm{d}m_R\right|^2\mathrm{d}m_{i,a}(y)\mathrm{d}\beta_{i,a}(\omega)\\
&=\sum_Rm_{i,a}(R)\int_{\Omega_{i,a}}\int_R\left|\frac{1}{N}\sum_{n=1}^{N}v(\tau_{i,a}(\omega)^{m!n}y)-\int_Rv\,\mathrm{d}m_R\right|^2\mathrm{d}m_R(y)\mathrm{d}\beta_{i,a}(\omega)\\
&\longrightarrow0.
\end{aligned} N 1 n = 1 ∑ N T i m ! n v − v 0 L 2 ( μ ) 2 = R ∑ ∫ π i − 1 ( Ω i , a × R ) N 1 n = 1 ∑ N T i m ! n v ( x ) − ∫ R v d m R 2 d μ ( x ) = R ∑ ∫ Ω i , a ∫ R N 1 n = 1 ∑ N v ( τ i , a ( ω ) m ! n y ) − ∫ R v d m R 2 d m i , a ( y ) d β i , a ( ω ) = R ∑ m i , a ( R ) ∫ Ω i , a ∫ R N 1 n = 1 ∑ N v ( τ i , a ( ω ) m ! n y ) − ∫ R v d m R 2 d m R ( y ) d β i , a ( ω ) ⟶ 0.
Here R R R ranges over the k k k components. Applying the mean ergodic theorem on X X X therefore gives
E ( v ∣ I ( T i m ! ) ) = v 0 ( m ≥ k ) . \mathbb{E}(v\mid\mathcal{I}(T_i^{m!}))=v_0\qquad(m\geq k). E ( v ∣ I ( T i m ! )) = v 0 ( m ≥ k ) .
Since K r a t ( T i ) = ⋁ m ≥ 1 I ( T i m ! ) \mathcal{K}_{\mathrm{rat}}(T_i)=\bigvee_{m\geq1}\mathcal{I}(T_i^{m!}) K rat ( T i ) = ⋁ m ≥ 1 I ( T i m ! ) , convergence of increasing conditional expectations proves Claim 1.
Claim 2. There is a measurable set X 0 ⊆ X ∗ X_0\subseteq X_* X 0 ⊆ X ∗ with μ ( X 0 ) = 1 \mu(X_0)=1 μ ( X 0 ) = 1 such that, for every x ∈ X 0 x\in X_0 x ∈ X 0 and every nonempty J ⊆ [ ℓ ] J\subseteq[\ell] J ⊆ [ ℓ ] for which ( p i ) i ∈ J (p_i)_{i\in J} ( p i ) i ∈ J is linearly independent,
∏ i ∈ J R i ⊆ pr J ( supp λ x ) , pr J ( H ) = ∏ i ∈ J G i ∘ . \prod_{i\in J}R_i\subseteq\operatorname{pr}_J(\operatorname{supp}\lambda_x),\qquad\operatorname{pr}_J(H)=\prod_{i\in J}G_i^\circ. i ∈ J ∏ R i ⊆ pr J ( supp λ x ) , pr J ( H ) = i ∈ J ∏ G i ∘ .
Here pr J \operatorname{pr}_J pr J denotes coordinate projection onto the indices in J J J .
On each M i , a M_{i,a} M i , a choose a countable base of nonempty open sets. For each basis set U U U , choose v U ∈ C ( M i , a ) v_U\in C(M_{i,a}) v U ∈ C ( M i , a ) with 0 ≤ v U ≤ 1 0\leq v_U\leq1 0 ≤ v U ≤ 1 , v U ≢ 0 v_U\not\equiv0 v U ≡ 0 , and supp v U ⊆ U \operatorname{supp}v_U\subseteq U supp v U ⊆ U . Write T i , a \mathcal{T}_{i,a} T i , a for this countable family.
Fix a nonempty set J J J for which ( p i ) i ∈ J (p_i)_{i\in J} ( p i ) i ∈ J is linearly independent. For i ∈ J i\in J i ∈ J , fix a i ∈ I i a_i\in I_i a i ∈ I i and v i ∈ T i , a i v_i\in\mathcal{T}_{i,a_i} v i ∈ T i , a i , and let v ^ i , v i , 0 \widehat v_i,v_{i,0} v i , v i , 0 be the functions from Claim 1 for the model ( i , a i ) (i,a_i) ( i , a i ) . Put
q = ∏ i ∈ J ∣ Comp ( M i , a i ) ∣ , w = ∏ i ∈ J v i , 0 , B N = 1 N ∑ n = 1 N ∏ i ∈ J T i p i ( n ) v ^ i . q=\prod_{i\in J}\lvert\operatorname{Comp}(M_{i,a_i})\rvert,\qquad
w=\prod_{i\in J}v_{i,0},\qquad
B_N=\frac{1}{N}\sum_{n=1}^{N}\prod_{i\in J}T_i^{p_i(n)}\widehat v_i. q = i ∈ J ∏ ∣ Comp ( M i , a i )∣ , w = i ∈ J ∏ v i , 0 , B N = N 1 n = 1 ∑ N i ∈ J ∏ T i p i ( n ) v i .
Since p i ( 0 ) = 0 p_i(0)=0 p i ( 0 ) = 0 , q ∣ p i ( q n ) q\mid p_i(qn) q ∣ p i ( q n ) . The component cycle calculation in Claim 1 and the identities on X ∗ X_* X ∗ give v i , 0 ( T i p i ( q n ) x ) = v i , 0 ( x ) v_{i,0}(T_i^{p_i(qn)}x)=v_{i,0}(x) v i , 0 ( T i p i ( q n ) x ) = v i , 0 ( x ) for x ∈ X ∗ x\in X_* x ∈ X ∗ . It is clear that the polynomials remain linearly independent after replacing t t t by q t qt q t and then by Claim 1 and [13 Theorem 2.10] ,
B N ( q ) : = 1 N ∑ n = 1 N ∏ i ∈ J T i p i ( q n ) v ^ i ⟶ w in L 2 ( μ ) . B_N^{(q)}:=\frac{1}{N}\sum_{n=1}^{N}\prod_{i\in J}T_i^{p_i(qn)}\widehat v_i\longrightarrow w\quad\text{in }L^2(\mu). B N ( q ) := N 1 n = 1 ∑ N i ∈ J ∏ T i p i ( q n ) v i ⟶ w in L 2 ( μ ) .
Define the measurable set E J a ( v ) = { x ∈ X ∗ : w ( x ) > 0 , lim N → ∞ B N ( x ) = 0 } E_J^{\mathbf a}(\mathbf v)=\{x\in X_*:w(x)>0,\ \lim_{N\to\infty}B_N(x)=0\} E J a ( v ) = { x ∈ X ∗ : w ( x ) > 0 , lim N → ∞ B N ( x ) = 0 } , where a = ( a i ) i ∈ J \mathbf a=(a_i)_{i\in J} a = ( a i ) i ∈ J and v = ( v i ) i ∈ J \mathbf v=(v_i)_{i\in J} v = ( v i ) i ∈ J . For every x ∈ E J a ( v ) x\in E_J^{\mathbf a}(\mathbf v) x ∈ E J a ( v ) , nonnegativity gives 0 ≤ B N ( q ) ( x ) ≤ q B q N ( x ) → 0 0\leq B_N^{(q)}(x)\leq qB_{qN}(x)\to0 0 ≤ B N ( q ) ( x ) ≤ q B q N ( x ) → 0 . Then by Fatou’s lemma
∫ E J a ( v ) ∣ w ∣ 2 d μ ≤ lim inf N → ∞ ∥ B N ( q ) − w ∥ 2 2 = 0. \int_{E_J^{\mathbf a}(\mathbf v)}\lvert w\rvert^2\,\mathrm{d}\mu\leq\liminf_{N\to\infty}\lVert B_N^{(q)}-w\rVert_2^2=0. ∫ E J a ( v ) ∣ w ∣ 2 d μ ≤ N → ∞ lim inf ∥ B N ( q ) − w ∥ 2 2 = 0.
Since w > 0 w>0 w > 0 on E J a ( v ) E_J^{\mathbf a}(\mathbf v) E J a ( v ) , μ ( E J a ( v ) ) = 0 \mu(E_J^{\mathbf a}(\mathbf v))=0 μ ( E J a ( v )) = 0 . Let
J = { J ⊆ [ ℓ ] : ( p i ) i ∈ J is linearly independent over Q , J ≠ ∅ } . \mathcal{J}=\{J\subseteq[\ell]:(p_i)_{i\in J}\text{ is linearly independent over }\mathbb{Q},\ J\neq\varnothing\}. J = { J ⊆ [ ℓ ] : ( p i ) i ∈ J is linearly independent over Q , J = ∅ } .
Hence J \mathcal{J} J is finite, and then let
X 0 = X ∗ ∖ ⋃ J ∈ J , a i ∈ I i ( i ∈ J ) v i ∈ T i , a i ( i ∈ J ) E J a ( v ) . X_0=X_*\setminus\bigcup_{\substack{J\in\mathcal{J},\,a_i\in I_i\ (i\in J)\\v_i\in\mathcal{T}_{i,a_i}\ (i\in J)}}E_J^a(\mathbf{v}). X 0 = X ∗ ∖ J ∈ J , a i ∈ I i ( i ∈ J ) v i ∈ T i , a i ( i ∈ J ) ⋃ E J a ( v ) .
The union is countable, so X 0 X_0 X 0 is a measurable set with μ ( X 0 ) = 1 \mu(X_0)=1 μ ( X 0 ) = 1 .
Fix x ∈ X 0 x\in X_0 x ∈ X 0 and J ∈ J J\in\mathcal{J} J ∈ J . If the asserted support inclusion fails, there are nonempty open sets U i ⊆ R i U_i\subseteq R_i U i ⊆ R i , i ∈ J i\in J i ∈ J , such that
( ∏ i ∈ J U i ) ∩ pr J ( supp λ x ) = ∅ . \left(\prod_{i\in J}U_i\right)\cap\operatorname{pr}_J(\operatorname{supp}\lambda_x)=\varnothing. ( i ∈ J ∏ U i ) ∩ pr J ( supp λ x ) = ∅ .
Choose v i ∈ T i , a i v_i\in\mathcal{T}_{i,a_i} v i ∈ T i , a i with supp v i ⊆ U i \operatorname{supp}v_i\subseteq U_i supp v i ⊆ U i . The Haar measures m R i m_{R_i} m R i give positive measure to every nonempty open set, so w ( x ) = ∏ i ∈ J ∫ R i v i d m R i > 0 w(x)=\prod_{i\in J}\int_{R_i}v_i\,dm_{R_i}>0 w ( x ) = ∏ i ∈ J ∫ R i v i d m R i > 0 . On the other hand, the product ∏ i ∈ J v i ( u i ) \prod_{i\in J}v_i(u_i) ∏ i ∈ J v i ( u i ) vanishes on supp λ x \operatorname{supp}\lambda_x supp λ x . Using π i ( T i n x ) = S i n π i ( x ) \pi_i(T_i^n x)=S_i^n\pi_i(x) π i ( T i n x ) = S i n π i ( x ) and λ N , x → λ x \lambda_{N,x}\to\lambda_x λ N , x → λ x ,
B N ( x ) = 1 N ∑ n = 1 N ∏ i ∈ J v i ( τ i p i ( n ) y i ) = ∫ ∏ i ∈ J v i ( u i ) d λ N , x ( u ) B_N(x)=\frac{1}{N}\sum_{n=1}^{N}\prod_{i\in J}v_i(\tau_i^{p_i(n)}y_i)=\int\prod_{i\in J}v_i(u_i)\,d\lambda_{N,x}(u) B N ( x ) = N 1 n = 1 ∑ N i ∈ J ∏ v i ( τ i p i ( n ) y i ) = ∫ i ∈ J ∏ v i ( u i ) d λ N , x ( u )
⟶ ∫ ∏ i ∈ J v i ( u i ) d λ x ( u ) = 0. \longrightarrow\int\prod_{i\in J}v_i(u_i)\,d\lambda_x(u)=0. ⟶ ∫ i ∈ J ∏ v i ( u i ) d λ x ( u ) = 0.
Thus x ∈ E J a ( v ) x\in E_J^a(\mathbf{v}) x ∈ E J a ( v ) , a contradiction. This proves the support inclusion.
We now prove the group equality. Put G J = ∏ i ∈ J G i ∘ G_J=\prod_{i\in J}G_i^\circ G J = ∏ i ∈ J G i ∘ , K = pr J ( H ) K=\operatorname{pr}_J(H) K = pr J ( H ) , R J = ∏ i ∈ J R i R_J=\prod_{i\in J}R_i R J = ∏ i ∈ J R i , and y J = ( y i ) i ∈ J y_J=(y_i)_{i\in J} y J = ( y i ) i ∈ J . Since each Haar measure m H z b m_{Hz_b} m H z b has full support on the compact orbit H z b Hz_b H z b ,
pr J ( supp λ x ) = ⋃ b = 1 r K z b , J , z b , J = ( z b , i ) i ∈ J . \operatorname{pr}_J(\operatorname{supp}\lambda_x)=\bigcup_{b=1}^{r}Kz_{b,J},\qquad z_{b,J}=(z_{b,i})_{i\in J}. pr J ( supp λ x ) = b = 1 ⋃ r K z b , J , z b , J = ( z b , i ) i ∈ J .
Since K K K is connected and R J R_J R J is a product of connected components,
K z b , J ∩ R J = { K z b , J , z b , J ∈ R J , ∅ , z b , J ∉ R J . Kz_{b,J}\cap R_J=
\begin{cases}
Kz_{b,J},&z_{b,J}\in R_J,\\
\varnothing,&z_{b,J}\notin R_J.
\end{cases} K z b , J ∩ R J = { K z b , J , ∅ , z b , J ∈ R J , z b , J ∈ / R J .
Then R J = R J ∩ ⋃ b = 1 r K z b , J = ⋃ 1 ≤ b ≤ r z b , J ∈ R J K z b , J R_J=R_J\cap\bigcup_{b=1}^{r}Kz_{b,J}=\bigcup_{\substack{1\le b\le r\\z_{b,J}\in R_J}}Kz_{b,J} R J = R J ∩ ⋃ b = 1 r K z b , J = ⋃ 1 ≤ b ≤ r z b , J ∈ R J K z b , J . Note that R J = G J y J R_J=G_Jy_J R J = G J y J , so for any b b b with z b , J ∈ R J z_{b,J}\in R_J z b , J ∈ R J we may choose d b ∈ G J d_b\in G_J d b ∈ G J such that d b y J = z b , J d_by_J=z_{b,J} d b y J = z b , J . Let D D D be the finite set of these elements. Then R J = ⋃ ξ ∈ D K ξ y J R_J=\bigcup_{\xi\in D}K\xi y_J R J = ⋃ ξ ∈ D K ξ y J .
Write y J = ( t i Γ i ) i ∈ J y_J=(t_i\Gamma_i)_{i\in J} y J = ( t i Γ i ) i ∈ J . The stabilizer of y J y_J y J in G J G_J G J is
Λ = { g ∈ G J : g y J = y J } = ∏ i ∈ J ( G i ∘ ∩ t i Γ i t i − 1 ) . \Lambda=\{g\in G_J:gy_J=y_J\}=\prod_{i\in J}(G_i^\circ\cap t_i\Gamma_it_i^{-1}). Λ = { g ∈ G J : g y J = y J } = i ∈ J ∏ ( G i ∘ ∩ t i Γ i t i − 1 ) .
It is discrete and countable. For g ∈ G J g\in G_J g ∈ G J , the preceding cover gives g y J = k ξ y J gy_J=k\xi y_J g y J = k ξ y J for some k ∈ K k\in K k ∈ K and ξ ∈ D \xi\in D ξ ∈ D . Thus ξ − 1 k − 1 g ∈ Λ \xi^{-1}k^{-1}g\in\Lambda ξ − 1 k − 1 g ∈ Λ , and G J = K D Λ G_J=KD\Lambda G J = K D Λ .
Choose compact sets F m ⊆ H F_m\subseteq H F m ⊆ H with H = ⋃ m ≥ 1 F m H=\bigcup_{m\ge1}F_m H = ⋃ m ≥ 1 F m . Then
G J = ⋃ m ≥ 1 , ξ ∈ D λ ∈ Λ pr J ( F m ) ξ λ . G_J=\bigcup_{\substack{m\ge1,\,\xi\in D\\\lambda\in\Lambda}}\operatorname{pr}_J(F_m)\xi\lambda. G J = m ≥ 1 , ξ ∈ D λ ∈ Λ ⋃ pr J ( F m ) ξ λ .
By Baire’s theorem, one of these compact sets has nonempty interior. Then there exists a nonempty open set U ⊆ K U\subseteq K U ⊆ K . Since 1 G J ∈ U − 1 U = ⋃ u ∈ U u − 1 U ⊆ K 1_{G_J}\in U^{-1}U=\bigcup_{u\in U}u^{-1}U\subseteq K 1 G J ∈ U − 1 U = ⋃ u ∈ U u − 1 U ⊆ K , the set U − 1 U U^{-1}U U − 1 U is an open neighborhood of 1 G J 1_{G_J} 1 G J . Hence
K = ⋃ k ∈ K k U − 1 U , G J ∖ K = ⋃ g ∈ G J ∖ K g K . K=\bigcup_{k\in K}kU^{-1}U,\qquad G_J\setminus K=\bigcup_{g\in G_J\setminus K}gK. K = k ∈ K ⋃ k U − 1 U , G J ∖ K = g ∈ G J ∖ K ⋃ g K .
Both sets are open. Since G J G_J G J is connected, K = G J K=G_J K = G J . This completes Claim 2.
Claim 3. For every x ∈ X 0 x\in X_0 x ∈ X 0 , the subgroup H H H contains ∏ i = 1 ℓ C i \prod_{i=1}^{\ell}C_i ∏ i = 1 ℓ C i . Fix x ∈ X 0 x \in X_0 x ∈ X 0 , i ∈ [ ℓ ] i \in[\ell] i ∈ [ ℓ ] , and a , b ∈ G i ∘ a,b \in G_i^\circ a , b ∈ G i ∘ . Choose J i , K i J_i,K_i J i , K i from the polynomial hypothesis. Claim 2 gives surjectivity of both pr J i ( H ) \operatorname{pr}_{J_i}(H) pr J i ( H ) and pr K i ( H ) \operatorname{pr}_{K_i}(H) pr K i ( H ) . We may therefore choose u , v ∈ H u,v \in H u , v ∈ H such that
u i = a , u j = 1 G j ( j ∈ J i ∖ { i } ) , u_i=a,\quad u_j=1_{G_j}\quad(j\in J_i\setminus\{i\}), u i = a , u j = 1 G j ( j ∈ J i ∖ { i }) ,
v i = b , v j = 1 G j ( j ∈ K i ∖ { i } ) . v_i=b,\quad v_j=1_{G_j}\quad(j\in K_i\setminus\{i\}). v i = b , v j = 1 G j ( j ∈ K i ∖ { i }) .
Since J i ∪ K i = [ ℓ ] J_i\cup K_i=[\ell] J i ∪ K i = [ ℓ ] , at least one of u j , v j u_j,v_j u j , v j is the identity for each j ≠ i j\ne i j = i . Consequently,
[ u , v ] = ( 1 G 1 , … , 1 G i − 1 , [ a , b ] , 1 G i + 1 , … , 1 G ℓ ) ∈ H . [u,v]=(1_{G_1},\ldots,1_{G_{i-1}},[a,b],1_{G_{i+1}},\ldots,1_{G_\ell})\in H. [ u , v ] = ( 1 G 1 , … , 1 G i − 1 , [ a , b ] , 1 G i + 1 , … , 1 G ℓ ) ∈ H .
These commutators generate C i C_i C i in the i i i th coordinate. Taking products over i i i gives ∏ i = 1 ℓ C i ⊆ H \prod_{i=1}^{\ell}C_i\subseteq H ∏ i = 1 ℓ C i ⊆ H .
For c i ∈ C i c_i\in C_i c i ∈ C i and Φ ∈ C ( ∏ i M i ) \Phi\in C(\prod_i M_i) Φ ∈ C ( ∏ i M i ) , by the invariance of Haar measure,
∫ Φ ( c 1 u 1 , … , c ℓ u ℓ ) d λ X ( u ) = 1 r ∑ b = 1 r ∫ H z b Φ ( c 1 u 1 , … , c ℓ u ℓ ) d m H z b ( u ) = 1 r ∑ b = 1 r ∫ H z b Φ ( u ) d m H z b ( u ) = ∫ Φ ( u ) d λ X ( u ) . \begin{aligned}
\int\Phi(c_1u_1,\ldots,c_\ell u_\ell)\,d\lambda_X(u)
&=\frac{1}{r}\sum_{b=1}^{r}\int_{H_{z_b}}\Phi(c_1u_1,\ldots,c_\ell u_\ell)\,dm_{H_{z_b}}(u)\\
&=\frac{1}{r}\sum_{b=1}^{r}\int_{H_{z_b}}\Phi(u)\,dm_{H_{z_b}}(u)
=\int\Phi(u)\,d\lambda_X(u).
\end{aligned} ∫ Φ ( c 1 u 1 , … , c ℓ u ℓ ) d λ X ( u ) = r 1 b = 1 ∑ r ∫ H z b Φ ( c 1 u 1 , … , c ℓ u ℓ ) d m H z b ( u ) = r 1 b = 1 ∑ r ∫ H z b Φ ( u ) d m H z b ( u ) = ∫ Φ ( u ) d λ X ( u ) .
This proves (iii). Since X 0 ⊆ X ∗ X_0\subseteq X_\ast X 0 ⊆ X ∗ , (i) and (ii) also hold on X 0 X_0 X 0 . □
Proof of the phase-factor theorem
Integer assertion of Theorem A. Take J i , K i J_i,K_i J i , K i from Lemma 3.1. For any distinct i i i and j j j , at least one of J i J_i J i or K i K_i K i contains both indices. Consequently, p i p_i p i and p j p_j p j are linearly independent. Choose s = s ( D , ℓ ) ≥ 1 s=s(D,\ell)\ge1 s = s ( D , ℓ ) ≥ 1 from [13 Theorem 2.8] . We begin by assuming E ( f 1 ∣ W s ( T 1 ) ) = 0 E(f_1\mid W_s(T_1))=0 E ( f 1 ∣ W s ( T 1 )) = 0 and proceed to show that this average tends to zero. By multilinearity, we may assume ∣ f i ∣ ≤ 1 |f_i|\le1 ∣ f i ∣ ≤ 1 .
Put h i = E ( f i ∣ Z s ( T i ) ) h_i=E(f_i\mid Z_s(T_i)) h i = E ( f i ∣ Z s ( T i )) . By [19 Theorem 1.2] , each Z s ( T i ) Z_s(T_i) Z s ( T i ) is generated by an increasing sequence of s s s -step nilsystem bundle factors. For ε > 0 \varepsilon>0 ε > 0 , choose this type of factor map π i : ( X , μ , T i ) → ( Y i , ν i , S i ) \pi_i:(X,\mu,T_i)\to(Y_i,\nu_i,S_i) π i : ( X , μ , T i ) → ( Y i , ν i , S i ) such that
π i − 1 Y i ⊆ Z s ( T i ) , F i = E ( f i ∣ π i − 1 Y i ) , \pi_i^{-1}\mathcal{Y}_i\subseteq Z_s(T_i),\qquad F_i=E(f_i\mid\pi_i^{-1}\mathcal{Y}_i), π i − 1 Y i ⊆ Z s ( T i ) , F i = E ( f i ∣ π i − 1 Y i ) ,
∑ i = 1 ℓ ∥ h i − F i ∥ 2 < ε , \sum_{i=1}^{\ell}\lVert h_i-F_i\rVert_2<\varepsilon, i = 1 ∑ ℓ ∥ h i − F i ∥ 2 < ε ,
where Y i \mathcal{Y}_i Y i is the sigma-algebra of Y i Y_i Y i . We identify functions on Y i Y_i Y i with their pullbacks to X X X . The choice of s s s gives ∥ A N p ( f 1 , … , f ℓ ) − A N p ( h 1 , … , h ℓ ) ∥ 2 → 0 \lVert A_N^p(f_1,\ldots,f_\ell)-A_N^p(h_1,\ldots,h_\ell)\rVert_2\to0 ∥ A N p ( f 1 , … , f ℓ ) − A N p ( h 1 , … , h ℓ ) ∥ 2 → 0 . Since ∣ h i ∣ , ∣ F i ∣ ≤ 1 |h_i|,|F_i|\le1 ∣ h i ∣ , ∣ F i ∣ ≤ 1 , expand the differences coordinate by coordinate, and using the property of measure preservation, obtain
∥ A N p ( h 1 , … , h ℓ ) − A N p ( F 1 , … , F ℓ ) ∥ 2 ≤ ∑ i = 1 ℓ ∥ h i − F i ∥ 2 . \lVert A_N^p(h_1,\ldots,h_\ell)-A_N^p(F_1,\ldots,F_\ell)\rVert_2\le\sum_{i=1}^{\ell}\lVert h_i-F_i\rVert_2. ∥ A N p ( h 1 , … , h ℓ ) − A N p ( F 1 , … , F ℓ ) ∥ 2 ≤ i = 1 ∑ ℓ ∥ h i − F i ∥ 2 .
Consequently,
lim sup N → ∞ ∥ A N p ( f 1 , … , f ℓ ) − A N p ( F 1 , … , F ℓ ) ∥ 2 < ε . \limsup_{N\to\infty}\lVert A_N^p(f_1,\ldots,f_\ell)-A_N^p(F_1,\ldots,F_\ell)\rVert_2<\varepsilon. N → ∞ lim sup ∥ A N p ( f 1 , … , f ℓ ) − A N p ( F 1 , … , F ℓ ) ∥ 2 < ε .
It remains to prove that ∥ A N p ( F 1 , … , F ℓ ) ∥ 2 → 0 \lVert A_N^p(F_1,\ldots,F_\ell)\rVert_2\to0 ∥ A N p ( F 1 , … , F ℓ ) ∥ 2 → 0 . Let V i , D i \mathcal{V}_i,\mathcal{D}_i V i , D i be the spaces in Lemma 3.4 for Y i Y_i Y i . For a unit-modulus function ϕ \phi ϕ on Y 1 Y_1 Y 1 and every q ≥ 1 q\ge1 q ≥ 1 ,
Δ T 1 s + 1 q ( ϕ ∘ π 1 ) = ( Δ S 1 s + 1 q ϕ ) ∘ π 1 . \Delta_{T_1}^{s+1}{}_q(\phi\circ\pi_1)=(\Delta_{S_1}^{s+1}{}_q\phi)\circ\pi_1. Δ T 1 s + 1 q ( ϕ ∘ π 1 ) = ( Δ S 1 s + 1 q ϕ ) ∘ π 1 .
Hence π 1 − 1 W s ( S 1 ) ⊆ W s ( T 1 ) \pi_1^{-1}W_s(S_1)\subseteq W_s(T_1) π 1 − 1 W s ( S 1 ) ⊆ W s ( T 1 ) . For h ∈ L 2 ( W s ( S 1 ) , ν 1 ) h\in L^2(W_s(S_1),\nu_1) h ∈ L 2 ( W s ( S 1 ) , ν 1 ) , it follows that
∫ Y 1 F 1 h ‾ d ν 1 = ∫ X f 1 h ∘ π 1 ‾ d μ = 0. \int_{Y_1}F_1\overline{h}\,d\nu_1=\int_X f_1\overline{h\circ\pi_1}\,d\mu=0. ∫ Y 1 F 1 h d ν 1 = ∫ X f 1 h ∘ π 1 d μ = 0.
By Lemma 3.4(iii), there are ψ k ∈ D 1 \psi_k\in\mathcal{D}_1 ψ k ∈ D 1 with $\lVert F_1-\psi_k\rVert_{L^2(\nu_1)}\to 0.
We first prove L 2 L^2 L 2 convergence to zero for V 1 ∈ D 1 V_1 \in\mathcal{D}_1 V 1 ∈ D 1 and V i ∈ V i V_i \in\mathcal{V}_i V i ∈ V i , 2 ≤ i ≤ ℓ 2 \le i \le\ell 2 ≤ i ≤ ℓ . By multilinearity and the definitions of these spaces, it suffices to take, on fixed pieces Ω i , a i × M i , a i \Omega_{i,a_i} \times M_{i,a_i} Ω i , a i × M i , a i ,
V 1 ( ω , y ) = b 1 ( ω ) ( v 1 ( c y ) − v 1 ( y ) ) , V i ( ω , y ) = b i ( ω ) v i ( y ) ( 2 ≤ i ≤ ℓ ) , V_1(\omega,y)=b_1(\omega)(v_1(cy)-v_1(y)), \qquad V_i(\omega,y)=b_i(\omega)v_i(y) \quad(2\le i\le\ell), V 1 ( ω , y ) = b 1 ( ω ) ( v 1 ( cy ) − v 1 ( y )) , V i ( ω , y ) = b i ( ω ) v i ( y ) ( 2 ≤ i ≤ ℓ ) ,
where b i ∈ L ∞ ( β i , a i ) b_i \in L^\infty(\beta_{i,a_i}) b i ∈ L ∞ ( β i , a i ) , v i ∈ C ( M i , a i ) v_i \in C(M_{i,a_i}) v i ∈ C ( M i , a i ) , and c ∈ [ G 1 , a 1 ∘ , G 1 , a 1 ∘ ] c \in[G^\circ_{1,a_1},G^\circ_{1,a_1}] c ∈ [ G 1 , a 1 ∘ , G 1 , a 1 ∘ ] . For all i i i , set V i = 0 V_i=0 V i = 0 outside Ω i , a i × M i , a i \Omega_{i,a_i}\times M_{i,a_i} Ω i , a i × M i , a i .
Let X 0 X_0 X 0 be the set in Lemma 3.5. If π i ( x ) ∉ Ω i , a i × M i , a i \pi_i(x) \notin\Omega_{i,a_i}\times M_{i,a_i} π i ( x ) ∈ / Ω i , a i × M i , a i for some i i i and x ∈ X 0 x\in X_0 x ∈ X 0 , then A N p ( V 1 , … , V ℓ ) ( x ) = 0 A_N^p(V_1,\ldots,V_\ell)(x)=0 A N p ( V 1 , … , V ℓ ) ( x ) = 0 for every N N N . We may therefore assume that π i ( x ) = ( ω i , y i ) ∈ Ω i , a i × M i , a i \pi_i(x)=(\omega_i,y_i)\in\Omega_{i,a_i}\times M_{i,a_i} π i ( x ) = ( ω i , y i ) ∈ Ω i , a i × M i , a i for every i i i . By Lemma 3.5 (ii) and (iii) we have
lim N → ∞ A N p ( V 1 , … , V ℓ ) ( x ) = ( ∏ i = 1 ℓ b i ( ω i ) ) ∫ ( v 1 ( c u 1 ) − v 1 ( u 1 ) ) ∏ i = 2 ℓ v i ( u i ) d λ x ( u ) = 0. \begin{aligned}
\lim_{N\to\infty} A_N^p(V_1,\ldots,V_\ell)(x)
=\left(\prod_{i=1}^{\ell}b_i(\omega_i)\right)\int(v_1(cu_1)-v_1(u_1))\prod_{i=2}^{\ell}v_i(u_i)\,d\lambda_x(u) \\
&=0.
\end{aligned} N → ∞ lim A N p ( V 1 , … , V ℓ ) ( x ) = ( i = 1 ∏ ℓ b i ( ω i ) ) ∫ ( v 1 ( c u 1 ) − v 1 ( u 1 )) i = 2 ∏ ℓ v i ( u i ) d λ x ( u ) = 0.
The last equality uses invariance under ( c , 1 G 2 , … , 1 G ℓ ) (c,1_{G_2},\ldots,1_{G_\ell}) ( c , 1 G 2 , … , 1 G ℓ ) . Using multilinearity and dominated convergence one can obtain the required L 2 L^2 L 2 limit for all V 1 ∈ D 1 V_1\in\mathcal{D}_1 V 1 ∈ D 1 and V i ∈ V i V_i\in\mathcal{V}_i V i ∈ V i , i ≥ 2 i\ge2 i ≥ 2 .
For any g 1 , … , g ℓ ∈ L ∞ ( μ ) g_1,\ldots,g_\ell\in L^\infty(\mu) g 1 , … , g ℓ ∈ L ∞ ( μ ) and every j ∈ [ ℓ ] j\in[\ell] j ∈ [ ℓ ] , measure preservation gives
sup N ∥ A N p ( g 1 , … , g ℓ ) ∥ 2 ≤ ∥ g j ∥ 2 ∏ i ≠ j ∥ g i ∥ ∞ . \sup_N\|A_N^p(g_1,\ldots,g_\ell)\|_2\le\|g_j\|_2\prod_{i\ne j}\|g_i\|_\infty. N sup ∥ A N p ( g 1 , … , g ℓ ) ∥ 2 ≤ ∥ g j ∥ 2 i = j ∏ ∥ g i ∥ ∞ .
Fix ψ ∈ D 1 \psi\in\mathcal{D}_1 ψ ∈ D 1 and apply Lemma 2.5 to the remaining ℓ − 1 \ell-1 ℓ − 1 coordinates. Since the spaces V i \mathcal{V}_i V i are dense in L 2 ( ν i ) L^2(\nu_i) L 2 ( ν i ) , A N p ( ψ , F 2 , … , F ℓ ) → 0 A_N^p(\psi,F_2,\ldots,F_\ell)\to0 A N p ( ψ , F 2 , … , F ℓ ) → 0 in L 2 L^2 L 2 . Taking ψ = ψ k \psi=\psi_k ψ = ψ k and using ∣ F i ∣ ≤ 1 |F_i|\le1 ∣ F i ∣ ≤ 1 yields
lim sup N → ∞ ∥ A N p ( F 1 , … , F ℓ ) ∥ 2 ≤ ∥ F 1 − ψ k ∥ 2 ⟶ 0. \limsup_{N\to\infty}\|A_N^p(F_1,\ldots,F_\ell)\|_2\le\|F_1-\psi_k\|_2\longrightarrow0. N → ∞ lim sup ∥ A N p ( F 1 , … , F ℓ ) ∥ 2 ≤ ∥ F 1 − ψ k ∥ 2 ⟶ 0.
By the preceding estimate we have lim sup N → ∞ ∥ A N p ( f 1 , … , f ℓ ) ∥ 2 ≤ ε \limsup_{N\to\infty}\|A_N^p(f_1,\ldots,f_\ell)\|_2\le\varepsilon lim sup N → ∞ ∥ A N p ( f 1 , … , f ℓ ) ∥ 2 ≤ ε . Since ε > 0 \varepsilon>0 ε > 0 is arbitrary, the limit is zero. The same argument applies to any coordinate j j j with E ( f j ∣ W s ( T j ) ) = 0 \mathbb{E}(f_j\mid\mathcal{W}_s(T_j))=0 E ( f j ∣ W s ( T j )) = 0 .
Finally let g i = E ( f i ∣ W s ( T i ) ) g_i=\mathbb{E}(f_i\mid\mathcal{W}_s(T_i)) g i = E ( f i ∣ W s ( T i )) . Then E ( f i − g i ∣ W s ( T i ) ) = 0 \mathbb{E}(f_i-g_i\mid\mathcal{W}_s(T_i))=0 E ( f i − g i ∣ W s ( T i )) = 0 , and by multilinearity we have
A N p ( f 1 , … , f ℓ ) − A N p ( g 1 , … , g ℓ ) = ∑ j = 1 ℓ A N p ( g 1 , … , g j − 1 , f j − g j , f j + 1 , … , f ℓ ) . A_N^p(f_1,\ldots,f_\ell)-A_N^p(g_1,\ldots,g_\ell)
=\sum_{j=1}^{\ell}A_N^p(g_1,\ldots,g_{j-1},f_j-g_j,f_{j+1},\ldots,f_\ell). A N p ( f 1 , … , f ℓ ) − A N p ( g 1 , … , g ℓ ) = j = 1 ∑ ℓ A N p ( g 1 , … , g j − 1 , f j − g j , f j + 1 , … , f ℓ ) .
Each term tends to zero in L 2 ( μ ) L^2(\mu) L 2 ( μ ) , as proved above. This proves the integer characteristic-factor assertion. □
Degree bounds for polynomial phases
In this section, we show the degree bounds 4 , 2 , 2 4,2,2 4 , 2 , 2 for the family t , t 2 , t 2 + t t,t^2,t^2+t t , t 2 , t 2 + t and then prove the integer upper bounds in Theorem B.
Recall the coefficientwise reduction P ↦ [ P ] P\mapsto[P] P ↦ [ P ] modulo Q \mathbb{Q} Q from § 2 . We need the Weyl’s theorem.
Lemma 4.1 (Weyl’s theorem). Let m ≥ 0 m\ge0 m ≥ 0 be an integer and P ( t ) = a 0 + a 1 t + ⋯ + a m t m ∈ R [ t ] P(t)=a_0+a_1t+\cdots+a_mt^m\in\mathbb{R}[t] P ( t ) = a 0 + a 1 t + ⋯ + a m t m ∈ R [ t ] . If
lim sup N → ∞ ∣ 1 N ∑ n = 1 N e ( P ( n ) ) ∣ > 0 , \limsup_{N\to\infty}\left|\frac{1}{N}\sum_{n=1}^{N}e(P(n))\right|>0, N → ∞ lim sup N 1 n = 1 ∑ N e ( P ( n )) > 0 , then a k ∈ Q a_k\in\mathbb{Q} a k ∈ Q for every 1 ≤ k ≤ m 1\le k\le m 1 ≤ k ≤ m . Equivalently, the polynomial [ P ] ∈ ( R / Q ) [ t ] [P]\in(\mathbb{R}/\mathbb{Q})[t] [ P ] ∈ ( R / Q ) [ t ] obtained by reducing the coefficients of P P P modulo Q \mathbb{Q} Q is constant.
Proof. If a k a_k a k is irrational for some k ≥ 1 k\ge1 k ≥ 1 , Weyl’s equidistribution theorem [26 ] gives
1 N ∑ n = 1 N e ( P ( n ) ) ⟶ 0 , \frac{1}{N}\sum_{n=1}^{N}e(P(n))\longrightarrow0, N 1 n = 1 ∑ N e ( P ( n )) ⟶ 0 , contrary to the assumption.
We apply this criterion to compute conditional expectations on phase factors.
Lemma 4.2. Let ( X , μ , S ) (X,\mu,S) ( X , μ , S ) be a measure-preserving system, let d , L ≥ 1 d,L\ge1 d , L ≥ 1 be integers, and let ϕ ∈ L ∞ ( μ ) \phi\in L^\infty(\mu) ϕ ∈ L ∞ ( μ ) satisfy ∣ ϕ ∣ = 1 |\phi|=1 ∣ ϕ ∣ = 1 and Δ S L ϕ = 1 \Delta_S^L\phi=1 Δ S L ϕ = 1 . Put
E = ⋂ k = d + 1 L − 1 { x : Δ S k ϕ ( x ) ∈ μ ∞ } , E=\bigcap_{k=d+1}^{L-1}\{x:\Delta_S^k\phi(x)\in\mu_\infty\}, E = k = d + 1 ⋂ L − 1 { x : Δ S k ϕ ( x ) ∈ μ ∞ } , with E = X E=X E = X when d ≥ L − 1 d\ge L-1 d ≥ L − 1 . Then E E E is S S S -invariant and E ( ϕ ∣ W d ( S ) ) = 1 E ϕ \mathbb{E}(\phi\mid W_d(S))=1_E\phi E ( ϕ ∣ W d ( S )) = 1 E ϕ .
Proof. If d ≥ L − 1 d\ge L-1 d ≥ L − 1 , then E = X E=X E = X and ϕ ∈ P d ( S ) \phi\in\mathcal{P}_d(S) ϕ ∈ P d ( S ) , so the conclusion holds. Suppose that d < L − 1 d<L-1 d < L − 1 . Put ψ k = Δ S k ϕ \psi_k=\Delta_S^k\phi ψ k = Δ S k ϕ and, for every r ≥ 1 r\ge1 r ≥ 1 , define E r = X S , ϕ ∩ ⋂ k = d + 1 L − 1 { x : ψ k ( x ) r ! = 1 } E_r=X_{S,\phi}\cap\bigcap_{k=d+1}^{L-1}\{x:\psi_k(x)^{r!}=1\} E r = X S , ϕ ∩ ⋂ k = d + 1 L − 1 { x : ψ k ( x ) r ! = 1 } , where X S , ϕ X_{S,\phi} X S , ϕ is the invariant conull set from Lemma 2.2 . For every x ∈ X S , ϕ x\in X_{S,\phi} x ∈ X S , ϕ , we have ψ k ( S x ) = ψ k ( x ) ψ k + 1 ( x ) \psi_k(Sx)=\psi_k(x)\psi_{k+1}(x) ψ k ( S x ) = ψ k ( x ) ψ k + 1 ( x ) and ψ L ( x ) = 1 \psi_L(x)=1 ψ L ( x ) = 1 . Descending induction on k k k gives S − 1 E r = E r S^{-1}E_r=E_r S − 1 E r = E r . Since r ! ∣ ( r + 1 ) ! r!\mid(r+1)! r ! ∣ ( r + 1 )! , we have E r ⊂ E r + 1 E_r\subset E_{r+1} E r ⊂ E r + 1 . Moreover, every root of unity has finite order, and there are only finitely many k k k . Hence E ∩ X S , ϕ = ⋃ r ≥ 1 E r E\cap X_{S,\phi}=\bigcup_{r\ge1}E_r E ∩ X S , ϕ = ⋃ r ≥ 1 E r . Since μ ( X S , ϕ ) = 1 \mu(X_{S,\phi})=1 μ ( X S , ϕ ) = 1 , E E E is S S S -invariant.
We first prove that 1 E ϕ 1_E\phi 1 E ϕ is W d ( S ) W_d(S) W d ( S ) -measurable. Fix r ≥ 1 r\ge1 r ≥ 1 and put M = r ! L ! M=r!L! M = r ! L ! . For n ∈ Z n\in\mathbb{Z} n ∈ Z and 1 ≤ k < L 1\le k<L 1 ≤ k < L ,
( M n k ) = r ! n L ! k ! ∏ b = 1 k − 1 ( M n − b ) ∈ r ! Z . \binom{Mn}{k}=r!n\frac{L!}{k!}\prod_{b=1}^{k-1}(Mn-b)\in r!\mathbb{Z}. ( k M n ) = r ! n k ! L ! b = 1 ∏ k − 1 ( M n − b ) ∈ r ! Z .
Hence, for x ∈ E r x\in E_r x ∈ E r , by Lemma 2.2 (i), ϕ ( S M n x ) = ∏ k = 0 d ψ k ( x ) ( M n k ) \phi(S^{Mn}x)=\prod_{k=0}^d\psi_k(x)^{\binom{Mn}{k}} ϕ ( S M n x ) = ∏ k = 0 d ψ k ( x ) ( k M n ) . The exponents have degree at most d d d in n n n , so Δ S M d + 1 ϕ = 1 \Delta_{S^M}^{d+1}\phi=1 Δ S M d + 1 ϕ = 1 on E r E_r E r . Since E r E_r E r is invariant, 1 E r ϕ + 1 E r c ∈ P d ( S ) 1_{E_r}\phi+1_{E_r^c}\in\mathcal{P}_d(S) 1 E r ϕ + 1 E r c ∈ P d ( S ) . Also E r ∈ I ( S ) ⊆ W d ( S ) E_r\in\mathcal{I}(S)\subseteq W_d(S) E r ∈ I ( S ) ⊆ W d ( S ) , so 1 E r ϕ ∈ L ∞ ( W d ( S ) ) 1_{E_r}\phi\in L^\infty(W_d(S)) 1 E r ϕ ∈ L ∞ ( W d ( S )) . As 1 E r ϕ → 1 E ϕ 1_{E_r}\phi\to1_E\phi 1 E r ϕ → 1 E ϕ almost everywhere, we obtain 1 E ϕ ∈ L ∞ ( W d ( S ) ) 1_E\phi\in L^\infty(W_d(S)) 1 E ϕ ∈ L ∞ ( W d ( S )) .
It remains to show that 1 E c ϕ 1_{E^c}\phi 1 E c ϕ is orthogonal to L 2 ( W d ( S ) ) L^2(W_d(S)) L 2 ( W d ( S )) . By Lemma 2.4 (iii), it suffices to take χ ∈ P d ( S ) \chi\in\mathcal{P}_d(S) χ ∈ P d ( S ) . Choose q ≥ 1 q\ge1 q ≥ 1 with Δ S q d + 1 χ = 1 \Delta_{S^q}^{d+1}\chi=1 Δ S q d + 1 χ = 1 . For almost every x ∈ E c x\in E^c x ∈ E c , Lemma 2.2 (ii) gives deg [ P S , ϕ , x ] > d \deg[P_{S,\phi,x}]>d deg [ P S , ϕ , x ] > d , while deg P S q , χ , x ≤ d \deg P_{S^q,\chi,x}\le d deg P S q , χ , x ≤ d . Therefore
deg [ P S , ϕ , x ( q t ) − P S q , χ , x ( t ) ] = deg [ P S , ϕ , x ] > d . \deg[P_{S,\phi,x}(qt)-P_{S^q,\chi,x}(t)]=\deg[P_{S,\phi,x}]>d. deg [ P S , ϕ , x ( q t ) − P S q , χ , x ( t )] = deg [ P S , ϕ , x ] > d .
By Lemma 2.2 (i) and Weyl’s theorem [26 ] ,
1 N ∑ n = 1 N ϕ ( S q n x ) χ ( S q n x ) ‾ = 1 N ∑ n = 1 N e ( P S , ϕ , x ( q n ) − P S q , χ , x ( n ) ) ⟶ 0. \frac{1}{N}\sum_{n=1}^{N}\phi(S^{qn}x)\overline{\chi(S^{qn}x)}
=\frac{1}{N}\sum_{n=1}^{N}e(P_{S,\phi,x}(qn)-P_{S^q,\chi,x}(n))\longrightarrow0. N 1 n = 1 ∑ N ϕ ( S q n x ) χ ( S q n x ) = N 1 n = 1 ∑ N e ( P S , ϕ , x ( q n ) − P S q , χ , x ( n )) ⟶ 0.
Since E E E is invariant and ∣ ϕ χ ‾ ∣ = 1 |\phi\overline{\chi}|=1 ∣ ϕ χ ∣ = 1 , by dominated convergence, we have
∫ X 1 E c ϕ χ ‾ d μ = lim N → ∞ ∫ X 1 E c ( x ) 1 N ∑ n = 1 N ϕ ( S q n x ) χ ( S q n x ) ‾ d μ ( x ) = 0. \int_X1_{E^c}\phi\overline{\chi}\,d\mu
=\lim_{N\to\infty}\int_X1_{E^c}(x)\frac{1}{N}\sum_{n=1}^{N}\phi(S^{qn}x)\overline{\chi(S^{qn}x)}\,d\mu(x)=0. ∫ X 1 E c ϕ χ d μ = N → ∞ lim ∫ X 1 E c ( x ) N 1 n = 1 ∑ N ϕ ( S q n x ) χ ( S q n x ) d μ ( x ) = 0.
Thus 1 E c ϕ ⊥ L 2 ( W d ( S ) ) 1_{E^c}\phi\perp L^2(W_d(S)) 1 E c ϕ ⊥ L 2 ( W d ( S )) . Together with the measurability of 1 E ϕ 1_E\phi 1 E ϕ , this proves the formula. □
Lemma 4.3. Let V V V be a vector space over Q \mathbb{Q} Q , let δ ∈ Q ∖ { 0 } \delta\in\mathbb{Q}\setminus\{0\} δ ∈ Q ∖ { 0 } , and let P 0 , P δ ∈ V [ u ] P_0,P_\delta\in V[u] P 0 , P δ ∈ V [ u ] and Q ∈ V [ u , w ] Q\in V[u,w] Q ∈ V [ u , w ] be formal polynomials with coefficients in V V V , with deg 0 = − ∞ \deg0=-\infty deg 0 = − ∞ . Suppose that, for β ∈ { 0 , δ } \beta\in\{0,\delta\} β ∈ { 0 , δ } , the polynomial
P β ( u + t ) + Q ( u , β + t 2 + t ) P_\beta(u+t)+Q(u,\beta+t^2+t) P β ( u + t ) + Q ( u , β + t 2 + t ) is unchanged when t t t is replaced by − t -t − t . Then
deg P 0 ≤ 4 , deg P δ ≤ 4 , deg w Q ≤ 2. \deg P_0\le4,\qquad\deg P_\delta\le4,\qquad\deg_w Q\le2. deg P 0 ≤ 4 , deg P δ ≤ 4 , deg w Q ≤ 2.
Proof. For β ∈ { 0 , δ } \beta\in\{0,\delta\} β ∈ { 0 , δ } , put
H β ( u , t ) = P β ( u + t ) + Q ( u , β + t 2 + t ) , H_\beta(u,t)=P_\beta(u+t)+Q(u,\beta+t^2+t), H β ( u , t ) = P β ( u + t ) + Q ( u , β + t 2 + t ) ,
R β ( u , w ) = ( 4 ( w − β ) + 1 ) Q w w ( u , w ) + 2 Q w ( u , w ) − Q u u ( u , w ) . R_\beta(u,w)=(4(w-\beta)+1)Q_{ww}(u,w)+2Q_w(u,w)-Q_{uu}(u,w). R β ( u , w ) = ( 4 ( w − β ) + 1 ) Q w w ( u , w ) + 2 Q w ( u , w ) − Q uu ( u , w ) .
The chain rule gives
( ∂ t 2 − ∂ u 2 ) H β ( u , t ) = R β ( u , β + t 2 + t ) . (\partial_t^2-\partial_u^2)H_\beta(u,t)=R_\beta(u,\beta+t^2+t). ( ∂ t 2 − ∂ u 2 ) H β ( u , t ) = R β ( u , β + t 2 + t ) .
The left-hand side is even in t t t . If R β R_\beta R β had positive w w w -degree m m m with leading coefficient a m ( u ) ≠ 0 a_m(u) \ne0 a m ( u ) = 0 , the coefficient of t 2 m − 1 t^{2m-1} t 2 m − 1 on the right would be m a m ( u ) ≠ 0 ma_m(u) \ne0 m a m ( u ) = 0 ; all lower w w w -degree terms have t t t -degree at most 2 m − 2 2m-2 2 m − 2 . Thus R 0 R_0 R 0 and R δ R_\delta R δ are independent of w w w . Since
R δ ( u , w ) − R 0 ( u , w ) = − 4 δ Q w w ( u , w ) , R_\delta(u,w)-R_0(u,w)=-4\delta Q_{ww}(u,w), R δ ( u , w ) − R 0 ( u , w ) = − 4 δ Q w w ( u , w ) ,
Q w w Q_{ww} Q w w is independent of w w w , and deg w Q ≤ 2 \deg_w Q \le2 deg w Q ≤ 2 . Consequently, Q ( u , β + t 2 + t ) Q(u,\beta+t^2+t) Q ( u , β + t 2 + t ) has no t 5 t^5 t 5 term. Taylor’s formula shows that the coefficient of t 5 t^5 t 5 in H β ( u , t ) H_\beta(u,t) H β ( u , t ) is P β ( 5 ) ( u ) / 5 ! P_\beta^{(5)}(u)/5! P β ( 5 ) ( u ) /5 ! . Evenness gives P β ( 5 ) = 0 P_\beta^{(5)}=0 P β ( 5 ) = 0 , so deg P β ≤ 4 \deg P_\beta\le4 deg P β ≤ 4 . □
Proposition 4.4. Let T 1 , T 2 , T 3 T_1,T_2,T_3 T 1 , T 2 , T 3 be invertible commuting measure-preserving transformations of a standard probability space ( X , X , μ ) (X,\mathcal{X},\mu) ( X , X , μ ) . Suppose that ϕ 1 , ϕ 2 , ϕ 3 ∈ L ∞ ( μ ) \phi_1,\phi_2,\phi_3\in L^\infty(\mu) ϕ 1 , ϕ 2 , ϕ 3 ∈ L ∞ ( μ ) and that there are integers q , L ≥ 1 q,L\ge1 q , L ≥ 1 such that, for each i = 1 , 2 , 3 i=1,2,3 i = 1 , 2 , 3 ,
∣ ϕ i ∣ = 1 , Δ T i q L ϕ i = 1. |\phi_i|=1,\qquad\Delta_{T_i^q}^{L}\phi_i=1. ∣ ϕ i ∣ = 1 , Δ T i q L ϕ i = 1. Let A N A_N A N be the average in (2) . Put ( d 1 , d 2 , d 3 ) = ( 4 , 2 , 2 ) (d_1,d_2,d_3)=(4,2,2) ( d 1 , d 2 , d 3 ) = ( 4 , 2 , 2 ) and, for each i = 1 , 2 , 3 i=1,2,3 i = 1 , 2 , 3 , define
h i = ϕ i − E ( ϕ i ∣ W d i ( T i ) ) . h_i=\phi_i-\mathbb{E}(\phi_i\mid\mathcal{W}_{d_i}(T_i)). h i = ϕ i − E ( ϕ i ∣ W d i ( T i )) . Then
∥ A N ( h 1 , ϕ 2 , ϕ 3 ) ∥ 2 ⟶ 0 , ∥ A N ( ϕ 1 , h 2 , ϕ 3 ) ∥ 2 ⟶ 0 , ∥ A N ( ϕ 1 , ϕ 2 , h 3 ) ∥ 2 ⟶ 0. \begin{aligned}
\|A_N(h_1,\phi_2,\phi_3)\|_2&\longrightarrow0,\\
\|A_N(\phi_1,h_2,\phi_3)\|_2&\longrightarrow0,\\
\|A_N(\phi_1,\phi_2,h_3)\|_2&\longrightarrow0.
\end{aligned} ∥ A N ( h 1 , ϕ 2 , ϕ 3 ) ∥ 2 ∥ A N ( ϕ 1 , h 2 , ϕ 3 ) ∥ 2 ∥ A N ( ϕ 1 , ϕ 2 , h 3 ) ∥ 2 ⟶ 0 , ⟶ 0 , ⟶ 0.
Proof. Fix j ∈ { 1 , 2 , 3 } j\in\{1,2,3\} j ∈ { 1 , 2 , 3 } . Apply Lemma 2.2 to each ( T i q , ϕ i ) (T_i^q,\phi_i) ( T i q , ϕ i ) . Set
X 0 = ⋂ g ∈ ⟨ T 1 , T 2 , T 3 ⟩ g − 1 ( ⋂ i = 1 3 X T i q , ϕ i ∩ ⋂ i , k = 1 3 { y ∈ X : T i T k y = T k T i y } ) . X_0=\bigcap_{g\in\langle T_1,T_2,T_3\rangle}g^{-1}\left(\bigcap_{i=1}^{3}X_{T_i^q,\phi_i}\cap\bigcap_{i,k=1}^{3}\{y\in X:T_iT_ky=T_kT_iy\}\right). X 0 = g ∈ ⟨ T 1 , T 2 , T 3 ⟩ ⋂ g − 1 i = 1 ⋂ 3 X T i q , ϕ i ∩ i , k = 1 ⋂ 3 { y ∈ X : T i T k y = T k T i y } .
Then μ ( X 0 ) = 1 \mu(X_0)=1 μ ( X 0 ) = 1 and g X 0 = X 0 gX_0=X_0 g X 0 = X 0 for every g ∈ ⟨ T 1 , T 2 , T 3 ⟩ g\in\langle T_1,T_2,T_3\rangle g ∈ ⟨ T 1 , T 2 , T 3 ⟩ . By Lemma 2.4 , W d j ( T j q ) = W d j ( T j ) \mathcal{W}_{d_j}(T_j^q)=\mathcal{W}_{d_j}(T_j) W d j ( T j q ) = W d j ( T j ) . Put V = R / Q V=\mathbb{R}/\mathbb{Q} V = R / Q , regarded as a Q \mathbb{Q} Q -vector space. For every y ∈ X 0 y\in X_0 y ∈ X 0 , write P i , y = P T i q , ϕ i , y P_{i,y}=P_{T_i^q,\phi_i,y} P i , y = P T i q , ϕ i , y for the polynomials from Lemma 2.2 . Lemma 2.2 (ii) and Lemma 4.2 therefore give
E ( ϕ j ∣ W d j ( T j ) ) ( y ) = 1 { deg [ P j , y ] ≤ d j } ϕ j ( y ) \mathbb{E}(\phi_j\mid\mathcal{W}_{d_j}(T_j))(y)=\mathbf{1}_{\{\deg[P_{j,y}]\le d_j\}}\phi_j(y) E ( ϕ j ∣ W d j ( T j )) ( y ) = 1 { d e g [ P j , y ] ≤ d j } ϕ j ( y )
for almost every y y y .
Let h j ( y ) = 1 { deg [ P j , y ] > d j } ϕ j ( y ) h_j(y)=\mathbf{1}_{\{\deg[P_{j,y}]>d_j\}}\phi_j(y) h j ( y ) = 1 { d e g [ P j , y ] > d j } ϕ j ( y ) for every y ∈ X 0 y\in X_0 y ∈ X 0 and set h j = 0 h_j=0 h j = 0 outside X 0 X_0 X 0 . Write ( p 1 ( t ) , p 2 ( t ) , p 3 ( t ) ) = ( t , t 2 , t 2 + t ) (p_1(t),p_2(t),p_3(t))=(t,t^2,t^2+t) ( p 1 ( t ) , p 2 ( t ) , p 3 ( t )) = ( t , t 2 , t 2 + t ) . Fix an integer 0 ≤ a < q 0\le a<q 0 ≤ a < q , put b i = p i ( − a ) b_i=p_i(-a) b i = p i ( − a ) , and define
Q i , y ( t ) = [ P i , T i b i y ( t − b i q ) ] ∈ V [ t ] . Q_{i,y}(t)=\left[P_{i,T_i^{b_i}y}\left(\frac{t-b_i}{q}\right)\right]\in V[t]. Q i , y ( t ) = [ P i , T i b i y ( q t − b i ) ] ∈ V [ t ] .
For h ∈ q Z h\in q\mathbb{Z} h ∈ q Z , Lemma 2.2 (ii) gives deg Q i , y < L \deg Q_{i,y}<L deg Q i , y < L , and Q i , T i h y ( t ) = Q i , y ( t + h ) Q_{i,T_i^h y}(t)=Q_{i,y}(t+h) Q i , T i h y ( t ) = Q i , y ( t + h ) . Since b j = p j ( − a ) b_j=p_j(-a) b j = p j ( − a ) and p j ∈ Z [ t ] p_j\in\mathbb{Z}[t] p j ∈ Z [ t ] , we have
m : = p j ( q n − a ) − b j q ∈ Z . m:=\frac{p_j(qn-a)-b_j}{q}\in\mathbb{Z}. m := q p j ( q n − a ) − b j ∈ Z .
Thus T j p j ( q n − a ) y = T j q n T j b j y T_j^{p_j(qn-a)}y=T_j^{qn}T_j^{b_j}y T j p j ( q n − a ) y = T j q n T j b j y . By Lemma 2.2 (ii), [ P j , T j p j ( q n − a ) y ] ( t ) = [ P j , T j b j y ] ( t + m ) [P_{j,T_j^{p_j(qn-a)}y}](t)=[P_{j,T_j^{b_j}y}](t+m) [ P j , T j p j ( q n − a ) y ] ( t ) = [ P j , T j b j y ] ( t + m ) . By the definition of Q j , y Q_{j,y} Q j , y ,
deg [ P j , T j p j ( q n − a ) y ] = deg [ P j , T j b j y ] = deg Q j , y . \deg[P_{j,T_j^{p_j(qn-a)}y}]=\deg[P_{j,T_j^{b_j}y}]=\deg Q_{j,y}. deg [ P j , T j p j ( q n − a ) y ] = deg [ P j , T j b j y ] = deg Q j , y .
Consequently,
h j ( T j p j ( q n − a ) y ) = 1 { deg Q j , y > d j } ϕ j ( T j p j ( q n − a ) y ) . h_j(T_j^{p_j(qn-a)}y)=\mathbf{1}_{\{\deg Q_{j,y}>d_j\}}\phi_j(T_j^{p_j(qn-a)}y). h j ( T j p j ( q n − a ) y ) = 1 { d e g Q j , y > d j } ϕ j ( T j p j ( q n − a ) y ) .
Put
B M , a ( y ) = 1 M ∑ n = 1 M h j ( T j p j ( q n − a ) y ) ∏ i ≠ j ϕ i ( T i p i ( q n − a ) y ) . B_{M,a}(y)=\frac{1}{M}\sum_{n=1}^{M}h_j(T_j^{p_j(qn-a)}y)\prod_{i\ne j}\phi_i(T_i^{p_i(qn-a)}y). B M , a ( y ) = M 1 n = 1 ∑ M h j ( T j p j ( q n − a ) y ) i = j ∏ ϕ i ( T i p i ( q n − a ) y ) .
Claim. The measurable set
E = { y ∈ X 0 : lim sup M → ∞ ∣ B M , a ( y ) ∣ > 0 } E=\{y\in X_0:\limsup_{M\to\infty}\lvert B_{M,a}(y)\rvert>0\} E = { y ∈ X 0 : M → ∞ lim sup ∣ B M , a ( y )∣ > 0 } has measure zero. Put F y ( n ) = ∑ i = 1 3 P i , T i b i y ( p i ( q n − a ) − b i q ) F_y(n)=\sum_{i=1}^{3}P_{i,T_i^{b_i}y}\left(\frac{p_i(qn-a)-b_i}{q}\right) F y ( n ) = ∑ i = 1 3 P i , T i b i y ( q p i ( q n − a ) − b i ) . For y ∈ E y\in E y ∈ E , we have deg Q j , y > d j \deg Q_{j,y}>d_j deg Q j , y > d j . Hence
B M , a ( y ) = 1 M ∑ n = 1 M ∏ i = 1 3 ϕ i ( T i p i ( q n − a ) y ) = 1 M ∑ n = 1 M e ( F y ( n ) ) , B_{M,a}(y)=\frac{1}{M}\sum_{n=1}^{M}\prod_{i=1}^{3}\phi_i\left(T_i^{p_i(qn-a)}y\right)=\frac{1}{M}\sum_{n=1}^{M}e(F_y(n)), B M , a ( y ) = M 1 n = 1 ∑ M i = 1 ∏ 3 ϕ i ( T i p i ( q n − a ) y ) = M 1 n = 1 ∑ M e ( F y ( n )) , where the last equality follows from Lemma 2.2 (i). By the definition of E E E ,
lim sup M → ∞ ∣ 1 M ∑ n = 1 M e ( F y ( n ) ) ∣ > 0. \limsup_{M\to\infty}\left\lvert\frac{1}{M}\sum_{n=1}^{M}e(F_y(n))\right\rvert>0. M → ∞ lim sup M 1 n = 1 ∑ M e ( F y ( n )) > 0. Lemma 4.1 therefore gives that [ F y ] [F_y] [ F y ] is constant. Substituting n = ( t + a ) / q n=(t+a)/q n = ( t + a ) / q , we obtain
[ F y ( t + a q ) ] = ∑ i = 1 3 [ P i , T i b i y ( p i ( t ) − b i q ) ] = ∑ i = 1 3 Q i , y ( p i ( t ) ) \left[F_y\left(\frac{t+a}{q}\right)\right]=\sum_{i=1}^{3}\left[P_{i,T_i^{b_i}y}\left(\frac{p_i(t)-b_i}{q}\right)\right]=\sum_{i=1}^{3}Q_{i,y}(p_i(t)) [ F y ( q t + a ) ] = i = 1 ∑ 3 [ P i , T i b i y ( q p i ( t ) − b i ) ] = i = 1 ∑ 3 Q i , y ( p i ( t )) = Q 1 , y ( t ) + Q 2 , y ( t 2 ) + Q 3 , y ( t 2 + t ) . =Q_{1,y}(t)+Q_{2,y}(t^2)+Q_{3,y}(t^2+t). = Q 1 , y ( t ) + Q 2 , y ( t 2 ) + Q 3 , y ( t 2 + t ) . The left-hand side is constant in t t t , since [ F y ] [F_y] [ F y ] is constant. Taking t = 0 t=0 t = 0 gives
Q 1 , y ( t ) + Q 2 , y ( t 2 ) + Q 3 , y ( t 2 + t ) = ∑ i = 1 3 Q i , y ( 0 ) . Q_{1,y}(t)+Q_{2,y}(t^2)+Q_{3,y}(t^2+t)=\sum_{i=1}^{3}Q_{i,y}(0). Q 1 , y ( t ) + Q 2 , y ( t 2 ) + Q 3 , y ( t 2 + t ) = i = 1 ∑ 3 Q i , y ( 0 ) . Then Q 1 , y ( t ) + Q 3 , y ( t 2 + t ) Q_{1,y}(t)+Q_{3,y}(t^2+t) Q 1 , y ( t ) + Q 3 , y ( t 2 + t ) is even in t t t .
Suppose that μ ( E ) > 0 \mu(E)>0 μ ( E ) > 0 . Recurrence for T 1 q T_1^q T 1 q gives δ ∈ q N \delta\in q\mathbb{N} δ ∈ q N with μ ( E ∩ T 3 − δ E ) > 0 \mu(E\cap T_3^{-\delta}E)>0 μ ( E ∩ T 3 − δ E ) > 0 . Recurrence for T 1 q T_1^q T 1 q on E ∩ T 3 − δ E E\cap T_3^{-\delta}E E ∩ T 3 − δ E gives a point x ∈ E ∩ T 3 − δ E x\in E\cap T_3^{-\delta}E x ∈ E ∩ T 3 − δ E and L L L distinct return times
0 = α 0 < α 1 < ⋯ < α L − 1 , α r ∈ q Z ≥ 0 , 0=\alpha_0<\alpha_1<\cdots<\alpha_{L-1},\qquad\alpha_r\in q\mathbb{Z}_{\geq0}, 0 = α 0 < α 1 < ⋯ < α L − 1 , α r ∈ q Z ≥ 0 , such that, x r : = T 1 α r x ∈ E ∩ T 3 − δ E x_r:=T_1^{\alpha_r}x\in E\cap T_3^{-\delta}E x r := T 1 α r x ∈ E ∩ T 3 − δ E for every 0 ≤ r < L 0\leq r<L 0 ≤ r < L . For 0 ≤ r < L 0\leq r<L 0 ≤ r < L , define
ℓ r ( u ) = ∏ 0 ≤ k < L k ≠ r u − α k α r − α k , Q ( u , w ) = ∑ r = 0 L − 1 ℓ r ( u ) Q 3 , x r ( w ) . \ell_r(u)=\prod_{\substack{0\leq k<L\\k\ne r}}\frac{u-\alpha_k}{\alpha_r-\alpha_k},\qquad Q(u,w)=\sum_{r=0}^{L-1}\ell_r(u)Q_{3,x_r}(w). ℓ r ( u ) = 0 ≤ k < L k = r ∏ α r − α k u − α k , Q ( u , w ) = r = 0 ∑ L − 1 ℓ r ( u ) Q 3 , x r ( w ) . For 0 ≤ r , s < L 0\leq r,s<L 0 ≤ r , s < L ,
ℓ r ( α s ) = { 1 , r = s , 0 , r ≠ s , Q ( α s , w ) = Q 3 , x s ( w ) . \ell_r(\alpha_s)=
\begin{cases}
1,&r=s,\\
0,&r\ne s,
\end{cases}
\qquad Q(\alpha_s,w)=Q_{3,x_s}(w). ℓ r ( α s ) = { 1 , 0 , r = s , r = s , Q ( α s , w ) = Q 3 , x s ( w ) . In particular, by α 0 = 0 \alpha_0=0 α 0 = 0 and x 0 = x x_0=x x 0 = x , one has Q ( 0 , w ) = Q 3 , x ( w ) Q(0,w)=Q_{3,x}(w) Q ( 0 , w ) = Q 3 , x ( w ) .
For β ∈ { 0 , δ } \beta\in\{0,\delta\} β ∈ { 0 , δ } , the identity Q i , T i h y ( t ) = Q i , y ( t + h ) Q_{i,T_i^h y}(t)=Q_{i,y}(t+h) Q i , T i h y ( t ) = Q i , y ( t + h ) (h ∈ q Z h\in q\mathbb{Z} h ∈ q Z ) and T 3 β x r = T 1 α r T 3 β x T_3^\beta x_r=T_1^{\alpha_r}T_3^\beta x T 3 β x r = T 1 α r T 3 β x , by commutativity, gives
Q 1 , T 3 β x ( t ) = Q 1 , T 3 β x ( α r + t ) , Q 3 , T 3 β x r ( w ) = Q 3 , x r ( w + β ) . Q_{1,T_3^\beta x}(t)=Q_{1,T_3^\beta x}(\alpha_r+t),\qquad Q_{3,T_3^\beta x_r}(w)=Q_{3,x_r}(w+\beta). Q 1 , T 3 β x ( t ) = Q 1 , T 3 β x ( α r + t ) , Q 3 , T 3 β x r ( w ) = Q 3 , x r ( w + β ) . Note that the Lagrange interpolation formula shows that a polynomial of degree less than L L L with coefficients in a Q \mathbb{Q} Q -vector space is determined by its values at L L L distinct rational points. Since Q 1 , T 3 β x ( u + t ) Q_{1,T_3^\beta x}(u+t) Q 1 , T 3 β x ( u + t ) has degree less than L L L in u u u , Lagrange interpolation gives
Q 1 , T 3 β x ( u + t ) + Q ( u , β + t 2 + t ) Q_{1,T_3^\beta x}(u+t)+Q(u,\beta+t^2+t) Q 1 , T 3 β x ( u + t ) + Q ( u , β + t 2 + t ) = ∑ r = 0 L − 1 ℓ r ( u ) ( Q 1 , T 3 β x ( t ) + Q 3 , T 3 β x r ( t 2 + t ) ) . =\sum_{r=0}^{L-1}\ell_r(u)\left(Q_{1,T_3^\beta x}(t)+Q_{3,T_3^\beta x_r}(t^2+t)\right). = r = 0 ∑ L − 1 ℓ r ( u ) ( Q 1 , T 3 β x ( t ) + Q 3 , T 3 β x r ( t 2 + t ) ) .
Each bracketed polynomial is even in t t t , since T 3 β x r ∈ E T_3^\beta x_r \in E T 3 β x r ∈ E . The coefficients ℓ r ( u ) \ell_r(u) ℓ r ( u ) do not depend on t t t , so the left-hand side is even for both β = 0 \beta= 0 β = 0 and β = δ \beta= \delta β = δ . Applying Lemma 4.3 with P 0 = Q 1 , x P_0 = Q_{1,x} P 0 = Q 1 , x and P δ = Q 1 , T 3 δ x P_\delta= Q_{1,T_3^\delta x} P δ = Q 1 , T 3 δ x gives
deg Q 1 , x ≤ 4 , deg w Q ≤ 2. \deg Q_{1,x} \le4,\qquad\deg_w Q \le2. deg Q 1 , x ≤ 4 , deg w Q ≤ 2.
Since Q ( 0 , w ) = Q 3 , x ( w ) Q(0,w) = Q_{3,x}(w) Q ( 0 , w ) = Q 3 , x ( w ) , we have deg Q 3 , x ≤ 2 \deg Q_{3,x} \le2 deg Q 3 , x ≤ 2 . Also, the identity at x x x gives
Q 2 , x ( t 2 ) = ∑ i = 1 3 Q i , x ( 0 ) − Q 1 , x ( t ) − Q 3 , x ( t 2 + t ) . Q_{2,x}(t^2) = \sum_{i=1}^{3} Q_{i,x}(0) - Q_{1,x}(t) - Q_{3,x}(t^2+t). Q 2 , x ( t 2 ) = i = 1 ∑ 3 Q i , x ( 0 ) − Q 1 , x ( t ) − Q 3 , x ( t 2 + t ) .
Its right-hand side has degree at most four, so deg Q 2 , x ≤ 2 \deg Q_{2,x} \le2 deg Q 2 , x ≤ 2 . Thus deg Q j , x ≤ d j \deg Q_{j,x} \le d_j deg Q j , x ≤ d j , contrary to x ∈ E x \in E x ∈ E . This proves the claim.
For each 0 ≤ a < q 0 \le a < q 0 ≤ a < q , the claim gives B M , a ( y ) ⟶ 0 B_{M,a}(y) \longrightarrow0 B M , a ( y ) ⟶ 0 for almost every y y y . Since there are only finitely many values of a a a , there is a conull set on which these limits hold for all a a a . Let
J N ( y ) = 1 N ∑ n = 1 N h j ( T j p j ( n ) y ) ∏ i ≠ j ϕ i ( T i p i ( n ) y ) . J_N(y) = \frac{1}{N}\sum_{n=1}^{N} h_j(T_j^{p_j(n)}y)\prod_{i\ne j}\phi_i(T_i^{p_i(n)}y). J N ( y ) = N 1 n = 1 ∑ N h j ( T j p j ( n ) y ) i = j ∏ ϕ i ( T i p i ( n ) y ) .
For every M ≥ 1 M \ge1 M ≥ 1 , the integers q n − a qn-a q n − a , where 1 ≤ n ≤ M 1 \le n \le M 1 ≤ n ≤ M and 0 ≤ a < q 0 \le a < q 0 ≤ a < q , are 1 , … , q M 1,\ldots,qM 1 , … , q M . Hence
J q M ( y ) = 1 q ∑ a = 0 q − 1 B M , a ( y ) ⟶ 0 J_{qM}(y) = \frac{1}{q}\sum_{a=0}^{q-1} B_{M,a}(y) \longrightarrow0 J q M ( y ) = q 1 a = 0 ∑ q − 1 B M , a ( y ) ⟶ 0
for almost every y y y .
For N ≥ q N \ge q N ≥ q , put M = ⌊ N / q ⌋ M = \lfloor N/q\rfloor M = ⌊ N / q ⌋ , so that 0 ≤ N − q M < q 0 \le N-qM < q 0 ≤ N − q M < q . Then
J N ( y ) = q M N J q M ( y ) + 1 N ∑ n = q M + 1 N h j ( T j p j ( n ) y ) ∏ i ≠ j ϕ i ( T i p i ( n ) y ) . J_N(y) = \frac{qM}{N}J_{qM}(y) + \frac{1}{N}\sum_{n=qM+1}^{N} h_j(T_j^{p_j(n)}y)\prod_{i\ne j}\phi_i(T_i^{p_i(n)}y). J N ( y ) = N q M J q M ( y ) + N 1 n = q M + 1 ∑ N h j ( T j p j ( n ) y ) i = j ∏ ϕ i ( T i p i ( n ) y ) .
Since ∣ h j ∣ ≤ 1 |h_j| \le1 ∣ h j ∣ ≤ 1 and ∣ ϕ i ∣ = 1 |\phi_i| = 1 ∣ ϕ i ∣ = 1 ,
∣ J N ( y ) ∣ ≤ q M N ∣ J q M ( y ) ∣ + N − q M N ≤ ∣ J q M ( y ) ∣ + q N |J_N(y)| \le\frac{qM}{N}|J_{qM}(y)| + \frac{N-qM}{N} \le|J_{qM}(y)| + \frac{q}{N} ∣ J N ( y ) ∣ ≤ N q M ∣ J q M ( y ) ∣ + N N − q M ≤ ∣ J q M ( y ) ∣ + N q
for almost every $y. Thus J N ( y ) → 0 J_N(y) \to0 J N ( y ) → 0 almost everywhere. Finally, ∣ J N ∣ ≤ 1 |J_N| \le1 ∣ J N ∣ ≤ 1 , so dominated convergence gives
lim N → ∞ ∥ J N ∥ 2 2 = lim N → ∞ ∫ X ∣ J N ∣ 2 d μ = 0. \lim_{N\to\infty}\|J_N\|_2^2 = \lim_{N\to\infty}\int_X |J_N|^2\,d\mu= 0. N → ∞ lim ∥ J N ∥ 2 2 = N → ∞ lim ∫ X ∣ J N ∣ 2 d μ = 0.
This completes the proof. □
Proof of the sharp-order theorem
Now we shall prove the integer upper bounds in Theorem B . The prime assertions are proved in § 6 .
Integer upper bounds in Theorem B and Corollary 1.2. Put ( d 1 , d 2 , d 3 ) = ( 4 , 2 , 2 ) (d_1,d_2,d_3)=(4,2,2) ( d 1 , d 2 , d 3 ) = ( 4 , 2 , 2 ) and ( p 1 ( t ) , p 2 ( t ) , p 3 ( t ) ) = ( t , t 2 , t 2 + t ) (p_1(t),p_2(t),p_3(t))=(t,t^2,t^2+t) ( p 1 ( t ) , p 2 ( t ) , p 3 ( t )) = ( t , t 2 , t 2 + t ) . Fix j ∈ { 1 , 2 , 3 } j\in\{1,2,3\} j ∈ { 1 , 2 , 3 } and suppose that E ( f j ∣ W d j ( T j ) ) = 0 \mathbb{E}(f_j\mid W_{d_j}(T_j))=0 E ( f j ∣ W d j ( T j )) = 0 . By multilinearity, we may assume ∣ f i ∣ ≤ 1 |f_i|\le1 ∣ f i ∣ ≤ 1 for i = 1 , 2 , 3 i=1,2,3 i = 1 , 2 , 3 . The polynomials t , t 2 , t 2 + t t,t^2,t^2+t t , t 2 , t 2 + t are pairwise linearly independent. Apply Corollary 1.1 with ℓ = 3 \ell=3 ℓ = 3 and D = 2 D=2 D = 2 , increase s s s to at least four if necessary, and put F i = E ( f i ∣ W s ( T i ) ) F_i=\mathbb{E}(f_i\mid W_s(T_i)) F i = E ( f i ∣ W s ( T i )) . By Corollary 1.1 and the tower property of conditional expectation, we have E ( F j ∣ W d j ( T j ) ) = E ( f j ∣ W d j ( T j ) ) = 0 \mathbb{E}(F_j\mid W_{d_j}(T_j))=\mathbb{E}(f_j\mid W_{d_j}(T_j))=0 E ( F j ∣ W d j ( T j )) = E ( f j ∣ W d j ( T j )) = 0 and
∥ A N ( f 1 , f 2 , f 3 ) − A N ( F 1 , F 2 , F 3 ) ∥ 2 ⟶ 0. \|A_N(f_1,f_2,f_3)-A_N(F_1,F_2,F_3)\|_2\longrightarrow0. ∥ A N ( f 1 , f 2 , f 3 ) − A N ( F 1 , F 2 , F 3 ) ∥ 2 ⟶ 0.
Next we will prove that ∥ A N ( F 1 , F 2 , F 3 ) ∥ 2 → 0 \|A_N(F_1,F_2,F_3)\|_2\to0 ∥ A N ( F 1 , F 2 , F 3 ) ∥ 2 → 0 .
Recall that P d ( T ) : = { ϕ ∈ L ∞ ( μ ) : ∣ ϕ ∣ = 1 , Δ T q d + 1 ϕ = 1 for some q ≥ 1 } \mathcal{P}_d(T):=\{\phi\in L^\infty(\mu):|\phi|=1,\ \Delta_{T^q}^{d+1}\phi=1\text{ for some }q\ge1\} P d ( T ) := { ϕ ∈ L ∞ ( μ ) : ∣ ϕ ∣ = 1 , Δ T q d + 1 ϕ = 1 for some q ≥ 1 } . Fix arbitrary U i ∈ A s ( T i ) : = span C P s ( T i ) U_i\in\mathcal{A}_s(T_i):=\operatorname{span}_{\mathbb{C}}\mathcal{P}_s(T_i) U i ∈ A s ( T i ) := span C P s ( T i ) and define V j = U j − E ( U j ∣ W d j ( T j ) ) V_j=U_j-\mathbb{E}(U_j\mid W_{d_j}(T_j)) V j = U j − E ( U j ∣ W d j ( T j )) , V i = U i V_i=U_i V i = U i for i ≠ j i\ne j i = j . Write each U i U_i U i as a finite sum U i = ∑ r = 1 m i c i , r ϕ i , r U_i=\sum_{r=1}^{m_i}c_{i,r}\phi_{i,r} U i = ∑ r = 1 m i c i , r ϕ i , r , where c i , r ∈ C c_{i,r}\in\mathbb{C} c i , r ∈ C and ϕ i , r ∈ P s ( T i ) \phi_{i,r}\in\mathcal{P}_s(T_i) ϕ i , r ∈ P s ( T i ) . For each i = 1 , 2 , 3 i=1,2,3 i = 1 , 2 , 3 , choose q i ≥ 1 q_i\ge1 q i ≥ 1 such that Δ T i q i s + 1 ϕ i , r = 1 \Delta_{T_i^{q_i}}^{s+1}\phi_{i,r}=1 Δ T i q i s + 1 ϕ i , r = 1 for every 1 ≤ r ≤ m i 1\le r\le m_i 1 ≤ r ≤ m i . Let q q q be the least common multiple of q 1 , q 2 , q 3 q_1,q_2,q_3 q 1 , q 2 , q 3 . Since q i ∣ q q_i\mid q q i ∣ q , Lemma 2.4 gives Δ T i s + 1 φ i , r = 1 \Delta_{T_i}^{s+1}\varphi_{i,r}=1 Δ T i s + 1 φ i , r = 1 for all i , r i,r i , r . Proposition 4.4 , applied with L = s + 1 L=s+1 L = s + 1 to each triple of phases, and multilinearity therefore give
∥ A N ( V 1 , V 2 , V 3 ) ∥ 2 ⟶ 0 \lVert A_N(V_1,V_2,V_3)\rVert_2 \longrightarrow0 ∥ A N ( V 1 , V 2 , V 3 ) ∥ 2 ⟶ 0
for every U i ∈ A s ( T i ) U_i\in\mathcal{A}_s(T_i) U i ∈ A s ( T i ) .
Since E ( F j ∣ W d j ( T j ) ) = 0 \mathbb{E}(F_j\mid\mathcal{W}_{d_j}(T_j))=0 E ( F j ∣ W d j ( T j )) = 0 , by the L 2 L^2 L 2 contraction property of conditional expectation, we have
∥ F j − V j ∥ 2 = ∥ ( F j − U j ) − E ( F j − U j ∣ W d j ( T j ) ) ∥ 2 ≤ ∥ F j − U j ∥ 2 . \lVert F_j-V_j\rVert_2=\lVert(F_j-U_j)-\mathbb{E}(F_j-U_j\mid\mathcal{W}_{d_j}(T_j))\rVert_2\leq\lVert F_j-U_j\rVert_2. ∥ F j − V j ∥ 2 = ∥( F j − U j ) − E ( F j − U j ∣ W d j ( T j )) ∥ 2 ≤ ∥ F j − U j ∥ 2 .
For i ≠ j i\neq j i = j , we have F i − V i = F i − U i F_i-V_i=F_i-U_i F i − V i = F i − U i . Hence, for i = 1 , 2 , 3 i=1,2,3 i = 1 , 2 , 3 , ∥ F i − V i ∥ 2 ≤ ∥ F i − U i ∥ 2 \lVert F_i-V_i\rVert_2\leq\lVert F_i-U_i\rVert_2 ∥ F i − V i ∥ 2 ≤ ∥ F i − U i ∥ 2 .
Now fix ε > 0 \varepsilon>0 ε > 0 . By Lemma 2.4 , first choose U 3 U_3 U 3 and define V 3 V_3 V 3 , then choose U 2 U_2 U 2 and define V 2 V_2 V 2 , and finally choose U 1 U_1 U 1 , so that
∥ F 3 − U 3 ∥ 2 < ε 3 , \lVert F_3-U_3\rVert_2<\frac{\varepsilon}{3}, ∥ F 3 − U 3 ∥ 2 < 3 ε ,
∥ F 2 − U 2 ∥ 2 < ε 3 ( 1 + ∥ V 3 ∥ ∞ ) , \lVert F_2-U_2\rVert_2<\frac{\varepsilon}{3(1+\lVert V_3\rVert_\infty)}, ∥ F 2 − U 2 ∥ 2 < 3 ( 1 + ∥ V 3 ∥ ∞ ) ε ,
∥ F 1 − U 1 ∥ 2 < ε 3 ( 1 + ∥ V 2 ∥ ∞ ) ( 1 + ∥ V 3 ∥ ∞ ) . \lVert F_1-U_1\rVert_2<\frac{\varepsilon}{3(1+\lVert V_2\rVert_\infty)(1+\lVert V_3\rVert_\infty)}. ∥ F 1 − U 1 ∥ 2 < 3 ( 1 + ∥ V 2 ∥ ∞ ) ( 1 + ∥ V 3 ∥ ∞ ) ε .
For any a 1 , a 2 , a 3 ∈ L ∞ ( μ ) a_1,a_2,a_3\in L^\infty(\mu) a 1 , a 2 , a 3 ∈ L ∞ ( μ ) , by measure preservation, we have
∥ A N ( a 1 , a 2 , a 3 ) ∥ 2 ≤ 1 N ∑ n = 1 N ∥ T i p i ( n ) a i ∥ 2 ∏ k ≠ i ∥ a k ∥ ∞ = ∥ a i ∥ 2 ∏ k ≠ i ∥ a k ∥ ∞ , for i = 1 , 2 , 3. \begin{aligned}
\lVert A_N(a_1,a_2,a_3)\rVert_2\leq\frac{1}{N}\sum_{n=1}^{N}\lVert T_i^{p_i(n)}a_i\rVert_2\prod_{k\neq i}\lVert a_k\rVert_\infty\\
&=\lVert a_i\rVert_2\prod_{k\neq i}\lVert a_k\rVert_\infty,\qquad\text{for }i=1,2,3.
\end{aligned} ∥ A N ( a 1 , a 2 , a 3 ) ∥ 2 ≤ N 1 n = 1 ∑ N ∥ T i p i ( n ) a i ∥ 2 k = i ∏ ∥ a k ∥ ∞ = ∥ a i ∥ 2 k = i ∏ ∥ a k ∥ ∞ , for i = 1 , 2 , 3.
Since ∣ F i ∣ ≤ 1 \lvert F_i\rvert\leq1 ∣ F i ∣ ≤ 1 , by multilinearity and the preceding estimate
∥ A N ( F 1 , F 2 , F 3 ) − A N ( V 1 , V 2 , V 3 ) ∥ 2 ≤ ∥ F 3 − V 3 ∥ 2 + ∥ F 2 − V 2 ∥ 2 ∥ V 3 ∥ ∞ + ∥ F 1 − V 1 ∥ 2 ∥ V 2 ∥ ∞ ∥ V 3 ∥ ∞ < ε . \begin{aligned}
\lVert A_N(F_1,F_2,F_3)-A_N(V_1,V_2,V_3)\rVert_2\leq\lVert F_3-V_3\rVert_2+\lVert F_2-V_2\rVert_2\lVert V_3\rVert_\infty\\
&+\lVert F_1-V_1\rVert_2\lVert V_2\rVert_\infty\lVert V_3\rVert_\infty\\
&<\varepsilon.
\end{aligned} ∥ A N ( F 1 , F 2 , F 3 ) − A N ( V 1 , V 2 , V 3 ) ∥ 2 ≤ ∥ F 3 − V 3 ∥ 2 + ∥ F 2 − V 2 ∥ 2 ∥ V 3 ∥ ∞ + ∥ F 1 − V 1 ∥ 2 ∥ V 2 ∥ ∞ ∥ V 3 ∥ ∞ < ε .
Since ε > 0 \varepsilon>0 ε > 0 is arbitrary and ∥ A N ( V 1 , V 2 , V 3 ) ∥ 2 ⟶ 0 \lVert A_N(V_1,V_2,V_3)\rVert_2\longrightarrow0 ∥ A N ( V 1 , V 2 , V 3 ) ∥ 2 ⟶ 0 , then ∥ A N ( F 1 , F 2 , F 3 ) ∥ 2 ⟶ 0 \lVert A_N(F_1,F_2,F_3)\rVert_2\longrightarrow0 ∥ A N ( F 1 , F 2 , F 3 ) ∥ 2 ⟶ 0 . Hence
∥ A N ( f 1 , f 2 , f 3 ) ∥ 2 ⟶ 0. \lVert A_N(f_1,f_2,f_3)\rVert_2\longrightarrow0. ∥ A N ( f 1 , f 2 , f 3 ) ∥ 2 ⟶ 0.
By Lemma 2.4 , we have W d j ( T j ) ⊆ Z d j ( T j ) \mathcal{W}_{d_j}(T_j)\subseteq\mathcal{Z}_{d_j}(T_j) W d j ( T j ) ⊆ Z d j ( T j ) . Thus E ( f j ∣ Z d j ( T j ) ) = 0 \mathbb{E}(f_j\mid\mathcal{Z}_{d_j}(T_j))=0 E ( f j ∣ Z d j ( T j )) = 0 implies E ( f j ∣ W d j ( T j ) ) = 0 \mathbb{E}(f_j\mid\mathcal{W}_{d_j}(T_j))=0 E ( f j ∣ W d j ( T j )) = 0 . Together with ∥ f j ∥ U d j + 1 ( T j ) = 0 \lVert f_j\rVert_{U^{d_j+1}(T_j)}=0 ∥ f j ∥ U d j + 1 ( T j ) = 0 if and only if E ( f j ∣ Z d j ( T j ) ) = 0 \mathbb{E}(f_j\mid\mathcal{Z}_{d_j}(T_j))=0 E ( f j ∣ Z d j ( T j )) = 0 , this proves the seminorm assertions.
Finally, for every i = 1 , 2 , 3 i=1,2,3 i = 1 , 2 , 3 , let g i = E ( f i ∣ W d i ( T i ) ) g_i=\mathbb{E}(f_i\mid\mathcal{W}_{d_i}(T_i)) g i = E ( f i ∣ W d i ( T i )) or g i = E ( f i ∣ Z d i ( T i ) ) g_i=\mathbb{E}(f_i\mid\mathcal{Z}_{d_i}(T_i)) g i = E ( f i ∣ Z d i ( T i )) . In either case, by the property of conditional expectation, E ( f i − g i ∣ W d i ( T i ) ) = 0 \mathbb{E}(f_i-g_i\mid\mathcal{W}_{d_i}(T_i))=0 E ( f i − g i ∣ W d i ( T i )) = 0 . By multilinearity,
A N ( f 1 , f 2 , f 3 ) − A N ( g 1 , g 2 , g 3 ) = A N ( f 1 − g 1 , f 2 , f 3 ) + A N ( g 1 , f 2 − g 2 , f 3 ) + A N ( g 1 , g 2 , f 3 − g 3 ) . \begin{aligned}
A_N(f_1,f_2,f_3)-A_N(g_1,g_2,g_3)=A_N(f_1-g_1,f_2,f_3)+A_N(g_1,f_2-g_2,f_3) \\
&+A_N(g_1,g_2,f_3-g_3).
\end{aligned} A N ( f 1 , f 2 , f 3 ) − A N ( g 1 , g 2 , g 3 ) = A N ( f 1 − g 1 , f 2 , f 3 ) + A N ( g 1 , f 2 − g 2 , f 3 ) + A N ( g 1 , g 2 , f 3 − g 3 ) .
By the previously proved result that the limit is zero, each term converges to zero in L 2 ( μ ) L^2(\mu) L 2 ( μ ) . Hence
∥ A N ( f 1 , f 2 , f 3 ) − A N ( g 1 , g 2 , g 3 ) ∥ 2 ⟶ 0. \lVert A_N(f_1,f_2,f_3)-A_N(g_1,g_2,g_3)\rVert_2\longrightarrow0. ∥ A N ( f 1 , f 2 , f 3 ) − A N ( g 1 , g 2 , g 3 ) ∥ 2 ⟶ 0.
A five-dimensional sharpness example
Proposition 5.1. There exist commuting Haar measure-preserving affine automorphisms T , S T,S T , S of T 5 \mathbb{T}^{5} T 5 , equipped with Haar probability m T 5 m_{\mathbb{T}^{5}} m T 5 , and f 1 , f 2 , f 3 ∈ L ∞ ( m T 5 ) f_{1}, f_{2}, f_{3} \in L^{\infty}(m_{\mathbb{T}^{5}}) f 1 , f 2 , f 3 ∈ L ∞ ( m T 5 ) such that
∥ f 1 ∥ U 4 ( T ) = 0 , ∥ f 2 ∥ U 2 ( S ) = ∥ f 3 ∥ U 2 ( S ) = 0 , \lVert f_{1}\rVert_{U^{4}(T)}=0,\qquad\lVert f_{2}\rVert_{U^{2}(S)}=\lVert f_{3}\rVert_{U^{2}(S)}=0, ∥ f 1 ∥ U 4 ( T ) = 0 , ∥ f 2 ∥ U 2 ( S ) = ∥ f 3 ∥ U 2 ( S ) = 0 , whereas for all N ≥ 1 N \ge1 N ≥ 1
∥ 1 N ∑ n = 1 N T n f 1 S n 2 f 2 S n 2 + n f 3 ∥ 2 = 1. \left\lVert\frac{1}{N}\sum_{n=1}^{N}T^{n}f_{1}\,S^{n^{2}}f_{2}\,S^{n^{2}+n}f_{3}\right\rVert_{2}=1. N 1 n = 1 ∑ N T n f 1 S n 2 f 2 S n 2 + n f 3 2 = 1. Furthermore, f 1 ∈ L ∞ ( W 4 ( T ) ) f_{1}\in L^{\infty}(W_{4}(T)) f 1 ∈ L ∞ ( W 4 ( T )) and f 2 , f 3 ∈ L ∞ ( W 2 ( S ) ) f_{2},f_{3}\in L^{\infty}(W_{2}(S)) f 2 , f 3 ∈ L ∞ ( W 2 ( S )) .
Proof. Fix α ∈ R ∖ Q \alpha\in\mathbb{R}\setminus\mathbb{Q} α ∈ R ∖ Q , and equip T 5 \mathbb{T}^{5} T 5 with Haar probability m T 5 m_{\mathbb{T}^{5}} m T 5 . With all coordinates taken modulo one, define
T x = ( x 1 + 6 α x 2 + 2 x 1 + 6 α x 3 + 3 x 2 + 3 x 1 + 6 α x 4 + 2 x 3 + 3 x 2 + 2 x 1 + 3 α x 5 + 6 x 4 ) , S x = ( x 1 x 2 + 2 α x 3 + x 1 x 4 + x 2 + α x 5 + 2 x 3 − 3 x 2 + 2 x 1 − 3 α ) . Tx=\begin{pmatrix}
x_{1}+6\alpha\\
x_{2}+2x_{1}+6\alpha\\
x_{3}+3x_{2}+3x_{1}+6\alpha\\
x_{4}+2x_{3}+3x_{2}+2x_{1}+3\alpha\\
x_{5}+6x_{4}
\end{pmatrix},
\qquad
Sx=\begin{pmatrix}
x_{1}\\
x_{2}+2\alpha\\
x_{3}+x_{1}\\
x_{4}+x_{2}+\alpha\\
x_{5}+2x_{3}-3x_{2}+2x_{1}-3\alpha
\end{pmatrix}. T x = x 1 + 6 α x 2 + 2 x 1 + 6 α x 3 + 3 x 2 + 3 x 1 + 6 α x 4 + 2 x 3 + 3 x 2 + 2 x 1 + 3 α x 5 + 6 x 4 , S x = x 1 x 2 + 2 α x 3 + x 1 x 4 + x 2 + α x 5 + 2 x 3 − 3 x 2 + 2 x 1 − 3 α .
Their linear parts are triangular integer matrices with diagonal entries one, so both maps are invertible and preserve m T 5 m_{\mathbb{T}^{5}} m T 5 . They commute, since substitution gives
T S x = S T x = ( x 1 + 6 α x 2 + 2 x 1 + 8 α x 3 + 3 x 2 + 4 x 1 + 12 α x 4 + 2 x 3 + 4 x 2 + 4 x 1 + 10 α x 5 + 6 x 4 + 2 x 3 + 3 x 2 + 2 x 1 + 3 α ) . TSx=STx=\begin{pmatrix}
x_{1}+6\alpha\\
x_{2}+2x_{1}+8\alpha\\
x_{3}+3x_{2}+4x_{1}+12\alpha\\
x_{4}+2x_{3}+4x_{2}+4x_{1}+10\alpha\\
x_{5}+6x_{4}+2x_{3}+3x_{2}+2x_{1}+3\alpha
\end{pmatrix}. T S x = S T x = x 1 + 6 α x 2 + 2 x 1 + 8 α x 3 + 3 x 2 + 4 x 1 + 12 α x 4 + 2 x 3 + 4 x 2 + 4 x 1 + 10 α x 5 + 6 x 4 + 2 x 3 + 3 x 2 + 2 x 1 + 3 α .
Define the circle-valued functions ℓ i : T 5 → T \ell_i:\mathbb{T}^{5}\to\mathbb{T} ℓ i : T 5 → T by
ℓ 1 = − x 4 , ℓ 2 = − x 5 , ℓ 3 = x 5 + 3 x 4 − x 3 , f i = e ( ℓ i ) . \ell_{1}=-x_{4},\qquad\ell_{2}=-x_{5},\qquad\ell_{3}=x_{5}+3x_{4}-x_{3},\qquad f_{i}=e(\ell_{i}). ℓ 1 = − x 4 , ℓ 2 = − x 5 , ℓ 3 = x 5 + 3 x 4 − x 3 , f i = e ( ℓ i ) .
A computation of their orbit increments gives
T n ℓ 1 − ℓ 1 = − 2 n x 3 − 3 n 2 x 2 − 2 n 3 x 1 − 3 α n 4 , T^{n}\ell_{1}-\ell_{1}=-2nx_{3}-3n^{2}x_{2}-2n^{3}x_{1}-3\alpha n^{4}, T n ℓ 1 − ℓ 1 = − 2 n x 3 − 3 n 2 x 2 − 2 n 3 x 1 − 3 α n 4 ,
S r ℓ 2 − ℓ 2 = − 2 r x 3 + 3 r x 2 − r ( r + 1 ) x 1 + 3 α r 2 , S^{r}\ell_{2}-\ell_{2}=-2rx_{3}+3rx_{2}-r(r+1)x_{1}+3\alpha r^{2}, S r ℓ 2 − ℓ 2 = − 2 r x 3 + 3 r x 2 − r ( r + 1 ) x 1 + 3 α r 2 ,
S r ℓ 3 − ℓ 3 = 2 r x 3 + r 2 x 1 . S^{r}\ell_{3}-\ell_{3}=2rx_{3}+r^{2}x_{1}. S r ℓ 3 − ℓ 3 = 2 r x 3 + r 2 x 1 .
These formulas follow by induction from the definitions of T T T and S S S . Substituting r = n 2 r=n^{2} r = n 2 and r = n 2 + n r=n^{2}+n r = n 2 + n into the second and third formulas, respectively, gives
( T n ℓ 1 − ℓ 1 ) + ( S n 2 ℓ 2 − ℓ 2 ) + ( S n 2 + n ℓ 3 − ℓ 3 ) = ( − 2 n − 2 n 2 + 2 ( n 2 + n ) ) x 3 + ( − 3 n 2 + 3 n 2 ) x 2 + ( − 2 n 3 − n 2 ( n 2 + 1 ) + ( n 2 + n ) 2 ) x 1 + ( − 3 n 4 + 3 n 4 ) α = 0. \begin{aligned}
(T^{n}\ell_{1}-\ell_{1})+(S^{n^{2}}\ell_{2}-\ell_{2})+(S^{n^{2}+n}\ell_{3}-\ell_{3})
={}&(-2n-2n^{2}+2(n^{2}+n))x_{3}\\
&+(-3n^{2}+3n^{2})x_{2}\\
&+(-2n^{3}-n^{2}(n^{2}+1)+(n^{2}+n)^{2})x_{1}\\
&+(-3n^{4}+3n^{4})\alpha=0.
\end{aligned} ( T n ℓ 1 − ℓ 1 ) + ( S n 2 ℓ 2 − ℓ 2 ) + ( S n 2 + n ℓ 3 − ℓ 3 ) = ( − 2 n − 2 n 2 + 2 ( n 2 + n )) x 3 + ( − 3 n 2 + 3 n 2 ) x 2 + ( − 2 n 3 − n 2 ( n 2 + 1 ) + ( n 2 + n ) 2 ) x 1 + ( − 3 n 4 + 3 n 4 ) α = 0.
Here the coefficient of x 1 x_{1} x 1 vanishes because
( n 2 + n ) 2 − n 2 ( n 2 + 1 ) = 2 n 3 . (n^{2}+n)^{2}-n^{2}(n^{2}+1)=2n^{3}. ( n 2 + n ) 2 − n 2 ( n 2 + 1 ) = 2 n 3 .
Thus
T n f 1 S n 2 f 2 S n 2 + n f 3 = e ( ℓ 1 + ℓ 2 + ℓ 3 ) = e ( 2 x 4 − x 3 ) T^{n}f_{1}\,S^{n^{2}}f_{2}\,S^{n^{2}+n}f_{3}=e(\ell_{1}+\ell_{2}+\ell_{3})=e(2x_{4}-x_{3}) T n f 1 S n 2 f 2 S n 2 + n f 3 = e ( ℓ 1 + ℓ 2 + ℓ 3 ) = e ( 2 x 4 − x 3 )
for every n ≥ 1 n \ge1 n ≥ 1 . Since the right-hand side has modulus one, it follows that
∥ 1 N ∑ n = 1 N T n f 1 S n 2 f 2 S n 2 + n f 3 ∥ L 2 ( m T 5 ) = 1 \left\| \frac{1}{N}\sum_{n=1}^{N}T^n f_1 S^{n^2}f_2 S^{n^2+n}f_3\right\|_{L^2(m_{\mathbb{T}^5})}=1 N 1 n = 1 ∑ N T n f 1 S n 2 f 2 S n 2 + n f 3 L 2 ( m T 5 ) = 1
for every N ≥ 1 N \ge1 N ≥ 1 .
For positive integers h 1 , h 2 , h 3 , h 4 , h , k h_1,h_2,h_3,h_4,h,k h 1 , h 2 , h 3 , h 4 , h , k , finite differences give
Δ T h 4 ⋯ Δ T h 1 f 1 = e ( − 72 α h 1 h 2 h 3 h 4 ) , \Delta_{T}^{h_4}\cdots\Delta_{T}^{h_1}f_1=e(-72\alpha h_1h_2h_3h_4), Δ T h 4 ⋯ Δ T h 1 f 1 = e ( − 72 α h 1 h 2 h 3 h 4 ) ,
Δ S k Δ S h f 2 = e ( − 2 h k ( x 1 − 3 α ) ) , \Delta_{S^k}\Delta_{S^h}f_2=e(-2hk(x_1-3\alpha)), Δ S k Δ S h f 2 = e ( − 2 hk ( x 1 − 3 α )) ,
Δ S k Δ S h f 3 = e ( 2 h k x 1 ) . \Delta_{S^k}\Delta_{S^h}f_3=e(2hkx_1). Δ S k Δ S h f 3 = e ( 2 hk x 1 ) .
The last two functions have integral zero, while irrationality of α \alpha α gives
lim H → ∞ 1 H ∑ h 4 = 1 H e ( − 72 α h 1 h 2 h 3 h 4 ) = 0 \lim_{H\to\infty}\frac{1}{H}\sum_{h_4=1}^{H}e(-72\alpha h_1h_2h_3h_4)=0 H → ∞ lim H 1 h 4 = 1 ∑ H e ( − 72 α h 1 h 2 h 3 h 4 ) = 0
for fixed h 1 , h 2 , h 3 h_1,h_2,h_3 h 1 , h 2 , h 3 . By (5) ,
∥ f 1 ∥ U 4 ( T ) = ∥ f 2 ∥ U 2 ( S ) = ∥ f 3 ∥ U 2 ( S ) = 0. \|f_1\|_{U^4(T)}=\|f_2\|_{U^2(S)}=\|f_3\|_{U^2(S)}=0. ∥ f 1 ∥ U 4 ( T ) = ∥ f 2 ∥ U 2 ( S ) = ∥ f 3 ∥ U 2 ( S ) = 0.
Also,
Δ T 5 f 1 = Δ S 3 f 2 = Δ S 3 f 3 = 1. \Delta_T^5f_1=\Delta_S^3f_2=\Delta_S^3f_3=1. Δ T 5 f 1 = Δ S 3 f 2 = Δ S 3 f 3 = 1.
Hence, by the definition of phase factor,
f 1 ∈ L ∞ ( W 4 ( T ) ) , f 2 , f 3 ∈ L ∞ ( W 2 ( S ) ) . f_1\in L^\infty(W_4(T)),\qquad f_2,f_3\in L^\infty(W_2(S)). f 1 ∈ L ∞ ( W 4 ( T )) , f 2 , f 3 ∈ L ∞ ( W 2 ( S )) .
Take T 1 = T T_1=T T 1 = T and T 2 = T 3 = S T_2=T_3=S T 2 = T 3 = S . This proposition shows that E ( f 1 ∣ Z 3 ( T 1 ) ) = 0 \mathbb{E}(f_1\mid Z_3(T_1))=0 E ( f 1 ∣ Z 3 ( T 1 )) = 0 and E ( f i ∣ Z 1 ( T i ) ) = 0 \mathbb{E}(f_i\mid Z_1(T_i))=0 E ( f i ∣ Z 1 ( T i )) = 0 for i = 2 , 3 i=2,3 i = 2 , 3 , but the average has L 2 L^2 L 2 -norm one. Hence none of the three Host–Kra orders can be lowered, even if the other two factors are unrestricted. Since W d ( T i ) ⊆ Z d ( T i ) W_{d}(T_i)\subseteq Z_{d}(T_i) W d ( T i ) ⊆ Z d ( T i ) , the phase-factor orders in Theorem B are also optimal. The summand does not depend on n n n , so the same example gives optimality over the primes; see § 6 .
Averages over the primes
In this section, we prove the prime assertions of Theorems A and B. We need the following the comparison theorem of Frantzikinakis, Host, and Kra [11 Proposition 3.6] .
Theorem 6.1. Let T 1 , … , T ℓ T_1,\ldots,T_\ell T 1 , … , T ℓ be commuting invertible measure-preserving transformations, let g 1 , … , g m ∈ L ∞ ( μ ) g_1,\ldots,g_m\in L^\infty(\mu) g 1 , … , g m ∈ L ∞ ( μ ) , and let q i , j ∈ Z [ t ] q_{i,j}\in\mathbb{Z}[t] q i , j ∈ Z [ t ] for 1 ≤ i ≤ ℓ 1\le i\le\ell 1 ≤ i ≤ ℓ and 1 ≤ j ≤ m 1\le j\le m 1 ≤ j ≤ m . For w > 2 w>2 w > 2 , put
W = ∏ p < w p , Λ w , b ′ ( n ) = φ ( W ) W Λ ′ ( W n + b ) , W=\prod_{p<w}p,\qquad\Lambda'_{w,b}(n)=\frac{\varphi(W)}{W}\Lambda'(Wn+b), W = p < w ∏ p , Λ w , b ′ ( n ) = W φ ( W ) Λ ′ ( W n + b ) , where Λ ′ ( n ) = 1 P ( n ) log n \Lambda'(n)=1_{\mathbb{P}}(n)\log n Λ ′ ( n ) = 1 P ( n ) log n . Then
lim w → ∞ lim sup N → ∞ max 1 ≤ b ≤ W ( b , W ) = 1 ∥ 1 N ∑ n = 1 N ( Λ w , b ′ ( n ) − 1 ) ⋅ ∏ j = 1 m ( ∏ i = 1 ℓ T i q i , j ( W n + b ) ) g j ∥ 2 = 0. \lim_{w\to\infty}\limsup_{N\to\infty}\max_{\substack{1\le b\le W\\(b,W)=1}}\left\|\frac{1}{N}\sum_{n=1}^{N}(\Lambda'_{w,b}(n)-1)\cdot\prod_{j=1}^{m}\left(\prod_{i=1}^{\ell}T_i^{q_{i,j}(Wn+b)}\right)g_j\right\|_2=0. w → ∞ lim N → ∞ lim sup 1 ≤ b ≤ W ( b , W ) = 1 max N 1 n = 1 ∑ N ( Λ w , b ′ ( n ) − 1 ) ⋅ j = 1 ∏ m ( i = 1 ∏ ℓ T i q i , j ( W n + b ) ) g j 2 = 0.
Lemma 6.2. Let T 1 , … , T ℓ T_1,\ldots,T_\ell T 1 , … , T ℓ be invertible commuting measure-preserving transformations, let p i ∈ Z [ t ] p_i\in\mathbb{Z}[t] p i ∈ Z [ t ] , and let f i ∈ L ∞ ( μ ) f_i\in L^\infty(\mu) f i ∈ L ∞ ( μ ) . Put a ( n ) = ∏ i = 1 ℓ T i p i ( n ) f i a(n)=\prod_{i=1}^{\ell}T_i^{p_i(n)}f_i a ( n ) = ∏ i = 1 ℓ T i p i ( n ) f i . If, for every m ≥ 1 m\ge1 m ≥ 1 and 1 ≤ b ≤ m 1\le b\le m 1 ≤ b ≤ m ,
∥ 1 N ∑ n = 1 N a ( m n + b ) ∥ 2 ⟶ 0 , \left\|\frac{1}{N}\sum_{n=1}^{N}a(mn+b)\right\|_2\longrightarrow0, N 1 n = 1 ∑ N a ( mn + b ) 2 ⟶ 0 , then ∥ A N , P ( f 1 , … , f ℓ ) ∥ 2 ⟶ 0 \|A_{N,\mathbb{P}}(f_1,\ldots,f_\ell)\|_2\longrightarrow0 ∥ A N , P ( f 1 , … , f ℓ ) ∥ 2 ⟶ 0 .
Proof. By scaling, assume ∣ f j ∣ ≤ 1 |f_j| \le1 ∣ f j ∣ ≤ 1 . For w > 2 w>2 w > 2 , put W = ∏ p < w , p ∈ P p W=\prod_{p<w,\ p\in\mathbb{P}}p W = ∏ p < w , p ∈ P p and Λ ′ ( n ) = 1 P ( n ) log n \Lambda'(n)=1_{\mathbb{P}}(n)\log n Λ ′ ( n ) = 1 P ( n ) log n . For 1 ≤ b ≤ W 1\le b\le W 1 ≤ b ≤ W with ( b , W ) = 1 (b,W)=1 ( b , W ) = 1 , define Λ w , b ′ ( n ) = φ ( W ) Λ ′ ( ( W n + b ) / W ) \Lambda'_{w,b}(n)=\varphi(W)\Lambda'((Wn+b)/W) Λ w , b ′ ( n ) = φ ( W ) Λ ′ (( W n + b ) / W ) , where φ ( W ) = # { 1 ≤ b ≤ W : ( b , W ) = 1 } \varphi(W)=\#\{1\le b\le W:(b,W)=1\} φ ( W ) = # { 1 ≤ b ≤ W : ( b , W ) = 1 } is Euler’s totient function. Theorem 6.1 , applied with polynomial p j p_j p j in direction T j T_j T j and zero polynomials in the other directions, gives
lim w → ∞ lim sup N → ∞ max 1 ≤ b ≤ W ( b , W ) = 1 ∥ 1 N ∑ n = 1 N ( Λ w , b ′ ( n ) − 1 ) a ( W n + b ) ∥ 2 = 0. \lim_{w\to\infty}\limsup_{N\to\infty}\max_{\substack{1\le b\le W\\(b,W)=1}}\left\|\frac{1}{N}\sum_{n=1}^{N}(\Lambda'_{w,b}(n)-1)a(Wn+b)\right\|_2=0. w → ∞ lim N → ∞ lim sup 1 ≤ b ≤ W ( b , W ) = 1 max N 1 n = 1 ∑ N ( Λ w , b ′ ( n ) − 1 ) a ( W n + b ) 2 = 0.
For every fixed w w w and every b b b with ( b , W ) = 1 (b,W)=1 ( b , W ) = 1 , the hypothesis gives
1 N ∑ n = 1 N a ( W n + b ) ⟶ 0 in L 2 ( μ ) . \frac{1}{N}\sum_{n=1}^{N}a(Wn+b)\longrightarrow0\quad\text{in }L^2(\mu). N 1 n = 1 ∑ N a ( W n + b ) ⟶ 0 in L 2 ( μ ) .
Since there are only finitely many such b b b , the convergence is uniform in b b b . Consequently,
lim w → ∞ lim sup N → ∞ max 1 ≤ b ≤ W ( b , W ) = 1 ∥ 1 N ∑ n = 1 N Λ w , b ′ ( n ) a ( W n + b ) ∥ 2 = 0. \lim_{w\to\infty}\limsup_{N\to\infty}\max_{\substack{1\le b\le W\\(b,W)=1}}\left\|\frac{1}{N}\sum_{n=1}^{N}\Lambda'_{w,b}(n)a(Wn+b)\right\|_2=0. w → ∞ lim N → ∞ lim sup 1 ≤ b ≤ W ( b , W ) = 1 max N 1 n = 1 ∑ N Λ w , b ′ ( n ) a ( W n + b ) 2 = 0.
Write L M = M − 1 ∑ n = 1 M Λ ′ ( n ) a ( n ) L_M=M^{-1}\sum_{n=1}^{M}\Lambda'(n)a(n) L M = M − 1 ∑ n = 1 M Λ ′ ( n ) a ( n ) . Fix w w w and, for M ≥ W M\ge W M ≥ W , put N = ⌊ M / W ⌋ N=\lfloor M/W\rfloor N = ⌊ M / W ⌋ . Then
∑ b = 1 W ∑ n = 1 N Λ ′ ( W n + b ) a ( W n + b ) = ∑ k = W + 1 W ( N + 1 ) Λ ′ ( k ) a ( k ) . \sum_{b=1}^{W}\sum_{n=1}^{N}\Lambda'(Wn+b)a(Wn+b)=\sum_{k=W+1}^{W(N+1)}\Lambda'(k)a(k). b = 1 ∑ W n = 1 ∑ N Λ ′ ( W n + b ) a ( W n + b ) = k = W + 1 ∑ W ( N + 1 ) Λ ′ ( k ) a ( k ) .
Since every prime greater than W W W is coprime to W W W and Λ w , b ′ ( n ) = φ ( W ) Λ ′ ( W n + b ) / W \Lambda'_{w,b}(n)=\varphi(W)\Lambda'(Wn+b)/W Λ w , b ′ ( n ) = φ ( W ) Λ ′ ( W n + b ) / W , we have
M L M = W φ ( W ) ∑ 1 ≤ b ≤ W ( b , W ) = 1 ∑ n = 1 N Λ w , b ′ ( n ) a ( W n + b ) + ∑ k = 1 W Λ ′ ( k ) a ( k ) − ∑ k = M + 1 W ( N + 1 ) Λ ′ ( k ) a ( k ) . ML_M=\frac{W}{\varphi(W)}\sum_{\substack{1\le b\le W\\(b,W)=1}}\sum_{n=1}^{N}\Lambda'_{w,b}(n)a(Wn+b)
+\sum_{k=1}^{W}\Lambda'(k)a(k)-\sum_{k=M+1}^{W(N+1)}\Lambda'(k)a(k). M L M = φ ( W ) W 1 ≤ b ≤ W ( b , W ) = 1 ∑ n = 1 ∑ N Λ w , b ′ ( n ) a ( W n + b ) + k = 1 ∑ W Λ ′ ( k ) a ( k ) − k = M + 1 ∑ W ( N + 1 ) Λ ′ ( k ) a ( k ) .
Since W N ≤ M WN\le M W N ≤ M and ∥ a ( k ) ∥ 2 ≤ 1 \|a(k)\|_2\le1 ∥ a ( k ) ∥ 2 ≤ 1 , we obtain
∥ L M ∥ 2 ≤ max 1 ≤ b ≤ W ( b , W ) = 1 ∥ 1 N ∑ n = 1 N Λ w , b ′ ( n ) a ( W n + b ) ∥ 2 + 2 W log ( M + W ) M . \|L_M\|_2\le\max_{\substack{1\le b\le W\\(b,W)=1}}\left\|\frac{1}{N}\sum_{n=1}^{N}\Lambda'_{w,b}(n)a(Wn+b)\right\|_2+\frac{2W\log(M+W)}{M}. ∥ L M ∥ 2 ≤ 1 ≤ b ≤ W ( b , W ) = 1 max N 1 n = 1 ∑ N Λ w , b ′ ( n ) a ( W n + b ) 2 + M 2 W log ( M + W ) .
For fixed w w w , the last term tends to zero as M → ∞ M\to\infty M → ∞ . Taking the upper limit in M M M and then letting w → ∞ w\to\infty w → ∞ gives ∥ L M ∥ 2 → 0 \|L_M\|_2\to0 ∥ L M ∥ 2 → 0 .
Finally, by the prime number theorem with θ ( N ) = ∑ p ≤ N log p \theta(N)=\sum_{p\le N}\log p θ ( N ) = ∑ p ≤ N log p , ∥ a ( n ) ∥ 2 ≤ 1 \|a(n)\|_2\le1 ∥ a ( n ) ∥ 2 ≤ 1 gives
∥ 1 π ( N ) ∑ p ≤ N a ( p ) − L N ∥ 2 ≤ ∣ 1 π ( N ) − log N N ∣ ∑ p ≤ N ∥ a ( p ) ∥ 2 + 1 N ∑ p ≤ N ( log N − log p ) ∥ a ( p ) ∥ 2 ≤ ∣ 1 π ( N ) − log N N ∣ π ( N ) + 1 N ∑ p ≤ N ( log N − log p ) = ∣ 1 − π ( N ) log N N ∣ + π ( N ) log N − θ ( N ) N ⟶ 0. \begin{aligned}
\left\|\frac{1}{\pi(N)}\sum_{p\le N}a(p)-L_N\right\|_2
\le\left|\frac{1}{\pi(N)}-\frac{\log N}{N}\right|\sum_{p\le N}\|a(p)\|_2+\frac{1}{N}\sum_{p\le N}(\log N-\log p)\|a(p)\|_2 \\
&\le\left|\frac{1}{\pi(N)}-\frac{\log N}{N}\right|\pi(N)+\frac{1}{N}\sum_{p\le N}(\log N-\log p) \\
&=\left|1-\frac{\pi(N)\log N}{N}\right|+\frac{\pi(N)\log N-\theta(N)}{N}\longrightarrow0.
\end{aligned} π ( N ) 1 p ≤ N ∑ a ( p ) − L N 2 ≤ π ( N ) 1 − N log N p ≤ N ∑ ∥ a ( p ) ∥ 2 + N 1 p ≤ N ∑ ( log N − log p ) ∥ a ( p ) ∥ 2 ≤ π ( N ) 1 − N log N π ( N ) + N 1 p ≤ N ∑ ( log N − log p ) = 1 − N π ( N ) log N + N π ( N ) log N − θ ( N ) ⟶ 0.
This proves the lemma. □
Prime assertion of Theorem A. Let s = s ( D , ℓ ) s=s(D,\ell) s = s ( D , ℓ ) be as in the integer assertion. Suppose E ( f j ∣ W s ( T j ) ) = 0 \mathbb{E}(f_j\mid W_s(T_j))=0 E ( f j ∣ W s ( T j )) = 0 for some j j j . Fix m ≥ 1 m\ge1 m ≥ 1 and 1 ≤ b ≤ m 1\le b\le m 1 ≤ b ≤ m , and put
p ~ i ( t ) = p i ( m t + b ) − p i ( b ) , f ~ i = T i p i ( b ) f i . \widetilde p_i(t)=p_i(mt+b)-p_i(b),\qquad\widetilde f_i=T_i^{p_i(b)}f_i. p i ( t ) = p i ( m t + b ) − p i ( b ) , f i = T i p i ( b ) f i .
The polynomials p ~ i \widetilde{p}_i p i have zero constant terms and degrees at most D D D . Moreover, for every nonempty J ⊆ [ ℓ ] J \subseteq[\ell] J ⊆ [ ℓ ] , r ~ p ( J ) = r p ( J ) \widetilde{r}_p(J)=r_p(J) r p ( J ) = r p ( J ) , ∣ J ∣ < 2 r ~ p ( J ) |J|<2\widetilde{r}_p(J) ∣ J ∣ < 2 r p ( J ) .
Since W s ( T ~ j ) W_s(\widetilde{T}_j) W s ( T j ) is T ~ j \widetilde{T}_j T j -invariant, E ( f j ∣ W s ( T ~ j ) ) = 0 \mathbb{E}(f_j\mid W_s(\widetilde{T}_j))=0 E ( f j ∣ W s ( T j )) = 0 . The integer assertion therefore gives
1 N ∑ n = 1 N ∏ i = 1 ℓ T i p i ( m n + b ) f i ⟶ 0 in L 2 ( μ ) . \frac{1}{N}\sum_{n=1}^{N}\prod_{i=1}^{\ell}T_i^{p_i(mn+b)}f_i\longrightarrow0\quad\text{in }L^2(\mu). N 1 n = 1 ∑ N i = 1 ∏ ℓ T i p i ( mn + b ) f i ⟶ 0 in L 2 ( μ ) .
As this holds for every m m m and b b b , Lemma 6.2 gives the prime zero-limit assertion. Finally, let g i = E ( f i ∣ W s ( T i ) ) g_i=\mathbb{E}(f_i\mid W_s(T_i)) g i = E ( f i ∣ W s ( T i )) . Applying the prime zero-limit assertion to f i − g i f_i-g_i f i − g i in each coordinate shows that
∥ A N , P p ( f 1 , … , f ℓ ) − A N , P p ( g 1 , … , g ℓ ) ∥ 2 ⟶ 0. \left\|A_{N,\mathbb{P}}^p(f_1,\ldots,f_\ell)-A_{N,\mathbb{P}}^p(g_1,\ldots,g_\ell)\right\|_2\longrightarrow0. A N , P p ( f 1 , … , f ℓ ) − A N , P p ( g 1 , … , g ℓ ) 2 ⟶ 0.
This proves prime characteristicity. □
Now we show the sharp orders over the primes.
Prime assertions and optimality in Theorem B and Corollary 1.2. Put ( d 1 , d 2 , d 3 ) = ( 4 , 2 , 2 ) (d_1,d_2,d_3)=(4,2,2) ( d 1 , d 2 , d 3 ) = ( 4 , 2 , 2 ) and a ( n ) = T 1 n f 1 T 2 n 2 f 2 T 3 n 2 + n f 3 a(n)=T_1^nf_1T_2^{n^2}f_2T_3^{n^2+n}f_3 a ( n ) = T 1 n f 1 T 2 n 2 f 2 T 3 n 2 + n f 3 . Suppose E ( f j ∣ W d j ( T j ) ) = 0 \mathbb{E}(f_j\mid W_{d_j}(T_j))=0 E ( f j ∣ W d j ( T j )) = 0 for some j j j . To apply Lemma 6.2, fix m ≥ 1 m\geq1 m ≥ 1 and 1 ≤ b ≤ m 1\leq b\leq m 1 ≤ b ≤ m . Give E = Z / m Z E=\mathbb{Z}/m\mathbb{Z} E = Z / m Z uniform probability m E m_E m E , let R z = z + 1 Rz=z+1 R z = z + 1 , and define T ~ 1 = T 1 × R \widetilde{T}_1=T_1\times R T 1 = T 1 × R , T ~ 2 = T 2 × id E \widetilde{T}_2=T_2\times\operatorname{id}_E T 2 = T 2 × id E and T ~ 3 = T 3 × id E \widetilde{T}_3=T_3\times\operatorname{id}_E T 3 = T 3 × id E . These transformations commute. For z ∈ Z / m Z z\in\mathbb{Z}/m\mathbb{Z} z ∈ Z / m Z , let F 1 ( x , z ) = f 1 ( x ) 1 { b + m Z } ( z ) F_1(x,z)=f_1(x)1_{\{b+m\mathbb{Z}\}}(z) F 1 ( x , z ) = f 1 ( x ) 1 { b + m Z } ( z ) , and F i ( x , z ) = f i ( x ) F_i(x,z)=f_i(x) F i ( x , z ) = f i ( x ) for i = 2 , 3 i=2,3 i = 2 , 3 .
We first check that E ( F j ∣ W d j ( T ~ j ) ) = 0 \mathbb{E}(F_j\mid W_{d_j}(\widetilde{T}_j))=0 E ( F j ∣ W d j ( T j )) = 0 . For Φ ∈ P d j ( T ~ j ) \Phi\in\mathcal{P}_{d_j}(\widetilde{T}_j) Φ ∈ P d j ( T j ) , choose q ≥ 1 q\geq1 q ≥ 1 with Δ T ~ j d j + 1 Φ = 1 \Delta_{\widetilde{T}_j}^{d_j+1}\Phi=1 Δ T j d j + 1 Φ = 1 . By Lemma 2.2(i), this identity also holds with q m qm q m in place of q q q . Since T ~ j q m = T j q m × id E \widetilde{T}_j^{qm}=T_j^{qm}\times\operatorname{id}_E T j q m = T j q m × id E , each section Φ ( ⋅ , z ) \Phi(\cdot,z) Φ ( ⋅ , z ) belongs to P d j ( T j ) \mathcal{P}_{d_j}(T_j) P d j ( T j ) . Since this holds for every phase generator Φ \Phi Φ , Lemma 2.4(iii) gives E ( F j ∣ W d j ( T ~ j ) ) = 0 \mathbb{E}(F_j\mid W_{d_j}(\widetilde{T}_j))=0 E ( F j ∣ W d j ( T j )) = 0 .
The integer assertion of Theorem B now gives
B N ( x , z ) : = 1 N ∑ n = 1 N 1 { z + n = b } a ( n ) ( x ) ⟶ 0 in L 2 ( μ ⊗ m E ) . B_N(x,z):=\frac{1}{N}\sum_{n=1}^{N}1_{\{z+n=b\}}a(n)(x)\longrightarrow0\quad\text{in }L^2(\mu\otimes m_E). B N ( x , z ) := N 1 n = 1 ∑ N 1 { z + n = b } a ( n ) ( x ) ⟶ 0 in L 2 ( μ ⊗ m E ) .
Since E = Z / m Z E=\mathbb{Z}/m\mathbb{Z} E = Z / m Z has uniform measure,
∥ B N ( ⋅ , 0 ) ∥ L 2 ( μ ) ≤ m ∥ B N ∥ L 2 ( μ ⊗ m E ) ⟶ 0. \|B_N(\cdot,0)\|_{L^2(\mu)}\leq\sqrt{m}\|B_N\|_{L^2(\mu\otimes m_E)}\longrightarrow0. ∥ B N ( ⋅ , 0 ) ∥ L 2 ( μ ) ≤ m ∥ B N ∥ L 2 ( μ ⊗ m E ) ⟶ 0.
Moreover,
1 N ∑ n = 1 N a ( m n + b ) = m B m N ( ⋅ , 0 ) + a ( m N + b ) − a ( b ) N . \frac{1}{N}\sum_{n=1}^{N}a(mn+b)=mB_{mN}(\cdot,0)+\frac{a(mN+b)-a(b)}{N}. N 1 n = 1 ∑ N a ( mn + b ) = m B m N ( ⋅ , 0 ) + N a ( m N + b ) − a ( b ) .
Since ∥ a ( n ) ∥ 2 ≤ ∏ i = 1 3 ∥ f i ∥ ∞ \|a(n)\|_2\leq\prod_{i=1}^{3}\|f_i\|_\infty ∥ a ( n ) ∥ 2 ≤ ∏ i = 1 3 ∥ f i ∥ ∞ ,
∥ 1 N ∑ n = 1 N a ( m n + b ) ∥ 2 ≤ m 3 / 2 ∥ B m N ∥ L 2 ( μ ⊗ m E ) + 2 N ∏ i = 1 3 ∥ f i ∥ ∞ ⟶ 0. \left\|\frac{1}{N}\sum_{n=1}^{N}a(mn+b)\right\|_2
\leq m^{3/2}\|B_{mN}\|_{L^2(\mu\otimes m_E)}+\frac{2}{N}\prod_{i=1}^{3}\|f_i\|_\infty\longrightarrow0. N 1 n = 1 ∑ N a ( mn + b ) 2 ≤ m 3/2 ∥ B m N ∥ L 2 ( μ ⊗ m E ) + N 2 i = 1 ∏ 3 ∥ f i ∥ ∞ ⟶ 0.
As m m m and b b b were arbitrary, Lemma 6.2 gives the prime zero-limit assertion. Put g i = E ( f i ∣ W d i ( T i ) ) g_i=\mathbb{E}(f_i\mid W_{d_i}(T_i)) g i = E ( f i ∣ W d i ( T i )) . Since E ( f i − g i ∣ W d i ( T i ) ) = 0 \mathbb{E}(f_i-g_i\mid W_{d_i}(T_i))=0 E ( f i − g i ∣ W d i ( T i )) = 0 , we have
∥ A N , P ( f 1 , f 2 , f 3 ) − A N , P ( g 1 , g 2 , g 3 ) ∥ 2 ⟶ 0. \|A_{N,\mathbb{P}}(f_1,f_2,f_3)-A_{N,\mathbb{P}}(g_1,g_2,g_3)\|_2\longrightarrow0. ∥ A N , P ( f 1 , f 2 , f 3 ) − A N , P ( g 1 , g 2 , g 3 ) ∥ 2 ⟶ 0.
Thus the phase factors are characteristic. Since W d i ( T i ) ⊆ Z d i ( T i ) W_{d_i}(T_i)\subseteq Z_{d_i}(T_i) W d i ( T i ) ⊆ Z d i ( T i ) , the Host–Kra factors are also characteristic. Moreover, if ∥ f j ∥ U d j + 1 ( T j ) = 0 \|f_j\|_{U^{d_j+1}(T_j)}=0 ∥ f j ∥ U d j + 1 ( T j ) = 0 for some j j j , the prime average tends to zero.
Finally, take the transformations and functions from Proposition 5.1, with T 1 = T T_1=T T 1 = T and T 2 = T 3 = S T_2=T_3=S T 2 = T 3 = S . They satisfy for any n ≥ 1 n\geq1 n ≥ 1 , T n f 1 S n 2 f 2 S n 2 + n f 3 = e ( 2 x 4 − x 3 ) T^nf_1S^{n^2}f_2S^{n^2+n}f_3=e(2x_4-x_3) T n f 1 S n 2 f 2 S n 2 + n f 3 = e ( 2 x 4 − x 3 ) , ∥ f 1 ∥ U 4 ( T ) = 0 \|f_1\|_{U^4(T)}=0 ∥ f 1 ∥ U 4 ( T ) = 0 and ∥ f 2 ∥ U 2 ( S ) = ∥ f 3 ∥ U 2 ( S ) = 0 \|f_2\|_{U^2(S)}=\|f_3\|_{U^2(S)}=0 ∥ f 2 ∥ U 2 ( S ) = ∥ f 3 ∥ U 2 ( S ) = 0 . Hence A N , P ( f 1 , f 2 , f 3 ) = e ( 2 x 4 − x 3 ) A_{N,\mathbb{P}}(f_1,f_2,f_3)=e(2x_4-x_3) A N , P ( f 1 , f 2 , f 3 ) = e ( 2 x 4 − x 3 ) and then ∥ A N , P ( f 1 , f 2 , f 3 ) ∥ 2 = 1 \|A_{N,\mathbb{P}}(f_1,f_2,f_3)\|_2=1 ∥ A N , P ( f 1 , f 2 , f 3 ) ∥ 2 = 1 for every N ≥ 2 N\geq2 N ≥ 2 . Thus all three Host–Kra orders are optimal. Since W d ⊆ Z d W_d\subseteq Z_d W d ⊆ Z d , the phase-factor orders are also optimal. □
Corollary 6.3. Let T 1 , … , T ℓ T_1,\ldots,T_\ell T 1 , … , T ℓ be commuting invertible measure-preserving transformations of a standard probability space ( X , X , μ ) (X,\mathcal{X},\mu) ( X , X , μ ) , where ℓ ≥ 3 \ell\ge3 ℓ ≥ 3 . Let p 1 , … , p ℓ ∈ Z [ t ] p_1,\ldots,p_\ell\in\mathbb{Z}[t] p 1 , … , p ℓ ∈ Z [ t ] have zero constant terms, degrees at most D D D , satisfy (4) , and d i = s ( D , ℓ ) d_i=s(D,\ell) d i = s ( D , ℓ ) for every i i i . Let A ∈ X A\in\mathcal{X} A ∈ X with μ ( A ) = δ > 0 \mu(A)=\delta>0 μ ( A ) = δ > 0 and E ( 1 A ∣ W d i ( T i ) ) = δ \mathbb{E}(\mathbf{1}_A\mid\mathcal{W}_{d_i}(T_i))=\delta E ( 1 A ∣ W d i ( T i )) = δ for all i ∈ [ ℓ ] i\in[\ell] i ∈ [ ℓ ] . Then
1 N ∑ n = 1 N μ ( A ∩ ⋂ i = 1 ℓ T i − p i ( n ) A ) ⟶ δ ℓ + 1 , 1 π ( N ) ∑ p ≤ N p ∈ P μ ( A ∩ ⋂ i = 1 ℓ T i − p i ( p ) A ) ⟶ δ ℓ + 1 . \frac{1}{N}\sum_{n=1}^{N}\mu\left(A\cap\bigcap_{i=1}^{\ell}T_i^{-p_i(n)}A\right)\longrightarrow\delta^{\ell+1},
\qquad
\frac{1}{\pi(N)}\sum_{\substack{p\leq N\\p\in\mathbb{P}}}\mu\left(A\cap\bigcap_{i=1}^{\ell}T_i^{-p_i(p)}A\right)\longrightarrow\delta^{\ell+1}. N 1 n = 1 ∑ N μ ( A ∩ i = 1 ⋂ ℓ T i − p i ( n ) A ) ⟶ δ ℓ + 1 , π ( N ) 1 p ≤ N p ∈ P ∑ μ ( A ∩ i = 1 ⋂ ℓ T i − p i ( p ) A ) ⟶ δ ℓ + 1 . Moreover, for every 0 < ε < δ ℓ + 1 0<\varepsilon<\delta^{\ell+1} 0 < ε < δ ℓ + 1 , the set E ε = { n ≥ 1 : μ ( A ∩ ⋂ i = 1 ℓ T i − p i ( n ) A ) > δ ℓ + 1 − ε } E_\varepsilon=\{n\geq1:\mu(A\cap\bigcap_{i=1}^{\ell}T_i^{-p_i(n)}A)>\delta^{\ell+1}-\varepsilon\} E ε = { n ≥ 1 : μ ( A ∩ ⋂ i = 1 ℓ T i − p i ( n ) A ) > δ ℓ + 1 − ε } satisfies
lim inf N → ∞ ∣ E ε ∩ [ 1 , N ] ∣ N ≥ ε δ − δ ℓ + 1 + ε , lim inf N → ∞ ∣ E ε ∩ P ∩ [ 1 , N ] ∣ π ( N ) ≥ ε δ − δ ℓ + 1 + ε . \liminf_{N\to\infty}\frac{|E_\varepsilon\cap[1,N]|}{N}\geq\frac{\varepsilon}{\delta-\delta^{\ell+1}+\varepsilon},
\qquad
\liminf_{N\to\infty}\frac{|E_\varepsilon\cap\mathbb{P}\cap[1,N]|}{\pi(N)}\geq\frac{\varepsilon}{\delta-\delta^{\ell+1}+\varepsilon}. N → ∞ lim inf N ∣ E ε ∩ [ 1 , N ] ∣ ≥ δ − δ ℓ + 1 + ε ε , N → ∞ lim inf π ( N ) ∣ E ε ∩ P ∩ [ 1 , N ] ∣ ≥ δ − δ ℓ + 1 + ε ε .
Proof. For f i = 1 A f_i=\mathbf{1}_A f i = 1 A (i ∈ [ ℓ ] i\in[\ell] i ∈ [ ℓ ] ), by Theorem A, both averages tend to δ ℓ \delta^\ell δ ℓ in L 2 L^2 L 2 . Multiply by 1 A \mathbf{1}_A 1 A and integrate to obtain the limit δ ℓ μ ( A ) = δ ℓ + 1 \delta^\ell\mu(A)=\delta^{\ell+1} δ ℓ μ ( A ) = δ ℓ + 1 .
Let t = δ ℓ + 1 − ε t=\delta^{\ell+1}-\varepsilon t = δ ℓ + 1 − ε and F N = [ 1 , N ] F_N=[1,N] F N = [ 1 , N ] or F N = P ∩ [ 1 , N ] F_N=\mathbb{P}\cap[1,N] F N = P ∩ [ 1 , N ] . Since 0 ≤ μ ( A ∩ ⋂ i = 1 ℓ T i − p i ( n ) A ) ≤ δ 0\leq\mu(A\cap\bigcap_{i=1}^{\ell}T_i^{-p_i(n)}A)\leq\delta 0 ≤ μ ( A ∩ ⋂ i = 1 ℓ T i − p i ( n ) A ) ≤ δ and μ ( A ∩ ⋂ i = 1 ℓ T i − p i ( n ) A ) ≤ t \mu(A\cap\bigcap_{i=1}^{\ell}T_i^{-p_i(n)}A)\leq t μ ( A ∩ ⋂ i = 1 ℓ T i − p i ( n ) A ) ≤ t for all n ∉ E ε n\notin E_\varepsilon n ∈ / E ε ,
1 ∣ F N ∣ ∑ n ∈ F N μ ( A ∩ ⋂ i = 1 ℓ T i − p i ( n ) A ) ≤ t + ( δ − t ) ∣ E ε ∩ F N ∣ ∣ F N ∣ . \frac{1}{|F_N|}\sum_{n\in F_N}\mu\left(A\cap\bigcap_{i=1}^{\ell}T_i^{-p_i(n)}A\right)\leq t+(\delta-t)\frac{|E_\varepsilon\cap F_N|}{|F_N|}. ∣ F N ∣ 1 n ∈ F N ∑ μ ( A ∩ i = 1 ⋂ ℓ T i − p i ( n ) A ) ≤ t + ( δ − t ) ∣ F N ∣ ∣ E ε ∩ F N ∣ .
The left-hand side tends to δ ℓ + 1 \delta^{\ell+1} δ ℓ + 1 . Hence
lim inf N → ∞ ∣ E ε ∩ F N ∣ ∣ F N ∣ ≥ δ ℓ + 1 − t δ − t = ε δ − δ ℓ + 1 + ε . \liminf_{N\to\infty}\frac{|E_\varepsilon\cap F_N|}{|F_N|}
\geq\frac{\delta^{\ell+1}-t}{\delta-t}
=\frac{\varepsilon}{\delta-\delta^{\ell+1}+\varepsilon}. N → ∞ lim inf ∣ F N ∣ ∣ E ε ∩ F N ∣ ≥ δ − t δ ℓ + 1 − t = δ − δ ℓ + 1 + ε ε .
:::
Remark 6.4. For ( p 1 , p 2 , p 3 ) = ( t , t 2 , t 2 + t ) (p_1,p_2,p_3)=(t,t^2,t^2+t) ( p 1 , p 2 , p 3 ) = ( t , t 2 , t 2 + t ) , by Theorem B one can get the same conclusions under the weaker assumptions
E ( 1 A ∣ W d i ( T i ) ) = δ , ( d 1 , d 2 , d 3 ) = ( 4 , 2 , 2 ) . \mathbb{E}(\mathbf{1}_A\mid\mathcal{W}_{d_i}(T_i))=\delta,\qquad(d_1,d_2,d_3)=(4,2,2). E ( 1 A ∣ W d i ( T i )) = δ , ( d 1 , d 2 , d 3 ) = ( 4 , 2 , 2 ) .
Consequences and a boundary example
In this section, we first give bounds for the integer and prime averages and a necessary condition for their nonvanishing. We then give a Heisenberg example where phase factors imply convergence to zero even when the Host–Kra seminorm is nonzero. Finally, we take a example to show that the strict inequality in (4) cannot in general be replaced by a non-strict inequality.
Bounds of polynomial averages
Proposition 7.1. Let T 1 , … , T ℓ T_1,\ldots,T_\ell T 1 , … , T ℓ be commuting invertible measure-preserving transformations, and let p 1 , … , p ℓ ∈ Z [ t ] p_1,\ldots,p_\ell\in\mathbb{Z}[t] p 1 , … , p ℓ ∈ Z [ t ] have zero constant terms and degrees at most D D D . Suppose that (4) holds, and put s = s ( D , ℓ ) s=s(D,\ell) s = s ( D , ℓ ) as in Theorem A. Then for every f 1 , … , f ℓ ∈ L ∞ ( μ ) f_1,\ldots,f_\ell\in L^\infty(\mu) f 1 , … , f ℓ ∈ L ∞ ( μ ) with ∣ f i ∣ ≤ 1 |f_i|\leq1 ∣ f i ∣ ≤ 1 for all i ∈ [ ℓ ] i\in[\ell] i ∈ [ ℓ ] , we have
lim sup N → ∞ ∥ A N P ( f 1 , … , f ℓ ) ∥ 2 ≤ min j ∈ [ ℓ ] ∥ E ( f j ∣ W s ( T j ) ) ∥ 2 , \limsup_{N\to\infty}\left\|A_N^{\mathbb{P}}(f_1,\ldots,f_\ell)\right\|_2\leq\min_{j\in[\ell]}\left\|\mathbb{E}(f_j\mid\mathcal{W}_s(T_j))\right\|_2, N → ∞ lim sup A N P ( f 1 , … , f ℓ ) 2 ≤ j ∈ [ ℓ ] min ∥ E ( f j ∣ W s ( T j )) ∥ 2 ,
lim sup N → ∞ ∥ A N , P ( f 1 , … , f ℓ ) ∥ 2 ≤ min j ∈ [ ℓ ] ∥ E ( f j ∣ W s ( T j ) ) ∥ 2 . \limsup_{N\to\infty}\left\|A_{N,\mathbb{P}}(f_1,\ldots,f_\ell)\right\|_2\leq\min_{j\in[\ell]}\left\|\mathbb{E}(f_j\mid\mathcal{W}_s(T_j))\right\|_2. N → ∞ lim sup ∥ A N , P ( f 1 , … , f ℓ ) ∥ 2 ≤ j ∈ [ ℓ ] min ∥ E ( f j ∣ W s ( T j )) ∥ 2 .
If either average does not tend to zero in L 2 L^2 L 2 , then for every j ∈ [ ℓ ] j\in[\ell] j ∈ [ ℓ ] there exists φ j ∈ P s ( T j ) \varphi_j\in\mathcal{P}_s(T_j) φ j ∈ P s ( T j ) such that
∫ X f j φ j ‾ d μ ≠ 0. \int_X f_j\overline{\varphi_j}\,d\mu\ne0. ∫ X f j φ j d μ = 0.
::: Proof. Put g i = E ( f i ∣ W s ( T i ) ) g_i=\mathbb{E}(f_i\mid\mathcal{W}_s(T_i)) g i = E ( f i ∣ W s ( T i )) . By Theorem A, replacing f i f_i f i by g i g_i g i does not change either average in the L 2 L^2 L 2 limit. Since ∣ g i ∣ ≤ 1 |g_i|\leq1 ∣ g i ∣ ≤ 1 , for every j ∈ [ ℓ ] j\in[\ell] j ∈ [ ℓ ] ,
∥ A N p ( g 1 , … , g ℓ ) ∥ 2 2 ≤ 1 N ∑ n = 1 N ∥ T j p j ( n ) g j ∥ 2 2 = ∥ g j ∥ 2 2 . \left\|A_N^{\mathrm{p}}(g_1,\ldots,g_\ell)\right\|_2^2\leq\frac{1}{N}\sum_{n=1}^{N}\left\|T_j^{p_j(n)}g_j\right\|_2^2=\|g_j\|_2^2. ∥ A N p ( g 1 , … , g ℓ ) ∥ 2 2 ≤ N 1 n = 1 ∑ N T j p j ( n ) g j 2 2 = ∥ g j ∥ 2 2 .
The same estimate holds for the prime average. Taking upper limits gives both inequalities. If either average does not tend to zero, then g j ≠ 0 g_j\neq0 g j = 0 for every j j j . By Lemma 2.4 (iii), there exists some φ j ∈ P s ( T j ) \varphi_j\in\mathcal{P}_s(T_j) φ j ∈ P s ( T j ) satisfies ∫ X f j φ j ‾ d μ ≠ 0 \int_X f_j\overline{\varphi_j}\,d\mu\neq0 ∫ X f j φ j d μ = 0 .
Remark 7.2. For ( p 1 , p 2 , p 3 ) = ( t , t 2 , t 2 + t ) (p_1,p_2,p_3)=(t,t^2,t^2+t) ( p 1 , p 2 , p 3 ) = ( t , t 2 , t 2 + t ) , Theorem B gives the same conclusions with W s ( T j ) \mathcal{W}_s(T_j) W s ( T j ) replaced by W d j ( T j ) \mathcal{W}_{d_j}(T_j) W d j ( T j ) , where ( d 1 , d 2 , d 3 ) = ( 4 , 2 , 2 ) (d_1,d_2,d_3)=(4,2,2) ( d 1 , d 2 , d 3 ) = ( 4 , 2 , 2 ) . In the phase-correlation conclusion, the corresponding identity is Δ T j d j + 1 φ j = 1 \Delta_{T_j}^{d_j+1}\varphi_j=1 Δ T j d j + 1 φ j = 1 .
An example with smaller characteristic factors
Let X = G / Γ X=G/\Gamma X = G /Γ be the standard Heisenberg nilmanifold, where G = R 3 G=\mathbb{R}^3 G = R 3 has multiplication ( x , y , z ) ( x ′ , y ′ , z ′ ) = ( x + x ′ , y + y ′ , z + z ′ + x y ′ ) (x,y,z)(x',y',z')=(x+x',y+y',z+z'+xy') ( x , y , z ) ( x ′ , y ′ , z ′ ) = ( x + x ′ , y + y ′ , z + z ′ + x y ′ ) and Γ = Z 3 \Gamma=\mathbb{Z}^3 Γ = Z 3 . Let T 2 T_2 T 2 be an ergodic left translation and let μ \mu μ be Haar measure. Its Kronecker factor K \mathcal{K} K is the horizontal quotient G / ( [ G , G ] Γ ) G/([G,G]\Gamma) G / ([ G , G ] Γ ) , and Z 2 ( T 2 ) = B X \mathcal{Z}_2(T_2)=B_X Z 2 ( T 2 ) = B X modulo completion. By [10 Propositions 2.1(i), 2.4 and Theorem 2.2] , every T 2 q T_2^q T 2 q is ergodic and has order-two quasi-eigenfunction factor K \mathcal{K} K . Hence
W 2 ( T 2 ) = K ⊊ Z 2 ( T 2 ) = B X . \mathcal{W}_2(T_2)=\mathcal{K}\subsetneq\mathcal{Z}_2(T_2)=B_X. W 2 ( T 2 ) = K ⊊ Z 2 ( T 2 ) = B X .
Choose a nonzero g ∈ L ∞ ( μ ) g\in L^\infty(\mu) g ∈ L ∞ ( μ ) with E ( g ∣ K ) = 0 \mathbb{E}(g\mid\mathcal{K})=0 E ( g ∣ K ) = 0 . Then ∥ g ∥ U 3 ( T 2 ) > 0 \|g\|_{U^3(T_2)}>0 ∥ g ∥ U 3 ( T 2 ) > 0 , since the U 3 ( T 2 ) U^3(T_2) U 3 ( T 2 ) null space is orthogonal to L 2 ( Z 2 ( T 2 ) ) = L 2 ( μ ) L^2(\mathcal{Z}_2(T_2))=L^2(\mu) L 2 ( Z 2 ( T 2 )) = L 2 ( μ ) . However, for any f 1 , f 3 ∈ L ∞ ( μ ) f_1,f_3\in L^\infty(\mu) f 1 , f 3 ∈ L ∞ ( μ ) and any T 1 , T 3 T_1,T_3 T 1 , T 3 commuting with each other and with T 2 T_2 T 2 , by Theorem B , we have
∥ A N ( f 1 , g , f 3 ) ∥ 2 ⟶ 0 , ∥ A N , P ( f 1 , g , f 3 ) ∥ 2 ⟶ 0. \|A_N(f_1,g,f_3)\|_2\longrightarrow0,\qquad\|A_{N,\mathbb{P}}(f_1,g,f_3)\|_2\longrightarrow0. ∥ A N ( f 1 , g , f 3 ) ∥ 2 ⟶ 0 , ∥ A N , P ( f 1 , g , f 3 ) ∥ 2 ⟶ 0.
Hence these averages can tend to zero even if the Host–Kra seminorm is nonzero.
A boundary example for the rank condition
In general, the strict inequality in (4) cannot be replaced by ∣ J ∣ ≤ 2 r p ( J ) |J|\leq2r_p(J) ∣ J ∣ ≤ 2 r p ( J ) . The following example shows that, even when all transformations are equal, no finite-order phase factor need be characteristic.
Proposition 7.3. Let p j ( t ) = t + j t 2 p_j(t)=t+jt^2 p j ( t ) = t + j t 2 for j = 0 , 1 , 2 , 3 j=0,1,2,3 j = 0 , 1 , 2 , 3 . Then there exist a totally ergodic nilrotation T T T on a Heisenberg nilmanifold X X X and functions f 0 , f 1 , f 2 , f 3 ∈ L ∞ ( μ ) f_0,f_1,f_2,f_3\in L^\infty(\mu) f 0 , f 1 , f 2 , f 3 ∈ L ∞ ( μ ) with ∣ f j ∣ ≤ 1 |f_j|\leq1 ∣ f j ∣ ≤ 1 such that, for every s ≥ 1 s\geq1 s ≥ 1 , 0 ≤ j ≤ 3 0\leq j\leq3 0 ≤ j ≤ 3
E ( f j ∣ W s ( T ) ) = 0 , \mathbb{E}(f_j\mid\mathcal{W}_s(T))=0, E ( f j ∣ W s ( T )) = 0 , whereas
∥ 1 N ∑ n = 1 N ∏ j = 0 3 T p j ( n ) f j − 1 12 ∥ 2 ⟶ 0. \left\|\frac{1}{N}\sum_{n=1}^{N}\prod_{j=0}^{3}T^{p_j(n)}f_j-\frac{1}{12}\right\|_2\longrightarrow0. N 1 n = 1 ∑ N j = 0 ∏ 3 T p j ( n ) f j − 12 1 2 ⟶ 0. Moreover, for every nonempty J ⊆ { 0 , 1 , 2 , 3 } J\subseteq\{0,1,2,3\} J ⊆ { 0 , 1 , 2 , 3 } ,
r p ( J ) = min { ∣ J ∣ , 2 } , ∣ J ∣ ≤ 2 r p ( J ) , r_p(J)=\min\{|J|,2\},\qquad|J|\leq2r_p(J), r p ( J ) = min { ∣ J ∣ , 2 } , ∣ J ∣ ≤ 2 r p ( J ) , with equality in the second inequality only when J = { 0 , 1 , 2 , 3 } J=\{0,1,2,3\} J = { 0 , 1 , 2 , 3 } .
Proof. Take the nilrotation T : u ↦ a u T:u\mapsto au T : u ↦ a u on the Heisenberg nilmanifold from Subsection 7.2, where a = ( α , β , γ ) a=(\alpha,\beta,\gamma) a = ( α , β , γ ) . Then T T T is totally ergodic, β \beta β is irrational, and W 2 ( T ) \mathcal{W}_2(T) W 2 ( T ) is equal to its Kronecker factor K \mathcal{K} K . Since the eigenfunctions of T T T generate K \mathcal{K} K , K ⊆ W 1 ( T ) ⊆ W 2 ( T ) = K \mathcal{K}\subseteq\mathcal{W}_1(T)\subseteq\mathcal{W}_2(T)=\mathcal{K} K ⊆ W 1 ( T ) ⊆ W 2 ( T ) = K . Thus W 1 ( T ) = W 2 ( T ) = K \mathcal{W}_1(T)=\mathcal{W}_2(T)=\mathcal{K} W 1 ( T ) = W 2 ( T ) = K , and Lemma 2.4 (ii) gives W 1 ( T q ) ⊆ W 2 ( T q ) ⊆ K \mathcal{W}_1(T^q)\subseteq\mathcal{W}_2(T^q)\subseteq\mathcal{K} W 1 ( T q ) ⊆ W 2 ( T q ) ⊆ K for all q ≥ 1 q\geq1 q ≥ 1 .
We first show this equality holds for all finite order. Fix R = T q R=T^q R = T q and ∣ ϕ ∣ = 1 |\phi|=1 ∣ ϕ ∣ = 1 with Δ R s + 1 ϕ = 1 \Delta_R^{s+1}\phi=1 Δ R s + 1 ϕ = 1 . Let r ≥ 0 r\geq0 r ≥ 0 be least with Δ R r + 1 ϕ = 1 \Delta_R^{r+1}\phi=1 Δ R r + 1 ϕ = 1 . If r ≥ 2 r\geq2 r ≥ 2 , then ψ = Δ R r − 2 ϕ \psi=\Delta_R^{r-2}\phi ψ = Δ R r − 2 ϕ belongs to P 2 ( R ) ⊆ L ∞ ( K ) \mathcal{P}_2(R)\subseteq L^\infty(\mathcal{K}) P 2 ( R ) ⊆ L ∞ ( K ) , and by ergodicity, λ = Δ R 2 ψ \lambda=\Delta_R^2\psi λ = Δ R 2 ψ is constant, and the minimality of r r r gives λ ≠ 1 \lambda\neq1 λ = 1 . Lemma 4.2 , with d = 1 d=1 d = 1 and L = 3 L=3 L = 3 , gives ψ = E ( ψ ∣ W 1 ( R ) ) = 1 { λ ∈ μ ∞ } ψ \psi=\mathbb{E}(\psi\mid\mathcal{W}_1(R))=\mathbf{1}_{\{\lambda\in\mu_\infty\}}\psi ψ = E ( ψ ∣ W 1 ( R )) = 1 { λ ∈ μ ∞ } ψ . Thus λ \lambda λ is a root of unity. But R ( Δ R ψ ) = λ Δ R ψ R(\Delta_R\psi)=\lambda\Delta_R\psi R ( Δ R ψ ) = λ Δ R ψ , contradicting total ergodicity. Hence r ≤ 1 r\le1 r ≤ 1 , so ϕ ∈ P 1 ( R ) \phi\in P_1(R) ϕ ∈ P 1 ( R ) . This proves W s ( T ) = K W_s(T)=\mathcal{K} W s ( T ) = K for all s ≥ 1 s\ge1 s ≥ 1 .
Use the fundamental domain [ 0 , 1 ) 3 [0,1)^3 [ 0 , 1 ) 3 , with coordinates ( x , y , z ) (x,y,z) ( x , y , z ) and Lebesgue measure. Put I = [ 0 , 1 / 2 ) I=[0,1/2) I = [ 0 , 1/2 ) and
( c 0 , c 1 , c 2 , c 3 ) = ( − 1 , 3 , − 3 , 1 ) , f j ( x , y , z ) = 1 I ( y ) e ( c j z ) . (c_0,c_1,c_2,c_3)=(-1,3,-3,1),\qquad f_j(x,y,z)=1_I(y)e(c_jz). ( c 0 , c 1 , c 2 , c 3 ) = ( − 1 , 3 , − 3 , 1 ) , f j ( x , y , z ) = 1 I ( y ) e ( c j z ) .
Since K \mathcal{K} K is the horizontal factor and c j ≠ 0 c_j\ne0 c j = 0 for every j j j , we have
E ( f j ∣ K ) ( x , y ) = 1 I ( y ) ∫ 0 1 e ( c j z ) d z = 0. \mathbb{E}(f_j\mid\mathcal{K})(x,y)=1_I(y)\int_0^1 e(c_jz)\,dz=0. E ( f j ∣ K ) ( x , y ) = 1 I ( y ) ∫ 0 1 e ( c j z ) d z = 0.
Denote C N = 1 N ∑ n = 1 N ∏ j = 0 3 T n + j n 2 f j C_N=\frac{1}{N}\sum_{n=1}^N\prod_{j=0}^3T^{n+jn^2}f_j C N = N 1 ∑ n = 1 N ∏ j = 0 3 T n + j n 2 f j . We shall prove that ∥ C N − 1 12 ∥ 2 → 0 \left\|C_N-\frac{1}{12}\right\|_2\to0 C N − 12 1 2 → 0 .
Fix n ≥ 1 n\ge1 n ≥ 1 and put t j = n + j n 2 t_j=n+jn^2 t j = n + j n 2 for 0 ≤ j ≤ 3 0\le j\le3 0 ≤ j ≤ 3 . Recall that for any t ∈ Z t\in\mathbb{Z} t ∈ Z
a t ( x , y , z ) = ( x + t α , y + t β , z + t γ + ( t 2 ) α β + t α y ) . a^t(x,y,z)=\left(x+t\alpha,\ y+t\beta,\ z+t\gamma+\binom{t}{2}\alpha\beta+t\alpha y\right). a t ( x , y , z ) = ( x + t α , y + tβ , z + t γ + ( 2 t ) α β + t α y ) .
Put Y j = y + β t j Y_j=y+\beta t_j Y j = y + β t j and k j = ⌊ Y j ⌋ k_j=\lfloor Y_j\rfloor k j = ⌊ Y j ⌋ . After reducing a t j ( x , y , z ) a^{t_j}(x,y,z) a t j ( x , y , z ) modulo Γ \Gamma Γ , its second coordinate is { Y j } \{Y_j\} { Y j } , and its third coordinate is q j : = z + γ t j + ( t j 2 ) α β + α t j y − ( x + α t j ) k j q_j:=z+\gamma t_j+\binom{t_j}{2}\alpha\beta+\alpha t_jy-(x+\alpha t_j)k_j q j := z + γ t j + ( 2 t j ) α β + α t j y − ( x + α t j ) k j modulo one.
If some { Y j } ∉ I \{Y_j\}\notin I { Y j } ∈ / I , then ∏ j = 0 3 T t j f j ( x , y , z ) = ∏ j = 0 3 1 I ( { Y j } ) ⋅ e ( ∑ j = 0 3 c j q j ) = 0 \prod_{j=0}^3T^{t_j}f_j(x,y,z)=\prod_{j=0}^3 1_I(\{Y_j\})\cdot e\left(\sum_{j=0}^3c_jq_j\right)=0 ∏ j = 0 3 T t j f j ( x , y , z ) = ∏ j = 0 3 1 I ({ Y j }) ⋅ e ( ∑ j = 0 3 c j q j ) = 0 . Now we consider the case of all { Y j } \{Y_j\} { Y j } belong to [ 0 , 1 / 2 ) [0,1/2) [ 0 , 1/2 ) . Since Y j Y_j Y j has degree one in j j j , Y j + 2 − 2 Y j + 1 + Y j = 0 Y_{j+2}-2Y_{j+1}+Y_j=0 Y j + 2 − 2 Y j + 1 + Y j = 0 . Then we have k j + 2 − 2 k j + 1 + k j = − ( { Y j + 2 } − 2 { Y j + 1 } + { Y j } ) = 0 k_{j+2}-2k_{j+1}+k_j=-\left(\{Y_{j+2}\}-2\{Y_{j+1}\}+\{Y_j\}\right)=0 k j + 2 − 2 k j + 1 + k j = − ( { Y j + 2 } − 2 { Y j + 1 } + { Y j } ) = 0 for j = 0 , 1 j=0,1 j = 0 , 1 . Thus k j = k 0 + j ( k 1 − k 0 ) k_j=k_0+j(k_1-k_0) k j = k 0 + j ( k 1 − k 0 ) .
Since t j t_j t j and k j k_j k j have degree at most one in j j j , q j q_j q j has degree at most two in j j j . By
( c 0 , c 1 , c 2 , c 3 ) = ( − 1 , 3 , − 3 , 1 ) , ∑ j = 0 3 c j q j = − q 0 + 3 q 1 − 3 q 2 + q 3 = 0. (c_0,c_1,c_2,c_3)=(-1,3,-3,1),\qquad\sum_{j=0}^3c_jq_j=-q_0+3q_1-3q_2+q_3=0. ( c 0 , c 1 , c 2 , c 3 ) = ( − 1 , 3 , − 3 , 1 ) , j = 0 ∑ 3 c j q j = − q 0 + 3 q 1 − 3 q 2 + q 3 = 0.
Consequently,
C N ( x , y , z ) = 1 N ∑ n = 1 N ∏ j = 0 3 1 I ( { y + n β + j n 2 β } ) = 1 N ∑ n = 1 N F ( y + n β , n 2 β ) , C_N(x,y,z)=\frac{1}{N}\sum_{n=1}^N\prod_{j=0}^3 1_I(\{y+n\beta+jn^2\beta\})=\frac{1}{N}\sum_{n=1}^NF(y+n\beta,n^2\beta), C N ( x , y , z ) = N 1 n = 1 ∑ N j = 0 ∏ 3 1 I ({ y + n β + j n 2 β }) = N 1 n = 1 ∑ N F ( y + n β , n 2 β ) ,
where
F ( s , t ) = ∏ j = 0 3 1 I ( { s + j t } ) . F(s,t)=\prod_{j=0}^3 1_I(\{s+jt\}). F ( s , t ) = j = 0 ∏ 3 1 I ({ s + j t }) .
Take 0 ≤ s < 1 0\le s<1 0 ≤ s < 1 and − 1 / 2 ≤ t < 1 / 2 -1/2\le t<1/2 − 1/2 ≤ t < 1/2 . If F ( s , t ) = 1 F(s,t)=1 F ( s , t ) = 1 , then { s + j t } ∈ [ 0 , 1 / 2 ) \{s+jt\}\in[0,1/2) { s + j t } ∈ [ 0 , 1/2 ) for 0 ≤ j ≤ 3 0\le j\le3 0 ≤ j ≤ 3 . For j = 0 , 1 , 2 j=0,1,2 j = 0 , 1 , 2 ,
⌊ s + ( j + 1 ) t ⌋ − ⌊ s + j t ⌋ = t − { s + ( j + 1 ) t } + { s + j t } ∈ ( − 1 , 1 ) . \lfloor s+(j+1)t\rfloor-\lfloor s+jt\rfloor=t-\{s+(j+1)t\}+\{s+jt\}\in(-1,1). ⌊ s + ( j + 1 ) t ⌋ − ⌊ s + j t ⌋ = t − { s + ( j + 1 ) t } + { s + j t } ∈ ( − 1 , 1 ) .
This difference is an integer, so it is zero. Since ⌊ s ⌋ = 0 \lfloor s\rfloor=0 ⌊ s ⌋ = 0 , we obtain ⌊ s + j t ⌋ = 0 \lfloor s+jt\rfloor=0 ⌊ s + j t ⌋ = 0 for all j j j . Hence F ( s , t ) = 1 F(s,t)=1 F ( s , t ) = 1 when 0 ≤ s + j t < 1 / 2 0\le s+jt<1/2 0 ≤ s + j t < 1/2 for 0 ≤ j ≤ 3 0\le j\le3 0 ≤ j ≤ 3 . Then
F ( s , t ) = 1 ⟺ 0 ≤ s < 1 2 , − s 3 ≤ t < 1 / 2 − s 3 , F(s,t)=1\Longleftrightarrow0\le s<\frac{1}{2},\qquad-\frac{s}{3}\le t<\frac{1/2-s}{3}, F ( s , t ) = 1 ⟺ 0 ≤ s < 2 1 , − 3 s ≤ t < 3 1/2 − s ,
∫ T 2 F ( s , t ) d s d t = ∫ 0 1 / 2 1 6 d s = 1 12 . \int_{\mathbb{T}^2}F(s,t)\,ds\,dt=\int_0^{1/2}\frac{1}{6}\,ds=\frac{1}{12}. ∫ T 2 F ( s , t ) d s d t = ∫ 0 1/2 6 1 d s = 12 1 .
Since β \beta β is irrational, Weyl’s theorem [26 ] gives equidistribution of ( y + n β , n 2 β ) (y+n\beta,n^2\beta) ( y + n β , n 2 β ) in T 2 \mathbb{T}^2 T 2 for every y y y . The function F F F is Riemann integrable, since its discontinuities lie on the lines s + j t ≡ 0 , 1 / 2 ( m o d 1 ) s+jt\equiv0,1/2\pmod1 s + j t ≡ 0 , 1/2 ( mod 1 ) . Thus C N ( x , y , z ) → 1 / 12 C_N(x,y,z)\to1/12 C N ( x , y , z ) → 1/12 . Since 0 ≤ C N ≤ 1 0\le C_N\le1 0 ≤ C N ≤ 1 , bounded convergence gives
∥ C N − 1 12 ∥ 2 → 0. \left\|C_N-\frac{1}{12}\right\|_2\to0. C N − 12 1 2 → 0.
Finally, since p j ( t ) = t + j t 2 p_j(t)=t+jt^2 p j ( t ) = t + j t 2 , any two distinct polynomials p i p_i p i and p j p_j p j are linearly independent over Q \mathbb{Q} Q . Hence r p ( J ) = min { ∣ J ∣ , 2 } r_p(J)=\min\{|J|,2\} r p ( J ) = min { ∣ J ∣ , 2 } , as required.
References
(Z. Liu) School of Mathematics (Zhuhai), Sun Yat-sen University, Zhuhai, Guangdong, 519000, P.R. China
Email address: liuzhw55@mail2.sysu.edu.cn